text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
arXiv:1705.03884v1 [math.GR] 10 May 2017
COPRODUCTS OF FINITE GROUPS
CHRIS HALL
Abstract. We show that for any pair of non-trivial finite groups, their
coproduct in the category of finite groups is not representable.
Given
a category C and a pair of objects X1 , X2 ∈ C, the coproduct
`
X1 X2 is representable in C iff
∃X ∈ C, ∃(ιX1 , ιX2 ) ∈ HomC (X1 , X) × HomC (X2 , X)
such that
∀Y ∈ C, ∀(f1 , f2 ) ∈ HomC (X1 , Y ) × HomC (X2 , Y ), ∃!f ∈ HomC (X, Y )
making the diagram
X1 ❆
⑤⑤
⑤⑤
⑤
~⑤
⑤
∃!f
X `❇❴ ❴ ❴ ❴
❇❇
❇❇
❇
ιX2 ❇❇
ιX 1
X2
❆❆ f
❆❆ 1
❆❆
❆
❴ ❴ ❴/ Y
⑥>
⑥⑥
⑥
⑥⑥
⑥⑥ f2
commute.
Let G, H be finite `
groups. If we regard them as members of the category
of all groups, then G H is representable: it is the free product G ∗ H, and
G → G ∗ H and H → G ∗ H are the canonical inclusions. The purpose of
this note is to prove the following theorem:
`
Theorem 1. If G, H are non-trivial groups, then the coproduct G H in
the category of finite groups is not representable.
The key will be the following proposition:
Proposition 2. Let G, H be non-trivial groups, and let g ∈ G r {1} and
h ∈ H r {1}. For every m ≥ 1, then there exist a finite group Tm and a
homomorphism qm : G ∗ H → Tm such that |h qm (gh) i| > m.
Before proving the proposition we show how it implies Theorem 1.
Proof of Theorem 1. Let F be a finite group and
ιG : G → F and ιH : H → F
We gratefully acknowledge Bob Guralnick for helpful discussions on this topic.
1
2
CHRIS HALL
be homomorphisms. Let
g ∈ G r {1} and h ∈ H r {1},
let m be the order of ιG (g)ιH (h), and let
qm : G ∗ H → Tm
be a homomorphism such that |h qm (gh) i| > m as in Proposition 2. Let
fG : G → Tm and fH : H → Tm
be the respective compositions of
G → G ∗ H and H → G ∗ H
with qm so that
G
⑥ ●●●●
⑥⑥
●●
⑥
●●
⑥⑥ fG
●
⑥
~
⑥
qm #
o
F ❆`
Tm
G; ∗ H
O
✇
❆❆
✇✇
❆❆
✇
✇
f
❆
ιH ❆❆H
✇✇
✇✇
ιG
H
commutes. Observe that
{ f ∈ Hom(F, Tm ) : fG = ιG f and fH = ιH f }
is empty since fG (g)fH (h) has order exceeding m while
(f ιG (g))(f ιH (h)) = f (ιG (g)ιH (h))
has order m. Therefore one cannot
` find a morphism F → Tm to complete
the above diagram, and hence G H is not representable in the category
of finite groups.
The proof of the proposition will occupy the remainder of this note.
Proof of Proposition 2.
To start we recall a special case of a result of Marciniak:
Theorem 3. If there exists a faithful representation G×H → GLn (Q), then
there exists a faithful representation G ∗ H → GLn (K) where K = Q(t).
Proof. See [1].
Recall that a group is residually finite iff every non-identity element is
contained in the complement of a finite-index normal subgroup.
Corollary 4. G ∗ H is residually finite.
Proof. Let Q[G] and Q[H] be the respective group algebras of G and H,
and let n = |G| + |H|. If V = Q[G] ⊕ Q[H], then V ≃ Qn and there exists
a canonical faithful representation G × H → GLn (Q). Let
ρ : G ∗ H → GLn (K)
COPRODUCTS OF FINITE GROUPS
3
be the corresponding representation given by Theorem 3. It is faithful and
the image is finitely generated, so the corollary is a consequence of the fact
that any finitely generated subgroup of GLn (C) is residually finite. Rather
than appeal to this general fact though, we prove directly that G ∗ H is
residually finite.
Let w ∈ G ∗ H be a non-identity element. We must show that there is a
finite-index normal subgroup of G ∗ H whose complement contains w.
The set ρ(G ∪ H) is finite and contained in GL(K), so the least common
multiple of the denominators of all the entries of the matrices in ρ(G ∪ H)
is a non-zero polynomial D ∈ Q[t]. Moreover, the greatest common divisor
of the numerators of all the entries of ρ(w) − In is a non-zero polynomial
N ∈ Q[t].
Let Z ⊂ Q be the finite set of zeros of D · N in Q and U := Q r Z. Let
S be the polynomial ring Q[t][1/DN ] and
(1)
GLn (S) → GLn (K)
be the homomorphism induced by the inclusion S → K.
Observe that the elements of ρ(G ∪ H) all have finite order and generate
ρ(G ∗ H). Therefore the elements of det(ρ(G ∪ H)) are roots of unity and
generate det(ρ(G ∗ H)), so the latter is contained in S × . In particular, ρ
factors through (1).
For each s ∈ U , let Fs be the quotient field S/(t − s)S and ρs be the
composition
G ∗ H → GLn (S) → GLn (Fs )
where the last homomorphism is induced by the quotient S → Fs . Observe that Fs is (canonically) isomorphic to Q, so we can regard ρs as a
homomorphism
ρs : G ∗ H → GLn (Q).
Observe also that the least common multiple of the denominators of the
entries of all matrices in ρs (G ∪ H) is a positive integer Ds ∈ N, and that
det(ρs (G ∗ H)) ⊆ Z[1/Ds ]× . Therefore the image of ρs is contained in the
image of the natural homomorphism
GLn (Z[1/Ds ]) → GLn (Q),
and hence there is a homomorphism
ρs,p : G ∗ H → GLn (Z/p)
for each p ∤ Ds . By construction, ρs,p (w) 6= In and the kernel of ρs,p has finite
index in G ∗ H. In particular, the complement of the latter is a finite-index
subgroup containing w, so G ∗ H is residually finite as claimed.
Observe that gh has infinite order in G ∗ H. Therefore, for each m ≥ 1,
the element wm := (gh)m! is not the identity, hence Corollary 4 implies that
there exists a finite-index normal subgroup Km ⊆ G ∗ H whose complement
contains wm .
4
CHRIS HALL
To complete the proof of the proposition, we let Tm := (G ∗ H)/Km and
qm : G ∗ H → Tm
be the canonical quotient. By definition,
qm (wm! ) = qm (wm ) 6= 1,
hence qm (w) has order exceeding m as claimed.
References
1. Zbigniew S. Marciniak, A note on free products of linear groups, Proc. Amer. Math.
Soc. 94 (1985), no. 1, 46–48. MR 781053
| 4math.GR
|
Preprint, Proceedings of the Second International Conference on Quantum, Nano, and Micro Technologies (ICQNM 2008),
IEEE Computer Society, pp. 66-71, 2008.
A Survey of Quantum Programming Languages:
History, Methods, and Tools
Donald A. Sofge, Member, IEEE
Abstract— Quantum computer programming is emerging as a
new subject domain from multidisciplinary research in quantum
computing, computer science, mathematics (especially quantum
logic, lambda calculi, and linear logic), and engineering attempts
to build the first non-trivial quantum computer. This paper
briefly surveys the history, methods, and proposed tools for
programming quantum computers circa late 2007. It is intended
to provide an extensive but non-exhaustive look at work leading
up to the current state-of-the-art in quantum computer
programming. Further, it is an attempt to analyze the needed
programming tools for quantum programmers, to use this
analysis to predict the direction in which the field is moving, and
to make recommendations for further development of quantum
programming language tools.
Index Terms— quantum computing, functional programming,
imperative programming, linear logic, lambda calculus
I. INTRODUCTION
T
HE importance of quantum computing has increased
significantly in recent years due to the realization that we
are rapidly approaching fundamental limits in shrinking the
size of silicon-based integrated circuits (a trend over the past
several decades successfully described by Moore’s Law).
This means that as we attempt to make integrated circuit
components ever smaller (e.g. below 25nm in feature sizes),
we will increasingly encounter quantum mechanical effects
that interfere with the classical operation of the circuits.
Quantum computing offers a path forward that specifically
takes advantage of quantum mechanical properties, such as
superposition and entanglement, to achieve computational
solutions to certain problems in less time (fewer
computational cycles) than is possible using classical
computing paradigms. Certain problems have been shown to
be solvable exponentially faster on a quantum computer than
has been achieved on a classical computer [1]. Furthermore,
quantum parallelism allows certain functions that have thus far
proven to be computationally intractable using classical
computation to be executed in reasonable time (e.g. factoring
large numbers using Shor’s algorithm). Quantum algorithms
may contain both classical and quantum components (as does
Shor’s algorithm), and can thus leverage the benefits of each.
Manuscript received September 12, 2007. This work was supported by the
Naval Research Laboratory under NRL Work Order N0001406WX30002.
Donald A. Sofge is with the Navy Center for Applied Research in
Artificial Intelligence, Naval Research Laboratory, Washington D.C. (phone:
202-404-4944; fax: 202-767-2166; e-mail: [email protected]).
However, existing classical (non-quantum) programming
languages lack both the data structures and the operators
necessary to easily represent and manipulate quantum data.
Quantum computing possesses certain characteristics that
distinguish it from classical computing such as the
superposition of quantum bits, entanglement, destructive
measurement, and the no-cloning theorem. These differences
must be thoroughly understood and even exploited in the
context of quantum programming if we are to truly realize the
potential of quantum computing. We need native quantum
computer programming languages that embrace the
fundamental aspects of quantum computing, rather than
forcing us to adapt and use classical programming languages
and techniques as ill-fitting stand-ins to develop quantum
computer algorithms and simulations.
Ultimately, a
successful quantum programming language will facilitate
easier coding of new quantum algorithms to perform useful
tasks, allow or provide a capability for simulation of quantum
algorithms, and facilitate the execution of quantum program
code on quantum computer hardware.
II. ORIGINS AND HISTORY OF QUANTUM COMPUTING
Prior surveys of quantum computing and quantum
programming [2], [3] trace the origins of quantum computing
and quantum programming to Feynman’s 1982 proposal for
constructing a quantum computer as a means of simulating
other quantum systems, noting that a quantum computer may
efficiently simulate a quantum system, whereas a classical
computer simulation of a quantum system would require
exponential resources both in memory space and
computational time [4]. Preskill credits Paul Benioff [5] with
making this proposal concomitantly with Feynman. However,
the roots of quantum programming go far deeper than this,
through the field of quantum information theory, in the work
of Birkhoff and von Neumann on quantum logic in 1936 [6].
This work also formed the basis for quantum mechanics as it
is currently practiced today.
Deutsch [7] investigated the computational power of
quantum computers and proposed a quantum version of the
Turing machine. He proposed one of the earliest quantum
algorithms as a means of demonstrating the solution of a
problem that would be difficult to solve using a classical
computer yet would be quite easy to solve using a quantum
algorithm.
Preprint, Proceedings of the Second International Conference on Quantum, Nano, and Micro Technologies (ICQNM 2008),
IEEE Computer Society, pp. 66-71, 2008.
The difficulty of solving certain problems using computers
is often characterized by analogy with hypothetical Turing
machine based solutions [8]. Alan Turing showed that the
capability of any general purpose computer could be simulated
with a Turing machine.
By studying the time (#
computational cycles) and space (memory requirement)
complexity of problems, we can classify problems according
to their Turing complexity (e.g. deterministic polynomial (P),
non-deterministic polynomial (NP), exponential (EXP)). The
Deutsch-Jozsa algorithm [1] was designed in 1992 to
maximally illustrate the computational advantage of quantum
computing over classical computing.
The algorithm
determines a property of a binary input string (either constant
or balanced) in a single step, whereas determining this
property on a classical computer grows exponentially in
computational complexity with the length of the bit string.
Unfortunately, this result has not yet resulted in the creation of
any particularly useful quantum algorithms for typical data
processing needs.
Since that time, substantial effort has been made to
characterize the complexity of problems using quantum
Turing machine models versus classical Turing machine
models. While there is still some question regarding certain
complexity equivalence classes (e.g. does BQP=NP?), there is
widespread agreement amongst researchers in this area that
quantum computing has not yet been proved to generally
move problems from one complexity class under classical
computing to a lower complexity class under quantum
computing [9]. However, this does not mean that nontrivial
quantum computers once built will not allow substantial
computational gains for solving certain problems over
currently known techniques using classical computers. An
important thread in quantum computing research is to define
problem classes and applications that result in significant
speed-ups using quantum instead of classical computers.
The invention of linear logic by Girard in 1987 [10] has also
played a significant role in the formulation of recent quantum
programming languages, specifically those based upon lambda
calculus (described in section III.b). Linear logic differs from
classical logic in that assumptions (states, or inputs) and
hypotheses are considered resources that may be consumed. It
provides a means for resource control. Linear logic differs
from usual logics such as classical or intuitionistic logic where
the governing judgement is of truth, which may be freely used
as many times as necessary.
For example, suppose A represents water, B represents cold
(or a freezing process), and C represents ice. Then A B
C. But after the process is applied, resource A is consumed
and is no longer available. Research in using linear logic is
still quite active, and linear logic is important element to many
current quantum programming language development efforts.
The first practical steps toward formulating a quantum
programming language were made by Knill in 1996 in his
proposal for conventions for a quantum pseudocode [11], and
his description of the quantum random access machine
(QRAM) model of a quantum computer. The QRAM model is
built upon the (probably accurate) assumption that any
practical quantum computer will in fact be a classical machine
with access to quantum computing components, such as qubit
registers. QRAM defines a set of specific operations to be
performed on computer hardware including preparation of
quantum states (from classical states), certain unitary
operations, and measurement. Knill’s quantum pseudocode
provides a syntax for describing qubits, qubit registers, and
operations involving both classical and quantum information.
While extremely useful, Knill’s proposal falls short of
possessing all of the needed characteristics of a real quantum
programming language due to its informal structure, lack of
strong typing, and representation of only some of the quantum
mechanical properties needed.
A variety of tools have been created for simulating quantum
circuits and modest quantum algorithms on classical
computers using well-known languages such as C, C++, Java,
and rapid prototyping languages such as Maple, Mathematica,
and Matlab. A good on-line reference for these simulators is
http://www.quantiki.org/wiki/index.php/List_of_QC_simulators
While simulators may provide an excellent means for
quickly learning some of the basics concepts of quantum
computing, they are not substitutes for actual quantum
programming languages since they are designed to run only on
classical computer architectures, and will not realize any of
the computational advantages of quantum computing.
III. A TAXONOMY OF QUANTUM PROGRAMMING LANGUAGES
Quantum programming languages may be taxonomically
divided into (A) imperative quantum programming languages,
(B) functional quantum programming languages, and (C)
others (may include mathematical formalisms not intended for
computer execution). In addition, Glendinning [12] maintains
an online catalog of quantum programming languages,
simulation systems, and other tools. Finally, a slightly more
dated but at the time quite comprehensive and still useful
survey of quantum computer simulators was provided by Julia
Wallace in 1999 [13].
Early quantum programming language development efforts
focused on exploring the Quantum Turing Machine (QTM)
model as proposed by Deutsch [7]. While interesting and
informative from the standpoint of understanding
computational complexity of problem classes with respect to
quantum computing, it did not result in practical tools for
programming quantum computers.
The quantum circuit model quickly became the driving
force in quantum programming. The Deutsch-Jozsa algorithm,
Quantum Fourier Transform, Shor’s factoring algorithm, and
Grover’s algorithm were all described using the quantum
circuit model [8]. In order to build this into a language
(instead of just designing circuits), Knill [11] proposed a
quantum programming pseudocode that, along with adapted
imperative programming languages such as C and C++,
resulted in the first imperative quantum programming
languages (e.g., QCL). These languages built upon the
QRAM model of quantum computation, assumed classical
flow control with both classical and quantum data, and
allowed interleaved measurements and quantum operations.
Preprint, Proceedings of the Second International Conference on Quantum, Nano, and Micro Technologies (ICQNM 2008),
IEEE Computer Society, pp. 66-71, 2008.
The use of imperative quantum programming languages
gave way to a multitude of functional quantum programming
languages such as QFC, QPL, and QML (described below),
mostly based upon the QRAM model, but also increasingly
utilizing the work in mathematical logic to define better
operational semantics for quantum computing, and also
including the tool of linear logic discussed previously.
communication. Mlnarik provides full formalized syntax for
the language, operational semantics, proves type soundness
(eliminating type errors) for the non-communicating part of
the language, and can be used for proving correctness of
implemented quantum algorithms. The author also provides a
publicly accessible simulator for LanQ at
http://lanq.sorceforge.net/
A. Imperative Programming Languages
Imperative programming languages, also known as procedural
languages, are fundamentally built upon the use of statements
to change the global state of a program or system of variables.
Common classical imperative languages include FORTRAN,
Pascal, C, and Java. These may be contrasted with functional
(or declarative) languages such as Lisp, APL, Haskell, J, and
Scheme, in which computation is based upon the execution of
mathematical functions.
Imperative quantum programming languages today are
largely descendents of Knill’s proposed quantum pseudocode
and the QRAM model of quantum computing. Arguably the
first “real” quantum programming language [2] was QCL,
developed and refined from 1998-2003 by Bernhard Ömer
[14]-[18]. QCL (Quantum Computation Language) utilizes a
syntax derived from C, and also provides a full quantum
simulator for code development and testing on a classical
computing platform. In support of the classical+quantum
model of computation as envisioned with the QRAM virtual
hardware model, QCL also provides a full classical
programming sublanguage. High-level quantum programming
features include automatic memory management, user defined
operators and functions, and computation of the inverse of a
user defined operator. QCL may be downloaded from the web
at http://tph.tuwien.ac.at/~oemer/qcl.html
Betelli et al. [19] devised an imperative language based
upon C++. The language was created in the form of a C++
library, and thus can be compiled. It is also maintained as a
downloadable source code through http://sra.itc.it/people/
serafini/qlang/
Important features of this language include construction and
optimization of quantum operators at run-time, classes for
basic quantum operations like QHadamard, QFourier, QNot,
QSwap, and Qop. The language also supports user definition
and construction of new operators. Another important feature
offered by this language is simulation of the noise parameters
in the simulator.
This language is occasionally called Q Language by the
authors, but another perhaps more prominent functional
language exists with the same name, (for which the Q stands
for eQuation), thus use of this name should probably be
avoided in at least one of the two contexts.
Another imperative quantum programming language, called
qGCL (quantum Guarded Command Language), was proposed
by Sanders & Zuliani [20] as a derivative of Dijkstra’s
guarded command language, intended more for algorithm
derivation and verification rather than programming.
Most recently (October 2007) Mlnarik [21] introduced the
imperative quantum programming language LanQ which uses
C-like syntax and supports both classical and quantum process
operations, including process creation and interprocess
B. Functional Quantum Programming Languages
Functional (or declarative) programming languages do not
rely upon the update of a global system state, but instead
perform mathematical transformations by executing mappings
from inputs to outputs. Most recent developments in quantum
programming have focused on the use of functional rather than
imperative languages. The languages in this category are
based upon the concept of a lambda calculus. Lambda calculi
are constructions from mathematical logic used to investigate
the properties of functions, such as computability, recursion,
and stopping.
Lambda calculi may be considered the smallest universal
programming languages. They consist of a single
transformation rule (variable substitution) and a single
function definition scheme. They are universal in the sense
that any computable function can be expressed and evaluated
using this formalism. It is thus equivalent to the Turing
machine formalism. However, lambda calculi emphasize the
use of transformation rules, and do not care about the actual
machine implementing them.
Lambda calculi were first proposed by Alonzo Church and
Stephen Cole Kleene in the 1930s, and used by Church in
1936 [22], [23] to address the decision problem
(entscheidungs-problem) challenge proposed by David
Hilbert. Lambda calculi can be used to define what a
computable function is. The question of whether two lambda
calculus expressions are equivalent cannot be solved by a
general algorithm. This was the first question, even before the
halting problem, for which undecidability could be proved.
Since their invention, lambda calculi have greatly influenced
classical functional programming languages such as Lisp, ML
and Haskell.
In 1996 Maymin [24] proposed a quantum lambda calculus
to investigate the Turing computability of quantum
algorithms. While Maymin’s lambda calculus was found to
efficiently solve NP-complete problems, it was unfortunately
found to be more expressive than any physically realizable
quantum computer [25].
In 2004 van Tonder [26] defined a quantum lambda
calculus for pure quantum computation (no measurements take
place), analyzed the non-duplicability of quantum states
through use of linear logic, and argued that the language has
the same equivalent computational capabilities as a quantum
Turing machine. Quantum algorithms are implemented in a
quantum simulator built upon the Scheme programming
language.
Also in 2004 Selinger [27] proposed a functional quantum
programming language called QFC (Quantum Flow Charts)
which represents programs via functional flow charts, and an
equivalent form which utilizes textual syntax called QPL
(Quantum Programming Language). These languages rely
Preprint, Proceedings of the Second International Conference on Quantum, Nano, and Micro Technologies (ICQNM 2008),
IEEE Computer Society, pp. 66-71, 2008.
upon the notion of using classical control and quantum data,
and build upon a lambda calculus model to handle both
classical and quantum data within the same formalism. These
languages may be compiled using the QRAM virtual quantum
computer model. However, they still lack many desirable
aspects including higher-order features and side-effects.
In 2004 Danos [28] studied a one-way (non-reversible)
model of quantum computation that included notation for
entanglement, measurement, and local corrections.
In 2005 Perdrix [29] defined a type system that reflects
entanglement of quantum states. This too was based upon a
lambda calculus. Altenkirch and Grattage [30], [31] developed
a functional quantum programming language called QML in
which control as well as data may be quantum. QML is based
upon a linear logic [10] (described previously in Section II),
but focuses on the elimination of weakening (discarding a
quantum state) instead of contraction (duplication of quantum
state). Several researchers have proposed domain-specific
functional quantum programming languages implemented in
Haskell, following in the style of Selinger’s QPL, using linear
logic and lambda calculi, and building upon the 2001 work of
Mu and Bird [32] in which quantum programming is modeled
in Haskell through definition of a data type for quantum
registers. Sabry [33] extended this model to include
representation of entangled states. Other related efforts in this
vein include the work of Vizzotta and da Rocha Costa [34],
Karczmarczuk [35], and Skibinski [36].
C. Other Quantum Programming Language Paradigms
A substantially different approach to quantum programming
was offered by Freedman, Kitaev, and Wong [37] based upon
the simulation of topological quantum field theories (TQFT’s)
by quantum computers. TQFT’s
provide a more robust
model of quantum computation by representing quantum
states as physical systems resistant to perturbations. Quantum
operations are determined by global topological properties,
such as paths that particles follow. This radically different
approach to quantum computing may provide new insights
and lead to the creation of new types of quantum algorithms.
However, as currently formulated it only deals with evolution
of state and does not include a measurement process.
A number of efforts have been made in recent years to
define languages to support quantum cryptographic protocols,
and specifically focus on the inclusion of communication
between quantum processes. Such processes may be local or
nonlocal, thus giving rise to distributed quantum programming
specifications. A prime example of this is Mauerer’s [38]
specification of the cQPL language, based upon Selinger’s
QPL, but with extensions added to support communication
between distributed processes.
Another thread of development has focused on the use of
quantum process algebras, such as QPAlg (quantum process
algebra) by Jorrand and Lalire [39] to describe interactions
between classical and quantum processes. Gay and Nagarajan
[40] define the language CQP (communicating quantum
processes) for modeling systems combining classical and
quantum communication, with particular emphasis on
applications such as quantum cryptography. CQP is designed
specifically to provide complete protocol analysis, prove type
soundness, and lead to methods for formal verification of
systems modeled in the language. Both QPAlg and CQP
influenced the design of the imperative programming language
LanQ described previously.
Adão and Mateus [41] give a process calculus for security
protocols built upon the QRAM computational model with an
added cost model. Udrescu et al. [42] describe a hardware
description language for designing quantum circuits similar to
those used for VLSI design.
The use of Girard’s linear logic and lambda calculus has
spurred a considerable amount of recent work in the
formulation of mathematical formalisms for quantum
computing that include such semantic features as
entanglement, communication, teleportation, partial and mixed
quantum states, and destructive measurement. While exciting
and worthwhile, these formalisms generally fall short of
actually specifying specific programming languages with full
syntax, well defined operators, and simulators and compilers
needed to implement Shor’s or Grover’s algorithms or to
create and test new quantum algorithms. These efforts do help
to address shortcomings in quantum theory itself as discussed
in the next section.
IV. CHALLENGES IN QUANTUM PROGRAMMING LANGUAGE
DEVELOPMENT
The difficulties in formulating useful, effective, and in some
sense universally capable quantum programming languages
arise from several root causes. First, quantum mechanics itself
(and by extension quantum information theory) is incomplete.
Specifically missing is a theory of measurement. Quantum
theory is quite successful in describing the evolution of
quantum states, and even in predicting probabilistic outcomes
after measurements have been made, but the process of state
collapse is (with a few exceptional cases) not covered. So
issues such as decoherence, diffusion, entanglement between
particles (or entangled state, of whatever physical
instantiation), and communication (including teleportation) are
not well defined from a quantum information (and by
extension quantum computation) perspective. Work with
semantic formalisms and linear logic attempt to redress this by
providing a firmer basis in a more complete logic consistent
with quantum mechanics. These logic-based formalisms
(once validated) may then be combined with language syntax
and other programming language features to more accurately
and completely reflect the potential capabilities of quantum
computing.
A second key source of difficulty is the lack of quantum
computing hardware for running quantum algorithms. Given
no specific set of quantum operations (e.g. specific quantum
gates, ways that qubits are placed in superposition or
entangled), then no guidance is available to computer
scientists designing quantum programming languages as to
what data structures should be implemented, what types of
operations to allow, what features to disallow (such as abiding
Preprint, Proceedings of the Second International Conference on Quantum, Nano, and Micro Technologies (ICQNM 2008),
IEEE Computer Society, pp. 66-71, 2008.
by the no-cloning theorem), and how to best design the
language to do what quantum computing does best (e.g., is an
imperative language like C or Java a better place to start, or a
functional language like LISP or APL?). A third source of
difficulty is the paucity of practical applications for quantum
computing. Shor’s algorithm generated tremendous
excitement over the potential of quantum computing in large
part because most educated individuals could immediately
recognize the value of the contribution, at least in the context
of code breaking. Grover’s algorithm also provides a speedup over classical computation for a specific class of problems
(searching unsorted lists), from O(N) to O(N1/2), which is still
quite impressive, especially for large N. Shor’s algorithm is
capable of achieving its impressive speedup by exploiting the
structure of the problem in the quantum domain, recasting the
factoring problem as a period-finding problem which may
then be solved in parallel. Grover’s algorithm is arguably
more general, and variants of Grover’s algorithm have been
proposed for a variety of problems [43], [44].
Finally, there is much on-going research to understand the
quantum physics behind quantum information, and by
extension quantum computing. We have previously discussed
several quantum computing conceptual models including
QRAM, the quantum Turing machine, and the quantum circuit
model. However, recent work by in the area of adiabatic
quantum computing by Dorit Aharonov [45] and weak
measurements by Jeff Tollaksen [46] and Yakir Aharonov
[47] arguably falls under none of these categories, yet these
methods represent potentially exciting new directions for
realizing quantum computing. But this wealth of models also
serves to further complicate the process of establishing a set of
practical tools for writing quantum programs, namely quantum
programming languages.
However, the question of what sorts of quantum programs will
be written and what applications they will address is still open,
and the answer to this question will serve as a major driving
factor in shaping the quantum programming languages to
come. This is supported analogically by comparison with the
numerous application domains for classical computing and the
even larger number of classical programming languages
created to develop products from them in the form of
application programs.
Perhaps the greatest potential for useful computational gains
of quantum computing over classical computing may be found
with the class of problems known hidden subgroup problems
[48], [49]. The factoring problem that Shor’s algorithm
addresses, as reformulated by Shor, falls within this class.
One strategy, then, would be to try to understand which
problems may be mapped into the hidden subgroup problem
class, and then investigate those problems for development of
quantum algorithms to address them.
Finally, the realization of quantum computing hardware will
significantly drive the design of quantum language operators
and data structures for quantum programs to be run on that
hardware. We can already make some guesses about early
hardware (e.g. superposition of a small number of locally
contiguous qubits, preparation of states using Hadamard gates,
rapid decoherence of quantum states). But if qubits are
represented by photons, such guesses may very well be wrong.
When considering the need for quantum algorithms, and
asking what kinds of quantum algorithms are needed, it is
useful to think of a (classical) programmer as possessing a
toolbox filled with a rich diversity of tools (algorithms) from
which he/she can construct solutions (new programs) to client
problems. From this basis we can proceed cautiously to
design small but interesting quantum algorithms, gradually
expanding the universe of basic quantum functions and
capabilities to put into our quantum programming toolbox.
V. A PATH FORWARD FOR QUANTUM PROGRAMMING
LANGUAGE DEVELOPMENT
Julia Wallace [13] in her quite thorough survey of quantum
simulators notes that most of the simulators and even some of
the languages developed could only really implement one
algorithm, usually Shor’s (some could do Grover’s as well,
and some could do more). Selinger [2] comments “each new
algorithm seems to rely on a unique set of ‘tricks’ to achieve
its goal”. The hope, of course, is that one can take a set of
known quantum algorithms, such as Shor’s, Grover’s,
Deutsch-Jozsa, and a few others, and from this set to grow a
methodology for quantum programming, or at least generate a
few new useful quantum algorithms. Unfortunately, this set of
examples is too small and specialized to grow a methodology
for quantum programming, and generating useful new
quantum algorithms has proven quite difficult.
The work in operational semantics, lambda calculi, and
linear logic (described above) and the inclusion of a complete
programming grammar and syntax will undoubtedly yield
multiple powerful languages for developing quantum
algorithms.
These languages will probably be of the
functional class, but they may take other forms as well.
REFERENCES
[1]
D. Deutsch and R. Jozsa, “Rapid solutions of problems by quantum
computation,” in Proceedings of the Royal Society of London, Series A,
vol. 439, pp. 553-558, 1992.
[2]
P. Selinger, “A brief survey of quantum programming languages,” in
Proceedings of the 7th International Symposium on Functional and
Logic Programming, volume 2998 of Lecture Notes in Computer
Science, Springer, 2004.
[3]
S. Gay, “Quantum Programming Languages: Survey and Bibliography,”
in Mathematical Structures in Computer Science, 16(4), Cambridge
University Press, 2006.
[4]
R. Feynman, “Simulating physics with computers,” in International
Journal of Theoretical Physics 21(6–7), pp. 467–488, 1982.
[5]
J. Preskill, “Lecture notes for physics 219: Quantum information and
computation,” 1998. Available from http://www.theory.caltech.edu
/people/preskill/ph219
[6]
G. Birkhoff and J. von Neumann, “The Logic of Quantum Mechanics,”
Annals of Mathematics, Vol. 37, pp. 823-843, 1936.
[7]
D. Deutsch, “Quantum theory, the Church-Turing principle and the
universal quantum computer,” in Proceedings of the Royal Society of
London A 400, pp. 97–117, 1985.
Preprint, Proceedings of the Second International Conference on Quantum, Nano, and Micro Technologies (ICQNM 2008),
IEEE Computer Society, pp. 66-71, 2008.
[8]
M. Nielsen and I. Chuang, Quantum Computation and Quantum
Information, Cambridge University Press, 2000.
[9]
A. Kitaev, A. Shen, and M. Vyalyi, Classical and Quantum
Computation, Number 47 in Graduate Series in Mathematics, AMS,
Providence, RI, 2002.
[10] J. Girard, “Linear logic,” Theoretical Computer Science, 50:1-102,
Elsevier Science Publishers Ltd., 1987.
[11] E. Knill, “Conventions for quantum pseudocode,” Technical Report
LAUR-96-2724, Los Alamos National Laboratory, 1996.
[12] I. Glendinning, “Quantum programming languages and tools,”
http:// www.vcpc.univie.ac.at/ian/hotlist/qc/programming.shtml
[13] J. Wallace, “Quantum Computer Simulation - A Review; ver. 2.1,”
University of Exeter Technical Report, 1999. See also the web site
http://www.dcs.ex.ac.uk/~jwallace/simtable.html, 2002.
[14] B. Ömer, “A Procedural Formalism for Quantum Computing,” Master’s
Thesis, Department of Theoretical Physics, Tech. Univ. of Vienna, 1998.
[15] B. Ömer, “Quantum Programming in QCL,” Master’s Thesis, Institute
of Information Systems, Technical University of Vienna, 2000.
[16] B. Ömer, “Procedural quantum programming,” in Proceedings of the
AIP Conference on Computing Anticipatory Systems, American Institute
of Physics, 2001.
[17] B. Ömer, “Classical concepts in quantum programming,” arXiv:quantph/0211100, 2002.
[30] T. Altenkirch and J. Grattage, “A functional quantum programming
language,” in Proceedings of the 20th Annual IEEE Symposium on
Logic in Computer Science, IEEE Computer Society, 2005.
[31] T. Altenkirch and J. Grattage, “QML: Quantum data and control,”
manuscript, 2005.
[32] S. Mu and R. Bird, “Functional quantum programming,” in Proceedings
of the 2nd Asian Workshop on Programming Languages and Systems,
Korea Advanced Institute of Science and Technology, 2001.
[33] A. Sabry, “Modelling quantum computing in Haskell,” in Proceedings
of the ACM SIGPLAN Workshop on Haskell, ACM Press, 2003.
[34] J. Vizzotto and A. da Rocha Costa, “Concurrent quantum programming
in Haskell,” in VII Congresso Brasileiro de Redes Neurais, Sessão de
Computação Quântica, 2005.
[35] J. Karczmarczuk, “Structure and interpretation of quantum mechanics —
a functional framework,” in Proceedings of the ACM SIGPLAN
Workshop on Haskell, ACM Press, 2003.
[36] J. Skibinski, “Haskell simulator of quantum computer,” 2001.
Available from http://web.archive.org/web/20010520121707/
www.numeric-quest.com/haskell.
[37] M. Freedman, A. Kitaev, and Z. Wong, “Simulation of topological field
theories by quantum computers,” arXiv:quant-ph/0001071/v3, 2000.
[38] W. Mauerer, “Semantics and Simulation of Communication in Quantum
Computing,” Master’s Thesis, University Erlangen-Nuremberg, 2005.
[18] B. Ömer, “Structured Quantum Programming,” Ph.D. thesis, Technical
University of Vienna, 2003.
[39] P. Jorrand and M. Lalire, "From Quantum Physics to Programming
Languages: A Process Algebraic Approach," Unconventional
Programming Paradigms, LNCS Series, Vol. 3566, Springer, 2005.
[19] S. Bettelli, T. Calarco, and L. Serafini, “Toward an architecture for
quantum programming,” in The European Physical Journal D 25, pp.
181–200, EDP Sciences/Springer, 2003.
[40] S. Gay and R. Nagarajan, "Communicating quantum processes," in
Proceedings of the 32nd ACM Annual Symposium on Principles of
Programming Languages, pp. 145-157, ACM Press, 2005.
[20] J. Sanders P. and Zuliani, “Quantum programming,” in Mathematics of
Program Construction, volume 1837, LNCS, Springer, 2000.
[41] P. Adão and P. Mateus, “A process algebra for reasoning about quantum
security,” in Proceedings of the 3rd International Workshop on
Quantum Programming Languages, Electronic Notes in Theoretical
Computer Science, Elsevier Science, 2005.
[21] H. Mlnarik, “Operational Semantics and Type Soundness of Quantum
Programming Language LanQ,” arXiv:quant-ph/0708.0890v1, 2007.
[22] A. Church, “An unsolvable problem of elementary number theory,” in
American Journal of Mathematics, 58, pp. 345–363, 1936.
[23] A. Church, “A note on the Entscheidungsproblem,” in Journal of
Symbolic Logic, 1, pp. 40–41, 1936.
[24] P. Maymin, “Extending the lambda calculus to express randomized and
quantumized algorithms,” arXiv:quant-ph/9612052, 1996.
[42] M. Udrescu, L. Prodan, and M. Vlâdutiu, “Using HDLs for describing
quantum circuits: a framework for efficient quantum algorithm
simulation,” in Proceedings of the 1st ACM Conference on Computing
Frontiers, ACM Press, 2004.
[43] M. Lanzagorta, R. Gomez, and J. Uhlmann, “Quantum Rendering,” in
Proceedings of SPIE Vol. 5105, Quantum Information and Computation,
E. Donkor, A. Pirich, H. Brandt (Eds.), SPIE Press, 2003.
[25] P. Maymin, “The lambda-q calculus can efficiently simulate quantum
computers,” arXiv:quant-ph/9702057, 1997.
[44] M. Lanzagorta and J. Uhlmann, “Quantum Computational Geometry,”
In Proceedings of SPIE Vol. 5436, Quantum Information and
Computation II, E. Donkor, A. Pirich, H. Brandt (Eds.), SPIE, 2004.
[26] A. van Tonder, “A lambda calculus for quantum computation, in SIAM
Journal on Computing 33(5), pp. 1109–1135, 2004. Also arXiv:quantph/0307150.
[45] D. Aharonov, W. van Dam, J. Kempe, Z. Landau, S. Lloyd, and O.
Regev, "Adiabatic Quantum Computation is Equivalent to Standard
Quantum Computation," arXiv:quant-ph/0405098, 2006.
[27] P. Selinger, “Towards a quantum programming language,” in
Mathematical Structures in Computer Science 14(4), pp. 527–586,
Cambridge University Press, 2004.
[46] J. Tollaksen and D. Ghoshal, “NP Problems, Post-selection and Weak
Measurements,” in Proceedings SPIE, Defense and Security Conference,
Florida, SPIE Press, 2006.
[28] V. Danos, E. Kashefi, and P. Panangaden, “Parsimonious and robust
realizations of unitary maps in the one-way model,” in Physical Review
A 72(064301), The American Physical Society, 2005. Also arXiv:quantph/0411071.
[47] Y. Aharonov and D. Rohrlich, Quantum paradoxes: quantum theory for
the perplexed, Wiley-VCH, 2005.
[29] S. Perdrix, “Quantum patterns and types for entanglement and
separability,” in Proceedings of the 3rd International Workshop on
Quantum Programming Languages, Electronic Notes in Theoretical
Computer Science, Elsevier Science, 2005.
[48] A. Kitaev, "Quantum measurements and the Abelian Stabilizer
Problem," quant-ph/9511026, 1995.
[49] S. Lomonaco (Jr.) and L. H. Kauffman, “Quantum Hidden Subgroup
Algorithms: A Mathematical Perspective,” AMS CONM/305, pp. 139202, 2002. Also arXiv:quant-ph/0201095.
| 6cs.PL
|
Puzzle Imaging: Using Large-scale Dimensionality Reduction
Algorithms for Localization
Joshua I. Glaser1* , Bradley M. Zamft2 , George M. Church2 , and Konrad P. Kording1,3,4
arXiv:1502.07816v3 [q-bio.NC] 21 Jun 2015
1
Department of Physical Medicine and Rehabilitation, Northwestern University and Rehabilitation Institute of
Chicago, Chicago, Illinois, USA
2
Department of Genetics, Harvard Medical School, Boston, Massachusetts, USA
3
Department of Physiology, Northwestern University, Chicago, Illinois, USA
4
Department of Applied Mathematics, Northwestern University, Chicago, Illinois, USA
*
Direct correspondence to [email protected]
Abstract
Current high-resolution imaging techniques require an intact sample that preserves spatial relationships. We here present a novel approach, “puzzle imaging,” that allows imaging a spatially scrambled
sample. This technique takes many spatially disordered samples, and then pieces them back together
using local properties embedded within the sample. We show that puzzle imaging can efficiently produce
high-resolution images using dimensionality reduction algorithms. We demonstrate the theoretical capabilities of puzzle imaging in three biological scenarios, showing that (1) relatively precise 3-dimensional
brain imaging is possible; (2) the physical structure of a neural network can often be recovered based
only on the neural connectivity matrix; and (3) a chemical map could be reproduced using bacteria with
chemosensitive DNA and conjugative transfer. The ability to reconstruct scrambled images promises to
enable imaging based on DNA sequencing of homogenized tissue samples.
Contents
1 Introduction
2
2 Results
2.1 Neural Voxel Puzzling . . . . . .
2.1.1 Overview . . . . . . . . .
2.1.2 Dimensionality Reduction
2.1.3 Performance . . . . . . .
2.2 Neural Connectomics Puzzling .
2.2.1 Overview . . . . . . . . .
2.2.2 Dimensionality Reduction
2.2.3 Performance . . . . . . .
2.3 Chemical Puzzling . . . . . . . .
2.3.1 Overview . . . . . . . . .
2.3.2 Dimensionality Reduction
2.3.3 Performance . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4
4
4
5
6
9
9
9
10
11
11
13
14
3 Discussion
3.1 Potential Uses of Puzzle Imaging
3.2 Simulation Limitations . . . . . .
3.3 Algorithms . . . . . . . . . . . .
3.4 Conclusion . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
17
17
17
17
18
1
4 Methods
4.1 Sparse Diffusion Maps Algorithm . . . . . . . . . . . . . . . .
4.2 Unweighted Landmark Isomap Algorithm . . . . . . . . . . .
4.3 Neural Voxel Puzzling . . . . . . . . . . . . . . . . . . . . . .
4.3.1 Similarity Matrix . . . . . . . . . . . . . . . . . . . . .
4.3.2 Simulations . . . . . . . . . . . . . . . . . . . . . . . .
4.4 Neural Connectomics Puzzling . . . . . . . . . . . . . . . . .
4.4.1 Similarity Matrix . . . . . . . . . . . . . . . . . . . . .
4.4.2 Simulations . . . . . . . . . . . . . . . . . . . . . . . .
4.5 Chemical Puzzling . . . . . . . . . . . . . . . . . . . . . . . .
4.5.1 Similarity Matrix . . . . . . . . . . . . . . . . . . . . .
4.5.2 Simulations . . . . . . . . . . . . . . . . . . . . . . . .
4.6 Computational Complexity . . . . . . . . . . . . . . . . . . .
4.6.1 Sparse Diffusion Maps . . . . . . . . . . . . . . . . . .
4.6.2 Unweighted Landmark Isomap . . . . . . . . . . . . .
4.6.3 Neural Voxel Puzzling . . . . . . . . . . . . . . . . . .
4.6.4 Neural Connectomics Puzzling . . . . . . . . . . . . .
4.6.5 Chemical Puzzling . . . . . . . . . . . . . . . . . . . .
4.7 Algorithm Comparison: ULI vs SDM . . . . . . . . . . . . . .
4.7.1 Domain . . . . . . . . . . . . . . . . . . . . . . . . . .
4.7.2 Reconstruction Accuracy . . . . . . . . . . . . . . . .
4.7.3 Speed . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.8 Algorithm Time Improvements . . . . . . . . . . . . . . . . .
4.8.1 Sparse Diffusion Maps vs. Diffusion Maps . . . . . . .
4.8.2 Unweighted Landmark Isomap vs. Landmark Isomap
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5 Acknowledgements
6 Supplementary Information
6.1 Chemical Puzzling Conjugation Efficiency
6.2 Reconstruction of Non-cubes . . . . . . .
6.3 Simulation Limitations . . . . . . . . . . .
6.3.1 Neural Voxel Puzzling . . . . . . .
6.3.2 Neural Connectomics Puzzling . .
6.3.3 Chemical Puzzling . . . . . . . . .
25
.
.
.
.
.
.
.
.
.
.
.
.
7 References
1
18
18
18
19
19
20
20
20
20
21
21
21
21
21
21
22
23
23
23
23
23
24
24
24
24
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
25
25
27
27
27
27
28
28
Introduction
Many biological assays require the loss/destruction of spatial information of samples, making it difficult to
create a high-resolution image of cellular properties. As a prime example, determining genetic properties of
a biological sample usually requires breaking apart that sample for DNA sequencing (but see [1]). This limits
the resolution of an image of genetic content to the precision of tissue sectioning prior to sequencing. Along
with determining gene expression, researchers are attempting to use genetic information to determine neural
connectivity [2], neural activity [3], other cellular properties [4], and chemical concentrations [5]. Being
able to image these types of properties at high resolution and large scale could therefore lead to expanded
measurement and recording applications. This would be possible if we could recover spatial information post
hoc.
In order to recover a sample’s spatial information, information about its relative spatial location could be
embedded and utilized. For example, imagine that each piece of genetic information was attached to a puzzle
piece (the embedded relative spatial information; Figure 1A). While the puzzle pieces by themselves don’t
2
provide spatial information, fitting the pieces together would lead to a spatially correct image of the genetic
information. Thus, the use of relative spatial information (how the puzzle pieces’ locations relate to one
another) could allow for higher-resolution imaging.
Using relative spatial information to reconstruct an image can be thought of as a dimensionality reduction
problem. If there are N puzzle pieces, one can construct an N × N similarity matrix S, where Sij determines
how close puzzle piece i and j are (higher similarity means shorter distance; Figure 1B). The goal is to map
this high dimensional similarity matrix to accurate 2- or 3-dimensional locations of each piece (Figure 1C).
Importantly, there is a whole class of dimensionality reduction methods that aims to preserve high dimensional distances in the reduced dimension (e.g. [6–8]). These types of techniques would allow a “piecing of
the puzzle back together.”
Here, we propose “puzzle imaging,” and develop two dimensionality reduction algorithms that would allow
large-scale puzzle imaging. We describe three concrete examples in which puzzle imaging would be beneficial:
(1) “Neural Voxel Puzzling,” in which a relatively high-resolution 3-dimensional brain map is reproduced
by giving DNA barcodes to neurons; (2) “Neural Connectomics Puzzling,” in which neural connections are
used to recover neural locations; and (3) “Chemical Puzzling,” in which a chemical map could be reproduced
using bacteria with chemosensitive DNA and conjugative transfer. Each of these examples leverage the
faster-than-Moore’s law advances in both the cost and speed of genetic sequencing [9], and therefore are
likely to become more relevant as this “genetic revolution” proceeds. In each example, we use our algorithms
on simulated data, and provide a preliminary demonstration of the capabilities of puzzle imaging.
3
Figure 1: Puzzle Imaging. There are many properties, such as genetic information, that are easier to
determine when the original spatial information about the sample is lost. However, it may be possible to
still image these properties using relative spatial information. (A) As an example, let us say that each piece
of genetic information is attached to a puzzle piece. While the puzzle pieces don’t provide absolute spatial
information, they provide relative spatial information: we know that nearby pieces should have similar
colors, so we can use color similarity to determine how close puzzle pieces should be to one another. (B) We
can make a similarity matrix of the puzzle pieces, which states how similar the puzzle pieces’ colors are to
each other, and thus how close the pieces should be to one another. (C) Through dimensionality reduction
techniques, this similarity matrix can be used to map each puzzle piece to its correct relative location.
2
2.1
2.1.1
Results
Neural Voxel Puzzling
Overview
The purpose of neural voxel puzzling is to create a 3-dimensional image of the brain at high resolution. This
image could provide useful neuroanatomical information by itself, or could be used in conjunction with other
technologies to determine the locations of genetically encoded neural properties in the brain [10].
The first step in voxel puzzling is to label each neuron in the brain with a unique DNA or RNA barcode
4
throughout its entire length (Figure 2A). Ongoing research aims to tackle this challenge [2, 11, 12]. Recently,
researchers have succeeded in having bacteria generate a large diversity of barcodes in vivo [11]. Next, the
brain is shattered into many voxels (Figure 2B; note that voxels are only squares for simplification), and the
DNA in each voxel is sequenced, yielding a record of which neurons were in which voxels. This provides us
with relative spatial information about voxel placement: voxels that share more neurons will likely be closer
to each other. We can use this relative spatial information to puzzle the voxels into their correct locations.
2.1.2
Dimensionality Reduction
More formally, puzzling together the brain is a dimensionality reduction problem, with each voxel represented
by an N -dimensional object, where N is the number of neurons. Dimension k corresponds to neuron k
being in that voxel (Figure 2C). These N -dimensional voxels must be mapped to their correct 3-dimensional
coordinates (or 2 dimensions in Figure 2’s simplified example). This can be done because voxels that are closer
in 3-dimensional space will have more neurons in common, and thus will be more similar in N -dimensional
space. With the knowledge of the similarity between all pairs of voxels, we can create a similarity matrix
(Figure 2D) and then puzzle the voxels back together (Figure 2E). Without any additional information, it
is impossible to know exact locations; only the relative locations of voxels can be known. Thus, the output
could be a flipped or rotated version of the input (Figure 2E). That being said, additional information could
be used to correctly orient the reconstruction. For example, the overall shape could be matched for a sample
that is asymmetric (like the brain), or a few samples with known locations could act as landmarks.
In order to convert the knowledge of which neurons are in which voxels (Figure 2C) to a reconstructed
puzzle (Figure 2E), a dimensionality reduction algorithm is needed. We demonstrate the performance of two
algorithms that are promising for large-scale problems. The first is a variant of Diffusion Maps [8], which
uses a sparse similarity (affinity) matrix instead of the standard calculation. We will refer to this method
as Sparse Diffusion Maps, or SDM. The second is a variant of Landmark Isomap [13, 14], which is faster
for unweighted graphs (binary similarity matrices). We will refer to this method as Unweighted Landmark
Isomap, or ULI. In the below simulations, we use 10 landmark points. For the full algorithms, see Methods.
5
Figure 2: Neural Voxel Puzzling Overview. (A) An example of 6 “neurons” (lines) going through 4
voxels. Each neuron has a unique DNA barcode (here color). (B) These voxels are broken apart. (C) A
coincidence matrix, X, is constructed describing which neurons are in which voxels. Gray signifies that a
neuron is in a particular voxel. (D) A similarity matrix is constructed describing how many neurons a pair
of voxels has in common. This matrix can be calculated as XXT . (E) The voxels are puzzled back together.
The reconstruction may be rotated or flipped, as shown here.
2.1.3
Performance
We tested the ability of both dimensionality reduction algorithms to determine the locations of 8000 simulated
voxels of varying dimensions. Voxels were not confined to be cubes; they could be any shape. In our simplistic
simulations, our “neurons” were long rods with cross-sectional areas of 1 µm2 (about the size of an axon [15])
and were assumed to fully go through each voxel they entered. We set the total number of neurons in our
simulations so that they would fill voxels of the chosen size. Neurons were oriented within the volume at
randomly determined angles. See Methods for simulation details.
In our simulations, we used two metrics to determine the quality of the reconstruction. The first metric was
the mean error in distance across all voxels. Because the final reconstruction will be a scaled, rotated, and
reflected (flipped) version of the initial voxel locations, we first transformed (scaled, rotated, and reflected)
the final reconstruction to match the original voxel locations. We used the transformation that minimized
the mean error.
Another metric we used is the correlation coefficient (R value) of the plot of reconstructed distances between
all points against their true distances. This metric determines how well relative distances between points are
preserved, and does not depend on any transformations following the reconstruction. A perfect reconstruction
(ignoring differences in scaling, rotation, and reflection) would have a linear relationship between the true
and reconstructed distances, and thus an R value of 1.
We first tested both methods on simulations with voxels with average sides of 5 µm (Figure 3A). The SDM
method led to a faithful reconstruction, with the exception that the reconstructed voxels tended to be
overrepresented around the outside of the cube and underrepresented in the middle. The ULI method also
leads to a faithful reconstruction (Figure 3B).
6
We next tested both methods of reconstruction while varying the voxel size (Figure 3C,D,E). While voxels
could be any shape, for ease of understanding, we report voxel sizes as the edge length of the cube corresponding to the average voxel size. For most voxel sizes, ULI leads to a slightly more accurate reconstruction
than SDM. In general, when looking at the error in terms of absolute distance, using smaller voxels increases
possible resolution. As an example of the excellent resolution that can be achieved, both methods achieve
mean errors below 6 µm using an average voxel size of 3 µm (Figure 3E).
Finally, we tested the performance of puzzle imaging when removing a fraction of the voxels. We did this
because in real neurons, there are cell bodies that would fill multiple voxels. These voxels would not add any
information to puzzle reconstruction, so we exclude them to simulate the existence of cell bodies. Moreover,
this generally simulates the robustness of neural voxel puzzling to missing data. In these simulations, we
used a voxel size of 5 µm (Figure 3F,G). For both methods, when no voxels were removed, the average mean
error rate was below 2 voxels. The average mean error rate was still below 2.5 voxels when 60% of the voxels
were removed. This demonstrates that these methods are robust to missing voxels (cell bodies).
7
Figure 3: Neural Voxel Puzzling Performance. (A) On the left, an example reconstruction of voxel
locations using the SDM method. Colors are based on initial locations: those with a larger initial x location
are redder, while those with a larger initial y location are bluer. In the middle, a 2-dimensional slice through
reconstructed volume. The distance errors are calculated following scaling and rotating the reconstructed
volume to match the original volume. On the right, one metric for the accuracy of reconstruction is shown by
plotting the reconstructed distances between all points against their true distances for the reconstruction in
this panel. The mean plus/minus the standard deviation (shaded) is shown. A perfect reconstruction would
be a straight line, corresponding to an R value of 1. (B) Same as panel A, except an example reconstruction
using the ULI method. (C) R values for simulations using the SDM method (blue) and ULI method (red), as
a function of the voxel size. While voxels were not confined to be cubes, for ease of understanding, we report
voxel sizes as the edge length of the cube corresponding to the average voxel size. Error bars represent the
standard deviation across simulations in each panel. (D) Mean distance errors in voxels for both methods as
a function of the voxel size. (E) Mean distance errors in microns for both methods as a function of the voxel
size. (F) Voxels are removed to represent voxels that do not contain location information (such as voxels
8
that contain a single cell body). R values for simulations using both methods are plotted as a function of
the percentage of voxels removed. (G) Mean distance errors are plotted as a function of the percentage of
voxels removed.
2.2
2.2.1
Neural Connectomics Puzzling
Overview
The purpose of neural connectomics puzzling is to estimate the locations of neurons based on their neural
connections. There is ongoing work to provide a unique DNA barcode to every neuron and link those
barcodes when neurons are synaptically connected [2]. This DNA could later be sequenced to determine
all neural connections. For this proposed technology, it would be useful to know the locations of connected
neurons, rather than simply knowing that two neurons are connected.
The first steps in connectomics puzzling are to label each neuron in the brain with a unique DNA barcode
and label neural connections via the pairing of barcodes (Figure 4A). One proposed method would be to
have viruses shuttle the barcodes across synapses, where they can be integrated [2]. Other techniques for
creating, pairing, and transporting barcodes have been proposed [10, 16]. Next, the brain is homogenized,
and the DNA barcode pairs are sequenced, yielding a record of which neurons are connected (Figure 4B).
2.2.2
Dimensionality Reduction
Each neuron can be treated as an N -dimensional object, where N is the number of neurons. Each dimension
corresponds to a connection with a given neuron (Figure 4C). These N -dimensional neurons must be mapped
to their correct 3-dimensional coordinates (or 2 dimensions in Figure 4’s simplified example). This can be
done because neurons are more likely to be connected to nearby neurons (e.g. [17,18]), and thus we can treat
the connectivity matrix as a similarity matrix. Using this similarity matrix, we can then puzzle the neurons
back into place (Figure 4D). As with voxel puzzling, it is impossible to know exact locations without any
additional information; only the relative locations of neurons can be known (Figure 4D).
In order to convert the knowledge of which neurons are connected (Figure 4C) to a reconstructed puzzle
(Figure 4D), we used SDM, which we previously used for Neural Voxel Puzzling. See Algorithm Comparison:
ULI vs SDM for an explanation on why ULI will not work for Neural Connectomics Puzzling.
9
Figure 4: Neural Connectomics Puzzling Overview. (A) An example of 9 connected neurons (circles).
Lines signify connections. (B) After the brain is homogenized, the only remaining information is a record of
the connections. Connections are shown here as adjacent circles. (C) A connectivity matrix is constructed
describing the connections between neurons. Gray signifies a connection. Since connections are correlated
with how close neurons are to one another, this connectivity matrix can be treated as the similarity matrix.
(D) The neurons are puzzled back together. The formation may be rotated or flipped, as shown here.
2.2.3
Performance
We tested the ability of connectomics puzzling to determine the locations of a simulated network of neurons.
Hellwig [17] described the probability of connections as a function of distance between pyramidal cells in
layers 2 and 3 of the rat visual cortex. We simulated 8000 neurons in a 400 µm edge-length cube (so they
were on average spaced ∼20 µm apart from each other in a given direction). Connections between neurons
were randomly determined based on their distance using the previously mentioned probability function
(Figure 5E).
We first simulated only connections within layer 3. As our example simulation shows (Figure 5A), the
locations of neurons were able to be estimated very accurately. As with voxel puzzling, we use mean error
and R values to describe the quality of reconstruction, although now the distance is between neurons rather
than between voxels. Next we simulated connections between layers 2 and 3; in our simulation, half of the
10
neurons were in layer 2 and half were in layer 3. In an example simulation (Figure 5B), it’s clear that the
locations of neurons could still be estimated accurately. In fact, the reconstruction clearly separated the
neurons from the two layers.
Looking quantitatively at the differences between layer 3 reconstructions and layer 2/3 reconstructions, R
values are slightly higher, and mean errors are slightly lower for layer 3 reconstructions. Median R values across simulations are 0.97 vs. 0.91 (layer 3 vs. 2/3), and median mean errors across simulations are
19 µm vs. 42 µm (Figure 5C,D). This disparity is largely due to the gap between layers in the reconstructions (Figure 5B); reconstructions within each layer are as accurate (in terms of R values) as the layer 3
reconstruction.
Next, in order to test when connectomics puzzling would be useful, we tested how reconstruction is dependent
on the parameters of the connection probability distribution. For these simulations, we assumed the layer
3 connection probability distribution, and then changed either the baseline connection probability (the
connection probability at a distance of 0) or the standard deviation of the connection probability distribution.
There was a great improvement in reconstruction accuracy when the baseline probability increased from
0.10 to 0.15, and accuracy continued to increase until a baseline probability of 0.35, where it plateaued
(Figure 5F). In general, there are high accuracy reconstructions for a wide range of baseline probabilities,
with the exception of low probabilities.
When looking at the effect of the standard deviation of the probability distribution, there was a general trend
that reconstruction accuracy decreased as the standard deviation increased (Figure 5G). This is because
connections become less closely related to distances when the standard deviation increases. For example, if
the standard deviation was infinite so that the probability of connection was the same for all distances, then
knowing the connections would no longer allow us to infer the distances between neurons.
One notable exception from the above trend is that there is a low reconstruction accuracy for the smallest
standard deviations considered (50-100 µm). We have found that the algorithm does not work well in
connectomics puzzling when the connection matrix is far too sparse. When we use a cube with 200 µm edges
and 1000 neurons instead of 400 µm edges and 8000 neurons (the same neuronal density), simulations with a
50 µm standard deviation perform very well. Thus, it’s important to note that the quality of reconstruction
will depend on the size of the cube being used.
2.3
2.3.1
Chemical Puzzling
Overview
Puzzle imaging has a number of potential applications besides neuroscience, including some that are potentially more tractable and therefore may have a higher probability of being implemented in the near future.
One example we develop here is that of reconstructing spatial chemical environments from populations of
recording cells that have been mixed, a process we call “chemical puzzling.”
One example of a chemical puzzling assay would consist of the initial spreading of many “pioneer” cells
across an environment containing a heterogeneous distribution of a particular chemical (Figure 6A,B). These
pioneer cells would be endowed with the ability to detect the presence of that chemical and to record its
concentration into a nucleotide sequence. This could be done through molecular ticker-tape methods using
DNA polymerases [3, 19, 20], similar strategies using RNA polymerases, or other mechanisms [4, 5, 21, 22], for
example involving chemically-induced methylation or recombination.
Some pioneer cells would also be given the ability to share genetic information with other cells — in the
case of prokaryotes, this could be by the introduction of barcoded F-like plasmids encoding the components
essential to conjugative transfer of the plasmid [23]. The pioneer cells would be placed at random places
within the chemical environment (Figure 6B), and would then begin to grow outwards (Figure 6C). When
the colonies become large enough to contact neighboring colonies, those that contain the F-plasmid (or
equivalent), denoted F+ , will copy it and transfer it to those without it (F- ) (Figure 6D). We can think
about the F-plasmid transfer between the colonies descended from two pioneer cells as these two pioneer
11
Figure 5: Neural Connectomics Puzzling Performance.
12
Figure 5: (A) On the left, an example reconstruction of neural locations based on a simulation of connections of pyramidal cells in layer 3 of cortex. Colors are based on initial locations: those with a larger
initial x location are more red, while those with a larger initial y location are more blue. The distance errors
are calculated following scaling and rotating the reconstructed volume to match the original volume. In the
middle, a 2-dimensional slice through reconstructed volume. On the right, one metric for the accuracy
of reconstruction is shown by plotting the reconstructed distances between all points against their true distances for the reconstruction in this panel. The mean plus/minus the standard deviation (shaded) is shown.
A perfect reconstruction would be a straight line, corresponding to an R value of 1. (B) Same as panel A,
except based on an example simulation of pyramidal cells in both layers 2 and 3 of cortex. (C) Boxplots of
R values for layer 3 simulations, and layer 2/3 simulations. The 5 lines (from bottom to top) correspond to
5%, 25%, 50%, 75%, and 95% quantiles of R values across simulations. (D) Boxplots (as in panel C) of mean
errors across simulations. (E) The probability of connection as a function of distance between pyramidal
cells, which is used in the simulations of the other panels [17]. (F) Using the parameters of the connectivity
probability distribution of layer 3, the baseline connection probability (the probability of connection at a
distance of 0) of the connectivity distribution is changed. R values and mean errors are shown as a function
of this baseline probability. Error bars represent the standard deviation across simulations in this and the
next panel. (G) Using the parameters of the connectivity probability distribution of layer 3, the standard
deviation of the connectivity distribution is changed. R values and mean errors are shown as a function of
the standard deviation.
cells becoming “connected” (just like a neural connection in the previous section). This sharing of genetic
information provides information about which pioneer cells are close to each other, which can then be used
to reconstruct where each cell was spatially.
Again, note that a benefit resulting from this strategy is that the spatial information can be destroyed and
reconstructed post hoc. In terms of this example, this would be equivalent to wiping the surface that the
bacteria are growing on and reconstructing its chemical spatial information after sequencing, in “one pot.”
Preparation of the DNA for sequencing could also be done in one pot, i.e. all of the cells could be lysed and
the DNA extracted at one time, if the genetic material carrying the connection information and the chemical
information were physically linked (e.g. if a barcode on the F-plasmid were inserted into the chemical record
region of the recipient cell’s genetic information). Otherwise, amplification of the connection information
and the chemical record could be done on each cell using emulsion-based methods [24].
2.3.2
Dimensionality Reduction
Each pioneer cell can be thought of as an N -dimensional object, where N is the number of pioneer cells. Each
dimension corresponds to a connection with another pioneer cell (whether the pioneer cell’s descendants are
involved in a conjugative transfer with another pioneer cell’s descendants) (Figure 6E). These N -dimensional
cells must be mapped to their correct 2-dimensional coordinates. This can be done because pioneer cells
will be more likely to be connected to nearby pioneer cells. We construct a similarity matrix by determining
all cells two connections away. This is because the matrix of connections doesn’t directly provide accurate
information about how close the pioneer cells are to each other, as F- ’s can’t be connected to F- ’s (and same
for F+ ’s). The matrix of mutual connections allows F- ’s to be connected to F- ’s. With the knowledge of
the similarity between all pairs of pioneer cells, we can puzzle the pioneers back into place (Figure 6G). The
chemical recording functionality of the cells then allows the chemical environment to be determined at the
locations of the pioneer cells (Figure 6H), and extrapolated beyond (Figure 6I).
In order to convert the knowledge of the similarity between cells (Figure 6F) to a reconstructed puzzle
(Figure 6G), a dimensionality reduction algorithm is needed. Here we use Unweighted Landmark Isomap, as
we used in Neural Voxel Puzzling, with only 5 landmark points. See Algorithm Comparison: ULI vs SDM
for an explanation on why SDM will not work well for Chemical Puzzling.
13
Figure 6: Chemical Puzzling Overview. (A) An example area contains a chemical whose concentration
is represented by a grayscale value. (B) Pioneer cells (X’s and O’s, each with a different color) are put on
the plate. X’s represent F+ cells that can transfer an F-plasmid into an F- cell (O’s). (C) The pioneer cells
replicate and spread. Descendants of a pioneer are shown in the same color. (D) When the cell colonies
become large enough to contact neighboring colonies, the F+ cells (X’s) will copy the F-plasmid and transfer
it to the F- cells (O’s). This is shown as the X’s color filling in the center of the O’s. In the inset (below),
the F-plasmid transfer (conjugation) is shown. (E) The DNA is sequenced to determine which pioneer cells
are “connected” (which had a conjugative transfer occur between their colonies). A connectivity matrix is
made from this data. (F) The matrix of connections doesn’t directly provide accurate information about
how close the original cells are to each other because O’s can’t be connected to O’s (and same for X’s).
As our similarity matrix, we thus use the matrix of mutual connections, which allows O’s to be connected
to O’s. (G) The location of the original cells is estimated from the matrix in panel F. (H) The chemical
concentrations at each of the original cells locations is known as the cells’ DNA acts as a chemical sensor. (I)
The chemical concentration everywhere is extrapolated based on the chemical concentrations at the known
pioneer cells.
2.3.3
Performance
To further demonstrate the potential for chemical puzzling, we performed a simulation of the chemical
puzzling problem. We used a complex chemical concentration described by the letter “P” (for Puzzle
Imaging), with the concentration also decreasing when moving outwards from the center, and a constant
background concentration (Figure 7A). The image size is 1000 x 1000 pixels, and each pixel is 1 µm2 (about
the size of a cell [25]). The corresponding size of the letter, then, is about 600 µm x 800 µm. Each pioneer
cell is randomly placed on a single pixel.
We simulated the placement and growth of thousands of pioneer cells in that environment (Figure 7B,C).
14
Each cell synthesized a DNA element in which the percentage GC composition of the synthesized DNA was
proportional to the chemical concentration at that spot. We were able to successfully reconstruct the letter
“P” in the image (Figure 7B,C), though this was dependent on the number of pioneer cells used and the
length of the incorporated DNA element. Fewer pioneer cells resulted in a decrease in the spatial resolution,
whereas the dynamic range greatly increased with longer DNA incorporations (Figure 7B,C). When only
two base pairs are used, the background concentration and the concentration decrease away from the center
were unable to be accurately detected. With 50 base pairs, the chemical concentrations were reconstructed
very accurately.
Lastly, the above simulations assumed that when a F+ and F- cell are in contact, transfer of genetic information occurs 100% of the time, which would likely not occur [26]. Still, relatively faithful chemical
reconstruction was accomplished with conjugation efficiencies as low as 30% (see Figure S1 and Supplementary Information). Overall, this technique holds the potential to determine chemical concentrations at very
high resolution.
15
Figure 7: Chemical Puzzling Performance. (A) The chemical concentration across the plate. It
is described by the letter “P,” with the concentration decreasing moving outwards from the center, and
a constant background concentration. (B) A simulation is done with an initial cell density of 1%. (C)
A simulation is done with an initial cell density of 0.1%. For panels B and C, the top row shows the
initial locations of the pioneer cells. They are color-coded by location. The second row shows example
reconstructed locations of the pioneer cells. The third row shows the reconstructed chemical concentrations
when 50 base pairs are used to detect the concentration. The bottom row shows the reconstructed chemical
concentrations when 2 base pairs are used to detect the concentration. Note that the black border represents
regions of unknown concentration.
16
3
Discussion
Here we proposed the concept of puzzle imaging. We developed two possible large-scale nonlinear dimensionality reduction algorithms for use in puzzle imaging, and demonstrated some of puzzle imaging’s abilities
and weaknesses in three possible applications. Using simplistic simulations, we showed that voxel puzzling
may allow locating neural structures within about 10 µm. In regards to connectomics puzzling, knowing only
the connections between neurons within a layer of cortex could be sufficient to localize neurons within about
20 µm. We also showed that chemical puzzling could be used to accurately determine chemical concentrations
at a resolution well below 1 mm. Lastly, we describe how Sparse Diffusion Maps is faster than Diffusion
Maps, and Unweighted Landmark Isomap is faster than Isomap (see Methods).
3.1
Potential Uses of Puzzle Imaging
For neuroscience, puzzle imaging could be a scalable way to localize important molecular information within
the brain. Neuronal RNA/DNA barcodes could be annotated with additional molecular information [10,16].
For example, molecular ticker tapes have been proposed that would record the activity of a neuron into
DNA [3, 19, 20]. It would be very valuable to know the location of the neurons that are being recorded
from. Additionally, RNA that is being expressed could be annotated to the barcodes. This could provide
information about what genes are being expressed in different locations of the brain, or about the distribution
of cell types throughout the brain. If engineered cells capable of recording and/or stimulating adjacent
neurons can circulate in and out of the brain, the concepts outlined here might help achieve input/output
to many/all neurons without surgery.
The applications of chemical puzzling go beyond that of determining the chemical composition of bacterial
cells growing on a surface. Indeed, biology often has the ability to survive and thrive in the harshest
of environments, including spaces too small or desolate for even robotic access. Biological sensing and
recording can open these areas to characterization and perhaps utilization. Applications could include those
in the fields of geology (in which the composition of fractured geologic strata, which contain millions of
microscopic cracks and pores, can be assayed) and medicine (in which the composition of the complex
biological environments of, for example, the gastrointestinal tract, can be assayed).
3.2
Simulation Limitations
In all our simulations, we made simplifications in order to provide a preliminary demonstration of the
feasibility of puzzle imaging. In neural voxel puzzling, our simulations assumed that there were equal neuronal
densities across the volume and that neurons were oriented at random angles. In neural connectomics
puzzling, our simulations used a maximum of two neuronal layers, and assumed that connection probability
distributions did not differ within a layer (although in reality there are different cell types [27]). For both
of these neuroscience examples, reconstruction errors due to these simplifying assumptions could likely be
remedied by using previously known neuroanatomical information (e.g. the orientation of axons in a brain
region) or molecular annotations (e.g. about whether a barcode is in an axon or dendrite, or about cell
type [28]). In chemical puzzling, our simulations assumed that cells were stationary, and we used a simple
model of outward cellular growth. For locations with viscous flow or rapidly moving cells, complex cell
growth and movement models would be necessary to achieve accurate reconstructions.
For a more in-depth discussion of limitations for all simulations, see Supplementary Information.
3.3
Algorithms
In this paper, we demonstrated two algorithms that could be used for puzzle imaging. Refinements of the
presented algorithms would be beneficial for puzzle imaging (to overcome the limitations in Results). For
instance, a different metric could be used to create the similarity matrix in both methods. In addition, the
17
number of landmark points in Unweighted Landmark Isomap can be optimally tuned for the specific use.
Moreover, novel algorithms for large-scale nonlinear dimensionality reduction would be beneficial for more
accurate puzzle imaging reconstruction.
While the algorithms presented here were designed for puzzle imaging, they could be generally used as faster
versions of Diffusion Maps and Landmark Isomap for large problems. Both methods can preserve relative
locations when reconstructing a swiss roll or helix, classical tests of nonlinear dimensionality reduction
techniques. Further research needs to be done to see how these methods compare to traditional methods for
other applications.
3.4
Conclusion
Here we have provided a preliminary demonstration that puzzle imaging may be possible. In order to make
puzzle imaging a reality, significant biological work still needs to be done. For example, having location
or neuron specific barcodes would be needed to make the neuroscience puzzle imaging approaches possible.
We hope that this paper will inspire experimentalists and theoreticians to collaborate to help make puzzle
imaging a reality.
4
4.1
Methods
Sparse Diffusion Maps Algorithm
Sparse Diffusion Maps is the same as the standard Diffusion maps algorithm [8] when using 1 timestep,
except we here use a sparse similarity matrix. Thus, in the below algorithm, only step 1 differs from
standard Diffusion maps. The algorithms take N high-dimensional points, and reduce each point to k
dimensions (where k is 2 or 3 in our applications). The algorithm follows in Table 1.
Step 1
Step 2
Sparse Diffusion Maps Algorithm
Efficiently create a sparse, symmetric, nonnegative, similarity matrix, S that is a connected graph.
Methods used to create S in our applications follow.
−1
Create matrix M by normalizing
P S so that each row sums to 1. That is, M = D S, where D is
a diagonal matrix with Di,i =
Si,j
j
Step 3
Step 4
Find the k+1 largest eigenvalues e1 , ..., ek+1 of M and their corresponding eigenvectors v1 , ..., vk+1 .
Each eigenvector is N -dimensional.
The final k-dimensional positions of the N points are the N rows in [e2 v2 , ..., ek+1 vk+1 ].
Table 1: Sparse Diffusion Maps Algorithm
4.2
Unweighted Landmark Isomap Algorithm
Unweighted Landmark Isomap is based on Landmark Isomap [13, 14]. The most important change is that
we compute geodesic distances more efficiently due to our graph being unweighted (Step 3). Additionally,
we create the similarity matrices uniquely for each application (Step 1). All other steps are identical to
Landmark Isomap. The algorithm follows in Table 2.
Note that the method of constructing the similarity matrix (Step 1) and the number of landmark points
(Step 2) are user options within this algorithm. We discuss our choices for these steps for our applications
below.
18
Step 1
Steps 2&3
Step 2
Unweighted Landmark Isomap Algorithm
Efficiently create a sparse, binary, symmetric, similarity matrix, S that is a connected graph.
Methods used to create S in our applications follow.
These steps select a landmark point (Step 2) and then calculate the distance from all points to
that landmark point (Step 3). In total, we will select ` landmark points, so these steps will be
repeated in alternation ` times.
Select a landmark point. We do this in the following way (MaxMin algorithm from [13]):
The first landmark point is selected randomly. Each additional landmark point is chosen in succession in order to maximize the minimum distance to the previously selected landmark points. In
other words, for every point i, we calculate the minimum distance mi to all the previously selected
landmark points (indexed by j): mi = min γi,j , where γi,j is the calculated geodesic distance
j
between the points in Step 3. We then choose the point that has the largest minimum distance:
landmark = max mi .
i
Step 3
Step 4
Step 5
Calculate the geodesic distance from all N points to the previously selected landmark point. The
geodesic distance γi,j from point i to landmark point j is the number of steps it would to get from
point i to landmark point j in the graph of S. This can be calculated efficiently using a breadth
first search (as opposed to using Dijkstra’s algorithm) because S is an unweighted graph.
Use classical multidimensional scaling (MDS) [6] to place the landmark points. This means that
the points are placed in order to minimize the sum squared error between the Euclidean distances
of the placed landmark points and their known geodesic distances. This is done in the following
way [13, 14]:
Let ∆` be the matrix of squared geodesic distances between all the landmark points. Let µi be
the mean of row i of ∆` and µ be the mean of all entries of ∆` . Create the matrix B with entries
Bj,k = [∆` ]j,k − µj − µk + µ. Find the first (largest) k eigenvalues and eigenvectors of B. Let Vi,j
be the ith component of the j th eigenvector. Let λj be the j th eigenvalue. The matrix
L has the
√
coordinates of the landmark points in its columns. L has the entries Li,j = Vj,i · λi .
For each point, triangulate its location using the known geodesic distances to each of the landmark
points. This is done in the following way [13, 14]:
V
Let L# i,j = √j,i
, and let ∆ be the matrix of squared geodesic distances between the landmark
λi
¯ ` , where
points and all points. The position of point a, xa , is calculated by xa = − 21 L# ∆a − ∆
∆a is row a of the matrix ∆ (the squared geodesic distances from point a to all the landmark
¯ ` is the column mean of the matrix ∆` .
points), and ∆
Table 2: Unweighted Landmark Isomap Algorithm
4.3
4.3.1
Neural Voxel Puzzling
Similarity Matrix
Let X be an n × p (voxels × neurons) coincidence matrix that states which neurons are in which voxel. To
construct the similarity matrix, we first compute A = XXT , an n×n matrix that gives the similarity between
voxels. We then threshold this matrix (an approximate nearest neighbors calculation). Thresholding makes
the matrix more sparse (beneficial for SDM), and makes points not connected to far-away points (important
for increasing resolution in ULI). We use different thresholding methods prior to ULI and SDM. The specific
thresholding methods below are what we used for our simulations; the method of thresholding is a tunable
parameter.
For ULI, we threshold each row independently. The threshold for a row is the maximum of that row divided
by 2. That is, for row i, Ti = 12 max Aij , where Aij is the element of A in row i and column j. After
j
thresholding, the entries of the non-symmetric
similarity matrix are Nij = Aij ≥ Ti . We create a symmetric
unweighted similarity matrix: S = N + NT > 0.
For SDM, the entire matrix is thresholded by the same value. We first create a vector w that contains the
19
maximum of each row of A. Its entries are wi = max Aij . The threshold is the minimum of w. That is,
j
T = min w. Thresholding all entries, we get S = A ≥ T .
4.3.2
Simulations
For our simulations, we initially chose the average voxel size. If the average voxel had a volume of x3 µm3 ,
we reported the voxel size as x µm (the length of an edge of a cube of comparable size). We next created a
cube with edges of length 20x µm (so it had a volume of 8000x3 µm3 ). We then randomly placed 8000 voxel
centers in the cube. We next added in long rectangular prisms (neurons) so that the entire volume would be
filled. We assumed the neurons fully went through each voxel they entered. As the cross-sectional area of a
voxel was on average x2 µm2 , and we assumed neurons with cross-sectional areas of 1 µm2 (about the size
of an axon), we added enough neurons so that each voxel would contain x2 neurons. Neurons were oriented
within the volume at randomly determined angles.
When doing simulations with removed voxels, we placed voxels and neurons in the regular way. Then, we
discarded voxels and aimed to reconstruct the locations of the other voxels. The R values and mean errors
listed are for the voxels that were reconstructed (removed voxels were not included in the analysis).
Additionally, in our simulations, voxels that did not contain any neurons, or that did not share any neurons
with another voxel, were excluded from analysis. When voxels were 2 µm, 3 µm, 4 µm, and 5 µm, 2.8%, 0.14%,
0.0046%, and 0.0003% of voxels were respectively excluded. Also, when removing voxels (Figure 2F,G), we
excluded the remaining voxels that did not make up the main connected component of the similarity graph.
When using ULI, when 80%, 85%, and 90% of voxels were removed, 0.36%, 1.7%, and 34.7% of remaining
voxels were excluded from analysis. When using SDM, when 80%, 85%, and 90% of voxels were removed,
0.05%, 0.09%, and 0.37% of remaining voxels were excluded from analysis.
In our simulations, when using ULI, we used 10 landmark points (see Unweighted Landmark Isomap Algorithm above).
4.4
4.4.1
Neural Connectomics Puzzling
Similarity Matrix
Let C be an n × n (neurons × neurons) connectivity matrix that states which neurons are connected to each
other. As mentioned in Results, this connectivity matrix can be directly used as the similarity matrix.
4.4.2
Simulations
We randomly placed 8000 points (neurons) in a 400 µm edge-length cube. Thus, the neurons were on average
∼20 µm apart from each other in a single direction. We initially assumed all neurons were pyramidal cells
in layer 3 of rat visual cortex and simulated connections according to the known connection probability
distribution as a function of distance (Figure 5E) [17]. Every pair of neurons was randomly assigned a
connection (or not) based on this probability distribution. Next, we simulated neurons in layers 2 and
3. The top half of the cube was assumed to be layer 2, and the bottom half was assumed to be layer 3.
Again, every pair of neurons was randomly assigned a connection (or not) based on the relevant probability
distribution (either between layer 2 and 2, layer 2 and 3, or layer 3 and 3; Figure 5E) [17].
20
4.5
4.5.1
Chemical Puzzling
Similarity Matrix
Let C be an n × n (pioneer cells × pioneer cells) connectivity matrix that states which pioneer cells are
“connected” to each other. The similarity matrix is calculated as CT C.
4.5.2
Simulations
We randomly placed pioneer cells on a pixel in the 1000 × 1000 pixel image. The number of pioneer cells
was 1 million (the number of pixels) times the initial density, so 10000 in Figure 3B, and 1000 in Figure 3C.
Each cell was randomly assigned to be F+ or F- . We assumed the cells grew out circularly over time with
approximately the same rate of growth. For every pixel, we determined which colony (progeny from which
pioneer cell) would reach the pixel first. This produced a final map of all pixels assigned to different pioneer
cells. On the borders of colonies, cells could conjugate with one another if one colony was F+ and the other
was F- . For each pixel on the border, a conjugation occurred according to the probability of conjugation
(100% probability for Figure 7; varying for Figure S1). More specifically, any time different colonies occupied
pixels horizontal or vertical of each other, a conjugation could occur for the cells in the bordering pixels.
We reconstructed the pioneer cells’ locations using ULI with 5 landmark points (see Unweighted Landmark
Isomap Algorithm above). We assumed that 3 cells were placed on the plate so that their locations were
known. We scaled, rotated, and reflected the reconstructed locations of the pioneer cells so that the locations
of the 3 known cells were matched (the average distance error was minimized).
The chemical concentration on the plate was between 0 and 1. If a pioneer cell was at a location with a
chemical concentration of ρ, each base pair has a probability of being a G or C with probability ρ. The
reconstructed chemical concentration was the total number of G’s and C’s divided by the number of base
pairs. Thus, if there were 2 base pairs, the reconstructed chemical concentration could be 0, 0.5, or 1. We
used linear interpolation to determine the chemical concentration of areas on the plate that did not have a
reconstructed concentration.
We also note that along with our approximate simulation method of outward circular growth, we also ran a
smaller, more realistic, simulation. In this stochastic growth simulation, during each round, cells randomly
divided into any single adjacent unoccupied pixel. When an F+ colony grew next to an F- colony, or vice
versa, cells could conjugate (according to the conjugation probability). F- cells turned F+ upon receipt of
the F-plasmid from an F+ cell. This growth simulation continued until the plate was full. On smaller plates
(when it was less time consuming to run the realistic simulation), both simulation types produced nearly
equivalent results.
4.6
4.6.1
Computational Complexity
Sparse Diffusion Maps
We list the complexity of Step 1 for the individual methods below. The run-time of this algorithm is
dominated by Step 3. For the below complexity explanation, S is an n × n sparse matrix with m non-zero
elements. The complexity of this algorithm is listed in Table 3.
4.6.2
Unweighted Landmark Isomap
We list the complexity of Step 1 for the individual methods below. The complexity of this algorithm is listed
in Table 4.
21
Step
Step 3
(and Total)
Sparse Diffusion Maps Computational Complexity
Worst-case
com- Further Explanation
plexity
O (m)
This is the complexity for the power iteration algorithm [29],
which is very efficient for solving for a small number of
eigenvalues/eigenvectors. In the power iteration method,
the matrix is continually multiplied by a vector, which has
complexity O (m), until convergence. The number of steps
to calculate eigenvalue ei depends on |ei /ei+1 |: a larger ratio
means quicker convergence.
If this was not a sparse matrix, as is the case for standard Diffusion Maps, then this step would be O n2 .
Table 3: Sparse Diffusion Maps Computational Complexity
Step
Step 2
Step 3
Step 4
Step 5
Total
Unweighted Landmark Isomap Computational Complexity
Worst-case
Further Explanation
complexity
O (` · n)
[13]
O (` · m)
A breadth first search has a complexity of O (m). This is done
for each of ` landmark points.
If the graph had been weighted (as assumed in standard
Landmark Isomap), then we would need to use Dijkstra’s algorithm.
The fastest general implementation is
O (m + n · log n) [30] for each landmark point, yielding
O (` · m + ` · n · log n).
O `3
[13]
O (k · ` · n)
[13]
O ` · m + k · ` · n + `3
Table 4: Unweighted Landmark Isomap Computational Complexity
4.6.3
Neural Voxel Puzzling
To compute the similarity matrix, we can take a shortcut instead of performing the sparse matrix multiplication XXT (which will be O n2 in the best case scenario [31]). The entries of XXT tell us how many
neurons are shared between particular voxels. This can also be calculated by drawing the bipartite graph
that X describes, with voxels on one side and neurons on the other (and edges when a neuron goes through
a voxel). For a given voxel, to determine which voxels it shares a neuron with, we only need to trace all the
paths to its connected neurons, and then the paths from those neurons back to voxels. This has a complexity
of O (αin · αout ), where αin is the largest number of neurons in a voxel, and αout is the largest number of
voxels a neuron goes through. Thus, for all voxels, the complexity is O (n · αin · αout ). Determining the
similarity matrix can take a comparable amount of time to the steps in ULI or SDM.
Note that this
step is faster
than the standard method of computing a similarity matrix in Diffusion Maps:
−kxi −xj k2 2
Si,j = exp
, where xi and xj are columns within X. For our sparse matrices, this would
2σ 2
take O (q · n), where q is the number of nonzero entries in X, due to the pairwise vector subtraction. The
Landmark Isomap algorithm does not give a method for computing the similarity matrix.
22
4.6.4
Neural Connectomics Puzzling
As no computation is required to construct the similarity matrix, the overall complexity of Neural Connectomics Puzzling is the complexity of the SDM algorithm.
4.6.5
Chemical Puzzling
As in neural voxel puzzling, we can take a shortcut to calculate the
similarity matrix instead of performing
the sparse matrix multiplication CT C (which would take O n2 at best for very sparse matrices [32]).
Again, we can construct a bipartite graph representing C, where the pioneer cells are now on both sides of
the graph (and there’s an edge between connected cells). We can determine if a pioneer cell is connected
within 2 steps of another pioneer cell by counting all the ways to get to the other side of the graph (via
connections) and back to cells on the same side. For a given cell, this has the complexity O β 2 , where β
is the largest number of connections a pioneer cell has (how many different cell colonies that pioneer cell’s
colony has conjugated with). Thus, the total complexity of determining the similarity matrix is O n · β 2 .
Determining the similarity matrix can take a comparable amount of time to the steps in ULI.
Additionally, chemical puzzling has the extra step of reconstructing the image of chemical concentrations.
The chemical concentrations are known at the reconstructed locations of the pioneer cells, but interpolation
needs to be used to estimate the concentrations at other locations. This interpolation step can dominate the
total time, depending on the desired resolution and interpolation method.
4.7
Algorithm Comparison: ULI vs SDM
There are pros and cons of both Sparse Diffusion Maps (SDM) and Unweighted Landmark Isomap (ULI),
which make them suitable for different applications.
4.7.1
Domain
To run ULI, the graph entered into the algorithm must contain only short-range connections. Practically,
ULI will not work when there is high similarity (a large value in the similarity matrix) between points that
are far away from one another. This is because having high similarity between far away points would make
those far away points near each other in the reconstruction, a problem known as “short circuiting” [33].
This limitation means that neural connectomics puzzling, which contains long-range connections that are
indistinguishable for short-range connections, is not compatible with ULI. SDM, on the other hand, is robust
to high similarity values between far-away points, as long as similarity values are generally higher between
nearby points. Thus, SDM does work for neural connectomics puzzling.
4.7.2
Reconstruction Accuracy
Reconstructions using SDM are generally biased towards having points around the perimeter, and therefore
don’t faithfully reconstruct the center of the volume. This perimeter bias makes the SDM method unsuitable
for use in Chemical puzzling. When used with Chemical puzzling, SDM leads to faulty chemical reconstructions near the center of the image. For neural voxel puzzling, as seen in our simulations (Figure 3), ULI was
slightly more accurate than SDM, except for when at least 85% of the voxels had been removed.
Another important note is that the SDM method will only accurately reconstruct the volume when the
volume is a cube (Figure S2). Thus, for the SDM method to be used in practice with neural voxel or
connectomics puzzling, the brain would need to first be sectioned into cubes, and then the cubes would need
to be pieced together. The ULI method, on the other hand has the benefit of accurately reconstructing
volumes that are not cubes (Figure S2).
23
4.7.3
Speed
Both SDM and ULI are designed to be fast dimensionality reduction methods for use with large datasets.
In practice, when directly comparing their speed in the neural voxel puzzling simulations using 8000 5 µm
voxels, SDM took about 1.5 seconds, while ULI (with 10 landmark points) took about 0.9 seconds on a
2.3 GHz processor running Matlab. The previous times did not include constructing the similarity matrix.
See Computational Complexity above, and Algorithm Time Improvements below, for the computational
complexity of larger problems.
Pros/Cons
Sparse Diffusion Maps
Only accurately reconstructs cubes
Biased towards exterior points
Is robust to problems that have high similarity
between some far-away points.
Summary
Unweighted Landmark Isomap
Can accurately reconstruct non-cubes
Not biased towards exterior points
Is generally not robust to problems that have high
similarity between some far-away points.
Table 5: Algorithm Comparison Summary: ULI vs. SDM
4.8
Algorithm Time Improvements
Our algorithms took previous algorithms and adapted them to be faster for large-scale puzzle imaging.
4.8.1
Sparse Diffusion Maps vs. Diffusion Maps
The difference between Sparse Diffusion Maps and Diffusion Maps is that we construct a sparse similarity
matrix. The most time consuming step in these algorithms (besides constructing the similarity matrix) is
finding the largest k eigenvalues and eigenvectors, where k is the dimension size you’re projecting into (2
or 3 for puzzle imaging). This computation will be significantly faster when the similarity matrix is sparse.
For instance, one fast algorithm, “power iteration,” [29] has a complexity of O (m), where m is the number
of non-zero elements in the similarity matrix S (see Computational Complexity).
Computing the similarity matrix is the other time consuming step in Diffusion Maps. Let’s say X is an
n × p matrix, and we want to compute the n × n similarity
matrix 2S that gives the similarity between the
−kxi −xj k2
columns of X. This is generally calculated as Si,j = exp
, where xi and xj are columns within
2σ 2
X. Computing S would have a complexity of O (q · n), where q is the number of nonzero entries in X (see
Computational Complexity).
In our scenarios where we use SDM, we compute a sparse similarity matrix more efficiently. For neural
connectomics puzzling, computing the similarity matrix takes no time, as it simply describes the connections.
For neural voxel puzzling, XXT (the main step of computing the similarity matrix) can be calculated with a
complexity of O (n · αin · αout ), where αin is the largest number of neurons in a voxel, and αout is the largest
number of voxels a neuron goes through (see Computational Complexity). This approach is thus significantly
faster than Diffusion Maps for our problem.
4.8.2
Unweighted Landmark Isomap vs. Landmark Isomap
The main difference between ULI and Landmark Isomap is that ULI uses an unweighted similarity matrix. One of the time-consuming steps in Landmark Isomap is computing the geodesic distance between
all points and each of the ` landmark points. For weighted graphs, this can be solved fastest with Dijkstra’s algorithm in O (` · m + ` · n · log n) [30]. For unweighted graphs, we can simply use a breadth first
search for each landmark point, which has a complexity of O (` · m). When we include the other steps
of the algorithm (other than constructing the similarity matrix), Landmark Isomap has a complexity of
24
O ` · m + ` · n · log n + k · ` · n + `3 , while ULI is faster, with a complexity of O ` · m + k · ` · n + `3 (see
Computational Complexity).
It is important to note that computing the similarity matrices in neural voxel puzzling and chemical puzzling
may take a comparable amount of time as ULI or Landmark Isomap. As mentioned above, computing the
similarity matrix for neural voxel puzzling has a complexity
of O (n · αin · αout ). For chemical puzzling,
computing the similarity matrix has a complexity of O n · β 2 , where β is the largest number of connections a
pioneer cell has (how many different cell colonies that pioneer cell’s colony has conjugated with). Additionally,
for chemical puzzling, using interpolation to estimate the chemical concentration can be a very time-intensive
step depending on the resolution desired.
5
Acknowledgements
We would like to thank Ted Cybulski and Mohammad Gheshlaghi Azar for very helpful comments on previous
versions of the manuscript. We would like to thank Josh Vogelstein for helpful initial discussions.
Joshua Glaser was supported by NIH grant 5R01MH103910 and NIH grant T32 HD057845. Bradley Zamft
was supported by NIH grant 5R01MH103910. George Church acknowledges support from the Office of Naval
Research and the NIH Centers of Excellence in Genomic Science. Konrad Kording is funded in part by the
Chicago Biomedical Consortium with support from the Searle Funds at The Chicago Community Trust, and
is also supported by NIH grants 5R01NS063399, P01NS044393, and 1R01NS074044.
6
6.1
Supplementary Information
Chemical Puzzling Conjugation Efficiency
The simulations in the main text (Figure 7) assumed that when an F+ and F- cell are in contact, transfer
of genetic information occurs 100% of the time. Evidence to the contrary exists, especially in the absence
of selection, where the fitness penalty associated with F+ strains carrying a plasmid, and the competition
between cell growth and conjugation, limit plasmid transfer [26]. Furthermore, recombination of the Fplasmid with the host genome can result in Hfr (high frequency of recombination) strains, which have the
ability to transfer genetic information (e.g. barcodes), but generally do not confer conjugative ability to the
recipient [34]. Our analysis, however, does not rely on every neighboring F+ /F- pair to mate. Rather, it
only requires that the descendants of nearby pioneer cells mate at least one once (if they’re an F+ /F- pair).
In simulations, relatively faithful chemical reconstruction was accomplished with a conjugation probability
of only 30% (Figure S1). This probability is likely very feasible, as neighboring pairs will have many chances
to mate.
25
Figure S1: Chemical Puzzling with Low Conjugation Efficiency. We do a simulation of chemical
puzzling, as in Figure 7B, except now with varying conjugation efficiencies. (A) The original cell locations
(left), and initial chemical concentration (right). (B-E) On the left, the reconstructed cells, and on the
right, the reconstructed chemical concentration, using (B) 100% conjugation probability, (C) 40% conjugation probability, (D), 30% conjugation probability, and (E) 20% conjugation probability. The reconstructed
chemical concentrations assumed the pioneer cells had 50 base pairs to encode the concentration. The black
area on the outside is a border, not a chemical concentration.
26
6.2
Reconstruction of Non-cubes
Figure S2: Reconstruction of Non-cubes. We do voxel puzzling as in Figure 3A,B, with 8000 5 µm
voxels. However, now the overall shape is a rectangular prism (height is half of the length and width) rather
than a cube. (A) The original voxel locations. (B) Reconstruction using SDM. (C) Reconstruction using
ULI.
6.3
6.3.1
Simulation Limitations
Neural Voxel Puzzling
In our simulations, we aimed to demonstrate some of the problems that could arise when using real data.
For instance, for voxel puzzling, we removed a large portion of the voxels to represent cell bodies taking up
multiple voxels. It would likely be possible to place these voxels when using real data, by looking at the
surrounding voxels that share a neuron tag.
In our voxel puzzling simulations, there were equal densities of neurons in all portions of the volume, which is
not the case in the brain. This should not be a large concern, however, due to our method of constructing the
similarity matrix. When we do thresholding, we are essentially taking the approximate nearest neighbors.
This step removes the effects of having more total connections between voxels in one region (due to a high
density of neurons) compared to another.
Our neurons were also all oriented at random angles, so that voxels that were horizontal of each other had
the same probability of sharing a neuron as voxels that were vertical of each other (or on top of each other).
If this is not the case (e.g. axons are oriented in a similar manner in many brain regions), the reconstruction
will be skewed. For instance, if most neurons were oriented vertically, in the reconstruction, the vertical
component will be compressed. Following reconstruction, if it’s seen that many neurons in a region are all
oriented in the same manner, the dimensions of this region could be scaled proportionally. Moreover, having
molecular annotations about whether axons or dendrites of a neuron are in a voxel could help with this
process.
6.3.2
Neural Connectomics Puzzling
In our simulations of neural connectomics puzzling, we demonstrated that accurate reconstructions were still
possible when there were multiple connection probability distributions (as a function of distance). However,
this simulation only had 2 layers, and the connection probability distributions were consistent in each layer.
It is possible that reconstructions could be less accurate when using more layers. This could be dealt
with by using a priori knowledge about neuroanatomy in order to ensure cubes were cut to contain a
homogenous population. It also could be problematic if connection probability distributions differed within
a layer across multiple cell types [27]. This could be dealt with by using annotated molecular information
about cell types [28]. In this scenario, separate reconstructions could be done using neuronal populations
with homogenous connectivity probability functions, and then these reconstructions could be combined.
27
Additionally, it may be possible to estimate cell types given connectivity information (by using clustering
methods on the connectivity data) [35], and then use these cell type estimates (as mentioned above) to help
localize the neurons.
It is also important to note that connectomics puzzling is based on the idea that there is a higher probability
of connection between neurons that are closer to each other. If this is not the case, then the connectomics
puzzling reconstruction may be inaccurate. For certain cell types or portions of the brain (e.g. layer 5
pyramidal cells), there may be a higher probability of long-range connections than short connections. Using
a priori knowledge about neuroanatomy or information about cell types based on molecular annotations,
these neurons could be excluded (or specific connections these neurons have could be excluded). Alternatively,
this information about long-range connections could be utilized in connectomics puzzling to help provide an
accurate reconstruction.
6.3.3
Chemical Puzzling
In our simulations of chemical puzzling, we used simplified models of bacterial growth and movement. In
terms of cell growth, we assumed that cell colonies keep growing outward and stop once they contact another
colony. In reality, cells will push on one another, causing the colonies to move. Additionally, in our two
dimensional simulations, we ignored the possibility of cells growing in the vertical dimension, which could
lead to additional conjugations. While we could incorporate a more sophisticated (and much more expensive
in terms of CPU time) simulation of bacterial growth, it is important to note that these simplifications will
have little effect on whether two colonies have a conjugation, which is the metric that influences our results.
Our simulations also assumed that cells are stationary; that is, once a cell occupies a particular spot in two
(or three) dimensional space, it stays there forever. In this case, the trajectory of a lineage is determined by
a two (or three) dimensional random walk with uniform and unity step size, where two trajectories do not
cross. This would certainly not be true in the case of motile cells or when there is sufficient mixing, as cells
arising from different lineages could cross before dividing. Further, if the particular chemical being sensed
induces a bias in cell motility (i.e. is a chemoattractant or chemorepellant), a bias will be further introduced
into the colony growth model. These movement affects could potentially be problematic for two reasons: 1)
it could allow far-away colonies to conjugate with one another; 2) pioneer cells could determine the chemical
concentration of a location different from the location where they start to divide. Detailed analysis of these
movement affects is beyond the scope of this paper, and we simply note that in the case that cell motility
is also an unbiased random walk (a good assumption in the absence of chemotaxis and viscous flow [36])
and there is no mixing, our analysis stands in the limit that the average distance covered by the cells during
one division time is much smaller than their size. Outside of this limit, our reconstructed images would be
blurred.
7
References
[1] Je Hyuk Lee, Evan R Daugharthy, Jonathan Scheiman, Reza Kalhor, Joyce L Yang, Thomas C Ferrante,
Richard Terry, Sauveur SF Jeanty, Chao Li, Ryoji Amamoto, et al. Highly multiplexed subcellular rna
sequencing in situ. Science, 343(6177):1360–1363, 2014.
[2] Anthony M Zador, Joshua Dubnau, Hassana K Oyibo, Huiqing Zhan, Gang Cao, and Ian D Peikon.
Sequencing the connectome. PLoS Biol, 10(10):e1001411, 2012.
[3] K. P. Kording. Of toasters and molecular ticker tapes. PLoS Comput Biol, 7(12):e1002291, 2011.
[4] Fahim Farzadfard and Timothy K Lu. Genomically encoded analog memory with precise in vivo dna
writing in living cell populations. Science, 346(6211):1256272, 2014.
[5] Felix Moser, Andrew Horwitz, Jacinto Chen, Wendell A Lim, and Christopher A Voigt. Genetic sensor
for strong methylating compounds. ACS synthetic biology, 2(10):614–624, 2013.
28
[6] Joseph B Kruskal. Multidimensional scaling by optimizing goodness of fit to a nonmetric hypothesis.
Psychometrika, 29(1):1–27, 1964.
[7] Joshua B Tenenbaum, Vin De Silva, and John C Langford. A global geometric framework for nonlinear
dimensionality reduction. Science, 290(5500):2319–2323, 2000.
[8] Ronald R Coifman and Stphane Lafon. Diffusion maps. Applied and computational harmonic analysis,
21(1):5–30, 2006.
[9] P.A. Carr and G.M. Church. Genome engineering. Nat Biotechnol, 27(12):1151–62, 2009.
[10] Adam H Marblestone, Evan R Daugharthy, Reza Kalhor, Ian D Peikon, Justus M Kebschull, Seth L
Shipman, Yuriy Mishchenko, Je Hyuk Lee, Konrad P Kording, Edward S Boyden, et al. Rosetta brains:
A strategy for molecularly-annotated connectomics. arXiv preprint arXiv:1404.5103, 2014.
[11] Ian D Peikon, Diana I Gizatullina, and Anthony M Zador. In vivo generation of dna sequence diversity
for cellular barcoding. Nucleic acids research, 42(16):e127–e127, 2014.
[12] Carmen Gerlach, Jan C Rohr, Lela Peri, Nienke van Rooij, Jeroen WJ van Heijst, Arno Velds, Jos
Urbanus, Shalin H Naik, Heinz Jacobs, Joost B Beltman, et al. Heterogeneous differentiation patterns
of individual cd8+ t cells. Science, 340(6132):635–639, 2013.
[13] Vin De Silva and Joshua B Tenenbaum. Sparse multidimensional scaling using landmark points. Report,
Technical report, Stanford University, 2004.
[14] Vin D Silva and Joshua B Tenenbaum. Global versus local methods in nonlinear dimensionality reduction. In Advances in neural information processing systems, pages 705–712, 2002.
[15] Jnos A Perge, Jeremy E Niven, Enrico Mugnaini, Vijay Balasubramanian, and Peter Sterling. Why do
axons differ in caliber? The Journal of Neuroscience, 32(2):626–638, 2012.
[16] Adam H Marblestone, Evan R Daugharthy, Reza Kalhor, Ian Peikon, Justus Kebschull, Seth L Shipman,
Yuriy Mishchenko, David A Dalrymple, Bradley M Zamft, Konrad P Kording, et al. Conneconomics:
The economics of large-scale neural connectomics. bioRxiv, 2013.
[17] Bernhard Hellwig. A quantitative analysis of the local connectivity between pyramidal neurons in layers
2/3 of the rat visual cortex. Biological cybernetics, 82(2):111–121, 2000.
[18] Beth L Chen, David H Hall, and Dmitri B Chklovskii. Wiring optimization can relate neuronal structure and function. Proceedings of the National Academy of Sciences of the United States of America,
103(12):4723–4728, 2006.
[19] Bradley M. Zamft, Adam H. Marblestone, Konrad Kording, Daniel Schmidt, Daniel Martin-Alarcon,
Keith Tyo, Edward S. Boyden, and George Church. Measuring cation dependent dna polymerase fidelity
landscapes by deep sequencing. PLoS One, 7(8):e43876, 2012.
[20] Joshua I Glaser, Bradley M Zamft, Adam H Marblestone, Jeffrey R Moffitt, Keith Tyo, Edward S
Boyden, George Church, and Konrad P Kording. Statistical analysis of molecular signal recording.
PLoS Comput Biol, 9(7):e1003145, 2013.
[21] Ari E Friedland, Timothy K Lu, Xiao Wang, David Shi, George Church, and James J Collins. Synthetic
gene networks that count. Science, 324(5931):1199–1202, 2009.
[22] George M Church and Jay Shendure. Nucleic acid memory device, 2003.
[23] Neil Willetts and Ron Skurray. The conjugation system of f-like plasmids. Annual review of genetics,
14(1):41–76, 1980.
[24] Michihiko Nakano, Jun Komatsu, Shun-ichi Matsuura, Kazunori Takashima, Shinji Katsura, and Akira
Mizuno. Single-molecule pcr using water-in-oil emulsion. Journal of biotechnology, 102(2):117–124,
2003.
29
[25] NEZCE Grossman, EZ Ron, and CL Woldringh. Changes in cell dimensions during amino acid starvation
of escherichia coli. Journal of bacteriology, 152(1):35–41, 1982.
[26] Jos M Ponciano, Leen De Gelder, Eva M Top, and Paul Joyce. The population biology of bacterial
plasmids: a hidden markov model approach. Genetics, 176(2):957–968, 2007.
[27] DTJ Liley and JJ Wright. Intracortical connectivity of pyramidal and stellate cells: estimates of synaptic
densities and coupling symmetry. Network: Computation in Neural Systems, 5(2):175–189, 1994.
[28] Dmitry Usoskin, Alessandro Furlan, Saiful Islam, Hind Abdo, Peter Lnnerberg, Daohua Lou, Jens
Hjerling-Leffler, Jesper Haeggstrm, Olga Kharchenko, Peter V Kharchenko, et al. Unbiased classification
of sensory neuron types by large-scale single-cell rna sequencing. Nature neuroscience, 2014.
[29] Gene H Golub and Charles F Van Loan. Matrix computations, volume 3. JHU Press, 2012.
[30] Michael L Fredman and Robert Endre Tarjan. Fibonacci heaps and their uses in improved network
optimization algorithms. Journal of the ACM (JACM), 34(3):596–615, 1987.
[31] Haim Kaplan, Micha Sharir, and Elad Verbin. Colored intersection searching via sparse rectangular matrix multiplication. In Proceedings of the twenty-second annual symposium on Computational geometry,
pages 52–60. ACM, 2006.
[32] Raphael Yuster and Uri Zwick. Fast sparse matrix multiplication. ACM Transactions on Algorithms
(TALG), 1(1):2–13, 2005.
[33] Laurens JP van der Maaten, Eric O Postma, and H Jaap van den Herik. Dimensionality reduction: A
comparative review. Journal of Machine Learning Research, 10(1-41):66–71, 2009.
[34] PK Gupta. Cell and Molecular biology. Rastogi Publications, 2009.
[35] Eric Jonas and Konrad Kording. Automatic discovery of cell types and microcircuitry from neural
connectomics. arXiv preprint arXiv:1407.4137, 2014.
[36] Jonathan Saragosti, Pascal Silberzan, and Axel Buguin. Modeling e. coli tumbles by rotational diffusion.
implications for chemotaxis. PloS one, 7(4):e35412, 2012.
30
| 5cs.CE
|
Socially Compliant Navigation through Raw Depth Inputs with
Generative Adversarial Imitation Learning
arXiv:1710.02543v2 [cs.RO] 26 Feb 2018
Lei Tai1
Jingwei Zhang2
Ming Liu1
Abstract— We present an approach for mobile robots to
learn to navigate in dynamic environments with pedestrians
via raw depth inputs, in a socially compliant manner. To
achieve this, we adopt a generative adversarial imitation
learning (GAIL) strategy, which improves upon a pre-trained
behavior cloning policy. Our approach overcomes the disadvantages of previous methods, as they heavily depend on the
full knowledge of the location and velocity information of
nearby pedestrians, which not only requires specific sensors,
but also the extraction of such state information from raw
sensory input could consume much computation time. In this
paper, our proposed GAIL-based model performs directly
on raw depth inputs and plans in real-time. Experiments
show that our GAIL-based approach greatly improves the
safety and efficiency of the behavior of mobile robots from
pure behavior cloning. The real-world deployment also shows
that our method is capable of guiding autonomous vehicles
to navigate in a socially compliant manner directly through
raw depth inputs. In addition, we release a simulation plugin
for modeling pedestrian behaviors based on the social force
model.
I. I NTRODUCTION
1) Socially compliant navigation: The ability to cope
with dynamic pedestrian environments are crucial for autonomous ground vehicles. In static environments, mobile
robots are required to reliably avoid collision with static
objects and plan feasible paths to their target locations;
while in dynamic environments with pedestrians, they are
additionally required to behave in socially compliant manners, where they need to understand the dynamic human
behaviors and react accordingly under specific socially
acceptable rules.
Traditional solutions can be classified into two categories: model-based and learning-based. Model-based
methods aim to extend the multi-robot navigation solutions
with socially compliant constraints [1]. However, force
parameters need to be carefully tuned for each specific
scenario. Learning-based methods, on the other hand, aim
to directly recover the expert policy through direct supervised learning (e.g., behavior cloning), or through inverse
reinforcement learning to recover the inner cost function.
Yet, all previous approaches require the knowledge of
the precise localization and velocity information of nearby
∗ This paper is supported by Shenzhen Science, Technology
and Innovation Commission (SZSTI) JCYJ20160428154842603 and
JCYJ20160401100022706; also supported by the Research Grant Council
of Hong Kong SAR Government, China, under Project No. 11210017 and
No. 16212815 and No. 21202816 awarded to Prof. Ming Liu.
1 Department of Electronic and Computer Engineering, The Hong Kong
University of Science and Technology; {ltai, eelium}@ust.hk
2 Department of Computer Science, Albert Ludwig University of
Freiburg; {zhang, burgard}@informatik.uni-freiburg.de
Wolfram Burgard2
Depth input
Fig. 1: A mobile robot navigating in a socially compliant
manner in an indoor environment. Our approach tackles
this scenario through raw depth images.
pedestrians. This limitation restricts these methods to be
only applicable for robots equipped with high precision
sensors, like 3D Lidars [2], [3]. Moreover, the estimation
of the state information of pedestrians is generally timeconsuming. Developing effective navigation strategies directly from raw sensor inputs is still of great importance.
2) Perception from visual input: Compared with 3D Lidars, vision sensors come at affordable prices, making them
more suitable to equip mobile agents that are beginning to
populate our social life. Thus, navigation solutions that can
directly operate on raw visual inputs are more feasible than
those depending on expensive Lidars.
In coping with visual information, deep learning approaches have become imperative due to their ability to
extract hierarchically more abstract feature representations.
Also, with the improvement of the computation power of
mobile platforms, such approaches have been utilized in
various robotics tasks such as obstacle avoidance [4]. In
this paper, we deploy neural networks for extracting useful
features from raw depth visual inputs, captured by onboard
depth cameras of the mobile robot.
We use depth inputs over RGB images because the visual
fidelity of the simulated color images are much worse than
that of the depth images. This bigger deviation of color
images from synthetic environments to real-world scenarios
makes it more challenging to transfer the model trained on
simulated images directly to the real world. The simulated
depth images, on the other hand, are more consistent with
the real domain and can greatly ease the transfer phase.
3) Learning-based navigation: Deep reinforcement
learning have recently gained much attention, and have
been successfully extended to learn autonomous navigation
from raw sensory inputs [5], [6], [7], [8].
In terms of socially compliant navigation, behavior
cloning methods are easy to deploy since they treat the
policy learning as a pure supervised learning task. Yet,
the learned model of behavior cloning completely ignores
the temporal correlation between samples in subsequent
frames, thus it cannot generalize well to scenarios that deviate too much from the training data. Agents are also able
to learn from expert demonstrations via an intermediate
step of learning the latent cost or the reward function [2],
[9], [10] through inverse reinforcement learning [11], [12].
Chen et al. [3] defined a complex reward to train the socially aware planning policies (e.g. passing by the right/left
side) through reinforcement learning. However, all of these
methods depend on accurate pedestrian information and
expensive sensors as mentioned before.
Generative adversarial imitation learning (GAIL) is an
effective alternative for learning from demonstrations [13].
InfoGAIL [14] successfully solved a simulated autonomous
driving task with raw visual input. However, only simulated
experiments are presented in those two papers.
In this paper, we effectively deploy behavior cloning for
learning an initial policy, Then, we apply GAIL on the
basis of this initial policy, to benefit the policy model by
taking the temporal correlations in the dataset into account.
Particularly, this paper presents the following contributions:
• We introduce an effective GAIL-based approach that
is able to learn and improve socially compliant navigation policies through raw depth inputs.
• We release a plugin for simulating pedestrians behaving in socially compliant manners, as well as a
dataset, where 10,000 socially compliant navigation
state-action pairs are recorded, based on the social
force model [1], in various social scenarios.
II. BACKGROUND
We consider a Markov decision process (MDP), where
an agent interacts with the environment through a sequence
of observations, actions and reward signals. The agent
executes an action at ∈ A at time step t from its current
state st ∈ S, according to its policy π : S → A. It then
receives a reward signal rt : S → R and transits to the next
state st+1 according to the dynamics of the environment.
We use πE to denote the expert policy.
A. Imitation learning
Imitation learning aims to learn policies directly from
experts, whose demonstrations are only provided in the
form of samples of trajectories. Main approaches for imitation learning can be categorized into behavior cloning (BC)
and inverse reinforcement learning (IRL) Behavior cloning
tackles this problem in a supervised manner, by directly
learning the mapping from the states in the recorded
trajectories to their corresponding labels: the expert policy.
It is conceptually simple and is able to work well with large
amounts of training data, but suffers from the compounding
error caused by covariate shift, due to its ignorance of
the temporal correlation in the recorded trajectories. Thus
pure behavior cloning tends to over-fit and is difficult to
generalize to unseen scenarios that deviate much from the
recorded dataset. Inverse reinforcement learning aims to
extract the latent reward or cost function under the optimal
expert demonstrations. It imitates by taking the underlying
MDP into account, learning from entire trajectories instead
of single frames. However, the requirement of running reinforcement learning in an inner loop makes IRL extremely
expensive to run [2], [9], [10].
B. Generitive adversarial imitation learning
Inspired by Generative adversarial networks (GAN) [15],
Ho and Ermon proposed generative adversarial imitation
learning (GAIL) [13], which surpasses the intermediate
step of learning a reward function, but is able to directly
learn a policy from expert demonstrations. In the GAIL
model, the generator πθ is forced to generate state-action
(S×A) pairs matching that from the expert demonstrations,
while the discriminator Dω learns to tell the generated
policy πθ (θ denotes the parameters of the generator of
the GAIL model) apart from the expert policy πE . The
objective of GAIL is to optimize the function below (H(π)
represents the causal entropy of the policy):
Eπθ [log(D(s, a))] + EπE [log(1 − D(s, a))] − λH(πθ )
(1)
Following this objective, the learning procedure of GAIL
interleaves between updating the parameters ω of the
discriminator Dω to maximize Eq. 1, and performing trust
region policy optimization (TRPO) [16] to minimize Eq. 1
with respect to θ, which parameterizes the policy generator πθ . Here, the discrimination scores of the generated
samples are regarded as costs (can be viewed as the
negative counterpart of rewards) of the state-action pairs
in the learning process of TRPO. As a state-of-the-art on
policy reinforcement learning method, TRPO constraints
the deviation of the updated policy from the original policy
according to their KL divergence.
We extend the GAIL framework with Wasserstein GAN
(WGAN) [17]. In WGAN, the classification formulation of
the discriminator network is substituted with a regression
problem. By eliminating the usage of softmax in traditional
GAN, WGAN directly maximizes the score of the real
data and minimizes the score of the generated data. It is
proved to improve the training stability of InfoGAIL [14]
The objective function of WGAN is:
Eπθ [D(s, a)] − EπE [D(s, a))]
(2)
C. Social force model
Helbing and Molnar [1] proposed the social force model
for pedestrian dynamics, which is broadly applied to socially compliant navigation scenarios. This model (Eq. 3)
computes the acceleration of pedestrians according to the
sum of various forces applied to them.
Fig. 2: The simulated environment used in this paper. Each
pedestrian is behaving under the social force model [1].
The Gazebo plugin is released1 .
d~vt
= F~desired + F~social + F~obs + F~fluct
dt
(3)
In Eq. 3, F~desired represents the desired force, which
drives the vehicle towards its navigation goal; F~social is the
social force to measure the influence of nearby pedestrians;
F~obs is the obstacle force to keep the agent from colliding
with static obstacles in the environment; and F~fluct is
the force caused by fluctuations, which comes from the
random variations of the environment and the stochastic
pedestrian behaviors. Among them, the desired force can be
simply represented as λ(pdesired − pt ), where pt and pdesired
represent the pose of the agent and the target respectively.
It is similar to the concept of prefered velocity in other
planning methods [18]. The calculation of the obstacle
force F~obs is based on [1] and we will not present the details
here. We note that both the estimation of F~obs and F~fluct are
omitted in this paper to prioritize the social aspects.
To successfully train our model in an interactive manner
with all considered social scenarios, an efficient simulation
environment that is capable of modeling pedestrians in socially acceptable behaviors is crucial. The existing robotics
simulation environments for pedestrians typically do not
come with embedded social force model, that the simulated
pedestrians are not able to behave in social compliant manner. As an additional contribution of this paper, we release a
plugin1 for simulating socially compliant pedestrians under
the framework of Gazebo, as is shown in Fig. 2.
III. M ETHODS
In this paper, we formulate the problem of navigation
in pedestrian environment as a Markov decision process
(MDP). The state st is composed of the depth image xt and
the force towards the desired target F~tdesired . Each action at
corresponds to a moving command ut to be executed by the
mobile robot. Two parameterised networks, the generator
(policy) network πθ and the discriminator network Dω are
updated in an interleaving manner in the training phase.
a. Passing (easy)
b. Overtaking (easy)
c. Crossing (easy)
d. Passing1 (hard)
e. Passing2 (hard)
f. Crossing (hard)
Fig. 3: Six basic scenarios considered in this paper to compare the navigation performance of the behavior cloning
policy and the GAIL policy. We test the algorithm on a
mobile robot from the view of the red agent.
A. Behavior cloning
We initialize the policy generator network πθ with the
weights pre-trained via behavior cloning. To collect the
training dataset and test the trained model, we identify the
following most commonly encountered and typical social
scenarios [3], which are shown in Fig. 3. Specifically, we
collect trajectories from three relatively easy scenarios:
(a) passing, (b) overtaking, (c) crossing, as well as three
difficult ones: (d) passing out of a group of pedestrians, (e)
passing between a group of pedestrians, and (f) crossing
with a group of pedestrians. The differences between scenarios (d) and (e) are intended to motivate socially aware
navigation strategies [2] that pedestrians walking close to
each other should be regarded as a group, and the agent is
expected to decide whether there is enough space between
them for it to make a crossing. In our socially compliant
pedestrian simulator, we collect data by mounting a depth
sensor onto one of the pedestrians, to the height matching
that of real-world setups. Then, the social force model, as
described by Eq. 3, is used to label each incoming depth
image with their corresponding social force. We note that
only the pedestrians within the field of view (FOV) of the
depth sensor (with a sensing range of 3.5m, and a vertical
sensing angles of ±35◦ ), are considered in the social force
calculation. Desired force, represented by the normalized
direction vector to the navigation target, are also collected
as another input source for the model. We assign target
locations for each episode, ensuring that the agent would
have to encounter with pedestrians in the environment
before reaching its targets.
As an additional contribution of this paper, we release
the collected dataset2 for further benchmarking for the
literature. The first person view depth image, RGB image,
social force, desired force and velocity of the agent are
all recorded and included in the dataset. The final dataset
contains 10,000 state-action pairs collected from social
scenario shown in Fig. 3. The initial configuration (the
1 https://github.com/onlytailei/gym
2 https://ram-lab.com/file/tailei/tai
ped sim
18icra dataset.html
State ( ) | 400 x 250
Algorithm 1 Asynchronous GAIL
Conv | 32 | 7x7 | 4
ResNet-50
(pre-trained)
Conv | 32 | 5x5 | 4
Conv | 32 | 20x20 | 30
Conv | 32 | 3x3 | 2
Concatenate
Conv | 32 | 3x3 | 1
Dense | 256
Dense | 128
State (
)|2
Merge
) | 2 Action (
Social Force (
Collect expert trajectories TE .
Initialize πθ with the behavior cloning policy θ0 .
Randomly initialize Dω with ω0 .
for iteration i = 0, 1, . . . do
for iteration k = 1, K do
Randomly choose a social scenario simulation.
Sample one trajectory: Tik ∼ πθi
end for
Ascending gradients of ωi on mini-batches (κi ∼ Ti ,
κE ∼ TE ):
∆ω = Eκi [∇ω Dω (s, a)] − EκE [∇ω Dω (s, a)]
Update to ωi+1 after clipping the weights to
(0.01, 0.01).
Update θi to θi+1 with the cost function Dωı+1 (s, a)
under the TRPO rule.
end for
)|2
the generalization ability of the model and prevent overfitting [20]. This setup also makes it possible to transfer our
trained model naturally to different motion planners, as the
social force prediction part can stay intact regardless of the
mobile platforms in use.
(a) Generator (policy) network.
State ( ) | 400 x 250
action prediction
social force prediction
Conv | 32 | 3x3 | 4
Conv | 64 | 3x3 | 4
Action (
)|2
Conv | 128 | 3x3 | 4
State (
)|2
Merge
Dense | 256
Dense | 128
Score (
evaluation loss
0.125
0.100
0.075
0.050
0.025
0.000
)|1
(b) Discriminator network.
Fig. 4: Network structures for the generator and the discriminator network. Every convolutional layer is represented by its type, channel size, kernel size and stride size.
Other layers are represented by their types and dimonsions.
position and velocity information) of the pedestrians and
the agent are perturbed by random noise for each episode.
The model for behavior cloning is the same as the
generator network of our GAIL model, which is depicted in
Fig. 4a. The model takes the depth image and the desired
force as input. The depth image is passed through three
steams of feature extraction layers, to benefit the resulting
representation from the residual learning of skip connection
structures [19]. The extracted features are merged with
the desired force and then used to predict the expert
policies, as well as for performing a subtask: social force
prediction. Learning on multiple tasks using the same set
of features enforces the extracted features to be an effective
and compact representation, which can greatly improve
0
50
100
150
training epochs
200
Fig. 5: Average loss ± one standard deviation for supervised behavior cloning of social force and action predition
in 200 epochs. It is used as the initial policy.
B. Algorithm
The original GAIL [13] requires training one model for
each specific task. The trajectories sampled from the environment are quite similar across episodes, which greatly
limits the generalization of the trained model. Inspired
by various asynchronous methods in the deep reinforcement learning literature (Async-DRL) [21], we propose
to perform generative adversarial imitation learning in a
modified training procedure. Specifically, during training,
our model interacts with several different social scenario
simulations in an interleaving manner. The resulting model
thus converges to a generalized policy across all social
scenarios and is able to perform well on all considered
tasks. We note that our proposed method differs from
the other Async-DRL methods in that, those previous
approaches have multiple instances of the environment of
a. Passing (easy)
b. Overtaking (easy)
c. Crossing (easy)
d. Passing1 (hard)
e. Passing2 (hard)
f. Crossing (hard)
2m
Start
End
Fig. 6: Navigation trajectories of test episodes. In every scenario, the upper part is the performance of the model purely
trained from behavior cloning and the lower part is the performance of the model optimized by GAIL. The thickness of
the path increases from the starting position of the agent to its ending point. Red paths are the trajectories executed by
the mobile robot under the trained policy. The other agents are moving under the social force model [1].
the same task, while for our approach, each instance of the
environment corresponds to a different social task.
We describe our approach in detail in Algorithm 1. We
initialize the generator (policy) network of our GAIL model
by the pre-trained weights θ0 from behavior cloning. The
discriminator network is initialized randomly with ω0 . In
every training step, the trajectories τi are sampled from different social scenario simulations. The sampled trajectories
τi and the sampled expert trajectories τE are then fed into
the discriminator network. The weight of the discriminator
network is clipped between (−0.01, 0.01) to update to
ωi+1 , to fulfill the constraint of WGAN [17]. After that, θi
is updated to θi+1 , by following the TRPO rule for updating
the policy parameters. The discrimination score of the state
action pair Dωi+1 (s, a) are used as the cost for policy
gradient optimization. As [14], in the trajectories sampling
procedure, an augmented cost punishing the collision with
pedestrians is added, in case that demonstrations under
social force model are not optimal enough.
As mentioned before, the generator model has the same
architecture as the model used for conducting behavior
cloning, which is shown in Fig. 4a. The discriminator,
on the other hand, contains only one stream for feature
extraction, the resulting features are concatenated with the
desired force of the corresponding step. Then, they are
merged with the selected action, which is then fed to
several fully-connected layers to estimate a score to tell
apart generated policies from expert policies. The details
of the discriminator architecture are shown in Fig. 4b.
IV. E XPERIMENTS
We begin our experiment by training the behavior
cloning model. First, we collect training data from different
social scenario simulations, as shown in Fig. 2. To add to
the variations of the training dataset, we collect sample
trajectories from the perspective of both the agent (shown
in red in Fig. 3) and the pedestrians (shown in blue in Fig.
3). The social force model, as is described in Sec. II-C, is
8
Passing
(easy)
Overtaking
(easy)
Crossing
(easy)
Passing1
(hard)
Passing2
(hard)
Crossing
(hard)
1.6
Passing
(easy)
Overtaking
(easy)
Crossing
(easy)
Passing1
(hard)
Passing2
(hard)
Crossing
(hard)
5
7
9
11
7
1.4
5
average travel-time ratio
average min-dist ratio
6
BC
GAIL
4
3
1.2
1.0
2
0.8
1
0
1
2
3
4
5
6
7
8
9
10
11
12
(a) min-dist: average minimum distance ratio
0.6
1
2
3
4
6
8
10
12
(b) travel time: average travel time ratio
Fig. 7: Average minimum distance to pedestrians and the average travel time of 10 episodes in all test scenarios. Green
boxes represent results from the BC policy, and the blue ones show results from the GAIL policy. In every scenario, all
the statistics are normalized by dividing the mean of the BC policy.
utilized to generate the expert policy. In total, 10,000 stateaction pairs are collected, containing trajectories from all
the considered social scenarios, with randomized starting
configurations. 2000 samples are separated for evaluating
the pre-trained model.
To train the behavior cloning model, we use RMSprop
with a learning rate of 1e−4, and decay it by a factor of 0.9
after every 20 epochs. We experiment with different weight
ratios for the two supervised tasks (social force prediction,
action prediction, as shown in Fig. 4a). We can observe
from Fig. 5 that those tasks are effectively learned.
Then we start the training of our model. As mentioned
before, we initialize the generator network of the GAIL
policy with the pre-trained behavior cloning model. In
every training step, 3 trajectories are sampled from 6
randomly chosen scenarios as Algorithm 1, which add
up to a total number of 1000 samples approximately.
This training strategy is highly effective for learning a
generalized policy over all considered tasks, and greatly
improves the generalization capability of our trained model
across the most commonly encountered social scenes.
We train GAIL for 300 iterations, in a duration of 10
hours, on a Tesla P100 graphics processing unit. The discriminator is optimized through RMSprop with a learning
rate of 5e−5. The generator (policy) network is optimized
following the TRPO rule.
In the following, we present both quantitative and qualitative results, which clearly show the effectiveness our
approach. In particular, we can observe that the pre-trained
policy via behavior cloning is only able to perform some
reactive actions, since it completely ignores the temporal
correlations in the sample trajectories, thus is only able to
make frame-wise decisions; while the improved policy via
GAIL is able to drive the agent to behave in a sociallycompliant manner, due to its underlying formulation that
takes the underlying MDP of the trajectories into account,
which enables it to make planning decisions on the horizon
of a whole trajectory.
A. Evaluation in simulated environments
To ensure a fair comparison between test runs, we
remove all the randomness that is included in the training
setups. A simulated Turtlebot3 waffle is used to execute 10
test episodes in each scenario, under both BC policy and
GAIL policy, taking the depth image and the desired force
as input. The other pedestrian agents in the environment
are navigating under the social force model, taking in also
the social force from the mobile robot.
1) Qualitative evaluation: Both the trajectories of BC
policy and GAIL policy are shown in Fig. 6. For each
scenario, we show the BC policy performance in the upper
part and the GAIL policy in the lower part. We increase
the thickness of the trajectory from the starting point to
the ending point. The trajectories of the robot are shown in
red, the trajectories of pedestrians are shown blue. We can
clearly observe that agent performing under the optimized
policy by GAIL is able to navigate in a more sociallycompliant manner in all considered scenarios. In particular,
in scenarios (a), (b), (c) and (f), the GAIL policy guides
the robot further away from the pedestrians. Also, the
GAIL policy successfully guides the mobile robot to pass
in between (d) and travel out of (e) a group of pedestrians,
which the BC policy fails to accomplish.
We note that limited by the small FOV, the robot is only
able to perceive the pedestrians when they are fairly nearby,
especially in the crossing scenario (c). This leads to some
sudden turnings shown in the trajectories.
2) Quantitative evaluation: We choose two metrics to
perform quantitative evaluation: (1) min-dist: the minimal
distance from the robot to other pedestrians in one episode
and (2) travel-time: the time taken by the robot to travel
from the starting location to the target. Those statistics are
shown in Fig. 7.
From Fig. 7, we can observe that our GAIL model
performs much better than the behavior cloning baseline.
Fig. 7a shows that it clearly converges to a much safer
navigation strategy as its average min-dist to pedestrians is
larger than that of the baseline in all six social scenarios;
as for the travel-time, it plans more efficient paths in most
scenarios as can be seen in Fig. 7b, but performs slightly
slower than that of behavior cloning in two easy scenarios.
Those two statistics show that the GAIL model learns to
navigate in a much more socially acceptable way, and is
able to plan paths that are both safe and efficient.
B. Real world experiments
We also conduct real-world experiments to test the
performance of our approach in realistic scenarios. We use
a Turtlebot waffle platform, which is shown in Fig. 8. It
navigates autonomously in an indoor office environment
as shown in Fig. 1, under the improved GAIL policy. A
low-cost laser range sensor is used to localize the robot in
the environment. Navigation targets are chosen randomly
on the collision-free areas of the map. The depth image,
captured by an obboard Intel Realsense R200, is cropped to
400 × 250 and fed into the trained model. A Nvidia Jetson
TX2 is mounted for real-time neural network processing.
The model control cycle runs in real-time at 15Hz. A video
showing the real-world performance of our trained policy
can be found in https://goo.gl/42yf6f.
Learned policy
desired force
Desired force
target
Target
desired force
Output velocity
Depth input
Turtlebot waffle
Fig. 8: Real world experiments in an indoor office environment, through a Turtlebot waffle platform by taking raw
depth images and desired forces as inputs.
V. C ONCLUSION
We presented an approach for agents to learn to navigate
in a socially compliant manner, from raw depth images.
Our algorithm is able to guide the agent to perform socially
compliant behaviors, as well as plan efficient paths to
reach its goal location. We validated our approach in both
simulated and real world experiments; moreover, we release a plugin for simulating pedestrians under social force
model, as well as a dataset collected from our simulation
environment.
The performance of the GAIL policy in real-world
environments is influenced by the limited maximum speed
and the FOV of the Turtlebot waffle. We leave it as
future work to evaluate the proposed algorithm on more
compatible platforms. Recent work [22] of real-to-sim
domain adaptation for visual control also makes it possible
to tackle this problem through raw RGB images.
R EFERENCES
[1] D. Helbing and P. Molnar, “Social force model for pedestrian
dynamics,” Physical review E, vol. 51, no. 5, p. 4282, 1995.
[2] M. Pfeiffer, U. Schwesinger, H. Sommer, E. Galceran, and R. Siegwart, “Predicting actions to act predictably: Cooperative partial
motion planning with maximum entropy models,” in 2016 IEEE/RSJ
International Conference on Intelligent Robots and Systems (IROS),
2016, pp. 2096–2101.
[3] Y. F. Chen, M. Everett, M. Liu, and J. P. How, “Socially aware
motion planning with deep reinforcement learning,” in IROS, Sept
2017, pp. 1343–1350.
[4] L. Tai, S. Li, and M. Liu, “A deep-network solution towards
model-less obstacle avoidance,” in 2016 IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS), Oct 2016,
pp. 2759–2764.
[5] P. Mirowski, R. Pascanu, F. Viola, H. Soyer, A. Ballard, A. Banino,
M. Denil, R. Goroshin, L. Sifre, K. Kavukcuoglu, et al., “Learning
to navigate in complex environments,” in ICLR, 2017.
[6] L. Tai, G. Paolo, and M. Liu, “Virtual-to-real deep reinforcement
learning: Continuous control of mobile robots for mapless navigation,” in 2017 IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS), Sep 2017, pp. 31–36.
[7] J. Zhang, J. T. Springenberg, J. Boedecker, and W. Burgard, “Deep
reinforcement learning with successor features for navigation across
similar environments,” in 2017 IEEE/RSJ International Conference
on Intelligent Robots and Systems (IROS), Sep 2017, pp. 2371–2378.
[8] J. Zhang, L. Tai, J. Boedecker, W. Burgard, and M. Liu, “Neural
slam: Learning to explore with external memory,” arXiv preprint
arXiv:1706.09520, 2017.
[9] H. Kretzschmar, M. Spies, C. Sprunk, and W. Burgard, “Socially
compliant mobile robot navigation via inverse reinforcement learning,” The International Journal of Robotics Research, vol. 35,
no. 11, pp. 1289–1307, 2016.
[10] B. Okal and K. O. Arras, “Learning socially normative robot
navigation behaviors with bayesian inverse reinforcement learning,”
in ICRA, May 2016, pp. 2889–2895.
[11] A. Y. Ng, S. J. Russell, et al., “Algorithms for inverse reinforcement
learning.” in ICML, 2000, pp. 663–670.
[12] P. Abbeel and A. Y. Ng, “Apprenticeship learning via inverse
reinforcement learning,” in ICML, 2004, p. 1.
[13] J. Ho and S. Ermon, “Generative adversarial imitation learning,” in
NIPS, 2016, pp. 4565–4573.
[14] Y. Li, J. Song, and S. Ermon, “Infogail: Interpretable imitation
learning from visual demonstrations,” in NIPS, 2017.
[15] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley,
S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial nets,”
in NIPS, 2014.
[16] J. Schulman, S. Levine, P. Abbeel, M. Jordan, and P. Moritz, “Trust
region policy optimization,” in ICML, 2015, pp. 1889–1897.
[17] M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein gan,” arXiv
preprint arXiv:1701.07875, 2017.
[18] J. Van den Berg, M. Lin, and D. Manocha, “Reciprocal velocity
obstacles for real-time multi-agent navigation,” in 2008 IEEE International Conference on Robotics and Automation, May 2008, pp.
1928–1935.
[19] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for
image recognition,” in CVPR, June 2016.
[20] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. MIT
Press, 2016, vol. 1.
[21] V. Mnih, A. P. Badia, M. Mirza, A. Graves, T. Lillicrap, T. Harley,
D. Silver, and K. Kavukcuoglu, “Asynchronous methods for deep
reinforcement learning,” in ICML, 2016.
[22] J. Zhang, L. Tai, Y. Xiong, M. Liu, J. Boedecker, and W. Burgard,
“Vr goggles for robots: Real-to-sim domain adaptation for visual
control,” arXiv preprint arXiv:1802.00265, 2018.
| 2cs.AI
|
arXiv:1705.07962v2 [cs.LG] 19 Sep 2017
pix2code: Generating Code from a Graphical User
Interface Screenshot
Tony Beltramelli
UIzard Technologies
Copenhagen, Denmark
[email protected]
Abstract
Transforming a graphical user interface screenshot created by a designer into computer code is a typical task conducted by a developer in order to build customized
software, websites, and mobile applications. In this paper, we show that deep
learning methods can be leveraged to train a model end-to-end to automatically
generate code from a single input image with over 77% of accuracy for three
different platforms (i.e. iOS, Android and web-based technologies).
1
Introduction
The process of implementing client-side software based on a Graphical User Interface (GUI) mockup
created by a designer is the responsibility of developers. Implementing GUI code is, however,
time-consuming and prevent developers from dedicating the majority of their time implementing the
actual functionality and logic of the software they are building. Moreover, the computer languages
used to implement such GUIs are specific to each target runtime system; thus resulting in tedious
and repetitive work when the software being built is expected to run on multiple platforms using
native technologies. In this paper, we describe a model trained end-to-end with stochastic gradient
descent to simultaneously learns to model sequences and spatio-temporal visual features to generate
variable-length strings of tokens from a single GUI image as input.
Our first contribution is pix2code, a novel approach based on Convolutional and Recurrent Neural
Networks allowing the generation of computer tokens from a single GUI screenshot as input. That
is, no engineered feature extraction pipeline nor expert heuristics was designed to process the input
data; our model learns from the pixel values of the input image alone. Our experiments demonstrate
the effectiveness of our method for generating computer code for various platforms (i.e. iOS and
Android native mobile interfaces, and multi-platform web-based HTML/CSS interfaces) without the
need for any change or specific tuning to the model. In fact, pix2code can be used as such to support
different target languages simply by being trained on a different dataset. A video demonstrating our
system is available online1 .
Our second contribution is the release of our synthesized datasets consisting of both GUI screenshots
and associated source code for three different platforms. Our datasets and our pix2code implemention
are publicly available2 to foster future research.
2
Related Work
The automatic generation of programs using machine learning techniques is a relatively new field of
research and program synthesis in a human-readable format have only been addressed very recently.
1
2
https://uizard.io/research#pix2code
https://github.com/tonybeltramelli/pix2code
A recent example is DeepCoder [2], a system able to generate computer programs by leveraging
statistical predictions to augment traditional search techniques. In another work by Gaunt et al. [5],
the generation of source code is enabled by learning the relationships between input-output examples
via differentiable interpreters. Furthermore, Ling et al. [12] recently demonstrated program synthesis
from a mixed natural language and structured program specification as input. It is important to note
that most of these methods rely on Domain Specific Languages (DSLs); computer languages (e.g.
markup languages, programming languages, modeling languages) that are designed for a specialized
domain but are typically more restrictive than full-featured computer languages. Using DSLs thus
limit the complexity of the programming language that needs to be modeled and reduce the size of
the search space.
Although the generation of computer programs is an active research field as suggested by these
breakthroughs, program generation from visual inputs is still a nearly unexplored research area. The
closest related work is a method developed by Nguyen et al. [14] to reverse-engineer Android user
interfaces from screenshots. However, their method relies entirely on engineered heuristics requiring
expert knowledge of the domain to be implemented successfully. Our paper is, to the best of our
knowledge, the first work attempting to address the problem of user interface code generation from
visual inputs by leveraging machine learning to learn latent variables instead of engineering complex
heuristics.
In order to exploit the graphical nature of our input, we can borrow methods from the computer vision
literature. In fact, an important number of research [21, 4, 10, 22] have addressed the problem of
image captioning with impressive results; showing that deep neural networks are able to learn latent
variables describing objects in an image and their relationships with corresponding variable-length
textual descriptions. All these methods rely on two main components. First, a Convolutional Neural
Network (CNN) performing unsupervised feature learning mapping the raw input image to a learned
representation. Second, a Recurrent Neural Network (RNN) performing language modeling on the
textual description associated with the input picture. These approaches have the advantage of being
differentiable end-to-end, thus allowing the use of gradient descent for optimization.
(b) Sampling
(a) Training
Figure 1: Overview of the pix2code model architecture. During training, the GUI image is encoded by
a CNN-based vision model; the context (i.e. a sequence of one-hot encoded tokens corresponding to
DSL code) is encoded by a language model consisting of a stack of LSTM layers. The two resulting
feature vectors are then concatenated and fed into a second stack of LSTM layers acting as a decoder.
Finally, a softmax layer is used to sample one token at a time; the output size of the softmax layer
corresponding to the DSL vocabulary size. Given an image and a sequence of tokens, the model (i.e.
contained in the gray box) is differentiable and can thus be optimized end-to-end through gradient
descent to predict the next token in the sequence. During sampling, the input context is updated for
each prediction to contain the last predicted token. The resulting sequence of DSL tokens is compiled
to the desired target language using traditional compiler design techniques.
3
pix2code
The task of generating computer code written in a given programming language from a GUI screenshot
can be compared to the task of generating English textual descriptions from a scene photography. In
both scenarios, we want to produce a variable-length strings of tokens from pixel values. We can
thus divide our problem into three sub-problems. First, a computer vision problem of understanding
the given scene (i.e. in this case, the GUI image) and inferring the objects present, their identities,
2
positions, and poses (i.e. buttons, labels, element containers). Second, a language modeling problem
of understanding text (i.e. in this case, computer code) and generating syntactically and semantically
correct samples. Finally, the last challenge is to use the solutions to both previous sub-problems by
exploiting the latent variables inferred from scene understanding to generate corresponding textual
descriptions (i.e. computer code rather than English) of the objects represented by these variables.
3.1
Vision Model
CNNs are currently the method of choice to solve a wide range of vision problems thanks to their
topology allowing them to learn rich latent representations from the images they are trained on
[16, 11]. We used a CNN to perform unsupervised feature learning by mapping an input image to a
learned fixed-length vector; thus acting as an encoder as shown in Figure 1.
The input images are initially re-sized to 256 × 256 pixels (the aspect ratio is not preserved) and the
pixel values are normalized before to be fed in the CNN. No further pre-processing is performed. To
encode each input image to a fixed-size output vector, we exclusively used small 3 × 3 receptive
fields which are convolved with stride 1 as used by Simonyan and Zisserman for VGGNet [18].
These operations are applied twice before to down-sample with max-pooling. The width of the first
convolutional layer is 32, followed by a layer of width 64, and finally width 128. Two fully connected
layers of size 1024 applying the rectified linear unit activation complete the vision model.
(b) Code describing the GUI written in our DSL
(a) iOS GUI screenshot
Figure 2: An example of a native iOS GUI written in our markup-like DSL.
3.2
Language Model
We designed a simple lightweight DSL to describe GUIs as illustrated in Figure 2. In this work we
are only interested in the GUI layout, the different graphical components, and their relationships;
thus the actual textual value of the labels is ignored. Additionally to reducing the size of the search
space, the DSL simplicity also reduces the size of the vocabulary (i.e. the total number of tokens
supported by the DSL). As a result, our language model can perform token-level language modeling
with a discrete input by using one-hot encoded vectors; eliminating the need for word embedding
techniques such as word2vec [13] that can result in costly computations.
In most programming languages and markup languages, an element is declared with an opening
token; if children elements or instructions are contained within a block, a closing token is usually
needed for the interpreter or the compiler. In such a scenario where the number of children elements
contained in a parent element is variable, it is important to model long-term dependencies to be
able to close a block that has been opened. Traditional RNN architectures suffer from vanishing
and exploding gradients preventing them from being able to model such relationships between data
points spread out in time series (i.e. in this case tokens spread out in a sequence). Hochreiter and
Schmidhuber proposed the Long Short-Term Memory (LSTM) neural architecture in order to address
this very problem [9]. The different LSTM gate outputs can be computed as follows:
3
it
ft
ot
ct
ht
= φ(Wix xt + Wiy ht−1 + bi )
= φ(Wf x xt + Wf y ht−1 + bf )
= φ(Wox xt + Woy ht−1 + bo )
= ft • ct−1 + it • σ(Wcx xt + Wcy ht−1 + bc )
= ot • σ(ct )
(1)
(2)
(3)
(4)
(5)
With W the matrices of weights, xt the new input vector at time t, ht−1 the previously produced
output vector, ct−1 the previously produced cell state’s output, b the biases, and φ and σ the activation functions sigmoid and hyperbolic tangent, respectively. The cell state c learns to memorize
information by using a recursive connection as done in traditional RNN cells. The input gate i is
used to control the error flow on the inputs of cell state c to avoid input weight conflicts that occur in
traditional RNN because the same weight has to be used for both storing certain inputs and ignoring
others. The output gate o controls the error flow from the outputs of the cell state c to prevent output
weight conflicts that happen in standard RNN because the same weight has to be used for both
retrieving information and not retrieving others. The LSTM memory block can thus use i to decide
when to write information in c and use o to decide when to read information from c. We used the
LSTM variant proposed by Gers and Schmidhuber [6] with a forget gate f to reset memory and help
the network model continuous sequences.
3.3
Decoder
Our model is trained in a supervised learning manner by feeding an image I and a contextual sequence
X of T tokens xt , t ∈ {0 . . . T − 1} as inputs; and the token xT as the target label. As shown on
Figure 1, a CNN-based vision model encodes the input image I into a vectorial representation p. The
input token xt is encoded by an LSTM-based language model into an intermediary representation qt
allowing the model to focus more on certain tokens and less on others [8]. This first language model
is implemented as a stack of two LSTM layers with 128 cells each. The vision-encoded vector p and
the language-encoded vector qt are concatenated into a single feature vector rt which is then fed into
a second LSTM-based model decoding the representations learned by both the vision model and the
language model. The decoder thus learns to model the relationship between objects present in the
input GUI image and the associated tokens present in the DSL code. Our decoder is implemented as
a stack of two LSTM layers with 512 cells each. This architecture can be expressed mathematically
as follows:
p = CN N (I)
qt = LST M (xt )
rt = (q, pt )
(6)
(7)
(8)
yt = sof tmax(LST M 0 (rt ))
xt+1 = yt
(9)
(10)
This architecture allows the whole pix2code model to be optimized end-to-end with gradient descent
to predict a token at a time after it has seen both the image as well as the preceding tokens in the
sequence. The discrete nature of the output (i.e. fixed-sized vocabulary of tokens in the DSL) allows
us to reduce the task to a classification problem. That is, the output layer of our model has the same
number of cells as the vocabulary size; thus generating a probability distribution of the candidate
tokens at each time step allowing the use of a softmax layer to perform multi-class classification.
3.4
Training
The length T of the sequences used for training is important to model long-term dependencies;
for example to be able to close a block of code that has been opened. After conducting empirical
experiments, the DSL input files used for training were segmented with a sliding window of size
48; in other words, we unroll the recurrent neural network for 48 steps. This was found to be a
satisfactory trade-off between long-term dependencies learning and computational cost. For every
4
token in the input DSL file, the model is therefore fed with both an input image and a contextual
sequence of T = 48 tokens. While the context (i.e. sequence of tokens) used for training is updated
at each time step (i.e. each token) by sliding the window, the very same input image I is reused
for samples associated with the same GUI. The special tokens < ST ART > and < EN D > are
used to respectively prefix and suffix the DSL files similarly to the method used by Karpathy and
Fei-Fei [10]. Training is performed by computing the partial derivatives of the loss with respect to
the network weights calculated with backpropagation to minimize the multiclass log loss:
L(I, X) = −
T
X
xt+1 log(yt )
(11)
t=1
With xt+1 the expected token, and yt the predicted token. The model is optimized end-to-end hence
the loss L is minimized with regard to all the parameters including all layers in the CNN-based vision
model and all layers in both LSTM-based models. Training with the RMSProp algorithm [20] gave
the best results with a learning rate set to 1e − 4 and by clipping the output gradient to the range
[−1.0, 1.0] to cope with numerical instability [8]. To prevent overfitting, a dropout regularization
[19] set to 25% is applied to the vision model after each max-pooling operation and at 30% after
each fully-connected layer. In the LSTM-based models, dropout is set to 10% and only applied to
the non-recurrent connections [24]. Our model was trained with mini-batches of 64 image-sequence
pairs.
Dataset type
iOS UI (Storyboard)
Android UI (XML)
web-based UI (HTML/CSS)
3.5
Table 1: Dataset statistics.
Training set
Synthesizable
Instances Samples
26 × 105
1500
93672
21 × 106
1500
85756
31 × 104
1500
143850
Test set
Instances Samples
250
15984
250
14265
250
24108
Sampling
To generate DSL code, we feed the GUI image I and a contextual sequence X of T = 48 tokens
where tokens xt . . . xT −1 are initially set empty and the last token of the sequence xT is set to the
special < ST ART > token. The predicted token yt is then used to update the next sequence of
contextual tokens. That is, xt . . . xT −1 are set to xt+1 . . . xT (xt is thus discarded), with xT set to
yt . The process is repeated until the token < EN D > is generated by the model. The generated
DSL token sequence can then be compiled with traditional compilation methods to the desired target
language.
(b) Micro-average ROC curves
(a) pix2code training loss
Figure 3: Training loss on different datasets and ROC curves calculated during sampling with the
model trained for 10 epochs.
5
4
Experiments
Access to consequent datasets is a typical bottleneck when training deep neural networks. To the best
of our knowledge, no dataset consisting of both GUI screenshots and source code was available at the
time this paper was written. As a consequence, we synthesized our own data resulting in the three
datasets described in Table 1. The column Synthesizable refers to the maximum number of unique
GUI configuration that can be synthesized using our stochastic user interface generator. The columns
Instances refers to the number of synthesized (GUI screenshot, GUI code) file pairs. The columns
Samples refers to the number of distinct image-sequence pairs. In fact, training and sampling are
done one token at a time by feeding the model with an image and a sequence of tokens obtained with
a sliding window of fixed size T . The total number of training samples thus depends on the total
number of tokens written in the DSL files and the size of the sliding window. Our stochastic user
interface generator is designed to synthesize GUIs written in our DSL which is then compiled to
the desired target language to be rendered. Using data synthesis also allows us to demonstrate the
capability of our model to generate computer code for three different platforms.
Table 2: Experiments results reported for the test sets described in Table 1.
Error (%)
Dataset type
greedy search beam search 3 beam search 5
iOS UI (Storyboard)
22.73
25.22
23.94
Android UI (XML)
22.34
23.58
40.93
web-based UI (HTML/CSS)
12.14
11.01
22.35
Our model has around 109 × 106 parameters to optimize and all experiments are performed with the
same model with no specific tuning; only the training datasets differ as shown on Figure 3. Code
generation is performed with both greedy search and beam search to find the tokens that maximize
the classification probability. To evaluate the quality of the generated output, the classification error
is computed for each sampled DSL token and averaged over the whole test dataset. The length
difference between the generated and the expected token sequences is also counted as error. The
results can be seen on Table 2.
Figures 4, 5, and 6 show samples consisting of input GUIs (i.e. ground truth), and output GUIs
generated by a trained pix2code model. It is important to remember that the actual textual value of the
labels is ignored and that both our data synthesis algorithm and our DSL compiler assign randomly
generated text to the labels. Despite occasional problems to select the right color or the right style
for specific GUI elements and some difficulties modelling GUIs consisting of long lists of graphical
components, our model is generally able to learn the GUI layout in a satisfying manner and can
preserve the hierarchical structure of the graphical elements.
(a) Groundtruth GUI 1
(b) Generated GUI 1
(c) Groundtruth GUI 2
Figure 4: Experiment samples for the iOS GUI dataset.
6
(d) Generated GUI 2
5
Conclusion and Discussions
In this paper, we presented pix2code, a novel method to generate computer code given a single GUI
image as input. While our work demonstrates the potential of such a system to automate the process
of implementing GUIs, we only scratched the surface of what is possible. Our model consists of
relatively few parameters and was trained on a relatively small dataset. The quality of the generated
code could be drastically improved by training a bigger model on significantly more data for an
extended number of epochs. Implementing a now-standard attention mechanism [1, 22] could further
improve the quality of the generated code.
Using one-hot encoding does not provide any useful information about the relationships between the
tokens since the method simply assigns an arbitrary vectorial representation to each token. Therefore,
pre-training the language model to learn vectorial representations would allow the relationships
between tokens in the DSL to be inferred (i.e. learning word embeddings such as word2vec [13]) and
as a result alleviate semantical error in the generated code. Furthermore, one-hot encoding does not
scale to very big vocabulary and thus restrict the number of symbols that the DSL can support.
(a) Groundtruth GUI 5
(b) Generated GUI 5
(c) Groundtruth GUI 6
(d) Generated GUI 6
Figure 5: Experiment samples from the web-based GUI dataset.
Generative Adversarial Networks GANs [7] have shown to be extremely powerful at generating
images and sequences [23, 15, 25, 17, 3]. Applying such techniques to the problem of generating
computer code from an input image is so far an unexplored research area. GANs could potentially be
used as a standalone method to generate code or could be used in combination with our pix2code
model to fine-tune results.
A major drawback of deep neural networks is the need for a lot of training data for the resulting
model to generalize well on new unseen examples. One of the significant advantages of the method
we described in this paper is that there is no need for human-labelled data. In fact, the network can
model the relationships between graphical components and associated tokens by simply being trained
on image-sequence pairs. Although we used data synthesis in our paper partly to demonstrate the
capability of our method to generate GUI code for various platforms; data synthesis might not be
needed at all if one wants to focus only on web-based GUIs. In fact, one could imagine crawling the
World Wide Web to collect a dataset of HTML/CSS code associated with screenshots of rendered
websites. Considering the large number of web pages already available online and the fact that new
websites are created every day, the web could theoretically supply a virtually unlimited amount of
7
training data; potentially allowing deep learning methods to fully automate the implementation of
web-based GUIs.
(a) Groundtruth GUI 3
(b) Generated GUI 3
(c) Groundtruth GUI 4
(d) Generated GUI 4
Figure 6: Experiment samples from the Android GUI dataset.
References
[1] D. Bahdanau, K. Cho, and Y. Bengio. Neural machine translation by jointly learning to align
and translate. arXiv preprint arXiv:1409.0473, 2014.
[2] M. Balog, A. L. Gaunt, M. Brockschmidt, S. Nowozin, and D. Tarlow. Deepcoder: Learning to
write programs. arXiv preprint arXiv:1611.01989, 2016.
[3] B. Dai, D. Lin, R. Urtasun, and S. Fidler. Towards diverse and natural image descriptions via a
conditional gan. arXiv preprint arXiv:1703.06029, 2017.
[4] J. Donahue, L. Anne Hendricks, S. Guadarrama, M. Rohrbach, S. Venugopalan, K. Saenko, and
T. Darrell. Long-term recurrent convolutional networks for visual recognition and description.
In Proceedings of the IEEE conference on computer vision and pattern recognition, pages
2625–2634, 2015.
[5] A. L. Gaunt, M. Brockschmidt, R. Singh, N. Kushman, P. Kohli, J. Taylor, and D. Tarlow. Terpret:
A probabilistic programming language for program induction. arXiv preprint arXiv:1608.04428,
2016.
[6] F. A. Gers, J. Schmidhuber, and F. Cummins. Learning to forget: Continual prediction with
lstm. Neural computation, 12(10):2451–2471, 2000.
[7] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and
Y. Bengio. Generative adversarial nets. In Advances in neural information processing systems,
pages 2672–2680, 2014.
[8] A. Graves. Generating sequences with recurrent neural networks.
arXiv:1308.0850, 2013.
arXiv preprint
[9] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–
1780, 1997.
[10] A. Karpathy and L. Fei-Fei. Deep visual-semantic alignments for generating image descriptions.
In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages
3128–3137, 2015.
[11] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional
neural networks. In Advances in neural information processing systems, pages 1097–1105,
2012.
[12] W. Ling, E. Grefenstette, K. M. Hermann, T. Kočiskỳ, A. Senior, F. Wang, and P. Blunsom.
Latent predictor networks for code generation. arXiv preprint arXiv:1603.06744, 2016.
8
[13] T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and J. Dean. Distributed representations of
words and phrases and their compositionality. In Advances in neural information processing
systems, pages 3111–3119, 2013.
[14] T. A. Nguyen and C. Csallner. Reverse engineering mobile application user interfaces with
remaui (t). In Automated Software Engineering (ASE), 2015 30th IEEE/ACM International
Conference on, pages 248–259. IEEE, 2015.
[15] S. Reed, Z. Akata, X. Yan, L. Logeswaran, B. Schiele, and H. Lee. Generative adversarial text
to image synthesis. In Proceedings of The 33rd International Conference on Machine Learning,
volume 3, 2016.
[16] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, and Y. LeCun. Overfeat: Integrated recognition, localization and detection using convolutional networks. arXiv preprint
arXiv:1312.6229, 2013.
[17] R. Shetty, M. Rohrbach, L. A. Hendricks, M. Fritz, and B. Schiele. Speaking the same language:
Matching machine to human captions by adversarial training. arXiv preprint arXiv:1703.10476,
2017.
[18] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image
recognition. arXiv preprint arXiv:1409.1556, 2014.
[19] N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: a
simple way to prevent neural networks from overfitting. Journal of Machine Learning Research,
15(1):1929–1958, 2014.
[20] T. Tieleman and G. Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running average of
its recent magnitude. COURSERA: Neural networks for machine learning, 4(2), 2012.
[21] O. Vinyals, A. Toshev, S. Bengio, and D. Erhan. Show and tell: A neural image caption
generator. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
pages 3156–3164, 2015.
[22] K. Xu, J. Ba, R. Kiros, K. Cho, A. C. Courville, R. Salakhutdinov, R. S. Zemel, and Y. Bengio.
Show, attend and tell: Neural image caption generation with visual attention. In ICML,
volume 14, pages 77–81, 2015.
[23] L. Yu, W. Zhang, J. Wang, and Y. Yu. Seqgan: sequence generative adversarial nets with policy
gradient. arXiv preprint arXiv:1609.05473, 2016.
[24] W. Zaremba, I. Sutskever, and O. Vinyals. Recurrent neural network regularization. arXiv
preprint arXiv:1409.2329, 2014.
[25] H. Zhang, T. Xu, H. Li, S. Zhang, X. Huang, X. Wang, and D. Metaxas. Stackgan: Text to
photo-realistic image synthesis with stacked generative adversarial networks. arXiv preprint
arXiv:1612.03242, 2016.
9
| 2cs.AI
|
On the Quantitative Hardness of CVP
Huck Bennett∗
[email protected]
Alexander Golovnev†‡
[email protected]
Noah Stephens-Davidowitz∗§
[email protected]
arXiv:1704.03928v2 [cs.CC] 5 Oct 2017
Abstract
For odd integers p ≥ 1 (and p = ∞), we show that the Closest Vector Problem in the `p
norm (CVPp ) over rank n lattices cannot be solved in 2(1−ε)n time for any constant ε > 0 unless
the Strong Exponential Time Hypothesis (SETH) fails. We then extend this result to “almost all”
values of p ≥ 1, not including the even integers. This comes tantalizingly close to settling the
quantitative time complexity of the important special case of CVP2 (i.e., CVP in the Euclidean
norm), for which a 2n+o(n) -time algorithm is known. In particular, our result applies for any
p = p(n) 6= 2 that approaches 2 as n → ∞.
We also show a similar SETH-hardness result for SVP∞ ; hardness of approximating CVPp
to within some constant factor under the so-called Gap-ETH assumption; and other quantitative
hardness results for CVPp and CVPPp for any 1 ≤ p < ∞ under different assumptions.
∗
Courant Institute of Mathematical Sciences, New York University.
Yahoo Research.
‡
Part of this work was done while the author was at Courant Institute of Mathematical Sciences, and was supported
by the National Science Foundation under Grant No. CCF-1320188. Any opinions, findings, and conclusions or
recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the
National Science Foundation.
§
Supported by the National Science Foundation (NSF) under Grant No. CCF-1320188, and the Defense Advanced
Research Projects Agency (DARPA) and Army Research Office (ARO) under Contract No. W911NF-15-C-0236.
†
1
Introduction
A lattice L is the set of all integer combinations of linearly independent basis vectors b1 , . . . , bn ∈ Rd ,
L = L(b1 , . . . , bn ) :=
n
nX
o
zi bi : zi ∈ Z .
i=1
We call n the rank of the lattice L and d the dimension or the ambient dimension.
The two most important computational problems on lattices are the Shortest Vector Problem
(SVP) and the Closest Vector Problem (CVP). Given a basis for a lattice L ⊂ Rd , SVP asks us to
compute the minimal length of a non-zero vector in L, and CVP asks us to compute the distance
from some target point t ∈ Rd to the lattice. Typically, we define length and distance in terms of
the `p norm for some 1 ≤ p ≤ ∞, given by
kxkp := (|x1 |p + |x2 |p + · · · + |xd |p )1/p
for finite p and
kxk∞ := max |xi | .
1≤i≤d
In particular, the `2 norm is the familiar Euclidean norm, and it is by far the best studied in this
context. We write SVPp and CVPp for the respective problems in the `p norm. CVP is known to
be at least as hard as SVP (in any norm, under an efficient reduction that preserves the rank and
approximation factor) [GMSS99] and appears to be strictly harder.
Starting with the breakthrough work of Lenstra, Lenstra, and Lovász in 1982 [LLL82], algorithms
for solving these problems in both their exact and approximate forms have found innumerable
applications, including factoring polynomials over the rationals [LLL82], integer programming [Len83,
Kan87, DPV11], cryptanalysis [Sha84, Odl90, JS98, NS01], etc. More recently, many cryptographic
primitives have been constructed whose security is based on the worst-case hardness of these or
closely related lattice problems [Ajt04, Reg09, GPV08, Pei08, Pei16]. Given the obvious importance
of these problems, their complexity is quite well studied. Below, we survey some of these results.
We focus on algorithms for the exact and near-exact problems since these are most relevant to
our work and because the best known algorithms for the approximate variants of these problems
typically use algorithms for the exact problems as subroutines [Sch87, GN08, MW16]. (Many of the
results described below are also summarized in Table 1.)
1.1
Algorithms for SVP and CVP
The current fastest known algorithms for solving
The AKS algorithm and its descendants.
SVPp all use the celebrated randomized sieving technique due to Ajtai, Kumar, and Sivakumar [AKS01]. The original algorithm from [AKS01] was the first 2O(n) -time algorithm for SVP, and
it worked for both p = 2 and p = ∞.
In the p = 2 case, a sequence of works improved upon the constant in the exponent [NV08,
PS09, MV10, LWXZ11], and the current fastest running time of an algorithm that provably solves
SVP2 exactly is 2n+o(n) [ADRS15].1 While progress has slowed, this seems unlikely to be the end
of the story. Indeed, there are heuristic sieving algorithms that run in time (3/2)n/2+o(n) [NV08,
1
The algorithm in [ADRS15] is quite a bit different than the other algorithms in this class, but it can still be
thought of as a sieving algorithm.
1
WLTB11, Laa15, BDGL16], and there is some reason to believe that the provably correct [ADRS15]
algorithm can be improved. In particular, there is a provably correct 2n/2+o(n) -time algorithm that
approximates SVP2 up to a small constant approximation factor [ADRS15].
A different line of work extended the randomized sieving approach of [AKS01] to obtain 2O(n) time algorithms for SVP in additional norms. In particular, Blömer and Naewe extended it to all `p
norms [BN09]. Subsequent work extended this further, first to arbitrary symmetric norms [AJ08]
and then to the “near-symmetric norms” that arise in integer programming [Dad12].
Finally, a third line of work extended the [AKS01] approach to approximate CVP. Ajtai, Kumar,
and Sivakumar themselves showed a 2O(n) -time algorithm for approximating CVP2 to within any
constant approximation factor strictly greater than one [AKS02]. Blömer and Naewe obtained the
same result for all `p norms [BN09], and Dadush extended it further to arbitrary symmetric norms
and again to “near-symmetric norms” [Dad12]. We stress, however, that none of these results apply
to exact CVP, and indeed, there are some barriers to extending these algorithms to exact CVP.
(See, e.g., [ADS15].)
Exact CVP appears to be a much more subtle problem than
Exact algorithms for CVP.
2
exact SVP. Indeed, progress on exact CVP has been much slower than the progress on exact SVP.
Over a decade after [AKS01], Micciancio and Voulgaris presented the first 2O(n) -time algorithm
for exact CVP2 [MV13], using elegant new techniques built upon the approach of Sommer, Feder,
and Shalvi [SFS09]. Specifically, they achieved a running time of 4n+o(n) , and subsequent work
even showed a running time of 2n+o(n) for CVP2 with Preprocessing (in which the algorithm
is allowed access to arbitrary advice that depends on the lattice but not the target vector; see
Section 2.1) [BD15]. Later, [ADS15] showed a 2n+o(n) -time algorithm for CVP2 , so that the current
best known asymptotic running time is actually the same for SVP2 and CVP2 .
However, for p =
6 2, progress for exact CVPp has been minimal. Indeed, the fastest known
algorithms for exact CVPp with p 6= 2 are still the nO(n) -time enumeration algorithms first developed
by Kannan in 1987 [Kan87, DPV11, MW15]. Both algorithms for exact CVP2 mentioned in the
previous paragraph use many special properties of the `2 norm, and it seems that substantial new
ideas would be required to extend them to arbitrary `p norms.
1.2
Hardness of SVP and CVP
Van Emde Boas showed the NP-hardness of CVPp for any p and SVP∞ in 1981 [vEB81]. Extending
this to SVPp for finite p was a major open problem until it was proven (via a randomized reduction)
for all 1 ≤ p ≤ ∞ by Ajtai in 1998 [Ajt98]. There has since been much follow-up work, showing the
hardness of these problems for progressively larger approximation factors, culminating in NP-hardness
of approximating CVPp up to a factor of nc/ log log n for some constant c > 0 [ABSS93, DKRS03]
and hardness of SVPp with the same approximation factor under plausible complexity-theoretic
assumptions [CN98, Mic01b, Kho05, HR12]. These results are nearly the best possible under
√
plausible assumptions, since approximating either problem up to a factor of n is known to be in
NP ∩ coNP [GG00, AR05, Pei08].
2
In particular, there can be arbitrarily many lattice points that are approximate closest vectors, which makes
sieving techniques seemingly useless for solving exact CVP. (See, e.g., [ADS15] for a discussion of this issue.) We
note, however, that hardness results (including ours) tend to produce CVP instances with a bounded number of
approximate closest vectors (e.g., 2O(n) ).
2
However, such results only rule out the possibility of polynomial-time algorithms (under reasonable complexity-theoretic assumptions). They say very little about the quantitative hardness of
these problems for a fixed lattice rank n.3
This state of affairs is quite frustrating for two reasons. First, in the specific case of CVP2 ,
algorithmic progress has reached an apparent barrier. In particular, both known techniques for
solving exact CVP2 in singly exponential time are fundamentally unable to produce algorithms whose
running time is asymptotically better than the current best of 2n+o(n) [MV13, ADS15].4 Second,
some lattice-based cryptographic constructions are close to deployment [ADPS16, BCD+ 16, NIS16].
In order to be practically secure, these constructions require the quantitative hardness of certain
lattice problems, and so their designers rely on quantitative hardness assumptions [APS15]. If, for
example, there existed a 2n/20 -time algorithm for SVPp or CVPp , then these cryptographic schemes
would be insecure in practice.
We therefore move in a different direction. Rather than trying to extend non-quantitative
hardness results to larger approximation factors, we show quantitative hardness results for exact (or
nearly exact) problems. To do this, we use the tools of fine-grained complexity.
1.3
Fine-grained complexity
Impagliazzo and Paturi [IP99] introduced the Exponential Time Hypothesis (ETH) and the Strong
Exponential Time Hypothesis (SETH) to help understand the precise hardness of k-SAT. Informally,
ETH asserts that 3-SAT takes 2Ω(n) -time to solve in the worst case, and SETH asserts that k-SAT
takes essentially 2n -time to solve for unbounded k. I.e., SETH asserts that brute-force search is
essentially optimal for solving k-SAT for large k.
Recently, the study of fine-grained complexity has leveraged ETH, SETH, and several other assumptions to prove quantitative hardness results about a wide range of problems. These include both
problems in P (see, e.g., [CLR+ 14, BI15, ABW15] and the survey by Vassilevska Williams [Wil15]),
and NP-hard problems (see, e.g., [PW10, CDL+ 12, CFK+ 15]). Although these results are all
conditional, they help to explain why making further algorithmic progress on these problems is
difficult—and suggest that it might be impossible. Namely, any non-trivial algorithmic improvement
would disprove a very well-studied hypothesis.
One proves quantitative hardness results using fine-grained reductions (see [Wil15] for a formal
definition). For example, there is an efficient mapping from k-SAT formulas on n variables to
Hitting Set instances with universes of n elements [CDL+ 12]. This reduction is fine-grained in the
sense that for any constant ε > 0, a 2(1−ε)n -time algorithm for Hitting Set implies a 2(1−ε)n -time
algorithm for k-SAT, breaking SETH.
Despite extensive effort, no faster-than-2n -time algorithm for k-SAT with unbounded k has
been found. Nevertheless, there is no consensus on whether SETH is true or not, and recently,
Williams [Wil16] refuted a very strong variant of SETH. This makes it desirable to base quantitative
hardness results on weaker assumptions when possible, and indeed our main result holds even
3
One can derive certain quantitative hardness results from known hardness proofs, but in most cases the resulting
lower bounds are quite weak. The only true quantitative hardness results known prior to this work were a folklore
ETH-hardness result for CVP and an unpublished result due to Samuel Yeom, showing that CVP cannot be solved in
−4
time 210 n under plausible complexity-theoretic assumptions [Vai15]. (In Section 6.2, we present a similar proof of a
stronger statement.)
4
Both techniques require short vectors in each of the 2n cosets of L mod 2L (though for apparently different
reasons).
3
Problem
Upper Bound
CVPp
CVP2
CVP∞ /SVP∞
CVPPp
nO(n) (2O(n) )
2n
O(n)
2
O(n)
n
(2O(n) )
SETH
2n
—
2n *
—
Lower Bounds
Max-2-SAT ETH
2ωn/3
2Ω(n)
2ωn/3
2Ω(n)
—
2Ω(n)
√
√
2Ω( n)
2Ω( n)
Notes
Gap-ETH
2Ω(n) *
2Ω(n) *
2Ω(n) *
—
“almost all” p ∈
/ 2Z
Table 1: Summary of known quantitative upper and lower bounds, with new results in blue. Upper
bounds in parentheses hold for any constant approximation factor strictly greater than one, and
lower bounds with a ∗ apply for some constant approximation factor strictly greater than one. ω is
the matrix multiplication exponent, satisfying 2 ≤ ω < 2.373. We have suppressed smaller factors.
assuming a weaker variant of SETH based on the hardness of Weighted Max-k-SAT (except for the
case of p = ∞).
1.4
Our contribution
We now enumerate our results. See also Table 1.
SETH-hardness of CVPp .
Our main result is the SETH-hardness of CVPp for any odd integer
p ≥ 1 and p = ∞ (and SVP∞ ). Formally, we prove the following. (See Sections 3 and 4 for finite p
and Section 6.3 for p = ∞.)
Theorem 1.1. For any constant integer k ≥ 2 and any odd integer p ≥ 1 or p = ∞, there is an
efficient reduction from k-SAT with n variables and m clauses to CVPp (or SVP∞ ) on a lattice of
rank n (with ambient dimension n + O(m)).
In particular, there is no 2(1−ε)n -time algorithm for CVPp for any odd integer p ≥ 1 or p = ∞
(or SVP∞ ) and any constant ε > 0 unless SETH is false.
Unfortunately, we are unable to extend this result to even integers p, and in particular, to the
important special case of p = 2. In fact, this is inherent, as we show that our approach necessarily
fails for even integers p ≤ k − 1. In spite of this, we actually prove the following result that
generalizes Theorem 1.1 to “almost all” p ≥ 1 (including non-integer p).
Theorem 1.2. For any constant integer k ≥ 2, there is an efficient reduction from k-SAT with n
variables and m clauses to CVPp on a lattice of rank n (with ambient dimension n + O(m)) for any
p ≥ 1 such that
1. p is an odd integer or p = ∞;
2. p ∈
/ Sk , where Sk is some finite set (containing all even integers p ≤ k − 1); or
3. p = p0 + δ(n) for any p0 ≥ 1 and any δ(n) 6= 0 that converges to zero as n → ∞.
In particular, if SETH holds then for any constant ε > 0, there is no 2(1−ε)n -time algorithm for
CVPp for any p ≥ 1 such that
1. p is an odd integer or p = ∞;
4
2. p ∈
/ Sk for some sufficiently large k (depending on ε); or
3. p = p0 + δ(n).
Notice that this lower bound (Theorem 1.2) comes tantalizingly close to resolving the quantitative
complexity of CVP2 . In particular, we obtain a 2n -time lower bound on CVP2+δ for any 0 6= δ(n) =
o(1), and the fastest algorithm for CVP2 run in time 2n+o(n) . But, formally, Theorems 1.1 and 1.2
say nothing about CVP2 . (Indeed, there is at least some reason to believe that CVP2 is easier than
CVPp for p 6= 2 [RR06].)
We note that our reductions actually work for Weighted Max-k-SAT for all finite p =
6 ∞, so
that our hardness result holds under a weaker assumption than SETH, namely, the corresponding
hypothesis for Weighted Max-k-SAT.
Finally, we note that in the special case of p = ∞, our reduction works even for approximate
CVP∞ , or even approximate SVP∞ , with an approximation factor of γ := 1 + 2/(k − 1). In
particular, γ is constant for fixed k. This implies that for every constant ε > 0, there is a constant
γε > 1 such that no 2(1−ε)n -time algorithm approximates SVP∞ or CVP∞ to within a factor of γε
unless SETH fails.
As we discussed above, many 2O(n) -time
Quantitative hardness of approximate CVP.
algorithms for CVPp only work for γ-approximate CVPp for constant approximation factors γ > 1.
However, the reduction described above only works for exact CVPp (except when p = ∞).5
So, it would be preferable to show hardness for some constant approximation factor γ > 1. One
way to show such a hardness result is via a fine-grained reduction from the problem of approximating
Max-k-SAT to within a constant factor. Indeed, in the k = 2 case, we show that such a reduction
exists, so that there is no 2o(n) -time algorithm for approximating CVPp to within some constant
factor unless a 2o(n) -time algorithm exists for approximating Max-2-SAT. We also note that a
2o(n) -time algorithm for approximating Max-2-SAT to within a constant factor would imply one for
Max-3-SAT as well. (See Proposition 2.12.)
We present this result informally here (without worrying about specific parameters and the
exact definition of approximate Max-2-SAT). See Section 5 for the formal statement.
Theorem 1.3. There is an efficient reduction from approximating Max-2-SAT with n variables
and m clauses to within a constant factor to approximating CVPp to within a constant factor on a
lattice of rank n (with ambient dimension n + O(m)) for any finite p ≥ 1.
Quantitative hardness of CVP with Preprocessing.
CVP with Preprocessing (CVPP) is
the variant of CVP in which we are allowed arbitrary advice that depends on the lattice, but
not the target vector. CVPP and its variants have potential applications in both cryptography
(e.g., [GPV08]) and cryptanalysis. And, an algorithm for CVPP2 is used as a subroutine in the
celebrated Micciancio-Voulgaris algorithm for CVP2 [MV13, BD15]. The complexity of CVPPp is
well studied, with both hardness of approximation results [Mic01a, FM04, Reg04, AKKV11, KPV14],
and efficient approximation algorithms [AR05, DRS14].
We prove the following quantitative hardness result for CVPPp . (See Section 6.1.)
5
One can likely show that our “exact” reductions actually work for γ-approximate CVPp with some approximation
factor γ = 1 + o(1). But, this is not very interesting because standard techniques for “boosting” the approximation
factor are useless for us. (They increase the rank far too much.)
5
√
Theorem 1.4. For any 1 ≤ p < ∞, there is no 2o( n) -time algorithm for CVPP unless there is a
(non-uniform) 2o(n) -time algorithm for Max-2-SAT. In particular, no such algorithm exists unless
(non-uniform) ETH fails.
Additional quantitative hardness results for CVPp .
We also observe the following weaker
hardness result for CVPp for any 1 ≤ p < ∞ based on different assumptions. The ETH-hardness
of CVPp was already known in folklore, and even written down by Samuel Yeom in unpublished
work [Vai15]. We present a slightly stronger theorem than what was previously known, showing a
reduction from Max-2-SAT on n variables to CVPp on a lattice of rank n. (Prior to this work, we
were only aware of reductions from 3-SAT on n variables to CVPp on a lattice of rank Cn for some
very large constant C > 1000.)
Theorem 1.5. For any 1 ≤ p < ∞, there is an efficient reduction from Max-2-SAT with n variables
to CVPp on a lattice of rank n (and dimension n + m, where m is the number of clauses).
In particular, for any constant c > 0, there is no (poly(n) · 2cn )-time algorithm for CVPp unless
there is a 2cn -time algorithm for Max-2-SAT, and there is no 2o(n) -time algorithm for CVPp unless
ETH fails.
The fastest known algorithm for the Max-2-SAT problem is the poly(n) · 2ωn/3 -time algorithm
due to Williams [Wil05], where 2 ≤ ω < 2.373 is the matrix multiplication exponent [Wil12, LG14].
This implies that a faster than 2ωn/3 -time algorithm for CVPp (and CVP2 in particular) would
yield a faster algorithm for Max-2-SAT. (See, e.g., [Woe08] Open Problem 4.7 and the preceding
discussion.)
1.5
Techniques
Max-2-SAT.
We first show a straightforward reduction from Max-2-SAT to CVPp for any
1 ≤ p < ∞. I.e., we prove Theorem 1.5. This simple reduction will introduce some of the high-level
ideas needed for our more difficult reductions.
Given a Max-2-SAT instance Φ with n variables and m clauses, we construct the lattice basis
B :=
Φ̄
2αIn
!
,
(1)
where α > 0 is some very large number and Φ̄ ∈ Rm×n is given by
Φ̄i,j :=
2
−2
0
if the ith clause contains xj ,
if the ith clause contains ¬xj ,
otherwise .
(2)
I.e., the rows of Φ̄ correspond to clauses and the columns correspond to variables. Each entry
encodes whether the relevant variable is included in the relevant clause unnegated, negated, or not
at all, using 2, −2, and 0 respectively. (We assume without loss of generality that no clause contains
repeated literals or a literal and its negation simultaneously.) The target t ∈ Rm+n is given by
t := (t1 , t2 , . . . , tm , α, α, . . . , α)T ,
(3)
ti := 3 − 2ηi ,
(4)
where
6
where ηi is the number of negated variables in the ith clause.
Notice that the copy of 2αIn at the bottom of B together with the sequence of α’s in the last
coordinates of t guarantee that any lattice vector Bz with z ∈ Zn is at distance at least αn1/p away
from t. Furthermore, if z ∈
/ {0, 1}n , then this distance increases to at least α(n − 1 + 3p )1/p . This is
a standard gadget, which will allow us to ignore the case z ∈
/ {0, 1}n (as long as α is large enough).
I.e., we can view z as an assignment to the n variables of Φ.
Now, suppose z does not satisfy the ith clause. Then, notice that the ith coordinate of Bz will
be exactly −2ηi , so that (Bz − t)i = 0 − 3 = −3. If, on the other hand, exactly one literal in the ith
clause is satisfied, then the ith coordinate of Bz will be 2−2ηi , so that (Bz−t)i = 2−3 = −1. Finally,
if both literals are satisfied, then the ith coordinate will be 4 − 2ηi , so that (Bz − t)i = 4 − 3 = 1.
In particular, if the ith clause is not satisfied, then |(Bz − t)i | = 3. Otherwise, |(Bz − t)i | = 1.
It follows that the distance to the target is exactly distp (t, L)p = αp n + S + 3p (m − S) =
p
α n − (3p − 1)S + 3p m, where S is the maximal number of satisfied clauses of Φ. So, the distance
distp (t, L) tells us exactly the maximum number of satisfiable clauses, which is what we needed.
Difficulties extending this to k-SAT. The above reduction relied on one very important fact:
that |4 − 3| = |2 − 3| < |0 − 3|. In particular, a 2-SAT clause can be satisfied in two different ways;
either one variable is satisfied or two variables are satisfied. We designed our CVP instance above so
that the ith coordinate of Bz − t is 4 − 3 if two literals in the ith clause are satisfied by z ∈ {0, 1}n ,
2 − 3 if one literal is satisfied, and 0 − 3 if the clause is unsatisfied. Since |4 − 3| = |2 − 3|, the
“contribution” of this ith coordinate to the distance kBz − tkpp is the same for any satisfied clause.
Since |0 − 3| > |4 − 3|, the contribution to the ith coordinate is larger for unsatisfied clauses than
satisfied clauses.
Suppose we tried the same construction for a k-SAT instance. I.e., suppose we take Φ̄ ∈ Rm×n
to encode the literals in each clause as in Eq. (2) and construct our lattice basis B as in Eq. (1) and
target t as in Eq. (3), perhaps with the number 3 in the definition of t replaced by an arbitrary
t∗ ∈ R. Then, the ith coordinate of Bz − t would be 2Si − t∗ , where Si is the number of literals
satisfied in the ith clause.
No matter how cleverly we choose t∗ ∈ R, some satisfied clauses will contribute more to the
distance than others as long as k ≥ 3. I.e., there will always be some “imbalance” in this contribution.
As a result, we will not be able to distinguish between, e.g., an assignment that satisfies all clauses
but has Si far from t∗ /2 for all i and an assignment that satisfies fewer clauses but has Si ≈ t∗ /2
whenever i corresponds to a satisfying clause.
In short, for k ≥ 3, we run into trouble because satisfying assignments to a clause may satisfy
anywhere between 1 and k literals, but k distinct numbers obviously cannot all be equidistant from
some number t∗ . (See Section 6.2 for a simple way to get around this issue by adding to the rank
of the lattice. Below, we show a more technical way to do this without adding to the rank of the
lattice, which allows us to prove SETH-hardness.)
A solution via isolating parallelepipeds.
To get around the issue described above for k ≥ 3,
we first observe that, while many distinct numbers cannot all be equidistant from some number t∗ ,
∗
∗
it is trivial to find many distinct vectors in Rd that are equidistant from some vector t∗ ∈ Rd .
We therefore consider modifying the reduction from above by replacing the scalar ±2 values in our
∗
∗
matrix Φ̄ with vectors in Rd for some d∗ . In particular, for some vectors V = (v 1 , . . . , v k ) ∈ Rd ×k ,
7
we define Φ̄ ∈ Rd
∗ m×n
as
Φ̄i,j
vs
−v s
:=
0
d
if xj is the sth literal in the ith clause,
if ¬xj is the sth literal in the ith clause,
otherwise ,
(5)
where we have abused notation and taken Φ̄i,j to be a column vector in d∗ dimensions. By defining
∗
t ∈ Rd m+n appropriately,6 we will get that the “contribution of the ith clause to the distance”
∗
kBz − tkpp is exactly kV y − t∗ kpp for some t∗ ∈ Rd , where y ∈ {0, 1}k such that ys = 1 if and only
if z satisfies the sth literal of the relevant clause. (See Table 2 for a diagram showing the output of
the reduction and Theorem 3.2 for the formal statement.) We stress that, while we have increased
the ambient dimension by nearly a factor of d∗ , the rank of the lattice is still n.
This motivates the introduction of our primary technical tool, which we call isolating paral∗
lelepipeds. For 1 ≤ p ≤ ∞, a (p, k)-isolating parallelepiped is represented by a matrix V ∈ Rd ×k and
∗
a shift vector t∗ ∈ Rd with the special property that one vertex of the parallelepiped V {0, 1}k − t∗ is
“isolated.” (Here, V {0, 1}k − t∗ is an affine transformation of the hypercube, i.e., a parallelepiped.) In
particular, every vertex of the parallelepiped, V y − t∗ for y ∈ {0, 1}k has unit length kV y − t∗ kp = 1
except for the vertex −t∗ , which is longer, i.e., kt∗ kp > 1. (See Figure 1.)
In terms of the reduction above, an isolating parallelepiped is exactly what we need. In particular,
if we plug V and t∗ into the above reduction, then all satisfied clauses (which correspond to non-zero
y in the above description) will “contribute” 1 to the distance kBz − tkpp , while unsatisfied clauses
(which correspond to y = 0) will contribute 1 + δ for some δ > 0. Therefore, the total distance will
be exactly kBz − tkpp = αp n + m+ (z) + (m − m+ (z))(1 + δ) = αp n − δm+ (z) + (1 + δ)m, where
m+ (z) is the number of clauses satisfied by z. So, the distance distp (t, L) exactly corresponds to
the maximal number of satisfied clauses, as needed.
Constructing isolating parallelepipeds.
Of course, in order for the above to be useful, we
must show how to construct these (p, k)-isolating parallelepipeds. Indeed, it is not hard to find
constructions for all p ≥ 1 when k = 2, and even for all k in the special case when p = 1 (see
Figure 1). Some other fairly nice examples can also be found for small k, as shown in Figure 2.
For p > 1 and large k, these objects seem to be much harder to find. (In fact, in Section 4.2, we
show that there is no (p, k)-isolating parallelepiped for any even integer p ≤ k − 1.) Our solution is
therefore a bit technical.
k
k
At a high level, in Section 4, we consider a natural class of parallelepipeds V ∈ R2 ×k , t∗ ∈ R2
parametrized by some weights α0 , α1 , . . . , αk ≥ 0 and a scalar shift t∗ ∈ R. These parallelepipeds
are constructed so that the length of the vertex kV y − t∗ kpp for y ∈ {0, 1}k depends only on
the Hamming weight of y and is linear in the αi for fixed t∗ . In other words, there is a matrix
Mk (p, t∗ ) ∈ R(k+1)×(k+1) such that Mk (p, t∗ )(α0 , . . . , αk )T encodes the value of kV y − t∗ kpp for each
possible Hamming weight of y ∈ {0, 1}k . (See Lemma 4.2.)
We show that, in order to find weights α0 , . . . , αk ≥ 0 such that V and t∗ define a (p, k)-isolating
parallelepiped, it suffices to find a t∗ such that Mk (p, t∗ ) is invertible. For each odd integer p ≥ 1
and each k ≥ 2, we show an algorithm that finds such a t∗ . (See Section 4.1.)
6
In particular, we replace the scalars ti in Eq. (4) with vectors
ti := t∗ −
X
∗
v s ∈ Rd ,
where the sum is over s such that the sth literal in the ith clause is negated.
8
(k, k)
v2
v1 + v2
(2, 2)
t∗
··
·
(1, 1)
t∗
(0, 0)
0
v1
Figure 1: (p, k)-isolating parallelepipeds for p = 2, k = 2 (left) and p = 1, k ≥ 1 (right). On the left,
the vectors v 1 , v 2 , and v 1 + v 2 are all at the same distance from t∗ , while 0 is strictly farther away.
On the right is the degenerate parallelepiped generated by k copies of the vector (1, 1). The vectors
(i, i) are all at the same `1 distance from t∗ for 1 ≤ i ≤ m, while (0, 0) is strictly farther away. The
(scaled) unit balls centered at t∗ are shown in red, while the parallelepipeds are shown in black.
1
V :=
·
1/3
2 · 12
121/3 121/3 121/3
1
1
−1
1
−1
1
1
−1
−1
,
−1
1
1
−1
1
−1
−1
−1
1
t∗ :=
1
2 · 121/3
2 · 121/3
2
2
·
2
.
2
2
2
Figure 2: A (3, 3)-isolating parallelepiped in seven dimensions. One can verify that kV y − t∗ k33 = 1
for all non-zero y ∈ {0, 1}3 , and kt∗ k33 = 3/2.
9
To extend this result to other p ≥ 1, we consider the determinant of Mk (p, t∗ ) for fixed k and t∗ ,
viewed as a function of p. We observe that this function has a rather nice form—it is a Dirichlet
P
polynomial. I.e., for fixed t∗ and k, the determinant can be written as exp(ai p) for some ai ∈ R.
Such a function has finitely many roots unless it is identically zero. So, we take the value of t∗ from
above such that, say, Mk (1, t∗ ) is invertible. Since Mk (1, t∗ ) does not have zero determinant, the
Dirichlet polynomial corresponding to det(Mk (p, t∗ )) cannot be identically zero and therefore has
finitely many roots. This is how we prove Theorem 1.2. (See Section 4.3.)
Extension to constant-factor approximation.
In order to extend our hardness results to
approximate CVPp for finite p, we can try simply using the same reduction with k-SAT replaced by
approximate Max-k-SAT. Unfortunately, this does not quite work. Indeed, it is easy to see that the
“identity matrix gadget” that we use to restrict our attention to lattice vectors whose coordinates
are in {0, 1} (Eq. (1)) cannot tolerate an approximation factor larger than 1 + O(1/n) (for finite p).
However, we observe that when k = 2, this identity matrix gadget is actually unnecessary. In
particular, even without this gadget, it “never helps” to consider a lattice vector whose coordinates
are not all in {0, 1}. It then follows immediately from the analysis above that Gap-2-SAT reduces
to approximate CVPp with a constant approximation factor strictly greater than one. We note that
we do not know how to extend this result to larger k > 2 (except when p = 1, see Theorem 5.3). We
show that the case k = 2 is sufficient for proving Gap-ETH-hardness (see Proposition 2.12), but we
suspect that one can just “remove the identity matrix gadget” from all of our reductions for finite p.
If this were true, it would show Gap-ETH-hardness of approximation for slightly larger constant
approximation factors and imply even stronger hardness results under less common assumptions.
1.6
Open questions
The most important question that we leave open is the extension of our SETH-hardness result
to arbitrary p ≥ 1. In particular, while our result applies to p = p(n) 6= 2 that approaches 2
asymptotically, it does not apply to the specific case p = 2. An extension to p = 2 would settle
the time complexity of CVP2 up to a factor of 2o(n) (assuming SETH). However, we know that
our technique does not work in this case (in that (2, k)-parallelepipeds do not exist for k ≥ 3), so
substantial new ideas might be needed to resolve this issue.
Another direction would be to strengthen our hardness of approximation results in one of two
possible directions. First, one could try to increase the approximation factor. (Prior techniques for
amplifying the approximation factor increase the rank of the lattice quite a bit, so they do not yield
very interesting quantitative hardness results.) Second, one could try to show a reduction from
Gap-k-SAT to approximate CVPp for k ≥ 3. For p ∈ {1, ∞}, we already have such a reduction, and
as we mentioned above, we suspect that we can simply “remove the identity matrix gadget” in our
current reduction to achieve this for 1 < p < ∞. But, we do not know how to prove that this works.
Finally, we note that our main reduction constructs lattices of rank n, but the ambient dimension
d can be significantly larger. (Specifically, d = n + O(m), where m is the number of clauses in the
relevant SAT instance, and where the hidden constant depends on k and can be very large.) Lattice
problems are typically parameterized in terms of the rank of the lattice (and for the `2 norm, one
can assume without loss of generality that d = n), but it is still interesting to ask whether we can
reduce the ambient dimension d.
10
Organization
In Section 2, we review some necessary background knowledge. In Section 3, we show how to
use a (p, k)-isolating parallelepiped (for finite p) to reduce any n-variable instance of k-SAT to a
CVPp instance with rank n, and we show that this immediately gives SETH-hardness for p = 1.
In Section 4, we show how to construct (p, k)-isolating parallelepipeds, first for odd integers p ≥ 1
and then for “almost all” p. In Section 5, we show 2Ω(n) -hardness of approximating √
CVPp up to a
Ω(
n) ETH- and
constant factor. In Section 6, we prove a number of additional hardness results: 2
Max-2-SAT-hardness of CVPPp (Section 6.1), ETH- and Max-2-SAT-hardness of CVPp (Section 6.2),
and SETH-hardness of CVP∞ and SVP∞ (Section 6.3).
2
Preliminaries
Throughout this paper, we work with lattice problems over Rd for convenience. Formally, we must
pick a suitable representation of real numbers and consider both the size of the representation
and the efficiency of arithmetic operations in the given representation. But, we omit such details
throughout to ease readability.
2.1
Computational lattice problems
Let distp (L, t) := minx∈L kx − tkp denote the `p distance of t to L. In addition to SVP and CVP,
we also consider a variant of CVP called the Closest Vector Problem with Preprocessing (CVPP),
which allows arbitrary preprocessing of a lattice.
Definition 2.1. For any γ ≥ 1 and 1 ≤ p ≤ ∞, the γ-approximate Shortest Vector Problem
with respect to the `p norm (γ-SVPp ) is the promise problem defined as follows. Given a lattice L
(specified by a basis B ∈ Rd×n ) and a number r > 0, distinguish between a ‘YES’ instance where
there exists a non-zero vector v ∈ L such that kvkp ≤ r, and a ‘NO’ instance where kvkp > γr for
all non-zero v ∈ L.
Definition 2.2. For any γ ≥ 1 and 1 ≤ p ≤ ∞, the γ-approximate Closest Vector Problem with
respect to the `p norm (γ-CVPp ) is the promise problem defined as follows. Given a lattice L
(specified by a basis B ∈ Rd×n ), a target vector t ∈ Rd , and a number r > 0, distinguish between a
‘YES’ instance where distp (L, t) ≤ r, and a ‘NO’ instance where distp (L, t) > γr.
When γ = 1, we refer to the problems simply as SVPp and CVPp , respectively.
Definition 2.3. The Closest Vector Problem with Preprocessing with respect to the `p norm (CVPPp )
is the problem of finding a preprocessing function P and an algorithm Q which work as follows.
Given a lattice L (specified by a basis B ∈ Rd×n ), P outputs a new description of L. Given P (L), a
target vector t ∈ Rd , and a number r > 0, Q decides whether distp (L, t) ≤ r.
When we measure the running time of a CVPP algorithm, we only count the running time of Q,
and not of the preprocessing algorithm P .
11
2.2
Satisfiability problems and the Max-Cut problem
A k-SAT formula Φ on n Boolean variables x1 , . . . , xn and m clauses C1 , . . . , Cm is a conjunction
k
Φ = ∧m
i=1 Ci of clauses Ci = ∨s=1 `i,s , where the literals `i,s denote a variable xj or its negation ¬xj .
The goal is to decide whether there exists an assignment a ∈ {0, 1}n to the variables of Φ such that
all clauses have at least one “true” literal, i.e., so that all clauses are satisfied.
The value of a k-SAT formula Φ, denoted val(Φ), is the maximum fraction of clauses satisfied
by an assignment to Φ.
Definition 2.4. Given a k-SAT formula Φ and constants 0 ≤ δ < ε ≤ 1, the (δ, ε)-Gap-k-SAT
problem is the promise problem defined as follows. The goal is to distinguish between a ‘YES’
instance in which val(Φ) ≥ ε, and a ‘NO’ instance in which val(Φ) < δ.
Definition 2.5. Given a k-SAT formula Φ with clauses C = {C1 , . . . , Cm }, a clause weight function
w : C → Z+ , and a weight threshold W , the Weighted Max-k-SAT problem is to decide whether there
P
exists an assignment a to the variables of Φ such that Ci is sat. by a w(Ci ) ≥ W .
Definition 2.6. Given an undirected graph G = (V, E), an edge weight function w : E → Z+ , and
a weight threshold W , the Weighted Max-CUT problem is defined as follows. The goal is to decide
P
whether V can be partitioned into sets V1 and V2 such that eij ∈E:vi ∈V1 ,vj ∈V2 w(eij ) ≥ W .
There exists a folklore reduction from an instance of Weighted Max-Cut on a graph with n
vertices to an instance of Weighted Max 2-SAT on a formula with n variables. See, e.g., [GHNR03].
2.3
Exponential Time Hypotheses
Impagliazzo and Paturi [IP99] introduced the following two hypotheses (ETH and SETH), which
are now widely used to study the quantitative hardness of computational problems.
Definition 2.7. The Exponential Time Hypothesis (ETH) is the hypothesis defined as follows. For
every k ≥ 3 there exists a constant ε > 0 such that no algorithm solves k-SAT formulas with n
variables in 2εn -time. In particular, there is no 2o(n) -time algorithm for 3-SAT.
Definition 2.8. The Strong Exponential Time Hypothesis (SETH) is the hypothesis defined as
follows. For every constant ε > 0 there exists k such that no algorithm solves k-SAT formulas with
n variables in 2(1−ε)n -time.
An important tool in the study of the exact complexity of k-SAT is the Sparisification Lemma of
Impagliazzo, Paturi, and Zane [IPZ01] which roughly says that any k-SAT formula can be replaced
with 2εn formulas each with O(n) clauses for some ε > 0.
Proposition 2.9 (Sparsification Lemma, [IPZ01]). For every k ∈ Z+ and ε > 0 there exists a
constant c = c(k, ε) such that any k-SAT formula Φ with n variables can be expressed as Φ = ∨ri=1 Ψi
where r ≤ 2εn and each Ψi is a k-SAT formula with at most cn clauses. Moreover, this disjunction
can be computed in 2εn -time.
In this paper we also consider the W-Max-SAT-SETH hypothesis, which corresponds to SETH
but with Weighted Max-k-SAT in place of k-SAT. Our main result only relies on this weaker variant
of SETH, and is therefore more robust.
Dinur [Din16] and Manurangsi and Raghavendra [MR17] recently introduced a “gap” version of
ETH, which asserts that Gap-3-SAT takes 2Ω(n) -time.
12
Definition 2.10. The (randomized) Gap-Exponential Time Hypothesis ((randomized) Gap-ETH)
is the hypothesis that there exist constants δ < 1 and ε > 0 such that no (randomized) algorithm
solves (δ, 1)-Gap-3-SAT instances with n variables in 2εn -time.
As Dinur [Din16] notes, one can sparsify a Gap-SAT instance simply by sampling clauses.
Therefore, we can assume (almost) without loss of generality that Gap-ETH applies only to formulas
with O(n) clauses. The caveat is that the sampling is randomized, so finding a 2o(n) -time algorithm
for sparse Gap-3-SAT only implies a randomized 2o(n) -time algorithm for general Gap-3-SAT.
We give a variant of Dinur’s sampling argument in Proposition 2.11. The idea is to show that
both the total number of sampled clauses and the number of sampled clauses that are satisfied by
any given assignment are highly concentrated around their expectation by using the Chernoff bound,
and then to take a union bound over the bad events where these quantities deviate substantially
from their expectation.
We will use the following multiplicative Chernoff bounds (see, e.g., [HP]). Let X1 , . . . , Xn be
independent identically distributed Bernoulli random variables with expectation p, so that the
P
expectation of ni=1 Xi is µ = pn. Then:
Pr[
n
X
2 /2
Xi < (1 − α)µ] < e−µα
,
(6)
.
(7)
i=1
Pr[
n
X
2 /4
Xi > (1 + α)µ] < e−µα
i=1
Proposition 2.11 (Sparsification for Gap-SAT). For any 0 < δ < δ 0 < 1, there is a polynomial-time
randomized reduction from a (δ, 1)-Gap-k-SAT instance Φ with n variables and m clauses to a
(δ 0 , 1)-Gap-k-SAT instance Φ0 with n variables and O(n) clauses.
Proof. Let Φ0 be the formula obtained by sampling each clause of Φ independently with probability
p := min{1, 10/(δα2 ) · n/m}, where α is fixed so that −(1 − δ 0 /δ)/(1 + δ 0 /δ) < α < 1. Clearly, if
val(Φ) = 1 then val(Φ0 ) = 1 as well. We analyze the case where val(Φ) < δ.
In expectation Φ0 has pm clauses. Furthermore, because val(Φ0 ) < δ, in expectation any fixed
assignment will satisfy fewer than δpm clauses of Φ0 . Therefore by Equation (6),
Pr[Number of clauses in Φ0 ≤ (1 − α)pm] ≤ e−α
2 pm/2
≤ e−2n .
(8)
Furthermore, by Equation (7), we have that for each fixed assignment a,
Pr[Number of clauses in Φ0 sat. by a ≥ (1 + α)δpm] ≤ e−α
2 δpm/4
≤ e−2n .
(9)
By applying Equations (8) and (9), and taking a union bound we get that the probability that
has at least (1 − α)pm clauses and that no assignment to Φ0 satisfies more than (1 + α)δpm
clauses is at least 1 − (e−2n + 2n e−2n ) ≥ 1 − 2e−n . Therefore,
Φ0
val(Φ0 ) ≤
(1 + α)pm
· δ < δ0
(1 − α)pm
with high probability.
13
x1
x2
···
xn−1
xn
v1
v2
···
0d∗
−v 3
t∗ − v 3
..
.
···
..
..
.
..
.
..
.
0d∗
−v 1
···
v2
v3
t∗ − v 1
x1
x2
..
.
2α1/p
0
..
.
0
2α1/p
..
.
···
···
..
.
0
0
..
.
0
0
..
.
α1/p
α1/p
..
.
xn−1
xn
0
0
0
0
···
···
2α1/p
0
0
2α1/p
α1/p
α1/p
(
C1
..
.
.
(
Cm
B
t
Table 2: A basis B and target vector t output by the reduction from Theorem 3.2 with some
∗
∗
(p, 3)-isolating parallelepiped given by V = (v 1 , v 2 , v 3 ) ∈ Rd ×3 and t∗ ∈ Rd . In this example, the
first clause is C1 ≡ x1 ∨ x2 ∨ ¬xn and the mth clause is Cm ≡ ¬x2 ∨ xn−1 ∨ xn . By the definition of
an isolating parallelepiped (Definition 3.1), the contribution of the first d coordinates to the distance
kBz − tkpp will be 1 for any assignment z ∈ {0, 1}n satisfying C1 , while non-satisfying assignments
contribute (1 + δ) for some δ > 0. For example, if z1 = 1, z2 = 0, zn = 1, the clause C1 is satisfied,
and the first d coordinates will contribute kv 1 − v 3 − (t∗ − v 3 )kpp = kv 1 − t∗ kpp = 1. On the other
hand, if z1 = 0, z2 = 0, zn = 1, then C1 is not satisfied, and k − v 3 − (t∗ − v 3 )kpp = kt∗ kpp = 1 + δ.
Additionally, we will use a reduction of Garey et al. [GJS76] from 3-SAT to Max-2-SAT which
also works as a reduction from Gap-3-SAT to Gap-2-SAT. The reduction works by outputting ten 1and 2-clauses for each 3-clause in the original formula. Any assignment which satisfies the original
clause corresponds to an assignment which satisfies 7 of the output clauses, and any assignment
which does not satisfy the original clause corresponds to an assignment which satisfies 6 of the
output clauses.
Proposition 2.12 ([GJS76, Theorem 1.1]). For every 0 ≤ δ < ε ≤ 1, there is a polynomial-time
reduction from every instance of (δ, ε)-Gap-3-SAT with n variables and m clauses to an instance of
((6 + δ)/10, (6 + ε)/10)-Gap-2-SAT with n + m variables and 10m clauses.
3
SETH-hardness from isolating parallelepipeds
We start by giving a reduction from instances of weighted Max-k-SAT on formulas with n variables
to instances of CVPp with rank n for all p that uses a certain geometric object, which we define
next. Let 1n and 0n denote the all 1s and all 0s vectors of length n respectively, and let In denote
the n × n identity matrix.
14
∗
∗
Definition 3.1. For any 1 ≤ p ≤ ∞ and integer k ≥ 2, we say that V ∈ Rd ×k and t∗ ∈ Rd define
a (p, k)-isolating parallelepiped if ktkp > 1 and kV x − t∗ kp = 1 for all x ∈ {0, 1}k \ {0k }.
In order to give the reduction, we first introduce some notation related to SAT. Let Φ be a
k-SAT formula on n variables x1 , . . . , xn and m clauses C1 , . . . , Cm . Let ind(`) denote the index of
the variable underlying a literal `. I.e., ind(`) = j if ` = xj or ` = ¬xj . Call a literal ` positive if
` = xj and negative if ` = ¬xj for some variable xj . Given a clause Ci = ∨ks=1 `i,s , let Pi := {s ∈ [k] :
`i,s is positive} and let Ni := {s ∈ [k] : `i,s is negative} denote the indices of positive and negative
literals in Ci respectively. Given an assignment a ∈ {0, 1}n to the variables of Φ, let Si (a) denote the
indices of literals in Ci satisfied by a. I.e., Si (a) := {s ∈ Pi : aind(`i,s ) = 1} ∪ {s ∈ Ni : aind(`i,s ) = 0}.
Finally, let m+ (a) denote the number of clauses of Φ satisfied by the assignment a, i.e., the number
of clauses i for which |Si (a)| ≥ 1.
Theorem 3.2. If there exists a computable (p, k)-isolating parallelepiped for some p = p(n) ∈ [1, ∞)
and integer k ≥ 2, then there exists a polynomial-time reduction from any (weighted-)Max-k-SAT
instance with n variables to a CVPp instance of rank n.
Proof. For simplicity, we give a reduction from unweighted Max-k-SAT, and afterwards sketch how
to modify our reduction to handle the weighted case as well. Namely, we give a reduction from any
Max-k-SAT instance (Φ, W ) to an instance (B, t∗ , r) of CVPp . Here, the formula Φ is on n variables
x1 , . . . , xn and m clauses C1 , . . . , Cm . (Φ, W ) is a ‘YES’ instance if there exists an assignment a
such that m+ (a) ≥ W .
∗
By assumption, there exist computable d∗ = d∗ (p, k) ∈ Z+ , V = [v 1 , . . . , v k ] ∈ Rd ×k , and
∗
t∗ ∈ Rd such that kt∗ kp = (1 + δ)1/p for some δ > 0 and kV z − t∗ kp = 1 for all z ∈ {0, 1}k \ {0k }.
We define the output CVPp instance as follows. Let d := md∗ + n. The basis B ∈ Rd×n and
target vector t ∈ Rd in the output instance have the form
B1
..
.
B=
∗
Bm
2α1/p · In
,
t=
t1
..
.
tm
α1/p · 1n
,
∗
with blocks Bi ∈ Rd ×n and ti ∈ Rd for 1 ≤ i ≤ m and α := m + (m − W )δ. Note that α is the
maximum possible contribution of the clauses C1 , . . . , Cm to kBy − tkpp when (Φ, W ) is a ‘YES’
instance. For every 1 ≤ i ≤ m and 1 ≤ j ≤ n, set the jth column (Bi )j of block Bi (corresponding
to the clause Ci = ∨ks=1 `i,s ) as
(Bi )j :=
v s if xj is the sth literal of clause i,
−v s if ¬xj is the sth literal of clause i,
0 ∗ otherwise,
d
and set ti := t∗ − s∈Ni v s . Set r := (α(n + 1))1/p .
Clearly, the reduction runs in polynomial time. We next analyze for which y ∈ Zn it holds that
kBy − tkp ≤ r. Given y ∈
/ {0, 1}n ,
P
kBy − tkpp ≥ k2α1/p In y − α1/p 1n kpp ≥ α(n + 2) > rp ,
15
so we only need to analyze the case when y ∈ {0, 1}n . Consider an assignment y ∈ {0, 1}n to the
variables of Φ. Then,
kBi y − ti kp =
X
yind(`i,s ) · v s −
s∈Pi
=
X
yind(`i,s ) · v s − t∗ −
s∈Ni
yind(`i,s ) · v s +
s∈Pi
=
X
X
X
s∈Si (a)
s∈Ni
1 − yind(`i,s ) · v s − t∗
p
p
s∈Ni
X
vs
v s − t∗ .
p
By assumption, the last quantity is equal to 1 if |Si (y)| ≥ 1, and is equal to (1 + δ)1/p otherwise.
Because |Si (y)| ≥ 1 if and only if Ci is satisfied, it follows that
kBy − tkpp =
m
X
kBi y − ti kpp + αn = m + (m − m+ (y))δ + αn .
i=1
Therefore, kBy − tkp ≤ r if and only if m+ (y) ≥ W , and therefore there exists y such that
kBy − tkp ≤ r if and only if (Φ, W ) is a ‘YES’ instance of Max-k-SAT, as needed.
To extend this to a reduction from weighted Max-k-SAT to CVPp , simply multiply each block
Bi and the corresponding target vector ti by w(Ci )1/p , where w(Ci ) denotes the weight of the clause
Ci . Then, by adjusting α to depend on the weights w(Ci ) we obtain the desire reduction.
Because the rank n of the output CVPp instance matches the number of variables in the input
SAT formula, we immediately get the following corollary.
Corollary 3.3. For any efficiently computable p = p(n) ∈ [1, ∞) if there exists a computable
(p, k)-isolating parallelepiped for infinitely many k ∈ Z+ , then, for every constant ε > 0 there
is no 2(1−ε)n -time algorithm for CVPp assuming W-Max-SAT-SETH. In particular there is no
2(1−ε)n -time algorithm for CVPp assuming SETH.
It is easy to construct a (degenerate) family of isolating parallelepipeds for p = 1, and therefore
we get hardness of CVP1 as a simple corollary. (See Figure 1.)
Corollary 3.4. For every constant ε > 0 there is no 2(1−ε)n -time algorithm for CVP1 assuming
W-Max-SAT-SETH, and in particular there is no 2(1−ε)n -time algorithm for CVP1 assuming SETH.
1
Proof. Let k ∈ Z+ , let V = [v 1 , . . . , v k ] with v 1 = · · · = v k := k−1
(1, 1)T ∈ R2 , and let t∗ :=
1
T
2
∗
k
∗
k−1 (k, 1) ∈ R . Then, kV x − t k1 = 1 for every x ∈ {0, 1} \ {0k }, and kt k1 = (k + 1)/(k − 1) > 1.
The result follows by Corollary 3.3.
4
Finding isolating parallelepipeds
∗
∗
We now show how to find a (p, k)-isolating parallelepiped given by V ∈ Rd ×k and t∗ ∈ Rd as in
Definition 3.1. We will first show a general strategy for trying to find such an object for any p ≥ 1
and integer k ≥ 2. In Section 4.1, we will show how to successfully implement this strategy in the
case when p is an odd integer. In Section 4.2, we show that (p, k)-isolating parallelepipeds do not
16
exist for even integers p ≤ k − 1. Finally, in Section 4.3 we show how to mostly get around this
issue in order to find (p, k)-isolating parallelepipeds for “almost all” p ≥ 1.
It will actually be convenient to find a slightly different object that “works with {±1}k instead
of {0, 1}k .” We observe below that this suffices.
∗
Lemma 4.1. There is an efficient algorithm that takes as input a matrix V ∈ Rd ×k and vector
∗
t∗ ∈ Rd such that kV y − t∗ kp = 1 for any y ∈ {±1}k \ {−1k } and k − V 1k − t∗ kp > 1 , and outputs
∗
∗
a matrix V 0 ∈ Rd ×k and vector (t∗ )0 ∈ Rd that form a (p, k)-isolating parallelepiped.
Proof. Define V 0 := 2V and (t∗ )0 = V 1k + t∗ . Now consider the affine transformation f : Rk → Rk
defined by f (x) := (2x − 1k ), which maps {0, 1}k to {±1}k and 0k to −1k . Then, for x ∈ {0, 1}k
and y = f (x) = 2x − 1k ∈ {±1}k , we have
kV 0 x − (t∗ )0 kp = V 0
y + 1k
− (t∗ )0
2
p
= V0
y
1k
+V0
− (t∗ )0
2
2
p
= kV y − t∗ kp ,
as needed.
Intuitively, a “reasonable” matrix V should act symmetrically on bit strings. I.e., if y, y 0 ∈ {±1}k
have the same number of positive entries, then V y should be a permutation of V y 0 . This implies
that any row of V must be accompanied by all possible permutations of this row. If we further
require that each row in V is α · v for some v ∈ {±1}k and α ∈ R, then we arrive at a very general
construction that is still possible to analyze.
k
For weights α0 , . . . , αk ≥ 0, we define V := V (α0 , . . . , αk ) ∈ R2 ×k as follows. The rows of V
1/p
are indexed by the strings {±1}k , and row v is αk−|v| v T , where
|v| := {# of positive entries in v}
k
is the number of positive entries in v. For a shift t∗ ∈ R, we set t∗ := t∗ (α0 , . . . , αk , t∗ ) ∈ R2
1/p
such that the coordinate of t∗ corresponding to v is αk−|v| t∗ . (Figure 2 is an example of this
construction. In particular, it shows V (α0 , α1 , α2 , α3 ) with α0 = 12, α1 = α2 = 1 and α3 = 0 and
t∗ (α0 , α1 , α2 , α3 , t∗ ) with t∗ = 2, where we have omitted the last row, whose weight is zero.)
In what follows, we will use the binomial coefficient ji extensively, and we adopt the convention
that
i
j
= 0 if j > i or j < 0 or j ∈
/ Z.
Lemma 4.2. For any y ∈ {±1}k , weights α0 , . . . , αk ≥ 0, and shift t∗ ∈ R,
kV y −
t∗ kpp
=
k
X
j=0
αk−j
k
X
|y|
`=0
`
!
!
k − |y|
· 2|y| + 2j − k − 4` − t∗
j−`
k
p
,
where V := V (α0 , . . . , αk ) ∈ R2 ×k and t∗ := t∗ (α0 , . . . , αk , t∗ ) as above.
In other words, kV y − t∗ kpp depends only on |y|, and if w ∈ (R≥0 )k+1 is the vector such that
wj = kV y 0 − t∗ kpp for all y 0 ∈ {±1} with |y 0 | = j, then
w = Mk (p, t )
∗
17
α0
α1
..
.
αk
,
where Mk (p, t∗ ) ∈ R(k+1)×(k+1) is given by
Mk (p, t∗ )i,j :=
k
X
`=0
i
`
!
!
k−i
· 2i + 2j − k − 4` − t∗
j−`
p
.
(10)
Proof. We have
kV y − t∗ kpp =
k
X
X
αj
j=0
hv, yi − t∗
p
.
|v|=k−j
Notice that hv, yi depends only on how many of the j negative entries of v align with the positive
entries of y. In particular,
∗ p
X
hv, yi − t
=
|v|=k−j
!
!
k
X
|y| k − |y|
`=0
j−`
`
k
X
|y|
!
=
`=0
`
· − ` + (|y| − `) + (j − `) − (k − |y| − j + `) − t∗
!
k − |y|
· 2|y| + 2j − k − 4` − t∗
j−`
p
p
,
as needed.
Lemma 4.3. For any t∗ ∈ R, the matrix Mk (p, t∗ ) defined in Eq. (10) is stochastic.
Mk (p, t∗ )1k+1 = λ(t∗ )1k+1 for some λ(t∗ ) ∈ R. Furthermore, λ(t∗ ) > 0.
I.e.,
Proof. We rearrange the sum corresponding to the ith entry of Mk (p, t∗ )1k+1 , setting r := (i+j)/2−`
to obtain
k X
k
X
j=0 `=0
i
`
!
!
k−i
· 2i + 2j − k − 4` − t∗
j−`
p
=
X
|4r − k − t∗ |p
r
=
X
k
X
j=0
∗ p
|4r − k − t |
r
k
X
j=0
i
(i + j)/2 − r
!
k−i
r + (j − i)/2
!
i
r + (i − j)/2
!
k−i
r + (j − i)/2
!
Finally, we recall Vandermonde’s identity, which says that
k
X
j=0
i
r + (i − j)/2
!
k−i
r + (j − i)/2
!
=
k
r
!
.
Therefore, the summation does not depend on i (and is clearly positive), as needed.
Lemma 4.3 tells us that for any t∗ ∈ R, Mk (p, t∗ )(1k+1 )/λ = 1k+1 for some λ > 0. We wish
to show that, for some t∗ ∈ R, we can find α0 , . . . , αk ≥ 0 such that Mk (p, t∗ )(α0 , α1 , . . . , αk )T =
1k+1 + εe0 for some ε > 0, where e0 := (1, 0, . . . , 0)T . In order to do this, it suffices to show that
Mk (p, t∗ ) is invertible. Then, we can take
(α0 , α1 , . . . , αk )T := 1k+1 /λ + εMk (p, t∗ )−1 e0 .
If ε := (λ · kMk (p, t∗ )−1 e0 k∞ )−1 > 0, then the αi must be non-negative. We make this formal in
the next proposition.
18
.
Proposition 4.4. There is an efficient algorithm that takes as input any p ≥ 1, an integer k ≥ 2,
and t∗ ∈ R such that det(Mk (p, t∗ )) 6= 0, where Mk (p, t∗ ) is defined as in Eq. (10) and outputs
k
k
V ∈ R2 ×k and t∗ ∈ R2 that define a (p, k)-isolating parallelepiped.
Proof. By Lemma 4.1, it suffices to construct a matrix that works for y ∈ {±1}k . The algorithm
behaves as follows on input k ≥ 2 and p ≥ 1 and t∗ ∈ R. By Lemma 4.3, Mk (p, t∗ )1k+1 = λ1k+1 for
some λ > 0. Since we are promised that det(Mk (p, t∗ )) 6= 0, we see that Mk (p, t∗ ) is invertible. The
algorithm therefore sets
(α0 , . . . , αk )T := 1k+1 /λ + εMk (p, t∗ )−1 e0 ,
(11)
where ε := (λ · kMk (p, t∗ )−1 e0 k∞ )−1 > 0 is chosen to be small enough such that the αi are all
non-negative. Finally, it outputs V := V (α0 , . . . , αk ) and t∗ := t∗ (α0 , . . . , αk , t∗ ) as defined above.
To prove correctness, we note that V and t∗ have the desired property. Indeed, it follows
from the definition of Mk (p, t∗ ) in Lemma 4.2 that kV y − t∗ kpp is the jth coordinate of w :=
Mk (p, t∗ )(α0 , . . . , αk )T , where j := |y|. But, by Eq. (11), we see that the jth coordinate of w is
1 + ε if j = 0 and is 1 otherwise, as needed.
4.1
Finishing the proof for odd integer p
We now handle the case when p ≥ 1 is an odd integer. Notice that, if p ≥ 1 is an integer, then
det(Mk (p, t∗ )) is some piecewise combination of polynomials of degree at most (k + 1)p in t∗ . In
particular, it is a polynomial in t∗ if we restrict our attention to the interval t∗ ∈ [−k, −k + 2].
We wish to argue that this is not the zero polynomial when p is odd. To prove this, it suffices to
show that the coefficient of (t∗ )(k+1)p is non-zero, which we do below by studying a matrix whose
determinant is this coefficient (when p is odd).
We first show an easy claim concerning matrices that can be written as sums of the identity
plus a certain kind of rank-one matrix.
Claim 4.5. For any matrix A ∈ Rd×d with constant columns given by Ai,j = aj for some
a0 , . . . , ad−1 ∈ R and any λ ∈ R,
det(A − λId ) = (−λ)d−1
X
aj − λ .
j
P
Proof. Notice that A is a rank-one stochastic matrix with one non-zero eigenvalue given by aj .
P
Therefore, the characteristic polynomial of A is det(A − λId ) = (−λ)d−1 ( aj − λ), as needed.
Lemma 4.6. For an integer k ≥ 1 and an odd integer p ≥ 1, the function t∗ 7→ det(Mk (p, t∗ )),
where Mk (p, t∗ ) is defined as in Eq. (10), is a polynomial of degree at most (k + 1)p when restricted
to the interval t∗ ∈ [−k, −k + 2]. Furthermore, the coefficient of (t∗ )(k+1)p of this polynomial is
exactly 2k (2 − 2k ).
In particular, t∗ 7→ det(Mk (p, t∗ )) is a non-zero polynomial of degree (k + 1)p on the interval
t∗ ∈ [−k, −k + 2] for k ≥ 2.
Proof. For any t∗ ∈ [−k, −k + 2], the matrix Mk (p, t∗ ) is given by
∗
Mk (p, t )i,j =
k
X
`=0
!
i
`
!
k−i
· 2i+2j−k−4`−t∗
j−`
p
19
=
k
X
`=0
!
δi+j−2`
i
`
!
k−i
·(2i+2j−k−4`−t∗ )p ,
j−`
where δr = −1 if r = 0 and 1 otherwise. (Here, we have used the fact that `i k−i
j−` is only non-zero
∗
when ` ≤ min(i, j). Therefore, 2i + 2j ≥ 4`, so that 2i + 2j − k − 4` − t ≥ 2i + 2j − 4` − 2 ≥ 0
unless 2i − 2j − 4` = 0.)
The coefficient of (t∗ )(k+1)p in the polynomial t∗ 7→ det(Mk (p, t∗ )) is therefore given by det(M 0 ),
where M 0 is defined as
0
(M )i,j := −
k
X
!
δ2`−i−j
`=0
!
k−i
j−`
!
k
X
k−i
i
−
(j − i)/2
`
`=0
!
k−i
k
−
(j − i)/2
j
i
=2
(i + j)/2
i
=2
(i − j)/2
i
`
!
!
!
!
k−i
j−`
!
,
where we have again applied Vandermonde’s identity. Notice that the first term is non-zero if and
only if i = j, in which case it is equal to 2. In other words, M 0 = A + 2Ik+1 , where Ai,j := − kj .
The result then follows from Claim 4.5.
Corollary 4.7. There is an efficient algorithm that takes as input an integer k ≥ 2 and odd integer
p ≥ 1 and outputs t∗ ∈ Q such that det(Mk (p, t∗ )) 6= 0, with Mk (p, t∗ ) defined as in Eq. (10).
Proof. The algorithm works as follows. It chooses (k + 1)p + 1 distinct points t0 , . . . , t(k+1)p ∈
[−k, −k + 2] arbitrarily. (E.g., it chooses ti = −k + 2i/((k + 1)p).) For each ti , it computes
det(Mk (p, ti )). It outputs the first ti such that the determinant is non-zero.
We claim that det(Mk (p, ti )) 6= 0 for at least one index i. Indeed, by Lemma 4.6, t∗ 7→
det(Mk (p, t∗ )) is a non-zero polynomial of degree (k + 1)p. The result then follows from the fact
that such a polynomial can have at most (k + 1)p roots.
Theorem 1.1 for finite p now follows immediately from Theorems 3.2 together with Proposition 4.4,
and Corollary 4.7.
4.2
Limitations of the approach
In the previous section, we showed that for every odd p ≥ 1 and every integer k ≥ 2, there exists a
(p, k)-isolating parallelepiped. This allowed us to conclude that CVPp is SETH-hard for odd values
of p. Now, we show that this approach necessarily fails for even p ≥ 2. Namely, we show that for
every even p, there is no (p, k)-isolating parallelepiped for any k > p.7 For simplicity, we show this
for p = 2, but a straightforward generalization works for all even p.
∗
Lemma 4.8. For any integer k ≥ 3 and vectors v 1 , . . . , v k , t∗ ∈ Rd , we have
X
(−1)|S| t∗ −
i∈S
S⊆[k]
7
X
2
vi
2
=0.
When k ≤ p, it is possible to construct (p, k)-isolating parallelepiped for even p. See, e.g., Figure 1.
20
Proof. We have
X
S⊆[k]
|S|
(−1)
∗
t −
X
i∈S
2
vi
2
=
|S|
X
kt∗ k22
(−1)
D
∗
−2 t ,
= kt∗ k22 ·
(−1)|S| − 2
X
S⊆[k]
X
vi +
kt∗ k22
X
·
X
(−1)
|S|
t∗ , v i ·
·0−2
2
vi
X
!
2
(−1)|S|
S3i
+2
X
hv i , v j i ·
i<j
S3i
X
X
i∈S
i∈[k]
kvi k22
i∈[k]
=
E
i∈S
S⊆[k]
+
X
X
(−1)|S|
S3i,j
∗
ht , v i i · 0
i∈[k]
+
X
kvi k22 · 0 + 2
X
hv i , v j i · 0
i<j
i∈[k]
=0,
where the penultimate equality uses the fact that
X
(−1)|S| = (1 − 1)n = 0
S⊆[n]
for n ≥ 1.
Corollary 4.9. There is no (2, k)-isolating parallelepiped for any integer k ≥ 3.
∗
Proof. Assume V = [v 1 , . . . , v k ] ∈ Rd ×k and t∗ ∈ Rd form a (2, k)-isolating parallelepiped. For
P
any S 6= ∅, kt∗ − i∈S v i k22 = 1 by the definition of an isolating parallelepiped. Thus, applying
Lemma 4.8, we have
kt∗ k22 =
X
(−1)|S|+1 t∗ −
X
i∈S
∅6=S⊆[k]
2
vi
2
=1,
which contradicts the assumption that V and t∗ form an isolating parallelepiped.
4.3
Extending our result to almost all p
We now wish to extend Theorem 1.1 to arbitrary p ≥ 1. Unfortunately, we know that we cannot
do this for all p, since we showed in Section 4.2 that no such construction is possible when p is an
even integer. However, we show a construction that works for “almost all values of p.” In particular,
for any fixed k, the construction works for all but finitely many choices of p. We also observe that
this implies that, for every fixed p0 , k, there is an ε > 0 such that the construction works for every
p ∈ (p0 − ε) or p ∈ (p0 + ε). In particular, for any non-zero δ = δ(n) = o(1), the construction works
for p = p0 + δ(n) for sufficiently large integers n.
In Section 4.1, we observed that the function t∗ 7→ det(Mk (p, t∗ )) is a piecewise polynomial when
p is an integer. This is what allowed us to analyze this case relatively easily (in both Section 4.1 and
in Section 4.2). For non-integer p, the function t∗ 7→ det(Mk (p, t∗ )) is much less nice. So, instead of
holding p fixed and varying t∗ , we will be interested in studying the function fk,t∗ (p) := det(Mk (p, t∗ ))
for fixed t∗ and k. We first observe that this function has a fairly nice structure.
21
Lemma 4.10. For any t∗ ∈ R, integer k ≥ 1, and p ≥ 1, let
fk,t∗ (p) := det(Mk (p, t∗ )) ,
(12)
where Mk (p, t∗ ) is as defined in Eq. (10). Then, for fixed k, t∗ , fk,t∗ (p) is a Dirichlet polynomial.
I.e., there are some real numbers b0 , . . . , br , c0 , . . . , cr ∈ R (depending on t∗ and k) such that
fk,t∗ (p) =
r
X
bi exp(ci p)
(13)
i=0
for some finite r.
Proof. To see that fk,t∗ (p) is a Dirichlet polynomial for fixed t∗ , k, it suffices to note that (1) each
entry of Mk (p, t∗ ) is a Dirichlet polynomial; (2) the determinant of a matrix can be written as a
polynomial in the coordinates; and (3) a polynomial of Dirichlet polynomials is itself a Dirichlet
polynomial.
Corollary 4.11. There is an efficient algorithm that takes as input k ≥ 2 and any p ≥ 1 and either
k
k
fails or outputs V ∈ R2 ×k and t∗ ∈ R2 that define a (p, k)-isolating parallelepiped. Furthermore,
for any fixed k ≥ 2, the algorithm only fails for finitely many choices of p ≥ 1.
Proof. By Corollary 4.7, for any k ≥ 2, we can find a t∗ ∈ Q such that, say, fk,t∗ (1) 6= 0, where
fk,t∗ (p) is defined as in Eq. (12). Clearly, fk,t∗ (p) is non-zero as a function of p for these values
of t∗ , k. Furthermore, by Lemma 4.10, fk,t∗ (p) is a Dirichlet polynomial. The result follows by
the fact that any non-zero Dirichlet polynomial has finitely many roots (see, e.g., Theorem 3.1
in [Jam06]).
Theorem 4.12. There is an efficient algorithm that takes as input an integer k ≥ 2 and any p ≥ 1
k
k
and either fails or outputs V ∈ R2 ×k and t∗ ∈ R2 that define a (p, k)-isolating parallelepiped.
Furthermore, for any fixed k ≥ 2, the algorithm only fails on finitely many values of p ≥ 1.8
Proof. The result follows immediately from Proposition 4.4 and Corollary 4.11.
Item 2 of Theorem 1.2 now follows immediately from Theorem 3.2 and Theorem 4.12.
We now provide what amounts to a different interpretation of the above.
Lemma 4.13. There is an efficient algorithm that takes as input any p0 ≥ 1 and an integer k ≥ 2
and outputs a value t∗ such that fk,t∗ (p0 + δ) and fk,t∗ (p0 − δ) are non-zero for sufficiently small
δ > 0, where fk,t∗ (p) is as defined in Eq. (12).
Proof. The algorithm simply calls the procedure from Corollary 4.7 with, say, p = 1 and outputs
the result. I.e., it suffices to choose any t∗ such that fk,t∗ (1) 6= 0. As in the proof of Corollary 4.11,
we observe that the function fk,t∗ (p) is zero on a finite set of values X. The result then follows by
taking δ := minx∈X\{p0 } |x − p0 |/2 if X \ {p0 } is non-empty, and δ := c for any c > 0 otherwise.
Finally, we derive the main theorem of this section.
8
As we observed in Section 4.2, the set of failure points necessarily includes all even integers p ≤ k − 1.
22
Theorem 4.14. For any efficiently computable δ(n) 6= 0 that converges to zero as n → ∞ and p0 ≥ 1,
there is an efficient algorithm that takes as input an integer k ≥ 2 and sufficiently large positive
k
k
integer n and outputs a matrix V ∈ R2 ×k and vector t∗ ∈ R2 that define a (p0 + δ(n), k)-isolating
parallelepiped.
Proof. The result follows immediately from Proposition 4.4 and Lemma 4.13. In particular, the
algorithm runs the procedure from Lemma 4.13, receiving as output some t∗ ∈ R such that
fk,t∗ (p0 ± ε) is non-zero for sufficiently small ε > 0. In particular, if n is sufficiently large, then
fk,t∗ (p0 + δ(n)) will be non-zero. The result then follows from Proposition 4.4.
Item 3 of Theorem 1.2 now follows from Theorem 3.2 and Theorem 4.14.
5
Gap-ETH-based Hardness of Approximation
In this section we prove Gap-ETH-based hardness of approximation for CVPp for every p ∈ [1, ∞).
We also show a stronger hardness of approximation result for CVP1 . (In Section 6.3, we additionally
show a stronger result for p = ∞.) The main idea behind our proof is to use the reduction from
Max-2-SAT to CVP described in Section 1.5 without the “identity matrix gadget” that we used
to force any closest vector to be a 0-1 combination of basis vectors. We will show that this is
permissible because the resulting CVP instance always has a 0-1 combination of basis vectors that
is at least as close to the target as any other lattice vector.
Theorem 5.1. For every 0 ≤ δ < ε ≤ 1 and every p ∈ [1, ∞), there is a polynomial-time reduction
from any (δ, ε)-Gap-2-SAT instance with n variables to an instance of γ-CVPp of rank n, where
γ :=
δ + (1 − δ) · 3p
ε + (1 − ε) · 3p
1/p
>1.
In particular, when ε = 1, the corresponding approximation factor is (δ + (1 − δ) · 3p )1/p .
Proof. Given a (δ, ε)-Gap-2-SAT instance with n variables x1 , . . . , xn and m clauses C1 , . . . , Cm , we
construct a CVPp instance (B, t, r) for some fixed p ∈ [1, ∞) as follows. Let B ∈ Zm×n be the basis
defined by
2 if Ci contains xj ,
−2 if Ci contains ¬xj ,
bi,j :=
0 otherwise.
Let t ∈ Rm be the target vector defined by ti := 3 − 2|Ni |, and let r := (εm + (1 − ε)m · 3p )1/p .
We claim that there is always a 0-1 combination of basis vectors which is a closest vector to t.
Assuming this claim, we analyze kBy − tkp only for y ∈ {0, 1}n without loss of generality, while
deferring the proof of the claim until the end.
Let y ∈ {0, 1}n be an assignment to the variables of Φ. For every 1 ≤ i ≤ m,
|(By − t)i | = 2
X
= 2
X
yind(`i,s ) −
s∈Pi
X
yind(`i,s ) − (3 − 2|Ni |)
s∈Ni
yind(`i,s ) +
s∈Pi
X
s∈Ni
= 2 · |Si (y)| − 3 .
23
(1 − yind(`i,s ) ) − 3
(14)
The last expression is equal to 1 if y satisfies Ci (i.e. if |Si (y)| ≥ 1) and 3 if not. We therefore have
kBy −
tkpp
=
m
X
|(By − t)i |p = m+ (y) + (m − m+ (y)) · 3p .
i=1
Therefore, if val(Φ) ≥ ε then there exists y ∈ {0, 1}n such that kBy − tkpp ≤ εm + (1 − ε)m · 3p = rp ,
and if val(Φ) < δ, then for every y ∈ {0, 1}n , it holds that kBy − tkpp > δm + (1 − δ)m · 3p = γ p rp .
It follows that the reduction achieves the claimed approximation factor of γ.
It remains to prove the claim that there is always a 0-1 combination of basis vectors which is a
closest vector to t. We show this by demonstrating that for every y ∈ Zn there exists χ(y) ∈ {0, 1}n
such that kB · χ(y) − tkp ≤ kBy − tkp . Given y ∈ Zn , let χ(y) ∈ {0, 1}n denote the vector whose
ith coordinate is set to 1 if yi ≥ 1 and is set to 0 otherwise.
Fix y ∈ Zn and 1 ≤ i ≤ n, and refer to Equation (14). If χ(y) satisfies Ci then |(By − t)i | ≥
|(B·χ(y)−t)i | = 1. On the other hand, if χ(y) does not satisify Ci , then yind(`i,s ) ≤ 0 for all s ∈ Pi and
1 − yind(`i,s ) ≤ 0 for all s ∈ Ni so that |(By − t)i | ≥ 3 and therefore |(By − t)i | ≥ |(B · χ(y) − t)i | = 3.
Combining these cases it follows that for all y ∈ Zn and 1 ≤ i ≤ n, |(B · χ(y) − t)i | ≤ |(By − t)i |,
and therefore kB · χ(y) − tkp ≤ kBy − tkp , proving the claim.
Corollary 5.2. For every p ∈ [1, ∞) and 0 < δ < δ 0 < 1, there is no 2o(n) -time algorithm for
γ-CVPp with
6 + δ 0 + (4 − δ 0 ) · 3p 1/p
γ :=
>1
7 + 3p+1
unless randomized Gap-ETH (with respect to (δ, 1)-Gap-3-SAT) fails.
Proof. Concatenate the reductions described in Proposition 2.11, Proposition 2.12, and Theorem 5.1.9
5.1
A stronger result for `1
In Theorem 5.1 we showed that there was no need to use the identity matrix gadget in our reduction
from Gap-2-SAT to CVP. An interesting question is whether the identity matrix gadget is also
unnecessary in our reduction in Theorem 3.2 from Gap-k-SAT to CVPp using (p, k)-isolating
parallelepipeds. If so, then we get stronger “Gap-SETH-hardness” for CVPp for all p for which
there exist (p, k)-isolating parallelepipeds for infinitely many k ∈ Z+ .
Although we do not know how to show this in general, we are able to get stronger hardness of
approximation for CVP1 in this way by using the family of (1, k)-isolating parallelepipeds described
in Corollary 3.4. The analysis is similar to the analysis in Theorem 5.1. In particular, a 0-1
combination of basis vectors will always be at least as close to the target vector as any other lattice
vector will.
Theorem 5.3. For every 0 ≤ δ < ε ≤ 1, there is a polynomial-time reduction from any (δ, ε)-Gapk-SAT instance with n variables to an instance of γ-CVP1 of rank n, where
γ :=
δ · (k − 1) + (1 − δ) · (k + 1)
ε · (k − 1) + (1 − ε) · (k + 1)
9
> 1.
Note that the Gap-2-SAT instance output by Proposition 2.12 contains 1-clauses as well as 2-clauses. We therefore
stress that the reduction described in Theorem 5.1 works for this case as well.
24
Proof. Given a (δ, ε)-Gap-k-SAT instance with n variables x1 , . . . , xn and m clauses C1 , . . . , Cm , we
construct a CVP1 instance (B, t, r) as follows. The basis B ∈ Z(2m)×n and target vector t ∈ Z2m in
the output instance have the form
B1
..
B = . ,
Bm
t1
..
t = . ,
tm
with blocks Bi ∈ Z2×n and ti ∈ Z2 for 1 ≤ i ≤ m. For every 1 ≤ i ≤ m and 1 ≤ j ≤ n, set the jth
column (Bi )j of block Bi (corresponding to the clause Ci = ∨ks=1 `i,s ) as
(1, 1)T
−(1, 1)T
(Bi )j :=
(0, 0)T
if Ci contains xj ,
if Ci contains ¬xj ,
otherwise,
and set ti := (k, 1)T − |Ni | · (1, 1)T . Set r := εm(k − 1) + (1 − ε) · m(k + 1).
We claim that there is always a 0-1 combination of basis vectors that is closest to t. Assuming
the claim, we analyze kBy − tk only for y ∈ {0, 1}n without loss of generality, while deferring the
proof of the claim until the end.
Let y ∈ {0, 1}n be an assignment to the variables of Φ. For every 1 ≤ i ≤ m,
kBi y − ti k1 =
=
X
yind(`i,s ) · (1, 1)T −
X
s∈Pi
s∈Ni
X
X
yind(`i,s ) · (1, 1)T +
s∈Pi
yind(`i,s ) · (1, 1)T − (k, 1)T − |Ni | · (1, 1)T
1 − yind(`i,s ) · (1, 1)T − (k, 1)T
s∈Ni
= |Si (y)| · (1, 1)T − (k, 1)T
1
1
1
(15)
.
The last expression is equal to k − 1 if y satisfies Ci (i.e. if |Si (y)| ≥ 1) and k + 1 if not. We
therefore have
kBy − tk1 =
m
X
kBi y − ti k1 = m+ (y) · (k − 1) + (m − m+ (y)) · (k + 1).
i=1
Therefore, if val(Φ) ≥ ε then there exists y ∈ {0, 1}n such that kBy − tk1 ≤ εm(k − 1) + (1 −
ε) · m(k + 1) = r, and if val(Φ) < δ then for every y ∈ {0, 1}n , it holds that kBy − tk1 >
δm(k − 1) + (1 − δ)m(k + 1) = γr. It follows that the reduction achieves the claimed approximation
factor of γ.
It remains to prove the claim that there is always a 0-1 combination of basis vectors that is a
closest vector to t. We show this by demonstrating that for every y ∈ Zn there exists χ(y) ∈ {0, 1}n
such that kBχ(y) − tk1 ≤ kBy − tk1 . Given y ∈ Zn , let χ(y) ∈ {0, 1}n denote the vector whose
coordinate is set to 1 if yi ≥ 1 and is set to 0 otherwise.
Note that kc(1, 1)T − (k, 1)T k1 = k − 1 for all c ∈ [k], and kc(1, 1)T − (k, 1)T k1 ≥ k + 1 for
all c ∈ Z \ [k]. Fix y ∈ Zn and 1 ≤ i ≤ n, and refer to Equation (15). If χ(y) satisfies Ci
then kBi · y − ti k1 ≥ kBi · χ(y) − ti k1 = k − 1. On the other hand, if χ(y) does not satisfy
Ci , then yind(`i,s ) ≤ 0 for all s ∈ Pi and 1 − yind(`i,s ) ≤ 0 for all s ∈ Ni so that kBi · y − ti k1 ≥
25
kBi · χ(y) − ti k1 = k + 1. Combining these cases it follows that for all y ∈ Zn and 1 ≤ i ≤ n,
kBi · χ(y) − ti k1 ≤ kBi · y − ti k1 , and therefore kB · χ(y) − tk1 ≤ kBy − tk1 , proving the claim.
6
Additional hardness results
In this section we prove a number of additional results about the quantitative hardness of CVP and
related problems. In Section 6.1, we give a reduction from Max-2-SAT to CVPPp for all p ∈ [1, ∞),
proving Theorem 1.4. In Section 6.2, we give a reduction from Max-k-SAT (and in particular
Max-2-SAT) to CVPp for all p ∈ [1, ∞), proving Theorem 1.5. Finally, in Section 6.3 we give a
reduction from k-SAT to CVP∞ and SVP∞ , proving the special case of Theorem 1.1 for p = ∞.
Our reductions all use the same high-level idea as the reduction given in Theorem 3.2, but each
uses new ideas as well. Throughout this section we adopt the notation from Section 3.
6.1
Hardness of CVPPp
In this section, we prove ETH-hardness of CVPP. To do this, for every n, we construct a single
lattice Ln ⊂ Rd of rank O(n2 ), such that for every n-variable instance of Max-2-SAT, there exists
an efficiently computable t ∈ Rd that is close to the lattice if and only if Φ is satisfiable. Clearly,
any efficient algorithm for CVPP on this lattice would imply a similarly efficient algorithm for
Max-2-SAT (and also 3-SAT, as described below).
Our basis Bn for Ln will encode all possible O(n2 ) clauses of a Max-2-SAT instance on n
variables, together with a gadget that will allow us to “switch on or off” each clause by only changing
the coordinates of the target vector t. (This gadget costs us a quadratic blow-up in the lattice rank.)
Then, given an instance (Φ, W ) of Max-2-SAT, we define the target vector t such that it “switches
on” all clauses from Φ and “switches off” all the remaining clauses.
Lemma 6.1. For every p ∈ [1, ∞), there is a pair of polynomial-time algorithms (P, Q) (in analogy
to the definition of CVPP) that behave as follows.
1. On input an integer n ≥ 1, P outputs a basis Bn ∈ Rd×N of a rank N lattice Ln ⊂ Rd , where
d = d(n) = O(n2 ) and N = N (n) = O(n2 ).
2. On input a Max-2-SAT instance with n variables, Q outputs a target vector t ∈ Rd and a
distance bound r ≥ 0 such that distp (t, Ln ) ≤ r if and only if the input is a ‘YES’ instance.
Proof. Let M = 4 n2 = O(n2 ) be the total possible number of 2-clauses on n variables, and let
C1 , . . . , CM denote those clauses.
The algorithm P constructs the basis Bn ∈ Rd×N , where d := n + 2M, N := n + M , as
(bT1 , cT1 )
..
.
B :=
,
(bTM , cTM )
2α1/p IN
with rows (bTi , cTi ) of B satisfying bi ∈ Rn and ci ∈ RM for 1 ≤ i ≤ M , and where α := 2p M .
For every 1 ≤ i ≤ M and 1 ≤ j ≤ n, set the jth coordinate of bi (corresponding to the clause
Ci = `i,1 ∨ `i,2 ) as
26
(bi )j :=
1 if xj appears in the ith clause,
−1 if ¬xj appears in the ith clause,
0 otherwise.
For every 1 ≤ i ≤ M , set cTi := (0T(i−1) , 1, 0T(M −i) ), i.e., set (c1 , . . . , cM ) = IM .
Given an instance (Φ, W ) of Max-2-SAT with m clauses, the algorithm Q outputs
r := (M − m + W ) · 1/2p + (m − W ) · (3/2)p + α(n + M − m)
and t ∈ Rd defined as
u1
..
.
1/p
uM
t := α1/p · 1n
v1
..
.
,
vM
where for 1 ≤ i ≤ M , ui = 3/2 − |Ni |, and vi = 0 if the clause Ci appears in the formula Φ and
vi = α1/p otherwise.
Clearly both algorithms are efficient. We now analyze for which y ∈ ZN we have kBy − tkp ≤ r.
Note that the vector v = (v1 , . . . , vM )T has exactly m coordinates equal to zero, and M − m
coordinates equal to α1/p . Given y ∈
/ {0, 1}N , we have
kBy − tkpp ≥ k2α1/p IN y − (α1/p 1Tn , v T )T kpp ≥ α(n + M − m + 1) ≥ 2p M + α(M + n − m) > rp .
Furthermore, if y ∈ {0, 1}N has a non-zero coordinate yn+M +i (for 1 ≤ i ≤ M ) at a position
corresponding to a tn+M +i = 0 in t (i.e., Ci ∈ Φ), then again kBy − tkpp ≥ α(n + M − m + 1) > rp .
So, we can restrict our attention to y ∈ {0, 1}N with yn+M +i = 0 whenever Ci ∈ Φ.
0 )T ∈ {0, 1}M ,
Consider an assignment a ∈ {0, 1}n to the n variables of Φ. Take (y 0 )T = (y10 , . . . yM
T
0
T
T
and set y := (a , (y ) ) . Then, for 1 ≤ i ≤ M ,
h(bTi , cTi )T , yi − ti =
X
yind(`i,s ) −
s∈Pi
=
X
s∈Pi
X
yind(`i,s ) + hci , y 0 i − (3/2 − |Ni |)
s∈Ni
yind(`i,s ) −
X
(1 − yind(`i,s ) ) + yi0 − 3/2
s∈Ni
= |Si (a)| + yi0 − 3/2 .
If Ci ∈
/ Φ, then there exists yi0 ∈ {0, 1}, such that |h(bTi , cTi )T , yi − ti | = 1/2. Moreover, the choice
0
of yi does not affect |h(bTi , cTi )T , yi − ti0 | for i0 6= i. If Ci ∈ Φ and |Si (a)| > 0 for yi0 = 0, then
|h(bTi , cTi )T , yi − ti | = 1/2. On the other hand, if Ci ∈ Φ and |Si (a)| = 0, then yi0 = 0 implies
|h(bTi , cTi )T , yi − ti | ≥ 3/2.
Because |Si (a)| ≥ 1 if and only if Ci is satisfied, we see that the following holds if and only if
the number of satisfied clauses m+ (a) is at least W :
27
There exists a y 0 such that, setting y := (a, y 0 ), we have
kBy −
tkpp
=
M
X
|h(bTi , cTi )T , yi − ti |p + α(n + M − m)
i=1
= (M − m + m+ (a)) · (1/2)p + (m − m+ (a)) · (3/2)p + α(n + M − m)
≤ (M − m + W ) · (1/2)p + (m − W ) · (3/2)p + α(n + M − m)
= rp .
Therefore, there exists y with kBy − tkp ≤ r if and only if (Φ, W ) is a ‘YES’ instance of
Max-2-SAT.
Proof of Theorem 1.4. The main statement follows from Lemma 6.1. The “in particular” part
follows from Lemma 6.1 and the existence of a reduction from 3-SAT with n variables clauses to
(many) Max-2-SAT instances with O(n) variables. Indeed, such a reduction follows by applying
the Sparsification Lemma (Proposition 2.9), and then reducing each sparse 3-SAT instance to a
Max-2-SAT instance with only a linear blow-up in the number of variables (see, e.g., the reduction
in [GJS76, Theorem 1.1]).
6.2
ETH- and Max-2-SAT-hardness for all p ∈ [1, ∞)
Theorem 6.2. For every p = p(n) ∈ [1, ∞) and k ≥ 2 there is a polynomial-time reduction from
any (Weighted-)Max-k-SAT instance with n variables and m clauses to a CVPp instance of rank
n + (k − 2)m.
Proof. For simplicity we give a reduction from unweighted Max-k-SAT. The modification sketched
for reducing from Weighted Max-k-SAT in Theorem 3.2 works here as well. Namely, we give a
reduction from a Max-k-SAT instance (Φ, W ) to an instance (B, t, r) of CVPp . Here the formula Φ
is on n variables x1 , . . . , xn and m clauses C1 , . . . , Cm . (Φ, W ) is a ‘YES’ instance if there exists an
assignment a such that m+ ≥ W . We assume without loss of generality that each variable appears
at most once per clause. We define the output CVPp instance as follows. Let d := n + (k − 1)m
and let N := n + (k − 2)m. The basis B ∈ Rd×N in the output instance has the form
(bT1 , cT1 )
..
.
B=
T T ,
(bm , cm )
2α1/p · IN
t=
t1
..
.
tm
α1/p · 1N
with rows (bTi , cTi ) of B satisfying bi ∈ Rn and ci ∈ Rm(k−2) for 1 ≤ i ≤ m, and where α :=
W · ( 12 )p + (m − W ) · ( 32 )p . For every 1 ≤ i ≤ m and 1 ≤ j ≤ n, set the jth coordinate of bi
(corresponding to the clause Ci = ∨ks=1 `i,s ) as
1 if xj is in the ith clause,
−1 if ¬xj is in the ith clause,
(bi )j :=
0 otherwise.
For every 1 ≤ i ≤ m, set cTi := (0T(i−1)(k−2) , −1Tk−2 , 0T(m−i)(k−2) ). I.e., each ci has a block of −1s of
length k−2, and ci , ci0 are coordinate-wise disjoint for i 6= i0 . For every 1 ≤ i ≤ m set ti := 3/2−|Ni |.
Finally, set r := (α(N + 1))1/p .
28
We next analyze for which y ∈ ZN it holds that kBy − tkp ≤ r. Given y ∈
/ {0, 1}N , kBy − tkpp ≥
k2α1/p IN y − α1/p 1N kpp ≥ α(N + 2) > rp , so we only need to analyze the case when y ∈ {0, 1}n .
Consider an assignment a ∈ {0, 1}n to the variables of Φ, take (y 0 )T = ((y 01 )T , . . . , (y 0m )T ) ∈
{0, 1}(k−2)m with y 0i ∈ {0, 1}k−2 for 1 ≤ i ≤ m, and set y T := (aT , (y 0 )T ). Then, for 1 ≤ i ≤ m,
h(bTi , cTi )T , yi − ti =
X
yind(`i,s ) −
s∈Pi
=
X
X
yind(`i,s ) + hci , y 0i i − (3/2 − |Ni |)
s∈Ni
yind(`i,s ) +
s∈Pi
= |Si (a)| −
X
(1 − yind(`i,s ) ) − ky 0i k1 − 3/2
s∈Ni
ky 0i k1
− 3/2 .
It follows that if |Si (a)| = 0 then |h(bTi , cTi )T , yi − ti | ≥ 32 . On the other hand, if |Si (a)| ≥ 1,
then there exists y 0i ∈ {0, 1}k−2 such that |h(bTi , cTi )T , yi − ti | = 12 . Indeed, picking any y 0i with
Hamming weight |Si (a)| − 2 or |Si (a)| − 1 achieves this. Moreover, the choice of y 0i does not affect
|h(bTi , cTi )T , yi − ti0 | for i0 6= i.
Because |Si (a)| ≥ 1 if and only if Ci is satisfied, we see that the following holds if and only if
the number of satisfied clauses m+ (a) is at least W :
There exists a y 0 such that, setting y := (a, y 0 ), we have
kBy − tkpp =
m
X
|h(bTi , cTi )T , yi − ti |p + αN
i=1
+
= m (a) · (1/2)p + (m − m+ (a)) · (3/2)p + αN
≤ W · (1/2)p + (m − W ) · (3/2)p + αN
= rp .
Therefore, there exists y such that kBy − tkp ≤ r if and only if (Φ, W ) is a ‘YES’ instance of
Max-k-SAT.
We remark that a straightforward modification of the preceding reduction gives a reduction from
an instance of Max-k-SAT with k ≥ 3 on n variables and m clauses to a CVPp instance of rank
n + (blog2 (k − 2)c + 1)m (as opposed to rank n + (k − 2)m). The idea is to encode the value k − 2
(corresponding to the row-specific blocks −1k−2 used in the reduction) in binary rather than unary.
Corollary 6.3. For every p ∈ [1, ∞) there is no 2o(n) -time algorithm for CVPp assuming ETH.
Proof. The claim follows by combining the Sparsification Lemma (Proposition 2.9) with the reduction
in Theorem 6.2.10
When k = 2, the rank of the CVPp instance output by the reduction in Theorem 6.2 is n.
Therefore, we get the following corollary.
Corollary 6.4. If there exists a 2(ω/3−ε)n -time (resp. 2(1−ε)n -time and polynomial space) algorithm
for CVPp for any p ∈ [1, ∞) and for any constant ε > 0, then there exists a 2(ω/3−ε)n -time (resp.
2(1−ε)n -time and polynomial space) algorithm for (Weighted-)Max-2-SAT and (Weighted-)Max-Cut.
10
As a technical point, we must reduce from k-SAT rather than Max-k-SAT to show hardness under ETH because
the Sparsification Lemma works for k-SAT.
29
6.3
The hardness of SVP∞ and CVP∞
Theorem 6.5. There is a polynomial-time reduction from a k-SAT instance with n variables to a
CVP∞ instance of rank n.
Proof. We give a reduction from a k-SAT formula Φ with n variables x1 , . . . , xn and m clauses
C1 , . . . , Cm to an instance (B, t, r) of CVP∞ . We assume without loss of generality that each
variable appears at most once per clause. We define the output CVP∞ instance as follows. Let
d := m + n. The basis B ∈ Rd×n in the output instance has the form
bT1
..
.
B=
bTm
(k + 1) · In
t1
..
.
,
t=
tm
(k + 1)/2 · 1n
,
with rows bi of B satisfying bi ∈ Rn for 1 ≤ i ≤ m. For every 1 ≤ i ≤ m, set bi as in the proof of
Theorem 6.2 and set ti := (k + 1)/2 − |Ni |. Set r := (k − 1)/2.
We next analyze for which y ∈ Zn it holds that kBy − tk∞ ≤ r. Given y ∈
/ {0, 1}n , kBy − tk∞ ≥
k(k − 1) · In y − (k − 1)/2 · 1n k∞ ≥ 3(k − 1)/2 > r, so we only need to analyze the case when
y ∈ {0, 1}n . Consider an assignment y ∈ {0, 1}n to the variables of Φ. Then
hbi , yi − ti =
X
yind(`i,s ) −
s∈Pi
=
X
X
yind(`i,s ) − ((k + 1)/2 − |Ni |)
s∈Ni
yind(`i,s ) −
s∈Pi
X
(1 − yind(`i,s ) ) − (k + 1)/2
s∈Ni
= |Si (a)| − (k + 1)/2 .
It follows that if |Si (a)| = 0 then |hbi , yi − ti | = (k + 1)/2, and otherwise |hbi , yi − ti | ≤ (k − 1)/2.
Because |Si (a)| ≥ 1 if and only if Ci is satisfied, we then have that kBy − tk∞ = max{|hb1 , yi −
t1 |, . . . , |hbm , yi − tm |, (k − 1)/2} = (k − 1)/2 = r if y satisfies Φ, and kBy − tk∞ = (k + 1)/2 > r
otherwise. Therefore there exists y such that kBy − tk∞ ≤ r if and only if Φ is satisfiable.
Lemma 6.6. There is a polynomial-time reduction from a k-SAT instance with n variables to an
SVP∞ instance of rank n + 1.
Proof. We give a reduction from a k-SAT formula Φ with n variables x1 , . . . , xn and m clauses
C1 , . . . , Cm to an instance (B 0 , r) of SVP∞ . Define the basis B 0 ∈ R(m+n+1)×(n+1) as
0
B :=
B
0Tn
!
−t
,
−(k − 1)/2
where B and t are as defined in the proof of Theorem 6.5, and set r := (k − 1)/2. We consider for
which y ∈ Zn+1 \ {0n+1 } it holds that kByk∞ ≤ r. It is not hard to check that if |yi | ≥ 2 for some
1 ≤ i ≤ n + 1, or if the signs of yi and yn+1 differ for some 1 ≤ i ≤ n, then kByk∞ > r. Therefore
we need only consider y of the form y = ±(aT , 1)T where a ∈ {0, 1}n . But for such a y we have
that kB 0 yk∞ = kBa − tk∞ , and kBa − tk∞ ≤ (k − 1)/2 if and only if a is a satisfying assignment
to Φ by the analysis in the proof of Theorem 6.5.
30
Corollary 6.7. For any constant ε > 0 there is no 2(1−ε)n -time algorithm for SVP∞ or CVP∞
assuming SETH, and there is no 2o(n) -time algorithm for SVP∞ or CVP∞ assuming ETH.
Proof. Combine Theorem 6.5 and Lemma 6.6.
Note that the preceding reduction in fact achieves an approximation factor of γ = γ(k) :=
1 + 2/(k − 1). This implies that for every constant ε > 0, there is a γε > 1 such that no 2(1−ε)n -time
algorithm that approximates SVP∞ or CVP∞ to within a factor of γε unless SETH fails.
Finally, we remark that the reduction given in Theorem 6.2 is parsimonious when used as a
reduction from 2-SAT. I.e., there is a one-to-one correspondence between satisfying assignments in
the input instance and close vectors in the output instance. The reductions given in Theorem 6.5
and Lemma 6.6 are also parsimonious.11
Because #2-SAT is #P-hard, our reductions therefore show that the counting version of CVPp
(called the Vector Counting Problem) is #P-hard for all 1 ≤ p ≤ ∞, and that the counting version of
SVP∞ is #P-hard. This improves (and arguably simplifies the proof of) a result of Charles [Cha07],
which showed that the counting version of CVP2 is #P-hard.
Acknowledgments
We would like to thank Oded Regev for many fruitful discussions and for helpful comments on an
earlier draft of this work. We also thank Vinod Vaikuntanathan for recommending that we consider
Gap-ETH.
References
[ABSS93] Sanjeev Arora, László Babai, Jacques Stern, and Z. Sweedyk. The hardness of approximate optima in lattices, codes, and systems of linear equations. In FOCS, 1993.
[ABW15]
Amir Abboud, Arturs Backurs, and Virginia Vassilevska Williams. Tight hardness
results for LCS and other sequence similarity measures. In FOCS, 2015.
[ADPS16] Erdem Alkim, Léo Ducas, Thomas Pöppelmann, and Peter Schwabe. Post-quantum key
exchange — A new hope. In USENIX Security Symposium, 2016.
[ADRS15] Divesh Aggarwal, Daniel Dadush, Oded Regev, and Noah Stephens-Davidowitz. Solving
the Shortest Vector Problem in 2n time via discrete Gaussian sampling. In STOC, 2015.
[ADS15]
Divesh Aggarwal, Daniel Dadush, and Noah Stephens-Davidowitz. Solving the Closest
Vector Problem in 2n time— The discrete Gaussian strikes again! In FOCS, 2015.
[AJ08]
V. Arvind and Pushkar S. Joglekar. Some sieving algorithms for lattice problems. In
IARCS Annual Conference on Foundations of Software Technology and Theoretical
Computer Science, 2008.
[Ajt98]
Miklós Ajtai. The Shortest Vector Problem in L2 is NP-hard for randomized reductions.
In STOC, 1998.
11
Actually, in the case of SVP∞ , each satisfying assignment to the SAT formula corresponds to a pair ±v of shortest
non-zero vectors, so that there are exactly twice as many such vectors as there are satisfying assignments.
31
[Ajt04]
Miklós Ajtai. Generating hard instances of lattice problems. In Complexity of computations and proofs, volume 13 of Quad. Mat., pages 1–32. Dept. Math., Seconda Univ.
Napoli, Caserta, 2004. Preliminary version in STOC’96.
[AKKV11] Mikhail Alekhnovich, Subhash Khot, Guy Kindler, and Nisheeth K. Vishnoi. Hardness
of approximating the Closest Vector Problem with Pre-processing. Computational
Complexity, 20, 2011.
[AKS01]
Miklós Ajtai, Ravi Kumar, and D. Sivakumar. A sieve algorithm for the Shortest Lattice
Vector Problem. In STOC, 2001.
[AKS02]
Miklós Ajtai, Ravi Kumar, and D. Sivakumar. Sampling short lattice vectors and the
Closest Lattice Vector Problem. In CCC, 2002.
[APS15]
Martin R. Albrecht, Rachel Player, and Sam Scott. On the concrete hardness of Learning
with Errors. J. Mathematical Cryptology, 9(3):169–203, 2015.
[AR05]
Dorit Aharonov and Oded Regev. Lattice problems in NP ∩ coNP. J. ACM, 52(5):749–
765, 2005. Preliminary version in FOCS 2004.
[BCD+ 16] Joppe W. Bos, Craig Costello, Léo Ducas, Ilya Mironov, Michael Naehrig, Valeria
Nikolaenko, Ananth Raghunathan, and Douglas Stebila. Frodo: Take off the ring!
Practical, quantum-secure key exchange from LWE. In CCS, 2016.
[BD15]
Nicolas Bonifas and Daniel Dadush. Short paths on the Voronoi graph and the Closest
Vector Problem with Preprocessing. In SODA, 2015.
[BDGL16] Anja Becker, Léo Ducas, Nicolas Gama, and Thijs Laarhoven. New directions in nearest
neighbor searching with applications to lattice sieving. In SODA, 2016.
[BI15]
Arturs Backurs and Piotr Indyk. Edit Distance cannot be computed in strongly
subquadratic time (unless SETH is false). In STOC, 2015.
[BN09]
Johannes Blömer and Stefanie Naewe. Sampling methods for shortest vectors, closest
vectors and successive minima. Theoret. Comput. Sci., 410(18):1648–1665, 2009.
[CDL+ 12] Marek Cygan, Holger Dell, Daniel Lokshtanov, Dániel Marx, Jesper Nederlof, Yoshio
Okamoto, Ramamohan Paturi, Saket Saurabh, and Magnus Wahlström. On problems
as hard as CNF-SAT. In CCC, 2012.
[CFK+ 15] Marek Cygan, Fedor V. Fomin, Lukasz Kowalik, Daniel Lokshtanov, Dániel Marx,
Marcin Pilipczuk, Michal Pilipczuk, and Saket Saurabh. Parameterized Algorithms.
Springer, 2015.
[Cha07]
Denis Xavier Charles. Counting lattice vectors. J. Comput. Syst. Sci., 73(6):962–972,
2007.
[CLR+ 14] Shiri Chechik, Daniel H. Larkin, Liam Roditty, Grant Schoenebeck, Robert Endre
Tarjan, and Virginia Vassilevska Williams. Better approximation algorithms for the
graph diameter. In SODA, 2014.
32
[CN98]
Jin-Yi Cai and Ajay Nerurkar. Approximating the SVP to within a factor (1 + 1/ dimε )
is NP-hard under randomized conditions. In CCC, 1998.
[Dad12]
Daniel Dadush. A O(1/ε2 )n -time sieving algorithm for approximate Integer Programming.
In LATIN, 2012.
[Din16]
Irit Dinur. Mildly exponential reduction from gap 3SAT to polynomial-gap label-cover.
Electronic Colloquium on Computational Complexity (ECCC), 23:128, 2016.
[DKRS03] Irit Dinur, Guy Kindler, Ran Raz, and Shmuel Safra. Approximating CVP to within
almost-polynomial factors is NP-hard. Combinatorica, 23(2):205–243, 2003.
[DPV11]
Daniel Dadush, Chris Peikert, and Santosh Vempala. Enumerative lattice algorithms in
any norm via M-ellipsoid coverings. In FOCS, 2011.
[DRS14]
Daniel Dadush, Oded Regev, and Noah Stephens-Davidowitz. On the Closest Vector
Problem with a distance guarantee. In CCC, pages 98–109, 2014.
[FM04]
Uriel Feige and Daniele Micciancio. The inapproximability of lattice and coding problems
with preprocessing. Journal of Computer and System Sciences, 69(1):45–67, 2004.
Preliminary version in CCC 2002.
[GG00]
Oded Goldreich and Shafi Goldwasser. On the limits of nonapproximability of lattice
problems. J. Comput. Syst. Sci., 60(3):540–563, 2000. Preliminary version in STOC
1998.
[GHNR03] Jens Gramm, Edward A. Hirsch, Rolf Niedermeier, and Peter Rossmanith. Worst-case
upper bounds for MAX-2-SAT with an application to MAX-CUT. Discrete Applied
Mathematics, 130(2):139–155, 2003.
[GJS76]
Michael R. Garey, David S. Johnson, and Larry Stockmeyer. Some simplified NPcomplete graph problems. Theoretical Computer Science, 1(3):237–267, 1976.
[GMSS99] Oded Goldreich, Daniele Micciancio, Shmuel Safra, and Jean-Pierre Seifert. Approximating shortest lattice vectors is not harder than approximating closest lattice vectors.
Information Processing Letters, 71(2):55 – 61, 1999.
[GN08]
Nicolas Gama and Phong Q. Nguyen. Finding short lattice vectors within Mordell’s
inequality. In STOC, 2008.
[GPV08]
Craig Gentry, Chris Peikert, and Vinod Vaikuntanathan. Trapdoors for hard lattices
and new cryptographic constructions. In STOC, pages 197–206, 2008.
[HP]
S. Har-Peled. Concentration of Random Variables – Chernoff’s Inequality. Available at
http://sarielhp.org/teach/13/b_574_rand_alg/lec/07_chernoff.pdf.
[HR12]
Ishay Haviv and Oded Regev. Tensor-based hardness of the Shortest Vector Problem to
within almost polynomial factors. Theory of Computing, 8(23):513–531, 2012. Preliminary
version in STOC’07.
33
[IP99]
Russell Impagliazzo and Ramamohan Paturi. On the complexity of k-SAT. In CCC,
pages 237–240, 1999.
[IPZ01]
Russell Impagliazzo, Ramamohan Paturi, and Francis Zane. Which problems have
strongly exponential complexity? J. Comput. Syst. Sci., 63(4):512–530, 2001.
[Jam06]
G. J. O. Jameson. Counting zeros of generalised polynomials: Descartes’ rule of signs
and Laguerre’s extensions. The Mathematical Gazette, 90(518):223–234, 2006.
[JS98]
Antoine Joux and Jacques Stern. Lattice reduction: A toolbox for the cryptanalyst.
Journal of Cryptology, 11(3):161–185, 1998.
[Kan87]
Ravi Kannan. Minkowski’s convex body theorem and Integer Programming. Math. Oper.
Res., 12(3):415–440, 1987.
[Kho05]
Subhash Khot. Hardness of approximating the Shortest Vector Problem in lattices.
Journal of the ACM, 52(5):789–808, September 2005. Preliminary version in FOCS’04.
[KPV14]
Subhash Khot, Preyas Popat, and Nisheeth K. Vishnoi. 2log n hardness for Closest
Vector Problem with Preprocessing. SIAM Journal on Computing, 43(3):1184–1205,
2014.
[Laa15]
Thijs Laarhoven. Sieving for shortest vectors in lattices using angular locality-sensitive
hashing. In CRYPTO, 2015.
[Len83]
Hendrik W. Lenstra, Jr. Integer Programming with a fixed number of variables. Math.
Oper. Res., 8(4):538–548, 1983.
[LG14]
François Le Gall. Powers of tensors and fast matrix multiplication. In ISAAC, 2014.
[LLL82]
Arjen K. Lenstra, Hendrik W. Lenstra, Jr., and László Lovász. Factoring polynomials
with rational coefficients. Math. Ann., 261(4):515–534, 1982.
1−ε
[LWXZ11] Mingjie Liu, Xiaoyun Wang, Guangwu Xu, and Xuexin Zheng. Shortest lattice vectors
in the presence of gaps. IACR Cryptology ePrint Archive, 2011:139, 2011.
[Mic01a]
Daniele Micciancio. The hardness of the Closest Vector Problem with Preprocessing.
IEEE Transactions on Information Theory, 47(3):1212–1215, 2001.
[Mic01b]
Daniele Micciancio. The Shortest Vector Problem is NP-hard to approximate to within
some constant. SIAM Journal on Computing, 30(6):2008–2035, March 2001. Preliminary
version in FOCS 1998.
[MR17]
Pasin Manurangsi and Prasad Raghavendra. A birthday repetition theorem and complexity of approximating dense csps. In 44th International Colloquium on Automata,
Languages, and Programming, ICALP 2017, July 10-14, 2017, Warsaw, Poland, pages
78:1–78:15, 2017.
[MV10]
Daniele Micciancio and Panagiotis Voulgaris. Faster exponential time algorithms for the
Shortest Vector Problem. In SODA, pages 1468–1480, 2010.
34
[MV13]
Daniele Micciancio and Panagiotis Voulgaris. A deterministic single exponential time
algorithm for most lattice problems based on Voronoi cell computations. SIAM Journal
on Computing, 42(3):1364–1391, 2013.
[MW15]
Daniele Micciancio and Michael Walter. Fast lattice point enumeration with minimal
overhead. In SODA, 2015.
[MW16]
Daniele Micciancio and Michael Walter. Practical, predictable lattice basis reduction.
In Eurocrypt, 2016.
[NIS16]
NIST post-quantum standardization call for proposals. http://csrc.nist.gov/
groups/ST/post-quantum-crypto/cfp-announce-dec2016.html, 2016. Accessed:
2017-04-02.
[NS01]
Phong Q. Nguyen and Jacques Stern. The two faces of lattices in cryptology. In
Cryptography and Lattices, pages 146–180. Springer, 2001.
[NV08]
Phong Q. Nguyen and Thomas Vidick. Sieve algorithms for the shortest vector problem
are practical. J. Math. Cryptol., 2(2):181–207, 2008.
[Odl90]
Andrew M. Odlyzko. The rise and fall of knapsack cryptosystems. Cryptology and
Computational Number Theory, 42:75–88, 1990.
[Pei08]
Chris Peikert. Limits on the hardness of lattice problems in `p norms. Computational
Complexity, 17(2):300–351, May 2008. Preliminary version in CCC 2007.
[Pei16]
Chris Peikert. A decade of lattice cryptography. Foundations and Trends in Theoretical
Computer Science, 10(4):283–424, 2016.
[PS09]
Xavier Pujol and Damien Stehlé. Solving the Shortest Lattice Vector Problem in time
22.465n . IACR Cryptology ePrint Archive, 2009:605, 2009.
[PW10]
Mihai Pătraşcu and Ryan Williams. On the possibility of faster SAT algorithms. In
SODA, 2010.
[Reg04]
Oded Regev. Improved inapproximability of lattice and coding problems with preprocessing. IEEE Transactions on Information Theory, 50(9):2031–2037, 2004. Preliminary
version in CCC’03.
[Reg09]
Oded Regev. On lattices, learning with errors, random linear codes, and cryptography.
Journal of the ACM, 56(6):Art. 34, 40, 2009.
[RR06]
Oded Regev and Ricky Rosen. Lattice problems and norm embeddings. In STOC, 2006.
[Sch87]
Claus P. Schnorr. A hierarchy of polynomial time lattice basis reduction algorithms.
Theoretical Computer Science, 53:201 – 224, 1987.
[SFS09]
Naftali Sommer, Meir Feder, and Ofir Shalvi. Finding the closest lattice point by
iterative slicing. SIAM J. Discrete Math., 23(2):715–731, 2009.
[Sha84]
Adi Shamir. A polynomial-time algorithm for breaking the basic Merkle-Hellman
cryptosystem. IEEE Trans. Inform. Theory, 30(5):699–704, 1984.
35
[Vai15]
Vinod Vaikuntanathan. Private communication, 2015.
[vEB81]
Peter van Emde Boas. Another NP-complete problem and the complexity of computing
short vectors in a lattice. Technical report, University of Amsterdam, Department of
Mathematics, Netherlands, 1981. Technical Report 8104.
[Wil05]
Ryan Williams. A new algorithm for optimal 2-Constraint Satisfaction and its implications. Theoretical Computer Science, 348(2-3):357–365, 2005.
[Wil12]
Virginia Vassilevska Williams. Multiplying matrices faster than Coppersmith-Winograd.
In STOC, 2012.
[Wil15]
Virginia Vassilevska Williams. Hardness of easy problems: Basing hardness on popular
conjectures such as the Strong Exponential Time Hypothesis (invited talk). In IPEC,
pages 17–29, 2015.
[Wil16]
Ryan Williams. Strong ETH breaks with Merlin and Arthur: Short non-interactive
proofs of batch evaluation. In CCC, 2016.
[WLTB11] Xiaoyun Wang, Mingjie Liu, Chengliang Tian, and Jingguo Bi. Improved Nguyen-Vidick
heuristic sieve algorithm for shortest vector problem. In ASIACCS, 2011.
[Woe08]
Gerhard J. Woeginger. Open problems around exact algorithms. Discrete Applied
Mathematics, 156(3):397–405, 2008.
36
| 8cs.DS
|
HIERARCHICALLY COCOMPACT CLASSIFYING
SPACES FOR MAPPING CLASS GROUPS OF
SURFACES
arXiv:1712.00496v1 [math.GR] 1 Dec 2017
BRITA NUCINKIS AND NANSEN PETROSYAN
Dedicated to Peter Kropholler on the occasion of his 60th birthday.
Abstract. We define the notion of a hierarchically cocompact
classifying space for a family of subgroups of a group. Our main
application is to show that the mapping class group ModpSq of any
connected oriented compact surface S, possibly with punctures and
boundary components and with negative Euler characteristic has
a hierarchically cocompact model for the family of virtually cyclic
subgroups of dimension at most vcd ModpSq` 1. When the surface
is closed, we prove that this bound is optimal. In particular, this
answers a question of Lück for mapping class groups of surfaces.
1. Introduction
Let G be a group and denote by F a family of subgroups of G, that
is a collection of subgroups of G closed under conjugation and finite
intersection. We denote by EF G the classifying space for the family F.
A G-CW-complex X is said to be a model for EF G if X H is contractible
for all H P F, and X H is empty otherwise. The minimal dimension of
a model for EF G, denoted by gdF G, is called the geometric dimension
of G for the family F.
Definition 1.1. We say a G-CW-complex X is a hierarchically cocompact (hierarchically finite type) model for EF G if there are families Fi
Ť
(i P I some indexing set) such that F “ iPI Fi , and that there are
cocompact (finite type) models for EFi G for all i P I.
By the universal property for classifying spaces for families, this is
equivalent to saying that there is a model for EF G which is a mapping
telescope of cocompact (finite type) models for EFi G. This follows from
an argument analogous to [18, Theorem 6.11].
Date: December 5, 2017.
2010 Mathematics Subject Classification. 20J06, 55R35.
The second author was supported by the EPSRC First Grant EP/N033787/1.
1
2
BRITA NUCINKIS AND NANSEN PETROSYAN
Definition 1.1 was motivated by the following conjecture of Juan-Pineda
and Leary.
Conjecture 1.2 (Juan-Pineda–Leary, [12]). Denote by Vc the family
of virtually cyclic subgroups of a group G, and let G be a group admitting
a cocompact model for EVc G. Then G is virtually cyclic.
This conjecture has been proved in many cases, such as, for example,
hyperbolic groups [12], elementary amenable groups [9], one-relator
groups, CAT(0)-groups, acylindrically hyperbolic groups, 3-manifold
groups [22], and linear groups [23]. In all of these examples it was
shown that these groups cannot admit a model for EVc G with a finite type 0-skeleton, a condition equivalent to G having a finite set of
virtually cyclic subgroups tH1 , ..., Hn u such that every virtually cyclic
subgroup is sub-conjugate to Hi for some 1 ď i ď n. T. von Puttkamer
and X. Wu [23] actually conjecture that any finitely presented group
satisfying this condition is already virtually cyclic. They also exhibit a
finitely generated example of type F0 . Note, that non-finitely generated
examples were already known, as for example [21, 6.4.6] the famous construction of Higman-Neumann-Neumann with one conjugacy class of
elements.
In line with convention, we write EG for a classifying space for proper
actions and EG for EVc G. Denote the geometric dimension for proper
action by gd G and gdVc G by gd G. We also say that a group admitting
a hierarchically cocompact model for EG is of type hF. Finally, we use
the notation hF8 for a group admitting a hierarchically finite type
model for EG.
In this paper we provide a method by which to construct hierarchically
cocompact (hierarchically finite-type) models for EG out of cocompact
(finite-type) models for EG, provided that commensurators of virtually
cyclic subgroups satisfy some further finiteness conditions. This enables
us to show that many classes of finitely presented groups that do not
satisfy the hypothesis of Conjecture 1.2, in particular are not of type
F0 , are still of type hF. Based on our observations of groups of type
hF, we ask the following question.
Question 1.3. Suppose a group G is of type hF. Are commensurators
of virtually cyclic subgroups of type F8 ?
CLASSIFYING SPACES FOR MAPPING CLASS GROUPS
3
The main result of the paper is the following.
Theorem 1.4. Let S be a compact connected orientable surface of
genus g, with a finite number of boundary components and punctures,
and with negative Euler characteristic. Then the mapping class group
ModpSq has a hierarchically cocompact model for EModpSq of dimension vcd ModpSq ` 1. Moreover, if S is closed and g ě 1, then
gd ModpSq “ cd ModpSq “ vcd ModpSq ` 1.
Since gd ModpSq “ vcd ModpSq (see [1]), we note that this answers a
question by Wolfgang Lück which asks for which countable groups G
the inequality
gd G ´ 1 ď gd G ď gd G ` 1
holds, see for example [15, Problem 9.51]. By Theorem 1.4, the above
inequality is always satisfied for the mapping class group of any compact connected orientable surface S with a finite number of boundary
components and punctures, and and χpSq ă 0.
Roughly stated, the hierarchically cocompact model for EModpSq can
be obtained by attaching certain fibred spaces to the Teichmüller space
of S. These are associated to the Weyl groups of the infinite cyclic subgroups H “ xf y ď ModpSq representing the set I of the complete
sub-conjugacy classes of infinite virtually cyclic subgroups of ModpSq.
These spaces fibre over products of a Euclidean space and a Teichmüller
space corresponding to the pseudo-Anosov and the trivial components
in the canonical reduction system of f P ModpSq respectively. Since the
attaching maps are equivariant, Teichmüller spaces could be replaced
by spines or the relevant minimal models for the local subgroups to obtain the desired hierarchically cocompact model of minimal dimension.
It is worth pointing out, that if one applies the above attaching construction to only a finite subset J Ă I, then the resulting space will be
a cocompact model for the classifying space of ModpSq for the family
of all virtually cyclic subgroups that are sub-conjugate to a subgroup
in J.
Finite dimensional models for EModpSq have been exhibited by Degrijse and the second author in [5] for closed surfaces S (the obtained
bound on dimension is 9g ´ 8) and later by Juan-Pineda and TrufiloNegrete in [13] for surfaces S that have negative Euler characteristic with possible punctures and boundary components (the obtained
bound is rModpSq : ModpSqrmsspvcd ModpSq`1q, m ě 3 where ModpSqrms
is the level m congruence subgroup). Apart from exhibiting models that
4
BRITA NUCINKIS AND NANSEN PETROSYAN
are mapping telescopes of cocompact models, our bounds substantially
improve on the bounds given there. In particular, for closed surfaces,
our bounds are optimal.
Furthermore, recently Bartels and Bestvina showed that mapping class
groups of oriented surfaces G of finite type satisfy the Farrell-Jones
conjecture [2]. For example, the assembly map, which is induced by
the G-map EG Ñ tptu,
–
Ñ Kn pZGq,
HnG pEG; KZ q Ý
is an isomorphism, where HnG p´; KZq is the Bredon homology theory
with the K-theory functor coefficients. In particular, since homology
commutes with colimits and when χpSq ă 0, G “ ModpSq is of type
hF, the left-hand side of the isomorphism is a colimit of, hopefully,
more computable homological terms. For a detailed introduction into
the Farrell-Jones conjecture the reader is referred to [15].
In the next section, we define what exactly is meant by a hierarchically
cocompact model for a classifying space, and give a recipe for constructing these models. We also exhibit some examples in Section 3. Since
we will need some results on Bredon cohomological dimensions later
when giving lower bounds for the dimensions of our models, we give a
brief introduction into Bredon cohomology in Section 4. Section 5 is
devoted to proving our main result, Theorem 1.4.
2. Constructing hierarchically cocompact models
The following construction, due to Lück and Weiermann [17], will play
a crucial role in what is to come:
Let F and G be families of subgroups of a given group G such that
F Ď G.
Definition 2.1. [17, (2.1)] A strong equivalence relation on G r F,
denoted „, is an equivalence relation on G r F satisfying:
‚ For H, K P G r F with H ď K we have H „ K.
‚ Let H, K P G r F and g P G, then H „ K ðñ gHg ´1 „
gKg ´1.
Denote by rG r Fs the equivalence classes of „ and define for all rHs P
rG r Fs the following subgroup of G:
NG rHs “ tg P G | rgHg ´1s “ rHsu.
CLASSIFYING SPACES FOR MAPPING CLASS GROUPS
5
Now define a family of subgroups of NG rHs by
FrHs “ tK ď NG rHs | K P G r F , rKs “ rHsu Y pF X NG rHsq,
where F X NG rHs is the family of subgroups of NG rHs that are in F.
Remark 2.2. A typical example of a pair of families that has a strong
equivalence relation is the finite and the virtually cyclic subgroups of
G. Here, the equivalence relation is the commensurability.
We need the following theorem of Lück and Weiermann.
Proposition 2.3. [17, Theorem 2.3] Let F Ď G be families with a
strong equivalence relation on G r F. Denote by I a complete set of
representatives of the conjugacy classes in rG r Fs. Then the G-CWcomplex given by the cellular G push-out
Ů
Ů
rHsPI
rHsPI
Ů
G ˆNG rHs EFXNG rHs NG rHs
i
/
EF G
idG ˆNG rH frHs
rHsPI G ˆNG rHs EFrHs NG rHs
/
X
where either i or the frHs are inclusions, is a model for EG G.
The condition on the two maps being inclusions is not a big restriction
as one can replace the spaces by the mapping cylinders, see [17, Remark
2.5].
We also need the following transitivity principle.
Proposition 2.4. [17, Proposition 5.1] Let F Ď G be families of subgroups of a group G. Assume that G admits a cocompact (finite type)
model XG for EG G and that every subgroup H P G admits a cocompact
(finite type) model XFXH pHq for EFXH H. Then G admits a cocompact
(finite type) model for EF G of dimension
dim XG ` suptdim XFXH pHq | H P Gu.
Theorem 2.5. Let F Ď G be families with a strong equivalence relation
on GrF. Suppose that for every H P GrF, NG rHs admits a cocompact
(finite type) model XFXH pNG rHsq for EFXH NG rHs. Further assume that
G admits a cocompact (finite type) model XF for EF G and that for
each H P G, there is a cocompact (finite type) model XFrHs pNG rHsq for
EFrHs NG rHs, then G admits a hierarchically cocompact (hierarchically
finite type) model for EG G of dimension
suptdim XF , dim XFXH pNG rHsq ` 1, dim XFrHs pNG rHsq | H P G r Fu.
6
BRITA NUCINKIS AND NANSEN PETROSYAN
Proof. Let S be a set of representatives of some finite number of conjugacy classes in rG r Fs and denote
p
FrSs
:“ tK ď G | K P G r F , Dg P G, H P S, rK g s “ rHsu Y F.
p
Consider the G-push-out of Proposition 2.3 replacing G by FrSs:
Ů
Ů
rHsPS
rHsPS
G ˆNG rHs XFXH pNG rHsq
i
XF
/
idG ˆNG rHs ιrHs
Ů
rHsPS G ˆNG rHs cylpfrHs q
/
X
where ιrHs is the natural inclusion, and frHs : XFXH pNG rHsq Ñ XFrHs pNG rHsq
is given by the universal property for classifying spaces for a family. X
is now a cocompact (finite type) model for EFrSs
G, as it is a G-push-out
p
of cocompact (finite type) complexes. The result now follows from the
Ť
p
fact that G “ SĎI,|S|ă8 FrSs
and the remark after Definition 1.1.
p
We denote FrHs
:“ tK ď G | K P G r F , Dg P G, rK g s “ rHsu Y F and
p
p
note that FrHs
“ FrSs
exactly when S “ tHu. We end this section
with a useful lemma.
Lemma 2.6. Suppose F Ď G are families of subgroups of a group
G, and that for each H P G r F, G admits cocompact (finite type)
models for both EFrHs
G and EFXH H. Then G admits a hierarchically
p
cocompact (finite type) model for EG G.
Proof. For any two families F1 and F2 of subgroups of G, we have unique
maps (up to G-homotopy) EF1 XF2 G Ñ EF1 G and EF1 XF2 G Ñ EF2 G.
The double mapping cylinder X gives a model for EF1 YF2 G, and that
if EF1 XF2 G, EF1 G and EF2 G are cocompact (finite type), then so is X.
p 1 s, and F2 “ FrH
p 2 s. Then,
Now, let rH1s ‰ rH2 s P I and F1 “ FrH
p 1 sqrH2s “: FrH
p 1 , H2 s. By
obviously F1 X F2 “ F and F1 Y F2 “ pFrH
assumption there are cocompact (finite type) models for EF1 G and
EF2 G, and the transitivity result in Proposition 2.4 yields that there is
also a cocompact (finite type) model for EF1 XF2 G. The double mapping
cylinder X gives a cocompact (finite type) model for EF1 YF2 G.
p 1 , ..., Hi s :“ FrH
p 1 , ..., Hi´1 srHi s. Proceed by
For each i ą 2, let FrH
induction, to construct a cocompact (finite type) model for EFrH
p 1 ,...,His G
from cocompact (finite type) models for EFrH
p 1 ,...,Hi´1s G and EFrH
p i s G as
above.
CLASSIFYING SPACES FOR MAPPING CLASS GROUPS
7
3. Some examples
In the next examples, let F and Vc be the families of finite and virtually
cyclic subgroups of a given group G, respectively, equipped with the
commensurability relation on Vc r F (see Remark 2.2).
3.1. Thompson groups. In [18] the authors introduced a similar condition for the family of finite subgroups, quasi-F8 , which asks for a
group to have, for any k P Zą0 , finitely many conjugacy classes of finite subgroups of order k, and that normalisers of all finite subgroups
are of type F8 . [19, Theorem 4.9] shows that generalised Thompson
groups, which are automorphism groups of valid, bounded and complete Cantor-algebras are quasi-F8 and hence are hF8 .
3.2. Hyperbolic groups. Following [17], a group is said to satisfy
pNMFfin ĎVc q, if every virtually cyclic subgroup is contained in a unique
maximal virtually cyclic group V , such that NG pV q “ V. Then, by
[17, Corollary 2.11], the push-out in Proposition 2.3 reduces to
G ˆV EV
EG
/
G{V
/
X,
G, provided there is a cocomwhere X is a cocompact model for EFy
fin rV s
pact model for EG. Now, apply Lemma 2.6, to conclude that any group
satisfying pNMFfin ĎVc q that has a cocompact model for EG is of type
hF. In particular, hyperbolic groups satisfy pNMFfin ĎVc q [12, Remark
7], and hence are of type hF.
3.3. Polycyclic groups. Polycyclic groups are of type hF. This follows from [17, Lemma 5.15] - for every infinite virtually cyclic group V ,
there exists a cyclic subgroup C commensurate to V such that NG rCs “
NG pCq, and hence a model for EFrV s NG rV s is given by EpNG pCq{Cq
by pulling back the action. Now, since G is polycyclic, we have a cocompact EG and a cocompact EpNG pCq{Cq for every C. Now apply
Theorem 2.3.
3.4. Soluble Baumslag-Solitar groups. One can also show that the
soluble Baumslag-Solitar groups G “ BSp1, nq are of type hF. The
group G is torsion-free and admits a cocompact 2´dimensional model
for EG “ EG. Next, we show that for all cyclic subgroups H ď G, we
8
BRITA NUCINKIS AND NANSEN PETROSYAN
have a finite model XFrHs pNG rHsq for EFrHs NG rHs and apply Theorem
2.5. Recall,
BSp1, nq – Zr n1 s ¸ Z.
Any infinite cyclic subgroup H not contained in Zr n1 s, is equal to its
commensurator, see [8, Lemma 5], and hence NG rHs has a point as a
model for EFrHs NG rHs, and R is a cocompact model for EG. Infinite
cyclic subgroups of Zr n1 s, on the other hand, are all commensurate to
K “ x n1 y ď Zr n1 s. K has the entire G as its commensurator. Since
G is an HNN-extension with the vertex group K, the associated BassSerre tree is a cocompact model for EFrKs G. Applying Theorem 2.5, we
obtain that BSp1, nq is of type hF for all n ą 0.
Note that when n ą 1, the normaliser NG px n1 yq “ Zr n1 s is not even
finitely generated. However, in all examples in this note, the commensurators of virtually cyclic subgroups are of type F8 . It is not clear
whether this is just an artefact of our construction, and hence we ask,
see Question 1.3:
Question. Suppose a group G is of type hF. Are commensurators of
virtually cyclic subgroups of type F8 ?
4. Bredon cohomology
In this section we introduce all necessary facts and results regarding
Bredon cohomology needed later on to determine lower bounds for
dimensions of our classifying spaces. All results in this section are well
known; a good introduction to the subject can be found in [7].
Let F denote a family of subgroups of a given group G as before. We
consider the category OF G, which has as objects the transitive G-sets
with stabilisers in F. Morphisms in OF G are G-maps between those
G-sets. Modules over the orbit category, called OF G-modules are contravariant functors from the orbit category to the category of abelian
groups. Exactness is defined pointwise: a sequence
AÑBÑC
of OF G-modules is exact at B if and only if the sequence of abelian
groups
ApG{Kq Ñ BpG{Kq Ñ CpG{Kq
is exact at BpG{Kq for every G{K P OF G . The trivial OF G-module,
which is denoted Zp´q, is the constant functor Z from OF G to the
category of abelian groups.
CLASSIFYING SPACES FOR MAPPING CLASS GROUPS
9
The category OF G-Mod of OF G-modules has enough projectives. Hence
we can consider a projective resolution
P˚ p´q Ñ Zp´q Ñ 0
of the trivial module Zp´q. The Bredon cohomology functors H˚F pG, ´q
are defined as derived functors of HomF pZp´q, ´q. In particular, for
each N “ Np´q P OF G-Mod,
H˚F pG, Nq “ H˚ pmorpP˚ , Nqq.
We can now define cohomological dimensions in the Bredon-setting
with analogous properties to ordinary cohomology. Let n ě 0. We say
G has Bredon-cohomological dimension cdF G ď n if there is a projective resolution of Zp´q of length n. This is equivalent to HFn`1 pG, Nq “ 0
for all OF G-modules N.
Let X be a model for EF G. Then the cellular chain complex C˚ pXq
gives rise to a free resolution
C˚ pXqp´q Ñ Zp´q Ñ 0
by putting C˚ pXqpG{Kq “ C˚ pX K q for all K P OF G . In particular,
cdF G ď gdF G.
Furthermore, [14, Theorem 13.19], if cdF G ě 3, then gdF G “ cdF G.
As before, for F the family of finite groups, we write cd G and gd G,
and for F the family of virtually cyclic subgroups we write cd G and
gd G respectively. Note that for torsion-free groups, cd G “ cd G and
gd G “ gd G and Bredon-cohomology for the family of finite subgroups
becomes ordinary cohomology over the group ring ZG.
We will quite often make use of this simple observation. When G is virtually torsion-free, then cd G ě vcd G. So, if G admits a model for EG
of dimension vcd G, then it must necessary be of minimal dimension,
i.e. realising gd G.
5. Mapping class groups
In this section, we prove our main theorem and show that the mapping
class group of any compact orientable surface S with possibly finitely
many punctures and boundary components, and with negative Euler
characteristic χpSq is of type hF with a hierarchically cocompact model
of dimension vcd ModpSq ` 1. We recall some necessary background
on mapping class groups of surfaces and refer the reader to [6] and [11]
for further details.
10
BRITA NUCINKIS AND NANSEN PETROSYAN
Let S be a connected compact oriented surface with finitely many punctures and χpSq ă 0. The mapping class group of S, denoted by ModpSq,
is the group of isotopy classes of orientation preserving diffeomorphisms
of S pointwise fixing the boundary BS
ModpSq “ Diff` pS, BSq{Diff0 pS, BSq,
where Diff0 pS, BSq is the subgroup of Diff` pS, BSq consisting of elements
that are isotopic to the identity.
Any diffeomorphism of S induces an automorphism of H1 pS, Z{mZq for
m ě 2. This gives a well-defined homomorphism
ModpSq Ñ AutpH1 pS, Z{mZqq
where the kernel is denoted by ModpSqrms and it is called the level m
congruence subgroup of ModpSq.
Suppose now that S has no boundary. Let tα1 , . . . , αn u be a collection of pairwise disjoint, homotopically distinct essential simple closed
curves in S. Denote by σ “ trα1 s, . . . , rαn su the corresponding isotopy
classes.
Inclusion homomorphism. ([6, Thm. 3.18]). Let Nσ be an open regular neighbourhood of Ynj“1αj in S. Denote Sσ “ S r Nσ and set Sσ “
Yki“1 Si where each Si is a connected subsurface. Let tβ1 , γ1 u, . . . , tβk , γk u
denote the pairs of boundary components of Sσ that bound the annuli
in Nσ . The inclusion Sσ ãÑ S induces a homomorphism
ησ : ModpSσ q “
k
ź
ModpSi q Ñ ModpSq
i“1
with kernel xTβ1 Tγ´1
, . . . , Tβk Tγ´1
y. The restriction ηi :“ ησ |ModpSi q is
1
k
the map induced by the inclusion Si ãÑ S.
Capping homomorphism. ([6, Prop. 3.19]) Let Spσ be the surface obtained from Sσ by capping the boundary components with once-punctured
disks and write Spσ “ Yki“1 Spi . The inclusion Sσ ãÑ Spσ induces a homomorphism
θσ : ModpSσ q ։
k
ź
i“1
ModpSpi , Ωi q Ď ModpSpσ q
with kernel xTβ1 , Tγ1 , . . . , Tβk , Tγk y. Here Ωi denotes the set of punctures
coming from the boundary components of Si . The image of the restriction θi :“ θσ |ModpSi q , denoted ModpSpi , Ωi q, is the subgroup of ModpSpi q
consisting of all the elements that fix Ωi pointwise. ModpSpi , Ωi q contains the pure mapping class group and hence is finite index in ModpSpi q.
CLASSIFYING SPACES FOR MAPPING CLASS GROUPS
11
Cutting homomorphism. ([6, Prop. 3.20]) Define ModpSqσ “ tg P ModpSq | gpσq “
σu. There is a well-defined homomorphism
ρσ : ModpSqσ Ñ ModpS r Ynj“1αj q “ ModpSpσ q
with free abelian kernel xTα1 , . . . , Tαn y generated by the Dehn twists
about the curves α1 , . . . , αn . Let ModpSq0σ be the finite index subgroup
of ModpSqσ consisting of all the elements that fix each curve αi with
orientation. Denote the restriction ρσ,0 “ ρσ |ModpSq0σ . Then θσ “ ρσ,0 ˝ησ
(see the diagram on page 91 of [6]).
The canonical form. ([6, Cor. 13.3]) Let f P ModpSq and let σ “ σpf q be
its canonical reduction system (see [6, §13.2.2]). Let Sk`1 , . . . , Sk`n be
the pairwise disjoint annuli that are the closed neighbourhoods of the
curves α1 , . . . , αn representing σ. Then there is a representative φ of f
that permutes the Sl , so that some power of φ leaves invariant each Sl ,
1 ď l ď n ` k. By applying Nielsen-Thurston Classification Theorem
to each Sl , one obtains that there exists p ą 0 so that φp pSl q “ Sl for
all 1 ď l ď n ` k and
(5.1)
p
f “
k
ź
i“1
ηi pfi q
n
ź
Tαnjj
j“1
where each fi P ModpSi q is either pseudo-Anosov or the identity and
nj P N for 1 ď i ď k, 1 ď j ď n.
Remark 5.1. By a result of Ivanov [11, Corollary 1.8], if f P ModpSqrms,
m ě 3, then the integer p in (5.1) can always be taken to be one.
Remark 5.2 ([13, 3.10]). By Theorem 1.2 in [11], ModpSqrmsσ Ď ModpSq0σ .
Since θσ “ ρσ,0 ˝ ησ , it follows that θi pfi q “ fˆi for 1 ď i ď k.
For simplicity, we will often denote G “ ModpSq. Let F and Vc the families of finite and virtually cyclic subgroups of G, respectively, equipped
with the commensurabilty relation on Vc r F (see Remark 2.2). The
Teichmüller space T pSq is a CAT(0)-space on which G acts properly
and cocompactly. So, it gives a cocompact model for EG. [1, Corollary
1.3] of Aramayona and Martı́nez-Pérez states that there is a cocompact
model for EG of minimal dimension gd G “ vcd G. We need the following minor generalisation of this result.
Proposition 5.3. Let S be a compact (possibly disconnected) surface
with possible punctures and boundary components. Then there is a
cocompact model for EModpSq of dimension gd ModpSq “ vcd ModpSq.
12
BRITA NUCINKIS AND NANSEN PETROSYAN
Proof. In [1], this has been proven for connected S. Suppose S decomposes into the disjoint union of diffeomorphic copies of its connected
components
S “ \m1 S1 ¨ ¨ ¨ \mq Sq .
ś
Subsequently, this implies that ModpSq Ď qi“1 ModpSi q ≀ Σmi and is of
finite index. Combining [10, Theorem 4.1] and the extension theorem
for duality groups in [4, Thm. 3.5], gives us that
vcd ModpSq “
q
ÿ
mi ¨ vcd ModpSi q.
i“1
The next lemma shows that there exists a cocompact model for EModpSq
of dimension vcd ModpSq. Since vcd ModpSq ď gd ModpSq, the result
follows.
Lemma 5.4. Let K be a group. Suppose K has a cocompact model X
for EK. Then the wreath product W “ K ≀ Σm has a cocompact model
for EW of dimension m ¨ dim X.
Proof. The wreath product W acts on Y “ X ˆ ¨ ¨ ¨ ˆ X which consists
of m-copies of X with a diagonal action of K m and a permutation
action of Σm . Note that W acts cocompactly on Y and it is a model
for EW of dimension m ¨ dim X.
To establish Theorem 1.4, it remains to show, see Theorem 2.5, that for
each infinite cyclic subgroup H Ď G, there are cocompact models for
EFXNG rHs NG rHs and EFrHs NG rHs where NG rHs is the commensurator
of H in G of the required dimensions.
Proposition 5.5 ([13, Prop. 4.8]). Let S be an orientable closed surface
with finitely many punctures and χpSq ă 0. Suppose f P G generates
an infinite cyclic subgroup H. Then for any integer l ą 0 such that
f l P ModpSqrms, m ě 3, NG rHs “ NG pf l q holds.
We need the following slight generalisation of [13, Proposition 4.12] of
Juan-Pineda and Trufilo-Negrete.
Proposition 5.6. Suppose S is an orientable closed surface with finitely
many punctures and χpSq ă 0. Let f P ModpSqrms, m ě 3, with the
canonical reduction system σ with fˆa`1 , . . . , fˆk pseudo-Anosov and
´
k
¯ ź
ˆ
ˆ
ρσ pf q :“ id1 , . . . , ida , fa`1 , . . . , fk P
ModpSpi , Ωi q,
i“1
CLASSIFYING SPACES FOR MAPPING CLASS GROUPS
13
then there is a central extension
(5.2)
ρσ
1 Ñ Zn Ñ CModpSq pf q0 ÝÑ
a
ź
i“1
ModpSpi , Ωi q
k
ź
Vj Ñ 1.
j“a`1
where Vj “ CModpSpj ,Ωj q pfˆj q is virtually cyclic for each a ` 1 ď j ď k.
Also,
ρσ
1 Ñ Zn Ñ NModpSq pf qÝÑQ Ñ 1,
(5.3)
ˆ
ˆ
such that Q Ď Modp\ai“1 Spi qˆA, where A Ď NModp\k
p ppfa`1 , . . . , fk qq
j“a`1 Sk q
śk
is a finite extension of j“a`1 Vj .
Proof. The extension (5.2) has already been established in [13, Proposition 4.12]. The kernel is generated by the Dehn twists about the curves
α1 , . . . , αn which are fixed by CModpSq pf q0 , and hence is central.
By Lemma 3.8 of [13], for every g P NModpSq pf q, we have gpσq “ σ.
Therefore, NModpSq pf q Ď ModpSqσ . Let Q “ ρσ pNModpSq pf qq. Note
ś
ś
that Q is a finite extension of ai“1 ModpSpi , Ωi q kj“a`1 Vj and it is
contained in NModpSpσ q pρσ pf qq. So, to obtain (5.3), it suffices to show
that any g P ModpSpσ q that normalises ρσ pf q is contained in the subgroup Modp\a Spi q ˆ Modp\k
Spk q of ModpSpσ q.
i“1
j“a`1
Suppose this is not the case, and say g maps Sp1 diffeomorphically onto
Spk . Then for any x P Sp1 , we have ρσ pf qgpxq “ gρσ pf q˘1 pxq “ gpxq.
This shows that fpk is the identity on Spk which is a contradiction.
Lemma 5.7. Let 1 Ñ Zn Ñ G Ñ Q Ñ 1 be an extension of groups
where Q is finitely generated with vcd Q “ k ă 8. Then vcd G “ n ` k.
Proof. We can assume Q is torsion-free. Now we apply a result of
Fel’dman, see [3, Theorem 5.5]
We will quite often make use of the following corollary.
Corollary 5.8. Let 1 Ñ Zn Ñ G Ñ Q Ñ 1 be an extension of groups
where Q has a cocompact model for EQ of dimension vcd Q “ k ă 8.
Then G has a cocompact model for EG of dimension gd G “ vcd G “
n ` k.
Proof. Let F be the family of finite subgroup of G and G be the family
of the preimages of all the finite subgroups of Q under the projection
of G onto Q. Applying Proposition 2.4 together with the general fact
that finitely generated virtually free abelian groups of rank m have
the Euclidean space of dimension m as a cocompact model for proper
14
BRITA NUCINKIS AND NANSEN PETROSYAN
actions, we obtain that G has a cocompact model for EG of dimension
n ` k. Lemma 5.7 finishes the claim.
The next proposition concerns surfaces with boundary. Suppose S has
b ‰ 0 boundary components β1 , . . . , βb . Note that the capping homop Ωq Ď ModpSq
p where Ω is the set of
morphism θS : ModpSq Ñ ModpS,
punctures of Sp that come from capping the boundary components.
Proposition 5.9. Let S be a compact orientable surface with nonempty
boundary, finitely many punctures and χpSq ă 0. Suppose f P ModpSq
generates an infinite cyclic subgroup H. Then there exists l P N such
that f l P ModpSqrms, m ě 3, with NModpSq rHs “ NModpSq pf l q and
(5.4)
θ
S
1 Ñ Zb Ñ NModpSq pf l q ÝÑ
NModpS,Ωq
pθS pf l qq Ñ 1,
p
where Zb “ xTβ1 , . . . , Tβb y and θS pf l q is either trivial or of infinite order.
Proof. By replacing f with a sufficiently large power, we can assume
that l “ 1 and f P ModpSqrms, m ě 3 with θS pf q either trivial or of
infinite order. Restricting the capping homomorphism to NModpSq rHs
we have
θ
S
1 Ñ Zb Ñ NModpSq rHs ÝÑ
Q Ñ 1,
p Ωq. Note that if θS pf q is trivial, then NModpSq rHs “
where Q Ď ModpS,
p Ωq as desired. Otherwise,
NModpSq pHq “ ModpSq and Q “ ModpS,
xθS pf qy is infinite cyclic and Q Ď NModpS,Ωq
rxθS pf qys “ NModpS,Ωq
pθS pf qq
p
p
by Proposition 5.5 applied to Sp (replacing f with a sufficiently large
power if necessary). Since NModpS,Ωq
pθS pf qq Ď Q (see for example
p
[13, eq. (25)]), we deduce Q “ NModpS,Ωq
pθS pf qq. Hence, NModpSq rHs “
p
NModpSq pf q.
Proposition 5.10. Let G “ ModpSq where S is a compact orientable
surface S with χpSq ă 0 possibly with finitely many punctures and
boundary components. Denote by F the family of finite subgroups of G.
Let H be an infinite cyclic subgroup of G. Then the commensurator
NG rHs has cocompact models for EFXNG rHs NG rHs and EFrHs NG rHs of
dimensions vcd NG rHs and vcd NG rHs ´ 1, respectively.
Proof. By Propositions 5.5 and 5.9, we can assume NG rHs “ NG pHq
where H “ xf y and f P ModpSqrms, m ě 3.
Suppose first that the boundary of S is empty. By the Nielsen-Thurston
Classification Theorem, f is either pseudo-Anosov or reducible.
CLASSIFYING SPACES FOR MAPPING CLASS GROUPS
15
If f is pseudo-Anosov, then by [20, Theorem 1], NG pf q is virtually
cyclic. So, a Euclidean line and a point are (cocompact) models for
EFXNG pf q NG pf q and EFrHs NG pf q, respectively.
If f is reducible, say with the canonical reduction system σ, then NG pf q
satisfies (5.3) of Proposition 5.6:
ρσ
1 Ñ Zn Ñ NG pf qÝÑQ Ñ 1,
such that Q is a finite index subgroup of
P :“ Modp\a Spi q ˆ A
i“1
ˆ
ˆ
where A Ď NModp\k
p ppfa`1 , . . . , fk qq is virtually free abelian. Let
j“a`1 Sk q
L :“ Modp\ai“1 Spi q. To show that NG pf q has a cocompact model for
EFXNG pf q NG pf q of dimension vcd NG pf q, by Corollary 5.8, it suffices to
show that Q has a cocompact model for EQ of dimension vcd Q.
Since Q is a finite index subgroup in P , it is enough to show that P
has cocompact models for EP of dimension vcd P “ vcd Q. Since A
is virtually free abelian, applying Lemma 5.7, we obtain that vcd P “
vcd L ` vcd A. By Proposition 5.3, there is a cocompact model XL
for EL of dimension vcd L. Since A is finitely generated virtually free
abelian, there is a cocompact model XA for EA of dimension vcd A.
Then XL ˆ XA is a cocompact model for the classifying space EP of
dimension vcd P as desired.
To establish the second claim, observe that under the natural projection
of NG pHq onto WG pHq “ NG pHq{H, a classifying space EWG pHq becomes a model for a classifying space of NG pHq for EFrHs NG pHq. Hence
its suffices to show that there is a cocompact model for EWG pHq of
dimension at most vcd NG pHq ´ 1.
By Proposition 5.5, we only need to consider two case: ρσ pf q is the
identity or it has infinite order. First, suppose ρσ pf q is the identity, that
is H ď Zn . By Corollary 4.9 of [13], we can assume that Zn {H – Zn´1 .
Then
(5.5)
ρ
σ
1 Ñ Zn´1 Ñ WG pf qÝÑQ
Ñ 1,
Again, using Corollary 5.8, there is a cocompact model of dimension
n ´ 1 ` vcd Q “ vcd NG pf q ´ 1.
Now, suppose ρσ pf q has infinite order. Then
ρ
σ
Ñ 1,
1 Ñ Zn Ñ WG pf qÝÑQ{Z
where ρσ pf q generates Z ď Q. By Corollary 5.8, it suffices to show
then that Q{Z has a cocompact model for EpQ{Zq of dimension at
16
BRITA NUCINKIS AND NANSEN PETROSYAN
most vcd Q. Note that, from (5.3), it follows that Z is normal in P .
Therefore, it remains to show that P {Z has a cocompact model for
EpP {Zq of dimension vcd P . But P {Z – L ˆ pA{Zq and a similar
argument as above gives a cocompact model XL ˆ XA{Z for EpP {Zq of
dimension vcd P ´ 1.
Suppose S has nonempty boundary. The proof easily reduces to the case
of empty boundary. By (5.4) of Proposition 5.9, we have the central
extension
(5.6)
θ
S
1 Ñ Zb Ñ NModpSq pHq ÝÑ
NModpS,Ωq
pθS pf qq Ñ 1.
p
When θS pf q has infinite order, then the quotient of (5.6) by H, gives
(5.7)
1 Ñ Zb Ñ WModpSq pHq Ñ WModpS,Ωq
pθS pf qq Ñ 1.
p
In case θS pf q is trivial, note that NModpSq pHq “ ModpSq. Replacing
H “ xf y with a commensurable subgroup if necessary, we can assume
that f P Zb such that Zb {H – Zb´1 . Thus, we obtain
(5.8)
(5.9)
θ
S
p Ωq Ñ 1,
1 Ñ Zb Ñ NModpSq pHq ÝÑ
ModpS,
p Ωq Ñ 1.
1 Ñ Zb´1 Ñ WModpSq pHq Ñ ModpS,
From the empty boundary case of the proposition applied to Sp and
Corollary 5.8 applied to (5.6), (5.7), (5.8), and (5.9), we obtain the
desired result.
We need the following complete computation of the virtual cohomological dimension of ModpSq for surfaces with negative Euler characteristic
by Harer.
r
Theorem 5.11 ([10, Thm. 4.1]). Let S “ Sg,b
be an oriented surface
of genus g, b boundary components and r punctures and recall χpSq “
2 ´ 2g ´ b ´ r. If χpSq ă 0, then
$
& 4g ` 2b ` r ´ 4 g ą 0, r ` b ą 0
vcd ModpSq “
4g ´ 5
r`b“0
%
2b ` r ´ 3
g “ 0.
We are now ready to prove our main theorem.
Proof of Theorem 1.4. Set G “ ModpSq and let F and Vc be the families of finite and virtually cyclic subgroups of G, respectively, equipped
with the commensurabilty relation on Vc r F. By [1, Cor. 1.3], there
is a cocompact model for EG of minimal dimension gd G “ vcd G.
By Proposition 5.10, for each infinite cyclic subgroup H ď G, there
CLASSIFYING SPACES FOR MAPPING CLASS GROUPS
17
are cocompact models for both EFXNG rHs NG rHs and EFrHs NG rHs of
dimension vcd G.
Applying Theorem 2.5, we obtain a hierarchically cocompact model for
EG of dimension vcd G ` 1.
To prove the second part of the theorem, assume we have a closed
surface S. If g “ 1, then ModpSq – SLp2, Zq. So, by [5, Lemma 5.2],
cdpModpSqq ě 2. Combining this with [12, Proposition 9] shows that
gd ModpSq “ cd ModpSq “ 2.
Now, suppose that S is closed and g ě 2. Then vcd ModpSq “ 4g ´ 5.
Let σ “ trαs, rβsu where α and β are essential curves that separate S
into a pair of pants and a surface of genus g ´ 1 (see Figure 1). We will
show that cd ModpSq0σ “ 4g´4. This will imply that cd ModpSq ě 4g´
4 and by the first part of the theorem, we will obtain that gd ModpSq “
cd ModpSq “ 4g ´ 4.
Using the cutting homomorphism, there is a short exact sequence
ρσ
3
1
q ˆ ModpSg´1,0
q Ñ 1,
1 Ñ Z2 Ñ ModpSq0σ ÝÑModpS0,0
where Kerpρσ q “ xTα , Tβ y. By Lemma 5.7,
0
1
vcd ModpSσ q0 “ 2 ` vcd ModpS3,0
q ` vcd ModpSg´1,0
q
“ 2 ` 0 ` 4pg ´ 1q ` 1 ´ 4
“ 4g ´ 5.
By Proposition 5.3, it follows that there is a cocompact model of for
EModpSσ q0 of dimension vcd ModpSσ q0 “ 4g ´ 5 which is therefore the
same as gd ModpSσ q0 .
Set C “ ModpSσ q0 XModpSqr3s, where ModpSqr3s is the level 3 congruence subgroup of ModpSq which is torsion-free [6, Theorem 6.9]. Hence
C is a finite index torsion-free subgroup of ModpSσ q0 . Let F and Vc
be the families of finite (in this case trivial) and virtually cyclic subgroups of C, respectively, equipped with the commensurabilty relation
on Vc r F. Define
M : OVc C Ñ Z - mod : C{H Ñ pZCqH .
The long exact cohomology sequence associated to the push-out of
Proposition 2.3 gives us
ź p4g´5q
ź
i˚
Ñ
H p4g´5q pNC rHs, ZCq
HFrHs pNC rHs, Mq‘H p4g´5q pC, ZCq Ý
rHsPI
rHsPI
p4g´4q
Ñ HVc
pC, Mq.
18
BRITA NUCINKIS AND NANSEN PETROSYAN
By Proposition 5.10, this reduces to
ź
i˚
p4g´4q
H p4g´5q pC, ZCq Ý
Ñ
H p4g´5q pNC rHs, ZCq Ñ HVc pC, Mq.
rHsPI
Since C is of type F , note that H p4g´5q pC, ZCq ‰ 0. Consider the
infinite cyclic subgroups H1 “ xTα Tβ y and H2 “ xTα2 Tβ y. Since every
element of C fixes the curves α and β, it must commute with both
H1 and H2 . Thus, NC rH1 s “ NC rH2 s “ C and H1 , H2 represent
distinct classes in I. The composition of i˚ with the projection of
ś
p4g´5q
pNC rHs, ZCq onto the two factors corresponding to these
rHsPI H
subgroups is the diagonal map
∆ : H p4g´5q pC, ZCq Ñ H p4g´5q pC, ZCq ‘ H p4g´5q pC, ZCq
which is not surjective. Therefore, i˚ cannot be surjective and we obtain
p4g´4q
that HVc pC, Mq ‰ 0.
β
α
Figure 1. The closed surface S of genus g ě 2 and the
essential curves α, β separating it into a pair of pants
and a surface of genus g ´ 1.
References
[1] J. Aramayona and C. Martı́nez-Pérez, The proper geometric dimension of the
mapping class group, Algebr. Geom. Topol. 14 (2014), no. 1, 217–227, DOI
10.2140/agt.2014.14.217.
[2] B. Bartels and B. Bestvina, The Farrell-Jones Conjecture for mapping class
groups (2016). Preprint arXiv:1606.02844.
[3] Robert Bieri, Homological dimension of discrete groups, Second, Queen Mary
College Mathematical Notes, Queen Mary College, Department of Pure Mathematics, London, 1981.
[4] R. Bieri and B. Eckmann, Groups with homological duality generalizing
Poincaré duality, Invent. Math. 20 (1973), 103–124, DOI 10.1007/BF01404060.
[5] D. Degrijse and N. Petrosyan, Bredon cohomological dimensions for groups
acting on CATp0q-spaces, Groups Geom. Dyn. 9 (2015), no. 4, 1231–1265.
CLASSIFYING SPACES FOR MAPPING CLASS GROUPS
19
[6] B. Farb and D. Margalit, A primer on mapping class groups, Princeton Mathematical Series, vol. 49, Princeton University Press, Princeton, NJ, 2012.
[7] M. Fluch, On Bredon (Co-)Homological Dimensions of Groups (2010). Ph.D
thesis,arXiv:1009.4633.
[8] Martin Fluch, Classifying spaces with virtually cyclic stabilisers for certain
infinite cyclic extensions, J. Pure Appl. Algebra 215 (2011), no. 10, 2423–2430.
[9] J. R. J. Groves and John S. Wilson, Soluble groups with a finiteness condition
arising from Bredon cohomology, Bull. Lond. Math. Soc. 45 (2013), no. 1, 89–
92.
[10] J. L. Harer, The virtual cohomological dimension of the mapping class group
of an orientable surface, Invent. Math. 84 (1986), no. 1, 157–176, DOI
10.1007/BF01388737.
[11] N. V. Ivanov, Subgroups of Teichmüller modular groups, Translations of Mathematical Monographs, vol. 115, American Mathematical Society, Providence,
RI, 1992. Translated from the Russian by E. J. F. Primrose and revised by the
author.
[12] D. Juan-Pineda and I. J. Leary, On classifying spaces for the family of virtually
cyclic subgroups, Recent developments in algebraic topology, 2006, pp. 135–145.
MR2248975 (2007d:19001)
[13] D. Juan-Pineda and A. Trufilo-Negrete, On classifying spaces for the family of virtually cyclic subgroups in mapping class groups (2016). Preprint
arXiv:1606.00306.
[14] Wolfgang Lück, Transformation groups and algebraic K-theory, Lecture Notes
in Mathematics, vol. 1408, Springer-Verlag, Berlin, 1989. Mathematica Gottingensis.
[15] W. Lück, Isomorphism Conjectures in K- and L-Theory, 2015. book, in preparation, http://131.220.77.52/lueck/data/ic.pdf.
[16] W. Lück and D. Meintrup, On the universal space for group actions with
compact isotropy, Geometry and topology: Aarhus (1998), 2000, pp. 293–305.
MR1778113 (2001e:55023)
[17] W. Lück and M. Weiermann, On the classifying space of the family of virtually
cyclic subgroups, Pure App. Math. Q. 8 (2012), no. 2, 479–555.
[18] C. Martinez-Pérez and B.E.A. Nucinkis, Bredon cohomological finiteness conditions for generalisations of Thompson groups, Groups, Geometry, Dynamics
7 (2013), 931–959.
[19] Conchita Martı́nez-Pérez, Francesco Matucci, and Brita E. A. Nucinkis, Cohomological finiteness conditions and centralisers in generalisations of Thompson’s group V , Forum Math. 28 (2016), no. 5, 909–921.
[20] J. D. McCarthy, Subgroups of surface mapping class groups, ProQuest LLC,
Ann Arbor, MI, 1983. Thesis (Ph.D.)–Columbia University.
[21] Derek J. S. Robinson, A course in the theory of groups, Second, Graduate
Texts in Mathematics, vol. 80, Springer-Verlag, New York, 1996.
[22] T. von Puttkamer and X. Wu, On the finiteness of the classifying space for the
family of virtually cyclic subgroups (2016). preprint,arXiv:1607.03790.
, Linear Groups, Conjugacy Growth, and Classifying Spaces for Fam[23]
ilies of Subgroups, International Mathematics Research Notices, to appear
(2017). arXiv:1704.05304.
20
BRITA NUCINKIS AND NANSEN PETROSYAN
Department of Mathematics, Royal holloway, University of London,
Egham, TW20 0EX, UNITED KINGDOM
E-mail address: [email protected]
School of Mathematics, University of Southampton, Southampton,
SO17 1BJ, UNITED KINGDOM
E-mail address: [email protected]
| 4math.GR
|
Rate-Optimal Streaming Codes for Channels
with Burst and Isolated Erasures
M. Nikhil Krishnan, Student Member, IEEE, P. Vijay Kumar, Fellow, IEEE
Electrical Communication Engineering, Indian Institute of Science, Bangalore -
arXiv:1801.05919v1 [cs.IT] 18 Jan 2018
560012
email: {nikhilkrishnan.m, pvk1729}@gmail.com
Abstract
Recovery of data packets from packet erasures in a timely manner is critical for many streaming
applications. An early paper by Martinian and Sundberg introduced a framework for streaming codes
and designed rate-optimal codes that permit delay-constrained recovery from an erasure burst of length
up to B. A recent work by Badr et al. extended this result and introduced a sliding-window channel
model C(N, B, W ). Under this model, in a sliding-window of width W , one of the following erasure
patterns are possible (i) a burst of length at most B or (ii) at most N (possibly non-contiguous) arbitrary
erasures. Badr et al. obtained a rate upper bound for streaming codes that can recover with a time delay
T , from any erasure patterns permissible under the C(N, B, W ) model. However, constructions matching
the bound were absent, except for a few parameter sets. In this paper, we present an explicit family of
codes that achieves the rate upper bound for all feasible parameters N , B, W and T .
I. I NTRODUCTION
In many multimedia streaming applications, where packet losses are a norm, fast recovery of
lost packets is often desirable. Conventional coding schemes like linear block codes, while helpful
in combating erasures, can have increased delays due to buffering of data. In [1], Martinian and
P. Vijay Kumar is also a Visiting Professor at the University of Southern California. This research is supported in part by
the National Science Foundation under Grant 1421848 and in part by an India-Israel UGC-ISF joint research program grant.
M. Nikhil Krishnan would like to acknowledge the support of Visvesvaraya PhD Scheme for Electronics & IT awarded by
Department of Electronics and Information Technology, Government of India.
January 19, 2018
DRAFT
Sundberg introduce a framework for convolutional codes with decoding delay as an explicit
parameter. We refer to these codes as streaming codes. The setting considered is as follows:
At each time instance t ∈ {0, 1, 2, . . .}, an encoder E receives a message packet s[t] from
a source stream. Each message packet s[t] consists of k symbols drawn from a finite field Fq
of size q. In other words, s[t] , (s0 [t] s1 [t] . . . sk−1 [t])| ∈ Fkq . E is a convolutional encoder
and emits a coded packet x[t] at time t, where x[t] , (x0 [t] x1 [t] . . . xn−1 [t])| ∈ Fnq . E is also
assumed to be causal, wherein each coded packet x[t] is a function of the message packets until
time t, i.e., s[0], s[1], . . . , s[t]. Between the encoder-decoder pair, there exists a channel which
introduces erasures at packet level. Let y[t] denote the packet received at the decoder end. We
have:
y[t] =
∗ if x[t] is erased,
x[t] otherwise.
Let ŝ[t] denote the decoded packet corresponding to s[t]. The decoder is delay-constrained
with a delay parameter T . i.e., each decoded message packet ŝ[t] is obtained as a function of
received coded packets {y[0], y[1], . . . , y[t + T ]}. Note that some of these packets can possibly
be erased by the channel. The rate R of the code is naturally defined as nk .
In [1], the authors consider a channel which introduces a burst of length at most B. Together
with the delay-constraint T , the streaming codes designed for this channel can in fact tolerate
multiple erasure bursts, each of length at most B, with a guard space of at least T between
consecutive bursts. However, as the codes presented in [1] are tuned for burst erasures, they are
sensitive to isolated erasures. In a recent work by Badr et al. [2], the authors introduce a slidingwindow based channel model which accounts for burst erasures and isolated erasures. Under the
sliding-window channel model, in any sliding-window of width W , the channel will have one of
the following erasure patterns; (a) a burst erasure of length at most B or (b) up to N erasures at
arbitrary locations within the sliding-window. The channel is denoted by C(N, B, W ). Clearly,
N ≤ B as burst erasures are a special case of arbitrary erasures. Also, T ≥ B as otherwise nonzero rates are not feasible for a streaming code. Taken together with the delay-constraint T , the
C(N, B, W ) model specializes to the burst-erasure-only model in [1], if one chooses W = T + 1
and N = 1. Furthermore, a rate upper bound on codes for this channel has also been derived
in [2]. There are several other follow-up works that study various low-delay communication
schemes like [3], [4], [5], [6] and references therein.
DRAFT
January 19, 2018
In an independent, concurrent work [7], the authors prove the existence of rate-optimal
streaming codes for all parameters.
Our Results: In this work, we prove the tightness of the rate upper bound derived in [2] for
codes that allow both burst erasures and isolated erasures, with a delay-constraint of T . We
prove this by constructing explicit codes that meet the bound with equality. The family of codes
that achieve the bound are based on linearized polynomials, and require a field-size exponential
in T . For various range of parameters, we obtain rate-optimal codes that require lower fieldsize, which is of O(T 2 ). In some cases, even linear and binary field-sizes suffice. We obtain
rate-optimal streaming codes by reducing the problem to the design of linear block codes with
certain properties. For the corner case of B = N , these codes specialize to MDS codes.
II. P RELIMINARIES
A. Streaming Capacity [2]
In the context of streaming codes with decoding delay-constraint T , a rate R is said to be
achievable over C(N, B, W ), if there exists a streaming code with delay parameter T and rate R
that tolerates all the erasure patterns permitted by C(N, B, W ). The supremum of all such rates
is termed as the streaming capacity. In [2], the authors obtain the following upper bound for R:
R≤
Teff − N + 1
,
B + Teff − N + 1
(1)
where Teff = min{T, W − 1}. We shall refer to Teff as the effective delay. Achievability of this
rate bound is not known in general, except for a limited set of parameters. For the burst-alone
case, i.e., N = 1, Maximally Short (MS) codes ([1], [8]) are known to meet the bound (1).
For the other extreme case of N = B, Strongly-MDS convolutional codes [9] are shown to
be optimal in [2]. [10] provides a family of codes with R = 0.5 that meets (1). [2] shows the
existence of near-optimal codes for all feasible parameters B, N, Teff that have a guaranteed rate
of at least
Teff −N
.
B+Teff −N
Let [u, v] , {u, u + 1, . . . , v}. We have the following lemma.
Lemma II.1. If there is a streaming code Cstr which permits recovery of any packet x[t] from
erasure with a delay of at most Teff , even in presence of either (i) a burst erasure (involving
time t) of length at most B or (ii) at most N isolated erasures (again, including time t), then
Cstr tolerates all the erasure patterns permitted by C(N, B, W ) with a delay-constraint T .
January 19, 2018
DRAFT
Proof. Consider any erasure pattern that arises from the C(N, B, W ) model. Let j ∈ {1, 2, . . .}, J ∈
{2, 3, . . .}. Denote by tj , the coordinate at which the j th erasure happened. Let the induction
assumption be that, for all j ≤ (J − 1), x[tj ] is recovered by the ‘time’ (tj + Teff ), i.e., with a
delay Teff ≤ T . Hence during the decoding of x[tJ ], we can assume that, strictly before time tJ ,
there are no erasures occurring among the coordinates [0, tJ + Teff ].
Consider the window of width W consisting of coordinates [tJ , tJ + W − 1]. The coded packet
x[tJ ] is erased by assumption. By definition of the sliding window channel model, in the worst
case, there will be either (i) (B − 1) erasures at coordinates [tJ + 1, tJ + B − 1] or (ii) (N − 1)
erasures within [tJ + 1, tJ + W − 1]. As Teff ≤ (W − 1), when restricted to the set of coordinates
[0, tJ + Teff ], Cstr will be observing either a burst of length ≤ B including coordinate tJ or ≤ N
arbitrary erasures involving tJ (note that the induction assumption removes all erasures occurring
before time tJ ). In either case, Cstr will be able to recover x[tJ ] by time (tJ + Teff ), with delay
Teff ≤ T .
The base case of the induction can be shown to be true by taking J = 1. Thus we have shown
that for all j ≥ 1, x[tj ] can recover from any erasure pattern permitted by C(N, B, W ), with a
delay of Teff ≤ T .
Hence in order to construct streaming codes for C(N, B, W ) with a delay parameter of T ,
it is enough to focus on streaming codes that can handle an erasure burst of length B or N
isolated erasures, with a delay of at most Teff . We drop the subscript from the notation Teff for
brevity and refer to T as the (effective) delay parameter.
B. Diagonal Interleaving
Diagonal interleaving is a known technique that enables one to convert a block code into a
convolutional code. It has been used in the context of streaming codes in works like [1], [8]
to reduce the problem of designing rate-optimal streaming codes for burst-erasure-only case
(N = 1) to that of designing block codes having appropriate features. Consider an [n, k] linear
block code C , where n, k denote the code-length and dimension, respectively. We summarize
the diagonal interleaving using Fig. 1. In the figure, we consider a systematic encoder for C
and hence take xi [t] = si [t] for 0 ≤ i ≤ k − 1. In order to stress the point that the last (n − k)
symbols of a coded packet are parity symbols at any time t, we use the notation pj [t] , xk+j [t],
DRAFT
January 19, 2018
_
𝐷𝑘 2
_
.....
.....
𝐷0
𝐷1
𝐷2
....
𝐷𝑘 3
Parity symbols
at time 𝑡
Systematic Encoder for
_
𝐷𝑘 1
_
𝐷𝑛 𝑘
Fig. 1: A depiction of the diagonal interleaving technique. Here Dj denotes a delay of j time
units.
for 0 ≤ j ≤ n − k − 1. As an example, consider the following parameters for C ; n = 5, k = 3
. The diagonal interleaving technique will result in a convolutional code as shown in Fig. 2.
.......
Fig. 2: An illustration of the diagonal interleaving technique for a [5, 3] code. Each diagonal of
the form (s0 [t], s1 [t + 1], s2 [t + 2], p0 [t + 3], p1 [t + 4]) is a codeword in C , where t ≥ 0. For
notational consistency, define si [t] = pi [t] , 0 for t < 0.
Consider an [n, k] linear block code C over Fq . Let c , (c0 , c1 , c2 , . . . , cn−1 ) denote a codeword
in C .
Definition II.1. For i ∈ [0, n − 1] and I ⊆ [0, n − 1], the ith coordinate of C is said to be
January 19, 2018
DRAFT
recoverable from ( the set of coordinates ) I, if there exist λj ’s chosen from Fq such that:
ci =
X
λj cj ∀c ∈ C .
j∈I
If i ∈ I, it is straightforward to see that ith coordinate is recoverable from I. For parameters
N , B as in the channel model C(N, B, W ) and the delay parameter T , let ∆i , min{i+T, n−1}.
Definition II.2. C is said to be conforming to C(N, B, W ) with delay-constraint T , if both the
following statements hold for all i ∈ [0, n − 1]:
•
ith coordinate of C is recoverable from [0, ∆i ]\N , for all N ⊆ [0, n−1] such that |N | ≤ N .
•
ith coordinate of C is recoverable from [0, ∆i ]\B, where B = [u, v], for all 0 ≤ u ≤ v ≤ n−1
such that (v − u + 1) ≤ B.
One can show without much difficulty that, if there is an [n, k] linear block code C which
conforms to C(N, B, W ) with delay-constraint T , by applying diagonal interleaving, it will result
in a streaming code Cstr that can tolerate an erasure burst of length B or N isolated erasures,
with delay-constraint T . By Lemma II.1, Cstr yields an achievable rate
k
n
over C(N, B, W ) with
effective delay-constraint T . Hence from here onwards we restrict our attention to designing
linear block codes that conform to C(N, B, W ) with (effective) delay-constraint T and have rate
R = k/n that meets (1) with equality. We refer to these codes as rate-optimal linear block codes
with delay-constraint T .
Let T = aB + δ, where a ≥ 0 and 1 ≤ δ ≤ B. If a = 0, δ = B (as B ≤ T ).
III. R ATE -O PTIMAL L INEAR B LOCK C ODES WITH D ELAY C ONSTRAINT T
FOR THE
C ASE
δ ≥ (B − N )
A. Construction-A
Given the channel C(N, B, W ) and effective delay constraint T , consider the parameters n =
(B + T − N + 1) and k = (T − N + 1) for the linear block code C to be constructed over Fq2 .
Let δ ≥ (B − N ) and m , [m0 m1 . . . mk−1 ] be the k-length message vector to be mapped to
the codeword c , (c0 , c1 , . . . , cn−1 ) ∈ C . Consider an [nMDS = T + 1, kMDS = k, dmin = N + 1]
MDS code CMDS over Fq ⊆ Fq2 , with a generator matrix GMDS having all the entries from
Fq ⊆ F2q . Let c0 = [c00 c01 . . . c0T ] = mGMDS . Assign these (T + 1) code-symbols {c0j }Tj=0 to the
coordinates {0, 1, . . . , T − 1} ∪ {n − 1} of c. i.e., ci = c0i for 0 ≤ i ≤ T − 1 and cn−1 = c0T . For
DRAFT
January 19, 2018
the (B − N ) coordinates {cT +j : 0 ≤ j ≤ (B − N − 1)}, take cT +j = αcj + cj+B + . . . + cj+aB .
{z
}
|
(a + 1) terms
Note that, since j ≤ (B − N − 1) ≤ δ − 1, all the (a + 1) code-symbols taking part in the
check-sum belong to the set of coordinates [0, T − 1]. Here α can be any field element from the
set Fq2 \ Fq . Existence of GMDS is guaranteed when q ≥ (T + 1). Hence C can be constructed
with a field-size of O(T 2 ).
Let N = 2, B = 4, T = 10. Hence a = 2, δ = 2. The dimension of C , k = T − N + 1 = 9
and length, n = B + T − N + 1 = 13. This example code is illustrated in Fig. 3.
Fig. 3: An illustration of construction-A for the parameters: N = 2, B = 4, T = 10. Here the
set {c0i }10
i=0 corresponds to code-symbols coming from an [nMDS = 11, kMDS = 9] MDS code.
Proposition III.1. Construction-A gives a rate-optimal family of linear block codes with delay
constraint T .
Proof. (Burst erasure recovery with delay constraint T ) Consider an erasure burst of length B.
Let indicate the number of erasures occurred among the coordinates in [T, T + B − N − 1],
as a part of the erasure burst. Thus, the set of non-erased symbols among the coordinates
[T, T +B−N −1] is given by {ci+T : ≤ i ≤ (B−N −1)}, where ci+T , αci +ci+B +. . .+ci+aB .
As ci+T is assumed to be a non-erased symbol, the burst of length B ends at some coordinate
≤ (T +i−1). Note that, at most one code-symbol in the sum given by each ci+T can be a part of
the burst erasure, as the code-symbols in the sum are chosen to be B apart. Suppose that none of
the code-symbols {ci , ci+B , . . . , ci+aB }, which constitute the sum given by ci+T , are part of the
erasure burst. As the erasure burst is of length B, the burst can start only on or after the coordinate
(i+aB +1). This means that the burst will end at a coordinate ≥ i+aB +B ≥ T +i (as δ ≤ B).
This is a contradiction, and hence each non-erased coordinate in the range [T + , T + B − N − 1]
will have lost precisely one code-symbol from the check-sum it contains. Each of these codesymbols can be recovered from the check-sums. Thus there will be (B − N − ) coordinates
recovered from the (B − ) erasures occurred across coordinates [0, T − 1] ∪ {T − N + B}. The
January 19, 2018
DRAFT
remaining number of erasures is B − − (B − N − ) = N . As C restricted to the coordinates
[0, T − 1] ∪ {T − N + B} is an N erasure correcting MDS code, the remaining erasures can be
corrected as well.
Furthermore, we need to show that all the erased coordinates in the erasure burst can be
recovered with a delay of at most T . Towards this, we need to consider only erasure bursts
involving at least one of the coordinates in [0, B − N − 1]. For all the remaining coordinates,
the delay constraint is trivially met as the last coordinate of C is (T + B − N ). Suppose
ci : i ∈ [0, B − N − 1] is erased (as part of a burst erasure of length B). Clearly, none of the
code-symbols {ci+B , ci+2B , . . . , ci+aB , cT +i } will be part of the erasure burst and hence ci can
be recovered with a delay of T .
(Recovery from N arbitrary erasures with delay constraint T ) C can recover from any N
arbitrary erasures, as it is formed by adding (B − N ) dependent coordinates to an N erasure
correcting MDS code. As in the case of recovery from burst erasures, in order to show that delay
constraints are met during the recovery, one needs to consider only those N -erasure patterns
involving at least one of the code-symbols from the set of coordinates [0, B − N − 1]. Suppose
ci is erased for some i ∈ [0, B − N − 1]. Among the set of T coordinates given by [0, i − 1] ∪
[i, T − 1] ∪ {T + i}, there can possibly be (N − 1) more erasures.
If the coordinate (T + i) is among the erased coordinates, there will be possibly (N − 1)
erasures among the coordinates [0, T − 1], inclusive of the coordinate i. As these coordinates
are part of an MDS code of dimension k = T − N + 1, the code-symbol ci can be recovered
with a delay < T .
Now consider the remaining case that (T + i) is not an erased coordinate. Let G1 be the
systematic k × (T + 1)-generator matrix for the code obtained by puncturing C to the coordinates [0, T − 1] ∪ {T + i}. For 0 ≤ j ≤ T − N , the j th column of G1 is taken as
ej , [ |0 .{z
. . 0} 1| 0 {z. . . 0} ]| . We remark that all the elements of the sub-matrix obtained by
j
(k − j)
restricting G1 to the coordinates corresponding to [0, T − 1], belong to Fq ⊆ Fq2 . Let G2 be the
(T − N ) × (T − 1) sub-matrix obtained from G1 after removing columns corresponding to the
coordinates {i, (T + i)} of C and the ith row. Note that G2 corresponds to the MDS generator
matrix of a code of length (T − 1) and dimension (T − N ). The column of G1 corresponding
to the coordinate (T + i) will have the form [∗| .{z
. . ∗} β ∗ . . . ∗]| , where ∗ indicates elements
| {z }
i
from Fq ⊆
DRAFT
F2q
and β ∈
F2q
\ Fq . Let
G01
(k − i)
be an (T − N + 1) × (T − N + 1) sub-matrix of G1 with
January 19, 2018
(T −N ) columns from [0, i−1]∪[i+1, T −1] and (T +i). After expanding the determinant of G01
based on the column corresponding to the coordinate (T +i), we have det(G01 ) = ∗+β ∗det(G02 ),
where ∗ denotes field elements from Fq ⊆ F2q and G02 is some (T − N ) × (T − N ) sub-matrix
of G2 . As G2 is an MDS generator matrix with all the entries from Fq , det(G02 ) 6= 0 ∈ Fq and
hence det(G01 ) 6= 0. Thus, ci can be recovered with a delay of T .
Corollary III.2. (Binary field-size) When N = 1 and δ ≥ (B − 1), C can be over F2 .
Remark III.1. When B = N , C reduces to an [n = T + 1, k = T − B + 1] MDS code.
Remark III.2. (Linear field-size for N = (B − 1)) For N = (B − 1), if the coordinates 0
and T are swapped in C , with α 6= 0 ∈ Fq , the resultant code will be rate-optimal with delay
constraint T . In this case, field-size requirement will be ≥ (T + 1).
IV. R ATE -O PTIMAL L INEAR B LOCK C ODES WITH D ELAY C ONSTRAINT T
FOR ALL
F EASIBLE PARAMETERS : δ, B, N, T
A. Linearized Polynomials
A linearized polynomial [11] of q-degree (k − 1) is a polynomial of the form f (x) =
Pk−1
qi
i=0 mi x , where mi ∈ Fq m and mk−1 6= 0. It satisfies the property that: f (a1 x1 + a2 x2 ) =
a1 f (x1 ) + a2 f (x2 ), where ai ∈ Fq , xi ∈ Fqm . Furthermore, a linearized polynomial of qdegree (k − 1) can be uniquely determined from evaluations at k points {θi }k−1
i=0 ⊆ Fq m , which
are linearly independent over Fq . Gabidulin codes [12] are constructed based on linearized
polynomial evaluations.
B. Construction-B
Given the channel C(N, B, W ) and effective delay constraint T , consider the parameters n =
(B + T − N + 1), k = (T − N + 1) for the linear block code C to be constructed over Fqm . The k
message symbols {mi }k−1
i=0 will be taken as coefficients of a linearized polynomial with q-degree
(k − 1) and evaluated at (T + 1) points {θi }Ti=0 . Here θi ∈ Fqm and {θi }Ti=0 is a collection of
independent field elements over Fq ⊆ Fqm . Let c00 , c01 , . . . , c0T be the code-symbols (evaluations)
thus obtained. Assign these code-symbols to the coordinates {0, 1, . . . , T − 1} ∪ {n − 1} of c.
i.e., cj = c0j for 0 ≤ j ≤ T − 1 and cn−1 = c0T . Let δ 0 , min{δ, (B − N )}. For coordinates
0 ≤ j ≤ δ 0 − 1, take cT +j = cj + cj+B + . . . + cj+aB . For coordinates in the range δ 0 ≤ j ≤
January 19, 2018
DRAFT
(B − N ) − 1, cT +j = cj + cB+j + . . . + c(a−1)B+j + γj,0 caB + γj,1 caB+1 + . . . + γj,δ−1 caB+δ−1 .
Here {γi,j }’s are selected in such a way that the (B − N − δ 0 ) × δ matrix of the form:
γδ0 ,0
γδ0 ,1
...
γδ0 ,δ−1
γ0
γδ0 +1,1 . . . γδ0 +1,δ−1
δ +1,0
Γ,
,
.
.
..
..
γB−N −1,0 γB−N −1,1 . . . γB−N −1,δ−1
is Cauchy, with all the entries belonging to Fq ⊆ Fqm . Γ can be constructed explicitly (for
example, see [13, Ch. 5]) for q ≥ (δ + B − N − δ) = (B − N ). The evaluation points {θi } can
be chosen as follows. Let β̂ denote a primitive element of the extension field FqT +1 . A natural
candidate for θi , for 0 ≤ i ≤ T will be θi = β̂ i−1 . Hence construction-B is possible with a
field-size ≥ q (T +1) , where q ≥ (B − N ).
Proposition IV.1. Construction-B gives a rate-optimal family of linear block codes, for all
feasible parameters δ, B, N, T .
Proof. (Recovery from N arbitrary erasures with delay constraint T ) Similar to the scenario in
Proposition III.1, C is N -erasure correcting, as C is constructed by adding (B − N ) dependent
coordinates to an N -erasure correcting MDS code (obtained from linearized polynomial evaluations). Hence we only need to show that delay requirements are met for arbitrary erasures
involving at least one of the coordinates in [0, (B − N − 1)]. For i ∈ [0, (B − N − 1)], let i be an
erased coordinate. Consider the set of coordinates Ri , [0, i − 1] ∪ [i + 1, T − 1] ∪ {T + i}. The
corresponding code-symbols will be evaluations of a linearized polynomial at the evaluation
P
points (field elements): Ai ∪ {θi + θ∈Bi θ}, where Ai , {θ0 , θ1 , . . . , θi−1 , θi+1 , . . . , θT −1 },
Bi , {θi , θB+i , . . . , θ(a−1)B+i , θaB+i }, if 0 ≤ i ≤ δ 0 − 1 or Bi , {θi , θB+i , . . . , θ(a−1)B+i , γi,0 θaB ,
γi,1 θaB+1 , . . . , γi,δ−1 θaB+δ−1 }, if δ 0 ≤ i ≤ (B − N − 1).
The set of (T − 1) evaluation points Ai is an independent set over Fq by construction. Adding
P
{θi + θ∈Bi θ} to Ai will not add any dependency, as θi ∈
/ Ai . Thus, access to evaluations at
P
any k = (T − N + 1) out of the T points in Ai ∪ {θi + j∈Bi θj }, will be sufficient to recover all
the k message symbols. This essentially implies that even if there are (N − 1) erasures among
the coordinates in Ri , ci can still be corrected with a delay of at most T .
(Burst erasure recovery with delay constraint T ) If δ ≥ (B − N ), proof exactly follows as
in the case of Proposition III.1. For the remaining case of δ < (B − N ), proof is moved to
Appendix A.
DRAFT
January 19, 2018
Example IV.1. Consider the parameters B = 7, N = 2, δ = 3, T = 10. Hence for C , we have
the parameters n = (B + T − N + 1) = 16 and k = (T − N + 1) = 9. Also, δ = δ 0 = 3. Let α̂
be a primitive element of F5 and
1 α̂ α̂2
.
Γ=
1 α̂2 α̂4
Consider the set of (T + 1) = 11 field elements {θi }10
i=0 , which forms an independent set over
F5 . For example, let β̂ be a primitive element of F5m , where m ≥ (T + 1) = 11. Then, θi can
be chosen as β̂ i .
Let {mi }8i=0 be the set of k = 9 message symbols drawn from F5m . Consider the corresponding
P
i
linearized polynomial f (x) = 8i=0 mi x5 . The code-symbols will be as follows: ci = f (θi ), for
0 ≤ i ≤ (T − 1) = 9, c10 = c0 + c7 = f (θ0 + θ7 ), c11 = c1 + c8 = f (θ1 + θ8 ), c12 = c2 + c9 =
f (θ2 + θ9 ), c13 = c3 + c7 + α̂c8 + α̂2 c9 = f (θ3 + θ7 + α̂θ8 + α̂2 θ9 ), c14 = c4 + c7 + α̂2 c8 + α̂4 c9 =
f (θ4 + θ7 + α̂2 θ8 + α̂4 θ9 ), c15 = f (θ10 ).
Consider the recovery of C from N = 2 arbitrary erasures, with a delay at most 10. Clearly,
C can tolerate 2 erasures. This is because, it is constructed by adding (B − N ) = 5 additional
parity checks (at coordinates 10, 11, 12, 13 and 14) to a 2-erasure correcting [11, 9]-MDS code.
As the last coordinate of C is 15, delay constraint of 10 will be trivially met during the recovery
of any cj : j ≥ 5, which is a part of 2 arbitrary erasures. Hence we need to consider only
those 2-erasure patterns which involve at least one coordinate from [0, 4]. Suppose c1 is erased.
Consider the T = 10 coordinates, R1 = {0} ∪ [2, 9] ∪ {11}. The corresponding 10 evaluation
points {θ0 , θ2 , θ3 , . . . , θ9 , (θ1 +θ8 )} form an independent set over F5 . Therefore, even if one more
coordinate is lost (since N = 2, and coordinate 1 is already assumed to be lost) from R1 , there
will still be k = 9 independent evaluation points available from the non-erased coordinates in
R1 . As the last coordinate in R1 is (T + 1) = 11, c1 can be recovered with a delay at most 10.
Similar arguments hold when c0 , c2 , c3 or c4 is part of a 2-erasure pattern.
For the burst erasure case, consider consecutive erasures of length B = 7, which erases the set
of coordinates [u, v], where 0 ≤ u ≤ v ≤ (n − 1) and v = u + B − 1. Let denote the number of
coordinates in [u, v] that intersect [T, T + B − N − 1] = [10, 14]. Clearly, 0 ≤ ≤ (B − N ) = 5.
•
= 5: If the burst erasure is such that = (B − N ) = 5, the [11, 9]-MDS code across
coordinates [0, 9] ∪ {15} will have B − (B − N ) = N = 2 coordinates erased and hence
all the erasures can be corrected.
January 19, 2018
DRAFT
•
= 0: Each symbol from [10, 14] will recover one unique erased code-symbol. Let [u, u +
6] ⊆ [0, 9] denote the B = 7 coordinates involved in the burst erasure. Clearly, there exists
j0 , j1 , . . . , j4 ∈ [u, u+6] such that j` mod B = `, for 0 ≤ ` ≤ B −N −1 = 4. Code-symbol
at coordinate 10 + ` will be recovering precisely the corresponding j` .
•
= 1: The erased coordinates belong to [4, 10]. As c11 = c1 +c8 and c12 = c2 +c9 , these two
code-symbols recover coordinates 8 and 9, respectively. From c13 (= c3 + c7 + α̂c8 + α̂2 c9 )
and c14 (= c4 +c7 + α̂2 c8 + α̂4 ), after removing the interference from the known code-symbols
{c3 , c8 , c9 }, we obtain the sums ĉ13 = c7 and ĉ14 = c4 + c7 , respectively. These two sums
essentially yield the symbols c4 and c7 .
•
= 2: The erased coordinates are in [5, 11]. As c12 = c2 + c9 , it recovers coordinate 9.
After removing the known code-symbols {c3 , c4 , c9 }, the check-sums provided by c13 and
c14 result in ĉ13 = c7 + α̂c8 and ĉ14 = c7 + α̂2 c8 , respectively. These sums would recover
the code-symbols c7 and c8 (uses the Cauchy property of Γ).
•
= 3: Here the erased coordinates are from the set [6, 12]. After removing the interference
from the known code-symbols {c3 , c4 }, ĉ13 = c7 + α̂c8 + α̂2 c9 = f (θ7 + α̂θ8 + α̂2 ), ĉ14 =
c7 + α̂2 c8 + α̂4 c9 = f (θ7 + α̂2 θ8 + α̂4 θ9 ). From the Cauchy property of Γ, the evaluation
points {θ7 + α̂θ8 + α̂2 θ9 , θ7 + α̂2 θ8 + α̂4 θ9 } are both independent. These evaluation points
along with {θ0 , θ1 , . . . , θ5 , θ10 } form an independent set of cardinality, k = 9.
•
= 4: The erased coordinates are in [7, 13]. After removing the interference from the known
code-symbols {c4 }, ĉ14 = c7 + α̂2 c8 + α̂4 c9 = f (θ7 + α̂2 θ8 + α̂4 θ9 ). θ7 + α̂2 θ8 + α̂4 θ9 along
with {θ0 , θ1 , . . . , θ6 , θ10 } result in an independent set of size k = 9.
Thus, for every , the (5 − ) non-erased coordinates [10 + , 14] will provide one linearized
polynomial evaluation each. These (5 − ) evaluations along with the 11 − (7 − ) evaluations
corresponding to the non-erased coordinates in [0, 10] ∪ {15}, result in linearized polynomial
evaluations over 9 independent evaluation points and hence all the erased symbols can be
recovered.
In order to show that the delay constraint of T = 10 is met during the recovery from burst
erasures, we need to consider only the burst erasures involving at least one of the coordinates
[0, B − N − 1] = [0, 4]. Suppose ci , where i ∈ [0, 2], is part of a burst-erasure. The coordinates
{7 + i, 10 + i} will be non-erased, as they are at least B = 7 apart from coordinate i. Hence
using the check-sum provided by c10+i , the code-symbol ci can be recovered, with a delay of
10 time units.
DRAFT
January 19, 2018
For ci , where i ∈ [3, 4], which is part of a burst erasure, the symbol c10+i will be non-erased.
There are three possible cases for (i, ); (3, 0), (4, 0) and (4, 1). If = 0, the code-symbols
{c7 , c8 , c9 } will be known by the time (10 + i). After removing the interference from the known
symbols at coordinates {7, 8, 9}, c10+i will yield the code-symbol ci . If = 1, we have already
seen that, c4 will be recovered at time 14.
A PPENDIX A
P ROOF OF P ROPOSITION IV.1: B URST E RASURES , C ASE : δ < (B − N )
Consider the set of erased coordinates, B , [u, v], where 0 ≤ u ≤ v ≤ n−1 and v−u+1 = B.
Let |B ∩ [T, T + B − N − 1]| = .
1) = (B − N ): The number of erasures across the coordinates {0, T − 1} ∪ {T − N + B}
will be B − (B − N ) = N , which can be corrected.
2) = 0: Each code-symbol from the set {cT , cT +1 , . . . , cT +δ−1 } will correct one erased symbol
(argument similar to that in the proof of Proposition III.1). This essentially means that all
the code-symbols which are part of the check-sums given by {cT , cT +1 , . . . , cT +δ−1 }, are
known by time T + δ − 1. In particular, we have {caB , caB+1 , . . . , caB+δ−1 } known by
time T + δ − 1, irrespective of whether some of them are erased or not. From each codesymbol cT +j , cj + cB+j + . . . + c(a−1)B+j + γj,0 caB + γj,1 caB+1 + . . . + γj,δ−1 caB+δ−1 , for
δ ≤ j ≤ B −N −1, the interference from {caB , caB+1 , . . . , caB+δ−1 } can be canceled. Hence
cT +j ’s can be thought of as simply, ĉT +j = cj +cB+j +. . .+c(a−1)B+j . For 0 ≤ i ≤ B−N −1,
let `i be the unique coordinate ∈ [u, v] such that `i mod B = i. The check-sum provided
by the code-symbol at coordinate (T + i) will be recovering precisely the erased symbol c`i .
Thus in total, the set of symbols {cT , cT +1 , . . . , cT +B−N −1 } will correct (B − N ) erasures.
3) 0 < < (B − N ): Let ζ indicate the number of non-erased coordinates lying in [T + , T +
δ − 1]. Clearly, ζ > 0 iff < δ. Assuming ζ > 0, each such non-erased ci+T that lies in
≤ i ≤ δ − 1, helps in recovering one unique erased symbol from the erasure burst. The
argument, again, is similar to the one in the proof of Proposition III.1. This also means that
the symbols {caB+j : ≤ j ≤ δ − 1} are known by time T + δ − 1.
Now let us see how non-erased symbols in the range cT +i : max{δ, } ≤ i ≤ (B − N − 1)
help in the recovery of erased symbols. Note that each cT +i in the range max{δ, } ≤ i ≤
(B − N − 1) takes the form: cT +i = ci + cB+i + . . . + c(a−1)B+i + γi,0 caB + γi,1 caB+1 +
. . . + γi,δ−1 caB+δ−1 . As we assume > 0, we can remove interference from the set of
January 19, 2018
DRAFT
non-erased symbols {ci , cB+i , . . . , c(a−2)B+i } (these can be easily verified to be non-erased
symbols, when > 0). For ζ > 0, we can remove the interference from the set of symbols
{caB+j : ≤ j ≤ δ − 1} (these are known by time T + δ − 1). Thus effectively, each
non-erased symbol cT +i in the range max{δ, } ≤ i ≤ (B − N − 1) provides the sum
ĉT +i = c(a−1)B+i + γi,0 caB + γi,1 caB+1 + . . . + γi,min{,δ}−1 caB+min{,δ}−1 . There are the
following two possibilities.
•
(B − ) ≤ (N + δ): This condition essentially means that none of the c(a−1)B+i terms
appearing in ĉT +i , for max{δ, } ≤ i ≤ (B − N − 1), are part of the burst erasure
considered. In this scenario, we can further remove the interfering c(a−1)B+i ’s from ĉT +i ’s,
where max{δ, } ≤ i ≤ (B − N − 1). Furthermore, if (B − ) < δ, not all of the symbols
{caB , caB+1 , . . . , caB+δ−1 } will be erased. Hence we can rewrite the effective check-sum
contributed by each ĉT +i as ĉT +i = γi,max{δ−B+,0} caB+max{δ−B+,0} + γi,1 caB+1 + . . . +
γi,min{,δ}−1 caB+min{,δ}−1 . The number of remaining erased symbols caB+i in the range
0 ≤ i ≤ min{, δ} − 1, {caB+max{δ−B+,0} , caB+max{δ−B+,0}+1 , . . . , caB+min{,δ}−1 } and
number of non-erased symbols in the coordinate range [T + max{, δ}, T + B − N − 1]
is given in Table I, for all the cases with (B − ) ≤ (N + δ). From Table I, for all the
cases, the number of remaining erased symbols ≥ number of code-symbols (evaluations)
in the range [T + max{, δ}, T + B − N − 1]. As γx,y ’s are chosen carefully as part of
a Cauchy matrix, all these equations (evaluations) will be independent. Thus the number
of independent equations provided by coordinates in the range [T + , T + B − N − 1]
is equal to (B − N − ).
•
(B − ) > (N + δ): Let α0 , (B − − N − δ). The number of erased symbols recovered
by the non-erased coordinates lying in [T + , T + δ − 1] is max{δ − , 0}. Note that
α0 < (B − N − δ). For each c(a−1)B+i such that i ∈ [B − N − α0 , B − N − 1], there
exists a non-erased cT +i which results in a ĉT +i = c(a−1)B+i + γi,0 caB + γi,1 caB+1 +
. . . + γi,min{,δ}−1 caB+min{,δ}−1 , after removing the known interfering symbols. To the
contrary, if cT +i is part of the erased symbols, this would mean that B ≥ (B − N −
i) + N + δ + i =⇒ δ ≤ 0, which is a contradiction. Thus cT +i for i ∈ [B − N −
α0 , B − N − 1] gives α0 independent equations as they all give one unique component
c(a−1)B+i each. For j ∈ [max{, δ}, B − N − α0 − 1], cT +j ’s are non-erased and results in
a ĉT +j = γj,0 caB + γj,1 caB+1 + . . . + γj,min{,δ}−1 caB+min{,δ}−1 , after removing the known
DRAFT
January 19, 2018
Condition
No. of remaining erased (unknown)
No. of equations
symbols among the coordinates
from the coordinates
[aB, aB + δ − 1]
[T + max{, δ}, T + B − N − 1]
(B − ) ≥ δ, ≥ δ
δ
(B − N − )
(B − ) ≥ δ, < δ
(B − N − δ)
(B − ) < δ, ≥ δ
(B − )
(B − N − )
(B − ) < δ, < δ
(B − δ)
(B − N − δ)
Table I: A summary of all the conditions, when (B − ) ≤ (N + δ).
interfering symbols. These corresponds to (B−N −α0 −max{, δ}) = min{, δ} equations
and min{, δ} unknowns and hence are all independent. Therefore, the total number of
independent evaluations supplied by the non-erased coordinates in [T + , T + B − N − 1]
is max{δ − , 0} + α0 + min{, δ} = δ + α0 = B − N − .
Therefore for all the cases of 0 ≤ ≤ (B − N ), (B − N − ) non-erased code-symbols in the
range [T +, T +B −N −1] provide that many evaluations, which are independent (over Fq ), with
respect to the T + 1 − (B − ) evaluation points corresponding to the (T − B + + 1) non-erased
locations in [0, T − 1] ∪ {T − N + B}. This essentially adds up to T − N + 1 evaluations over
independent evaluation points. As k = T − N + 1, all the erasures can be corrected.
In order to prove that the delay constraint of T is met during the recovery from burst erasures,
we need to consider only the burst erasures involving at least one of the coordinates [0, B −
N − 1]. It is enough to show that for any erasure burst of length B that starts at coordinate
i ∈ [0, B − N − 1], the code-symbol ci can be recovered with a delay of at most T .
If ci , where i ∈ [0, δ − 1], is part of a burst-erasure, we fall back on the proof of the
corresponding part in Proposition III.1 to see that delay conditions are met. In the following
we consider the case of i ∈ [δ, B − N − 1].
Consider an erasure burst which starts at i, where i ∈ [δ, B −N −1]. The symbols {ci+B , ci+2B ,
. . . , ci+(a−1)B , cT +i } will be non-erased. Let be as defined previously. If = 0, all the codesymbols in {caB , caB+1 , . . . , caB+δ−1 } will be known by the time (T + i). This follows from the
fact that all the symbols cT , cT +1 , . . . , cT +δ−1 are non-erased. After removing the interference
January 19, 2018
DRAFT
from known symbols at coordinates {i+B, i+2B, . . . , i+(a−1)B, aB, aB +1, . . . , aB +δ −1},
the code-symbol cT +i will yield ci .
For > 0, we first note that, for a burst starting at coordinate i ∈ [δ, B − N − 1], a is
forced to be equal to 1. i.e., T = B + δ. From the max{δ − , 0} non-erased coordinates in
[T + , T + δ − 1], we can obtain code-symbols in the coordinate range [aB + , aB + δ − 1 =
T − 1]. Note that i = T − B + = δ + . Number of non-erased coordinates in the range
[T + max{δ, }, T + i] = i − max{δ, } + 1. For each j ∈ [max{δ, }, i − 1], cj is a nonerased symbol. This follows from our assumption that, the error burst is starting at time i. For
j ∈ [max{δ, }, i], cT +j , cj + γj,0 cB + γj,1 cB+1 + . . . + γj,δ−1 cB+δ−1 (as a = 1). After removing
the known symbols from the coordinate range [B + , B + δ − 1] and [max{δ, }, i − 1], we have
the resultant sums:
ĉT +j =
γj,0 cB + γj,1 cB+1 + . . . + γj,min{,δ}−1 cB+min{,δ}−1 if max{δ, } ≤ j ≤ i − 1,
cj + γj,0 cB + γj,1 cB+1 + . . . + γj,min{,δ}−1 cB+min{,δ}−1 j = i.
For the i − max{δ, } = δ + − max{δ, } = min{δ, } equations in the range max{δ, } ≤
j ≤ i − 1, there are that many unknowns. As γx,y ’s are carefully chosen to be part of a
Cauchy matrix, these equations can be solved (at time (T + i − 1)) to obtain the code-symbols:
{cB , cB+1 , . . . , cB+min{,δ} }. Thus at time (T + i), the contribution from these code-symbols can
be removed from ĉT +i to obtain ci . This completes the proof.
R EFERENCES
[1] E. Martinian and C. W. Sundberg, “Burst erasure correction codes with low decoding delay,” IEEE Trans. Inf. Theory,
vol. 50, no. 10, pp. 2494–2502, 2004.
[2] A. Badr, P. Patil, A. Khisti, W. Tan, and J. G. Apostolopoulos, “Layered Constructions for Low-Delay Streaming Codes,”
IEEE Trans. Inf. Theory, vol. 63, no. 1, pp. 111–141, 2017.
[3] Z. Li, A. Khisti, and B. Girod, “Correcting erasure bursts with minimum decoding delay,” in Proc. Asilomar Conference
on Signals, Systems and Computers, 2011, pp. 33–39.
[4] Ö. F. Tekin, T. Ho, H. Yao, and S. Jaggi, “On erasure correction coding for streaming,” in Proc. Information Theory and
Applications Workshop, 2012, pp. 221–226.
[5] D. Leong, A. Qureshi, and T. Ho, “On coding for real-time streaming under packet erasures,” in Proc. Int. Symp. Inf.
Theory, 2013, pp. 1012–1016.
[6] N. Adler and Y. Cassuto, “Burst-Erasure Correcting Codes With Optimal Average Delay,” IEEE Trans. Inf. Theory, vol. 63,
no. 5, pp. 2848–2865, 2017.
[7] S. L. Fong, A. Khisti, B. Li, W. Tan, X. Zhu, and J. Apostolopoulos, “Optimal Streaming Codes for Channels with Burst
and Arbitrary Erasures,” arXiv preprint arXiv:1801.04241, 2018.
DRAFT
January 19, 2018
[8] E. Martinian and M. Trott, “Delay-Optimal Burst Erasure Code Construction,” in Proc. Int. Symp. Inf. Theory.
IEEE,
2007, pp. 1006–1010.
[9] H. Gluesing-Luerssen, J. Rosenthal, and R. Smarandache, “Strongly-MDS convolutional codes,” IEEE Trans. Inf. Theory,
vol. 52, no. 2, pp. 584–598, 2006.
[10] A. Badr, A. Khisti, W. Tan, and J. G. Apostolopoulos, “Streaming codes for channels with burst and isolated erasures,” in
Proc. INFOCOM.
IEEE, 2013, pp. 2850–2858.
[11] R. Lidl and H. Niederreiter, Introduction to Finite Fields and Their Applications.
New York, NY, USA: Cambridge
University Press, 1986.
[12] E. Gabidulin, “Theory of Codes with maximum rank distance,” Problems of Information Transmission, vol. 21, no. 7, pp.
1–12, 1985.
[13] R. M. Roth, Introduction to coding theory.
January 19, 2018
Cambridge University Press, 2006.
DRAFT
| 7cs.IT
|
1
Distributed Frequency Control with Operational
Constraints, Part II: Network Power Balance
arXiv:1703.00083v1 [cs.SY] 28 Feb 2017
Zhaojian Wang, Feng Liu, Steven H. Low, Fellow, IEEE, Changhong Zhao, and Shengwei Mei Fellow, IEEE
Abstract—In Part I of this paper we propose a decentralized
optimal frequency control of multi-area power system with operational constraints, where the tie-line powers remain unchanged
in the steady state and the power mismatch is balanced within
individual control areas. In Part II of the paper, we propose a
distributed controller for optimal frequency control in the network power balance case, where the power mismatch is balanced
over the whole system. With the proposed controller, the tieline powers remain within the acceptable range at equilibrium,
while the regulation capacity constraints are satisfied both at
equilibrium and during transient. It is revealed that the closedloop system with the proposed controller carries out primal-dual
updates with saturation for solving an associated optimization
problem. To cope with discontinuous dynamics of the closed-loop
system, we deploy the invariance principle for nonpathological
Lyapunov function to prove its asymptotic stability. Simulation
results are provided to show the effectiveness of our controller.
Index Terms—Power system dynamics, frequency control;
network power balance; distributed control.
I. I NTRODUCTION
In Part I of the paper we have investigated the optimal
frequency control of multi-area power system with operational
constraints [1]. In that case, the tie-line powers are required to
be unchanged in the steady state after load disturbances, which
implies the power mismatch in each area has to be balanced
individually. It is referred to as the per-node power balance
case. In Part II of the paper, we consider the transmission
congestion in the distributed optimal frequency design.
The per-node balance case in Part I mainly considers the
situation where the power delivered from one area to another is
fixed, e.g. contract power, which should not be violated in normal operation. However, in some circumstances, control areas
may cooperate for better frequency recovery or regulation cost
reduction. In this case, power mismatch may be balanced by all
generations and controllable loads among all control areas in
cooperation. Similar situations also appear in one control area
with multiple generators and controllable loads that cooperate
to eliminate the power mismatch in the area. It is referred
to as the network power balance case. Compared with the
This work was supported by the National Natural Science Foundation of
China ( No. 51377092, No. 51677100, No. 51621065), Foundation of Chinese
Scholarship Council (CSC No. 201506215034), the US National Science
Foundation through awards EPCN 1619352, CCF 1637598, CNS 1545096,
ARPA-E award DE-AR0000699, and Skoltech through Collaboration Agreement 1075-MRA.
Z. Wang, F. Liu and S. Mei are with the Department of Electrical Engineering, Tsinghua University, Beijing, China, 100084 e-mail:
([email protected]).
S. H. Low and C. Zhao are with the Department of Electrical Engineering, California Institute of Technology, Pasadena, CA, USA, 91105 email:([email protected])
per-node balance case, the most challenging problem in this
case is that the tie-line powers may change and congestions
may occur. In addition, local information is not sufficient
and neighboring information turns to be helpful. As for the
constraints, the tie-line power constraints are not hard limits,
which only need to be satisfied at equilibrium. The capacity
limits on the generations and controllable loads also required
to be satisfied both in steady state and during transient.
In the recent literature of frequency control, tie-line power
constraints are considered in [2]–[7]. In [2], tie-line power
constraints are included in the load-side secondary frequency
control. A virtual variable is used to estimate the tie-line
power, whose value is identical to the tie-line power at
equilibrium. In [3], [4], an optimal economic dispatch problem
including tie-line power constraints is formulated, then the
solution dynamics derived from a primal-dual algorithm is
shaped as a port-Hamiltonian form. The power system dynamics also have a port-Hamiltonian form, which are interconnected with the solution dynamics to constitute a closedloop Hamiltonian system. Then, the optimality and stability are
proved. In [6], a unified method is proposed for primary and
secondary frequency control, where the congestion management is implemented in the secondary control. In [7], a realtime control framework is proposed for tree power networks,
where transmission capacities are considered.
Similar to the per-node balance case, hard limits, such as
capacity constraints of power injections on buses, are enforced
only in the steady-state in the literature, which may fail if such
constraints are violated in transient. Here we construct a fully
distributed control to recover nominal frequency while eliminating congestion. Differing from the literature, it enforces
regulation capacity constraints not only at equilibrium, but
also during transient. We show that the controllers together
with the physical dynamics serve as primal-dual updates with
saturation for solving the optimization problem. The optimal
solution of the optimization problem and the equilibrium point
of closed-loop system are identical.
The enforcement of capacity constraints during transient and
tie-line power limits simultaneously makes the stability proof
difficult. Specifically, the Lyapunov function is not continuous
anymore, as in the per-node case in Part I of the paper. In
this situation, the conventional LaSalle’s invariance principle
does not apply. To overcome the difficulty, we construct a
nonpathological Lyapunov function to mitigate the impacts of
nonsmooth dynamics. The salient features of the controller
are:
1) Control goals: the controller restores the nominal frequency and balance the power mismatch in the whole
2
system after unknown load disturbance while minimizing the regulation costs;
2) Constraints: the regulation capacity constraints are always enforced even during transient and the congestions
can be eliminated automatically;
3) Communication: only neighborhood communication is
needed in the network balance case;
4) Measurement: the controller is adaptive to unknown
load disturbances automatically with no need of load
measurement.
The rest of this paper is organized as follows. In Section II, we describe our model. Section III formulates the
optimal frequency control problem in the network balance
case, presents the distributed frequency controller and proves
the optimality, uniqueness and stability of the closed-loop
equilibrium. Simulation results are given in Section IV. Section
V concludes the paper.
nominal or scheduled values so that, e.g., ω j (t) = 0 means the
frequency is at its nominal value.
In this paper, all nodes cooperate to rebalance power over
the entire network after a disturbance. The power flows Pi j on
the tie lines may deviate from their scheduled values and we
require that they satisfy line limits, i.e.,
∀(i, j) ∈ E
Pi j ≤ Pi j ≤ Pi j
for some upper and lower bounds Pi j , Pi j .
In DC approximation the power flow on line (i, j) is given
by Pi j = Bi j (θi − θ j ). Hence line flow constraints in the pernode balance case is θi = θ j for all (i, j) ∈ E and in the network
balance case is:
∀(i, j) ∈ E
θ i j ≤ θi − θ j ≤ θ i j
(3)
II. N ETWORK MODEL
We summarize the notation used in Part I [1]. The power
network is model by a directed graph G := (N, E) where N =
{0, 1, 2, ...n} is the set of nodes (control areas) and E ⊆ N ×
N is the set of edges (tie lines). If a pair of nodes i and j
are connected by a tie line directly, we denote the tie line
by (i, j) ∈ E. Let m := |E| denote the number of tie lines.
Use (i, j) ∈ E or i → j interchangeably to denote a directed
edge from i to j. Assume the graph is connected and node 0
is a reference node. For each node j ∈ N, θ j (t) denotes the
rotor angle at node j at time t and ω j (t) is the frequency.
Pjg (t) denotes the (aggregate) generation at node j at time t
and ugj (t) is its generation control command. Pjl (t) denotes
the (aggregate) controllable load and ulj (t) is its load control
command. p j is the (aggregate) uncontrollable load.
The power system dynamics for each node j ∈ N is
θ̇ j = ω j (t)
where θ i j = Pi j /Bi j , θ i j = Pi j /Bi j are lower and upper bounds
on angle differences.
As the generation Pjg and load Pjl in each area can increase
or decrease, and a line flow Pi j can in either direction, we
make the following assumption.
A1:
g
l
1) Pgj < 0 < P j and Plj < 0 < P j for ∀ j ∈ N.
2) θ i j ≤ 0 ≤ θ i j for (i, j) ∈ E.
3) θ0 (t) := 0 and φ0 (t) := 0 for all t ≥ 0.
Here φ is a vector variable that represents virtual phase angles
in the network balance case in Section III. The assumption
θ0 ≡ φ0 ≡ 0 amounts to using (θ0 (t), φ0 (t)) as reference angles.
It is made merely for notational convenience: as we will see,
the equilibrium point will be unique with this assumption (or
unique up to reference angles without this assumption).
(1a)
M j ω̇ j = Pjg (t) − Pjl (t) − p j − D j ω j (t)
Bi j (θi (t) − θ j (t)) −
B jk (θ j (t) − θk (t))
i:i→ j
k: j→k
= −Pjg (t) + ugj (t) − ω j (t)/R j
= −Pjl (t) + ulj (t)
+
T jg Ṗjg
T jl Ṗjl
∑
∑
(1b)
(1c)
(1d)
where D j > 0 are damping constants, R j > 0 are droop
parameters, and B jk > 0 are line parameters that depend on
the reactance of the line ( j, k). Let x := (θ , ω, Pg , Pl ) denote
the state of the network and u := (ug , ul ) denote the control.
The capacity constraints are:
Pgj
Pjg (t)
≤
Plj
≤ Pjl (t)
g
≤ Pj,
l
≤ P j,
III. C ONTROLLER D ESIGN FOR T HE N ETWORK P OWER
BALANCE C ASE
j∈N
(2a)
j∈N
(2b)
Here (2a) and (2b) are hard limits on the regulation capacities
of generation and controllable load at each node, which should
not be violated at any time even during transient.
The system operates in a steady state initially, i.e., the
generation and the load are balanced and the frequency is at
its nominal value. All variables represent deviations from their
In the per-node balance case, individual control areas rebalance power within their own areas after disturbances. However,
in many circumstances, it may be more efficient for all control
areas to eliminate power imbalance of the overall system in a
coordinated manner. This can be modeled as the condition:
∑ j Pjg
=
∑j
Pjl + p j
(4)
In this case the tie-line flows may not be restored to their predisturbance values. To ensure that they are within operational
limits, the constraints (3) are imposed.
Even though the philosophy of the controller design as well
as the proofs are similar to the per-node case, the details are
much more complicated. Our presentation will however be
brief where there is no confusion.
3
A. Control goals
In the network power balance case, the control goals are
formalized as the following optimization problem.
NBO: min
z2j
α j 2
β j 2
Dj
∑ 2 Pjg + ∑ 2 Pjl + ∑ 2 ω 2j + ∑ 2
j
j
j
j
(5a)
g
l
g
l
over
x := (θ , φ , ω, P , P ) and u := (u , u )
s. t.
(2),
Pjg = Pjl + p j +U j (θ , ω)
Pjg
= Pjl
+ p j + Û j (φ )
θ i j ≤ φi − φ j ≤ θ i j ,
Pjg
Pjl
=
=
ugj ,
ulj ,
j∈N
Even though neither the network balance condition (4) nor
the line limits (3) are explicitly enforced in (5), they are
satisfied at optimality (Theorem 2 below). Indeed, the virtual
phase angles φ and the conditions (5b)–(5d) are carefully
designed to enforce these conditions as well as to restore
the nominal frequency ω ∗ = 0 at optimality. This technique
is previously used in [2].
(5b)
j∈N
(5c)
(i, j) ∈ E
(5d)
j∈N
(5e)
j∈N
(5f)
where α j > 0, β j > 0 are constant weights; z j is a shorthand
defined for convenience as
zj
2) Even though any feasible point of (5) has z j = 0, the
objective function is augmented with z2j to improve
convergence (see [8]).
:= Pjg − Pjl − p j − Û j (φ )
U(θ , ω) := Dω +CBCT θ , and Û(φ ) := CBCT φ .
While θ represents the phase angles in the physical power
network, φ is a cyber quantity that can be interpreted as virtual
phase angles (see remarks below). The matrices D, C and B
are defined as in the previous section.
As in the per-node case, we define the variables θ̃i j := θi −
θ j , or in vector form, θ̃ := CT θ . As we fix θ0 := 0 to be
a reference angle under assumption A1, θ̃ = CT θ defines a
bijection between θ and θ̃ . Similarly we define φ̃i j := φi − φ j
or φ̃ := CT φ , and φ0 := 0 so there is a bijection between φ
and φ̃ . Note that both θ̃ and φ̃ are restricted to the column
space of CT . We will use (θ , φ ) and (θ̃ , φ̃ ) interchangeably.
For instance we will abuse notation and write Û(φ ) := CBCT φ
or Û(φ̃ ) := CBφ̃ .
We now summarize some of the interesting properties
of NBO (5) that will be proved formally in the next two
subsections. We first compare NBO (5) with PBO in [1] for
the per-node balance case.
Remark 1 (Comparison of NBO and PBO).
1)
Intuitively the network balance condition (4) is a
relaxation of the per-node balance condition (3) in [1],
and hence we expect that the optimal cost of NBO
lower bounds that of PBO. This is indeed the case,
as we now argue. Constraint (5c) implies that any
feasible point of (5) has z j = 0 and hence these two
optimization problems have the same objective function.
Their variables and constraints are different in that PBO
directly enforces the per-node balance condition while
NBO (5) has the additional variable φ and constraints
∗
(5c)(5d). Any optimal point (θ ∗ , ω ∗ , Pg∗ , Pl ) for PBO
however defines a feasible point (θ ∗ , φ , ω ∗ , Pg∗ , Pl∗ )
for NBO (5) with the same cost where φ = θ ∗ . The
point (θ ∗ , φ , ω ∗ , Pg∗ , Pl∗ ) satisfies (5c)(5d) because
∗
(θ ∗ , ω ∗ , Pg∗ , Pl ) satisfies (5b), ω ∗ = 0 and θi∗ = θ j∗ by
Theorem 2 in [1], and θ i j ≤ 0 ≤ θ i j by assumption A1.
Remark 2 (Virtual phase angles φ ).
1) Under mild conditions, ω ∗ = 0 at optimality for both PBO and NBO.
For NBO, this is a consequence of the constraint (5c)
on φ ; see Lemma A.1. Summing (5c) over all j ∈ N
also implies the network balance condition (4) since
1T Û(φ ) = 1T CBCT φ = 0.
2) In PBO, θi∗ = θ j∗ at optimality (i.e., tie-line flows are
restored Pi∗j = 0) and U(θ ∗ , ω ∗ ) = 0. This does not
necessarily hold in NBO. However φ is regarded as
virtual phase angles because, at optimality, φ ∗ differs
from the real phase angles θ ∗ only by a constant,
φ ∗ − θ ∗ = 1(φ0 − θ0 ) (Lemma A.1 in the appendix).
Hence CT φ ∗ − CT θ ∗ = CT · 1(φ0 − θ0 ) = 0, implying
φ̃i∗j = θ̃i∗j . Then the constraints (5d) are exactly the flow
constraints (3). In other words, we impose the flow
constraints on θ̃ indirectly by enforcing such constraints
on the virtual angle φ̃ . .
B. Distributed controller
Our control laws are:
λ̇ j = γ λj Pjg (t) − Pjl (t) − p j − Û j (φ̃ (t)) , j ∈ N
(6a)
η̇i+j
=
γiηj [φ̃i j (t) − θ i j ]+
η+
∀(i, j) ∈ E
(6b)
η̇i−j
= γiηj [θ i j − φ̃i j (t)]+
η−
∀(i, j) ∈ E
(6c)
φ̃˙i j
ij
ij
φ̃
= γi j (Bi j [λi (t) − λ j (t) + zi (t) − z j (t)]
+ ηi−j (t) − ηi+j (t)
∀(i, j) ∈ E
(6d)
h
iPgj
ugj (t) = Pjg (t) − γ gj α j Pjg (t) + ω j (t) + z j (t) + λ j (t)
g
Pj
+ ω j (t)/R j ,
j∈N
(6e)
iPlj
h
ulj (t) = Pjl (t) − γ lj β j Pjl (t) − ω j (t) − z j (t) − λ j (t)
l
Pj
j∈N
(6f)
where γ λj , γiηj , γi j , γ gj , γ lj are positive constants. For any xi , ai ∈
R, the operator [xi ]+
ai is defined by
φ̃
[xi ]+
ai
:=
xi if ai > 0 or xi > 0;
0, otherwise.
For a vector case, [x]+
a is defined accordingly componentwise [9].
4
Here we assume that each node i updates a set of internal
1
states (λi (t), ηi+j (t), η −
ji (t), φ̃i j (t)) according to (6a)–(6d). In
contrast to the completely decentralized control derived in the
per-node balance case, here the control is distributed where
each node i updates (λi (t), ηi+j (t), ηi−j (t)) using only local
measurements or computation but requires the information
(λ j (t), z j (t)) from its neighbors j to update φ̃i j (t). Note that
zi (t) is not a variable but a shorthand for (function) Pjg (t) −
Pjl (t) − p j − Û j (φ̃ (t)). The control inputs (ugi (t), uli (t)) in
(6e)(6f) are functions of the network state (Pig (t), Pil (t), ωi (t))
and the internal state (λi (t), ηi+j (t), ηi−j (t), φ̃i j (t)). We write ugj
and ulj as functions of (Pjg , Pjl , ω j , λ j ): for j ∈ N
(7a)
ugj (t) := ugj Pjg (t), ω j (t), λ j (t), z j (t)
ulj (t) := ulj Pjl (t), ω j (t), λ j (t), z j (t)
(7b)
where the functions are defined by the right-hand sides of
(6e)(6f).
Now we comment on the implementation of the control (6).
Remark 3 (Implementation).
1) As discussed above,
communication is needed only between neighboring
nodes (areas) to update the variables φ̃i j (t).
2) Similar to the per-node power balance case, we can
avoid measuring the load change p j by using (1b) and
the definition of z j (t) to replace (6a) with
z j (t) = M j ω̇ j + D j ω j (t) −
∑ Pi j (t) + ∑
i:i→ j
Pjk (t)
k: j→k
− Û j (φ̃ (t))
= γ λj z j (t)
λ̇ j
C. Design rationale
The controller design (6) is also motivated by a (partial)
primal-dual algorithm for (5), as for the per-node power
balance case.
Primal-dual algorithms. The optimization problem in the
network balance case differs from that in the per-node balance case in the inequalities (5d) on φ̃ . Consider a general
constrained convex optimization with inequality constraints:
min f (x) s.t. g(x) = 0, h(x) ≤ 0
where Γx , Γρ1 , Γρ2 are strictly positive diagonal gain matrices.
Here, if a is a scalar then (a)+ := max{a, 0} and if a is a
vector then (a)+ is defined accordingly componentwise. For a
dual algorithm, (8a) is replaced by
x(t) := min L(x; ρ(t))
x∈X
(8d)
As for the per-node balance case, all variables in x(t) are
updated according to (8a) except ω(t) which is updated
according to (8d), as we see below.
The set X in (8a) is defined by the constraints (2):
n
o
g l
(9)
X := (Pg , Pl ) : (Pg , Pl ) ≤ (Pg , Pl ) ≤ (P , P )
Controller (6) design. Let ρ1 := (λ , µ) be the Lagrange multipliers associated with constraints (5c) and (5b) respectively,
ρ2 := (η + , η − ) the multipliers associated with constraints
(5d), and ρ := (ρ1 , ρ2 ). Define the Lagrangian of (5) by (10).
Note that it is only a function of (x, ρ) and independent of
u := (ug , ul ) as we treat u as a function of (x, ρ) defined by
the right-hand sides of (6e)(6f).
The closed-loop dynamics (1)(6) carry out an approximate
primal-dual algorithm (8) for solving (5) in real time over the
coupled physical power network and cyber computation. Since
the reasoning is similar to the per-node balance case, we only
provide a summary. Rewrite the Lagrangian L2 in vector form
1 g T g g
L2 (x; ρ) =
(P ) A P + (Pl )T Al Pl + ω T Dω + zT z
2
+ λ T Pg − Pl − p −CBφ̃
+ µ T Pg − Pl − p − Dω −CBθ̃
+ (η + )T φ̃ − θ + (η − )T θ − φ̃
(11)
where Al := diag(β j , j ∈ N), B := diag(Bi j , (i, j) ∈ E).
First, the control (6b)(6c) can be interpreted as a continuoustime version of the dual update (8c) on the dual variable ρ2 :=
(η + (t), η − (t)):
+
η̇ + = Γη ∇η + L2 (x(t); ρ(t)) η + (t)
+
−
η̇ − = Γη ∇η − L2 (x(t); ρ(t)) η −
x∈X
Rn
where f :
→ R, g : Rn → Rk1 , h : Rn → Rk2 , and X ⊆ Rn
is closed and convex. Here an inequality constraint h(x) ≤ 0
is imposed explicitly. Let ρ1 ∈ Rk1 be the Lagrange multiplier
associated with the equality constraint g(x) = 0, ρ2 ∈ Rk2 that
associated with the inequality constraint h(x) ≤ 0, and ρ :=
(ρ1 , ρ2 ). Define the Lagrangian L(x; ρ) := f (x) + ρ1T g(x) +
ρ2T h(x). A standard primal-dual algorithm takes the form:
x(t + 1) := ProjX (x(t) − Γx ∇x L(x(t); ρ(t))) (8a)
ρ1 (t + 1) := ρ1 (t) + Γρ1 ∇ρ1 L(x(t); ρ(t))
ρ2 (t + 1) :=
1 For
+
ρ2 (t) + Γ ∇ρ2 L(x(t); ρ(t))
ρ2
(8b)
where Γη := diag(γiηj , (i, j) ∈ E).
Second, the control (6a) carries out the dual update (8b) on
λ (t):
λ̇
= Γλ ∇λ L2 (x(t), ρ(t))
(12a)
where Γλ := diag(γ λj , j ∈ N). The swing dynamic (1b) carries
out the dual update (8b) on µ(t) because, as in the per-node
balance case, we can identify µ(t) ≡ ω(t) so that
µ̇
= ω̇
= M −1 ∇µ L2 (x(t); ρ(t))
(12b)
(8c)
each (directed) link (i, j) ∈ E we assume that only node i maintains
the variables (ηi+j (t), η −
ji (t), φ̃i j (t)). In practice, node j will probably maintain
symmetric variables to reduce communication burden or for other reasons
outside our mathematical model here.
where M := diag(M j , j ∈ N).
Finally we show that (1a), (1c), (1d), and (6d) implement
a mix of the primal updates (8a) and (8d) on the primal variables x := (θ̃ (t); φ̃ (t); ω(t); Pg (t); Pl (t)). Setting ω(t) ≡ µ(t)
5
2
∑ α j Pjg +
1
L2 (x; ρ) =
2
j∈N
2
∑ β j Pjl +
j∈N
!
∑ D j ω 2j + ∑ z2j
j∈N
!
+
j
∑ µj
Pjg − Pjl
− p j − D jω j +
j∈N
∑
Bi j θi j −
i:i→ j
∑
B jk θ jk
k: j→k
!
+
∑ λj
Pjg − Pjl
− pj +
j∈N
∑
Bi j φi j −
i:i→ j
∑
B jk φ jk
k: j→k
∑
ηi−j θ i j − φi j (t) +
g
(12c)
ηi+j φi j (t) − θ i j
∑
(10)
(i, j)∈E
(i, j)∈E
is equivalent to the primal update (8d) on ω(t), as in the pernode balance case. Moreover the control laws (6e)(6f) are then
equivalent to
T g Ṗg = [Pg (t) − Γg ∇Pg L2 (x(t), ρ(t))]PPg − Pg (t)
h
iPl
T l Ṗl = Pl (t) − Γl ∇Pl L2 (x(t), ρ(t)) l − Pl (t)
+
2) The nominal frequency is restored, i.e., ω ∗j = 0 for all
j ∈ N; moreover φ̃i∗j = θ̃i∗j for all (i, j) ∈ E.
3) The network balance condition (4) is satisfied by x∗ .
4) The line limits (3) are satisfied by x∗ , implying Pi j ≤
Pi j ≤ Pi j on every tie line (i, j) ∈ E.
where Γφ̃ := diag(γi j , (i, j) ∈ E) .
Theorem 2 shows that the equilibrium point has a simple
yet intuitive structure. Moreover, Theorem 2 implies that the
closed-loop system can autonomously eliminate congestions
on tie lines. This feature has important implications. It means
our distributed frequency control is capable of serving as a
corrective re-dispatch without the coordination of dispatch
centers if a congestion arises. This can enlarge the feasible
region for economic dispatch, since corrective re-dispatch has
been naturally taken into account.
The proofs of Theorem 1 and 2 are given in Appendix A.
D. Optimality of equilibrium point
E. Asymptotic stability
P
(12d)
i.e., the generator and controllable load at each node j carry
out the primal update (8a). For (θ̃ , φ̃ ), (1a) and (6d) are
equivalent to the primal update (8a):
θ̃˙
φ̃˙
= −B−1 ∇θ̃ L2 (x(t), ρ(t))
φ̃
= −Γ ∇φ̃ L2 (x(t), ρ(t))
(12e)
(12f)
φ̃
In this subsection, we address the optimality of the equilibrium point of the closed-loop system (1)(6). Given an
(x, ρ) := (θ̃ , φ̃ , ω, Pg , Pl ), (λ , µ), (η − , η + ) , recall that the
control input u(x, ρ1 , ρ2 ) is given by (7).
Definition 1. A point (x∗ , ρ ∗ ) := (θ̃ ∗ , φ̃ ∗ , ω ∗ , Pg∗ , Pl∗ , λ ∗ , η +∗ ,
η −∗ , µ ∗ ) is an equilibrium point or an equilibrium of the
closed-loop system (1)(6) if
1) The right-hand side of (1) vanishes at x∗ and u(x∗ , ρ ∗ ).
2) The right-hand side of (6a)–(6d) vanishes at (x∗ , ρ ∗ ).
Definition 2. A point (x∗ , ρ ∗ ) is primal-dual optimal if
(x∗ , u(x∗ , ρ ∗ )) is optimal for (5) and ρ ∗ is optimal for its dual
problem.
We make the following assumption:
A2: The problem (5) is feasible.
The following theorem characterizes the correspondence
between the equilibrium of the closed-loop system (1)(6) and
the primal-dual optimal solution of (5).
Theorem 1. Suppose A2 holds. A point (x∗ , ρ ∗ ) is primal-dual
optimal if and only if (x∗ , ρ ∗ ) is an equilibrium of closed-loop
system (1)(6) satisfying (2) and µ ∗ = 0.
Next result says that, at equilibrium, the network balance
condition (4) and line limits (3) are satisfied and the nominal
frequency is restored. Moreover the equilibrium is unique.
(x∗ , ρ ∗ )
Theorem 2. Suppose A1 and A2 hold. Let
be primaldual optimal. Then
1) The equilibrium (x∗ , µ ∗ ) is unique, with (θ ∗ , φ ∗ ) being
unique up to (equilibrium) reference angles (θ0 , φ0 ).
In this subsection, we address the asymptotic stability of the
closed-loop system (1)(6), under an additional assumption:
A3: The initial state of the closed-loop system (1)(6) is finite,
and pgj (0), plj (0) satisfy constraint (2).
As in the per-node balance case the closed-loop system
(1)(6) satisfies constraint (2) even during transient.
Lemma 3. Suppose A1 and A3 hold. Then constraint (2) is
satisfied for all t > 0, i.e. (Pg (t), Pl (t)) ∈ X for all t ≥ 0 where
X is defined in (9).
The proof is exactly the same as that for Lemma 3 in [1]
and omitted.
Similar to the per-node balance case, we first rewrite the
closed-loop system using states θ̃ , φ̃ instead of θ , φ (they are
equivalent under assumption A1). Setting µ ≡ ω, the closedloop system (1)(6) is equivalent to (in vector form):
θ̃˙ (t) = CT ω(t)
(13a)
−1
g
l
ω̇(t) = M
P (t) − P (t) − p − Dω(t) −CBθ̃ (t)
(13b)
g
g −1
l
l −1
Ṗ (t) = (T )
Ṗ (t) = (T )
g
g
(−P (t) + û (t))
−Pl (t) + ûl (t)
η̇ + (t) = Γη [φ̃ (t) − θ ]+
η+
−
η
− φ̃ (t)]+
η−
η̇ (t) = Γ [θ
λ̇ (t) = Γλ Pg (t) − Pl (t) − p −CBφ̃ (t)
φ̃˙ (t) = Γφ̃
(13c)
(13d)
(13e)
(13f)
(13g)
BCT λ (t) + BCT z(t) + η − (t) − η + (t) (13h)
6
where z(t) := Pg (t) − Pl (t) − p −CBφ̃ (t) and
iPg
h
ûg (t) = Pg (t) − Γg Ag Pjg (t) + ω(t) + z(t) + λ (t)
g
P
h
iPl
ûlj (t) = Pl (t) − Γl Al Pjl (t) − ω(t) − z(t) − λ (t)
l
P
:= (θ̃ , ω, Pg , Pl , λ , η + , η − , φ̃ ).
Denote w
Note that the right-hand sides of (13e)(13f) are discontinuous due to projection to the nonnegative quadrant for
(η + (t), η − (t)). The system (13) is called a projected dynamical system and we adopt the concept of Caratheodory
solutions for such a system where a trajectory (w(t),t ≥ 0)
is called a Caratheodory solution, or just a solution, to (13)
if it is absolutely continuous in t and satisfies (13) almost
everywhere. The result in [10, Theorems 2 and 3] implies that,
given any initial state, there exists a unique solution trajectory
to the closed-loop system (13) as the unprojected system is
Lipschitz and the nonnegative quadrant is closed and convex.
See [11, Theorem 3.1] for extension of this result to the Hilbert
space.
With regard to system (13), we first define two sets, σ + and
−
σ , as follows [8].
σ+
:= {(i, j) ∈ E | ηi+j = 0, φ̃i j − θ i j < 0}
σ−
:= {(i, j) ∈ E | ηi−j = 0, θ i j − φ̃i j < 0}
Then (6b) and (6c) are equivalent to
η
γi j (φ̃i j − θ i j ),
η̇i+j =
0,
η
γi j (θ i j − φ̃i j ),
η̇i−j =
0,
if(i, j) ∈
/ σ +;
if(i, j) ∈ σ + .
(14a)
if(i, j) ∈
/ σ −;
if(i, j) ∈ σ − .
(14b)
In a fixed σ + , σ − , define F(w).
−B1/2CT ω
−M −1/2 Pg − Pl − p − Dω −CBθ̃
(T g )−1 (Ag Pg + ω + z + λ )
(T l )−1 Al Pl − ω − z − λ
F(w) = −(Γη )1/2 [φ̃ − θ ]+
η+
+
η
1/2
−(Γ ) [θ − φ̃ ]η −
−(Γλ )1/2 Pg − Pl − p −CBφ̃
−(Γφ̃ )1/2 BCT λ + BCT z + η − − η +
(15)
If σ + and σ − do not change, F(w) is continuously differentiable in w.
Similarly, we define S := Rm+n+1 × X × R2m+n+1+m , where
the closed convex set X is defined in (9). Then for any w we
define the projection of w − F(w) onto S as
H(w) := ProjS (w − F(w)) := arg min ky − (w − F(w))k2
y∈S
Then the closed-loop system (13) is equivalent to
ẇ(t) = Γ2 (H(w(t)) − w(t))
where the positive definite gain matrix is :
Γ2 = diag B−1/2 , M −1/2 , (T g )−1 , (T l )−1 ,
λ 1/2
η 1/2
η 1/2
φ̃ 1/2
(Γ ) , (Γ ) , (Γ ) , (Γ )
(16)
Note that the projection operation H has an effect only on
(Pg ; Pl ) and Lemma 3 indicates that w(t) ∈ S for all t > 0,
justifying the equivalence of (13) and (16).
A point w∗ ∈ S is an equilibrium of the closed-loop system
(16) if and only if it is a fixed point of the projection H(w∗ ) =
w∗ . Let E2 := { w ∈ S | H(w(t)) − w(t) = 0 } be the set of
equilibrium points. Then we have the following theorem.
Theorem 4. Suppose A1, A2 and A3 hold. Starting from any
initial point w(0), w(t) remains in a bounded set for all t and
w(t) → w∗ as t → ∞ for some equilibrium w∗ ∈ E2 that is
optimal for problem (5).
For any equilibrium point w∗ , we define the following
function taking the same form as the per-node case.
1
Ṽ2 (w) = −(H(w) − w)T F(w) − ||H(w) − w||22
2
1
∗ T −2
+ k(w − w ) Γ2 (w − w∗ )
2
(17)
where k is small enough such that Γ2 − kΓ−1
2 > 0 is strictly
positive definite.
For any fixed σ + and σ − , Ṽ2 is continuously differentiable
as F(w) is continuously differentiable in this situation. Similar
to V1 (w) used in Part I of the paper, we know Ṽ2 (w) ≥ 0 on
S and Ṽ2 (w) = 0 holds only at any equilibrium w∗ = H(w∗ )
[12]. Moreover, Ṽ2 is nonincreasing for fixed σ + and σ − , as
we prove in Appendix B.
It is worth to note that the index sets σ + and σ − may change
sometimes, resulting in discontinuity of Ṽ2 (w). To circumvent
such an issue, we slightly modify the definition of V2 (w) at
the discontinuous points as:
1) V2 (w) := Ṽ2 (w), if Ṽ2 (w) is continuous at w;
2) V2 (w) := lim sup Ṽ2 (v), if Ṽ2 (w) is discontinuous at w.
v→w
Then V2 (w) is upper semi-continuous in w, and V2 (w) ≥ 0 on
S and V2 (w) = 0 holds only at any equilibrium w∗ = H(w∗ ).
As V2 (w) is not differentiable for w at discontinuous points,
we use the Clarke gradient as the gradient at these points [13].
Note that Ṽ2 is continuous almost everywhere except the
switching points. Hence both V2 (w) is nonpathological [14],
[15]. With these definitions and notations above, we can prove
Theorem 4. The detail of proof is provided in Appendix B.
IV. C ASE STUDIES
A. System configuration
A four-area system based on Kundur’s four-machine, twoarea system [16] [17] is used to test our optimal frequency
controller. There are one (aggregate) generator (Gen1∼Gen4),
one controllable (aggregate) load (L1c∼L4c) and one uncontrollable (aggregate) load (L1∼L4) in each area, which is
shown in Fig.1. The parameters of generators and controllable
loads are given in Table I. The total uncontrollable load in
each area are identically 480MW. At time t = 10s, we add
step changes on the uncontrollable loads in four areas to test
the performance of our controllers.
All the simulations are implemented in PSCAD [18] with
8GB memory and 2.39 GHz CPU. We use the detailed
electromagnetic transient model of three-phase synchronous
Bus7
Bus5
Bus8
Bus9
Bus10 Bus11 Bus3
T3
machines to simulate generators with both governors
and
exciters.
The uncontrollable load L1-L4 are modelled
Gen1
Gen3 by
the fixed load
in Bus6
PSCAD, while controllable load
L1c-L4c
T2
T4
Load1
Load2
are formulated
by the
self-defined controlled
current
Bus2
Bus4 source.
Gen4 We need
The closed-loop Gen2
system diagram is shown in Fig.2.
measure loacal frequency, generation, controllable load and
tie-line power flows to compute control demands. Only φ̃i j
are exchanged between neighbors. All variables are added by
their initial steady state values to explicitly show the actual
values.
T1
Bus1
Bus5
Bus7
Bus6
Bus8
Bus9
Bus10 Bus11 Bus2
T1
T2
Gen1
1) Stability and optimality: The dynamics of local frequencies and tie-line power flows are illustrated in Fig.3.
The frequencies are well restored in all four control areas
while the tie line powers are remained within their acceptable
ranges. The generations and controllable loads are different
from that before disturbance, indicating that the system is
stabilized at a new steady state. The resulting equilibrium
point is given in Table IV, which is identical to the optimal
solution of (5) computed by centralized optimization using
CVX. These simulation results confirm that our controller
can autonomously guarantee the frequency stability while
achieving optimal operating point in the overall system.
Gen2
L1c
L2
L1
Area2
L2c
Bus15 Bus16 Bus17 Bus4
Bus3 Bus14 Bus13 Bus12
T4
T3
Gen4
Gen3
Area3
L3c
L3
Area4
L4c
L4
60
50.2
Frequency (Hz)
Area1
80
50.3
Area 1
Area 2
Area 3
Area 4
50.1
50
49.9
49.8
49.7
0
Fig. 1: Four-area power system
Tie line power (MW)
Bus1
7
40
20
0
Line 21
Line 31
Line 32
Line 42
-20
-40
-60
20
40
60
80
100
120 140
160
-80
0
20
40
Time (s)
ij
Controller
Controller
u lj
u gj
uig
TABLE I:
S YSTEMPARAMETERS
, P g , Pl , P , P
, P , P , Pij , Pjk
u
l
j
i
Area j
1
2
3
4
Controller
g
j
u lj
Dj
0.04
0.045
0.05
0.055
Rj
0.04
0.06
0.05
0.045
i
i
ij
100
120 140
160
Pjg∗ (MW)
Pjl∗ (MW)
Area 1
620
23.6
Area 2
596
59.8
Area 3
660
23.6
Area 4
580
39.7
Tie line
Pi∗j (MW)
(2,1)
-39.94
(3,1)
13.35
(3,2)
53.27
(4,2)
-59.6
2) Dynamic performance: In this subsection,
j , Pjgwe
, Pjlanalyze
, Pij , Pjk
the impacts of operational (capacity and line power) constraints on the dynamic property. Similarly, we compare the
responses of frequency controllers with and without
considerController
ing input saturations. The trajectories of mechanical power of
g
l
turbines and controllable loads are shown in uFig.4
and Fig.5, u j
j
respectively. In this case, the system frequency is restored, and
the same optimal equilibrium point is achieved.
ki
T jg
αj
βj
T jl
2
2.5
4
4
2.5
4
6
5
Controller
1.5
2.5
5
4
3
5.5 l 5
g 3
ui
80
Time (s)
TABLE III:
E QUILIBRIUM P OINTS
uil
Fig. 2: Closed-loop system diagram
g
j
60
Fig. 3: Dynamics of frequency (left) and tie-line flows (right)
in network balance case
i , Pi g , Pi l , Pij , Pki
j , Pjg , Pjl , Pij , Pjk
ui
B. Network Power Balance Case
In this case, the generations in each area are initiated as
(560.9, 548.7, 581.2, 540.6) MW and the controllable loads
(70.8, 89.6, 71.3, 79.4) MW. The load changes are identical to
those in Table II, which are also unknown to the controllers.
We use method in Remark 4 to estimate the load changes.
Operational constraints on generations, controllable loads and
tie lines are shown in Table II.
900
Mechanical pwoer (MW)
j
DFIG
Bus12
Tw
850
Area 1
Area 2
Area 3
Area 4
800
750
Area 1
Area 2
Area 3
Area 4
700
650
600
550
500
TABLE II:
450
0
C APACITY LIMITS IN NETWORK CASE
Area 2
[530, 680]
Area 3
[550, 700]
Area 4
[530, 670]
[Plj ,P j ] (MW)
[20, 80]
[60, 100]
[20, 80]
[35, 80]
Tie line
[Pi j ,Pi j ] (MW)
(2,1)
[-65, 65]
(3,1)
[-65, 65]
(3,2)
[-65, 65]
(4,2)
[-65, 65]
l
40
60
80
Time (s)
Area 1
[550, 710]
g
[Pgj ,P j ] (MW)
20
100
120 140 160 0
20
40
60
80
100
120 140 160
Time (s)
Fig. 4: Mechanical outputs with(left)/without(right) capacity
constraints
3) Congestion analysis: In this scenario, we reduce tieline power constraints to Pi j = −Pi j = 50MW, which causes
congestions in tie-line (2,3) and (2,4). The steady states
8
120
Area 1
Area 2
Area 3
Area 4
Pl (MW)
100
Area 1
Area 2
Area 3
Area 4
80
60
40
20
0
0
20
40
60
80
100
120
140
160 0
20
40
Time (s)
60
80
100
120
140
160
Time (s)
Fig. 5: Controllable loads with(left)/without(right) capacity
constraints
under the distributed control are listed in Table IV. Note that
∗ +
Pjg∗ − p j + Pjl∗ − ∑ Pjk
∑ Pi∗j = 0 hold in each area.
i:i→ j
k: j→k
The dynamics of tie-line powers in two different scenarios
shown in Fig.6 indicate that (2,4) reaches the limit in steady
state. However, by adopting the proposed fully distributed
optimal frequency control, the congestion is eliminated and all
the tie line powers remain within the limits. Thus, congestion
control is achieved optimally in a distributed manner.
TABLE IV:
S IMULATION R ESULTS WITH C ONGESTION
Pjg∗ (MW)
Pjl∗ (MW)
Area 1
618
25.1
Area 2
595
60.7
Area 3
658
25.1
Area 4
585
34.9
Tie line
Pi∗j (MW)
(2,1)
-36.4
(3,1)
13.1
(3,2)
49.5
(4,2)
-49.9
80
Tie line power (MW)
60
40
20
Line 21
Line 31
Line 32
Line 42
0
-20
Line 21
Line 31
Line 32
Line 42
-40
-60
-80
0
20
40
60
80
Time (s)
100
120 140
160 0
20
40
60
80
100
120 140
160
Time (s)
Fig. 6: Tie line power with(left)/without(right) capacity constraints
V. C ONCLUDING R EMARKS
In this paper, we have devised a distributed optimal frequency control in the network balance case, which can autonomously restore the nominal frequencies after unknown
load disturbances while minimizing the regulation costs. The
capacity constraints on the generations and controllable loads
can also be satisfied even during transient. In addition, congestions can be eliminated automatically, implying tie-line powers
can be remained within given ranges. Only neighborhood
communication is required in this case. Like the per-node
case, here the closed-loop system again carries out a primaldual algorithm with saturation to solve the associated optimal
problem. To cope with the discontinuity introduced due to
enforcing different types of constraints, we have constructed
a nonpathological Lyapunov function to prove the asymptotically stability of the closed-loop systems. Simulations on a
modified Kundur’s power system validate the effectiveness of
our controller.
This approach is also applicable to other problem involving
frequency regulation, e.g. standalone microgrid or demand
side management. We highlight two crucial implications of
our work: First, our distributed frequency control is capable
of serving as an automatically corrective re-dispatch without
the coordination of dispatch center when certain congestion
happens; Second, the feasible region of economic decisions
can be enlarged benefiting from the corrective re-dispatch.
In this sense, our work may provide a systematic way to
bridge the gap between the (secondary) frequency control in a
fast timescale and the economic dispatch in a slow timescale,
hence breaking the traditional hierarchy of the power system
frequency control and economic dispatch.
R EFERENCES
[1] Z. Wang, F. Liu, S. H. Low, C. Zhao, and S. Mei, “Distributed frequency
control with operational constraints, part i: Per-node power balance,”
https://arxiv.org/abs/1702.07965, 2017.
[2] E. Mallada, C. Zhao, and S. H. Low, “Optimal load-side control for
frequency regulation in smart grids,” IEEE Trans. Autom. Control, to
appear, 2017.
[3] T. Stegink, C. D. Persis, and A. van der Schaft, “A unifying energybased approach to stability of power grids with market dynamics,” arXiv
preprint arXiv:1604.05200, 2016.
[4] T. Stegink, C. D. Persis, and A. van der Schaft, “A port-hamiltonian
approach to optimal frequency regulation in power grids,” in Proc. 54th
IEEE Conference on Decision and Control (CDC), Dec 2015, pp. 3224–
3229.
[5] P. Yi, Y. Hong, and F. Liu, “Distributed gradient algorithm for constrained optimization with application to load sharing in power systems,”
Syst. Control Lett., vol. 83, pp. 45–52, Sept 2015.
[6] C. Zhao, E. Mallada, S. Low, and J. Bialek, “A unified framework
for frequency control and congestion management,” in Power Systems
Computation Conference (PSCC), 2016. IEEE, 2016, pp. 1–7.
[7] X. Zhang and A. Papachristodoulou, “A real-time control framework for
smart power networks: Design methodology and stability,” Automatica,
vol. 58, pp. 43–50, 2015.
[8] D. Feijer and F. Paganini., “Stability of primal-dual gradient dynamics
and applications to network optimization,” Automatica, vol. 46, no. 12,
pp. 1974–1981, Dec 2010.
[9] A. Cherukuri, E. Mallada, and J. Corts, “Asymptotic convergence of
constrained primaldual dynamics,” Syst. Control Lett., vol. 87, pp. 10 –
15, 2016.
[10] P. Dupuis and A. Nagurney, “Dynamical systems and variational inequalities,” Annals of Operations Research, vol. 44, pp. 9–42, 1993.
[11] M.-G. Cojocaru and L. B. Jonker, “Existence of solutions to projected
differential equations in hilbert spaces,” Proc. Amer. Math. Soc., vol.
132, pp. 183–193, 2004.
[12] M. Fukushima, “Equivalent differentiable optimization problems and
descent methods for asymmetric variational inequality problems,” Math.
programming, vol. 53, no. 1-3, pp. 99–110, Jan 1992.
[13] F. H. Clarke, Optimization and nonsmooth analysis. Siam, 1990, vol. 5.
[14] A. Bacciotti and F. Ceragioli, “Nonpathological lyapunov functions and
discontinuous carathodory systems,” Automatica, vol. 42, no. 3, pp. 453
– 458, 2006.
[15] A. Bacciotti and F. Ceragioli, “Stability and stabilization of discontinuous systems and nonsmooth lyapunov functions,” ESAIM: Control,
Optimisation and Calculus of Variations, vol. 4, pp. 361–376, 1999.
[16] J. Fang, W. Yao, Z. Chen, J. Wen, and S. Cheng, “Design of anti-windup
compensator for energy storage-based damping controller to enhance
power system stability,” IEEE Trans. Power Syst., vol. 29, no. 3, pp.
1175–1185, May 2014.
[17] P. Kundur, Power System Stability and Control. McGraw-hill New
York, 1994, vol. 7.
[18] https://hvdc.ca/pscad/.
9
Since (x∗ , ρ ∗ ) is a saddle point, we must have
∇φ̃ L2 (x∗ , ρ ∗ ) = 0, implying φ̃˙ = 0.
A PPENDIX A
P ROOFS OF T HEOREM 1 AND T HEOREM 2
We start with a lemma.
Lemma A.1. Suppose (x∗ , u∗ ) is optimal for (5). Then
1) ω ∗ = 0, i.e., the nominal frequency is restored;
2) the network balance condition (4) is satisfied by x∗ ;
3) φ ∗ − θ ∗ = (φ0∗ − θ0∗ ) 1;
4) θ i j ≤ θi∗j ≤ θ i j , i.e., the line limits (3) are satisfied.
Proof. Suppose (x∗ , u∗ ) is optimal but ω ∗ 6= 0. Then (5c)
implies
Pg∗ − Pl∗ − p
= CBCT φ ∗
(A.1)
Consider x̂ := (θ̂ , φ ∗ , ω̂, Pg∗ , Pl∗ ) with θ̂ := φ ∗ and ω̂ := 0.
Then x̂ satisfies (5b)(5c) due to (A.1). Hence (x̂, u∗ ) is feasible
for (5) but has a strictly lower cost, contradicting the optimality
of (x∗ , u∗ ). Hence ω ∗ = 0.
Multiplier both sides of (A.1) by 1T yields the network
balance condition (4), proving 2).
To prove 3), setting ω ∗ = 0 in (5b) and combining with (5c)
yield
CBCT θ ∗
= Pg∗ − Pl∗ − p = CBCT φ ∗
Since CBCT is an (n + 1) × (n + 1) matrix with rank n, its null
space has dimension 1 and is spanned by 1 because CT 1 = 0.
Hence CBCT (φ ∗ − θ ∗ ) = 0 implies that φ ∗ − θ ∗ = (φ0∗ − θ0∗ ) 1.
To prove 4), note that φ̃ = CT φ and θ̃ = CT θ and hence
φ̃ ∗ − θ̃ ∗
We have the following result.
Lemma A.2. Suppose (x∗ , ρ ∗ ) is primal-dual optimal. Then
h
iPgj
g∗
g∗
g
g∗
∗
∗
∗
ug∗
=
P
=
P
−
γ
α
P
+
ω
+
z
+
λ
j j
j
j
j
j
j
j
j
g
Pj
=
⇐: Suppose now (x∗ , ρ ∗ ) is an equilibrium of the closed-loop
system (1)(6) that satisfies (2) and µ ∗ = 0. Since (5) is convex
with linear constraints, (x∗ , ρ ∗ ) is a primal-dual optimal if and
only if (x∗ , u(x∗ , ρ ∗ )) is primal feasible and satisfies
x∗ = arg min{L2 (x; ρ ∗ )|(x, u(x, ρ ∗ )) satisfies(2), (5e), (5f)}
x
(A.2)
This is because ∇ρ1 L2 (x∗ , ρ ∗ ) = 0 since µ̇ = λ̇ = 0, η +∗ ≥
0, η −∗ ≥ 0, and the complementary slackness condition
−∗
∗
+
∗
ηi+∗
j (φ̃i j − θ i j ) = 0, ηi j (θ i j − φ̃i j ) = 0 is satisfied since η̇ =
−
η̇ = 0.
To show that (x∗ , u(x∗ , ρ ∗ )) is primal feasible, note that
since (x∗ , u(x∗ , ρ ∗ )) is an equilibrium of (1), it satisfies ω ∗ = 0
and hence (5e)(5f), in addition to (2). Moreover ω̇ = 0 means
(x∗ , u(x∗ , ρ ∗ )) satisfies (5b), λ̇ = 0 implies (5c), η̇ + = η̇ − = 0
implies (5d).
To show that (x∗ , ρ ∗ ) satisfies (A.2), note that (5e)(5f) and
(6e)(6f) imply that
Pjl∗
h
iPlj
= Pjl∗ − γ lj β j Pjl∗ − ω ∗j − z∗j − λ j∗
l
Pj
for any γ gj > 0 and γ lj > 0.
Lemma A.2 shows that the saturation of control input does
not impact the optimal solution of optimization problem (5).
With Lemma 3, Lemma A.1 and Lemma A.2, we now can
prove Theorem 1 and Theorem 2.
Proof of Theorem 1. ⇒: Suppose (x∗ , ρ ∗ ) is primal-dual optimal. Then x∗ satisfies the operational constraints (2). Moreover
the right-hand side of (1) vanishes because:
• θ̇ = 0 since ω ∗ = 0 from Lemma A.1.
• ω̇ = 0 since constraint (5b) holds for x∗ .
• Ṗg = Ṗl = 0 since ω ∗ = 0 and x∗ satisfies (5e) and (5f).
• λ̇ = 0 since (5c) holds for x∗ .
• η̇ + = η̇ − = 0 since (5d) holds for x∗ .
• From (12f) we have
φ̃˙
= −Γφ̃ ∇φ̃ L2 (x∗ , ρ ∗ )
h
iPgj
Pjg∗ − γ gj α j Pjg∗ + ω ∗j + z∗j + λ j∗
g
Pjg∗
=
Pjl∗
h
iPlj
= Pjl∗ − γ lj β j Pjl∗ − ω ∗j − z∗j − λ j∗
l
= CT (φ ∗ − θ ∗ ) = (φ0∗ − θ0∗ )CT 1 = 0
i.e. φ̃ ∗ = θ̃ ∗ . We conclude from (5d) that θ i j ≤ θi∗j ≤ θ i j . This
completes the proof.
ul∗
j
Hence (x∗ , ρ ∗ ) is an equilibrium of the closed-loop system
(1)(6) that satisfies the operational constraints (2). Moreover
µ ∗ = ω ∗ = 0 since ∂∂ ωL2j (x∗ , ρ ∗ ) = D j (ω ∗j − µ ∗j ) = 0 and D j > 0
for all j ∈ N.
Pj
Pj
The rest of the proof follows the same line of argument as that
in Theorem 1 in [1]. This proves that (x∗ , ρ ∗ ) is primal-dual
optimal and completes the proof of Theorem 1.
Next we prove Theorem 2.
Proof of Theorem 2. Let (x∗ , ρ ∗ ) = (θ̃ ∗ , φ̃ ∗ , ω ∗ , Pg∗ , Pl∗ ,
λ ∗ ,η +∗ , η −∗ , µ ∗ ) be primal-dual optimal. For the uniqueness
of x∗ , (ω ∗ , Pg∗ , Pl∗ ) are unique because the objective function
in (5) is strictly convex in (ω, Pg , Pl ). Hence µ ∗ = ω ∗ is unique
as well. Assumption A1 that φ0∗ := 0 and (A.1) imply that φ ∗
is uniquely determined by the equilibrium (ω ∗ , Pg∗ , Pl∗ ). Since
θ ∗ −φ ∗ = (θ0∗ −φ0∗ )1, assumption A1 that θ0∗ := 0 then implies
that θ ∗ is unique. This proves the uniqueness of (x∗ , µ ∗ ).
The remaining three parts of the theorem follow from
Lemma A.1.
A PPENDIX B
P ROOF T HEOREM 4
We start with a lemma.
Lemma B.1. Suppose A1, A4 and A5 hold. Then
1) V̇2 (w(t)) ≤ 0 in a fixed σ + and σ − .
2) The trajectory w(t) is bounded, i.e., there exists w such
that kw(t)k ≤ w for all t ≥ 0.
10
Q=
−BCT
D
I|N|
−I|N|
0
0
0
0
0
CB
0
0
0
0
0
0
0
−I|N|
Ag + I|N|
−I|N|
−I|N|
0
0
−BCT
0
I|N|
−I|N|
Al + I|N|
I|N|
0
0
BCT
0
0
0
0
0
0
0
−BCT
Proof of Lemma B.1 . Given fixed σ + , σ − , for all (i, j) ∈
/ σ +,
−
(i, j) ∈
/ σ , we have
V̇2 (w) ≤ k(H(w) − w∗ )T · Γ−1
2 (H(w) − (w − F(w)))
T
− (H(w) − w) Γ2 · Q · Γ2 (H(w) − w)
T
− (H(w) − (w − F(w)))
(B.1a)
(B.1b)
(Γ2 − kΓ−1
2 )(w − H(w))
(B.1c)
− k(w − w∗ )T · Γ−1
2 F(w)
(B.1d)
where Q is a semi-definite positive matrix and Γ2 Q = ∇w F(w),
which given in (B.2). Here, the subscript of I means its
dimension, and |A| means the cardinality of set A.
Given fixed σ + and σ − , F(w) is continuous differentiable.
In this case, (B.1a) and (B.1c) are nonpositive due to as
discussed in the proof of Theorem 4 in [1]. (B.1b) is also
nonpositive as Q is semi-definite positive (see Eq. (B.2)).
Next, we prove that (B.1d) is nonpositive. Similar to the
per-node case, substituting µ(t) ≡ ω(t) into the Lagrangian
L2 (x, ρ) in (11) we obtain a function L̂2 (w) defined as follows.
L̂2 (w) := L2 θ̃ , ω, Pg , Pl , λ , µ, η + , η − , φ̃
µ=ω
1 g T g g
l T l l
T
=
(P ) A P + (P ) A P − ω Dω + zT z
2
+ λ T Pg − Pl − p −CBφ̃ + ω T Pg − Pl − p −CBθ̃
+ (η − )T θ − φ̃ + (η + )T φ̃ − θ
In addition, denote w1 = (θ̃ , Pg , Pl , φ̃ ), w2 = (λ , ω, η + , η − ).
Then L̂2 is convex in w1 and concave in w2 .
In F(w), [φ̃ −θ ]+
and [θ − φ̃ ]+
have unknown dimensions
η+
η−
+
−
(up to σ and σ ). Fortunately, we have
+
+∗ T
(η + − η +∗ )T [φ̃ − θ ]+
η + ≤ (η − η ) (φ̃ − θ )
= (η + − η +∗ )T ∇η + L̂2
where the inequality holds since ηi+j = 0 ≤ ηi+∗
j and φ̃i j −θ i j <
0 for (i, j) ∈ σ + , i.e., (ηi+j − ηi+∗
θ
)
·
(
φ̃
−
ij
i j ) ≥ 0. Similarly,
j
−
(η − η
−∗ T
) [θ
− φ̃ ]+
η−
−
≤ (η − η
−∗ T
−
−∗ T
= (η − η
) (θ − φ̃ )
) ∇η − L̂2 .
Consequently, it can be verified that
∗ T
(w − w∗ )T Γ−1
2 F(w) ≤ (w − w ) ∇w L̂2 (w)
∇w1 L̂2
= (w − w∗ )T
(w1 , w2 )
−∇w2 L̂2
0
0
0
0
0
0
0
I|E|−|σ + |
0
0
0
0
0
−CB
0
CB
0
CB
0
−I|E|−|σ + |
0
I|E|−|σ − |
−I|E|−|σ − | BCT CB
(B.2)
∇θ̃ L̂2
∇λ L̂2
∇ω L̂2
∇Pg L̂2
where, ∇w1 L̂2 :=
∇Pl L̂2 and ∇w2 L̂2 := ∇η + L̂2 .
∇φ̃ L̂2
∇η − L̂2
Then we have
∗ T
−1
− k(w − w∗ )T Γ−1
2 F(w) ≤ −k(w − w ) · Γ2 F(w)
= − k(w1 − w∗1 )T ∇w1 L̂2 (w1 , w2 ) + k(w2 − w∗2 )T ∇w2 L̂2 (w1 , w2 )
≤ k L̂2 (w∗1 , w2 ) − L̂2 (w1 , w2 ) + L̂2 (w1 , w2 ) − L̂2 (w1 , w∗2 )
= k L̂2 (w∗1 , w2 ) − L̂2 (w∗1 , w∗2 ) + L̂2 (w∗1 , w∗2 ) − L̂2 (w1 , w∗2 )
|
{z
} |
{z
}
≤0
≤0
≤0
(B.3)
where the first inequality holds because L̂2 is convex in w1
and concave in w2 and the second inequality follows because
(w∗1 , w∗2 ) is a saddle point. Therefore (B.1d) is nonpositive,
proving the first assertion.
To prove the second assertion, we further investigate the
situation that σ + or σ − changes. We only consider the set σ +
since it is the same to σ − . We have the following observations:
• The set σ + is reduced, which only happens when φ̃i j −θ i j
goes through zero, from negative to positive. Hence an
extra term will be added to V2 . As this term is initially
zero, there is no discontinuity of V2 in this case.
+
• The set σ + is enlarged when ηi j goes to zero from
positive while φ̃i j < θ i j . Here V2 will lose a positive term
(γiηj )2 (φ̃i j − θ i j )2 /2, causing discontinuity.
In the context, we conclude that V2 is always nonincreasing
along the trajectory even when σ + or σ − changes and
discontinuity occurs.
To prove that the trajectory w(t) is bounded note that [12,
Theorem 3.1] proves that V̂2 (w) := − (H(w) − w)T F(w) −
1
2
2 ||H(w) − w||2 satisfies V̂2 (w) ≥ 0 over S. Therefore, we have
1
∗
k(w(t) − w∗ )T Γ−2
2 (w(t) − w ) ≤ V2 (w(t)) ≤ V2 (w(0))
2
indicating the trajectory w(t) is bounded.
Lemma B.2. Suppose A1, A4 and A5 hold. Then
1) The trajectory w(t) converges to the largest weakly invariant subset W2∗ contained in W2 := {w ∈ S|V̇2 (w) = 0}.
2) Every point w∗ ∈ W2∗ is an equilibrium point of (16).
Proof of Lemma B.2 . Given an initial point w(0) there is a
compact set Ω0 := Ω(w(0)) ⊂ S such that w(t) ∈ Ω0 for t ≥ 0
and V̇2 (w) ≤ 0 in Ω0 .
11
Invoking the proof of Lemma B.1, V2 is radially unbounded
and positively definite except at equilibrium. As V2 and V̇2 are
nonpathological, we conclude that any trajectory w(t) starting
from Ω0 converges to the largest weakly invariant subset W2∗
contained in W2 = { w ∈ Ω0 | V̇2 (w) = 0 } [14, Proposition 3],
proving the first assertion.
For the second assertion, We fix w(0) ∈ W2∗ and then prove
that w(0) must be an equilibrium point.
From (B.1b), direct computing yields
V̇2 (w(t)) ≤ −ω̇ T Dω̇ − (Ṗg )T Ag Ṗg
− (Ṗl )T Al Ṗl − (CBφ̃˙ − Ṗg + Ṗl )T (CBφ̃˙ − Ṗg + Ṗl )
≤0
(B.4)
Since Ag , Al and D are positively definite diagonal matrices,
V̇2 (w) = 0 holds only when Ṗg = Ṗl = ω̇ = 0. Therefore, for
any w(0) ∈ W2∗ , the trajectory w(t) satisfies
Ṗg (t) = Ṗl (t) = ω̇(t) = 0,
Pg (t),
t ≥0
(B.5)
Pl (t)
Hence
and ω(t) are all constants due to the
boundedness property guaranteed by Lemma B.1.
On the other hand, for V̇2 (w) = 0, both terms in (B.3) have
to be zero, implying that
L̂2 (w1 (t), w∗2 ) = L̂2 (w∗1 , w∗2 )
must hold in W2 . Differentiating with respect to t gives
T
∂
∗
L̂2 (w1 (t), w2 ) · ẇ1 (t) = 0 = −φ̃˙ T (Γφ̃ )−1 φ̃˙ (B.6)
∂ w1
The second equality holds due to Eq. (B.5) and (13h). Then we
can conclude φ̃˙ = 0 immediately, implying φ̃ is also constant
in W ∗ due to its boundedness.
Invoking the close-loop dynamics (13), θ̃˙ (t), η̇ + (t), η̇ − (t)
and λ̇ (t) must be constants in W2∗ as Pg (t), Pl (t), ω(t) and
φ̃ (t) are all constants. Then we conclude that θ̃˙ (t) = η̇ + (t) =
η̇ − (t) = λ̇ (t) = 0 holds for all t ≥ 0 due to the boundedness
property of w(t) (Lemma B.1). This implies that any w(0) ∈
W2∗ must be an equilibrium point, completing the proof.
Proof of Theorem 4 . Fix any initial state w(0) and consider
the trajectory (w(t),t ≥ 0) of the closed-loop system closeloop dynamics (13). As mentioned in the proof of Lemma B.2,
w(t) stays entirely in a compact set Ω0 . Hence there exists an
infinite sequence of time instants tk such that w(tk ) → ŵ∗ as
tk → ∞, for some ŵ∗ ∈ W2∗ . Lemma B.2 guarantees that ŵ∗
is an equilibrium point of the closed-loop system (13), and
hence ŵ∗ = H(ŵ∗ ). Thus, using this specific equlibrium point
ŵ∗ in the definition of V2 , we have
V2∗ = lim V2 (w(t)) = lim V2 (w(tk ))
t→∞
tk →∞
=
lim V2 w(tk ) = V2 (ŵ∗ ) = 0
w(tk )→ŵ∗
Here, the first equality uses the fact that V2 (t) is nonincreasing
in t; the second equality uses the fact that tk is the infinite
sequence of t; the third equality uses the fact that w(t) is
absolutely continuous in t; the fourth equality is due to the
upper semi-continuity of V2 (w), and the last equality holds as
ŵ∗ is an equilibrium point of V2 .
∗
The quadratic term (w− ŵ∗ )T Γ−2
2 (w− ŵ ) in V2 then implies
∗
that w(t) → ŵ as t → ∞, which completes the proof.
| 3cs.SY
|
Cell Detection in Microscopy Images with Deep
arXiv:1708.03307v3 [cs.CV] 21 Feb 2018
Convolutional Neural Network and Compressed Sensing
Yao Xue and Nilanjan Ray
Computing Science, University of Alberta, Canada
Abstract
The ability to automatically detect certain types of cells or cellular subunits in
microscopy images is of significant interest to a wide range of biomedical research and
clinical practices. Cell detection methods have evolved from employing hand-crafted
features to deep learning-based techniques. The essential idea of these methods is that
their cell classifiers or detectors are trained in the pixel space, where the locations
of target cells are labeled. In this paper, we seek a different route and propose a
convolutional neural network (CNN)-based cell detection method that uses encoding
of the output pixel space. For the cell detection problem, the output space is the
sparsely labeled pixel locations indicating cell centers. We employ random projections
to encode the output space to a compressed vector of fixed dimension. Then, CNN
regresses this compressed vector from the input pixels. Furthermore, it is possible
to stably recover sparse cell locations on the output pixel space from the predicted
compressed vector using L1 -norm optimization. In the past, output space encoding
using compressed sensing (CS) has been used in conjunction with linear and non-linear
predictors. To the best of our knowledge, this is the first successful use of CNN with CSbased output space encoding. We made substantial experiments on several benchmark
datasets, where the proposed CNN + CS framework (referred to as CNNCS) achieved
the highest or at least top-3 performance in terms of F1-score, compared with other
state-of-the-art methods.
Keywords: Cell Detection, Convolutional Neural Network, Compressed Sensing.
1
1
Introduction
Automatic cell detection is to find whether there are certain types of cells present in an
input image (e.g. microscopy images) and to localize them in the image. It is of significant
interest to a wide range of medical imaging tasks and clinical applications. An example is
breast cancer, where the tumor proliferation speed (tumor growth) is an important biomarker
indicative of breast cancer patients’ prognosis. In practical scenario, the most common
method is routinely performed by pathologists, who examine histological slides under a
microscope based on their empirical assessments, which could be really accurate in several
cases, but generally is slow and prone to fatigue induced errors.
Cell detection and localization constitute several challenges that deserves our attention.
First, target cells are surrounded by clutters represented by complex histological structures
like capillaries, adipocytes, collagen etc. In many cases, the size of the target cell is small, and
consequently, it can be difficult to distinguish from the aforementioned clutter. Second, the
target cells can appear very sparsely (only in tens), moderately densely (in tens of hundreds)
or highly densely (in thousands) in a typical 2000-by-2000 pixel high resolution microscopy
image as shown in Fig. 1. Additionally, significant variations in the appearance among the
targets can also be seen. These challenges render the cell detection/localization/counting
problems far from being solved at the moment, in spite of significant recent progresses in
computer vision research.
In recent years, object detection has been significantly advanced following the big success
by deep learning. However, cell detection or localization task is not simply a sub-task of a
general object detection, which typically deals with extended objects, such as humans and
vehicles that occupy a significant portion of the field of view in the image. Extended object
detection and localization have witnessed much progress in the computer vision community.
For example, Region-based Convolutional Neural Networks (R-CNN) [17] and its variants
[16], [31], Fully Convolutional Networks (FCN) [34] with recent optimization [30] have become the state-of-the-art algorithms for the extended object detection problem. These solutions cannot be easily translated to cell detection, since assumptions and challenges are
different for the latter. For example, for an extended object, localization is considered suc-
2
Figure 1: Left picture shows a microscopy image with two target cells annotated by yellow
crosses on their centers. Right top pictures give details about the two target cells whose
nuclei are in mitotic phase. Right bottom pictures provide more examples of mitotic figures.
cessful if a detection bounding box is 50% overlapping with the actual bounding box. For
cell detection, tolerance is typically on a much tighter side in order for the localization to be
meaningful.
Conventional cell detection approaches
In the last few decades, different cell recognition methods had been proposed [15]. Traditional computer vision based cell detection systems adopt classical image processing techniques, such as intensity thresholding, feature detection, morphological filtering, region accumulation, and deformable model fitting. For example, Laplacian-of-Gaussian (LoG) [24]
operator was a popular choice for blob detection; Gabor filter or LBP feature [29] offers
many interesting texture properties and had been attempted for a cell detection task [3].
Conventional cell detection approaches follow a “hand-crafted feature representation”+“classifier”
framework. First, detection system extracts one (or multiple) kind of features as the representation of input images. Image processing techniques offer a range of feature extraction
algorithms for selection. After that, machine learning based classifiers work on the feature vectors to identify or recognize regions containing target cells. “Hand-crafted feature
representation”+“classifier” approaches suffer from the following limitations:
3
(1) It is a non-trivial and difficult task for humans to select suitable features. In many
cases, it requires significant prior knowledge about the target cells and background.
(2) Most hand-crafted features contain many parameters that are crucial for the overall
performance. Consequently, users need to perform a lot of trial-and-error experiments to
tune these parameters.
(3) Usually, one particular feature is not versatile enough. The feature may often be
tightly coupled with a particular type of target cell and may not work well when presented
with a different type of target cell.
(4) The performance of a hand-crafted feature-based classifier soon reaches an accuracy
plateau, even when trained with plenty of training data.
Deep learning based cell detection approaches
In comparison to the conventioal cell detection methods, deep neural networks recently
has been applied to a variety of computer vision problems, and has achieved better performance on several benchmark vision datasets [25], [17], [34]. The most compelling advantage
of deep learning is that it has evolved from fixed feature design strategies towards automated
learning of problem-specific features directly from training data [26]. By providing massive
amount of training images and problem-specific labels, users do not have to go into the
elaborate procedure for the extraction of features. Instead, deep neural network (DNN) is
subsequently optimized using a mini-batch gradient descent method over the training data,
so that the DNN allows autonomic learning of implicit relationships within the data. For
example, shallow layers of DNN focus on learning low-level features (such as edges, lines,
dots), while deep layers of DNN form more abstract high-level semantic representations (such
as probability maps, or object class labels).
With the advent of deep learning in the computer vision community, it is no wonder
that the state-of-the-art methods in cell detection are based on deep neural networks. The
essential idea behind all these methods is that detectors are trained as a classifier in the
image pixel space, either as a pixel lableing [4] or as a region proposal network [8]. Thus,
these methods predict the {x, y}-coordinates of cells directly on a 2-D image. Because, target
cell locations are sparse in an image, the classifiers in these methods face the class imbalance
issue. Moreover, target cells are often only subtly different from other cells. Thus, these
4
methods tend produce significant amount of false positives.
Compressed sensing-based output encoding
In this work, deviating from past approaches, we introduce output space encoding in the
cell detection and localization problem. Our observation is that the output space of cell
detection is quite sparse: an automated system only needs to label a small fraction of the
total pixels as cell centroid locations. To provide an example, if there are 5000 cells present in
an image of size 2000-by-2000 pixels, this fraction is 5000/(2000 ∗ 2000) = 0.00125, signifying
that even a dense cell image is still quite sparse in the pixel space.
Based on the observation about sparse cell locations within a microscopy image, we are
motivated to apply compressed sensing (CS) [12] techniques in the cell detection task. First,
a fixed length, compressed vector is formed by randomly projecting the cell locations from
the sparse pixel space. Next, a deep CNN is trained to predict the encoded, compressed
vector directly from the input pixels (i.e., microscopy image). Then, L1 norm optimization
is utilized to recover sparse cell locations. We refer to our proposed cell detection framework
as CNNCS (convolutional neural network + compressed sensing).
Output space encoding/representation/transformation sometimes yields more accurate
predictions in machine learning [11], [36]. In the past, CS-based encoding was used in conjunction with linear and non-linear predictions [19], [23], [21]. We believe, the proposed
CNNCS is the first such attempt to solve cell detection and localization that achieved competitive results on benchmark datasets. There are several advantages of using CS-based
output encoding for cell detection and localization. First, the compressed output vector is
much shorter in length than the original sparse pixel space. So, the memory requirement
would be typically smaller and consequently, there would be less risk of overfitting. Next,
there are plenty of opportunities to apply ensemble averages to improve generalization. Furthermore, CS-theory dictates that pairwise distances are approximately maintained in the
compressed space [12], [13]. Thus, even after output space encoding, the machine learner still
targets the original output space in an equivalent distance norm. From earlier research, we
also point out a generalization error bound for such systems. Our contribution is summarized
below:
First, this is the first attempt to combine deep learning with CS-based output encoding
5
to solve cell detection and localization problem.
Second, we try to overcome the aforementioned class imbalance issue by converting a
classification problem into a regression problem, where sparse cell locations are distributed
by a random projection into a fixed length vector as a target for the regression.
Third, we introduce redundancies in the CS-based output encoding that are exploited by
CNN to boost generalization accuracy in cell detection and localization. This redundancies
also help to reduce false detections.
Fourth, we demonstrate that the proposed CNNCS framework achieves competitive results compared to the state-of-the-art methods on several benchmark datasets and challenging cell detection contests.
2
Background and Related Work
2.1
General Object Detection
Prior to deep learning, general object detection pipeline consisted of feature extraction followed by classifiers or detectors. Detection had traditionally been addressed using the handcrafted features such as SIFT [27], HOG [10], LBP [29], etc. At that time, progress in object
detection greatly depended on the invention of more discriminative hand-crafted features.
Following the big success of Convolutional Neural Network (CNN) in image classification task
[25], deep learning model have been widely adopted in the computer vision community. For
example, Region-based Convolutional Neural Networks (R-CNN) [17] and its variants [16],
[31], Fully Convolutional Networks (FCN) [34] with recent optimization [30] have become
the state-of-the-art algorithms for the extended object detection problems.
Once again, cell detection or localization task is not simply a sub-task of a general object
detection, those state-of-the-art solutions for general object detection are able to provide
useful clues to right direction, but cannot be easily translated to cell detection.
6
2.2
Cell Detection and Localization
The state-of-the-art methods in detection and localization today include deep learning techniques for cell detection and localization. Recently a deep Fully Convolutional Network
(FCN) [34] was proposed for the image segmentation problem and had shown remarkable
performance. Soon after the FCN is proposed, [38] presented a FCN-based framework for
cell counting, where their FCN is responsible for predicting a spatial density map of target
cells, and the number of cells can be estimated by an integration over the learned density
map. Slightly similar to [38], a cascaded network [8] has been proposed for cell detection.
[8] uses a FCN for candidate region selection, and then a CNN for further discrimination
between target cells and background.
In [9], a mitosis detection method has been proposed by CNN-based prediction followed
by ad-hoc post processing. As the winner of ICPR 2012 mitosis detection competition, [9]
used deep max-pooling convolutional neural networks to detect mitosis in breast histology
images. The networks were trained to classify each pixel in the images, using a patch centered
on the pixel as context. Then post processing was applied to the network output. In [33],
expectation maximization has been utilized within deep learning framework in an end-to-end
fashion for mitosis detection. This work presents a new concept for learning from crowds
that handle data aggregation directly as part of the learning process of the convolutional
neural network (CNN) via additional crowd-sourcing layer. It is the first piece of work where
deep learning has been applied to generate a ground-truth labeling from non-expert crowd
annotation in a biomedical context.
2.3
Compressed Sensing
During the past decade, compressed sensing or compressive sensing (CS) [12] has emerged as a
new framework for signal acquisition and reconstruction, and has received growing attention,
mainly motivated by the rich theoretical and experimental results as shown in many reports
[13], [14], [12], and so on. As we know, the Nyquist-Shannon sampling theorem states that
a certain minimum sampling rate is required in order to reconstruct a band-limited signal.
However, CS enables a potentially large reduction in the sampling and computation costs for
7
sensing/reconstructing signals that are sparse or have a sparse representation under some
linear transformation (e.g. Fourier transform).
Under the premise of CS, an unknown signal of interest is observed (sensed) through a
limited number of linear observations. Many works [13], [14], [12] have proven that it is
possible to obtain a stable reconstruction of the unknown signal from these observations,
under the general assumptions that the signal is sparse (or can be represented sparsely with
respect to a linear basis) and matrix discoherence. The signal recovery techniques typically
rely on a convex optimization with a penalty expressed by L1 norm, for example orthogonal
matching pursuit (OMP) [5] and dual augmented Lagrangian (DAL) method [35].
3
Proposed Method
3.1
System Overview
The proposed detection framework consists of three major components: (1) cell location
encoding phase using random projection, (2) a CNN based regression model to capture the
relationship between a cell microscopy image and the encoded signal y, and (3) decoding
phase for recovery and detection. The flow chart of the whole framework is shown in Fig. 2.
During training, the ground truth location of cells are indicated by a pixel-wise binary
annotation map B. We propose two cell location encoding schemes, which convert cell
location from the pixel space representation B to a compressed signal representation y.
Then, training pairs, each consisting of a cell microscopy image and the compressed signal
y, train a CNN to work as a multi-label regression model. We employ the Euclidean loss
function during training, because it is often more suitable for a regression task. Image
rotations may be performed on the training sets for the purpose of data augmentation as
well as making the system more robust to rotations.
During testing, the trained network is responsible for outputting an estimated signal ŷ
for each test image. After that, a decoding scheme is designed to estimate the ground truth
cell location by performing L1 minimization recovery on the estimated signal ŷ, with the
known sensing matrix.
8
Figure 2: The system overview of the proposed CNNCS framework for cell detection and
localization.
3.2
3.2.1
Cell Location Encoding and Decoding Scheme
Encoding Schemes
In the CNNCS framework, we employ two types of random projection-based encodings as
described below.
Scheme-1: Encoding by Reshaping
For the cell detection problem, cells are often annotated by pixel-level labels. The most
common way is to attach a dot or cross at the center of every cell, instead of a bounding
box around the cell. So, let us suppose there is a pixel-wise binary annotation map B of
size w-by-h, which indicates the location of cells by labeling 1 at the pixels of cell centroids,
otherwise labeling 0 at background pixels. To vectorize the annotation map B, the most
intuitive scheme is to concatenate every row of B into a binary vector f of length wh. Thus,
9
a positive element in B with {x, y} coordinates will be encoded to the [x + h(y − 1)]-th
position in f . f is also a k-sparse signal, so, there are at most k non-zero entries in f . Here,
we refer this intuitive encoding scheme as ”Scheme-1: Encoding by Reshaping”.
After the vector f is generated, we apply a random projection. CS theory guarantees
that f could be fully represented by linear observations y:
y = Φf,
(1)
provided the sensing matrix Φ satisfies a restricted isometry property (RIP) condition [13],
[14]. In many cases, Φ is typically a M × N (M N = hw) random Gaussian matrix.
Here, the number of observations M is much smaller than N , and obeys: M ≥ CM klog(N ),
where CM is a small constant greater than one.
Scheme-2: Encoding by Signed Distances
For the encoding scheme-1, the space complexity of the interim result f is O(wh). For
example, to encode the location of cells in a 260-by-260 pixel image, scheme-1 will produce f
as a 67,600-length vector; so that in the subsequent CS process, a huge sensing matrix in size
of M -by-67600 is required in order to match the dimension of f , which will make the system
quite slow, even unacceptable for larger images. To further optimize the encoding scheme,
we propose a second scheme, where the coordinates of every cell centroid are projected onto
multiple observation axes. We refer the second encoding scheme as ”Scheme-2: Encoding by
Projection.”
To encode location of cells, we create a set of observation axes OA = {oal } , l = 1, 2, . . . , L,
where L indicates the total number of observation axes used. The observation axes are
uniformly-distributed around an image (See Fig. 3, left-most picture) For the l-th observation
√
axis oal , the location of cells is encoded into a R-length (R = w2 + h2 ) sparse signal,
referred as fl (See Fig. 3, third picture). We calculate the perpendicular signed distances (fl )
from cells to oal . Thus, fl contains signed distances, which not only measure the distance,
but also describe on which side of oal cells are located. After that, the encoding of cell
locations under oal is yl , which is obtained by the following random projection:
yl = Φfl ,
10
(2)
We repeat the above process for all the L observation axes and obtain each yl . After concatenating all the yl , l = 1, 2, . . . , L, the final encoding result y is available, which is the joint
representation of cells location. The whole encoding process is illustrated by Fig. 3.
Figure 3: Cell location encoding by signed distances (Scheme-2).
√
For encoding scheme-2, the size of the sensing matrix Φ is M -by- w2 + h2 . In comparison, encoding scheme-1 requires a much larger sensing matrix of size M -by-wh. The
first advantage of encoding scheme-2 is that it dramatically reduces the size of the sensing
matrix, which is quite helpful for the recovery process, especially when the size of images is
large. Secondly, the encoding result y carries redundant information about cell locations.
In the subsequent decoding phase, averaging over the redundant information makes the final
detection more reliable. More details can be found in experiments section. A final point
is that in case more than one cell locations are projected to the same bin in a particular
observation axis, such a conflict will not occur for the same set of cells at other observation
axes.
3.2.2
Decoding Scheme
Accurate recovery of f can be obtained from the encoded signal y by solving the following
L1 norm convex optimization problem:
fˆ = arg min kf k1
f
subject to
y = Φf
(3)
After fˆ is recovered, every true cell is localized L times, i.e. with L candidate positions
predicted. The redundancy information allows us to estimate more accurate detection of a
true cell.
11
The first two images of Fig. 4 from left present examples of the true location signal f
and decoded location signal fˆ, respectively. The noisy signed distances of fˆ are typically
very close to each observation axis. That is why we create observation axes outside of the
image space, so that these noisy distances can be easily distinguished from true candidate
distances. This separation is done by mean shift clustering, which also groups true detections
into localized groups of detections. Two such groups (clusters) are shown in Fig. 4, where the
signed distances formed circular patterns of points (in green) around ground truth detections
(in yellow). Averaging over these green points belonging to a cluster provides us a predicted
location (in red) as shown in Fig. 4.
Figure 4: Cell Location Decoding Scheme. From left to right: true location signal f , decoded
location signal fˆ and detection results. Yellow crosses indicate the ground-truth location of
cells, green crosses are the candidates points, red crosses represent the final detected points.
3.3
Signal Prediction by Convolutional Neural Network
We utilize a CNN to build a regression model between a cell microscopy image and its cell
location representation: compressed signal y. We employ two kinds of CNN architectures.
One of them is AlexNet [25], which consists of 5 convolution layers + 3 fully connected layers;
the other is the deep residual network (ResNet) [18] where we use its 152-layer model. In
both the architectures, the loss function is defined as the Euclidean loss. The dimension of
output layer of AlexNet and ResNet has been modified to the length of compressed signal
y. We train the AlexNet model from scratch, in comparison, we perform fine-tuning on the
weights in fully-connected layer of the ResNet.
To prepare the training data, we generate a large number of square patches from training
12
Figure 5: An illustration of the process of signal prediction by convolutional neural network.
The bottom row presents the feature maps learned from Convolutional (Conv) layers of the
CNN with training process going on. The current CNN follows the AlexNet architecture.
These feature maps come from the Conv1, Conv1, Conv2, Conv3, Conv3, Conv4 and Conv5
respectively. The top-right picture shows the ground-truth compressed signal (red) and
compressed signal (blue) predicted from the Fullly-connected (Fc) layer of the CNN. From
the picture, we can observe that the predicted signal approximate the pattern of ground
truth signal well.
13
images. Along with each training patch, there is a signal (i.e. the encoding result: y),
which indicates the location of target cells present in the patch. After that, patch rotation is
performed on the collected training patches for data augmentation and making the system
rotation invariant.
The trained CNN not only predicts the signal from its output layer, the feature maps
learned from its Conv layers also provide rich information for recognition. Fig. 5 visualizes
the learned feature maps, which represents the probabilistic score or activation maps of
target cell regions (indicated by green boxes in the left image) during training process. It
can be observed that higher scores are fired on the target regions of score masks, while most
of the non-target regions have been suppressed more and more with training process going
on.
To further optimize our CNN model, we apply Multi-Task Learning (MTL) [7]. During
training a CNN, two kinds of labels are provided. The first kind is the encoded vector:
y, which carries the pixel-level location information of cells. The other kind is a scalar:
cell count (c), which indicates the total number of cells in a training image patch. We
concatenate the two kinds of labels into the final training label by label = {y, λc}, where λ
is a hyper parameter. Then, Euclidean loss is applied on the fusion label. Thus, supervision
information for both cell detection and cell counting can be jointly used to optimize the
parameters of our CNN model.
3.4
Theoretical Justification
Equivalent Targets for Optimization
We first show that from the optimization standpoint, compressed vector is a good proxy
for the original, sparse output space. This result directly follows from the CS theory. As
mentioned before, f indicates the cell location represented in pixel space, and y is the cell
location represented in compressed signal space. They follow the relationship: y = Φf ,
where Φ is the sensing matrix. Let us assume that fp and fg are respectively the prediction
and ground-truth vectors in the pixel space. Similarly, we have yp and yg as their compressed
counterparts, respectively.
Claim: kyg − yp k and kfg − fp k are approximately equivalent targets for optimization.
14
Proof: According to the CS theory, a sensing matrix Φ ∈ Rm×d should satisfy the
(k, δ) − restricted isometry property ((k, δ) − RIP ), which states that for all k−sparse f ∈
Rd , δ ∈ (0, 1), the following holds [13], [14], [12]:
(1 − δ) kf k ≤ kΦf k ≤ (1 + δ) kf k.
(4)
Note that if the sensing matrix Φ satisfies (2k, δ)-RIP, then (4) also holds good. Now
replace f with (fg − fp ) and note that (fg − fp ) is 2k-sparse. Thus,
(1 − δ) kfg − fp k ≤ kyg − yp k ≤ (1 + δ) kfg − fp k .
(5)
From the right hand side inequality, we note that if kfg − fp k is small, then kyg − yp k would
be small too. In the same way, if kyg − yp k is large, then the inequality implies that kfg − fp k
would be large too. Similarly, from the left hand side inequality, we note that if kfg − fp k
is large then kyg − yp k will be large, and if kyg − yp k is small then kfg − fp k will small too.
These relationships prove the claim that from the optimization perspective kyg − yp k and
kfg − fp k are approximately equivalent.
A Bound on Generalization Prediction Error
In this section we mention a powerful result from [19]. Let h be the predicted compressed
vector by the CNN, f be the ground truth sparse vector, fˆ be the reconstructed sparse vector
from prediction, and Φ be the sensing matrix. Then the generalization error bound provided
in [19] is as follows:
kfˆ − f k22 ≤ C1 · kh − Φf k22 + C2 · sperr(fˆ, f ),
(6)
where C1 and C2 are two small constants and sperr measures how well the reconstruction
algorithm has worked [19]. This result demonstrates that expected error in the original space
is bound by the expected errors of the predictor and that of the reconstruction algorithm.
Thus, it makes sense to apply a very good machine learner such as deep CNN that can
minimize the first term in the right hand side of (6). On the other hand, DAL provides one
of the best L1 recovery algorithms to minimize the second term in the right side of (6).
15
4
Experiments
4.1
Datasets and Evaluation Criteria
We utilize seven cell datasets, on which CNNCS and other comparison methods are evaluated. The 1st dataset [22] involves 100 H&E stained histology images of colorectal adenocarcinomas. The 2nd dataset [6] consists of 200 highly realistic synthetic emulations of
fluorescence microscopic images of bacterial cells. The 3rd dataset [39] comprises of 55 high
resolution microscopic images of breast cancers double stained in red (cytokeratin epithelial
marker) and brown (nuclear proliferative marker). The 4th dataset is the ICPR 2012 mitosis
detection contest dataset [32] including 50 high-resolution (2084-by-2084) RGB microscope
slides of Mitosis. The 5th dataset [1] is the ICPR 2014 grand contest of mitosis detection,
which is a follow-up and an extension of the ICPR 2012 contest on detection of mitosis.
Compared with the contest in 2012, the ICPR 2014 contest is much more challenging, which
contains way more images for training and testing. The 6th dataset is the AMIDA-2013
mitosis detection dataset [37], which contains 676 breast cancer histology images belonging
to 23 patients. The 7th dataset is the AMIDA-2016 mitosis detection dataset [2], which is
an extension of the AMIDA 2013 contest on detection of mitosis. It contains 587 breast
cancer histology images belonging to 73 patients for training, and 34 breast cancer histology
images for testing with no ground truth available. For each dataset, the annotation that
represents the location of cell centroids is shown in Fig.6, details of datasets are summarized
in Table. 1.
Figure 6: Dataset examples and their annotation.
For evaluation, we adopt the criteria of the ICPR 2012 mitosis detection contest [32],
which is also adopted in several other cell detection contests. A detection would be counted
as true positive (T P ) if the distance between the predicted centroid and ground truth cell
16
Table 1: Size is the image size; Ntr/Nte is the number of images selected for training and
testing; AC indicates the average number of cells per image.
Cell Dataset
Size
Ntr/Nte
AC
Nuclei [22]
500×500
50/50
310.22
Bacterial [6]
256×256
100/100
171.47
Ki67 Cell [39]
1920×2560
45/10
2045.85
ICPR 2012 [32]
2084×2084
35/15
5.31
ICPR 2014 [1]
1539×1376
1136/496
4.41
AMIDA 2013 [37]
2000×2000
447/229
3.54
AMIDA 2016 [2]
2000×2000
587/34
2.13
centroid is less than ρ. Otherwise, a detection is considered as false positives (F P ). The
missed ground truth cells are counted as false negatives (F N ). In our experiments, ρ is set
to be the radius of the smallest cell in the dataset. Thus, only centroids that are detected
to lie inside cells are considered correct. The results are reported in terms of Precision:
P = T P/(T P + F P ) and Recall: R = T P/(T P + F N ) and F1 -score: F1 = 2P R/(P + R) in
the following sections.
4.2
Experiments with Encoding Scheme-1
To evaluate, we carry out performance comparison experiment between CNNCS and three
state-of-the-art cell detection methods (“FCN-based” [38], “Le.detect” [4], “CasNN” [8]). In
this experiment, the scheme-1: encoding by reshaping is applied in CNNCS.
For the four methods to provide different values of Precision-Recall as shown in Fig. 7,
we tune hyper parameters of every method. With scheme-1, CNNCS has a threshold T
to apply on the recovered sparse signal fˆ before re-shaping it to a binary image B. T
is used to perform cell vs. non-cell binary classification and can be treated as a hyper
parameter during training. In “FCN-based” [38], there is also a threshold applied to the
local probability-maximum candidate points to make final decision about cell or non-cell.
Similarly, in the first step of “Le.detect” [4], researchers use a MSER-detector (a stability
threshold involved here) to produce a number of candidate regions, on which their learning
17
procedure determines which of these candidates regions correspond to cells. In the first
experiment, we analyze the three methods using Precision-Recall curves by varying their
own thresholds.
Fig. 7 presents Precision-Recall curves on three cell datasets. All the four methods give
reliable detection performances in the range of recall=[0.1-0.4]. After about recall=0.6, the
precision of “FCN-based” [38] drops much faster. This can be attributed to the fact that
“FCN-based” [38] works by finding local maximum points on a cell density map. However,
the local maximum operation fails in several scenarios, for example when two cell density
peaks are close to each other, or large peak may covers neighboring small peaks. Consequently, to obtain the same level of recall, “FCN-based” [38] provides many false detections.
Figure 7: Precision and recall curves of four methods on three datasets.
Furthermore, it also can be observed that CNNCS has an improvement over “Le.detect”
18
[4] (red line clearly outperforms black line under varying recall values). This can be largely
explained by the fact that traditional methods (no matter if [4] or [38] is used) always try
to predict the coordinates of cells directly on a 2-D image. The coordinates are sensitive
to system prediction bias or error, considering the nature of cell detection that cells are
small and quite dense in most cases. It is not surprising that “Le.detect” [4] will miss
some cells and/or detect other cells in wrong locations. In comparison, CNNCS transfers
the cell detection task from pixel space to compressed signal space, where the location
information of cells is no longer represented by {x, y}-coordinates. Instead, CNNCS performs
cell detection by regression and recovery on a fixed length compressed signal. Compared to
{x, y}-coordinates representation, the compressed signal is more robust to system prediction
errors. For example, as shown in the right top corner of Fig. 5, even though there are
differences between the ground-truth compressed signal and predicted compressed signal,
the whole system can still give reliable detection performance as shown in Fig. 7.
To get a better idea of the CNNCS method, we visualize a set of cell images with their
detected cells and ground-truth cells in Fig. 8. It can be observed that CNNCS is able to
accurately detect most cells under a variety of conditions.
4.3
4.3.1
Experiments with Encoding Scheme-2
Experiment on ICPR 2012 mitosis detection dataset
To evaluate the performance of encoding scheme 2, we carry out the second group of performance comparison experiments. In the first experiment, we apply the proposed method on
the ICPR 2012 mitosis detection contest dataset, which consists of 35 training images and
15 testing images. For the training process, we extracted image sub-samples (260-by-260)
with no overlap between each other from the 35 training images. After that every 90◦ image
rotation is performed on each sub-sample for data augmentation, resulting in a total of 8,960
training dataset. In addition, we perform random search to tune the three hyper parameters
in scheme-2: (1) the number of rows in sensing matrix: M , (2) the number of observation
lines: L and (3) the importance (λ) of cell count during MTL. After that, the best performance is achieved when M = 112, L = 27, λ = 0.20. Furthermore, we trained five CNN
19
Figure 8: Detection results. Ground-truth: red, Prediction: blue. Left part shows the
ground truth signal and the predicted sparse signal that carries the location information of
cells; right part shows the ground-truth and detected cells.
models to reduce the performance variance introduced by a single model and to improve
the robustness of the whole system. Recently, deep residual network (ResNet) introduces
residual connections into deep convolutional networks and has yielded state-of-the-art performance in the 2015 ILSVRC challenge [18]. This raises the question of whether there is any
benefit in introducing and exploiting more recent CNN architectures into the cell detection
task. Thus, in the experiment, we have explored the performance of CNNCS with different neural network architectures (AlexNet and ResNet). Finally, CNNCS gets the highest
F1-score among all the comparison methods, details are summarized in Table 2.
Compared to the state-of-the-art method: CasNN-average [8], CNNCS with ResNet and
MTL achieved a better performance with F1 -score 0.837. It can be observed from Table 2
that the precision of our method outperforms the previous best precision by 0.06-0.07, and
recall also has recorded about 0.02 improvement. This phenomenon can be attributed to the
20
Table 2: Results of ICPR 2012 grand challenge of mitosis detection.
Method
Precision
Recall
F1 -score
UTRECHT
0.511
0.680
0.584
NEC [28]
0.747
0.590
0.659
IPAL [20]
0.698
0.740
0.718
DNN [9]
0.886
0.700
0.782
RCasNN [8]
0.720
0.713
0.716
CasNN-single [8]
0.738
0.753
0.745
CasNN-average [8]
0.804
0.772
0.788
CNNCS-AlexNet
0.860
0.788
0.823
CNNCS-ResNet
0.867
0.801
0.833
CNNCS-ResNet-MTL
0.872
0.805
0.837
detection principle of our method, where every ground-truth cell is localized with multiple
candidate points guaranteed to be around the true location, then the average coordinates of
these candidates is computed as the final detection. As a result, localization closer to the true
cell becomes more reliable compared to other methods, thus leading to a higher precision.
In addition, an improvement of F1 -score from 0.833 to 0.837 achieved by MTL demonstrates
that the knowledge jointly learned from cell detection and cell counting provides further
benefits at negligible additional computations.
4.3.2
Experiment on ICPR 2014 mitosis detection dataset
In the second experiment, we evaluated CNNCS on the ICPR 2014 contest of mitosis detection dataset (also called MITOS-ATYPIA-14), which is a follow-up and an extension of the
ICPR 2012 contest on detection of mitosis. Compared with the contest in 2012, the ICPR
2014 contest is much more challenging, which contains more images for training and testing.
It provides 1632 breast cancer histology images, 1136 images for training, 496 images for
testing. Each image is in the size of 1539×1376. We divide the training images into training
set (910 images) and validation set (226 images). We perform random search on the validation set to optimize the hyper parameters. The best performance on MITOS-ATYPIA-14
dataset is achieved when M = 103, L = 30, λ = 0.24. On the test dataset, we have achieved
21
the highest F1-score among all the participated teams. The F1-score of all the participated
teams are shown in Table 3. As we see, the CNNCS method shows significant improvement
compared to the results of other teams in all the histology slice groups. On an average,
CNNCS has almost doubled the F1-score of teams at the second place.
Table 3: Results of ICPR 2014 contest of mitosis detection in breast cancer histological
images. F1-scores of participated teams are shown.
4.3.3
Slice group
CUHK
MINES
YILDIZ
STRAS
CNNCS
A06
0.119
0.317
0.370
0.160
0.783
A08
0.333
0.171
0.172
0.024
0.463
A09
0.593
0.473
0.280
0.072
0.660
A19
0.368
0.137
0.107
0.011
0.615
Average
0.356
0.235
0.167
0.024
0.633
Experiment on AMIDA 2013 mitosis detection dataset
The third experiment was performed on the AMIDA-2013 mitosis detection dataset, which
contains 676 breast cancer histology images, belonging to 23 patients. Suspicious breast
tissue is annotated by at least two expert pathologists, to label the center of each cancer
cell. We train the proposed CNNCS method using 377 images, validate on 70 training images
and test it on the testing set of AMIDA-2013 challenge that has 229 images from the last
8 patients. We employ ResNet as the network architecture with data balancing and MTL
in the training set. Similar to previous experiments, we perform random search on the
validation set to optimize the hyper parameters. The best performance on AMIDA-2013
dataset is achieved when M = 118, L = 25, λ = 0.32. Finally among all the 17 participated
teams, we achieve the third highest F1-score=0.471, which is quite close to the second
place, and has a significant improvement over the fourth place method [33]. For details,
Table 4 summarizes the comparison between CNNCS and other methods.
22
Table 4: Results of AMIDA-2013 MICCAI grand challenge of mitosis detection.
4.3.4
Method
Precision
Recall
F1 -score
IDSIA [9]
0.610
0.612
0.611
DTU
0.427
0.555
0.483
AggNet [33]
0.441
0.424
0.433
CUHK
0.690
0.310
0.427
SURREY
0.357
0.332
0.344
ISIK
0.306
0.351
0.327
PANASONIC
0.336
0.310
0.322
CCIPD/MINDLAB
0.353
0.291
0.319
WARWICK
0.171
0.552
0.261
POLYTECH/UCLAN
0.186
0.263
0.218
MINES
0.139
0.490
0.217
SHEFFIELD/SURREY
0.119
0.107
0.113
SEOUL
0.032
0.630
0.061
UNI-JENA
0.007
0.077
0.013
NIH
0.002
0.049
0.003
CNNCS
0.3588
0.5529
0.4352
Experiment on AMIDA 2016 mitosis detection dataset
In the fourth experiment, we participated in the AMIDA-2016 mitosis detection challenge
(also called TUPAC16), which is a follow-up and an extension of the AMIDA-2013 contest
on detection of mitosis. Its training dataset has 587 breast cancer histology images in size
of 2000×2000, belonging to 73 patients. Its test dataset contains 34 breast cancer histology
images in the same size without publicly available ground truth labels.
We train the proposed CNNCS method using randomly chosen 470 training images and
validate on the remaining 117 training images. Additionally, we apply the following ensemble
averaging technique to further increase precision and recall values. Originally, we have
partitioned every test image into about 100 non-overlapping patches. Instead of starting
the partitioning from the top-left corner, now we set the starting point of the first patch
from {offset, offset}. The offset values are set as 0, 20, 40,..., 160, and 180 (i.e. every
20 pixel) resulting in a total of 10 different settings. Under every offset setting, CNNCS
23
Table 5: Results of AMIDA-2016 MICCAI grand challenge of mitosis detection.
Team
F1 -score
Lunit Inc.
0.652
IBM Research Zurich and Brazil
0.648
Contextvision (SLDESUTO-BOX)
0.616
The Chinese University of Hong Kong
0.601
Microsoft Research Asia
0.596
Radboud UMC
0.541
University of Heidelberg
0.481
University of South Florida
0.440
Pakistan Institute of Engineering and Applied Sciences
0.424
University of Warwick
0.396
Shiraz University of Technology
0.330
Inha University
0.251
CNNCS (on validation set)
0.634
method is run on all the generated patches and provides detection results. Then, we merge
detection results from all the offset settings. The merging decision rule is that if there are
6 or more detections within a radius of 9 pixels, then we accept average of these locations
as our final detected cell center. Other implementation settings are similar to the settings
in the experiment of AMIDA-2013. Finally, we achieved F1-score=0.634 on the validation
set (becuase of the lack of publicly available test set), which is the third highest in all the
15 participated teams. Table 5 provides more details of the contest results. Furthermore,
Fig.9 provides twelve examples of our detection results in the AMIDA-2016 grand challenge
of mitosis detection.
5
Conclusion
This is the first attempt demonstrating that deep convolutional neural network can work
in conjunction with compressed sensing-based output encoding schemes toward solving a
significant medical image processing task: cell detection and localization from microscopy
images. In this work, we made substantial experiments on several mainstream datasets and
24
Figure 9: Results on AMIDA-2016 dataset. Yellow cross indicates the ground-truth position
of target cells. Green cross indicates cell position predicted by an observation axis. Red
cross indicates the final detected cell position, which is the average of all green crosses.
challenging cell detection contests, where the proposed CNN + CS framework (referred to
as CNNCS) achieved very competitive (the highest or at least top-3 in terms of F1-score)
results compared to the state-of-the-art methods in cell detection task. In addition, the
CNNCS framework has the potential to be trained in End-to-End manner, which is our near
future plan and could further boost performance.
25
References
[1] https://mitos-atypia-14.grand-challenge.org/home/.
[2] http://tupac.tue-image.nl/.
[3] Yousef Al-Kofahi, Wiem Lassoued, William Lee, and Badrinath Roysam. Improved
automatic detection and segmentation of cell nuclei in histopathology images. IEEE
Transactions on Biomedical Engineering, 57:841–852, 2010.
[4] Carlos Arteta, Victor Lempitsky, J. Alison Noble, and Andrew Zisserman. Learning
to detect cells using non-overlapping extremal regions. International Conference on
Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 348–
356, 2012.
[5] T. Tony Cai and Lie Wang. Orthogonal matching pursuit for sparse signal recovery
with noise. IEEE Transactions on Information Theory., 2011.
[6] Alison Noble Carlos Arteta, Victor Lempitsky and Andrew Zisserman. Learning to
count objects in images. Neural Information Processing Systems., 2010.
[7] Rich Caruana. Multitask learning. Machine Learning, 28(1):41–75, Jul 1997.
[8] Hao Chen, Qi Dou, Xi Wang, Jing Qin, and Pheng-Ann Heng. Mitosis detection in
breast cancer histology images via deep cascaded networks. Proceedings of the Thirtieth
Conference on Artificial Intelligence (AAAI), 2016.
[9] Dan C. Cirean, Alessandro Giusti, Luca M. Gambardella, and Jrgen Schmidhuber.
Mitosis Detection in Breast Cancer Histology Images with Deep Neural Networks. 2013.
[10] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection.
Computer Vision and Pattern Recognition, 2005.
[11] T. G. Dietterich and G. Bakiri. Solving multiclass learning problems via error-correcting
output codes. Journal of Artificial Intelligence Research, pages 263–286, 1995.
26
[12] David L. Donoho. Compressed sensing. IEEE Transactions on Information Theory,
2006.
[13] Justin Romberg Emmanuel Candes. Practical signal recovery from random projections.
IEEE Transactions on Signal Process, 2005.
[14] Terence Tao Emmanuel Candesy, Justin Rombergy. Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information. IEEE Transactions on Information Theory., 2006.
[15] Meijering Erik. Cell segmentation: 50 years down the road [life sciences]. IEEE Signal
Process. Mag., 2012.
[16] Ross Girshick. Fast r-cnn. International Conference on Computer Vision, 2015.
[17] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. Computer Vision and
Pattern Recognition, 2014.
[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for
image recognition. Computer Vision and Pattern Recognition., 2015.
[19] Daniel Hsu, Sham M. Kakade, John Langford, and Tong Zhang. Multi-label prediction
via compressed sensing. arXiv:0902.1284v2 [cs.LG], 2009.
[20] Humayun Irshad. Automated mitosis detection in histopathology using morphological
and multi-channel statistics features. Journal of Pathology Informatics, 2013.
[21] Arnaud Joly. Exploiting random projections and sparsity with random forests and
gradient boosting methods. arXiv:1704.08067, 2016.
[22] Y.W Tsang I.A. Cree D.R.J. Snead K. Sirinukunwattana, S.E.A. Raza and N.M. Rajpoot. Locality sensitive deep learning for detection and classification of nuclei in routine
colon cancer histology images. IEEE Transactions on Medical Imaging., 2016.
[23] Ashish Kapoor, Raajay Viswanathan, and Prateek Jain. Multilabel classification using
bayesian compressed sensing. Advances in Neural Information Processing Systems, 2012.
27
[24] Hui Kong, Hatice Cinar Akakin, and Sanjay E. Sarma. A generalized laplacian of gaussian filter for blob detection and its applications. IEEE Transactions on Cybernetics,
43:1719–1733, 2013.
[25] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classication with
deep convolutional neural networks. Neural Information Processing Systems, 2012.
[26] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521:436–444,
2015.
[27] David G. Lowe. Object recognition from local scale-invariant features. In Proceedings
of the International Conference on Computer Vision-Volume 2 - Volume 2, ICCV ’99,
pages 1150–, Washington, DC, USA, 1999. IEEE Computer Society.
[28] Christopher D. Malon and Eric Cosatto. Classification of mitotic figures with convolutional neural networks and seeded blob features. Journal of Pathology Informatics,
2013.
[29] T. Ojala, M. Pietikinen, and D. Harwood. A comparative study of texture measures
with classification based on feature distributions. Pattern Recognition, 29:51–59, 1996.
[30] Joseph Redmon and Ali Farhadi. Yolo9000: Better, faster, stronger. arXiv preprint
arXiv:1612.08242, 2016.
[31] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards realtime object detection with region proposal networks. Advances in Neural Information
Processing Systems, 2015.
[32] Ludovic Roux, Daniel Racoceanu, N. Loménie, Maria S. Kulikova, Humayun Irshad,
J. Klossa, Frédérique Capron, Catherine Genestie, Gilles Le Naour, and Metin N. Gurcan. Mitosis detection in breast cancer histological images, an icpr 2012 contest. Journal
of Pathology Informatics, 4, 05/2013 2013.
[33] Albarqouni Shadi, Baur Christoph, Achilles Felix, Belagiannis Vasileios, Demirci Stefanie, and Navab Nassir. Aggnet: Deep learning from crowds for mitosis detection in
breast cancer histology images. IEEE Transactions on Medical Imaging., 2016.
28
[34] Evan Shelhamer, Jonathan Long, and Trevor Darrell. Fully convolutional models for
semantic segmentation. The IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016.
[35] Ryota Tomioka, Taiji Suzuki, and Masashi Sugiyama. Super-linear convergence of dual
augmented lagrangian algorithm for sparsity regularized estimation. The Journal of
Machine Learning Research, 2011.
[36] Grigorios Tsoumakas, Ioannis Katakis, and Ioannis Vlahavas. Random k-labelsets for
multi-label classification. IEEE Transactions on Knowledge and Data Engineering,
pages 1079–1089, 2011.
[37] Mitko Veta, Paul J van Diest, Stefan M Willems, Haibo Wang, Anant Madabhushi, Angel Cruz-Roa, Fabio Gonzalez, Anders B L Larsen, Jacob S Vestergaard, Anders B Dahl,
Dan C Cireşan, Jürgen Schmidhuber, Alessandro Giusti, Luca M Gambardella, F Boray
Tek, Thomas Walter, Ching-Wei Wang, Satoshi Kondo, Bogdan J Matuszewski, Frederic Precioso, Violet Snell, Josef Kittler, Teofilo E de Campos, Adnan M Khan, Nasir M
Rajpoot, Evdokia Arkoumani, Miangela M Lacle, Max A Viergever, and Josien P W
Pluim. Assessment of algorithms for mitosis detection in breast cancer histopathology
images. Medical image analysis, 20:237–48, 02/2015 2015.
[38] Weidi Xie, J. Alison Noble, and Andrew Zisserman. Microscopy cell counting with fully
convolutional regression networks. Deep Learning Workshop in MICCAI, 2015.
[39] Yao Xue, Nilanjan Ray, Judith Hugh, and Gilbert Bigras. Cell counting by regression
using convolutional neural network. ECCV 2016 workshop on BioImage Computing,
2016.
29
| 1cs.CV
|
A NONAMENABLE TYPE F∞ GROUP OF PIECEWISE PROJECTIVE
HOMEOMORPHISMS.
arXiv:1408.3127v4 [math.GR] 21 May 2017
YASH LODHA
Abstract. We prove that the group of homeomorphisms of the circle introduced by the author in joint
work with Justin Moore (Groups, Geometry and Dynamics 2015) is of type F∞ . This provides the first
example of a type F∞ group which is nonamenable and does not contain non abelian free subgroups.
To prove our result we provide a certain generalisation of cube complexes, which we refer to as cluster
complexes. We also obtain a normal form, or a canonical unique choice of words for the elements of the
group.
Contents
1. Introduction
2
2. Proof Strategy
3
3. Preliminaries
4
3.1. Finiteness properties of groups.
4
3.2. Nonpositively curved cube complexes
5
3.3. Coset graphs
5
3.4. Hyperplane arrangements and the face complex
6
3.5. Binary sequences and the group F
6
3.6. The group G
7
4. Cluster complexes
11
5. Normal Forms
13
5.1. Step 1 of the procedure
14
5.2. Step 2 of the procedure
16
6. The 1-skeleton of X
18
6.1. The actions of F on Ω and X (1)
22
6.2. A detour on expansion moves and the expansion Lemmas
25
7. The complex X
28
7.1. Descriptions of graphs in H and the G-action.
29
7.2. H is closed under intersection
30
Date: January 30, 2018.
2010 Mathematics Subject Classification. Primary: 43A07; Secondary: 20F05.
Key words and phrases. finiteness properties, amenable, finitely presented, free group, piecewise, projective, Thompson’s
group, torsion free.
This research was supported in part by NSF grant DMS–1262019 at Cornell University and an EPFL-Marie Curie
postdoctoral fellowship.
1
2
YASH LODHA
7.3. Graphs in H as 1-skeletons of clusters
33
7.4. Higher dimensional cells in X
35
8. The action of G on X
36
8.1. Stabilisers of cells
36
8.2. The quotient X/G
39
9. X is simply connected.
39
10. X is aspherical
42
10.1. The parallel equivalence relation on the 1-cells of X
42
10.2. Orthogonal 1-clusters
45
10.3. Amplification of pre-clusters
46
10.4. Weakly balanced systems of pre-clusters
50
10.5. Balanced systems of clusters
54
10.6. Balanced systems and nonpositive curvature
58
10.7. Proof of asphericity
60
11. A question
60
References
61
1. Introduction
The von Neumann-Day problem asks whether every nonamenable group contains nonabelian free subgroups (see [21] and [11]). In 1980, Ol’shanskii solved the von Neumann-Day problem by producing a
counterexample [19]. Soon after, Adyan showed that certain Burnside groups are also counterexamples
[2, 3]. Ol’shanskii and Sapir constructed the first finitely presented counterexample in 2003 [20].
In [18] Monod produced a remarkable family of counterexamples. In particular he showed that the group
of piecewise PSL2 (R) homeomorphisms of RP1 that fix infinity is nonamenable and does not contain
nonabelian free subgroups. However, Monod’s examples are not finitely presentable.
In [16] we constructed a finitely presentable nonamenable subgroup of Monod’s group with 3 generators
and 9 relations. This group denoted by G is generated by a(t) = t + 1 together with the following two
homeomorphisms of R:
t
if t ≤ 0
(
t
1
2t
if
0
≤
t
≤
if 0 ≤ t ≤ 1
2
b(t) = 1−t 1
c(t) = 1+t
1
3
−
if
≤
t
≤
1
t
otherwise
t
2
t+1
if 1 ≤ t
The finiteness properties type Fn are natural topological generalisations of the properties of finite generation and finite presentability. A group is said to be of type F∞ if it is of type Fn for all n, or equivalently
if it is the fundamental group of a connected, aspherical CW complex with finitely many cells in every
dimension. The study of counterexamples to the von Neumann-Day problem has an interesting historical
connection with finiteness properties of groups. In the 1970s Thompson discovered a remarkable finitely
presented group, which is now known as Thompson’s group F . The elements a, b from above generate a
copy of F acting on the real line by piecewise projective homeomorphisms. In 1979 Geoghegan made the
following conjectures about F (See [13].)
A NONAMENABLE TYPE F∞ GROUP
(1)
(2)
(3)
(4)
3
F has type F∞ .
F has no nonabelian free subgroups.
F is nonamenable.
All homotopy groups of F at infinity are trivial.
Conjectures (1) and (4) were proved by Brown and Geoghegan [7]. Conjecture (2) was proved by Brin
and Squier [8]. The status of (3) still remains open. There is considerable interest in Conjecture (3)
especially because if F is nonamenable it would be an elegant counterexample to the von Neumann-Day
problem.
It is natural to inquire about the higher finiteness properties satisfied by the group G, especially in light
of Geoghegan’s conjectures about Thompson’s group F . In this article we establish that G is of type F∞ .
This makes the group G the first type F∞ group which is nonamenable, and does not contain nonabelian
free subgroups. More particularly, we prove the following.
Theorem 1.1. The group G acts on a connected cell complex X by cell permuting homeomorphisms such
that the following holds.
(1) X is contractible.
(2) The quotient X/G has finitely many cells in each dimension.
(3) The stabilizer of each cell is of type F∞ .
It follows that the group G is of type F∞ .
Our complex X is a special type of CW complex built out of CW subdivisions of Euclidean cubes.
Although there is no natural metric on such a complex, notions of metric and curvature shall play an
essential role in proving that the complex is contractible. Finally, we remark that it is observed in [23]
that all homotopy groups of G at infinity are trivial. So G satisfies all four statements of Geoghegan’s
conjecture for Thompson’s group F , and is the first such example.
2. Proof Strategy
In this section we provide an outline of the proof. Our complex X is a special type of CW complex, which
we call a complex of clusters. A cluster is a certain CW subdivision of a cube. An n-cluster is a CW
subdivision of an n dimensional cube [0, 1]n , which is obtained by considering the intersection pattern
with [0, 1]n ⊂ Rn of a family of hyperplanes of Rn . We give a precise definition of clusters in Section 4.
For instance, a 2-cluster can be a square, or a square with a diagonal 1-cell, and two triangular 2-cells
that fill it. The latter cluster is obtained from the intersection pattern of [0, 1]2 with the hyperplane
x = y in R2 . A sub-cluster of the latter 2-cluster is either a 0-cell, or one of the five 1-cells (including the
diagonal 1-cell).
To describe a complex of clusters, we draw an analogy with cube complexes. A cube complex is a CW
complex obtained by gluing cubes along their faces, i.e. gluing along facial “sub-cubes”. A complex of
clusters is a CW complex obtained by gluing clusters along sub-clusters. The key difference is that in
a cluster complex, sub-clusters could be placed diagonally in a cluster. For instance, we can take two
2-clusters described in the previous paragraph, and glue along the diagonal 1-cells. This way we obtain
a complex with four triangular 2-cells, or four triangles all glued along one edge.
To give the reader a more concrete idea of what our complex X looks like, we give a brief description of X.
A detailed description will appear in Sections 6 and 7. The 0-skeleton X (0) is defined as the set of right
cosets of F in G, on which G acts on the right. Two cosets F τ1 , F τ2 are connected by an edge in X (1)
if the double coset F τ1 τ2−1 F is one of four distinguished double cosets. It is then immediate from the
definition that G acts on X (1) in a way that preserves the edge relation, since F τ1 g(τ2 g)−1 F = F τ1 τ2−1 F .
In order to describe the cluster complex structure on X, we identify an infinite family H of finite, connected
subgraphs of X (1) , whose elements are in a natural way 1-skeletons of clusters. We demonstrate that H
4
YASH LODHA
is closed under finite intersections, i.e. the intersection of a finite number of graphs in H is also a graph
in H if it is non empty.
Then we “fill” the 1-skeleton of each graph in H by adding higher cells to obtain a cluster. We demonstrate
that these fillings are “compatible”, i.e. given two graphs Γ1 , Γ2 ∈ H and their respective fillings, the
filling of the intersection Γ1 ∩ Γ2 agrees with the induced fillings obtained from the restrictions of Γ1 , Γ2
respectively. Moreover, we demonstrate that this filling is G-equivariant. The union of X (1) together
with the fillings of graphs in H will define our complex of clusters X.
The structure of the paper is as follows. In Section 4 we define the notion of a cluster, and a cluster
complex. In Section 5 we construct a normal form for elements of G. This normal form will play a small
but essential role in the rest of the paper, but is of independent interest since it can be a useful tool to
study the group.
In Section 6 we define X (1) and study the action of G on X (1) . In Section 6, we shall also prove the
expansion lemmas, which shall be needed later in the paper, in particular during the proof of asphericity
of X. However, we prove these Lemmas here since their proofs shall emerge naturally from the ideas in
section 6. In Section 7, we define H and the higher cells of X
In Section 8 we prove that X/G has finitely many cells in every dimension, and that StabG (e) for each
cell e is a group of type F∞ . In fact, we demonstrate that the stabiliser of a cell is a product of HigmanThompson groups. In Section 9 we prove that X is simply connected, and in Section 10 we prove that
X is aspherical. Then we conclude that X is contractible. At the end of the article we state a question,
suggested to the author by Gromov, that could provide an interesting direction for further research.
A complex of clusters in general does not admit a natural piecewise Euclidean metric. For instance,
one cannot simply declare all the 1-cells to be isometric to [0, 1], since each 1-cell in our complex will
occur as a diagonal 1-cell of clusters of any given dimension. However, our proof that the complex is
aspherical shall involve notions of metric and curvature. More particularly, we demonstrate that every
finite subcomplex Y of X is contained in a sub complex Y of X such that Y is homeomorphic to a non
positively curved cube complex.
3. Preliminaries
In this section we will review some terminology needed later in the paper. Readers may wish to skim or
skip the material and refer back to it only as necessary. In this article, all actions shall be right actions.
However, when we make use of function notation we shall differ from this convention, i.e. for instance we
shall write x · f g = g(f (x)).
3.1. Finiteness properties of groups. The classical finiteness properties of groups are that of being
finitely generated and finitely presented. These notions were generalized by C.T.C. Wall [22]. In this
paper we are concerned with the properties type Fn . These properties are quasi-isometry invariants of
groups [1]. In order to discuss these properties first we need to define Eilenberg-Maclane complexes.
An Eilenberg-Maclane complex for a group G, or a K(G, 1), is a connected CW-complex X such that
e is contractible. It is a fact that for any group G, there is an Eilenberg-Maclane
π1 (X) = G and X
complex X which is unique up to homotopy type. A group is said to be of type Fn if it admits an
Eilenberg-Maclane complex with a finite n-skeleton. Clearly, a group is finitely generated if and only if
it is of type F1 , and finitely presented if and only if it is of type F2 . (For more details see [12].) A group
is said to be of type F∞ if it is of type Fn for all n ∈ N. Equivalently, a group is of type F∞ if it admits
a K(G, 1) with finitely many cells in each dimension.
The following is a special case of a well known result (See Proposition 1.1 in [6]. Note that the property
discussed there is type FP∞ , but since for the class of finitely presentable groups type FP∞ and type
F∞ are the same, we state this proposition in terms of type F∞ ).
Proposition 3.1. Let Γ be a finitely presentable group that acts on a cell complex X by cell permuting
homeomorphisms such that the following holds:
A NONAMENABLE TYPE F∞ GROUP
5
(1) X is contractible.
(2) The quotient X/Γ has finitely many cells in each dimension.
(3) The stabilizers of each cell are of type F∞ .
Then Γ is of type F∞ .
3.2. Nonpositively curved cube complexes. Although the complex we construct in this article is
not a cube complex, cube complexes will play an important role in the proof of contractibility of our
complex.
By a regular n-cube n we mean a cube which is isometric to the cube [0, 1]n in Rn . A cube complex is a
cell complex of regular Euclidean cubes glued along their faces by homeomorphisms. The metric on such
cube complexes is the piecewise Euclidean metric. (see [9] for details).
Given a vertex v of a regular cube n , the link Lk(v, n ) is the set of unit tangent vectors at v that
point in n . This is a subset of the unit sphere S n−1 which is homeomorphic to a simplex of dimension
n − 1. This admits a natural spherical metric, in which the dihedral angles are right angles.
Now let v be a 0-cell in a cube complex X and let
S
S = {C | C is a cube in X that contains v as a vertex }
The link Lk(v, X) = C∈S Lk(v, C). This is a complex of spherical “all right” simplices glued along their
faces by isometries. This admits a natural piecewise spherical metric. (see [9] for details).
A simplicial complex Z is called a “flag” complex if any set v1 , ..., vn of vertices of Z that are pairwise
connected by an edge span a simplex. This is also known as the “no empty triangles” condition. Now we
are ready to state the main definition.
Definition 3.2. A cube complex X is said to be nonpositively curved if the link of each vertex is a flag
complex.
For basic results concerning these spaces, we refer the reader to [9]. We shall only use the fact that
nonpositively curved cube complexes are aspherical, which is a corollary of the following theorem of
Gromov.
Theorem 3.3. (Gromov) A cube complex X is CAT(0) if and only if it is nonpositively curved and
simply connected.
This means that the universal cover of a nonpositively curved cube complex is CAT (0), and hence
contractible. So we have the following.
Corollary 3.4. Nonpositively curved cube complexes are aspherical.
3.3. Coset graphs. We shall use the notion of a coset graph. We remark that this notion may not be
familiar to a group theorist, although it is more likely familiar to a graph theorist. This is a natural
generalisation of Cayley graphs. Let G be a group and H be a subgroup of G. Let S be a finite set of
elements of G such that:
(1) S ⊆ G \ H.
(2) For each s ∈ S we have s−1 ∈ S.
(3) S ∪ H generates G.
Then we can form the so called coset graph Cos(G, H, S) as follows. The vertices of Cos(G, H, S) are the
right cosets of H in G, and two cosets Hg1 , Hg2 are connected by an edge if
[
g1 g2−1 ∈ H(S)H =
HsH
s∈S
Here
HsH = {h1 sh2 | h1 , h2 ∈ H}
6
YASH LODHA
is a double coset.
The group G acts on the graph Cos(G, H, S) on the right, and this
T action is vertex transitive. Moreover
the action of G on this graph is faithful if and only if the core g∈G g −1 Hg is trivial. By the following
theorem (see Theorem 3.8 in [17]), in fact every vertex transitive graph, i.e. a graph whose group of
automorphisms is vertex transitive, can be described in this way.
Theorem 3.5. [17] Let Ω be some vertex transitive graph. Then Ω is isomorphic to some coset graph
Cos(G, H, S).
The 1-skeleton of our complex X will emerge as a vertex transitive graph in this way.
3.4. Hyperplane arrangements and the face complex. Consider Rn endowed with the usual orthonormal basis, and with variables x1 , ..., xn representing coordinates in this basis. An affine hyperplane
is an (n − 1)-dimensional subspace comprising of solutions of an equation of the form
a1 x1 + ... + an xn = an+1
where ai ∈ R. The hyperplane itself is denoted as the set
{a1 x1 + ... + an xn = an+1 }
A finite hyperplane arrangement
A is a finite set of affine hyperplanes in Rn . A region of A is a connected
S
n
component of R \ H∈A H. Let R(A) denote the set of regions of A. It is an elementary exercise to
prove that every region is open, convex and thus homeomorphic to the interior of an n-dimensional ball.
The set of flats of the arrangement are the affine subspaces of Rn obtained by taking an intersection of
hyperplanes in A. The trivial intersection is also a flat, which equals Rn . Given a flat T , we denote the
hyperplane arrangement A ↾ T consisting of the set of hyperplanes
{T ∩ H | H ∈ A, T 6⊆ H}
in the subspace T . We define S
the regions of T in a similar way as above, i.e. R(A) ↾ T equals the set of
connected components of T \ H∈A↾T H.
The union of the sets of regions
[
R(A) ↾ T
T is a flat of A
is called
the face complex of the arrangement A. The face complex provides a cellular structure on
S
Rn ∂Rn = Bn .
3.5. Binary sequences and the group F . We will take N to include 0. Let 2N denote the collection of
all infinite binary sequences and let 2<N denote the collection of all finite binary sequences. For s ∈ 2<N ,
s(i) denotes the i’th digit of s. If i ∈ N and u is a binary sequence of length at least i, we will let u ↾ i
denote the initial part of u of length i. We denote by |s| the length of s, which is the number of digits in
s.
If s and t are finite binary sequences, then we will write s ⊆ t if s is an initial segment of t and s ⊂ t if s
is a proper initial segment of t. If neither s ⊆ t nor t ⊆ s, then we will say that s and t are independent.
A list of finite binary sequences s1 , ..., sn is said to be independent, if they are pairwise independent.
The set 2<N is equipped with an order defined by s < t if t ⊂ s or s and t are independent and s(i) < t(i)
where i is the smallest number such that s(i) 6= t(i). If u, s1 , ..., sn ∈ 2<N , we say that u dominates
s1 , ..., sn if for each 1 ≤ i ≤ n, either si , u are incompatible or si ⊂ u.
The finite binary sequences s, t are said to be consecutive if there is a binary sequence u and numbers
n1 , n2 ∈ N such that s = u01n1 and t = u10n2 . A list of finite binary sequences s1 , ..., sn is said to be
consecutive if each pair si , si+1 is consecutive for 1 ≤ i ≤ n − 1. Note that if s1 , ..., sn are consecutive
then they are automatically independent.
A NONAMENABLE TYPE F∞ GROUP
7
If ξ and η are infinite binary sequences, then we will say that ξ and η are tail equivalent if there are s, t,
and ζ such that ξ = sζ and η = tζ. We use 0∞ , 1∞ to denote the constant infinite sequences 000.... and
111... respectively. More generally, given a finite binary sequence s, s∞ denotes the sequence sss.....
We denote by T the collection of all finite rooted binary trees. A tree T in T will be denoted by a set of
finite binary sequences s1 , ..., sn which are the addresses of leaves in T . The indices are ordered so that
if i < j then si < sj . We view elements T of T as prefix sets. So we view T as a set of finite binary
sequences with the property that every infinite binary sequence has a unique initial segment in T .
A tree diagram is a pair (L, R) of elements of T with the property that |L| = |R|. A tree diagram
describes a map of infinite binary sequences as follows:
si ξ 7→ ti ξ
where si and ti are the ith elements of L and R, respectively, in the order as defined above and ξ is
any binary sequence. The collection of all such functions from 2N to 2N defined in this way, under the
operation of composition, is Thompson’s group F . The function associated to a tree diagram is also
defined on any finite binary sequence u such that u has a prefix in L. So the group F admits a partial
action on 2<N . Given f ∈ F and s ∈ 2<N , we say that f acts on s if f (s) is defined. Similarly, given
s1 , ..., sn ∈ 2<N we say that f acts on s1 , ..., sn if f (s1 ), ..., f (sn ) are all defined.
We direct the reader to the standard reference [4] for the definition and properties of Thompson’s group
F ; additional information can be found in [10]. We shall mostly follow the notation and conventions of
[16].
3.6. The group G. To prove that the group G is of type F∞ , the action of the group on the real line by
piecewise projective homeomorphisms does not appear to be useful. In [16] we describe a combinatorial
model for G by means of a faithful action of G by homeomorphisms of the cantor set 2N . This model
was used to prove that G is finitely presentable. This combinatorial model will be used throughout this
paper, and we shall not refer to the functions a, b, c : R → R which were defined in the introduction.
We recall the combinatorial description.
if ξ
0η
x(ξ) = 10η if ξ
11η if ξ
We start with the following two primitive functions:
= 00η
if ξ = 00η
0y(η)
y(ξ) = 10y −1 (η) if ξ = 01η
= 01η
= 1η
11y(η)
if ξ = 1η
From these functions, we define families of functions xs (s ∈ 2<N ) and ys (s ∈ 2<N ) which act just as x
and y, but localised to those binary sequences which extend s.
(
(
sx(η) if ξ = sη
sy(η) if ξ = sη
xs (ξ) =
ys (ξ) =
ξ
otherwise
ξ
otherwise
If s is the empty-string, it will be omitted as a subscript. Let
S = {xt , ys | s, t ∈ 2<N , s 6= 0k , s 6= 1k , s 6= ∅}
Our group is generated by functions in the set S. In fact, G is generated by x, x1 , y10 which correspond
respectively to the functions a, b, c that were defined in the introduction. The infinite generating set S is
much more desirable for the purpose of our proofs and will be used throughout this article. We now list
an infinite set of relations R satisfied by the generators in S.
(1)
(2)
(3)
(4)
(5)
If xs (t) is defined, then xt xs = xs xxs (t)
x2s = xs0 xs xs1 .
If xs (t) is defined, then yt xs = xs yxs (t) .
if s and t are independent, then ys yt = yt ys .
−1
ys = xs ys0 ys10
ys11 .
Note that in the above there is no occurrence of the functions ys± where s = 1k , s = ∅ or s = 0k . We
proved in [16] that:
8
YASH LODHA
Theorem 3.6. The group G ∼
= hS | Ri is finitely presented, nonamenable, and does not contain nonabelian free subgroups.
The generators {xs | s ∈ 2<N } together with the relations
(1) If xs (t) is defined, then xt xs = xs xxs (t) .
(2) x2s = xs0 xs xs1
describe an infinite presentation for Thompson’s group F .
Recall that in [10]
−1
−1
{x−1 , x−1
1 , x11 , x111 , ...} ⊆ X
is used as the infinite generating set for F . We now recall the normal form for Thompson’s group F
in this generating set (See Corollary-Definition 2.7 in [10]). We rephrase the normal form to suit our
conventions, which differ from those used in [10]. We shall refer to this normal form as the F -normal
form.
Theorem 3.7. Every non-trivial element of F can be expressed in unique normal form:
−bn an
a1 a0
1 −b2
x−b0 x−b
1 x12 ...x1n x1n ...x1 x
such that:
(1) a0 , ..., an and b0 , ..., bn are nonnegative numbers.
(2) Exactly one of an , bn is nonzero.
(3) If for some 0 ≤ k ≤ n − 1 it is true that ak , bk > 0, then ak+1 > 0 or bk+1 > 0.
Furthermore, every such normal form word is nontrivial in F .
Now we recall from [16] the following definitions.
Definition 3.8. An X-word is a word in the letters
{xts | s ∈ 2<N , t ∈ Z}
A Y -word is a word in the letters
{yst | s ∈ 2<N , s 6= 0k , s 6= 1k , s 6= ∅, t ∈ Z}
An S-word is a word in the generators
{yst , xvu | s, u ∈ 2<N , s 6= 0k , s 6= 1k , s 6= ∅, and t, v ∈ Z}
The elements of the set
{yst | s ∈ 2<N , s 6= 0k , s 6= 1k , s 6= ∅, t ∈ Z}
are called percolating elements.
Definition 3.9. An S-word f yst11 ...ystnn is in standard form if it is the concatenation of an X-word f
followed by a Y -word yst11 ...ystnn with the property that si < sj if i < j. We will write standard form to
mean an S-word in standard form.
It is convenient to denote a standard form as f λ, where f is an X-word and λ is a Y -word in standard
form. We will mostly use the convention of denoting X-words by letters f, g, h, often with subscripts,
and Y -words in standard form with either greek letters, or explicitly as yst11 ...ystnn .
The depth of f yst11 ...ystnn equals inf {|si | | 1 ≤ i ≤ n}. We use the convention that standard forms that are
X-words have infinite depth.
The following Lemma was proved in [16].
Lemma 3.10. If Ω is any S-word and l ∈ N, then Ω can be converted into a standard form f λ using the
relations, so that f λ has depth at least l.
A NONAMENABLE TYPE F∞ GROUP
9
We also define a slightly weaker notion than a standard form, which will be useful in phrasing our
arguments.
Definition 3.11. An S-word f yst11 ...ystnn is in standard form upto commutation if it is the concatenation
of an X-word f followed by a Y -word yst11 ...ystnn with the property that if sj ⊂ si then i < j.
Note that this notion is slightly weaker than the standard form. For instance y100 y101 is in standard form
but not y101 y100 , although the latter is in standard form upto commutation.
The definition implies the following elementary fact.
Lemma 3.12. A standard form upto commutation can be converted into a standard form by the application of a finite sequence of commuting relations, which are relations (4) of R.
Definition 3.13. Associated with a standard form (or a standard form up to commutation) f yst11 ...ystnn
and a sequence σ ∈ 2ω is the notion of a calculation. This is an infinite string in letters y, y −1 , 0, 1. The
evaluation of f yst11 ...ystnn on σ comprises of a prefix replacement that is determined by the transformation
f ∈ F followed by an infinite sequence of applications of the transformations described by the percolating
elements. The latter is encoded as an infinite string with letters y, y −1 , 0, 1 and denoted as the calculation
of yst11 ...ystnn on f (σ). The calculation is equipped with the following substitutions:
y00 → 0y
−1
−1
y01 → 10y −1
y1 → 11y
−1
y 0 → 00y
y 10 → 01y
y −1 11 → 1y −1
−1
For example, for the word y100 y10 and the binary sequence 1001111... the calculation string is 10y0y −11111....
The output of the evaluation of the word on the binary string is the limit of the strings obtained from
performing these substitutions. In this way, we define the calculation of a standard form (up to commutation) yst11 ...ystnn on an infinite binary sequence σ. The calculation of a standard form (up to commutation)
f yst11 ...ystnn on an infinite binary sequence σ is defined to be the calculation of yst11 ...ystnn on f (σ).
The set of all such strings will be denoted as {0, 1, y, y −1}N . A calculation is an element of this set with
the property that there are only finitely many occurrences of y ± .
A calculation has a potential cancellation if upon performing a finite set of substitutions we encounter
a substring of the form yy −1 or y −1 y. When a calculation has no potential cancellations, we say that
it has exponent n if n is the number of occurrences of the symbols y ± . Note that there is no potential
cancellation in the example above, and the exponent is 2.
The following was proved in [16]. (Lemma 5.9.)
Lemma 3.14. Suppose that Λ is in {0, 1, y, y −1}N and contains no potential cancellations. Then advancing any occurrence of a y ± results in a word with no potential cancellations.
It is also natural to consider finite strings in letters y, y −1 , 0, 1, equipped with the same substitutions
as above. Of course, for such strings one may not be able to advance the rightmost occurrence of y ±1 .
However, such strings will be useful to consider in the following context. Let f yst11 ...ystnn be a standard
form and let u be a finite binary sequence such that f acts on u and f (u) dominates s1 , ..., sn . Given
any infinite binary sequence ψ, we know that f (uψ) = f (u)ψ. Also, the calculation of f yst11 ...ystnn on uψ,
which equals the calculation of yst11 ...ystnn on f (u)ψ, contains a tail ψ. Upon deleting this tail from this
calculation, we obtain a calculation of f yst11 ...ystnn on u. Note that the choice of the sequence ψ does not
affect the definition of this calculation. We denote the set of all such strings as {0, 1, y, y −1}<N . The
following lemma was proved in [16] (Lemma 5.10.)
Lemma 3.15. Suppose that Λ is in {0, 1, y, y −1}<N and contains no potential cancellations, and contains
n occurrences of y ± . Then there is a finite binary sequence u such that the following holds. We can
perform a sequence of substitutions on the calculation Λu to obtain a calculation of the form vy n , where
v is a finite binary sequence.
We now describe a family of basic operations involving standard forms. Notice that each of these substitutions corresponds either to a relation in R or to a group-theoretic identity.
10
YASH LODHA
Definition 3.16. We will say that an S-word Ω1 is derived from an S-word Ω0 , written symbolically as
Ω0 → Ω1 , if it is the result of applying moves of the following forms:
±1 i
<N
(1) (Rearranging) yti x±1
are such that t.xs is defined.
s → xs yx±1 (t) where s, t ∈ 2
s
(2)
(3)
(4)
(5)
(6)
−1
−1
−1
<N
(Expansion) ys → xs ys0 ys10
ys11 and ys−1 → x−1
.
s ys00 ys01 ys1 where s ∈ 2
N
(Commuting) yu yv ↔ yv yu where u, v ∈ 2 and u, v are independent.
xi+j ↔ xi xj
y i+j ↔ y i y j
(Cancellation) Delete an occurrence of y i y −i
−1
The percolating elements ys0 , ys10
, ys11 obtained upon performing the expansion move on ys are called
−1
−1
offsprings of ys . Similarly, the elements ys00
, ys01 , ys1
obtained upon performing an expansion move
−1
−1
on ys are offsprings of ys . If we perform a sequence of such moves, we will also refer to offsprings
−1
of offsprings are offsprings themselves. For instance, consider the move ys → xs ys0 ys10
ys11 produces
−1
−1
offsprings ys0 , ys10 , ys11 of ys . Now upon performing a subsequent expansion move on ys10 , we obtain
−1
−1
−1
−1
−1
xs ys0 x−1
s10 ys1000 ys1001 ys101 ys11 = xs xs10 ys0 ys1000 ys1001 ys101 ys11
−1
−1
So we say that ys0, ys1000
, ys1001 , ys101
, ys11 are offsprings of ys . In our arguments is shall be useful to
keep track of the offsprings of percolating elements under expansion.
An important combination of moves that will be used frequently in our arguments is expansion followed
by rearrangement. For example, let f (yst11 ....ystnn )yu be an S-word such that xu acts on each sequence in
the set {s1 , ..., sn }. We can perform an expansion move on the word to obtain
−1
yu11 )
f (yst11 ....ystnn )(xu yu0 yu10
followed by rearranging moves to obtain
−1
f xu (yst1′ ....ystn′n )(yu0 yu10
yu11 )
1
s′i
where = xu (si ). It is easy to check that if our original word was a standard form, then the resulting
word is also a standard form. This combination of moves shall be used frequently in the proofs in the
paper. We now extend the notion of potential cancellation for a calculation to an analogous notion for
standard forms.
Definition 3.17. (Potential cancellation) A standard form f yst11 ...ystnn is said to have a potential cancellation if there is an infinite binary sequence τ such that the calculation of yst11 ...ystnn on τ contains a potential
cancellation.
Lemma 3.18. Let f λ1 be a standard form which does not have a potential cancellation. Let gλ2 be a
standard form obtained by performing a sequence of expansion followed by rearranging moves on f λ1 .
Then gλ2 does not have a potential cancellation.
Proof. Let λ1 = yst11 ...ystnn , such that ti ∈ {1, −1}. Note that this is for notational convenience, i.e. for
instance we write y 2 as yy. Since the existence of potential cancellations in f yst11 ...ystnn is not affected by
f ∈ F , we assume that this is the empty word.
Assume that xsi acts on each s1 , ..., si−1 , and that ti = 1. The case when ti = −1 is completely analogous.
Applying expansion followed by rearrangement, we obtain
...ystnn )
y
)(ysti+1
)(xsi ysi 0 ys−1
(yst11 ...ysti−1
i+1
i−1
i 10 si 11
t
...ystnn )
= (xsi )(yst1′ ...ysi−1
y
)(ysti+1
)(ysi 0 ys−1
′
i+1
i 10 si 11
i−1
1
where s′j = xsi (sj ) for 1 ≤ j ≤ i − 1.
Now assume that we have introduced a potential cancellation after performing this move. Let τ be an
infinite binary sequence for which the associated calculation of
t
...ystnn )
(xsi )(yst1′ ...ysi−1
y
)(ysti+1
)(ysi 0 ys−1
′
i+1
i 10 si 11
1
i−1
A NONAMENABLE TYPE F∞ GROUP
11
denoted by Λ contains a potential cancellation. It must be the case that τ contains as a prefix a sequence
which equals either si 0, si 10, or si 11.
Now let Λ′ be the calculation produced by the evaluation of yst11 ...ystnn on x−1
si (τ ). The calculation Λ is in
fact obtained from Λ′ by advancing a symbol y. By our hypothesis, there is no potential cancellation in
Λ′ . So by Lemma 3.14 there cannot be a potential cancellation in Λ. Hence we obtain a contradiction.
Therefore, upon performing a sequence of expansion followed by rearrangement moves, we obtain a
standard form with no potential cancellations.
We end this subsection with a discussion about the action of standard forms and tail equivalence. The
action of F on infinite binary sequences preserves tail equivalence, yet the elements ysn do not. This
is a fundamental and important distinction between the groups F and G. The advantage of standard
forms without potential cancellations is that we can identify precisely where the relation is not preserved.
Recall that the support of a homeomorphism φ is the set of elements in the domain that are moved by
the homeomorphism. We denote this as Supp(φ).
Lemma 3.19. Let f yst11 ...ystnn be a standard form that does not contain potential cancellations. Let U ⊂ 2N
be the support of yst11 ...ystnn .
(1) f yst11 ...ystnn does not preserve tail equivalence on a dense subset of f −1 (U ). More particularly, given
an open set U ′ in f −1 (U ), there is an infinite binary sequence τ ∈ U ′ such that f yst11 ...ystnn (τ ) and
τ are not tail equivalent.
(2) f yst11 ...ystnn preserves tail equivalence on 2N \ f −1 (U ).
Proof. It suffices to prove the statement in the case when f is trivial, since f maps f −1 (U ) to the support
of the Y -word yst11 ...ystnn , and outside f −1 (U ) the word f yst11 ...ystnn acts like an element of F , and hence
preserves tail equivalence.
Let U ′ be an open subset of U . Let τ be an finite binary sequence such that the following holds:
(1) [τ 0∞ , τ 1∞ ] ⊂ U ′
(2) τ dominates s1 , ..., sn .
The associated calculation Λ of yst11 ...ystnn on τ does not contain potential cancellations, by our hypothesis.
By Lemma 3.15, there is a finite binary sequence u such that one can perform substitutions on Λu to
obtain vy m where v is a finite binary sequence and m 6= 0. We assume that m > 0, the other case is
m
m
analogous. It follows that the associated calculation of yst11 ...ystnn on τ u02 102 1... equals
m
m
m
m
vy m 02 102 1...
This calculation produces the output
v012 012 ...
which is not tail equivalent to the input.
4. Cluster complexes
We consider Rn with variables x1 , ..., xn that represent coordinates in the usual orthonormal basis. In
particular, we fix an order on the set of coordinates, described by the indices, i.e. xi < xj if i < j. In
this section, we shall use the concepts concerning hyperplane arrangements described in Subsection 3.4
from the Preliminaries.
We consider two types of affine hyperplanes in Rn .
(1) An affine hyperplane of Rn is said to be of type 1 if it is of the form {xi = 0} or {xi = 1} for
some 1 ≤ i ≤ n.
(2) A hyperplane of Rn is said to be of type 2 if it is of the form {xi = xi+1 } for some 1 ≤ i ≤ n − 1.
12
YASH LODHA
Let A be a hyperplane arrangement in Rn . We say that A is admissible if the following holds.
(1) Each hyperplane in A is of type 1 or of type 2.
(2) A contains all the hyperplanes of type 1.
Now we define the notion of an n-cluster. Recall the notion of the face complex of a finite hyperplane
arrangement from subsection 3.4.
Definition 4.1. An n-cluster is a CW subdivision of [0, 1]n obtained by restricting the face complex of
an admissible hyperplane arrangement to [0, 1]n .
Let A be an admissible hyperplane arrangement in Rn . We denote by C(A) as the restriction of the face
complex of A to [0, 1]n , which is a subcomplex. In particular, we say that C(A) is the n-cluster (or simply
cluster) associated with A.
Hence we have a correspondence:
{admissible hyperplane arrangements in Rn } ↔ {n-clusters}
In this correspondence, the admissible arrangement A given by the set of all type 1 hyperplanes corresponds to the standard CW structure of the regular Euclidean cube [0, 1]n . If A consists of all type 1
hyperplanes of R2 and the type 2 hyperplane {x1 = x2 }, then the resulting 2-cluster is a square with a
diagonal connecting (0, 0) and (1, 1). This has four 0-cells, five 1-cells, and two triangular 2-cells.
Our next step is to define the notion of a subcluster of a cluster. First we shall define a codimension 1
subcluster, and then subsequently define lower dimensional clusters.
Definition 4.2. Let A be an admissible hyperplane arrangement, and let C(A) ⊂ Rn be the associated
n-cluster. A codimension 1 subcluster of C(A) is obtained by taking the intersection of a hyperplane H
in A with C(A).
A codimension 1 subcluster automatically inherits a definition in terms of an admissible hyperplane
arrangement in Rn−1 . To see this, let H ∈ A be a hyperplane. Now H is of one of the following forms
{xk = xk+1 }
{xk = 1}
{xk = 0}
In each respective case, we endow H with a set of coordinates y1 , ..., yn−1 where yj = xj if j < k and
yj = xj+1 if k ≤ j ≤ n − 1. The restriction A ↾ H provides an admissible hyperplane arrangement for
H∼
= Rn−1 .
This generalises to the following notion of subclusters.
Definition 4.3. A subcluster of C(A) is a subcomplex obtained by an intersection of C(A) with a flat T
in A.
We categorize subclusters of a given cluster as facial and diagonal. This is done
T by separating the flats
of A into two types. We say that a flat T of A is of type 1, if it is of the form H∈A H, where A ⊆ A is
a collection of type 1 hyperplanes.
T
We say that a flat T of A is of type 2, if is of the form H∈A H, where A ⊆ A has the following property.
A contains at least one hyperplane of the form {xk = xk+1 } so that T is not contained in any of the
following hyperplanes:
{xk = 0}
{xk = 1}
{xk+1 = 0}
{xk+1 = 1}
Definition 4.4. Consider a subcluster obtained by taking an intersection of the cluster with a flat T .
If T is a type 1 flat, then the subcluster is said to be a facial subcluster. If T is a type 2 flat, then the
subcluster is said to be a diagonal subcluster.
The following Lemma is a corollary to a similar statement concerning the face complex.
A NONAMENABLE TYPE F∞ GROUP
13
Lemma 4.5. Each cell e of a cluster is the interior of a convex polytope. The set of 0-cells incident to
e is convex independent, i.e. no one element lies in the convex span of the others. Moreover, the closure
of e equals the convex span of the 0-cells incident to e.
The notion of a nonpositvely curved cube complex admits a natural generalisation in this setting, despite
the absence of a natural metric. One may view this notion as a combinatorial criterion that guarantees
asphericity.
Definition 4.6. Let C be a complex of clusters. We say that C is nonpositively curved if for every finite
subcomplex Y of C, there is a nonpositively curved cube complex Z, and continuous maps f : Y → Z, g :
Z → C so that the following diagram commutes:
Z
f
Y
g
C
The following is an immediate consequence of the definition.
Proposition 4.7. If C is a nonpositively curved cluster complex, then C is apsherical.
We shall demonstrate that the complex X we construct in this article is nonpositively curved in this
sense.
5. Normal Forms
For our group presentation G = hS, Ri, we shall describe a unique, canonical choice of word for each
group element. We shall describe a procedure that converts a given word into such a word. The normal
form described in this section (or slightly weaker variants of it) will be used throughout the paper, in
particular in the proof that the complex X is contractible. However, aside from the goals of this paper,
this normal form is a useful tool to study the group and is therefore of independent interest. The reader
is encouraged to develop a technical familiarity of the moves described in 3.16, before reading the proofs
in this section.
To help phrase the arguments, it is useful to imagine a Y -standard form yst11 ...ystnn as a set of decorations
on the infinite rooted binary tree. This is done by means of adding integer labels ti on the nodes si . The
two central notions in this section are potential cancellation and potential contraction. In particular, our
normal form will be a standard form f yst11 ...ystnn such that f is in the usual normal form for Thompson’s
group F , and yst11 ...ystnn does not contain any potential cancellations or potential contractions.
We have already defined the notion of potential cancellations. We shall now provide a reformulation of
the definition that will be useful in the proofs in this section.
t
Definition 5.1. For a standard form f yst11 ...ystnn , we say that the pair ysjj , ystii is adjacent if si ⊂ sj and
for each sequence u satisfying si ⊂ u ⊂ sj , u does not equal sk for any 1 ≤ k ≤ n. This means that if we
walk along the geodesic in the tree connecting si and sj , we do not encounter any labels besides those at
the first and the last nodes.
Such an adjacent pair is said to be a potential cancellation if the associated calculation of the subword
t
ysjj ystii on some (or any) infinite binary sequence that extends sj contains a potential cancellation.
Lemma 5.2. A standard form contains a potential cancellation if and only if it contains an adjacent
pair that is a potential cancellation.
Proof. A standard form f yst11 ...ystnn contains a potential cancellation if there is an infinite binary sequence
σ such that the associated calculation Ω of yst11 ...ystnn on σ contains a potential cancellation. This means
that Ω contains a substring of the form y ti uy tj , where u ∈ 2<N and so that advancing an occurrence of
14
YASH LODHA
y ±1 from y ti to y tj along u produces a cancellation. Conversely, the same cancellation is witnessed when
t
t
processing the calculation of the subword ystii ysjj on σ, where ystii , ysjj is an adjacent pair in yst11 ...ystnn with
a potential cancellation.
Definition 5.3. A standard form is said to contain a potential contraction if either of the following holds.
−1
±
(1) It contains an occurrence of a subword of the form ys0 ys10
ys11 but no occurrences of ys1
.
−1
−1
±
(2) It contains an occurrence of a subword of the form ys00 ys01 ys1 but no occurrences of ys0
.
−1
If a standard form f yst11 ...ystnn contains a potential contraction of the form ys0 ys10
ys11 (in particular, no
±
−1
v
occurrences of ys1 ), we can replace the subword ys0 ys10 ys11 with the word x−1
y
.
s If yu is a percolating
s
element occurring to the left of x−1
in
the
resulting
word,
then
either
u,
s
are
independent
or s ⊂ u. Since
s
±
there are no occurrences of ys1
, the element x−1
must
act
on
u.
So
we
apply
rearranging
substitutions
s
to move the x−1
to
the
left
of
all
percolating
elements.
The
resulting
word
is
a
standard
form.
We call
s
−1
−1
this a contraction move. The contraction move for the other case, i.e. subwords of the form ys00
ys01 ys1
±
(but no occurrences of ys0 ), is defined in a similar way.
The procedure for converting a standard form f yst11 ...ystnn into a normal form using moves involves three
steps:
that does
(1) Step 1: Using a sequence of moves, convert f yst11 ...ystnn into a standard form f ′ yuv11 ...yuvm
m
not contain potential cancellations.
(2) Step 2: Using a sequence of contraction moves, convert f ′ yuv11 ...yuvm
into a standard form
m
f ′′ yrq11 ...yrqkk which does not contain potential contractions. Moreover, the moves performed do
not introduce any potential cancellations.
(3) Step 3 Convert f ′′ into a word g which is in F -normal form as described in Theorem 3.7.
Output: A standard form gyrq11 ...yrqkk such that g is in F -normal form as in 3.7, and yrq11 ...yrqkk does not
contain any potential cancellations or potential contractions. This word shall be the unique normal form
as stated in the main theorem below.
Theorem 5.4. For each element in G, there is a unique word f yst11 ...ystnn such that
(1) f yst11 ...ystnn is a standard form with no potential contractions or potential cancellations.
(2) f ∈ F is an X-word in F -normal form in the sense of 3.7.
Moreover, given any word that represents this element in G, we can convert it into the word f yst11 ...ystnn
using a sequence of moves from 3.16.
The goal of the rest of this section is to describe Steps 1, 2 of the procedure and then prove the main
theorem. For Step 3 we refer the reader to [10].
5.1. Step 1 of the procedure. First we make a fundamental observation concerning potential cancellations in standard forms. This will provide a part of the reduction in Step 1.
Lemma 5.5. Let (yst11 ...ystnn )yst be a standard form such that the following holds.
(1) t1 , ..., tn , t ∈ {1, −1}.
(2) The sequences s1 , ..., sn are independent.
(3) Each pair ystii , yst is an adjacent pair that is a potential cancellation.
Then we can perform a sequence of expansion, rearrangement and cancellation moves on this word to
produce a standard form f yuv11 ...yuvm
such that u1 , ..., um are independent and v1 , ..., vm ∈ {1, −1}.
m
Proof. We assume that t = 1, the other case is completely analogous. Performing an expansion move on
−1
ys produces a word xs ys0 ys10
ys11 . We claim that xs acts on each sequence si for 1 ≤ i ≤ n, unless s = si
−ti
and ys = ysi .
A NONAMENABLE TYPE F∞ GROUP
15
Let Λ be the associated calculation of the action of (yst11 ...ystnn )ys on the sequence si u for any choice of
infinite binary sequence u. (The choice of such a sequence does not affect the argument.) There are only
two occurrences of y ±1 in Λ. The first corresponds to ys and the other to ystii . Advancing the symbol
y in Λ that corresponds to ys involves acting by x on the suffix of sys in Λ. Since there is a potential
i
. Therefore, xs acts on each
cancellation in Λ, we can advance this symbol, unless s = si and ys = ys−t
i
i
si , unless s = si and ys = ys−t
.
i
Applying the rearrangement move, we obtain
−1
xs (yst1′ ...ystn′n )(ys0 ys10
ys11 )
1
where
s′i
= xs (si ).
Now each sequence s′i contains as an initial segment exactly one of the three sequences s0, s10, s11,
and is independent from the remaining two. Also, note that the distance in the tree between s′i and
the prefix of s′i in {s0, s10, s11} is smaller than the distance between si and s. In this way, we can
−1
continue to apply expansion followed by rearrangement moves on ys0 , ys10
, ys11 and their offsprings, until
this distance is 0 for each such pair. Then we can perform the required cancellations. At this stage, any
remaining percolating elements yuv11 , ..., yuvnn are offsprings of ys and hence have the property that u1 , ..., un
are pairwise independent.
Now we describe the final part of the first step.
Lemma 5.6. Any standard form can be converted to a standard form with no potential cancellations
using the relations.
Proof. We prove this by induction on the number of percolating elements in the standard form. The base
case is trivial. Let this be true for n − 1. Let f yst11 ...ystnn be a standard form such that ti ∈ {1, −1}. We
remark that the condition for ti ’s is a matter of convenient notation, i.e. for example we write ys2 as ys ys .
t
By the inductive hypothesis it follows that we can convert f yst11 ...ysn−1
n−1 into a standard form with no
potential cancellations using the relations. Moreover, from applying Lemmas 3.10 and 3.18 it follows that
with no potential cancellations and depth
this can further be converted into a standard form hypq11 ...ypqm
m
tn
is
a
standard
form up to commutation, since none
y
strictly larger than |sn |. It follows that hypq11 ...ypqm
m sn
of the sequences pi are initial segments of sn .
Let τ be an infinite binary sequence for which the associated calculation Λ is a potential cancellation.
The potential cancellation in this calculation must occur between the first two occurrences of y ± in τ ,
since hypq11 ...ypqm
does not have any potential cancellation. Such an occurrence of y ± corresponds to a
m
potential cancellation of an adjacent pair of the form ypqii , ystnn for some 1 ≤ i ≤ m.
y tn into
Now we make a record of all such adjacent pairs. Using commuting moves we convert hypq11 ...ypqm
m sn
k1
kl
v1
vo tn
a word h(yr1 ...yrl )(yu1 ...yuo )ysn such that
(1)
(2)
(3)
(4)
The word h(yrk11 ...yrkll )(yuv11 ...yuvoo )ystnn is a standard form up to commutation.
v1 , ..., vo ∈ {1, −1}.
v
Each pair yujj , ystnn is an adjacent pair that is a potential cancellation.
v
No adjacent pair of the form yrkii , ystnn or yrkii , yujj is a potential cancellation.
Now we apply a sequence of expansion and rearranging moves to convert the standard form hyrk11 ...yrkll
into a standard form h′ λ1 so that:
(1) h′ ∈ F , λ1 is a Y -word.
(2) h′ λ1 is a standard form of depth at least max{|u1 |, ..., |uo |} + 3.
Upon substitution, we obtain the word
h′ λ1 (yuv11 ...yuvoo )ystnn
16
YASH LODHA
Conditions (1) − (4) still hold for this new word, with λ1 replacing (yrk11 ...yrkll ) in the respective parts.
By Lemma 5.5, we can perform a sequence of expansion, rearranging and cancelling substitutions on the
word (yuv11 ...yuvoo )ystnn to obtain a standard form gλ2 so that g ∈ F , λ2 is a Y -word and gλ2 has no potential
cancellations. The proof of Lemma 5.5 guarantees the following:
(1) Any percolating element yst that occurs in λ2 satisfies that |s| ≤ max{|u1 |, ..., |uo |} + 1.
(2) For any percolating element yst that occurs in λ1 , g acts on the sequence s.
Upon substitution, we obtain a word h′ λ1 gλ2 . Let λ3 be the standard form obtained by replacing each
t
percolating element yst of λ1 by yg(s)
. We obtain
h′ λ1 gλ2 = h′ gλ3 λ2
Thanks to the condition on the depths of λ1 and λ2 , the word (h′ g)λ3 λ2 is a standard form up to
commutation. Thanks to condition (4) above and Lemma 3.14, this word does not contain an adjacent
pair that is a potential cancellation. The conditions ensure that this is also a standard form up to
commutation. After applying a sequence of commutation moves, we obtain a standard form with no
potential cancellations.
5.2. Step 2 of the procedure.
Lemma 5.7. Any standard form can be converted into a standard form with no potential contractions
and no potential cancellations.
Proof. By Lemma 5.6 we can convert our standard form into a standard form with no potential cancellations. On the resulting standard form, we perform contraction moves (defined in 5.3), one by one, until
no contraction moves can be performed. This process terminates, since upon performing a contraction
substitution we obtain a standard form which has fewer occurrences of y ± . Moreover, thanks to Lemma
3.14 we do not introduce any potential cancellations in this process.
Before we proceed to prove the main theorem, we shall make a fundamental observation concerning
exponents of calculations. (Recall the Definition of exponents of calculations from 3.13.) This provides
an invariant that distinguishes different standard forms.
Lemma 5.8. Let f yst11 ...ystnn and gypq11 ...ypqm
be standard forms that represent the same group element. Let
m
u be a finite binary sequence such that:
(1) f (u), g(u) are defined and dominate s1 , ..., sn and p1 , ..., pm respectively.
(2) Let Θ and Λ be calculations of yst11 ...ystnn on f (u) and ypq11 ...ypqm
on g(u) respectively, such that Θ
m
and Λ do not admit potential cancellations.
Then the exponents of Θ and Λ are the same.
Proof. Assume without loss of generality that the exponent of Λ is greater than that of Θ. Assume that
k > 0 is the exponent of Λ. (Recall that the exponent is always nonnegative.) By Lemma 3.15, there is a
finite binary sequence v such that we can perform moves on Λv to obtain a calculation of the form wy k ,
where w is a finite binary sequence.
k
k
k
k
k
Now consider the sequence uv02 102 102 1.... We know that
k
k
k
k
(uv02 102 102 1...)
f yst11 ...ystnn (uv02 102 102 1...) = gypq11 ...ypqm
m
In particular, we know that the calculations
k
k
k
Θv02 102 102 1...
and
k
k
k
k
k
k
k
k
k
Λv02 102 102 1... = wy k 02 102 102 1... = w012 012 012 ...
produce the same output.
A NONAMENABLE TYPE F∞ GROUP
k
k
17
k
The calculation Θv02 102 102 1... has fewer than k occurrences of y ± . It is a pleasant visual exercise for
the reader that this calculation cannot produce an output that is tail equivalent to a sequence of the form
k
k
k
w012 012 012 .... This means that our assumption that the exponents are different must be false.
Proof of Theorem 5.4
Proof. From Lemmas 5.6 and 5.7, it follows that for any given group element we can find a word satisfying
the hypothesis of the theorem. It remains to show that this is unique. By way of contradiction assume
that
gypq11 ...ypqm
f yst11 ...ystnn
m
are two distinct standard forms that satisfy conditions (1) and (2) of the Theorem, and represent the
same element of G. Assume without loss of generality that pm ≤ sn and that qm > 0.
There are three cases:
(1) sn = pm and either tn , qm < 0 or tn , qm > 0.
(2) sn = pm and either tn < 0 < qm or qm < 0 < tn .
(3) sn > pm .
Case (1): We can perform a cancellation on both sides of the equality, until we land in one of the remaining
cases.
Now we shall demonstrate that in both case (2) and (3), the following holds.
Claim: There is a finite binary sequence u such that:
a. pm ⊆ u and u dominates s1 , ..., sn .
b. The associated calculation Λ of (yst11 ...ystnn )yp−1
on u does not contain a potential cancellation.
m
Note that if u is a sequence that satisfies the above, then any sequence that extends u also satisfies the
above.
does not
Proof of the claim for case (2): In this case it is obvious that the standard form (yst11 ...ystnn )yp−1
m
admit any potential cancellations. This means that any finite binary sequence u satisfying condition a in
the claim also satisfies condition b in the claim.
Proof of the claim for case (3): Let us assume that for each binary sequence u that satisfies condition a
of the claim, condition b fails. Since yst11 ...ystnn does not contain any potential cancellations, the associated
on u must contain a potential cancellation that corresponds to an adjacent
calculation of (yst11 ...ystnn )yp−1
m
−1
ti
pair of the form ysi , ypm .
This means that there is a set of indices k1 , ..., kl ∈ {1, ..., n} such that:
(1) sk1 , ..., skl are leaves of a finite binary tree rooted at pm .
tk
(2) Each pair yskii , yp−1
that is a potential cancellation.
is an adjacent pair in (yst11 ...ystnn )yp−1
m
m
t
tk
In this situation, one can perform a sequence of contraction moves on the subword yskk11 ...yskll of yst11 ...ystnn
to obtain the subword ypm . This means that the standard form yst11 ...ystnn admits a potential contraction,
contradicting our hypothesis. So the claim holds in case (3).
Now extend the finite binary sequence u, if necessary, so that the following also holds. f −1 (u) and
g(f −1 (u)) are defined, and g(f −1 (u)) dominates p1 , ..., pm . (This is merely a matter of choosing a
sufficiently long sequence that extends u.)
Let σ = f −1 (u). Recall from the above the following.
on f (σ).
(1) Λ is the calculation of (yst11 ...ystnn )yp−1
m
We consider the following additional calculations:
18
YASH LODHA
−1
on g(σ).
(2) Θ is the calculation of ypq11 ...ypqm
m
′
t1
tn
(3) Λ is the calculation of ys1 ...ysn on f (σ).
(4) Θ′ is the calculation of ypq11 ...ypqm
on g(σ).
m
According to our assumptions, the calculations Λ, Λ′ , Θ, Θ′ do not contain any potential cancellations.
First observe that:
(1) The exponent of Λ is the same as that of Θ.
(2) The exponent of Λ′ is the same as that of Θ′ .
This follows from Lemma 5.8 since
−1
(σ)
(σ) = gypq11 ...ypqm
f (yst11 ...ystnn )yp−1
m
m
and
(σ)
f yst11 ...ystnn (σ) = gypq11 ...ypqm
m
However, it also holds that:
(1) The exponent of Λ is greater than that of Λ′ .
(2) The exponent of Θ is less than or equal to that of Θ′ .
This is a contradiction. Therefore, the words
yst11 ...ystnn
ypq11 ...ypqm
m
must be equal. Upon cancellation, we obtain that f = g, and since these words are in the normal form
for F , they must also be equal. Therefore, the words
f yst11 ...ystnn
gypq11 ...ypqm
m
are equal.
The following is an immediate consequence, and provides a normal form representative for right cosets
of F in G. We leave the proof as an elementary exercise for the reader.
is also Y -word in normal form
Corollary 5.9. Let yst11 ...ystnn be a Y -word in normal form. If yuv11 ...yuvm
m
such that
∈ F (yst11 ...ystnn )
yuv11 ...yuvm
m
tn
t1
vm
v1
then yu1 ...yum and ys1 ...ysn are the same words.
6. The 1-skeleton of X
In this section, we describe and study the nodes and edges of X (1) .
Definition 6.1. The 0-skeleton X (0) is defined to be the set of right cosets of F in G. The action of G
on X (0) is the usual right action on right cosets.
We observe that this action is faithful.
Lemma 6.2. The action of G on the right cosets of F in G is faithful.
T
Proof. The kernel of the action is the normal core g∈G g −1 F g. Let f be a nontrivial element of the
T
normal core. This necessarily must be an element of F , since g∈G g −1 F g ⊂ F . Since f is a non trivial
element of F , there is a non-constant finite binary sequence s such that s and f (s) are independent.
Now from our assumption it follows that f ∈ ys−1 F ys . So there is an element h ∈ F such that ys f ys−1 = h.
After applying a rearrangement move, we obtain h = f yf (s) ys−1 . Since s, f (s) are independent, the
element f yf (s) ys−1 cannot be an element of F . In particular, because yf (s) ys−1 does not preserve the tail
equivalence relation on the sequence s011011011..... This means that our assumption must be false, and
hence the normal core is trivial.
A NONAMENABLE TYPE F∞ GROUP
19
To define the edge relation, it will be useful to consider certain types of standard forms, which we shall
call special forms.
Definition 6.3. A standard form yst11 ...ystnn is called a special form, if the following holds.
(1) s1 , ..., sn are consecutive.
(2) t1 , ..., tn ∈ {−1, 1} and ti+1 = −ti .
The definition implies that each pair of percolating elements in the special form commute. Moreover,
there is a finite rooted binary tree T such that s1 , ..., sn are consecutive leaves in T .
−1
Given a percolating element ys , if we apply an expansion move, we obtain xs ys0 ys10
ys11 . The word
−1
ys0 ys10 ys11 is a special form. Now we apply an expansion move
−1
−1
−1
ys10
→ x−1
s10 ys1000 ys1001 ys101
to obtain the word
−1
−1
xs ys0 (x−1
s10 ys1000 ys1001 ys101 )ys11
Applying rearrangement moves, we obtain
−1
−1
xs x−1
s10 (ys0 ys1000 ys1001 ys101 ys11 )
The Y -word
−1
−1
ys0 ys1000
ys1001 ys101
ys11
is also a special form. This generalises to the following in a straightforward manner.
Lemma 6.4. Let s ∈ 2<N such that s 6= 0k , 1k , ∅. If we apply a sequence of expansion and rearrangement
moves on ys±1 , to obtain a standard form f yst11 ...ystnn , then the Y -word yst11 ...ystnn is a special form. Moreover,
the support of f ∈ F lies in the interval [s0∞ , s1∞ ].
Another basic observation about special forms is the following.
Lemma 6.5. Let yst11 ...ystnn be a special form, and let f ∈ F be an element such that f acts on each
sequence in the set {s1 , ..., sn }. It follows that yft1(s1 ) ...yftn(sn ) is also a special form.
Proof. The proof is elementary, and uses the fact that tree diagrams for elements of F map consecutive
pairs of sequences, upon which they act, to consecutive pairs.
Now we are ready to define the edges in X (1) .
Definition 6.6. The following equivalent conditions determine the edge relation in X (1) .
(1) (E1) F τ1 , F τ2 ∈ X (0) are connected by an edge if the double coset F τ1 τ2−1 F equals one of the
following double cosets.
−1
F y10 F
F y10
F
−1
F y100
y101 F
−1
F y100 y101
F
(2) (E2) F τ1 , F τ2 ∈ X (0) are connected by an edge if the element τ1 τ2−1 admits a standard form
f yst11 ...ystn1 such that yst11 ...ystn1 is a special form.
The advantage of the formulation in (E1) is that it provides a succint definition of the 1-skeleton. However,
in practise, (E2) will be more useful.
Lemma 6.7. The two conditions (E1) and (E2) in Definition 6.6 are equivalent.
Proof. First we will show that (E1) =⇒ (E2). Assume that F τ1 , F τ2 satisfy (E1). We show that (E2)
holds in the following cases:
(1) F τ1 τ2−1 F = F y10 F
−1
(2) F τ1 τ2−1 F = F y100
y101 F .
20
YASH LODHA
The proof for the remaining cases is symmetric.
Case 1: It follows that τ1 τ2−1 = f1 y10 f2 for some f1 , f2 ∈ F . We apply a sequence of expansion and
rearrangement substitutions on y10 to obtain a standard form f3 yst11 ...ystnn such that f2 acts on s1 , ..., sn .
By Lemma 6.4, yst11 ...ystnn must be a special form.
So
τ1 τ2−1 = (f1 f3 )(yst11 ...ystnn )f2
Applying a rearrangement move, we obtain
τ1 τ2−1 = (f1 f3 f2 )(yut11 ...yutnn )
where ui = f2 (si ). By Lemma 6.5 yut11 ...yutnn is also a special form, so we are done.
−1
Case 2: There are elements f1 , f2 ∈ F such that τ1 τ2−1 = f1 (y100
y101 )f2 . We apply a sequence of
−1
expansion and rearrangement substitutions on y100
and y101 to obtain standard forms f3 yst11 ...ystnn and
f4 yuv11 ...yuvm
respectively, so that f2 acts on each ui , sj . By Lemma 6.4,
m
Supp(f4 ) ⊂ [1010∞, 1011∞]
So f4 commutes with each percolating element ystii since
Supp(ystii ) ⊂ [1000∞, 1001∞ ]
Now
−1
)
y100
y101 = f3 (yst11 ...ystnn )f4 (yuv11 ...yuvm
m
)
= f3 f4 (yst11 ...ystnn )(yuv11 ...yuvm
m
so that
yst11 ...ystnn yuv11 ...yuvm
m
is a special form.
This way we obtain
τ1 τ2−1 = (f1 f3 f4 f2 )(yst1′ ...ystn′n yuv1′ ...yuvm
′ )
m
1
1
where s′i = f2 (si ) and u′i = f2 (ui ). By Lemma 6.5,
yst1′ ...ystn′n yuv1′ ...yuvm
′
m
1
1
is a special form, and we are done.
Next we show that if a pair F τ1 , F τ2 ∈ X (0) satisfies (E2), then it also satisfies (E1). By our assumption,
there is a special form yst11 ...ystnn and f ∈ F such that τ1 τ2−1 = f yst11 ...ystnn . By definition, the double coset
F τ1 τ2−1 F satisfies (E1) if and only if the double coset F τ2 τ1−1 F also satisfies this property. So we can
assume without loss of generality that t1 = 1. We consider two cases.
Case (1): n is odd. Each time we apply an expansion move on a percolating element ys±1 , the percolating
element is deleted and three new percolating elements appear. This means that the number of percolating
elements in the new word increases by 2.
So upon applying a sequence of n−1
2 expansion (followed by rearrangement) moves to y10 , we can obtain a
standard form of the type hyuv11 ...yuvnn . By Lemma 6.4, yuv11 ...yuunn is a special form. Also, by our assumption
on t1 , it follows that ti = vi for each 1 ≤ i ≤ n.
Now we construct a pair of finite rooted binary trees T1 , T2 such that:
(1)
(2)
(3)
(4)
T1 , T2 has the same number of leaves.
s1 , ..., sn are consecutive leaves of T1 .
u1 , ..., un are consecutive leaves of T2 .
The tree diagram (T1 , T2 ) maps si to ui for each 1 ≤ i ≤ n.
The construction of such a tree pair is an elementary exercise in the group F . The main idea is the
following. Start with trees T3 , T4 such that s1 , ..., sn are leaves in T3 , and u1 , ..., un are leaves in T4 . Now
add carets to leaves of T3 , T4 to obtain trees T1 , T2 so that:
A NONAMENABLE TYPE F∞ GROUP
21
(1) s1 , ..., sn are leaves in T1 and u1 , ..., un are leaves in T2 .
(2) The number of leaves to the left of s1 in T1 is the same as the number of leaves to the left of u1
in T2 .
(3) The number of leaves to the right of sn in T1 is the same as the number of leaves to the right of
un in T2 .
Let g = (T1 , T2 ) ∈ F . It follows that
τ1 τ2−1 g = hyst11 ...ystnn g = hgyut11 ...yutnn
= hgyuv11 ...yuvnn
Therefore,
F τ1 τ2−1 F = F yuv11 ...yuvnn F = F y10 F
Case 2: n is even. First we apply a sequence of expansion and rearranging moves to the percolating
−1
elements of y100 y101
, and their offsprings, to obtain hyuv11 ...yuvnn where yuv11 ...yuvnn is a special form. Once
again, at each step the expansion move increases the number of percolating elements by a factor of 2. So
we need to apply expansion moves n−2
times. Now just as in case (1) we construct an element g ∈ F
2
such that g(si ) = ui . And we obtain
τ1 τ2−1 g = hyst11 ...ystnn g = hgyuv11 ...yuvnn
−1
It follows that F τ1 τ2−1 F = F yuv11 ...yuvnn F = F y100 y101
F.
1
n
is not a special form. However
...ys−t
Remark 6.8. If λ = yst11 ...ystnn is a special form, then λ−1 = ys−t
1
n
−t
j
−1
i
commute
in
G,
the
group
element
described
by
λ
can
be represented by the special
,
y
since ys−t
sj
i
−t1
−tn
t1
tn
form ys1 ...ysn . Throughout this article, given a special form λ = ys1 ...ysn , we shall refer to the special
n
1
.
...ys−t
form λ−1 as the special form ys−t
n
1
are equivalent if they belong to
Definition 6.9. We say that two special forms yst11 ....ystnn and yuv11 ...yuvm
m
the same right coset of F in G.
Our analysis of X shall involve studying special forms upto equivalence. Hence it is natural to specialise
basic manipulations of standard forms to special forms up to equivalence.
Definition 6.10. Given a special form λ = yst11 ...ystnn , an expansion move at ystii entails one of the following:
(1) Replacing ystii by ysi 0 ys−1
y
if ti = 1.
i 10 si 11
−1
(2) Replacing ystii by ys−1
y
y
if
ti = −1.
s
01
i
si 1
i 00
A contraction move in λ entails one of the following:
−1
(1) Replacing a subword ys0 ys10
ys11 by ys .
−1
−1
(2) Replacing a subword ys00 ys01 ys1
by ys−1
When we use the phrase expansion move or contraction move, it will be clear from the context whether
we mean this in the sense of Definition 3.16 or Definition 6.10.
Lemma 6.11. The following conditions hold for special forms.
(1) Performing contraction and expansion moves in the sense of 6.10 on a special form produces
special forms that are equivalent.
(2) Given any pair of equivalent special forms, one can be obtained from the other by performing a
sequence of expansion and contraction moves in the sense of 6.10.
Proof. Let yst11 ...ystnn be a special form. If a substitution
−1
ys → xs ys0 ys10
ys11
or
−1
−1
ys−1 = x−1
s ys00 ys01 ys1
22
YASH LODHA
is performed, then xs (or x−1
s ) commutes with every percolating element that occurs on the left. This
means that xs (or x−1
)
can
be deleted from the word to obtain a special form that lies in the same coset.
s
Consider two special forms that lie in the same coset. Special forms cannot have potential cancellations,
by definition. So we apply a sequence of potential contractions to both, until we obtain normal forms
yst11 ...ystnn
ypq11 ...ypqm
m
Since both these special, normal forms lie in the same coset, by Theorem 5.4 they must in fact be the
same words.
Remark 6.12. Let yst11 ...ystnn be a special form. Then we can obtain a special normal form yuv11 ...yuvm
m
by applying a sequence of contraction moves in the sense of 6.10 to yst11 ...ystnn . Special normal forms
are characterised by the property that they are special forms upon which no contraction move can be
performed. Given a special normal form, any equivalent special form can be obtained by performing a
sequence of expansion moves.
Definition 6.13. We define Ω to be the set of right cosets of F in G which admit a coset representative
which is a special form.
Remark 6.14. Let yst11 ...ystnn be a special form and let f ∈ F . We can apply a sequence expansion moves
with the property that f
in the sense of 6.10 to yst11 ...ystnn to obtain an equivalent special form yuv11 ...yuvm
m
acts on u1 , ..., um . This means that
) · f = F yfv1(u1 ) ...yfvm
(F yst11 ...ystnn ) · f = (F yuv11 ...yuvm
(um )
m
Clearly, F yfv1(u)1 ...yfvm
(u)m ∈ Ω. Therefore, it makes sense to consider the right action of F on Ω.
6.1. The actions of F on Ω and X (1) . Our main goal for the remainder of the section is to understand
the action of G on X (1) and demonstrate that this is cocompact and the stabilisers are of type F∞ .
Any 1-cell contains in its G-orbit a 1-cell of the form F, F yst11 ...ystnn where yst11 ...ystnn is a special form. To
understand the stabiliser of this cell, an important step is to understand the stabiliser of F yst11 ...ystnn under
the action of F . Indeed it turns out that the two stabilisers are the same. For this reason, it is crucial to
study the action of F on Ω in more detail. We do this below.
Lemma 6.15. Consider elements
∈Ω
F yst11 ...ystnn , F yuv11 ...yuvm
m
are special forms. These elements are in the same F -orbit if and only if
where yst11 ...ystnn and yuv11 ...yuvm
m
both conditions below hold:
(1) n, m have the same parity.
(2) t1 = v1 .
In particular, the right action of F on Ω has precisely four orbits.
Proof. If yuv11 ...yuvm
is a special form and f ∈ F , then by performing expansion moves in the sense of
m
6.10 we obtain an equivalent special form yst11 ...ystnn on which f acts. Since f (s1 ), ..., f (sn ) are consecutive,
yft1(s1 ) ...yftn(sn ) is also special form. Observe that both the parity of the number of percolating elements,
as well as the sign of the exponent of the leftmost percolating element, are preserved under this action
of F .
Now let
yst11 ...ystm
m
ypq11 ...ypqnn
be two special forms such that n, m have the same parity and t1 = q1 . Upon performing expansion moves
we can assume that n = m. We construct an f ∈ F such that f (si ) = pi for 1 ≤ i ≤ n, to obtain:
F (yst11 ...ystnn ) · f = F (ypq11 ...ypqnn )
This is done in precisely the same manner as in the proof of Lemma 6.7.
A NONAMENABLE TYPE F∞ GROUP
23
Recall the Higman-Thompson groups F3 , and F3,r for r ∈ N>0 . For an introduction to these groups, we
refer the reader to [6]. It has been shown in [6] that these are of type F∞ .
We now state the main Proposition concerning the action of G on X (1) . In particular, we show that this
is cocompact and the stabilizers of cells are of type F∞ .
Proposition 6.16. The action of G on the set of 1-cells of X (1) satisfies the following:
(1) There are precisely two orbits of 1-cells.
(2) For each 1-cell e in X (1) , StabG (e) is isomorphic to F3 × F × F .
Proof of part (1):
Further, observe that the 1-cells
F, F yst11 ...ystnn
n
1
...ys−t
F, F ys−t
n
1
are in the same G-orbit. So we conclude that any 1-cell contains in its G-orbit a 1-cell of the form
F, F yst11 ...ystnn where t1 = 1.
Consider two 1-cells
e2 = F, F yuv11 ...yuvm
e1 = F, F yst11 ...ystnn
m
such that t1 = v1 = 1. First we claim that any element of G that maps e1 to e2 must be an element of F .
Indeed, this would follow if any element that maps e1 to e2 maps the trivial coset F to itself. Assume by
way of contradiction that there is an element g ∈ G such that:
F (yst11 ...ystnn ) · g = F
F · g = F yuv11 ...yuvm
m
In standard form, such an element must be of the form g = f yuv11 ...yuvm
for some f ∈ F .
m
It follows that
=F
F yst11 ...ystnn · f yuv11 ...yuvm
m
be a special form obtained from yst11 ...ystnn by expansion moves so that f acts on p1 , ..., pk
Now let
and each f (pi ) dominates u1 , ..., um . Note that for such a special form it must also hold that q1 = 1.
ypq11 ...ypqkk
It follows that
= F yfq1(p1 ) ...yfqk(pk ) yuv11 ...yuvm
F yst11 ...ystnn · f yuv11 ...yuvm
m
m
=F
In order for this to hold, it must be true that the adjacent pair yfq1(p1 ) , yuv11 is a potential cancellation, and
that f (p1 ) = u1 0l for some l ∈ N. This is impossible, since v1 = q1 = 1 and since advancing a y along a
sequence of 0’s does not change the sign of y in a calculation.
. By Lemma 6.15, such an element exists if and
It follows that g ∈ F and maps F yst11 ...ystnn to F yuv11 ...yuvm
m
only if the parity of n, m is the same. It follows that the action of G on the 1-cells has precisely two
orbits.
Proof of Part (2):
Now we observe that the statement about the stabilisers reduces to Proposition 6.17, which is stated
below. As in the proof of part (1), since G acts transitively on X (0) , it follows that each 1-cell contains in
its G-orbit a 1-cell of the form F, F yst11 ...ystnn where yst11 ...ystnn is a special form. By the above, it suffices to
understand the stabiliser of a 1-cell F, F yst11 ...ystnn where yst11 ...ystnn is a special form and t1 = 1. Moreover,
also from the proof of part (1), any element g ∈ G that maps the 1-cell F, F yst11 ...ystnn onto itself must map
F to F and F yst11 ...ystnn to F yst11 ...ystnn . It follows that
StabG (F, F yst11 ...ystnn ) = StabF (F yst11 ...ystnn )
This reduces the proof of part (2) to Proposition 6.17.
Proposition 6.17. Let yst11 ...ystnn be a special form. Then
StabF (F yst11 ...ystnn ) ∼
= F3 × F × F
24
YASH LODHA
Before we proceed to the proof of the above, we define some concepts that will be used in the proof. The
key idea is to provide a precise connection between tree diagrams for the group F3 and the stabiliser
above.
Let s1 , ..., sn be consecutive finite binary sequences. We consider the set of finite sub-forests of the infinite
rooted binary tree, with roots s1 , ..., sn . Such a forest is determined by its leaves, which is a set of finite
binary sequences u1 , ..., um satisfying:
(1) For each ui there is an sj such that sj ⊂ ui .
(2) Let ν be an infinite binary sequence which contains an element of {s1 , ..., sn } as a prefix. Then
there is a unique element of the set {u1 , ..., um } which is a prefix of ν.
In other words, s1 , ..., sn are the roots and u1 , ..., um are the leaves of the forest. We also consider labels
on the set of leaves of such a forest with elements of the set {−1, 1}, so that any pair of leaves that are
consecutive binary sequences have distinct labels. The set of all such forests with labeled leaves and roots
s1 , ..., sn will be denoted by F (s1 , ..., sn ).
determines an element
Consider a special normal form yst11 ...ystnn . Each equivalent special form yuv11 ...yuvm
m
of F (s1 , ..., sn ), which has leaves u1 , ..., um labeled respectively by t1 , ..., tm . We shall now characterise
precisely which elements of F (s1 , ..., sn ) correspond to special forms equivalent to yst11 ...ystnn . Let C1 and C2
be finite binary trees rooted at the empty string, and with leaves {00, 01, 0} and {0, 10, 11} respectively.
The forest T in F (s1 , ..., sn ) corresponding to the special form yst11 ...ystnn is simply the forest with n roots
labelled by t1 , ..., tn , and no edges.
If ti = 1, then performing an expansion move at ystii in yst11 ...ystnn produces
)(ysi 0 ys−1
...ystnn )
y
)(ysti+1
(yst11 ...ysti−1
i−1
i+1
i 10 si 11
The forest corresponding to this special form is obtained by gluing the root of C2 at si in T , deleting the
label on si , and labelling si 0, si 10, si 11 by 1, −1, 1 respectively.
If ti = −1, then performing an expansion move at ystii in yst11 ...ystnn produces
)(ys−1
y
y −1 )(ysti+1
...ystnn )
(yst11 ...ysti−1
i−1
i+1
i 00 si 01 si 1
The forest corresponding to this special form is obtained by gluing the root of C1 at si in T , deleting the
label on si , and labelling si 00, si 01, si 1 by −1, 1, −1 respectively.
So each expansion move corresponds to attaching exactly one of C1 or C2 to the leaf where the expansion
move is performed, ans relabelling. This is determined by the exponent of the percolating element upon
which it is performed. We make this precise below. In this context it makes sense to assume that the
special form we begin with is a special normal form. (Recall the discussion in 6.12.)
Lemma 6.18. Let yst11 ...ystnn be a special normal form. The elements of F (s1 , ..., sn ) that correspond to
equivalent special forms are determined recursively by the following:
(1) (The base case) The forest T with no edges, and n roots s1 , ..., sn labelled with t1 , ..., tn respectively.
(2) (Expansion) Let T1 be an element of F (s1 , ..., sn ) that corresponds to a special form equivalent to
yst11 ...ystnn .
(a) Let s be a leaf of T1 labeled by 1. We obtain an element T2 ∈ F (s1 , ..., sn ) by gluing the
root of C2 at s, deleting the label at s, and labelling s0, s10, s11 by 1, −1, 1 respectively. T2
corresponds to a special form equivalent to yst11 ...ystnn .
(b) Let s be a leaf of T1 labeled by −1. We obtain an element T2 ∈ F (s1 , ..., sn ) by gluing the
root of C1 at s, deleting the label at s, and labelling s00, s01, s1 by −1, 1, −1 respectively. T2
corresponds to a special form equivalent to yst11 ...ystnn .
Every element of F (s1 , ..., sn ) that corresponds to a special form equivalent to yst11 ...ystnn can be obtained
by applying a sequence of expansions described in (2) to the forest T described in the base case (1).
A NONAMENABLE TYPE F∞ GROUP
25
We denote the set of 3-branching forests with n roots by F3,n . The set of forests obtained in Lemma
6.18 are in natural bijective correspondence with F3,n . We leave this a straightforward exercise for the
reader. The main idea is that both trees C1 , C2 can be associated with a single ternary caret, i.e. a tree
with a root connected with three leaves. Since each addition of C1 or C2 is determined by the label, this
determines the correspondence.
Now consider two elements T1 , T2 of F (s1 , ..., sn ) such that T1 , T2 have the same number of leaves. The
forest pair (T1 , T2 ) determines a prefix replacement homeomorphism of the rational interval [s1 0∞ , sn 1∞ ].
The set of all such forest pairs is closed under composition and inverses, and hence forms a group of
homeomorphisms of [s1 0∞ , sn 1∞ ]. In fact, the following holds.
Lemma 6.19. Consider forest pairs (T1 , T2 ) ∈ Homeo+ ([s1 0∞ , sn 1∞ ]) such that:
(1) T1 , T2 are forests in F (s1 , ..., sn ) corresponding to special forms equivalent to yst11 ...ystnn .
(2) T1 , T2 have the same number of leaves.
The set of all such pairs forms a group under composition. This group is isomorphic to F3,n . The isomorphism is described by the aforementioned bijection between forests pairs in F (s1 , ..., sn ) that correspond
to equivalent special forms and forests diagrams for F3,n .
Proof. The idea behind the proof is to mimic the tree diagrams for F3,n for this group while replacing
the role of the ternary caret with C1 and C2 . The reader familiar with tree diagrams for F3,n will find
this elementary.
Recall that the groups F3,n for n ∈ N>0 , are all isomorphic to the group F3,1 = F3 (see [6]). With this is
mind, we proceed to the proof of our proposition.
Proof of Proposition 6.17:
Proof. Let yst11 ...ystnn be a special normal form. Consider f ∈ StabF (F yst11 ...ystnn ). If f does not act on
s1 , ..., sn , we perform a sequence of expansion moves to obtain a special form yuv11 ...yuvm
such that f acts
m
on u1 , ..., um . By our assumption,
· f = F yfv1(u1 ) ...yfvm
F yuv11 ...yuvm
m
(um )
= F yst11 ...ystnn
tn
t1
It follows that we can perform a sequence of contraction moves on yfv1(u1 ) ...yfvm
(um ) to obtain ys1 ...ysn .
We observe that the following holds for the element f :
(1) f stabilises the intervals [0∞ , s1 0∞ ], [s1 0∞ , sn 1∞ ] and [sn 1∞ , 1∞ ].
(2) The restriction of f to [s1 0∞ , sn 1∞ ] admits a forest diagram (T1 , T2 ), where T1 , T2 are forests in
F (s1 , ..., sn ) that correspond to special forms yuv11 ...yuvm
and yfv1(u1 ) ...yfvm
(um ) respectively.
m
This defines a monomorphism of StabF (F yst11 ...ystnn ) into F × F3,n × F . Thanks to Lemma 6.19 the
projection to F3,n is surjective. Observe that we also have an epimorphism, since if we modify the
restrictions to [0∞ , s1 0∞ ] and [sn 1∞ , 1∞ ] to any element of the copies of F supported on these respective
intervals, the new element is also in StabF (F yst11 ...ystnn ). Therefore, this is an isomorphism.
6.2. A detour on expansion moves and the expansion Lemmas. We now take a detour to prove
two Lemmas concerning this notion of expansion moves which we shall make crucial use of in subsequent
parts of the paper. We prove it here since this is the most natural place to prove it, even though we shall
only make use of it later in the paper in Subsection 7.2 and also in the final Section. In particular, the
reader will benefit from the intuition and some notation from the previous subsection while reading these
proofs.
26
YASH LODHA
In this section we use both notions of expansion coming from Definitions 6.10 and 3.16. The two notions
are closely related, and we explore this relationship in some depth in this subsection. The notion of
expansion moves as defined for special forms in 6.10 can be defined for any word yst11 ...ystnn with the
property that s1 , ..., sn are independent and ti ∈ {−1, 1}. And the contents of the Lemma 6.11 also hold
for such words with the same convention that equivalence for such words means that they lie in the same
coset.
Now expansion moves in the sense of 6.10 on percolating elements of yst11 ...ystnn correspond to a subset
of elements of F (s1 , ..., sn ) as in the discussion above. Similar ideas as above apply in the situation
where yst11 ...ystnn is not a special form, but satisfies that s1 , ..., sn are independent and ti ∈ {−1, 1}. In
particular, the set of equivalent Y -words obtained by applying such expansion moves on yst11 ...ystnn is in
natural bijective correspondence with finite 3-ary trees with n roots.
Before we proceed, we state the elementary relationship between the two notions of expansion.
Lemma 6.20. Consider a word yst11 ...ystnn with the property that s1 , ..., sn are independent and ti ∈ {−1, 1}.
be obtained from performing expansion moves on yst11 ...ystnn in the sense of Definition 6.10.
Let yuv11 ...yuvm
m
Then there is an f ∈ F such that f yuv11 ...yuvm
can be obtained from performing expansion (followed by
m
rearrangement) moves on yst11 ...ystnn in the sense of Definition 3.16. Moreover, Supp(f ) ⊆ Supp(yst11 ...ystnn )
and f −1 acts on u1 , ..., un .
Proof. This follows immediately from the definitions.
Now we state and prove the weak expansion Lemma.
Lemma 6.21. (Weak expansion Lemma) Let yst11 ...ystnn and yuv11 ...yuvm
be words satisfying that:
m
(1) s1 , ..., sn are independent and t1 , ..., tn ∈ {1, −1}.
(2) u1 , ..., um are independent and v1 , ..., vm ∈ {1, −1}.
(3) F yst11 ...ystnn = F yuv11 ...yuvm
m
Then we can perform expansion moves (in the sense of 6.10) on both words to obtain the same word
ypq11 ...ypqll such that p1 , ..., pl are independent and q1 , ..., ql ∈ {1, −1}.
Proof. Performing a sequence of contraction moves on both words produces the same Y -word yrj11 ...yrjkk in
normal form, thanks to Corollary 5.9.
The set of Y -words obtained by applying a sequence of expansion moves on yrj11 ...yrjkk corresponds to finite
3-ary forests with k roots. (Recall the discussion in the previous subsection.) Two such forests correspond
to yst11 ...ystnn and yuv11 ...yuvm
. By adding ternary carets to both these forests, we produce the same forest,
m
which corresponds to a Y -word ypq11 ...ypqll with the desired properties.
The Strong expansion Lemma will provide the core technical step in the proof of asphericity in the last
section. Before we state and prove the Lemma, we consider additional elementary feature of expansions.
The reader familiar with the arguments in [16] will already be aware of this, but for the sake of
completeness we discuss first the following feature. As a motivating example, consider the expansion
−1
−1
−1 −1
ys = xs ys0 ys10
ys11 . Now taking inverses, we obtain ys−1 = ys11
ys10 ys0
xs . Now
x−1
s (s11) = s1
x−1
s (s10) = s01
x−1
s (s0) = s00
So upon applying a rearrangement move, followed by commutation moves we get
−1
−1 −1
−1
−1
−1
−1 −1
ys11
ys10 ys0
xs = x−1
s ys1 ys01 ys00 = xs ys00 ys01 ys1
−1
−1
And xs−1 ys00
ys01 ys1
is precisely the result of applying an expansion move (in the sense of 3.16) to ys−1 .
Using an elementary inductive argument this generalises to the following:
A NONAMENABLE TYPE F∞ GROUP
27
Lemma 6.22. Let yst be such that t ∈ {1, −1}. Let f yst11 ...ystnn be a standard form obtained from applying
expansion moves (in the sense of 3.16) to yst . Then f −1 acts on s1 , ..., sn and
f −1 yft1−1 (s1 ) ...yftn−1 (sn )
can be obtained from applying expansion moves (in the sense of 3.16) to ys−t .
Definition 6.23. Let yst and yuv be a pair of percolating elements such that t, v ∈ {1, −1}. Then the
pair is said to contain no common offsprings if the following holds: For any pair yst11 ...ystnn and yuv11 ...yuvm
m
obtained from performing expansion moves (in the sense of 6.10) on yst and yuv respectively, each pair
v
ystii , yujj is distinct.
For the negation of the above, we say that the pair yst and yuv contains common offsprings. That is, yst and
yuv contain common offsprings if there are yst11 ...ystnn and yuv11 ...yuvm
obtained from performing expansion
m
v
moves on yst and yuv respectively (in the sense of 6.10) such that ystii = yujj for some 1 ≤ i ≤ n, 1 ≤ j ≤ m.
Note that for a pair yst and yuv , if s, u are independent, then the pair contains no common offsprings.
However, the case where u ⊆ s is more subtle. The following criterion is useful in detecting if such a pair
does not have common offsprings.
Lemma 6.24. Let yst and yuv be such that t, v ∈ {1, −1} and u ⊆ s. If yst and yuv have common offsprings
then the standard form yst yu−v contains a potential cancellation.
Proof. Assume that yst and yuv have common offsprings. Then we can apply expansion moves in the sense
v
of 6.10 to obtain words yst11 ...ystnn and yuv11 ...yuvm
such that ystii = yujj for some pair i, j. Now there are
m
elements f, g ∈ F such that f yst11 ...ystnn and gyuv11 ...yuvm
can be obtained from yst and yuv respectively using
m
expansions in the sense of 3.16.
By Lemma 6.22 we know that g −1 acts on u1 , ..., um . Note that g −1 acts on si since si = uj . We also
ensure that g −1 acts on elements of the set s1 , ..., sn by performing expansion moves (in the sense of 3.16)
to percolating elements
, ...ystnn
, ysti+1
yst11 , ..., ysti−1
i+1
i−1
in the word f yst11 ...ystnn if necessary. Note that the resulting word also witnesses the common offspring
since it contains the percolating element ystii . For notational convenience we also denote the resulting
word by f yst11 ...ystnn .
−vm
−v
1
using expansions
By Lemma 6.22, we know that the word g −1 yg−v
−1 (u ) ...yg −1 (u ) can be obtained from yu
1
m
in the sense of 3.16. So we get
−vm
1
yst yu−v = (f yst11 ...ystnn )g −1 yg−v
−1 (u ) ...yg −1 (u )
1
m
−vm
1
= f g −1 (ygt1−1 (s1 ) ...ygtn−1 (sn ) )yg−v
−1 (u ) ...yg −1 (u )
1
m
v
j
t −v
Now since ygti−1 (si ) = yg−1
(uj ) we can see that there is a potential cancellation in ys yu .
The Strong expansion Lemma, which we state and prove below, shall be play a role in of the proof of
asphericity of X in the final section.
Lemma 6.25. (Strong expansion Lemma) Let yst11 ...ystnn and yuv11 ...yuvm
be words such that s1 , ..., sn and
m
u1 , ..., um are independent, and each ti , vj ∈ {1, −1}. Then there are words ypq11 ...ypqll and yrw11 ...yrwkk such
that:
(1) ypq11 ...ypqll is obtained from yst11 ...ystnn by applying expansion moves as in 6.10.
(2) yrw11 ...yrwkk is obtained from yuv11 ...yuvm
by applying expansion moves as in 6.10.
m
w
w
w
(3) For each pair ypqii , yrjj it holds that either ypqii = yrjj or ypqii , yrjj contains no common offsprings.
as the words
Proof. We show this by induction on n. For the base case, we consider yst and yuv11 ...yuvm
m
such that u1 , ..., um are independent, and each t, vj ∈ {1, −1}. We apply expansion moves to yuv11 ...yuvm
m
28
YASH LODHA
to obtain a special form yrw11 ...yrwkk with depth greater than |s|. It follows that the word (yrw11 ...yrwkk )ys−t is
a standard form up to commutation.
Now using Step 1 in the proof of Theorem 5.4, we reduce the word (yrw11 ...yrwkk )ys−t to a word with no
potential cancellations. In this process, we obtain
1
k
...yfw(r
)(ypq11 ...ypqll )
(yrw11 ...yrwkk )f (ypq11 ...ypqll ) = f (yfw(r
1)
k)
such that:
(1) f (ypq11 ...ypqll ) is obtained from performing a sequence of expansion moves (in the sense of 3.16) on
ys−t .
q
q
i
i
ypjj does
, ypjj satisfies that either f (ri ) = pj , wi = −qj or the standard form yfw(r
(2) Each pair yfw(r
i)
i)
not contain a potential cancellation.
We claim that the words
yrw11 ...yrwkk
−ql
1
yf−q
−1 (p ) ...yf −1 (p )
1
l
witness the statement of the Lemma. First, thanks to Lemma 6.22 note that
−ql
1
f −1 yf−q
−1 (p ) ...yf −1 (p )
1
l
is obtained by yst using expansion moves (in the sense of 3.16). In particular,
−ql
1
yf−q
−1 (p ) ...yf −1 (p )
1
l
is obtained by yst using expansion moves (in the sense of 6.10).
−q
Next, if there is a pair yrwii , yf −1j (pj ) which is not equal and have a common offspring, then by Lemma
q
6.24 the word yrwii yfj−1 (pj ) has a potential cancellation. Since f acts on both ri and f −1 (pj ), we get
q
i
y qj
yrwii yfj−1 (pj ) · f = f yfw(r
i ) pj
q
i
ypjj has a potential cancellation. This contraSince f is merely a prefix replacement map, the word yfw(r
i)
dicts conclusion (2) from the above.
Now the inductive step is essentially the same as the base case. Given yst11 ...ystnn and yuv11 ...yuvm
, we
m
t
v1
vm
tn
and
y
...y
,
and
then
for
y
perform the process for yst11 ...ysn−1
and
the
resulting
Y
-word
obtained
n−1
u1
um
sn
by expansions. In the latter step, only the part of the word whose support intersects
from yuv11 ...yuvm
m
that of ystnn is expanded, and so the resulting offsprings from those expansions have disjoint support with
t
yst11 ...ysn−1
n−1 .
7. The complex X
The goal of this section is to define the complex X by adding higher dimensional cells to the 1-skeleton
X (1) defined in the previous section. We define a class of finite subgraphs in X (1) that shall be the 1skeletons of clusters in X. The set of all such graphs will be denoted throughout the paper as H. Higher
cells will be then added to graphs in H, and hence to X (1) in a well defined and G-equivariant manner.
Before we state the main definition of H, we state a few preliminary definitions. This language will be
useful in describing elements of H.
are said to be independent if each pair si , pj is
Definition 7.1. Special forms yst11 ...ystnn and ypq11 ...ypqm
m
independent. A list of special forms τ1 , ..., τn is said to be independent if the special forms are pairwise
independent.
A list τ1 , ..., τn is said to be sorted if it satisfies the following:
(1) τ1 , ..., τn are independent.
(2) For each i < j, the pair τi , τj satisfies that if yst and yuv are percolating elements that appear in
τi and τj respectively, then s < u.
A NONAMENABLE TYPE F∞ GROUP
29
Informally, the second condition here means that τ1 , ..., τn appear in the order left to right as visualised
in the infinite rooted binary tree.
The following is a basic observation that will be useful throughout the paper. The proof of this is an
elementary exercise concerning expansion moves in the sense of 6.10, and the partial action of elements
of F on the set of finite binary sequences.
Lemma 7.2. Let τ1 , ..., τn be a sorted list of special forms. The following holds.
(1) If τ1′ , ..., τn′ are special forms such that τi , τi′ are equivalent, then τ1′ , ..., τn′ is also a sorted list.
(2) Let f ∈ F be such that for each percolating element yst in any τi , f acts on s. Let τi′ be the special
form obtained from replacing each yst in τi with yft (s) . Then τ1′ , ..., τn′ is a sorted list.
Throughout the paper,
we fix the
Q
Q following convention for a sorted list τ1 , ..., τn . Given X ⊆ {1, ..., n},
we denote τX = i∈X τi . Here i∈X τi denotes a formal word obtained by concatenation of the τi for
i ∈ X in the increasing order of indices. So for instance, if X = {1, 3, 4}, then τX = τ1 τ3 τ4 .
Definition 7.3. An element C of H is a subgraph of X (1) which is determined by the following:
(1) A sorted list of special forms τ1 , ..., τn .
(2) A coset F τ .
Then C is the induced subgraph of the vertex set
{F τX τ | X ⊆ {1, ..., n}}
(1)
in X . The above information is said to be a description of C. We say that C is described by the base
F τ and parameters τ1 , ..., τn . It will be often convenient to choose τ above as a Y -word, and later in the
paper we shall make this a convention for certain definitions.
Observe that each closed 1-cell in X (1) is an element of H. To see this, recall that each 1-cell F τ1 , F τ2
satisfies that
F τ2 τ1−1 = F ν
where ν is a special form. It follows that F τ2 = F ντ1 , and so the 1-cell admits a description with base
F τ1 and parameter ν.
7.1. Descriptions of graphs in H and the G-action. A graph in H can be described in many different
ways. For example, the closed 1-cell F y10 , F can be described with base F and parameter y10 , or with
−1
base F y10 and parameter y10
. Since graphs in H shall emerge as 1-skeletons of clusters in X, we shall
describe clusters in X by describing their 1-skeleton in H. Therefore, it shall be useful to understand
different kinds of descriptions of graphs in H. Much of the technical difficulty in this article arises from
dealing with non-uniqueness of the descriptions of clusters.
Given a graph in H, we can find a description with any given node of the graph as base. This follows
immediately from the definitions and is captured in the following Lemma.
Lemma 7.4. Let C ∈ H be described with base F τ and parameters τ1 , ..., τn . Then for any X ⊆ {1, ..., n},
C admits a description with base F τX τ , and parameters τ1s1 , ..., τnsn , where si = −1 if i ∈ X and si = 1
otherwise.
Next, we observe that replacing parameters by equivalent special forms describe the same graph.
Lemma 7.5. Let C ∈ H be described with base coset F τ and parameters τ1 , ..., τn . Let τ1′ , ..., τn′ be
special forms such that τi′ is equivalent to τi . Then C also admits the description with base coset F τ and
parameters τ1′ , ..., τn′ .
Proof. Since they are equivalent, it follows that τi′ can be obtained from τi by a sequence of expansion
and contraction moves in the sense of 6.10. In particular, from Lemma 6.20 we know that there are
fi ∈ F such that:
30
YASH LODHA
(1) τi = fi τi′ for 1 ≤ i ≤ n.
(2) The support of fi is contained in the support of τi , which equals the support of τi′ .
This means that fi , τj′ commute if i 6= j. It follows that for each X ⊆ {1, ..., n}, we have F τX τ =
′
F τX
τ.
A nice consequence of the above is that when describing the graph C, we can conveniently choose parameters with desired features. For instance, in some cases it may be useful to choose parameters with
sufficiently large depth.
We end this section with the following useful remark about the choice of base coset.
Remark 7.6. Let C be a cluster described with base F τ and parameters τ1 , ..., τn . The word τ may not
be a Y -word, but we can find a different description for C whose base coset representative is a Y -word.
This is done as follows.
First we convert the word τ into a standard form f ν1 , where f ∈ F and ν1 is a Y -word. Let τi′ be a
special form obtained by applying expansion moves on τi such that for each yst that occurs in τi′ , f acts
on s. Let τi′′ be the special form obtained by replacing each percolating element yst in τi′ by yft (s) . It
follows from Lemma 7.2 that τ1′′ , ..., τn′′ is a sorted list.
So for each X ⊆ {1, ..., n}
′′
F τX τ = F τX f ν1 = F τX
ν1
We conclude that C admits a description with base at F ν1 and with parameters τ1′′ , ..., τn′′ .
Lemma 7.7. Let C ∈ H. For an element ν ∈ G, C · ν is also an element of H.
Proof. Let C be described with base F τ and parameters τ1 , ..., τn . Then C · ν is a cluster that admits a
description with base F (τ ν) and parameters τ1 , ..., τn .
Lemma 7.7 is a simple observation, but has useful consequences. For instance, while formulating our
arguments we shall often assume that a given cluster is based at the trivial coset, using transitivity of
the G-action on X (0) .
7.2. H is closed under intersection. Our goal now is to show the following.
Proposition 7.8. Let C1 , C2 ∈ H. If C1 ∩ C2 is nonempty, then it is an element of H.
First we observe the following. If F τ is a node in C1 ∩ C2 , acting upon this by τ −1 , we obtain
(C1 ∩ C2 ) · τ −1 = (C1 · τ −1 ) ∩ (C2 · τ −1 )
Thanks to Lemma 7.7, C1 ∩C2 ∈ H if and only if (C1 ·τ −1 )∩(C2 ·τ −1 ) ∈ H. The advantage of considering
the latter is that the node given by the trivial coset F lies in (C1 · τ −1 ) ∩ (C2 · τ −1 ).
Before we supply a full proof of Proposition 7.8, we need to achieve a preliminary understanding of the
situation. In this subsection we describe conceptual ingredients that are needed in the proof.
Definition 7.9. Let τ1 , ..., τn and λ1 , ..., λm be sorted lists. Such a pair of sorted lists is said to be
compatible if the following holds. For each X ⊆ {1, ..., n}, Y ⊆ {1, ..., m} if τX , λY are equivalent, then
τX and λY are equal as words.
Lemma 7.10. Let τ1 , ..., τn and λ1 , ..., λm be sorted lists. Then there is a compatible pair of sorted lists
τ1′ , ..., τn′ and λ′1 , ..., λ′m such that τi , τi′ and λi , λ′i are equivalent for each 1 ≤ i ≤ n.
Proof. This follows from an application of Lemma 6.21 on the words τ{1,...,n} , λ{1,...,m} .
A NONAMENABLE TYPE F∞ GROUP
31
Definition 7.11. Let τ1 = yst11 ...ystnn and τ2 = yuv11 ...yuvm
be special forms. τ1 , τ2 are said to be consecutive,
m
if sn , u1 are consecutive. τ1 , τ2 are said to be alternating if they are consecutive, and also satisfy that
v1 = −tn . It follows from the definition that τ1 , τ2 are alternating if and only if the product word τ1 τ2 is
a special form. A list τ1 , ..., τn of special forms is said to be alternating if each pair τi , τi+1 is alternating
for 1 ≤ i ≤ n − 1. It follows that τ1 , ..., τn is alternating if and only if the word τ1 ...τn is a special form.
Consider C ∈ H described at base F and with parameters τ1 , ..., τn . Observe that for a nonempty set
X ⊆ {1, ..., n}, there is an edge connecting F and F τX if and only if X = {i, i + 1, i + 2, ..., j} so that
τi , τi+1 , ..., τj are alternating. This motivates the following.
Definition 7.12. Let τ1 , ..., τn be as above. A subset
{τi , τi+1 , τi+2 , ..., τj } ⊆ {τ1 , ..., τn }
is said to be an alternating block if τi , τi+1 , τi+2 , ..., τj are alternating. We abuse notation and also refer
to the subset of indices {i, i + 1, ..., j} ⊆ {1, ..., n} as an alternating block. This special usage will be clear
from the context.
Given the sorted list above, a block decomposition is given by the sets
{1, ..., l1 }, {l1 + 1, l1 + 2, ..., l2 }, ..., {lk + 1, ..., n}
where each set is an alternating block and 1 ≤ l1 < ... < lk < n. Clearly, any sorted list admits a unique
partition into maximal alternating blocks. This is called the maximal block decomposition of the sorted
list.
−1
−1
Example 7.13. Consider the sorted list y010 , y011
, y10
, y110 . The maximal block decomposition of this
is
−1
−1
{y010 , y011
}, {y10
, y110 }
−1
−1
Note that y011
, y10
are consecutive (since 011, 10 are consecutive binary sequences), but they are not
alternating.
Lemma 7.14. Let τ1 , ..., τn and λ1 , ..., λm be sorted lists of special normal forms such that τ{1,...,n} and
λ{1,...,m} are equal as words. Then for each maximal alternating block U ⊆ {1, ..., n} of the former, there
is a maximal alternating block V ⊆ {1, ..., m} of the latter such that τU = λV . In other words, there is a
natural bijection between the maximal block decompositions of τ1 , ..., τn and λ1 , ..., λm .
Proof. We read both words from left to right and partition them as we go along.
By definition, the set of nodes of a graph in H naturally forms a boolean algebra in the sense of an
algebra of subsets of a given finite set. We recall the analogous notion for a collection of subsets.
Definition 7.15. Let X be a collection of subsets of the set {1, ..., n}. We say that X is boolean, if there
are elements X1 , ..., Xk ∈ X such that:
(1) X1 , ...,S
Xk are pairwise disjoint.
(2) X = { i∈Z Xi | Z ⊆ {1, ..., k}}.
Under the partial ordering of inclusion, the sets X1 , ..., Xk are the minimal elements of X. The minimal elements of a boolean collection will be referred to as the atoms of the boolean collection. Any subcollection
X ⊆ 2{1,...,n} that is closed under taking unions and set difference is a boolean subcollection.
Lemma 7.16. Let τ1 , ..., τn and λ1 , ..., λm be a compatible pair of sorted lists. Let
X = {X ⊆ {1, ..., n} | τX = λY for some Y ⊆ {1, ..., m}}
Y = {Y ⊆ {1, ..., m} | τX = λY for some X ⊆ {1, ..., n}}
Then both X, Y are boolean sub collections.
32
YASH LODHA
Proof. We show this for X, the proof for Y is the same. In what follows below equality denotes equality
as words. Let X1 , X2 ⊆ X. Then there are Y1 , Y2 ∈ Y such that τX1 = λY1 and τX2 = λY2 . It follows
that τX1 ∪X2 = λX1 ∪X2 and τX1 \X2 = λX1 \X2 . Since X is closed under unions and set difference, we are
done.
Next we demonstrate that subgraphs of C ∈ H that are elements of H correspond precisely to certain
boolean subcollections.
Lemma 7.17. Let C ∈ H and let C ′ ⊂ C be an element of H. Then C admits a description with bases
at F τ and parameters τ1 , ..., τn such that the following holds. There is a boolean collection X ⊂ 2{1,...,n}
such that C ′ admits the following description.
(1) Base at F τ .
(2) The elements of the set
{τX | X is an atom of X}
are all special forms that are the parameters.
Proof. By the G-action, we can assume that the trivial coset F lies in C ∩ C ′ . We find a compatible pair
of sorted lists τ1 , ..., τn and ν1 , ..., νm that are parameters for descriptions of C, C ′ respectively with both
bases at F . Let
X = {X ⊆ {1, ..., n} | F τX = F λY for some Y ⊆ {1, ..., m}}
Since the sorted lists are compatible,
X = {X ⊆ {1, ..., n} | τX = λY for some Y ⊆ {1, ..., m}}
Thanks to Lemma 7.16, X is a boolean subcollection. In particular,
{τX | X is an atom of X}
are all special forms that are the parameters of C ′ with base F .
Remark 7.18. The following converse of the above Lemma follows immediately from the definitions.
Let C be the cluster described with base F τ and parameters τ1 , ..., τn . Consider a boolean collection
X ⊂ 2{1,...,n} which satisfies that the elements of
{τX | X is an atom of X}
are all special forms. Then there is a subgraph C ′ ⊂ C which lies in H, and is described with base F τ
and parameters in the set above.
Proof of Proposition 7.8
Proof. Let C1 , C2 be our clusters in H. Thanks to Lemmas 7.5 and 7.10, we can assume that C1 , C2
are both described with base at the trivial coset F and parametrised by a compatible pair of sorted lists
τ1 , ..., τn and ν1 , ..., νm respectively. Let
X = {X ⊆ {1, ..., n} | F τX ∈ C1 ∩ C2 }
= {X ⊆ {1, ..., n} | F τX = F λY for some Y ⊆ {1, ..., m}}
= {X ⊆ {1, ..., n} | τX = λY for some Y ⊆ {1, ..., m}}
Thanks to Lemma 7.16, X is a boolean subcollection.
We will show that each atom U of X satisfies that U is an alternating block, i.e. τU is a special form. Let
V ⊆ {1, ..., m} be such that τU = νV . Let U = {u1 , ..., uk } and V = {v1 , ..., vl } be the elements listed in
increasing order. Applying Lemma 7.14 to the sorted lists (of normal special forms)
τu1 , τu2 ..., τuk
νv1 , νv2 , ..., νvl
we conclude that any maximal alternating block U1 ⊆ U corresponds to a maximal alternating block
V1 ⊆ V satisfying equality of words τU1 = νV1 . But this means that F τU1 = F νV1 and so F τU1 ∈ C1 ∩ C2 .
Therefore U1 ∈ X. By minimality of U it follows that U1 = U and so U is an alternating block.
A NONAMENABLE TYPE F∞ GROUP
33
This means that C1 ∩ C2 is described with base F and parameters in the set
{τU | U is an atom of X}
and hence is an element of H as desired.
7.3. Graphs in H as 1-skeletons of clusters. We demonstrate that an element C ∈ H is the 1-skeleton
of a cluster in the sense of Definition 4.1. This identification emerges in a natural way by considering
descriptions of graphs in H that we call proper descriptions. A proper description for C involves a certain
elementary condition on the parameters which we describe below.
Definition 7.19. Let τ1 , ..., τn be a sorted list of special forms. We say that the sorted list is proper, if
whenever a pair τi , τi+1 is consecutive, then it is also alternating.
Observe that given any pair of consecutive special forms τ1 , τ2 , exactly one of τ1 , τ2−1 or τ1 , τ2 is alternating.
(Recall the discussion in Remark 6.8 concerning the notation τ2−1 .) Indeed, given any sorted list of special
forms τ1 , ..., τn , there are numbers s1 , ..., sn ∈ {1, −1} such that τ1s1 , ..., τnsn is proper. (To see this, perform
an elementary induction on n.)
Lemma 7.20. Any element C ∈ H admits a description where the parameters are proper. Moreover, we
can also arrange it so that the parameters are all normal forms.
Proof. Let C be described with base F τ where τ is a Y -word, and parameters τ1 , ..., τn . Let s1 , ..., sn ∈
{1, −1} be such that τ1s1 , ..., τnsn is proper. Then C admits the following description:
(1) Base F τX τ , where X = {i ∈ {1, ..., n} | si = −1}.
(2) Parameters τ1s1 , ..., τnsn .
Finally, by applying a sequence of contraction moves to τ1s1 , ..., τnsn , we obtain a proper list of special
normal forms. To see this, one simply checks the following. If ν1′ , ν2′ are obtained from ν1 , ν2 respectively
by applying contraction and/or expansion moves, then ν1′ , ν2′ are alternating if and only if ν1 , ν2 are
alternating.
Definition 7.21. We call a description of C ∈ H with proper parameters as a proper description.
The advantage of a proper description for C is that the edge relation between nodes of C is easier to
characterise.
Lemma 7.22. Let C ∈ H be described with base F τ where τ is a Y -word, and proper parameters τ1 , ..., τn .
Then the nodes F τX τ, F τY τ for X, Y ⊆ {1, ..., n} form an edge if and only if one of the following holds:
(1) X ⊂ Y and τY \X is a special form.
(2) Y ⊂ X and τX\Y is a special form.
Proof. If F τX τ, F τY τ form an edge, this means that
Y
Y
τj−1
τi
τX τY−1 =
i∈X\Y
j∈Y \X
is a special form. If both X \ Y and Y \ X are nonempty, then there must be an alternating pair of the
−1
form τi , τi+1
or τi−1 , τi+1 . In either case, this contradicts the assumption that the parameters are proper.
The converse follows immediately from the definition of the edge relation in X (1) .
Now we are ready to provide the identification between elements of H and 1-skeletons of clusters.
Proposition 7.23. Let C ∈ H be described by base F τ where τ is a Y -word, and proper parameters
τ1 , ..., τn . Then C is isomorphic to the 1-skeleton of an n-cluster in the sense of Definition 4.1.
34
YASH LODHA
Proof. Without loss of generality (thanks to Lemma 7.7) we can assume that C is based at the trivial
coset. Define
Y = {i ∈ {1, ..., n − 1} | τi τi+1 is a special form}
Now define the hyperplane arrangement A in Rn consisting of the following hyperplanes:
(1) xi = 0, xi = 1, 1 ≤ i ≤ n.
(2) xi = xi+1 for i ∈ Y.
Clearly, A is admissible. We claim that the 1-skeleton of C(A) is isomorphic to C.
The 0-cells of C(A) are naturally in bijection with subsets of {1, ..., n} as follows. A subset X ⊆ {1, ..., n}
corresponds to the 0-cell given by the coordinates (y1 , ..., yn ) where yi = 0 if i ∈
/ X and yi = 1 if i ∈ X.
Observe that the 1-cells of C(A) are in a natural bijection with partitions of {1, ..., n} into three sets
X1 , X2 , X3 such that:
(1) X1 , X2 , X3 are pairwise disjoint and X1 ∪ X2 ∪ X3 = {1, ..., n}.
(2) X3 = {j, j + 1, ..., k} ⊂ {1, ..., n} such that for each j ≤ i ≤ k − 1, i ∈ Y.
The 1-cell corresponding to such a pair is the affine line segment connecting the 0-cells (y1 , ..., yn ) and
(z1 , ..., zn ) given by the following coordinates:
(y1 , ..., yn )
yi = 0 if i ∈ X1 ∪ X3 and yi = 1 if i ∈ X2
(z1 , ..., zn )
zi = 0 if i ∈ X1 and yi = 1 if i ∈ X2 ∪ X3
This is precisely the intersection of the following collection of hyperplanes with [0, 1]n :
{{xi = 1} | i ∈ X1 } ∪ {{xi = 0} | i ∈ X2 } ∪ {{xi = xj } | i, j ∈ X3 }
(1)
We now define the graph isomorphism between C(A)
and C. First, we identify the 0-cell given by
X ⊆ {1, ..., n} in C(A) with the 0-cell F τX of C. Then we identify the 1-cell of C(A) given by a pair
X1 , X2 (satisfying the above) with the 1-cell connecting the pair F τX1 , F τX1 ∪X2 . This identification
(1)
provides the isomorphism between C(A) and C.
We now show that the identification above is independent of the choice of proper description.
Lemma 7.24. Let C ∈ H and let C(A) be the cluster identified with a proper description of C in
Proposition 7.23. Then this identification is the same for any other choice of proper parameters describing
C.
Proof. By G-equivariance, we can assume the the base of the given description is the trivial coset,
with proper parameters τ1 , ..., τn . Let A1 , ..., Ak ⊆ {1, ..., n} be the maximal block decomposition for
{τ1 , ..., τn }.
With Lemmas 7.4 and 7.5 in mind, observe that any proper description of C (upto replacing equivalent
special forms as parameters) is determined by a set of numbers s1 , ..., sn ∈ {1, −1} such that:
(1) si = sj whenever i, j ∈ Al for some 1 ≤ l ≤ k.
(2) The base coset is F τX where X = {1 ≤ i ≤ n | si = −1}.
(3) The parameters are τ1s1 , ..., τnsn .
Let A1 be the hyperplane arrangement in Rn given by:
(1) {xi = 0}, {xi = 1}, 1 ≤ i ≤ n.
(2) {xi = xi+1 } for i ∈ {1 ≤ j ≤ n − 1 | τj τj+1 is a special form}.
Let A2 be the hyperplane arrangement in Rn given by:
(1) {xi = 0}, {xi = 1}, 1 ≤ i ≤ n.
s sj+1
is a special form}.
(2) {xi = xi+1 } for i ∈ {1 ≤ j ≤ n − 1 | τj j τj+1
A NONAMENABLE TYPE F∞ GROUP
35
It follows that both arrangements are the same and hence describe the same cluster. By symmetries of
this n-cluster, we are done.
7.4. Higher dimensional cells in X. Proposition 7.23 provides an identification of an element C ∈ H
with the 1-skeleton of an n-cluster C(A).
Definition 7.25. Given C ∈ H, let C(A) be the n-cluster whose 1-skeleton is identified with C in
Proposition 7.23. A filling C of C is obtained by adding the higher cells of C(A) to C in X (1) . The
complex X is obtained from X (1) by “filling” each graph in H. In other words,
[
X=
C
C∈H
We show that the fillings are well defined and indeed produce a CW complex X upon which G acts.
Lemma 7.26. The fillings of H are G-equivariant. In other words, for each C ∈ H, ν ∈ G, if C1 = C · ν
then C · ν = C1 In particular, the G-action on X (1) naturally extends to the G-action on X.
Proof. Let C be described with base F τ , where τ is Y -word, and proper parameters τ1 , ..., τn . Let ν ∈ G.
Then C · ν admits a description with base F (τ ν) and proper parameters τ1 , ..., τn . The hyperplane
arrangements A1 for C and A2 for C · ν as described in the proof of Proposition 7.23 are both:
(1) {xi = 0}, {xi = 1}, 1 ≤ i ≤ n.
(2) {xi = xi+1 } for i ∈ {1 ≤ j ≤ n − 1 | τj τj+1 is a special form}.
Therefore, the group action preserves the identification.
We end the section by demonstrating that X is indeed a CW complex. This follows immediately once we
combine Proposition 7.8 with the following.
Proposition 7.27. Let C1 , C2 ∈ H such that C1 ⊂ C2 . Then C1 agrees with the subcomplex of C2
determined by C1 ⊂ C2 .
Proof. By considering the G-action, we assume without loss of generality that C1 , C2 both contain the
trivial coset as a node. Let C2 be described with base F and parameters τ1 , ..., τn . By Lemma 7.17, there
is a boolean collection X ⊂ 2{1,...,n} such that C1 admits the following description.
(1) Base at F .
(2) The elements of the set {τX | X is an atom of X} are all special forms that are the parameters.
Let X1 , ..., Xk ⊂ {1, ..., n} be the minimal elements of this boolean collection appearing in increasing
order. If i, i + 1 ∈ Xj for some 1 ≤ j ≤ k, then τi , τi+1 are alternating. Using this fact, it is easy to find
numbers s1 , ..., sn such that:
(1) τ1s1 , ..., τnsn is proper.
(2) si = si+1 if i, i + 1 ∈ Xj for some 1 ≤ j ≤ k.
tk
t1
In particular, the list τX
, ..., τX
is proper, where tj equals sl for any l ∈ Xj . This provides proper
1
k
descriptions of C1 , C2 with base F τY where Y = {i ∈ {1, ..., n} | si = −1}, and proper parameters
tk
t1
τ1s1 , ..., τnsn and τX
, ..., τX
respectively.
1
k
Let A1 be the arrangement of hyperplanes:
(1) {xi = 0}, {xi = 1}, 1 ≤ i ≤ n.
(2) {xi = xi+1 } whenever τi τi+1 is a special form.
Now define a k-dimensional flat U of A1 as the intersection of the following hyperplanes.
S
(1) {xi = 0} for i ∈
/ 1≤i≤k Xk .
36
YASH LODHA
(2) {xi = xi+1 } with i for which there is a 1 ≤ j ≤ k such that i, i + 1 ∈ Xj .
Let A2 be the arrangement of hyperplanes obtained from restricting the arrangement A1 to the flat U .
It is an elementary exercise in hyperplane arrangements to show that A2 is naturally isomorphic to the
arrangement in Rk given by:
(1) {xi = 0}, {xi = 1}, 1 ≤ i ≤ k.
(2) {xi = xi+1 } for i such that τXi τXi+1 is a special form.
This proves that the restriction of C2 to C1 is indeed equal to C1 .
8. The action of G on X
For the rest of the article, we shall identify elements of H with their fillings, and denote these as clusters.
So given C ∈ H, we shall simply use C to denote the filled cluster. In this section we study the action of
G on X. We prove the following:
Proposition 8.1. The stabiliser StabG (e) of any cell e of X is a group of type F∞ . More particularly,
it is a finite product of copies of the groups F3 and F .
Proposition 8.2. The quotient X/G has finitely many cells in each dimension.
8.1. Stabilisers of cells. Given any n-cell e in X, by the construction of X there is an n-cluster C ∈ H
that contains e. Let C be described with base F τ , where τ is a Y -word and proper parameters τ1 , ..., τn .
Then the cluster C1 = C · τ −1 admits a description with base F and the same parameters τ1 , ..., τn as C.
It follows that there is an n-cell e′ of C1 such that e′ = e · τ −1 , and hence StabG (e) ∼
= StabG (e′ ). So it
suffices to understand the stabiliser of e′ . For the rest of this subsection we fix the description of C1 and
e′ . First, we make an elementary observation.
Lemma 8.3. The closure of e′ contains F and F τ{1,...,n} .
Proof. The cluster C1 ∈ H is isomorphic to the cluster (in the sense of definition 4.1) described by the
following hyperplanes in Rn .
(1) {xi = 1}, {xi = 0}, 1 ≤ i ≤ n.
(2) {xi = xi+1 } for i such that τi τi+1 is a special form.
The closure of an n-cell is an intersection of certain half spaces of the form
{xi ≥ 0}
{xi ≤ 1}
{xi ≤ xi+1 }
{xi ≥ xi+1 }
All these half-spaces contain the points (0, ..., 0) and (1, ..., 1), which correspond to the nodes F and
F τ{1,...,n} of C1 .
An important step in understanding the stabiliser StabG (e′ ) is to demonstrate that StabF (e′ ) = StabG (e′ ).
The stabiliser StabF (e′ ) is much easier to understand, since we already have developed a theory of the
action of F on Ω. In order to show that StabF (e′ ) = StabG (e′ ), we first need to prove the following
generalisation of Proposition 6.17.
Lemma 8.4. Let τ1 , ..., τn be a proper list of special forms as above, and let X1 , ..., Xk ⊂ {1, ..., n} be
pairwise disjoint sets. Then
\
StabF (F τXi )
1≤i≤k
is a finite direct product of copies of F and F3 .
Proof. For a given X ⊂ {1, ..., n}, let Y1 , ..., Yl be the maximal block decomposition of X. First we claim
that
\
StabF (F τXi ) =
StabF (F τYj )
1≤j≤l
A NONAMENABLE TYPE F∞ GROUP
37
It is clear that the latter is a subset of the former. So we need to show that if f ∈ StabF (F τX ), then
f ∈ StabF (F τYi ) for each 1 ≤ i ≤ l.
We show this for the case l = 2. The general case follows from a straightforward induction that uses the
same idea as in the base case. For notational simplicity, we denote τY1 = ψ1 and τY2 = ψ2 . Let ψ1′ , ψ2′ be
special forms obtained by applying expansion moves on ψ1 , ψ2 such that for each percolating element yst
of ψ1 or ψ2 , f acts on s. Let ψ1′′ , ψ2′′ be special forms obtained by replacing each percolating element yst
by yft (s) . Since
F ψ1 ψ2 · f = F ψ1′′ ψ2′′ = F ψ1 ψ2
we can apply a sequence of contraction and expansion moves on ψ1′′ ψ2′′ to obtain ψ1 ψ2 .
Now since Y1 , Y2 were assumed to be maximal alternating blocks, we know that ψ1 , ψ2 are not alternating.
Therefore, it must be the case that we can apply a sequence of contraction and expansion moves to ψ1′′
to obtain ψ1 and a sequence of contraction and expansion moves on ψ2′′ to obtain ψ2 . It follows that
f ∈ StabF (F ψ1 ) ∩ StabF (F ψ2 )
So our claim holds.
Now we apply the above claim to each Xi in the sets X1 , ..., Xk from the statement of the Lemma.
Reasoning along the lines of Proposition 6.17, we obtain that the restriction of the element f on the
support of each maximal alternating block of each Xi must be a tree diagram in natural correspondence
with a tree diagram of an element of F3 . (We do not recall the notation and details of this here, as they
have been spelled out in subsection 6.1.) In this way, we obtain the required direct product decomposition.
For our cell e′ , we define
Y = {Y ⊆ {1, ..., m} | F τY lies in the closure of e′ }
Let X be the boolean subcollection obtained by taking the closure of Y in 2{1,...,m} under unions and
set difference. Further, let X1 , ..., Xk be the atoms of X. For the remainder of this subsection, we fix e′ ,
Y, X and X1 , ..., Xk as above.
Lemma 8.5.
\
StabF (F τY ) =
Y ∈Y
In particular,
\
StabF (F τXi )
1≤i≤k
\
StabF (F τY )
Y ∈Y
is a finite product of copies of F and F3 , and is of type F∞ .
Proof. Let U, V ⊆ {1, ..., m}. We claim that
StabF (F τU ) ∩ StabF (F τV )
= StabF (F τU\V ) ∩ StabF (F τV \U ) ∩ StabF (F τU∩V )
It is obvious that the latter is contained in the former. It suffices to show that if
f ∈ StabF (F τU ) ∩ StabF (F τV )
then
f ∈ StabF (F τU\V ) ∩ StabF (F τV \U ) ∩ StabF (F τU∩V )
Let τi′ be a special form obtained by performing expansion moves on τi such that for each percolating
element yst of τi′ , f acts on s. Now let τi′′ denote special form obtained by replacing each percolating
element yst of τi′ with yft (s) . It follows from Lemma 7.2 that τ1′′ , ..., τn′′ is a sorted list of special forms.
It follows from our assumption that
F τU · f = F τU′′
F τV · f = F τV′′
38
YASH LODHA
By our assumption the word τU′′ , τV′′ are both products of independent special forms, and are equivalent
to τU , τV respectively. This means that we can perform a sequence of contraction and expansion moves
on τU′′ , τV′′ to obtain τU , τV respectively. So indeed upon performing contraction and expansion moves on
′′
′′
the subwords τU∩V
, τU\V
, τV′′ \U we must obtain τU∩V , τU\V , τV \U respectively. In particular, it follows
that f fixes the cosets F τU∩V , F τU\V , F τV \U . This proves our assertion.
Using this idea, we conclude the proof of the main assertion
\
\
StabF (F τX ) =
StabF (F τXi )
1≤i≤k
X∈X
using a standard induction argument of the boolean algebra of sets X obtained from Y by taking the
closure under unions and set difference.
Lemma 8.6. Given ν ∈ StabG (e′ ), for each X ∈ X it holds that
ν(Supp(τX )) = Supp(τX )
Proof. Given ν ∈ StabG (e′ ), ν induces a permutation of the 0-cells {F τY | Y ∈ Y}. In particular, a
power ν k fixes each 0-cell in this set. In particular, ν k ∈ F since it fixes the trivial coset.
Since ν k lies in
\
StabF (F τY )
\
StabF (F τX )
Y ∈Y
by Lemma 8.5 it also lies in
X∈X
This implies that for each X ∈ X,
ν k (Supp(τX )) = Supp(τX )
Since ν is a k-th root of ν in Homeo+ (2N ), it follows that
ν(Supp(τX )) = Supp(τX )
′
′
Lemma 8.7. StabG (e ) = StabF (e ) =
T
1≤i≤k
StabF (F τXi )
Proof. Let ν ∈ StabG (e′ ). Since ν permutes the 0-cells incident to e′ , it follows that F · ν = F τX for
some X ∈ Y. In particular, it holds that there is an f ∈ F such that ν = f τX . Now assume by way of
contradiction that X is nonempty.
By Lemma 8.6 ν(Supp(τX )) = Supp(τX ). It follows that f (Supp(τX )) = Supp(τX ).
be a special form obtained by applying expansion moves on yst11 ...ystnn
Now let τX = yst11 ...ystnn . Let yuv11 ...yuvm
m
such that:
(1) f acts on each ui for 1 ≤ i ≤ m.
(2) |f (ui )| > |sj | for each 1 ≤ i ≤ m, 1 ≤ j ≤ n.
Since the interval Supp(τX ) is invariant under the action of f , it follows that f (u1 ) = s1 1l for some l ∈ N.
Also, note that v1 = t1 .
So we obtain that
F τX · (f τX ) = F (yst11 ...ystnn )(f yst11 ...ystnn )
)(f yst11 ...ystnn )
= F (yuv11 ...yuvm
m
tn
t1
= F (yfv1(u1 ) ...yfvm
(um ) )(ys1 ...ysn )
Since f (Supp(τX )) = Supp(τX ), for the infinite binary sequence s1 1∞ the associated calculation Λ of the
standard form
tn
t1
(yfv1(u1 ) ...yfvm
(um ) )(ys1 ...ysn )
A NONAMENABLE TYPE F∞ GROUP
39
contains two occurrences of y t1 = y v1 . In particular, there is no potential cancellation and the exponent
of this calculation equals 2. Since F τX · (f τX ) is equal to a coset of the form F τU for some U ∈ Y, this
is a contradiction. Therefore, X must be the empty set, and so ν ∈ F .
To finish the proof, we must show the second equality in the statement of our Lemma. We already know
from Lemma 8.6 that ν(Supp(τX )) = Supp(τX ) for each X ∈ X.
For any Y ∈ Y, F τY · ν = F τY ′ where Y ′ ∈ Y. Now Y is a union of minimal elements of X. Since ν
fixes the support of each τX for such a minimal X, it must be the case that F τX · ν = F τX and hence
F τY · ν = F τY . This means that ν fixes each 0-cell in the closure of e′ . We conclude by applying Lemma
8.5 to obtain
\
\
StabF (F τY ) =
StabF (F τXi )
Y ∈Y
1≤i≤k
Proof of Proposition 8.1 From the discussion at the beginning of this subsection, it suffices to consider
a cell of the form e′ described there. We denote by Y, X and X1 , ..., Xk the same sets as above.
By Lemma 8.7 we know that
StabG (e′ ) =
\
StabF (F τXi )
1≤i≤k
T
Moreover, by Lemma 8.4 it follows that 1≤i≤k StabF (F τXi ) is a finite product of copies of F and F3 .
These groups are of type F∞ , which is a property closed under taking finite products.
8.2. The quotient X/G. Now we show that the quotient X/G has finitely many cells in each dimension.
Proof of Proposition 8.2
We show this for a given dimension n ∈ N. Any n-cell is a maximal dimensional cell of an n-cluster by
construction. Given an n-cluster, it contains in its G-orbit an n-cluster described with base the trivial
coset F and proper parameters. So it suffices to show that there are finitely many G-orbits of n-clusters
described with base F and proper parameters.
Let C1 , C2 be n-clusters described with base F and proper parameters τ1 , ..., τn and ν1 , ..., νn respectively.
Assume the following holds for each 1 ≤ i ≤ n.
(1) The parity of the number of percolating elements in νi and τi is the same.
(2) τi , τi+1 are consecutive if and only if νi , νi+1 are consecutive. (For 1 ≤ i ≤ n − 1.)
(3) If yst , yuv are the leftmost percolating elements occurring in λi , νi respectively, then t = u.
Then we can apply expansion moves to the respective special forms to obtain proper parameters τ1′ , ..., τn′
and ν1′ , ..., νn′ such that in addition to the above, the following holds.
(4) For each 1 ≤ i ≤ n, the word length of τi′ , λ′i is the same.
Let
′
yst11 ...ystkk = τ{1,...,n}
′
yuv11 ...yuukk = ν{1,...,n}
Given these conditions, it is an elementary exercise in the group F to construct an element f ∈ F such
that f (si ) = ui for each 1 ≤ i ≤ k. It follows that C1 · f = C2 . Since the number of combinatorial
conditions in (1) − (3) above is finite, we are done.
9. X is simply connected.
In this section we will show that X is simply connected. This is shown by providing an explicit homotopy
for a given loop to a trivial loop. Recall that in [16], we provide a procedure that takes as input a
word that represents the identity, and using the moves described in the preliminaries in Definition 3.16,
40
YASH LODHA
converts it into the empty word. A direct consequence of the result is that we can use the same procedure
to do the following. If we input a word in the infinite generating set
{yv , xu | u, v ∈ 2<N , v 6= 0k , 1k , ∅}
of G that represents an element of the subgroup F , then using the moves we can convert it into a word
in the generators
{xu | u ∈ 2<N }
In what follows the reader is not expected to have the knowledge of this procedure, but needs to assume
that indeed it is possible to do the above using the combinatorial moves described in the preliminaries.
We provide a visual interpretation of the combinatorial moves inside the complex. We will show that
each move provides a homotopy between two paths, and combine this with the existence of the procedure
to demonstrate that X is simply connected.
Given a loop l in X, (up to homotopy) we can assume that l is a path along the 1-skeleton that is
described as a path along a sequence of nodes
F τ, F (νk τ ), F (νk−1 νk τ ), ..., F (ν1 ...νk τ ), F τ
where
(1) τ is a Y -word.
(2) ν1 , ..., νk are special forms.
(3) F (νk ...ν1 ) = F ψ for a special form ψ.
Note that in general, the special forms ν1 , ..., νk are completely arbitrary. The product word ν1 ...νk
(viewed as a word obtained by concatenating ν1 , ..., νk ) may not be a special form, and pairs νi , νj are
not necessarily independent. However, since the last two nodes in the sequence are connected by an edge
we deduce that F ν1 ...νk = F ψ where ψ is a special form.
Let l1 be the loop l · τ −1 . By construction of X, l is homotopic to the trivial loop if and only if l1 has
this property. l1 is described as a path along a sequence of nodes:
F, F (νk ), F (νk−1 νk ), ..., F (ν1 ...νk ), F
where as above:
(1) ν1 , ..., νk are special forms.
(2) F (νk ...ν1 ) = F ψ for a special form ψ.
We describe a list of homotopies that correspond to moves in the analysis of standard forms. We list
them below. In what follows, τ ∈ G is an arbitrary element.
−1
(1) The expansion move yσ → xσ yσ0 yσ10
yσ11 corresponds to a homotopy between the paths F τ, F (yσ τ )
and
−1
−1
F τ, F (yσ11 τ ), F (yσ10
yσ11 τ ), F (yσ0 yσ10
yσ11 τ )
−1
−1
Similarly the expansion move yσ−1 → x−1
σ yσ00 yσ01 yσ1 corresponds to a homotopy between paths
−1
F τ, F (yσ τ ) and
−1
−1
−1
−1
F τ, F (yσ1
τ ), F (yσ01 yσ1
τ ), F (yσ00
yσ01 yσ1
τ)
These paths are homotopic in X because they are homotopic in a 3-cluster of X. For instance,
−1
the first homotopy is performed in the cluster with base F τ and parameters yσ0 , yσ10
, yσ11 .
t
(2) Consider an edge of the form F τ, F (ys f τ ). This is an edge since it satisfies (E2) of Definition
6.6, which can be seen by performing an expansion followed by rearrangement move on yst f .
Performing expansion moves on yst we obtain a special form yst11 ...ystnn equivalent to yst such that
f acts on s1 , ..., sn . The single edge path F τ, F (yst f τ ) is homotopic to the path described by the
sequence
y tn f τ ), ..., F (yst11 ...ystnn f τ )
F τ, F (ystnn f τ ), F (ystn−1
n−1 sn
A NONAMENABLE TYPE F∞ GROUP
41
in the n-cluster described with base F τ and parameters yst11 , ..., ystnn . Applying the rearranging
move
yst11 ...ystnn f = f yft1(s1 ) ...yftn(sn ) = yrt11 ...yrtnn
produces a new description of this path as
y tn τ ), ..., F (yrt11 ...yrtnn τ )
F, F (yrtnn τ ), F (yrtn−1
n−1 rn
(3) A commuting move yst ypq = ypq yst for t, q ∈ {−1, 1} and s, p independent, corresponds to a homotopy between paths of the form
F τ, F (yst τ ), F (ypq yst τ )
and
F τ, F (ypq τ ), F (yst ypq τ )
This can be performed in the complex since these paths are homotopic in the 2-cluster described
with base F τ and parameters yst , ypq .
(4) A cancellation move ys±1 ys∓1 → ∅ corresponds to shrinking a path of the form
F τ, F (ys±1 τ ), F (ys∓1 ys±1 τ )
to a trivial path. This is possible since the path is obtained by traversing the edge F τ, F (ys±1 τ )
forwards and then backwards.
Now we observe the following.
Lemma 9.1. Consider a loop l described by the path
F, F (λn ), F (λn−1 λn ), ..., F (λ1 ...λn ), F
where λ1 , ..., λn are special forms. It is homotopic to a loop described by a path of the form
), F
y tm ), ..., F (yst11 ...ystm
), F (ystm−1
F, F (ystm
m
m−1 sm
m
Proof. We show this by induction on n. For n = 1, let λ1 = ypq11 ...ypqkk . Now the 1-cell {F τ, F (λ1 τ )} is the
cross diagonal 1-cell of the k-cluster at τ parametrized by the special forms ypq11 , ..., ypqkk . It follows that
this 1-cell is homotopic to the path
F τ, F (ypqkk τ ), F (ypqk−1
y qk τ ), ..., F (ypq11 ...ypqkk τ )
k−1 pk
The inductive step is essentially the same as the base case, since by the inductive hypothesis we replace
the path
F τ, F (λn τ ), F (λn−1 λn τ ), ..., F (λ2 ...λn τ ), F (λ1 ...λn τ )
by a path
F τ, F (yuvll τ ), F (yuvl−1
y vl τ ), ..., F (λ1 yuv11 ...yuvll τ )
l−1 ul
and then argue the last edge
F (yuv11 ...yuvll τ ), F (λ1 yuv11 ...yuvll τ )
traversed in the path is homotopic to path of a suitable sequence of edges as in the base case.
Proposition 9.2. The complex X is simply connected.
Proof. Let L be a loop described as a path in X (1) . By considering the group action we can assume that
the loop L begins and ends at F . By the previous Lemma this loop is homotopic to a loop of the form
y tn ), ..., F (yst11 ...ystnn ) = F
F, F (ystnn ), F (ystn−1
n−1 sn
Since the 0-cells F = F (yst11 ...ystnn ) it follows that the word yst11 ...ystnn represents an element of F . From [16]
we know that this word can be reduced to an X-word by applying a sequence of expansion, commuting,
cancellation and rearranging substitutions. Since an application of each substitution produces a loop
homotopic to L, we observe that this process provides an explicit homotopy between L and the trivial
loop.
42
YASH LODHA
10. X is aspherical
In this section, we shall demonstrate that the complex X is a nonpositively curved cluster complex in
the sense of Definition 4.6, and hence is aspherical. In particular, we will show the following. For each
finite subcomplex Y of X which is a union of clusters, there is a subcomplex Y of X such that:
(1) Y ⊂ Y.
(2) Y is homeomorphic to a non positively curved cube complex.
We describe a process that takes as an input Y as produces as an output Y. There are two main structural
concepts defined and used in this section. These are the notions parallel 1-clusters and orthogonal 1clusters. A key algorithmic ingredient is what we refer to as amplification of clusters. An amplification
takes as an input a cluster, and produces as an output a larger cluster that contains the input as a
subcluster.
In this section whenever we refer to a description of a 1-cluster with base F τ , it shall be assumed that τ
is a Y -word (possibly empty) (unless specified otherwise). We shall write F τ, F ντ to denote a 1-cluster
with base F τ and parameter ν. It is important that the reader keeps this in mind, since this will be
crucial to formulate the definitions in this section.
10.1. The parallel equivalence relation on the 1-cells of X. Consider the regular Euclidean cube
[0, 1]n ⊂ Rn . Two 1-dimensional faces, i.e. edges of the cube are parallel if they are segments of parallel
lines in Rn . Now consider an n-cluster C, viewed as a CW subdivision of [0, 1]n . The facial 1-subclusters
of C are in natural correspondence with the closed edges of [0, 1]n . So we may view a pair of facial
1-subclusters of C as parallel if the corresponding closed edges of [0, 1]n are parallel. More concretely, let
C be described with base F τ and parameters τ1 , ..., τn . The set of facial 1-subclusters of C consists of:
{F τX τ, F τi τX τ | X ∪ {i} ⊆ {1, ..., n}, i ∈
/ X}
We say that two facial 1-subclusters of C are parallel, if they are of the form:
F τX τ, F τi (τX τ )
F τY τ, F τi (τY τ )
where X, Y ⊆ {1, ..., i − 1, i + 1, ..., n}.
We would like to extend this notion of parallel 1-clusters to pairs that are not necessarily facial 1subclusters of the same cluster. As a motivating example, consider the situation below.
Let C be as above, but for simplicity of notation, let us assume that τ is the empty word, and hence C
is described with base as the trivial coset and parameters τ1 , ..., τn . Let
e1 = F τX , F τi τX
e2 = F τY , F τi τY
be parallel, facial 1-subclusters of C.
Now consider a cluster D described with base F τY and parameters ν1 , ..., νm such that νj = τi for some
1 ≤ j ≤ m. Consider a 1-cell
e3 = F (νZ τY ), F νj (νZ τY )
of D such that Z ⊆ {1, ..., j − 1, j + 1, ..., m}. Note that in this description of D
e2 = F τY , F νj τY
It follows that e2 , e3 are parallel, facial 1-subclusters of the cluster D. It is natural to view e1 , e3 above as
in a certain sense, parallel. Indeed, there is a natural notion of parallel 1-clusters in X. This is precisely
the transitive closure of the relation obtained by declaring 1-clusters e1 , e2 as parallel if they are parallel,
facial 1-subclusters of a cluster.
We provide a more useful, concrete formulation of this definition below. In this formulation, it is not
immediately apparent that the resulting relation is an equivalence relation. This will be shown in a
subsequent Lemma.
A NONAMENABLE TYPE F∞ GROUP
43
Definition 10.1. Two 1-clusters e1 , e2 in X are said to be parallel if they admit descriptions
e1 = F τ1 , F ντ1
e2 = F τ2 , F ντ2
so that there is a sequence of special forms ν1 , ..., νn satisfying:
(1) νi , ν are independent for each 1 ≤ i ≤ n.
(2) ν1 ...νn = τ2 τ1−1 .
Remark 10.2. In the above definition, it is not required that ν1 , ..., νn are independent, sorted etc. In
fact in most cases they shall not be independent.
Note that the definition is symmetric. The special forms ν1−1 , ..., νn−1 satisfy that νi−1 , ν are independent
and νn−1 ...ν1−1 = τ1 τ2−1 . Also, note that the definition is G-invariant. That is, if e1 , e2 are parallel then
it follows from the definition that e1 · g, e2 · g are parallel for any g ∈ G.
Finally, note that if the descriptions
e1 = F τ1 , F ντ1
e2 = F τ2 , F ντ2
satisfy Definition 10.1, then so do the descriptions
e1 = F (ντ1 ), F ν −1 (ντ1 )
e2 = F (ντ2 ), F ν −1 (ντ2 )
In fact, the same ν1 , ..., νn as above witness the conditions of the Definition. Indeed it holds that νi , ν −1
are independent for each 1 ≤ i ≤ n. And also,
ν1 ...νn = νν −1 (ν1 ...νn ) = ν(ν1 ...νn )ν −1 = ν(τ2 τ1−1 )ν −1 = (ντ2 )(ντ1 )−1
Before we show that the parallel relation is an equivalence relation, we show that the definition is independent of the choice of description.
Lemma 10.3. Let e1 , e2 be parallel 1-clusters in X. Let e1 = F τ1 , F ντ1 be any choice of description.
Then there is a description e2 = F τ2 , F ντ2 , so that these descriptions satisfy Definition 10.1.
Proof. We shall find a Y -word τ2 such that e2 admits a description e2 = F τ2 , F ντ2 satisfying the required
property. Since e1 , e2 are parallel, by definition there are descriptions
e1 = F ψ1 , F ηψ1
e2 = F ψ2 , F ηψ2
and special forms η1 , ..., ηn satisfying that ηi , η are independent and η1 ...ηn ψ1 = ψ2 . Thanks to the last
paragraph of Remark 10.2, we can also find the above descriptions keeping the same base coset, so that
F ψ1 = F τ1 .
There is an f ∈ F such that f ψ1 = τ1 . Let η ′ , η1′ , ..., ηn′ be special forms obtained by applying expansion
moves on η, η1 , ..., ηn respectively such that for each percolating element yst that occurs, f −1 acts on s.
Let η ′′ , η1′′ , ..., ηn′′ be special forms obtained by replacing each percolating element yst of η ′ , η1′ , ..., ηn′ by
yft −1 (s) .
It follows that
F ηψ1 = F η ′ (f −1 f )ψ1 = F η ′′ τ1
and hence e1 = F τ1 , F η τ1 . Moreover,
′′
ψ2 = η1 ...ηn ψ1 = η1 ...ηn (f −1 f )ψ1 = f −1 η1′′ ...ηn′′ τ1
and hence
η1′′ ...ηn′′ τ1
We define τ2 = f ψ2 . It follows that τ2 =
So in particular τ2 is expressible as a Y -word, which
we fix and refer to as τ2 for the remainder of the proof.
f ψ2 =
η1′′ ...ηn′′ τ1 .
Now F η ′′ τ1 = F ντ1 and so F η ′′ = F ν. Since ν, η ′′ are special forms that lie in the same F -coset, ν can
be obtained from η ′′ by a sequence of expansion and contraction moves and hence has the same support.
In particular, it follows that the pairs ηi′′ , ν are independent. Therefore, the descriptions
F τ1 , F ντ1
F τ2 , F ντ2
44
YASH LODHA
satisfy the condition of Definition 10.1 with the special forms η1′′ , ..., ηn′′ witnessing the conditions.
Lemma 10.4. The parallel 1-cluster relation is an equivalence relation. The relation is precisely the
transitive closure of the relation obtained by declaring 1-clusters e1 , e2 as parallel if they are parallel,
facial 1-subclusters of a cluster.
Proof. In remark 10.2 we discussed why Definition 10.1 is symmetric. Therefore, it suffices to show that
this is transitive, i.e. given 1-cells e1 , e2 , e3 such that e1 , e2 and e2 , e3 are parallel, then e1 , e3 are parallel.
Fix a description F τ1 , F ντ1 of e1 . By two applications of Lemma 10.3, first on e1 , e2 and then on e2 , e3
we find descriptions
e1 = F τ1 , F ντ1
e2 = F τ2 , F ντ2
e3 = F τ3 , F ντ3
that satisfy the following conditions of Definition 10.1:
(1) There are special forms η1 , ..., ηn such that each pair ν, ηi is independent and η1 ...ηn = τ1 τ2−1 .
(2) There are special forms ψ1 , ..., ψm such that each pair ν, ψi is independent and ψ1 ...ψm = τ2 τ3−1 .
It follows that e1 , e3 are parallel since each pair ν, ηi and ν, ψi is independent and
η1 ...ηn ψ1 ...ψm = τ1 τ2−1 τ2 τ3−1 = τ1 τ3−1
This finishes the proof that our relation is an equivalence relation.
We leave the claim concerning the transitive closure as a pleasant exercise for the reader. This requires
translating the definition into a sequence of clusters witnessing the transitive closure. This statement will
not be used in the rest of the article, and is stated to provide motivation for the concept.
Lemma 10.5. The set of facial 1-subclusters of an n-cluster comprise of exactly n parallel equivalence
classes. A set of representatives for these equivalence classes is given by the set of facial 1-subclusters
incident to any given 0-cell in the cluster.
Proof. Consider an n-cluster C described with base at F τ and parameters ν1 , ..., νn . By G-invariance of
Definition 10.1, we can assume that τ is the empty word. First note that each facial 1-subcluster in C is
parallel to a facial 1-subcluster of the form F, F νi for some 1 ≤ i ≤ n. We will show that a pair
e1 = F, F νi
e2 = F, F νj
is not parallel if i 6= j. Assume by way of contradiction that e1 , e2 are parallel. By Lemma 10.3 there is
a description e2 = F ψ, F νi ψ, where ψ is a Y -word, such that the descriptions
F, F νi
F ψ, F νi ψ
satisfy Definition 10.1. So there are special forms η1 , ..., ηm such that ηk , νi are independent for each
1 ≤ k ≤ m, and η1 ...ηm = ψ.
Since we are considering two different descriptions of e2 :
e2 = F, F νj
e2 = F ψ, F νi ψ
we have to consider two possibilities:
(1) F = F ψ and F νj = F νi ψ.
(2) F = F νi ψ and F νj = F ψ.
Before we analyse each case separately, we observe the following. Since η1 ...ηm = ψ, and for each
1 ≤ j ≤ m, ηj , νi are independent, it follows that:
F νi ψ = F νi (η1 ....ηn ) = F (η1 ...ηn )νi = F ψνi
Case 1: Since F = F ψ, we conclude that ψ ∈ F (even though ψ is formally a Y -word, it is an element of
F ). It follows that
F νi ψ = F ψνi = F νi
But this means that F νj = F νi . This leads to a contradiction since νi , νj are independent special forms
and so the element νj νi−1 ∈
/ F.
A NONAMENABLE TYPE F∞ GROUP
45
Case 2: In this case, observe that:
F = F νi ψ = F ψνi = F νj νi
This means that F = F νj νi , which is a contradiction since νi , νj are independent special forms, so the
element νj νi ∈
/ F.
This means that our assumption that e1 , e2 are parallel must be false, and hence we conclude that the
set of facial 1-subclusters incident to the trivial coset F in C are pairwise non-parallel.
While working with finite sets of (parallel) 1-clusters, it will often be useful to consider descriptions that
share the same parameter.
Definition 10.6. (Common description) Consider the descriptions:
e1 = F τ1 , F ντ1
e2 = F τ2 , F ντ2
...
en = F τn , F ντn
This shall be denoted as a common description for the 1-cells e1 , ..., en , and ν is denoted as a common
parameter.
Remark 10.7. If a set of 1-clusters e1 , ..., en are pairwise parallel, then by repeated application of Lemma
10.3, as in the proof of Lemma 10.4, we can find a common description for these 1-clusters satisfying
Definition 10.1.
10.2. Orthogonal 1-clusters. As discussed before, our proof of asphericity involves finding a subcomplex which is homeomorphic to a nonpositively curved cube complex. The idea behind proving nonpositive
curvature is Gromov’s link condition, and we shall establish the link condition by means of the following
notion of orthogonality of 1-clusters.
Definition 10.8. (Orthogonal 1-clusters) Let e1 , e2 be 1-clusters that are both incident to a 0-cell u.
They are said to be orthogonal if they admit descriptions:
e1 = F τ, F ν1 τ
e2 = F τ, F ν2 τ
such that F τ = u and ν1 , ν2 are independent special forms. This is equivalent to requiring that there is
a 2-cluster C which contains e1 , e2 as facial 1-subclusters incident to u.
A list e1 , ..., en of 1-clusters that are all incident to a 0-cell u are said to be orthogonal if they are pairwise
orthogonal.
Remark 10.9. The notion of orthogonality is clearly G-invariant, i.e. e1 , e2 are orthogonal if and only if
e1 · g, e2 · g are orthogonal for any g ∈ G. We leave it as an elementary exercise for the reader to check
that the notion does not depend on the choice of descriptions, as long as the base is the same. In this
case, the parameters shall automatically be independent.
Lemma 10.10. Let e1 , ..., en be 1-clusters that are all incident to a 0-cell u. If they are orthogonal, then
there is an n-cluster that contains e1 , ..., en as facial 1-cells.
Proof. By G-invariance of the notion of orthogonality, we can assume without loss of generality that u is
the trivial coset F . Now we find descriptions for our 1-clusters with base F :
e1 = F, F τ1
....
en = F, F τn
for special forms τ1 , ..., τn . Since e1 , ..., en are pairwise orthogonal, it follows that τ1 , ..., τn are pairwise
independent. So the cluster C can be described with base F and parameters τ1 , ..., τn .
The above Lemma provides a bijective correspondence between the set of n-clusters incident to a given
0-cell u, and the set of n-tuples of pairwise orthogonal 1-clusters incident to u.
{ n-clusters incident to u} ↔ {n-tuples of orthogonal 1-clusters incident to u}
This discussion motivates the following definition.
46
YASH LODHA
Definition 10.11. A pre-cluster C consists of 1-clusters e1 , ..., en that are all incident to a 0-cell u, and
are orthogonal. The 0-cell u is said to be the base of the pre-cluster C.
Pre-clusters admit descriptions just like clusters. For instance, a pre-cluster C described with base F τ ,
and parameters τ1 , ..., τn consists of the 1-clusters:
F τ, F τ1 τ
...
F τ, F τn τ
with base the 0-cell given by F τ . The list of parameters τ1 , ..., τn is always assumed to be a sorted list
for convenience.
The closure C̄ of the n-pre-cluster C is the cluster described with base F τ and parameters τ1 , ..., τn .
Clearly, the 1-clusters in C are all facial 1-subclusters of C̄, and provide a set of representatives of the
parallel equivalence classes of the facial 1-subclusters of C̄.
Note that in the above definition we assume that a 1-cluster is also a 1-pre-cluster, whose closure is itself.
10.3. Amplification of pre-clusters. An amplifying move is an operation that takes as an input a
pre-cluster C1 and produces as an output a “larger” pre-cluster C2 such that C̄1 is a subcluster of C̄2 .
The idea behind the amplifying move is elementary, yet technical difficulties in explaining and implementing the move on a given pre-cluster arise from the fact that we need to choose a description. We will show
that in a certain precise sense, the concept is independent of the choice of description. We first provide
an elementary example to illustrate the motivation behind this concept.
Example 10.12. Consider the 1-cluster F, F y10 . We can apply an expansion move on the special form
−1
y10 to obtain the special form y100 y1010
y1011 . Consider the pre-cluster C1 consisting of the 1-clusters
F, F y100
−1
F, F y1010
F, F y1011
−1
The closure is a 3-cluster C̄1 described with base F and parameters y100 , y1010
, y1011 . This cluster contains
the 1-cluster F, F y10 as a cross diagonal.
−1
Now consider the expansion move y100 → y1000 y10010
y10011 . Consider the 5-pre-cluster C2 comprising of
−1
−1
{F, F τ | τ ∈ {y1000 , y10010
, y10011 , y1010
, y1011 }}
Then C̄2 is described with base F and parameters
−1
−1
y1000 , y10010
, y10011 , y1010
, y1011
and contains C̄1 as a diagonal subcluster, and also the 1-cluster F, F y10 as a cross diagonal. Continuing in
this fashion, we can produce arbitrarily large (pre)-clusters, each new cluster (in the closure) containing
the previous ones as subclusters.
Definition 10.13. (Amplifying a 1-cluster) Consider a 1-cluster C1 = F τ, F ντ with base u = F τ . Each
sorted list of special forms ν1 , ..., νn satisfying F ν = F ν1 ...νn , determines an amplification of C1 at u,
which produces a pre-cluster C2 consisting of:
F τ, F ν1 τ
F τ, F ν2 τ
...
F τ, F νn τ
Such an amplification is informally denoted as C1 → C2 . The 1-clusters of C2 shall be referred to as the
offsprings of the 1-cluster C1 .
Remark 10.14. Note that the definition is G-invariant in the following sense. Let C1 , C2 be as above. For
any g ∈ G, the pre-cluster C2 · g consisting of
F (τ g), F ν1 (τ g)
F (τ g), F ν2 (τ g)
...
F (τ g), F νn (τ g)
is produced by an amplification of C1 · g = F (τ g), F ν(τ g) given by F ν = F ν1 ...νn .
Now we show that the same amplifying move can be obtained from any description of a given 1-cluster,
with the same base coset. In this sense, the notion is independent of the choice of description.
A NONAMENABLE TYPE F∞ GROUP
47
Lemma 10.15. Let C1 be a 1-cluster F τ, F ντ . Let C2 be the amplification consisting of
e1 = F τ, F ν1 τ
...
en = F τ, F νn τ
where ν1 , ..., νn is a sorted list of special forms such that F ν = F ν1 ...νn .
Consider a different description C1 = F ψ, F ηψ, so that F ψ = F τ and η is a special form. Then there is
a sorted list of special forms η1 , ..., ηn such that:
(1) F η = F η1 ...ηn .
(2) ei = F ψ, F ηi ψ for each 1 ≤ i ≤ n.
Proof. Since F ψ = F τ , there is an f ∈ F such that ψ = f τ . Let νi′ be a special form obtained by
applying expansion moves on νi such that for each percolating element yst of νi′ , f −1 acts on s. Let νi′′
be the special form obtained by replacing each percolating element yst of νi′ with yft −1 (s) .
It follows that
F ν1 ...νn τ = F ν1′ ...νn′ τ = F ν1′ ...νn′ f −1 (f τ )
= F ν1′′ ...νn′′ (f τ ) = F ν1′′ ...νn′′ ψ
In particular F ν1′′ ...νn′′ = F η. By Lemma 7.2 we know that ν1′′ , ..., νn′′ is a sorted list of special forms.
Since
ei = F τ, F νi τ = F ψ, F νi′′ ψ
we conclude the statement of the Lemma with ηi = νi′′ .
We obtain the following corollary.
Corollary 10.16. Given a 1-cluster, and a base 0-cell incident to it, the set of pre-clusters obtained
by amplifying moves on a given description (with the given 0-cell as base) is the same as the set of
pre-clusters obtained by amplifying moves on any other description (with the given 0-cell as base).
To conclude our discussion, we show that the set of clusters obtained from taking closures of amplifications
of a 1-cluster are the same regardless of the choice of base 0-cell and description.
Lemma 10.17. Let C1 be a 1-cluster with incident 0-cells u, v. If C2 is a precluster obtained by performing an amplification of C1 at u, then there is a precluster C3 obtained by performing an amplification of
C1 at v such that C¯3 = C¯2 .
Proof. Let C1 be a 1-cluster F τ, F ντ . Let C2 be the amplification consisting of
F τ, F ν1 τ
...
F τ, F νn τ
where ν1 , ..., νn is a sorted list of special forms such that F ν = F ν1 ...νn . The closure C¯2 is the cluster
described with base F τ and parameters ν1 , ..., νn .
Now consider the description C1 = F ν −1 (ντ ), F (ντ ) with base coset F (ντ ) and parameter ν −1 . Since ν
is equivalent to ν1 ν2 ...νn , it follows from applying Lemma 6.22 that ν −1 is equivalent to ν1−1 ν2−1 ...νn−1 .
Now consider the amplification C3 given by
F (ντ ), F ν1−1 (ντ )
...
F (ντ ), F νn−1 (ντ )
The closure C¯3 is the cluster described with base F (ντ ) and parameters ν1−1 , ..., νn−1 .
Clearly, C¯2 = C¯3 since the above descriptions are different descriptions for the same cluster.
Now we define the amplifying move for an n-pre-cluster for n > 1. An amplifying move, or an amplification
of a pre-cluster produces another pre-cluster.
48
YASH LODHA
Definition 10.18. (Amplifying an n-pre-cluster) Consider a pre-cluster C1 with base u. Let E ⊆ C1 be
a set of 1-clusters E = {e1 , ..., ek }. An amplification C1 → C2 along E is given by replacing e1 , ..., ek in
C1 by amplifications D1 , ..., Dk of e1 , ..., ek respectively, at u.
More concretely
C2 = (C1 \ {e1 , ..., ek })
[
D1 ∪ ... ∪ Dk
We say that C2 is obtained from amplifying C1 along e1 , ..., ek . The output is also a pre-cluster with the
same base.
Let us consider a concrete example.
Example 10.19. Let C1 be a pre-cluster described as
e1 = F τ, F τ1 τ
...
en = F τ, F τn τ
Fix j, k ∈ {1, ..., n}, j 6= k. Let ν1 , ..., νl and ψ1 , ..., ψm be a sorted lists such that
F τj = F ν1 ...νl
F τk = F ψ1 ...ψm
Given this data, we now describe a (n + l + m − 2)-pre-cluster C2 which is an amplification of C1 along
{ej , ek }. This pre-cluster C2 is the following union:
[
[
({e1 , ..., en } \ {ej , ek }) {F τ, F νi τ | 1 ≤ i ≤ l} {F τ, F ψi τ | 1 ≤ i ≤ m}
Note that Definition 10.18 is G-invariant in the same sense as Definition 10.13. Also, this is independent
of the choice of description just as in 10.13. The above discussion combined with an analogue of Lemma
10.15 provides the following.
Lemma 10.20. Let C be a pre-cluster endowed with two different descriptions. Then the set of preclusters obtained from performing amplifying moves using one description is the same as the set of clusters
obtained from performing amplifying moves using the other description.
Proof. This proof is almost verbatum the same as the proof of Lemma 10.15.
Remark 10.21. Let C2 be an amplification of a pre-cluster C1 , and let C3 be an amplification of C2 . Then
C3 can be obtained from a single amplification of C1 .
We now extend the notion of amplification of a 1-cluster to a set of pairwise parallel 1-clusters.
Definition 10.22. (Common amplification for parallel 1-clusters) Let e1 , e2 be parallel 1-clusters. Let
e1 = F τ1 , F ντ1
e2 = F τ2 , F ντ2
be two descriptions satisfying the conditions of Definition 10.1. Let ν1 , ..., νn be a sorted list of special
forms such that F ν = F ν1 ...νn .
Now consider the amplification of e1 , e2 to obtain preclusters C1 , C2 respectively, as:
(1) C1 consists of
F τ1 , F ν1 τ1
F τ1 , F ν2 τ1
...
F τ1 , F νn τ1
F τ2 , F ν1 τ2
F τ2 , F ν2 τ2
...
F τ2 , F νn τ2
(2) C2 consists of
Such an amplification of parallel 1-clusters is said to be a common amplification.
A second variant of this definition is considered for the description e2 = F ντ, F ν −1 (ντ ). First, note that
since F ν = F ν1 ...νn , it follows from applying Lemma 6.22 that F ν −1 = F ν1−1 ...νn−1 . Now consider the
amplification C1 of e1 with base F τ1 as before. The corresponding common amplification of e2 with base
F (ντ2 ) is C2′ which consists of:
F (ντ2 ), F ν1−1 (ντ2 )
Note that C¯2 = C¯2′ .
F (ντ2 ), F ν2−1 (ντ2 )
...
F (ντ2 ), F νn−1 (ντ2 )
A NONAMENABLE TYPE F∞ GROUP
49
Remark 10.23. Note that in the above definition, for each 1 ≤ i ≤ n, the 1-clusters
F τ1 , F νi τ1
F τ2 , F νi τ2
are parallel. In the second variant of the definition the 1-clusters
F τ1 , F νi τ1
F (ντ2 ), F νi−1 (ντ2 )
are parallel
The above definition can be extended to any finite number of 1-clusters that are pairwise equivalent. This
is done by first finding common descriptions that satisfy the conditions of Definition 10.1, using Lemma
10.3. And then we perform common amplifications as above.
Definition 10.22 has a natural extension to pre-clusters.
Definition 10.24. (Common amplification of pre-clusters along parallel subsets) Let C1 , C2 be preclusters based at u, v respectively. Let E1 ⊆ C1 , E2 ⊆ C2 be sets of 1-clusters such that:
(1) E1 = {e1 , ..., ek } and E2 = {e′1 , ..., e′k }.
(2) ei , e′i are parallel for each 1 ≤ i ≤ k.
Fix an amplification C1 → C1′ along E1 by
e1 → D1
...
ek → Dk
at v.
We define a common amplification C2 → C2′ using the amplifications
e′1 → D1′
...
e′k → Dk′
at u where e′i → Di′ and ei → Di are common amplifications for each 1 ≤ i ≤ k.
We denote such a common amplification as:
C1 → C1′
C2 → C2′
along E1 , E2
Now we define a notion that is in a certain strong sense the opposite of the notion of parallel.
Definition 10.25. A pair of 1-clusters e1 , e2 is said to be disparate if the following holds:
(1) e1 , e2 are not parallel.
(2) Any pair of amplifications
e1 → C1
e2 → C2
satisfy that for each pair e′1 ∈ C1 , e′2 ∈ C2 , e′1 , e′2 are not parallel.
As a simple example of disparate 1-clusters, consider any pair of orthogonal 1-clusters incident to a
common 0-cell. We now demonstrate that being disparate is a property of parallel equivalence classes of
1-clusters.
Lemma 10.26. Given 1-clusters e1 , e2 , e3 such that e1 , e2 are parallel, and e1 , e3 are disparate, it also
holds that e2 , e3 are disparate.
Proof. To see this, observe that if e2 , e3 were not disparate, then we can perform amplifications to obtain
pre-clusters C2 , C3 such that there is a pair e′2 ∈ C2 , e′3 ∈ C3 such that they are parallel. However,
performing the common amplification for e1 , as for e2 , we obtain a pre-cluster C1 with e′1 ∈ C1 which is
parallel to e′2 and hence e′3 . This means that e1 , e3 were not disparate to begin with.
Lemma 10.27. Let e1 , e2 be disparate, and let C be an amplification of e1 . Then for each e3 ∈ C, e2 , e3
are disparate.
Proof. This follows immediately from the definitions.
Lemma 10.28. Let C be a cluster. Any pair of facial 1-cells of C is either parallel or disparate.
50
YASH LODHA
Proof. Let C be a cluster described with base at F τ , and parameters τ1 , ..., τn . Thanks to Lemma 10.5
and Remark 10.26 it suffices to show that any pair of 1-clusters
e1 = F τ, F τi τ
e2 = F τ, F τj τ
is disparate if i 6= j. From Lemma 10.5 we know that they are not parallel. Now if C1 , C2 are preclusters obtained from amplifications of e1 , e2 respectively (performed at the base F τ ), then for each
e ∈ C1 , e′ ∈ C2 the pair e, e′ is orthogonal. It follows again from Lemma 10.5 that e, e′ are not parallel.
10.4. Weakly balanced systems of pre-clusters.
Definition 10.29. Consider a pair e, u comprising of a 1-cluster e and a 0-cell u. e, u are said to be
disparate if for any e′ incident to u, the 1-clusters e, e′ are disparate. e, u are said to be compatible, if
there is a 1-cluster e′ incident to u such that e, e′ are parallel.
Note that this definition is G-invariant. So if e, u is compatible or disparate, then for each g ∈ G, e · g, u · g
also satisfies the same. This follows from the fact that the definition of parallel and disparate 1-clusters
is G-invariant.
The following elementary observation follows immediately from the definitions and Lemma 10.27.
Lemma 10.30. Let e be a 1-cluster and u be a 0-cell. Then the following holds:
(1) Let e, u be disparate (or compatible) and let e, e′ be parallel 1-clusters. Then e′ , u are disparate
(respectively, compatible).
(2) Let e, u be disparate (or compatible) and let C be a pre-cluster obtained from performing an
amplification of e. Then for each e′ ∈ C, e′ , u are disparate (respectively, compatible).
Definition 10.31. Let C1 , ..., Cn be pre-clusters with bases u1 , ..., un respectively. We say that this is a
weakly balanced system if for any pair
[
e∈
Ci
uj ∈ {u1 , ..., un }
1≤i≤n
e, uj are either disparate or compatible.
Lemma 10.32. Let C1 , ..., Cn be a weakly balanced system with bases u1 , ..., un . Let
C1 , ..., Cn → C1′ , ..., Cn′
be any amplification. Then the system C1′ , ..., Cn′ is also weakly balanced
Proof. Consider a pair e, uj where 1 ≤ j ≤ n and e ∈ Ci′ for some 1 ≤ i ≤ n. Then e is an offspring of a
1-cluster e′ ∈ Ci . Note that e′ , uj are compatible or disparate by our hypothesis. Since both properties
are inherited for the offspring e, thanks to Lemma 10.30, we are done.
Our goal in this subsection is to prove the following Proposition.
Proposition 10.33. (Weakly balancing procedure) Let C1 , ..., Cn be a system of clusters. There is an
amplification
C1 , ..., Cn → C1′ , ..., Cn′
′
′
such that C1 , ..., Cn is a weakly balanced system.
The technical core of the proof of this Proposition is the following Proposition and the more general
corollary.
Proposition 10.34. Let e be a 1-cluster and let u be a 0-cell. Then there is an amplification of e that
produces a pre-cluster C such that the following holds. For each e′ ∈ C, the pair e′ , u is either disparate
or compatible.
Corollary 10.35. Let e be a 1-cluster incident to a 0-cell u and let u1 , ..., un be 0-cells. Then there is
an amplification of e that produces a pre-cluster C such that the following holds. For each e′ ∈ C and
each 1 ≤ i ≤ n, the pair e′ , ui is disparate or compatible.
A NONAMENABLE TYPE F∞ GROUP
51
Before we provide proofs of the above, we study the notions of disparate and compatible for a pair e, u
in depth. Given suitable descriptions of e, we would like to detect whether e, u are disparate, or whether
there is a 1-cluster e′ incident to u such that e, e′ are parallel. To describe such situations we require
certain restrictions on descriptions of e. As a consequence of the G-invariance of Definition 10.29, we can
assume that u is the trivial coset in our arguments below.
Definition 10.36. Let F τ, F ντ be the description of a 1-cluster. We say that such a description is tame,
if τ is a Y -word in standard form with no potential cancellations and ντ is in standard form.
Remark 10.37. It is straightforward to produce tame descriptions for any given 1-cluster. For instance,
choose a normal form representative for the base coset, and choose the parameter to have sufficiently
large depth (by performing expansion moves).
Consider such a tame description F τ, F ντ . We know that τ is a standard form with no potential
cancellations, however ντ is a standard form that may admit potential cancellations. Any such potential
cancellation must occur between a pair of occurrences of y ± , one of which lies in ν and the other in τ .
Lemma 10.38. Let u be the trivial coset F . Suppose that a 1-cluster e admits a tame description F τ, F ντ
that satisfies either one of the following:
(1) ν, τ have disjoint supports.
(2) The restriction of ντ on Supp(ν) equals the restriction of an element of F on Supp(ν).
Then there is a 1-cluster e′ incident to u such that e, e′ are parallel.
Proof. Assume that (1) holds. Let
τ −1 = yst11 ...ystnn
where equality denotes equality as words. We know that
Supp(τ −1 ) ∩ Supp(ν) = ∅
Since the given description of e is tame, τ is a standard form with no potential cancellations. So
Supp(ystii ) ⊂ Supp(τ )
for each 1 ≤ i ≤ n. In particular,
Supp(ystii ) ∩ Supp(ν) = ∅
for each 1 ≤ i ≤ n.
It follows that the 1-clusters
F, F ν
F τ, F ντ
t1
are parallel with the sequence of special forms ys1 , ..., ystnn witnessing the conditions of Definition 10.1.
Now we consider the case when (2) holds. We shall produce another tame description of e which will land
us in case (1). We first convert ντ into a normal form f ψ where f ∈ F and ψ is a Y -word. Recall from
Lemma 3.19 that f ψ does not preserve tail equivalence on f −1 (Supp(ψ)) and preserves tail equivalence
on the complement of this set.
Since ντ = f ψ and since ντ preserves tail equivalence on Supp(ν), it follows that
Supp(ν) ∩ f −1 (Supp(ψ)) = ∅
This means that
f (Supp(ν)) ∩ Supp(ψ) = ∅
Now consider the description of e
F (ντ ), F ν −1 (ντ )
Now let η1 be a special form obtained by performing expansion moves on ν −1 such that for each percolating
element yst of η1 , f acts on s. Let η2 be the special form obtained by replacing each yst in η1 by yft (s) .
Moreover, by further expansions on η1 we may guarantee that η2 has sufficiently large depth so that η2 ψ
is a standard form.
52
YASH LODHA
It follows that
F ν −1 (f ψ) = F η1 (f ψ) = F η2 ψ
So we get a new tame description of our 1-cluster e
F ψ, F η2 ψ
Since
Supp(η2 ) = f (Supp(η1 )) = f (Supp(ν −1 )) = f (Supp(ν))
from the above it follows that
Supp(η2 ) ∩ Supp(ψ) = ∅
Using this new description we land in Case (1) above and therefore conclude that there is a 1-cluster e′
incident to u such that e, e′ are parallel.
Lemma 10.39. Let u be the trivial coset F . Suppose that a 1-cluster e admits a tame description F τ, F ντ
satisfying that:
(1) Supp(ν) ⊆ Supp(τ ).
(2) ντ restricted to Supp(ν) does not preserve tail equivalence on a dense subset of Supp(ν).
Then e, u are disparate.
Proof. Thanks to Lemma 10.17 it suffices to consider amplifications of e with a preferred choice of base,
which will be F τ . Now assume by way of contradiction that there is an amplification C of e with base
F τ given by
F τ, F ν1 τ
...
F τ, F νn τ
(with F ν = F ν1 ...νn ) so that for some 1 ≤ l ≤ n, the pair
e′ = F τ, F νl τ
u=F
is compatible. We can also assume that νl τ is a standard form (by replacing νl with an equivalent special
form, if necessary).
In other words, there is a 1-cell e′′ incident to F such that e′ , e′′ are parallel. By Lemma 10.3, there is a
description
e′′ = F η, F νl η
for a Y -word η, and a sequence of special forms ψ1 , ..., ψm such that:
(1) ψi , νl have disjoint supports for each 1 ≤ i ≤ m.
(2) (ψ1 ...ψm )τ = η.
Now since the trivial coset F is incident to u, we have two possible cases:
Case (1) F η = F .
Case (2) F νl η = F .
Proof for Case (1): Recall from the above that (ψ1 ...ψm )τ = η. In this case it follows that F (ψ1 ...ψm )τ =
F η = F and so (ψ1 ...ψm )τ ∈ F . Since τ is a Y -word with no potential cancellations, by Lemma 3.19 it
does not preserve tail equivalence on a dense subset of its support. Since Supp(ν) ⊆ Supp(τ ), this implies
that the action of τ on Supp(ν) does not preserve tail equivalence on a dense subset. In particular, the
action of τ on Supp(νl ) ⊂ Supp(ν) does not preserve tail equivalence on a dense subset.
Since ψi , νl have disjoint supports, the action of (ψ1 ...ψm )τ restricted to Supp(νl ) equals that of τ , and
hence does not preserve tail equivalence. So this cannot act as an element of F , and we get a contradiction.
Proof for Case (2): Recall from the above that (ψ1 ...ψm )τ = η and hence
νl η = νl (ψ1 ...ψm )τ = (ψ1 ...ψm )νl τ
A NONAMENABLE TYPE F∞ GROUP
53
This means that
F (ψ1 ...ψm )(νl τ ) = F νl η = F
and so (ψ1 ...ψm )(νl τ ) ∈ F . Since ψi , νl have disjoint supports, the action of (ψ1 ...ψm )(νl τ ) restricted to
Supp(νl ) equals the action of νl τ on Supp(νl ). If we show that the action of νl τ does not preserve tail
equivalence on Supp(νl ) we shall obtain the desired contradiction.
Recall from the above that ντ does not preserve tail equivalence on a dense subset of Supp(ν) and
hence on a dense subset of Supp(νl ). Since (ν1 ...νn ) and ν are equivalent special forms, this means that
(ν1 ...νn )τ does not preserve tail equivalence on a dense subset of Supp(νl ).
Now note that the actions of (ν1 ...νn )τ and νl τ agree on Supp(νl ) since ν1 , ..., νn are independent and
have disjoint supports. Hence νl τ does not preserve tail equivalence on Supp(νl ).
Proof of Proposition 10.34
Proof. Since the notion of parallel 1-clusters is invariant under the G action, we can assume that u is the
trivial coset F for the sake of this proof. We fix a tame description e = F τ, F ντ .
We perform a sequence of expansion moves on percolating elements of ν to obtain an equivalent special
form yst11 ...ystnn such that for each 1 ≤ i ≤ n, either Supp(ystii ) ⊂ Supp(τ ) or Supp(ystii ) ∩ Supp(τ ) = ∅. We
use this data for an amplification of e to obtain the pre-cluster:
e1 = F τ, F yst11 τ
...
en = F τ, F ystnn τ
If Supp(ystii ) ∩ Supp(τ ) = ∅, then by Lemma 10.38 ei , u are compatible.
We claim that for any given 1 ≤ i ≤ n, if Supp(ystii ) ⊂ Supp(τ ), then exactly one of the following holds:
(1) ystii τ does not preserve tail equivalence on a dense subset of Supp(ystii ).
(2) ystii τ acts like an element of F on Supp(ystii ).
It is clear that the above are mutually exclusive. We show that if (2) do not hold then (1) holds.
Since the description of e is tame, we know that τ is a standard form with no potential cancellations, and
that ντ is a standard form. Furthermore, since ystii is an offspring of a percolating element of ν, ystii τ is a
standard form.
For any infinite binary sequence of the form si σ, the associated calculation of ystii τ (after performing
cancellations in the calculation if necessary) has the same exponent. So if ystii τ does not act like F on
Supp(ystii ), then it does not preserve tail equivalence some sequence in this set. It follows that all such
calculations have exponent greater than 0. And hence it does not preserve tail equivalence on a dense
subset of Supp(ystii ) = [si 0∞ , si 1∞ ].
By applications of Lemmas 10.38 and 10.39, we conclude that in case (1) the pair ei , u is disparate and
in case (2) the pair is compatible.
Proof of Corollary 10.35
Proof. We proceed by induction on n. The base case is precisely Proposition 10.34.
Now assume that we have a pre-cluster C from an amplification e → C such that the statement holds for
C and u1 , ..., un−1 . Now apply Proposition 10.34 to each pair e ∈ C, un to obtain amplifications e → Ce
such that for each e′ ∈ Ce the pair e′ , un is disparate or compatible. Note that it is possible that e → Ce
is the trivial amplification and e = Ce .
Now consider
the amplification C → C ′ which is given by replacing each e ∈ C by Ce . In particular,
S
′
C = e∈C Ce . The pre-cluster C ′ satisfies the desired property. Since iterated amplifications of a
1-cluster can be viewed as a single amplification of the 1-cluster, we are done.
54
YASH LODHA
Proof of Proposition 10.33.
S
Proof. For each 1-cluster e ∈ 1≤i≤n Ci , using Corollary 10.35 we find an amplification e → Ce (at the
same base) such that for each ui and each e′ ∈ Ce , the pair , ui is compatible or disparate. (Note that
this could possibly be the trivial amplification, i.e. it could be that Ce = e.)
Our amplification
Ci → Ci′ is then the amplification obtained by replacing each e ∈ Ci with Ce , i.e.
S
Ci′ = e∈Ci Ce .
10.5. Balanced systems of clusters.
Definition 10.40.
We say that a weakly balanced system C1 , ..., Cn is a balanced system if any pair of
S
1-clusters in 1≤i≤n Ci is either parallel or disparate.
In this subsection we will provide a procedure that takes input a weakly balanced system of pre-clusters
C1 , ..., Cn and produces a balanced system of pre-clusters C1′ , ..., Cn′ such that each Ci′ is obtained by a
sequence of amplifications of Ci . Hence automatically, it follows that C̄i is a subcluster of C̄i′ for each
1 ≤ i ≤ n. Our goal now is to prove the following.
Proposition 10.41. (Balancing procedure) Let C1 , ..., Cn be a weakly balanced system of clusters in X.
We can perform a sequence of amplifications on Ci to obtain pre-clusters Ci′ such that C1′ , ..., Cn′ is a
balanced system.
Balanced systems are not as well behaved as weakly balanced systems. Performing ad hoc amplifications
of a balanced system will in general not produce a balanced system. In particular, the analog of Lemma
10.32 fails. However if we perform an amplification of a balanced system C1 , ..., Cn carefully enough, this
produces another balanced system.
Definition 10.42. (Common amplification for a balanced system) Let C1 , ..., Cn be a balanced system,
and let Ej be a set of 1-clusters in Cj for some 1 ≤ j ≤ n. Fix an amplification
Cj → Cj′
along Ej
We define amplifications
C1 → C1′
...
Cn → Cn′
as follows. For each 1 ≤ i ≤ n, i 6= j, define
Ei = {e ∈ Ci | ∃e′ ∈ Ej such that e, e′ are parallel }
Then Ci → Ci′ is defined as follows.
Let Ei = {e1 , ..., ek } and for each 1 ≤ l ≤ k let e′l ∈ Ej be the 1-cluster such that el , e′l are parallel.
Consider the amplification e′l → Dl′ which is the restriction of Cj → Cj′ to e′l . Let el → Dl be the
corresponding common amplification.
Then we define
Ci′ = (Ci \ {e1 , ..., ek })
[
Dl
1≤l≤k
The resulting amplification is denoted as
C1 , ..., Cn → C1′ , ..., Cn′
along Ej
and is said to be a common amplification for a balanced system.
Lemma 10.43. Let C1′ , ..., Cn′ be a system obtained from a common amplification of a balanced system
C1 , ..., Cn along Ej ⊆ Cj for some 1 ≤ j ≤ n. Then C1′ , ..., Cn′ is a balanced system.
A NONAMENABLE TYPE F∞ GROUP
55
Proof. First observe that by Lemma 10.32, C1′ , ..., Cn′ is weakly balanced. Now we show that it it balanced.
Consider 1-clusters
e′i ∈ Ci′ , e′l ∈ Cl′
for some 1 ≤ i, l ≤ n. Let
be 1-clusters such that
e′i , e′l
ei ∈ Ci , el ∈ Cl
are offsprings of ei , el in the amplifications
Ci → Ci′ , Cl → Cl′
respectively. (Note that it is possible that ei = e′i or el = e′l , in the case when these 1-clusters are
unchanged in the amplification.)
By our hypothesis, ei , el are either parallel or disparate. If they are disparate, then so are e′i , e′l thanks
to Lemma 10.27.
If ei , el are parallel, and e′i = ei and e′l = e′l , then e′l , e′i are parallel. If one of ei , el is amplified in the above
amplification, then there is a parallel 1-cluster ej ∈ Ej such that ei , el , ej are pairwise parallel. Following
Definition 10.42, in this case both ei , el are amplified as parallel 1-clusters in the sense of Definition 10.22.
So the union of their sets of offsprings must satisfy that each pair is parallel or disparate. In particular,
e′i , e′l are either parallel or disparate.
We now prove the balancing lemma which is an important technical step in the proof of Proposition 10.41.
Lemma 10.44. (Balancing Lemma) Let C1 , C2 be pre-clusters with a common base u. The we can
perform amplifications C1 → D1 and C2 → D2 such that D1 , D2 is a balanced system.
Proof. Note that since both clusters share the same base u, the system is trivially weakly balanced.
In this proof we shall use the word offspring for percolating elements obtained from expansion moves on
special forms, as well as elements of pre-clusters obtained from amplifying a pre-cluster. The usage will
be clear from the context.
Let u = F τ . Let C1 comprise of 1-clusters
F τ, F τ1 τ
...
F τ, F τo1 τ
F τ, F ψ1 τ
...
F τ, F ψo2 τ
and let C2 comprise of
Further, let
ψ{1,...,o2} = yuv11 ...yuvm
τ{1,...,o1 } = yst11 ...ystnn
m
where equality denotes equality of words. Note that since τ1 , ..., τo2 are independent and ψ1 , ..., ψo2 are
independent, it follows that s1 , ..., sn and u1 , ..., um are independent, and each ti , vj ∈ {1, −1}.
By the Strong expansion Lemma 6.25, there are words ypq11 ...ypqll and yrw11 ...yrwkk such that:
(1) ypq11 ...ypqll is obtained from yst11 ...ystnn by applying expansion moves as in 6.10.
by applying expansion moves as in 6.10.
(2) yrw11 ...yrwkk is obtained from yuv11 ...yuvm
m
w
w
w
(3) For each pair ypqii , yrjj it holds that either ypqii = yrjj or ypqii , yrjj contains no common offsprings.
Let D1 be given by the amplification of C1 consisting of the 1-clusters
F τ, F ypq11 τ
...
F τ, F ypqll τ
and D2 be the amplification of C2 consisting of the 1-clusters
F τ, F yrw11 τ
...
F τ, F yrwkk τ
It remains to check that the 1-clusters in each pair
F τ, F yrwii τ
F τ, F yrwjj τ
56
YASH LODHA
are either equal (hence parallel) or disparate.
w
If 1-clusters in such a pair are not equal and also not disparate, we claim that yrwii , yrjj have a common
offspring. (To see this, observe that there are amplifications of such a pair of 1-clusters using the descriptions above that have a common offspring in the sense of 10.13. Moreover, the parameters of the two
descriptions of the common offspring are equal.) This contradicts condition (3) above.
For the rest of this subsection, we fix a weakly balanced system C1 , ..., Cn . Thanks to Lemma 10.32 since
performing any amplifications on a weakly balanced system produces a weakly balanced system, we shall
omit the usage of the phrase “weakly balanced” in the remainder of the proof. It shall be evident that
all systems under consideration are weakly balanced.
Our goal will now be to perform amplifications on this system to obtain a balanced system. The proof is
an inductive argument, where the induction is performed on the number n of pre-clusters in the system
C1 , ..., Cn . Note that the base case is trivial, since 1-clusters in a pre-cluster are pairwise disparate, and
hence the system is balanced.
Inductive hypothesis A system of weakly balanced pre-clusters C1 , ..., Cn such that C1 , ..., Cn−1 is
balanced.
Inductive Step The inductive step will be done in n stages. First, declare
(0)
C1
(0)
= C1
C2
= C2
...
Cn(0) = Cn
Stage i, for 1 ≤ i ≤ n involves performing the following:
(i−1)
(i)
(1) An amplification of Cn
to obtain Cn .
(2) A common amplification of the balanced system
(i−1)
C1
(i)
(i−1)
(i)
(i)
, ..., Cn−1 → C1 , ..., Cn−1
(i)
Output at stage i: A system C1 , ..., Cn such that:
(i)
(i)
(1) C1 , ..., Cn−1 is a balanced system.
(i)
(i)
(i)
(2) C1 , ..., Ci , Cn is a balanced system.
(n)
(n)
Output at stage n: A balanced system C1 , ..., Cn .
Stage i
(i−1)
Let E = {e1 , ..., ek } be the set of 1-clusters in Cn
such that for each 1 ≤ j ≤ k, the pair ej , ui is
compatible. For each 1 ≤ j ≤ k, let e′j be a 1-cluster incident to ui such that ej , e′j are parallel. Let D1
(i−1)
be the pre-cluster with base ui that consists of the 1-clusters e′1 , ..., e′k . Note that by definition Cn
forms a balanced system.
(i−1)
Using the balancing Lemma 10.44, there are sets Ei ⊆ Ci
(i−1)
Ci
→ J1
, L ⊆ D1 and amplifications
D1 → J2 along Ei , L
such that J1 , J2 is a balanced system. Let
En = {e ∈ Cn(i−1) | ∃e′ ∈ L such that e, e′ are parallel}
We perform the common amplification
Cn(i−1) → J3
to obtain a balanced system J3 , J2 .
Lemma 10.45. J1 , J3 is a balanced system.
D1 → J2
along En , L
, D1
A NONAMENABLE TYPE F∞ GROUP
(i−1)
Proof. Since Cn
(i−1)
, Ci
57
is weakly balanced, it follows that J1 , J3 is weakly balanced.
Now for any e ∈ J3 , either one of the following holds.
(i−1)
(1) e ∈ Cn
\ En .
(2) e is an offspring of a 1-cluster e′ ∈ En .
Now given any e1 ∈ J2 we claim that e, e1 are parallel or disparate.
(i−1)
If e ∈ Cn
\ En , then e, ui are disparate, and hence e, e1 are disparate. If e is an offspring of a 1-cluster
e′ ∈ En , then there is an e′′ ∈ J2 such that e, e′′ are parallel. Now recall from the above that J2 , J3
is a balanced system. This means that either e′′ = e′ or e′′ , e′ are disparate. As a consequence e, e′ are
parallel or disparate.
Let Ei be as above, and for 1 ≤ j ≤ n − 1, j 6= i let
(i−1)
Ej = {e ∈ Cj
| ∃e′ ∈ Ei such that e, e′ are parallel }
Finally, we perform the common amplification of the system
(i−1)
C1
(i−1)
(i)
(i)
, ..., Cn−1 → C1 , ..., Cn−1
along Ei
in the sense of Definition 10.42. Note that in particular,
(i)
Ci
= J1
We set
Cn(i) = J3
To finish the proof of Proposition 10.41, it suffices to show the following:
Lemma 10.46. The system
(i)
(i)
C1 , ..., Ci , Cn(i)
is balanced.
Proof. The system
(i)
(i)
C1 , ..., Cn−1
is balanced, thanks to Lemma 10.43, since it was obtained from performing a common amplification of a
(i)
(i)
balanced system. Moreover, Ci , Cn is a balanced system thanks to Lemma 10.45 since we have
(i)
Ci
Cn(i) = J3
= J1
Consider a pair
(i)
e2 ∈ Cn(i)
e1 ∈ Cj
for some 1 ≤ j ≤ i − 1. We will show that e1 , e2 are parallel or disparate. This will finish the proof of
the Lemma.
There are two possibilities for e1 :
(i−1)
(a1) e1 ∈ Cj
\ Ej .
(i−1)
(a2) e1 is an offspring of some e′1 ∈ Cj
in the amplification
(i−1)
Cj
(i)
→ Cj
along Ej
There are two possibilities for e2 :
(i−1)
(b1) e2 ∈ Cn
\ En .
(b2) e2 is an offspring of some e′2 ∈ En in the amplification
(i−1)
Ci
(i)
→ Ci
along En
58
YASH LODHA
Case a1, b1 Our claim follows by the hypothesis at the end of Step i − 1 that the system
(i−1)
C1
(i−1)
, ..., Ci−1 , Cn(i−1)
is balanced.
Cases a2, b1 and a2, b2 In both the cases, e′1 ∈ Ej and there is a e3 ∈ Ei such that e′1 , e3 are parallel.
Recall that the amplification
(i−1)
Ci
→ J1 along Ei
D1 → J2 along L
is done using Lemma 10.44. So in particular, there is an e4 ∈ L such that e3 , e4 are neither parallel nor
disparate. It follows that e4 , e′1 are neither parallel, nor disparate.
(i−1)
(i−1)
(i−1)
are
, e′1 ∈ Cj
such that e4 , e′4 are parallel. This means that e′4 ∈ Cn
There is a e′4 ∈ En ⊆ Cn
neither parallel nor disparate. This contradicts the hypothesis at the end of Step i − 1 that the system
(i−1)
C1
(i−1)
, ..., Ci−1 , Cn(i−1)
is balanced.
Case a1, b2 By the hypothesis at the end of Step i − 1, e′2 , e1 are either parallel or disparate. If they are
disparate, so are e1 , e2 . We assume that e′2 , e1 are parallel. Also in this case, it is true that there is a
e3 ∈ L ⊆ D1 such that e′2 , e3 are parallel. In particular, e1 , e3 are parallel.
Following the amplification
(i−1)
Ci
→ J1 along Ei
D1 → J2 along L
(i−1)
such that e3 , e4 are neither parallel
done using Lemma 10.44, we observe that there is an e4 ∈ Ei ⊂ Ci
nor disparate. Since e1 , e3 are parallel, it follows that e1 , e4 are neither parallel nor disparate.
This contradicts the hypothesis that the system
(i−1)
C1
(i−1)
, ..., Cn−1
is balanced.
10.6. Balanced systems and nonpositive curvature. In this subsection we show that given a balanced system C1 , ..., Cn , there is a subcomplex C of X such that the following holds:
S
(1) 1≤i≤n C̄i is a subcomplex of C.
(2) C is homeomorphic to a nonpositively curved cube complex. In this step, we provide an explicit
such structure on C.
The main idea behind this is the following.
Definition 10.47. (Parallel closure) Let C1 , ..., Cn be a balanced system
S of pre-clusters in X. Let E be
the set of 1-clusters in X such that for each e ∈ E, there is an e′ ∈ 1≤i≤n Ci such that e and e′ are
parallel. Denote by C as the union of the following set of clusters in X:
{C is a cluster in X | Each facial 1-subcluster of C is an element of E}
We call C the parallel closure of C1 , ..., Cn .
Proposition 10.48. Let C be the parallel closure of a balanced system C1 , ..., Cn . Then the following
holds:
S
(1) 1≤i≤n C̄i ⊂ C.
(2) C is homeomorphic to a non positively curved cube complex.
Proof. Each facial 1-subcluster of C̄i is parallel to a 1-cluster in Ci . So the first claim follows immediately
from the definition of C.
A NONAMENABLE TYPE F∞ GROUP
59
We shall prove the second claim in two parts. First we demonstrate that C is homeomorphic to a cube
complex, which is locally finite and finite dimensional. Then we shall demonstrate that this is non
positively curved in the sense of Gromov. In the first part, we show that the intersection of each pair of
clusters in C is a facial subcluster of both.
First observe that if e1 and e2 are facial 1-subclusters of clusters C1 and C2 , respectively, in C, then e1 , e2
are parallel or disparate. This follows immediately from the definition of C.
Assume by way of contradiction that there are two clusters D1 , D2 in C such that D1 ∩ D2 is a diagonal
subcluster of D1 .
Now we find descriptions satisfying:
(1) D1 is described with base F τ and parameters ψ1 , ..., ψn .
(2) D2 is described with base F τ and parameters η1 , ..., ηm .
(3) There is a facial 1-subcluster e of D1 ∩ D2 , incident to F τ , which is a diagonal subcluster of D1 .
By (3) above there are sets
X ⊆ {1, ..., n}, |X| > 1
Y ⊆ {1, ..., m}
satisfying:
e = F τ, F ψX τ = F τ, F ηY τ
We claim that each pair of 1-clusters
F τ, F ψi τ
F τ, F ηj τ
is distinct for any i ∈ X, j ∈ Y . By way of contradiction, assume that
e′ = F τ, F ψi τ = F τ, F ηj τ
for some i ∈ X, j ∈ Y . This implies that e′ is a facial 1-subcluster of D1 ∩D2 since it is a facial 1-subcluster
of both D1 and D2 . This contradicts our hypothesis that e is a facial 1-subcluster of the intersection
D1 ∩ D2 since e, e′ cannot be orthogonal. (In particular, they are not disparate.)
Since F ψX = F ηY , and since both ψX , ηY are products of independent special forms, there is a Y -word
yst11 ...ystnn which can be obtained from both ηY , ψX using expansion moves. This means that there are
j ∈ X, k ∈ Y such that we can perform expansion moves on ψj , ηk respectively to obtain special forms
ψj′ , ηk′ satisfying that some ystii occurs in both ψj′ , ηk′ .
It follows that we can perform an amplification of the 1-clusters
e1 = F τ, F ψj τ
e2 = F τ, F ηk τ
to obtain pre-clusters that both contain the 1-cluster F τ, F ystii τ . This means that e1 , e2 are not disparate.
Since they are distinct 1-clusters incident to the same 0-cell, F τ, F ψj τ and F τ, F ηk τ cannot be parallel.
But since they are both facial 1-clusters of clusters in C, they must then be disparate. This is a contradiction. Therefore, our original assumption must be false, and it must be the case that D1 ∩ D2 is a facial
subcluster of both D1 and D2 .
Note that C is naturally homeomorphic to a cube complex. We shall abuse notation and also refer to C
as a cube complex itself. For the rest of the proof, each n-cluster in C is assumed to be replaced by a
regular Euclidean n-cube, facial subclusters are declared as subcubes, and the facial intersections provide
the gluing maps.
We now observe that C is locally finite and finite dimensional. Consider a 0-cell u in C. Let E be the set
of closed 1-cells incident to u in C. S
Each closed 1-cell in E, by definition, corresponds to a 1-cluster in
X which is parallel to a 1-cluster in 1≤i≤n Ci . We abuse notation and also denote this set of 1-clusters
as E.
60
YASH LODHA
Since
S any two 1-clusters in E are incident to the same 0-cell in X, they cannot be parallel. Since the
set 1≤i≤n Ci is finite, the number of 1-clusters incident to u in X that are parallel to a 1-cluster in
S
1≤i≤n Cn is finite. It follows that E is finite.
This observation implies that the cube
P complex C is locally finite, and the largest dimension of a cube in
C is bounded above by the quantity 1≤i≤n |Ci |.
Next, we demonstrate that C is nonpositively curved. It is easy to see that the link of each 0-cell in C
is a simplicial complex. We will show that the link of each 0-cell in C does not contain empty simplices,
and hence is a flag simplicial complex. To see this, let e1 , ..., en be closed 1-cells in C incident to a 0-cell
u with the property that each pair ei , ej are 1-faces of a square incident to u.
The closed 1-cells e1 , ..., en correspond to 1-clusters e′1 , ..., e′n in X with the property that each pair e′i , e′j
are facial 1-subclusters of a 2-cluster. In particular, e′1 , ..., e′n are pairwise orthogonal, and hence by
Lemma 10.10 there is an n-cluster D in X which contains e′1 , ..., e′n as facial 1-subclusters.
S
Each 1-cluster e′1 , ..., e′n is parallel to some 1-cluster in 1≤i≤n Ci , And each facial 1-subcluster of D is
parallel to some e′i . So it follows that D corresponds to a cube D′ in C, so that the 1-faces of D′ incident
to u are precisely e1 , ..., en . Therefore, the link at the vertex u in C contains an n-simplex whose vertices
correspond to e1 , ..., en .
Now we conclude the proof of asphericity.
10.7. Proof of asphericity. Any image of a sphere S n in X under a continuous map is contained in
some finite subcomplex Y . Since X is defined as a union of clusters, Y itself is contained
S in a finite union
of clusters. In particular, we find a finite list of pre-clusters C1 , ..., Cn such that Y ⊂ 1≤i≤n C̄i . Using
Propositions 10.33 and 10.41, we obtain a balanced system C1′ , ..., Cn′ such that
[
[
C̄i ⊆
C̄i′
1≤i≤n
1≤i≤n
Using Proposition 10.48, we conclude that the parallel closure C of C1′ , ..., Cn′ satisfies that:
S
S
(1) Y ⊆ 1≤i≤n C̄i ⊆ 1≤i≤n C̄i′ ⊆ C.
(2) C is homeomorphic to a nonpositively curved cube complex, hence is aspherical.
It follows that the image of the sphere is nullhomotopic in C and hence in X. Therefore, X is apsherical.
11. A question
The following question, suggested to the author by Mikhael Gromov, could provide an interesting direction
for future research.
Question 11.1. Is there a group G satisfying the following?
(1) G is of type F.
(2) G is nonamenable.
(3) G does not contain non abelian free subgroups.
Recall that a group is of type F if it admits a finite Eilenberg-Mclane complex. Examples of such groups
include finitely generated free abelian groups, torsion free hyperbolic groups (including finitely generated
free groups), torsion free subgroups of finite index in finitely generated Coxeter groups, torsion free
subgroups of finite index in arithmetic groups, and torsion free subgroups of finite index in the outer
automorphism group of a finitely generated free group. (See [12] for further details.) None of the above
examples can be sources of such a group.
The group studied in this article is not of type F, since it contains subgroups isomorphic to Z∞ . Interesting
subgroups of the group of piecewise projective homeomorphisms of the real line seem to always contain
A NONAMENABLE TYPE F∞ GROUP
61
Z∞ subgroups, hence may not provide such an example. This includes the case of Thompson’s group
F , for instance. Containing torsion elements is also an obstruction for type F. Hence nonamenable
Burnside groups, torsion Tarski monsters, Golod-Shafarevich groups as well as the finitely presentable
nonamenable group of Olshanskii-Sapir cannot provide a source of such a group.
References
[1] J. Alonso, “Finiteness conditions on Groups and Quasi-isometries.”, Journal of Pure and Applied Algebra 95 (1994)
121-129
[2] S.I. Adyan. The Burnside Problem and Identities in Groups. volume 95 of Ergebnisse der Mathematik und ihrer
Grenzgebiete. Springer, Berlin, 1979.
[3] S.I. Adyan. Random walks on free periodic groups. Izv. Akad. Nauk SSSR Ser. Mat. 46(6):1139–1149, 1982.
[4] J. Belk. Thompson’s group F . Ph.D. Thesis, Cornell University, 2004. arXiv:0708.3609.
[5] M. Bestvina and N. Brady, ‘ Morse theory and finiteness properties of groups Invent Math. 129 (1997) 445-40.
[6] K. Brown. Finiteness properties of groups. J. Pure Appl. Algebra 44 (1987), no. 1-3, 4575.
[7] K.S. Brown, and R. Geoghegan, An infinite-dimensional torsion-free F P∞ group. Invent Math. 77 (1984): 367-382.
[8] M.G. Brin, C.C. Squier. Groups of piecewise linear homeomorphisms of the real line. Invent. Math., 79(3):485–498,
1985.
[9] M.R. Bridson and A. Haefliger. Metric spaces of non-positive curvature Springer-Verlag, Berlin, 1999.
[10] J.W. Cannon, W.J. Floyd, W.R. Parry. Introductory notes on Richard Thompson’s groups. Enseign. Math. (2), 42(34):215–256, 1996.
[11] M.M. Day. Means for the bounded functions and ergodicity of the bounded representations of semi-groups. Trans.
Amer. Math. Soc., 69:276–291, 1950.
[12] R. Geoghegan Topological Methods in Group Theory Graduate Texts in Mathematics (Volume 243, Springer 2008)
[13] Gersten, S. M. Selected problems. Combinatorial group theory and topology (Alta, Utah, 1984), 545–551, Ann. of
Math. Stud., 111, Princeton Univ. Press, Princeton, NJ, 1987.
[14] M. Gromov, Hyperbolic groups. Essays in group theory. Mathematical Sciences Research Institute Publications 8(ed)
S.M. Gersten, Spriger, New York, 1987.
[15] S.V. Ivanov. Embedding free Burnside groups in finitely presented groups. Geometriae dedicata, 111:87–105, 2005.
[16] Y. Lodha, J.T. Moore, A nonamenable finitely presented group of piecewise projective homeomorphisms. Groups,
Geometry and Dynamics. Volume 10, Issue 1, 2016, pp. 177-200
[17] J. Lauri, R Scapellato Topics in Graph Automorphisms and Reconstruction. London Mathematical Society Student
texts 54.
[18] N. Monod. Groups of piecewise projective homeomorphisms. Proc. Natl. Acad. Sci., 110(12):4524–4527, 2013
[19] A.Y. Ol’shanskii. On the question of the existence of an invariant mean on a group. Uspekhi Mat. Nauk, 35(4):199–200,
1980.
[20] A.Y. Ol’shanskii, M.V. Sapir. Non-amenable finitely presented torsion-by-cyclic groups. Publ. Math. Inst. Hautes
Études Sci., 96:43–169, 2003.
[21] J. von Neumann. Zur allgemeinen Theorie des Maßes. Fund. Math., 13:73–116, 1929.
[22] C.T.C. Wall, “Finiteness Conditions for C.W. Complexes”, The Annals of Mathematics, 2nd Ser., Vol 81, No. 1. (Jan.,
1965), 56-69.
[23] M. Zaremsky. HNN decompositions of the Lodha-Moore groups, and topological applications J. Topol. Anal, Vol. 8,
No. 4, 627-653, 2016
Institute of Mathematics, EPFL, SB MATHGEOM EGG, Station 8, MA B3 514, Lausanne, CH-1015, Switzerland
E-mail address: [email protected]
| 4math.GR
|
arXiv:1509.03354v3 [math.AC] 30 Mar 2017
VALUATION SEMIRINGS
PEYMAN NASEHPOUR
Abstract. The main scope of this paper is to introduce valuation semirings
in general and discrete valuation semirings in particular. In order to do that,
first we define valuation maps and investigate them. Then we define valuation
semirings with the help of valuation maps and prove that a multiplicatively
cancellative semiring is a valuation semiring if and only if its ideals are totally
ordered by inclusion. We also prove that if the unique maximal ideal of a
valuation semiring is subtractive, then it is integrally closed. We end this
paper by introducing discrete valuation semirings and show that a semiring is
a discrete valuation semiring if and only if it is a multiplicatively cancellative
principal ideal semiring possessing a nonzero unique maximal ideal. We also
prove that a discrete valuation semiring is a Gaussian semiring if and only if
its unique maximal ideal is subtractive.
In memory of Prof. Dr. Manfred Kudlek
0. Introduction
Semirings not only have significant applications in different fields such as automata theory in theoretical computer science, (combinatorial) optimization theory, and generalized fuzzy computation, but are fairly interesting generalizations
of two broadly studied algebraic structures, i.e., rings and bounded distributive
lattices. Valuation theory for rings was introduced by Krull in [18] and has been
proven to be a very useful tool in ring theory. Also valuation theory represents a
nice interplay between ring theory and ordered Abelian groups. The main purpose
of this paper is to generalize valuation theory for semirings.
In this paper, all semirings are commutative with a zero and a nonzero identity.
In the first section of the present paper, similar to the classical concept of valuations
in Bourbaki ([2, VI, 3.1]), we define the concept of valuation maps for semirings
with values in totally ordered commutative monoids (tomonoids for short). Note
that our notion for valuation maps should not to be confused with the concept of
valuation functions for semirings in [12, Definition 6.4].
By an M -valuation on S, we mean a map v : S → M∞ with the following
properties:
(1) S is a semiring and M∞ is a tomonoid with the greatest element +∞, which
has been obtained from the tomonoid M with no greatest element,
(2) v(xy) = v(x) + v(y), for all x, y ∈ S,
(3) v(x + y) ≥ min{v(x), v(y)}, for all x, y ∈ S,
2010 Mathematics Subject Classification. 16Y60, 13B25, 13F25, 06D75.
Key words and phrases. Semiring, Bounded distributive lattice, Semiring polynomials, Monoid
semiring, Unique factorization semiring, Valuation semiring, Valuation map, Discrete valuation
semiring, Gaussian semiring.
1
2
PEYMAN NASEHPOUR
(4) v(1) = 0 and v(0) = +∞.
When there is an M -valuation v on S, it is easy to see that the set Sv = {s ∈
S : v(s) ≥ 0} is a subsemiring of S and plays an important role in our paper. If the
map v is surjective, we show that the semiring S is a semifield if and only if M is
an Abelian group and the set of units U (Sv ) of Sv is equal to {s ∈ Sv : v(s) = 0}
(See Theorem 1.8).
Let us recall that a nonempty subset I of a semiring S is said to be an ideal of
S, if a, b ∈ I implies a + b ∈ I and sa ∈ I for any s ∈ S. An ideal I of a semiring S
is said to be subtractive (in some references k-ideal), if a + b ∈ I and a ∈ I implies
b ∈ I for all a, b ∈ S. At last, an ideal I of a semiring S is said to be proper, if
I 6= S.
Let v be an M -valuation on S. We define v to have min-property, whenever
v(x) 6= v(y), then v(x + y) = min{v(x), v(y)} for any x, y ∈ S. Though this
property holds for valuation maps on rings, this is not the case for semirings (Check
Remark 1.13). We say “T is a V -semiring with respect to the triple (S, v, M )”, if
T is a semiring and there exists an M -valuation v on S such that the semiring S
contains T as a subsemiring and T = Sv = {s ∈ S : v(s) ≥ 0} (Definition 1.6).
Actually in Corollary 1.15, we show that if T is a V -semiring with respect to the
triple (K, v, M ), where K is a semifield and v is a surjective map, then the following
statements are equivalent:
(1) The valuation map v has the min-property;
(2) The maximal Kv -ideal Pv = {x ∈ Sv : v(x) > 0} is subtractive.
Similar to ring theory, an ideal A of a semiring S is said to be principal if A
is generated by one element, i.e., A = (a) = {sa : s ∈ S} for some a ∈ S. In
Proposition 1.18, we prove that if Sv is a V -semiring with respect to the triple
(K, v, M ), where K is a semifield and v is surjective, then the principal ideals of
Sv are totally ordered by inclusion. This is the base for the definition of valuation
semirings given in section 2.
Let us recall a notation. If S is a multiplicatively cancellative semiring (for short
MC semiring), we denote its semifield of fractions by F (S) (Refer to Remark 1.9).
In section 2, we define a semiring S to be a valuation semiring, if there exists an
M -valuation v on K, where K is a semifield containing S as a subsemiring, v is
surjective, and S = Kv = {s ∈ K : v(s) ≥ 0}. One of the main theorems of the
paper is the following (Theorem 2.4):
An MC semiring S is a valuation semiring, if and only if one of the following
equivalent conditions hold:
(1) For any element x ∈ F (S), either x ∈ S or x−1 ∈ S,
(2) For any ideals I, J of S, either I ⊆ J or J ⊆ I,
(3) For any elements x, y ∈ S, either (x) ⊆ (y) or (y) ⊆ (x).
Let S be an MC semiring and F (S) its semifield of fractions. The element
u ∈ F (S) is said to be integral over S if there exist a1 , a2 , . . . , an and b1 , b2 , . . . , bn
in S such that un + a1 un−1 + · · · + an = b1 un−1 + · · · + bn . The semiring S is said
to be integrally closed if the set of elements of F (S) that are integral over S is the
set S [5, p. 88].
In Theorem 2.7, we show that if S is a valuation semiring such that its unique
maximal ideal J is subtractive, then S is integrally closed.
VALUATION SEMIRINGS
3
Section 3 is devoted to discrete valuation semirings. We define a semiring S to
be a discrete valuation semiring (DVS for short), if S is a V -semiring with respect
to the triple (K, v, Z), where K is a semifield and v is surjective (Definition 3.1).
Before explaining briefly what we do in section 3, we need to recall some concepts.
A semiring is said to be quasi-local if it has only one maximal ideal. A semiring
satisfies ACCP if any ascending chain of principal ideals of S is stationary. A
semiring S is said to be a principal ideal semiring, if each ideal of S is principal.
In this section, we characterize discrete valuation semirings as follows (See Theorem 3.6): A semiring S is a discrete valuation semiring if and only if one of the
following equivalent conditions satisfies:
(1) S is a principal ideal MC semiring possessing a unique maximal ideal J =
6
(0),
(2) S is a quasi-local
T∞ MC semiring whose unique maximal ideal J 6= (0) is
principal and n=1 J n = (0),
(3) S is a quasi-local MC semiring whose unique maximal ideal J 6= (0) is
principal, which satisfies ACCP.
Let us recall that if S is a semiring, for a polynomial f ∈ S[X], the content of
f , denoted by c(f ), is defined to be the S-ideal generated by the coefficients of f .
A semiring S is called Gaussian if c(f g) = c(f )c(g) for all polynomials f, g ∈ S[X]
([22, Definition 7]). Finally in Theorem 3.8, we prove the following:
Let S be a discrete valuation semiring. Then the following statements are equivalent:
(1)
(2)
(3)
(4)
The unique maximal ideal of S is subtractive,
The valuation map v : F (S) → Z∞ satisfies the min-property,
Each ideal of the semiring S is subtractive,
The semiring S is Gaussian.
A connoisseur of valuation theory in commutative algebra already have noticed
that some of the definitions and results, mentioned above, are extensions of their
ring version ones. Therefore, throughout the paper, some examples are given to
show that some objects really satisfy the hypotheses of certain definitions and
theorems, while the others are given to show that the hypotheses of certain theorems
in the paper cannot be greatly weakened.
Though we have tried our paper to be self-contained in definitions and terminology, but many of them can be found in the book [11]. One may refer to the books
[12], [11] and [23] for more on semirings.
1. Valuation Maps on Semirings
The main scope of this section is to generalize the concept of valuation maps on
rings and investigate the properties of these maps. While according to the classical
definition of valuation maps in Bourbaki ([2, VI, 3.1]), the values of the elements of a
ring belong to a totally ordered Abelian group, we define valuation maps in such way
that the values of the elements of a semiring belong to an arbitrary totally ordered
commutative monoid. We recall that by a totally ordered commutative monoid
(tomonoid for short) (M, +, 0, ≤), it is meant a commutative monoid (M, +, 0)
such that (M, ≤) is a totally ordered set and x ≤ y implies x + z ≤ y + z for any
z ∈ M . We also need to recall the following:
4
PEYMAN NASEHPOUR
In monoid theory, there is a routine technique that one can obtain a monoid
with the greatest element from a monoid with no greatest element:
Let (M, +, 0, ≤) be a tomonoid with no greatest element. One can configure a
set M∞ by adjoining a new element to the monoid M , denoted by +∞, in this way
that the total order on M induces a total order on M∞ such that +∞ is the greatest
element, i.e., for all m ∈ M , we have that m < +∞ and the monoid structure on
M induces a monoid structure on M∞ with the following rule:
m + (+∞) = (+∞) + m = +∞, ∀ m ∈ M∞ .
It is immediate that M∞ is also a tomonoid with the greatest element +∞. Now
similar to the classical concept of valuations in Bourbaki ([2, VI, 3.1]), we define
the concept of valuation maps for semirings with values in tomonoids (not to be
confused with the concept of valuation functions for semirings in [12, Definition
6.4]).
Definition 1.1. By “an M -valuation on S”, we mean a map v : S → M∞ with
the following properties:
(1) S is a semiring and M∞ is a tomonoid with the greatest element +∞, which
has been obtained from the tomonoid M with no greatest element,
(2) v(xy) = v(x) + v(y), for all x, y ∈ S,
(3) v(x + y) ≥ min{v(x), v(y)}, for all x, y ∈ S,
(4) v(1) = 0 and v(0) = +∞.
When there is an M -valuation v on S, we set Sv = {s ∈ S : v(s) ≥ 0} and
Pv = {s ∈ Sv : v(s) > 0}. If M = Z, we say that v is a discrete valuation on S.
Remark 1.2. Let (S, +, ·) be a semiring and (M, ·, ≤) a tomonoid. It is possible
to give a dual definition of an M -valuation on S as follows:
We annex an element O to the tomonoid M to configure the set Mmax = M ∪{O}
and extend ordering and monoid multiplication by the rules:
O ≤ x and O · x = x · O = O, for all x ∈ Mmax .
Imagine we can define a function v : S → Mmax with following properties:
• v(x · y) = v(x) · v(y), for all x, y ∈ S,
• v(x + y) ≤ max{v(x), v(y)}, for all x, y ∈ S,
• v(1) = 1 and v(0) = O.
Then it is easy to see that v defines an M -valuation on S. On the other hand,
if v is an M -valuation on S, by reversing the ordering, one can get a function
v : S → M∞ such that the three above conditions of the current remark are
satisfied.
Remark 1.3.
(1) If, in Definition 1.1, S is a ring and M is a totally ordered
Abelian group, then our definition for valuation maps given in Definition
1.1 coincides with the definition of a Bourbaki valuation on S with values
in M∞ given in [2]. In fact, our definition for valuation maps, which is
from an arbitrary semiring into a tomonoid and is based on Bourbaki’s
classical definition, is somehow equivalent to the definition of valuation
maps defined in Definitions 2.1 and 2.3 in [13], that is from a semiring into
a so-called commutative bipotent semiring. Also note that our definition is
more general than the definition of valuation maps in [15], which is taken
VALUATION SEMIRINGS
5
from [13], but it is only from an additively idempotent semiring into R ∪
{+∞}.
(2) By considering what it has been explained in Remark 1.2, our definition for
valuation maps is more general than what is defined in Definition 1.2 in [25],
a paper on valuations over idempotent (i.e., characteristic one) semirings,
since in the same paper, by a valuation on a semiring S, it is meant a surjective map v : S → Mmax , for some totally ordered abelian group M , having
the three conditions given in Remark 1.2 plus the following extra condition:
v(x) ≤ max{v(x + y), v(y)} for all x, y ∈ S.
Before starting to investigate the properties of valuation maps on semirings, let
us give a couple of examples. Note that a semiring S is called entire if ab = 0
implies either a = 0 or b = 0 for all a, b ∈ S.
Example 1.4.
(1) Let S be an entire semiring. One can construct a trivial M valuation v on S, by defining v(s) = 0 for any s ∈ S − {0} and v(0) = +∞,
where M = {0}.
(2) Let (S, +, 0, ≤) be a tomonoid with no greatest element. Adjoin the greatest
element +∞ to S and call the new set S∞ . Define on S the two operations
a ⊕ b = min{a, b} and a ⊙ b = a + b. One can easily check that (S∞ , ⊕, ⊙)
is a semiring and v : S∞ → S∞ defined by v(s) = s, is an S-valuation on
S∞ !
(3) Let N0 denote the set of non-negative integers and p ∈ N0 be a fixed prime
number. Obviously one can uniquely write any natural number x in the
form of x = pn · y, where n ≥ 0 and y is a natural number and has no factor
of p. Now we define the map v : N0 → N0 ∪ {+∞} by v(x) = v(pn · y) = n,
if x is nonzero, and v(0) = +∞. A simple calculation shows that v is an
N0 -valuation on N0 .
(4) Let K be a semifield. By a non-archimedean absolute value on K, we mean
a function | · | : K −→ R, satisfying the following properties for all x, y ∈ K:
• |x| ≥ 0, and |x| = 0 if and only if x = 0,
• |xy| = |x||y|,
• |x + y| ≤ max{|x|, |y|}.
It is clear that the map v : K −→ R ∪ {+∞}, defined as v(x) = − ln |x|,
gives us an R-valuation on K (Refer to [7] and [14]).
Now let us bring the following straightforward proposition only for the sake of
reference. Note that an ideal P of a semiring S is said to be a prime ideal of S, if
P 6= S and ab ∈ P implies either a ∈ P or b ∈ P for all a, b ∈ S.
Proposition 1.5. Let there exist an M -valuation v on S. The following statements
hold:
(1) If s is multiplicatively invertible element of S, then v(s−1 ) = −v(s).
(2) If sn = 1, then v(s) = 0 for any s ∈ S and natural number n.
(3) The set v −1 (+∞) is a prime ideal of S.
(4) The semiring S is an entire semiring, if v(s) = +∞ implies s = 0 for all
s ∈ S.
(5) The set Sv = {s ∈ S : v(s) ≥ 0} is a subsemiring of S and Pv = {s ∈ Sv :
v(s) > 0} is a prime ideal of Sv .
(6) If s is a unit of the semiring Sv , then v(s) = 0.
6
PEYMAN NASEHPOUR
The statement (5) in Proposition 1.5 suggests us to give the following definition:
Definition 1.6. We say “T is a V -semiring with respect to the triple (S, v, M )”,
if T is a semiring and there exists an M -valuation v on S such that the semiring S
contains T as a subsemiring and T = Sv = {s ∈ S : v(s) ≥ 0}. When there is no
fear of ambiguity, we may drop the expression “with respect to the triple (S, v, M )”
and simply say that T is a V -semiring.
If we denote the units of a semiring S by U (S), then U (Sv ) ⊆ {x ∈ Sv : v(x) = 0}
(Proposition 1.5). One may ask if these two sets are equal in general. In the
following, we give simple but useful examples to show that the mentioned two sets
can be sometimes different:
Example 1.7.
(1) Let S = N0 be the set of all nonnegative integers and define
v : N0 −→ {0, +∞} to be a map that sends all s > 0 to 0 and sends 0 to +∞.
Then, it is easy to check that U (Sv ) = {1}, while {s ∈ Sv : v(s) = 0} = N.
(2) Put T = N0 [X][X −1] to be the entire semiring of all Laurent polynomials
on N0 . Define the map v from T to Z∞ by v(am X m + am+1 X m+1 + · · · +
an X n ) = m, where m, n ∈ Z and m ≤ n and am 6= 0 and v(0) = +∞. It
is easy to see that v is a discrete valuation on the entire semiring T , Sv =
N0 [X] and U (Sv ) = {1}, while {f ∈ N0 [X] : v(f ) = 0} = N + X · N0 [X]!
Now the question arises under what conditions the equality U (Sv ) = {x ∈ Sv :
v(x) = 0} holds. This is what we are going to show in Theorem 1.8:
Theorem 1.8. Let T be a V -semiring with respect to the triple (S, v, M ). The
following statements hold:
(1) The set Pv = {s ∈ Sv : v(s) > 0} is the only maximal ideal of the semiring
Sv if and only if U (Sv ) = {x ∈ Sv : v(x) = 0}.
(2) If v is a surjective map, then the semiring S is a semifield if and only if M
is an Abelian group and U (Sv ) = {s ∈ Sv : v(s) = 0}.
Proof. (1): It is straightforward to see that Pv = {s ∈ Sv : v(s) > 0} is a prime
ideal of Sv (Proposition 1.5). Now assume that Pv is the only maximal ideal of
the semiring Sv . We claim that any s ∈ Sv with v(s) = 0 is a unit. In contrary,
if s is a nonunit element of Sv , then the principal ideal (s) 6= Sv is a subset of
the only maximal ideal Pv of Sv and therefore v(s) > 0. On the other hand, if
U (Sv ) = {s ∈ Sv : v(s) = 0} and I is a proper ideal of Sv , then I cannot contain a
unit and this means that v(s) > 0 for any nonzero s ∈ I and therefore I ⊆ Pv .
(2): Let M be an Abelian group and U (Sv ) = {s ∈ Sv : v(s) = 0}. Let x be
a nonzero element of S such that v(x) = g for some g ∈ M . Since v is surjective,
there exists a nonzero y ∈ S such that v(y) = −g. From this, we have v(xy) = 0
and this means that xy is unit of Sv and therefore x is multiplicatively invertible.
On the other hand, let S be a semifield. Since v is surjective, M is a homomorphic
image of S − {0}, and the monoid M is in fact a group. Also if x ∈ Sv is a unit,
then v(x) = 0 (Proposition 1.5). Now let x ∈ S and v(x) = 0. So there exists a
y ∈ S such that xy = 1. This implies that v(y) = 0 and the inverse of x is in fact
in Sv and so x ∈ U (Sv ).
Remark 1.9. Let us recall two important points:
(1.9.a) From commutative semigroup theory, we know that any cancellative monoid
M can be embedded into an Abelian group gp(M ), known as the group of
VALUATION SEMIRINGS
7
the differences of M , in this sense that there is a monoid monomorphism
ι : M → gp(M ) defined by ι(m) = [(m, 0)]∼ such that any element g ∈
gp(M ) can be written in the form of g = ι(x) − ι(y) for some x, y ∈ M
[3, p. 50]. Now let M be a cancellative tomonoid. For simplification, we
denote every element of gp(M ) by (x − y), where x, y ∈ M . The total order
on M induces a total order on gp(M ) as follows:
We define the relation ≤gp on gp(M ) by (x1 − x2 ) ≤gp (y1 − y2 ), if
x1 +y2 ≤ x2 +y1 . One can easily check that (gp(M ), ≤gp ) becomes a totally
ordered Abelian group such that its order is preserved by the monomorphism ι and particularly 0 ≤ x if and only if 0 ≤gp ι(x) for each x ∈ M .
(1.9.b) Let S be a multiplicatively cancellative semiring. The semiring S can be
embedded into a semifield F (S), known as the semifield of fractions of S.
Note that there is a semiring monomorphism ε : S → F (S) such that any
element f ∈ F (S) can be considered in the form of f = ε(a) · ε(b)−1 with
a, b ∈ S and b 6= 0 [10, p. 20]. For simplification, we denote every element
of F (S) by a/b, where a ∈ S and b ∈ S − {0}.
Now by considering Remark 1.9, we have the following:
Theorem 1.10. Let T be a V -semiring with respect to the triple (S, v, M ), where S
is an MC semiring and M is a cancellative tomonoid. Then the following statements
hold:
(1) The map v ′ : F (S) → gp(M )∞ defined by v ′ (x/y) = (v(x) − v(y)) for
nonzero elements x, y ∈ S and v ′ (0/1) = +∞ is a gp(M )-valuation on
F (S).
(2) The V -semiring Sv can be considered as a subsemiring of the V -semiring
Sv ′ .
Proof. (1): A simple calculation shows that v ′ is a well-defined map from F (S) to
gp(M )∞ such that v ′ ((a/b) · (c/d)) = v ′ (a/b) + v ′ (c/d) and v ′ (1/1) = 0.
Now consider that
v ′ ((a/b) + (c/d)) = v ′ ((ad + bc)/bd) = (v(ad + bc) − v(bd))
≥ (min{v(a) + v(d), v(b) + v(c)} − (v(b) + v(d)))
= min{(v(a) − v(b)), (v(c) − v(d))} = min{v ′ (a/b), v ′ (c/d)}.
So we have already proved that v ′ is the gp(M )-valuation on the semifield F (S).
(2): It is easy to see that the map ε|Sv : Sv → {z/1 ∈ F (S) : z ∈ Sv }, defined
by ε|Sv (z) = z/1, is the semiring isomorphism that we need. On the other hand,
{z/1 ∈ F (S) : z ∈ Sv } is a subsemiring of Sv′ , since v ′ (z/1) = v(z) ≥ 0 for
any z ∈ Sv and this means that z/1 ∈ Sv′ . Therefore Sv can be considered as a
subsemiring of Sv′ and this finishes the proof.
Remark 1.11. The reader, who is familiar with valuation ring theory, knows that
if G is a totally ordered Abelian group and κ is a field and v : κ → G∞ is a valuation
map, then v has this property that v(x) 6= v(y) implies v(x + y) = min{v(x), v(y)}
for all x, y ∈ κ ([7, Statement 1.3.4, p. 20]). Though in Remark 1.13, by giving
a suitable example, we will show that, in general, this property does not hold for
valuation maps on semirings, but later in Corollary 1.15 and Theorem 3.8, we will
see that this property holds for some important families of semirings. Therefore, it
is justifiable to give a name to this property.
8
PEYMAN NASEHPOUR
Definition 1.12. Let there exist an M -valuation v on S. We define v to have minproperty, whenever v(x) 6= v(y), then v(x + y) = min{v(x), v(y)} for any x, y ∈ S.
In the following remark, we give some examples related to min-property for
valuation maps on semirings. The first example shows that a valuation map, may
not satisfy the min-property. The second and third examples are good examples of
valuation maps on entire semirings satisfying min-property.
Remark 1.13.
(1) Let T be an entire semiring. Consider the entire semiring
T [X][X −1] of Laurent polynomials and define the map deg : T [X][X −1] →
Z∞ by deg(am X m + am+1 X m+1 + · · · + an X n ) = n, where m, n ∈ Z and
m ≤ n and an 6= 0 and v(0) = +∞. It is easy to check that the map “ deg ”
is surjective with this property that deg(f g) = deg(f ) + deg(g). Let us
mention that a semiring S is called zerosumfree, if x + y = 0 implies x =
y = 0 for all x, y ∈ S. Our claim is that deg(f + g) ≥ min{deg(f ), deg(g)}
for f, g ∈ T [X][X −1] if and only if T is a zerosumfree entire semiring and
the proof is as follows:
If T is zerosumfree, then one can easily see that for all f, g ∈ T [X][X −1],
we have deg(f + g) = max{deg(f ), deg(g)} ≥ min{deg(f ), deg(g)}. And if
T is not a zerosumfree semiring, then there exists two nonzero elements
a, b ∈ T such that a + b = 0 and if one sets f = 1 + aX and g = bX, then
obviously deg(f ) = deg(g) = 1, while deg(f + g) = 0. From all we said we
get that if T is a zerosumfree entire semiring, then “ deg ” is a Z-valuation
on the entire semiring T [X][X −1], while the min-property does not hold for
“ deg ”.
Now we give two important examples of valuation maps on entire semirings satisfying the min-property:
(2) Let us recall that if M is a commutative monoid and T is a semiring, one
can define the monoid semiring T [M ] constructed from the monoid M and
the semiring T similar to the standard definition of monoid rings. We
write each element of f ∈ T [M ] as polynomials f = t1 X m1 + · · · + tn X mn ,
where t1 , . . . , tn ∈ T and m1 , . . . , mn are distinct elements of M . Note
that this representation of f is called the canonical form of f [9, p. 68].
One can easily check that if M is a tomonoid and T is an entire semiring,
then the map v : T [M ] → M defined by v(s1 X m1 + · · · + sn X mn ) =
min{m1 , . . . , mn }, if s1 X m1 + · · · + sn X mn 6= 0 and v(0) = +∞ is an
M -valuation on T [M ] satisfying the min-property. This example can be
interesting in another perspective because v(f ) = v(g) implies v(f + g) =
v(f ), for all f, g ∈ T [M ] if and only if T is a zerosumfree semiring.
(3) For any given entire semiring T , define the entire semiring of Laurent
P∞ power
series T [[X]][X −1] to be the set of all elements of the form n≥m an X n ,
whereP
m ∈ Z and ai ∈ T . It is, then, easy to check that the map v, defined
n
−1
], is the
by v( ∞
n≥m an X ) = m if am 6= 0 and v(0) = +∞ on T [[X]][X
−1
Z-valuation on the entire semiring T [[X]][X ] satisfying the min-property.
Note that v(f ) = v(g) implies v(f + g) = v(f ), for all f, g ∈ T [[X]][X −1] if
and only if T is a zerosumfree semiring.
Let S be a semiring and (M, +, 0) be a commutative additive monoid. The
monoid M is said to be an S-semimodule if there is a function, called scalar product,
VALUATION SEMIRINGS
9
λ : S × M −→ M , defined by λ(s, m) = sm such that the following conditions are
satisfied:
(1) s(m + n) = sm + sn for all s ∈ S and m, n ∈ M ;
(2) (s + t)m = sm + tm and (st)m = s(tm) for all s, t ∈ S and m ∈ M ;
(3) s · 0 = 0 for all s ∈ S and 0 · m = 0 and 1 · m = m for all m ∈ M .
A nonempty subset N of an S-semimodule M is said to be an S-subsemimodule
of M , if m + n ∈ N for all m, n ∈ N and sn ∈ N for all s ∈ S and n ∈ N . An
S-subsemimodule K of an S-semimodule N is said to be subtractive if x + y ∈ K
and x ∈ K imply that y ∈ K for any x, y ∈ N . For more on semimodules over
semirings, one may refer to [11, Chap. 14]. The following theorem introduces
some subtractive ideals of V -semirings. Also we will use this theorem to find some
equivalent conditions for a valuation map to satisfy the min-property.
Theorem 1.14. Let M be a tomonoid and v be an M -valuation on the semiring
T . The following statements hold:
(1) For any α ∈ M , the sets Kα = {x ∈ T : v(x) > α} and Lα = {x ∈ T :
v(x) ≥ α} are both Sv -subsemimodules of T and if v is a surjective map
and β > α is another element of M , then Kβ ⊂ Lβ ⊂ Kα ⊂ Lα .
(2) If α ≥ 0 is an element of M , then Iα = {x ∈ Sv : v(x) > α} and Jα =
{x ∈ Sv : v(x) ≥ α} are both Sv -ideals of Sv and if v is a surjective map
and β > α is another element of M , then Iβ ⊂ Jβ ⊂ Iα ⊂ Jα .
(3) For any x, y ∈ T (∈ Sv ), either Lv(x) ⊆ Lv(y) (Iv(x) ⊆ Iv(y) ) or Lv(y) ⊆
Lv(x) (Iv(y) ⊆ Iv(x) ).
(4) If the valuation map v has the min-property, then Kα and Lα (Iα and
Jα ) are both subtractive Sv -subsemimodules of T (ideals of Sv ) for any
(nonnegative) α ∈ M .
Proof. The assertions (1), (2) and (3) are straightforward.
For (4), we only prove one of the claims, since the proof of the other ones is
similar. Now let x, y ∈ T be such that x + y ∈ Kα and x ∈ Kα . If v(x) = v(y),
then from the assumption x ∈ Kα , we get that y ∈ Kα . If v(x) 6= v(y), then from
the assumptions x + y ∈ Kα and v(x + y) = min{v(x), v(y)} (min-property) and
this point that v(y) ≥ min{v(x), v(y)}, we get that again y ∈ Kα and this proves
that Kα is subtractive and the proof is complete.
Corollary 1.15. Let T be a V -semiring with respect to the triple (K, v, M ), where
K is a semifield and v is a surjective map. The following statements are equivalent:
(1) The valuation map v has the min-property;
(2) For each nonnegative element α ∈ M , the Kv -ideal Iα is subtractive;
(3) The maximal Kv -ideal Pv = {x ∈ Sv : v(x) > 0} is subtractive.
Proof. By Theorem 1.14, (1) implies (2) and obviously (2) implies (3). For proving
(3) ⇒ (1), we proceed as follows:
Since K is a semifield, by Theorem 1.8, M is an Abelian group, U (Kv ) = {x ∈
Kv : v(x) = 0}, and Pv = {x ∈ Kv : v(x) > 0} is the only maximal ideal of
Kv . Now let x, y ∈ K − {0} be such that v(x) < v(y) and g = v(y) − v(x).
Obviously g is a positive element of M and since v is a surjective map, there exists
a nonzero z ∈ Kv such that v(z) = g. This implies that v(yx−1 z −1 ) = 0 and so
yx−1 z −1 ∈ U (Kv ). Therefore there is a u ∈ U (Kv ) such that y = xzu. Now we
consider that x+y = x(1+zu). Obviously zu ∈ Pv and since Pv is subtractive, 1+zu
10
PEYMAN NASEHPOUR
needs to be a unit and v(x + y) = v(x). This means that v has the min-property
and the proof is complete.
Proposition 1.16. Let T be a V -semiring with respect to the triple (S, v, M ). Then
the following statements hold:
(a) For any x ∈ S, the cyclic Sv -subsemimodule (x) of Sv -semimodule S is a
subset of Lv(x) .
(b) For any x ∈ Sv , the principal Sv -ideal (x) of the semiring Sv is a subset of
Jv(x) .
Moreover if S is a semifield and the map v is surjective, then the following
statements hold:
(c) For any x ∈ S, the cyclic Sv -subsemimodule (x) of Sv -semimodule T is
equal to Lv(x) .
(d) For any x ∈ Sv , the principal Sv -ideal (x) of the semiring Sv is equal to
Jv(x) .
Proof. The proof of (a) and (b) is straightforward. For the proof of (c), let S be
a semifield and v surjective. By Theorem 1.8, the monoid M is in fact an Abelian
group and U (Sv ) = {x ∈ Sv : v(x) = 0}. If x = 0, then there is nothing to prove.
So let x 6= 0 and y ∈ Lv(x) . So v(y) ≥ v(x). Since M is an Abelian group, we can
choose m ∈ M such that m = v(y) − v(x) and since v is a surjective map, there is
a nonzero z ∈ T such that m = v(z). From this, we get that v(y) − v(x) − v(z) = 0
and so v(yx−1 z −1 ) = 0. This means that yx−1 z −1 = u is a unit element of Sv and
y = uzx. Since v(uz) ≥ 0, we have y ∈ (x) and this completes the proof. The proof
of (d) is similar to the proof of (c) and therefore, it is omitted.
In the following example, we show that the condition of being a semifield for S
in the statements (c) and (d) in Proposition 1.16 is necessary, in this sense that if
S is not a semifield, the equality (x) = Jv (x) may not hold:
Example 1.17. Let N0 be the set of nonnegative integers. Obviously any natural
number x > 0 can be written uniquely in the form of x = 5n ·y, where n ≥ 0 and y ∈
N is relatively prime to 5. We define v5 : N0 → N0 ∪{+∞} by v5 (x) = v5 (5n ·y) = n,
if x > 0 and v(0) = +∞. It is straightforward to check that v5 is an N0 -valuation
on the semiring N0 . Obviously v(2) = 0 and Jv(2) = {y ∈ N : v(y) ≥ 0} = N0 , while
(2) = 2N0 .
Proposition 1.18. Let T be a V -semiring with respect to the triple (K, v, M ),
where K is a semifield and v is surjective. Then for the cyclic Sv -subsemimodules
L1 , L2 of the Sv -semimodule K, either L1 ⊆ L2 or L2 ⊆ L1 . In particular for the
principal Sv -ideals J1 , J2 of the semiring Sv , either J1 ⊆ J2 or J2 ⊆ J1 .
Proof. If one of the cyclic subsemimodules L1 and L2 is zero, then there is nothing to
prove. If not, then there are two nonzero elements x1 , x2 ∈ K such that L1 = (x1 )
and L2 = (x2 ). Since K is a semifield and v is surjective, by Proposition 1.16,
L1 = Lv(x1 ) and L2 = Lv(x2 ) and by Theorem 1.14, either L1 ⊆ L2 or L2 ⊆ L1 .
The same proof works for ideals and this is what we wished to prove.
Let us recall that an integral domain D is said to be a valuation ring, if every
element x of its field of fractions K satisfies this property: x ∈
/ D implies x−1 ∈ D
([21, Chap. 4]). This is equivalent to this statement that all ideals of D are totally
VALUATION SEMIRINGS
11
ordered by inclusion ([19, Proposition 5.2]). Theorem 1.18 states that if T is a
V -semiring with respect to the triple (K, v, M ), where K is a semifield and v is
surjective, then the principal ideals of Sv (= T ) are totally ordered by inclusion.
The question arises if this causes all ideals of Sv to be totally ordered. As we will
see in Theorem 2.4, the answer to this question is affirmative. Now we pass to the
next section to define valuation semirings and investigate their properties.
2. Valuation Semirings
The main task of this section is to define the concept of valuation semirings,
which is an extension of the concept of valuation rings, and show that similar
facts for valuation rings hold for valuation semirings. A classical result in valuation
theory shows that D is a valuation ring if and only if D is a V -semiring with respect
to the triple (F (D), v, G∞ ), where F (D) is the field of fractions of the domain D,
G is the quotient group of F (D)∗ modulo U (D) and v : F (D) → G∞ is defined by
v(x) = xU , if x 6= 0, and v(0) = +∞ ([19, Definition 5.12 and Proposition 5.13]).
Before giving our definition for valuation semirings, we need to explain one more
point: From ring theory, we know that a ring D is said to be an integral domain if
one of the following equivalent conditions hold:
(1) The multiplication of the ring D has cancelation property, i.e., if ab = ac
and a 6= 0, then b = c for all a, b, c ∈ D,
(2) The ring D is entire, i.e., if ab = 0, then either a = 0 or b = 0.
In semiring theory, these two statements are not equivalent. In fact, if S is an
MC semiring, then S is an entire semiring but the inverse is not true. For instance,
the semiring (L = {0, 1, . . . , n}, max, min) for n ≥ 2 is not an MC semiring, since
if u ∈ L − {0, 1}, then min{u, u} = min{u, 1}, but u 6= 1, while obviously it is an
entire semiring.
Now imagine T is a V -semiring with respect to the triple (S, v, M ). The question
is: Does the assumption T is entire imply that the ideals of Sv are totally ordered
by inclusion? The following interesting example shows that this is not the case:
Example 2.1. Let B = {0, 1} be the Boolean semiring and B[X][X −1 ] the entire
semiring of Laurent polynomials. Define the surjective map v : B[X][X −1 ] → Z∞
by v(am X m + · · · + an X n ) = m, when m ≤ n are integer numbers and am 6= 0
and v(0) = +∞. Obviously B[X] is a V -semiring with respect to the triple
(B[X][X −1 ], v, Z), while even its principal ideals are not totally ordered under inclusion, since the ideals (X) and (X + 1) of B[X] are not comparable.
According to this introductory note and all we have seen in Theorem 1.10, Corollary 1.15, and Proposition 1.16, it seems the assumptions “K is a semifield” and “v
is a surjective map” are useful in defining valuation semirings. Therefore, we give
the following definition:
Definition 2.2. We define a semiring S to be a valuation semiring, if there exists
an M -valuation v on K, where K is a semifield containing S as a subsemiring, v is
surjective, and S = Kv = {s ∈ K : v(s) ≥ 0}.
Remark 2.3.
(1) In Definition 2.2, according to Theorem 1.8, M has to be an
Abelian group and S needs to be an MC semiring, since it is a subsemiring
of the semifield K.
12
PEYMAN NASEHPOUR
(2) In Definition 2.2, K needs to be isomorph to the semifield of fractions of
S. Define ϕ : F (S) → K by ϕ(a/b) = a · b−1 . An easy calculation shows
that ϕ is a semifield monomorphism. We only prove that ϕ is surjective.
Let x ∈ K. If v(x) ≥ 0, then x ∈ S and therefore ϕ(x/1) = x. If v(x) < 0,
then v(x−1 ) = −v(x) > 0 and x−1 ∈ S and obviously ϕ(1/(x−1 )) = x.
(3) Any valuation semiring is a quasi-local semiring (Theorem 1.8). Note that
a semiring is called to be quasi-local if it has only one maximal ideal ([11,
Example 6.1]).
(4) Let T be a V -semiring with respect to the triple (S, v, M ), where S is an
MC semiring and M is a cancellative tomonoid. Then Sv can be embedded
into a valuation semiring (Theorem 1.10).
While the proof of the following important theorem is somehow similar to its
ring version, we bring its proof here only for the completeness of the paper:
Theorem 2.4. For an MC semiring S, the following statements are equivalent:
(1)
(2)
(3)
(4)
S is a valuation semiring,
For any element x ∈ F (S), either x ∈ S or x−1 ∈ S,
For any ideals I, J of S, either I ⊆ J or J ⊆ I,
For any elements x, y ∈ S, either (x) ⊆ (y) or (y) ⊆ (x).
Proof. (1) ⇒ (2): Let S be a valuation semiring. According to Definition 2.2,
S = Kv = {s ∈ K : v(s) ≥ 0}. Now take x ∈ F (S). By considering the point (2)
mentioned in Remark 2.3, if x ∈
/ S, then v(x) < 0 and so v(x−1 > 0 and this means
−1
that x ∈ S.
(2) ⇒ (3): Imagine I * J and choose a ∈ I − J. Take b ∈ J − {0}. So a/b ∈
/ S,
because if a/b ∈ S, then (a/b) · b ∈ B, which is in contradiction with a ∈
/ J. So by
(2), b/a ∈ S and (b/a) · a ∈ I. Thus J ⊆ I.
(3) ⇒ (4): Obvious.
(4) ⇒ (1): Let U be the units of the semiring S. Definitely U is a subgroup
of F (S)∗ , the multiplicative group of nonzero elements of fraction semifield of S.
Define G = F (S)∗ /U to be the quotient group of F (S)∗ modulo U and write the
operation on G additively, i.e., aU +bU = (ab)U . Define a relation on G by yU ≤ xU
if for the cyclic S-subsemimodules (x), (y) of F (S), we have (x) ⊆ (y). It is obvious
that this is an order on G. On the other hand the cyclic S-subsemimodules of
F (S) are totally ordered, for the principal ideals of S are totally ordered and this
causes ≤ to be a totally order on G. Now we prove that G with the order ≤ is in
fact a totally ordered Abelian group. Let yU ≤ xU . So (x) ⊆ (y) and obviously
(zx) ⊆ (zy) and this means that zU + yU ≤ zU + xU .
Finally define v : F (S) → G∞ by v(x) = xU , for x 6= 0 and v(0) = +∞. From
the definition of G and v, it is clear that v(xy) = (xy)U = xU + yU = v(x) + v(y)
and so v(1) = U , the neutral element of the group G.
Also let x ∈ K. So x ∈ S if and only if (x) = (1) and this implies that x ∈ S
if and only if U ≤ xU , which means that x ∈ S if and only if U ≤ v(x). Now
we use this to prove that v(x + y) ≥ min{v(x), v(y)}. If either x = 0 or y = 0,
the inequality holds vividly. So let x, y are both nonzero element of F (S) and
v(y) ≤ v(x). From this, we get that yU ≤ xU and according to the definition of
order, (x) ⊆ (y). So there exists a nonzero s ∈ S such that x = sy and therefore
x/y ∈ S. Obviously this implies that x/y + 1 ∈ S and v(x/y + 1) ≥ U . At last
VALUATION SEMIRINGS
13
we get that v(x + y) = v(y(x/y + 1) = v(y) + v(x/y + 1) ≥ v(y) and the proof is
finished.
After seeing this theorem, the reader may ask if there is a semiring, which is neither a ring nor an MC semiring, but still its ideals are totally ordered by inclusion.
In the following, we show that the fuzzy semiring has this property:
Example 2.5. Let (I = [0, 1], max, min) be the fuzzy semiring and A an ideal of
I. It is easy to see that a ∈ A if and only if [0, a] ⊆ A for any a ∈ I. Therefore if
sup(A) ∈ A, then A = [0, sup(A)]. Now let sup(A) ∈
/ A. Obviously A ⊆ [0, sup(A)).
Our claim is that [0, sup(A)) ⊆ A. In order to prove the claim, we assume that x ∈
/
A. Then [x, 1] ∩ A = ∅, because if y ∈ [x, 1] ∩ A, then x ≤ y and x = min{x, y} ∈ A.
On the other hand
S we know that a ∈ A if andSonly if [0, a] ⊆ A for any a ∈ I.
This S
means that a∈A [0, a] ⊆ A and so [x, 1] ∩ ( a∈A [0, a]) = ∅. From this, we get
that a∈A ([0, a] ∩ [x, 1]) = ∅, which implies that [0, a] ∩ [x, 1] = ∅ for any a ∈ A.
Consequently a < x for any a ∈ A and this means that x ∈
/ [0, sup(A)). Hence, we
have already proved that if A is any ideal of I, then either A = [0, a] for some a ∈ A
or A = [0, b) for some b ∈ A. Finally an easy case-by-case discussion shows that all
ideals of I are totally ordered. Note that I is an entire semiring, while it is not an
MC semiring.
We finalize this section by generalizing another important classical result in
valuation ring theory, which states that any valuation ring is integrally closed ([19,
Proposition 5.5]). Actually, we prove that if the only maximal ideal of a valuation
semiring S is subtractive, then S is integrally closed in the sense of the following
definition borrowed from the paper [5, p. 88]. Note that if S is a ring, then the
following definition is equivalent to the standard definition of integrally closed rings
given in Definition 4.2 in the book [19] (For more on this, also check Remark 2.8).
Definition 2.6. Let S be an MC semiring and F (S) its quotient semifield. The
element u ∈ F (S) is said to be integral over S if there exist a1 , a2 , . . . , an and
b1 , b2 , . . . , bn in S such that un + a1 un−1 + · · · + an = b1 un−1 + · · · + bn . The
semiring S is said to be integrally closed, if the set of elements of F (S) that are
integral over S is the set S [5, p. 88].
Proposition 2.7. Let S be a valuation semiring such that its unique maximal ideal
J is subtractive. Then S is integrally closed.
Proof. Clearly any element of S is integral over S. Now let u ∈ F (S) be integral
over S. By Definition 2.6, there exist a1 , a2 , . . . , an and b1 , b2 , . . . , bn in S such that
un + a1 un−1 + · · · + an = b1 un−1 + · · · + bn .
(E1 )
Our claim is that u ∈ S. In contrary, suppose that u ∈
/ S. By Theorem 2.4,
u−1 ∈ S. On other hand, u−1 ∈
/ U (S), since if u−1 ∈ U (S), then u ∈ U (S). This
means that u−1 ∈ J. By multiplying the both sides of the equation E1 by u−n , we
get
1 + a1 u−1 + · · · + an u−n = b1 u−1 + · · · + bn u−n .
(E2 )
Obviously a1 u−1 +· · ·+an u−n , b1 u−1 +· · ·+bn u−n ∈ J and since J is subtractive,
we have 1 ∈ J, a contradiction. Therefore, u ∈ S and this finishes the proof.
Remark 2.8. For more on the notions “integral elements”, “integral closure”,
and “integrally closed rings or ideals” in commutative algebra, one can refer to
14
PEYMAN NASEHPOUR
[26, Definition 5.1.1]), [26, Definition 5.1.24], [1, p. 63], [24, Definition 1.1.1], [4,
Definition 10.2.1], and [6, Exercise 4.14].
Also in the paper [20], the concept of integral closure for elements and ideals
in idempotent semirings is introduced, and how it corresponds to its namesake in
commutative algebra is established. Finally the term “integrally closed ideals for
semirings” in Definition 1.4 of the paper [20] is used, but the relationship of these
notions in semirings to the notion of “integrally closed MC-semirings” given in page
88 of the paper [5] is unclear for the author.
3. Discrete Valuation Semirings
Let us recall that in Definition 1.1, we defined an M -valuation v on S to be
discrete, if M = Z. We start this section by defining discrete valuation semirings.
Definition 3.1. We define a semiring S to be a discrete valuation semiring (DVS
for short), if S is a V -semiring with respect to the triple (K, v, Z), where K is a
semifield and v is surjective.
Now we go further to give some examples for discrete valuation semirings that
are not rings. For this, we need to bring some concepts and definitions.
Recall that if S is a semiring, for a, b ∈ S, it is written a | b and said that “a
divides b”, if b = sa for some s ∈ S. This is equivalent to say that (b) ⊆ (a). Also
it is said that a and b are associates if a = ub for some unit u ∈ U (S) and if S is
an MC semiring, then this is equivalent to say that (a) = (b). A nonzero, nonunit
element s of a semiring S is said to be irreducible if s = s1 s2 for some s1 , s2 ∈ S,
then either s1 or s2 is a unit. This is equivalent to say that (s) is maximal among
proper principal ideals of S. An element p ∈ S − {1} is said to be a prime element,
if the principal ideal (p) is a prime ideal of S, which is equivalent to say if p | ab,
then either p | a or p | b.
An MC semiring S is called a unique factorization (or sometimes factorial) semiring if the following conditions are satisfied:
UF1 Each irreducible element of S is a prime element of S.
UF2 Any nonzero, nonunit element of S is a product of irreducible elements of
S.
First we construct a nice and general example of a discrete valuation semiring
as follows:
Example 3.2. Let S be a factorial semiring and K = F (S) its semifield of fractions.
Let p ∈ S be a fixed prime element of S. According to the definition of factorial
semirings, any nonzero x ∈ S can be uniquely written in the form of x = pn · x1 ,
where n ≥ 0 and x1 is a nonzero element of S such that it has no factor of p.
Therefore, any nonzero x/y ∈ K, can be uniquely written in the form of x/y =
pm · x1 /y1 , where m ∈ Z and x1 and y1 are nonzero elements of S, which have no
factor of p. Now we define a map vp : K → Z∞ with vp (x/y) = vp (pm · x1 /y1 ) = m,
when x/y is a nonzero element of K, and vp (0) = +∞. Obviously vp (1) = 0 and
vp (x/y · z/t) = vp (x/y) + vp (z/t). Now let x/y = pm · x1 /y1 and z/t = pn · z1 /t1 ,
where m, n ∈ Z and x1 , y1 , z1 and t1 all have no factor of p. Without loss of
generality, we can suppose that m ≤ n. Now consider the following:
vp (x/y + z/t) = vp (pm ((x1 t1 + pn−m y1 z1 )/y1 t1 ) =
m + vp (x1 t1 + pn−m y1 z1 ) ≥ m = min{vp (x/y), vp (z/t)}.
VALUATION SEMIRINGS
15
From all we said, we get that Sp = {x/y ∈ K : vp (x/y) ≥ 0} = {x/y ∈ K : p ∤ y}
is a discrete valuation semiring. Now the question arises if there are factorial
semirings, which are not rings. The first example that may come to one’s mind
is N0 . In Proposition 3.10, we also prove that if D is a Dedekind domain, then
(Id(D), +, ·) is a factorial semiring, where by Id(D), we mean the set of all ideals
of D. Obviously (Id(D), +, ·) is not a ring.
Lemma 3.3. If S is a discrete valuation semiring, then there exists a nonzero and
nonunit element t ∈ S such that any nonzero ideal I of S is of the form I = (tn )
for some n ≥ 0.
Proof. Let S be a discrete valuation semiring. By definition, S is a V -semiring with
respect to a suitable triple (K, v, Z), where K is a semifield and v is surjective. Now
let t ∈ S such that v(t) = 1. Our claim is that if I is a nonzero proper ideal of S,
then there exists a natural number n such that I = (tn ). Put n = min{v(s) : s ∈ I}.
Obviously n is a positive integer. Now let s ∈ I such that v(s) = n. It is clear that
v(s · t−n ) = 0 and by Theorem 1.8, s · t−n is a unit and (tn ) ⊆ I. On the other
hand, if s ∈ I − {0}, then v(s) ≥ n by choice of n and therefore s = utm , where
u is a unit and m ≥ n. This causes s ∈ (tn ). So we have already proved that any
nonzero proper ideal of S is of the form (tn ).
Before bringing the first important theorem of this section, which characterizes
discrete valuation semirings, we prove the following useful lemma. Let us recall
that, similar to ring theory, we say a semiring S satisfies ACCP if any ascending
chain of principal ideals of S is stationary.
Lemma 3.4. Let S beTan MC semiring, which satisfies ACCP. Then t ∈ S is a
∞
nonunit if and only if n=1 (tn ) = (0).
T∞
Proof. If t is a unit element of
S, then obviously n=1 (tn ) = S. Now let t be a
T∞
nonunit element of S and s ∈ n=1 (tn ). So for each natural number n, there exists
an sn ∈ S such that s = sn tn . This gives us the ascending chain (s1 ) ⊆ (s2 ) ⊆
· · · ⊆ (sn ) ⊆ · · · , which must stop somewhere, because S satisfies ACCP. Therefore
there is a natural number m such that (sm ) = (sm+1 ). This means that there is
an r ∈ S such that sm+1 = rsm and so we have s = sm+1 tm+1 = r(sm tm )t = rts.
Obviously this implies that s = 0, since if s 6= 0, then t is a unit and the proof is
finished.
Let us recall that a semiring is called to be quasi-local if it has only one maximal
ideal ([11, Example 6.1]).
Proposition 3.5. Let S be a semiring. Then the following statements hold:
S
(1) The set of all units U (S) of the semiring S is equal to S − m∈Max(S) m,
where by Max(S), we mean the set of all maximal ideals of S.
(2) The semiring S is quasi-local if and only if S − U (S) is an ideal of S.
Proof. (1) : Let S be a semiring and take U (S) to be the set of all units of S. If
s ∈ U (S), then s cannot be an element of a maximal ideal of S. On the other hand,
if s is not invertible, then the principal ideal (s) of S is proper and by Proposition
6.59 in [11], (s) is contained in a maximal ideal m of S and therefore s ∈ m.
(2) : If S is quasi-local and m is its unique maximal ideal, then by (1), S −U (S) =
m. On the other hand, if S − U (S) is an ideal of S, then any proper ideal of S is
16
PEYMAN NASEHPOUR
contained in S − U (S), since all elements of a proper ideal cannot be unit. This
implies that S−U (S) is the unique maximal ideal of S and the proof is complete.
Theorem 3.6. The following statements for a semiring S are equivalent:
(1) S is a discrete valuation semiring,
(2) S is a principal ideal MC semiring possessing a unique maximal ideal J 6=
(0),
(3) S is a unique factorization semiring with a unique (up to associates) irreducible element t,
(4) S is a quasi-local
T∞ MC semiring whose unique maximal ideal J 6= (0) is
principal and n=1 J n = (0),
(5) S is a quasi-local MC semiring, which satisfies ACCP and its unique maximal ideal J 6= (0) is principal.
(6) S is an MC semiring and there exists a nonzero and nonunit element t ∈ S
such that any nonzero ideal I of S is of the form I = (tn ) for some n ≥ 0.
Proof. First we prove (1) ⇒ (4) ⇒ (2) ⇒ (3) ⇒ (1).
(1) ⇒ (4): Let S be a discrete valuation semiring. It is clear that S is an MC
semiring. Also by Lemma 3.3, we know that S is quasi-local and its unique maximal
ideal J is nonzero and principal generated by an element t ∈ S such that v(t) = 1.
On the other hand, any ideal of S is principal. So by [11, Proposition 6.16], S is
Noetherian and in particular
T∞ it satisfies ACCP. Now since t is nonunit (because
v(t) = 1), by Lemma 3.4, n=1 J n = (0).
(4) T
⇒ (2): Let J = (t) be the maximal ideal of S and I its nonzero proper ideal.
n
n
Since ∞
n=1 J = (0), there exists a nonnegative number n such that I ⊆ J , but
n+1
n
n+1
I *J
. Our claim is that J ⊆ I. Let a ∈ I − J
. So we can suppose that
a = u · tn for some u ∈ S, where u ∈
/ J. By Proposition 3.5, u is a unit of S. This
means that (a) = (tn ) = J n for any a ∈ I − J n+1 . Now let x ∈ J n = (a). So
x = s · a. But a ∈ I, so x ∈ I. This means that all nonzero proper ideals of S are
in the form of J n = (tn ), where n ≥ 1.
(2) ⇒ (3): Let S be a principal ideal MC semiring such that the nonzero ideal
J = (t) is its unique maximal ideal. First we prove that t is an irreducible element
of S. On contrary, let t be reducible, then t = s1 s2 , where both s1 and s2 are
nonunit and this causes (t) ⊂ (s1 ) ⊂ S, which contradicts the maximality of J.
Now let t′ be another irreducible element of S and I = (m) be an arbitrary ideal
of S that contains (t′ ). So t′ = s · m for some s ∈ S and by definition either s or m
is a unit and this means that either (t′ ) = (m) or (m) = S. From this, we get that
(t′ ) is also a maximal ideal of S and by hypothesis, (t′ ) = (t), which means that S
has a unique (up to associates) irreducible element t. Since S is a principal ideal
semiring, it is Noetherian. So S satisfies ACCP. Now let s ∈ S be nonzero and
nonunit. Therefore by Lemma 3.4, there is a natural number n such that s ∈ (tn ),
while s ∈
/ (tn+1 ). Obviously this implies that s = utn , where u is a unit element of
S. So S is a unique factorization semiring.
(3) ⇒ (1): Let S be a unique factorization semiring with a unique (up to associates) irreducible element t. So any element of S can be written uniquely in the
form of utn for a unit u and a nonnegative integer n. This means that any element
of the fraction semifield F (S) of S can be written uniquely in the form of utn ,
where u is a unit and n is an integer number. Now define a map v : F (S) → Z∞
by v(utn ) = n for any n ∈ Z and v(0) = +∞. It is a routine discussion that S is a
VALUATION SEMIRINGS
17
V -semiring with respect to the triple (F (S), v, Z).
Now we show (4) ⇔ (5) ⇔ (6).
(4) ⇒ (5): It is clear that (2) implies (5) and since (4) is equivalent to (2), it
means that (4) implies (5).
(5) ⇒ (4): Let J = (t)
of S. Since t is
T∞be the nonzero unique maximalTideal
∞
nonunit, by Lemma 3.4, n=1 (tn ) = (0) and it means that n=1 J n = (0).
(5) ⇒ (6): By Lemma 3.3, (1) implies (6) and since (5) is equivalent to (1), it is
clear that (5) implies (6).
(6) ⇒ (5): Straightforward.
Let us recall that a semiring S is called an Euclidean semiring if there is a
function δ : S − {0} → N0 ∪ {+∞} satisfying the following condition:
For a, b ∈ S with b 6= 0, there exist q, r ∈ S satisfying a = qb + r with r = 0 or
δ(r) < δ(b) [11, Chap. 12, p. 136].
Proposition 3.7. Any discrete valuation semiring is an Euclidean semiring.
Proof. Let S be a V -semiring with respect to the triple (K, v, Z), where K is a
semifield and v is a surjective map. Define δ : S − {0} → N0 ∪ {+∞} by δ(s) = v(s)
for any nonzero s ∈ S and suppose that a, b ∈ S and b 6= 0. If δ(a) < δ(b), then
a = 0 · b + a. If δ(a) ≥ δ(b) then q = ab−1 ∈ S and therefore a = qb + 0 and this
completes the proof.
Now we are ready to prove another important theorem of this section. Before that
we need to recall the concept of Gaussian semirings, introduced in [22, Definition
7]: Let S be a semiring. For f ∈ S[X], the content of f , denoted by c(f ), is defined
to be the S-ideal generated by the coefficients of f . A semiring S is called Gaussian
if c(f g) = c(f )c(g) for all polynomials f, g ∈ S[X]. Also note that a semiring S is
called a subtractive semiring if each ideal of the semiring S is subtractive.
Theorem 3.8. Let S be a discrete valuation semiring. Then the following statements are equivalent:
(1) The unique maximal ideal of S is subtractive,
(2) The valuation map v : F (S) → Z∞ satisfies the min-property,
(3) S is a subtractive semiring,
(4) S is a Gaussian semiring.
Proof. (1) ⇒ (2): This is obvious by Corollary 1.15.
(2) ⇒ (3): Let the valuation map v : F (S) → Z∞ satisfy the min-property. Let
I be a nonzero proper ideal of S. By Lemma 3.3, there exists a t ∈ S with v(t) = 1
and n ∈ N such that I = (tn ) = {x ∈ S : v(x) ≥ n} and by Theorem 1.14, I is
subtractive. So we have already proved that each ideal of S is subtractive.
(3) ⇒ (4): Let S be a subtractive semiring and f, g ∈ S[X] be two arbitrary polynomials. By Dedekind-Mertens lemma for semirings ([22, Theorem 3]),
c(f )m+1 c(g) = c(f )m c(f g), where m = deg(g). If f = 0, then there is nothing to
prove. So suppose f 6= 0. Our claim is that c(f ) is a nonzero principal ideal of S.
In fact, since in valuation semirings principal ideals are totally ordered by inclusion
(Theorem 2.4), each finitely generated ideal has to be principal. But S is an MC
semiring, so c(f ) is a cancelation ideal of S and c(f )c(g) = c(f g) ([22, Proposition
15]).
18
PEYMAN NASEHPOUR
(4) ⇒ (1): If S is a Gaussian semiring, then by [22, Theorem 3], S is a subtractive
semiring. Therefore any ideal of S, in particular the unique maximal ideal of S is
subtractive and this finishes the proof.
Example 3.9. Let Q≥0 be the semifield of all nonnegative rational numbers. Every
nonzero element a/b ∈ Q≥0 can be written uniquely in the form of pn · a1 /b1 , where
n ∈ Z and a1 , b1 are both natural numbers and relatively prime to the given prime
number p. Now we define vp (a/b) = pn · a1 /b1 = n (Refer to Example 3.2). It is,
then, easy to see that vp is a discrete valuation on Q≥0 satisfying the min-property.
≥0
Therefore the semiring Q≥0
: p ∤ b} is a simple but good example for
p = {a/b ∈ Q
a subtractive discrete valuation semiring.
Now we bring the following proposition which gives us a general example for
subtractive discrete valuation semirings.
Proposition 3.10. Let D be a Dedekind domain and m an arbitrary maximal ideal
of D. Then the following statements hold:
(1) The semiring Id(D) is a Gaussian factorial semiring;
(2) The subsemiring {I/J ∈ F (Id(D)) : m ∤ J} of the semifield of fractions
F (Id(D)) is a subtractive discrete valuation semiring.
Proof. (1): Let D be a Dedekind domain. Then by [19, Theorem 6.16], every
nonzero proper ideal of D can be written as a product of prime ideals of D in
one and only one way, expect for the order of the factors. Also by [19, Corollary
6.17], every nonzero prime ideal of S is a maximal ideal. On the other hand, by
[19, Theorem 6.19], every nonzero ideal of D is invertible. From all we said we get
that (Id(D), +, ·) is a factorial semiring. Now we prove that Id(D) is a Gaussian
semiring. Let I, J ∈ Id(D). Since D is a Dedekind domain, any nonzero ideal of
D is invertible. Also since I ⊆ I + J, by [19, Theorem 6.20 (8)], there exists an
ideal K of D such that I = K(I + J). So I belongs to the principal ideal of the
semiring Id(D), generated by I + J, i.e., I ∈ (I + J). Similarly it can be proved
that J ∈ (I + J). So we have (I, J) = (I + J) and by [22, Theorem 8], Id(D) is a
Gaussian semiring.
(2): By considering the Example 3.2 and Theorem 3.8, the semiring Sm =
{I/J ∈ F (Id(D)) : m ∤ J} is a discrete valuation semiring and so we only need
to prove that the valuation map vm : F (Id(D)) → Z∞ has min-property. So let
I/J = mp I1 /J1 and K/L = mq K1 /L1 be arbitrary elements of F (Id(D)) such that
m ∤ I1 , J1 , K1 , L1 . Without loss of generality one may assume that p < q. Therefore
we have: vm (mp ((I1 L1 + m(q−p) J1 K1 )/J1 L1 ) = p + vm (I1 L1 + m(q−p) J1 K1 ). Our
claim is that m ∤ I1 L1 + m(q−p) J1 K1 . In contrary, if m | I1 L1 + m(q−p) J1 K1 , so
I1 L1 + m(q−p) J1 K1 is an element of the principal ideal (m). But we have proved
in above that Id(D) is a Gaussian semiring and by [22, Theorem 3], Id(D) is a
subtractive semiring and in particular (m) is a subtractive ideal of Id(D) and this
implies that I1 L1 ∈ (m), which contradicts our assumption that m ∤ I1 , L1 . So
vm (I1 L1 + m(q−p) J1 K1 ) = 0 and vm (I/J + K/L) = min{vm (I/J), vm (K/L)} and
the proof is complete.
Remark 3.11. Let Λ be a nonempty set. The set of all functions f : Λ → N0
L
(Λ)
(Λ)
with finite support is denoted by N0 (= λ∈Λ N0 ). We put SΛ = N0 ∪ {+∞}
(Λ)
and define the operations “min” and “+” on N0 componentwise and extend these
operations on SΛ as follows:
VALUATION SEMIRINGS
19
• min{f, +∞} = min{+∞, f } = f , for all f ∈ SΛ ,
• f + (+∞) = (+∞) + f = +∞, for all f ∈ SΛ .
It is easy to verify that (SΛ , min, +) is a semiring such that +∞ is its zero and
the constant zero function, i.e., f (λ) = 0 for all λ ∈ Λ, its 1SΛ . Our claim is that
if D is a Dedekind domain, then (Id(D), +, ·) and (SΛ , min, +) are isomorphic as
semirings for Λ = Max(D) and the routine proof is based on an interesting result
αn
1
mentioned in Proposition 6.8 in [17], which states that if I = mα
1 · · · mn and
min{α1 ,β1 }
min{αn ,βn }
β1
βn
J = m1 · · · mn , then I + J = m1
· · · mn
, where αi , βi are nonnegative integers, {mi } are distinct maximal ideals of D, and m0i is taken to be the
ideal D.
The question may arise if Proposition 3.11 can be generalized for the semiring
(Λ)
(SΛ , min, +), where SΛ = N0 ∪{+∞} and Λ is an arbitrary nonempty set. In fact,
a wonderful result by Luther Claborn, mentioned and proved in Theorem 15.18 in
[8], states that if Λ is an arbitrary nonempty set, then there is a Dedekind domain
D with a bijection σ : Max(D) → Λ.
Acknowledgements
The author was partly supported by Department of Engineering Science at University of Tehran and wishes to thank the anonymous referee for an extensive list
of comments and suggestions, which greatly helped to improve the paper. This
paper is dedicated to the memory of Prof. Dr. Manfred Kudlek, since he was the
professor who introduced the algebraic structure of semiring to me during my stay
in Tarragona, Spain.
References
[1] M. F. Atiyah and I. G. Macdonald, Introduction to Commutative Algebra, Addison-Wesley,
Reading, Massachusetts, 1969.
[2] N. Bourbaki, Commutative Algebra. Chapters 1–7, Elements of Mathematics, Springer, 1998.
[3] W. Bruns and J. Gubeladze, Polytopes, Rings, and K-Theory, Springer, Dordrecht, 2009.
[4] W. Bruns and J. Herzog, Cohen-Macaulay Rings, 2nd edn., Cambridge studies in advanced
mathematics: Vol. 39, Cambridge Univ. Press, Cambridge, 1998.
[5] D. J. Dulin and J. R. Mosher, The Dedekind property for semirings, J. of Aust. Math. Soc.,
14 (1972), 82–90.
[6] D. Eisenbud, Commutative Algebra with a View Toward Algebraic Geometry, SpringerVerlag, New York, 1995.
[7] A. J. Engler and A. Prestel, Valued Fields, Springer, Berlin, 2005.
[8] R. M. Fossum, The Divisor Class Group of a Krull Domain, Ergebnisse der Mathematik
und ihrer Grenzgebiete, Band 74, Springer-Verlag, Berlin, 1973.
[9] R. Gilmer, Commutative Semigroup Rings, The University of Chicago Press, 1984.
[10] J. S. Golan, Power Algebras over Semirings, Springer, 1999.
[11] J. S. Golan, Semirings and Their Applications, Kluwer Academic Publishers, Dordrecht,
1999.
[12] U. Hebisch and H. J. Weinert, Semirings - Algebraic Theory and Applications in Computer
Science, World Scientific, Singapore, 1998.
[13] Z. Izhakian, M. Knebusch, and L. Rowen, Supertropical semirings and supervaluations, J.
Pure Appl. Algebra 215, No. 10 (2011), 2431–2463.
[14] N. Jacobson, Basic Algebra II, 2nd ed., W. H. Freeman and Company, New York, 1989.
[15] J. Jun, Valuations of Semirings, arXiv preprint arXiv:1503.01392 (2015).
[16] I. Kaplansky, Commutative Rings, Allyn and Bacon, Boston, 1970.
[17] G. Karpilovsky, Topics in Field Theory, North-Holland Mathematics Studies Vol. 155, Amsterdam: North-Holland Publishing Company, 1989.
20
PEYMAN NASEHPOUR
[18] W. Krull, Beiträge zur Arithmetik kommutativer Integrittsbereiche. VI. Der allgemeine
Diskrimminantensatz. Unverzweigte Ringerweiterungen, Math. Z., Vol. 45 (1939), 1–19.
[19] M. D. Larsen and P. J. McCarthy, Multiplicative Theory of Ideals, Academic Press, New
York, 1971.
[20] A. W. Macpherson, On the difference between tropical functions and real-valued functions,
Preprint, arXiv:1507.00545, 2015.
[21] H. Matsumura, Commutative Ring Theory, 2nd ed., Cambridge Studies in Advanced Mathematics, 8, Cambridge University Press, Cambridge, 1989.
[22] P. Nasehpour, On the content of polynomials over semirings and its applications, J. Algebra
Appl., 15, No. 5 (2016), 1650088 (32 pages).
[23] G. Pilz, Near-Rings, The Theory and Its Aplications, Revised edition, North-Holland Publishing Company, Amsterdam-New York-Oxford, 1983.
[24] I. Swanson and C. Huneke, Integral Closure of Ideals, Rings, and Modules, London Mathematical Society Lecture Note Series 336, Cambridge University Press, Cambridge, 2006.
[25] J. Tolliver, Extension of valuations in characteristic one, Preprint, arXiv:1605.06425, 2016.
[26] F. Wang and H. Kim, Foundations of Commutative Rings and Their Modules, Algebra and
Applications: Vol. 22, Springer Nature, Singapore, 2016.
Peyman Nasehpour, Department of Engineering Science, Faculty of Engineering,
University of Tehran, Tehran, Iran
E-mail address: [email protected]
| 0math.AC
|
A Note on Extending Taylor’s Power Law for Characterizing Human
Microbial Communities: Inspiration from Comparative Studies on
the Distribution Patterns of Insects and Galaxies, and as a Case
Study for Medical Ecology and Personalized Medicine
Zhanshan (Sam) Ma
Computational Biology and Medical Ecology Laboratory
State Key Laboratory of Genetic Resources and Evolution
Kunming Institute of Zoology
Chinese Academy of Sciences, P. R. China
[email protected]
very useful for analyzing the distribution patterns of
human microbes within the human body. Finally, I suggest
that the parameters of the power law model built at
community level, especially when associated with
community metadata (or environmental covariates) and
when time series data from longitudinal studies are utilized,
can reveal important dynamic properties of human
microbiome such as community stability, which can be
invaluable in investigating etiology of some diseases
associated with human microbial communities. Overall, I
hope that the approach to human microbial community
with the power law offers an example that ecological
theories can play an important role in the emerging
medical ecology, which aims at studying the ecology of
human microbiome and its implications to human
diseases and health, as well as in personalized medicine.
ABSTRACT—Many natural patterns, such as the
distributions of blood particles in a blood sample, proteins
on cell surfaces, biological populations in their habitat,
galaxies in the universe, the sequence of human genes,
and the fitness in evolutionary computing, have been
found to follow power law (e.g., Kendal 2004, Illian et al.
2008, Venter 2001, Ma 2012). Taylor’s power law (Taylor
1961: Nature, vol. 189:732–735.) is well recognized as one
of the fundamental models in population ecology, thanks
to its wide applicability in describing the spatial
distribution patterns of biological populations. A
fundamental property of biological populations, which
Taylor’s power law reveals, is the near universal
heterogeneity (also known as aggregation or nonrandomness) of population abundance distribution in
habitat. Obviously, the heterogeneity also exists at the
community level, where not only the distributions of
population abundances but also the proportions of the
species composition in the community are often
heterogeneous or non-random. Indeed, community
heterogeneity is simply a reciprocal term of community
evenness, the major dimension what community diversity
tries to measure. Nevertheless, existing community
diversity indexes such as Shannon index and Simpson index
can only measure “local” or “static” diversity in the sense
that they are computed for each habitat at a specific time
point, but the indexes alone do not reflect the diversity
changes across habitats or over time. This inadequacy is
particularly problematic if the research objective is to
study the dynamics of community diversity, which is critical
for understanding the possible mechanisms that maintain
the community diversity and stability. In this note, I
propose to extend the application scope of Taylor’s power
law to the studies of human microbial communities,
specifically, the community heterogeneity at both
population (single species) and community (multiple
species) levels. I further suggested that population
dispersion models such as Taylor (1980, Nature: 286, 5355), which is known to generate population distribution
patterns consistent with the power law, should also be
Keywords: Power Law, Human Microbiome, Human
Microbial Communities, Medical Ecology, Personalized
Medicine, Point Process Statistics, Metagenomics.
BACKGROUND AND JUSTIFICATIONS
Maurer (1999) pointed out two possibilities for the lack of
generalizability in the study of community ecology. First,
generalizations may be impossible due to the extreme
complexity of ecological world. Second, patterns do not
exist at the scale of the local community where the
observation is made. Perhaps a third possibility can be
added to this list: that the tool for detecting the patterns may
be unsuitable at the research scale. While little may be done
with respect to the first possibility, ecologists can strive to
avoid the second and third possibilities by carefully
choosing a right scale level and developing powerful
approaches that are suitable for the chosen scale. In this
note, I propose to extend the application scope of Taylor’s
power law (Taylor 1961, 1977, 1984, 2007) from
population scale (level) to community scale (level) in order
to obtain an appropriate and powerful quantitative approach
for characterizing the microbial communities. The
inspiration comes from a phenomenon in population
1
Scott 1952, Neyman 1958). Even more inspiring is that the
studies on the distribution patterns of insects and microbes
in their habitat, as well as the studies on the distribution of
galaxies in the universe, performed by Neyman and other
scientists, evolved similarly in terms of the research
approaches developed in astronomy and biology,
respectively. In both astronomy and biology, three
mathematical approaches with slightly different names were
advanced; they are (i) frequency distribution—fitting the
occurrence frequency data of subjects (galaxies or insects)
per sampling unit to a probability distribution such as
Neyman distribution (Neyman 1939) and negative binomial
distribution (NBD); (ii) aggregation index (also known as
dispersion index)—using some statistics to measure the
aggregation (dispersion) degree of galaxies or insect
populations; (iii) power law modeling.
ecology where Taylor’s power law has been found to
describe the spatial distribution patterns of biological
populations in their habitat remarkably well. What Taylor’s
power law captures at the population scale is essentially the
heterogeneity of population abundance distribution in space
and/or time, and heterogeneity is certainly a fundamental
property at community level (scale). Therefore, an
interesting question is: can Taylor’s power law play a
similar role at the community level?
Taylor (1961) discovered that the Power Law model
[Equation (1)] fits population spatial distribution data
almost ubiquitously well from bacteria, plants, to mammals,
V = aM b ,
(1)
where M and V are population mean (density or abundance)
and variance, respectively, and a and b are parameters.
According to Taylor (1961, 1984): b>1 corresponds to
aggregated distribution (Fig. 1c) b=1 to random distribution
and (Fig. 1b) b<1 to regular distribution (Fig. 1a).
While Taylor’s power law [Equation (1)] describes the
distribution patterns of biological populations, the form of
power law for describing the distribution of galaxies takes a
slightly different form:
s
⎛r ⎞
1 + g (r ) = ⎜ 0 ⎟
(2)
⎝r⎠
where g(r) is the so-termed pair correlation function, s and
r0 are parameters. According to Illian et al. (2008), this
power law model [Equation (2)] was proposed by Totsuji &
Kihara (1969) to simplify Neyman et al’s (Neyman &Scott
1952, 1958, Neyman et al. 1953) computation and to relieve
the statistical difficulties Neyman et al (1953) encountered.
Neyman et al. (1953) statistical difficulty was that they
obtained different model parameters for different galaxy
catalogues (files of coordinates of galaxies); therefore, their
model was a probability distribution model that is a
superposition of groups of galaxies of varying size,
including super clusters. In fact, in biology, the discovery
and application of Taylor’s power law played a similar role
in the study of the distribution of biological populations,
especially in the study of insects (Taylor 1961, 1984). In
entomology, from 1920’s to 1950’s, several probability
distributions including negative binomial, Neyman, Thomas
Double Poisson, Polya distributions, had been developed to
fit the frequency data of insect population distribution. A
practical problem with the probability distribution approach
in entomology was that for each data set of population
distribution, there is a set of parameters, even for the same
probability distribution model, not to mention the
parameters from different probability models. A then wellaccepted solution was to find the common-K of the negative
binomial distribution fitted to different data sets, hoping
that the common-K to synthesize the information from
multiple samples and to represent the species-specific
characteristic of the biological species under investigation.
Later, it was found that common-K is a dubious index for
characterizing population spatial patterns and could not
assume the role hoped for by entomologists (Taylor et al.
1979, Perry et al. 1986). In a series of studies (Taylor 1961,
Taylor & Taylor 1977, Taylor et al. 1978, 1979, 1984,
1986), Taylor and his colleagues confirmed the wide
applicability of Taylor’s (1961) power law [Equation (1)] in
describing the ‘fat tails’ of the distribution of abundance of
Fig. 1. Three Distribution Patterns of Biological Populations:
(a): Regular, (b) Random, and (c) Aggregated.
Taylor's power-law accurately depicts the scaling
relationship that describes the ‘fat tails’ of the distribution
of abundance of nearly all organisms, and it has been found
to hold at multiple scale levels of population ecology.
Taylor’s power law is one form of the power law that is
found to underlie many natural processes such as the
distribution of galaxies in the universe, blood cells in a
blood sample. In addition, the three patterns displayed in
Figure 1 are considered as typical patterns of point process,
which is studied in a branch of spatial statistics, i.e, point
process statistics.
The point process and power law have been studied in
many fields of science; for example, in astronomy, both
spatial random fields (e.g., the density fields, the velocity
fields) and spatial point patterns (e.g., the position of
galaxies) have been studied extensively (Illian et al. 2008).
Astronomers had observed that matters in the universe is
distributed in aggregation (clusters) on small scales,
although, theoretically, the cosmological principle Einstein
assumed in 1917 considered that the universe is isotropic
and homogenous on large scale.
It is inspiring to note that Jerzy Neyman (1894-1981), a
distinguished mathematical statistician from the University
of California at Berkeley, studied the spatial distribution
patterns of insects, bacteria, and galaxies, apparently with
similar statistical approaches. Perhaps his study on galaxies
distribution was inspired by his experience in studying
bacteria and insects since his work in biology (Neyman
1939) was earlier than his work in astronomy (Neyman &
2
nearly all organisms across multiple scales in both space
and time; they also demonstrated the rarity of random
distribution of biological populations. They further found
that the parameter (exponent) b of Taylor’s power law
ranges from 1 to 2 for most species (hundreds of aphids,
moths, and birds in Great Britain) they investigated.
reached. Ma (1988, 1991) studied the relationship between
aggregation degree and population density and proposed the
concept of population aggregation critical density (PACD).
PACD can be computed with the following equation:
⎛ ln(a) ⎞
⎟⎟
(5)
m0 = exp⎜⎜
⎝ (1 − b) ⎠
where a & b are parameters of Taylor’s power law
[Equation (1)]. Based on PACD and the notion that
population aggregation is density-dependent, Ma (1988,
1991) reinterpreted Taylor’s power law, which consists of a
set of rules (which are based on the relationship among,
population density, m0, a, b, rather than b along as in the
original Taylor’s power law) for determining the spatial
distribution patterns of biological populations.
Back to astronomy, Totsuji & Kihara (1969) obtained
r0=15.3x106 light years and s=1.8. According to Illian et al
(2008), today the power law of the pair correlation function
g(r) is generally accepted in astronomy (Martinez & Saar
2002), and modern estimates of the exponent (s) range from
1.7 to 1.8. The range of s [Equation (2)] is surprisingly
similar to the range of b-value [Equation (10)] in biology
(between 1 and 2) as Taylor et al discovered (Taylor 1986).
The interpretation of the power law in astronomy is that
there are no characteristic scales (Illian et al. 2008), which
is a fundamental property of scale-free network. In a scalefree network, the degree distribution P(k) of network nodes
(k) is found to manifestly departure from Poisson
distribution predicted by random network theory and
instead follows power law (Albert & Bararasi 1999,
Newman & Barabasi et al. 2006, Barabasi 2009):
P(k ) ∝ k −λ .
(3)
It is well-known that the study of scale-free networks
initiated by Albert & Bararasi (1999) and others in the
1990s is part of the still explosively growing network
analysis studies in various disciplines towards a new
discipline of network science (Albert & Bararasi 1999,
Barabasi 2009). Obviously, the power laws Taylor (1961)
and Totsuji & Kihara (1969) discovered in the 1960s are
simply a manifestation of power law in the distributions of
organism and galaxies, respectively.
Nevertheless, I
believe that there is a need to investigate the earlier
discovered power laws (Taylor 1961, Totsuji & Kihara
1969) in the context of current network science since recent
advances have produced some powerful mathematical
methods and computational tools to analyze the
mechanisms underlying power law relationships. It is noted
that there is a practical difficulty to build network topology,
as adopted in the current studies on scale free networks,
with existing ecological data at the population level. But
this difficult is largely gone at the community level, which
explains the existing extensive studies on food web network
in community ecology.
In astronomy, power law is built with inter-point distance,
and in entomology, power law is built with population
density. Nevertheless, in entomology, distance has indeed
been utilized in modeling insect dispersion patterns. It is
generally agreed upon that insect dispersions generate
distribution patterns that follow power law. For example,
Taylor (1979, 1980) proposed a series of insect dispersion
models, which relate the distance (x) from dispersion center
with population density (N). Taylor found that the most
general form of insect dispersion model was
N = exp[a + bx c + d ln( x)],
(6)
where a, b, c, & d are model parameters. I suggest that
similar models to equation (6) should also be useful for
studying human microbiome.
From above brief comparative review on the studies in both
astronomy and biology, two points seem clear: (i) The
power laws in astronomy and biology are isomorphic and
they capture essentially the same or similar mechanisms
underlying the distributions of galaxies and insects,
respectively; (ii) Although in biology Taylor’s power law
has been applied only to the study of population distribution
patterns, in the studies on the distribution of galaxies in the
universe, there was not a similar limit, that is, power law
has been applied to study the distribution of various
galaxies, which are more similar to the community concept
in biology. Therefore, it seems natural to investigate the
applicability of Taylor’s power law at the community scale
in biology.
Furthermore, it is my opinion, there are at least three
incentives to pursue the extending of power law for
characterizing human microbial community: (i) The current
surge of interests in studying human microbial communities
initiated by US-NIH’s HMP (Human Microbiome Project),
fueled by the low-cost and high throughput new generation
of sequencing technology, has generated unprecedented
amount of data on human microbial communities. Given
that each individual subject carries a microbiome consisting
of multiple microbial communities (gut, skin, vaginal, oral,
nasal, etc), even for a project that only investigates one type
of microbial habitat (e.g., gut), there are multiple
communities to be analyzed as long as there are repetitions
of subjects (which are a must to conduct any statistical
analysis) in a study. Then, one must compare and hopefully
There are also parallel studies in astronomy and biology
regarding some kind of threshold values of aggregation
changes. In astronomy, r0 is such a threshold value, which
is equal to 15.3x106 light years and is a threshold value
termed minimum inter-point or hardcore distance. In
biology, Taylor (1981) proposed a Δ - model that tried to
explain the mechanism of Taylor’s power law, among other
objectives. The model is with the following form:
⎡⎛ R ⎞ s ⎛ R ⎞t ⎤
Δ = ε ⎢⎜ 0 ⎟ − ⎜ 0 ⎟ ⎥
(4)
⎢⎣⎝ R ⎠ ⎝ R ⎠ ⎥⎦
where Δ is a displacement (movement) of an animal, R is
the separation between animal individuals, ε , s, t are
parameters, R0 is a threshold across which an equilibrium is
3
synthesize the community diversity information from many
communities. Existing diversity indexes borrowed from
macro ecology are weak or even lack the capability for
synthesizing the information from multiple communities,
except for simple statistical comparisons of the values of a
diversity index. In addition, it is difficult to study the
dynamics of community diversities with the existing
diversity indexes. Power law can overcome these
limitations of existing diversity indexes, as it has been
demonstrated at the population level in biology and in
astronomy. (ii) Taylor’s power law essentially measures
variability and heterogeneity, which exist at both population
and community levels. Existing diversity indexes such as
Shannon index, Simpson index, essentially measures the
same properties (evenness and heterogeneity are simply
reciprocal terms), but do not possess the capability to
synthesize information from multiple communities.
Therefore, even purely acting as a diversity index, power
law should be superior to many of the existing diversity
indexes. (iii) Power law parameters for the microbial
community of a human individual have the potential to
become a personalized (individualized) metric related to
diseases and health, which can be invaluable for designing
and implementing personalized medicine schemes for
diseases associated with human microbial communities,
since the values of power law parameters should be
individual-dependent, perhaps more like finger printing. In
other words, power-law approach to human microbial
communities potentially offers a promising case study for
medical-ecology-supported personalized medicine.
Bioinformatics processing of the
metagenomic sequence data, which
generates species abundance table of
gene reads (Table 1 & 2).
PCR processing of the samples &
Sequencing of the 16s-RNA Gene.
Recruit volunteers & taking human
microbiome samples for either crosssectional or longitudinal study of a
cohort of individual subjects.
Figure 2. Pipeline of the metagenomic study of human
microbiome.
Table 1. Assumed abundance table of 16sRNA gene reads of
various bacterial species in human microbial communities
generated from a typical cross-sectional study of 100 individual
human subjects (numbered from #400 to #499).
Species-N
Subject
Species-1 Species-2 … …
N=500, Reads
Number
Reads
Reads
#400
100
1200
0
#401
200
800
1000
…
…
…
…
#499
1000
100
50
Table 2. Assumed abundance table of 16sRNA gene reads of
various bacterial species in human microbial communities
generated from a typical longitudinal study of 100 individual
st
st
human subjects (#400—#499) from January 1 to December 1 .
Species-N
SpeciesSpeciesSubject
Date
…
1 Reads
2 Reads
Number
… N=500,Reads
#400
Jan 1
100
1200
… 0
…
…
…
…
… 1000
…
Dec 1
500
2000
… 10
#499
Jan 1
200
1000
… 10
…
…
…
…
… …
…
Dec 1
700
900
… 0
A PROPOSAL FOR THE EXTENSIONS OF
TAYLOR’S POWER LAW
Figure 2 shows a typical flow of metagenomic study of
human microbiome. The product of the sequencingbioinformatics pipeline is usually a species abundance table
of gene reads in the format of Table (1) or Table (2) for
cross-sectional experiments or longitudinal study,
respectively. Table (1) contains the assumed gene reads
data of 100 individual subjects from a cross-sectional study,
i.e., each individual subject is sampled only once. In Table
(1), there would be abundance data of S=500 bacterial
species from 100 samples, which represent 100 microbial
communities. Table (2) contains the assumed gene reads
data of 100 individuals from a longitudinal study, sampled
from January 1st to December 1st, say every 10 days. Then
36 samples for each individual subject would have been
taken, and totally 3600 samples (100x36) or 3600 microbial
communities to be analyzed.
Table 3. Assumed abundance data of abundance distribution at
different locations of human vagina, with reference to a study
conducted by Kim et al. (2009)
Upper 1/3 Middle 1/3
Lower 1/3
Outer
Species
Abundance of
Species 1
100
900
1500
2000
20
1600
100
70
…
Abundance of
Species n
In the study of human microbiome, it is also possible to get
the abundance data at different locations in an organ, e.g.,
population abundance of bacterial species at different
locations of human vagina (e.g., Kim et al. 2009). It is then
possible to apply the dispersion models such as [Equation
(6)] to describe the location-dependent abundance
distribution. Table (3) shows an assumed data set with
human vaginal microbial community as an example.
With the assumed data formats displayed in Table (1) and
(2), for cross-sectional and longitudinal studies,
respectively, I envision five ways in which Taylor’s power
law models can be built to characterize the properties of
human microbial communities such as community diversity
and stability.
From a pure statistic perspective, building Taylor’s power
law model is a two-step process. The first step is to
4
compute the pairs of mean [Mi] and variance [Vi] from the
abundance data (which is in matrix form), and the second
step is to fit the model, which can be as simple as
performing a simple linear regression in the following form:
ln(V ) = ln(a ) + b ln( M ).
(7)
More complex algorithms to directly fit Taylor’s power law
[Equation (1)] are available in the literature [e.g.,
Marquardt’s algorithm, see Ma (1990, 1992)], but the
simple linear regression usually suffices.
individual subjects, as well as “synthesized” information
across N bacterial species. To some extent, these power law
parameters are similar to a “common” aggregation index of
N bacterial species, since the term aggregation largely
means skewed variation in the study of the spatial
distribution of population abundance, and the term
“common” again refers to some kind of “synthesis” of
information from many (N) bacterial species, similar to the
usage of term “common” in the common-K of negative
binomial distribution as mentioned previously.
It is noted that, when referring to the parameters of Taylor’s
power law, b is the most informative biologically and a is
the least informative since a is mainly influenced by
sampling schemes. Parameter m0, i.e., PACD (population
aggregation critical density), offers insights on the dynamic
relationship between aggregation degree and population
density in population ecology, but its biological
significance should be investigated further when the power
law is extended to community scale.
In summary, when Taylor’s power law is applied to crosssectional studies of human microbial communities, it may
offer insights on “common” diversity measure of microbial
communities for (across) multiple subjects, and “common”
aggregation measure for (across) multiple bacterial species.
The Applications to Longitudinal Studies
Obviously, cross-sectional data format is a special case of
longitudinal data format, i.e., a cross-sectional snapshot of
the longitudinal data at a fixed time. Another way to
examine the same data is to replace the time-series
observations for each subject with their mean, and then the
whole dataset of a longitudinal study is turned into a crosssectional dataset. The above discussed two applications of
Taylor’s power law to cross-sectional data can be applied to
the converted (taking the mean of time-series observations)
longitudinal data in the same manner. However, as it will
become clear from the following discussion, that the power
law models built from longitudinal data, even after
conversion, should reflect some dynamic properties of
microbial communities. I classify this “mean-conversion”
(i.e., using the means of time-series observations) based
application of Taylor’s power law as the third approach to
apply the power law for characterizing human microbial
communities.
The Applications to Cross-sectional Studies
For a cross-sectional study, two ways for applying Taylor’s
power [Equation (1)] law should be possible: one approach
is to compute pairs of the mean (M) and variance (V) across
columns (one pair for each row) and another one is to
compute M & V across rows (one pair for each column).
With the first way, the mean (Mi) and variance (Vi) reflect
the average abundance and variability of all bacterial
species detected in the communities, respectively. There is
a pair of mean (Mi) and variance (Vi) for the community of
each subject, and totally S pairs of M and V, that is,
i=1,2,…S, where S is the number of subjects recruited in the
study. The second step is simply to build a regression
model with S pairs of means and variances based on
Equation (7), and any statistical software would be able to
perform the task. In the assumed case of Table (1), S would
be 100 subjects, numbered from 400 to 499. The power
law parameters obtained in this application manner would
reflect the abundance variability of various bacterial species
across S subjects. In other words, the parameters are
determined by the variability of N bacterial species, and
furthermore, the variability information among S subjects is
also synthesized. To some extent, the parameters of
Taylor’s power law obtained in this type of application are
similar to a “common” diversity index that synthesizes the
diversity information of many subjects, similar to the
common-K of the negative binomial distribution in the
literature of the spatial distribution pattern of insect
population (Taylor et al. 1979, Perry et al. 1986), but
without the deficiency of the common-K.
A next immediate question is, whether the longitudinal
datasets can be used to fit Taylor’s power law directly,
without using the “mean-conversion” transformation. The
answer should be a sound yes. First, the longitudinal data
in the format of Table 2 can be used to fit Taylor’s power
law without combing time-series observations into mean
values, by simply treat time-series observations as
repetitions (repeated observations) of the subjects. I expect
that the results from this approach may be similar to the
above discussed “mean-conversion”-based analysis, but
further investigation with real world metagenomic data
should be performed to find out the difference, if there is
any. Nevertheless, I believe that the most interesting and
also the valuable applications of Taylor’s power law to
analyze longitudinal data should be performed on the
individual-based as discussed below. In the following
proposed application scheme, the dataset of each individual
is used to build the power law model separately. This
should be quite natural since all the time-series observations
for each individual subject belong to the same microbial
community. In other words, each individual carries a
microbial community at a specific habitat (e.g., gut).
Therefore, from community ecology perspective,
individual-based power-law modeling is truly at the
In contrast, the second approach of applying Taylor’s
power law to cross-sectional study computes Mj & Vj across
rows, where j=1, 2, … N species, i.e., a pair of Mj and Vj
for each bacterial species, and totally N pairs of Mj and Vj.
The follow step would be to obtain the parameters of
Taylor’s power law model by a simple linear regression
analysis with the N pairs of data [Equation (7)]. The power
law parameters obtained in this application manner reflect
the variability of bacterial species abundance among
5
community scale, and the previously discussed applications
of the power law are more about meta-communities.
In summary, when Taylor’s power law is applied to
longitudinal studies of human microbial communities, it
may offer insights on the stability of the community
diversity over time series, and on the stability of “common”
species aggregation.
Characterizing the stability of community diversity over
time with Taylor’s power law
As stated previously, the dataset used for performing
Taylor’s power law analysis is individual based, that is, we
apply the power law to characterize a single dynamic
community, or various snapshots of the same community.
For example, for the subject #400 in Table 2, we get an
abundance dataset consisting of 12x3=36 rows (3
observations per month for 12 months) of the observations
of N=500 bacterial species. Obviously, we can perform the
same power law analysis to each of the subjects (S=100 in
Table 2).
It is further noted that the power law modeling of human
microbial communities when applied to longitudinal studies
generates a set of power law parameters for each individual
subject, which characterizes the dynamic properties of
species aggregation and community diversity, or the
stability of microbial community. This kind of insights can
be invaluable for design and implementing medicalecology-supported personalized medicine, as being argued
in the next section.
When Taylor’s power law is applied to the dataset of an
individual subject, there are also two ways that the
application can be done, depending on how mean (M) and
variance (V) are computed. In both kinds of applications,
the computational procedures for building the power law
model are actually the same as those used for the previously
discussed modeling with cross-sectional data. Similar to the
case of cross-sectional data, the computation of mean (M)
and variance (V) can be cross-columns or cross-rows,
corresponding to two ways Taylor’s power law are applied
for analyzing individual-based datasets.
The Application of Dispersion Models
The dispersion models such as Equation (6) proposed by
Taylor (1979, 1980) can be used to describe location
dependent population abundance data of bacterial species
such as assumed in Table (3). It may be necessary to
develop new dispersion models for human microbiome,
other than directly applying Taylor’s (1979, 1980) models,
since unlike Taylor’s power law, the existing dispersion
models in literature are empirical and lack the necessary
generality.
PERSPECTIVES: A CASE STUDY FOR
MEDICAL ECOLOGY AND PERSONALIZED
MEDICINE
For example, with the assumed datasets displayed in Table
2, for subject #400, crossing 500 columns (corresponding to
N=500 bacterial species), we can compute 36 pairs (12
months, each month with 3 samples) of mean (M) and
variance (V). The 36 pairs of mean-variance data can be
used to fit Taylor’s power law [Equation (7)]. Since each
pair of the mean and variance reflects the heterogeneity
(diversity) of the community at each time point, the
parameters of the power law then should reflect the
dynamic property of the community diversity over time
series. In other words, the power law model built with cross
time-series mean-variance pairs should reveal the stability
of community diversity over time. I classify the Taylor’s
power built in this manner the fourth type of power law
modeling of microbial community with a mission to reveal
the stability of community diversity over time.
According to the website: www.medicalecology.org/, which
is maintained by Dr. Dickson Despommier and Dr. Steven
Chen of Columbia University, the term Medical Ecology
was first coined by eminent microbiologist Rene Dubos,
who believed that natural world could offer many of our
needs if explored fully. Dubos was apparently inspired by
events such as the discoveries of penicillin and gramicidin
(his own discovery), in which soil microbes played a
critical role, as well as the treatment of malaria with quinine.
Recent redefinition of medical ecology has much broader
meanings. For example, www.medicalecology.org/ stated
that “Medical Ecology is an emerging science that defines
those aspects of the environment that have a direct bearing
on human health. The concept of ecosystem functions and
services helps to describe global processes that contribute
to our well-being, helping to cleanse the air we breathe, the
water we drink, and the food we eat. Environmental
degradation often leads to alterations in these aspects,
leading to various states of ill health.”
Characterizing the stability of “common” species
aggregation with Taylor’s power law
The mean/variance pairs can also be computed across rows
for each subject. For example, for subject#400, we can
compute one pair of mean and variance for each bacterial
species across all of the 36 time-series observations; we can
obtain as many as 500 pairs of the mean-variance values,
i.e., each species with one pair. The 500 pairs of meanvariance data can then be fitted to the power law [Eq. (7)].
Since each pair of the mean and variance computed across
time series reflect the dynamic aggregation of each species,
the power law model built in this way should represent
aggregation dynamics common to all 500 species in the
community. The term “common” species aggregation
stability (i.e., the stability of aggregation common to all
species is obtained with this type of power law modeling.
In the post-HGP (Human Genomic Project) and HMP
(Human Microbiome Project) era, it is high time to expand
the scope of medical ecology again. It is my opinion that the
study of the relationship between human microbiome and
human host as well as its implications to human health and
diseases should be considered as a core component of
newly expanded medical ecology. I further believe that the
emerging medical ecology is at the stage when medical
genetics was emerging in the 1960s, and it should
ultimately assume a similar role in medicine as today’s
6
medical genetics assumes. In the following discussion, I
temporarily ignore the existing aspects of medical ecology;
instead, I focus exclusively on the human-microbiomecentered medical ecology.
likely and important fields. It is widely recognized that
ecology, evolution and behavior are three interwoven
disciplines. Therefore, in Figure 4, I pick three research
fields, each of which focuses on one of the three disciplines.
These fields are: (i) applications and extensions of
ecological theories from macro-ecology; (ii) evolution of
symbiosis, cooperation and mutualism of “human
microbiome/host” relationships, especially the role of
human immune system; (iii) communication behavior of
human microbiome and the communication between
microbiome and host. Of course, this list of three is far
from complete, and it is largely an impromptu attempt to
present a glimpse of the research topics of emerging
medical ecology. I will discuss the details of these topics in
a separate paper.
One may wonder if it is really necessary to redefine medical
ecology rather than putting the related research topics in the
context of some more familiar disciplines, such as
microbial ecology, clinical microbiology, bioinformatics,
and/or eco-informatics. In my opinion, no existing scientific
disciplines can accommodate all of the interesting research
themes that are waiting for medical ecology to embrace.
Furthermore, to advance medical ecology, there will be
some new theories and techniques to emerge, and these new
theories and techniques need the umbrella of a redefined
medical ecology. Figure 3 is a diagram showing the
relationship between redefined medical ecology and its core
parent fields, as well as its supporting and supported fields.
In Figure 3, medical ecology is depicted as a crossdisciplinary and trans-disciplinary subject with “native
parents” of medicine, ecology, and microbiology. Its
advances are directly dependent on various -omics research
(in particular metagenomics), computational biology,
bioinformatics, system biology and system ecology, etc. It
is my belief that medical ecology should become a
foundation of personalized medicine, and that personalized
medicine is likely to be the biggest beneficiary of research
in medical ecology.
Finally, I argue that the application of Taylor’s power law
for characterizing human microbial communities can serve
as a case study of medical ecology and personalized
medicine for the following reasons: (i) According to Figure
4, the study falls into the first category of the research
topics of medical ecology, i.e., the application of ecological
theories, mostly from macro-ecology. Taylor’s power law is
one of a countable few laws in population ecology of
animals and plants. (ii) As explained previously, the power
law parameters, when obtained from longitudinal studies of
human microbial communities, capture individual-specific
and dynamic properties (community diversity and species
aggregation) of an individual’s microbial community. It
should be possible to establish functional relationship
between the individual-specific power law parameters and
the metadata covariates that reflect the health/illness states
of an individual. Understanding this kind of functional
relationship can offer deep insights for practicing
personalized medicine, especially for those diseases that are
associated with human microbiome.
Personalized Medicine & Public Health
Key Supported Fields by Medical Ecology
Medical
Science
Ecological
Science
(i) Ecology Focus: Applying and extending existing macroecological theories; create new ecological theories and
discover new laws specific to human microbiome.
(ii) Evolution Focus: Study the evolution of humanmicrobiome symbiosis (mutualism, cooperation),
especially the role of human immune system, as well as
the possible phase transitions between healthy, subhealthy and disease states of human body.
(iii) Behavior focus: Study the communication behavior of
human microbiome system, e.g., quorum sensing,
communication with host, especially with immune system.
Medical
Ecology
Medical
Microbiology
Bioinformatics and Computational Biology;
Omics: Genomics, Metagenomics, Proteomics,
Ecology
Metabolomics, Transcriptomics, Personal Genomics,
Pharmacogenomics, Pharmacomicrobiomics etc;
Evolution
Behavior
Network
Science …
Human Microbiome / Host Ecosystem
System Biology and System Ecology
Key Supporting Fields of Medical Ecology
Suggested Focal Fields of Medical Ecology
Figure 3. Emerging Human-Microbiome-Centered Medical
Ecology, its parents, supporting and supported fields.
Figure 4. Suggested focal fields of Medical Ecology
In fact, scholars have already explored the application of
Taylor’s power law in epidemiology and evolutionary
ecology of parasites (Morand & Krasnov 2008). For
example, Morand & Krasnov (2008) suggested a possible
Research in medical ecology should develop some unique
fields and theories in the near future. Obviously, predicting
them is not an easy task, but here I try to identify three
7
Morand, S. and B. Krasnov. 2008. Why apply ecological laws to
epidemiology? Trends in Parasitology, 24(7):304-309
link between host defense against parasite and the value of
the parameter b of Taylor’s power law (when applied to
measure the aggregation of parasite populations). In
personalized medicine, frailty analysis can be an effective
statistical modeling approach (Ma et al. 2011), where the
distribution of frailty (disease risk) can follow the PVF
(power variance function) distributions. PVF is a family of
probability distributions and is different from Taylor’s
power law. But in this distribution family, the variance and
mean do satisfy power law function (which is the
mathematical essence of Taylor’s power law). It will be
interesting to investigate the underlying biomedical process
that may lead to the PVF frailty distribution. In a separate
front, investigating the existence and maintenance of scalefree networks (of which power law is a fundamental
property) in human microbiome should also be an important
endeavor. Putting power law analysis in the context of
network science (e.g., Bascompte 2007, Ings 2009,
Barabasi 2009) should produce synergetic advances in
applying power law to the study of human microbiome and
medical ecology.
Newman, M, Barabasi, AL, Watts, DJ. (eds). 2006. The Structure and
Dynamics of Networks. Princeton University Press. 583pp.
Neyman, J. 1939. On a new class of 'contagious' distributions, applicable
in entomology and bacteriology. Annals of Math. Stat., vol. 10, pp. 35-57.
Neyman, J., and Scott, E.L. (1952) A theory for the spatial distribution of
galaxis. Astrophys. J. 116, 144–163.
Neyman, J., Scott, E. L., and Shane, C. D. 1953. On the spatial distribution
of galaxies: a specific model. Astrophys. J., 117-92-133.
Neyman, J., and Scott, E.L. 1958. Statisical approach to problems of
cosmology (with discussion). J. Roy. Statist. Soc. B 20, 1–43.
Taylor, L. R. 1961. Aggregation, variance and the mean. Nature 189:732–
735.
Taylor, L. R., and R. A. J. Taylor. 1977. Aggregation, migration and
population mechanics. Nature 265:415–421.
Taylor, L. R., I. P. Woiwod and J. N. Perry. 1978. The DensityDependence of Spatial Behaviour and the Rarity of Randomness. Journal
of Animal Ecology, Vol. 47, No. 2, pp. 383-406
REFERENCE
Barabasi, AL, Albert, A. 1999. Emergence of scaling in random networks.
Science 286, 509-512
Taylor, L. R. I. P. Woiwod and J. N. Perry. 1979. The Negative Binomial
as a Dynamic Ecological Model for Aggregation, and the Density
Dependence of k. Journal of Animal Ecology, Vol. 48, No. 1:pp. 289-304
Barabasi, AL. 2009. Scale-free networks: a decade and beyond. Science,
325:412-413.
Taylor, L. R. 1984. Assessing and interpreting the spatial distribution of
insect populations. Annual Review of Entomology, 29, 321-357.
Bascompte, J. 2007. Networks in ecology. Basic and Applied Ecology,
8:485-490.
Taylor, L. R. 1986. Synoptic dynamics, migration and the Rothamsted
Insect Survey. Journal of Animal Ecology 55:1–38.
Illian, J. A. Penttinen, H. Stoyan, and D. Stoyan. 2008. Statistical Analysis
and Modeling of Spatial Point Patterns. Wiley.
Taylor, R. A. J. 1978. The relationship between density and distance of
dispersing insects. Ecological Entomology, 3, 63-70.
Ings, T. C. et al. 2009. Review: Ecological networks – beyond food webs.
J. Animal Ecology, 78(1): 253–269, January 2009.
Taylor, R. A. J. 1980. A family of regression equations describing the
density distribution of dispersing organisms. Nature, London, 286,53-55.
Kendal, W. S. 2004. Taylor’s ecological power law as a consequence of
scale invariant exponential dispersion models. Ecological Complexity.
1:193–209
Taylor, R. A. J. 1981. The behavioural basis of redistribution. I. The Deltamodel concept. Journal of Animal Ecology, 50:573–586
Taylor, R. A. J. 2007. Obituary: Roy (L. R.) Taylor (1924–2007). Journal
of Animal Ecology. 76, 630–631
Kim, T. K. et al. 2009. Heterogeneity of Vaginal Microbial Communities
within Individuals. J. Clin. Microbiol. vol. 47 no. 4 1181-1189
Totsuji, H., and Kihara, T. (1969) The correlation function for the
distribution of galaxies. Publ. Astron. Soc. Japan 21, 221–229.
Ma, Z. S. 1988. Revised Taylor’s Power Law and population aggregation
critical density, in: Proceedings of Annual National Conference of the
Entomological Society China. 1988, Nanjing, China (in Chinese), 1988.
Perry, J. N. & L. R. Taylor. 1986. Stability of Real Interacting Populations
in Space and Time: Implications, Alternatives and the Negative Binomial
kc quick view. Journal of Animal Ecology, Vol. 55, No. 3, pp. 1053-1068
Ma, Z. S. 1990. Fitting Taylor’s power law model with Marquardt’s
algorithm, J. Beijing Forestry University. 12(1):49-54
Venter, J.C., Adams, M.D., Myers, E.W., Li, P.W., Mural, R.J., Sutton, G.G., Smith, H.O., Yandell, M.,
Evans, C.A., Holt, R.A., Gocayne, J.D., Amanatides, P., Ballew, R.M., Huson, D.H., Wortman, J.R., Zhang,
Q., Kodira, C.D., Zheng, X.H., Chen, L., Skupski, M., Subramanian, G., Thomas, P.D., Zhang, J., Gabor
Miklos, G.L., Nelson, C., Broder, S., Clark, A.G., Nadeau, J., McKusick, V.A., Zinder, N., Levine, A.J.,
Roberts, R.J., Simon, M., Slayman, C., Hunkapiller, M., Bolanos, R., Delcher, A., Dew, I., Fasulo, D.,
Flanigan, M., Florea, L., Halpern, A., Hannenhalli, S., Kravitz, S., Levy, S., Mobarry, C., Reinert, K.,
Remington, K., Abu-Threideh, J., Beasley, E., Biddick, K., Bonazzi, V., Brandon, R., Cargill, M.,
Chandramouliswaran, I., Charlab, R., Chaturvedi, K., Deng, Z., Francesco, V.D., Dunn, P., Eilbeck, K.,
Evangelista, C., Gabrielian, A.E., Gan, W., Ge, W., Gong, F., Gu, Z., Guan, P., Heiman, T.J., Higgins, M.E.,
Ji, R.-R., Ke, Z., Ketchum, K.A., Lai, Z., Lei, Y., Li, Z., Li, J., Liang, Y., Lin, X., Lu, F., Merkulov, G.V.,
Milshina, N., Moore, H.M., Naik, A.K., Narayan, V.A., Neelam, B., Nusskern, D., Rusch, D.B., Salzberg, S.,
Shao, W., Shue, B., Sun, J., Wang, Z.Y., Wang, A., Wang, X., Wang, J., Wei, M.-H., Wides, R., Xiao, C.,
Yan, C., Yao, A., Ye, J., Zhan, M., Zhang, W., Zhang, H., Zhao, Q., Zheng, L., Zhong, F., Zhong, W., Zhu,
S.C., Zhao, S., Gilbert, D., Baumhueter, S., Spier, G., Carter, C., Cravchik, A., Woodage, T., Ali, F., An, H.,
Awe, A., Baldwin, D., Baden, H., Barnstead, M., Barrow, I., Beeson, K., Busam, D., Carver, A., Center, A.,
Cheng, M.L., Curry, L., Danaher, S., Davenport, L., Desilets, R., Dietz, S., Dodson, K., Doup, L., Ferriera, S.,
Garg, N., Gluecksmann, A., Hart, B., Haynes, J., Haynes, C., Heiner, C., Hladun, S., Hostin, D., Houck, J.,
Howland, T., Ibegwam, C., Johnson, J., Kalush, F., Kline, L., Koduru, S., Love, A., Mann, F., May, D.,
McCawley, S., McIntosh, T., McMullen, I., Moy, M., Moy, L., Murphy, B., Nelson, K., Pfannkoch, C.,
Pratts, E., Puri, V., Qureshi, H., Reardon, M., Rodriguez, R., Rogers, Y.-H., Romblad, D., Ruhfel, B., Scott,
R., Sitter, C., Smallwood, M., Stewart, E., Strong, R., Suh, E., Thomas, R., Tint, N.N., Tse, S., Vech, C.,
Wang, G., Wetter, J., Williams, S., Williams, M., Windsor, S., Winn-Deen, E., Wolfe, K., Zaveri, J., Zaveri,
K., Abril, J.F., Guigo, R., Campbell, M.J., Sjolander, K.V., Karlak, B., Kejariwal, A., Mi, H., Lazareva, B.,
Hatton, T., Narechania, A., 2001. The sequence of the human genome. Science 291,
Ma, Z. S. 1991. Further interpreted Taylor’s Power Law and population
aggregation critical density, Trans. Ecol. Soc. China (1991) 284–288.
Ma, Z. S. 1992. Optimization of nonlinear ecological models with the
accelerated simplex algorithm. J. of Biomathematics, 7(2):160-167.
Ma, Z. S., A. Zaid, L. J. Forney. 2011. Caring about trees in the forest:
incorporating frailty in risk analysis for personalized medicine.
Personalized Medicine, 8(6):681-688.
Ma, Z. S. 2012. Chaotic populations in genetic algorithms. J. Applied Soft
Computing. doi:10.1016/j.asoc.2012.03.001
Martínez, V.J., and Saar, E. 2002. Statistics of the Galaxy Distribution.
Chapman & Hall/CRC, Boca Raton, FL.
1304–1351.
Maurer, B. A. 1999. Untangling Ecological Complexity. The University of
Chicago Press.
8
| 5cs.CE
|
arXiv:1703.03963v1 [cs.DS] 11 Mar 2017
On Solving Travelling Salesman Problem with Vertex
Requisitions∗
Anton V. EREMEEV
Omsk Branch of Sobolev Institute of Mathematics SB RAS,
Omsk State University n.a. F.M. Dostoevsky
[email protected]
Yulia V. KOVALENKO
Sobolev Institute of Mathematics SB RAS,
[email protected]
Received: / Accepted:
Abstract: We consider the Travelling Salesman Problem with Vertex Requisitions, where
for each position of the tour at most two possible vertices are given. It is known that
the problem is strongly NP-hard. The proposed algorithm for this problem has less time
complexity compared to the previously known one. In particular, almost all feasible instances
of the problem are solvable in O(n) time using the new algorithm, where n is the number of
vertices. The developed approach also helps in fast enumeration of a neighborhood in the
local search and yields an integer programming model with O(n) binary variables for the
problem.
Keywords: Combinatorial optimization, System of vertex requisitions, Local search, Integer
programming.
MSC: 90C59, 90C10.
1
INTRODUCTION
The Travelling Salesman Problem (TSP) is one of the well-known NP-hard combinatorial optimization problems [1]: given a complete arc-weighted digraph with n vertices, find
a shortest travelling salesman tour (Hamiltonian circuit) in it.
The TSP with Vertex Requisitions (TSPVR) was formulated by
A.I. Serdyukov in [2]: find a shortest travelling salesman tour, passing at i-th position
This research of the authors is supported by the Russian Science Foundation Grant (project no. 15-1110009).
∗
1
a vertex from a given subset X i , i = 1, . . . , n. A special case where |X i | = n, i = 1, . . . , n, is
equivalent to the TSP.
This problem can be interpreted in terms of scheduling theory. Consider a single machine
that may perform a set of operations X = {x1 , . . . , xn }. Each of the identical jobs requires
processing all n operations in such a sequence that the i-th operation belongs to a given
subset X i ⊆ X for all i = 1, . . . , n. A setup time is needed to switch the machine from
one operation of the sequence to another. Moreover, after execution of the last operation
of the sequence the machine requires a changeover to the first operation of the sequence to
start processing of the next job. The problem is to find a feasible sequence of operations,
minimizing the cycle time.
TSP with Vertex Requisitions where |X i | ≤ k, i = 1, . . . , n, was called k-TSP with
Vertex Requisitions (k-TSPVR) in [2]. The complexity of k-TSPVR was studied in [2]
for different values of k on graphs with small vertex degrees. In [3] A.I. Serdyukov proved
the NP-hardness of 2-TSPVR in the case of complete graph and showed that almost all
feasible instances of the problem are solvable in O(n2) time. In this paper, we propose an
algorithm for 2-TSPVR with time complexity O(n) for almost all feasible problem instances.
The developed approach also has some applications to local search and integer programming
formulation of 2-TSPVR.
The paper has the following structure. In Section 2, a formal definition of 2-TSPVR is
given. In Section 3, an algorithm for this problem is presented. In Section 4, a modification
of the algorithm is proposed with an improved time complexity and it is shown that almost
all feasible instances of the problem are solvable in time O(n). In Section 5, the developed
approach is used to formulate and enumerate efficiently a neighborhood for local search. In
Section 6, this approach allows to formulate an integer programming model for 2-TSPVR
using O(n) binary variables. The last section contains the concluding remarks.
2
PROBLEM FORMULATION AND ITS HARDNESS
2-TSP with Vertex Requisitions is formulated as follows.
Let G = (X, U)
be a complete arc-weighted digraph, where X = {x1 , . . . , xn } is the set of vertices,
U = {(x, y) : x, y ∈ X, x 6= y} is the set of arcs with non-negative arc weights ρ(x, y),
(x, y) ∈ U . Besides that, a system of vertex subsets (requisitions) X i ⊆ X, i = 1, . . . , n,
is given, such that 1 ≤ |X i | ≤ 2 for all i = 1, . . . , n.
Let F denote the set of bijections from Xn := {1, . . . , n} to X, such that f (i) ∈ X i , i =
1, . . . , n, for all f ∈ F . The problem consists in finding such a mapping f ∗ ∈ F that ρ(f ∗ ) =
n−1
P
min ρ(f ), where ρ(f ) =
ρ(f (i), f (i + 1)) + ρ(f (n), f (1)) for all f ∈ F . Later on the
f ∈F
i=1
symbol I is used for the instances of this problem.
Any feasible solution uses only the arcs that start in a subset X i and end in X i+1 for
some i ∈ {1, . . . , n} (we assume n + 1 := 1). Other arcs are irrelevant to the problem and we
assume that they are not given in a problem input I.
2-TSPVR is strongly NP-hard [3]. The proof of this fact in [3] is based on a reduction of
Clique problem to a family of instances of 2-TSPVR with integer input data, bounded by a
polynomial in problem length. Therefore, in view of sufficient condition for non-existence of
Fully Polynomial-Time Approximation Scheme (FPTAS) for strongly NP-hard problems [4],
the result from [3] implies that 2-TSPVR does not admit an FPTAS, provided that P6=NP.
The k-TSPVR with k ≥ 3 cannot be approximated with any constant or polynomial factor
of the optimum in polynomial time, unless P=NP, as follows from [5].
3
SOLUTION METHOD
Following the approach of A.I. Serdyukov [3], let us consider a bipartite graph Ḡ = (Xn , X, Ū)
where the two subsets of vertices of bipartition Xn , X have equal sizes and the set of edges
is Ū = {{i, x} : i ∈ Xn , x ∈ X i }. Now there is a one-to-one correspondence between the set
of perfect matchings W in the graph Ḡ and the set F of feasible solutions to a problem
instance I: Given a perfect matching W ∈ W of the form {{1, x1 }, {2, x2 }, . . . , {n, xn }}, this
mapping produces the tour (x1 , x2 , . . . , xn ).
An edge {i, x} ∈ Ū is called special if {i, x} belongs to all perfect matchings in the graph Ḡ.
Let us also call the vertices of the graph Ḡ special, if they are incident with special edges.
Supposing that Ḡ is given by the lists of adjacent vertices, the special edges and edges
that do not belong to any perfect matching in the graph Ḡ may be efficiently computed by
the Algorithm 1 described below. After that all edges, except for the special edges and those
adjacent to them, are slit into cycles. Note that the method of finding all special edges and
cycles in the graph Ḡ was not discussed in [3].
Algorithm 1. Finding special edges in the graph Ḡ
Step 1 (Initialization). Assign Ḡ′ := Ḡ.
Step 2. Repeat Steps 2.1-2.2 while it is possible:
Step 2.1 (Solvability test). If the graph Ḡ′ contains a vertex of degree 0 then problem I
is infeasible, terminate.
Step 2.2 (Finding a special edge). If the graph Ḡ′ contains a vertex z of degree 1, then
store the corresponding edge {z, y} as a special edge and remove its endpoints y and z from Ḡ′ .
Each edge of the graph Ḡ is visited and deleted at most once (which takes O(1) time).
The number of edges |Ū | ≤ 2n. So the time complexity of Algorithm 1 is O(n).
Algorithm 1 identifies the case when problem I is infeasible. Further we consider only
feasible instances of 2-TSPVR and bipartite graphs corresponding to them.
After the described preprocessing the resulting graph Ḡ′ is 2-regular (the degree of each
vertex equals 2) and its components are even cycles. The cycles of the graph Ḡ′ can be
computed in O(n) time using the Depth-First Search algorithm (see e.g. [6]). Note that there
are no other edges in the perfect matchings of the graph Ḡ, except for the special edges and
edges of the cycles in Ḡ′ . In what follows, q(Ḡ) denotes the number of cycles in the graph Ḡ
(and in the corresponding graph Ḡ′ ).
Each cycle j, j = 1, . . . , q(Ḡ), contains exactly two maximal (edge disjoint) perfect
matchings, so it does not contain any special edges. Every perfect matching in Ḡ is
uniquely defined by a combination of maximal matchings chosen in each of the cycles and the
graph G = ( X8, X, U )
1
x1
2
x2
3
x3
4
x4
5
x5
6
x6
7
x7
8
x8
cycle 1
special
edges
cycle 2
Figure 1: An instance I with n = 8 and system of vertex requisitions: X 1 = {x1 , x2 },
X 2 = {x1 , x2 }, X 3 = {x3 }, X 4 = {x3 , x4 }, X 5 = {x5 , x6 }, X 6 = {x6 , x7 }, X 7 = {x7 , x8 },
X 8 = {x6 , x8 }. Here the edges drawn in bold define one maximal matching of a cycle, and the
rest of the edges in the cycle define another one. The special edges are depicted by dotted lines.
The edges depicted by dashed lines do not belong to any perfect matching. The feasible solutions of the instance are f 1 = (x1 , x2 , x3 , x4 , x5 , x6 , x7 , x8 ), f 2 = (x1 , x2 , x3 , x4 , x5 , x7 , x8 , x6 ),
f 3 = (x2 , x1 , x3 , x4 , x5 , x6 , x7 , x8 ), f 4 = (x2 , x1 , x3 , x4 , x5 , x7 , x8 , x6 ).
set of all special edges (see Fig. 1). Therefore, 2-TSPVR is solvable by the following algorithm.
Algorithm 2. Solving 2-TSPVR
Step 1. Build the bipartite graph Ḡ, identify the set of special edges and cycles and find
all maximal matchings in cycles.
Step 2. Enumerate all perfect matchings W ∈ W of Ḡ by combining the maximal matchings of cycles and joining them with special edges.
Step 3. Assign the corresponding solution f ∈ F to each W ∈ W and compute ρ(f ).
Step 4. Output the result f ∗ ∈ F , such that ρ(f ∗ ) = min ρ(f ).
f ∈F
To evaluate the Algorithm 2, first note that maximal matchings in cycles are found easily
in O(n) time. Now |F | = |W| = 2q(Ḡ) so the time complexity of Algorithm 2 of solving
2-TSPVR is O(n2q(Ḡ) ), where q(Ḡ) ≤ ⌊ n2 ⌋ and the last inequality is tight.
4
IMPROVED ALGORITHM
In [3], it was shown that almost all feasible instances of 2-TSPVR have not more than n
feasible solutions and may be solved in quadratic time. To describe this result precisely, let
us give the following
Definition 1 [3] A graph Ḡ = (Xn , X, Ū) is called “good” if it satisfies the inequality
q(Ḡ) ≤ 1.1 ln(n).
Note that any problem instance I, which corresponds to a “good” graph Ḡ, has at most
2
< n0.77 feasible solutions.
Let χ̄n denote the set of “good” bipartite graphs Ḡ = (Xn , X, Ū ) , and let χn be the set
of all bipartite graphs Ḡ = (Xn , X, Ū ). The results of A.I. Serdyukov from [3] imply
1.1 ln n
Theorem 1 |χ̄n |/|χn | −→ 1 as n → ∞.
The proof of Theorem 1 from [3] is provided in the appendix for the sake of completeness.
According to the frequently used terminology (see e.g. [7]), this theorem means that almost
all feasible instances I have at most n0.77 feasible solutions and thus they are solvable in
O(n1.77 ) time by Algorithm 2.
Using the approach
from [8] we will now modify Algorithm 2 for solving 2-TSPVR in
q(Ḡ)
O q(Ḡ)2
+ n time. Let us carry out some preliminary computations before enumerating
all possible combinations of maximal matchings in cycles in order to speed up the evaluation
of objective function. We will call a contact between cycle j and cycle j ′ 6= j (or between
cycle j and a special edge) the pair of vertices (i, i + 1) (we assume n + 1 := 1) in the left-hand
part of the graph Ḡ, such that one of the vertices belongs to the cycle j and the other one
belongs to the cycle j ′ (or the special edge). A contact inside a cycle will mean a pair of
vertices in the left-hand part of a cycle, if their indices differ exactly by one, or these vertices
are (n, 1).
Consider a cycle j. If a contact (i, i + 1) is present inside this cycle, then each of the
two maximal matchings w 0,j and w 1,j in this cycle determines the i-th arc of a tour in the
graph G. Also, if the cycle j has a contact (i, i + 1) to a special edge, each of the two maximal
matchings w 0,j and w 1,j also determines the i-th arc of a tour in the graph G. For each of the
matchings w k,j , k = 0, 1, let the sum of the weights of arcs determined by the contacts inside
the cycle j and the contacts to special edges be denoted by Pjk .
If cycle j contacts to cycle j ′ , j ′ 6= j, then each combination of the maximal matchings
of these cycles determines the i-th arc of a tour in the graph G for any contact (i, i + 1)
between the cycles. If a maximal matching is chosen in each of the cycles, one can sum up
the weights of the arcs in G determined by all contacts between cycles j and j ′ . This yields
(0,0)
(0,1)
(1,0)
(1,1)
four values which we denote by Pjj ′ , Pjj ′ , Pjj ′ and Pjj ′ , where the superscripts identify
the matchings chosen in each of the cycles j and j ′ respectively.
(0,0)
(0,1)
(1,0)
(1,1)
Parameters Pj0 , Pj1 , Pjj ′ , Pjj ′ , Pjj ′ and Pjj ′ can be found as follows. Suppose that
intermediate values of Pj0 , Pj1 for j = 1, . . . , q(Ḡ) are stored in one-dimensional arrays of
(0,0)
(0,1)
(1,0)
(1,1)
size q(Ḡ), and intermediate values of Pjj ′ , Pjj ′ , Pjj ′ and Pjj ′ for j, j ′ = 1, . . . , q(Ḡ) are
stored in two-dimensional arrays of size q(Ḡ) × q(Ḡ). Initially, all of these values are assumed
to be zero and they are computed in an iterative way by the consecutive enumeration of pairs
of vertices (i, i + 1), i = 1, . . . , n − 1, and (n, 1) in the left-hand part of the graph Ḡ. When
we consider a pair of vertices (i, i + 1) or (n, 1), at most four parameters (partial sums) are
updated depending on whether the vertices belong to different cycles or to the same cycle, or
one of the vertices is special. So the overall time complexity of the pre-processing procedure
is O(q 2 (Ḡ) + n).
Now all possible combinations of the maximal matchings in cycles may be enumerated
using a Grey code (see e.g. [9]) so that the next combination differs from the previous one by
altering a maximal matching only in one of the cycles. Let the binary vector δ = (δ1 , . . . , δq(Ḡ) )
define assignments of the maximal matchings in cycles. Namely, δj = 0, if the matching w 0,j is
chosen in the cycle j; otherwise (if the matching w 1,j is chosen in the cycle j), we have δj = 1.
This way every vector δ is bijectively mapped into a feasible solution fδ to 2-TSPVR.
In the process of enumeration, a step from the current vector δ̄ to the next vector δ changes the maximal matching in one of the cycles j. The new value of objective function ρ(fδ ) may be computed via the current value ρ(fδ̄ ) by the formula ρ(fδ ) =
P
P
(δj ,δ ′ )
(δ̄j ,δ̄ ′ )
δ̄
δ
ρ(fδ̄ ) − Pj j + Pj j −
Pjj ′ j , where A(j) is the set of cycles contacting to
Pjj ′ j +
j ′ ∈A(j)
j ′ ∈A(j)
the cycle j. Obviously, |A(j)| ≤ q(Ḡ), so updating the objective function value for the next
solution requires O(q(Ḡ)) time, and the
overall time complexity of the modified algorithm for
q(Ḡ)
solving 2-TSPVR is O q(Ḡ)2
+n .
In view of Theorem 1 we conclude that using this modification of Algorithm 2 almost all
feasible instances of 2-TSPVR are solvable in O(n0.77 ln n + n) = O(n) time.
5
LOCAL SEARCH
A local search algorithm starts from an initial feasible solution. It moves iteratively from one
solution to a better neighboring solution and terminates at a local optimum. The number of
steps of the algorithm, the time complexity of one step, and the value of the local optimum
depend essentially on the neighborhood. Note that neighborhoods, often used for the classical
TSP (e.g. k-Opt, city-swap, Lin-Kernighan [10]), will contain many infeasible neighboring
solutions if applied to 2-TSPVR because of the vertex requisition constraints.
A local search method with a specific neighborhood for 2-TSPVR may be constructed
using the relationship between the perfect matchings in the graph Ḡ and the feasible solutions.
The main idea of the algorithm consists in building a neighborhood of a feasible solution to
2-TSPVR on the basis of a Flip neighborhood of the perfect matching, represented by the
maximal matchings in cycles and the special edges.
Let the binary vector δ = (δ1 , . . . , δq(Ḡ) ) denote the assignment of the maximal matchings
to cycles as above. The set of 2q(Ḡ) vectors δ corresponds to the set of feasible solutions by a
one-to-one mapping fδ . We assume that a solution fδ′ belongs to the Exchange neighborhood
of solution fδ iff the vector δ ′ is within Hamming distance 1 from δ, i.e. δ ′ belongs to the Flip
neighborhood of vector δ.
Enumeration of the Exchange neighborhood takes O(q 2 (Ḡ)) time if the preprocessing described in Section 4 is carried out before the start of the local search (without the preprocessing
it takes O(nq(Ḡ)) operations). Therefore, for almost all feasible instances I, the Exchange
neighborhood may be enumerated in O(ln2 (n)) time.
6
MIXED INTEGER LINEAR PROGRAMMING
MODEL
The one-to-one mapping between the maximal matchings in cycles of the graph Ḡ and feasible solutions to 2-TSPVR may be also exploited in formulation of a mixed integer linear
programming model.
Recall that Pj0 (Pj1 ) is the sum of weights of all arcs of the graph G determined by the
contacts inside the cycle j and the contacts of the cycle j with special edges, when the maximal
(k,l)
matching w 0,j (w 1,j ) is chosen in the cycle j, j = 1, . . . , q(Ḡ). Furthermore, Pjj ′ is the sum
of weights of arcs in the graph G determined by the contacts between cycles j and j ′ , if the
′
maximal matchings w k,j and w l,j are chosen in the cycles j and j ′ respectively, k, l = 0, 1,
j = 1, . . . , q(Ḡ) − 1, j ′ = j + 1, . . . , q(Ḡ). These values are computable in O(n + q 2 (Ḡ)) time
as shown in Section 4.
Let us introduce the following Boolean variables:
0, if matching w 0,j is chosen in the cycle j,
dj =
1, if matching w 1,j is chosen in the cycle j,
j = 1, . . . , q(Ḡ).
The objective function combines the pre-computed arc weights for all cycles, depending
on the choice of matchings in d = (d1 , . . . , dq(Ḡ) ):
q(Ḡ)−1 q(Ḡ)
X (0,0)
(0,1)
Pjj ′ (1 − dj )(1 − dj ′ ) + Pjj ′ (1 − dj )dj ′
X
j=1
j ′ =j+1
q(Ḡ)−1 q(Ḡ)
+
X
X
(1,0)
Pjj ′ dj (1
− dj ′ ) +
(1,1)
Pjj ′ dj dj ′
j ′ =j+1
j=1
(1)
q(Ḡ)
+
X
j=1
Pj0 (1 − dj ) + Pj1 dj → min,
dj ∈ {0, 1}, j = 1, . . . , q(Ḡ).
(2)
Let us define supplementary real variables in order to remove non-linearity of the objective
(k)
function: for k ∈ {0, 1} we assume that pj ≥ 0 is an upper bound on the sum of weights of
arcs in the graph Ḡ determined by the contacts of the cycle j, if matching w k,j is chosen in
this cycle, i.e. dj = k, j = 1, . . . , q(Ḡ) − 1.
Then the mixed integer linear programming model has the following form:
q(Ḡ)−1
X
q(Ḡ)
(0)
pj
+
(1)
pj
j=1
+
X
j=1
Pj0 (1 − dj ) + Pj1 dj → min,
q(Ḡ)
q(Ḡ)
(0)
pj
≥
X
j ′ =j+1
(0,0)
Pjj ′
(1 − dj − dj ′ ) +
X
j ′ =j+1
(0,1)
Pjj ′ (dj ′ − dj ) ,
(3)
j = 1, . . . , q(Ḡ) − 1,
q(Ḡ)
q(Ḡ)
(1)
pj
≥
X
(4)
(1,0)
Pjj ′
j ′ =j+1
(dj − d ) +
j′
X
(1,1)
Pjj ′ (dj + dj ′ − 1) ,
j ′ =j+1
j = 1, . . . , q(Ḡ) − 1,
(5)
pj ≥ 0, k = 0, 1, j = 1, . . . , q(Ḡ) − 1,
(6)
dj ∈ {0, 1}, j = 1, . . . , q(Ḡ).
(7)
(k)
Note that if matching w 0,j is chosen for the cycle j in an optimal solution of problem (3)(0)
(1)
(7), then inequality (4) holds for pj as equality and pj = 0. Analogously, if matching w 1,j
(1)
(0)
is chosen for the cycle j, then inequality (5) holds for pj as equality and pj = 0. Therefore,
problems (1)-(2) and (3)-(7) are equivalent because a feasible solution of one problem corresponds to a feasible solution of another problem, and an optimal solution corresponds to an
optimal solution.
The number of real variables in model (3)-(7) is (2q(Ḡ) − 2), the number of Boolean
variables is q(Ḡ). The number of constraints is O(q(Ḡ)), where q(Ḡ) ≤ ⌊ n2 ⌋. The proposed
model may be used for computing lower bound of objective function or in branch-and-bound
algorithms, even if the graph Ḡ is not “good”.
Note that there are a number of integer linear programming models in the literature
on the classical TSP, involving O(n2 ) Boolean variables. Model (3)–(7) for 2-TSPVR has
at most ⌊ n2 ⌋ Boolean variables and for almost all feasible instances the number of Boolean
variables is O(ln(n)).
7
CONCLUSION
We presented an algorithm for solving 2-TSP with Vertex Requisitions, that reduces
the time complexity bound formulated in [3]. It is easy to see that the same approach is
applicable to the problem 2-Hamiltonian Path of Minimum Weight with Vertex
Requisitions, that asks for a Hamiltonian Path of Minimum Weight in the graph G, assuming the same system of vertex requisitions as in 2-TSP with Vertex Requisitions.
Using the connection to perfect matchings in a supplementary bipartite graph and some
preprocessing we constructed a MIP model with O(n) binary variables and a new efficiently
searchable Exchange neighborhood for problem under consideration.
Further research might address the existence of approximation algorithms with constant
approximation ratio for 2-TSP with Vertex Requisitions.
References
[1] Garey, M.R., Johnson, D.S.: Computers and intractability. A guide to the theory of NPcompleteness. W.H. Freeman and Company, San Francisco, CA (1979)
[2] Serdyukov, A.I.: Complexity of solving the travelling salesman problem with requisitions
on graphs with small degree of vertices. Upravlaemye systemi. 26, 73–82 (1985) (In Russian)
[3] Serdyukov, A.I.: On travelling salesman problem with prohibitions. Upravlaemye systemi.
17, 80–86 (1978) (In Russian)
[4] Garey, M.R., Johnson, D.S.: Strong NP-completeness results: Motivation, examples, and
implications. Journal of the ACM. 25, 499–508 (1978)
[5] Serdyukov A.I.: On finding Hamilton cycle (circuit) problem with prohibitions. Upravlaemye systemi. 19, 57–64 (1979) (In Russian)
[6] Cormen, T.H., Leiserson, C.E., Rivest, R.L., Stein, C.: Introduction to Algorithms, 2nd
edition. MIT Press (2001)
[7] Chvatal, V.: Probabilistic methods in graph theory. Annals of Operations Research. 1,
171–182 (1984)
[8] Eremeev, A., Kovalenko, J.: Optimal recombination in genetic algorithms for combinatorial optimization problems: Part II. Yugoslav Journal of Operations Research. 24 (2),
165-186 (2014)
[9] Reingold, E.M., Nievergelt, J., Deo, N.: Combinatorial algorithms: Theory and Practice.
Englewood Cliffs, Prentice-Hall (1977)
[10] Kochetov, Yu. A.: Computational bounds for local search in combinatorial optimization.
Computational Mathematics and Mathematical Physics. 48 (5), 747-763 (2008)
[11] Feller, W.: An Introduction to Probability Theory and Its Applications. Vol. 1. John
Wiley & Sons, New York, NY (1968)
[12] Riordan, J.: An Introduction to Combinatorial Analysis. John Wiley & Sons, New York,
NY (1958)
APPENDIX
Note that A.I. Serdyukov in [3] used the term block instead of term cycle, employed in Section 3
of the present paper. A block was defined in [3] as a maximal (by inclusion) 2-connected
subgraph of graph Ḡ with at least two edges. However, in each block of the graph Ḡ, the
degree of every vertex equals 2 (otherwise F = Ø because the vertices of degree 1 do not
belong to blocks and the vertex degrees are at most 2 in the right-hand part of Ḡ). So, the
notions block and cycle are equivalent in the case of considered bipartite graph Ḡ. We use the
term block in the proof of Theorem 1 below, as in the original paper [3], in order to avoid a
confusion of cycles in Ḡ with cycles in permutations of the set {1, . . . , n}.
Theorem 1 [3] |χ̄n |/|χn | −→ 1 as n → ∞.
Proof. Let Sn be the set of all permutations of the set {1, . . . , n}. Consider a random
permutation s from Sn . By ξ(s) denote the number of cycles in permutation s. It is known
n
P
1
and the
(see e.g. [11]) that the expectation E[ξ(s)] of random variable ξ(s) is equal to
i
i=1
variance V ar[ξ(s)] equals
n
P
i=1
i−1
.
i2
Let S̄n denote the set of permutations from Sn , where the
number of cycles is at most 1.1 ln(l). Then, using Chebychev’s inequality [11], we get
|S̄n |/|Sn | −→ 1 as n → ∞.
(8)
Now let Sn′ denote the set of permutations from Sn , which do not contain the cycles of
(i)
length 1, and let Sn be the set of permutations from Sn , which contain a cycle with element
i, i = 1, . . . , n. Using the principle of inclusion and exclusion [12], we obtain
|Sn \
Sn′ |
=
[
Sn(i)
1≤i≤n
X
Sn(i)
1≤i6=j6=k≤n
\
Sn(j)
\
=
n
X
Sn(i) −
i=1
X
1≤i6=j≤n
Sn(i)
\
Sn(j) +
Sn(k) − . . . = n! − Cn2 (n − 2)! + Cn3 (n − 3)! − . . . ≤
n! n!
2
2
+
= n! = |Sn |.
2
6
3
3
Therefore,
1
|Sn′ | ≥ |Sn |.
3
(9)
Combining (8) and (9), we get
|S̄n′ |/|Sn′ |
3|Sn′ \S̄n′ |
3|Sn \S¯n |
|Sn′ \S̄n′ |
≥
1
−
≥
1
−
−→ 1,
=1−
|Sn′ |
|Sn |
|Sn | n→+∞
(10)
where S̄n′ = Sn′ ∩ S¯n .
The values |χ̄n | and |χn \χ̄n | may be bounded, using the following approach. We assign
any permutation s ∈ Sl′ , l ≤ n, a set of bipartite graphs χn (s) ⊂ χn as follows. First of
all let us assign an arbitrary set of n − l edges to be special. Then the non-special vertices {i1 , i2 , . . . , il } ⊂ Xn of the left-hand part, where ij < ij+1 , j = 1, . . . , l − 1, are now
partitioned into ξ(s) blocks, where ξ(s) is the number of cycles in permutation s. Every
cycle (t1 , t2 , . . . , tr ) in permutation s corresponds to some sequence of vertices with indices
{it1 , it2 , . . . , itr } belonging to the block associated with this cycle. Finally, it is ensured that
for each pair of vertices {itj , itj+1 }, j = 1, . . . , r − 1, as well as for the pair {itr , it1 }, there
exists a vertex in the right-hand part X which is adjacent to both vertices of the pair. Except
for special edges and blocks additional edges are allowed in graphs from class χn (s). These
edges are adjacent to the special vertices of the left-hand part such that the degree of any
vertex of the left-hand part is not greater than two. Moreover, additional edges should not
lead to creating new blocks.
There are n! ways to associate vertices of the left-hand part to vertices of the righthand part, therefore the number of different graphs from class χn (s), s ∈ Sl′ , l ≤ n, is
|χn (s)| = Cnl 2ξn!
h(n, l), where function h(n, l) depends only on n and l, and ξ1 (s) is the
1 (s)
number of cycles of length two in permutation s. Division by 2ξ1 (s) is here due to the fact
that for each block that corresponds to a cycle of length two in s, there are two equivalent
ways to number the vertices in its right-hand part.
Let s = c1 c2 . . . cξ(s) be a permutation from set Sl′ , represented by cycles ci , i = 1, . . . , ξ(s),
and let cj be an arbitrary cycle of permutation s of length at least three, j = 1, . . . , ξ(s).
Permutation s may be transformed into permutation s1 ,
s1 = c1 c2 . . . cj−1 c−1
j cj+1 . . . cξ(s),
(11)
by reversing the cycle cj . Clearly, permutation s1 induces the same subset of bipartite graphs
in class χn as the permutation s does. Thus any two permutations s1 and s2 from set Sl′ ,
l ≤ n, induce the same subset of graphs in χn , if one of these permutations may be obtained
from the other one by several transformations of the form (11). Otherwise the two induced
subsets of graphs do not intersect. Besides that χn (s1 ) ∩ χn (s2 ) = ∅ if s1 ∈ Sl′1 , s2 ∈ Sl′2 ,
l1 6= l2 .
On one hand, if s ∈ S̄l′ , l ≤ n, then χn (s) ⊆ χ̄n . On the other hand, if s ∈ S̃l′ := Sl′ \S̄l′ ,
l < n, then either χn (s) ⊆ χ̄n or, alternatively, χn (s) ⊆ χn \χ̄n may hold. Therefore,
|χ̄n | ≥
n X
X
l=2 s∈S̄l′
≥
n
X
|S̄l′ |
·
l=2
|χn \χ̄n | ≤
n!
Cnl ξ1 (s) ξ(s)−ξ1 (s) h(n, l)
2
2
n!
Cnl 1.1ln(l) h(n, l)
2
n
X
l=⌊1.1ln(n)⌋
Now assuming ψ(n) =
X
s∈S̃l′
max
l=⌊1.1ln(n)⌋,...,n
n
X
≥
n!
Cnl ξ(s)
2
=
n X
X
Cnl
l=2 s∈S̄l′
|S̄l′ | · Cnl
l=⌊1.1ln(n)⌋
≤
n
X
n!
h(n, l) ≥
2ξ(s)
n!
21.1ln(l)
|S̃l′ | · Cnl
l=⌊1.1ln(n)⌋
(12)
h(n, l),
n!
21.1ln(l)
h(n, l).
(13)
|S̃l′ |/|S̄l′ | and taking into account (12), (13) and (10),
we obtain
|χn \χ̄n |
≤ ψ(n) → 0 as n → +∞.
|χ̄n |
Finally, the statement of the theorem follows from (14). Q.E.D.
(14)
| 8cs.DS
|
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
1
Distributed area coverage control with imprecise
robot localization: Simulation and experimental
studies
arXiv:1612.04704v1 [cs.SY] 14 Dec 2016
Sotiris Papatheodorou, Anthony Tzes Senior Member, IEEE, Konstantinos Giannousakis and Yiannis Stergiopoulos
Abstract—This article examines the area coverage problem for
a network of mobile robots with imprecise agents localization.
Each robot has uniform radial sensing ability, governed by first
order kinodynamics. The convex-space is partitioned based on
the Guaranteed Voronoi (GV) principle and each robot’s area
of responsibility corresponds to its GV-cell, bounded by hyperbolic arcs. The proposed control law is distributed, demanding
the positioning information about its GV-Delaunay neighbors.
Simulation and experimental studies are offered to highlight the
efficiency of the proposed control law.
Index Terms—Multi-agent systems, Multi-robot systems, Mobile robots, Cooperative systems, Autonomous agents.
I. I NTRODUCTION
REA coverage of a planar region by a swarm of robotic
agents is an active field of research with great interest and
potential. The task involves the dispersal inside a region of interest of autonomous, sensor–equipped robotic agents in order
to achieve sensor coverage of that region. Distributed control
algorithms are a popular choice for the task because of their
computational efficiency compared to centralized controllers,
their ability to adapt in real–time to changes in the robot
swarm and their reliance only on local information. Thus they
can be implemented on robots with low computational power
and low power radio transceivers, since each agent is required
to communicate only with its neighbors. The downside of
distributed control schemes is that they lead to local optima of
their objective function because of their lack of information
on the state of all agents.
Area coverage can be either static [1]–[3] in which case the
agents converge to some static final configuration or sweep
[4]–[6] in which the agents constantly move in order to satisfy
a constantly changing coverage objective.
Other factors that need to be taken into account are the
type of the region surveyed [7], [8], the sensing pattern and
performance of the sensors [9]–[12], or the particular dynamics
of the agents [13].
There have been varied approaches to area coverage such
as distributed optimization [14], [15], model predictive control
[16], [17], game theory [18] and optimal control [19].
A
All localization systems have an inherent uncertainty in
their measurements, thus there has been significant work on
taking this uncertainty into account either by data fusion [20],
safe trajectory planning [21] or spatial probability [22]. Our
approach creates a suitable space partioning scheme in lieu of
the agents’ positioning uncertainty and a distributed gradient–
based control law.
This article examines the problem of persistent area coverage of a convex planar region by a network of homogeneous
agents equipped with isotropic sensors. In order to take into
account the localization uncertainty, each agent is assumed
to lie somewhere inside a circular positioning uncertainty
region and Guaranteed Voronoi partitioning [23] of the region of interest is constructed from these uncertain regions.
Subsequently, a distributed control law is derived based on
the aforementioned partitioning so that a coverage objective
increases monotonously. Because of the complexity of that
law, a suboptimal one is proposed and both laws are compared
through simulations. Additionally, the suboptimal control law
is evaluated by two experiments using differential drive robots.
This article is an extension of [24] which did not contain
simulation results from the optimal control law and lacked
the experimental evaluation of the suboptimal control law.
The layout of the article is as follows. The required mathematical preliminaries are presented in Section II. The Voronoi
and the Guaranteed Voronoi space partitions are examined
in Section III. Section IV contains the definition of the
coverage objective and the derivation of the gradient ascent
based control law that maximizes it. Additionally it contains
the proposed suboptimal control law. Simulation studies are
presented in Section V in order to compare and evaluate both
control laws. Section VI contains experimental results from the
implementation of the suboptimal control law and is followed
by concluding remarks.
II. M ATHEMATICAL P RELIMINARIES
Let us assume a compact convex region Ω ∈ R2 under
surveillance and a network of n identical dimensionless mobile
agents (nodes). Each agent’s dynamics are governed by
q̇i = ui , ui ∈ R2 , qi ∈ Ω, i ∈ In
This work has received funding from the European Union’s Horizon 2020
Research and Innovation Programme under the Grant Agreement No.644128,
AEROWORKS.
The authors are with the Electrical & Computer Engineering Department,
University of Patras, Rio, Achaia 26500, Greece. Corresponding author’s
email: [email protected]
(1)
where ui is the control input for each agent and In =
{1, 2, . . . , n}.
The agents’ positions are not known precisely and thus in
the general case each agent has a positioning uncertainty of
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
2
different measure. We assume that an upper bound riu for the
positioning uncertainty of each agent is known and thus its
center lies within a disk
Ciu (qi , riu ) = {q ∈ Ω : k q − qi k≤ riu } , i ∈ In
(2)
where k · k is the Euclidean metric and qi the agent’s position
as reported by its localization equipment.
All agents are equipped with identical omnidirectional sensors with circular sensing footprints and as such the sensed
region of each agent can be defined as
When constructing the Voronoi cell of a point qi , only its
Delaunay neighbors need to be considered, thus simplifying
the definition of its Voronoi cell into
Vi = q ∈ Ω : kq − qi k ≤ q − q j , ∀ j ∈ Ni , i ∈ In .
Using this property, a Voronoi diagram can be constructed
by first finding the Delaunay triangulation of the input points
and then for each cell Vi by finding the intersection of the
halfplanes defined by point qi and all points in Ni .
Cis (qi , rs ) = {q ∈ Ω : k q − qi k≤ rs } , i ∈ In
where rs is the common range of the sensors.
We also define for each agent a ‘Guaranteed Sensed Region’
(GSR) as the region that is sensed by the agent for all of its
possible positions inside Ciu . The GSRs are defined as
(
)
Cigs (Ciu ,Cis ) =
\
Cis (qi , rs ), ∀qi ∈ Ciu , i ∈ In
qi
Ciu
Cis
and since
and
are disks, the previous expression can be
further simplified into
Cigs (qi , rs − riu ) = {q ∈ Ω :k q − qi k≤ rs − riu } , i ∈ In ,
Fig. 1. Voronoi diagram (left) for 6 dimensionless nodes, with its equivalent
Guaranteed Voronoi diagram (right) in the case of disks (centered on those
points).
(3)
provided that rs ≥ riu .
When an agent’s position is known precisely, i.e. riu = 0,
its GSR is equivalent to its sensed region, whereas when an
agent’s positioning uncertainty is too large compared to its
/
sensing capabilities, i.e. riu > rs , then Cigs = 0.
III. S PACE PARTITIONING
Most distributed area coverage schemes assign an area
of responsibility to each agent based on information from
its neighboring agents. Then each agent is responsible for
maximizing the coverage solely on its own responsibility
region. The most common partitioning scheme is the Voronoi
partitioning, described in Section III-A. However in this article, because of the localization uncertainty of the agents, the
guaranteed Voronoi diagram described in Section III-B is used
for the assignment of areas of responsibility.
A. Voronoi Diagram
The Voronoi diagram for a set of points
Q = {q1 , q2 . . . qn }, qi ∈ Ω is defined as follows
Vi = q ∈ Ω : kq − qi k ≤ q − q j , ∀ j ∈ In , j 6= i , i ∈ In .
Each Vi ⊂ Ω is called the Voronoi cell of node i. A Voronoi
diagram of 6 points constrained in a subset Ω of R2 is shown
in Figure 1 (left), where the cell boundaries are shown in blue.
It is important to note that the Voronoi diagram is a complete
S
tessellation of Ω, that is i∈In Vi = Ω.
Another useful property of Voronoi diagrams is their duality to Delaunay triangulations. Let us define the Delaunay
neighbors of a point qi as
Ni = j ∈ In , j 6= i : Vi ∩V j 6= 0/ , i ∈ In ,
(4)
B. Guaranteed Voronoi Diagram
The Guaranteed Voronoi (GV) diagram is defined for a
set of uncertain regions D = {D1 , D2 , . . . , Dn }, Di ⊂ Ω. Each
uncertain region Di contains all the possible positions of a
point qi ∈ R2 whose localization cannot be measured precisely.
Thus the GV diagram is defined as
Vig
= {q ∈ Ω : max kq − qi k ≤ min q − q j
∀ j ∈ In , j 6= i, qi ∈ Di , q j ∈ D j , i ∈ In .
(5)
An equivalent definition of the GV diagram is as follows
Vig =
\
Hi j
i∈In
where
Hi j = q ∈ Ω : max kq − qi k ≤ min q − q j , ∀qi ∈ Di , ∀q j ∈ D j .
This definition is analogous to the construction of the classic
Voronoi cell by the halfplane intersection method described
previously, with the difference that the regions being intersected are no longer halfplanes.
As such, each cell Vig contains the points on the plane that
are closer to the corresponding point qi of region Di for all
possible configurations of the uncertain points. A Guaranteed
Voronoi diagram of 6 regions constrained in a subset Ω of
R2 is shown in Figure 1 (right), where similarly the cell
boundaries are shown in blue.
Remark 1. In contrast to the classic Voronoi diagram, the
GuaranteedSVoronoi diagram does not constitute a tessellation
of Ω,Ssince i∈In Vig ⊂ Ω. We define a neutral region O so that
O ∪ i∈In Vig = Ω.
Similarly to the Voronoi diagram, we can define the Guaranteed Delaunay neighbors Nig of a region Di so that in order
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
3
to construct Vig , only the regions in Nig need be considered.
Thus the Guaranteed Delaunay neighbors are defined as
(6)
Nig = j ∈ In , j 6= i : ∂ Hi j ∩ ∂Vig 6= 0/ .
Having defined the Guaranteed Delaunay neighbors, the
definition of the GV diagram can be simplified into
Vig =
[
Hi j .
(7)
(a)
(b)
(c)
(d)
i∈Ni
It is important to note that when two or more uncertain
regions overlap, none of them are assigned GV cells. In
addition, if the uncertain regions degenerate into points, the
GV diagram converges to the classic Voronoi diagram.
C. Guaranteed Voronoi Diagram of Disks
The GV diagram is examined in the case when the uncertain
regions Di are disks Ciu as defined in Equation (2). In that case
it has been shown that Hi j are regions on the plane bounded
by hyperbolic branches. For any two nodes i and j, ∂ Hi j and
∂ H ji are the two branches of a hyperbola with foci at qi and
q j and semi-major axis equal to the sum of the radii of Ciu
and Cuj . The parametric equation of these hyperbola branches
can be found in the Appendix.
The dependence of the GV cells of two disks on the distance
of their centers di j = d(qi , q j ) is as follows. When the disks
Ciu ,Cuj , overlap both of the cells Vig ,V jg are empty as shown
in Fig. 2 (a). If the disks represent the possible positions of
the agents, this is an undesirable configuration as it can lead
to collisions. When the disks Ciu ,Cuj are outside tangent (i.e.,
d(qi , q j ) = riu + ruj ), the resulting cells are rays starting from
the centers of the disks qi , q j and extending along the direction
of qi q j as seen in Fig 2 (b). When the disks are disjoint, the
GV cells are bounded by the two branches of a hyperbola. As
di j increases further, the eccentricity of the hyperbola increases
and the distance of the disk centers from the hyperbola vertices
(the points of the hyperbola closest to its center) increases as
seen in Fig 2 (c), (d). The distance between the hyperbola’s
vertices remains constant at ri + r j as di j increases.
The GV cells of two disks also depend on the sum of
their radii ri + r j as seen in Figure 3. As ri + r j decreases,
the eccentricity of the hyperbola increases and the result on
the cells is the same as if the distance of the disks’ centers
increased, as explained previously. When ri + r j = 0, the GV
cells are the classic Voronoi cells.
It has been also shown that the Guaranteed Delaunay
neighbors Nig are a subset of the classic Delaunay neighbors
of the uncertain disks’ centers. The two sets Ni and Nig are
equal when the GV diagram is constructed for the whole plane,
however when the GV diagram is constrained within a subset
Ω of the plane, Nig can be a subset of Ni . Such a case is shown
in Figure 1 where nodes 1 and 5 are Delaunay neighbors
but not Guaranteed Delaunay neighbors. This way, the GV
diagram of disks can be constructed in a manner similar to
the classic Voronoi diagram. First the Delaunay triangulation
of the disks’ centers is constructed and then each cell Vig is
constructed by intersecting the regions bound by hyperbolic
arcs Hi j ∀ j ∈ Nig . This process is shown in Figure 4.
Fig. 2. Dependence of the GV cells on qi − q j .
Fig. 3. Dependence of the GV cells on ri + r j .
It is useful to define Guaranteed-Sensing Voronoi (GSV)
cells as
Vigs = Vig ∩Cigs , i ∈ In ,
(8)
which are the parts of the GV cells that are guaranteed to be
sensed by their respective nodes. It can be shown that since
GV cells are disjoint, GSV cells are also disjoint regions.
IV. P ROBLEM S TATEMENT - D ISTRIBUTED C ONTROL L AW
By taking into account the GV partitioning of the space, the
following area coverage criterion is constructed
H =
Z
Z
S
φ (q) dq =
gs
i∈In Vi
∑
gs
i∈In Vi
φ (q) dq =
∑ Hi ,
(9)
i∈In
Fig. 4. The construction of a GV cell by the hyperbolic region intersection
method.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
4
where the function φ : R2 → R+ is related to the a priori
knowledge of importance of a point q ∈ Ω indicating the
probability of an event to take place at q in a coverage
scenario. This criterion expresses the area that is guaranteed
to be closest to and covered by the nodes.
A distributed control law is designed in order to maximize
the coverage criterion (9) while taking into account the nodes’
guaranteed sensing pattern (3) and the GV partitioning (7). It
is shown that this control law leads to a monotonic increase
of the coverage criterion.
The boundary ∂ Ṽigs can be decomposed in disjoint sets as
∂Vigs = ∂Vigs ∩ ∂ Ω ∪ ∂Vigs ∩ ∂Cigs ∪ ∂Vigs ∩ ∂ O . (14)
These sets represent the parts of ∂Vigs that lie on the boundary
of Ω, the boundary of the node’s guaranteed sensing region,
and the boundary of the unassigned neutral region of Ω,
respectively. This decomposition is shown in Figure 5 with
∂Vigs ∩ ∂ Ω in green, ∂Vigs ∩ ∂Cigs in solid red and ∂Vigs ∩ ∂ O
in solid blue.
Theorem 1. For a network of mobile nodes with uncertain
localization as in (2), uniform range–limited radial sensing
performance as in (3), governed by the individual agent’s
kinodynamics described in (1), and the GV–partitioning of Ω
defined in (7), the coordination scheme
Z
ui = αi
"Z
αi
∑g
j∈Ni
gs
gs
gs
∂Vi ∩∂Ci
∂Vi ∩∂ Hi j
υii ni φ
ni φ dq +
#
Z
dq +
gs
∂V j ∩∂ H ji
υ ij n j φ
dq
(10)
where ni is the outward unit normal on ∂Vigs and αi a positive
constant, maximizes the performance criterion (9) along the
nodes’ trajectories in a monotonic manner, leading in a locally
area–optimal configuration of the network.
Fig. 5. The decomposition of ∂Vigs into disjoint sets (solid red, green and
blue).
Proof. In order to guarantee monotonous increase of the
coverage criterion (9), we evaluate its time derivative
Hence
∂ Hi
∂ Hi
∂H
=∑
q̇i = ∑
ui
∂t
∂
q
i
i∈In
i∈In ∂ qi
∂H
∂ qi
(11)
Z
gs
Vi
∂
∑g ∂ qi
j∈Ni
Z
gs
φ dq.
Vj
Considering the second summation term, infinitesimal motion
of qi may only affect ∂V jgs at ∂V jgs ∩ ∂ O, where O is the
neutral region defined in Remark 1, since both hyperbola
branches are affected by alteration of one of the foci. In
addition, only the Guaranteed Delaunay neighbors (6) of i are
considered in the summation, as a major property of GV–
partitioning. Therefore, the former expression can be written
via the generalized Leibniz integral rule [25] (by converting
surface integrals to line ones) as
∂H
=
∂ qi
Z
υ i n φ dq +
gs i i
∂Vi
Z
∑g
j∈Ni
gs
∂Vi ∩∂ O
(12)
monotonous increase of H is achieved.
We now evaluate the partial derivative of H with respect
to qi as
φ dq +
gs
gs
∂Vi ∩∂ Ω
Z
∂ Hi
ui =
, i ∈ In ,
∂ qi
Z
can be written as
=
By using the gradient–based control law
∂
∂H
=
∂ qi
∂ qi
∂H
∂ qi
∂V j ∩∂ O
υ ij n j φ dq,
where υii , υ ij stand for the transpose Jacobian matrices with
respect to qi of the points q ∈ ∂Vigs , q ∈ ∂V jgs , respectively,
i.e.
T
4 ∂q
, q ∈ ∂V jgs , i, j ∈ In .
(13)
υ ij (q) =
∂ qi
υii ni φ dq +
υii ni φ dq +
Z
gs
gs
∂Vi ∩∂Ci
Z
∑g
j∈Ni
gs
∂V j ∩∂ O
υii ni φ dq +
υ ij n j φ dq.
(15)
It is apparent that υii = 0 at q ∈ ∂Vigs ∩ Ω since all q ∈ ∂ Ω
remain unaltered by infinitesimal motions of qi , assuming no
alteration of the environment. Considering the second integral,
for any point q ∈ ∂ Ṽigs ∩ ∂Cigs it holds that υii (q) = I2 , where
I stands for the identity matrix, since they translate along the
direction of motion of qi at the same rate. As far as the sets
∂Vigs ∩ ∂ O and ∂V jgs ∩ ∂ O, j ∈ Nig are concerned, they can be
merged in pairs via utilization of the left and right hyperbolic
branches, as introduced in GV–partitioning (7), as follows
Z
gs
∂Vi ∩∂ O
υii ni φ
"Z
∑g
j∈Ni
gs
∂Vi ∩∂ Hi j
Z
dq +
∑g
gs
∂V j ∩∂ O
j∈Ni
υii ni φ
#
Z
dq +
υ ij n j φ dq =
gs
∂V j ∩∂ H ji
υ ij n j φ
dq
.
(16)
However, for any two Delaunay neighbors i, j it can be
observed that
• The hyperbolic branches ∂ Hi j and ∂ H ji are symmetric
with respect to the perpendicular bisector of qi , q j , and
are governed by the same set of parametric equations (left
and right branch).
• The vectors n j are mirrored images of the corresponding
ni with respect to the perpendicular bisector of qi , q j .
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
∂H
∂ qi
Taking into account the above,
∂H
∂ qi
"Z
∑g
j∈Ni
5
can be written as
A. Constraining Nodes Inside Ω
Z
=
gs
gs
gs
∂Vi ∩∂Ci
∂Vi ∩∂ Hi j
υii ni φ
ni φ dq +
#
Z
dq +
gs
∂V j ∩∂ H ji
υ ij n j φ
dq
.
(17)
The unit normal vectors ni , n j and Jacobian matrices υii , υ ij
in the second part of (17) can be evaluated via utilization
of the parametric representations of the sets over which the
integration takes place. These sets are parts of the left (∂ Hi j )
and right (∂ H ji ) branch of the hyperbola that assigns the space
among two arbitrary nodes i, j. The second term
of the sum
S
in (17) requires knowledge of all the nodes in j∈N g N gj ⊆ In
i
and so the control law is distributed.
The decomposition of the set ∂Vigs ∩ ∂ O of node i into
mutually disjoint hyperbolic arcs ∂Vigs ∩ ∂ Hi j and ∂Vigs ∩ ∂ Hik
is shown in Figure 6. Figure 6 also shows the other domains of
integration used in (17) ∂Vigs ∩ ∂Cigs , ∂V jgs ∩ ∂ H ji and ∂Vkgs ∩
∂ Hki .
When using this control law, it is possible that a portion
of some node’s positioning uncertainty region lies outside the
region Ω, i.e. that Ciu ∩ Ω 6= Ciu for some node i. As a result,
it would be possible for that node to be outside the region Ω
given that it may be anywhere within Ciu .
In order to prevent this situation, instead of Ω, a subset
Ωsi ⊆ Ω will be used, in the general case different for each
node since their uncertainty radii are allowed to differ. This
subset is calculated as the Minkowski difference of Ω with the
disk C (riu ) = {q ∈ Ω : k q k≤ riu }.
Ωsi = {q ∈ Ω | q +C(riu ) ⊆ Ω} , i ∈ In .
Thus by constraining the center of each uncertainty disk
qi inside Ωsi , it is guaranteed that no part of the uncertainty
disk will ever be outside Ω. This is achieved by projecting
the velocity control input ui on ∂ Ωsi . This is only needed if
qi ∈ ∂ Ωsi and the contorl input ui points towards the outside
of Ωsi . As a result, the control law becomes
(
proj∂ Ωs ui ifqi ∈ ∂ Ωsi ∧ qi + ε ui ∈
/ Ωsi
i
ũi =
(18)
ui
otherwise
where ε is an infinitesimally small positive constant.
B. Suboptimal Control Law
Because of the high complexity of the integrals over Hi j and
H ji of the control law (10), the large amount of information
required in order to implement it, as described in Remark 2, as
well as the problems described in Section IV-A, a simplified
version of the control law is proposed.
Fig. 6. The domains of integration of the control law (10).
The proposed law (10) leads to a gradient flow of H along
the nodes trajectories, while H increases monotonically, since
dH
∂H
=∑
· q̇i = ∑ kui k2 ≥ 0.
dt
i∈In ∂ qi
i∈In
Conjecture 1. The computationally efficient suboptimal control law
Z
ui = αi
n φ dq
(19)
gs
gs i
∂Vi ∩∂Ci
leads the nodes to suboptimal trajectories.
The control law enhancement described in Section IV-A is
used in conjunction with this simplified control law.
V. S IMULATION S TUDIES
Remark 2. It is not immediately clear from the control law
(10) the information node i requires in order to implement
it. The integrals over ∂Vigs ∩ ∂Cigs and ∂Vigs ∩ ∂ Hi j require
knowledge of the Guaranteed Delaunay neighbors Nig in order
to compute Vigs and Hi j . However the integral over ∂V jgs ∩
∂ H ji requires knowledge of the whole cell V jgs of node i and
as such requires informations form the Guaranteed Delaunay
neighbors N gj of node j. Thus node i requires information from
all nodes in
[ g
Ñig =
Nj ,
Simulations have been conducted in order to evaluate the
efficiency of the control law as well as compare the complete
(10) with the simplified (19) law. All nodes have common
positioning uncertainty regions Ci , ∀qi ∈ Ω as well as sensing
performance Cigs . The a priori importance of points inside the
region was φ (q) = 1, ∀q ∈ Ω. In this case, the maximum
possible value for H is
H max = ∑ A Cigs ,
i∈In
g
j∈Ni
which are the Guaranteed Delaunay neighbors of all the
Guaranteed Delaunay neighbors of node i.
where A (·) is the area function, given that n guaranteed
sensing disks can be packed inside Ω. In all simulations, the
control law extension as described in Section IV-A is used.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
6
A. Case Study I
In this case study the convergence speed between the
optimal and suboptimal laws is examined. A network of three
nodes in a large region is simulated. Figure 7 shows the initial
and final states of the network, with the uncertainty disks
shown in black, the guaranteed sensing disks in red and the
GV cells in blue. In Figure 8 the trajectories of the nodes
when using the optimal control law are shown in red and
when using the suboptimal control law in blue, while the initial
positions are represented by circles and the final by squares.
It is observed that the final configurations in both cases are
very similar since the average distance between nodes’ final
states is 0.57% of the region’s diameter. This is due to the low
number of nodes in the network and is a result that can not
be generalized, as seen in the following case study. Figure 9
shows the evolution of the coverage objective H with time,
again shown in red for the optimal control law and in blue for
the suboptimal. As it is expected, the optimal law converges
faster than the suboptimal. The final value of the objective
function is the same for both control laws in this case since
the network reaches a globally optimal state in both cases.
Fig. 8. The trajectories of the nodes using the optimal (red) and suboptimal
(blue) control laws.
(a)
Fig. 9. The coverage objective H using the optimal (red) and suboptimal
(blue) control laws.
(b)
(c)
Fig. 7. Case Study I: (a) Initial state, (b) Final state using the optimal control
law, (c) Final state using the suboptimal control law.
B. Case Study II
In this case study network of ten nodes is examined. Figure
10 shows the initial and final states of the network, with the
uncertainty disks shown in black, the guaranteed sensing disks
in red and the GV cells in blue. In Figure 11 the trajectories
of the nodes when using the optimal control law are shown in
red and when using the suboptimal control law in blue, while
the initial positions are represented by circles and the final
by squares. It is observed that the final configurations in both
cases seem very similar at first glance, however the average
distance between nodes’ final states is 10.12% of the region’s
diameter. Careful examination of the trajectories shows that
not all nodes starting from the same initial position converge to
the same final position, explaining the large average distance.
. Figure 12 shows the evolution of the coverage objective H
with time, again shown in red for the optimal control law and
in blue for the suboptimal. As it is expected, the optimal law
converges faster than the suboptimal. The final value of the
objective function is the same for both control laws in this case
because of the similarity of their final configurations. However,
since both control laws converge to local maxima, there are
no guarantees that one or the other will always achieve better
performance.
This case study highlights the need for the control enhancement in Section IV-A, especially in the case of the
optimal control law. It can be seen in Figure 11 that a node
moving under the optimal law (red trajectory) was stopped
from moving outside the region on the lower left corner.
VI. E XPERIMENTAL S TUDIES
In order to evaluate the simplified control law (19), two
experiments were planned and conducted. The experiments
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
7
(a)
Fig. 12. The coverage objective H using the optimal (red) and suboptimal
(blue) control laws.
(b)
(c)
Fig. 10. Case Study I: (a) Initial state, (b) Final state using the optimal control
law, (c) Final state using the suboptimal control law.
pose estimation server. ArUco [26] is a minimal library for
Augmented Reality applications and provides pose estimation
of some predefined sets of fiducial markers. Hende, by attaching a fiducial marker on top of each robot, the ArUco library
provides an estimation of each robot’s pose.
In order to estimate the positioning uncertainty of the
experimental setup, a fiducial marker was placed on the
centroid and on each of the vertices of the region Ω. A
3D triangular mesh was created from these measurements as
shown in Figure 13. The maximum uncertainty value from that
mesh rumax = 0.032 m was used for the experiments.
Fig. 11. The trajectories of the nodes using the optimal (red) and suboptimal
(blue) control laws.
consist of 3 robots, an external pose tracking system and
a computer, communicating with the robots and the pose
tracking system via a WiFi router.
The robots used are the differential-drive AmigoBots by
ActiveMedia Robotics, which provide a high level command
set, such as setting their translational and rotational velocities
directly. These robots are also equipped with encoders, used
for self-estimating their pose, but due to the drifting error,
an external tracking system was implemented. Although the
external pose estimation system is used for the control law
implementation, its performance is compared with that of
the encoders. Since the robots are not equipped with actual
sensors, circular sensing patters with a radius rs = 0.3 m were
assumed in the control law implementation.
The pose estimation system consists of a camera and an
ODROID-XU4 octa core computer, running an ArUco based
Fig. 13. Triangular mesh from positioning uncertainty measurements.
The control law was implemented algorithmically as a loop
with a period Ts = 0.1 sec. After the end of each iteration the
computer sends velocity commands to the robots. At first, the
robots’ current positions are used to create the Guaranteed
Voronoi diagram of the uncertain disks and calculate the
control inputs according to the simplified control law (19).
Subsequently, a target point qti ∈ R2 is created for each robot
based on its current position qi and its control input ũi . Once
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
8
the target point for each robot has been found, and given each
robot’s position qi and orientation θi , the translational vi and
rotational ωi velocity control inputs are sent to each robot.
t
kqi − qi k
vi = min
, vmax cos(dθi )
Ts
|dθi |
ω = min
, ωmax sin(dθi )
Ts
where dθi = ∠(qti − qi ) − θi and vmax , ωmax are constraints on
the maximum translational and rotational velocity respectively
of the Amigobot robots. Once all robots are within a predefined distance dt = 0.02 m of their respective target points, the
Guaranteed Voronoi diagram and the control law are calculated
again and new target points are produced.
The region shown in all figures regarding the experiments is
Ωs ⊆ Ω i.e. the region inside which the centers of the uncertain
disks are constrained, as explained in Section IV-A.
Both experiments are evaluated against simulations with
the same initial robot positions, positioning uncertainty and
sensing performance.
A. Experiment I
The initial and final positions of the robots for both the
experiment and the simulation are shown in Figures 14 and
15 respectively with the uncertainty disks shown in black, the
guaranteed sensing disks in red and the GV cells in blue.
Additionally, photos of the experiment initial and final configurations can be seen on Figure 16. It can be seen that in the
final configuration all guaranteed sensing disks are contained
within their respective GV cells in both the experiment and
the simulation, thus leading the network to a globally optimal
configuration. By comparing the final configurations, as well
as the robot trajectories for the experiment and simulation
shown in Figure 17, it can be seen that they differ significantly.
The mean distance between the nodes final positions in the
experiment and simulation is 8.68% of the diameter of Ωs .
Because in this current case there are multiple globally optimal
configurations the network can possibly reach and since the
implemented control law differs from the theoretical one, it
is natural that the final configurations between the experiment
and the simulation differ. Nevertheless, a globally optimal configuration was reached in both cases. The coverage objective
H did not increase monotonously due to the implemented
control law. It did increase however from 83.7% to 99.9% of
its maximum possible value. The positioning data from the
ArUco library and the robot encoders are compared in Figure
19 [Left]. As it is expected, the further a robot moves and the
more it rotates, the larger the positioning error of the encoders
grows. The mean error between the ArUco and encoder final
robot positions is 4.04% of the diameter of Ωs . Figure 19
[Right] shows the trajectories of the target points used in the
control law implementation.
Fig. 14. Experiment I: Initial configuration.
Fig. 15. Experiment I: Experiment [Left] and simulation [Right] final
configuration.
Ciu . The guaranteed sensing disks however were calculated
using only the actual uncertainty and not the one increased
by the robot radius. This was done to guarantee that the
robots would remain inside the region Ω. The initial and
final positions of the robots for both the experiment and the
simulation are shown in Figures 20 and 21 respectively with
the uncertainty disks shown in black, the guaranteed sensing
disks in red and the GV cells in blue. Additionally, photos
of the experiment initial and final configurations can be seen
on Figure 22. It can be seen that in the final configuration
all guaranteed sensing disks are almost completely contained
within their respective GV cells in both the experiment and the
Fig. 16. Experiment I: Photos from the initial [Left] and final [Right] robot
positions in the experiment.
B. Experiment II
In this experiment, half the diameter of the robots was
added to their positioning uncertainty, so that each robot was
completely contained within its positioning uncertainty disk
Fig. 17. Experiment I: Comparison of experiment (red) and simulation (green)
robot trajectories. Initial positions marked with circles and final positions with
squares.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
9
Fig. 20. Experiment II: Initial configuration.
Fig. 18. Experiment I: Coverage objective H as a function of time.
Fig. 21. Experiment II: Experiment [Left] and simulation [Right] final
configuration.
Fig. 19. Experiment I: [Left] ArUco (green) and encoder (red) reported
trajectories. [Right] Target point trajectories. Initial positions marked with
circles and final positions with squares.
simulation. By comparing the final configurations, as well as
the robot trajectories for the experiment and simulation shown
in Figure 23, it can be seen that they do not differ significantly.
The mean distance between the nodes final positions in the
experiment and simulation is 1.54% of the diameter of Ωs .
Because of the increased positioning uncertainty in this case,
which results in smaller GV cells, the number of local maxima
of the objective function is significantly smaller than in the
previous experiment. Thus it is quite possible that, despite their
differences, the theoretical and implemented control laws converge to the same final configuration. The coverage objective
H did not increase monotonously due to the implemented
control law. It did increase however from 38.8% to 98.3% of
its maximum possible value. The positioning data from the
ArUco library and the robot encoders are compared in Figure
25 [Left]. As it is expected, the further a robot moves and the
more it rotates, the larger the positioning error of the encoders
grows. The mean error between the ArUco and encoder final
robot positions is 1.54% of the diameter of Ωs . Figure 25
[Right] shows the trajectories of the target points used in the
control law implementation.
A video of the simulations and the experiments can be
found on http://anemos.ece.upatras.gr/images/stories/videos/
PTGS IEEETAC.mp4
VII. C ONCLUSION
This article examines the area coverage problem by a homogeneous team of mobile agents with imprecise localization.
A gradient ascent based control law was designed based on
a Guaranteed Voronoi partition of the region. Because of the
complexity of the resulting control law, a simpler suboptimal
control law is proposed and the performance of both is
compared in simulation studies. Additionally, two experiments
were conducted to highlight the efficiency of the suboptimal
control law.
A PPENDIX
We assume n disks Di , i ∈ In with centers qi = [xi , yi ]T ∈ R2
and radii riu , i ∈ In .
Fig. 22. Experiment II: Photos from the initial [Left] and final [Right] robot
positions in the experiment.
Fig. 23. Experiment II: Comparison of experiment (red) and simulation
(green) robot trajectories. Initial positions marked with circles and final
positions with squares.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
10
Fig. 24. Experiment II: Coverage objective H as a function of time.
Fig. 26. A hyperbola with its parameters 2a and 2c shown.
B. Hyperbola Parametric Equation
Assuming the foci of the hyperbola are qi = [−c 0]T and
q j = [c 0]T , its parametric equation is
±a cosh(t)
γ(t) =
, t ∈R
b sinh(t)
Fig. 25. Experiment II: [Left] ArUco (green) and encoder (red) reported
trajectories. [Right] Target point trajectories. Initial positions marked with
circles and final positions with squares.
A. Hyperbola Definition and Properties
We will use the GV diagram definition given in (5) for two
disks Di and D j and shown that the boundaries of their GV
cells are hyperbola branches.
Let us find the boundary of the cell of node i, Vig = Hi j ,
since we examine only two nodes. We have that
∂ Hi j = q ∈ Ω : max kq − bi k = min q − b j , ∀bi ∈ Di , ∀b j ∈ D j ,
however since Di and D j are disks, the maximum and minimum distances can be calculated and thus
∂ Hi j = q ∈ Ω : kq − qi k + riu = q − q j − riu ,
where qi and q j are the centers of Di and D j respectively. The
equation
∂ Hi j = q ∈ Ω : q − q j − kq − qi k = ri + r j ,
defines one branch of a hyperbola since it is the locus of points
whose difference of distances from two foci qi and q j is equal
to a positive constant ri +r j . The other branch of the hyperbola
is
∂ H ji = q ∈ Ω : kq − qi k − q − q j = ri + r j ,
and corresponds to the boundary of V jg .
In a hyperbola, the positive constant is called the major axis
and is usually denoted as 2a whereas the distance between
its foci is denoted as 2c. A third parameter, the minor axis
denoted as 2b is then defined by the equation a2 + b2 = c2 . As
such in this case we have that 2a = ri + r j and 2c =k qi − q j k.
Additionally, the closest points on the two branches are called
the vertices of the hyperbola and their distance is 2a.
where a, b and c are those defined in Appendix A and the
+(−) sign corresponds to the West (East) branch of the
hyperbola which is H ji (Hi j ).
By rotating the hyperbola foci around the origin by an angle
θ , the parametric equation becomes
cos(θ ) − sin(θ )
±a cosh(t)
γ(t) =
, t ∈R
sin(θ ) cos(θ )
b sinh(t)
Finally, given two foci qi = [xi yi ]T and q j = [x j y j ]T
anywhere on the plane, the general hyperbola parametric
equation is
" xi +x j #
cos(θ ) − sin(θ )
±a cosh(t)
2
, t ∈R
γ(t) =
+ yi +y
j
sin(θ ) cos(θ )
b sinh(t)
2
(20)
y −y
where θ = arctan( x jj −xii ).
C. Outward Unit Normal Vectors
For any curve γ(t) the normal vector is defined as
γ̇
γ̇
n̂(t) = γ̈ − γ̈ ·
.
k γ̇ k k γ̇ k
The magnitude of this vector is the curvature at that particular
point on the curve and it points towards the center of curvature.
Since the hyperbolic branches define convex regions on the
plane, the outward unit normal vector is given by
n(t) = −
n̂
.
k n̂ k
(21)
However, since the resulting expressions are too large to show
here, an indicative plot of the outward unit normal vectors of
one branch of a hyperbola is shown in Figure 27.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
11
R EFERENCES
Fig. 27. The outward unit normal vectors of one branch of the hyperbola are
shown in black.
D. Jacobian Matrix
T
i
The Jacobian
matrix υ j shows the change in the curve
γ jx (t)
γ j (t) =
caused by the movement of node i, thus its
γ jy (t)
transpose is
" ∂γ
#
∂γ
υ ij =
jx
jx
∂ xi
∂ γ jy
∂ xi
∂ yi
∂ γ jy
∂ yi
(22)
The expressions for the Jacobian matrix elements are too large
to show here and as such some indicative plots of them with
respect to the parameter t are shown in Figure 28.
(a)
(b)
(c)
(d)
Fig. 28. The Jacobian matrix
t. (a)
∂ γ jx
∂ xi ,
(b)
∂ γ jx
∂ yi ,
(c)
∂ γ jy
∂ xi ,
υii
(d)
(blue) and
∂ γ jy
∂ yi .
υ ij
(red) elements as functions of
[1] H. Mahboubi and A. G. Aghdam, “Distributed deployment algorithms
for coverage improvement in a network of wireless mobile sensors:
Relocation by virtual force,” IEEE Transactions on Control of Network
Systems, vol. 61, no. 99, p. 1, 2016.
[2] C. F. van Eeden, G. Lee, S. Katsuno, and H. Yonekura, “Minimised
moving distance algorithm for geometric deployment of robot swarms,”
in Proc. 55th Annual Conf. of the Society of Instrument and Control
Engineers of Japan (SICE), Sep. 2016, pp. 1195–1200.
[3] A. Simonetto, T. Keviczky, and D. V. Dimarogonas, “Distributed solution
for a maximum variance unfolding problem with sensor and robotic
network applications,” in Proc. and Computing (Allerton) 2012 50th
Annual Allerton Conf. Communication, Control, Oct. 2012, pp. 63–70.
[4] C. Song, L. Liu, G. Feng, Y. Wang, and Q. Gao, “Persistent awareness
coverage control for mobile sensor networks,” Automatica, vol. 49, no. 6,
pp. 1867–1873, 2013.
[5] C. Franco, D. M. Stipanović, G. López-Nicolás, C. Sagüés, and
S. Llorente, “Persistent coverage control for a team of agents with
collision avoidance,” European Journal of Control, vol. 22, pp. 30–45,
2015.
[6] C. Zhai and Y. Hong, “Decentralized sweep coverage algorithm for
multi-agent systems with workload uncertainties,” Automatica, vol. 49,
no. 7, pp. 2154–2159, 2013.
[7] Y. Stergiopoulos, M. Thanou, and A. Tzes, “Distributed collaborative
coverage-control schemes for non-convex domains,” IEEE Transactions
on Automatic Control, vol. 60, no. 9, pp. 2422–2427, 2015.
[8] R. J. Alitappeh, K. Jeddisaravi, and F. G. Guimarães, “Multi-objective
multi-robot deployment in a dynamic environment,” Soft Computing, pp.
1–17, 2016.
[9] Y. Stergiopoulos and A. Tzes, “Convex Voronoi-inspired space partitioning for heterogeneous networks: A coverage-oriented approach,” IET
Control Theory and Applications, vol. 4, no. 12, pp. 2802–2812, 2010.
[10] O. Arslan and D. E. Koditschek, “Voronoi-based coverage control of heterogeneous disk-shaped robots,” in 2016 IEEE International Conference
on Robotics and Automation (ICRA). IEEE, 2016, pp. 4259–4266.
[11] L. Pimenta, V. Kumar, R. Mesquita, and G. Pereira, “Sensing and
coverage for a network of heterogeneous robots,” in Proc. of the 47th
Conference on Decision and Control, Cancun, Mexico, 2008, pp. 3947–
3952.
[12] Y. Stergiopoulos and A. Tzes, “Cooperative positioning/orientation
control of mobile heterogeneous anisotropic sensor networks for area
coverage,” in Proc. IEEE International Conference on Robotics &
Automation, Hong Kong, China, 2014, pp. 1106–1111.
[13] J. M. Luna, R. Fierro, C. Abdallah, and J. Wood, “An adaptive coverage
control algorithm for deployment of nonholonomic mobile sensors,” in
Proc. 49th IEEE Conf. Decision and Control (CDC), Dec. 2010, pp.
1250–1256.
[14] J. Cortés and F. Bullo, “Coordination and geometric optimization via
distributed dynamical systems,” SIAM Journal on Control and Optimization, vol. 44, no. 5, pp. 1543–1574, 2005.
[15] J. Cortés, S. Martinez, and F. Bullo, “Spatially-distributed coverage optimization and control with limited-range interactions,” ESAIM: Control,
Optimisation and Calculus of Variations, vol. 11, no. 4, pp. 691–719,
2005.
[16] M. T. Nguyen and C. S. Maniu, “Voronoi based decentralized coverage
problem: From optimal control to model predictive control,” in Proc.
24th Mediterranean Conf. Control and Automation (MED), Jun. 2016,
pp. 1307–1312.
[17] F. Mohseni, A. Doustmohammadi, and M. B. Menhaj, “Distributed
receding horizon coverage control for multiple mobile robots,” IEEE
Systems Journal, vol. 10, no. 1, pp. 198–207, Mar. 2016.
[18] V. Ramaswamy and J. R. Marden, “A sensor coverage game with improved efficiency guarantees,” in Proc. American Control Conf. (ACC),
Jul. 2016, pp. 6399–6404.
[19] M. T. Nguyen, L. Rodrigues, C. S. Maniu, and S. Olaru, “Discretized
optimal control approach for dynamic multi-agent decentralized coverage,” in Proc. IEEE Int. Symp. Intelligent Control (ISIC), Sep. 2016, pp.
1–6.
[20] J. A. Hage, M. E. E. Najjar, and D. Pomorski, “Fault tolerant collaborative localization for multi-robot system,” in Proc. 24th Mediterranean
Conf. Control and Automation (MED), Jun. 2016, pp. 907–913.
[21] B. Davis, I. Karamouzas, and S. J. Guy, “C-opt: Coverage-aware trajectory optimization under uncertainty,” IEEE Robotics and Automation
Letters, vol. 1, no. 2, pp. 1020–1027, Jul. 2016.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
[22] J. Habibi, H. Mahboubi, and A. G. Aghdam, “Distributed coverage
control of mobile sensor networks subject to measurement error,” IEEE
Transactions on Automatic Control, vol. 61, no. 11, pp. 3330–3343, Nov.
2016.
[23] W. Evans and J. Sember, “Guaranteed Voronoi diagrams of uncertain
sites,” in 20th Canadian Conference on Computational Geometry, 2008,
pp. 207–210.
[24] S. Papatheodorou, Y. Stergiopoulos, and A. Tzes, “Distributed area
coverage control with imprecise robot localization,” in Proc. 24th
Mediterranean Conf. Control and Automation (MED), Jun. 2016, pp.
214–219.
[25] H. Flanders, “Differentiation under the integral sign,” American Mathematical Monthly, vol. 80, no. 6, pp. 615–627, 1973.
[26] S. Garrido-Jurado, R. M. noz Salinas, F. Madrid-Cuevas, and M. Marı́nJiménez, “Automatic generation and detection of highly reliable fiducial
markers under occlusion,” Pattern Recognition, vol. 47, no. 6, pp. 2280
– 2292, 2014.
12
| 3cs.SY
|
1
Subspace Identification of Large-Scale 1D
Homogeneous Networks
arXiv:1702.03539v1 [cs.SY] 12 Feb 2017
Chengpu Yu, Michel Verhaegen, and Anders Hansson
Abstract—This paper considers the identification of large-scale
1D networks consisting of identical LTI dynamical systems. A
new subspace identification method is developed that only uses
local input-output information and does not rely on knowledge
about the local state interaction. The identification of the local
system matrices (up to a similarity transformation) is done
via a low dimensional subspace retrieval step that enables
the estimation of the Markov parameters of a locally lifted
system. Using the estimated Markov parameters, the state-space
realization of a single subsystem in the network is determined.
The low dimensional subspace retrieval step exploits various key
structural properties that are present in the data equation such
as a low rank property and a two-layer Toeplitz structure in the
data matrices constructed from products of the system matrices.
For the estimation of the system matrices of a single subsystem,
it is formulated as a structured low-rank matrix factorization
problem. The effectiveness of the proposed identification method
is demonstrated by a simulation example.
Index Terms—Large-scale 1D distributed systems, nuclearnorm optimization, two-layer Toeplitz structure.
I. I NTRODUCTION
In recent years, the fundamental system theoretical research
on large-scale interconnected systems has been intensified both
on the topic of distributed control, see e.g. [1]–[4], as well
as on the topic of distributed system identification, see e.g.
[5]–[7]. These fundamental developments are inspired by the
increasing interest from the application side, such as in fluid
mechanics [8], flexible structures [9], large adaptive telescope
mirrors [10], wind turbine farms [11], and so on. In this paper,
we focus on the identification of large-scale 1D homogeneous
networked systems. Such systems may characterize a platoon
of cars or may result by discretizing dynamical systems
described via partial differential equations [3], [12], [13].
When considering the subspace identification of local system dynamics using local data only, there is not only the
missing information about the local system state, but also the
missing state information from the neighboring systems. This
makes the local identification problem more difficult. In [14] a
subspace identification (SID) scheme for large-scale circulant
systems was developed. The specific property of circularity
C. Yu and M. Verhaegen are with the Delft Center for Systems and
Control, Delft University, Delft 2628CD, Netherlands ([email protected],
[email protected])
A. Hansson is with the Division of Automatic Control, Department of Electrical Engineering, Linkoping University, Sweden ([email protected])
Part of the research of M. Verhaegen was done while he was a visiting
professor of the Division of Automatic Control, Department of Electrical
Engineering, Linkoping University, Sweden.
The research leading to these results has received funding from the
European Research Council under the European Union’s Seventh Framework
Programme (FP7/2007-2013) / ERC grant agreement no. 339681.
was exploited to decompose the large model description into
simple subsystems through a global transformation of the
input and output data. In [15] a network of identical systems
interacting in a known interconnection pattern, indicated as
so-called decomposable systems, was considered. The identification of the dynamics of a single subsystem in the network
using transformed input-output data sequences gives rise to
a challenging Bilinear Matrix Inequality (BMI). In [12] the
property that the inverse of the observability Grammian is offdiagonally decaying was used to approximate the neighboring
states, influencing the local system dynamics to be identified,
via an (unknown) linear combination of locally neighboring
inputs and outputs. The selection of these neighboring input
and output quantities requires however an exhaustive search,
which is quite computationally demanding. As a complementary work to [12], a nuclear norm identification solution
was provided in [16] to separate the local dynamics and
global dynamics by exploiting their distinct rank and order
properties; however, it did not consider identification of the
interconnections between the subsystems.
In this paper, we consider the subspace identification of
a local cluster of identical LTI systems connected in a 1D
network. Due to the unknown neighboring-state information,
the considered local identification is a blind identification
problem, where some input sequences are totally unknown. In
this scenario, the existing SID approaches that aim at retrieving
a matrix whose row or column space is of interest, the latter,
e.g., being the extended observability matrix, cannot work. As
a result, we aim to develop a new identification method for
identifying the system matrices of a single subsystem (up to a
similarity transformation) in the network that uses local inputoutput data only.
The subspace identification method presented in this paper
has two major features. First, a novel identification method
is developed that uses similar data equations as many existing
SID variants, but aims for a low dimensional subspace retrieval
that allows the identification of the finite number of Markov
parameters present in this data equation instead. This is a new
alternative to classical SID methods [17] that would fail in
retrieving subspace like the column space of the extended
observability matrix of the system to be identified. This novel
SID approach is applied to the identification of a locally lifted
system with the missing neighboring-state information. By
fully exploiting the two-layer structure of the block Toeplitz
matrices in the data equation of a lifted state-space system,
a low-rank optimization problem is formulated for which the
optimal solution can yield (parts of) the true Markov parameters of the locally lifted system. Second is the retrieval of
2
ݑଵ ሺ݇ሻ
ȭଵ
ݕଵ ሺ݇ሻ
ݔିோିଵ ሺ݇ሻ
ݑିோ ሺ݇ሻ
ݑାோ ሺ݇ሻ
ݑ ሺ݇ሻ
ȭାோ
ȭ
ȭିோ
ݕିோ ሺ݇ሻ
ݔାோାଵ ሺ݇ሻ
ݕାோ ሺ݇ሻ
ݕ ሺ݇ሻ
ݑே ሺ݇ሻ
ȭே
ݕே ሺ݇ሻ
ȭ୧ Ǥ
Fig. 1. Illustration of a cluster of subsystems in a neighborhood of the
subsystem Σi with radius R. The states xi−R−1 (k) and xi+R+1 (k) are
explicitly indicated. They are like all other states unmeasurable.
the system matrices, describing the local LTI dynamics, of an
individual subsystem and its interaction with its neighboring
systems (up to a similarity transformation) from the reliably
estimated Markov parameters. The retrieval of the system
matrices of an individual subsystem is inherently a challenging
structured state-space realization problem [18] for which the
optimal solution can yield the estimates of system matrices
up to a similarity transformation. The second feature also
includes the exploitation of the shifting structure of a timevarying generalized observability matrix.
The rest of the paper is organized as follows. Section II
describes the concerned identification problems and shows the
challenge of dealing with the identification of a small cluster
of subsystems in a large-scale networked system. Section III
shows that the existing subspace identification methods break
down in dealing with the concerned identification problem.
Section IV presents a method for identifying the Markov
parameters of the locally lifted state-space models. Section V
provides a solution to the state-space realization of a single
subsystem. Section VI provides a numerical example. The
conclusions are provided in Section VII.
The following notations are adopted throughout the paper.
The lowercase (uppercase) x (X) is used to denote a vector
(matrix). The superscripts T and −1 are transpose and inverse
operators, respectively. vec (X) represents the vector stacked
by columns of X. tr (X) stands for the trace of X. In denotes
an n × n identity matrix and Om,n is an m × n zero matrix. ⊗
stands for the Kronecker product. kxk2 represents the norm of
x. kXkF and kXk∗ denote the Frobenius norm and nuclear
norm of the matrix X, respectively.
II. P RELIMINARIES
AND
P ROBLEM D EFINITION
{Σi }N
i=1
We consider the LTI systems
connected in a
homogeneous 1D network as shown in Fig. 1. The lifted statespace model of the concerned network is written as:
Σ1 :
Σi :
x1 (k + 1)
y1 (k)
xi (k + 1)
=
=
=
yi (k)
=
i = 2, · · · , N − 1,
ΣN : xN (k + 1) =
yN (k)
=
Ax1 (k) + Ar x2 (k) + Bu1 (k)
Cx1 (k) + e1 (k)
Axi (k) + Al xi−1 (k) + Ar xi+1 (k)
+Bui (k)
Cxi (k) + ei (k)
AxN (k) + Al xN −1 (k) + BuN (k)
CxN (k) + eN (k)
(1)
where xi (k) ∈ Rn×1 , ui (k) ∈ Rm×1 , yi (k) ∈ Rp×1 and
ei (k) ∈ Rp×1 are the state, input, output and measurement
noise of the i-th subsystem, xi−1 (k) and xi+1 (k) are the
neighboring states of the i-th subsystem. For xi (k), the
subscript i denotes the spatial index and k is referred to as the
time index. For a large-scale distributed system, we always
assume that N ≫ n and n > max{p, m}.
By lifting all states xi (k) into the vector x(k) as x(k) =
T
T
x1 (k) · · · xTN (k) and doing the same for the inputs,
outputs and noises defining resp. the vectors u(k), y(k) and
e(k), the global system in Fig. 1 has the following state space
model:
x(k + 1) = Ax(k) + Bu(k)
(2)
y(k) = Cx(k) + e(k),
where A, B, C are N × N
following forms:
A
Al
A=
B=
block matrices which have the
A
..
.
B
B
..
.
B
Ar
..
.
..
.
Al
, C =
,
Ar
A
C
C
..
.
C
.
To identify the system model in (2), the existing SID methods only estimate the triplet (A, B, C) up to a similarity transformation, thereby not preserving the special block-diagonal
and block-tridiagonal structure of these system matrices. In addition, the computational complexity of SID methods, which is
at least O(N 3 ) with N being the number of subsystems in the
network, may easily disqualify their use for the identification
of large-scale networks.
To deal with the high computational complexity for identifying the global system model in (2), we consider the
identification of a cluster of subsystems {Σj }i+R
j=i−R in a
neighborhood of Σi with radius R satisfying R < i < N − R
and R ≪ N , as shown in Fig. 1. The lifted state-space model
of this cluster is represented as
xi (k + 1) = AR xi (k) + B R ui (k) + DR v i (k),
y i (k) = C R xi (k) + ei (k),
(3)
with xi (k), ui (k), y i (k) the subparts of x(k), u(k), y(k) respectively from the block-rows (i−R) to (i+R); AR , B R , C R
are (2R + 1) × (2R + 1) block matrices which have forms
similar to A, B, C in (2), respectively; the (2R + 1) × 2 block
matrix DR and the vector v i (k) are defined as:
Al 0
0
0
xi−R−1 (k)
..
.
.. , v i (k) =
DR = .
.
xi+R+1 (k)
0
0
0 Ar
For the identification of the cluster model in (3), the
following two problems will be addressed.
3
Problem 1: The Markov parameters of the locally lifted
state-space model in (3) will be identified using only the local
input ui (k) and the local output yi (k).
This problem will be addressed in Section IV. The difficulty
of this problem w.r.t. classical subspace identification methods
in [19], [20] or their more recent variants [21], [22] is that
the adjacent state sequences xi−R−1 (k) and xi+R+1 (k) are
missing or the input sequence v i (k) of the considered cluster
model is unknown.
Problem 2: The order the local subsystem Σi , denoted
by n, as well as the local system matrices A, Ar , Al , B, C
are to be identified based on the identified Markov parameters in Problem 1. More specifically, the system matrices are to be identified up to a similarity transformation,
i.e., the estimates Â, Âr , Âl , B̂, Ĉ of A, Ar , Al , B, C satisfy
 = Q−1 AQ, Âl = Q−1 Al Q, Âr = Q−1 Ar Q, B̂ =
Q−1 B, Ĉ = CQ with Q ∈ Rn×n being a non-singular
ambiguity matrix.
The second problem will be addressed in Section V. It is
noteworthy that, when the system order n is known, the second problem is inherently a structured state-space realization
problem as studied in [18], which turns out to be a challenging
non-convex optimization problem.
The developed method in this paper can be applied to more
general classes of large-scale identification problems than just
a 1D chain of identical LTI systems, such as 1D heterogeneous
networks consisting of clusters of identical LTI dynamical
systems [23].
In addressing the locally lifted identification problem in (3),
we stipulate the following assumption.
Assumption A.1 The global system (A, B, C) and the locally
lifted system (AR , B R , C R ) are assumed to be minimal.
The persistent excitation of the input signal u(k), which
will be used for the identifiability analysis in the sequel, is
defined below.
Definition 1. A time sequence u(k) ∈ RN m is persistently
exciting of order s if there exists an integer h such that the
(block-) Hankel matrix
u(k)
u(k + 1) · · ·
u(k + h − 1)
u(k + 1)
u(k + 2) · · ·
u(k + h)
,
.
..
..
.
.
.
.
.
.
.
u(k + s − 1) u(k + s) · · · u(k + s + h − 2)
has full row rank for any positive integer k.
III. B REAK
DOWN OF EXISTING SUBSPACE
IDENTIFICATION METHODS
The data equation of the local state-space model (3) is given
as follows:
i
i
i
i
+ Es,h
,
+ TsDR Vs,h
Ys,h
= Os xih + TsBR Us,h
i
In this equation, the (block-) Hankel matrix Ys,h
y i (1) · · ·
yi (h)
..
..
i
.
Ys,h = .
..
.
y i (s) · · · y i (h + s − 1)
(4)
is defined as
with the superscript i being the spatial index of the subsystem
Σi , the subscripts s, h respectively being the number of block
i
,
rows and the number of block columns. Analogous to Ys,h
i
i
i
we define the block-Hankel matrices Us,h , Vs,h , Es,h from the
sequences ui (k), v i (k), ei (k), respectively. The matrix TsBR is
a block Toeplitz matrix defined from the triplet (AR , B R , C R )
as
0
C RBR
0
TsBR =
,
..
.
.
.
.
.
.
.
s−2
BR · · · C RBR 0
C R AR
and TsDR is defined in a similar way from the triplet
(AR , DR , C R ). The final matrix definitions in (4) are
CR
C R AR
Os =
..
.
s−1
C R AR
and
xih = [xi (k) · · · xi (k + h − 1)] .
Existing subspace identification (SID) methods, see e.g.
[17], break down when retrieving a matrix with a subspace
i
i
like Os or xih from the available data matrices Ys,h
and Us,h
in (4).
To explain this deficiency of existing SID methods, we
consider the simple noise-free case e(k) ≡ 0. Let
−1
i,T
i,T
i
i
Π⊥
Us,h
,
U = I − Us,h Us,h Us,h
i
Then the subspace revealing matrix [17] is Ys,h
Π⊥
U , or
h
iT
i,T
i,T
if we consider the approach in [22].
Ys,h Us,h
When the unknown system input v i (k) in (3) is absent,
i
≡ 0), the column space of
namely v i (k) ≡ 0 (or Vh,s
i
Os can be retrieved from that of Ys,h
Π⊥
U , while the row
i
space of xh can be retrieved from that of compound matrix
iT
h
i,T
i,T
. However, for the case v i (k) 6= 0, the
Ys,h
Us,h
i
presence of the bilinear term TsDR Vs,h
destroys the important
subspace revealing property in existing SID methods.
Next, we shall show that the identification method in [21]
cannot handle the concerned identification problem. One could
consider the unknown input v i (k) in (3) as a ”missing input
variable”. However, for the simple case e(k) ≡ 0 and using
the notation Hv to denote the set block Hankel matrices of
i
the same structure as the matrix Vs,h
, the rank minimization
problem of [21] can be written as
i
Ys,h
i
.
min rank Us,h
(5)
V ∈Hv
V
By the following inequality
i
i
Ys,h
i
≥ rank Ys,h
rank Us,h
,
i
Us,h
V
4
we can see that V = 0 is an optimal solution to (5). It implies
that the whole missing input sequence, not just a few missing
entries, cannot be recovered using the low-rank optimization
of [21].
To resolve this deficiency of SID methods, when considering
the identification of local clusters in a large homogeneous
network, a novel SID approach is developed in this paper.
This method differs from the existing SID methods in two
major ways.
First, the low-rank subspace retrieval step that is characteristic for many SID methods is unable to find an accurate (or
consistent) estimate of the key subspace of Os or xih . In this
paper, the low-rank optimization will be used for the accurate
estimation of (parts of) the Markov parameters in the block
Toeplitz matrix TsBR .
Second, contrary to existing SID methods that are unable
to preserve the structures in the estimated system matrices,
the proposed subspace identification method is capable of
accurately estimating the non-zero block entries (up to a
similarity transformation) of the structured system matrices
in (3), as stated in Problem 2.
IV. I DENTIFYING
THE M ARKOV PARAMETERS OF THE
LOCALLY LIFTED STATE SPACE MODEL (3)
In this section, the Markov parameters of the state-space
model in (3) will be estimated by exploiting the low-rank
property of the unmeasurable-state related terms in (4) and
the specific block Toeplitz structure of TsBR . Subsection IV-A
formulates a low rank optimization problem without considering the specific structure of TsBR . It is shown in Lemma 2 that
this low-rank optimization is unable to recover the true Markov
parameters. Subsection IV-B reformulates another low-rank
optimization problem by incorporating the specific Toeplitz
structure of TsBR . It is shown in Theorem 1 that, under some
mild conditions, the optimal solution to the proposed lowrank optimization problem can yield (parts of) the true Markov
parameters.
A. Formulation of a low-rank optimization problem
Due to the unmeasurable state sequences in a network, the
i
in (4) is unknown. However,
matrix sum Os xih + TsDR Vs,h
this matrix sum has a low-rank property that will be exploited
as a solution in the new subspace identification method.
i
Lemma 1. For the data equation (4), when h > ps or Ys,h
is
i
DR i
a fat matrix, the sum Os xh + Ts Vs,h satisfies the following
rank property
i
rank Os xih + TsDR Vs,h
≤ (2R + 1)n
(6)
+ min{(s − 1)sp, 2(s − 1)n},
where (s − 1)sp and 2(s − 1)n denote the number of non-zero
rows and columns of TsDR , respectively.
Proof: From the structures of Os and TsDR , we can
get that
rank Os xih ≤ rank (Os ) ≤ (2R + 1)n and
i
≤ rank TsDR ≤ min{(s−1)sp, 2(s−1)n}.
rank TsDR Vs,h
Thus, the result in the lemma is straightforward.
From Lemma 1, we can derive a condition to select the
parameter s in the data equation (4) and the cluster radius R,
i
defined above equation (3), such that the sum Os xih +TsDR Vs,h
is of low rank (or rank deficient). This condition reads:
(2R + 1)sp > (2R + 1)n + min{(s − 1)sp, 2(s − 1)n}. (7)
The above condition means that the number of the rows of the
i
is larger than an upper bound of its
matrix Os xih + TsDR Vs,h
rank. In practice, by fixing a value of s satisfying that s > np ,
we can always find a value of R such that the above inequality
holds. Therefore, in the sequel, we assume that the matrix sum
i
Os xih + TsDR Vs,h
is of low-rank (or rank deficient).
Denote the noise-free output as ŷi (k) = y i (k)−ei (k) and its
i
related block Hankel matrix Ŷs,h
. Based on the rank property
discussed above, a low-rank regularized optimization problem
is then proposed as follows:
B
min
i ∈H
Θs R ∈T ,Ŷs,h
h+s−1
X
kŷi (t) − y i (t)k2
t=1
(8)
i
h
i
i
R
+ λ · rank Ŷs,h
− ΘB
s Us,h ,
where T and H denote respectively the set of block Toeplitz
and block Hankel matrices with appropriate block sizes, and
the regularization parameter λ allows to make a trade-off
between the two terms in the cost function. It is remarked
that the block Toeplitz structure of T corresponds to the first
block-Toeplitz layer that is described in Subsection IV-B.
In the absence of measurement noise, it will be shown in
the following lemma that the optimal solution to (8) is nonunique.
Lemma 2. Consider the optimization problem in (8). Suppose
that the following assumptions are satisfied:
1) The global system input u(k) to (2) is persistently
exciting of order N n+s with s being the SID dimension
parameter in (4);
2) The measurement noise is absent, i.e., ŷi (k) = yi (k);
3) The matrix pairs (Al , B) and (Ar , B) have full row
rank.
Then, the optimal solution to the following rank optimization
problem is non-unique:
i
i
R
for R + s < i < N − R − s.
min rank Ys,h
− ΘB
s Us,h
B
Θs R ∈T
(9)
The proof to the above lemma is provided in Appendix A.
The above lemma indicates that the Markov parameters, as the
block entries of TsBR , cannot be recovered even if the global
optimal solution to the low-rank optimization problem in (9)
can be found.
In the next subsection, a solution will be provided by further
constraining the structure of block entries of TsBR , due to the
specific structures of the system matrices (AR , B R , C R ) in
(3).
B. Structure constrained low-rank optimization
The block matrix TsBR in (4) has a two-layer block Toeplitz
structure, which will be utilized in the proposed identification
5
method. The first layer is the block Toeplitz structure of
TsBR with respect to its block entries C R AjR B R . The second
layer is the partial block Toeplitz structure inside the block
entries C R AjR B R , as highlighted in the following example
and lemma.
Example 1. If we take R = 3 and assume that each block in
AR , B R , C R has size 2 × 2, then a visual illustration of the
structures of the matrices {M j = C R AjR B R }3j=1 is given in
Figure 2.
Fig. 2. Illustration of the partial Toeplitz structure of the matrices {M j =
C R AjR B R }3j=1 Top-left for j = 1 with block-bandwidth 1; top-right for
j = 2 with block-bandwidth 2; bottom for j = 3 with block-bandwidth
3. The deep blue color represents zero entries. The parts surrounded by red
curves have block Toeplitz structures.
Lemma 3. Based on the matrices AR , B R , C R defined in (3),
the following hold about the matrix product M j = C R AjR B R :
1) M j is a banded block matrix, with block-bandwidth j.
2) The submatrices of M j , for j < 2R + 1, at the
l−th block row and q−th block column with l, q ∈
{1, · · · , 2R + 1}, are inside the partial block-Toeplitz
region for the index-pair (l, q) satisfying:
i + 1 ≤ l + q ≤ 4R + 3 − i.
Proof: Since B R and C R are block diagonal matrices
with constant diagonal blocks, M j has the same structure
pattern as AjR . The matrix AR can be written as
AR = I ⊗ A + J− ⊗ Al + J+ ⊗ Ar
with J− and J+ block-column shifting matrices such that
post-multiplication of a matrix with J− results in shifting the
block-columns of that matrix to the left and adding a zero
block column in the last column, while post-multiplication of
a matrix with J+ results in the opposite shift and adding a
zero block column in the first column.
The proof is completed by induction. For j = 1 and j = 2
this can be easily checked. Let us assume that AjR has a block
Toeplitz structure for the matrix blocks with the index-pair
(l, q) satisfying that j + 1 ≤ l + q ≤ 4R + 3 − j, and using
as
the above expression for AR , we can express Aj+1
R
Aj+1
= AjR AR = AjR (I ⊗ A + J− ⊗ Al + J+ ⊗ Ar ). (10)
R
By several trivial manipulations of the above matrix product,
j+1
has a partial Toeplitz structure
it can be derived that AR
for the matrix blocks with the index-pair (l, q) satisfying that
(j + 1) + 1 ≤ l + q ≤ 4R + 3 − (j + 1). This completes the
proof.
Let Tf denote the set of two-layer block Toeplitz matrices
that was outlined above for the matrix TsBR . More clearly, the
structure of TsBR is illustrated in the following example.
Example 2. If we take R = 2 and s = 4, the Toeplitz matrix
TsBR can be parameterized in terms of the parameter set
p×m
{ϕj }11
as follows:
j=1 with ϕj ∈ R
0
M0
0
,
(11)
TsBR =
M1 M0
0
M2 M1 M0 0
ϕ1
ϕ1
,
ϕ1
where
M0
=
ϕ1
ϕ1
ϕ3 ϕ4
ϕ2 ϕ3 ϕ4
, M
ϕ2 ϕ3 ϕ4
=
=
M1
2
ϕ2 ϕ3 ϕ4
ϕ2 ϕ3
ϕ10 ϕ8 ϕ9
ϕ6 ϕ7 ϕ8 ϕ9
ϕ5 ϕ6 ϕ7 ϕ8 ϕ9 .
ϕ5 ϕ6 ϕ7 ϕ8
ϕ5 ϕ6 ϕ11
Then we specialize the rank-optimization problem in (8) as
B
min
i ∈H
Θs R ∈Tf ,Ŷs,h
h+s−1
X
kŷi (t) − y i (t)k2
t=1
(12)
i
h
i
i
R
+ λ · rank Ŷs,h
U
− ΘB
s
s,h .
The difference between (8) and (12) is in the definition of the
constrained Toeplitz sets indicated by T and Tf , respectively.
Though it is only a minor notational difference, the impact
on the solution is huge. As will be outlined in Theorem 1, it
enables uniqueness of part of the solution.
For the uniqueness property, use will be made of the
following time-varying observability matrix Oj,k , which is a
sub-matrix of the extended observability matrix Oj defined in
(4), consisting of the block rows corresponding to the secondlayer block Toeplitz part of TsBR .
Definition 2. Let Gj be a j × (j + 2) block Toeplitz matrix
of the form
Al A Ar
Al A Ar
(13)
Gj =
.
..
..
..
.
.
.
Al A Ar
6
Denote C j = Ij ⊗ C. A time-varying observability matrix
Oj,k , for j > 2(k − 1), is defined in terms of the matrix pair
(C j , Gj ) as [24, Chapter 3]:
Cj
C j−2 Gj−2
C j−4 Gj−4 Gj−2
Oj,k =
.
..
.
C j−2(k−1) Gj−2(k−1) · · · Gj−2
Theorem 1. Suppose that the following assumptions are
satisfied:
1) Assumption A.1 holds and νo is the observability index
of the pair (AR , C R );
2) The cluster radius R of the lifted model in (3) and the
dimension parameter s in (4) satisfy
s > νo ,
R ≥ s − 1;
3) The conditions 1)-3) of Lemma 2 hold;
4) The time-varying observability matrix O2R+1,s−1 , defined in Definition 2, has full column rank.
Then the submatrices of the Markov parameters M j =
C R AjR B R , for j = 0, 1, · · · , s − 2, contained in the matrix
TsBR in (4), for which Lemma 3 has shown that they preserve
the second layer of block-Toeplitz structure, can be computed
in a unique manner from the following low-rank optimization
problem:
min
B
Θs R ∈Tf
h
i
i
i
R
rank Ys,h
− ΘB
s Us,h
for R + s < i < N − R − s.
(14)
The proof of the above theorem can be found in the
Appendix B, where we can see that the (unique) global optimal
solution can yield correct recovery of certain parts of the
Markov parameters M j = C R AjR B R , for j = 0, 1, · · · , s − 2.
In addition, based on the conditions 2) and 4) of the above
theorem, it can be established that the inequality (7) holds,
i
namely the matrix sum Os xih + TsDR Vs,h
is indeed of lowrank (or rank deficient). This in turn indicates that the low-rank
cost function in (12) is a reasonable choice.
Lemma 2 and Theorem 1 demonstrate that the second-layer
Toeplitz structure, as pointed out in Lemma 3, is crucial for
enforcing a (unique) solution to the identification problem 1
of this paper. The considered identification of the Markov
parameters boils down to solving the low-rank regularized
optimization problem in (12).
Since the optimization problem in (12) is non-convex,
it is difficult to obtain an optimal solution under a mild
computational burden. In this paper, the reweighted nuclear
norm optimization method [25] is adopted, which is regarded
as an iterative heuristic for the rank minimization problem
(12).
V. S TATE - SPACE REALIZATION OF
A SINGLE SUBSYSTEM
In this section, we shall study the final realization of
the system matrices {C, A, Al , Ar , B} from the estimated
submatrices of the Markov parameters M j = C R AjR B R , for
j = 0, 1, · · · , s − 2, contained in the matrix TsBR in (4), for
which Lemma 3 has shown that they preserve the second-layer
of block-Toeplitz structure.
Remark 1. The realization problem considered in this section
is a structured state-space realization problem, which turns out
to be a challenging non-convex optimization problem [18],
[20], [26]. This problem is usually solved by the gradientbased optimization methods, which are sensitive to the selection of the initial condition. Here, we transform the realization
problem into a structured low-rank matrix factorization problem for which the optimal solution enables the estimation of
system matrices (C, Al , A, Ar , B) up to a similarity transformation. This factorization problem is then reformulated into
an equivalent low-rank optimization problem, which is then
numerically solved using the optimization method developed
in [27].
We start the solution by developing expressions of the
submatrices in the second-layer Toeplitz regions in terms of
the system matrices {C, A, Al , Ar , B}. This is done in the
following Lemma.
Lemma 4. Consider the block matrices AR , B R and C R
defined in (3). Let the sequence of non-zero block entries from
left to right of the (j +1)-th block row of the matrix C R AjR B R
be denoted as {Fj,−j , Fj,1−j , · · · , Fj,j−1 , Fj,j }, then these
matrix entries satisfy the following relationship:
j
X
Fj,k z −k = C(Al z −1 + A + Ar z)j B,
(15)
k=−j
where z ∈ C.
Proof: The above result can be derived using the filter
bank theory in [28].
As Fj,k are the Markov parameters inside the second-layer
block Toeplitz part of TsBR , the values of Fj,k for j ∈
{0, 1, · · · , s−2}, k ∈ {−j, · · · , j} are assumed to be available
in this section. Based on these values Fj,k , we address the
problem of estimating the matrices {C, A, Al , Ar , B} up to a
similarity transformation.
Dual to Definition 2, we shall define a time-varying controllability matrix Cj,k , which is a sub-matrix of the extended
controllability matrix determined by (AR , B R ).
Definition 3. Let Γj be a (j + 2) × j block Toeplitz matrix of
the form
Ar
A Ar
.
..
Al A
(16)
Γj =
.
..
.
A
A
r
l
..
. A
Al
Denote B j = Ij ⊗ B. A time-varying controllability matrix
Cj,k , for j > 2(k − 1), is defined in terms of the matrix pair
(Γj , B j ):
Cj,k = [ B j
Γj−2 B j−2
···
Γj−2 · · · Γj−2(k−1) B j−2(k−1) ] .
7
In the sequel, we let s be an even integer such that s/2
is an integer as well. The solution to the realization of the
system matrices {C, Al , A, Ar , B} is done in two phases.
In the first phase, the structured time-varying observability
matrix O2R+1,s/2 and the structured time-varying controllability matrix C2R+1,s/2 are to be estimated from the available
matrix values Fi,j . In the second phase, the system matrices {C, A, Al , Ar , B} are derived from these time-varying
observability and controllability matrices. It is remarked that
the subscript s/2 in O2R+1,s/2 (or C2R+1,s/2 ) means that
the maximum moment of the block entries in O2R+1,s/2 (or
C2R+1,s/2 ) is s/2−1. The subscript s/2 is adopted because the
sum of the maximum moments of O2R+1,s/2 and C2R+1,s/2 is
equal to the maximum moment, s− 2, of the available Markov
parameters M j .
A. Determining the time-varying observability and controllability matrices
In this subsection, the determination of O2R+1,s/2 and
C2R+1,s/2 will be formulated as a structured low-rank matrixfactorization problem. More importantly, we show that the
optimal solution to this matrix-factorization problem can
yield the estimates of O2R+1,s/2 and C2R+1,s/2 up to a
block-diagonal ambiguity matrix with identical block-diagonal
entries. This is crucial to identify the system matrices
{C, A, Al , Ar , B} up to a similarity transformation, as stated
in Problem 2.
First, by the definitions of O2R+1,s/2 and C2R+1,s/2 , we
can find that the product of O2R+1,s/2 and C2R+1,s/2 is equal
to a matrix constructed from {Fj,k }jk=−j for j = 0, 1, · · · , s−
2. This is demonstrated in the following example.
Example 3. When R = 1 and s = 4, the product of O3,2 and
C3,2 can be expressed as
C
0
0
B 0 0 Ar B
0
C
0
0 B 0 AB
O3,2 C3,2 =
0
0
C
0 0 B Al B
CAl CA CAr
F0,0
0
0
F1,1
0
F1,0
F0,0
0
.
=
0
0
F0,0 F1,−1
F1,−1 F1,0 F1,1 F2,0
The above equation provides a simple example, showing that
the product of O3,2 and C3,2 can be expressed in terms of
{Fj,k }jk=−j for j = 0, 1, 2.
Here, the product of O2R+1,s/2 and C2R+1,s/2 is represented as:
O2R+1,s/2 C2R+1,s/2 = H2R+1,2R+1 ,
(17)
where H2R+1,2R+1 is a (2R + 1) × (2R + 1) block matrix
that is assumed to be known.
Given the matrix H2R+1,2R+1 , the problem of interest is
to determine O2R+1,s/2 and C2R+1,s/2 from equation (17).
According to Definitions 2 and 3, we can see that O2R+1,s/2
and C2R+1,s/2 are structured matrices. These structures are
instrumental to analyzing the properties of the optimal solution
to (17), as shown in Theorem 2. In order to show these
structures, use will be made of the following matrix definition.
For j = 0, 1, · · · , s/2 and z ∈ C, the matrix sequences
{Wj,l }jl=−j and {Ej,l }jl=−j are defined as
j
X
Wj,l z −l = C(Al z −1 + A + Ar z)j
l=−j
j
X
(18)
Ej,l z
−l
= (Al z
−1
j
+ A + Ar z) B.
l=−j
From the defined quantities Wj,l and Ej,l in (18), the matrices
Wj and Ej are defined as
T
T
E0,0
W0,0
T
E1,−1
W1,−1
T
E1,0
W1,0
T
E1,1
W1,1
(19)
Wj =
Ej =
.
T
E2,−2
W2,−2
..
..
.
.
Wj−1,j−1
T
Ej−1,j−1
It can be seen that Oj,k (or Cj,k ) is a block matrix constructed
from the block components of Wk (or Ek ). This is illustrated
in the next example.
Example 4. The time-varying observability matrix O5,3 is
represented in terms of W3 as follows
W0,0
W0,0
W
0,0
W
0,0
W
O5,3 =
0,0
.
W1,−1 W1,0
W
1,1
W
W
W
1,−1
1,0
1,1
W1,−1 W1,0 W1,1
W2,−2 W2,−1 W2,0 W2,1 W2,2
Denote by O2R+1,s/2 the set of block matrices having the
same structure as O2R+1,s/2 , as illustrated in Example 4 , and
C2R+1,s/2 the set of block matrices having the same structure
as C2R+1,s/2 . It is remarked that the definitions of both the
sets O2R+1,s/2 and C2R+1,s/2 require knowledge of the system
order n. We then propose the following structured low-rank
matrix factorization problem:
min kH2R+1,2R+1 − OCk2F
O,C
(20)
s.t. O ∈ O2R+1,s/2 , C ∈ C2R+1,s/2 .
It will be shown in the following theorem that the optimal
solution to (20) can yield the estimates of O or C up to a
block-diagonal ambiguity matrix with identical block-diagonal
entries.
Theorem 2. Consider the optimization problem in (20). Suppose that the following assumptions are satisfied:
1) The values of R and s satisfy R ≥ s − 2, and s is a
positive even integer;
2) The system order of a subsystem, n, is known;
8
3) The matrices Oj,s/2 and Cj,s/2 , for any j ≥ min{2s −
3, 2R}, have full column and row rank, respectively;
4) The matrix H2R+1,2R+1 satisfying equation (17) is
known exactly.
Then, any optimal solution pair {Ô, Ĉ} to the optimization
(20) satisfies
Ô = O2R+1,s/2 Q
Ĉ = Q
−1
C2R+1,s/2 ,
(21)
where Q = I2R+1 ⊗ Q with Q ∈ Rn×n being a nonsingular
ambiguity matrix.
The proof of this theorem is given in Appendix C.
B. Rank-constrained form of the optimization problem (20)
In this subsection, the structured low-rank matrix factorization problem in (20) is reformulated into a rank-constrained
optimization problem.
From the matrix quantities defined in (18) and (19), we
can see that O2R+1,s/2 is affine in terms of Ws/2 , while
C2R+1,s/2 is affine in terms of Es/2 . Then, it can be established that product O2R+1,s/2 C2R+1,s/2 is affine in terms of
block entries of Ws/2 Es/2 [29]. This affine operator H(·) is
defined by
O2R+1,s/2 C2R+1,s/2 = H(Ws/2 Es/2 ).
(22)
In the sequel, the affine operator H(·) is assumed to be known.
Instead of O2R+1,s/2 and C2R+1,s/2 , we regard the product
X = Ws/2 Es/2 as the variable to be determined. Then, the
affine function H(X) works on the block entries of X. By
taking into account the low-rank property of the matrix product Ws/2 Es/2 , we propose the following rank-constrained
optimization problem:
min kH2R+1,2R+1 − H(X)k2F
X
s.t. rank(X) = n,
(23)
where n is the system order of a subsystem in (1) and X has
the same size as Ws/2 Es/2 .
Next, we will show that the rank-constrained optimization
problem in (23) is an equivalent formulation of (20). To show
this result, the following lemma is required.
Lemma 5. Suppose that Oj,s/2 and Cj,s/2 , for j > s − 2,
have full column and row rank, respectively. Then, Ws/2 and
Es/2 have full column and row rank, respectively, and both
ranks are n.
Proof: Since Oj,s/2 has full column rank, the block
iT
h
T
T
T
, which is constructed
vector W0,0
, W1,−1
, · · · , Ws/2,−s/2
by stacking the non-zero block entries in the first block column
of Oj,s/2 , has full column rank as well. Since this block vector
is a sub-part of Ws/2 , it can be derived that Ws/2 has full
column rank. Similarly, it can be proven that the block vector
Es/2 has full row rank as well.
The next theorem shows the equivalence between the optimization problems in (20) and (23).
Theorem 3. Consider the optimization problems in (20) and
(23). Suppose that all the conditions of Theorem 2 hold. Then,
the optimization problems (20) and (23) are equivalent in
the
can yield the estimate
sense that their optimal solutions
Ŵs/2 , Ês/2 of Ws/2 , Es/2 up to an ambiguity matrix,
i.e., there exist a nonsingular matrix Q ∈ Rn×n such that
Ŵs/2 = Ws/2 Q, Ês/2 = Q−1 Es/2 .
Proof: First, we will show that the optimal
solution
to (20) can yield the estimate of Ws/2 , Es/2 up to an
ambiguity matrix.
From the matrix quantities in (18) and (19), we can see
that structured matrices O2R+1,s/2 and C2R+1,s/2 are linearly
parameterized by the block components of Ws/2 and Es/2 ,
respectively. By the result of Theorem 2, it can be derived
that the optimal
solution to (20) can yield the estimate of
Ws/2 , Es/2 up to an ambiguity matrix.
Second, we will show that the optimal
solution to (23) can
yield the estimate of Ws/2 , Es/2 up to an ambiguity matrix.
By equation (22), the optimization problem (20) can be
reformulated as
min
W,E
kH2R+1,2R+1 − H(WE)k2F ,
(24)
where the variables W and E have the same sizes of Ws/2 and
Es/2 , respectively. It is noted that equation (24) is a parameterized form of the structured low-rank matrix factorization
problem in (20).
By Theorem 2 and Lemma
5, the optimal solution
(Ŵ, Ê) to (24) satisfies rank ŴÊ = n and H(ŴÊ) =
H2R+1,2R+1 . It can be derived that X = ŴÊ is an optimal
solution to (23) and the criterion of (23) becomes zero.
Let X̂ be an optimal solution to (23). It should satisfy
H(X̂) = H2R+1,2R+1 and rank X̂ = n. Let the SVD
decomposition of X̂ be given by X̂ = Û ΣV̂ T , where Û and
V̂ are constructed by n orthogonal columns, and Σ ∈ Rn×n
is a diagonal matrix with positive diagonal entries. Then,
(Ŵ, Ê) = (Û , ΣV̂ T ) is an optimal solution
to (24). Therefore,
(Û , ΣV̂ T ) is an estimate of Ws/2 , Es/2 up to an ambiguity
matrix. The proof is then completed.
The rank-constrained optimization problem in (23) is nonconvex and NP-hard. Following our previous work [27], the
rank-constrained optimization problem (23) is recast into
a difference-of-convex optimization problem which is then
solved using the sequential convex programming method.
Let X̂ be an optimal solution to (23). The SVD decomposition of X̂ is given by
T
Σ1
V1
,
(25)
X̂ = U1 U2
Σ2
V2T
where U1 and V1 consists of n orthogonal columns, and the
diagonal matrix Σ1 ∈ Rn×n has larger diagonal entries than
Σ2 . The estimates of Ws/2 and Es/2 can then be obtained as
follows:
Ŵs/2 = U1 ,
Ês/2 = Σ1 V1T .
9
Since O2R+1,s/2 and C2R+1,s/2 are respectively affine in
terms of Ws/2 and Es/2 , the variables O and C in (20) can
be estimated from Ŵs/2 and Ês/2 , respectively.
C. Determining the system matrices {A, Al , Ar , B, C}
In view of the theoretical result in Theorem 2, we assume
that the obtained estimates Ô2R+1,s/2 and Ĉ2R+1,s/2 satisfy
Ô2R+1,s/2 = O2R+1,s/2 Q,
Ĉ2R+1,s/2 = Q−1 C2R+1,s/2 ,
(26)
where Q = I2R+1 ⊗ Q with Q ∈ Rn×n being nonsingular.
Based on these estimates, we will address the identification
of the system matrices {A, Al , Ar , B, C} up to a similarity
transformation.
First, the shifting structure of the time-varying observability
matrix O2R+1,s/2 will be explored. Denote
C j−2k1 Gj−2k1 · · · Gj−2
C j−2(k +1) Gj−2(k1 +1) · · · Gj−2
1
Oj,k1 :k2 =
..
.
C j−2k2 Gj−2k2 · · · Gj−2
where 0 ≤ k1 < k2 ≤ s/2 − 2 and 2k2 ≤ j ≤ 2R + 1. The
matrix Oj,k1 :k2 above is constructed by the block rows of Oj,k
with block-row indices from k1 to k2 . Then, the structureshifting property of O2R+1,s/2 can be represented as
O2R−1,0:s/2−2 G2R−1 = O2R+1,1:s/2−1 ,
(27)
where O2R−1,0:s/2−2 and O2R+1,1:s/2−1 are sub-matrices of
O2R+1,s/2 , and G2R−1 is a block Toeplitz matrix defined in
Definition 2. This is illustrated in the following example.
Example 5. When R = 2 and s = 6, the structure-shifting
property in (27) can be explicitly written as
Lemma 6. Let Ô2R+1,s/2 satisfy equation (26). Assume
that Ô2R−1,0:s/2−2 has full column rank. Then, the optimal
solution Ĝ to the optimization problem in (28) satisfies
(29)
Ĝ = I2R−1 ⊗ Q−1 G2R−1 (I2R+1 ⊗ Q) ,
where Q ∈ Rn×n is a nonsingular matrix.
The above lemma can be derived straightforwardly based
on equation (26) and the optimization formulation (28).
The condition that the time-varying observability matrix
Ô2R−1,0:s/2−2 has full column rank is similar to condition 3)
of Theorem 2. Lemma 6 implies that the matrices Al , A, Ar
can be determined up to a similarity transformation, i.e.,
Âl = Q−1 Al Q, Â = Q−1 AQ, Âr = Q−1 Ar Q.
In addition, according to equation (26), the estimates Ĉ and B̂
can be extracted respectively from Ô2R+1,s/2 and Ĉ2R+1,s/2 ,
satisfying that
Ĉ = CQ, B̂ = Q−1 B.
To ease the reference, the proposed local network identification algorithm is summarized in Algorithm 1.
Algorithm 1: Local identification for 1D distributed systems
Construct a spatially stacked state-space model (3) and its
Step 1
temperally stacked equation (4) based on local observations;
B
Step 2
Estimate Ts R from the optimization problem (12);
Step 4
Estimate O2R+1,s/2 and C2R+1,s/2 using
the method described in Subsection V-B;
Step 5
Extract the estimates of C and B from the estimates of
O2R+1,s/2 and C2R+1,s/2 , respectively;
Step 6
Estimate Al , A, Ar by solving (28).
VI. N UMERICAL
SIMULATION
In this section, numerical simulations are provided to
demonstrate
the effectiveness of the proposed identification
C
method
–
Algorithm
1. In the simulation, the distributed
A
A Ar
C
l
system
is
constructed
by
connecting 40 identical subsystems
Al A Ar
C
Al A Ar
in
a
line,
and
the
identification
for the 20-th subsystem
CAl CA CAr
is performed. The system matrices (A, Al , Ar , B, C) with
CAl
CA
CAr
3×3
, B ∈ R3×2 and C ∈ R2×3 are randomly
CAl
CA
CAr
A, Al , Ar ∈ R
,
=
generated
such
that
Assumption A.1 is satisfied and the 1D
CAl
CA
CAr
CA2l C(AAl + Al A)
∗
C(AAr + Ar A) CA2r
networked system is stable.
To construct an augmented state-space model in (3), we set
where ∗ is used to represent the term C(Al Ar + A2 + Ar Al ). R = 5 and s = 8. The system input and the measurement noise
Based on equation (27), we formulate the following struc- are generated as white noise sequences and the data length is
tured least-squares optimization problem to identify the ma- set to 800.
To evaluate the identification performance against the noise
trices Al , A, Ar based on the estimate Ô2R+1,s/2 :
effect, the criterion signal-to-noise ratio (SNR) is adopted,
which is defined as
min kÔ2R−1,0:s/2−2 G − Ô2R+1,1:s/2−1 k2F
G
(28)
var(yi (k) − ei (k))
s.t. G ∈ G2R−1 ,
.
SNR = 10 log
var(ei (k))
where G2R−1 denotes a set of matrices having the same
In the sequel, we shall carry out numerical simulations with
structure as G2R−1 , as shown in Definition 2.
The optimal solution to (28) has properties shown in the the SNR ranging from 0 dB to 95 dB.
We use the criterion impulse-response fitting to evaluate
following lemma.
the performance of the proposed identification method. The
10
normalized fitting error of the impulse-response sequence
CAi B is defined by
T P10
1 X i=0 kCAi B − Ĉj Âij B̂j kF
,
P10
i
T j=1
i=0 kCA BkF
Normalized fitting error
10-1
10-2
10-3
0.02
Normalized fitting error
where T is the number of randomly generated networked
systems and {Ĉj , Âj , B̂j } are the estimates of {C, A, B} of
the j-th generated networked system, respectively. Similarly,
we can also define the normalized fitting error for the impulseresponse sequences CAil B and CAir B.
Fig. 3 shows the identification performance of the proposed
method against the SNR. The normalized fitting errors are
calculated by averaging the results of 100 randomly generated
networked systems, and the regularization parameter λ in (8)
is set to λ = 10−3 . It can be observed from Fig. 3 that, using
the proposed identification method, the normalized fitting error
decreases along with the increase of the SNR. When the SNR
is larger than 50 dB, the normalized fitting errors can be as
small as 10−4 .
0.025
0.015
0.01
0.005
0
10-8
10
20
30
40
50
60
70
80
90
VII. C ONCLUSION
The local identification of 1-D large-scale distributed systems has been studied. Compared with the classical system
identification problems, the challenging point of the local
system identification is that there are two unknown system
inputs which are the states of its neighboring subsystems. By
exploiting both the spatial and temporal structures of the distributed system, especially the two-layer Toeplitz structure of
the Markov-parameter matrix, a low-rank optimization problem has been provided for identifying the Markov parameters
of a local cluster of identical subsystems, where the associated
optimal solution can yield (parts of) the true Markov parameters. Moreover, the system realization of a structured statespace model is formulated as a structured low-rank matrix
100
A PPENDIX A
P ROOF OF L EMMA 2
100
Fig. 4 shows the impulse-response fidelity of individual
subsystems against the regularization parameter λ, where
SNR is set to SNR=40 dB. The normalized fitting errors are
computed by averaging the results of 100 randomly generated
networked systems. We can observe from Fig. 4 that good
identification performance can be achieved by choosing the
regularization parameter λ around 10−2 .
10-2
factorization problem, showing that the system matrices can
be determined up to a similarity transformation by enforcing
the structure of the generalized observability/controllability
matrix.
Although we only consider the identification of 1D homogeneous networked systems, it can be extended to 2D homogeneous networks using the same identification framework,
namely exploiting both the spatial and temporal structures
of the concerned distributed system. In our future work,
the identification of large heterogeneous networks will be
investigated.
SNR (dB)
Fig. 3. Normalized fitting errors under different noise levels with λ = 10−3 .
10-4
Fig. 4. Normalized fitting errors against the regularization parameter λ with
SNR=40 dB.
10-4
0
10-6
The proof of the lemma is divided
into three steps.
xih
i
has full row rank under
Step I: we shall show that Vs,h
i
Us,h
the given assumptions in the lemma.
It can be derived from (2) that
I
···
A
···
= .
· · · ..
As−1
x(k + s − 1) · · ·
x(k)
x(k + 1)
..
.
×
B
..
.
As−2 B
x(k)
u(k)
..
.
..
.
···
···
···
B
.
···
u(k + s − 2) · · ·
(30)
Let
Ξ = [h O(2R+1)n,(i−R−1)n I(2R+1)n O(2R+1)n,(N−i−R)n ], i
O
In
On,(2R+1)n
On,n
On,(N −i−R−1)n
Ω = On,(i−R−2)n
On,n
On,(2R+1)n
In
On,(N −i−R−1)n
n,(i−R−2)n
and
Π = [ O(2R+1)m,(i−R−1)m I(2R+1)m O(2R+1)m,(N−i−R)m ]
11
be selection matrices such that
Ξ
Ω
ΩA
ΩB
i
..
..
..
.
xh
.
.
i
s−1
s−2
Vs,h = ΩA
ΩA
B
·
·
·
ΩB
i
Us,h
0
Π
..
..
.
.
0
Π
0
x(k)
···
u(k)
···
.
..
×
.
·
·
·
u(k + s − 2) · · ·
u(k + s − 1) · · ·
and
0
(31)
Π
By Lemma 10.4 in [20], under the assumption that u(k) is
persistently exciting of order nN + s, it can be established
that
x(k)
···
u(k)
···
..
.
···
u(k + s − 1) · · ·
has full row rank. In addition, by explicitly unfolding the
expressions of ΩAj B, it can be verified that, when (Al , B)
and (Ar , B) have full row rank and R ≥ s − 1, the coefficient
matrix on the right-hand side of (31) has full row rank. For
the sake of brevity, we only show this for the case with
s = 3. The associated coefficient matrix in (31) is shown
in (32) with ∗ denoting unimportant entries for analyzing the
concerned rank property. Under the condition that (Al , B) and
(Ar , B) have full row rank, it is easy to see that the submatrix stacked by the first five and last two block rows of
(32) has full row rank. It can be further verified that the whole
coefficient matrix in (32) has full row rank. This
is a conse
2
Al B 0
quence of Al Al B B = Al B
0 0 I
Ar B 0
2
with
and Ar Ar B B = Ar B
0 0 I
Al B and Ar B having full row rank.
Now that the coefficient matrix in (31)has full row rank,
xih
i
has full row rank
Sylvesters’ inequality shows that Vs,h
i
Us,h
as well.
R
R
Step II: Let TB
∈ T be the true value of ΘB
s
s . Next, we
BR
shall show that Ts is an optimal solution
(9). Let ∆s ∈ T .
i to
xh
i
, we can derive
Under the full row rank property of Vs,h
i
Us,h
that
i
i
R
rank Ys,h
− TB
s Us,h
i
xh
Os TsDR
=rank
i
Vs,h
=rank Os TsDR
i
i
R
− ∆s Us,h
rank Ys,h
− TB
s
i
xh
i
=rank Os TsDR ∆s Vs,h
i
Us,h
DR
∆s .
=rank Os Ts
By the inequality
rank Os TsDR ≤ rank Os
TsDR
∆s
we can obtain that
i
i
R
rank Ys,h
− TB
s Us,h
i
i
R
− ∆s Us,h
− TB
≤ rank Ys,h
s
(33)
,
(34)
R
is an optimal solution to (9).
for any ∆s ∈ T . Thus, TB
s
Step III: In this step, we shall show that the optimal solution
to (9) is non-unique. Let
∆s = TsDR Gs ,
(35)
R
where Gs = Is ⊗GR with GR ∈ R2n×(2R+1)n . Denote TB
=
s
BR
BR
Ts − ∆s . It is easy to verify that Ts ∈ T .
Since ∆s = TsDR Gs , it can be established that
i
i
R
rank [Yk,s,h
− TB
s Uk,s,h
= rank Os TsDR ∆s
I 0 0
DR
Os Ts
(36)
= rank
0 I Gs
= rank Os TsDR
i
i
R
= rank [Ys,h
− TB
s Us,h .
R
It can then be observed from the above equation that both TB
s
BR
and Ts are optimal solutions to (9); therefore, the optimal
solution to (9) is non-unique.
A PPENDIX B
P ROOF OF T HEOREM 1
Following the proof of Lemma 2, the following matrix in
(31) has full row rank:
i
xh
i
.
Vs,h
i
Us,h
Let Tf denote the set of the two-layer block Toeplitz
R
matrices, defined in Subsection IV-B. Let TB
∈ Tf be the
s
R
true value of the estimate of ΘB
in
(14).
Since
the matrix set
s
Tf belongs to the matrix set T , the minimal value of the cost
function in (9) is smaller than or equal to that in (14), i.e.
i
i
R
min rank Ys,h
− ΘB
s Us,h
B
Θs R ∈T
≤
i
i
R
min rank Ys,h
− ΘB
s Us,h .
B
Θs R ∈Tf
R
∈ Tf is an
As shown in Lemma 2, the true value TB
s
optimal solution to (9). Then, it can be observed from the
R
above inequality that TB
∈ Tf is also an optimal solution to
s
12
0
0
0
0
0
A2l
0
0
0
0
Al
0
∗
0
0
In
0
A
0
∗
0
| In
| 0
| 0
| Ar
| 0
| ∗
| 0
|
|
∗
∗
∗
∗
∗
∗
∗
In
0
0
0
Al
0
∗
| 0
| 0
| In
| 0
| A
| 0
| ∗
|
|
0
0
0
0
Ar
0
∗
0
0
0
0
0
0
A2r
0
0
Al B
0
B
0
AB
0
|
|
|
0 |
∗
0 |
∗
Ar B |
∗
0 |
∗
| I(2R+1)m
|
0
|
|
|
| 0
| 0
| 0
| Al B
|
|
0
B
0
AB
0
0
0
Ar B
0
0
|
|
|
|
|
0|
∗
0|
∗
|
0
| I(2R+1)m
B
0
|
|
|
|
|
|0
|0
|
|
0
0
0
B
(32)
(14). As a consequence, the corresponding minimal value of
the cost function of (14) is
rank
TsDR
Os
.
BR
R
R
Let TB
∈ Tf be such that ∆s = TB
is a perturbation
s
s −Ts
we seek of the true value such that it retains the minimal value
of the cost function in (14). Using the full row rank condition
of the matrix in (31), we obtain that
i
i
R
rank Ys,h
− TB
s Us,h
i
i
+ ∆s Us,h
= rank Os xih + TsDR Vs,h
= rank Os TsDR ∆s .
(37)
∗
..
.
∗
0
0
0
0
.
∗
..
.
(38)
∗
(b) Permutation of ∆s : Let Ls denote the matrix that
permutes all zero block rows of TsDR to top such that
Ls TsDR =
2
TsDR
0
T2
,
Ls Os =
Os1
∗
following special structures
0
∆0,1
..
.
···
0
..
.
∆1,s−3
∗
..
.
∆0,s−2
,
0
,
(40)
Now, we seek that part of ∆s ∈ Tf that retains the minimal
value of the cost function in (14) only when it is zero. For that
reason, we permute the block rows of ∆s compatible with the
zero block-row pattern of the matrix TsDR .
(a) Structure of TsDR : The structure of TsDR is determined
by C R AjR DR for j = 0, 1, · · · , s − 2. By the definitions of
AR , C R and DR in (3), we can find that C R AjR DR is a (2R+
1) × 2 block matrix. When R ≥ (s − 1), the block rows of
C R AjR D R indexed from (j + 2) to (2R − j) are zero. For
instance, C R AjR D R has the block structure as shown below
with ∗ representing non-zero block entries:
∆s and Os , yielding the
1
∆
Ls ∆s =
∗
0
∆
0,0
∆
1,0
=
..
.
∆s−2,0
(39)
where T is a sub-matrix of
by removing the corresponding zero block rows, as illustrated in (38). It is noteworthy that
the zero block rows of TsDR correspond to the second-layer
block Toeplitz part of TsBR , as shown in Lemma 3.
Then, the permutation matrix Ls is applied to the matrices
such that Os1 = O2R+1,s and ∆j,l , for j = 0, 1, · · · , s − 2 and
l = 0, · · · , s − 2 − j, is a [2(R − j − l) − 1] × [2R + 1] block
Toeplitz matrix in the form
0 δj,−j δj,1−j · · ·
δj,j
0
..
..
..
..
..
..
∆j,l =
,
.
.
.
.
.
.
0
δj,−j
δj,1−j
···
δj,j
0
with {δj,k }jk=−j being block entries of appropriate sizes. Note
that ∆j,l for l ≥ 1 is a sub-matrix of ∆j,0 , which is constructed
by stacking the block rows indexed from (l+1) to (2R−1−l).
Now, we are ready to show for what values of ∆j,l the
following rank constraint holds
rank Os TsDR ∆s = rank Os TsDR .
(41)
It can be derived from the above equality that
Q1
DR
= Ls ∆s ,
Ls Os Ts
Q2
or
1
∆
O2R+1,s 0
Q1
=
,
Q2
∗
T2
∗
(42)
(43)
where Q1 and Q2 are coefficient matrices of appropriate sizes,
and ∗ denotes non-zero block entries caused by the block-row
permutation. So we focus on the top part of (43) and rewrite
it as
(44)
∆1 = O2R+1,s Q1,0 Q1,1 · · · Q1,s−1 ,
s−1
where {Q1,j }j=0
are sub-matrices of Q1 satisfying that Q1 =
[Q1,0 Q1,1 · · · Q1,s−1 ].
By considering the last block column of equation (44),
under the assumption that O2R+1,s−1 has full column rank, we
can obtain that Q1,s−1 = 0. Similarly, by considering the second to the last block column, it can be derived that Q1,s−2 = 0
13
and ∆0,s−2 = 0, further ∆0,j = 0 for j = 0, 1, · · · , s − 3.
Next, for the third to the last block column, since ∆0,s−3 = 0
and O2R+1,s−1 has full column rank, we can obtain that
Q1,s−3 = 0 and ∆1,j = 0 for j = 0, · · · , s − 3. By repeating
this procedure, it can be established that equation (44) holds
only when ∆1 = 0 and Q1,j = 0 for j = 0, 1, · · · , s − 1. In
other words, the matrix ∆s with a nontrivial sub-matrix
∆1
DR
cannot be linearly represented by Os Ts
.
Since the block rows of ∆1 correspond to the secondlayer Toeplitz part of ∆s , we can see that the rank
R
Os TsD
∆s
is strictly larger than that of
of
DR
Os Ts
as long as the second-layer Toeplitz part of
∆s is non-zero. Thus, the proof is completed.
To illustrate this, the first block-column equation of (47) can
be equivalently written as follows:
Ĉ
ˆl
CA
0
0
0
0
0
0
C
CAl
0
0
=
0
0
0
0
0
CA
C
0
0
0
CAl
0
0
CAr
0
C
0
0
CA
CAl
0
0
0
0
C
0
CAr
CA
0
0
0
0
0
C
0
CAr
Q1,1
Q2,1
Q3,1
.
Q4,1
Q5,1
It is obvious that, since O4,2 (or O2R,s/2 ) has full column
rank, it has
A PPENDIX C
P ROOF OF T HEOREM 2
Q2,1 = Q3,1 = Q4,1 = Q5,1 = 0,
ˆ l = CAl Q1,1 .
Ĉ = CQ1,1 , CA
Assume that the matrix pair {Ô2R+1,s/2 , Ĉ2R+1,s/2 } is one
of the optimal solution pairs to (20), it can be established that
Ô2R+1,s/2 and O2R+1,s/2 has the same structure and the same
column space, which can be algebraically represented as
Ô2R+1,s/2 = O2R+1,s/2 Q,
(45)
where Q ∈ R(2R+1)n×(2R+1)n is a (2R + 1, 2R + 1) block
matrix which is nonsingular. In the sequel, we denote Qi,j the
(i, j)-th block entry of Q. In order to prove the theorem, it is
sufficient to prove that the estimate Ŵs/2 of Ws/2 satisfies
that
Ŵs/2 = Ws/2 Q,
(46)
where Q ∈ Rn×n is a nonsingular matrix and Ws/2 is defined
in (19). In the sequel, we denote by Ŵi,j , for j = −i, · · · , i,
the estimate of Wi,j which is a block component of Ws/2 .
To illustrate the structure of equation (45), it is expanded
into equation (47) by setting R = 2 and s = 4.
The following proof will be divided into three steps.
Step 1. We will show that, under the assumption that
O2R,s/2 has full column rank, the following relations can be
derived from equation (45):
i = 0 : Ŵi,−i = Wi,−i Q1,1 ,
Wi,−i Q1,j = 0 for j = 2, · · · , s − 1;
i = 1 : Ŵi,−i = Wi,−i Q1,1 ,
Ŵi,2−i = Wi,−i Q1,3 + Wi,1−i Q1,2 + Wi,2−i Q1,1 ,
Wi,−i Q1,j+2 + Wi,1−i Q1,j+1 + Wi,2−i Q1,j = 0
for j = 2, · · · , s − 3;
···
Qi,2 = 0 for i = 3, · · · , 2R + 1;
CQ1,2 = 0;
Ŵi,1−i = Wi,−i Q1,2 + Wi,1−i Q1,1 for i = 1, · · · , s/2 − 1.
(51)
To illustrate this, by using the relations in (50), the second
block-column equation of (47) can be equivalently written as
ˆ
CA
0
CQ1,1
0
0
0
CAl Q1,1
0
CA
l
C
=
CA
CAr
C
C
C
C
CAl
CA
CAl
CAr
CA
CAr
Q1,2
Q2,2
Q3,2
Q4,2
Q5,2
We can observe that the bottom subpart of the vector on the
left-hand side equals the product of the first block column of
O4,2 and Q1,1 . By the assumption that O4,2 (or O2R,s/2 ) has
full column rank, we can derive that
Q3,2 = Q4,2 = Q5,2 = 0; CQ1,2 = 0;
ˆ = CAl Q1,2 + CAQ1,1 .
CA
i = s/2 − 1 : Ŵi,−i = Wi,−i Q1,1 ,
···
Ŵi,s−2−i = Wi,−i Q1,s−1 + · · · + Wi,s−2−i Q1,1 .
(48)
Let’s consider the equation corresponding to the first blockcolumn of (45). Since O2R,s/2 has full column rank, it can be
derived that
Ŵi,−i = Wi,−i Q1,1 for i = 0, 1, · · · , s/2 − 1.
Next, by substituting the relations in (49) into the equation
corresponding to the second block-column equation of (45)
and under the assumption that O2R,s/2 has full column rank,
we can obtain that
(52)
Ŵi,1−i = Wi,−i Q1,2 + Wi,1−i Q1,1 ,
Qi,1 = 0 for i = 2, 3, · · · , 2R + 1;
(50)
(49)
Q2,2 = Q1,1 ;
(53)
By iteratively considering from the first to the (s − 1)-th
block-column equation of (45) using the same strategy, we
can derive the relations in equation (48).
Step 2. Under the assumption that O2R,s/2 has full column
rank, the following relations can be derived from equation (45)
.
14
Ĉ
ˆ
CAl
Ĉ
Ĉ
Ĉ
Ĉ
ˆ
CA
ˆl
CA
ˆr
CA
ˆ
CA
ˆl
CA
ˆr
CA
ˆ
CA
ˆr
CA
C
=
CA
l
C
C
C
C
CA
CAl
CAr
CA
CAl
as well:
CAr
CA
CAr
Q1,1
Q2,1
Q3,1
Q4,1
Q5,1
Q1,2
Q2,2
Q3,2
Q4,2
Q5,2
Q1,3
Q2,3
Q3,3
Q4,3
Q5,3
Q1,4
Q2,4
Q3,4
Q4,4
Q5,4
Q1,5
Q2,5
Q3,5
Q4,5
Q5,5
(47)
equations (55)-(56) can be explicitly written as
i = 0 : Ŵi,i = Wi,i Q2R+1,2R+1 ,
Wi,i Q2R+1,j = 0 for j = 2R + 3 − s, · · · , 2R;
i = 1 : Ŵi,i = Wi,i Q2R+1,2R+1 ,
Ŵi,i−1 = Wi,i Q2R+1,2R + Wi,i−1 Q2R+1,2R+1 ,
Ŵi,i−2 = Wi,i Q2R+1,2R−1 + Wi,i−1 Q2R+1,2R
+ Wi,i−2 Q2R+1,2R+1 ,
Wi,i Q2R+1,j + Wi,i−1 Q2R+1,j+1
+ Wi,i−2 Q2R+1,j+2 = 0
for j = 2R + 3 − s, · · · , 2R − 2;
···
i = s/2 − 1 : Ŵi,i = Wi,i Q2R+1,2R+1 ,
···
Ŵi,i−s+2 = Wi,i Q2R+1,2R+3−s + · · ·
+ Wi,i−s+2 Q2R+1,2R+1 .
(54)
The relations in (54) are derived using the same strategy in
Step 1 by iteratively considering from the (2R + 1)-th to the
(2R + 3 − s)-th block-column equation of (45) in a reverse
order.
Step 3. We will show Ŵs/2 = Ws/2 Q1,1 using the relations
in (48) and (54).
Denote by Ô2s−3,s/2 (:, s−1) the (s−1)-th block column of
the estimate Ô2s−3,s/2 , which is a sub-matrix of Ô2R+1,s/2 .
The relations in (48) can be compactly written as:
Q1,s−1
..
.
Q1,1
,
Ô2s−3,s/2 (:, s − 1) = O2s−3,s/2
(55)
0
..
.
0
while those in (54) can be compactly written as:
0
..
.
0
Ô2s−3,s/2 (:, s − 1) = O2s−3,s/2
Q2R+1,2R+1
..
.
Q2R+1,2R+3−s
(56)
To illustrate this using the specific example shown in (47), the
0
0
Ĉ
0
0
ˆr
CA
ˆ
CA
ˆl
CA
Q1,3
Q1,2
= O5,2
Q1,1
0
0
,
0
0
Ĉ
0
0
ˆr
CA
ˆ
CA
ˆl
CA
0
0
= O5,2
Q5,5
Q
5,4
Q5,3
.
(57)
In view of equations (55) and (56), we can obtain that
Q1,s−1
0
..
..
.
.
Q1,2
0
O2s−3,s/2 Q1,1 = O2s−3,s/2 Q2R+1,2R+1
0
Q2R+1,2R
..
..
.
.
0
Q2R+1,2R+3−s
By assumption 3) of the theorem, the matrix O2s−3,s/2 has
full column rank. As a result, it can be established from the
above equation that
Q1,1 = Q2R+1,2R+1 ;
Q1,2 = · · · = Q1,s−1 = 0;
Q2R+1,2R+3−s = · · · = Q2R+1,2R = 0.
(58)
Since the block vector Ô2s−3,s/2 (:, s − 1) contains all the
block components of Ŵs/2 , we can obtain that
Ŵs/2 = Ws/2 Q1,1 .
Therefore, the proof is completed.
R EFERENCES
[1] B. Bamieh, F. Paganini, and M. Dahleh, “Distributed control of spatially
invariant systems,” Automatic Control, IEEE Transactions on, vol. 47,
pp. 1091–1107, Jul 2002.
[2] N. Motee and A. Jadbabaie, “Optimal control of spatially distributed
systems,” Automatic Control, IEEE Transactions on, vol. 53, pp. 1616–
1629, Aug 2008.
[3] J. Rice and M. Verhaegen, “Distributed control in multiple dimensions:
A structure preserving computational technique,” Automatic Control,
IEEE Transactions on, vol. 56, pp. 516–530, March 2011.
[4] A. Mahajan, N. C. Martins, M. Rotkowitz, and S. Yüksel, “Information
structures in optimal decentralized control.,” in CDC, pp. 1291–1306,
2012.
[5] M. Wozny and W. Carpenter, “A characteristics approach to parameter
identification in distributed systems,” Automatic Control, IEEE Transactions on, vol. 14, pp. 422–422, Aug 1969.
[6] D. Gorinevsky and M. Heaven, “Performance-optimized applied identification of separable distributed-parameter processes,” Automatic Control,
IEEE Transactions on, vol. 46, pp. 1584–1589, Oct 2001.
15
[7] J. Rice and M. Verhaegen, “Efficient system identification of heterogeneous distributed systems via a structure exploiting extended kalman
filter,” Automatic Control, IEEE Transactions on, vol. 56, pp. 1713–
1718, July 2011.
[8] T. Bewley and S. Liu, “Optimal and robust control and estimation of
linear paths to transition,” Journal of Fluid Mechanics, vol. 365, pp. 305–
349, Jun 25 1998.
[9] F. Wu and S. Yildizoglu, “Distributed parameter-dependent modeling
and control of flexible structures,” Journal of Dynamic Systems Measurement and Control-Transactions of The ASME, vol. 127, pp. 230–239,
Jun 2005.
[10] D. G. MacMynowski, R. Heimsten, and T. Andersen, “Distributed Force
Control of Deformable Mirrors,” European Journal of Control, vol. 17,
pp. 249–260, May-Jun 2011.
[11] R. R. Pena, R. D. Fernandez, and R. J. Mantz, “Passivity control via
Power Shaping of a wind turbine in a dispersed network,” International
Journal of Hydrogen Energy, vol. 39, pp. 8846–8851, May 2014.
[12] A. Haber and M. Verhaegen, “Subspace identification of large-scale
interconnected systems,” Automatic Control, IEEE Transactions on,
vol. 59, pp. 2754–2759, Oct 2014.
[13] J. Kulkarni, R. DAndrea, and B. Brandl, “Application of distributed
control techniques to the adaptive secondary mirror of cornells large atacama telescope,” in SPIE Astronomical Telescopes and Instrumentation
Conf., pp. 750–756, 2002.
[14] P. Massioni and M. Verhaegen, “Subspace identification of circulant
systems,” Automatica, vol. 44, no. 11, pp. 2825 – 2833, 2008.
[15] P. Massioni and M. Verhaegen, “Subspace identification of distributed,
decomposable systems,” in Decision and Control, 2009 held jointly
with the 2009 28th Chinese Control Conference. CDC/CCC 2009.
Proceedings of the 48th IEEE Conference on, pp. 3364–3369, Dec 2009.
[16] N. Matni and A. Rantzer, “Low-rank and low-order decompositions for
local system identification,” arXiv:1403.7175, 2014.
[17] M. Verhaegen, “Subspace techniques in system identification,” Encyclopedia of Systems and Control, pp. 1386–1396, 2015.
[18] G. Mercere, O. Prot, and J. Ramos, “Identification of parameterized
gray-box state-space systems: From a black-box linear time-invariant
representation to a structured one,” Automatic Control, IEEE Transactions on, vol. 59, pp. 2873–2885, Nov 2014.
[19] L. Ljung, System Identification: Theory for the user (Second Edition).
Prentice Hall PTR, 1999.
[20] M. Verhaegen and V. Verdult, Filtering and System Identification: A
Least Squares Approach. Cambridge University Press, 2007.
[21] Z. Liu, A. Hansson, and L. Vandenberghe, “Nuclear norm system
identification with missing inputs and outputs,” Systems and Control
Letters, vol. 62, no. 8, pp. 605 – 612, 2013.
[22] M. Verhaegen and A. Hansson, “Nuclear norm subspace identification
(n2sid) for short data batches,” CoRR abs/1401.4273, 2014.
[23] C. Yu and M. Verhaegen, “Subspace identification of distributed clusters
of homogeneous systems,” IEEE Transactions on Automatic Control,
vol. 62, no. 1, pp. 463–468, 2017.
[24] P. Dewilde and A.-J. Van der Veen, Time-varying systems and computations. Springer Science & Business Media, 2013.
[25] K. Mohan and M. Fazel, “Reweighted nuclear norm minimization with
application to system identification,” in American Control Conference
(ACC), 2010, pp. 2953–2959, IEEE, 2010.
[26] L.-L. Xie and L. Ljung, “Estimate physical parameters by black box
modeling,” in Proceedings of the 21st Chinese Control Conference,
pp. 673–677, 2002.
[27] C. Yu, M. Verhaegen, S. Kovalsky, and R. Basri, “Identification of
structured lti mimo state-space models,” in Decision and Control (CDC),
2015 IEEE 54th Annual Conference on, pp. 2737–2742, IEEE, 2015.
[28] G. Strang, “Fast transforms: Banded matrices with banded inverses,”
Proceedings of the National Academy of Sciences of the United States
of America, vol. 107, no. 28, pp. 12413 – 12416, 2009.
[29] S. Choudhary and U. Mitra, “Identifiability scaling laws in bilinear
inverse problems,” arXiv preprint arXiv:1402.2637, 2014.
| 3cs.SY
|
Symmetric Complete Sum-free Sets in Cyclic Groups
Ishay Haviv∗
Dan Levy†
arXiv:1703.04118v2 [math.CO] 1 May 2017
May 2, 2017
Abstract
We present constructions of symmetric complete sum-free sets in general finite cyclic groups.
It is shown that the relative sizes of the sets are dense in [0, 31 ], answering a question of Cameron,
and that the number of those contained in the cyclic group of order n is exponential in n. For
primes p, we provide a full characterization of the symmetric complete sum-free subsets of Z p
of size at least ( 31 − c) · p, where c > 0 is a universal constant.
1
Introduction
This paper studies symmetric complete sum-free subsets of finite cyclic groups. Let G be an
abelian additive group. A set S ⊆ G is symmetric if x ∈ S implies − x ∈ S. It is complete if for
every z ∈ G \ S there exist x, y ∈ S for which x + y = z, and it is sum-free if there are no x, y, z ∈ S
satisfying x + y = z.
Sum-free sets were first studied by Schur [29], who proved in 1916 that the set of positive
integers cannot be partitioned into a finite number of sum-free sets. This result initiated a fruitful
line of research, which now, over a century later, is central in the area of additive combinatorics,
with applications to several areas of mathematics such as extremal graph theory, Ramsey theory,
projective geometry, and model theory (see, e.g., [7, 9, 26, 6]).
A fundamental question on sum-free sets is how large can a sum-free subset of a group G be.
It is easy to see that the size of a largest sum-free subset of Z contained in [1, n] is dn/2e, attained,
for example, by the set of all odd numbers in this range. For the cyclic group G = Z p where p is a
prime, the well-known Cauchy-Davenport theorem [12, 14] yields that every sum-free set has size
at most b( p + 1)/3c. The sets attaining this bound were explicitly characterized in the late sixties
by Yap [31, 32], Diananda and Yap [17], and Rhemtulla and Street [27].
Theorem 1.1 ([31, 17, 32, 27]). Let p be a prime, and let S ⊆ Z p be a sum-free set of maximum size.
1. If p = 3k + 1 for an integer k then |S| = k and S is equal, up to an automorphism, to one of the
following sets:
[k + 1, 2k],
∗ School
† School
[k, 2k − 1], and for k ≥ 4, {k} ∪ [k + 2, 2k − 1] ∪ {2k + 1}.
of Computer Science, The Academic College of Tel Aviv-Yaffo, Tel Aviv 61083, Israel.
of Computer Science, The Academic College of Tel Aviv-Yaffo, Tel Aviv 61083, Israel.
1
(1)
2. If p = 3k + 2 for an integer k then |S| = k + 1 and S is equal, up to an automorphism, to the set
[k + 1, 2k + 1].
Note that for p = 3k + 2 the sum-free subsets of Z p of maximum size are symmetric and complete,
whereas for p = 3k + 1 only sets of the third type in (1) have these properties. For a general finite
abelian group G, the largest size of a sum-free set was determined in 2005 by Green and Ruzsa [22].
Another question of interest is that of counting sum-free sets. Cameron and Erdős conjectured in [11] that the number of sum-free subsets of the integers [1, n] is O(2n/2 ). This conjecture
was confirmed independently by Green [20] and by Sapozhenko [28] (see also [2]), improving on
Alon [1], Calkin [4], and Erdős and Granville (see [11]). Counting sum-free sets was also considered for finite abelian groups. For the cyclic group Z p with p prime, an essentially tight upper
bound of 2 p(1/3+o(1)) was proved by Green and Ruzsa [21], who later extended their result to general finite abelian groups [22]. The task of counting sum-free sets that are maximal with respect to
inclusion was addressed in [3].
1.1
Our Contribution
In this paper we present families of complete sum-free subsets of various sizes of the cyclic group
Zn of order n. Such sets have found several applications in the literature (see, e.g., [7, 23, 9, 8, 6]
and Section 1.2).
The first family, studied in Section 3, consists of subsets of Zn whose size is linear in n. These
sets include a central symmetric interval and additional elements located symmetrically on its
3 n −1
left and right. Specifically, for s ∈ [ n+
4 , 3 ] and a set of integers T ⊆ [0, 2t − 1], where t is an
appropriately chosen parameter, we consider the set ST ⊆ Zn defined as
ST = [n − 2s + 1, 2s − 1] ∪ ±(s + T )
(see Definition 3.1 and Figure 1). We prove sufficient and necessary conditions on the set T for
which the set ST is sum-free and complete (see Lemmas 3.2 and 3.3) and refer to the sets T of
size t satisfying these conditions as t-special (see Section 3.2). We use the analysis of the sets ST to
establish a variety of results, described below.
Firstly, we provide a full characterization of the ‘large’ symmetric complete sum-free subsets
of Z p for primes p. Namely, we show that for every s ≥ ( 13 − c) · p where c > 0 is a universal
constant and p is a sufficiently large prime, the sets ST ⊆ Z p of size s for t-special sets T are, up
to automorphisms, all the symmetric complete sum-free subsets of Z p of size s (see Theorem 3.8).
This significantly generalizes the characterization of the symmetric complete sum-free sets of maximum size derived from Theorem 1.1. Our proof makes use of a result due to Deshouillers and
Lev [16] on the structure of large sum-free subsets of Z p (see Theorem 2.2; see also [25, 15]).
Secondly, we consider counting aspects of our characterization result. It is shown that counting
the symmetric complete sum-free subsets of Z p of a given size reduces to counting t-special sets
for a certain value of t. In the following theorem, g(t) stands for the number of t-special sets (see
Section 3.4).
Theorem 1.2. There exists a constant c > 0 such that for every sufficiently large prime p and every
1 ≤ r ≤ c · p the following holds.
2
1. If p = 3k + 1 for an integer k then the number of symmetric complete sum-free subsets of Z p of size
p −1
k − 2r is 2 · g(3r + 1).
2. If p = 3k + 2 for an integer k then the number of symmetric complete sum-free subsets of Z p of size
p −1
k − 2r + 1 is 2 · g(3r ).
Lastly, we observe that there is an abundance of t-special sets and derive the following.
Theorem 1.3. There exists a constant c > 0 such that for every sufficiently large n there exist at least 2cn
symmetric complete sum-free subsets of Zn .
In Section 4, we present for a general n a family of symmetric complete sum-free subsets of
Zn of various sizes. The construction combines sets ± A, ± B, C, where A is an interval, B is an
arithmetic progression, and C is a symmetric interval (see Figure 2). A key property of these sets,
used to prove completeness and sum-freeness, is that the sums of the elements from A with the
elements from B precisely fill the gaps between the elements of − B.
We use our construction to prove that the relative sizes of the symmetric complete sum-free
subsets of finite cyclic groups are dense in [0, 13 ]. This answers an open question of Cameron [10]
related to a natural probability measure on sum-free sets of positive integers (see [8, 9, 5, 6] and
Section 1.2).
Theorem 1.4. For every 0 ≤ α ≤ 31 and ε > 0, for every sufficiently large integer n there exists a
symmetric complete sum-free set S ⊆ Zn whose size satisfies
α−ε ≤
|S|
≤ α + ε.
n
We also obtain, for a general n, a symmetric complete sum-free subset of Zn of size propor√
tional to n. This matches, up to a multiplicative constant, the smallest possible size of such
a set, and extends a result of Hanson and Seyffarth [23] that holds for cyclic groups Zn where
n = m2 + 5m + 2 for an integer m.
Theorem 1.5. There exists a constant c > 0 such that for every sufficiently large integer n there exists a
√
symmetric complete sum-free subset of Zn of size at most c · n.
1.2
Applications
We gather here several applications of our results.
Random sum-free sets. Cameron introduced in [8] a probability measure on the sum-free subsets of the positive integers. This measure is defined via a process that constructs a random sumfree set R as follows: Go over the positive integers in turn, and join to R, independently with
probability 1/2, every element that cannot be expressed as a sum of two elements that are already
in R.
The question of identifying sets M for which the probability that M contains a random sumfree set is a constant bounded away from 0 was addressed in [8, 9, 5, 6]. For example, it was
shown in [8] that the probability of a random sum-free set to consist entirely of odd numbers is
3
about 0.218 (the exact constant is unknown). Conditioned on this event, the density1 of a random
sum-free set is almost surely 1/4. The intuitive reason is that if all the elements of the set are odd,
then every odd number is unconstrained in the process and is thus joined to the set independently
with probability 1/2.
For a subset S of Zn , let MS be the set of all positive integers that are equal modulo n to
elements from S. Cameron proved in [8] that for every complete sum-free subset S of Zn the
probability that a random sum-free set is contained in MS is nonzero (see also an extension by
Calkin [5]). Conditioned on this event, it can be seen that the density of a random sum-free set is
almost surely |S|/(2n). Cameron asked in [10] (see also [9]) whether the set
n |S|
2n
S is a complete sum-free subset of Zn , n ≥ 1
o
is dense in the interval [0, 61 ]. Our Theorem 1.4 answers this question in the affirmative.
Regular triangle-free graphs with diameter 2. A nice application of symmetric complete sumfree sets of ‘small’ size comes from the following graph theory question: For an integer n, what is
the smallest d for which there exists an n vertex d-regular triangle-free graph of diameter 2? This
question was studied by Hanson and Seyffarth [23], who were motivated by a problem of Erdős,
Hajnal, and Moon [18] on k-saturated graphs. It was observed in [23] that if S is a symmetric
complete sum-free subset of an abelian group G then the Cayley graph associated with G and S is
an n vertex d-regular triangle-free graph of
pdiameter 2 for n = | G | and d = |S|. The completeness
of S easily yields a lower bound of |S| ≥ 2| G | − O(1). As already mentioned, our Theorem 1.5
extends a result of [23] and shows that this lower bound is attained, up to a multiplicative constant,
for every cyclic group Zn (see [13] for a construction for a non-cyclic group).
Dioid partitions of groups. The motivation for the current work has arrived from the concept
of dioid partitions of groups introduced by the authors in [24]. For a group G, written here in
multiplicative notation, a partition Π of G is a dioid partition if it satisfies the following properties:
1. For every π1 , π2 ∈ Π, π1 π2 is a union of parts of Π, where π1 π2 is a setwise product.
2. There exists e ∈ Π satisfying eπ = πe = π for every π ∈ Π.
3. For every π ∈ Π, π −1 ∈ Π, where π −1 = { g−1 | g ∈ π }.
It is shown in [24] that dioid partitions of a group G naturally define algebraic structures known as
dioids (see, e.g., [19]). Dioid partitions of G are provided, for example, by the set of all conjugacy
classes of G and by the set of all double cosets of any subgroup of G. For a cyclic group of prime
order these examples yield only the two ‘trivial’ partitions { G } and {{ g} | g ∈ G }. However, for
a prime p ≥ 5, every symmetric complete sum-free subset S of Z p yields the non-trivial 3-part
dioid partition {{0}, S, (S + S) \ {0}} (see [24]). Our results in the current paper illuminate some
of the structural and counting aspects of these partitions.
1 The
density of a set of positive integers is the limit of the proportion of the numbers in [1, n] that belong to the set,
as n tends to infinity. We assume in this discussion that the density of a random sum-free set exists almost surely and
refer the interested reader to [10] for more details.
4
2
Preliminaries
Symmetry, completeness, and sum-freeness. For an additive group G and sets A, B ⊆ G we
denote A + B = { x + y | x ∈ A, y ∈ B}, − A = {− x | x ∈ A}, and A − B = A + (− B).
We also denote ± A = A ∪ (− A). The set A is said to be symmetric if A = − A. It is complete if
G ⊆ A ∪ ( A + A), and it is sum-free if A ∩ ( A + A) = ∅. Note that each of these three properties
is preserved by any automorphism of G. Notice that A is complete and sum-free if and only if
A + A = G \ A.
The following claim will be used to prove sum-freeness and completeness of sets.
Claim 2.1. Let G be an additive group, let G1 ⊆ G be a set satisfying G = G1 ∪ (− G1 ), and let S ⊆ G be
a symmetric set. Then,
1. If every x, y ∈ G1 ∩ S satisfy x + y ∈
/ S then S is sum-free.
2. If G1 \ S ⊆ S + S then S is complete.
Proof: For the first item, let x, y ∈ S and denote z = x + y. We prove that z ∈
/ S. If x, y ∈ G1
then by assumption z ∈
/ S. If − x, −y ∈ G1 then by assumption −z ∈
/ S, which by symmetry of S
implies z ∈
/ S. Otherwise, without loss of generality, x, −y ∈ G1 . Assume by contradiction that
z ∈ S. If z ∈ G1 then x, which equals the sum of −y and z is not in S, a contradiction. Otherwise,
−z ∈ G1 and thus −y, which equals the sum of x and −z, is not in S, again a contradiction.
For the second item, let z ∈ G \ S. We prove that z ∈ S + S. If z ∈ G1 then by assumption
z ∈ S + S. Otherwise, z ∈ − G1 , thus (−z) ∈ G1 \ S, implying that −z ∈ S + S, which by symmetry
of S implies that z ∈ S + S as well.
Cyclic groups. For an arbitrary integer n ≥ 1, we consider the cyclic group Zn = {0, 1, . . . , n −
1}, with addition modulo n. Every element a ∈ Zn has infinitely many representatives in Z, those
of the form a + m · n for m ∈ Z. For integers a ≤ b we denote [ a, b] = {z ∈ Z | a ≤ z ≤ b}.
We use the same notation for intervals in Zn , viewing the elements of [ a, b] as representatives of
elements of Zn . Note that we use the same addition symbol for both integer addition and addition
modulo n. In case of possible ambiguity, we will explicitly mention what the notation means. For
a prime p, the group automorphisms of Z p are realized as multiplication modulo p by an element
of Z p \ {0}. For d ∈ Z p \ {0} and A ⊆ Z p the set d · A ⊆ Z p is said to be a dilation of A.
Large sum-free sets of cyclic groups of prime order. We state below a theorem of Deshouillers
and Lev [16] on the structure of large sum-free subsets of Z p for a prime p.
Theorem 2.2 ([16]). For every sufficiently large prime p, every sum-free subset S of Z p of size |S| ≥
0.318p is contained in a dilation of the interval [|S|, p − |S|].
Theorem 2.2 improves previous results of Lev [25] and Deshouillers and Freiman [15]. Dainiak
and Sisask independently proved that the constant 0.318 in the theorem cannot get below 0.25
(see, e.g., [30]). Finding the smallest constant for which the statement holds is an interesting open
question.
5
3
Large Symmetric Complete Sum-free Sets
In this section we present and analyze a construction of ‘large’ symmetric complete sum-free subsets of the cyclic group Zn .
3.1
The Sets ST
Consider the following definition.
Definition 3.1. Let n and s be integers for which t = (n − 3s + 1)/2 is a positive integer and n ≤ 4s − 3.
For a set T ⊆ [0, 2t − 1], let ST ⊆ Zn be the set defined as
ST = [n − 2s + 1, 2s − 1] ∪ ±(s + T ).
1
s
n − 2s + 1
n−s
2s − 1
s+T
n−1
−(s + T )
Figure 1: An illustration of the elements of the set ST .
A few remarks are in order. The set ST ⊆ Zn defined above depends on a parameter s and a
set T. The definition requires the parameters n and s to satisfy
n+3
n−1
≤s≤
,
4
3
where the upper bound on s follows from t ≥ 1. Notice that the interval [n − 2s + 1, 2s − 1]
included in ST is symmetric, thus the set ST is symmetric as well. The set s + T includes | T |
elements, which all belong to the interval [s, s + 2t − 1]. By the definition of t we have s + 2t −
1 = n − 2s, hence the elements of s + T are located on the left of the interval [n − 2s + 1, 2s − 1]
(see Figure 1). By symmetry, the elements of −(s + T ) are located on its right. This implies that
ST ⊆ [s, n − s] and that the size of ST is
|ST | = 4s − n − 1 + 2| T |.
In the case of | T | = t, which will be of particular interest for us, we have |ST | = s.
We consider now the question of what conditions on the set T make the set ST a complete
sum-free subset of Zn . The following lemma gives a sufficient and necessary condition on T for
sum-freeness of ST .
Lemma 3.2 (Sum-freeness of ST ). Let n and s be integers for which t = (n − 3s + 1)/2 is a positive
integer and n ≤ 4s − 3, and let T ⊆ [0, 2t − 1] be a set. Then, the set ST ⊆ Zn is sum-free if and only if
2t − 1 ∈
/ T + T + T,
where the addition in (2) is over the integers.
6
(2)
Proof: Denote A = [n − 2s + 1, 2s − 1], and recall that ST = A ∪ ±(s + T ) ⊆ [s, n − s]. Let us
consider the elements of ST + ST . Observe that
A + A = [2n − 4s + 2, 4s − 2] = [0, 4s − n − 2] ∪ [2n − 4s + 2, n − 1],
(3)
where for the second equality we have used n ≤ 4s − 3. By t ≥ 1 we have
4s − n − 2 = s − (n − 3s + 1) − 1 = s − 2t − 1 < s,
hence, using the symmetry of A and ST , it follows that A + A is disjoint from ST .
Denote B = [s, s + 2t − 1], and notice that s + T ⊆ B. The set
A + B = [n − s + 1, 3s + 2t − 2] = [n − s + 1, n − 1] = −[1, s − 1]
is disjoint from ST , because ST is symmetric and has no intersection with [1, s − 1]. This implies
that the sum of A and s + T is disjoint from ST as well.
It remains to consider the sums of pairs of elements in ±(s + T ). The sum of s + T and −(s + T )
is contained in ±[0, 2t − 1], which is disjoint from ST because
2t − 1 < 2t = n − 3s + 1 < s.
We turn to show that the condition 2t − 1 ∈
/ T + T + T holds if and only if the sum of every
two elements in s + T does not belong to ST . This would imply that this condition is equivalent to
the sum-freeness of ST and thus would complete the proof of the lemma.
First, assume that 2t − 1 ∈
/ T + T + T, and let s + `1 and s + `2 be two elements of s + T where
`1 , `2 ∈ T. Set `3 = 2t − 1 − (`1 + `2 ), and note that the assumption 2t − 1 ∈
/ T + T + T implies
that `3 ∈
/ T. The sum of s + `1 and s + `2 is
2s + `1 + `2 = 2s + (2t − 1 − `3 ) = −(n − 2s − 2t + 1 + `3 ) = −(s + `3 ),
(4)
so by symmetry of ST it suffices to show that s + `3 ∈
/ ST . If `3 ≥ 0 then `3 ∈ [0, 2t − 1], implying
s + `3 ∈ B. However, B ∩ ST = s + T, so by `3 ∈
/ T it follows that s + `3 is not in ST , as required.
Otherwise, `3 < 0, and we claim that in this case s + `3 ∈ [0, s − 1]. Indeed, s + `3 < s, and by
`3 ≥ −(2t − 1) we obtain that
s + `3 ≥ s − (2t − 1) = 4s − n > 0.
Since the interval [0, s − 1] is disjoint from ST , it follows that s + `3 is not in ST , and we are done.
For the other direction, assume that 2t − 1 ∈ T + T + T, that is, there exist `1 , `2 , `3 ∈ T such
that `3 = 2t − 1 − (`1 + `2 ). By the definition of ST , the elements s + `1 , s + `2 and s + `3 are in ST .
As in (4), the sum of the first two is −(s + `3 ), which by symmetry of ST does belong to ST . Hence,
the sum of two elements of s + T belongs to ST , as required.
We now consider the completeness of ST , that is, the property that every element of Zn belongs
to either ST or to ST + ST (or both). The following lemma provides a sufficient and necessary
condition on the set T, assuming that s is sufficiently large compared to n. We use here the notation
min( T ) for the smallest integer in T.
7
Lemma 3.3 (Completeness of ST ). Let n and s be integers for which t = (n − 3s + 1)/2 is a positive
integer and n ≤ 7s/2 − 1, and let T ⊆ [0, 2t − 1] be a nonempty set. Then, the set ST ⊆ Zn is complete if
and only if
[0, 2t − 1 + min( T )] \ (2t − 1 − T ) ⊆ T + T,
(5)
where the addition in (5) is over the integers.
Proof: Denote m = min( T ) and A = [n − 2s + 1, 2s − 1] as before. By (3), we have
[0, 4s − n − 2] ⊆ A + A ⊆ ST + ST .
(6)
In addition, we have s + m ∈ s + T ⊆ ST and
(s + m) + A = [n − s + m + 1, 3s + m − 1] = −[n − 3s − m + 1, s − m − 1].
By symmetry of ST this implies that
[n − 3s − m + 1, s − m − 1] ⊆ ST + ST .
(7)
n − 3s − m + 1 ≤ n − 3s + 1 ≤ 4s − n − 1,
(8)
Observe that
where the second inequality follows from the assumption n ≤ 7s/2 − 1. Combining (6), (7) and (8),
we obtain that
[0, s − m − 1] ⊆ ST + ST .
Since the interval A starts at n − 2s + 1, using the symmetry of ST , it follows that
Zn ⊆ ST ∪ (ST + ST ) if and only if [s − m, n − 2s] ⊆ ST ∪ (ST + ST ).
The interval [s − m, n − 2s] does not intersect the following sets:
• A, because n − 2s < n − 2s + 1;
• −(s + T ), because −(s + T ) ⊆ [n − s − (2t − 1), n − s] and
n − 2s < 2s = n − s − (2t − 1);
• A + A, by (3) and the fact that 4s − n − 2 = s − 2t − 1 < s − (2t − 1) ≤ s − m;
• (s + T ) + (s + T ), because this set is contained in [2s, 2s + 4t − 2], and n − 2s < 2s;
• (s + T ) − (s + T ), because this set is contained in ±[0, 2t − m − 1], and
2t − m − 1 = n − 3s − m < s − m;
• (s + T ) + A, because this set is contained in
s + [m, 2t − 1] + A = [s + m, s + 2t − 1] + A = [n − s + m + 1, 3s + 2t − 2],
and n − 2s < n − s + m + 1;
8
• −(s + T ) + A, because this set is contained in
−(s + [m, 2t − 1] + A) = [n − 3s − 2t + 2, s − m − 1],
and s − m − 1 < s − m.
Therefore, the only way for the elements of [s − m, n − 2s] to be in ST ∪ (ST + ST ) is to belong
to either s + T or to
−(s + T ) − (s + T ) = −2s − ( T + T ) = n − 2s − ( T + T ).
The condition
[s − m, n − 2s] ⊆ (s + T ) ∪ (n − 2s − ( T + T ))
is equivalent to
[0, n − 3s + m] ⊆ ( T + m) ∪ (n − 3s + m − ( T + T )),
which by the definition of t is equivalent to
[0, 2t − 1 + m] ⊆ ( T + m) ∪ (2t − 1 + m − ( T + T )),
and the latter, using the fact that [0, a] = a − [0, a], is equivalent to the condition in the lemma, so
we are done.
3.2
Special Sets
We now turn to identify the sets T for which ST is a complete sum-free subset of Zn of size s (see
Definition 3.1). Lemmas 3.2 and 3.3 supply the required conditions on T for sum-freeness and
completeness of ST . Recall that the size of ST is s whenever | T | = t. This leads us to the notion of
t-special sets given below.
Definition 3.4. For an integer t ≥ 1 we say that a set T ⊆ [0, 2t − 1] is t-special if it satisfies
1. | T | = t,
2. 2t − 1 ∈
/ T + T + T, and
3. [0, 2t − 1 + min( T )] \ (2t − 1 − T ) ⊆ T + T.
Note that the addition is over the integers.
The following simple claim shows that in case that 0 ∈ T the first and second conditions in
Definition 3.4 imply the third.
Claim 3.5. For every integer t ≥ 1 and a set T ⊆ [0, 2t − 1] satisfying 0 ∈ T, if | T | = t and 2t − 1 ∈
/
T + T + T then T is t-special.
Proof: Assume that 0 ∈ T, | T | = t and 2t − 1 ∈
/ T + T + T. We claim that for every ` ∈ [0, t − 1]
exactly one of the elements ` and 2t − 1 − ` belongs to T. Indeed, the elements ` and 2t − 1 − `
cannot both belong to T as their sum, together with 0 ∈ T, is 2t − 1, which does not belong to
T + T + T. In addition, if T does not contain one of ` and 2t − 1 − ` for some ` ∈ [0, t − 1] then
its size cannot reach t. It follows that every ` ∈ [0, 2t − 1] for which 2t − 1 − ` ∈
/ T satisfies
` = ` + 0 ∈ T + T. This implies that [0, 2t − 1] \ (2t − 1 − T ) ⊆ T + T, hence T is t-special.
9
Remark 3.6. It is easy to notice that for every t ≥ 1 there exist t-special sets. For example, the set
{0} ∪ [t, 2t − 2] and the set of even elements in [0, 2t − 1] are t-special, as is easy to verify using Claim 3.5.
Combining Lemmas 3.2 and 3.3 with the fact that |ST | = s if and only if | T | = t, we obtain the
following result.
Theorem 3.7. Let n and s be integers for which t = (n − 3s + 1)/2 is a positive integer and n ≤ 7s/2 − 1,
and let T ⊆ [0, 2t − 1] be a set. Then, T is t-special if and only if ST is a complete sum-free subset of Zn of
size s.
3.3
Characterizing the Large Symmetric Complete Sum-free Sets in Z p
Consider the cyclic group Z p for a prime p. We provide a characterization, based on Definitions 3.1
and 3.4, of the symmetric complete sum-free subsets of Z p of a given size s for large values of s. By
‘large’ we mean that s ≥ ( 13 − c) · p for a universal constant c > 0 (the proof will give s ≥ 0.318p).
Recall that the maximum possible size of such a set is b( p + 1)/3c, making the restriction on s
quite natural.
Theorem 3.8. There exists a constant c > 0 such that for every sufficiently large prime p and an integer s,
for which t = ( p − 3s + 1)/2 is a positive integer satisfying t ≤ c · p, the following holds. The symmetric
complete sum-free subsets of Z p of size s are precisely all the dilations of the sets ST for t-special sets
T ⊆ [0, 2t − 1].
Note that Theorem 3.8 characterizes the symmetric complete sum-free subsets of Z p of size s
for every possible size s of such a set satisfying
p(1 − 2c) + 1
p−1
≤s≤
,
3
3
(9)
where c > 0 is the constant in Theorem 3.8. Indeed, the size of every symmetric sum-free subset
of Z p , for a prime p > 2, is even. For an even s satisfying (9) the number t = ( p − 3s + 1)/2 is an
integer with 1 ≤ t ≤ c · p.
The fact that the sets ST for t-special sets T ⊆ [0, 2t − 1] are symmetric complete sum-free
subsets of Z p already follows from Theorem 3.7. This immediately implies that their dilations
have these properties as well. Hence, to complete the proof of Theorem 3.8 it suffices to show that
these are the only sets with those properties. This is the goal of the following lemma.
Lemma 3.9. There exists a constant c > 0 such that for every sufficiently large prime p and an integer s,
for which t = ( p − 3s + 1)/2 is a positive integer satisfying t ≤ c · p, the following holds. Every symmetric
complete sum-free subset of Z p of size s is a dilation of a set ST where T ⊆ [0, 2t − 1] is t-special.
Proof: Let S ⊆ Z p be a symmetric complete sum-free set of size |S| = s. Define c = 0.023 (the
constant satisfying 31 · (1 − 2c) = 0.318). By the definition of t, the assumption t ≤ c · p implies
that
p − 2t + 1
p(1 − 2c) + 1
s=
≥
≥ 0.318p.
3
3
10
Theorem 2.2 implies that the sum-free set S is contained in a dilation of the interval C = [s, p − s].
As the statement of our lemma is invariant under dilations, we can assume that S ⊆ C. Observe
that
C + C = [2s, 2p − 2s] = [0, p − 2s] ∪ [2s, p − 1].
By S ⊆ C we have S + S ⊆ C + C, so the completeness of S implies that Z p \ (C + C ) ⊆ S, that is,
[ p − 2s + 1, 2s − 1] ⊆ S.
Thus we have identified 4s − p − 1 elements of S. The remaining s − (4s − p − 1) = 2t elements
of S belong to
[s, p − 2s] ∪ [2s, p − s],
so by symmetry they are fully defined by the t elements of S in [s, p − 2s]. Denote the elements
of S ∩ [s, p − 2s] by s + T where T ⊆ [0, 2t − 1]. This yields that S is equal to a set ST as given in
Definition 3.1 where | T | = t. Applying Theorem 3.7 we obtain that T is t-special, as desired.
3.3.1
Symmetric Complete Sum-free Sets of Almost Maximum Size
For a prime p, Theorem 3.8 essentially reduces the task of computing the symmetric complete
sum-free subsets of Z p of a given size to that of computing the corresponding t-special sets. We
demonstrate this reduction below and derive an explicit characterization of the symmetric complete sum-free subsets of Z p of ‘almost maximum size’, that is, size smaller by 2 than the maximum
size of such a subset.
For a prime p = 3k + 1 set s = k − 2, and notice that t = ( p − 3s + 1)/2 = 4. It is easy to verify
that the 4-special sets T ⊆ [0, 7] are
{0, 4, 5, 6}, {0, 2, 4, 6}, {0, 3, 5, 6}, {1, 2, 6, 7}.
For every such T this gives us the following symmetric complete sum-free subset of Z p :
((k − 2) + T ) ∪ [k + 6, 2k − 5] ∪ ((2k + 3) − T ).
Assuming that p is sufficiently large, the dilations of these sets are the only symmetric complete
sum-free subsets of Z p of size k − 2.
Similarly, for a prime p = 3k + 2 set s = k − 1, and notice that t = ( p − 3s + 1)/2 = 3. The
3-special sets T ⊆ [0, 5] are
{0, 2, 4}, {0, 3, 4}.
For every such T this gives us the following symmetric complete sum-free subset of Z p :
((k − 1) + T ) ∪ [k + 5, 2k − 3] ∪ ((2k + 3) − T ).
Again, assuming that p is sufficiently large, the dilations of these sets are the only symmetric
complete sum-free subsets of Z p of size k − 1.
11
3.4
Counting the Large Symmetric Complete Sum-free Sets
We now consider the question of estimating the number of symmetric complete sum-free subsets
of Z p of a given size s. For every t ≥ 1 denote the number of t-special sets by g(t). We prove
below Theorem 1.2 that reduces counting the symmetric complete sum-free subsets of Z p of size
s to computing g(t) for an appropriately chosen value of t.
Proof of Theorem 1.2: Let p = 3k + 1 be a sufficiently large prime. For an integer r ≥ 1, denote
s = k − 2r and t = ( p − 3s + 1)/2 = 3r + 1. Assuming that r ≤ c · p for a sufficiently small c,
Theorem 3.8 implies that the symmetric complete sum-free subsets of Z p of size s are the dilations
of the sets ST for the t-special sets T.
We first prove that for every 2 ≤ d ≤ p − 2 and every t-special sets T and T 0 (including the
case T = T 0 ) it holds that ST 6= d · ST 0 . By the symmetry of ST we may assume 2 ≤ d ≤ ( p − 1)/2.
Consider the interval D = [−(s − 1), s − 1], and notice that
| D | = 2s − 1 = 2(k − 2r ) − 1 ≥ 2(k − 2cp) − 1 ≥
2
3
− 5c p ≥ d,
where the last inequality holds for a sufficiently small c. Consider the set
E = d · [ p − 2s + 1, 2s − 1] ⊆ d · ST 0 ,
whose elements form an arithmetic progression of difference d. As D is disjoint from ST , it suffices
to show that D intersects E. Assume by contradiction that D does not include any element of E. By
| D | ≥ d it follows that D does not include any element located between two consecutive elements
of E as well. This implies that [d( p − 2s + 1), d(2s − 1)] is an interval disjoint from D. As its size is
d(4s − p − 2) + 1, it follows that the total number of elements in the group is at least
d(4s − p − 2) + 1 + | D | ≥ 2(4s − p − 2) + 1 + (2s − 1) = 10s − 2p − 4
= 10(k − 2r ) − 2(3k + 1) − 4 = 4k − 20r − 6 ≥
4
3
− 21c p > p,
where for the last two inequalities we again assume that c is sufficiently small. This clearly gives
a contradiction.
Now, for every t-special set T we have to count the ( p − 1)/2 dilations of ST , hence the total
number of symmetric complete sum-free subsets of Z p of size s is ( p − 1)/2 times the number of
(3r + 1)-special sets, as required.
The proof of the second item is essentially identical. Here, for p = 3k + 2 and s = k − 2r + 1
we have t = ( p − 3s + 1)/2 = 3r.
3.4.1
Counting Special Sets
Motivated by Theorem 1.2, we would like to understand the behavior of the number g(t) of tspecial sets as a function of t. While g(t) is trivially bounded from above by (2tt) < 22t , the following claim bounds g(t) from below and shows that it grows exponentially in t.
Claim 3.10. For every integer t ≥ 1, g(t) ≥ 2bt/3c .
12
Proof: For every set I ⊆ [d2t/3e, t − 1] consider the set TI defined as
TI = {0} ∪ I ∪ {2t − 1 − i | d2t/3e ≤ i ≤ t − 1, i ∈
/ I } ∪ [2t − d2t/3e, 2t − 2].
Equivalently, the set TI consists of {0} ∪ I and every element 2t − 1 − i for i ∈ [0, t − 1] with
i∈
/ {0} ∪ I.
As there are 2bt/3c possibilities for the set I, each of which gives a different TI , it suffices to
verify that those sets TI are t-special. The set TI includes 0 and has size t. By Claim 3.5, it suffices
to prove that no three elements of TI have sum 2t − 1. If the three elements are nonzero then their
sum is at least 2t. If exactly one of them is zero, then the sum of the other two cannot equal to
2t − 1 because TI includes exactly one of i and 2t − 1 − i for every i ∈ [0, t − 1]. Finally, if at least
two of the elements are zeros then their sum with the third is at most 2t − 2, and we are done.
It is easy to derive now that the number of symmetric complete sum-free subsets of Zn is
exponential in n, confirming Theorem 1.3.
Proof of Theorem 1.3: For a sufficiently large n, let s be the smallest integer for which n ≤ 7s/2 −
1 and t = (n − 3s + 1)/2 is a positive integer. Notice that for this choice we have bt/3c ≥ c · n for
some universal constant c (any c < 1/42 would suffice). By Theorem 3.7, for every t-special set
T ⊆ [0, 2t − 1] the set ST is a symmetric complete sum-free subset of Zn . By Claim 3.10 there are
at least 2bt/3c ≥ 2cn t-special sets T, and since each of them defines a different ST , we are done.
4
Symmetric Complete Sum-free Sets of Various Sizes
In this section we present a construction of symmetric complete sum-free subsets of the cyclic
group Zn that yields such sets with various sizes, confirming Theorems 1.4 and 1.5.
Let t ≥ 1, d ≥ 2 and k ≥ 4 be integers, and let n be one of the following two integers
4dk + 6t − 11, 4dk + 6t − 14.
(10)
Note that every sufficiently large integer n can be expressed as one of these two numbers for
certain choices of the parameters. Consider the following three subsets of Zn :
• A = [ 21 (d n2 e + t + 1), 12 (d n2 e + t + 1) + d − 2],
n
o
• B = 12 (d n2 e + t + 1) + 2d − 2 + i · d 0 ≤ i ≤ k − 4 , and
• C = [b n2 c − t, d n2 e + t].
Define the set
(n)
St,d,k = (± A) ∪ (± B) ∪ C.
Note that A is an interval of length d − 1 and that B is an arithmetic progression of difference d
that includes k − 3 elements, all located on the right of the elements of A (see Figure 2). The set C
is a symmetric interval in Zn , whose size |C | is 2t + 1 if n is even and 2t + 2 if n is odd. It follows
(n)
from the definition that the set St,d,k is also symmetric in Zn .
13
A
1
n
4
+
d
d
d
d
n
2
t
2
C
2d
2d
n
2
−t
d
d
d
d
−A
+t
−B
B
3n
4
−
t
2
n−1
(n)
Figure 2: An illustration of the elements of the set St,d,k . Some additive constants are omitted.
A simple calculation shows, for the two possible values of n given in (10), that the last element
of B, denoted bl , satisfies
jnk
− t − 2d + 2.
(11)
bl =
2
Indeed, for n = 4dk + 6t − 11 we have dn/2e = 2dk + 3t − 5 and bn/2c = 2dk + 3t − 6, hence
1 l n m
bl =
+ t + 1 + 2d − 2 + (k − 4) · d
2 2
1
=
(2dk + 4t − 4) + 2d − 2 + (k − 4) · d
2
jnk
− t − 2d + 2.
= 2dk + 2t − 2d − 4 =
2
For n = 4dk + 6t − 14 the equality is obtained similarly. It follows that bl < bn/2c − t, hence the
elements of B are all located on the left of those of C (see Figure 2). In particular, we obtain that
the sets ± A, ± B, C, considered modulo n, are pairwise disjoint, implying that
(n)
|St,d,k | = 2| A| + 2| B| + |C | = 2(d + k − 4) + |C |.
(12)
Observe that the sets − A and − B can be explicitly written as
1
3n
• − A = [ 21 (b 3n
2 c − t − 1) − d + 2, 2 (b 2 c − t − 1)] and
n
o
• − B = d n2 e + t + 2d − 2 + i · d 0 ≤ i ≤ k − 4 ,
where for − B we have used (11).
The main result of this section is the following.
Theorem 4.1. For all integers t ≥ 1, d ≥ 2, k ≥ 4 and n ∈ {4dk + 6t − 11, 4dk + 6t − 14} such that
(n)
|C | ≥ d (that is, d ≤ 2t + 1 if n is even and d ≤ 2t + 2 if n is odd), the set St,d,k is a symmetric complete
sum-free subset of Zn .
In order to prove the theorem one has to show that every element of Zn belongs to exactly
(n)
(n)
(n)
one of St,d,k and St,d,k + St,d,k . We start with the following claim that shows that the sum A + B
precisely fills the gaps between the elements of the arithmetic progression − B.
Claim 4.2. The sets − B and A + B are disjoint and their union is
hl n m
i
1 j 3n k
+ t + 2d − 2,
−t−1 −d+1 .
2
2
2
14
Proof: The set − B is an arithmetic progression of difference d and length k − 3 starting at d n2 e +
t + 2d − 2. The set A is a d − 1 length interval, so adding it to the arithmetic progression B results
in a pairwise disjoint union of k − 3 shifts of A. The first element of A + B is
1 l n m
lnm
1 l n m
+t+1 +
+ t + 1 + 2d − 2 =
+ t + 2d − 1,
2 2
2 2
2
i.e., the consecutive to the first element of − B. Hence, A + B precisely consists of the intervals
of length d − 1 that fill the gaps between the elements of − B and one more interval that follows
the last element of − B. Hence, the sets − B and A + B are disjoint, and their union is an interval
starting at the first element of − B. The last element of this interval is the last element of A + B
which equals
1 l n m
2
j n k
1 j 3n k
+t+1 +d−2 +
− t − 2d + 2 =
− t − 1 − d + 1,
2
2
2
2
where we have used (11) and the fact that
1
2
· d n2 e + b n2 c =
1
2
· b 3n
2 c.
In the following claim we consider the set B + C.
Claim 4.3. If |C | ≥ d then the set B + C is
h 1 j 3n k
2
i
− t + 1 + 2d − 2, n − 2d + 2 .
2
Proof: As | B| = k − 3 and C is an interval, B + C is a union of k − 3 intervals, each of which is a
shift of C by an element of B. Since B is an arithmetic progression of difference d, each of these
intervals is a shift by d of the previous interval, and since |C | ≥ d, these shifts overlap, so B + C is
itself an interval. Its first element is
1 l n m
j n k
1 j 3n k
+ t + 1 + 2d − 2 +
−t =
− t + 1 + 2d − 2,
2 2
2
2
2
and the last one, using (11), is
j n k
l n m
− t − 2d + 2 +
+ t = n − 2d + 2,
2
2
so we are done.
We are now ready to prove Theorem 4.1.
(n)
Proof of Theorem 4.1. Denote S = St,d,k = (± A) ∪ (± B) ∪ C. Assume that |C | ≥ d. The proof is
given in the following two lemmas.
Lemma 4.4. The set S ⊆ Zn is complete.
Proof: We shall prove that Zn ⊆ S ∪ (S + S). By Claim 2.1, using the symmetry of S and the fact
that 0 ∈
/ S, it suffices to prove that every element of [b n2 c, n − 1] is in either S or S + S. This is
shown via the following steps:
• [b n2 c, d n2 e + t]: This set is contained in C ⊆ S.
15
• [d n2 e + t + 1, d n2 e + t + 2d − 3]: Observe that this set is equal to A + A ⊆ S + S.
• [d n2 e + t + 2d − 2, 21 (b 3n
2 c − t − 1) − d + 1]: By Claim 4.2, this set is the union of − B and
A + B, and is thus contained in S ∪ (S + S).
1
3n
• [ 12 (b 3n
2 c − t − 1) − d + 2, 2 (b 2 c − t − 1)]: This set is equal to − A ⊆ S.
1
3n
• [ 12 (b 3n
2 c − t + 1), 2 (b 2 c − t + 1) + d + |C | − 3]: Observe that this set is equal to A + C ⊆
S + S.
• [ 12 (b 3n
2 c − t + 1) + d + |C | − 2, n − 2d + 2]: By Claim 4.3, using |C | ≥ d, this interval is
contained in B + C ⊆ S + S.
• [n − 2d + 3, n − d] : For d = 2 this set is empty. For d ≥ 3 it is contained in the interval
[n − 2d + 2, n − d], obtained by adding to A the negative of the first element of B, and is thus
contained in S + S.
• [n − d + 1, n − 1] : Observe that by |C | ≥ d we have
jnk
2
− 2t ≤ n − d + 1,
2
hence this set is contained in [2b n2 c − 2t, n − 1] ⊆ C + C ⊆ S + S.
This completes the proof of the lemma.
Lemma 4.5. The set S ⊆ Zn is sum-free.
Proof: By Claim 2.1, using the symmetry of S and C, it suffices to verify that each of the sets
A + A, A + B, B + B, B + C, A + C, and C + C
has an empty intersection with S. We verify this below:
• The set A + A is equal to [d n2 e + t + 1, d n2 e + t + 2d − 3], hence its elements are located immediately after C and before the first element −bl of − B, so they do not belong to S.
• By Claim 4.2, the set A + B is equal to [d n2 e + t + 2d − 2, 12 (b 3n
2 c − t − 1) − d + 1] \ (− B ),
hence its elements are located after C, before − A, and do not include elements of − B, so
they do not belong to S.
• The set B + B is a sum of an arithmetic progression of difference d and length k − 3 with itself,
hence it forms an arithmetic progression of difference d of length 2k − 7. The first element
of B + B is d n2 e + t + 4d − 3, located on the right of C, and its last element, using (11), is
2b n2 c − 2t − 4d + 4 < n. Hence, it suffices to verify that B + B does not intersect − B and
− A. The first element of − B is d n2 e + t + 2d − 2. Observe that it differs from the first element
of B + B by 2d − 1, which is not divisible by d. Since − B is an arithmetic progression of
difference d, it follows that B + B does not intersect it. Now, let a0 denote the first element of
A. Using (11), the sum of the first and last elements of B is
1 l n m
j n k
1 j 3n k
+ t + 1 + 2d − 2 +
− t − 2d + 2 =
− t + 1 = − a0 + 1.
2 2
2
2
2
16
This implies that the element consecutive to the interval − A = [− a0 − d + 2, − a0 ] lies in
B + B. Since the size of − A is d − 1 and the arithmetic progression B + B has difference d,
we get that B + B does not intersect − A.
• By Claim 4.3, the set B + C is equal to [ 12 (b 3n
2 c − t + 1) + 2d − 2, n − 2d + 2], hence its elements
are located on the right of − A and do not belong to S.
1
n
• The set A + C is equal to [ 21 (b 3n
2 c − t + 1), 2 (3d 2 e + 3t + 1) + d − 2]. It can be verified that
the values of n in (10) satisfy that the right edge of the interval is smaller than n. Hence, its
elements are located on the right of − A and do not belong to S.
• The set C + C is equal to [0, |C | − 1] ∪ [n − |C | + 1, n − 1]. By (10), it is clear that n ≥ 6t + 18,
hence the first element of A satisfies
1 l n m
1
a0 =
+ t + 1 ≥ (4t + 10) = 2t + 5 > |C |.
2 2
2
It follows that the elements of C + C are located on the left of A and on the right of − A, so
they do not belong to S.
This completes the proof of the lemma.
We can now show that for every sufficiently large integer n there exist symmetric complete
sum-free subsets of Zn of various sizes.
Theorem 4.6. There exist constants c1 , c2 , c3 > 0 such that for every sufficiently large integer n there
exists a collection of symmetric complete sum-free subsets of Zn whose sizes form an arithmetic progression
√
√
√
with first element at most c1 · n, difference at most c2 · n, and last element at least n3 − c3 · n.
√
Proof: Let n be a sufficiently large integer. Define d0 to be the unique integer in b nc − {0, 1, 2}
√
√
that satisfies d0 = 1 (mod 3), and notice that n − 3 ≤ d0 ≤ n. Define a = 11 if n is odd and
a = 14 if n is even. As n is sufficiently large, one can write n = 2m − a for a positive integer m. Let
m0 be the integer satisfying m0 = m (mod 2d0 ) and 0 ≤ m0 < 2d0 . The choice of d0 guarantees that
3 divides one of m0 + 2d0 , m0 + 4d0 , m0 + 6d0 . Denote by 3t0 the one divided by 3, and notice that
√
√
2d0 ≤ 3t0 ≤ 8d0 , implying 2 n/3 − 2 ≤ t0 ≤ 8 n/3 and d0 ≤ 2t0 + 1. As m = 3t0 (mod 2d0 ), it
follows that there exists an integer k0 for which m = 2d0 k0 + 3t0 , thus
n = 4d0 k0 + 6t0 − a.
(13)
Using again the assumption that n is sufficiently large, we have t0 ≥ 1, d0 ≥ 2, k0 ≥ 4 and
√
k0 ≤ n/(4d0 ) ≤ n/3. Applying Theorem 4.1 to the integers t0 , d0 , k0 , we get that there exists a
symmetric complete sum-free subset of Zn , whose size s, using (12), satisfies s = 2(d0 + k0 + t0 ) − 7
√
if n is even and s = 2(d0 + k0 + t0 ) − 6 if n is odd. In both cases we have s ≤ c1 · n for a universal
constant c1 > 0.
j
k
Now, define b = k03−4 . For every integer 0 ≤ i ≤ b denote
k i = k0 − 3 · i and ti = t0 + 2d0 · i.
Observe, using (13), that for every such i we have n = 4d0 k i + 6ti − a, k i ≥ k0 − 3b ≥ 4, and
d0 ≤ 2ti + 1. Applying Theorem 4.1 to the integers ti , d0 , k i , we get that there exists a symmetric
17
complete sum-free subset of Zn , whose size, as follows from (12), is s + i · 2(2d0 − 3). Denote
√
r = 2(2d0 − 3), and notice that r ≤ c2 · n for a universal constant c2 > 0. The integers s + i · r, for
0 ≤ i ≤ b, form an arithmetic progression, each of its elements is a size of a symmetric complete
sum-free subset of Zn . To complete the proof, observe that its last element satisfies
s+b·r ≥
√
k0 − 6
1
n
· 2(2d0 − 3) = (4d0 k0 − 24d0 − 6k0 + 36) ≥ − c3 · n
3
3
3
for a universal constant c3 > 0, and we are done.
Theorem 1.5 is an immediate consequence of Theorem 4.6. We turn to derive Theorem 1.4.
Proof of Theorem 1.4: Let 0 ≤ α ≤ 31 and ε > 0 be fixed constants. Let n be a sufficiently large
integer. By Theorem 4.6, for universal constants c1 , c2 , c3 > 0, there exists a collection of symmetric
complete sum-free subsets of Zn whose sizes form an arithmetic progression with first element at
√
√
√
most c1 · n, difference at most c2 · n, and last element at least n3 − c3 · n. Let S be a set in this
collection whose size is closest to α · n. Then, for c = max(c1 , c2 /2, c3 ), we have
α·n−c·
√
n ≤ |S| ≤ α · n + c ·
√
n.
√
Assuming that n is sufficiently large, so that ε ≥ c/ n, it follows that
α−ε ≤
|S|
≤ α + ε,
n
as required.
Acknowledgements
We thank Peter Cameron, Grahame Erskine, David Grynkiewicz, Sarah Hart, Karen Seyffarth and
Olof Sisask for their kind help with the literature and for useful discussions.
References
[1] N. Alon. Independent sets in regular graphs and sum-free subsets of abelian groups. Israel
Journal of Mathematics, 73(2):247–256, 1991.
[2] N. Alon, J. Balogh, R. Morris, and W. Samotij. A refinement of the Cameron-Erdős conjecture.
Proc. London Math. Soc., 108(1):44–72, 2014.
[3] J. Balogh, H. Liu, M. Sharifzadeh, and A. Treglown. A sharp bound on the number of maximal
sum-free sets. Electronic Notes in Discrete Mathematics, 49:57–64, 2015. Preliminary version in
EuroComb 2015.
[4] N. J. Calkin. On the number of sum-free sets. Bull. London Math. Soc., 22(2):141–144, 1990.
[5] N. J. Calkin. On the structure of a random sum-free set of positive integers. Discrete Mathematics, 190(1-3):247–257, 1998.
18
[6] N. J. Calkin and P. J. Cameron. Almost odd random sum-free sets. Combinatorics, Probability
and Computing, 7(1):27–32, 1998.
[7] P. J. Cameron. Four lectures on projective geometry. In C. Baker and L. Batten, editors, Finite
Geometries, volume 103, pages 27–63. New York: Marcel Dekker, 1984.
[8] P. J. Cameron. On the structure of a random sum-free set. Probability Theory and Related Fields,
76(4):523–531, 1987.
[9] P. J. Cameron. Portrait of a typical sum-free set. In C. Whitehead, editor, Surveys in Combinatorics 1987, volume 123, pages 13–42. Cambridge Univ. Press, 1987.
[10] P. J. Cameron. Another sum-free set problem (Peter Cameron’s Blog), 2010. Available at
https://cameroncounts.wordpress.com/2010/03/07/another-sum-free-set-problem/.
[11] P. J. Cameron and P. Erdős. On the number of sets of integers with various properties. In
R. A. Mollin, editor, Number Theory: Proceedings of the First Conference of the Canadian Number
Theory Association, pages 61–79. 1990.
[12] A. Cauchy. Recherches sur les nombres. J. École Polytech, 40:99–116, 1813.
[13] W. E. Clark and J. Pedersen. Sum-free sets in vector spaces over GF (2). Journal of Combinatorial
Theory, Series A, 61(2):222 – 229, 1992.
[14] H. Davenport. On the addition of residue classes. J. London Math. Soc., 10:30–32, 1935.
[15] J.-M. Deshouillers and G. A. Freiman. On sum-free sets modulo p. Funct. Approx. Comment.
Math., 35(1):51–59, 2006.
[16] J.-M. Deshouillers and V. F. Lev. A refined bound for sum-free sets in groups of prime order.
Bull. Lond. Math. Soc., 40(5):863–875, 2008.
[17] P. Diananda and H. Yap. Maximal sum-free sets of elements of finite groups. Proc. Japan Acad.,
45:1–5, 1969.
[18] P. Erdős, A. Hajnal, and J. W. Moon. A problem in graph theory. The American Mathematical
Monthly, 71(10):1107–1110, 1964.
[19] M. Gondran and M. Minoux. Graphs, Dioids and Semirings: New Models and Algorithms. Operations Research/Computer Science Interfaces Series, Springer, 2010.
[20] B. Green. The Cameron-Erdős conjecture. Bull. London Math. Soc., 36(6):769–778, 2004.
[21] B. Green and I. Z. Ruzsa. Counting sumsets and sum-free sets modulo a prime. Studia Scientiarum Mathematicarum Hungarica, 41(3):285–293, 2004.
[22] B. Green and I. Z. Ruzsa. Sum-free sets in abelian groups. Israel Journal of Mathematics,
147(1):157–188, 2005.
[23] D. Hanson and K. Seyffarth. k-saturated graphs of prescribed maximum degree. Congressus
Numerantium, 44:127–138, 1984.
19
[24] I. Haviv and D. Levy. Dioid partitions of groups. 2017. In preparation.
[25] V. F. Lev. Large sum-free sets in Z/pZ. Israel Journal of Mathematics, 154(1):221–233, 2006.
[26] G. Payne. A solution to a problem of Cameron on sum-free complete sets. Journal of Combinatorial Theory, Series A, 70(2):305 – 312, 1995.
[27] A. H. Rhemtulla and A. P. Street. Maximal sum-free sets in finite abelian groups. Bull. Austral.
Math. Soc., 2:289–297, 1970.
[28] A. Sapozhenko. The Cameron-Erdős conjecture. Dokl. Akad. Nauk, 393(6):749–752, 2003.
[29] I. Schur. Über die kongruenz x m + ym ≡ zm (mod p). Jahresbericht der Deutschen MathematikerVereinigung, 25:114–117, 1916.
[30] O. Sisask. Combinatorial properties of large subsets of abelian groups. PhD thesis, University
of Bristol, 2009.
[31] H. P. Yap. The number of maximal sum-free sets in C p . Nanta Math., 2(1):68–71, 1968.
[32] H. P. Yap. Structure of maximal sum-free sets in C p . Acta Arithmetica, 17(1):29–35, 1970.
20
| 4math.GR
|
Local Search Methods for Fast Near Neighbor Search
E.S. Telleza,b , G. Ruizc , E. Chavez∗,d , M. Graffa,b
arXiv:1705.10351v2 [cs.DS] 31 May 2017
a CONACyT
Consejo Nacional de Ciencia y Tecnologı́a, Dirección de Cátedras, Insurgentes
Sur 1582, Crédito Constructor 03940, Ciudad de México, México.
b INFOTEC Centro de Investigación e Innovación en Tecnologı́as de la Información y
Comunicación, Circuito Tecnopolo Sur No 112, Fracc. Tecnopolo Pocitos II, Aguascalientes
20313, México.
c Universidad Michoacana de San Nicolás de Hidalgo, México
d Centro de Investigación Cientı́fica y de Educación Superior de Ensenada, Baja California,
México.
Abstract
Near neighbor search (NNS) has been traditionally addressed from an algorithmic perspective, that is, given a dataset and a distance function, the goal
is to build a data structure along with discarding rules that quickly find the
near neighbor of a query under the given distance function. In this manuscript,
we take another approach. We restate NNS as a combinatorial optimization
problem, being the goal to minimize the distance from the query to the data
set. This cost function is well defined, and the minimum is realized precisely
with the nearest object to the query. Adopting this new view allows the use of
a rich collection of optimization algorithms with a long tradition.
We present three local search algorithms that solve the NNS problem from
the combinatorial optimization point of view. Two of the algorithms can be
described as the hyper-parameter optimization of two known indexing methods,
simplifying their adoption for final users. The third method achieves excellent
search tradeoffs among speed, memory, and quality of the approximation. For
instance, for moderately large dimensions, our indexes can reach above 0.99
recall reviewing less than 1% of the database.
We conducted extensive experimentation with five real-world standard benchmark and five synthetic datasets to prove our claims.
Key words: Near neighbor search, Combinatorial optimization, Approximate
algorithms.
1. Introduction
Nearest neighbor search is a pervasive problem in computer science. The
applications range from pattern recognition to textual and multimedia infor∗ Corresponding
author: E. Chavez, email: [email protected].
Email addresses: [email protected] (E.S. Tellez), [email protected] (G.
Ruiz), [email protected] (E. Chavez), [email protected] (M. Graff)
Preprint submitted to Elsevier
June 1, 2017
mation retrieval, machine learning, compression, biometric identification and
authentication, and bioinformatics [37].
Given a query object q, the nearest neighbor search consists in retrieving
the set k nn(q) of k closer items to the query from a finite set S ⊂ U from
a possibly infinite set U , let n = |S|. The result set of k nn(q) is a ball of
radius rk , B(q, rk ) ⊆ S, containing those items in S being close to q; ties are
arbitrarily broken. Closeness between objects is measured with some criterion,
say a distance function d : U × U → R+ which is usually required to obey the
metric properties: d(u, v) = 0 ⇐⇒ u = v, d(u, v) = d(v, u), and the triangle
inequality |d(u, v) − d(v, w)| ≤ d(u, w) ≤ d(u, v) + d(v, w). A search can be
trivially solved just evaluating all distances from q to each object in S; however,
this approach is not of use when S is large, or whenever the cost of evaluating
the distance function makes this approach prohibitive.
A vector space V of dimension δ is a set of tuples of δ numbers (a subset
of Rδ ). There are many possible distances for vectors, being the Euclidean or
Manhattan distances the most popular. Both are Minkowski distances (Lp )
defined in general as follows:
Lp (u, v) =
δ
X
!1/p
|ui − vi |
p
(1)
i=1
The Euclidean distance is obtained for p = 2, and the Manhattan distance for
p = 1. In the limit where p → ∞, we have the Chebyshev distance L∞ (u, v) =
max1≤i≤δ |ui − vi |.
In other domains, for example in Information Retrieval and Natural Language Processing, TFIDF vectors are traditional representations of documents.
TFIDF vectors use the cosine similarity,
Pδ
simC (u, v) = qP
δ
i=1
i=1
u2i ·
ui vi
qP
δ
i=1
(2)
vi2
Metric methods should use ∠(u, v) = cos−1 simC (u, v), that is, the angle between
vectors. We use this distance in our Documents benchmark, see §4.
Another distance of interest in this manuscript is the Levenshtein or edit
distance, defined between strings of symbols. It computes the number of edit
operations needed to convert one string into another. Edit operations are insertion, deletion, and substitution. For strings u, v of sizes i, j respectively it is
defined as follows:
max(i, j) if min(i, j) = 0
ed(ui−1 , vj )
+1
ed(ui , vj−1 )
+ 1(
ed(ui , vj ) =
min
0 if last(ui ) = last(vj )
ed(ui−1 , vj−1 ) +
1 if last(ui ) 6= last(vj )
2
where last(·) is a function that returns the last symbol of a string. The edit
distance is used for the Wiktionary benchmark in §4. A related distance, the
Longest Common Subsequence (LCS) distance can be defined in terms of ed
just setting to 2 the cost when last(ui ) 6= last(vj ). The effect is to allow only
insertions and deletions.
Indexes for multidimensional data like quad-trees, oct-trees, or kd-trees,
make use of the coordinates to partition either the data or the space in such
a way that items are pruned by checking numeric values per dimension. This
approach is convenient whenever δ remains low since it can achieve logarithmic
times on n when δ is considered a constant value [34]. Partition by dimension
does not work in high dimensions because the algorithms have an exponential
dependency on δ, this is folklore and is often called the curse of dimensionality
in the literature.
On the other hand, metric access methods claim to be independent of the
representation of the objects. They use the distance function to partition the
data. The metric approach is useful, in principle, on datasets of high explicit
dimension and small intrinsic dimension. In high dimensional metric spaces the
distribution of the distance among items is concentrated around the mean, and
hence, the partitions have a lot of overlap. Therefore, the curse of dimensionality appears again, in disguise. This effect has been studied both in theory
and in practice and has been reported in the literature. The net effect of the
curse of dimensionality, either in vector or metric spaces, is a slowdown in the
performance at query time. Even for highly selective queries, it is necessary to
check virtually the entire database.
1.1. Approximate and Probabilistic Indexes
On intrinsically high dimensional datasets, when the curse of dimensionality
prevents the use of exact algorithms, a more promising approach to have a
fast search is the use of approximate and probabilistic algorithms. Trading
accuracy for speed gives roominess in the design of algorithms. For example,
instead requesting exactly the k nearest neighbors, we allow two types of error,
i) Not reporting some relevant results, or ii) Reporting objects not on the query
outcome. Since the results will contain errors, we need some way to measure the
quality of the result set. For this purpose, we define the recall of an approximate
result k nn∗ (q) as
recall(k nn∗ (q)) =
|k nn∗ (q) ∩ k nn(q)|
|k nn(q)|
It is not hard to modify the definition of recall to take into account distance
accuracy (how far the result is from the actual answer). This alternate definition
is of use when comparing results with a large number of ties, e.g. when the
distance is discrete.
There exist a number of generic methods to turn an exact metric index into
an approximate and/or probabilistic index [12, 43, 8, 36]. However, indexes
designed to be approximate scale better with both the intrinsic dimension and
3
the database size [15, 10, 20, 33, 39, 2, 16, 38, 40, 41, 42]. Below, we describe
some relevant examples.
Locality Sensitive Hashing (LSH) is a set of fast approximate searching techniques giving probabilistic guarantees on the quality of the result set, Gionis et
al. [20]. The general idea of an LSH index is to find hashing functions such that
close items share the same bucket with high probability, while distant objects
share the same bucket with low probability.
More formally, a family of hashing functions H = {g1 , g2 , . . . , gh }, gi : U →
{0, 1} is called (p1 , p2 , r1 , r2 )-sensitive if, for any p, q ∈ U ,
— if d(p, q) < r1 then P r[hash(p) = hash(q)] > p1 ,
— if d(p, q) > r2 then P r[hash(p) = hash(q)] < p2 ,
where hash(u) is the concatenation of the output of individual hashing functions
gi following a fixed order, that is, hash(u) = g1 (u)g2 (u) · · · gh (u).
Notice that LSH gives probabilistic guarantees about the maximum distance
between the actual answer and the answer provided by the algorithm, and it
cannot guarantee the quality regarding the recall. A large number of empty
buckets will appear to achieve the desired p1 and p2 values (for potential queries
of U that are far from any element in S). This behavior stress the memory
footprint of the algorithm, because to solve it several hashing functions (a so
called hashing family) are needed. Still, high-quality indexes usually need high
amounts of memory.
LSH is well defined for general metric spaces, but the process of finding
suitable hashing functions gi is not trivial and has to be engineered for each case.
There exist simple LSH functions for the Euclidean (L2 ), Hamming, Jaccard,
and Cosine distances [3, 20]. Also, there exist several approaches extending the
concept of LSH to be data sensitive. In general, the problem consist in selecting
hashing functions suitable for the data and the query set [19].
The Fast Library for Approximate Nearest Neighbors, FLANN [28], is a popular library that implements different indexing strategies and selects the more
suitable one, based on parameters like the search time, the construction time,
and the memory available. With those parameters, together with the parameters of the indexes, a cost function is defined. FLANN builds and tries different
instances of the indexes using optimization techniques to find a local minimum
to the cost function. The instance corresponding to the local minimum will be
the one selected.
The indexing methods used by the FLANN are: Randomized k-d forest,
priority search k-means tree, and the sequential search. The randomized k-d
forest consists of multiple k-d trees searched in parallel [35] where the number
of visited leafs on each tree is limited to control the search cost. The k-means
tree [28] is a data structure computing k-means clustering on the dataset producing k compact regions. At each region, the procedure is repeated recursively,
producing a tree of arity k. The recursion is stopped when regions have less than
k items. The search procedure has a best first strategy, and the algorithm uses
a priority queue to order the paths. The performance is adjusted controlling the
4
expansion factor (number of visited regions). As commented, their parameters
are optimized, and the user only needs to specify the desired recall.
Recently, a set of small and fast indexes have appeared, among the remarkable ones we found CNAPP [42], PP-Index [17], MIF [1], and the quantized
permutations [26, 27]. In these indexes, a similar structure arises: every node is
associated with a set of k nearest references (KNR), where the set of references
is a sample of the database. The similarity between items is hinted by a similarity function over the shared neighbors. In this approach, there is no navigation
and proximity queries are solved using an inverted index. This structural similarity was systematically explored in [11], adding several new indexes to the
list. In the same paper, the authors include experimental comparisons showing
the performance of KNR indexes as compared against near neighbor techniques
based on hashing, in particular, LSH [23, 4], and DSH [19]. The best indexes
in the KNR framework are KNR-Cos and KNR-LCS. Those indexes use the
cosine similarity and the longest common sub-sequence as similarity functions
as proximity predictors, respectively. The experimental evidence suggests that
KNR-Cos surpasses the majority of the state of the art KNR indexes, so we will
use it as the canonical member of the KNR family in our experimental section.
In [21] near neighbors are found using a random walk over a precomputed
table of neighbors. The required preprocessing time and memory for the table
are quadratic, preventing the scalability of the approach for large sets. In [7] the
Cover Trees (CT) are introduced. A CT is composed of levels, where nodes at
some level cover lower levels using a notion of proximity. The SASH data structure [7] is constructed by sampling the database, connecting the non-sampled
elements to their approximate nearest neighbors and repeating in the next level.
In [22], the authors introduced the Rank Cover Trees (RCT) incorporating
ideas from the above algorithms. A tree is built using ordered rank for pruning,
instead of rules derived from distances and the triangle inequality. Node descendants in the tree are obtained using a rank order. Since only rank information is
used for navigation, the number of nodes visited is known a priori. The authors
include proofs, or sketches, for complexity and accuracy bounds.
Malkov et al. [24] introduced the Approximate Proximity Graph (APG), an
index with excellent searching performance. The authors suggest that the speed
and accuracy improvement in APG is the consequence of the incremental construction. They claim it produces a graph with small world properties. APG
can trade speed and accuracy for memory usage. If memory is scarce, speed
and accuracy will also suffer.
Construction and searching are entangled in APG. The construction is incremental and consists of a simple rule: To insert the j-th element, simply
find the (approximate) t-nearest neighbors among the j − 1 elements already
indexed; then, the new item is linked (in both directions) to its t near neighbors. The authors encourage the preservation of links to distant items since
they boost the search performance. Notice that these links naturally arise using
the incremental construction.
Search is also greedy. A starting point is selected and greedily follows the
neighbor minimizing the distance to the query, and repeat until no further
5
Algorithm 1 The search algorithm for APG as described in [25]. We use our
notation N (u) to describe the connected vertices of u, see Section 2 for more
details.
Name: Search algorithm APG
Input: A transition function N , the database S, and the query q, the number
of restarts m
Output: The set res of near neighbors of q.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
Let res be an empty min-queue of fixed size k
Let candidates be an empty min-queue
Let visited be an empty set of object identifiers
for i = 1 to m do
Randomly select c ∈ S \ visited
Append c into visited and (d(q, c), c) into candidates and res
loop
Let (rb , best) be the nearest pair in candidates
Remove best from candidates
Let r be the biggest radius in res, empty res defines r = ∞
if rb > r then
break loop
end if
for u ∈ N (best) do
if u 6∈ visited then
Add u to visited and (d(q, u), u) to candidates and res
end if
end for
end loop
end for
improvement is possible. This simple procedure performs poorly. However, it is
likely to boost the recall by performing m searches from random starting points.
Many restarts slow down the overall searching time, yet the authors claim that
using m = O(log n) restarts achieve recall close to one.
The parameters governing accuracy and speed are the number of starting
points m and t, the number of direct neighbors in the graph. Since it is necessary
to restart many times to achieve high accuracy, then the speed is penalized.
Increasing t could also boost the speed at the expense of more memory.
In a follow-up paper, in [25], the searching procedure is updated as follows.
As the previous algorithm, the search is repeated m times, but now the restarts
are persistent. In other words, the closest open node is selected as the new
restarting point. In other words, a set of visited items and a set of candidates
are kept along the entire search process. At the beginning of each restart, the
algorithm appends a random item to the list of candidates to add diversity to
the search process. Algorithm 1 formally describes this algorithm.
6
2. Proximity Search as an Optimization Problem
Combinatorial optimization can be stated as follows. Given a set of states
Ω and a cost function f : Ω → R, find ω ∈ Ω such that for , f (ω) is optimal.
An optimal state is one that reaches either the minimum [or the maximum]. In
other words, the goal is to obtain ω such that
f (ω) = min f (x) [or, max f (x) in the case of maximization].
x∈Ω
x∈Ω
For our indexing problem we set Ω = S, and for a fixed query q ∈ U we set
f (ω) = d(w, q), and we will seek for the minimum of this function.
It should be clear that f (nn(q)) is minimal, hence optimizing f (·) implies
finding the nearest neihbor of q.
For the optimization problem, we will preserve the following restrictions,
derived from the metric space approach.
• The distance function is the only way to compare any pair of items.
• Since objects are seen as a black box, items (states) are given, and cannot
be generated (for example from averaging).
• The gradient is not defined, i.e., the notion of direction is not well-defined.
• There are no assumptions about the distance function d. e.g. uniqueness
nor convexity are assumed.
• The size of the database can be large, and the intrinsic dimension is high.
The above restrictions narrow the alternatives in the selection of optimization techniques. We cannot use gradient or steepest descent. There are, however, many options based on local search. Those techniques require some binding
to navigate between states (which are objects of the database in our case). In
other words, we need a graph with vertex the objects of the database and with a
suitable definition for the edges. Edge definition is critical for the performance
of the optimization algorithms. The three main components of our approach are
the transition function N , the function to be minimized f , and the minimizing
algorithm A.
Definition 1 (State transitions N ). N : Ω → Ω+ is a function associating
each state u with a set of possible transitions.
Definition 2 (The function to be optimized dq ). Each query q defines a
function to minimize, dq (u) = d(q, u) for u ∈ Ω.
Definition 3 (The minimizing algorithm A). A is an algorithm to minimize dq . It explores Ω using the transitions N to find a state minimizing dq for
a given q.
7
2.1. About A
Combinatorial optimization is a well-established area of computer science.
There are many alternatives to choose from. Many meta-heuristics are discussed for example in citeBurkeKendall2014. We selected local search due to
the properties of our problem statement and restrictions.
(u)
d(u, q)
q
u
d(u’, q) = min {d(v, q) | v ∈
(u)}
u’
Figure 1: An object u and its neighborhood N (u), u0 is the item closer to q in the neighborhood.
May we remind that in the problem (S, d) is a metric database, and N
describes the edges of a graph with the vertices in S. With a proper graph, it is
possible to minimize dq starting from any random vertex u ∈ S. The property
of the edges is to connect a vertex with neighbors locally minimizing the target
cost function at each step. There should be at least one v ∈ N (u) minimizing
dq . Figure 1 illustrates a step of the minimization procedure. At each point u,
we must explore its neighborhood N (u), and try to minimize dq . For example,
let u0 ∈ N (u) be an object such that d(q, u0 ) = min{d(q, v) | v ∈ N (u)∪{u}}, we
move to u0 and repeat the process, until we reach a local minimum where there is
no object closer to q than the current node. While typical metric access methods
will try to discard objects using the triangle inequality, see §1, the optimization
approach will seek to minimize dq . In this process, we will find many local
minima before we reach the global minimum or a good approximation.
We review some prominent meta-heuristics that allow us to use them as part
of our algorithms.
2.1.1. Greedy Search.
Perhaps the simpler working algorithm for local search is the Greedy search
algorithm, it works as follows to solve a nearest neighbor query (nn(q)):
i) Start in random state from Ω.
8
ii) Let s be the current best answer to nn(q). Compute d(q, u) for each u ∈
N (s).
iii) If there exists some u that decreases dq then u becomes the best answer
and repeat step ii). Otherwise stop the search.
iv) Our best guess of nn(q) is the item producing the minimum known value
of dq .
The greedy search will retrieve a local minimum for dq . The quality of the
answer can be improved by using multiple restarts of greedy search; this is a
standard amplification method.
m
800
600
400
200
0
P
P
P
P
m
40
35
30
25
20
15
10
5
=0.99
=0.95
=0.90
=0.80
0.02 0.04 0.06 0.08 0.1
P
P
P
P
P
0.1
(a) P ranges from 0.005 to 0.1
0.2
0.3
=0.99
=0.95
=0.90
=0.80
0.4
0.5
P
(b) P ranges from 0.1 to 0.5
Figure 2: The number of steps m for several desired recall (P ? ) as a function of the expected
recall per step P . Observe how low P values need extremely large m values. This behavior is
consequent with the curse of dimension.
If each search has probability P of finding the exact nn(q), then we can
achieve the desired probability P ? with m independent searches with
m=
log (1 − P ? )
;
log (1 − P )
(3)
this expression supposes independence of N (·), which can be hard to achieve.
Notice that this formula does not predict the closeness of the retrieved item at
any given m; however, it is faster than an exact method solving nn(a).
From the optimization perspective, a low P implies a high number of local
minimums. This issue is probably yet another incarnation of the curse of dimension in the proximity search problem. Figure 2 illustrates the effect. Notice
that our method is practical when
cost = m × hops × |N (·)| + 1 < n,
(4)
where hops is the expected length of the path (number of transitions) from a
random starting point to the local minimum found, and |N (·)| the expected
number of transitions per item. We show in our experimental section that the
number of transitions is low in practice.
9
2.1.2. Tabu Search.
Tabu search is a meta-heuristic allowing to take non-greedy decisions at each
step, i.e., not always decreasing dq at each transition. This strategy recognizes
the existence of local minima and plateaus; the idea is to avoid loops and to
follow the same path twice. Tabu search solves the problem adding memory to
mark already visited states as tabu, and then ignoring them on later stages of the
search process. We will store visited states in short-term, medium-term, or longterm memory. The desired behavior, along with the problem, and the available
resources, determine the kind of memory used in a given implementation of tabu
search.
2.1.3. Beam Search.
Beam search is a strategy that optimizes dq using a beam of states B ⊂ Ω of
size b. At each step, beam search computes B 0 = ∪s∈B N (s) and selects the best
b from B and B 0 . The search stops whenever is impossible to improve B. Tabu
search can be used along with beam search to avoid non-essential computations.
At the beginning of the search process, B is populated randomly from Ω. Notice
that the low memory requirement is the main advantage especially for large
datasets or systems with limited memory.
2.2. About N
A fundamental component of the local search meta-heuristic is the transition
function N , if properly defined, it can ensure the convergence to the global minimum or at least to a suitable local minimum. To give an idea of the complexity
of setting a good N , lets think in A be the Greedy Search algorithm. The
function N defines a graph where the vertices are the objects of the database,
and a vertex u is connected with the elements of N (u). Note that the greedy
search starting at a point u will always stay in the path connected component
of u. Therefore, a natural requirement for the graph is to be path connected.
Since our approach is approximate, then our transition function does not
need to ensure that A finds the global optimum dq . From the similarity search
perspective, we allow to lose some results, and also, some retrieved items will
not be part of the exact query. Nonetheless, in practice, the error rate will be
low, and the performance gain is significant. Those claims are supported with
experimentally.
The proposal of Malkov et al. [24, 25] introduce a working N construction
for its APG, already explained in §1. The underlying graph N can be used with
improved versions of greedy search and best-first search. Our contribution is
more about A. As the APG, we will use the undirected version of the graph.
3. Direct Improvements on APG
As detailed on Section 1, search in APG is repeated m times to improve
the expected recall. Authors suggest in [25], that m must be adjusted for each
dataset. This requirement supposes an additional parameter selection, which
10
Algorithm 2 The search algorithm for APG*. The main difference is that m
is not longer necessary, yet a small value σ is needed to guess a good number
of restarts. The main idea is to apply the search procedure repeatedly while
the covering radius of res improves. In practice, it produces a competitive local
minimum.
Name: Search algorithm APG*
Input: A transition function N , the database Ω = S, the query k − nn(q), a
minimum number of repeats σ (pretty small, e.g., 4)
Output: The set res of near neighbor of q.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
Let res be an empty min-queue of fixed size k
Let candidates be an empty min-queue
Let visited be an empty set of object identifiers
repeat
Let cov∗ ← cov(res)
for i = 1 to σ do
{This for loop is identical to that of APG (Algorithm 1)}
Randomly select c ∈ S \ visited
Append c into visited and (d(q, c), c) into candidates and res
loop
Let (rb , best) be the nearest pair in candidates
Remove best from candidates
if rb > cov(res) then
break loop
end if
for u ∈ N (best) do
if u 6∈ visited then
Add u to visited and (d(q, u), u) to candidates and res
end if
end for
end loop
end for
until cov∗ = cov(res) {Stops when there is no improvement over dq , i.e.,
cov(res)}
can be difficult because the best m ideally should be adapted as n grows. A
significant overestimation of m impact the construction time because it is incremental, based on searches. Searching is also negatively affected because of the
number of restarts.
Before we introduce our new algorithms, we will define an important function. Let, result be a min priority queue (min-queue) of fixed size k. The
cov(result) function is defined as follows: if result is under its full capacity
then it evaluates to the maximum possible distance value; if it is full, then
cov(result) is the radius of the furthest item.
11
Our first index contribution is Algorithm 2 (APG*). The objective is to
remove parameter m, checking the changes over the global result set, via the
covering radius cov(res). In some sense, we are minimizing dq , and stopping
the algorithm when our guess of dq cannot be improved after σ tries. All our
new algorithms will use the same technique. Notice that σ should be small (in
the range of 2 to 4). The idea is to relax the stopping condition, avoiding the
introduction of an extra parameter. Larger values of σ should be avoided in
favor of other parameters.
Algorithm 3 The search algorithm for APG*-R. As APG*, it tries to adjust
the number of restarts automatically.
Name: Search algorithm APG*-R
Input: A transition function N , the database Ω = S, the query k − nn(q), a
minimum number of repeats σ (pretty small, e.g., 4),
Output: The set res of near neighbor of q.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
Let res be an empty min-queue of fixed size k
Let visited ← ∅
repeat
Let cov∗ ← cov(res)
for i = 1 to σ do
Let res∗ be an empty min-queue of fixed size k
Randomly select sp ∈ S \ visited
Append sp into visited and (d(q, sp), sp) into res∗
Let evaluated ← ∅
repeat
for v ∈ N (sp) do
if v 6∈ evaluated then
Add (d(q, v), v) into res∗
end if
end for
Update sp to be the closer item in res∗ not in visited. If there is no
item satisfying the condition, then sp is set as undefined.
Add sp into visited
until sp is undefined
res ← res ∪ res∗ {The properties of the min-queue of fixed size should
be preserved}
end for
until cov∗ = cov(res) {Stops when dq does not improve, i.e., cov(res) stays
unchanged}
Our second contributed index is named APG*-R (Algorithm 3). The idea is
that APG and APG* can get stuck on a local minimum since they are limited
to use always the best-known candidate. APG*-R share structural similarities
with APG and APG*; however, there is a conceptual difference. On the one
12
hand, APG*-R uses random starting points, not the best known in APG, and
also, res∗ is local to each step. We also added a local memory evaluated to
avoid additional evaluations, this scheme is different to that found whenever
the candidates set has a global scope, and so, the improvement policy is also
global.
Algorithm 4 Beam Search over Ω and N .
Name: Beam Search
Input: A transition function N , the database S, the query q, the size of the
beam b
Output: The set res of near neighbor of q.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
Let res be an empty min-queue of fixed size k
Let beam be an empty min-queue of fixed size b
Set visited ← {}
Set beam ← {}
for i = 1 to b do
Randomly select u from S
Add u to visited and (d(u, q), u) to res and beam
end for
repeat
Let cov∗ ← cov(res)
for i = 1 to σ do
Let beam∗ ← {} {Fixed sized priority min-queue of size b as beam}
for c ∈ beam do
for u ∈ N (c) do
if u 6∈ visited then
Add u to visited and (d(q, u), u) to res and beam∗
end if
end for
end for
beam ← beam∗
end for
until cov∗ = cov(res) {Stops when there is no improvement over dq , i.e.,
cov(res)}
Our final index is BS (Algorithm 4), an index using beam search over S
and N . As in our previous approaches, we ensure at least σ steps to take the
decision to stop the search. So, if after some minimization steps, dq does not
change, then we achieved a (good) local minimum. The critical parameter is
the size of the beam b, which can be adjusted at any time. Notice that beam
search does not need to restart; however, it needs to know b. As we will see
in our experimental section, the use of beam search as an alternative to other
schemes produce significant improvements in both recall and search speed.
13
dimension
3 × n5
106
3 × 106
16
32
64
0.035s
0.050s
0.076s
0.137s
0.202s
0.342s
0.320s
0.438s
0.708s
Table 1: Time spent on a sequential scan.
4. Experimental Results
In this section, we compare the experimental performance of our technique
against APG, KNR, and the RCT. We fixed our attention in the recall and
search speed. All experiments were performed on a 24-core Intel Xeon 2.60
GHz workstation with 256GB of RAM, running CentOS 7. We do not use
the multiprocessing capabilities in the search process. Both, the index and the
database were maintained in memory.
Our benchmarks consist of both synthetic and real databases. In particular,
we selected the following databases:
— Nasa. This database is a collection of 40, 150 vectors of 20 coordinates obtained from the SISAP project (http://www.sisap.org). It uses L2 as
distance function. A sequential search completes on 0.005 seconds on our
testing machine.
— Colors. The second benchmark is a set of 112, 682 color histograms (112dimensional vectors) from SISAP, under the L2 distance. A sequential
search uses 0.053 seconds.
— Wiktionary. A third benchmark is an English dictionary with 736, 639 entries
with Levenshtein’s distance as metric.1 A sequential search completes in
0.700 seconds.
— Documents. This database is a collection of 25,157 short news articles from
the TREC-3 collection of the Wall Street Journal 1987-1989. We use their
tf-idf vectors, taken from the SISAP project [18], www.sisap.org. We
use the angle between the vectors as the distance measure [5]. We remove
100 random documents from the collection and use them as queries (thus
these 100 documents are not indexed in the database). The objects are
vectors of hundred thousands coordinates. This dataset has a very high
intrinsic dimension in the sense of Chavez et al. [13]. Most exact metric
indexes will need to evaluate the entire dataset to solve nearest neighbor
queries. As a reference, a sequential scan needs 0.185 seconds.
1 Available
at https://www.wiktionary.org/
14
— BigANN-1M. A one million subset of the TEXMEX corpus.2 The original
dataset contains a billion SIFT points (128 coordinates). We also extract
256 queries from the standard query set (not in the dataset). In average,
a sequential search of the nn under L2 needs 0.441 seconds to be solved.
— RVEC. On the one hand, for measuring the effect of varying dimension, we
used datasets of 16, 32, 64, 128 and 256 dimensions, each one with 105
randomly generated items. A query is solved by exhaustive search in
0.015, 0.014, 0.024, 0.044, and 0.081 seconds, respectively. On the other
hand, the performance as the size increases is measured using 16, 32 and
64-dimensional datasets of size 3×105 , 106 , and 3×106 items each. Table 4
shows the search time spent by an exhaustive evaluation of the dataset. All
these datasets use the L2 distance. The importance of synthetic data with
a uniform distribution is to study the the performance of the algorithms
for a fixed size and dimensionality.
To compare against LSH we used the E2 LSH tool3 by Alex Andoni. It automatically optimizes an LSH index for the available memory. We fix the index
memory to be 1, 3, and 5GB, which is many times higher than the necessary
by other indexes. To be fair, since LSH offers distance-based rather than recallbased guarantees, we used it to search by radius (i.e., return any element within
distance r to the query), setting r as the average radius of the 30-th nearest
neighbor. To avoid extreme or uninteresting cases, we removed queries yielding empty answers and those with more than 1000 answers. For the RCT [22]
we used the C++ implementation, kindly shared by the authors. Notice that
LSH, RCT, and FLANN do not offer an implementation for Levenshtein and
the cosine similarity, so they only are compared on datasets with the L2 metric.
Therefore, we include LSH and RCT only on experiments over vector datasets
fulfilling this requirement. Also, LSH is not present in RVEC benchmarks since
its performance is not preserved under this type of datasets [23, 4].
While our implementations are written in C# and run under the Mono
virtual machine, E2 LSH, RCT, and FLANN are written in C++, and run in
native form. For example, their sequential search has an advantage of around
2 to 3X in the case of E2 LSH and RCT, and 5X times in the case of FLANN
due to their code’s optimizations. To give a full picture of the performance,
we report both the absolute times and speedup. The latter is computed on the
sequential scan in its implementation. We also give the actual times to give the
full context to the reader.
4.1. Performance per Database
In the first batch of experiments, we present the indexes APG*, APG*-R,
BS, with different values for t (the number of neighbors in the construction) for
the Colors, Nasa, Wikionary, Documents, and BigANN-1M databases.
2 Available
at http://http://corpus-texmex.irisa.fr/
~andoni/LSH/
3 http://www.mit.edu/
15
The APG, and BS are labeled with t = 8, 16, and 32, for each value of t,
the parameter m for APG and b for BS varies from 8, 16, and 32. The APG*
and APG*-R takes t = 8, 16, and 32. Each KNR uses k = 7, 12 near references
among 2048 total references (larger values of k will produce better recalls but
at the cost of higher memory consumption). For each k, the index is allowed
to review at most 3% and 10% of the database (these will be the two points in
each curve). For the RCT we used a covering factor of 32, 64, and 128 (larger
values will produce better recalls but slower searches).
On Figure 3(a) we show the recall and speedup of the methods for the Colors
database. The APG’s for t = 8 are the fastest but with less recall, as we increase
the number of neighbors in the construction, the recall gets better, but the speed
is affected. When we compare this with the line of the APG*, the APG* is very
close to the corresponding APG. Now, the APG*-R achieves a higher recall
but at the cost of losing speed. Next, we have the BS, it has the best overall
performance, that is, for a given recall, it has the best speed. Note how it gets
the most recall. The RCT, KNR, and LSH are left behind.
The results of the Nasa database are shown on Figure 3(b). This figure
corresponds to the smallest database among our benchmarks, and all the indexes
reach competitive recall. The results for the APG are similar to the Colors
database, and the recall increases with the amount of memory; however, we can
find faster alternatives. The APG* is still very close the corresponding APG.
Note that APG*-R achieve a significant improvement as compared with APG
and APG*. All the BS perform with high recall at a competitive search speed.
The RCT has more appeal here than on Colors being the faster index. While
KNR is quite fast, it gets stuck and has difficulties to improve its recall just
adjusting the k parameter; this behavior is shared with LSH since their recall
values are below others.
The Wiktionary works with a costly distance function, in this type of datasets
it is critical to lower the number of computed distances. On Figure 3(c) almost
every index is hundreds of times faster than a sequential scan. The performance
of the APG is again below the others and the APG* is again essentially the
same. On the other hand, the APG*-R is above all others. For the correct
amount of neighbors, the APG*-R is an absolute improvement of the APG*.
The BS excels in performance, and it reaches the best recall. KNR achieve its
maximum performance rapidly, and it is left behind for all local search methods.
The Figure 3(d) shows the Documents database. Note how most of the
indexes have the same curve of tendency. Remember that this database has
high intrinsic dimension, so, we can have a high recall or a high search speed,
but it is quite difficult to be good on both. The APG* is practically the same as
the APG, however without hassle about the m parameter. KNR indexes exhibit
a good trade between recall and search speed, but APG*-R reaches better recall
values. As before, BS shows the best performances among the alternatives.
The results for the BigANN-1M are shown on Figure 3(e). The results are
striking similar to the Documents experiment. The principal difference is the
higher values for the speedup because BigANN-1M is much larger than Documents. Apart from that, the APG are the same as the APG*; the APG*-R
16
have higher recalls values; the BS instances have higher recalls at the cost of
the speedup; the KNR, RCT and the LSH have below the average speed.
In summary, APG*-R is better in general than the APG* which is similar
to APG but computing m online. Also, all the BS have good recall and speed,
making it the rule of thumb for application needing high quality results.
4.2. The effect of the dimension on the Search Performance
We compared indexes to see how dimensionality of datasets affects them.
Here, we selected RVEC datasets with a wide range of dimensions, specifically,
16, 32, 64, 128, and 256. We measured recall, the search cost, and the speed up.
Each line of points represents the results for each dimension.
The Figure 4 shows the cost of the indexes. We compared the recall with the
cost of the searches. Figure 4(a) shows the indexes using 1 to 10 integers per
element of the database, that is, on the APG the number of neighbors is small.
The APG and the APG* are similar and produce low recall. This performance
contrasts with the APG*-R that shows a significant improvement on it. These
results demonstrate the effectiveness of this variant. Moreover, the BS has a
broad range of action, note how the recall is much better than the APG at
the cost of computing more distances. Here, the KNR has a good performance
with one of the configurations reaching the higher recall and showing an almost
constant cost on the dimension. The RCT, like the BS, is all spread over but
with more distances computed.
Figure 5(a) shows the speedup. The APG and APG* are very fast because
of the small number of distances they have to compute but the low recall makes
them useless. The APG*-R appears in the middle of the other indexes. The
different instances of the BS show how it gains recall but loses speed when we
increase the beam. The KNR gets good recalls but is slower than the average.
Finally, the RCT is fast and with good recall for these smaller indexes. Note
how the recall for the APG’s and BS increases when adding more memory.
For Figure 4(b) we let the methods take more neighbors for the elements;
they compute more distances than smaller indexes, but they reach higher recall
values. The APG and APG* cannot get a good recall but the improve of the
APG*-R makes them comparable to the BS. Figure 5(b) contains the speeds
of those indexes. The PG and APG* are very fast, although with a low recall.
The APG*-R and BS with b = 8 are very similar, one a little faster and the
other with better recall. As expected, the other configurations of BS have higher
recalls.
Figures 4(c) and 4(d) are very similar; they show how the APG and APG*
become to get a good recall for the first dimensions but continue to have a low
value for the higher ones. With the APG*-R and BS, we see that they can have
very high recalls even for the big dimensions.
Figures 5(c) and 5(d) are again very similar. APG and APG* are not very
fast. APG*-R is faster for dimensions 16, 32, and 64 but the BS with b = 8
becomes faster for the larger dimensions.
17
90.0
60.0
50.0
40.0
30.0
16.0
14.0
12.0
10.0
8.0
6.0
20.0
4.0
10.0
0.0
APG t=8
APG t=16
APG t=32
APG*
APG*-R
BS t=8
BS t=16
BS t=32
KNR-Cos k=7
KNR-Cos k=12
RCT
LSH
FLANN
18.0
speedup
70.0
speedup
20.0
APG t=8
APG t=16
APG t=32
APG*
APG*-R
BS t=8
BS t=16
BS t=32
KNR-Cos k=7
KNR-Cos k=12
RCT
LSH
FLANN
80.0
2.0
0.60.7
0.8
0.9
0.95
0.0
0.96
0.60.7
0.8
0.9
0.95 0.96 0.97
recall
(a) Colors
0.99
35.0
APG t=8
APG t=16
APG t=32
APG*
APG*-R
BS t=8
BS t=16
BS t=32
KNR-Cos k=7
KNR-Cos k=12
300.0
250.0
200.0
APG t=8
APG t=16
APG t=32
APG*
APG*-R
BS t=8
BS t=16
BS t=32
KNR-Cos k=7
KNR-Cos k=12
30.0
25.0
speedup
350.0
20.0
15.0
150.0
10.0
100.0
5.0
50.0
0.9
0.95 0.96
0.97
0.98
0.0
0.99
0.60.7
0.8
0.9
0.95 0.96 0.97
recall
0.98
recall
(c) Wiktionary
(d) Documents
250.0
APG t=8
APG t=16
APG t=32
APG*
APG*-R
BS t=8
BS t=16
BS t=32
KNR-Cos k=7
KNR-Cos k=12
RCT
LSH
FLANN
200.0
speedup
0.0
0.999
(b) Nasa
400.0
speedup
0.98
recall
150.0
100.0
50.0
0.0
0.60.7
0.8
0.9
0.95 0.96 0.97
0.98
0.99
0.999
recall
(e) BigANN-1M
Figure 3: Recall and speedup performance for our real world datasets.
18
0.99
0.18
0.12
0.10
0.08
0.06
APG m=8
APG m=16
APG m=32
APG*
APG-R*
BS b=8
BS b=16
BS b=32
0.25
0.20
# distances / n
0.14
# distances / n
0.30
APG m=8
APG m=16
APG m=32
APG*
APG-R*
BS b=8
BS b=16
BS b=32
KNR-Cos k=7
KNR-Cos k=12
RCT c=32
RCT c=64
RCT c=128
0.16
0.15
0.10
0.04
0.05
0.02
0.00
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0.00
1
0.6 0.7
0.8
recall
(a) 1 to 10 integers
0.95
0.97
0.99
(b) 10 to 30 integers
0.50
0.60
APG m=8
APG m=16
APG m=32
APG*
APG-R*
BS b=8
BS b=16
BS b=32
0.40
0.35
0.30
APG m=8
APG m=16
APG m=32
APG*
APG-R*
BS b=8
BS b=16
BS b=32
0.50
0.40
# distances / n
0.45
# distances / n
0.9
recall
0.25
0.20
0.30
0.20
0.15
0.10
0.10
0.05
0.00
0.6 0.7
0.8
0.9
0.95
0.97
0.00
0.990.999
recall
0.6 0.7
0.8
0.9
0.95
0.97
0.990.999
recall
(c) 30 to 100
(d) 100 to 300 integers
Figure 4: Performance comparison (recall vs search cost) among our indexes and several state
of the art techniques for fixed n = 105 over several dimensions. Each curve corresponds
contains points that characterize the performance in a dimension, e.g., 16, 32, 64, 128, and
256; as a hint to review figures, large dimensions correspond to smaller recalls and scale is
exponential.
19
30
20
15
10
10
8
6
4
5
0
APG m=8
APG m=16
APG m=32
APG*
APG-R*
BS b=8
BS b=16
BS b=32
12
speedup
25
speedup
14
APG m=8
APG m=16
APG m=32
APG*
APG-R*
BS b=8
BS b=16
BS b=32
KNR-Cos k=7
KNR-Cos k=12
RCT c=32
RCT c=64
RCT c=128
2
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0
1
0.6 0.7
0.8
recall
(a) 1 to 10 integers
9
0.95
0.97
0.99
(b) 10 to 30 integers
7
APG m=8
APG m=16
APG m=32
APG*
APG-R*
BS b=8
BS b=16
BS b=32
7
6
APG m=8
APG m=16
APG m=32
APG*
APG-R*
BS b=8
BS b=16
BS b=32
6
5
speedup
8
speedup
0.9
recall
5
4
3
4
2
3
1
2
1
0.6 0.7
0.8
0.9
0.95
0.97
0
0.990.999
recall
0.6
0.7
0.8
0.9
0.95
0.97
0.99 0.999
recall
(c) 30 to 100 integers
(d) 100 to 300 integers
Figure 5: Recall and speedup comparison in four different classes of memory among our
indexes and several state of the art techniques for fixed n = 105 over several dimensions.
Each point in curves corresponds to a different dimension 16, 32, 64, 128, and 256; as a hint to
review figures, large dimensions correspond to smaller recalls and scale is exponential.
20
4.3. Scalability
A central property of any search index is its capacity to scale with the size of
the database. Even a static index can be dynamized using the Bentley and Saxe
[6] approach, as it has been shown for metric indexes in [29]. This dynamization
comes with a small penalization at search time, in addition to a constant factor introduced in the construction. This technique is beyond the scope of this
document, and static indexes are a single large index without additional search
overheads. Notice that the APG’s and our BS has an incremental construction.
Hence they can handle a growing database naturally. In any case, most indexes
will be affected by database growth, and we expect the performance to be different in this setup. We used the randomly generated database of dimensions
16, 32, and 64, each with sizes 3 × 105 , 106 , and 3 × 106 . Each point in a curve
is produced by a different database size; speedups are larger as n increases. As
the database grows, so does the size of the indexes and it could not fit in main
memory. Because of that, we set the number of initial neighbors to 8 for the
APG’s and BS.
Figure 6(a) shows the recall and the search speed for the different sizes of
the database for 16-dimension. The APG and APG* scored a recall below 0.7
on all the sizes, even when using 32 searches for the query. As shown in the
previous experiments, the APG*-R gets a boost on the recall compared to the
APG*. Note how the lines of the BS are practically vertical meaning that the
index gains in speed much more than it loses in the recall. The RCT does not
get a recall as good as the BS and is considerably slower. On this and the other
dimensions, the sizes corresponding to the lines of each method are 3 × 105 , 106 ,
and 3 × 106 from right to left except for the KNR. For this method, the recall
gets better as the size of the database increases. In Figure 6(b) we see the recall
and the number of distances for the queries. This general behavior is the same
in the next figures.
Figure 6(c) shows the same tendency: the APG*-R has been an improvement
over the APG’s, the BS having very competitive recalls and speeds, the KNR
been slow, and the RCT behind the BS. The Figure 6(d) shows the same but
with the distances computed. The same can be said to the figures 6(e) and
6(f), but the interesting part is to note how the change of dimension affects
more the indexes than the size of the database for reaching high recalls. For
16 dimensions, high recall values can be achieved for all the sizes; for 32 and
64-dimension, the BS always gets better recalls.
As a summary, we arrange the results for the methods with best recall values
on Tables 2, 3, and 4 for 16, 32, and 64 dimensions respectively.
5. Tuning APG*, APG*-R, and BS
The flexibility of the combinatorial optimization framework for proximity
search is one of the sources of their great potential, but also may be problematic for an end-user. In this section, we provide a simple guide for finding a
competitive tuning, although it should be noticed that the optimal selection of
the parameters cannot be guaranteed.
21
Table 2: The achieved performance for several sizes of the random database of dimension 16.
Index
APG*-R
BS b=8
BS b=16
BS b=32
KNR-Cos k=7
KNR-Cos k=12
RCT
105
Cost Recall
1868.84
0.95
2063.57
0.96
3137.97
0.98
4937.93
0.99
4553.08
0.85
4950.76
0.92
3 × 105
Cost Recall
3561.40
0.97
2297.48
0.95
3566.05
0.98
5621.50
0.99
9998.99
0.92
10948.57
0.96
15836.4
0.95
106
Cost
4002.21
2473.05
3925.69
6200.69
28686.36
31941.04
26863.9
Recall
0.96
0.95
0.98
0.99
0.95
0.98
0.94
3 × 106
Cost Recall
4208.77
0.96
2705.92
0.94
4217.09
0.97
6637.78
0.98
83004.65
0.97
91945.43
0.99
42947
0.94
Table 3: The achieved performance for several sizes of the random database of dimension 32.
Index
APG*-R
BS b=8
BS b=16
BS b=32
KNR-Cos k=12
RCT
105
Cost Recall
4484.40
0.86
4405.64
0.87
6565.40
0.94
9450.75
0.97
5105.69
0.65
3 × 105
Cost Recall
9276.67
0.88
5529.78
0.83
8105.05
0.91
12122.14
0.95
11100.57
0.71
19229.8
0.71
106
Cost
12865.79
6480.63
10078.76
15067.28
32088.94
34085.0
Recall
0.82
0.78
0.88
0.94
0.78
0.67
3 × 106
Cost Recall
14168.41
0.82
7309.23
0.73
11446.03
0.84
17256.64
0.91
92088.18
0.83
52872.1
0.63
Table 4: The achieved performance for several sizes of the random database of dimension 64.
Index
APG*-R
BS b=8
BS b=16
BS b=32
105
Cost
10357.14
8181.72
11692.90
16510.76
Recall
0.69
0.71
0.84
0.91
3 × 105
Cost Recall
21958.30
0.74
11582.28
0.62
16907.20
0.76
24060.58
0.86
22
106
Cost
35579.05
16384.80
24275.12
35590.51
Recall
0.61
0.50
0.66
0.79
3 × 106
Cost Recall
51994.29
0.57
22398.66
0.40
32656.85
0.56
49443.82
0.71
180
120
100
80
APG m=8
APG m=16
APG m=32
APG*
APG* R
BS b=8
BS b=16
BS b=32
KNR-Cos k=7
KNR-Cos k=12
RCT c=128
100000
# distances / n
140
speedup
1e+06
APG m=8
APG m=16
APG m=32
APG*
APG* R
BS b=8
BS b=16
BS b=32
KNR-Cos k=7
KNR-Cos k=12
RCT c=128
160
10000
60
1000
40
20
0
0.60.7
0.8
0.9
0.95
0.97
0.98
100
0.99
0.6
0.7
0.8
0.9
recall
(a) Search speedup / 16 dimensions
120
0.97
0.98
0.99
(b) Search cost / 16 dimensions
1e+06
APG m=8
APG m=16
APG m=32
APG*
APG* R
BS b=8
BS b=16
BS b=32
KNR-Cos k=7
KNR-Cos k=12
RCT c=128
80
60
APG m=8
APG m=16
APG m=32
APG*
APG* R
BS b=8
BS b=16
BS b=32
KNR-Cos k=7
KNR-Cos k=12
RCT c=128
100000
# distances / n
100
speedup
0.95
recall
40
10000
20
0
0.60.7
0.8
0.9
1000
0.95
0.60.7
0.8
recall
(c) Search speedup / 32 dimensions
70
1e+06
40
30
APG m=8
APG m=16
APG m=32
APG*
APG* R
BS b=8
BS b=16
BS b=32
KNR-Cos k=7
KNR-Cos k=12
RCT c=128
100000
# distances / n
speedup
50
0.95
(d) Search cost / 32 dimensions
APG m=8
APG m=16
APG m=32
APG*
APG* R
BS b=8
BS b=16
BS b=32
KNR-Cos k=7
KNR-Cos k=12
RCT c=128
60
0.9
recall
10000
20
10
0
0.6
0.7
1000
0.8
recall
0.6
0.7
0.8
recall
(e) Search speedup / 64 dimensions
(f) Search cost / 64 dimensions
Figure 6: Speedup and search cost as a function of the expected recall of our indexes and
state of the art. We use datasets with dimensions 16, 32, and 64. Each dataset has three
instances with sizes of 3 × 105 , 106 , and 3 × 106 . We also limited memory to be under eight
direct neighbors per item, i.e. 16 undirected neighbors in average.
23
First of all, a guide for index selection. We first notice that APG* performs
as good as APG; however, APG*-R, and BS perform better. Both APG*-R
and BS need to know the number of direct near-neighbors t. APG*-R has no
additional parameters since it adjusts the number of repetitions based on a
simple iterative optimization; then the robustness is entirely based on t. For
this reason, t should be relatively large; our suggestion is to use values greater
than 16 but smaller than 64. Since the underlying graph is undirected, the
memory usage will be n(2t + 1) integers. Thus, the right selection of t should
consider this amount of main memory for the adequate performance. For the
case of BS, t can be smaller, i.e., 8 or 16; larger values of t should be used for
very high dimensions, as described in our experimental section. After fixing t,
the free parameter b (the size of the beam) should be adjusted to obtain the
desired performance. A significant b will produce high-quality search results
while increasing search time. In contrast, faster searches will be achieved for
small b values, yet the quality of the results could decrease. As a rule of thumb,
we recommend fixing b to the largest value producing a search time below the
user threshold. Nonetheless, good recall values are achieved for most of our
datasets for b values between 16 and 64. A more precise setup of b could be
accomplished with the help of a query set Q. The idea is to adjust b to obtain
a minimum recall on Q; this fitting should be applied after a fixed number of
insertions.
6. Summary and Perspectives
In this manuscript, we approach proximity search as a combinatorial optimization problem. To this end, we define the necessary framework to allow
the use of combinatorial optimization algorithms studied in the literature. In
particular, this contribution focuses on local search algorithms.
Using our new perspective, we introduced three near-neighbor indexes, namely,
APG*, APG*-R, and BS. Those indexes use an underlying graph and several
variants of local search to navigate them. As part of our contribution, we compare our techniques against state of the art indexes. We found that our indexes
significantly surpass the performance of the majority of the alternatives, in almost all of our benchmarks. In particular, we believe that our indexes are of
practical use due to their excellent trade-off among speed, recall, and index size.
In addition to our core contributions, this manuscript also gives an extensive
comparison of state of the art on the nearest neighbor problem. We tested
algorithms working in both general metrics, and those specialized on vector
spaces, in particular for the Euclidean distance. We also include E2 LSH, a well
known and used near neighbor index based on hashing. Our intention was to
provide a better and wider picture of the alternatives for near neighbor search,
disregarding the underlying theoretical basis with a comprehensive guide to
real-world applications in mind.
In the present work, we fixed the construction of N to be that defined by
APG while we focus on finding better algorithms for navigating the graph, i.e.,
A. However, new N definitions with improved or additional properties should
24
be explored. The indexes introduced in this manuscript have some parameters
to be determined experimentally. This situation is not uncommon; however,
this complicates the usage of the new indexes in real world tasks. Even when
we give guidelines for tuning our indexes, it is necessary to look forward to
construction algorithms that determine their parameters automatically using,
for example, the topology and geometry of the given dataset.
To be effectively used in production systems, our indexes should support
some degree of persistence. A secondary memory implementation could be a
solution; however, the high number of random seeks arising on graph-based
algorithms is a problem in disk-based storage. A possible solution is to use
exclusively solid state storage as secondary memory; however, I/O operations
are still far slower than main memory. An alternative scheme is a pure memory
index having a secondary memory mirror, such that any modification is performed in both memory and disk storage. Therefore, the idea is that the data
persist and the index become available as soon the process is launched in the
system. This scheme is quite attractive as it is implemented in high-performance
databases like Redis [9].4
Another important task is to support dynamic operations. This feature is
useful for applications using metric databases to represent knowledge, like those
using incremental learning. Since our indexes are incrementally constructed, the
insertion operation is native. However, the deletion algorithm is not yet studied.
The literature in metric indexes used to solve the deletion operation marking
those items as unavailable for most operations [13]. While this is enough for
workloads with just a few deletions, a high rate of removals could yield to a
significant unnecessary extra memory. In other approaches, the metric indexes
can afford real deletions. However, to maintain the invariants of the indexes, a
partial or even a complete reconstruction of the index is needed ([31, 32, 14]).
A recent approach uses the strategy of Bentley and Saxe ([6, 30]) to produce
dynamic structures using a list of log n static structures. This approach gives
the possibility of using any static indexes at the expense of a log n factor in
the search cost. Also, since only amortized times are ensured, incremental
constructions cannot be used (i.e. a single insertion or deletion operation can
be lengthy). We believe that our combinatorial optimization approach is robust
enough to support deletions, and the primary task becomes on determining
optimal replacements of the removed items. This operation deserves attention
in future research on the area.
The current construction of APG and our indexes is a sequential process,
i.e., the i-th item is connected with its neighbors among the previous i − 1
items. In the case of APG and APG*, each search is composed of several greedy
searches. Each greedy search is not independent of others since they use the
best-known candidate at each step. A possible point of parallelization is at each
hop since 2t distances are computed. In the case of BS, at each hop 2tb distance
4 Redis is an in-memory data structure store, used as a database, cache and message broker,
https://redis.io/
25
computations are performed. In any case, the necessary distance calculations
can be executed in parallel. However, the intrinsic cost of thread creation and
management can be too high to take advantage of parallelism. APG*-R, on
the other hand, can be improved with independent searches. Nonetheless, there
exists space for clever and better parallel construction algorithms.
Finally, we also should mention the lack of theoretical guarantees for both the
search time and the expected recall. This task is quite relevant since theoretical
assurances allow the usage of our indexes in those applications where a minimum
guarantee is required.
References
[1] Amato, G., Gennaro, C., Savino, P., 2014. Mi-file: using inverted files for
scalable approximate similarity search. Multimedia Tools and Applications
71 (3), 1333–1362.
[2] Amato, G., Savino, P., 2008. Approximate similarity search in metric spaces
using inverted files. In: InfoScale ’08: Proceedings of the 3rd international
conference on Scalable information systems. ICST (Institute for Computer
Sciences, Social-Informatics and Telecommunications Engineering), ICST,
Brussels, Belgium, Belgium, pp. 1–10.
[3] Andoni, A., Indyk, P., January 2008. Near-optimal hashing algorithms for
approximate nearest neighbor in high dimensions. Communications ACM
51, 117–122.
URL http://doi.acm.org/10.1145/1327452.1327494
[4] Andoni, A., Indyk, P., January 2008. Near-optimal hashing algorithms for
approximate nearest neighbor in high dimensions. Communications of the
ACM 51, 117–122.
[5] Baeza-Yates, R. A., Ribeiro-Neto, B. A., 1999. Modern Information Retrieval. ACM Press / Addison-Wesley.
URL citeseer.ist.psu.edu/baeza-yates99modern.html
[6] Bentley, J. L., Saxe, J. B., 1980. Decomposable searching problems i. staticto-dynamic transformation. Journal of Algorithms 1 (4), 301–358.
[7] Beygelzimer, A., Kakade, S., Langford, J., 2006. Cover trees for nearest
neighbor. In: Proceedings of the 23rd international conference on Machine
learning. ACM, pp. 97–104.
[8] Bustos, B., Navarro, G., 2004. Probabilistic proximity search algorithms
based on compact partitions. Journal of Discrete Algorithms 2 (1), 115–
134.
[9] Carlson, J. L., 2013. Redis in Action. Manning Publications Co., Greenwich, CT, USA.
26
[10] Chavez, E., Figueroa, K., Navarro, G., Sep. 2008. Effective proximity retrieval by ordering permutations. IEEE Transactions on Pattern Analysis
and Machine Intelligence 30 (9), 1647–1658.
[11] Chávez, E., Graff, M., Navarro, G., Téllez, E., 2015. Near neighbor
searching with K nearest references . Information Systems 51 (0), 43–61.
URL
http://www.sciencedirect.com/science/article/pii/
S0306437915000241
[12] Chávez, E., Navarro, G., 2003. Probabilistic proximity search: Fighting the
curse of dimensionality in metric spaces. Information Processing Letters 85,
39–46.
[13] Chávez, E., Navarro, G., Baeza-Yates, R., Marroquı́n, J. L., 2001. Searching
in metric spaces. ACM Comput. Surv. 33 (3), 273–321.
[14] Ciaccia, P., Patella, M., Zezula, P., 1997. M-tree: An efficient access
method for similarity search in metric spaces. In: Proceedings of the 23rd
International Conference on Very Large Data Bases. VLDB ’97. Morgan
Kaufmann Publishers Inc., San Francisco, CA, USA, pp. 426–435.
URL http://dl.acm.org/citation.cfm?id=645923.671005
[15] Clarkson, K. L., 1999. Nearest neighbor queries in metric spaces. Discrete
and Computational Geometry 22, 63–93.
[16] Esuli, A., 2009. Pp-index: Using permutation prefixes for efficient and scalable approximate similarity search. In: Proceedings of the 7th Workshop on
Large-Scale Distributed Systems for Information Retrieval (LSDS-IR’09).
Boston, USA, pp. 17–24.
[17] Esuli, A., 2012. Use of permutation prefixes for efficient and scalable approximate similarity search. Information Processing & Management 48 (5),
889–902.
[18] Figueroa, K., Chavez, E., Navarro, G., 2009. Technical description of the
metric spaces library. Technical Report. SISAP Project.
[19] Gao, J., Jagadish, H. V., Lu, W., Ooi, B. C., 2014. DSH: Data sensitive hashing for high-dimensional k-nnsearch. In: Proc. ACM International
Conference on Management of Data (SIGMOD). pp. 1127–1138.
[20] Gionis, A., Indyk, P., Motwani, R., 1999. Similarity search in high dimensions via hashing. In: Proceedings of the 25th International Conference on
Very Large Data Bases. VLDB ’99. Morgan Kaufmann Publishers Inc., San
Francisco, CA, USA, pp. 518–529.
URL http://portal.acm.org/citation.cfm?id=645925.671516
[21] Goyal, N., Lifshits, Y., Schütze, H., 2008. Disorder inequality: a combinatorial approach to nearest neighbor search. In: Proceedings of the 2008
International Conference on Web Search and Data Mining. ACM, pp. 25–
32.
27
[22] Houle, M. E., Nett, M., 2013. Rank cover trees for nearest neighbor search.
In: Similarity Search and Applications. Springer, pp. 16–29.
[23] Indyk, P., 2004. Handbook of Discrete and Computational Geometry., 2nd
Edition. CRC press, chapter 39.
[24] Malkov, Y., Ponomarenko, A., Logvinov, A., Krylov, V., 2012. Scalable
distributed algorithm for approximate nearest neighbor search problem in
high dimensional general metric spaces. In: Proc. 5th International Conference on Similarity Search and Applications (SISAP). pp. 132–147.
[25] Malkov, Y., Ponomarenko, A., Logvinov, A., Krylov, V., 2014. Approximate nearest neighbor algorithm based on navigable small world graphs.
Information Systems 45, 61–68.
[26] Mohamed, H., 2014. Scalable approximate k-nn in multidimensional big
data. phd thesis. Ph.D. thesis, University of Geneva.
[27] Mohamed, H., Marchand-Maillet, S., 2013. Quantized ranking for
permutation-based indexing. In: Brisaboa, N., Pedreira, O., Zezula, P.
(Eds.), Similarity Search and Applications. Vol. 8199 of Lecture Notes in
Computer Science. Springer Berlin Heidelberg, pp. 103–114.
URL http://dx.doi.org/10.1007/978-3-642-41062-8_11
[28] Muja, M., Lowe, D., Nov 2014. Scalable nearest neighbor algorithms for
high dimensional data. Pattern Analysis and Machine Intelligence, IEEE
Transactions on 36 (11), 2227–2240.
[29] Naidan, B., Hetland, M. L., 2012. Static-to-dynamic transformation for
metric indexing structures. In: Navarro, G., Pestov, V. (Eds.), Similarity Search and Applications - 5th International Conference, SISAP 2012,
Toronto, ON, Canada, August 9-10, 2012. Proceedings. Vol. 7404 of Lecture Notes in Computer Science. Springer, pp. 101–115.
URL http://dx.doi.org/10.1007/978-3-642-32153-5_8
[30] Naidan, B., Hetland, M. L., 2013. Static-to-dynamic transformation for
metric indexing structures (extended version). Information Systems.
[31] Navarro, G., Reyes, N., 2008. Dynamic spatial approximation trees. ACM
Journal of Experimental Algorithmics 12, 1.5:1–1.5:68.
URL http://doi.acm.org/10.1145/1227161.1322337
[32] Navarro, G., Reyes, N., 2009. Dynamic spatial approximation trees for
massive data. In: Second International Workshop on Similarity Search and
Applications, SISAP 2009, 29-30 August 2009, Prague, Czech Republic.
pp. 81–88.
URL http://doi.ieeecomputersociety.org/10.1109/SISAP.2009.28
[33] Patella, M., Ciaccia, P., 2009. Approximate similarity search: A multifaceted problem. Journal of Discrete Algorithms 7 (1), 36–48.
28
[34] Samet, H., 2006. Foundations of Multidimensional and Metric Data Structures, 1st Edition. The morgan Kaufman Series in Computer Graphics and
Geometic Modeling. Morgan Kaufmann Publishers, University of Maryland
at College Park.
[35] Silpa-Anan, C., Hartley, R., June 2008. Optimised kd-trees for fast image
descriptor matching. In: Computer Vision and Pattern Recognition, 2008.
CVPR 2008. IEEE Conference on. pp. 1–8.
[36] Skopal, T., Nov. 2007. Unified framework for fast exact and approximate
search in dissimilarity spaces. ACM Transactions on Database Systems
32 (4).
URL http://doi.acm.org/10.1145/1292609.1292619
[37] Skopal, T., 2010. Where are you heading, metric access methods?: a
provocative survey. In: Proceedings of the Third International Conference
on SImilarity Search and APplications. SISAP ’10. ACM, New York, NY,
USA, pp. 13–21.
URL http://doi.acm.org/10.1145/1862344.1862347
[38] Tellez, E. S., Chavez, E., 2010. On locality sensitive hashing in metric
spaces. In: Proceedings of the Third International Conference on SImilarity
Search and APplications. SISAP 2010. ACM, New York, NY, USA, pp. 67–
74.
URL http://doi.acm.org/10.1145/1862344.1862355
[39] Tellez, E. S., Chavez, E., Camarena-Ibarrola, A., November 2009. A brief
index for proximity searching. In: Proceedings of 14th Iberoamerican
Congress on Pattern Recognition CIARP 2009. Lecture Notes in Computer
Science. Springer Verlag, Berlin, Heidelberg, pp. 529–536.
[40] Tellez, E. S., Chavez, E., Graff, M., 2011. Scalable pattern search analysis.
In: 3rd Mexican Congress on Pattern Recognition, MCPR 2011. Springer
Verlag, Lecture Notes in Computer Science.
[41] Tellez, E. S., Chávez, E., Navarro, G., 2011. Succinct nearest neighbor
search. In: Proc. 4th International Conference on Similarity Search and
Applications, SISAP 2011. ACM Press.
[42] Tellez, E. S., Chávez, E., Navarro, G., 2012. Succinct nearest neighbor
search. Information Systems.
[43] Zezula, P., Amato, G., Dohnal, V., Batko, M., 2006. Similarity Search - The
Metric Space Approach, 1st Edition. Vol. 32. Springer, Series: Advances in
Database Systems.
29
| 8cs.DS
|
ESTIMATING PERIMETER USING GRAPH CUTS
arXiv:1602.04102v2 [math.ST] 14 Aug 2016
NICOLÁS GARCÍA TRILLOS1 , DEJAN SLEPČEV2 AND JAMES VON BRECHT3
A BSTRACT. We investigate the estimation of the perimeter of a set by a graph cut of a
random geometric graph. For Ω ⊂ D = (0, 1)d , with d ≥ 2, we are given n random
i.i.d. points on D whose membership in Ω is known. We consider the sample as a random
geometric graph with connection distance ε > 0. We estimate the perimeter of Ω (relative
to D) by the, appropriately rescaled, graph cut between the vertices in Ω and the vertices in
D\Ω. We obtain bias and variance estimates on the error, which are optimal in scaling with
respect to n and ε. We consider two scaling regimes: the dense (when the average degree of
the vertices goes to ∞) and the sparse one (when the degree goes to 0). In the dense regime
there is a crossover in the nature of approximation at dimension d = 5: we show that in
low dimensions d = 2, 3, 4 one can obtain confidence intervals for the approximation error,
while in higher dimensions one can only obtain error estimates for testing the hypothesis
that the perimeter is less than a given number.
1. I NTRODUCTION
This paper investigates the use of random-graph cuts to obtain empirical estimates of
the perimeter of a domain Ω ⊂ D := (0, 1)d for d ≥ 2. Let x1 , . . . , xn , . . . denote a
sequence of independent random points uniformly distributed on the unit cube D and let
Vn := {x1 , . . . , xn }.
The problem of estimating the perimeter of Ω based on knowing which points of Vn
belong to Ω is a classical question, see [4, 8, 9, 17, 23, 24] for recent contributions, and see
Subsection 1.2 below which contains a discussion about related work. Here we consider an
estimator of the perimeter that is based on a geometric graph constructed from the point
cloud Vn . More precisely, we select εn > 0, and connect two points in the cloud if they are
within distance εn of each other; then we consider an appropriately scaled ‘cut’ determined
by the number of edges in the graph that connect points in Ω with points that belong to
Ωc . This type of estimator is natural to consider, since graph cuts arise as a discretization
of the perimeter in many applications such as clustering [3, 7, 14, 15, 18, 25, 26, 27, 28].
Our choice of the estimator is therefore based on its use in various statistical and machine
learning applications.
We focus on estimating the approximation error of the perimeter of an arbitrary (but
fixed) set Ω; the error estimates that we obtain are uniform on a class of sets where certain
geometric quantities are controlled (see Remark 1.7 below). One of the important features
of our estimator is that it has a small bias. Indeed, the expectation of our estimator provides
a second order approximation (in terms of the natural parameter εn ) of the true perimeter of
Ω under some regularity conditions on the boundary of the set (see (8) below); this turns
out to be a sharp estimate for the bias. We also obtain precise estimates for the variance
of the estimator; for these estimates to hold, we do not need any regularity assumptions
Date: August 16, 2016.
Key words and phrases. perimeter, nonparametric estimation, graph cut, point cloud, random geometric graph,
concentration inequality.
1
2
NICOLÁS GARCÍA TRILLOS1 , DEJAN SLEPČEV2 AND JAMES VON BRECHT3
on the boundary of the set Ω except that it has finite perimeter in the most general sense.
Furthermore, we show that our estimator converges a.s. for remarkably sparse graphs (and
indeed in settings which are sparser than for previously considered estimators). The a.s.
convergence holds with no regularity assumptions on Ω (other than the fact that it has finite
perimeter in the most general sense). Finally, assuming some smoothness on the boundary
1
of Ω, we are able to establish (in the dense graph regime n1/d
ε 1) the asymptotic
distribution of the error. These estimates lead to asymptotic confidence intervals (which
we refer to simply as confidence intervals) and bounds on the type I and type II errors for
hypothesis tests associated to the perimeter of a set.
1.1. Set-up and main results. Let us now be more precise about the setting we consider
in this paper. We consider random geometric graphs with vertex set Vn and radius εn > 0.
That is, graphs where xi and xj are connected by an edge if ||xi − xj || ≤ εn . The graph
cut between A ⊆ Vn and Ac is given by
X X
Cutεn (A, Ac ) :=
1{kxi −xj k≤εn } .
xi ∈A xj ∈V \A
We define the graph perimeter as a rescaling of the graph cut: For any Ω ⊆ D
2
(1)
GPern,εn (Ω) :=
Cut(Vn ∩ Ω, Vn ∩ Ωc ).
n(n − 1)εd+1
n
See Figure 1 for an illustration of this construction. The scaling is chosen so that GPern,εn (Ω)
becomes a consistent estimator for the true (continuum) perimeter.
Ω
Ωc
F IGURE 1. Edges of the cut between Ω and Ωc are represented by bold
lines, while other edges are dashed lines. Total number of vertices is
n = 200 and connectivity radius ε = 0.13.
One of our interests lies in determining how well does GPern,εn (Ω) estimate the relative
perimeter of Ω in D. We first investigate for which scaling of εn on n does the convergence
ESTIMATING PERIMETER USING GRAPH CUTS
3
hold almost surely as n → ∞. In other words, we want to understand the relation between
edge-sparsity of random geometric graphs and point-wise convergence of the graph perimeter to the continuum perimeter in the almost sure sense. We consider this question for a very
broad family of sets Ω ⊂ D, which are only assumed to have finite relative perimeter in D
in the general sense of [2]. That is, we define the relative perimeter of Ω with respect to D
to be
Z
∞
d
(2)
Per(Ω) = sup
div(v) dx : (∀x ∈ D) kv(x)k ≤ 1, v ∈ Cc (D, R ) .
Ω
If Ω has a smooth relative boundary then Per(Ω) is nothing but the surface area of ∂Ω ∩ D.
We remark that the notion of the perimeter we use is more general than the notion of
Hausdorff measure of the boundary, Hd−1 (∂Ω ∩ D), and than the Minkowski content,
which are the ones more typically used in the statistics literature [4, 8, 9]. In particular, as
we see below, we work with consistent nonparametric estimators in the most general setting
available.
As we recall below in (17 –19), it is known that when εn → 0 as n → ∞ then the bias
of the estimator vanishes in the limit:
(3)
E(GPern,εn (Ω)) → σd Per(Ω)
as
n → ∞.
The scaling factor σd satisfies
Z
(4)
|z1 | dz =
σd :=
||z||≤1
2sd−2
,
(d + 1)(d − 1)
where z1 denotes the first component of the vector z ∈ Rd and sd−2 is the area of the
(d − 2)-dimensional unit sphere (the boundary of the unit ball in Rd−1 ). We refer to the
normalizing quantity σd as the surface tension.
We obtain the following estimates on the deviation of the graph perimeter GPern,εn (Ω)
from its mean. Let
1
1
√nε
if n1/d
≤ εn
n
(5)
f (n, εn ) :=
1
1
1
(d+1)/2
if n2/(d+1)
≤ εn ≤ n1/d
.
nεn
Theorem 1.1. Let p ≥ 1 and let Ω ⊆ D be a set with finite perimeter. Assume εn → 0 as
n → ∞. Then,
(6)
E(| GPern,εn (Ω) − E(GPern,εn (Ω))|p ) ≤ Cp,d (max{1, Per(Ω)}f (n, εn ))p
where Cp,d is a constant that depends only on p and dimension d. In particular, if
2
n− (d+1) εn 1, then
GPern,εn (Ω) → σd Per(Ω), almost surely as n → ∞.
The last part of the previous theorem follows from (3), the moment estimates (6),
Markov’s inequality, and Borel-Cantelli Lemma which imply that
GPern,εn (Ω) − E (GPern,εn (Ω)) → 0
a.s.
Remark 1.2. We note that the a.s. convergence holds for rather sparse graphs (see Figure 2).
Namely the typical degree of a node is ωd nεd , where ωd is the volume of the unit ball in d
2
1
dimensions. When n− (d+1) εn n− d the a.s. convergence holds, while the average
4
NICOLÁS GARCÍA TRILLOS1 , DEJAN SLEPČEV2 AND JAMES VON BRECHT3
(a) n = 400 and ε = 0.045
(b) n = 1000 and ε = 0.027
F IGURE 2. Here we illustrate the “sparse” regime when 1/n1/d ε
(1/n)2/d+1 when the average degree goes to zero. Nevertheless since the
number of edges in the cut still increases as n does the convergence of
the cut to the perimeter still holds.
degree of a vertex converges to zero. The convergence is still possible because the expected
number of edges crossing ∂Ω is still a quantity converging to infinity.
Remark 1.3. Given that we show the almost-sure consistency of our estimators for any
arbitrary (but fixed) set for which the perimeter is finite, our construction provides a universal
strongly convergent estimator which was the desired property listed as an open problem in
[8] (for the estimator they considered).
We turn to estimating the bias of the empirical approximation: |E(GPern,εn (Ω)) −
σd Per(Ω)|. We first characterize the mean of the graph perimeter, E(GPern,εn (Ω)), as the
non-local perimeter Perεn (Ω) of Ω, defined as
Z Z
2
(7)
Perε (Ω) := d+1
1{kx−yk≤ε} dxdy.
ε
Ω D\Ω
The non-local nature of the functional essentially has to do with the fact that it involves
averages of finite differences as opposed to a local approach where one considers derivatives.
We then proceed to estimate |Perεn (Ω) − σd Per(Ω)| explicitly. It proves straightforward
to check that |Perεn (Ω) − σd Per(Ω)| = O (εn ) for general subsets Ω ⊆ D with smooth
relative boundary. However, we show that the error is actually quadratic in εn
(8)
|E(GPern,εn (Ω)) − σd Per(Ω)| = |Perεn (Ω) − σd Per(Ω)| = O(ε2n )
under the extra condition that dist(Ω, ∂D) > 0. This is the content of the next lemma,
whose proof may be found in Appendix A.
Lemma 1.4. Let Ω be a set with smooth boundary, such that dist(Ω, ∂D) > 0. Let
0 < ε < dist(Ω, ∂D) and let Perε (Ω) be defined by (7). Then
(9)
Perε (Ω) = σd Per(Ω) + O(ε2 ).
ESTIMATING PERIMETER USING GRAPH CUTS
5
Remark 1.5. The assumption dist(Ω, ∂D) > 0 in the above Lemma is needed in order to
obtain bias of order ε2 . If Ω touches the boundary ∂D the error of order ε2 is not expected, as
can be seen for example by considering the rectangle Ω := {x = (x1 , . . . , xd ) ∈ D : x1 ≤ 1/2},
for which the error is of order ε; in this situation the error is completely due to the region
where ∂Ω meets ∂D transversally. Thus, in general, for Ω ⊆ D with smooth relative
boundary, the bias is of order ε. On the other hand, the smoothness of the boundary of Ω is
only needed in the previous lemma to guarantee that curvature and its derivatives are well
defined. Finally, the constant involved in the term O(ε2 ) depends on the reach of the set
Ω, and the intrinsic curvature of ∂Ω together with its derivatives; this can be seen from our
computations in Appendix A.
Combining the bias and variance of estimates allows us to obtain the rates of convergence
for the error |GPern,εn (Ω) − σd Per(Ω)|. In particular we estimate the ‘standard deviation’
1/2
std(n) := E (GPern,εn (Ω) − σd Per(Ω))2
,
which we may quantify precisely by using the variance-bias decomposition
2
std2 (n) = Var(GPern,εn (Ω)) + (E(GPern,εn (Ω)) − σd Per(Ω)) .
Using the special case p = 2 of Theorem 1.1 to estimate the variance and using Lemma 1.4
to estimate the bias we obtain the following.
2
Theorem 1.6. Let Ω ⊂ D be an open set with smooth boundary. Assume n− d+1
εn 1 and consider f (n, εn ) defined via (5). The error of approximating σd Per(Ω) by
GPern,εn (Ω) satisfies
std(n) = O(f (n, εn ) + εn ).
If we furthermore assume that Ω does not touch the boundary of D, that is dist(Ω, ∂D) > 0,
then a better estimate holds:
std(n) = O(f (n, εn ) + ε2n ).
A simple calculation using (5) allows one to choose a scaling of εn on n so that the error
of the approximation is as small as possible. In particular, for a set Ω which touches the
boundary (dist(Ω, ∂D) = 0) the optimal scaling of εn is
(
(
n−1/3
if d ≤ 3
n−1/3
if d ≤ 3
εn ∼
giving
std(n)
=
−2/(d+3)
−2/(d+3)
n
if d ≥ 3
n
if d ≥ 3.
We note that for d < 3 the optimal εn is achieved in the regime n−1/d . εn 1, while if
d > 3 it is in the sparse regime. If we consider sets Ω with smooth boundary but such that
dist(Ω, ∂D) > 0, then the optimal scaling of εn on n is as follows
(
(
n−2/5
if d ≤ 5
n−1/5
if d ≤ 5
εn ∼
giving
std(n)
=
−4/(d+5)
−2/(d+5)
n
if d ≥ 5
n
if d ≥ 5.
Again we note that the optimal εn is in the sparse regime if d > 5. This has implications to
how well is the perimeter estimated by graph cuts in the graphs considered in most machine
learning applications. Namely if d ≥ 5 and the graph has average degree bounded from
below, that is when n−1/d . εn 1, then, since the bias bound is sharp, most of the error
is due to the bias term.
6
NICOLÁS GARCÍA TRILLOS1 , DEJAN SLEPČEV2 AND JAMES VON BRECHT3
Remark 1.7. The constant appearing in the moment estimates of Theorem 1.1 depends
exclusively on the power p, the dimension d, and the true perimeter of the set Ω; in
particular, taking p = 2 we see that the estimates for the variance of GPern,εn (Ω) are
uniform on the class of sets Ω whose perimeter is bounded above by some fixed constant.
For general sets with large perimeter we obtain uniform estimates for moments of relative
error | GPern,εn (Ω) − E(GPern,εn (Ω))|/ Per(Ω) instead of the absolute error. Thus, by
combining Theorem 1.1 and the proof of Lemma 1.4 one can derive that the corresponding
error estimates for |GPern,εn (Ω) − σd Per(Ω)| (either relative or absolute) are uniform on
the class of sets Ω with smooth boundary satisfying the following conditions: The reach
of Ω is bounded below by a fixed positive constant; The distance to the boundary ∂D is
bounded below by a fixed positive constant; The curvature and first derivatives of curvature
are bounded from above by a fixed constant. This last requirement comes from the O(ε2 )
terms following Taylor’s theorem in the proof of our bias estimates.
We now consider obtaining confidence intervals for the value of the true perimeter Per(Ω)
based on the estimator σ1d GPern,εn (Ω). We focus on the dense regime (See Figures 3
1
and 4), n1/d
εn 1, and first obtain the asymptotic distribution of GPern,εn (Ω) −
E(GPern,εn (Ω)).
(a) n = 100, ε = 0.17
(b) n = 300 and ε = 0.12
F IGURE 3. Here we illustrate the “very dense” regime, 1 ε
(ln(n)/n)1/d , when the graphs are connected with high probability.
Since the term E (GPern,εn (Ω)) = Perεn (Ω) depends on the set Ω itself, which is
unknown (we only assume we have an oracle access to it), we need the bias to be negligible
compared to the standard deviation of our estimator. This allows us to construct confidence
intervals for Per(Ω) without using any additional information about the bias (e.g. upper
1
bounds). From (6) the standard deviation of GPern,εn (Ω) scales as √nε
while by (8) the
n
2
bias scales as εn ; these estimates lead to restrictions on the dimensions for which the bias is
negligible with respect to the standard deviation. Namely this is possible for d = 2, 3, and
4.
ESTIMATING PERIMETER USING GRAPH CUTS
(a) n = 100 and ε = 0.12
7
(b) n = 400 and ε = 0.07
F IGURE 4. Here we illustrate an intermediate, ”dense”, regime
(ln(n)/n)1/d ε (1/n)1/d . The average degree still increases,
but the graphs are disconnected with high probability.
Theorem 1.8. Let Ω ⊆ D be an open set with smooth boundary such that dist(Ω, ∂D) > 0.
Let εn be such that
1
εn 1
1/d
n
Then,
r
nεn
w
(GPern,εn (Ω) − Perεn (Ω)) −→ N (0, 1),
4Cd Per(Ω)
where Cd is given by
Z 1
2
(10)
Cd := 2
Bd (0, 1) ∩ x = (x1 , . . . , xd ) ∈ Rd : xd ≥ t
dt.
0
If in addition the dimension d is either 2, 3 or 4 and if
1
(11)
n1/d
εn
1
n1/5
then,
r
(12)
nεn
w
(GPern,εn (Ω) − σd Per(Ω)) −→ N (0, 1).
4Cd Per(Ω)
Naturally, the previous theorem implies that one can obtain confidence intervals for the
value of Per(Ω) when the dimension d is 2, 3 or 4. Let us fix α ∈ (0, 0.5) and let Zα be the
1 − α quantile of the standard normal distribution. That is, Zα is such that
P (N (0, 1) ≤ Zα ) = 1 − α.
Then provided that
1
n1/d
εn
1
n1/5
,
8
NICOLÁS GARCÍA TRILLOS1 , DEJAN SLEPČEV2 AND JAMES VON BRECHT3
+
it follows that with probability converging to 1 − α, Per(Ω) ∈ (a−
n , an ), where
s
4Cd Per(Ω)
1
a±
(GPern,εn (Ω) ±
Zα/2 ).
n :=
σd
nεn
For d ≥ 5, the confidence intervals can not be obtained unless one has some extra quantitative information on the smoothness of ∂Ω, such as curvature bounds or upper bounds on the bias. Nevertheless, since it is known that the nonlocal perimeter
Perεn (Ω) = E(GPern,εn (Ω)) is less than σd Per(Ω) (see (30)), one can construct a test for
the hypothesis that Per(Ω) is less than a certain number ρ without using any quantitative
estimates on the smoothness of ∂Ω. In this case, we have automatic upper bounds for the
bias and may consider the hypothesis testing of
(13)
H0 : Per(Ω) ≤ ρ,
vs HA : Per(Ω) > ρ,
based on our estimator GPern,εn (Ω). We consider the statistic:
r
nεn
(14)
ln :=
(GPern,εn (Ω) − σd ρ) .
4Cd ρ
The test consists on
(15)
Accept H0 if ln ≤ Zα ,
Proposition 1.9. Assume d ≥ 2 and
1
n1/d
reject otherwise.
εn 1. Then, the type I error satisfies
lim sup PH0 (ln > Zα ) ≤ α,
n→∞
i.e., the type I error is asymptotically below α. The type II error satisfies
1
PHA (ln ≤ Zα ) = O √
.
nεn
Remark 1.10 (Extensions). The estimates obtained in Theorem 1.1 and Theorem 1.6 are
not exclusive to the case where the points are uniformly distributed in the unit cube and
to geometric graphs. In fact, with slight modifications to the proofs of Theorem 1.1 and
Theorem 1.6, we can extend these results to more general situations. For example, if the
data points are distributed according to some smooth density p that is supported on a regular,
bounded domain D ⊂ Rd with p bounded below and above by positive constants, then the
results still hold. In this case, the limiting value of GPern,εn (Ω) is a weighted perimeter
Per(Ω, p2 ) (see for example [10]). Convergence is guaranteed for the same scaling for εn
as in the uniform case. The fact that the weight is p2 and not p (as may be a priori expected)
ultimately comes from the fact that a graph cut is a double sum.
Furthermore if instead of weights 1{kxi −xj k≤εn } in the definition of the graph cut,
one considers edge weights η(kxi − xj k/εn ) where η is nonnegative, integrable and nonincreasing, the results still hold, provided we change the surface tension σd with a surface
tension associated to η defined as in [10].
Remark 1.11. The previous results allow us to construct asymptotic confidence intervals
for Per(Ω) for an arbitrary set Ω with smooth boundary and dist(Ω, ∂D) > 0, using the
estimator GPern,εn (Ω). Nevertheless, since very small sets (in the sense of volume) may
have very large perimeter the error estimates are not uniform in Ω and thus the asymptotic
confidence intervals may not be of direct practical use. This issue is unavoidable without
further assumptions on the set Ω or a change of framework for testing. One possible way
ESTIMATING PERIMETER USING GRAPH CUTS
9
to restrict the class of Ω considered is pointed out in Remark 1.7. In the class of sets Ω to
which Remark 1.7 applies, the error of approximation is uniformly controlled.
Alternatively, one may consider the property testing framework used in [23, 21], where
one could test the property “the set Ω is close to a set whose perimeter is less than ρ” reliably.
In that framework we would not be using the estimator GPern,εn (Ω) to test whether Per(Ω)
is below ρ or not, but rather whether Ω is close to a set whose perimeter is below ρ or not.
We notice that in this paper we have taken the classical hypothesis testing approach.
Outline. In Subsection 1.3 we establish (3) and give an outline of the argument behind
our main results. We present the proof of Theorem 1.1 in Section 2, while in Subsection
2.1 we show that the scaling is sharp (up to logarithmic corrections) in the sense that
if n2 εd+1
→ 0 then GPern,εn (Ω) converges in probability to zero and hence does not
n
converge almost surely to the (rescaled) relative perimeter. In Section 3 we prove the results
on the asymptotic distribution of the error stated in Theorem 1.8. In Subsection 3.1 we study
the type I and type II errors of the hypothesis test of Proposition 1.9. Finally, Appendix A
deals with the bias estimate from Lemma 1.4.
1.2. Discussion. Here we discuss the connections between our work and related works
in the literature. First we relate it to other estimators of perimeter based on a random
sample. Then we contrast the type of the convergence and the scaling regimes considered in
this paper with the ones needed for the convergence of graph-cut based machine learning
algorithms for clustering and related tasks.
The problem of estimating the perimeter of a set, Ω, based on knowing which points
of a random sample Vn belong to Ω, has been considered by a number of works. Cuevas,
Fraiman, and Rodrı́guez-Casaet [9], considered estimators of the Minkowksi content, which
agrees with perimeter for regular enough sets, but is a less general notion of the perimeter
than the one we consider, (2). Their estimator is based on counting the vertices near the
boundary (relative to a parameter εn ), while we “count” edges of a graph. The error bound
obtained was of order O(n−1/2d ). Pateiro-López and Rodrı́guez-Casal [24] consider a
similar estimator and improve the bounds to O((ln n/n)1/(d+1) ) under qualitative regularity
conditions (rolling ball conditions). Cuevas, Fraiman, and Györfi [8], obtain the convergence
of estimators similar to those of [9], when εn 1/n1/d and under weaker conditions on
the regularity of Ω, although still not in the full generality we consider in this paper.
Armendáriz, Cuevas, and Fraiman [4] consider a similar set-up to that in [9, 24] but
with different sampling rates for Ω and Ωc : let n denote the number of sample points in
Ω and k the number of sample points in Ωc . Under mild assumptions on the regularity
of ∂Ω, and under some conditions on n and k which include 1 εn,k (1/n1/3 ) and
k (n/εn,k )d/2 they obtain the asymptotic distribution of the error, under the same
scaling in n and ε that we consider in Theorem 1.8. We notice that our estimator is different,
and also that in [4] a very large number of points k in Ωc is needed for the consistency
to hold. This allows the authors to obtain the asymptotic distribution of the total error
in any dimension, while in our setting we only obtain it in low dimensions. A further
difference between their work and ours, is that we allow for a wider range in ε, namely
1 εn (1/n)2/(d+1) .
Jiménez and Yukich [17] give a different use to the point cloud and instead of considering
a parameter ε to count points close to the boundary of the set or to define a geometric graph,
they consider a new estimator based on the Delaunay triangulation induced by the cloud.
They obtain results not only on estimating the perimeter of the set, but also integrals of
functions over ∂Ω.
10
NICOLÁS GARCÍA TRILLOS1 , DEJAN SLEPČEV2 AND JAMES VON BRECHT3
Kothari, Nayyeri, and O’Donnell [21] and Neeman [23] consider an estimator essentially
based on the following procedure: pick n random points uniformly distributed on D and to
each point associate a random direction for a ‘needle’ based at the point with length of order
√
εn ; then count how many of the needles touch the boundary of Ω. Their main motivation
is to consider the perimeter estimation from the viewpoint of property testing as introduced
in [19]. In that setting, the idea is to produce an algorithm that requires a small number
of samples (essentially independent of the dimension d) in order to determine if a given
set has a small perimeter or is ‘far away’ from a set that has small perimeter. The authors
show completeness and soundness of the test they design (the notion of completeness and
soundness is as in [19, 21]). The notion of testing used in their work is one of the main
differences with our work since we consider the perimeter testing in the more classical
framework of hypothesis testing (Proposition 1.9). We note that the completeness of [21] is
analogous to the type I error, but the soundness is fundamentally different from estimating
the type II error.
It is also worth mentioning the work of Belkin, Narayanan and Niyogi [5] where they
consider an algorithm that requires as few samples as possible in order to estimate the
perimeter of a convex body. Their results show that there is an algorithm that uses O(d4 γ −2 )
samples to obtain an estimator for the true perimeter of the convex set, with an error of
approximation of γ; this statement holds with high probability. We notice the polynomial
dependence on dimension in their estimates.
Let us now contrast the type of the convergence and the scaling regimes that we consider
in this paper with the ones needed for the convergence of graph-cut based machine learning
algorithms for clustering and related tasks.
Graph-cut based algorithms for tasks such as clustering have played an important role in
machine learning [3, 7, 14, 15, 18, 25, 26, 27, 28]. Data clustering algorithms are called
consistent if as the sample size n → ∞ their outputs converge to a desired partitioning of
the underlying measure being sampled. It is of interest to understand under what scaling
of εn on n does the consistency hold. Here we showed that for a fixed set Ω the value of
graph perimeter (and indeed of the graph-cut-based objective functionals such as Cheeger,
ratio, or normalized cuts) converges to the perimeter of Ω; this result holds even for rather
sparse graphs (Theorem 1.6). In particular, partitioning such sparse graphs (as on Figure
2) does not provide almost any information about the clusters present in the data. We
conclude that the convergence of the graph perimeter of a fixed set towards the continuum
perimeter does not provide the needed information on the asymptotic properties of graphcut based clustering algorithms. To obtain consistency of such algorithms one needs a
stronger notion of convergence of graph based functionals towards continuum functionals.
Recently the authors [10], and together with Laurent and Bresson [11] have developed the
appropriate notion of convergence (based on Γ-convergence from the calculus of variations),
and have applied it to consistency of Cheeger, ratio, sparsest, and normalized cut based
point cloud clustering. More specifically, in [11] it is shown that consistency holds if
(log(n))1/d
ε 1 (for d ≥ 3), in that regime the graphs are connected with high
n1/d
probability as n → ∞.
1.3. Outline of the argument. In order to understand the asymptotic behavior of the graph
perimeter, following [3] we first define a symmetric kernel φε : D × D → (0, ∞) by
φεn (x, y) =
1{||x−y||≤εn }
εd+1
n
|1Ω (x) − 1Ω (y)|.
ESTIMATING PERIMETER USING GRAPH CUTS
11
Using the kernel φεn , we can then write GPern,εn (Ω) as
(16)
n
n
X
X
2
GPern,εn (Ω) =
φε (xi , xj ),
n(n − 1) i=1 j=i+1 n
which is a U-statistic in the terminology of [16]. A simple computation shows that the mean
of this U -statistic, is the non-local perimeter Perε (Ω) defined in (7), that is,
Z Z
(17)
E(GPern,εn (Ω)) = Perεn (Ω) =
φε (x, y) dxdy.
D
D
Additionally, Remark 4.3 in [10] establishes that the non-local perimeter Perεn (Ω) approaches a constant multiple of the relative perimeter of Ω as the parameter εn goes to zero.
More precisely, if εn → 0 as n → ∞ then
(18)
Perεn (Ω) → σd Per(Ω)
as n → ∞
for σd the surface tension (4). This convergence also follows from the estimates in Appendix
A in the special case that Ω has a smooth boundary. Combining (18) with (17) we conclude
that if εn converges to zero as n → ∞ then
(19)
E(GPern,εn (Ω)) → σd Per(Ω)
as n → ∞.
Since the graph perimeter GPern,εn (Ω) is a U -statistic of order two we can use the general
theory of U -statistics to obtain moment estimates for GPern,εn (Ω). Let us first note that
Hoeffding’s decomposition theorem for U -statistics of order two (see [20]) implies that
GPern,εn (Ω) can be written as:
(20)
GPern,εn (Ω) − Perεn (Ω) = 2Un,1 + Un,2 ,
where Un,1 is a U-statistic of order one ( just a sum of centered independent random
variables) and Un,2 is a U-statistic of order two which is canonical or completely degenerate
(see [20]). In order to define the variables Un,1 and Un,2 , let us introduce the functions
Z
φ̄εn (x) :=
φεn (x, z) dz,
x ∈ D,
D
(21)
gn,1 (x) := φ̄εn (x) − Perεn (Ω),
x ∈ D,
gn,2 (x, y) := φεn (x, y) − φ̄εn (x) − φ̄εn (y) + Perεn (Ω),
x, y ∈ D.
With the previous definitions, we can now define
n
Un,1 =
(22)
Un,2 =
1X
gn,1 (xi ),
n i=1
X
2
n(n − 1)
gn,2 (xi , xj ).
1≤i<j≤n
R
R
We remark that D gn,1 (z)dz = 0 and that D gn,2 (x, z)dz = 0 for all x ∈ D. Because
of this, Un,1 and Un,2 are said to be canonical statistics of order one and two respectively
(see [20]). Now, Bernstein’s inequality [6] implies that
Cp
p
,
(23)
E(|Un,1 |p ) ≤ p max Apn,1 , Bn,1
n
where
√
(24)
An,1 := ||gn,1 ||∞ , Bn,1 := n||gn,1 ||2 .
12
NICOLÁS GARCÍA TRILLOS1 , DEJAN SLEPČEV2 AND JAMES VON BRECHT3
and Cp is a universal constant. See also [12] for a slight generalization of the previous
result.
On the other hand some of the moment estimates in [12] for canonical U -statistics of
order two can be used to prove that
(25)
E (|Un,2 |p ) ≤
Cp
p
p
, Cn,2
,
max Apn,2 , Bn,2
2p
n
where
(26)
An,2 := ||gn,2 ||∞ ,
Bn,2 := n||gn,2 ||2 ,
(Cn,2 )2 := n||
Z
2
gn,2
(·, y) dy||∞ .
D
and Cp is a universal constant. From the decomposition (20) it follows that for p ≥ 1
E (| GPern,εn (Ω) − Perεn (Ω)|p ) ≤ Cp (E (|Un,1 |p ) + E (|Un,2 |p )).
Thus in order to obtain the moment estimates for GPern,εn (Ω) in Theorem 1.1, we focus
on finding estimates for the quantities in (24) and (26).
Remark 1.12. The estimates on Un,1 and Un,2 exhibit a crossover in the nature when the
1
1
εn n1/d
parameter εn transitions between the sparse ( n2/(d+1)
) and the dense regime
1
( n1/d εn 1). From the theory of U -statistics the crossover in the nature of the bounds,
is connected to the different nature of the two components in the canonical decomposition
for U -statistics. Under the dense regime the biggest source of error comes from the term
Un,1 while in the sparse regime the biggest source of error comes from the term Un,2 . The
two variables Un,1 and Un,2 exhibit a different nature. In fact, one can think that Un,1
is a global quantity as it is the sum of averages, while Un,2 is simply the sum of pure
interactions concentrating on the boundary of the set Ω. We believe that there is a deeper
geometric and analytic reason for the scaling of the error that appears in the sparse regime
1
1
( n2/(d+1)
εn n1/d
). Moreover, we believe that such understanding would allow us to
complete the study of the asymptotic distribution in Theorem 1.8 for the sparse regime. We
expect such distribution to be of the Gaussian chaos type.
The bias estimates in Appendix A are obtained by a series of computations whose
starting point is writing Perε (Ω) in terms of an iterated integral, the outer one taken over
the manifold ∂Ω and the inner one taken along the normal line to ∂Ω at an arbitrary point
x ∈ ∂Ω. Such computations show that the first order term of Perεn (Ω) on εn vanishes.
Finally, Theorem 1.8 is obtained by using the canonical decomposition of U -statistics
1
and by noticing that in the dense regime n1/d
εn 1, the variable Un,2 is negligible in
relation to Un,1 . We make use of the CLT for triangular arrays after computing the variances
of the involved variables.
2. P ROOF OF T HEOREM 1.1
We first compute the moments of Un,1 and so we start computing the quantities An,1
and Bn,1 from (24). Denote by Tε the ε-tube around ∂Ω, that is, consider the set
(27)
Tε := x ∈ Rd : dist(x, ∂Ω) ≤ ε .
We also consider the half tubes Tε− and Tε+ ,
(28)
Tε− := {x ∈ Ω : dist(x, ∂Ω) ≤ ε} ,
Tε+ := {x ∈ Ωc : dist(x, ∂Ω) ≤ ε} .
ESTIMATING PERIMETER USING GRAPH CUTS
With these definitions it is straightforward to check that
d+1
|Bd (x, εn ) ∩ Ω|/εn
c
(29)
φ̄εn (x) = |Bd (x, εn ) ∩ Ω |/εd+1
n
0
13
if x ∈ Tε+n
if x ∈ Tε−n
if x 6∈ Tεn .
Since |Bd (x, εn ) ∩ Ω| and |Bd (x, εn ) ∩ Ωc | are bounded by αd εdn , where αd is the volume
of the d-dimensional unit ball, we deduce that
1
An,1 = O
.
εn
In order to compute the quantity Bn,1 we use the following lemma, whose proof may be
found in Appendix B.
Lemma 2.1. Let p ≥ 1 and let Ω ⊆ D, be a set with finite perimeter. Then, for all ε > 0
we have
Z
αp−1 σd
φ̄pεn (x)dx ≤ dp−1 Per(Ω).
ε
D
In particular, taking p = 1 in the previous expression, we obtain
Perε (Ω) ≤ σd Per(Ω).
(30)
R
Using the previous lemma with p = 2 we deduce that D φ̄2εn (x)dx = O ε1n , and
since
Z
Z
2
2
gn,1 (x)dx =
φ̄2εn (x)dx − (Perεn (Ω)) ,
D
D
we conclude that
r
Bn,1 = O
n
εn
.
From the previous computations, we deduce that
p
p
E(|Un,1 | ) ≤ Cp,d max{1, Per(Ω)} max
1
1
p ,
p
n εn , np/2 εp/2
n
where Cp,d depends on p and d, but is independent of the set Ω. If
in particular nε1n is o(1), then
(31)
E(|Un,1 |p ) ≤
Cp,d max{1, Per(Ω)}p
p/2
np/2 εn
1
n2/(d+1)
,
≤ εn , so that
.
R
Remark 2.2. Later on, in Lemma 3.1, we provide an explicit computation of D φ̄2εn (x)dx
up to order ε1n , which is useful when studying the asymptotic distribution of a rescaled
version of Un,1 .
Now we turn to the task of obtaining moment estimates for Un,2 . We estimate the
quantities An,2 , Bn,2 and Cn,2 from (26). Let us start by estimating An,2 . Note that for
any (x, y) ∈ D × D, φ̄εn (x) and φ̄εn (y) are of order ε1n and that Perεn (Ω) is of order one.
Thus, it is clear from the definition of gn,2 in (21) that
1
An,2 = O
.
εd+1
n
14
NICOLÁS GARCÍA TRILLOS1 , DEJAN SLEPČEV2 AND JAMES VON BRECHT3
1
φεn (x, y), we obtain that for every x ∈ D,
On the other hand, using φ2εn (x, y) = εd+1
n
Z
Z
2
gn,2
(x, y)dy =
φ2εn (x, y)dy − φ̄2εn (x) + 2θn φεn (x)
D
D
Z
Z
− 2 φεn (x, y)φ̄εn (y)dy +
φ̄2εn (y)dy − θn2
D
D
(32)
1
= d+1 φ̄εn (x) − φ̄2εn (x) + 2θn φεn (x)
εn
Z
Z
−2
φεn (x, y)φ̄εn (y)dy +
φ̄2εn (y)dy − θn2 ,
D
D
where we are using θn := Perεn (Ω). From this, it follows that
r
n
Cn,2 = O
.
εd+2
n
Finally, upon integration of (32) and direct computations, we obtain
Z
θn
2
||gn,2 ||2 = d+1 − 2
φ̄2εn (y)dy + θn2 ,
εn
D
which implies that
n
Bn,2 = O
.
(d+1)/2
εn
Thus, from (25) we deduce that
1
1
1
p
E(|Un,2 | ) ≤ Kp,d max
,
,
,
p(d+1)
p(d+1)/2
p(d+2)/2
n2p εn
np εn
n3p/2 εn
where Kp,d = Cp,d (max{1, Per(Ω)})p for Cp,d some constant that does not depend on the
1
set Ω. Hence, if n2/(d+1)
≤ εn , we have
E(|Un,2 |p ) ≤
(33)
Kp,d
p(d+1)/2
np εn
.
Combining (31) and (33) and using the canonical decomposition (20), we obtain (6).
2.1. Sharpness of the Rate in Theorem 1.1. A very simple argument shows that the rates
for εn that guarantee the almost sure convergence of the graph perimeter to the actual
perimeter in Theorem 1.1 are optimal in terms of scaling, up to logarithmic corrections.
In fact, suppose n2 εd+1
= o(1) and let en denote the random variable that counts the
n
number of edges that cross in the interface between Ω and its complement. In other words,
we define
n
n
X
X
en := εd+1
φεn (xi , xj ).
n
i=1 j=i+1
As a consequence, if Ω has finite perimeter then we have
(34)
GPern,εn (Ω) =
2
en ,
n(n − 1)εd+1
n
E(en ) =
n(n − 1)εd+1
n
Perεn (Ω).
2
Note that en takes integer values in the range {0, 1, . . . , N } for N = n(n − 1)/2, so
that
N
X
E(en ) =
kpnk
pnk := P(en = k).
k=1
ESTIMATING PERIMETER USING GRAPH CUTS
15
The fact that pn0 + · · · + pnN = 1 implies
E(en ) =
N
X
kpnk ≥
k=1
N
X
pnk = (1 − pn0 ).
k=1
In particular, from (34) and (18) we deduce that if n2 εd+1
→ 0 and Ω has finite perimeter
n
then
(1 − pn0 ) ≤ E(en ) = o(1).
On the other hand, note that for any given γ > 0 it is true that GPern,εn (Ω) > γ implies
that en 6= 0. In turn
P (GPern,εn (Ω) > γ) ≤ P (en 6= 0) = 1 − pn0 = o(1).
We conclude that if n2 εd+1
→ 0 then GPern,εn (Ω) converges in probability to zero.
n
Therefore, if Ω has a non-zero, finite perimeter then GPern,εn (Ω) does not converge to
σd Per(Ω) in probability (nor almost surely, either).
3. P ROOF OF T HEOREM 1.8
The proof of Theorem 1.8 relies on the following lemma, whose proof may be found in
Appendix C.
Lemma 3.1. Asssume that εn → 0 as n → ∞, and that Ω ⊂ D is an open set with smooth
boundary so that dist(Ω, ∂D) > 0. Then
(35)
Var(gn,1 (X1 )) =
Cd Per(Ω)
+ O(1),
εn
where Cd is given by (10).
Now we turn our attention to the proof of Theorem 1.8.
Proof of Theorem 1.8. Note that from (20), (21) and (22) we obtain
r
r
n
X
nεn
εn
(GPern,εn (Ω) − Perεn (Ω)) =
gn,1 (xi )
4Cd Per(Ω)
nCd Per(Ω) i=1
r
nεn
+
Un,2 .
4Cd Per(Ω)
From the moment estimates (33), we deduce that
r
nεn
P
(36)
Un,2 → 0.
4Cd Per(Ω)
On the other hand, we note that from (35)
√
std(gn,1 (x1 )) εn
p
→ 1,
Cd Per(Ω)
as n → ∞,
where std(gn,1 (x1 )) is the standard deviation of gn,1 (x1 ). Lyapunov’s condition which is
sufficient to allow us to use the central limit theorem for triangular arrays is easily checked
from Lemma 2.1. We deduce that
r
n
X
εn
w
gn,1 (Xi ) −→ N (0, 1).
nCd Per(Ω) i=1
16
NICOLÁS GARCÍA TRILLOS1 , DEJAN SLEPČEV2 AND JAMES VON BRECHT3
Combining with (36) and the Slutsky’s theorem, we obtain the desired result. Finally, to
obtain the last statement in the theorem, we note that from the bias estimates in Lemma 1.4,
√
nεn | Perεn (A) − σd Per(A)| = O(n1/2 ε5/2
n )
√
Under the condition (11), we conclude that nεn | Perεn (A) − ση Per(A)| → 0. This
implies (12).
3.1. Application to Perimeter Testing. Here we prove Proposition 1.9. We assume that
Ω ⊆ D is an open set with smooth boundary such that Ω ⊂ D. Note that under the null
hypothesis, if ln > Zα , then,
r
nεn
(37)
Zα < ln ≤
(GPern,εn (Ω) − Perεn (Ω)) ,
4Cd Per(Ω)
where we used that Perεn (Ω) ≤ σd Per(Ω) by Lemma 2.1. Thus, using Theorem 1.8, we
deduce that asymptotically, the type I error of our test is
r
nεn
PH0 (ln > Zα ) ≤ P
(GPern,εn (Ω) − Perεn (Ω)) > Zα
4Cd Per(Ω)
−→ P (N (0, 1) > Zα ) = α,
which establishes the first part of Proposition 1.9. In order to compute the type II error of
our test, suppose that Per(Ω) = ρ0 where ρ0 > ρ. In that case,
(38)
r
nεn
(σd ρ − Pern,εn (Ω)) ≥ −Zα
PHA (ln ≤ Zα ) = PHA
4Cd ρ
r
r
nεn
nεn
= PHA
(Perεn (Ω) − Pern,εn (Ω)) ≥ −Zα +
(Perεn (Ω) − σd ρ) ,
4Cd ρ
4Cd ρ
Now recall that limn→∞ Perεn (Ω) = σd Per(Ω) = σd ρ0 > σd ρ. In particular, we deduce
that
r
nεn
lim
(Perεn (Ω) − σd ρ) = +∞.
n→∞
4Cd ρ
Thus, for large enough n,
r
r
nεn
1
nεn
−Zα +
(Perεn (Ω) − σd ρ) ≥
(σd ρ0 − σd ρ).
4Cd ρ
2 4Cd ρ
Hence, for large enough n,
r
r
nεn
1
nεn
PHA (ln ≤ Zα ) ≤ PHA
(Perεn (Ω) − Pern,εn (Ω)) ≥
(σd ρ0 − σd ρ).
4Cd ρ
2 4Cd ρ
Using the moment estimates from Theorem 1.1, and Markov’s inequality, we deduce that
1
.
PHA (ln ≤ Zα ) = O √
nεn
That is, the type II error is of order
√1 .
nεn
ESTIMATING PERIMETER USING GRAPH CUTS
17
A PPENDIX A. P ROOF OF L EMMA 1.4
Since Ω ⊂⊂ D and Ω has smooth boundary the relative perimeter of Ω with respect to
D in the generalized sense (2) simply corresponds to the usual perimeter of ∂Ω in the sense
that
Z
dHd−1 = Hd−1 (∂Ω).
Per(Ω) =
∂Ω
Additionally, for all ε ≤ δ := dist(Ω, ∂D) we have that
Z
2
Perε (Ω) = d+1
|Bd (x, ε) ∩ Ωc | dx,
ε
Ω
where Bd (x, r) denotes the ball of radius r in Rd centered at x and Ωc denotes the complement of Ω in all of space. Moreover, since ∂Ω is a compact smooth manifold, we can
assume without the loss of generality ( by taking ε small enough) that for every x ∈ Tε
there is a unique point P (x) in ∂Ω closest to x. Furthermore, we can assume that the map
P is smooth. We may further write
Perε (Ω) =
2
εd+1
Z
Tε−
|Bd (x, ε) ∩ Ωc | dx,
where Tε− is defined in (28). This reformulation makes it natural to write the previous
integral as an iterated integral; the outer integral is taken over the manifold ∂Ω and the inner
integral is taken along the normal line to ∂Ω at an arbitrary point x along the boundary.
To make this idea precise, we first let N(x) denote the outer unit normal to ∂Ω at x ∈ ∂Ω
and then consider the transformation (x, t) ∈ ∂Ω×(0, 1) 7→ x−tεN(x) for all ε sufficiently
small. The Jacobian of this transformation equals ε det(I + tεSx ), where Sx denotes the
shape operator (or second fundamental form) of ∂Ω at x, see [13] for instance. For all ε
sufficiently small, we may therefore conclude that
Z
Z Z 1
1
|Bd (x, ε)∩Ωc | dx =
|Bd (x − tεN(x), ε) ∩ Ωc | det(I + tεSx )dt dHd−1 (x).
ε Tε−
∂Ω
0
As a consequence, we also have that
(39)
Z Z 1
2
|Bd (x − tεN(x), ε) ∩ Ωc | det(I + tεSx ) dt dHd−1 (x).
Perε (Ω) = d
ε ∂Ω
0
With the expression (39) in hand, we may now proceed to establish (9) by expanding
Perε (Ω) in terms of ε and appealing to some elementary computations that show that the
first order term in ε vanishes.
For a fixed x ∈ ∂Ω, we first wish to understand the behavior of the function
Z 1
1
gx (ε) := d
|Bd (x − tεN(x), ε) ∩ Ωc | det(I + tεSx ) dt
ε
0
for ε in a neighborhood of zero. Without loss of generality, we may assume that x = 0,
that N(x) = ed and that around x the boundary ∂Ω coincides with the graph x̂ =
(x1 , . . . , xd−1 ) 7→ (x̂, f (x̂)) ∈ Rd of a smooth function f (x̂) that satisfies both f (0) = 0
and ∇f (0) = 0 simultaneously. By symmetry of the shape operator Sx , there exists an orthonormal basis for Rd−1 (where we identify Rd−1 with the hyperplane {(x̂, xd ) : xd = 0})
consisting of eigenvectors of the shape operator. We let v1 , . . . , vd−1 denote the eigenvectors of Sx and κ1 , . . . , κd−1 the corresponding eigenvalues ( also known as principal
18
NICOLÁS GARCÍA TRILLOS1 , DEJAN SLEPČEV2 AND JAMES VON BRECHT3
curvatures). In particular, whenever kx̂k ≤ ε we have that
d−1
f (x̂) =
(40)
1X
κi hx̂, vi i2 + O(ε3 )
2 i=1
where curvatures κi = κi (x) and the O(ε3 ) error term canp
be uniformly bounded.
With these reductions in place, we first define u(ŷ) := ε2 − kŷk2 and then let
if f (ŷ) + εt < −u(ŷ)
2u(ŷ)
h(ŷ, t; ε) := u(ŷ) − εt − f (ŷ) if − u(ŷ) ≤ f (ŷ) + εt ≤ u(ŷ)
0
otherwise.
A direct calculation then shows that
(41)
c
Z
|Bd (x − tεN(x), ε) ∩ Ω | =
h(ŷ, t; ε) dŷ,
Bd−1 (0,ε)
p
and an application of (40) shows that h(ŷ, t; ε) = 2 ε2 − kŷk2 only if
kŷk2 = ε2 − O(ε4 )
It therefore follows that
Z
h(ŷ, t; ε) dŷ ≤ O(ε2 )
Bd−1 (0,ε)∩{f (ŷ)+εt<−u(ŷ)}
u(ŷ) = O(ε2 ).
and
Z
√
Bd−1 (0,ε)∩{kŷk≥
dŷ = O(εd+3 ).
ε2 −O(ε4 )}
We then let Aεt denote the set Aεt := {ŷ ∈ Bd−1 (0, ε) : −u(ŷ) ≤ f (ŷ) + εt ≤ u(ŷ)} and
use the previous estimate in (41) to uncover
Z
(42) |Bd (x − tεN(x), ε) ∩ Ωc | =
(u(ŷ) − εt − f (ŷ)) dŷ + O(εd+3 ).
Bd−1 (0,ε)∩Aεt
We may then note that
det(I + εtSx ) = (1 + tεκ1 ) . . . (1 + tεκd−1 ) = 1 + tεHx + O(ε2 ),
Pd−1
where Hx := i=1 κi represents the mean curvature. Using this fact in (42) then yields
!
Z 1 Z
1
gx (ε) = d
u(ŷ) − εt − f (ŷ) dŷ (1 + tεHx ) dt + O(ε2 ).
ε 0
Bd−1 (0,ε)∩Aεt
Now let f ε (z) := 1ε f (εz) and define the corresponding subset Ctε of (0, 1) × Bd−1 (0, 1) as
n
o
p
p
Ctε := (t, z) ∈ (0, 1) × Bd−1 (0, 1) : − 1 − kzk2 ≤ f ε (z) + t ≤ 1 − kzk2 ,
then make the change of variables ŷ = εz to see that
Z p
gx (ε) =
1 − kzk2 − t − f ε (z) (1 + tεHx ) dzdt + O(ε2 ).
Ctε
Recalling (40) shows that
d−1
(43)
f ε (z) =
εX
κi hz, vi i2 + O(ε2 ),
2 i=1
ESTIMATING PERIMETER USING GRAPH CUTS
19
which then allows us to obtain an expansion of gx (ε) in terms of ε according to the relation
Z p
gx (ε) =
1 − kzk2 − t dtdz
Ctε
d−1
p
1X
tHx ( 1 − kzk2 − t) −
κi hz, vi i2
2 i=1
Z
(44)
+ε
Ctε
!
dtdz + O(ε2 ).
The bias estimate (9) then directly follows after computing each of these terms individually.
We begin by considering the first term in the expansion, i.e.
Z p
I :=
1 − kzk2 − t dtdz.
Ctε
p
Given ε > 0 p
and z ∈ Bd−1 (0, 1) define c(z) := max{− 1 − kzk2 − f ε (z), 0} and
C(z) := min{ 1 − kzk2 − f ε (z), 1}, so that we may easily write
Z
p
C(z) + c(z)
I=
(C(z) − c(z))
1 − kzk2 −
dz.
2
Bd−1 (0,1)
As the set where c(z) 6= 0 has measure at most O(ε2 ), we easily conclude that
Z
p
C(z)
I=
C(z)
1 − kzk2 −
dz + O(ε2 ).
2
Bd−1 (0,1)
p
If C(z) = 1 then 1 − kzk2 − C(z)
= 12 (1 − kzk2 ) + O(ε2 ) as well. In any case, it
2
follows that
Z
1
σd
I=
(45)
(1 − kzk2 ) dz + O(ε2 ) =
+ O(ε2 ).
2 Bd−1 (0,1)
2
We now proceed to compute the second term in the expansion
Z p
II :=Hx
t 1 − kzk2 − t2 dtdz
Ctε
Z
=Hx
p
2
C (z)
Bd−1 (0,1)
1 − kzk C(z)
−
2
3
!
dz + O(ε2 )
and the third term in the expansion
Z
Z
d−1
d−1
1X
1X
III :=
κi
hz, vi i2 dtdz =
κi
hz, vi i2 C(z) dz + O(ε2 )
2 i=1
2 i=1
Ctε
Bd−1 (0,1)
p
in a similar fashion. We always have C(z) = 1 − kzk2 + O(ε), so that
Z
Hx
(46)
II =
(1 − kzk2 )3/2 dz + O(ε)
6 Bd−1 (0,1)
Z
Hx vol(S d−2 ) 1
(47)
=
(1 − r2 )3/2 rd−2 dr + O(ε)
6
0
The third term follows similarly by appealing to spherical coordinates, in that we have
Z
d−1
p
1X
III =
κi
1 − kzk2 hz, vi i2 dz + O(ε)
2 i=1
Bd−1 (0,1)
Z
Hx vol(S d−2 ) 1 p
=
1 − r2 rd dr + O(ε) = II + O(ε)
2(d − 1)
0
20
NICOLÁS GARCÍA TRILLOS1 , DEJAN SLEPČEV2 AND JAMES VON BRECHT3
thanks to an integration by parts in the final term. We therefore have that I = σd /2 + O(ε2 )
and II − III = O(ε), so that gx (ε) = σd /2 + O(ε2 ) and
Z
Perε (Ω) = 2
gx (ε) dHd−1 = σd Per(Ω) + O(ε2 )
∂Ω
as desired.
We may also show that when Ω is a fixed ball, say Ω = Bd (xc , 31 ) for xc ∈ Rd the
center point of D, that the absolute value of the difference between Perε (Ω) and σd Per(Ω)
remains bounded from below by cε2 for c > 0 some positive constant. The proof proceeds
similarly to the proof of the bias estimate above. In particular, this shows that the bound in
Lemma 1.4 is optimal in terms of scaling for general sets with smooth boundary.
A PPENDIX B. P ROOF OF L EMMA 2.1
The proof follows the same argument used in the proof of Theorem 4.1 in [10] or
Theorem 6.2 in [1]. We assume that dist(Ω, ∂D) > 0. Such assumption implies that the
perimeter of Ω with respect to D, that is Per(Ω) defined in (2), is equal to the perimeter
of Ω with respect to Rd . We remark that a slight modification of the argument we present
below proves the result in the general case and hence we omit the details (see for example
the proof of Theorem 4.1 in [10]).
First we prove that for any function u : Rd → [0, 1] with u ∈ W 1,1 (Rd ) ∩ C ∞ (Rd ) and
for all ε > 0 we have
p
Z
Z Z
1kx−yk≤ε
αdp−1 σd
|u(y) − u(x)|dy dx ≤ p−1
k∇u(x)kdx,
(48)
εd+1
ε
Rd
Rd
Rd
Inequality (48) follows from
p
Z Z
1kx−yk≤ε
|u(y)
−
u(x)|dy
dx
εd+1
Rd
Rd
!p
Z
Z
1
|u(x + εh) − u(x)|dh dx
= p
ε Rd
Bd (0,1)
Z Z
αp−1
≤ dp
|u(x + εh) − u(x)|p dh dx
ε
Rd Bd (0,1)
Z Z
αp−1
≤ dp
|u(x + εh) − u(x)|dh dx
ε
Rd Bd (0,1)
Z Z
Z 1
αdp−1
= p−1
∇u(x + tεh) · hdt dh dx
ε
Rd Bd (0,1)
0
Z Z
Z 1
αdp−1
|∇u(x + tεh) · h|dt dh dx
≤ p−1
ε
Rd Bd (0,1) 0
Z Z
Z
αdp−1 1
= p−1
|∇u(x) · h|dx dh dt
ε
0
Bd (0,1) Rd
Z Z
Z
αdp−1 1
∇u(x)
= p−1
k∇u(x)k
· h dh dx dt
ε
0
Rd
Bd (0,1) k∇u(x)k
Z
αdp−1 σd
= p−1
k∇u(x)kdx
ε
Rd
ESTIMATING PERIMETER USING GRAPH CUTS
21
where in the first equation we used the change of variables h = x−y
ε , in the first inequality
we used Jensen’s inequality and in the second inequality the fact that u takes values in [0, 1].
Now, for any set Ω ⊆ D as in the statement, we can find a sequence of functions
{uk }k∈N with uk : Rd → [0, 1] , uk ∈ W 1,1 (Rd ) ∩ C ∞ (Rd ) and such that
Z
L1 (Rd )
(49)
uk −→ 1Ω , lim
k∇uk (x)kdx = Per(Ω).
k→∞
Rd
Such sequence can be obtained for example with the aid of standard mollifiers (see Theorem
13.9 in [22]). It follows from (48) and from (49) that
p
Z Z
1kx−yk≤ε
αdp−1 σd
dx
≤
|1
(y)
−
1
(x)|dy
Per(Ω).
Ω
Ω
εd+1
εp−1
Rd
Rd
Finally, notice that
Z
Z
φ̄pεn (x)dx ≤
D
Rd
Z
Rd
1kx−yk≤ε
|1Ω (y) − 1Ω (x)|dy
εd+1
p
dx ≤
αdp−1 σd
Per(Ω).
εp−1
A PPENDIX C. P ROOF OF L EMMA 3.1
The proof is based on similar computations to the ones in Appendix A and thus we
simply highlight the main ideas. First of all note that
Z
2
Var (gn,1 (X1 )) =
φ̄2εn (x)dx − (Perεn (Ω)) .
D
Since Perεn (Ω) = O(1), our task reduces to computing the integral in the above expression.
It follows from (29) that for all εn small enough (so that Tεn ⊆ D),
Z
Z
Z
1
1
|Bd (x, εn ) ∩ Ωc |2 dx + 2(d+1)
|Bd (x, εn ) ∩ Ω|2 dx.
φ̄2εn (x)dx = 2(d+1)
Tε−n
Tε+n
D
εn
εn
We compute the first of the integrals from the above expression. As in the proof of Lemma
1.4, we write
Z
1
|Bd (x, εn ) ∩ Ωc |2 dx
2(d+1)
Tε−n
εn
Z
Z 1
1
1
c 2
=
|B
(x
−
tε
N(x),
ε
)
∩
Ω
|
det(I
+
tε
S
)
dt
dHd−1 (x).
d
n
n
n x
εn ∂Ω ε2d
0
n
For x ∈ ∂Ω we study the expression
Z 1
|Bd (x − tεn N(x), εn ) ∩ Ωc |2
det(I + tεn Sx )dt.
hεn (x) :=
ε2d
0
n
Note that det(I +tεn Sx ) = (1+tεn κ1 )···(1+tεn κd−1 ) = 1+O(εn ), where κ1 , . . . , κd−1
are the principal curvatures (eigenvalues of the shape operator Sx ). Hence,
Z 1
|Bd (x − tεn N(x), εn ) ∩ Ωc |2
hεn (x) =
dt + O(εn ).
ε2d
0
n
Without loss of generality, we may assume that x = 0, that N(x) = ed and that around x the
boundary ∂Ω coincides with the graph x̂ = (x1 , . . . , xd−1 ) 7→ (x̂, f (x̂)) ∈ Rd of a smooth
function f (x̂) that satisfies both f (0) = 0 and ∇f (0) = 0 simultaneously, and we denote by
v1 , . . . , vd−1 the eigenvectors of Sx (just as in Appendix A). Then, f : Bd−1 (0, εn ) → R,
Pd−1
2
satisfies f (x̂) = i=1 κi hx̂, vi i + O(ε3n ).
NICOLÁS GARCÍA TRILLOS1 , DEJAN SLEPČEV2 AND JAMES VON BRECHT3
22
Now for fixed t ∈ [0, 1] we define H1−t to be the hyperplane
H1−t := {x = (x̂, xd ) : xd ≥ 1 − t} ,
and we let
At := |Bd (0, 1) ∩ H1−t |.
With these definitions we can write
(50)
Z
d−1
X
|Bd (x − tεn N(x), εn ) ∩ Ωc | = εdn At +
κi
= εdn At +
Bd−1 (0,εn
i=1
Pd−1 Z
i=1 κi
d−1
2
√
hx̂, vi i dx̂ + O(εd+2
n )
1−t)
Bd−1 (0,εn
√
kx̂k2 dx̂ + O(εd+2
n )
1−t)
= εdn At + O(εd+1
n ),
where the second
equality holds due to symmetry, and where the last equality follows after
R
computing Bd−1 (0,εn √1−t) kx̂k2 dx̂ using polar coordinates. From the above, it follows
that
|Bd (x − tεn N(x), εn ) ∩ Ωc |2
= A2t + O(εn ).
ε2d
n
We conclude that
Z
1
A2t dt + O(εn ).
hεn (x) =
0
Therefore,
1
2(d+1)
εn
Z
Tε−n
|Bd (x, εn ) ∩ Ωc |2 dx =
1
εn
Analogously, we can obtain a similar expression for
from this we deduce (35).
Z
1
A2t dt Per(Ω) + O(1).
0
1
2(d+1)
εn
R
Tε+n
|Bd (x, εn ) ∩ Ω|2 dx and
Acknowledgements. DS and NGT are grateful to NSF for its support (grants DMS1211760 and DMS-1516677). JvB was supported by NSF grant DMS 1312344/DMS
1521138. The authors are grateful to ICERM (supported by NSF grant 0931908), where
part of the research was done during the research cluster: Geometric analysis methods for
graph algorithms. Furthermore they are grateful to Ki-Net (NSF Research Network Grant
RNMS11-07444) for opportunities provided. The authors would like to thank the Center
for Nonlinear Analysis of the Carnegie Mellon University for its support.
R EFERENCES
[1] G. A LBERTI AND G. B ELLETTINI, A non-local anisotropic model for phase transitions: asymptotic
behaviour of rescaled energies, European J. Appl. Math., 9 (1998), pp. 261–284.
[2] L. A MBROSIO , N. F USCO , AND D. PALLARA, Functions of bounded variation and free discontinuity
problems, Oxford Mathematical Monographs, The Clarendon Press Oxford University Press, New York,
2000.
[3] E. A RIAS -C ASTRO , B. P ELLETIER , AND P. P UDLO, The normalized graph cut and Cheeger constant: from
discrete to continuous, Adv. in Appl. Probab., 44 (2012), pp. 907–937.
[4] I. A RMEND ÁRIZ , A. C UEVAS , AND R. F RAIMAN, Nonparametric estimation of boundary measures and
related functionals: asymptotic results, Adv. in Appl. Probab., 41 (2009), pp. 311–322.
[5] M. B ELKIN , H. NARAYANAN , AND P. N IYOGI, Heat flow and a faster algorithm to compute the surface
area of a convex body, in FOCS ’06: Proceedings of the 47th Annual IEEE Symposium on Foundations of
Computer Science (FOCS’06), Washington, DC, USA, 2006, IEEE Computer Society, pp. 47–56.
ESTIMATING PERIMETER USING GRAPH CUTS
23
[6] S. B ERNSTEIN, On a modification of Chebyshev’s inequality and of the error formula of Laplace, Ann. Sci.
Inst. Sav. Ukraine, Sect. Math, 1 (1924), pp. 38–49.
[7] X. B RESSON , T. L AURENT, D. U MINSKY, AND J. H. VON B RECHT, Convergence and energy landscape
for cheeger cut clustering, in Advances in Neural Information Processing Systems (NIPS), P. L. Bartlett,
F. C. N. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger, eds., 2012, pp. 1394–1402.
[8] A. C UEVAS , R. F RAIMAN , AND L. G Y ÖRFI, Towards a universally consistent estimator of the Minkowski
content, ESAIM Probab. Stat., 17 (2013), pp. 359–369.
[9] A. C UEVAS , R. F RAIMAN , AND A. RODR ÍGUEZ -C ASAL, A nonparametric approach to the estimation of
lengths and surface areas, Ann. Statist., 35 (2007), pp. 1031–1051.
[10] N. G ARC ÍA T RILLOS AND D. S LEP ČEV, Continuum limit of total variation on point clouds, Archive for
Rational Mechanics and Analysis, (2015), pp. 1–49. online first.
[11] N. G ARC ÍA T RILLOS , D. S LEP ČEV, J. VON B RECHT , T. L AURENT, AND X. B RESSON, Consistency of
Cheeger and ratio graph cuts. to appear in Journal of Machine Learning Research, 2015.
[12] E. G IN É , R. L ATAŁA , AND J. Z INN, Exponential and moment inequalities for U -statistics, in High
dimensional probability, II (Seattle, WA, 1999), vol. 47 of Progr. Probab., Birkhäuser Boston, Boston, MA,
2000, pp. 13–38.
[13] A. G RAY, Tubes, Progress in Mathematics, Birkhäuser Basel, 2004.
[14] L. H AGEN AND A. K AHNG, New spectral methods for ratio cut partitioning and clustering., IEEE Trans.
Computer-Aided Design, 11 (1992), pp. 1074 –1085.
[15] M. H EIN AND S. S ETZER, Beyond spectral clustering - tight relaxations of balanced graph cuts, in Advances
in Neural Information Processing Systems (NIPS), J. Shawe-Taylor, R. Zemel, P. Bartlett, F. Pereira, and
K. Weinberger, eds., 2011, pp. 2366–2374.
[16] W. H OEFFDING, Probability inequalities for sums of bounded random variables, Journal of the American
statistical association, 58 (1963), pp. 13–30.
[17] R. J IM ÉNEZ AND J. E. Y UKICH, Nonparametric estimation of surface integrals, Ann. Statist., 39 (2011),
pp. 232–260.
[18] R. K ANNAN , S. V EMPALA , AND A. V ETTA, On clusterings: Good, bad and spectral, Journal of the ACM
(JACM), 51 (2004), pp. 497–515.
[19] M. K EARNS AND D. RON, Testing problems with sublearning sample complexity, Journal of Computer and
System Sciences, 61 (2000), pp. 428 – 456.
[20] V. S. KOROLJUK AND Y. V. B OROVSKICH, Theory of U -statistics, vol. 273 of Mathematics and its
Applications, Kluwer Academic Publishers Group, Dordrecht, 1994. Translated from the 1989 Russian
original by P. V. Malyshev and D. V. Malyshev and revised by the authors.
[21] P. KOTHARI , A. NAYYERI , R. O’D ONNELL , AND C. W U, Testing surface area, in Proceedings of the
Twenty-Fifth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ’14, SIAM, 2014, pp. 1204–
1214.
[22] G. L EONI, A first course in Sobolev spaces, vol. 105 of Graduate Studies in Mathematics, American
Mathematical Society, Providence, RI, 2009.
[23] J. N EEMAN, Testing surface area with arbitrary accuracy, in Proceedings of the 46th Annual ACM Symposium on Theory of Computing, STOC ’14, New York, NY, USA, 2014, ACM, pp. 393–397.
[24] B. PATEIRO -L ÓPEZ AND A. RODR ÍGUEZ -C ASAL, Length and surface area estimation under smoothness
restrictions, Adv. in Appl. Probab., 40 (2008), pp. 348–358.
[25] J. S HI AND J. M ALIK, Normalized cuts and image segmentation, Pattern Analysis and Machine Intelligence,
IEEE Transactions on, 22 (2000), pp. 888–905.
[26] A. S ZLAM AND X. B RESSON, Total variation and cheeger cuts., in ICML, J. Faenkranz and T. Joachims,
eds., Omnipress, 2010, pp. 1039–1046.
[27] U. VON L UXBURG, A tutorial on spectral clustering, Statistics and computing, 17 (2007), pp. 395–416.
[28] Y.-C. W EI AND C.-K. C HENG, Towards efficient hierarchical designs by ratio cut partitioning, in ComputerAided Design, 1989. ICCAD-89. Digest of Technical Papers., 1989 IEEE International Conference on, IEEE,
1989, pp. 298–301.
1
D IVISION OF A PPLIED M ATHEMATICS , P ROVIDENCE , RI, 02912, USA., EMAIL : NICOLAS GARCIA TRILLOS @ BROWN . EDU
2
D EPARTMENT OF M ATHEMATICAL S CIENCES , C ARNEGIE M ELLON U NIVERSITY, P ITTSBURGH , PA,
15213, USA., TEL . +412 268-2545, EMAIL : SLEPCEV @ MATH . CMU . EDU
3 D EPARTMENT OF M ATHEMATICS AND S TATISTICS , C ALIFORNIA S TATE U NIVERSITY, L ONG B EACH
L ONG B EACH , CA 90840, USA., EMAIL : JAMES . VON B RECHT @ CSULB . EDU
| 10math.ST
|
MONOIDS OF MODULES AND ARITHMETIC OF DIRECT-SUM
DECOMPOSITIONS
arXiv:1401.6553v1 [math.AC] 25 Jan 2014
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
Abstract. Let R be a (possibly noncommutative) ring and let C be a class of finitely generated (right)
R-modules which is closed under finite direct sums, direct summands, and isomorphisms. Then the set
V(C) of isomorphism classes of modules is a commutative semigroup with operation induced by the direct
sum. This semigroup encodes all possible information about direct sum decompositions of modules in C.
If the endomorphism ring of each module in C is semilocal, then V(C) is a Krull monoid. Although this
fact was observed nearly a decade ago, the focus of study thus far has been on ring- and module-theoretic
conditions enforcing that V(C) is Krull. If V(C) is Krull, its arithmetic depends only on the class group
of V(C) and the set of classes containing prime divisors. In this paper we provide the first systematic
treatment to study the direct-sum decompositions of modules using methods from Factorization Theory
of Krull monoids. We do this when C is the class of finitely generated torsion-free modules over certain
one- and two-dimensional commutative Noetherian local rings.
1. Introduction
The study of direct-sum decompositions of finitely generated modules is a classical topic in Module
Theory dating back over a century. In the early 1900s, Wedderburn, Remak, Krull, and Schmidt proved
unique direct-sum decomposition results for various classes of groups (see [66], [59], [53], and [65]). In the
middle of the last century, Azumaya [2] proved uniqueness of (possibly infinite) direct-sum decomposition
of modules provided that each indecomposable module has a local endomorphism ring. In the commutative setting, Evans [20] gave an example due to Swan illustrating a non-unique direct-sum decomposition
of a finitely-generated module over a local ring. The past decade has seen a new semigroup-theoretical
approach. This approach was first introduced by Facchini and Wiegand [29] and has been used by several
authors (for example, see [3], [4], [7], [8], [19], [22], [25], [26], [27], [28], [29], [46], [49], and [57]). Let R be
a ring and let C be a class of right R-modules which is closed under finite direct sums, direct summands,
and isomorphisms. For a module M in C, let [M ] denote the isomorphism class of M . Let V(C) denote
the set of isomorphism classes of modules in C. (We assume here that V(C) is indeed a set, and note that
this hypothesis holds for all examples we study.) Then V(C) is a commutative semigroup with operation
defined by [M ] + [N ] = [M ⊕ N ] and all information about direct-sum decomposition of modules in C
can be studied in terms of factorization of elements in the semigroup V(C). In particular, the direct-sum
decompositions in C are (essentially) unique (in other words, the Krull-Remak-Schmidt-Azumaya Theorem — briefly, KRSA — holds) if and only if V(C) is a free abelian monoid. This semigroup-theoretical
point of view was justified by Facchini [22] who showed that V(C) is a reduced Krull monoid provided
that the endomorphism ring EndR (M ) is semilocal for all modules M in C. This result allows one to
describe the direct-sum decomposition of modules in terms of factorization of elements in Krull monoids,
a well-studied class of commutative monoids.
2010 Mathematics Subject Classification. 13C14, 16D70, 20M13.
Key words and phrases. Krull monoids, sets of lengths, direct-sum decompositions, indecomposable modules.
The first author was a Fulbright-NAWI Graz Visiting Professor in the Natural Sciences and supported by the AustrianAmerican Education Commission. The second author was supported by the Austrian Science Fund FWF, Project Number
P21576-N18.
1
2
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
However, thus far much of the focus in this direction has been on the study of module-theoretic
conditions which guarantee that all endomorphism rings are semilocal, as well as on trying to describe
the monoid V(C) in terms of various ring- and module-theoretic conditions. Although some factorizationtheoretic computations have been done in various settings (eg. the study of elasticity in [4], [7], and [8] and
the study of the ω-invariant in [19]), the general emphasis has not been on the arithmetic of the monoid
V(C). Our intent is to use known module-theoretic results along with factorization-theoretic techniques
in order to give detailed descriptions of the arithmetic of direct-sum decompositions of finitely generated
torsion-free modules over certain one- and two-dimensional local rings. We hope that this systematic
approach will not only serve to inspire others to consider more detailed and abstract factorizationtheoretic approaches to the study of direct-sum decompositions, but to provide new and interesting
examples for zero-sum theory over torsion-free groups. We refer to [23] and to the opening paragraph in
the recent monograph [56] for broad information on the Krull-Remak-Schmidt-Azumaya Theorem, and
to two recent surveys [26] and [9] promoting this semigroup-theoretical point of view. More details and
references will be given in Section 3.
Krull monoids, both their ideal theory and their arithmetic, are well-studied; see [37] for a thorough
treatment. A reduced Krull monoid is uniquely determined (up to isomorphism) by its class group G, the
set of classes GP ⊂ G containing prime divisors, and the number of prime divisors in each class. Let V(C)
be a monoid of modules and suppose V(C) is Krull with class group G and with set of classes containing
prime divisors GP . We are interested in determining what this information tells us about direct-sum
decompositions of modules. Let M be a module in C and let M = M1 ⊕ · · · ⊕ Mℓ where M1 , . . . , Mℓ
are indecomposable right R-modules. Then ℓ is called the length of this factorization (decomposition
into indecomposables), and the set of lengths L(M ) ⊂ N is defined as the set of all possible factorization
lengths. Then KRSA holds if and only if |G| = 1. Moreover, it is easy to check that |L(M )| = 1 for all M
in C provided that |G| ≤ 2. Clearly, sets of lengths are a measure how badly KRSA fails. Assuming that
V(C) is Krull, M has at least one direct-sum decomposition in terms of indecomposable right R-modules,
and, up to isomorphism, only finitely many distinct decompositions. In particular, all sets of lengths are
finite and nonempty. Without further information about the class group G and the subset GP ⊂ G, this
is all that can be said. Indeed, there is a standing conjecture that for every infinite abelian group G there
is a Krull monoid with class group G and set GP such that every set of lengths has cardinality one (see
[35]). On the other hand, if the class group of a Krull monoid is infinite and every class contains a prime
divisor, then every finite subset of N≥2 occurs as a set of lengths (see Proposition 6.2).
Thus an indispensable prerequisite for the study of sets of lengths (and other arithmetical invariants)
in Krull monoids is detailed information about not only the class group G, but also on the set GP ⊂ G
of classes containing prime divisors. For the monoid V(C), this is of course a module-theoretic task which
depends on both the ring R and the class C of R-modules. Early results gave only extremal sets GP
and thus no further arithmetical investigations were needed. In Sections 4 and 5 we determine, based
on deep module-theoretic results, the class group G of V(C). We then exhibit well-structured sets GP
providing a plethora of arithmetically interesting direct-sum decompositions. In particular, we study
the classes of finitely generated modules, finitely generated torsion-free modules, and maximal CohenMacaulay modules over one- and two-dimensional commutative Noetherian local rings. We restrict, if
necessary, to specific families of rings in order to obtain explicit results for GP , since it is possible that even
slightly different sets GP can induce completely different behavior in terms of the sets of lengths. Given
this information, we use transfer homomorphisms, a key tool in Factorization theory and introduced in
Section 3, which make it possible to study sets of lengths and other arithmetical invariants of general
Krull monoids instead in an associated monoid of zero-sum sequences (see Lemma 3.4). These monoids
can be studied using methods from Additive (Group and Number) Theory (see [34]).
Factorization Theory describes the non-uniqueness of factorizations of elements in rings and semigroups
into irreducible elements by arithmetical invariants such as sets of lengths, catenary, and tame degrees.
We will define each of these invariants in Section 2. The goal is to relate the arithmetical invariants
MONOIDS OF MODULES AND ARITHMETIC OF DIRECT-SUM DECOMPOSITIONS
3
with algebraic parameters (such as class groups) of the objects under consideration. The study of sets of
lengths in Krull monoids is a central topic in Factorization Theory. However, since much of this theory
was motivated by examples in number theory (such as holomorphy rings in global fields), most of the
focus so far has been on Krull monoids with finite class group and with each class containing a prime
divisor. This is in contrast to Krull monoids stemming from Module Theory which often have infinite
class group (see Section 4). A key result in Section 6 shows that the arithmetic of these two types of
Krull monoids can have drastically different arithmetic.
In combination with the study of various arithmetical invariants of a given Krull monoid, the following
dual question has been asked since the beginning of Factorization Theory: Are arithmetical phenomena
characteristic for a given Krull monoid (inside a given class of Krull monoids)? Affirmative answers have
been given for the class of Krull monoids with finitely generated class groups where every class contains
a prime divisor. Since sets of lengths are the most investigated invariants in Factorization Theory, the
emphasis in the last decade has been on the following question: Within the class of Krull monoids having
finite class group and such that every class contains a prime divisor, does the system of sets of lengths of
a monoid H characterize the class group of H? A survey of these problems can be found in [37, Sections
7.1 and 7.2]. For recent progress, see [63], [62], and [10]. In Corollary 6.8 we exhibit that for many Krull
monoids stemming from the Module Theory of Sections 4 and 5, the system of sets of lengths and the
behavior of absolutely irreducible elements characterizes the class group of these monoids.
In Section 2 we introduce some of the main arithmetical invariants studied in Factorization Theory
as well as their relevance to the study of direct-sum decompositions. Our focus is on sets of lengths and
on parameters controlling their structure, but we will also need other invariants such as catenary and
tame degrees. Section 3 gives a brief introduction to Krull monoids, monoids of modules, and transfer
homomorphisms. Sections 4 and 5 provide explicit constructions stemming from Module Theory of class
groups and distribution of prime divisors in the classes. Finally, in Section 6, we present our results on
the arithmetic of direct-sum decomposition in the Krull monoids discussed in Sections 4 and 5.
We use standard notation from Commutative Algebra and Module Theory (see [56]) and we follow the
notation of [37] for Factorization Theory. All monoids of modules V(C) are written additively, while all
abstract Krull monoids are written multiplicatively. This follows the tradition in Factorization Theory,
and makes sense also here because our crucial tool, the monoid of zero-sum sequences, is written multiplicatively. In particular, our arithmetical results in Section 6 are written in a multiplicative setting but
they are derived for the additive monoids of modules discussed in Sections 4 and 5. Since we hope that
this article is readable both for those working in Ring and Module Theory as well as those working in
Additive Theory and Factorization Theory, we often recall concepts of both areas which are well known
to the specialists in the respective fields.
2. Arithmetical Preliminaries
In this section we gather together the concepts central to describing the arithmetic of non-factorial
monoids. In particular, we exhibit the arithmetical invariants which will be studied in Section 6 and which
will give a measure of non-unique direct sum decompositions of classes of modules studied in Sections 4
and 5. When possible, we recall previous work in the area of direct-sum decompositions for which certain
invariants have been studied. For more details on non-unique factorization, see [37]. First we record some
preliminary terminology.
Notation. We denote by N the set of positive integers and set N0 = N ∪ {0}. For every n ∈ N, Cn
denotes a cyclic group of order n. For real numbers a, b ∈ R we set [a, b] = {x ∈ Z : a ≤ x ≤ b}. We use
the convention that sup ∅ = max ∅ = min ∅ = 0.
4
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
Subsets of the integers. Let L, L′ ⊂ Z. We denote by L + L′ = {a + b : a ∈ L, b ∈ L′ } the sumset of
L and L′ . If ∅ 6= L ⊂ N, we call
o
nm
sup L
: m, n ∈ L =
∈ Q≥1 ∪ {∞}
ρ(L) = sup
n
min L
the elasticity of L. In addition, we set ρ({0}) = 1. Distinct elements k, l ∈ L are called adjacent if
L ∩ [min{k, l}, max{k, l}] = {k, l}. A positive integer d ∈ N is called a distance of L if there exist
adjacent elements k, l ∈ L with d = |k − l|. We denote by ∆(L) the set of distances of L. Note that
∆(L) = ∅ if and only if |L| ≤ 1, and that L is an arithmetical progression with difference d ∈ N if and
only if ∆(L) ⊂ {d}.
Monoids and Rings. By a monoid H we always mean a commutative semigroup with identity 1 which
satisfies the cancellation law; that is, if a, b, and c are elements of the H with ab = ac, then b = c.
Let H be a monoid. We denote by A(H) the set of atoms (irreducible elements) of H, by q(H) a
quotient group of H with H ⊂ q(H) = {a−1 b : a, b ∈ H}, and by H × the set of invertible elements of H.
We say that H is reduced if H × = {1}, and we denote by Hred = H/H × = {aH × : a ∈ H} the associated
reduced monoid. Let H ′ ⊂ H be a subset. We say that H ′ is divisor-closed if a ∈ H ′ and b ∈ H with
b | a implies that b ∈ H ′ . Denote by [H ′ ] ⊂ H the submonoid generated by H ′ .
A monoid F is called free abelian with basis P ⊂ F if every a ∈ F has a unique representation of the
form
Y
a=
pvp (a) with vp (a) ∈ N0 and vp (a) = 0 for almost all p ∈ P .
p∈P
If F is free abelian with basis P, we set F = F (P) and call
X
|a| =
vp (a) the length of a and supp(a) = {p ∈ P : vp (a) > 0} the support of a .
p∈P
(P)
The multiplicative monoid F (P) is, of course, isomorphic to the additive monoid (N0 , +).
Throughout this manuscript, all rings have a unit element and, apart from a a few motivating remarks
in Section 3, all rings are commutative. Let R be a ring. Then we let R• = R \ {0} denote the nonzero
elements of R and let R× denote its group of units. Note that if R is a domain, then R• is a monoid as
defined above. By the dimension of a ring we always mean its Krull dimension.
Abelian groups. Let G be an additive abelian group and let G0 ⊂ G a subset. Then −G0 = {−g : g ∈
G0 }, G•0 = G0 \ {0}, and hG0 i ⊂ G denotes the subgroup generated by G0 . A family (ei )i∈I of elements
of G is said to be independent if ei 6= 0 for all i ∈ I and, for every family (mi )i∈I ∈ Z(I) ,
X
mi e i = 0
implies
mi ei = 0 for all i ∈ I .
i∈I
L
rank r∗ (G) is the supremum of
The family (ei )i∈I is called a basis for G if G =
i∈I hei i. The total P
the cardinalities of independent subsets of G. Thus r∗ (G) = r0 (G) + p∈P rp (G), where r0 (G) is the
torsion-free rank of G and rp (G) is the p-rank of G for every prime p ∈ P.
Factorizations. Let H be a monoid. The free abelian monoid Z(H) = F A(Hred ) is called the
factorization monoid of H, and the unique homomorphism
π : Z(H) → Hred
satisfying π(u) = u
for each u ∈ A(Hred )
is called the factorization homomorphism of H. For a ∈ H and k ∈ N,
•
•
•
•
ZH (a) = Z(a) = π −1 (aH × ) ⊂ Z(H) is the set of factorizations of a,
Zk (a) = {z ∈ Z(a)
: |z| = k} is the set of factorizations of a of length k,
LH (a) = L(a) = |z| : z ∈ Z(a) ⊂ N0 is the set of lengths of a, and
L(H) = {L(b) : b ∈ H} is the system of sets of lengths of H.
MONOIDS OF MODULES AND ARITHMETIC OF DIRECT-SUM DECOMPOSITIONS
5
By definition, we have Z(a) = {1} and L(a) = {0} for all a ∈ H × . If H is assumed to be Krull, as
is the case in the monoids of modules we study, and a ∈ H, then the set of factorizations Z(a) is finite
and nonempty and hence L(a) is finite and nonempty.
Suppose that there is a ∈ H with |L(a)| > 1 with
distinct k, l ∈ L(a). Then for all N ∈ N, L aN ⊃ {(N − i)k + il : i ∈ [0, N ]} and hence |L aN | > N .
Thus, whenever there is an element a ∈ H that has at least two factorizations of distinct lengths, there
exist elements of H having arbitrarily many factorizations of distinct lengths. This motivates the need
for more refined measures of nonunique factorization.
Several invariants such as elasticity and the ∆-set measure non-uniqueness in terms of sets of lengths.
Other invariants such as the catenary degree provide an even more subtle measurement in terms of the
distinct factorizations of elements. However, these two approaches cannot easily be separated and it is
often the case that a factorization-theoretical invariant is closely related to an invariant of the set of
lengths. Thus the exposition that follows will introduce invariants as they are needed and so that the
relations between these invariants can be made as clear as possible.
The monoid H is called
• atomic if Z(a) 6= ∅ for all a ∈ H,
• factorial if |Z(a)| = 1 for all a ∈ H (equivalently, Hred is free abelian), and
• half-factorial if |L(a)| = 1 for all a ∈ H.
Let z, z ′ ∈ Z(H). Then we can write
z = u1 · . . . · ul v1 · . . . · vm
and z ′ = u1 · . . . · ul w1 · . . . · wn ,
where l, m, n ∈ N0 and u1 , . . . , ul , v1 , . . . , vm , w1 , . . . , wn ∈ A(Hred ) are such that
{v1 , . . . , vm } ∩ {w1 , . . . , wn } = ∅ .
′
Then gcd(z, z ) = u1 · . . . · ul , and we call
d(z, z ′ ) = max{m, n} = max{|z gcd(z, z ′ )−1 |, |z ′ gcd(z, z ′ )−1 |} ∈ N0
the distance between z and z ′ . If π(z) = π(z ′ ) and z 6= z ′ , then clearly
2 + |z| − |z ′ | ≤ d(z, z ′ ) .
For subsets X, Y ⊂ Z(H), we set
d(X, Y ) = min{d(x, y) : x ∈ X, y ∈ Y } ,
and thus d(X, Y ) = 0 if and only if (X ∩ Y 6= ∅, X = ∅, or Y = ∅).
From this point on, we will assume all monoids to be atomic. Since the the monoids described in Sections
4 and 5 are of the form V(C) for C a subclass of finitely generated modules over a commutative Noetherian
ring, they are Krull and hence atomic.
The set of distances and chains of factorizations. We now recall the ∆-set of a monoid H, an
invariant which describes the sets of lengths of elements in H, and illustrate its relationship with distances
between factorizations of elements in H. We denote by
[
∆(L) ⊂ N
∆(H) =
L∈L(H)
the set of distances of H. By definition, ∆(H) = ∅ if and only if H is half-factorial. For a more thorough
investigation of factorizations in H, we will need a distinguished subset of the set of distances. Let
∆∗ (H) denote the set of all d = min ∆(S) for some divisor-closed submonoid S ⊂ H with ∆(S) 6= ∅. By
definition, we have ∆∗ (H) ⊂ ∆(H).
Suppose that H is not factorial. Then there exists
an element a ∈ H with |Z(a)| > 1, and so there
exist distinct z, z ′ ∈ Z(a). Then for N ∈ N, Z aN ⊃ {z N −i (z ′ )i : i ∈ [0, N ]}. Although d z N , (z ′ )N =
N d(z, z ′ ) > N suggests that the factorizations z N and (z ′ )N of aN are very different,
d z N −i (z ′ )i , z N −i+1 (z ′ )i−1 = d(z, z ′ )
6
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
for each i ∈ [1, N ]. This illustrates that the distance alone is too coarse of an invariant, and motivates
the study of the catenary degree as a way of measuring how distinct two factorizations are. As will be
described below, there is a structure theorem for the set of lengths of a Krull monoid. However, except
in very simple situations, there is no known structure theorem for the set of factorizations of an element
in a Krull monoid. Thus we use the catenary degree, its many variations, the tame degree, and other
invariants help to measure the subtle distinctions between factorizations.
Let a ∈ H and N ∈ N0 ∪ {∞}. A finite sequence z0 , . . . , zk ∈ Z(a) is called a (monotone) N -chain
of factorizations if d(zi−1 , zi ) ≤ N for all i ∈ [1, k] and (|z0 | ≤ · · · ≤ |zk | or |z0 | ≥ · · · ≥ |zk | resp.).
We denote by c(a) (or by cmon (a) resp.) the smallest N ∈ N0 ∪ {∞} such that any two factorizations
z, z ′ ∈ Z(a) can be concatenated by an N -chain (or by a monotone N -chain resp.). Then
c(H) = sup{c(b) : b ∈ H} ∈ N0 ∪ {∞} and cmon (H) = sup{cmon(b) : b ∈ H} ∈ N0 ∪ {∞}
denote the catenary degree and the monotone catenary degree of H. The monotone catenary degree
is studied by using the two auxiliary notions of the equal and the adjacent catenary degrees. Let ceq (a)
denote the smallest N ∈ N0 ∪ {∞} such that any two factorizations z, z ′ ∈ Z(a) with |z| = |z ′ | can be
concatenated by a monotone N -chain. We call
ceq (H) = sup{ceq (b) : b ∈ H} ∈ N0 ∪ {∞}
the equal catenary degree of H. We set
cadj (a) = sup{d Zk (a), Zl (a) : k, l ∈ L(a) are adjacent} ,
and the adjacent catenary degree of H is defined as
cadj (H) = sup{cadj (b) : b ∈ H} ∈ N0 ∪ {∞} .
Obviously, we have
c(a) ≤ cmon (a) = sup{ceq (a), cadj (a)} ≤ sup L(a)
for all a ∈ H ,
and hence
c(H) ≤ cmon (H) = sup{ceq (H), cadj (H)} .
Note that cadj (H) = 0 if and only if H is half-factorial, and if H is not half-factorial, then 2 + sup ∆(H) ≤
c(H). Moreover, ceq (H) = 0 if and only if for all a ∈ H and all k ∈ L(a) we have |Zk (a)| = 1. A recent
result of Coykendall and Smith implies that if D is a domain, we have that ceq (D• ) = 0 if and only if D•
is factorial [18, Corollary 2.12].
We call
∼H,eq = (x, y) ∈ Z(H) × Z(H) : π(x) = π(y) and |x| = |y|
the monoid of equal-length relations of H. Let Z ⊂ Z(H) be a subset. We say that an element
x ∈ Z is
minimal in Z if for all elements y ∈ Z with y | x it follows that x = y. We denote by Min Z the set of
minimal elements
in Z. Let x ∈ Z. Since the number of elements y ∈ Z with y | x is finite, there exists
an x∗ ∈ Min Z with x∗ | x. We will need the following lemma.
Lemma 2.1. Let H be an atomic monoid.
(1) ceq (H) ≤ sup |x| : (x, y) ∈ A(∼H,eq ) for some y ∈ Z(H) \ {x} .
(2) For d ∈ ∆(H) let Ad = x ∈ Z(H) : |x| − d ∈ L(π(x) . Then
cadj (H) ≤ sup{|x| : x ∈ Min(Ad ), d ∈ ∆(H)}.
Proof. See [12, Proposition 4.4].
Unions of sets of lengths and the refined elasticities. We now return to studying sets of lengths.
We note that the elasticity of certain monoids of modules were studied in [7] and [8], but that in Section
MONOIDS OF MODULES AND ARITHMETIC OF DIRECT-SUM DECOMPOSITIONS
7
6 we will provide results which generalize these results to larger classes of Krull monoids. In addition, we
will fine tune these results by also computing the refined elasticities. Let k, l ∈ N. If H 6= H × , then
[
L
Uk (H) =
k∈L,L∈L(H)
is the union of all sets of lengths containing k. In other words, Uk (H) is set of all m ∈ N for which there
exist u1 , . . . , uk , v1 , . . . , vm ∈ A(H) with u1 · . . . · uk = v1 · . . . · vm . When H × = H, we set Uk (H) = {k}.
In both cases, we define ρk (H) = sup Uk (H) ∈ N ∪ {∞} and λk (H) = min Uk (H) ∈ [1, k]. Clearly, we
have U1 (H) = {1}, k ∈ Uk (H), and since Uk (H) + Ul (H) ⊂ Uk+l (H), it follows that
λk+l (H) ≤ λk (H) + λl (H) ≤ k + l ≤ ρk (H) + ρl (H) ≤ ρk+l (H) .
The elasticity ρ(H) of H is defined as
ρ(H) = sup{ρ(L) : L ∈ L(H)} ∈ R≥1 ∪ {∞} ,
and it is not difficult to verify that
n ρ (H)
o
ρk (H)
k
ρ(H) = sup
: k ∈ N = lim
.
k→∞
k
k
The structure of sets of lengths. To describe the structure of sets of lengths and of their unions, we
need the concept of arithmetical progressions as well as various generalizations. Let l, M ∈ N0 , d ∈ N,
and {0, d} ⊂ D ⊂ [0, d]. We set
Pl (d) = dZ ∩ [0, ld] = {0, d, 2d, . . . , ld} .
Thus a subset L ⊂ Z is an arithmetical progression (with difference d ∈ N and length l ∈ N0 ) if
L = min L + Pl (d). A subset L ⊂ Z is called an almost arithmetical multiprogression (AAMP for short)
with difference d, period D, and bound M , if
L = y + (L′ ∪ L∗ ∪ L′′ ) ⊂ y + D + dZ
∗
∗
where
∗
• L is finite and nonempty with min L = 0 and L = (D + dZ) ∩ [0, max L∗ ],
• L′ ⊂ [−M, −1] and L′′ ⊂ max L∗ + [1, M ], and
• y ∈ Z.
Note that an AAMP is finite and nonempty, and that an AAMP with period {0, d} and bound M = 0 is
a (usual) arithmetical progression with difference d.
The ω-invariant and the tame degrees. We now study the ω-invariant as well as local and global
tame degrees. We note that these notions have been studied in specific noncommutative module-theoretic
situations in terms of the so-called semi-exchange property (see [19]). Moreover, when describing the sets
of lengths of elements within a Krull monoid H in terms of AAMPs (see Proposition 6.2), the bound M
(described above) is a tame degree related to the monoid H. We begin with the definition. For an atom
u ∈ H, let ω(H, u) denote the smallest N ∈ N ∪ {∞} having the following property :
For any multiple a of u and any factorization a = v1 · . . . · vn of a, there exists a subset Ω ⊂ [1, n]
such that |Ω| ≤ N and
Y
u
vν .
ν∈Ω
Furthermore, we set
ω(H) = sup{ω(H, u) : u ∈ A(H)} ∈ N ∪ {∞} .
An atom u ∈ H is prime if and only if ω(H, u) = 1, and thus H is factorial if and only if ω(H) = 1. If
H satisfies the ACC on divisorial ideals (in particular, H is a Krull monoid or a Noetherian domain),
then ω(H, u) < ∞ for all u ∈ A(H) [38, Theorem 4.2].QRoughly speaking, the tame degree t(H, u) is the
maximum of ω(H, u) and a factorization length of u−1 ν∈Ω vν in H. More precisely, for an atom u ∈ H,
the local tame degree t(H, u) is the smallest N ∈ N0 ∪ {∞} having the following property:
8
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
For any multiple a of u and any factorization a = v1 · . . . · vn of a which does not contain u, there is
a short subproduct which is a multiple of u, say v1 ·. . .·vm , and a refactorization of this subproduct
which contains u, say v1 · . . . · vm = uu2 · . . . · uℓ , such that max{ℓ, m} ≤ N .
Thus the local tame degree t(H, u) measures the distance between any factorization of a multiple a of u
and a factorization of a which contains u. As before, we set
t(H) = sup{t(H, u) : u ∈ A(H)} ∈ N0 ∪ {∞} .
We conclude this section with the following lemma (see [37, Chapter 1] and [39]) which illustrates how
the primary invariants measure the non-uniqueness of factorizations and show that all of these invariants
are trivial if the monoid is factorial.
Lemma 2.2. Let H be an atomic monoid.
(1) H is half-factorial if and only if ρ(H) = 1 if and only if ρk (H) = k for every k ∈ N.
(2) H is factorial if and only if c(H) = t(H) = 0 if and only if ω(H) = 1.
(3) c(H) = 0 or c(H) ≥ 2, and if c(H) ≤ 2, then H is half-factorial.
(4) c(H) ≤ ω(H) ≤ t(H) ≤ ω(H)2 , and if H is not factorial, then max{2, ρ(H)} ≤ ω(H).
(5) If c(H) = 3, then every L ∈ L(H) is an arithmetical progression with difference 1.
3. Krull monoids, monoids of modules, and transfer homomorphisms
The theory of Krull monoids is presented in detail in the monographs [44] and [37]. Here we gather the
terminology required for our treatment. We then present an introduction to monoids of modules — the
key objects of our study. Finally, we recall important terminology and results about monoids of zero-sum
sequences and transfer homomorphisms — the key tools in our arithmetical investigations.
Krull monoids. Let H and D be monoids. A monoid homomorphism ϕ : H → D is called
• a divisor homomorphism if ϕ(a) | ϕ(b) implies that a | b for all a, b ∈ H.
• cofinal if for every a ∈ D there exists some u ∈ H such that a | ϕ(u).
• a divisor theory (for H) if D = F (P) for some set P, ϕ is a divisor homomorphism,
and for every
a ∈ F (P), there exists a finite nonempty subset X ⊂ H satisfying a = gcd ϕ(X) .
We call C(ϕ) = q(D)/q(ϕ(H)) the class group of ϕ, use additive notation for this group, and for a ∈ q(D),
we denote by [a] = a q(ϕ(H)) ∈ q(D)/q(ϕ(H)) the class containing a. Clearly D/H = {[a] : a ∈ D} ⊂
C(ϕ) is a submonoid with quotient group C(ϕ). The homomorphism ϕ is cofinal if and only if C(ϕ) = D/H
and, by definition, every divisor theory is cofinal. Let ϕ : H → D = F (P) be a divisor homomorphism.
Then ϕ(H) = {a ∈ D : [a] = [1]} and GP = {[p] = pq(ϕ(H)) : p ∈ P} ⊂ C(ϕ) is called the set of classes
containing prime divisors. Moreover, we have hGP i = C(ϕ) and [GP ] = {[a] : a ∈ D}.
The monoid H is called a Krull monoid if it satisfies one of the following equivalent conditions :
(a) H is completely integrally closed and satisfies the ACC on divisorial ideals.
(b) H has a divisor theory.
(c) H has a divisor homomorphism into a free abelian monoid.
If H is a Krull monoid, then a divisor theory is unique up to unique isomorphism, and the class group
associated to a divisor theory depends only on H. It is called the class group of H and will be denoted
by C(H). Moreover, a reduced Krull monoid H with divisor theory H ֒→ F (P) is uniquely determined
up to isomorphism by its characteristic (G, (mg )g∈G ) where G is an abelian group together with an
isomorphism Φ : G → C(H) and with family (mg )g∈G of cardinal numbers mg = |P ∩ Φ(g)| (see [37,
Theorem 2.5.4], and the forthcoming Lemma 3.4).
It is well-known that a domain R is a Krull domain if and only if its multiplicative monoid R• is a
Krull monoid, and we set the class group of R to be C(R) = C(R• ). Property (a) shows that a Noetherian
MONOIDS OF MODULES AND ARITHMETIC OF DIRECT-SUM DECOMPOSITIONS
9
domain is Krull if and only if it is integrally closed. In addition, many well-studied classes of commutative
monoids such as regular congruence monoids in Krull domains and Diophantine monoids are Krull. The
focus of the present paper is on Krull monoids stemming from Module Theory.
Monoids of modules. Let R be a (not necessarily commutative) ring and C a class of (right) Rmodules. We say that C is closed under finite direct sums, direct summands, and isomorphisms provided
the following holds: Whenever M, M1 , and M2 are R-modules with M ∼
= M1 ⊕ M2 , we have M ∈ C if
and only if M1 , M2 ∈ C. We say that C satisfies the Krull-Remak-Schmidt-Azumaya Theorem (KRSA
for short) if the following holds:
If k, l ∈ N and M1 , . . . , Mk , N1 , . . . , Nl are indecomposable modules in C with M1 ⊕ · · · ⊕ Mk ∼
=
N1 ⊕ · · · ⊕ Nl , then l = k and, after a possible reordering of terms, Mi ∼
= Ni for all i ∈ [1, k].
Suppose that C is closed under finite direct sums, direct summands, and isomorphisms. For a module
M ∈ C, we denote by [M ] its isomorphism class, and by V(C) the set of isomorphism classes. (For our
purposes here, we tacitly assume that this is actually a set. For the classes of modules studied in Sections
4 and 5 this is indeed the case. For the involved set-theoretical problems in a more general context, see
[26, Section 2].) Then V(C) is a commutative semigroup with operation [M ] + [N ] = [M ⊕ N ] and all
information about direct-sum decompositions of modules in C can be studied in terms of factorizations in
the semigroup V(C). By definition, C satisfies KRSA if and only if V(C) is a free abelian monoid, which
holds if EndR (M ) is local for each indecomposable M in C (see [56, Theorem 1.3]).
If the endomorphism ring EndR (M ) is semilocal for all modules M in C, then V(C) is a Krull monoid
([22, Theorem 3.4]). There is an abundance of recent work which provides examples of rings and classes
of modules over these rings for which all endomorphism rings are semilocal (see [24], [25], and [26]. For
monoids of modules, a characterization of when the class group is a torsion group is given in [27]).
Suppose that V(C) is a Krull monoid. Then to understand the structure of direct-sum decompositions
of modules in C is to understand the arithmetic of the reduced Krull monoid V(C). Since any reduced
Krull monoid H is uniquely determined by its class group and by the distribution of prime divisors (that
is, the characteristic of H), one must study these parameters.
In the present paper we will focus on the following classes of modules over a commutative Noetherian
local ring S, each closed under finite direct sums, direct summands, and isomorphisms. For a commutative
Noetherian local ring S, we denote by
• M(S) the semigroup of isomorphism classes of finitely generated S-modules, by
• T (S) the semigroup of isomorphism classes of finitely generated torsion-free S-modules, and by
• C(S) the semigroup of isomorphism classes of maximal Cohen-Macaulay (MCM) S-modules.
Note that in order to make C(S) a semigroup, we insist that [0S ] ∈ C(S), even though the zero module
is not MCM. We say that a commutative Noetherian local ring S has finite representation type if there
are, up to isomorphism, only finitely many indecomposable MCM S-modules. Otherwise we say that S
has infinite representation type.
b m)
b
Throughout, let (R, m) be a commutative Noetherian local ring with maximal ideal m, and let (R,
b
denote its m-adic completion. Let V(R) and V(R) be any of the above three semigroups. If M is an
b is an R-module
b
c] ∈ V(R),
b and every
c ∼
with [M
R-module such that [M ] ∈ V(R), then M
= M ⊗R R
b
b has finite
such R-module is called extended. Note that R has finite representation type if and only if R
b
representation type (see [56, Chapter 10]), and that the dimension of R is equal to the dimension of R.
The following crucial result shows that the monoid V(R) is Krull.
b m)
b
Lemma 3.1. Let (R, m) be a commutative Noetherian local ring with maximal ideal m, and let (R,
denote its m-adic completion.
b
b
(1) For each indecomposable finitely generated R-module
M , EndRb (M ) is local, and therefore M(R),
b
b
T (R), and C(R) are free abelian monoids.
10
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
b is a divisor homomorphism. It is cofinal if and only if every finitely
(2) The embedding M(R) ֒→ M(R)
b
generated R-module is a direct summand of an extended module.
b and C(R) ֒→ C(R)
b are divisor homomorphisms.
(3) The embeddings T (R) ֒→ T (R)
In particular, M(R), T (R), and C(R) are reduced Krull monoids. Moreover, the embeddings in (2) and
b
(3) are injective and map R-modules onto the submonoid of extended R-modules.
Proof. Property (1) holds by the Theorem of Krull-Remak-Schmidt-Azumaya (see [56, Chapter 1]).
In [67], Wiegand proved that the given embedding is a divisor homomorphism (see also [9, Theorem
3.6]). The characterization of cofinality follows from the definition and thus (2) holds.
Let M, N be R-modules such that either [M ], [N ] ∈ V(R) where V(R) denotes either T (R) or C(R)
c] divides [N
b ] in V(R).
b Then we have divisibility in M(R),
b and hence in M(R) by
and suppose that [M
(2). Since V(R) ⊂ M(R) is divisor-closed, it follows that [M ] divides [N ] in V(R), proving (3).
Together, (2) and (3) show that M(R), T (R), and C(R) satisfy Property (c) in the definition of Krull
c] are injective.
monoids. Since each of these monoids is reduced, the maps induced by [M ] 7→ [M
b is not necessarily cofinal, as is shown in [47] and [30]. In
Note that the embedding M(R) ֒→ M(R)
Sections 4 and 5 we will study in detail the class group and the distribution of prime divisors of these
Krull monoids, in the case of one-dimensional and two-dimensional commutative Noetherian local rings.
Monoids of zero-sum sequences. We now introduce Krull monoids having a combinatorial flavor
which are used to model arbitrary Krull monoids. Let G be an additive abelian group and let G0 ⊂ G be
a subset. Following the tradition in Additive Group and Number Theory, we call the elements of F (G0 )
sequences over G0 . Thus a sequence S ∈ F (G0 ) will be written in the form
Y
S = g1 · . . . · gl =
g vg (S) .
g∈G0
We will use all notions (such as the length) as in general free abelian monoids. We set −S = (−g1 ) · . . . ·
(−gl ), and call σ(S) = g1 + · · · + gl ∈ G the sum of S. The monoid
B(G0 ) = {S ∈ F (G0 ) : σ(S) = 0}
is called the monoid of zero-sum sequences over G0 , and its elements are called zero-sum sequences
over G0 . Obviously, the inclusion B(G0 ) ֒→ F (G0 ) is a divisor homomorphism, and hence B(G0 ) is a
reduced Krull monoid by Property (c) in the definition of Krull monoids. By definition, the inclusion
B(G0 ) ֒→ F (G0 ) is cofinal if and only if for every g ∈ G0 there is an S ∈ B(G0 ) with g | S; equivalently,
there is no proper subset G′0 ( G0 such that B(G′0 ) = B(G0 ). If |G| 6= 2, then C(B(G)) ∼
= G, and every
class contains precisely one prime divisor.
For every arithmetical invariant ∗(H), as defined for a monoid H in Section 2, it is usual to write
∗(G0 ) instead of ∗(B(G0 )) (whenever the meaning is clear from the context). In particular, we set
A(G0 ) = A(B(G0 )), L(G0 ) = L(B(G0 )), cmon (G0 ) = cmon (B(G0 )), etc.
The study of sequences, subsequence sums, and zero-sums is a flourishing subfield of Additive Group
and Number Theory (see, for example, [31], [40], and [43]). The Davenport constant D(G0 ), defined as
D(G0 ) = sup{|U | : U ∈ A(G0 )} ∈ N0 ∪ {∞} ,
is among the most studied invariants in Additive Theory and will play a crucial role in the computations
of arithmetical invariants (see the discussion after Lemma 3.4). We will need the following two simple
lemmas which we present here so as to not clutter the exposition of Section 6.
Lemma 3.2. Suppose that the inclusion B(G0 ) ֒→ F (G0 ) is cofinal. The following are equivalent.
(a) There exist nontrivial submonoids H1 , H2 ⊂ B(G0 ) such that B(G0 ) = H1 × H2 .
(b) There exist nonempty subsets G1 , G2 ⊂ G0 such that G0 = G1 ⊎ G2 and B(G0 ) = B(G1 ) × B(G2 ).
(c) There exist nonempty subsets G1 , G2 ⊂ G0 such that G0 = G1 ⊎ G2 and A(G0 ) = A(G1 ) ⊎ A(G2 ).
MONOIDS OF MODULES AND ARITHMETIC OF DIRECT-SUM DECOMPOSITIONS
11
Proof. Clearly (b) implies (a). The converse follows from [37, Proposition 2.5.6]. The implication (b)
implies (c) is obvious. We now show that (c) implies (b). Let B ∈ B(G0 ). Since B(G0 ) is a Krull
monoid, it is atomic and hence B = U1 · . . . · Ul with U1 , . . . , Ul ∈ A(G0 ). After renumbering (if
necessary), we can find k ∈ [0, l] such that U1 , . . . , Uk ∈ A(G1 ) and Uk+1 , . . . , Ul ∈ A(G2 ). Thus
B(G0 ) = B(G1 )B(G2 ). If B ∈ B(G1 ) ∩ B(G2 ), then B is a product of atoms from A(G1 ) and a product
of atoms from A(G2 ). Since their intersection is empty, both products are empty. Therefore B = 1 and
hence B(G0 ) = B(G1 ) × B(G2 ).
Lemma 3.2.(c) shows that B(G• ) is not a direct product of submonoids. Suppose that 0 ∈ G0 . Then
0 ∈ B(G0 ) is a prime element and B(G0 ) = B({0}) × B(G•0 ). But B({0}) = F ({0}) ∼
= (N0 , +), and thus
all the arithmetical invariants measuring the non-uniqueness of factorizations of B(G0 ) and of B(G•0 )
coincide. Therefore we can assume that 0 ∈
/ G0 whenever it is convenient.
Lemma 3.3. Let G be an abelian group and let G0 ⊂ G be a subset such that 1 < D(G0 ) < ∞.
(1) We have ρ(G0 ) ≤ D(G0 )/2, k ≤ ρk (G0 ) ≤ kρ(G0 ), and ρ(G0 )−1 k ≤ λk (G0 ) ≤ k for all k ∈ N.
(2) Suppose that ρ2 (G0 ) = D(G0 ). Then ρ(G0 ) = D(G0 )/2, and for all k ∈ N, we have
ρ2k (G0 ) = kD(G0 )
and
kD(G0 ) + 1 ≤ ρ2k+1 (G0 ) ≤ kD(G0 ) +
D(G0 )
.
2
Moreover, if j, l ∈ N0 such that lD(G0 ) + j ≥ 1, then
2j
≤ λlD(G0 )+j (G0 ) ≤ 2l + j .
2l +
D(G0 )
Proof. By definition, λk (G0 ) ≤ k ≤ ρk (G0 ). Since ρ(G0 ) = sup{ρk (G0 )/k : k ∈ N}, it follows that
ρk (G0 ) ≤ kρ(G0 ) and k ≤ ρ(G0 )λk (G0 ). Furthermore, 2ρk (G0 ) ≤ kD(G0 ) for all k ∈ N implies that
ρ(G0 ) ≤ D(G0 )/2. This gives (1).
We now prove (2). Since ρk (G0 ) + ρl (G0 ) ≤ ρk+l (G0 ) for every k, l ∈ N, (1) implies that
kD(G0 ) = kρ2 (G0 ) ≤ ρ2k (G0 ) ≤ (2k)
D(G0 )
= kD(G0 ) ,
2
and hence
D(G0 )
.
2
Let j, l ∈ N0 such that lD(G0 ) + j ≥ 1. For convenience, set ρ0 (G0 ) = λ0 (G0 ) = 0. Since
2
2l =
lD(G0 ) ≤ λlD(G0 ) (G0 ) and ρ2l (G0 ) = lD(G0 ) ,
D(G0 )
kD(G0 ) + 1 = ρ2k (G0 ) + ρ1 (G0 ) ≤ ρ2k+1 (G0 ) ≤ (2k + 1)ρ(G0 ) ≤ kD(G0 ) +
it follows that λlD(G0 ) (G0 ) = 2l, and hence
2l +
2
2j
=
lD(G0 ) + j = ρ(G0 )−1 lD(G0 ) + j
D(G0 )
D(G0 )
≤ λlD(G0 )+j (G0 ) ≤ λlD(G0 ) (G0 ) + λj (G0 ) ≤ 2l + j .
Transfer homomorphisms. Transfer homomorphisms are a central tool in Factorization Theory. In
order to study a given monoid H, one constructs a transfer homomorphism θ : H → B to a simpler
monoid B, studies factorizations in B, and then lifts arithmetical results from B to H. In the case
of Krull monoids, transfer homomorphisms allow one to study nearly all of the arithmetical invariants
introduced in Section 2 in an associated monoid of zero-sum sequences. We now gather the basic tools
necessary for this approach.
12
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
A monoid homomorphism θ : H → B is called a transfer homomorphism if it has the following
properties:
(T 1) B = θ(H)B × and θ−1 (B × ) = H × .
(T 2) If u ∈ H, b, c ∈ B and θ(u) = bc, then there exist v, w ∈ H such that u = vw, θ(v) ≃ b
and θ(w) ≃ c.
The next result provides the link between the arithmetic of Krull monoids and Additive Group and
Number Theory. This interplay is highlighted in the survey [34].
Lemma 3.4. Let H be a Krull monoid, ϕ : H → D = F (P) a cofinal divisor homomorphism, G = C(ϕ)
e : D → F (GP ) denote the
its class group, and GP ⊂ G the set of classes containing prime divisors. Let β
e
unique homomorphism defined by β(p) = [p] for all p ∈ P.
e ◦ ϕ : H → B(GP ) is a
(1) The inclusion B(GP ) ֒→ F (GP ) is cofinal, and the homomorphism β = β
transfer homomorphism.
(2) For all a ∈ H, LH (a) = LB(GP ) (β(a)). In particular, L(H) = L(GP ), ∆(H) = ∆(GP ), Uk (H) =
Uk (GP ), ρk (H) = ρk (GP ), and λk (H) = λk (GP ) for each k ∈ N.
(3) Suppose that H is not factorial. Then c(H) = c(GP ), cadj (H) = cadj (GP ), cmon (H) = cmon (GP ),
∆∗ (H) = ∆∗ (GP ), and ω(H) ≤ D(GP ).
Proof. See [37, Section 3.4] for details pertaining to most of the invariants. For the statements on the
monotone catenary degree see [36]. Roughly speaking, all of the statements in (2) are straightforward,
but the statements in (3) are more subtle. Note that a statement corresponding to (3) does not hold true
for the tame degree (see [33]).
In summary, if the monoid of modules V(R) is Krull with class group G and set GP of classes containing
prime divisors, then the arithmetic of direct-sum decompositions can be studied in the monoid B(GP )
b as in Lemma 3.1 and all
of zero-sum sequences over GP . In particular, if H = M(R) and D = M(R)
notation is as in Lemma 3.4, then
c∼
D(GP ) = sup{l : M
= N1 ⊕ · · · ⊕ Nl such that [M ] ∈ A(H) and [Ni ] ∈ A(D) for each i ∈ [1, l]} .
4. Monoids of modules: Class groups and distribution of prime divisors I
Throughout this section we use the following setup:
(S) (R, m) denotes a one-dimensional analytically unramified commutative Noetherian local ring with
b its m-adic completion, and spl(R) =
unique maximal ideal m, k = R/m its residue field, R
b
| spec(R)| − | spec(R)| the splitting number of R.
In this section we investigate the characteristic of the Krull monoids M(R) and T (R) for certain
one-dimensional local rings. This study is based on deep module-theoretic work achieved over the past
several decades. We gather together module-theoretic information and proceed using a recent construction
(see 4.4) to obtain results on the class group and on the set GP of classes containing prime divisors.
The literature does not yet contain a systematic treatment along these lines. Indeed, early results (see
Theorem 4.2 below) indicated only the existence of extremal sets GP which imply either trivial direct-sum
decompositions or that all arithmetical invariants describing the direct-sum decompositions are infinite.
In either case there was no need for further arithmetical study. Here we reveal that finite and wellstructured sets GP occur in abundance. Thus, as we will see in Section 6, the arithmetical behavior of
direct-sum decompositions is well-structured.
b and R are both reduced
We first gather basic ring and module-theoretic properties. By definition, R
and the integral closure of R is a finitely generated R-module. Moreover, we have C(R) = T (R). Let
M be a finitely generated R-module. If p is a minimal prime ideal of R, then Rp is a field, Mp is a
finite-dimensional Rp -vector space, and we set rankp (M ) = dimRp (Mp ). If p1 , . . . , ps are the minimal
MONOIDS OF MODULES AND ARITHMETIC OF DIRECT-SUM DECOMPOSITIONS
13
prime ideals of R, then rank(M ) = (r1 , . . . , rs ) where ri = rankpi (M ) for all i ∈ [1, s]. The module M is
said to have constant rank if r1 = · · · = rs .
We start with a beautiful result of Levy and Odenthall, which gives us a tool to determine which
b
finitely generated R-modules
are extended from R-modules.
b
Proposition 4.1. [57, Theorem 6.2] Let M be a finitely generated torsion-free R-module.
Then M is
b with
extended if and only if rankp (M ) = rankq (M ) whenever p and q are minimal prime ideals of R
p ∩ R = q ∩ R. In particular, if R is a domain, then M is extended if and only if its rank is constant.
We start our discussion with a result which completely determines the characteristic of the Krull
monoid M(R). The arithmetic of this monoid is studied in Proposition 6.2.2.
Theorem 4.2. [46, Theorem 6.3] Let G denote the class group of M(R) and let GP ⊂ G denote the set
of classes containing prime divisors.
(1) If R is not Dedekind-like, then G is free abelian of rank spl(R) and each class contains |k|ℵ0 prime
divisors.
(2) If R is a DVR, then G = 0.
(3) If R is Dedekind-like but not a DVR, then either
(a) spl(R) = 0 and G = 0, or
(b) spl(R) = 1, G is infinite cyclic with G = hei and GP = {−e, 0, e}. Each of the classes e and −e
contain ℵ0 prime divisors and the class 0 contains |k|ℵ0 prime divisors.
Thus, for the rest of this section, we focus our attention on T (R). To determine if the divisor homob is a divisor theory, we will require additional information. For now, we easily
morphism T (R) ֒→ T (R)
show that it is always cofinal.
b is a cofinal divisor homomorphism.
Proposition 4.3. The embedding T (R) ֒→ T (R)
Proof. By Lemma 3.1 the embedding is a divisor homomorphism. If M is a finitely generated torsion-free
b
R-module,
we can consider its rank, rank(M ) = (r1 , . . . , rt ), where t is the number of minimal primes
b If r1 = · · · = rt , then by Proposition 4.1 M is extended, say M = N
b for some finitely generated
of R.
torsion-free R-module N and the result is trivial. If the rank of M is not constant, set r = max{r1 , . . . , rt }
r−r1
r−rt
b
b 1
b t
and consider the R-module
L = R/q
⊕ · · · ⊕ R/q
, where q1 , . . . , qt denote the minimal
b Then rank(N ⊕ L) = (r, . . . , r) is constant and hence N ⊕ L is extended, say N ⊕ L ∼
primes of R.
= Pb
for some finitely generated torsion-free R-module P . Clearly M is isomorphic to a direct summand of Pb
and the result follows.
(P)
b is free abelian, we can identify it with the free abelian monoid N
Since T (R)
where P is an index
0
b
set for the isomorphism classes of indecomposable finitely generated torsion-free R-modules. We then
use Proposition 4.1 to give a detailed description of T (R). The following construction has been used
numerous times (see, for example, [7], [8], and [28]).
Construction 4.4.
• Let p1 , . . . , ps be the distinct minimal prime ideals of R. For each i ∈ [1, s], let qi,1 , . . . , qi,ti be the
b lying over pi . Note that spl(R) = Ps (ti − 1).
minimal primes of R
i=1
b
• Let P be the set of isomorphism classes of indecomposable finitely generated torsion-free R-modules.
• Let A(R) be the spl(R) × |P| matrix whose column indexed by the isomorphism class [M ] ∈ P is
T
r1,1 − r1,2 · · · r1,1 − r1,t1 · · · rs,1 − rs,2 · · · rs,1 − rs,ts
where ri,j = rankqi,j (M ).
(P)
Then T (R) ∼
= ker(A(R)) ∩ N(P) ⊂ N0 is a Diophantine monoid.
14
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
b lie over the minimal prime
If one has a complete description of how the minimal prime ideals of R
b
ideals of R together with the ranks of all indecomposable finitely generated torsion-free R-modules,
then
Construction 4.4 completely describes the monoid T (R). In certain cases (eg. Section 4.1) we are able to
obtain all of this information. Other times we know only some of the ranks that occur for indecomposable
b
R-modules
and thus have only a partial description for T (R). However, as was shown in [8], the ranks
b
of indecomposable cyclic R-modules
gives enough information about the columns of A(R) to prove that
b is nearly always a divisor theory. First we recall that if q1 , . . . , qt are the minimal primes
T (R) ֒→ T (R)
b and E ⊂ [1, t]. Then
of R,
!
(
b
1, i ∈ E
R
rank T
= (r1 , . . . , rt ), where ri =
0, i 6∈ E.
i∈E qi
Thus every nontrivial t-tuple of zeros and ones can be realized as the rank of a nonzero (necessarily
b
indecomposable) cyclic R-module.
Thus we have the following:
Construction 4.5. Let all notation be as in Construction 4.4. After renumbering if necessary,
P there is
p ∈ [0, s] such that t1 , . . . , tp ≥ 2 and such that ti = 1 for each i ∈ [p + 1, s]. Then spl(R) = pj=1 tj − p.
For each i ∈ [1, p], let Ai be the set of (ti − 1) × 1 column vectors all of whose entries are either 0 or 1,
and let Bi be the set of (ti − 1) × 1 column
Q vectors all of whose entries are either 0 or −1.
We now define T to be the spl(R) × pi=1 (2ti − 1) matrix, each of whose columns has the form
T1
..
where Ti ∈ Ai ∪ Bi for each i ∈ [1, p].
. ,
Tp
With the notation as in Constructions 4.4 and 4.5, we give a realization result which shows that the
matrix T occurs as a submatrix of A(R).
Proposition 4.6. [8, Proposition 3.7] For each column α of T , there exist nonnegative integers ri,j and
b
an indecomposable torsion-free R-module
Mα of rank
(r1,1 , . . . , r1,t1 , . . . , rp,1 , . . . , rp,tp , rp+1,1 , . . . , rs,1 )
such that
α = r1,1 − r1,2
···
···
r1,1 − r1,t1
rp,1 − rp,2
···
rp,1 − rp,tp
T
.
In particular, the matrix A(R) nearly always satisfies the hypotheses of the following lemma.
Lemma 4.7. [8, Lemma 4.1] Fix an integer q ≥ 1, and let Iq denote the q ×q identity matrix. Let
P be an
index set, and let D be a q × |P| integer matrix whose columns are indexed by P. Assume D = D1 | D2 ,
where D1 is the q × (2q + 2) integer matrix
1 −1
..
.. ,
I −I
.
.
q
q
1 −1
and D2 is an arbitrary integer matrix with q rows (and possibly infinitely many columns). Let H =
(P)
ker(D) ∩ N0 .
(1)
(2)
(3)
(4)
The map D : Z(P) → Z(q) is surjective.
(P)
The natural inclusion H ֒→ N0 is a divisor theory.
ker(D) = q(H).
C(H) ∼
= Z(q) , and this isomorphism maps the set of classes containing prime divisors onto the set of
distinct columns of D.
MONOIDS OF MODULES AND ARITHMETIC OF DIRECT-SUM DECOMPOSITIONS
15
In particular, we observe the following: Given a fixed column α of D, the cardinality of {β : β is a column of
D and β = α} is equal to the cardinality of prime divisors in the class corresponding to α. Therefore,
the characteristic of the Krull monoid H is completely given by the matrix D.
Based on the previous results, one easily obtains the following theorem which provides the framework
for our study of the characteristic of T (R).
Theorem 4.8.
b is free abelian.
(1) If spl(R) = 0, then T (R) ∼
= T (R)
b is a divisor theory. Moreover,
(2) If spl(R) ≥ 2 then the embedding T (R) ֒→ T (R)
(P)
∼
(a) T (R) = ker(A(R)) ∩ N0 ,
(b) C(T (R)) ∼
= Z(spl(R)) , and this isomorphism maps the set of classes containing prime divisors
onto the set of distinct columns of A(R).
b is a divisor theory if and only if the defining
Suppose that spl(R) = 1. The embedding T (R) ֒→ T (R)
matrix A(R) contains at least two positive and at least two negative entries (see Proposition 6.1.2).
b
In many cases, computing the ranks of indecomposable R-modules
and hence the columns of the
defining matrix A(R) is difficult. However, an additional hypotheses on R implies that the set of classes
containing prime divisors satisfies GP = −GP , a crucial property for all arithmetical investigations (see
Proposition 6.2 and the subsequent remarks).
∼ S/(f ) where (S, n) is a hypersurface i.e., a regular Noeb=
Corollary 4.9. Suppose in addition that R
b and
therian local ring of dimension two and where 0 6= f ∈ n. If G is the class group of T (R) ֒→ T (R)
GP is the set of classes containing prime divisors, then GP = −GP .
Proof. With the hypotheses given, we can apply [8, Proposition 6.2] to see that if M is any indecomposable
b
b
R-module
with rank (r1 , . . . , rt ), then there is an indecomposable R-module
N with rank
(m − r1 , m −
r2 , . . . , m − rt ) for some m ≥ max{r1 , . . . , rt }. Using Construction 4.4 we see that if α = a1 · · · aq is
the column of A(R) indexed by M , then −α is the column indexed by N . Therefore, since GP corresponds
to the distinct columns of A(R), GP = −GP .
Remark 4.10. Although the system of equations developed in Construction 4.4 is somehow natural, it
is not the only system of equations which can be used to define T (R). Indeed, the matrix A(R) can be
adjusted by performing any set of elementary row operations. If J is an elementary matrix corresponding
(P)
(P)
to such a set of row operations, then T (R) ∼
= ker(JA(R)) ∩ N0 . Moreover, this
= ker(A(R)) ∩ N0 ∼
isomorphism gives rise to an automorphism of C(T (R)) mapping the set of classes containing prime
divisors to another set of classes containing prime divisors. Example 4.20 illustrates the usefulness of
considering an alternate defining matrix for T (R).
4.1. Finite Representation Type. Throughout this subsection, let R be as in Setup (S), and suppose
in addition that R has finite representation type.
Decades with of work going back to a paper of Green and Reiner [42] and including papers by R. and
S. Wiegand, Çimen, Arnuvut, Luckas, and Baeth (see [68], [16], [1], and [3]) culminated in a paper by
Baeth and Luckas [6] which classified precisely the tuples that can occur as the ranks of indecomposable
b have at most
torsion-free R-modules. We note that since R has finite representation type, both R and R
three minimal primes (see [17, Theorem 0.5]).
Proposition 4.11. [6, Main Theorem 1.2]
b is a domain, then every indecomposable finitely generated torsion-free R-module
b
(1) If R
has rank 1, 2
or 3.
b has exactly two minimal prime ideals, then every indecomposable finitely generated torsion-free
(2) If R
b
R-module has rank (0, 1), (1, 0), (1, 1), (1, 2), (2, 1) or (2, 2).
16
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
b has exactly two minimal prime ideals, then every indecomposable finitely generated torsion-free
(3) If R
b
R-module has rank (0, 0, 1), (0, 1, 0), (1, 0, 0), (0, 1, 1), (1, 0, 1), (1, 1, 0), (1, 1, 1) or (2, 1, 1).
b
Note the lack of symmetry in case (3): With a predetermined order on the minimal prime ideals of R,
there is an indecomposable module of rank (2, 1, 1), but not of rank (1, 2, 1) or (1, 1, 2). As is stated in [6,
Remark 5.2], even for a fixed number of minimal primes, not each of these tuples will occur as the rank
of an indecomposable module for each ring. However, since when applying Construction 4.4 we cannot
distinguish between an indecomposable of rank (2, 1) and one of rank (1, 0), and since all nontrivial tuples
of zeros and ones occur as ranks of indecomposable cyclic modules, we have [7, Proposition 3.3]:
(1) If spl(R) = 1, then A(R) = 1
· · · 1 −1 · · ·
0 −1 1 −1 1
(2) If spl(R) = 2, then A(R) =
−1 0 1 −1 0
−1 0 · · · 0 .
0 0 1 ···
.
1 0 1 ···
When spl(R) = 1, we are guaranteed at least one entry for each of 1, −1, and 0, coming from the
b
ranks of indecomposable cyclic R-modules.
If we have at most one 1 or at most one −1 in the defining
b has exactly two minimal primes p and q,
matrix A(R), then it must be the case that R is a domain, R
b is the only indecomposable torsion-free R-module
b
and up to isomorphism either R/p
of rank (r, s) with
b is the only indecomposable torsion-free R-module
b
r − s = 1 or the R/q
of rank (r, s) with r − s = −1.
If this is the situation, we say that R satisfies condition (†). In case spl(R) = 2, we are guaranteed that
each column listed appears at least once as a column of A(R).
We then have the following refinement of Theorem 4.8 when R has finite representation type. The
arithmetic of this monoid is studied in Proposition 6.2.2, Theorem 6.4, and Corollary 6.10.
Theorem 4.12. [7, Proposition 3.3]
b is not a divisor theory but T (R) is
(1) If spl(R) = 1 and R satisfies condition (†) then T (R) ֒→ T (R)
free abelian.
b is a divisor theory with
(2) If spl(R) = 1 and R does not satisfy condition (†), then T (R) ֒→ T (R)
infinite cyclic class group G = hei, and GP = {−e, 0, e}.
(2)
b is a divisor theory and C(T (R)) ∼
(3) If spl(R) = 2, then T (R) ֒→ T (R)
= Z . Moreover,
this
isomor
1
−1
1
−1
0
0
phism maps the set of classes containing prime divisors onto
,
,
,
,
,
.
1
−1
0
0
1
−1
4.2. Infinite Representation Type. Throughout this subsection, let R be as in Setup (S), and suppose
in addition that R has infinite representation type.
Unfortunately, in this case, there is no known complete list of the tuples that can occur as ranks of
indecomposable finitely generated torsion-free R-modules. Thus we cannot give a full description of T (R)
b has infinite representation
using Construction 4.4. However, with the additional assumption that R/q
b we can produce a wide variety of interesting ranks and can
type for some minimal prime ideal q of R,
provide a partial description of T (R). This information is enough to show that, very much unlike the
finite representation type case of Section 4.1, all of the arithmetical invariants we study are infinite.
Proposition 4.13. [61, Theorem 3.4.1] Let S be a one-dimensional analytically unramified commutative
Noetherian local ring with residue field K, and with t minimal prime ideals q1 , . . . , qt such that S/q1 has
infinite representation type. Let (r1 , . . . , rt ) be a nonzero t-tuple of nonnegative integers with ri ≤ 2r1 for
all i ∈ [2, t].
(1) There exists an indecomposable torsion-free S-module of rank (r1 , . . . , rt ).
(2) If the residue field K is infinite, then the set of isomorphism classes of indecomposable torsion-free
S-modules of rank (r1 , . . . , rt ) has cardinality |K|.
MONOIDS OF MODULES AND ARITHMETIC OF DIRECT-SUM DECOMPOSITIONS
17
b
By Proposition 4.13 the conditions of Lemma 4.7 are satisfied. Therefore the map T (R) ֒→ T (R)
is a divisor theory and the class group C(T (R)) is free abelian of rank spl(R). Our main result of this
subsection is a refinement of Theorem 4.8. Its arithmetical consequences are given in Proposition 6.2.1,
strongly improving the arithmetical characterizations given in [8].
b such
Theorem 4.14. Suppose that spl(R) ≥ 1 and that there is at least one minimal prime ideal q of R
b has infinite representation type. Then C(T (R)) is free abelian of rank spl(R) and the set of
that R/q
classes containing prime divisors contains an infinite cyclic subgroup.
Proof. Let p1 , . . . , ps denote the minimal primes of R and, for each i ∈ [1, s], let qi,1 , . . . , qi,ti denote the
b lying over pi . Without loss of generality, assume that R/q
b 1,1 has infinite
set of minimal primes of R
representation type. If t1 = 1, then without loss of generality, t2 > 1. From Proposition 4.13 there
b
is, for each pair (r, s) of nonnegative integers (not both zero), an indecomposable R-module
M with
rankq2,1 (M ) = r, rankq2,2 (M ) = s, and rankqi,j (M ) = 0 for all (i, j) 6∈ {(1, 1), (2, 1), (2, 2)}. Now suppose
that t1 > 1. Then we have, from Proposition 4.13, for each pair (r, s) of nonnegative integers (not both
b
zero) satisfying r − s ≥ −s, an indecomposable R-module
M with rankq1,1 (M ) = r, rankq1,2 (M ) = s,
andnrankqi,j (M ) = 0 for all (i,oj) 6∈ {(1, 1), (1, 2)}. In either case, using Construction 4.4 we see that the
T
set x 0 · · · 0 : x ∈ Z occurs as a set of columns for A(R) and hence occurs as a subset of the
set of classes containing prime divisors.
4.3. Divisor-closed submonoids of T (R). Suppose that R has infinite representation type but, in
b has finite representation type for each minimal prime q of
contrast to Theorem 4.14, suppose that R/q
b
R. Then there is no known classification of all ranks of indecomposable finitely generated torsion-free
R-modules. Specific rings have been studied in the literature, but even in these settings, a complete
solution has been unattainable. We now give such an example which we will return to in Section 4.4.
Example 4.15. Let K be an algebraically closed field of characteristic zero. Consider the ring S =
K[[x, y]]/(x4 − xy 7 ) which has exactly two minimal primes xS and (x3 − y 7 )S. Detailed constructions
in [51] and [61] show that S has indecomposable modules of ranks (m, m), (m + 1, m), and (m + 2, m)
for each positive integer m. Moreover, [8, Proposition 6.2] guarantees indecomposable modules of ranks
(s − (m + 1), s − m) and (t − (m + 2), t − m) where s ≥ m + 1 and t ≥ m + 2 are positive integers.
Determining what other tuples occur as ranks of indecomposable torsion-free S-modules appears to be
quite difficult.
Thus, since studying T (R) as a whole is out of reach at the present state of knowledge, we pick finitely
many R-modules M1 , . . . , Mn , and study the direct-sum relations among them. In more technical terms,
instead of studying the full Krull monoid T (R), we focus on divisor-closed submonoids. Suppose that H is
a Krull monoid and H ֒→ F (P) a cofinal divisor homomorphism. If H ′ ⊂ H is a divisor-closed submonoid,
then H ′ ֒→ H ֒→ F (P) is a divisor homomorphism. For each of the arithmetical invariants ∗(·) introduced
in Section 2, we have ∗(H ′ ) ≤ ∗(H ′ ) resp. ∗(H ′ ) ⊂ ∗(H); for example we have c(H ′ ) ≤ c(H), L(H ′ ) ⊂
L(H), and so on. Moreover, if H ′ is the smallest divisor-closed submonoid containing finitely many
elements a1 , . . . , ak ∈ H, it is also the smallest divisor-closed submonoid containing a1 · . . . · ak .
For the rest of Section 4, we study divisor-closed submonoids of T (R) generated by a single R-module
M , regardless of whether R has finite or infinite representation type. We denote this monoid by add(M ).
Before discussing specific examples in Section 4.4, we carefully recall the consequences of our main
Construction 4.4 for such submonoids.
b be as in Construction 4.4. Let M be a finitely generated torsion-free
Construction 4.16. Let R and R
R-module. Then add(M ) consists of all isomorphism classes [N ] ∈ T (R) such that N is isomorphic to a
direct summand of M (n) for some finite positive integer n.
c = L(n1 ) ⊕ · · · ⊕ L(nk ) where the Li are pairwise nonisomorphic indecomposable finitely genWrite M
1
k
b
b ] ∈ add(M
c) and
erated torsion-free R-modules
and the ni are positive integers. If [N ] ∈ add(M ), then [N
18
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
(a )
(a )
b N
b ∼
thus, since direct-sum decomposition is essentially unique over R,
= L1 1 ⊕ · · · ⊕ Lk k with each ai
(k)
a nonnegative integer at most ni . Thus there is a divisor homomorphism Ψ : add(M ) → N0 given by
(k)
[N ] 7→ (a1 , . . . , ak ). We identify add(M ) with the saturated submonoid Γ(M ) = Ψ(add(M )) of N0 .
Moreover,
is the transpose of the
if A(M ) is the spl(R)×k integer-valued matrix for which the lth column
row-vector r1,1 − r1,2 · · · r1,1 − r1,t1 · · · rs,1 − rs,2 · · · rs,1 − rs,ts where ri,j = rankqi,j (Vl ),
∼ Γ(M ) = ker(A(M )) ∩ N(k) .
then add(M ) =
0
We now state a corollary of Theorem 4.8 for add(M ).
Corollary 4.17. Let M be a finitely generated torsion-free R-module as in Construction 4.16.
c) is free abelian.
(1) If spl(R) = 0, then add(M ) ∼
= add(M
(k)
(2) If spl(R) ≥ 1 and A(M ) satisfies the conditions of Lemma 4.7, then the inclusion Γ(M ) ⊂ N0 is a
divisor theory. Moreover,
(k)
(a) add(M ) ∼
= ker(A(M )) ∩ N0 ,
(b) C(add(M )) ∼
= Z(spl(R)) , and this isomorphism maps the set of classes containing prime divisors
onto the set of distinct columns of A(M ).
Before considering explicit examples, we give a realization result (see also [56, Chapter 1]).
Proposition 4.18. Let H be a reduced Krull monoid with free abelian class group G of rank q and let
GP ⊂ G denote the set of classes containing prime divisors. Suppose that GP is finite and that G has a
basis (e1 , . . . , eq ) such that G0 = {e0 = e1 + · · · + eq , e1 , . . . , eq , −e0 , . . . , −eq } ⊂ GP . Then there exists
an analytically unramified commutative Noetherian local domain S and a finitely generated torsion-free
S-module M such that add(M ) ∼
= H.
Proof. Let Φ : G → Z(q) denote the isomorphism which maps (e1 , . . . , eq ) onto the standard basis of Z(q) .
b
Let S be an analytically unramified Noetherian local domain with completion
S having q + 1 minimal
b 0 has infinite representation type. For s = s1 · · · sq ∈ Φ(GP ), set
primes Q0 , . . . , Qq such thatPS/Q
r0 = s1 + · · · + sq and ri = j6=i sj for each i ∈ [1, q]. By Theorem 4.13 there exists an indecomposable
L
b
finitely generated torsion-free S-module
Ns such that rank(Ns ) = (r0 , . . . , rq ). Set N =
s∈Φ(GP ) Ns
(a−ai )
Lq b
and write rank(N ) = (a0 , . . . , aq ). Set a = max{a0 , . . . , aq } and L = i=0 S/Qi
. Then N ⊕ L
b
is a finitely generated torsion-free S-module
with constant rank and is thus extended from a finitely
generated torsion-free S-module M . By Construction 4.16 and Corollary 4.17 we see that add(M ) has
class group isomorphic to Z(q) and this isomorphism maps the set of prime divisors onto the elements of
the set Φ(GP ).
4.4. Examples. In this section we provide the constructions of naturally occurring monoids add(M )
where M is a finitely generated torsion-free R-module. In particular, we construct specific modules M ,
c is often a direct sum of indecomposable cyclic R-modules
b
whose completion M
and we determine the class
group G of add(M ) and the set of classes GP ⊂ G containing prime divisors. Note that the Krull monoids
M(R) of all finitely generated R-modules and T (R) of all finitely generated torsion-free R-modules have
class groups G′ ⊃ G and a set G′P of classes containing prime divisors such that G′P ⊃ GP . Since add(M )
is a divisor-closed submonoid of both M(R) and of T (R), a study of the arithmetic of add(M ) provides
a partial description of M(R) and T (R). Moreover, the values of arithmetical invariants of add(M ) give
lower bounds on the same arithmetical invariants of M(R) and T (R).
b
In each of the following examples we construct an R-module
L = Ln1 1 ⊕ · · · ⊕ Lnk k of constant rank,
b
where L1 , . . . , Lk are pairwise non-isomorphic indecomposable R-modules.
Then, by Corollary 4.17 with
MONOIDS OF MODULES AND ARITHMETIC OF DIRECT-SUM DECOMPOSITIONS
19
c∼
M
= L for some R-module M ,
(k)
(k)
add(M ) ∼
= ker A(M ) ∩ N0 ⊂ N0 ∼
= add(L) .
In particular, we do so in such a way that the natural map add(M ) ֒→ add(L) is a divisor theory with
class group isomorphic to Z(spl(R)) and where the set of classes containing prime divisors maps onto the
distinct columns of A(M ).
Example 4.19. We now construct a monoid of modules whose arithmetic will be studied in Proposition
6.12. Let S be as in Example 4.15. Then there are indecomposable torsion-free S-modules M1 , M−1 , M2 ,
M−2 , N1 , N−1 , N2 , and N−2 with ranks (respectively) (2, 1), (1, 2), (3, 1), (1, 3), (3, 2), (3, 2), (2, 3), (4, 2),
and (2, 4). Set L to be the direct sum of these eight indecomposable S-modules. By Lech’s Theorem [54],
b∼
there exists a Noetherian local domain (R, m) with m-adic completion R
= S. Since L has constant rank,
(8)
∼
L is extended from some R-module M , and add(M ) ֒→ add(L) = N0 is a divisor theory with infinite
cyclic class group G and with GP = {−2e, −e, e, 2e} where G = hei.
Example 4.20. We now provide an example that illustrates the convenience of choosing an alternate
defining matrix for add(M ), as is described in Remark 4.10. Its arithmetic is given in Theorem 6.4.
b has five minimal prime ideals q(1,1) ,
Suppose that R has two minimal prime ideals p1 , p2 and that R
q(1,2) , q(1,3) , q(2,1) , and q(2,2) , with q(i,j) lying over pi for each i ∈ [1, 2] and for each j. Set
L =
b
R
b
R
⊕
⊕
b
R
⊕
q(1,2) ∩ q(2,1)
q(1,1) ∩ q(1,3) ∩ q(2,2)
q(1,1) ∩ q(1,2) ∩ q(2,1)
b
b
b
R
R
R
⊕
⊕
.
q(1,1)
q(1,3) ∩ q(2,2)
q(1,2) ∩ q(1,3) ∩ q(2,1) ∩ q(2,2)
(6)
c∼
Since L has constant rank 3, there is an R-module M such that M
= L. Then add(M ) ∼
= ker(A(M ))∩N0
where
1 0 −1 1 0 −1
1 0 −1 1 0 −1
1 0 −1 1 0 −1
A(M ) = 0 1 0 1 −1 −1 ∼ −1 1 1 0 −1 0 ∼ 0 1 0 1 −1 −1 = JA(M ).
−1 1 1 0 −1 0
0 0 0 0 0
0
0 0 0 0 0
0
1 0 −1 1 0 −1
(6)
(6)
Thus add(M ) ∼
∩ N0 . Since the matrix A(M ) has
= ker
= ker(JA(M )) ∩ N0 ∼
0 1 0 1 −1 −1
rank two, the representation of add(M ) as a Diophantine matrix defined by two equations more clearly
describes this monoid. Moreover, since the map from Z(6) to Z(2) is surjective (the map A(M ) : Z(6) →
Z(3) is not surjective), we immediately see that C(add(M )) ∼
= Z(2) , and this isomorphism maps the set of
classes containing prime divisors onto
1
0
−1
0
1
−1
,
,
,
,
,
.
0
1
0
−1
1
−1
Example 4.21. We now consider a monoid add(M ) which generalizes the monoid T (R) when R has
finite representation type, and its arithmetic is studied in Theorem 6.7 and Corollary 6.10. Suppose that
b has q + 1 minimal primes q1 , . . . , qq+1 , and set
R
L=
M
∅6=I⊂[1,q+1]
b
R
∩i∈I qi
.
b
R
b
From the symmetry of the set of ranks of the indecomposable cyclic R-modules
∩i∈I qi we immediately see
that L has constant rank (2q , . . . , 2q ) and is therefore extended from some R-module M . Then add(M ) ∼
=
20
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
T
(q)
ker(A(M ))∩N0 where A(M ) is an q × 2q+1 − 1 integer-valued matrix with columns ǫ1 · · · ǫq where
either ǫi ∈ {0, 1} for all i ∈ [1, q] or ǫi ∈ {0, −1} for all i ∈ [1, q].
(q)
Since the columns of A(M ) contain a basis for Z(q) , add(M ) ֒→ add(L) ∼
= N0 is a divisor theory with
(q)
∼
class group C(add(M )) = Z , and this isomorphism maps the set of classes containing prime divisors
onto
o
o n
n
T
T
ǫ1 · · · ǫq
: ǫi ∈ {0, −1} \ 0 · · · 0 .
: ǫi ∈ {0, 1} ∪ ǫ1 · · · ǫq
Example 4.22. In this example we construct a monoid add(M ) which generalizes the monoid of Example
4.21 by including all vectors having entries in {−1, 0, 1} in the set GP . This larger set of classes containing
prime divisors adds much complexity to the arithmetic. Suppose that R has q minimal primes and that
b has 2q minimal primes
R
q(1,1) , q(1,2) , q(2,1) , . . . , q(q,2)
where q(i,j) ∩ R = q(i′ ,j ′ ) ∩ R if and only if i = i′ . As in the previous example, let
M
L=
∅6=I⊂{(1,1),...,(q,2)}
b
R
∩(i,j)∈I q(i,j)
.
b
From the symmetry of the set of ranks of the indecomposable cyclic R-modules
2q−1
b
R
∩{i,j}∈I qi,j
we immediately
2q−1
see that L has constant rank (2
,...,2
) and is therefore extended from some R-module M . Then
(q)
add(M ) ∼
= ker(A(M )) ∩ N0 where A(M ) is an q × 22q − 1 integer-valued matrix with columns of the
T
form r(1,1) − r(1,2) r(2,1) − r(2,2) · · · r(q,1) − r(q,2) where (r(1,1) , r(1,2) , . . . , r(q,2) ) is the rank of one
b
of the 22q − 1 indecomposable cyclic R-modules
— that is, any one of the q-tuples of 1s and 0s (not all
T
0). In other words, the columns of A(M ) are exactly the 3q columns ǫ1 · · · ǫq where ǫi ∈ {−1, 0, 1}
for all i ∈ [1, q], repeated with some multiplicity. For example, the column of all zeros occurs for each of
b
R
b
the indecomposable cyclic R-modules
∩(i,j)∈I q(i,j) where (i, 1) ∈ I if and only if (i, 2) ∈ I.
(22q −1)
is a divisor theory
Since the columns of A(M ) contain a basis for Z(q) , add(M ) ֒→ add(L) ∼
= N0
(q)
Z
,
and
this
isomorphism
maps
the
set
of
classes
containing
prime
divisors onto
whose class group G ∼
=
n
o
T
ǫ1 · · · ǫq
: ǫi ∈ {−1, 0, 1} .
Example 4.23. In this example we consider add(M ) when the completion of M is isomorphic to a
b
direct sum of some (but not all) of the indecomposable cyclic R-modules.
In this case, the example is
constructed in such a way that B(GP ) is a direct product of non-trivial submonoids (see Lemma 3.2).
b has 3q minimal primes
Suppose that R has q minimal primes and that R
q(i,j) : i ∈ [1, q], j ∈ [1, 3]
b
where q(i,j) ∩ R = q(i′ ,j ′ ) ∩ R if and only if i = i′ . Let L be the R-module
q
M
b (i,1) ⊕ R/q
b (i,2) ⊕ R/q
b (i,3) ⊕ R/
b q(i,1) ∩ q(i,2) ⊕ R/
b q(i,1) ∩ q(i,3) ⊕ R/
b q(i,2) ∩ q(i,3) .
R/q
i=1
We see immediately that L has constant rank (3, . . . , 3) and thus L is extended from some R-module M .
(2q)
Then add(M ) ∼
= ker(A(M )) ∩ N0 where A(M ) is an 2q × 6q integer-valued matrix with columns
{e2k−1 , e2k , e2k−1 + e2k , −e2k−1 , −e2k , −e2k−1 − e2k : k ∈ [1, q]}
where (e1 , . . . , e2q ) denotes the canonical basis of Z(2q) .
For k ∈ [1, q], we set Gk = {e2k−1 , e2k , e2k−1 +e2k , −e2k−1 , −e2k , −e2k−1 −e2k }. Then GP = ⊎k∈[1,q] Gk
is the set of classes containing prime divisors and B(GP ) = B(G1 ) × · · · × B(Gq ). From Proposition 6.1
MONOIDS OF MODULES AND ARITHMETIC OF DIRECT-SUM DECOMPOSITIONS
21
we will see that B(Gk ) ֒→ F (Gk ) is a divisor theory, whence B(GP ) ֒→ F (GP ) and add(M ) ֒→ add(L)
are divisor theories. The arithmetic of this monoid is studied in Proposition 6.12 and Corollary 6.15.
Example 4.24. As in Example 4.23, suppose that R has q minimal primes and suppose that the comb of R has 3q minimal primes
pletion R
q(i,j) : i ∈ [1, q], j ∈ [1, 3]
b = S/(f ) where (S, n) is
where q(i,j) ∩ R = q(i′ ,j ′ ) ∩ R if and only if i = i′ . Further suppose that R
b (i,j) has infinite
a regular Noetherian local ring of dimension two and where 0 6= f ∈ n and that R/q
representation type for all pairs (i, j). By Proposition 4.13, for each k ∈ [1, q] there are indecomposable
b
finitely generated torsion-free R-modules
Mk and Nk of ranks (r1,1 , . . . , rq,3 ) and (s1,1 , . . . , sq,3 ) where
0, i 6= k
0, i 6= k
3, i = k, j = 1
ri,j = 2, i = k, j ∈ [1, 2]
and si,j =
.
2, i = k, j = 2
0, i = k, j = 3
0, i = k, j = 3
Moreover, by Corollary 4.9, for each k ∈ [1, q] there are constant tk ≥ 2 and t′k ≥ 3 and indecomposable
′
′
b
finitely generated torsion-free R-modules
Mk′ and Nk′ having ranks (r1,1
, . . . , rq,3
) and (s′1,1 , . . . , s′q,3 )
where
′
tk , i 6= k
t′ − 3, i = k, j = 1
tk , i 6= k
′
k
and s′i,j =
ri,j
= tk − 2, i = k, j ∈ [1, 2]
.
′
t
i = k, j = 2
k − 2,
tk , i = k, j = 3
′
tk , i = k, j = 3
Let
!
q
M
′
′
b
b
b
b
b
L=R⊕
(Mk ⊕ Nk ⊕ Mk ⊕ Nk ) ⊕ R/(q(i,1) ∩ q(i,3) ) ⊕ R/(q(i,1) ∩ q(i,2) ) ⊕ R/q(i,2) ⊕ R/q(i,3) .
k=1
(2q)
Since L has constant rank, L is extended from an R-module M . Then add(M ) ∼
= ker(A(M )) ∩ N0
where A(M ) is an 2q × (8q + 1) integer-valued matrix with columns 0 and
{e2k−1 , e2k , 2e2k , e2k−1 + 2e2k , −e2k−1 , −e2k , −2e2k , −e2k−1 − 2e2k : k ∈ [1, q]}
where (e1 , . . . , e2q ) denotes the canonical basis of Z(2q) .
For k ∈ [1, q], we set Gk = {e2k−1 , e2k , 2e2k , e2k−1 + 2e2k , −e2k−1 , −e2k , −2e2k , −e2k−1 − 2e2k }. Then
GP = ⊎k∈[1,q] Gk is the set of classes containing prime divisors and B(GP ) = B(G1 ) × · · · × B(Gq ). From
Proposition 6.1 we will see that B(Gk ) ֒→ F (Gk ) is a divisor theory, whence B(GP ) ֒→ F (GP ) and
add(M ) ֒→ add(N ) are divisor theories. The arithmetic of this monoid is studied in Proposition 6.13 and
Corollary 6.15.
Example 4.25. In our final example we construct a tuple (G, GP ) which generalizes the monoid T (R)
when R has finite representation type (see Theorem 4.12). The arithmetic of such Krull monoids is
b has q + 1 minimal primes q1 , . . . , qq+1 , and
studied in Theorem 6.4 and Corollary 6.10. Suppose that R
set
q+1
M
b j
b ∩i6=j qi ⊕ R/q
.
R/
L=
j=1
Note that L has constant rank (q, . . . , q) and is hence extended from some R-module M . Then add(M ) ∼
=
(q)
ker(A(M )) ∩ N0 where A(M ) is an q × 2q integer-valued matrix with columns
e1 , . . . , eq , e0 = e1 + e2 + · · · + eq , −e1 , . . . , −eq , −e0 .
22
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
(2q)
By Proposition 6.1, add(M ) ֒→ add(L) ∼
= Z(q) ,
= N0 is a divisor theory with class group C(add(M )) ∼
and this isomorphism maps the set of classes containing prime divisors onto
{e0 = e1 + · · · + eq , e1 , . . . , eq , −e0 , . . . , −eq }.
5. Monoids of modules: Class groups and distribution of prime divisors II
In this section we investigate the characteristic of the Krull monoids T (R) and C(R) for two-dimensional
Noetherian local Krull domains (see Theorem 5.4). We will show that, apart from a well-described exb
ceptional case, their class groups are both isomorphic to the factor group C(R)/ι(C(R)),
where ι : C(R) →
b
b
C(R) is the natural homomorphism between the class groups of R and R. In a well-studied special case
b is a hypersurface with finite representation type, this factor group is a finite
where R is factorial and R
cyclic group (see Theorem 5.5). This is in strong contrast to the results on one-dimensional rings in the
previous section where all class groups are torsion-free.
Let S be a Krull domain and let Iv∗ (S) denote the monoid of nonzero divisorial ideals. Then ϕ : S →
∗
Iv (S), defined by a 7→ aS, is a divisor theory. In this section we view C(S) as the class group of this
specific divisor theory. First we give a classical result (see [13, Chapter VII, Section 4.7]).
Lemma 5.1. Let S be a Noetherian Krull domain. One can associate to each finitely generated S-module
M a class c(M ) ∈ C(S) in such a way that
(1) If 0 → M ′ → M → M ′′ → 0 is an exact sequence of finitely generated S-modules, then c(M ) =
c(M ′ ) + c(M ′′ ).
(2) If I is a fractional ideal of S and Iv the divisorial ideal generated by I, then c(I) = c(Iv ).
Note that if S is any Noetherian domain, every ideal of S is obviously an indecomposable finitely
generated torsion-free S-module. If, in addition, the ring has dimension two, then we have the following
stronger result.
Lemma 5.2. Let S be a Noetherian local Krull domain of dimension two.
(1) [21, Lemma 1.1 and Theorem 3.6] Every divisorial ideal of S is an indecomposable MCM S-module.
(2) [60, Proposition 3] In addition, assume that the m-adic completion Sb of S is a Krull domain. Then
b
a finitely generated torsion-free S-module
N is extended from an S-module if and only if c(N ) is in
b
the image of the natural homomorphism ι : C(S) → C(S).
We now give a result on abstract Krull monoids which encapsulates the structure of the monoids of
modules described in Theorem 5.4.
Lemma 5.3. Let D = F (P) be a free abelian monoid, G an additive abelian group, ψ : D → G a
homomorphism, and H = ψ −1 (0) ⊂ D.
(1) If H ⊂ D is cofinal, then the inclusion H ֒→ D is a divisor homomorphism and ψ : D/H → ψ(D) ⊂ G
given by ψ ([a]) = ψ(a) is an isomorphism.
(2) The inclusion H ֒→ D is a divisor theory if and only if hψ(P)i = [ψ(P\{q})] for every q ∈ P. If this
is the case, then ψ : D/H → ψ(D) is an isomorphism and, for every g ∈ ψ(D), the set P ∩ ψ −1 (g)
−1
is the set of prime divisors in the class ψ (g).
(3) If the restriction ψ |P : P → G of ψ to P is an epimorphism, then H ֒→ D is cofinal. Moreover, it
is a divisor theory apart from the following exception:
G = {0, g}
and
|P ∩ ψ −1 (g)| = 1 .
If H ֒→ D is not a divisor theory, then H is factorial.
MONOIDS OF MODULES AND ARITHMETIC OF DIRECT-SUM DECOMPOSITIONS
23
Proof. For the proofs of (1) and (2), see [37, Proposition 2.5.1]. We now consider the proof of (3).
Let a ∈ D. Since ψ |P : P → G is an epimorphism, there exists p ∈ P ⊂ D such that ψ(p) = −ψ(a).
Therefore ap ∈ H and the inclusion H ֒→ D is cofinal. In order to show that H ֒→ D is a divisor theory
we distinguish three cases. First suppose that |G| = 1. Then |D/H| = 1, and hence H = D. Next
suppose that |G| > 2. By (2) we must verify that
ψ(q) ∈ [ψ(P\{q})] for every q ∈ P .
Let q ∈ P. Since |G| > 2, there exist g1 , g2 ∈ G \ {0, ψ(q)} with ψ(q) = g1 + g2 . Since the restriction
ψ |P : P → G is an epimorphism, there exist p1 , p2 ∈ P \ {q} with ψ(pi ) = gi for i ∈ [1, 2]. Therefore
ψ(q) = g = g1 + g2 = ψ(p1 ) + ψ(p2 ) ∈ [ψ(P\{q})] .
Finally, suppose that |G| = 2. Then H ֒→ D is a divisor theory if and only if hψ(P)i = [ψ(P\{q})]
for every q ∈ P if and only if there exist distinct q1 , q2 ∈ P such that ψ(qi ) 6= 0 for i ∈ [1, 2]. Clearly, if
q ∈ P is the unique element of P with ψ(q) 6= 0, then H is free abelian with basis P \ {q} ∪ {q 2 }.
We are now able to determine both the class group and the set of classes containing prime divisors for
the monoids T (R) and C(R). This generalizes and refines the results of [4]. Since each divisorial ideal
over a two-dimensional local ring is MCM and thus finitely generated and torsion-free, Theorem 5.4 can
be stated in parallel both for T (R) and C(R).
Theorem 5.4. Let (R, m) be a Noetherian local Krull domain of dimension two whose m-adic completion
b is also a Krull domain. Let V(R) (respectively V(R))
b denote either T (R) (resp. T (R))
b or C(R) (resp.
R
b
b
C(R)), and let ι : C(R) → C(R) be the natural map.
b is a cofinal divisor homomorphism. The class group of this divisor ho(1) The embedding V(R) ֒→ V(R)
b
momorphism is isomorphic to G = C(R)/ι(C(R))
and every class contains a prime divisor. Moreover
b satisfies the following condition:
the embedding is a divisor theory except if R
b
(E) |G| = 2 and, up to isomorphism, there is precisely one non-extended indecomposable R-module
b
M with [M ] ∈ V(R).
b satisfies (E).
In particular, V(R) satisfies KRSA if and only if either |G| = 1 or R
b
b
(2) Suppose that the embeddings T (R) ֒→ T (R) and C(R) ֒→ C(R) are both divisor theories. Then their
class groups are isomorphic. If (G, (mg )g∈G ) is the characteristic
P of T (R) and (G, (ng )g∈G ) is the
characteristic of C(R), then mg ≥ ng for all g ∈ G. Moreover, g∈G mg infinite.
b By (1) of Lemma 5.1 there is a homomorphism
Proof. We set D = F (P) = V(R).
b → C(R)/ι(C(R))
b
ψ : V(R)
=G
given by [M ] 7→ c(M ) + ι(C(R)) .
b is an indecomposable MCM R-module.
b
By (1) of Lemma 5.2, every divisorial ideal of R
That is, the class
b
b
b
of each divisorial ideal of R in C(R) is the image of some [M ] ∈ V(R), where M is an indecomposable MCM
b
b
R-module.
In other words, ψ restricted to P = A(D) is an epimorphism and ψ(A(D)) = C(R)/ι(C(R)).
−1
By Lemma 5.3, the inclusion H = ψ (0) ⊂ D is a cofinal divisor homomorphism. By (2) of Lemma
b and thus the embedding V(R) ֒→ V(R)
b is a divisor
5.2, H is the image of the embedding V(R) ֒→ V(R),
theory if and only if the inclusion H ֒→ D is a divisor theory. By Lemma 5.3 this always holds apart from
the described Exception (E). A Krull monoid is factorial if and only if its class group is trivial. Thus, if
b is a divisor theory then KRSA holds for V(R) if and only if |G| = 0. If V(R) ֒→ V(R)
b is
V(R) ֒→ V(R)
not a divisor theory, then the inclusion H ֒→ D is not a divisor theory. By Lemma 5.3, H is factorial,
whence V(R) is factorial.
Since each MCM R-module is finitely generated and torsion-free, it is clear that mg ≥ ng for each
g ∈ G. From [11] we know that there infinitely many nonisomorphic indecomposable finitely generated
P
b
torsion-free R-modules,
and therefore g∈G mg infinite.
24
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
Let
•
•
R be as in the above theorem and assume in addition that
R contains a field and k = R/m is algebraically closed with characteristic zero.
b is a hypersurface i.e., R
b is isomorphic to a three-dimensional regular Noetherian local ring
R
modulo a regular element.
• R has finite representation type.
Such rings were classified in [15] and [52] and are given, up to isomorphism, in Table 1. Note that since
b has finite representation type and each divisorial ideal of R
b is and indecomposable MCM R-module,
b
R
b
C(R) and hence C(R) is finite.
Table 1: Two-dimensional Rings with Finite Representation Type
(An )
(Dn )
(E6 )
(E7 )
(E8 )
k[[x, y, z]]/(x2 + y 2 + z n+1 )
n≥1
k[[x, y, z]]/(x2 z + y 2 + z n−1 ) n ≥ 4
k[[x, y, z]]/(x3 + y 2 + z 4 )
k[[x, y, z]]/(x3 + xz 3 + y 2 )
k[[x, y, z]]/(x2 + y 3 + z 5 )
An amazing theorem of Heitmann [48] gives the existence of a local factorial domain whose completion
is a ring as in Table 1. In this situation we can determine the characteristic of C(R).
b isomorphic
Theorem 5.5. Let (R, m) be a Noetherian local factorial domain with m-adic completion R
to a ring in Table 1.
b is a ring of type (An ), then C(C(R)) is cyclic of order n + 1 and each class contains exactly one
(1) If R
prime divisor.
b is a ring of type (Dn ).
(2) Suppose R
(a) If n is even, then C(C(R)) ∼
= C2 ⊕ C2 . The trivial class contains n2 prime divisors. Two nontrivial classes each contain a single prime divisor and their sum contains n−2
2 prime divisors.
(b) If n is odd, then C(C(R)) is cyclic of order four. The classes of order four each contain a single
prime. The remaining classes each contain n−1
2 prime divisors.
b is a ring of type (E6 ), then C(C(R)) is cyclic of order three. The trivial class contains three
(3) If R
prime divisors, while each remaining class contains two prime divisors.
b is a ring of type (E7 ), then C(C(R)) is cyclic of order two. The trivial class contains five prime
(4) If R
divisors and the non-trivial class contains three prime divisors.
b is a ring of type (E8 ), then C(C(R)) is trivial, with the trivial class containing all nine prime
(5) If R
divisors.
b for R
b a ring listed in Table 1 were given in [14]. Since R is factorial,
Proof. The class groups C(R)
b
Following the proof of [4, Theorem 4.3] one can
C(R) = 0 and by Theorem 5.4, C(C(R)) ∼
= C(R).
b
b by using the Auslander-Reiten sequence for
compute the class of each indecomposable R-module in C(R)
b The result follows by considering the map ψ defined in the proof of Theorem 5.4.
R.
The above theorem completely determines the characteristic of the monoid C(R). With this information, in addition to being able to completely describe the arithmetic of C(R) as we do in Theorem 6.8, we
can easily enumerate the atoms of C(R) (the nonisomorphic indecomposable MCM modules). We now
illustrate this ability with an example. If β : C(R) → B(GP ) is the transfer homomorphism of Lemma
b is a ring of type (Dn ) with n even. Then R
b has
3.4, then A(C(R)) = β −1 (A(B(GP )). Suppose that R
exactly n + 1 nonisomorphic indecomposable MCM modules. If C2 ⊕ C2 = {0, e1 , e2 , e1 + e2 }, then
A(C2 ⊕ C2 ) = {0, e21 , e22 , (e1 + e2 )2 , e1 e2 (e1 + e2 )} ,
MONOIDS OF MODULES AND ARITHMETIC OF DIRECT-SUM DECOMPOSITIONS
25
and hence R has exactly
n
n−2 n−2
n−2
n2 + 8
+1·1+1·1+
·
+1·1·
=
2
2
2
2
4
nonisomorphic indecomposable MCM modules.
|A(C(R))| =
We conclude this section by noting that a two-dimensional local Krull domain (R, m) having completion
isomorphic to a ring in Table 1 may not be factorial. However, Theorem 5.4 implies that C(C(R)) is a
factor group of a group given in Theorem 5.5. In particular, C(C(R)) is a finite cyclic group such that
every class contains a prime divisor, and thus the arithmetic of C(R) is described in Theorem 6.8.
6. The arithmetic of monoids of modules
In this section we study the arithmetic of the Krull monoids that have been discussed in Sections 4
and 5. Thus, using the transfer properties presented in Section 2, we describe the arithmetic of directsum decompositions of modules. Suppose that H is a Krull monoid having a divisor homomorphism
ϕ : H → F (P) and let GP ⊂ C(ϕ) be the set of classes containing prime divisors. The first subsection deals
with quite general sets GP and provides results on the finiteness or non-finiteness of various arithmetical
parameters. The second subsection studies three specific sets GP , provides explicit results on arithmetical
parameters, and establishes a characterization result (Theorems 6.4, 6.7, 6.8, and Corollary 6.10) . The
third subsection completely determines the system of sets of lengths in case of small subsets GP . It
shows that small subsets in torsion groups and in torsion-free groups can have the same systems of sets
of lengths, and it reveals natural limits for arithmetical characterization results (Corollary 6.15).
6.1. General sets GP of classes containing prime divisors. In this subsection we consider the
algebraic and arithmetic structure of Krull monoids with respect to GP . We will often assume that
GP = −GP , a property which has a strong influence on the arithmetic of H. Recall that GP = −GP
holds in many of the (finite and infinite representation type) module-theoretic contexts described in
Sections 4 and 5. More generally, all configurations (G, GP ) occur for certain monoids of modules (see
Proposition 4.18, [49], and [56, Chapter 1]) and, by Claborn’s Realization Theorem, all configurations
(G, GP ) occur for Dedekind domains (see [37, Theorem 3.7.8]). In addition, every abelian group can
be realized as the class group of a Dedekind domain which is a quadratic extension of a principal ideal
domain, and in this case we have GP = −GP (see [55]).
Proposition 6.1. Let H be a Krull monoid, ϕ : H → F (P) a divisor homomorphism with class group
G = C(ϕ), and let GP ⊂ G denote the set of classes containing prime divisors.
(1) If GP is finite, then A(GP ) is finite and hence D(GP ) < ∞. If G has finite total rank, then GP is
finite if and only if A(GP ) is finite if and only if D(GP ) < ∞.
(2) If GP = −GP , then [GP ] = G. Moreover, the map ϕ : H → D and the inclusion B(GP ) ֒→ F (GP )
are both cofinal.
(3) Suppose that G is infinite cyclic, say G = hei, and that {−e, e} ⊂ GP . Then B(GP ) ֒→ F (GP ) is a
divisor theory if and only if there exist k, l ∈ N≥2 such that −ke, le ∈ GP .
(4) Let r, α ∈ N with r + α > 2. Let (e1 , . . . , er ) ∈ GrP be independent and let e0 ∈ GP such that
αe0 = e1 + · · · + er , {−e0 , . . . , −er } ⊂ GP , and he0 , . . . , er i = G.
(a) The map ϕ : H → F (P) and the inclusion B(GP ) ֒→ F (GP ) are divisor theories with class group
isomorphic to G.
(b) If 0 ∈
/ GP , then B(GP ) is not a direct product of nontrivial submonoids.
26
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
Proof. (1) follows from [37, Theorem 3.4.2].
If GP = −GP , then [GP ] = hGP i = G. By Lemma 3.4, (2) follows once we verify that ϕ is cofinal. If
p ∈ P, then there is a q ∈ P with q ∈ −[p], whence there is an a ∈ H with ϕ(a) = pq, and so ϕ is cofinal.
If {−ke : k ∈ N} ∩ GP = {−e} or {ke : k ∈ N} ∩ GP = {e}, then B(GP ) is factorial. Since F (GP ) 6=
B(GP ), the inclusion B(GP ) ֒→ F (GP ) is not a divisor theory. Conversely, suppose that there exist
m
k
m
k, l ∈ N≥2 such that −ke, le ∈ GP . Let m ∈ N. If me ∈ GP , then
me
=
gcd
(me)(−e)
,
(me)
(−ke)
,
m
l
m
and if −me ∈ GP , then −me = gcd (−me)e , (−me) (le) . Thus every element of GP is a greatest
common divisor of a finite set of elements from B(GP ) and hence B(GP ) ֒→ F (GP ) is a divisor theory.
We now suppose that G and GP are as in (4). To prove (a) it is sufficient to show that B(GP ) ֒→
F (GP ) is a divisor theory. By [37, Proposition 2.5.6] we need only verify that hGP i = [GP \ {g}]
for every g ∈ GP . Let g ∈ G. We will show that [GP \ {g}] = [GP ] = [e0 , . . . , er , −e0 , . . . , −er ]. If
g∈
/ {e0 , . . . , er , −e0 , . . . , −er }, the assertion is clear. By symmetry it suffices to consider the case where
g ∈ {e0 , . . . , er }. If g = ei for some i ∈ [1, r], then ei = αe0 +(−e1 )+· · ·+(−ei−1 )+(−ei+1 )+· · ·+(−er ) ∈
[{±eν : ν ∈ [0, r] \ {i}], and hence [GP \ {g}] ⊃ [{±eν : ν ∈ [0, r] \ {i}}] = [±e0 , . . . , ±er ] = G. If g = e0 ,
then e0 = e1 + · · · + er + (α − 1)(−e0 ) ∈ [GP \ {e0 }], and hence [GP \ {g}] = [GP ] = G.
To prove (b) we use Lemma 3.2 and suppose that 0 ∈
/ GP with GP = G1 ⊎ G2 such that A(GP ) =
A(G1 ) ⊎ A(G2 ). We must show that either G1 or G2 is empty. Suppose that V = e1 · . . . · er (−e0 )α ∈
A(G1 ). Since (−e0 )e0 , . . . , (−er )er ∈ A(GP ), it follows that {±e0 , . . . , ±er } ⊂ G1 . Let g ∈ GP . Since
GP ⊂ G = [±e0 , . . . , ±er ], there exists U ∈ A(GP ) such that g ∈ supp(U ) ⊂ {g, ±e0 , . . . , ±er }, and
hence g ∈ G1 . Thus G1 = GP and G2 = ∅.
For our characterization results, we need to recall the concept of an absolutely irreducible element, a
classical notion in algebraic number theory. An element u in an atomic monoid H is called absolutely
irreducible if u ∈ A(H) and |Z(un )| = 1 for all n ∈ N; equivalently, the divisor-closed submonoid of H
generated by u is factorial. Suppose that H ֒→ F (P) is a divisor theory with class group G and that
u = pk11 · . . . · pkmm where m, k1 , . . . , km ∈ N and where p1 , . . . , pm ∈ P are pairwise distinct. Then u is
absolutely irreducible. if and only if (k1 , . . . , km ) is a minimal element of the set
Γ = {(s1 , . . . , sm ) ∈ Nl0 : ps11 · . . . · psmm ∈ H} \ {0}
with respect to the usual product ordering, and the torsion-free rank of h[p1 ], . . . , [pm ]i in G is m − 1 (see
ord([p1 ])
is absolutely irreducible,
[37, Proposition 7.1.4]). In particular, if [p1 ] ∈ G has finite order, then p1
and if [p1 ] ∈ G has infinite order, then p1 q1 is absolutely irreducible for all q1 ∈ P ∩ (−[p1 ]).
Proposition 6.2. Let H be a Krull monoid, ϕ : H → F (P) a cofinal divisor homomorphism with class
group G, and let GP ⊂ G denote the set of classes containing prime divisors.
(1) Suppose that GP is infinite.
(a) If GP has an infinite subset G0 such that G0 ∪ (−G0 ) ⊂ GP and hG0 i has finite total rank, then
Uk (H) is infinite for each k ≥ 2. Moreover, D(GP ) = ρk (H) = ω(H) = t(H) = ∞.
(b) If there exists e ∈ G such that both GP ∩ {ke : k ∈ N} and GP ∩ {−ke : k ∈ N} are infinite,
then ∆(H) is infinite and c(H) = cmon (H) = ∞.
(c) If GP contains an infinite group, then every finite subset L ⊂ N≥2 occurs as a set of lengths.
(2) Suppose that GP is finite and that H is not factorial.
(a) The set ∆(H) is finite and there is a constant M1 ∈ N such that every set of lengths is an
AAMP with difference d ∈ ∆∗ (H) and bound M1 .
(b) There is a constant M2 ∈ N such that, for every k ≥ 2, the set Uk (H) is an AAMP with period
{0, min ∆(H)} and bound M2 .
(c) c(H) ≤ ω(H) ≤ t(H) ≤ 1 +
cmon (H) <
|G•P |
2
D(GP )(D(GP )−1)
2
+2
and
|G• |+1
(2|G•P | + 2)(|G•P | + 2)(D(GP ) + 1) P
.
MONOIDS OF MODULES AND ARITHMETIC OF DIRECT-SUM DECOMPOSITIONS
27
(d) Suppose that GP = −GP . Then ω(H) = D(GP ), ρ(H) = D(GP )/2, ρ2k (H) = kD(GP ) < ∞,
and λkD(GP )+j (H) = 2k + j for all k ∈ N and j ∈ [0, 1]. If G is torsion-free, then D(GP ) is the
maximal number s of absolutely irreducible atoms u1 , . . . , us such that 2 ∈ L(u1 · . . . · us ).
(e) If, in particular, GP = −GP and D(GP ) = 2, then Uk (H) = {k} for all k ∈ N and cmon (H) =
c(H) = ω(H) = t(H) = 2.
Proof. Throughout the proof we implicitly assume the results of Lemma 2.2 and of Lemma 3.4. In
particular, we have ρ(H) ≤ ω(H) and c(H) ≤ ω(H) ≤ D(GP ).
For (1), suppose that GP is infinite. We first prove (a). Theorem 3.4.2 in [37] implies that A(G0 ) and
D(G0 ) are infinite. Thus, for every k ∈ N, there is Uk ∈ A(G0 ) with |Uk | ≥ k and hence L Uk (−Uk ) ⊃
{2, |Uk |}. This implies that U2 (G0 ) is infinite and thus Uk (G0 ) is infinite for all k ≥ 2. Therefore
ρk (H) = ρk (GP ) = ∞ for all k ≥ 2 and, since ρ(H) ≤ ω(H) ≤ t(H), each of these invariants is infinite.
(b) follows from [36, Theorem 4.2].
(c) is a realization result is due to Kainrath. See [50] or [37, Theorem 7.4.1].
Now, in order to prove (2), we suppose that GP is finite and that H is not factorial. Then D(GP ) > 1
and 2 ≤ c(H) ≤ ω(H) ≤ t(H). By Proposition 6.1, B(GP ) is finitely generated and D(GP ) < ∞. The
respective upper bounds given in (c) for cmon (H) and t(H) can be found in [41, Theorem 3.4] and [37,
Theorem 3.4.10].
We now consider (a). Since 2 + sup ∆(H) ≤ c(H) < ∞, (c) implies that ∆(H) is finite. Since B(GP )
is finitely generated, the assertion on the structure of sets of lengths follows from [37, Theorem 4.4.11].
Since t(H) < ∞ and ∆(H) is finite, [32, Theorems 3.5 and 4.2] imply the assertion in (b) on the
structure of the unions of sets of lengths.
In order to prove (d), we suppose that GP = −GP . The statements about ρ2k (H), ρ(H), and
λkD(GP )+j (H) follow from Lemma 3.3, and it remains to show that ω(H) = D(GP ). We have ω(H) ≤
D(GP ) < ∞. If D(GP ) = 2, then ω(H) = D(GP ). Suppose that D(GP ) ≥ 3. If V = g1 · . . . · gl ∈ A(GP )
with |V | = l = D(GP ) and Ui = (−gi )gi for all i ∈ [1, l], then V | U1 · . . . · Ul but yet V divides no proper
subproduct of U1 · . . . · Ul . Thus D(GP ) ≤ ω(GP ) ≤ ω(H).
Let t denote the maximal number of absolutely irreducible atoms with the required property. Since
ρ(H) = D(GP )/2, it follows that t ≤ D(GP ). Let V = g1 · . . . · gl ∈ A(GP ) with |V | = l = D(GP ). For
i ∈ [1, l] choose an element pi ∈ P ∩ gi and an element qi ∈ P ∩ (−gi ). Since G is torsion-free, the element
ui = pi qi ∈ H is absolutely irreducible for each i ∈ [1, l] and, by construction, we have 2 ∈ L(u1 · . . . · ul ).
The statement in (e) follows immediately from (c) and (d).
Let all notation be as in Proposition 6.2. We note that if GP is infinite but without a subset G0 as in
(1a), then none of the conclusions of (1a) need hold. A careful analysis of the case where G is an infinite
cyclic groups is handled in [36]. We also note that the description of the structure of sets of lengths given
in (2a) is best possible (see [64]).
By Lemma 3.4, many arithmetical phenomena of a Krull monoid H are determined by the tuple
(G, GP ). We now provide a first result indicating that conversely arithmetical phenomena give us back
information on the class group. Indeed, our next corollary characterizes arithmetically whether the class
group of a Krull monoid is torsion-free or not. To do so we must study the arithmetical behavior of
elements similar to absolutely irreducible elements. Note that such a result cannot be accomplished via
sets of lengths alone (see Propositions 6.12 and 6.13 and (1c) of Proposition 6.2; in fact, there is an open
conjecture that every abelian group is the class group of a half-factorial Krull monoid [35]).
Proposition 6.3. Let H be a Krull monoid with class group G. Then G has an element of infinite order
if and only if there exists an irreducible element u ∈ H having the following two arithmetical properties.
(a) Whenever there are v ∈ H \ H × and m ∈ N with v | um , then u | v n for some n ∈ N.
N
for each
(b) There exist l ≥ 2 and a1 , . . . , al ∈ H such that u | a1 · . . . · al but yet u ∤ a−1
ν (a1 · . . . · al )
ν ∈ [1, l] and for every N ∈ N.
28
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
Proof. We may assume that H is reduced. Consider a divisor theory H ֒→ F (P) and denote by GP ⊂ G
the set of classes containing prime divisors.
First suppose that G is a torsion group and let u ∈ A(H) have Property (a). Then u = pk11 · . . . · pkmm
for some m, k1 , . . . , km ∈ N and pairwise distinct elements p1 , . . . , pm ∈ P. Then (a) implies that k = 1
and hence u is absolutely irreducible. Thus Property (b) cannot hold for any l ≥ 2.
Conversely, suppose that G is not a torsion group. Since [GP ] = G there exists a p ∈ P such that [p] ∈ G
has infinite order, and there is an element u′ ∈ A(H) with p | u′ . Suppose that u′ = p1 · . . . · pn · q1 · . . . · qr ,
where p = p1 , p2 , . . . , pn , q1 , . . . , qr ∈ P, [p1 ], . . . , [pn ] have infinite order, and [q1 ], . . . , [qr ] have finite order
each of which divides some integer N . Then (q1 · . . . · qr )N ∈ H, whence (p1 · . . . · pn )N ∈ H. After a
possible reordering there is an atom u = pk11 · . . . · pkmm ∈ A(H) dividing a power of (p1 · . . . · pn )N such
that there is no atom v ∈ A(H) with suppP (v) ( {p1 , . . . , pm }. Thus u satisfies Property (a). Since
km−1
H ֒→ F (P) is a divisor theory, there exist b1 , . . . , bs ∈ H such that pk11 · . . . · pm−1
= gcd(b1 , . . . , bs ).
Hence there is an i ∈ [1, s], say i = 1, such that pm ∤ b1 . Similarly, there are c1 , . . . , ct ∈ H such that
pkmm = gcd(c1 , . . . , ct ). Without loss of generality, there exists i ∈ [1, m − 1] such that pi ∤ c1 . Therefore
N
u | b1 c1 , but yet u ∤ bN
1 and u ∤ c1 for any N ∈ N, and so Property (b) is satisfied.
Propositions 6.1, 6.2, and 6.3 provide abstract finiteness and non-finiteness results. To obtain more
precise information on the arithmetical invariants, we require specific information on GP . In the next
subsection we will use such specific information to give more concrete results.
6.2. Specific sets GP of classes containing prime divisors and arithmetical characterizations.
We now provide an in-depth study of the arithmetic of three classes of Krull monoids studied in Sections
4 and 5. Theorem 6.4 describes the arithmetic of the monoids discussed in Examples 4.12, 4.20, 4.25 and
in Theorem 4.12. Its arithmetic is simple enough that we can more or less give a complete description.
Theorem 6.4. Let H be a Krull monoid with class group G and suppose that
GP = {e0 , . . . , er , −e0 , . . . , −er } ⊂ G
is the set of classes containing prime divisors, where r, α ∈ N with r + α > 2 and (e1 , . . . , er ) is an
independent family of elements each having infinite order such that e1 + · · · + er = αe0 . Then:
(1) A(GP ) = {V, −V, Uν : ν ∈ [0, r]}, where V = (−e0 )α e1 · . . . · er and Uν = (−eν )eν for all ν ∈ [0, r].
In particular, D(GP ) = r + α.
(2) Suppose that
r
Y
eki i (−ei )li ∈ F (GP ) ,
S=
i=0
where k0 , l0 , . . . , kr , lr ∈ N0 . Then S ∈ B(GP ) if and only if li = α−1 (k0 − l0 ) + ki for all i ∈ [1, r].
If S ∈ B(GP ) with k0 ≥ l0 and k ∗ = min{k1 , . . . , kr }, then
r
o
n
Y
−1
Uiki −ν : ν ∈ [0, min{α−1 l0 , k ∗ }]
Z(S) = V ν (−V )α (k0 −l0 )+ν U0l0 −αν
i=1
and
L(S) = {α−1 (k0 − l0 ) + l0 + k1 + · · · + kr − (r + α − 2)ν : ν ∈ [0, min{α−1 l0 , k ∗ }] } .
(3) The system of sets of lengths of H can be described as follows.
(a) ∆(H) = {r + α − 2}.
(b) ρ(H) = D(GP )/2.
(c) For each k ∈ N, the set Uk (H) is an arithmetical progression with difference r + α − 2.
MONOIDS OF MODULES AND ARITHMETIC OF DIRECT-SUM DECOMPOSITIONS
29
(d) For each k ∈ N and each j ∈ [0, 1], ρ2k+j (H) = kD(GP ) + j.
(e) For each l ∈ N0 , λlD(GP )+j (H) = 2l + j whenever j ∈ [0, D(GP ) − 1] and lD(GP ) + j ≥ 1.
(f ) Finally,
o
n
L(H) = m + {2k ∗ + (r + α − 2)λ : λ ∈ [0, k ∗ ]} : m, k ∗ ∈ N0 .
(4) c(H) = cmon (H) = ω(H) = t(H) = D(GP ) = r + α.
Proof. By Lemma 3.4, all assertions on lengths of factorizations and on catenary degrees can be proved
working in B(GP ) instead of H.
Obviously, {Uν : ν ∈ [0, r]} ⊂ A(GP ) and to prove (1) it remains to verify that if W ∈ A(GP ) with
W 6= Uν , then W = V . Note that e0 ∈ he1 , . . . , er i but that he0 i ∩ hei : i ∈ Ii = {0} for any proper subset
I ( [1, r]. Thus, if
Y
W =
eki i (−e0 )k0 ∈ A(GP ) \ {Uν : ν ∈ [0, r]} ,
i∈I
P
where ∅ 6= I ⊂ [1, r] and k0 , ki ∈ N for all i ∈ I, then k0 e0 = i∈I ki ei ∈ hei : i ∈ Ii and hence I = [1, r].
Assume to the contrary that there is i ∈ [1, r] such that ki > 1. Since V ∈ B(GP ) and W ∈ A(GP ), it
follows that k0 ∈ [1, α − 1]. Then
0 6= (k1 − 1)e1 + · · · + (kr − 1)er = (k0 − α)e0 ∈ [e1 , . . . , er ] ∩ [−e1 , . . . , −er ] = {0} ,
a contradiction. Thus k1 = · · · = kr = 1 and we obtain that k0 = α, whence W = V ∈ A(GP ).
To prove (2), suppose that S ∈ B(GP ) and that l0 ≥ k0 . Then
S ′ = (−e0 )l0 −k0
r
Y
eki i (−ei )li ∈ B(GP ) ,
i=1
whence l0 − k0 = αm0 ∈ αN0 , S ′′ =
Qr
ki −m0
(−ei )li
i=1 ei
∈ B(GP ), and
l0 − k0
k0 − l0
=
+ ki for each i ∈ [1, r] .
α
α
The same holds true if l0 ≤ k0 . Conversely, if l1 , . . . , lr satisfy the asserted equations, then obviously
σ(S) = 0.
Suppose that S ∈ B(GP ) and that k0 ≥ l0 . Then
li = ki − m0 = ki −
S = ek00 (−e0 )l0
r
Y
−1
eki i (−ei )ki +α
(k0 −l0 )
i=1
= (−e0 )e0
= (−e0 )e0
l0
r
ki
α−1 (k0 −l0 ) Y
(−ei )ei
e0 α (−e1 ) · . . . · (−er )
i=1
l0 −αν
−1
= U0l0 −αν (−V )α
α
e0 α (−e1 ) · . . . · (−er )
−1
(k0 −l0 )+ν
Vν
r
Y
(k0 −l0 )+ν
(−e0 )α e1 · . . . · er
r
ν Y
(−ei )ei
i=1
ki −ν
Uiki −ν
i=1
−1
for each ν ∈ [0, min{α
∗
l0 , k }]. Therefore Z(S) and hence L(S) have the given forms.
We now consider the statements of (3). The assertion on ∆(GP ) follows immediately from (2). Since
∆(GP ) = {r + α − 2}, all sets Uk (GP ) are arithmetical progressions with difference r + α − 2. The
assertion on each ρ2k (GP ) and each ρ(GP ) follow from Proposition 6.2.
30
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
In order to determine L(GP ), let S ∈ B(GP ) be given with all parameters as in (2). First suppose
that l0 ≥ αk ∗ . Then
L(S) = (α−1 (k0 − l0 ) + l0 + k1 + · · · + kr ) − (r + α)k ∗ + {(r + α)k ∗ − (r + α − 2)ν : ν ∈ [0, k ∗ ]}
= (α−1 (k0 − l0 ) + l0 + k1 + · · · + kr ) − (r + α)k ∗ + {2k ∗ + (r + α − 2)λ : λ ∈ [0, k ∗ ]} .
Thus L(S) has the form
L(S) = m + {2k ∗ + (r + α − 2)λ : λ ∈ [0, k ∗ ]}
for some m, k ∗ ∈ N0 . Conversely, for every choice of m, k ∗ ∈ N0 , there is an S ∈ B(GP ) such that L(S)
has the given form.
Now suppose that l0 ≤ αk ∗ − 1 and set m0 = ⌊ lα0 ⌋. Then
L(S) = (α−1 (k0 − l0 ) + l0 + k1 + · · · + kr ) − (r + α)m0 + {(r + α)m0 − (r + α − 2)ν : ν ∈ [0, m0 ]}
= (α−1 (k0 − l0 ) + l0 + k1 + · · · + kr ) − (r + α)m0 + {2m0 + (r + α − 2)λ : λ ∈ [0, m0 ]} ,
and hence L(S) has the form
L(S) = m + {2m0 + (r + α − 2)λ : λ ∈ [0, m0 ]}
for some m ∈ N and m0 ∈ N0 .
Next we verify that, for every k ∈ N, ρ2k+1 (GP ) ≤ kD(GP ) + 1. By [37, Proposition 1.4.2], for all
k ∈ N, ρk (GP ) = sup{sup L : L ∈ L(GP ), k = min L}. Thus we may choose k ∈ N and L ∈ L(GP ) with
min L = 2k + 1. Then, there exist l, m, k ∗ ∈ N0 such that
L = m + {2k ∗ + (r + α − 2)λ : λ ∈ [0, k ∗ ]}
with m = 2l + 1 and 2k + 1 = min L = 2(k ∗ + l) + 1. Now
max L = m + (r + α)k ∗ = 2l + 1 + (r + α)(k − l) = (r + α)k + 1 − (r + α − 2)l ≤ kD(GP ) + 1 ,
and thus ρ2k+1 (GP ) ≤ kD(GP ) + 1.
It remains to verify the assertions on the λlD(GP )+j (GP ). Let l ∈ N0 and j ∈ [0, D(GP ) − 1]. Then
Lemma 3.3 implies λlD(Gp )+j (GP ) ≤ 2l + j, and that equality holds if j ∈ [0, 1]. It remains to verify that
λlD(Gp )+j (GP ) ≥ 2l + j when j ∈ [2, D(GP ) − 1]. Let L ∈ L(GP ) with lD(GP ) + j ∈ L. Then there exist
m, k ∗ ∈ N0 such that
L = m + {2k ∗ + (r + α − 2)λ : λ ∈ [0, k ∗ ]} = m + k ∗ D(GP ) − {(D(GP ) − 2)ν : ν ∈ [0, k ∗ ]} .
Suppose lD(GP ) + j = max L − ν(D(GP ) − 2) = m + k ∗ D(GP ) − ν(D(GP ) − 2) for some ν ∈ [0, k ∗ ]. Then
j ≡ m + 2ν mod D(GP ) and hence m + 2ν ≥ j. This implies
(k ∗ − ν)D(GP ) + j ≤ m + k ∗ D(GP ) − ν(D(GP ) − 2) = lD(GP ) + j ,
and hence l ≥ k ∗ − ν. Therefore we obtain
min L = lD(GP ) + j − (k ∗ − ν)(D(GP ) − 2) = (l − k ∗ + ν)D(GP ) + j + 2(k ∗ − ν) ≥ 2l + j ,
and thus λlD(Gp )+j (GP ) ≥ 2l + j.
Finally we consider the catenary degrees of H and prove the statements given in (4). Using Proposition
6.2 we infer
D(GP ) = r + α = 2 + max ∆(GP ) ≤ c(GP ) = c(H) ≤ ω(H) ≤ t(H) .
Since c(GP ) ≤ cmon (GP ), it remains to show that cmon (GP ) ≤ D(GP ) and that t(H) ≤ D(GP ).
We proceed in two steps. First we verify that cmon (GP ) = max{ceq (GP ), cadj (GP )} ≤ r + α. Since
A(∼B(GP ),eq ) = {(−V, −V ), (V, V ), (Uν , Uν ), (−Uν , −Uν ) : ν ∈ [0, r]} ,
it follows that ceq (GP ) = 0. If Ar+α−2 = {x ∈ Z(GP ) : |x| − (r + α − 2) ∈ L(π(x))}, then Min(Ar+α−2 ) =
{U0α U1 · . . . · Ur }, and hence cadj (GP ) ≤ r + α by Lemma 2.1.
In order to show that t(H) ≤ D(GP ), we must verify the following assertion (see [38, Theorem 3.6]).
MONOIDS OF MODULES AND ARITHMETIC OF DIRECT-SUM DECOMPOSITIONS
31
(A) Let j ∈ N and w, w1 , . . . , wj ∈ A(H) such that w divides the product w1 · . . . · wj but yet w divides
no proper subproduct of w1 · . . . · wj . Then min L(w−1 w1 · . . . · wj ) ≤ D(GP ) − 1.
Proof of (A). We use the transfer homomorphism β : H → B(GP ) as defined in Lemma 3.4. Set
W = β(w) and Wi = β(wi ) for each i ∈ [1, j]. Then j ≤ |W | and W, W1 , . . . , Wj ∈ A(GP ). Clearly
min L(w−1 w1 · . . . · wj ) ≤ max L(W −1 W1 · . . . · Wj ) ≤
|W1 · . . . · Wj | − |W |
.
2
Thus, if |W | = 2, then
|W1 | + |W2 | − |W |
≤ D(GP ) − 1 .
2
It remains to consider the case W ∈ {−V, V }, and by symmetry we may suppose that W = V . If
|W1 | = · · · = |Wj | = 2, then j = |V | and w−1 w1 · . . . · wj ∈ A(H). Suppose there is ν ∈ [1, j], say ν = 1,
such that β(w1 ) ∈ {−V, V }. Since w does not divide a subproduct of w1 · . . . · wj and gcd(V, −V ) = 1, it
follows that β(w1 ) = V . Then L(w−1 w1 · . . . · wj ) = L(W −1 W1 · . . . · Wj ) = L(W2 · . . . · Wj ) and hence
min L(w−1 w1 · . . . · wj ) ≤
min L(w−1 w1 · . . . · wj ) ≤ j − 1 ≤ |V | − 1 = D(GP ) − 1 .
The next corollary again reveals that certain arithmetical phenomena characterize certain algebraic
properties of the class group.
Corollary 6.5. Let H be a Krull monoid as in Theorem 6.4 with class group G and set GP of classes
containing prime divisors. Then r + 1 is the minimum of all s ∈ N having the following property :
(P) There are absolutely irreducible elements w1 , . . . , ws ∈ A(H) such that 2, D(GP ) ∈ L(w1k1 · . . . · wsks )
for some (k1 , . . . , ks ) ∈ Ns0 .
Proof. First we verify that r + 1 satisfies property (P). For i ∈ [0, s], let pi ∈ P ∩ ei and qi ∈ P ∩ (−ei )
and set wi = pi qi . Then w0 , . . . , ws are absolutely irreducible elements and, by Theorem 6.4, it follows
that 2, D(GP ) ∈ L(w0α w1 · . . . · wr ).
Conversely, let s ∈ N, and let w1 , . . . , ws and k1 , . . . , ks be as above. For i ∈ [1, s], we set Wi = β(wi ).
Ps
Since ρ(GP ) = D(GP )/2 and 2, D(GP ) ∈ L(W1k1 · . . . · Wsks ), it follows that
i=1 ki |Wi | = D(GP ),
k1
|W1 | = · · · = |Ws | = 2, Wi = (−gi )gi for i ∈ [1, s], and that S = g1 · . . . · gsks ∈ A(GP ). Now
Theorem 6.4 implies S = (−e0 )α e1 · . . . · er , whence {W1 , . . . , Ws } = {(−e0 )e0 , . . . , (−er )er }. Thus
|{w1 , . . . , ws }| ≥ |{W1 , . . . , Ws }| = r + 1 and so r + 1 is minimal with property (P).
We now begin collecting information in order to study the arithmetic of the Krull monoid presented
in Example 4.21. In spite of the simple geometric structure of GP (the set consists of the vertices of the
unit cube and their negatives), the arithmetic of this Krull monoid is highly complex. We get only very
limited information. Nevertheless, this will be sufficient to give an arithmetical characterization.
Lemma 6.6. Let G be an abelian group and let (en )n≥1 be a family of independent elements each having
infinite order. For r ∈ N, set
−
+
G+
r = {a1 e1 + · · · + ar er : a1 , . . . , ar ∈ [0, 1]}, Gr = −Gr ,
and
−
Gr = G+
r ∪ Gr .
(1) Let s ∈ [2, r], f0 = e1 + · · · + es , and fi = f0 − ei for all i ∈ [1, s]. Then (f1 , . . . , fs ) is independent,
f1 + · · · + fs = (s − 1)f0 , and ∆({f0 , . . . , fs , −f0 , . . . , −fs }) = {2s − 3}.
(2) Let s ∈ [3, r], f0 = e1 + · · · + es , fi = f0 − ei for each i ∈ [1, s − 1], and set fs′ = −es . Then
(f1 , . . . , fs−1 , fs′ ) is independent, f1 + · · · + fs−1 + fs′ = (s − 2)f0 , and
∆({f0 , . . . , fs−1 , fs′ , −f0 , . . . , −fs−1 , −fs′ }) = {2s − 4}.
(3) If s ≤ [1, r − 1], then D(Gr ) ≥ D(Gs ) + D(Gr−s ) − 1. In particular, D(G1 ) = 2 and D(Gr ) > D(Gr−1 )
for r ≥ 2.
32
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
Proof. Since (e1 , . . . , es ) is a basis, there is a matrix As with (f1 , . . . , fs ) = (e1 , . . . , es )As . Since det(As ) 6=
0, it follows that (f1 , . . . , fs ) is independent. By definition, we have f1 +· · ·+fs = (s−1)f0 . The assertion
on the set of distances then follows from Theorem 6.4 and we have proved (1).
We now consider (2). Note that fs′ = fs − f0 . Using (1) we infer that
0 = (f1 + · · · + fs ) − (s − 1)f0 = (f1 + · · · + fs−1 ) + (fs − f0 ) − (s − 2)f0
and hence f1 + · · · + fs−1 + fs′ = (s − 2)f0 . Since (f1 , . . . , fs−1 , −es ) = (e1 , . . . , es )Bs for some matrix Bs
with det(Bs ) = (−1)2s det(As−1 ) 6= 0, it follows that (f1 , . . . , fs−1 , fs′ ) is independent. The assertion on
the set of distances follows from Theorem 6.4.
It is clear that D(G1 ) = 2 and that D(Gr ) > D(Gr−1 ) whenever r ≥ 2. To prove the remaining
statements of (3), suppose that s ∈ [1, r − 1]. After a change of notation, we may suppose that Gr−s ⊂
hes+1 , . . . , er i such that hGr i = hGs i ⊕ hGr−s i. If U = a1 · . . . · ak ∈ A(Gs ) with k = D(Gs ) and
V = b1 · . . . · bl ∈ A(Gr−s ) with l = D(Gr−s ), then W = (a1 + b1 ) · a2 · . . . · ak b2 · . . . · bl ∈ A(Gr ), and
hence D(Gr ) ≥ |W | = k + l − 1 = D(Gs ) + D(Gr−s ) − 1.
In Theorem 6.7 we restrict to class groups of rank r ≥ 3 because when r ≤ 2 we are in the setting of
Theorem 6.4 where we have precise information about arithmetical invariants. For r ∈ N0 , we denote by
Fr the rth Fibonacci number. That is, F0 = 0, F1 = 1, and Fr = Fr−1 + Fr−2 for all r ≥ 2.
Theorem 6.7. Let H be a Krull monoid with free abelian class group G of rank r ≥ 3 and let GP ⊂ G
denote the set of classes containing prime divisors. Suppose that there is a basis (e1 , . . . , er ) of G such
−
+
−
+
that G•P = G+
P ∪ GP , where GP = {ǫ1 e1 + · · · + ǫr er : ǫ1 , . . . , ǫr ∈ [0, 1]} and GP = −GP .
(1) Fr+2 ≤ D(GP ).
(2) c(H) ≤ ω(H) = D(GP ), ρ(H) = D(GP )/2, and ρ2k (H) = kD(GP ) for each k ∈ N.
(3) [1, 2r − 3] ⊂ ∆∗ (H) ⊂ ∆(H) ⊂ [1, c(H) − 2].
Proof. See [5] for the proof of (1). (2) follows from Proposition 6.2.
Note that for every s ∈ [2, r] we have 2s − 3 ∈ ∆∗ (H) and, by Lemma 6.6, for all s ∈ [3, r] we have
2s − 4 ∈ ∆∗ (H). This implies that the interval [1, 2r − 3] is contained in ∆∗ (H), giving (3).
The third class of Krull monoids studied in this subsection are Krull monoids with finite cyclic class
group having prime divisors in each class. Thus Theorem 6.8 describes the arithmetic of the monoids
constructed in Theorem 5.5. Holomorphy rings in global fields are Krull monoids with finite class group
and prime divisors in all classes. For this reason this class of Krull monoids has received a great deal of
attention.
Theorem 6.8. Let H be a Krull monoid with finite cyclic class group G of order |G| = n ≥ 3, and
suppose that every class contains a prime divisor. Then:
(1) c(H) = ω(H) = D(G) = n and ∆(H) = [1, n − 2].
(2) For every k ∈ N the set Uk (H) is a finite interval, whence Uk (H) = [λk (H), ρk (H)]. Moreover, for
all l ∈ N0 with ln + j ≥ 1,
(
2l + j
for j ∈ [0, 1]
ρ2k+j (H) = kn + j for j ∈ [0, 1] and λln+j (H) =
2l + 2
for j ∈ [2, n − 1].
(3)
max ∆∗ (H) = n − 2
and
jnk
max ∆∗ (H) \ {n − 2} =
− 1.
2
Proof. The proof of (1) can be found in [37, Theorem 6.7.1] and the proof of (3) can be found in [37,
Theorem 6.8.12]. For (2) see [34, Corollary 5.3.2].
MONOIDS OF MODULES AND ARITHMETIC OF DIRECT-SUM DECOMPOSITIONS
33
Much recent research is devoted to the arithmetic of Krull monoids discussed in Theorem 6.8. We
briefly address some open questions. Let H be as above and suppose that n ≥ 5. The precise values of
t(H) and of cmon (H) are unknown. It is easy to check that D(G) = n < t(H) (in contrast to what we have
in Theorem 6.4). For recent results on lower and upper bounds of the tame degree, see [33]. We remark
that there is a standing conjecture that the monotone catenary degree is that n = c(H) = cmon (H) (this
coincides what we have in Theorem 6.4; see [41]). For recent progress on ∆∗ (H) we refer to [58].
Having at least a partial description of the arithmetic of the three monoids described in Theorems 6.4,
6.7, and 6.8, we now work to show that except for in a small number of exceptions, these monoids have
vastly different arithmetic. After some preliminary work this distinction is made clear in Corollary 6.10.
Lemma 6.9. Let G be an abelian group with finite total rank and let G0 ⊂ G be a subset with G0 = −G0 .
Suppose that L(G0 ) = L(Cn ) for some n ≥ 5. Then there exists an absolutely irreducible element
U ∈ A(G0 ) with |U | = D(G0 ).
Proof. First observe that D(G0 ) = ρ2 (G0 ) = ρ2 (Cn ) = D(Cn ) = n and, by [37, Theorem 3.4.2], A(G0 ) is
finite, say A(G0 ) = {U1 , −U1 , . . . , Uq , −Uq }. If g ∈ Cn with ord(g) = n, then for all k ∈ N we have
Lk = {2k + ν(n − 2) : ν ∈ [0, k]} = L g nk (−g)nk ∈ L(Cn ) = L(G0 ) .
(q)
Since ρ(Lk ) = ρ(G0 ) = D(G0 )/2, there exists, for every k ∈ N, a tuple (k1 , . . . , kq ) ∈ N0 such that
k1 + · · · + kq = k and
Lk = L (−U1 )k1 U1k1 · . . . · (−Uq )kq Uqkq .
Therefore there exists λ ∈ [1, q] such that L (−Uλ )k Uλk = Lk for every k ∈ N. Set U = Uλ and note
that for every V ∈ A(G0 ) with V | (−U )k U k for some k ∈ N, it follows that |V | ∈ {2, n}. After changing
notation if necessary, we may suppose that there is no V ∈ A(G0 ) such that |V | = n, supp(V ) ( supp(U ),
and V | U k (−U )k for some k ∈ N.
In order to show that U is absolutely irreducible, it remains to verify that the torsion-free rank
of hsupp(U )i is | supp(U )| − 1. Assume to the contrary that there exist t ∈ [2, | supp(U )| − 1] and
g1 , . . . , gt ∈ supp(U ) which are linearly dependent. Then there are s ∈ [1, t], m1 , . . . , ms ∈ N, and
ms+1 , . . . , mt ∈ −N such that
m1 g1 + · · · + ms gs + (−ms+1 )(−gs+1 ) + · · · + (−mt )(−gt ) = 0 .
Then
V = g1m1 · . . . · gsms (−gs+1 )−ms+1 · . . . · (−gt )−mt ∈ B(G0 ) .
Without restriction we may suppose that the above equation is minimal and that V ∈ A(G0 ). Since
V | U k (−U )k for some k ∈ N and |V | > 2, we obtain a contradiction to the minimality of supp(U ).
The following corollary highlights that the observed arithmetical phenomena in our case studies —
Theorems 6.4, 6.7, and 6.8 — are characteristic for the respective Krull monoids. In particular, this
illustrates that the structure of direct-sum decompositions over the one-dimensional Noetherian local rings
with finite representation type studied in Section 4 can be quite different from the structure of direct-sum
decompositions over the two-dimensional Noetherian local Krull domains with finite representation type
studied in Section 5. As characterizing tools we use the system of sets of lengths along with the behavior
of absolutely irreducible elements.
Corollary 6.10. For i ∈ [1, 3], let Hi and Hi′ be Krull monoids with class groups Gi and G′i . Further
suppose that
• G1 and G′1 are finitely generated and torsion-free of rank r1 and r1′ with sets of classes containing
prime divisors as in Theorem 6.4 (with parameters α, α′ ∈ N such that α + r1 ≥ α′ + r1′ > 2).
• G2 and G′2 are finitely generated and torsion-free of rank r2 ≥ r2′ ≥ 3 with sets of classes containing
prime divisors as in Theorem 6.7.
34
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
• G3 and G′3 are finite cyclic of order |G3 | ≥ |G′3 | ≥ 5 such that every class contains a prime divisor.
Then:
(1) L(H1 ) = L(H1′ ) if and only if r1 + α = r1′ + α′ . If this holds, then the arithmetic behavior of the
absolutely irreducible elements of H1 and H1′ coincide in the sense of Corollary 6.5 if and only if
r1 = r1′ .
(2) L(H2 ) = L(H2′ ) if and only if r2 = r2′ .
(3) L(H3 ) = L(H3′ ) if and only if |G3 | = |G′3 |.
(4) L(H1 ) 6= L(H2 ) and L(H1 ) 6= L(H3 ).
(5) For i ∈ [2, 3], let si denote the maximal number of absolutely irreducible elements u1 , . . . , usi ∈ Hi
such that 2 ∈ L(u1 · . . . · usi ). Then either L(H2 ) 6= L(H3 ) or s2 6= s3 .
Proof. The if and only if statement in (1) follows immediately from Theorem 6.4. Suppose that L(H1 ) =
L(H1′ ). Then the assertion in (1) on the arithmetic behavior of absolutely irreducible elements follows
from Corollary 6.5.
To prove (2), first note that one implication is clear, both for H2 and H3 . Suppose that L(H2 ) = L(H2′ ),
and let GP ⊂ G2 and G′P ⊂ G′2 denote the set of classes containing prime divisors. Theorem 6.7 implies
that
D(GP ) = ρ2 (H) = ρ2 (H ′ ) = D(G′P ) ,
and thus Lemma 6.6 implies r2 = r2′ . Now consider (3). If L(H3 ) = L(H3′ ), then Theorem 6.8 implies
that
|G3 | − 2 = max ∆(H3 ) = max ∆(H3′ ) = |G′3 | − 2 .
For (4), note that L(H1 ) is distinct from both L(H2 ) and L(H3 ) since |∆(H1 )| = 1, |∆(H2 )| > 1, and
|∆(H3 )| > 1.
For (5) we assume that L(H2 ) = L(H3 ) and let GP ⊂ G2 denote the set of classes containing prime
divisors. Theorems 6.7 and 6.8 imply that
D(GP ) = ρ2 (H2 ) = ρ2 (H3 ) = |G3 | .
By Proposition 6.3 we obtain that D(GP ) = s2 . Now assume to the contrary that s2 = s3 . If |G3 | = n,
then there are absolutely irreducible elements u1 , . . . , un and atoms v1 , v2 ∈ A(H3 ) such that v1 v2 =
u1 · . . . · un . Without restriction, we suppose H3 is reduced and we consider a divisor theory H ֒→ F (P).
Since a minimal zero-sum sequence of length n over G3 consists of one element of order n repeated n times,
the factorization of the atoms v1 , v2 , u1 , . . . , un in F (P) must have the following form: v1 = p1 · . . . · pn ,
v2 = q1 · . . . · qn , and ui = pi qi for all i ∈ [1, n], where p1 , . . . , pn , q1 , . . . , qn ∈ P, [p1 ] = · · · = [pn ] ∈ G3 ,
and [q1 ] = · · · = [qn ] = [−p1 ]. But [37, Proposition 7.1.5] implies that the elements u1 , . . . , un are not
absolutely irreducible, a contradiction.
Remark 6.11. Let H2 and H3 be as in Corollary 6.10. We set n = |G3 |, r = r2 , and let GP,r ⊂ G2
denote the set of classes containing prime divisors. Assume that L(H2 ) = L(H3 ). Then
Fr+2 ≤ D(GP,r ) = ρ2 (H2 ) = ρ2 (H3 ) = n .
That is, the orders of the cyclic groups for which L(H2 ) = L(H3 ) grow faster than the sequence of
Fibonacci numbers. We conjecture that L(H2 ) and L(H3 ) are always distinct but have not further investigated this (rather delicate combinatorial) problem which would require a more detailed investigation of
D(GP,r ).
Now suppose that H is a Krull monoid with class group G such that every class contains a prime
divisor. If L(H) = L(H3 ), then following Theorem 6.8, one can show that G is isomorphic to the finite
cyclic group G3 (see [34, Corollary 5.3.3]). Therefore sets of lengths characterize Krull monoids with
finite cyclic class group having the property that every class contains a prime divisor.
MONOIDS OF MODULES AND ARITHMETIC OF DIRECT-SUM DECOMPOSITIONS
35
6.3. Small sets GP of classes containing prime divisors and limits of arithmetical characterizations. In this final subsection we study the arithmetic of Krull monoids having small sets of classes
containing prime divisors. This study pertains to the monoids of Theorem 4.12, Example 4.19, Example
4.20, and Theorem 5.5. The most striking phenomenon here is that these systems of sets of lengths
are additively closed (see Proposition 6.14). As a consequence, if L(H) is such a system and H ′ is a
monoid with L(H ′ ) ⊂ L(H), then L(H × H ′ ) = L(H) (see Example 4.23, Example 4.24, and Corollary
6.15). These phenomena are in strong contrast to the results in the previous subsection, and they show
up natural limits for obtaining arithmetical characterization results. Recall that, for l ∈ N0 and d ∈ N,
Pl (d) = {0, d, . . . , ld}.
Proposition 6.12. Let H be a Krull monoid with infinite cyclic class group G and suppose that
GP = {−2e, −e, 0, e, 2e} ⊂ G = hei
is the set of classes containing prime divisors. Then there is a transfer homomorphism θ : H → B(C3 ),
and hence
L(H) = L(C3 ) = L(C2 ⊕ C2 ) = y + 2k + Pk (1) : y, k ∈ N0 .
Moreover, L(H) coincides with the system of sets of lengths of the Krull monoid studied in Theorem 6.4
with parameters r = 2 and α = 1.
Proof. By Lemma 3.4 there is a transfer homomorphism β : H → B(GP ). Since the composition of two
transfer homomorphisms is a transfer homomorphism, it is sufficient to show that there is a transfer
homomorphism θ′ : B(GP ) → B(C3 ). Write C3 = {0, g, −g}. Since B(GP ) = F ({0}) × B(G•P ) and
B(C3 ) = F ({0}) × B({−g, g}), it suffices to show that there is a transfer homomorphism θ : B(G•P ) →
B({−g, g}). In this case, L(H) = L(GP ) = L(C3 ). Moreover, L(C3 ) = L(C2 ⊕ C2 ) has the form given
in [37, Theorem 7.3.2] and this coincides with the system of sets of lengths in Theorem 6.4, provided
(r, α) = (2, 1).
Note that A(G•P ) = {V, −V, U1 , U2 }, where V = e2 (−2e), U1 = (−e)e, and U2 = (−2e)(2e), and
A({−g, g}) = {V , −V , U }, where V = g 3 and U = (−g)g. Then there is a monoid epimorphism
e = θ(−2e)
e
e
e
θe: F (G•P ) → F ({−g, g}) satisfying θ(e)
= g and θ(−e)
= θ(2e)
= −g. If
′
′
A = ek1 (−e)k1 (2e)k2 (−2e)k2 ∈ F (G•P ) with
k1 , k1′ , k2 , k2′ ∈ N0 ,
then A ∈ B(G•P ) if and only if k1 − k1′ + 2(k2 − k2′ ) = 0. If this holds, then k1 + k2′ − (k1′ + k2 ) ≡ 0 mod 3
and hence
′
′
e
θ(A)
= g k1 +k2 (−g)k1 +k2 ∈ B({−g, g}) .
Thus θ = θe |B(G•P ) : B(G•P ) → B({−g, g}) is a monoid epimorphism satisfying θ(V ) = V , θ(−V ) = −V ,
θ(U1 ) = θ(U2 ) = U , and θ−1 (1) = {1} = B(G•P )× .
Thus in order to show that θ is a transfer homomorphism, it remains to verify Property (T2). Let
A ∈ B(G•P ) be as above and suppose that
eC
e
θ(A) = B
′
e C
e ∈ B({−g, g}) and B
e = g m (−g)m such that m ∈ [0, k1 + k ′ ], m′ ∈ [0, k ′ + k2 ] and m ≡ m′
with B,
2
1
e and θ(C) = C.
e Clearly it is
mod 3. Our goal is to find B, C ∈ B(G•P ) such that A = BC, θ(B) = B,
•
e that is, to find parameters
sufficient to find B ∈ B(G ) with B | A and θ(B) = B,
P
m1 ∈ [0, k1 ], m′1 ∈ [0, k1′ ], m2 ∈ [0, k2 ], and m′2 ∈ [0, k2′ ]
such that
(C1)
m1 + m′2 = m, m′1 + m2 = m′ ,
and m1 − m′1 + 2(m2 − m′2 ) = 0 .
e If |B|
e = |A|, then k1 + k ′ + k2 + k ′ = |A| = |B|
e = m + m′ and
To do so we proceed by induction on |B|.
1
2
′
′
′
hence m = k1 + k1 and m = k1 + k2 . Thus we set
m1 = k1 , m′1 = k1′ , m2 = k2 , and m′2 = k2′ ,
36
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
and the assertion is satisfied with B = A. Suppose now that the quadruple (m1 , m′1 , m2 , m′2 ) satisfies
e by an atom of B({−g, g}) (if possible) shows that we
(C1) with respect to the pair (m, m′ ). Dividing B
must verify that there are solutions to (C1) with respect to each of the pairs
(m − 1, m′ − 1), (m − 3, m′ ), and (m, m′ − 3)
(2)
in N0 .
One checks respectively that at least one of the following quadruples satisfy (C1).
• (m1 − 1, m′1 − 1, m2 , m′2 ) or (m1 , m′1 , m2 − 1, m′2 − 1)
• (m1 − 2, m′1 , m2 , m′2 − 1) or (m1 − 3, m′1 − 1, m2 + 1, m′2 )
• (m1 , m′1 − 2, m2 − 1, m′2 ) or (m1 − 1, m′1 − 3, m2 , m′2 + 1)
Now the assertion follows by the induction hypothesis.
Proposition 6.13. Let H be a Krull monoid with free abelian class group G of rank 2. Let (e1 , e2 ) be a
basis of G and suppose that
GP = {0, e1 , e2 , 2e2 , e1 + 2e2 , −e1 , −e2 , −2e2, −e1 − 2e2 }
is the set of classes containing prime divisors. Then there is a transfer homomorphism θ : H → B(C4 )
and hence
L(H) = L(C4 ) = y + k + 1 + Pk (1) : y, k ∈ N0 ∪ y + 2k + Pk (2) : y, k ∈ N0 ⊃ L(C3 ) .
Proof. As in Proposition 6.12, it suffices to show that there is a transfer homomorphism θ : B(G•P ) →
B(C4• ). Then L(H) = L(GP ) = L(C4 ) and L(C4 ) has the form given in [37, Theorem 7.3.2]. Proposition
6.12 shows that L(C3 ) ⊂ L(C4 ).
We note that A(G•P ) = {W, −W, V1 , −V1 , V2 , −V2 , U1 , U2 , U3 , U4 }, where
W = e1 e2 e2 (−e1 − 2e2 ), V1 = e1 (2e2 )(−e1 − 2e2 ), V2 = e2 e2 (−2e2 ),
U1 = (−e1 )e1 , U2 = (−e2 )e2 ), U3 = (−e1 − 2e2 )(e1 + 2e2 ), and U4 = (−2e2 )(2e2 ) .
We set C4 = {0, g, 2g, −g} and observe that A(C4• ) = {W , −W , V , −V , U1 , U2 }, where
W = g 4 , V = g 2 (2g), U1 = (−g)g,
and U2 = (2g)(2g) .
There is a monoid epimorphism θe: F (G•P ) → F (C4• ) satisfying
e 1 ) = θ(e
e 2 ) = θ(−e
e
e
e
e
e
e
θ(e
1 − 2e2 ) = g, θ(−e1 ) = θ(−e2 ) = θ(e1 + 2e2 ) = −g, and θ(2e2 ) = θ(−2e2 ) = 2g .
If
′
′
′
′
A = ek11 (−e1 )k1 ek22 (−e2 )k2 (2e2 )k3 (−2e2 )k3 (e1 + 2e2 )k4 (−e1 − 2e2 )k4 ∈ F (G•P ) ,
with k1 , k1′ , . . . , k4 , k4′ ∈ N0 , then A ∈ B(G•P ) if and only if
k1 − k1′ + k4 − k4′ = 0
and k2 − k2′ + 2k3 − 2k3′ + 2k4 − 2k4′ = 0 .
If this holds, then
k1 − k1′ + k2 − k2′ − (k4 − k4′ ) + 2k3 + 2k3′ ≡ 0 mod 4
and hence
′
′
′
′
e
θ(A)
= g k1 +k2 +k4 (−g)k1 +k2 +k4 (2g)k3 +k3 ∈ B(C4• ) .
Thus θ = θe |B(G•P ) : B(G•P ) → B(C4• ) is a monoid epimorphism satisfying θ(W ) = W , θ(−W ) = −W ,
θ(V1 ) = θ(V2 ) = V , θ(−V1 ) = θ(−V2 ) = −V , θ(U1 ) = θ(U2 ) = θ(U3 ) = U1 , θ(U4 ) = U2 , and θ−1 (1) =
{1} = B(G•P )× .
Thus in order to show that θ is a transfer homomorphism, it remains to verify Property (T2). Let
A ∈ B(G•P ) be as above and suppose that
eC
e
θ(A) = B
′
′′
e C
e ∈ B(C • ) and B
e = g m (−g)m (2g)m such that
with B,
4
m ∈ [0, k1 + k2 + k4′ ], m′ ∈ [0, k1′ + k2′ + k4 ], m′′ ∈ [0, k3 + k3′ ], and m − m′ + 2m′′ ≡ 0
mod 4 .
MONOIDS OF MODULES AND ARITHMETIC OF DIRECT-SUM DECOMPOSITIONS
37
e and θ(C) = C.
e It will suffice to find
Our goal is to find B, C ∈ B(G•P ) such that A = BC, θ(B) = B,
•
e
B ∈ B(GP ) with B | A and θ(B) = B. Thus we must find parameters
mν ∈ [0, kν ] and m′ν ∈ [0, kν′ ] for ν ∈ [1, 4]
such that
m1 + m2 + m′4 = m, m′1 + m′2 + m4 = m′ , m3 + m′3 = m′′ ,
m1 − m′1 + m4 − m′4 = 0 and m2 − m′2 + 2m3 − 2m′3 + 2m4 − 2m′4 = 0 .
(C2)
e = m + m′ + m′′ . If |B|
e = |A|, then we set mν = kν and m′ = k ′ for all
We proceed by induction on |B|
ν
ν
ν ∈ [1, 4], and the assertion is satisfied with B = A. Suppose now that the octuplet (m1 , m′1 , . . . , m4 , m′4 )
e by an element of A(C • ) (if possible)
satisfies (C2) with respect to the triple (m, m′ , m′′ ). Dividing B
4
shows that we must verify that there are solutions to (C2) with respect to each of the triples
(m−1, m′ −1, m′′ ), (m−2, m′ , m′′ −1), (m, m′ −2, m′′ −1), (m, m′ , m′′ −2), (m−4, m′ , m′′ ), and (m, m′ −4, m′′ )
(8)
provided that they lie in N0 . As in proof of the previous proposition, one finds the required solutions
and hence the assertion follows by the induction hypothesis.
Let L be a family of subsets of Z. We say that L is additively closed if the sumset L + L′ ∈ L for all
L, L′ ∈ L.
Proposition 6.14. Let G be a finite cyclic group. Then L(G) is additively closed if and only if |G| ≤ 4.
Proof. We suppose that |G| = n and distinguish four cases.
First assume that n ≤ 2. Since B(G) is factorial, it follows that L(G) = {{m} : m ∈ N0 } which is
obviously additively closed.
Next assume that n = 3. By Proposition 6.12 we have L(C3 ) = y + 2k + Pk (1) : y, k ∈ N0 . If
y1 , y2 , k1 , k2 ∈ N0 . Then
y1 + 2k1 + Pk1 (1) + y2 + 2k2 + Pk2 (1) = (y1 + y2 ) + 2(k1 + k2 ) + Pk1 +k2 (1) ∈ L(C3 ) ,
and hence L(C3 ) is additively closed.
Now assume that n = 4. By Proposition 6.13 we have
L(C4 ) = y + k + 1 + Pk (1) : y, k ∈ N0 ∪ y + 2k + Pk (2) : y, k ∈ N0 .
Clearly, the sumset of two sets of the first form is of the first form again, and the sumset of two sets of
the second form again the second form. Thus it remains to consider the sumset L1 + L2 where L1 has
the first form, L2 has the second form, and both L1 and L2 have more than one element. If y1 , y2 ∈ N0
and k1 , k2 ∈ N, then
y1 + k1 + 1 + Pk1 (1) + y2 + 2k2 + Pk2 (2) = (y1 + y2 ) + (k1 + 2k2 ) + 1 + Pk1 +2k2 (1) ∈ L(C4 ) .
Finally, assume that n ≥ 5 and assume to the contrary that L(G) is additively closed. Let d ∈ [1, n−2].
Then {2, d + 2} ∈ L(G) by [37, Theorem 6.6.2], and hence the k-fold sumset
{2, d + 2} + · · · + {2, d + 2} = 2k + Pk (d)
lies in L(G) for all k ∈ N. Then [37, Corollary 4.3.16] implies that n − 3 divides some d ∈ ∆∗ (G). By
Theorem 6.8 we have
jnk
max ∆∗ (G) = n − 2 and max ∆∗ (G) \ {n − 2} =
−1,
2
a contradiction to n ≥ 5.
Corollary 6.15.
(1) Let H be an atomic monoid such that L(H) is additively closed, and let H ′ be an atomic monoid
with L(H ′ ) ⊂ L(H). Then L(H × H ′ ) = L(H).
38
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
(2) Let H be an atomic monoid with L(H) = L(Cn ) for n ∈ [3, 4]. For k ∈ N and i ∈ [1, k], let Hi be an
atomic monoid with L(Hi ) ⊂ L(Cn ). Then L(H × H1 × · · · × Hk ) = L(Cn ).
Proof. Since L(H × H ′ ) = {L + L′ : L ∈ L(H), L′ ∈ L(H ′ )}, (1) follows.
For (2), we set H ′ = H1 × · · ·× Hk . Since L(Cn ) is additively closed by Proposition 6.14, it follows that
L(H ′ ) = {L1 + · · · + Lk : Li ∈ L(Hi ), i ∈ [1, k]} ⊂ L(Cn ). Finally 1. implies that L(H × H ′ ) = L(H).
We conclude this manuscript by suggesting a rich program for further study. Any progress in these
directions will lead to a better understanding of direct-sum decompositions of classes of modules where
each module has a semilocal endomorphism ring. Moreover, this program could stimulate new studies
in combinatorial Factorization Theory where much of the focus has been on Krull monoids having finite
class group.
Program for Further Study.
A. Module-Theoretic Aspect. Let R be a ring and let C be a class of right R-modules which is closed
under finite direct sums, direct summands, and isomorphisms, and such that the endomorphism ring
EndR (M ) is semilocal for each module M in C (such classes of modules are presented in a systematical
way in [24]). Then V(C), the monoid of isomorphism classes of modules in C is a reduced Krull monoid
with class group G and set GP ⊂ G of classes containing prime divisors.
Since the long-term goal — to determine the characteristic of V(C) — is out of reach in most cases, the
focus of study should be on those properties of GP which have most crucial influence on the arithmetic
of direct-sum decompositions. In particular,
• Is GP finite or infinite ?
• Is GP well-structured in the sense of Proposition 6.2 ?
B. Arithmetical Aspect of Direct-Sum Decompositions. Let H be a Krull monoid with finitely
generated class group G and let GP ⊂ G denote the set of classes containing prime divisors.
1. Finiteness conditions.
(a) Characterize the finiteness of arithmetical invariants (as introduced in Section 2) and the validity
of structural finiteness results (as given in Proposition 6.2, 2.(a) and 2.(b)).
For infinite cyclic groups much work in this direction can be found done in [36].
(b) If GP contains an infinite group, then every finite subset L ⊂ N≥2 occurs as a set of lengths in
H (see Proposition 6.2) and hence ∆(H) = N, and Uk (H) = N≥2 for all k ≥ 2. Weaken the
assumption on GP to obtain similar results.
A weak condition on GP enforcing that ∆(H) = N can be found in [45].
2. Upper bounds and precise formulas. Suppose that G is torsion-free, say GP ⊂ G = Z(q) ⊂ (R(q) , | · |),
where | · | : R(q) → R≥0 is a Euclidean norm.
(a) If GP ⊂ {x ∈ R : |x| ≤ M } for some M ∈ N, then derive upper bounds for the arithmetical
invariants in terms of M .
(b) If GP has a simple geometric structure (e.g., the set of vertices in a cube; see Examples 4.21 and
4.22), derive precise formulas for the arithmetical invariants, starting with the Davenport constant.
A first result in this direction can be found in [5].
(c) Determine the extent to which the arithmetic of a Krull monoid with GP as in (b) is characteristic
for GP . In particular, determine how this compares with the arithmetic of a Krull monoid H ′ where
G′P has the same geometric structure as GP with different parameters and how this compares with
the arithmetic of a Krull monoid having finite class group and prime divisors in all classes.
MONOIDS OF MODULES AND ARITHMETIC OF DIRECT-SUM DECOMPOSITIONS
39
References
[1] M. Arnavut, M. Luckas, and S. Wiegand, Indecomposable modules over one-dimensional Noetherian rings, J. Pure
Appl. Algebra 208 (2007), 739 – 760.
[2] G. Azumaya, Corrections and supplementaries to my paper concerning Krull-Remak-Schmidt’s theorem, Nagoya Math.
J. 1 (1950), 117 – 124.
[3] N.R. Baeth, A Krull-Schmidt theorem for one-dimensional rings with finite Cohen-Macaulay type, J. Pure Appl.
Algebra 208 (2007), 923 – 940.
[4]
, Direct sum decompositions over two-dimensional local domains, Commun. Algebra 37 (2009), 1469 – 1480.
[5] N.R. Baeth, A. Geroldinger, D.J. Grynkiewicz, and D. Smertnig, A semigroup-theoretical view of direct-sum decompositions and associated combinatorial problems, manuscript.
[6] N.R. Baeth and M. Luckas, Bounds for indecomposable torsion-free modules, J. Pure Appl. Algebra 213 (2009), 1254
– 1263.
, Monoids of torsion-free modules over rings with finite representation type, J. Commut. Algebra 3 (2011), 439
[7]
– 458.
[8] N.R. Baeth and S. Saccon, Monoids of modules over rings of infinite Cohen-Macaulay type, J. Commut. Algebra 4
(2012), 297 – 326.
[9] N.R. Baeth and R. Wiegand, Factorization theory and decomposition of modules, Am. Math. Mon. 120 (2013), 3 – 34.
[10] P. Baginski, A. Geroldinger, D.J. Grynkiewicz, and A. Philipp, Products of two atoms in Krull monoids and arithmetical characterizations of class groups, Eur. J. Comb. 34 (2013), 1244 – 1268.
[11] H. Bass, Torsion free and projective modules, Trans. Am. Math. Soc. 102 (1962), 319 – 327.
[12] V. Blanco, P. A. Garcı́a-Sánchez, and A. Geroldinger, Semigroup-theoretical characterizations of arithmetical invariants
with applications to numerical monoids and Krull monoids, Illinois J. Math. 55 (2011), 1385 – 1414.
[13] N. Bourbaki, Commutative Algebra, Springer, 1988.
[14] E. Brieskorn, Rationale Singularitäten komplexer Flächen, Invent. Math. 4 (1968), 336 – 358.
[15] R.-O. Buchweitz, G. M. Greuel, and F.-O. Schreyer, Cohen-Macaulay modules on hypersurface singularities II, Invent.
Math. 88 (1987), 165 – 182.
[16] N. Çimen, One-dimensional rings of finite Cohen-Macaulay type, J. Pure Appl. Algebra 132 (1998), 275 – 308.
[17] N. Çimen, R. Wiegand, and S. Wiegand, One-dimensional rings of finite representation type, Abelian Groups and
Modules (A. Facchini and C. Menini, eds.), vol. 343, Kluwer Academic, 1995, pp. 95 – 121.
[18] J. Coykendall and W.W. Smith, On unique factorization domains, J. Algebra 332 (2011), 62 – 70.
[19] L. Diracca, On a generalization of the exchange property to modules with semilocal endomorphism rings, J. Algebra
313 (2007), 972 – 987.
[20] E.G. Evans, Krull-Schmidt and cancellation over local rings, Pacific J. Math. 46 (1973), 115 – 121.
[21] E.G. Evans and P. Griffith, Syzygies, London Math. Soc. Lecture Note Ser., vol. 106, Cambridge University Press,
1985.
[22] A. Facchini, Direct sum decomposition of modules, semilocal endomorphism rings, and Krull monoids, J. Algebra 256
(2002), 280 – 307.
, The Krull-Schmidt Theorem, Handbook of Algebra, vol. 3, North-Holland, Amsterdam, 2003, pp. 357 – 397.
[23]
, Geometric regularity of direct-sum decompositions in some classes of modules, Fundam. Prikl. Mat. 10 (2004),
[24]
231 – 244.
[25]
, Krull monoids and their application in module theory, Algebras, Rings and their Representations (A. Facchini,
K. Fuller, C. M. Ringel, and C. Santa-Clara, eds.), World Scientific, 2006, pp. 53 – 71.
[26]
, Direct-sum decompositions of modules with semilocal endomorphism rings, Bull. Math. Sci. 2 (2012), 225 –
279.
[27] A. Facchini and F. Halter-Koch, Projective modules and divisor homomorphisms, J. Algebra Appl. 2 (2003), 435 –
449.
[28] A. Facchini, W. Hassler, L. Klingler, and R. Wiegand, Direct-sum decompositions over one-dimensional CohenMacaulay local rings, Multiplicative Ideal Theory in Commutative Algebra (J.W. Brewer, S. Glaz, W. Heinzer, and
B. Olberding, eds.), Springer, 2006, pp. 153 – 168.
[29] A. Facchini and R. Wiegand, Direct-sum decomposition of modules with semilocal endomorphism rings, J. Algebra
274 (2004), 689 – 707.
[30] A. Frankhild, S. Sather-Wagstaff, and R. Wiegand, Ascent of module structures, vanishing of Ext, and extended
modules, Michigan Math. J. 57 (2008), 321 – 337.
[31] W. Gao and A. Geroldinger, Zero-sum problems in finite abelian groups : a survey, Expo. Math. 24 (2006), 337 – 369.
[32]
, On products of k atoms, Monatsh. Math. 156 (2009), 141 – 157.
[33] W. Gao, A. Geroldinger, and W. A. Schmid, Local and global tameness in Krull monoids, Commun. Algebra, to appear.
[34] A. Geroldinger, Additive group theory and non-unique factorizations, Combinatorial Number Theory and Additive
Group Theory (A. Geroldinger and I. Ruzsa, eds.), Advanced Courses in Mathematics CRM Barcelona, Birkhäuser,
2009, pp. 1 – 86.
40
NICHOLAS R. BAETH AND ALFRED GEROLDINGER
[35] A. Geroldinger and R. Göbel, Half-factorial subsets in infinite abelian groups, Houston J. Math. 29 (2003), 841 – 858.
[36] A. Geroldinger, D.J. Grynkiewicz, G.J. Schaeffer, and W.A. Schmid, On the arithmetic of Krull monoids with infinite
cyclic class group, J. Pure Appl. Algebra 214 (2010), 2219 – 2250.
[37] A. Geroldinger and F. Halter-Koch, Non-Unique Factorizations. Algebraic, Combinatorial and Analytic Theory, Pure
and Applied Mathematics, vol. 278, Chapman & Hall/CRC, 2006.
[38] A. Geroldinger and W. Hassler, Local tameness of v-noetherian monoids, J. Pure Appl. Algebra 212 (2008), 1509 –
1524.
[39] A. Geroldinger and F. Kainrath, On the arithmetic of tame monoids with applications to Krull monoids and Mori
domains, J. Pure Appl. Algebra 214 (2010), 2199 – 2218.
[40] A. Geroldinger and I. Ruzsa, Combinatorial Number Theory and Additive Group Theory, Advanced Courses in Mathematics - CRM Barcelona, Birkhäuser, 2009.
[41] A. Geroldinger and P. Yuan, The monotone catenary degree of Krull monoids, Result. Math. 63 (2013), 999 – 1031.
[42] E. Green and I. Reiner, Integral representations and diagrams, Michigan Math. J. 25 (1978), 53 – 84.
[43] D.J. Grynkiewicz, Structural Additive Theory, Developments in Mathematics, Springer, 2013.
[44] F. Halter-Koch, Ideal Systems. An Introduction to Multiplicative Ideal Theory, Marcel Dekker, 1998.
[45] W. Hassler, Factorization properties of Krull monoids with infinite class group, Colloq. Math. 92 (2002), 229 – 242.
[46] W. Hassler, R. Karr, L. Klingler, and R. Wiegand, Big indecomposable modules and direct-sum relations, Ill. J. Math.
51 (2007), 99 – 122.
[47] W. Hassler and R. Wiegand, Extended modules, J. Commut. Algebra 1 (2009), 481 – 506.
[48] R. Heitmann, Characterization of completions of unique factorization domains, Trans. Am. Math. Soc. 337 (1993),
379 – 387.
[49] D. Herbera and P. Přı́hoda, Big projective modules over noetherian semilocal rings, J. Reine Angew. Math. 648 (2010),
111 – 148.
[50] F. Kainrath, Factorization in Krull monoids with infinite class group, Colloq. Math. 80 (1999), 23 – 30.
[51] R. Karr and R. Wiegand, Direct-sum behavior of modules over one-dimensional rings, Commutative Algebra: Noetherian and Non-Noetherian Perspectives (M. Fontana, S.-E. Kabbaj, B. Olberding, and I. Swanson, eds.), Springer,
2011, pp. 251 – 275.
[52] H. Knörrer, Cohen-Macaulay modules on hypersurface singularities I, Invent. Math. 88 (1987), 153 – 164.
[53] W. Krull, Über verallgemeinerte endliche Abelsche Gruppen, Math. Z. 23 (1925), 161 – 196.
[54] C. Lech, A method for constructing bad Noetherian local rings, Algebra, Algebraic Topology and their Interactions,
Lecture Notes in Mathematics, vol. 1183, Springer, 1986, pp. 241 – 247.
[55] C.R. Leedham-Green, The class group of Dedekind domains, Trans. Am. Math. Soc. 163 (1972), 493 – 500.
[56] G.J. Leuschke and R. Wiegand, Cohen-Macaulay representations, Mathematical Surveys and Monographs, vol. 181,
American Mathematical Society, Providence, RI, 2012.
[57] L. Levy and C. Odenthal, Krull-Schmidt theorems in dimension 1, Trans. Am. Math. Soc. 348 (1996), 3391 – 3455.
[58] A. Plagne and W.A. Schmid, On congruence half-factorial Krull monoids with cyclic class group, submitted.
[59] R. Remak, Über die Zerlegung der endlichen Gruppen in direkte unzerlegbare Faktoren, J. Reine Angew. Math. 139
(1911), 293 – 308.
[60] C. Rotthaus, D. Weston, and R. Wiegand, Indecomposable Gorenstein modules of odd rank, J. Algebra 214 (1999),
122 – 127.
[61] S. Saccon, One-dimensional local rings of infinite Cohen-Macaulay type, Ph.D. thesis, University of Nebraska–Lincoln,
Department of Mathematics (2010).
[62] W.A. Schmid, Arithmetical characterization of class groups of the form Z/nZ ⊕ Z/nZ via the system of sets of lengths,
Abh. Math. Semin. Univ. Hamb. 79 (2009), 25 – 35.
, Characterization of class groups of Krull monoids via their systems of sets of lengths : a status report, Number
[63]
Theory and Applications : Proceedings of the International Conferences on Number Theory and Cryptography (S.D.
Adhikari and B. Ramakrishnan, eds.), Hindustan Book Agency, 2009, pp. 189 – 212.
[64]
, A realization theorem for sets of lengths, J. Number Theory 129 (2009), 990 – 999.
[65] O. Schmidt, Über unendliche Gruppen mit endlicher Kette, Math. Z. 29 (1929), 34 – 41.
[66] J. Wedderburn, On the direct product in the theory of finite groups, Ann. Math. 10 (1909), 173 – 176.
[67] R. Wiegand, Direct-sum decompositions over local rings, J. Algebra 240 (2001), 83 – 97.
[68] R. Wiegand and S. Wiegand, Bounds for one-dimensional rings of finite Cohen-Macaulay type, J. Pure Appl. Algebra
93 (1994), 311 – 342.
Department of Mathematics and Computer Science, University of Central Missouri, Warrensburg, MO
64093, USA; [email protected]
Institut für Mathematik und Wissenschaftliches Rechnen, Karl-Franzens-Universität Graz, Heinrichstraße
36, 8010 Graz, Austria; [email protected]
| 0math.AC
|
arXiv:1802.09158v1 [cs.GT] 26 Feb 2018
Surrogate Scoring Rules and a Dominant Truth Serum
for Information Elicitation
YANG LIU, Harvard University
YILING CHEN, Harvard University
We study information elicitation without verification (IEWV) and ask the following question: Can we achieve
truthfulness in dominant strategy in IEWV? This paper considers two elicitation settings. The first setting is
when the mechanism designer has access to a random variable that is a noisy or proxy version of the ground
truth, with known biases. The second setting is the standard peer prediction setting where agents’ reports are
the only source of information that the mechanism designer has. We introduce surrogate scoring rules (SSR)
for the first setting, which use the noisy ground truth to evaluate quality of elicited information, and show
that SSR achieve truthful elicitation in dominant strategy. Built upon SSR, we develop a multi-task mechanism, dominant truth serum (DTS), to achieve truthful elicitation in dominant strategy when the mechanism
designer only has access to agents’ reports (the second setting). The method relies on an estimation procedure to accurately estimate the average bias in the reports of other agents. With the accurate estimation, a
random peer agent’s report serves as a noisy ground truth and SSR can then be applied to achieve truthfulness in dominant strategy. A salient feature of SSR and DTS is that they both quantify the quality or value of
information despite lack of ground truth, just as proper scoring rules do for the with verification setting. Our
work complements both the strictly proper scoring rule literature by solving the case where the mechanism
designer only has access to a noisy or proxy version of the ground truth, and the peer prediction literature
by achieving truthful elicitation in dominant strategy.
1
INTRODUCTION
Eliciting private information about events and tasks from human agents is an almost universal
problem in our society when there is uncertainty and asymmetric information. For example, financial experts are asked to make predictions on whether the S & P 500 index will go up next
week, seismologists attempt to predict future earthquake activities, students doing peer grading
are asked to assess the correctness of others’ solutions. From the perspective of incentive alignment, an important goal of a mechanism designer is to incentivize agents to truthfully reveal their
private information so that high-quality information is obtained.1
Incentive design for information elicitation has been extensively studied. However, proposed
solutions have achieved different notions of truthfulness depending on which one of the following two categories an elicitation problem falls into: information elicitation with verification and
information elicitation without verification.
Information elicitation with verification refers to settings where the mechanism designer will
have access to the ground truth at some point (e.g. whether S & P 500 index actually went up
and whether a predicted earthquake actually happened) and can use the ground truth to evaluate
the quality of elicited information. Strictly proper scoring rules [Brier, 1950, Gneiting and Raftery,
2007, Matheson and Winkler, 1976a, Murphy and Winkler, 1970] have been proposed and studied
to incentivize truthful reporting of private predictions (probability assessment of event outcome).
Facing a strictly proper scoring rule, an agent strictly maximizes his expected score by truthfully
revealing his prediction. Hence, truthfulness in dominant strategy is achieved here. Moreover, the
score of an agent measures the quality of his prediction.
1 Many
also consider incentivizing both effort exertion to improve the quality of private information and truthful information revelation.
2
1 INTRODUCTION
Information elicitation without verification refers to settings where the mechanism designer
does not have access to the ground truth (e.g. whether a student’s homework solution is correct is unavailable as otherwise peer grading is not needed). An elegant family of incentive
mechanisms [Dasgupta and Ghosh, 2013, Jurca and Faltings, 2006, Miller et al., 2005, Prelec, 2004,
Radanovic and Faltings, 2013, Witkowski and Parkes, 2012], collectively called peer prediction, has
been developed for this category to elicit private signals (e.g. a peer grader’s judgement on the
correctness of a homework solution). Peer prediction leverages the correlation of agents’ private
signals and scores an agent’s report based on how it compares with the reports from other agents.
Agents play a Bayesian game in peer prediction and there exists a Bayesian Nash Equilibrium
(BNE) where every agent reports truthfully. Truthfulness at BNE is clearly a weaker incentive
guarantee than truthfulness in dominant strategy as the former is sensitive to errors and mistakes
of other agents. Peer prediction also suffers from multiplicity of equilibria as there are other nontruthful BNE.2 In experiments, the truthful equilibrium was often not reached [Gao et al., 2014].
In addition, the peer prediction score of an agent measures the correlation of his report with those
of others, but it doesn’t reflect the quality of the agent’s report.
We focus on information elicitation without verification in this paper and ask:
Can we achieve truthfulness in dominant strategy in information elicitation without verification?
We provide a positive answer to this question for binary information elicitation. Moreover, in our
proposed mechanisms, the score of an agent measures the quality of his report as in information
elicitation with verification.
Two specific information elicitation without verification settings are considered in this paper.
The first setting is when the mechanism designer has access to a random variable that is a noisy
or proxy version of the ground truth, with known biases. The second setting is the standard peer
prediction setting where agents’ reports are the only source of information that the mechanism
designer has.
For the first setting, we introduce surrogate scoring rules (SSR), which use the noisy ground
truth to evaluate quality of elicited information, and show that SSR achieve truthful elicitation in
dominant strategy without access to the ground truth. These surrogate scoring rules are inspired
by the use of surrogate loss functions in machine learning [Angluin and Laird, 1988, Bylander,
1994, Natarajan et al., 2013, Scott, 2015, Scott et al., 2013] and they remove bias from the noisy
random variable such that in expectation a report is as if evaluated against the ground truth.
Built upon SSR, in the second setting where the mechanism designer only has access to agents’
reports, we develop a multi-task mechanism, dominant truth serum (DTS), to achieve truthful elicitation in dominant strategy. The method relies on an estimation procedure to accurately estimate
the average bias in the reports of other agents. With the accurate estimation, a random peer
agent’s report serves as a noisy ground truth and SSR can then be applied to achieve truthfulness
in dominant strategy. We further show that DTS can be extended to achieve truthful elicitation
in dominant strategy in settings where agents’ signals are correlated conditioned on the ground
truth. This allows us to overcome the failure of classic peer prediction in the presence of weak,
correlated signals, observed by [Gao et al., 2016].
A salient feature of SSR and DTS is that they both quantify the quality or value of information
despite lack of ground truth, just as proper scoring rules do for the with verification setting. In this
sense, SSR and DTS unify the with and without verification settings by always rewarding agents
2 Several
recent peer prediction mechanisms [Dasgupta and Ghosh, 2013, Kong and Schoenebeck, 2016a, Shnayder et al.,
2016] have made truthful equilibrium focal in the sense that it leads to the highest expected payoff to agents among all
equilibria. But there is at least one other equilibrium that gives the same expected payoff to agents.
1.1 Related work
3
based on their quality of information (and hence achieving truthfulness in dominant strategy). In
addition, SSR and DTS apply to both signal and prediction elicitation.
The rest of the paper is organized as follows. In Section 2, we describe our model and the two
elicitation settings. Section 3 introduces SSR for the first elicitation setting. Section 4 details DTS
for the second elicitation setting. Then, in Section 5, we extend DTS to correlated signals. Section 6
concludes with discussions on some additional directions.
1.1
Related work
The most relevant literature to our paper is strictly proper scoring rule and peer prediction. Scoring
rules were developed for eliciting truthful prediction (probability) [Brier, 1950, Gneiting and Raftery,
2007, Jose et al., 2006, Matheson and Winkler, 1976b, Savage, 1971, Winkler, 1969]. For example,
the pioneer works [Brier, 1950] proposed Brier scoring to verify the qualities of forecasts. A full
characterization result is given for strictly proper scoring rules in [Gneiting and Raftery, 2007].
Our work complements the literature of scoring rule by proposing surrogate scoring rules to cope
with the case when there is only access to a noisy or proxy version of the ground truth.
Our surrogate scoring rules can be viewed as a class of proxy scoring rules defined in [Witkowski et al.,
2017]. [Witkowski et al., 2017] proposed to design proper proxy scoring rules by using an unbiased
proxy random variable as the ground truth in a strictly proper scoring rule. Our proxy random
variable is biased. We focus on designing a surrogate scoring function that removes bias in any
noisy ground truth, and thus can deal with a more general setting.
Motivated by the complexity of reporting the probability for events with large outcome space,
property elicitation, such as for mean, variance etc of probability distributions, has been studied
more recently [Frongillo and Kash, 2015, Lambert et al., 2008]. In this paper we do not cover the
property elicitation setting.
The core idea of peer prediction is to score each agent based on another reference report elicited
from the rest of agents, and to leverage on the stochastic correlation between different agents’ information. This line of research started with the celebrated Bayesian Truth Serum work [Prelec,
2004], where a surprisingly popular answer methodology is shown to be able to incentivize agents
to truthfully report even they believe they hold the minority answer (but more likely to be true
in their own opinion). The seminal work [Miller et al., 2005] established that strictly proper scoring rule [Gneiting and Raftery, 2007] can be adopted in the peer prediction setting for eliciting
truthful reports (but the mechanism designer need to know details of agents’ model); a sequence
of followed up works have been done to relax the assumptions that have been imposed therein
[Radanovic and Faltings, 2013, Witkowski and Parkes, 2012]. More recently, [Dasgupta and Ghosh,
2013, Witkowski et al., 2013] formally introduced and studied an effort sensitive model for binary signal data elicitation. Particularly [Dasgupta and Ghosh, 2013] proposed a multi-task peer
prediction mechanism that can help remove undesirable equilibria that lead to low quality reports. These results are further strengthened and extended to a non-binary signal setting in
[Kong and Schoenebeck, 2016a, Shnayder et al., 2016]. An information theoretical framework for
studying this type of elicitation problems has been proposed in [Kong and Schoenebeck, 2016a,b].
[Kong and Schoenebeck, 2016b] also proposed a conceptual mechanism to achieve truthfulness in
dominant strategy but it requires infinite number of tasks. Our work is the first one to establish
truthfulness in dominant strategy with a finite number of tasks for a peer prediction setting.
It is worth to mention that our work borrows ideas from the machine learning literature on
learning with noisy data [Bylander, 1994, Menon et al., 2015, Natarajan et al., 2013, Scott, 2015,
van Rooyen and Williamson, 2015]. From a high level’s perspective, our goal in this paper aligns
with the goal in learning from noisy labels - both aim to evaluate a prediction when the ground
4
2 INTRODUCTION
truth is missing, but instead a noisy signal of the ground truth is available. Our work addresses
additional challenge that the error rate of the noisy signal remains unknown a priori.
2 OUR MODEL
Prior work on information elicitation typically considers one of two types of information: private
signals and private beliefs (i.e. probabilistic predictions). In this paper, we develop mechanisms
that apply to both cases. Hence, in Section 2.1, we first introduce our model of information and
notations to unify these two types of information. Then, Section 2.2 and 2.3 will introduce our
model and design goals for two elicitation settings without accessing the ground truth. In the
first setting (Section 2.2), the mechanism designer has access to a random variable that is a noisy
version of the ground truth with known bias. The second setting (Section 2.3) is the typical setting
in the peer prediction literature where the mechanism designer only has access to reports from
agents. For both settings, our design goal is to achieve truthful elicitation in dominant strategy.
2.1
Model of information
Suppose we are interested in eliciting information about a binary event y ∈ {0, 1} from a set of
human agents, indexed by [N ] := {1, 2, ..., N }. The realization of the event y has prior distribution
P0 := Pr[y = 0], P1 = Pr[y = 1]. We assume a non-trivial prior distribution that 0 < P0 , P1 < 1.
Each of the N agents holds a noisy observation of y, denoting as si . Agents’ observations are
conditionally independent: Pr[si , s j |y] = Pr[si |y] · Pr[s j |y]. We short-hand the following error
rates:
e 1,i := Pr[si = 0|y = 1], e 0,i := Pr[si = 1|y = 0],
i.e., e 1,i , e 0,i are the “error probabilities” or the subjective “human bias” of agent i’s observation for
y. We do not assume homogeneous agents, that is we allow agents to have different e 1,i , e 0,i . The
error rates can also model subjectivity in agents’ private belief and observation. Based on signal
si , each agent can form a posterior belief about y, denoting as pi := Pr[y = 1|si ]. When there
are multiple tasks, we assume the prior, P0 and P1 , and each agent’s error rates, e 1,i and e 0,i , are
homogeneous across tasks.
We consider eliciting either si or pi from agents,3 but we are not able to access the ground truth
y to verify the reported information from agents. We call the elicitation of these two types of
information as signal elicitation and prediction elicitation respectively. The literature on proper
scoring rules has been primarily considering prediction elicitation, while the literature on peer
prediction has focused on signal elicitation. To unify both types of information, we denote the
information space as I, and each agent’s information as Ii . For signal elicitation, Ii = si and
I = {0, 1}. For prediction elicitation, Ii = pi and I = [0, 1]. Denote agent i’s report to a mechanism
as ai ∈ I . Then, for for signal elicitation and prediction elicitation, we have ai : si → {0, 1} and
ai : pi → [0, 1] respectively.
2.2
Model of elicitation with noisy ground truth
The first elicitation setting is when the mechanism designer has access to the realization of a binary
random variable z ∈ {0, 1}, which is a noisy or proxy version of the ground truth with known bias.
The bias of z is again captured by the error rates:
e 1,z := Pr[z = 0|y = 1], e 0,z := Pr[z = 1|y = 0].
3 For
clarity of presentation, we assume that we elicit the same type of information from all agents. But our work can be
extended to a mixed elicitation setting.
2.3 Model of elicitation with peer reports
5
We cannot expect to do much if z is independent of y and hence assume that z and y are stochastically relevant, an assumption commonly adopted in the information elicitation literature [Miller et al.,
2005].
Definition 2.1. z and y are stochastically relevant if there exists an s ∈ {0, 1} such that
Pr[y = s |z = 0] , Pr[y = s |z = 1].
The following lemma shows that the stochastic relevance requirement directly translates to a
constraint on the error rates, that is, e 1,z + e 0,z , 1.
Lemma 2.2. y and z are stochastically relevant if and only if e 1,z + e 0,z , 1. And y and z are
stochastically irrelevant (independent) if and only if e 1,z + e 0,z = 1.
We consider both signal elicitation and prediction elicitation. Our goal is to design a scoring
function φ(ai , z) : I × {0, 1} → R+ such that when agent i’s report ai is evaluated against the
realization of the noisy ground truth z, agent i strictly maximizes his expected score by reporting
his true information ai = Ii . Formally, we want to achieve
Strict properness with noisy ground truth: E[φ(Ii , z)] > E[φ(ai , z)], ∀ai , Ii ,
(1)
where the expectation is taken with respect to the agent’s subjective belief Pr[z|si ].
It’s worth noting that prediction elicitation in this model, as in the literature, doesn’t need agents
to form their prediction pi according to the signal model described in Section 2.1, and instead only
requires that agents each have a subjective belief pi . For simplicity of presentation, we use the
signal model in Section 2.1 throughout the paper.
2.3
Model of elicitation with peer reports
The second elicitation setting is close to the standard setting considered in the peer prediction
literature. Here the only source of information that the mechanism designer has access to is the
agents’ reports. In this setting, we assume ”differentiable” prior such that P0 , P1 . This is a
technical condition that we need. In practice, with multiple tasks, we can perturb the priors by
adding some known tasks to (or deleting task from) one of the two classes.
2.3.1 Informativeness of signals. We adopt the standard assumption requiring that each agent’s
signal si and the ground truth y are stochastically relevant, i.e. e 1,i +e 0,i , 1. Moreover, we assume
that signals are Bayesian informative of y, that is, Pr[y = s |si = s] > Pr[y = s], ∀i, s. In other
words, the posterior probability for y = s is greater than the prior probability of y = s if an agent
receives a signal s. This assumption has been adopted by [Dasgupta and Ghosh, 2013, Miller et al.,
2005], and has been shown by [Liu and Chen, 2017] to be equivalent to a constraint on the error
rates of agent signals.
Lemma 2.3. (Lemma 2.1 in [Liu and Chen, 2017]) si is Bayesian informative of y if and only if
e 1,i + e 0,i < 1.
The case e 1,i + e 0,i > 1 can be viewed as negative Bayesian informative as always flipping
one’s observation will return a Bayesian informative signal. In this paper we focus only on settings where every agent’s signal is Bayesian informative of y. But this assumption is to simplify
presentation as our results hold for negative Bayesian informative agents, as well as for a mixed
population.
6
3 OUR MODEL
2.3.2 Knowledge of agents and the mechanism designer. For signal elicitation, agents do not need
to be aware of their own values of e 1,i and e 0,i , but they need to know that their own signals are
Bayesian informative and the error rates satisfy e 1,i + e 0,i < 1. For prediction elicitation, agents
need to know their own values of e 1,i and e 0,i to form their predictions. The above knowledge does
not need to be common knowledge among agents. This is because we seek to achieve dominant
strategy truthfulness and hence agents are as if not playing a Bayesian game.
The mechanism designer knows the priors of ground truth P0 , P1 , but not the realized y, nor
the private information si ’s or pi ’s. Further the designer does not know any of the biases of agents’
information, e 1,i and e 0,i , a-priori. This can capture scenarios when we know the prior distribution
of answers for a certain type of questions, but not specifically the answer for a particular question,
nor how human agents make wrong prediction of such events. In our proposed mechanism, we
will show how to infer the true error rates of agents’ reports.
2.3.3 Design goal. We again aim to achieve truthful elicitation in dominant strategy for both
signal elicitation and prediction elicitation. Denote by D−i the set of information collected from
agents j , i, possibly for multiple tasks, i.e. D−i := {a j }j,i For this part, our goal is to find a
reference signal z D−i ∈ {0, 1}, as a function of D−i , and a scoring function
φ D−i (ai , z) : I × {0, 1} → R+ ,
also as a function of D−i , such that it is a dominant strategy for agent i to truthfully report his
private information:
Truthfulness in dominant strategy with peer reports: Ii = argmaxai E[φ D−i (ai , z D−i )], (2)
where the expectation is taken with respect to the agent’s subjective belief Pr[z D−i |si ]. When it’s
clear in context, we’ll use z as shorthand for z D−i and φ for φ D−i .
3 SURROGATE SCORING RULES
In this section we propose surrogates scoring rules (SSR) to address the question of achieving strict
properness when only having access to a noisy ground truth. SSR is built upon strictly proper
scoring rules. Thus, we first briefly discuss strictly proper scoring rules. Then, we formally define
SSR, prove its strict properness and give an implementation of it.
3.1
Strictly proper scoring rules
As mentioned earlier, when ground truth is available, strictly proper scoring rules were designed
to elicit predictions, i.e. the pi ’s. A scoring function S : [0, 1] × {0, 1} → R+ is strictly proper if
and only if
E[S(pi , y)] > E[S(p̃i , y)], ∀p̃i , pi ,
where the expectation is taken with respect to the agent’s belief Pr[y|si ]. There is a rich family of
strictly proper scoring rules, including Brier (S(pi , y) = 1 − (pi − y)2 ), logarithmic and spherical
scoring rules [Gneiting and Raftery, 2007].
Though not enjoying much attention in the literature, the above idea of defining strictly proper
scoring rules also applies to signal elicitation, that is to design a function S : {0, 1}2 → R+ such
that
E[S(si , y)] > E[S(s̃i , y)], ∀s̃i , yi ,
where the expectation is taken with respect to the agent’s belief Pr[y|si ]. For instance, with knowledge of the prior of y (Pr[y = s], s ∈ {0, 1}), the following prior dependent output agreement
3.3 An implementation of SSR
7
scoring function is strictly proper (1/Prior):
S(s̃i , y) =
1
· 1(s̃i = y).
Pr[y = s̃i ]
Lemma 3.1. 1/Prior scoring function is strictly proper.
To unify prediction and signal elicitation, we will denote a strictly proper scoring rule for eliciting agent i’s information Ii as S(ai , y) : I × {0, 1} → R+ .
3.2
Surrogate scoring rules
We define surrogate scoring rules (SSR), when we only have access to z, a noisy and proxy signal
of the ground truth y, and e 1,z + e 0,z , 1.
Definition 3.2 (Surrogate Scoring Rules). φ : I × {0, 1} → R+ is a surrogate scoring rule if there
exists a strictly proper scoring rule S : I × {0, 1} → R+ and a strictly increasing function f : R+ →
R+ such that ∀ai ,
Ey,z [φ(ai , z)] = f (Ey [S(ai , y)]).
(3)
If Eqn .(3) is true, the strict properness of φ(·) holds immediately because of the strict properness
of S and monotonicity of f (·). For instance, a simple such f to aim for can be an affine function
that there exist two constants c 1 > 0, c 2 such that
Ey,z [φ(ai , z)] = c 1 · Ey [S(ai , y)] + c 2 .
(4)
Theorem 3.3. SSR is strictly proper for eliciting Ii .
The above approach seeks a surrogate scoring function φ(·) that helps us remove the bias in z
to return us a strictly proper score in expectation.
SSR can be viewed as a particular class of proxy scoring rules [Witkowski et al., 2017]. But the
approach of [Witkowski et al., 2017] to achieve properness is to plug in an unbiased proxy ground
truth to a strictly proper scoring rule. SSR on the other hand directly works with biased proxy and
the scoring function is designed to de-bias the noise.
3.3 An implementation of SSR
Consider any strictly proper scoring function S for eliciting Ii using y, the ground truth. We define
the following SSR, which we name as SSR alpha:
(SSR alpha) : φ(ai , z = s) =
(1 − e 1−s,z ) · S(ai , s) − es,z · S(ai , 1 − s)
.
1 − e 1,z − e 0,z
(5)
From above we note that the knowledge of the error rates e 1,z , e 0,z is crucial for defining a SSR.
The above scoring function is inspired by the work ”learning with noisy label” [Natarajan et al.,
2013]. Intuitively speaking, the linear transform will ensure that in expectation, ai is scored against
y:
Lemma 3.4. When e 1,z + e 0,z < 1, we have for (SSR alpha): Ez |y [φ(ai , z)] = S(ai , y), ∀y ∈ {0, 1}.
This can be proved fairly straightforwardly via spelling out the expectation. Interested readers
are also referred to Lemma 1, [Natarajan et al., 2013]. In above, the scoring function φ(·) takes
the form of linear combination of the scoring function evaluated at the noisy outcome z and the
one evaluated at the opposite signal 1 − z. The second term is to account for the fact that the
reference signal is merely a noisy proxy for the ground truth signal. The linear weights, carefully
defined over the bias level of the reference, balance the bias in the noisy signal with respect to
8
4 SURROGATE SCORING RULES
the ground truth. Also note the above unbiasedness is achieved w.r.t. each possible realization of
y, being independent of its prior distribution. This helps us remove the requirement of common
knowledge over the priors among agents, and thus our solution works for the subjective prior
setting. Now we establish the strict properness of SSR:
Theorem 3.5. (SSR alpha) is strictly proper.
Proof. Consider agent i. Consider the case that e 1,z + e 0,z < 1. The proof is straightforward
following the “unbiasedness” property established for φ(·) in Lemma 3.4:
E[φ(ai , z)] = E E[φ(ai , z)|y] = E[S(ai , y)].
The theorem follows immediately from the strictly properness of S.
Now consider the case with e 1,z + e 0,z > 1. We cannot directly apply Lemma 3.4. Now let’s
define the “flip signal” ẑ of z: ẑ = 1 − z. Easy to see that
e 1, ẑ := Pr[ẑ = 0|z = 1] = Pr[z = 1|z = 1] = 1−e 1,z , e 0, ẑ := Pr[ẑ = 1|z = 0] = Pr[z = 0|z = 0] = 1−e 0,z ,
and e 1, ẑ + e 0, ẑ < 1. The scoring function for agent i then becomes:
(1 − e 1−s,z ) · S(ai , s) − es,z · S(ai , 1 − s)
.
1 − e 1,z − e 0,z
[1 − (1 − es,z )]S(ai , 1 − s) − (1 − e 1−s,z )S(ai , s)
=
1 − (1 − e 1,z ) − (1 − e 0,z )
(1 − e 1−(1−s), ẑ )S(ai , 1 − s) − e 1−s, ẑ S(ai , s)
=
1 − e 1, ẑ − e 0, ẑ
= φ(ai , ẑ = 1 − s).
φ(ai , z = s) =
(6)
Then we have
E[φ(ai , z)] = E[φ(ai , ẑ)] = E E[φ(ai , ẑ)|y] = E[S(ai , y)],
where the last equality is due to the unbiasedness of φ(·) with respect to ẑ (instead of z), and the
fact that e 1, ẑ + e 0, ẑ < 1, so Lemma 3.4 can be applied. Again it is easy to see it would be agent
i’s best interest to tell the truth, due to the strict properness or truthfulness of S(·), finishing the
proof.
Remark 1. We provide a simple yet effective scoring function to unify the two elicitation scenarios.
Intuitively, the reason we can achieve above is that the core step of surrogate function φ(·) only operates
with the reference answer z, and it is independent of the format of the information being elicited.
4
DOMINANT TRUTH SERUM
The results in previous section are built upon the fact that there exists a reference signal for the
ground truth and we know its error rates. In this section, we apply the idea of SSR to the peer
prediction setting. A reasonable way to do so is to take agents’ reports as the source for this noisy
copy of the ground truth. Yet the mechanism designer cannot assume the knowledge of the noise
in agents’ reports. In this section we discuss how to select the reference answer z and how to learn
e 1,z , e 0,z accurately so that φ(·) will be well defined. Then applying Theorem 3.5, we will be able to
establish a dominant strategy argument in the peer prediction setting.
4.2 Bias learning with matching
4.1
9
Reference report
An intuitive way of selecting a reference report z is to uniformly randomly select a report from
agents j , i, i.e., from D−i . Denote by ẽ 1,i , ẽ 0,i the error rate of agent i’s reported information
ai . For the signal elicitation setting, when agents truthfully report, i.e. ai = si , we have ẽ 1,i =
e 1,i , ẽ 0,i = e 0,i , and when they revert their signals we have ẽ 1,i = 1 − e 1,i , ẽ 0,i = 1 − e 0,i . For
the case with prediction elicitation, we can similarly define a signal from agent i, via drawing a
random sample: asi ∼ Bernoulli(p̃i ). In this case, we will still call such a asi as agent i’s “signal”
but we should keep in mind this signal is randomly drawn from his reported prediction, instead
of being deterministic. We similarly define ẽ 1,i , ẽ 0,i , and assume conditional independence among
asi s.
e 1,z and e 0,z can then be characterized fairly easily as follows:
Í
Í
j,i ẽ 1, j
j,i ẽ 0, j
e 1,z :=
, e 0,z :=
.
(7)
N −1
N −1
This can be obtained by the following argument: Denote the following even
Wj := {agent j selected as the reference agent}.
Then we have
Pr[z = 0|y = 1] =
Õ
j,i
=
Õ
j,i
Pr z = 0, Wj |y = 1
Pr z = 0|y = 1, Wj · Pr Wj =
Õ
1
·
ẽ 1, j .
N − 1 j,i
We have similar interpretation for e 0,z . The next subsection will detail the steps towards estimating
e 0,z , e 1,z . Note if we know the average error rates when agents truthfully report, the SSR already
gives us an equilibrium implementation of truthful elicitation - as in the equilibrium argument,
every other agent will truthfully report.
4.2 Bias learning with matching
Our learning algorithm for inferring e 0,z , e 1,z relies on establishing three equations towards characterizing e 1,z , e 0,z . We will first show that the three equations, with knowing their true parameters,
together will uniquely define e 1,z , e 0,z . Then we argue that with estimated and imperfect parameters from agents’ report, the solution from the perturbed set of equations will approximate the true
values of e 1,z , e 0,z , with guaranteed accuracy.
(1) Posterior distribution: The first equation is based on on the posterior distribution of 0/1
label information collected from D−i , denoting as P0, −i := Pr[z = 0], P1, −i := Pr[z = 1].
For P0, −i we have
P0, −i = P1e 1,z + P0 (1 − e 0,z ).
(8)
Note the above equation is also equivalent with P1, −i = P1 (1 − e 1,z ) + P0e 0,z .
(2) Matching between two signals: The second equation is derived from a second order
statistics, namely the matching probability. Consider the following experiments: draw
two independent signals (for the same task) from the reference agents 4 , denote them as
4 Note in practice,
we won’t obtain independent copies of answers for the same question from the same agents for inferring
the matching probability. However when the number of agent is large enough, we will show that drawing two or three
agents randomly from the population (without replacement) can approximate these independent matching probabilities
with small and diminishing errors (as a function of number of agents N ).
10
4 DOMINANT TRUTH SERUM
z 1 , z 2 . Denote the matching-on-1 probability of the two signals as
Pr[z 1 = z 2 = 1] = q 2, −i .
Further the matching probability can be written as a function of e 0,z , e 1,z :5
2
q 2, −i = P1 (1 − e 1,z )2 + P0e 0,z
.
(9)
(3) Matching among three signals: The third equation is obtained by going one order
higher that, we check the matching-on-1 probability over three independent signals drawn
from reference agents. Similarly as defined for (2) matching between two signals, draw
three independent signals (for the same task) from the reference agents, denote them as
z 1 , z 2 , z 3 . Denote the matching-on-1 probability of the three signals as Pr[z 1 = z 2 = z 3 =
1] = q 3, −i . Then we also have that
2
q 3, −i = P1 (1 − e 1,z )3 + P0e 0,z
.
(10)
The reasons we selected the above three equations are two-fold. The first reason is that, as we
will show below, the above three equations uniquely define e 0,z , e 1,z . The second reason is that
in practice we only need to know P1, −i , q 2, −i , q 3, −i in order to establish the three equations, since
we have assumed the knowledge of priors P0 , P1 . All three parameters P1, −i , q 2, −i , q 3, −i can be
estimated from agents’ reports, without the need of knowing any ground truth labels.
As promised, we first establish the following theorem:
Theorem 4.1. (e 0,z , e 1,z ) is the unique pair of solution to Eqn.(8, 9, 10), when P0 , P1 .
Proof. First of all, there are at most two solutions from Eqn. (8) and (9), since the equations
are at most order two. Further we prove the following properties of its solutions:
Lemma 4.2. For the solutions to Eqn.(8) and (9), we have
(1) When the reports are uninformative such that e 0,z + e 1,z = 1, the above equation return
exactly one solution satisfying that e 0,z + e 1,z = 1.
(2) When the reports are informative such that e 0,z + e 1,z , 1, Eqn.(8) and (10) jointly return two
pair of solutions
[e 0,z (1), e 1,z (1)], [e 0,z (2), e 1,z (2)]
and exactly one of them satisfies e 0,z (1)+e 1,z (1) < 1 and the other one that e 0,z (2)+e 1,z (2) > 1.
Further their summations of the two pairs have the same distance to the 1 (uninformative):
1 − (e 0,z (1) + e 1,z (1)) = (e 0,z (2) + e 1,z (2)) − 1.
We shall shorthand the following: x 1 := e 0,z , x 2 := 1 − e 1,z .
Proof. Eqn.(8) and (9) become equivalent with the following
(I) : P0x 1 + P1x 2 = P1, −i , (II) : P0x 12 + P1x 22 = q 2, −i .
From the first equation (I) we know that x 1 =
x1 − x2 =
P 1, −i −P1 ·x 2
.
P0
Then
P1, −i − P1 · x 2
P1, −i − x 2
− x2 =
P0
P0
5 Note here we implicitly assumed that agents’ reported signals are also conditional
independent of the ground truth signal.
In practice, without knowing the ground truth, it is hard for agents to coordinate based on the ground truth. In fact if
the reporting strategies are functions of each agent’s observation, we will have the conditional independence established
immediately, according to the conditional independence of agents’ observations.
4.3 Statistical consistency results
11
Further plug x 1 into the second equation (II) we have
P1x 22 − 2P1 P1, −i x 2 + P1,2 −i − P0q 2, −i = 1 ⇒ x 2 = P1, −i ± ∆,
q
(2P1P1, −i )2 − 4P1 (P1,2 −i − P0q 2, −i )
.
where ∆ :=
2P1
From above we have derived the following e 0,z + e 1,z − 1 = x 1 − x 2 = ± P∆0 . When e 0,z + e 1,z = 1,
we know we must have ∆ = 0 so we land at an unique solution. When e 0,z + e 1,z , 1, we know
that ∆ > 0 and we have two solutions – but both solutions satisfy that
|e 0,z + e 1,z − 1| =
∆
,
P0
finishing the proof.
We now show the following fact:
Eqn.(10) − Eqn.(8) · Eqn.(9)
⇔P0x 13 + P1x 23 − (P0x 1 + P1x 2 )(P0x 12 + P1x 22 ) = q 3,i − q 2,i P1
⇔P0 P1 (x 13 + x 23 − x 12x 2 − x 1x 22 ) = q 3, −i − q 2, −i P1, −i
⇔P0 P1 (x 1 − x 2 )2 (x 1 + x 2 ) = q 3, −i − q 2, −i P1, −i
(q 3, −i − q 2, −i P1, −i )P0
∆2 P1
The above equation, together with equation (I), will uniquely identify the solution for x 1 , x 2 when
P0 , P1 :
q 3, −i − q 2, −i P1, −i
1
P1 − P1, −i ,
x1 =
P1 − P0
q 2, −i − P1,2 −i
q 3, −i − q 2, −i P1, −i
1
x2 =
P
P1, −i −
0 .
P1 − P0
q 2, −i − P1,2 −i
⇔x 1 + x 2 =
This completes the proof.
This result implies that when there is no ground truth data, knowing how frequently human
agents reach consensus with each other will help us characterize their (average) subjective biases.
We conjecture that the above estimation method can be extended to the case with unknown
priors of the ground truth. The idea is very straightforward following our approach: We can relax
the knowledge of the priors to have two additional unknown variables. Then we will go two orders
up to have additional two equations characterizing the matchings among four and five reference
signals, to cop with the two more unknown variables. This merits future studies.
4.3 Statistical consistency results
Now we detail the learning procedure. Our method relies on multiple tasks (thus our mechanism
is naturally a multi-task one). We will collect information on multiple independently drawn events
from priors y(1), y(2), · · · ∼ P0 , P1 by randomly assigning them to agents. Each task will be randomly assigned to three agents. Suppose there are K tasks that have been assigned to agents j , i
- this is not hard to guarantee ∀i when we assign more than K tasks to the entire set of agents
randomly, and when N is large.
The algorithm first estimates the following three quantities for each agent i: q̃ 2, −i , q̃ 3, −i and
P̃1,i −i based on above information. Denote the three agents that are assigned task k = 1, ..., K as
12
4 DOMINANT TRUTH SERUM
r 1 (k), r 2 (k), r 3 (k) , i, and their reports for task k as ỹr idx (k) (k), idx = 1, 2, 3. Then we estimate the
three parameters as follows:
ÍK
1(ỹr 1 (k) (k) = 1)
,
P̃1, −i = k=1
K
ÍK
1(ỹr 1 (k) (k) = ỹr 2 (k) (k) = 1)
,
q̃ 2, −i = k=1
K
ÍK
1(ỹr 1 (k) (k) = ỹr 2 (k) (k) = ỹr 3 (k) (k) = 1)
.
q̃ 3, −i = k=1
K
Mechanism 1 Learning of e 1,z , e 0,z
1. Estimate q̃ 2, −i , q̃ 3, −i and P̃1, −i using {ỹr idx (k) (k), k = 1, 2, ..., K, idx = 1, 2, 3}.
2. Compute the following:
q̃ 3, −i − q̃ 2, −i P̃1, −i
1
ẽ 0,z : =
· P1 − P̃1, −i ,
P1 − P0 q̃ 2, −i − (P̃1, −i )2
q̃ 3, −i − q̃ 2, −i P̃1, −i
1
· P0 ,
ẽ 1,z : = 1 −
P̃1, −i −
P1 − P0
q̃ 2, −i − (P̃1, −i )2
(11)
(12)
Remark 2. We notice that the sensitivity in computing e 0,z , e 1,z is inversely proportional to P1 −P0 .
This implies that the difficulty in selecting the right pair of root increases when the priors are less
differentiable.
We would like to show that the estimation error in estimating reports’ error rate can be bounded
as a function of K and N . The first source of errors is the imperfect estimations of q 2, −i , q 3, −i , P1, −i .
The second is due to estimation errors for matching probability with heterogeneous agents. Formally we have the following theorem:
0.12
1.8
1.7
0.08
Reward gap
Estimation error
0.1
0.06
0.04
1.5
1.4
0.02
0
500
1.6
1000
1500
Number of tasks
2000
1.3
500
1000
1500
Number of tasks
2000
Fig. 1. Convergence of the estimation. Left: convergence with different number of tasks. e 0,z = 0.85, e 1,z =
0.83. Right: gap in reward functions (1/Prior scores) between truth-telling and mis-reporting.
Theorem 4.3. The mechanism designer can learn noisy copies of e 1,z , e 0,z , φ(·) for each agent i
using data collected from agents j , i, satisfying that
4.4 Truth-telling is a dominant strategy
13
(1) |ẽ 1,z − e 1,z | ≤ ϵ, |e˜0,z − e 0,z | ≤ ϵ with probability at least 1 − δ .
(2) For the scoring function φ(·) defined using ẽ 1,z , ẽ 0,z , we have with probability at least 1 − δ 1
that |φ̃(t, y) − φ(t, y)| ≤ ϵ1 , ∀t, y.
q
q
log K
log K
1
1
),
ϵ
=
O
+
,
δ
=
O(
, δ 1 = O( K1 ) can be made arbitrarily
All terms ϵ = O N1 +
1
K
K
N
K
small with increasing number of samples K and N .
Proof can be found in Appendix. The main steps of the proof are based on Theorem 4.1 and
sensitivity analysis of the solutions Eqn. (11,12). A set of simple simulation results are shown in
Figure 1. We simulated a scenario with e 0,z = 0.85, e 1,z = 0.83, i.e., majority people are wrong
from the reference set of agents. Nevertheless the Left side figure shows a nice decay in the errors
of estimating e 0,z , e 1,z . The Right side figure shows the gap in expected score between truthful
reporting and misreporting (denoted as ”Reward gap”) with the estimated e 0,z , e 1,z (for the same
simulation setting). Indeed we show truthful reporting is better (positive gap, increasing with more
accurate estimations of e 0,z , e 1,z ) even when majority of people are wrong (as e 0,z = 0.85, e 1,z =
0.83).
4.4 Truth-telling is a dominant strategy
The above results readily imply that our strategy is asymptotically (in K, N ) dominant. To see this,
suppose both ẽ 1,z , ẽ 0,z have converged to their true values. We define dominant truth serum as
follows:
Mechanism 2 Dominant Truth Serum (DTS)
1. For each task agent i received and reported information, randomly select one of the rest two
reference reports (recall each task is randomly assigned to three agents) as the reference z.
2. When z is uninformative that ẽ 1,z + ẽ 0,z = 1:
• Score agent i zero regardless of his report.
3. When z is informative that ẽ 1,z + ẽ 0,z , 1:
• Score agents via (SSR alpha) defined using z, ẽ 1,z , ẽ 0,z . (with possible scalings)
Note again in above, if the collected reports are in the format of prediction, we will draw a signal
according to the reported prediction - this is the z that we will use.
Theorem 4.4. Asymptotically it is a dominant strategy for agent i to report truthfully under DTS.
Proof. First for the case that ẽ 1,z + ẽ 0,z = 1, it is indifferent for agent i to truthfully report, or
to misreport, or to randomize between the two strategies. Thus truth-telling is a weakly dominant
strategy. When ẽ 1,z +ẽ 0,z , 1, the dominant strategy argument follows from the strictly properness
of (SSR alpha).
Remark 3. Several remarks follow.
• We would like to emphasize again that both z and φ come from D−i : z will be decided by
agents j , i’s reports D−i . φ not only has z as input, but its definition also depends on e 1,z
and e 0,z , which will be learned from D−i .
• When making decisions on reporting, we show under our mechanisms agents can choose to
be oblivious of how much error presents in others’ reports. This removes the practical concern
of implementing a particular Nash Equilibrium.
• Another salient feature of our mechanism is that we have migrated the cognitive load for
having prior knowledge from agents to the mechanism designer. Yet we do not assume the
14
4 DOMINANT TRUTH SERUM
designer has direct knowledge neither; instead we will leverage the power of estimation from
reported data to achieve our goal.
Now we argue the dominant strategy in finite sample regime, under noisy estimations. We first
define the informative region. When e 0,z + e 1,z is arbitrarily close to 1, we have the difficulty in
determining the number of samples needed for the learning process. With this in mind, we will
modify our mechanism to the following:
• Pre-select a small constant 0 < κ < 1 (but close to 0).
• When |ẽ 1,z + ẽ 0,z − 1| ≤ κ (instead of setting it to be exactly 0; this characterizes our
informative region), pay agent i nothing;
• Else follow our original mechanism and pay according to (SSR alpha).
κ can then help us quantify the number of K, N needed. With this we prove the following theorem:
Theorem 4.5. When set κ small enough, and K, N large enough, we have that with finite number
of tasks:
(1) For signal elicitation, it is a dominant strategy for each agent i to report si truthfully under
DTS;
(2) For prediction elicitation, it is a dominant strategy to report pi truthfully when S(·) is strictly
concave in report ai and Lipschitz.
(3) For prediction elicitation, it is an ϵ(K, Nq)-dominant strategy to report pi truthfully for any
log K
1
Lipschitz S(·), where ϵ(K, N ) = O N1 +
K + K is a diminishing term in both K and N .
Above we adopt the classical approximate-dominant strategy definition that truth-telling is ϵdominant strategy if E[φ(Ii , z)] > E[φ(ai , z)] − ϵ, ∀ai , Ii .
The proof for the signal elicitation case is rather straightforward. We need to show that when
the estimation errors of the error rates e 0,z , e 1,z are small enough, the scores under each basic
reporting strategy (truthful, always reverting the observation, always reporting 1 and always reporting 0) will be close to the true score. Then the truthfulness follows from the strict properness
of S. It is slightly more challenging for the prediction elicitation setting (reporting pi s), due to
the continuous reporting space. The main idea is that when the estimation error drops below the
regularization term that corresponds to S’s strict concavity, any possible gain from deviating from
truthful reporting due to the error terms is less than the loss due to the drop in the regularization
term. Note it is not hard to find strictly concave scoring function S - both Brier and logarithmic
scores are strictly concave.
4.5
Weak dominance v.s. strong dominance
We have shown when we cannot learn an informative enough reference report from the rest of
the agents, there is very little we can do regarding that. With this note, our deterministic payment
strategy is only weakly dominant. We argue that if agents’ reports can be modeled as being from
certain distributions, the weak dominance case happens rarely.
Í
Denote by X i = ẽ 1,i +ẽ 0,i , the weakly dominant case happens only when j,i X j = 1. Potentially
there are infinitely many strategies that lead to this state. However, suppose agents’ reporting
strategies are drawn from a distribution defined over a continuous space (mixed strategy space),
Í
Í
then we can easily argue that {ω : j,i X j (ω) = 1} (via taking j,i X j as a random variable) is a
zero measure event.
Also we would like to point out that, as long as each agent believes that this uninformative state
happens with probability < 1, DTS induces strong dominant strategy in truthful reporting, as a
non-trivial mix between weak and strong dominance returns strong dominance.
5.1 Learning the error rate
15
Í
Also, as a matter of fact, j,i X j = 1 corresponds to the case that the reference signal is
stochastically irrelevant to the ground truth. Theoretically speaking, there is little one can do for
eliciting private signals using such a reference signal, via a peer prediction method. In practice,
one can use screening tasks to remove some workers to move away from this uninformative state.
5
CORRELATED SIGNALS
Earlier on, we have assumed that agents’ private signals are conditional independent (on ground
truth). In this section, we discuss the generalization of our method to the case when agents’ observations are correlating with each other. The intuition for why we can hope for this extension is
that we do not explore the correlation structure among signals explicitly. Technically this is due
to our scoring system’s dependence on the reference answer z through e 1,z , e 0,z , which is independent of whether z is drawn from a set of independent signals or a set of correlated ones. Being
correlated or not, as long as we can learn the average error rate from the rest of agents, we will
have an unbiased score for each agent w.r.t. the ground truth signal. To extend our results, the
only challenge we need to address is to accurately estimate Pr[z = 1|y = 0], Pr[z = 0|y = 1].
5.1
Learning the error rate
However, learning Pr[z = 1|y = 0], Pr[z = 0|y = 1] with correlated signals is not as straightforward
as before. The reason is in Eqn. (9, 10) of our proposed method in Section 4.2, the second and thirdorder terms for characterizing the matching probability will not hold. Correlation among signals
will violate the product structure of the matching equations we had earlier. Nonetheless we would
like to stick with our idea of using the matching probabilities among agents to obtain additional
information. We show this is not entirely hopeless:
Lemma 5.1. The second order matching probabilities are the same for the cases with correlated and
independent signals:
Pr[z 1 = 1, z 2 = 1|y = 0] + Pr[z 1 = 0, z 2 = 0|y = 0]
= Pr[z 1 = 1|y = 0] Pr[z 2 = 1|y = 0] + Pr[z 1 = 0|y = 0] Pr[z 2 = 0|y = 0]
Pr[z 1 = 1, z 2 = 1|y = 1] + Pr[z 1 = 0, z 2 = 0|y = 1]
= Pr[z 1 = 1|y = 1] Pr[z 2 = 1|y = 1] + Pr[z 1 = 0|y = 1] Pr[z 2 = 0|y = 1]
With above lemma, we know the correlation structure does not affect estimating the second
order matching probability. Similar to what we have detailed earlier that, the first order and second
order equation together will help us reduce the possible solutions to at most two pairs. Particularly
we can do the following:
1. Estimate q̃i using {ỹr idx (k) (k), k = 1, 2, ..., K, idx = 1, 2, 3}:
ÍK
1(ỹr 1 (k) (k) = ỹr 2 (k) (k))
.
q̃ −i = k=1
K
i.e., q̃ −i is the empirical matching probability (on both 0 & 1).
2. Solve the following set of equations.
2
2
(I): P0 [e 0,z
+ (1 − e 0,z )2 ] + P1 [e 1,z
+ (1 − e 1,z )2 ] = q̃ −i ,
(II): P0e 0,z + P1 (1 − e 1,z ) = P̃1, −i .
First notice that Eqn. (I) doesn’t contain additional information than
2
P0e 0,z
+ P1 (1 − e 1,z )2 = q 2, −i (please refer to Eqn. (9) in previous estimation)
16
5 CORRELATED SIGNALS
, in light of P0e 0,z + P1 (1 − e 1,z ) as:
2
2
2
P0 [p0,z
+ (1 − p0,z )2 ] + P1 [p1,z
+ (1 − p1,z )2 ] = 2[P0p0,z
+ P1 (1 − p1,z )2 ]
− 2[P0p0,z + P1 (1 − p1,z )] + 1
So they necessarily lead to the same set of solutions, and thus Lemma 4.2 will hold too.
We need a third equation for picking the exact solution. The idea is to similarly check one order
higher matching, but we will use Pr[z 1 = z 2 , z 3 = 1] instead of Pr[z 1 = z 2 = z 3 = 1]. To let our
presentation stay focused, we make the following assumption:
Assumption 1. With three drawn reference signals z 1 , z 2 , z 3 from j , i, we have Pr[z 1 , z 2 , z 3 |y] =
Pr[z 1 , z 2 |y] · Pr[z 3 |y].
That is the conditional dependency is up to order two. This is not entirely unreasonable to assume: when signals’ correlations are due to a common set of observed attributes (e.g., essay length
to essay quality), we can assume that one another signal z 2 already encodes sufficient correlational
information with z 1 , and thus the conditional independence between (z 1 , z 2 ) and z 3 . We then note
the following fact: adopting the short-hand notation x 1 := e 0,z , x 2 := 1 − e 1,z we show
Pr[z 1 = z 2 , z 3 = 1] = P0 Pr[z 1 = z 2 , z 3 = 1|y = 0] + P1 Pr[z 1 = z 2 , z 3 = 1|y = 1]
= P0 Pr[z 1 = z 2 |y = 0] Pr[z 3 = 1|y = 0] + P1 Pr[z 1 = z 2 |y = 1] Pr[z 3 = 1|y = 1]
= P0 (x 12 + (1 − x 1 )2 )x 1 + P1 (x 22 + (1 − x 2 )2 )x 2 .
The first equation is based on the Bayesian structure of ground truth signal. The second is due
to Assumption 1. The third equation is due to Lemma 5.1. The following Lemma can then be
established:
Lemma 5.2.
x1 + x2 =
1
Pr[z 1 = z 2 , z 3 = 1] − Pr[z 1 = z 2 ] Pr[z 1 = 1]
+
.
2P0 P1 (x 1 − x 2 )2
P0 P1 (x 1 − x 2 )2
Similarly with knowing |x 1 − x 2 | (jointly determined by the first order and second order equations, as detailed in Lemma 4.2), we are able to formulate another equation regarding x 1 , x 2 .
All three parameters Pr[z 1 = z 2 , z 3 = 1], Pr[z 1 = z 2 ], Pr[z 1 = 1] are estimable from agents’
reports. Together with P0x 1 + P1x 2 = P1, −i , we are able to solve for x 1 , x 2 . The rest of arguments
follow the proof of Theorem 4.1.
5.2
Application: removing curse of weak signals
Extending our results to the correlated signal case not only removes one technical assumption,
but also helps remove a curse for peer prediction introduced by the existence of weak but strongly
correlated signals. Consider the case where besides the quality signal (i.e., the true signal we really
want to elicit) si , each agent i also observes a cheap signal, denoting as sil . Such cheap signals often
exist in practice. For example in peer grading, the length of an assignment is one of such signals.
Cheap signals are often less accurate than the quality signals in that
Pr[sil = 1|y = 0] > Pr[si = 1|y = 0], Pr[sil = 0|y = 1] > Pr[si = 0|y = 1],
and further such weak signals do not always satisfy conditional independence in that Pr[sil , s lj |y] ,
Pr[sil |y] Pr[s lj |y]. Nonetheless it is also often true that weak signals correlate with each other more
than the quality signals: Pr[sil , s lj |y] > Pr[si , s j |y].
The existence of weak signals has been observed to be harmful in that reporting such weak yet
more correlated signals may lead to higher utilities for agents [Gao et al., 2016], as peer prediction
17
essentially checks the correlation among signals. Suppose si and sil are conditionally independent
given y, formally we prove the following results.
Theorem 5.3. When S is taken as the 1/Prior scoring rule, reporting si (instead of the weak signal
sil ) is a dominant strategy with DTS.
6
DISCUSSIONS AND CONCLUDING REMARKS
Recovering the ground truth. Though being a fundamental step, truthful elicitation is also often
not the ultimate goal, as the mechanism designer may want to infer the ground truth answer in
the end. As a by-product, our framework for inferring the error rates from agents would allow us
to do so. The reasoning is as follows: first ask each agent j , i to report a label. If the ground
truth label is 0, the posterior distribution of agents’ reports being 1 will converge to e 0,z with high
probability; while on the other hand, if the ground truth is 1, this posterior distribution converges
to 1 − e 1,z . When e 0,z , 1 − e 1,z (reference answer being informative), we will know the ground
truth label with high probability via posterior distribution testing.
Unbiased estimator of the true score. Our paper proposes a novel and robust way for quantifying
the quality of information when there is no ground truth verification, using peer evaluation, as well
as an estimation procedure for learning the bias in peer evaluation. This may find applications in a
sequential elicitation setting, where such scores can provide informative feedbacks to contributed
information before the outcome reveals.
Effort exertion. So far our discussion focuses on eliciting truthful reports, our results extend
to the scenario for eliciting high quality data. For instance, when workers can choose to exert
costly effort (consider a binary effort level case, and denote the cost as c for exerting efforts) to
improve the quality of their answers, and once the reference signal is informative (either positively
or negatively), we can scale up the scoring functions to cover the cost c, as similarly done in
[Miller et al., 2005], to establish the dominance of reporting a high-effort signal.
Heterogeneous error rates. In order to estimate error rates correctly, we needed to make the assumption that agents have homogeneous error rates over multiple tasks. We now discuss the
applicability of our method in light of the heterogeneity issue. First we would like to emphasize that in practice, for each set of experiments we can choose to group tasks according to their
types (e.g., image labeling, solving puzzles, objects recognition), and run our mechanism over each
group separately. For this setting, we can assume the homogeneity more comfortably. When it is
not quite possible to group tasks together, suppose that each human agent’s error rates are also
task contingent in that for each possible task, we show our solution can be viewed as a linear
approximation for this task contingent case. We have more details on this matter in the Appendix.
A Machine Learning method. To echo recent works on using machine learning techniques [Liu and Chen,
2017] to learn a machine learning model to generate a reference answer, instead soliciting from
other agents, we show this idea is also ready to be plugged into our SSR solution framework. Consider the current task that needs to be elicited and denote its feature vector as x ∈ Rd . Suppose
we have learned (following the results in [Liu and Chen, 2017], such a classifier is learnable purely
from agent’s reported noisy data) a good classifier f˜ ∗ (x) for predicting its true outcome. Replace
the reference answer z with f˜ ∗ (x) and plug in its error rates. The rest of job is to reason about
e 1,z , e 0,z (for f˜ ∗ (x)). There are possibly many different ways to do so. We sketch its possibility
when there is again no ground truth label being available in Appendix.
18
6 DISCUSSIONS AND CONCLUDING REMARKS
Conclusion. In this paper we propose surrogate scoring rules (SSR), which complement the literature of strictly proper scoring rules by considering the setting when there is only access to a
noisy copy of the ground truth. We further extend SSR to the peer prediction setting where the
noisy ground truth can only be inferred via collecting reports from other reference agents. This
returns us a dominant truth serum (DTS), which achieves truthful elicitation in dominant strategy.
REFERENCES
Dana Angluin and Philip Laird. 1988. Learning from noisy examples. Machine Learning 2, 4 (1988), 343–370.
Glenn W Brier. 1950. Verification of forecasts expressed in terms of probability. Monthey Weather Review 78, 1 (1950), 1–3.
Tom Bylander. 1994. Learning linear threshold functions in the presence of classification noise. In Proceedings of the seventh
annual conference on Computational learning theory. ACM, 340–347.
Anirban Dasgupta and Arpita Ghosh. 2013. Crowdsourced judgement elicitation with endogenous proficiency. In Proceedings of the 22nd international conference on World Wide Web. International World Wide Web Conferences Steering
Committee, 319–330.
Rafael Frongillo and Ian A Kash. 2015. Vector-valued property elicitation. In Conference on Learning Theory. 710–727.
Alice Gao, James R Wright, and Kevin Leyton-Brown. 2016. Incentivizing evaluation via limited access to ground truth:
Peer-prediction makes things worse. arXiv preprint arXiv:1606.07042 (2016).
Xi Alice Gao, Andrew Mao, Yiling Chen, and Ryan Prescott Adams. 2014. Trick or treat: putting peer prediction to the test.
In Proceedings of the fifteenth ACM conference on Economics and computation. ACM, 507–524.
Tilmann Gneiting and Adrian E. Raftery. 2007. Strictly Proper Scoring Rules, Prediction, and Estimation. J. Amer. Statist.
Assoc. 102, 477 (2007), 359–378.
Victor Richmond Jose, Robert F. Nau, and Robert L. Winkler. 2006. Scoring Rules, Generalized Entropy and utility maximization. (2006). Working Paper, Fuqua School of Business, Duke University.
R. Jurca and B. Faltings. 2006. Minimum payments that reward honest reputation feedback. In Proceedings of the 7th ACM
conference on Electronic commerce (EC ’06). ACM, 190–199.
Yuqing Kong and Grant Schoenebeck. 2016a. Equilibrium selection in information elicitation without verification via
information monotonicity. arXiv preprint arXiv:1603.07751 (2016).
Yuqing Kong and Grant Schoenebeck. 2016b. A Framework For Designing Information Elicitation Mechanisms That Reward
Truth-telling. arXiv preprint arXiv:1605.01021 (2016).
N.S. Lambert, D.M. Pennock, and Y. Shoham. 2008. Eliciting properties of probability distributions. In Proceedings of the
9th ACM Conference on Electronic Commerce (EC ’08). ACM, 129–138.
Yang Liu and Yiling Chen. 2017. Machine Learning aided Peer Prediction. ACM EC (June 2017).
Yang Liu and Mingyan Liu. 2015. An Online Learning Approach to Improving the Quality of Crowd-Sourcing. In Proceedings
of the 2015 ACM SIGMETRICS International Conference on Measurement and Modeling of Computer Systems (SIGMETRICS
’15). ACM, New York, NY, USA, 217–230. DOI:http://dx.doi.org/10.1145/2745844.2745874
James E Matheson and Robert L Winkler. 1976a. Scoring rules for continuous probability distributions. Management science
22, 10 (1976), 1087–1096.
James E. Matheson and Robert L. Winkler. 1976b. Scoring Rules for Continuous Probability Distributions. Management
Science 22, 10 (1976), 1087–1096.
Aditya Menon, Brendan Van Rooyen, Cheng Soon Ong, and Bob Williamson. 2015. Learning from corrupted binary labels
via class-probability estimation. In International Conference on Machine Learning. 125–134.
Nolan Miller, Paul Resnick, and Richard Zeckhauser. 2005. Eliciting Informative Feedback: The Peer-Prediction Method.
Management Science 51, 9 (2005), 1359 –1373.
Allan H Murphy and Robert L Winkler. 1970. Scoring rules in probability assessment and evaluation. Acta psychologica 34
(1970), 273–286.
Nagarajan Natarajan, Inderjit S Dhillon, Pradeep K Ravikumar, and Ambuj Tewari. 2013. Learning with noisy labels. In
Advances in neural information processing systems. 1196–1204.
Dražen Prelec. 2004. A Bayesian Truth Serum for Subjective Data. Science 306, 5695 (2004), 462–466.
G. Radanovic and B. Faltings. 2013. A Robust Bayesian Truth Serum for Non-Binary Signals. In Proceedings of the 27th
AAAI Conference on Artificial Intelligence (AAAI ’13).
Leonard J Savage. 1971. Elicitation of personal probabilities and expectations. J. Amer. Statist. Assoc. 66, 336 (1971), 783–801.
Clayton Scott. 2015. A Rate of Convergence for Mixture Proportion Estimation, with Application to Learning from Noisy
Labels.. In AISTATS.
Clayton Scott, Gilles Blanchard, Gregory Handy, Sara Pozzi, and Marek Flaska. 2013. Classification with Asymmetric Label
Noise: Consistency and Maximal Denoising.. In COLT. 489–511.
19
V. Shnayder, A. Agarwal, R. Frongillo, and D. C. Parkes. 2016. Informed Truthfulness in Multi-Task Peer Prediction. ACM
EC (March 2016). arXiv:cs.GT/1603.03151
Brendan van Rooyen and Robert C Williamson. 2015. Learning in the Presence of Corruption. arXiv preprint
arXiv:1504.00091 (2015).
Robert L. Winkler. 1969. Scoring rules and the evaluation of probability assessors. J. Amer. Statist. Assoc. 64, 327 (1969),
1073–1078.
Jens Witkowski, Pavel Atanasov, Lyle H Ungar, and Andreas Krause. 2017. Proper proxy scoring rules.
Jens Witkowski, Yoram Bachrach, Peter Key, and David C. Parkes. 2013. Dwelling on the Negative: Incentivizing Effort in
Peer Prediction. In Proceedings of the 1st AAAI Conference on Human Computation and Crowdsourcing (HCOMP’13).
J. Witkowski and D. Parkes. 2012. A Robust Bayesian Truth Serum for Small Populations. In Proceedings of the 26th AAAI
Conference on Artificial Intelligence (AAAI ’12).
20
6 DISCUSSIONS AND CONCLUDING REMARKS
APPENDIX
PROOF OF LEMMA 2.2
Proof. Suppose not we will have
Pr[y = 0|si = 0] = Pr[y = 0|si = 1],
(13)
Pr[y = 1|si = 0] = Pr[y = 1|si = 1].
(14)
From Eqn. (13) we know that
Pr[y = 0, si = 1] Pr[y = 0, si = 0]
=
Pr[si = 1]
Pr[si = 0]
Pr[y = 0](1 − e 0,i )
Pr[y = 0]p0,i
=
,
⇔
Pr[si = 1]
Pr[si = 0]
when Pr[y = 0] , 0 we know that
e 0,i
Pr[si = 1]
.
=
Pr[si = 0] 1 − e 0,i
(15)
Pr[si = 1] 1 − e 1,i
.
=
Pr[si = 0]
e 1,i
(16)
Similarly from Eqn. (14) we know
Therefore we know
e 0, i
1−e 0, i
=
1−e 1, i
e 1, i
from which we know e 0,i + e 1,i = 1. Contradiction.
PROOF OF LEMMA 3.1
Proof. This can be fairly easily established by noting the following: by truthful reporting agent
i’s expected score is
Pr[y = si |si ] Pr[y = si ]
>
= 1.
E[S(si , y)|si ] =
Pr[y = si ]
Pr[y = si ]
On the other hand, by reverting the observation, the agent has
E[S(1 − si , y)|si ] =
Pr[y = 1 − si |si ] Pr[y = 1 − si ]
<
= 1.
Pr[y = 1 − si ]
Pr[y = 1 − si ]
Both of above inequalities are due to Bayesian informativeness (and correspondingly negative
Bayesian informativeness).
PROOF OF THEOREM 4.3
Proof. Estimation error due to heterogeneous agents: The first challenge lies in the fact that the
higher order equations doesn’t capture the true matching probability with heterogeneous workers.
If we assign one task (to be labeled) to two different workers, due to the asymmetric error rate,
the LHS of Eqn. (I) is not precise– drawing a worker without replacement leads to a different
labeling accuracy, and will complicate the solution for the system of equations. We show that our
estimation, though being ignoring above bias, will not affect our results by too much: denoting by
e 1 the accuracy of first drawn sample (for checking the matching) when y = 1 (we can similarly
argue for y = 0). Then we have E[e 1 ] = e 1,z . And the average accuracy of the second drawn
21
sample e 2 (conditional on e 1 ) is given by
following:
e 1, z (N −1)−e 1
.
N −2
Further the matching probability satisfies the
E[e 1 · e 2 ] = E E[e 1 · e 2 |e 1 ]
e 1,z (N − 1) − e 1
=E e 1 ·
N −2
e 1,z (N − 1) E[e 12 ]
=E[e 1 ] ·
−
N −2
N −2
E[e 12 ]
N −1 2
e −
=
N − 2 1,z N − 2
Note both e 1,z and E[e 12 ] are no more than 1. Then
|
2 |
|e 1,z
E[e 12 ]
E[e 12 ]
N −1
2
2
e 1,z −
− e 1,z
|≤
+
≤
.
N −2
N −2
N −2 N −2 N −2
That the estimated quantity is bounded away from the true parameter by at most Θ(1/N ). Similarly
for the third order equation we have
3
|E[e 1 · e 2 · e 3 ] − e 1,z
|≤
3
.
N −3
Estimation errors due to finite estimation samples: The second sources of errors come from the
estimation errors of q̃ 2, −i , q̃ 3, −i and P̃1, −i . We have the following lemma:
Lemma 6.1. When there are K samples for estimating q̃ 2, −i , q̃ 3, −i and P̃1, −i respectively (total budgetlog 6/δ
ing 3K) and K ≥ 2ϵ 2 for any ϵ, δ > 0, we have with probability at least 1 − δ
|q̃ 2, −i − q 2, −i | ≤ ϵ +
3
2
, |q̃ 3, −i − q 3, −i | ≤ ϵ +
, |P̃1, −i − P1, −i | ≤ ϵ.
N −2
N −3
The above lemma can be easily established using Chernoff bound, and our arguments in estimating with heterogeneous agents above. We now bound the error in estimating p0,z , p1,z , under
the (ϵ, δ )-event proved in Lemma 6.1. First of all
|e˜0,z
q̃ 3, −i − q̃ 2, −i P̃1, −i
q 3, −i − q 2, −i P1
1
1
P1 − P̃1, −i −
− e 0,z | = |
P1 − P1, −i
P1 − P0 q̃ 2, −i − (P̃1, −i )2
P1 − P0
q 2, −i − P12
q̃ 3, −i − q̃ 2, −i P̃1, −i q 3, −i − q 2, −i P1
1
|
|P
+
|
P̃
−
P
|
≤
−
1
1, −i
1, −i .
P1 − P0 q̃ 2, −i − (P̃1, −i )2
q 2, −i − P12
According to Lemma 7 of [Liu and Liu, 2015], which we reproduce as follows:
Lemma 6.2. For k ≥ 1 and two sequences {li }im=1 and {qi }im=1 and 0 ≤ li , qi ≤ 1, ∀i = 1, ..., k., we
have
|
m
Ö
i =1
li −
m
Ö
j=1
qj | ≤
m
Õ
i =1
|li − qi | .
(17)
22
6 DISCUSSIONS AND CONCLUDING REMARKS
We know the following facts
2
N −2
− q 2, −i P1, −i |
|q̃ 2, −i − (P̃1, −i )2 ) − (q 2, −i − P1,2 −i )| ≤ |(P̃1, −i )2 − P1,2 −i | + |q̃ 2, −i − q 2, −i | ≤ 3ϵ +
|(q̃ 3, −i − q̃ 2, −i P̃1, −i ) − (q 3, −i − q 2, −i P1, −i )| ≤ |q̃ 3, −i − q 3, −i | + |q̃ 2, −i P̃1, −i
2
3
≤ 3ϵ +
+
.
N −2 N −3
First we prove that
q 2, −i − P1,2 −i =P0x 12 + P1x 22 − (P0x 1 + P1x 2 )2 = P0 P1 (x 1 − x 2 )2 ≥ P0 P1κ 2 .
Let 3ϵ +
2
N −2
|
≤ P0 P1κ 2 /2, we know that (using mean-value theorem/inequality)
q̃ 3, −i − q̃ 2, −i P̃1, −i
q̃ 2, −i − (P̃1, −i
)2
−
q 3, −i − q 2, −i P1, −i
q̃ 3, −i − q̃ 2, −i P̃1, −i q 3, −i − q 2, −i P1, −i
−
|
|≤|
2
q 2, −i − P1
q̃ 2, −i − (P̃1, −i )2
q̃ 2, −i − (P̃1, −i )2
q 3, −i − q 2, −i P1, −i q 3, −i − q 2, −i P1, −i
|
+|
−
q 2, −i − P1,2 −i
q̃ 2, −i − (P̃1, −i )2
≤2
ϵ+
2
3
N −2 + N −3
κ 2 P02 P12
+2
2
N −2
.
(κ 2 P02 P12 )2
ϵ+
Together we proved that
|e˜0,z
ϵ + N2−2 + N3−3
ϵ + N2−2
1
2P1
+ 2P1 2 2 2 2 + ϵ .
− e 0,z | ≤
P1 − P0
κ 2 P02 P12
(κ P0 P1 )
Similarly we are able to work out sensitivity analysis for e 1,z that
ϵ + N2−2
ϵ + N2−2 + N3−3
1
|e˜1,z − e 1,z | ≤
+ 2P0 2 2 2 2 + ϵ .
2P0
P1 − P0
κ 2 P02 P12
(κ P0 P1 )
q
log K
. With the error rates bounds for
Summarizing and set δ = O( K1 ) we have ϵ = O N1 +
K
e 0,z , e 1,z , from Lemma 5.4 [Liu and Chen, 2017], which we reproduce in our contexts as follows:
Lemma 6.3. When K, N are large enough s.t. ϵ ≤ (1 −e 0,z −e 1,z )/4 := ∆e , with probability at least
e
1 − 2δ , |φ̃(t, y) − φ(t, y)| ≤ 1+∆
· ϵ.
2∆2
e
We then obtain a bound on estimating φ(·).
PROOF OF THEOREM 4.5
Proof. Suppose |e 0,z + e 1,z − 1| ≥ 2κ. Then when the estimation errors are small enough such
that |ẽ 0,z − e 0,z | + |ẽ 1,z − e 1,z | ≤ κ, we know that |ẽ 0,z + ẽ 1,z − 1| ≥ κ. Denote by
∆φ = E[S(si , y)] − max E[S(ai , y)],
ai ,si
the minimum gap between the scores for truthfully reporting and mis-reporting. We can easily
show that ∆φ > 0 for signal elicitation case, as we can take ai to be the three basis of mis-reporting:
always reverting the observation, always reporting 1 and always reporting 0. Then with a noisy
estimation of φ(·), we have (using Theorem 4.3)
|E[φ̃(ai , z)] − E[φ(ai , z)]| ≤ ϵ1 + δ 1 · max φ̃, ∀ai .
23
Above, we implicitly assumed the boundedness of φ̃(·): notice with bounded scoring function
S
. Since E[φ(si , z)] = E[S(si , y)], choose ϵ1 , δ 1 such that
S, indeed we know that max φ̃ ≤ 2 max
κ
ϵ1 + δ 1 · max φ̃ < ∆φ /2 we will have (for ai , si )
E[φ̃(si , z)] > E[S(si , y)] − ∆φ /2
> E[S(ai , y)] + ∆φ /2 > E[φ̃(ai , z)].
i.e., the strict properness will preserve, under the noisy estimations. From above results we also
observe that a larger ∆φ will allow more noisy estimations. We thus can trade more payment with
sample complexity, via designing the strictly proper scoring functions to increase ∆φ .
Now consider the prediction elicitation case. Suppose that S(p, y) is strictly concave w.r.t. p ∀y
with parameter λ.
˜ i , y) = S(ai , y) + ϵ(ai , y) ,
S(a
where ϵ(ai , y) indicates the error term. First notice that
E[S(ai , y)] − E[S(ai′ , y)] ≥ λ|ai − ai′ |.
Further we notice that
˜ i , y) − S(ai , y)
ϵ(ai , y) = S(a
1 − ẽ 1−y,z
1 − e 1−y,z
−
)S(ai , y)
1 − ẽ 1,z − ẽ 0,z 1 − e 1,z − e 0,z
ey,z
ey,z
−
)S(ai , −y).
−(
1 − e 1,z − e 0,z 1 − e 1,z − e 0,z
Due to the sample complexity results we know that with probability at least 1 − δ that
1 − ẽ 1−y,z
1 − e 1−y,z
|
−
| ≤ ϵ1 ,
1 − ẽ 1,z − ẽ 0,z 1 − e 1,z − e 0,z
ey,z
ey,z
−
| ≤ ϵ1 .
|
1 − e 1,z − e 0,z 1 − e 1,z − e 0,z
=(
Suppose S(p, y) is also Lipschitz w.r.t. p ∀y with parameter L. By Lipschitz conditions we know
that
δ
|E[ϵ(ai , y)] − E[ϵ(ai , y)]| ≤ ϵ1 L · |ai − ai′ | + 2 L · |ai − ai′ |.
κ
Therefore when ϵ1 , δ are small enough such that
δ
ϵ1 L + 2 L < λ,
κ
no deviation is profitable.
q
log K
1
The ϵ(K, N ) = ϵ1 L + κδ2 L = O N1 +
K + K -dominant strategy argument follows naturally
from above error term analysis. We will not repeat the details.
HETEROGENEOUS ERROR RATES ACROSS TASKS
In order to estimate error rates correctly, we needed to make the assumption that human workers
have homogeneous error rates across multiple tasks. We now discuss the applicability of our
method in light of the heterogeneity issue. Before starting, we would like to emphasize that in
practice, for each set of experiments we can choose to group tasks according to their types (e.g.,
image labeling, solving puzzles, objects recognition), and run our mechanism over each group
separately. For this setting, we can assume the homogeneity more comfortably.
24
6 DISCUSSIONS AND CONCLUDING REMARKS
Nonetheless when it is not quite possible to group tasks together, suppose that each human
agent’s error rates are also task contingent in that for each possible task x with y ∼ P := {P0 , P1 }
we have e 0,i (y), e 1,i (y). Redefine the error rate of the reference answer as follows:
Í
Í
j,i Ex,y |y=0 [ẽ 0, j (x)]
j,i Ex,y |y=1 [ẽ 1, j (x)]
, e 0,z =
.
e 1,z =
N −1
N −1
Again e 1,z , e 0,z captures the error rate of the reference answer. With knowing e 1,z , e 0,z , the dominant strategy argument in Theorem 3.4 holds.
So far so good. However it becomes less clear in how to estimate e 1,z , e 0,z . If we follow the idea
in Section 4.2, the first order equation holds as before
Í
Í
j,i Ex,y |y=0 [ẽ 0, j (x)]
j,i Ex,y |y=1 [ẽ 1, j (x)]
P0
+ P1 (1 −
) = P1, −i ,
N −1
N −1
i.e., P0e 0,z + P1 (1 − e 1,z ) = P1, −i . However the higher order matching statistics appear to be
very different. For example, the second order matching probability between two agents on label 1
becomes
Í
Í
2
2
j,i Ex,y |y=0 [ẽ 0, j (x)]
j,i Ex,y |y=1 [1 − ẽ 1, j (x)]
P0
+ P1
= P1, −i ,
N −1
N −1
Instead we would like a equation as a function of
Í
2 Í
2
j,i Ex,y |y=0 [ẽ 0, j (x)]
j,i Ex,y |y=1 [1 − ẽ 1, j (x)]
,
N −1
N −1
in order to identify the true error rates on average. To get around of this issue, we can adopt the
following approximation for ẽ 0,2 j (x) using Taylor expansion:
2
2
ẽ 0,2 j (x) ≈ e 0,z
+ 2e 0,z (ẽ 0, j (x) − e 0,z ) = 2e 0,z · ẽ 0, j (x) − e 0,z
.
Then we claim that
Í
j,i
Ex,y |y=0 [ẽ 0,2 j (x)]
N −1
≈
Í
j,i
2 ]
Ex,y |y=0 [2e 0,z · ẽ 0, j (x) − e 0,z
N −1
2
2
= 2e 0,z · e 0,z − e 0,z
= e 0,z
Similarly we have
Í
j,i
Ex,y |y=1 [1 − ẽ 1,2 j (x)]
≈ (1 − e 1,z )2 .
N −1
From above, we see we are able to recover the matching equations defined in our earlier solution
for task homogeneous cases; and it can be viewed as a linear approximation for this task contingent
case.
PROOF OF LEMMA 5.1
Proof. W.l.o.g., consider y = 0. We first derive the following difference term
Pr[z 1 = 1, z 2 = 1|y = 0] − Pr[z 1 = 1|y = 0] Pr[z 2 = 1|y = 0]
= (Pr[z 1 = 1|z 2 = 1, y = 0] − Pr[z 1 = 1|y = 0]) Pr[z 2 = 1|y = 0]
= (Pr[z 1 = 1|z 2 = 1, y = 0] − Pr[z 1 = 1|z 2 = 1, y = 0] Pr[z 2 = 1|y = 0]
− Pr[z 1 = 1|z 2 = 0, y = 0] Pr[z 2 = 0|y = 0]) Pr[z 2 = 1|y = 0]
= Pr[z 2 = 0|y = 0] Pr[z 2 = 1|y = 0](Pr[z 1 = 1|z 2 = 1, y = 0] − Pr[z 1 = 1|z 2 = 0, y = 0]).
25
On the other hand,
Pr[z 1 = 0, z 2 = 0|y = 0] − Pr[z 1 = 0|y = 0] Pr[z 2 = 0|y = 0]
= (Pr[z 1 = 0|z 2 = 0, y = 0] − Pr[z 1 = 0|y = 0]) Pr[z 2 = 0|y = 0]
= (Pr[z 1 = 0|z 2 = 0, y = 0] − Pr[z 1 = 0|z 2 = 0, y = 0] Pr[z 2 = 0|y = 0]
− Pr[z 1 = 0|z 2 = 1, y = 0] Pr[z 2 = 1|y = 0]) Pr[z 2 = 0|y = 0]
= Pr[z 2 = 0|y = 0] Pr[z 2 = 1|y = 0](Pr[z 1 = 0|z 2 = 1, y = 0] − Pr[z 1 = 0|z 2 = 0, y = 0]).
The above two terms add up to 0. The proof for y = 1 is symmetric. Proved.
PROOF OF LEMMA 5.2
Proof.
Pr[z 1 = z 2 , z 3 = 1] − Pr[z 1 = z 2 ] Pr[z 1 = 1] = P0 (x 12 + (1 − x 1 )2 )x 1 + P1 (x 22 + (1 − x 2 )2 )x 2
− P0 (x 12 + (1 − x 1 )2 ) + P1 (x 22 + (1 − x 2 )2 ) (P0x 1 + P1x 2 )
= P0 P1 (x 1 − x 2 )2 (2(x 1 + x 2 ) − 1).
(18)
Similarly with knowing |x 1 −x 2 | (jointly determined by the first order and second order equations,
as detailed in Lemma 4.2), we are able to formulate another equation regarding x 1 , x 2 that
x1 + x2 =
1
Pr[z 1 = z 2 , z 3 = 1] − Pr[z 1 = z 2 ] Pr[z 1 = 1]
+
.
2P0 P1 (x 1 − x 2 )2
P0 P1 (x 1 − x 2 )2
PROOF OF THEOREM 5.3
Proof. We only need to prove for the case when si , sil , there is no incentive for agents to
report sil . With accurately learning e 0,z , e 1,z , we know
E[φ(ai , z)|si , sil ] = E[S(ai , y)|si , sil ] =
Pr[y = ai |si , sil ]
Pr[y = ai ]
Suppose si = 1 − ŝ, sil = ŝ. Next we prove that
Pr[y = 1 − ŝ|si = 1 − ŝ, sil = ŝ] > P1−ŝ
Pr[y = ŝ |si = 1 − ŝ, sil = ŝ] < Pŝ
following which, our claim is clearly true. First
ˆ sil = ŝ] =
Pr[y = ŝ |si = 1 − s,
=
ˆ si = 1 − ŝ, sil = s]
ˆ
Pr[y = s,
Pr[si = 1 − ŝ, sil = ŝ |y = ŝ]Py=ŝ
Pr[si = 1 − ŝ, sil = ŝ]
ˆ
Pr[si = 1 − ŝ, sil = s]
.
26
6 DISCUSSIONS AND CONCLUDING REMARKS
Then
ˆ < Pŝ
Pr[y = ŝ |si = 1 − ŝ, sil = s]
ˆ < Pr[si = 1 − s,
ˆ sil = ŝ]
⇔ Pr[si = 1 − ŝ, sil = ŝ |y = s]
Õ
Pŷ · Pr[si = 1 − ŝ|y = ŷ] Pr[sil = ŝ |y = ŷ]
⇔ Pr[si = 1 − ŝ |y = ŝ] Pr[sil = ŝ|y = ŝ] <
ŷ
⇔ Pr[si = 1 − ŝ |y =
⇔ Pr[si = 1 − ŝ |y =
ŝ](1 − Pr[sil = 1 − ŝ |y = ŝ])
ŝ] < Pr[sil = 1 − ŝ|y = ŝ],
< (1 − Pr[si = 1 − ŝ |y = ŝ]) Pr[sil = 1 − ŝ|y = ŝ]
which holds by the definition of weak signal (a higher error rate).
A MACHINE LEARNING METHOD
Suppose that each task is associated with a corresponding feature vector. Consider the current
task that needs to be elicited and denote its feature vector as x ∈ Rd . Suppose we have learned
(following the results in [Liu and Chen, 2017], such a classifier is learnable purely from agent’s
reported noisy data) a good classifier f˜ ∗ (x) for predicting its true outcome. Our ML-aided SSR
simply works in the following ways: replacing the reference answer z with f˜ ∗ (x) and plug in its
error rates. The rest of job is to reason about e 1,z , e 0,z (for f˜ ∗ (x)). There are possibly many different
ways of doing so. We demonstrate its possibility when there is again no ground truth label being
available with the following simple estimation procedure:
• Assign K tasks x1 , x2 , ..., xK to three randomly drawn peers, denote them as ziref (xk ), i =
1, 2, 3, k = 1, 2, ..., K.
• Estimate Pr[ f˜ ∗ (xi ) = z 1ref (xi ) = 1] and Pr[ f˜ ∗ (xi ) = 1].
• Estimate Pr[z 1ref (xi ) = 1|y = 0] (follow Mechanism 1) using the three reference answers
ziref (xk ), i = 1, 2, 3, k = 1, 2, ..., K.
Then note the following fact:
Pr[ f˜ ∗ (xi ) = z ref (xi ) = 1] = P0 Pr[ f˜ ∗ (xi ) = 1|y = 0] Pr[z ref (xi ) = 1|y = 0]
1
1
+ P1 Pr[ f˜ ∗ (xi ) = 1|y = 1] Pr[z 1ref (xi ) = 1|y = 1]
Together with
Pr[ f˜ ∗ (xi ) = 1] = P0 Pr[ f˜ ∗ (xi ) = 1|y = 0] + P1 Pr[ f˜ ∗ (xi ) = 1|y = 1],
we will be able to solve for e 1,z , e 0,z when we have a good estimates of Pr[ f˜ ∗ (xi ) = z 1ref (xi ) = 1] and
Pr[ f˜ ∗ (xi ) = 1], and that Pr[z 1ref (xi ) = 1|y = 0] , Pr[z 1ref (xi ) = 1|y = 1]. With above argument, we
see that we will be able to achieve dominant strategy scorings without even the need of reassigning
all tasks.
| 2cs.AI
|
Learning from Survey Training Samples:
Rate Bounds for Horvitz-Thompson Risk Minimizers
arXiv:1610.03316v1 [math.ST] 11 Oct 2016
Stéphan Clémençon∗, Patrice bertail†, Guillaume Papa∗
October 12, 2016
Abstract
The generalization ability of minimizers of the empirical risk in the context of binary classification has been investigated under a wide variety of complexity assumptions for the collection
of classifiers over which optimization is performed. In contrast, the vast majority of the works
dedicated to this issue stipulate that the training dataset used to compute the empirical risk
functional is composed of i.i.d. observations and involve sharp control of uniform deviation of
i.i.d. averages from their expectation. Beyond the cases where training data are drawn uniformly without replacement among a large i.i.d. sample or modelled as a realization of a weakly
dependent sequence of r.v.’s, statistical guarantees when the data used to train a classifier are
drawn by means of a more general sampling/survey scheme and exhibit a complex dependence
structure have not been documented in the literature yet. It is the main purpose of this paper
to show that the theory of empirical risk minimization can be extended to situations where statistical learning is based on survey samples and knowledge of the related (first order) inclusion
probabilities. Precisely, we prove that minimizing a (possibly biased) weighted version of the
empirical risk, refered to as the (approximate) Horvitz-Thompson risk (HT risk), over a class
of controlled complexity lead to a rate for the excess risk of the order OP ((κN (log N)/n)1/2 )
with κN = (n/N)/ mini≤N πi , when data are sampled by means of a rejective scheme of (deterministic) size n within a statistical population of cardinality N ≥ n, a generalization of basic
sampling without replacement with unequal probability weights πi > 0. Extension to other
sampling schemes are then established by a coupling argument. Beyond theoretical results,
numerical experiments are displayed in order to show the relevance of HT risk minimization
and that ignoring the sampling scheme used to generate the training dataset may completely
jeopardize the learning procedure.
1
Introduction
Whereas statistical learning techniques crucially exploit data that can serve as examples to train
a decision rule, they may also make use of weights individually assigned to the observations, resulting from survey sampling stratification. Such weights could correspond either to true inclusion
probabilities or else to calibrated or post-stratification weights, minimizing some discrepancy under certain margin constraints for the inclusion probabilities. In the context of statistical inference
based on survey data, the asymptotic properties of specific statistics such as Horvitz-Thompson
∗
LTCI,
Télécom
ParisTech,
Université
Paris-Saclay,
75013,
Paris,
France.
{stephan.clemencon,guillaume.papa}@telecom-paristech.fr
†
Université Paris Ouest-Nanterre-La Défense, MODALX, Nanterre, France. [email protected]
1
estimators (cf [14]), whose computation involves not only the observations but also the weights,
have been widely investigated: in particular, mean estimation and regression have been the subject
of much attention, refer to [13], [20], [3] for instance, and a comprehensive functional limit theory
for distribution function estimation is progressively documented in the statistical literature, see [8],
[7], [21]. At the same time, the last decades have witnessed a rapid development of the field of
machine-learning. Revitalized by different breakout algorithms (e.g. SVM, boosting methods), its
practice is now supported by a sound probabilistic theory based on recent non asymptotic results
in the study of empirical processes, see [17], [5]. However, most papers dedicated to theoretical
results grounding the Empirical Risk Minimization approach (ERM in short), the main paradigm
of statistical learning, assume that the training of a decision rule is based on a dataset formed of
independent replications of a generic random vector Z, a collection of N ≥ 1 i.i.d. observations
Z1 , . . . , ZN namely. In contrast, few results are available in situations where the training dataset is
generated by a more complex sampling scheme. One may refer to [1] for concentration inequalities
permitting to study the generalization ability of empirical risk minimizers when the training data
are obtained by standard sampling without replacement (SWOR in abbreviated form) or to [24] in
the case where the decision rule is learnt from a path of a weakly dependent stochastic process.
It is the goal of this paper to extend the ERM theory to situations where the training dataset is
generated by means of a more general sampling scheme, with possibly unequal probability weights.
We first consider the case of rejective sampling (sometimes refered to as conditional Poisson sampling), an important generalization of basic SWOR. The rate bound results obtained by means
of properties of so-termed negatively associated random variables in this case are next shown to
extend to a class of more general sampling schemes by a coupling argument. In addition, numerical
experiments have been carried out in order to provide empirical evidence of the approach developed.
They show in particular that statistical accuracy of the ERM approach may go down the drain if
the sampling scheme underlying the training dataset is ignored.
The paper is organized as follows. In section 2, the probabilistic framework of the present
study is described at length and basic results of the probabilistic theory of classification are briefly
recalled, together with some important notions of survey theory. The main theoretical results are
stated in section 3, while illustrative numerical experiments are presented in section 4. Certain
proofs are sketched in the Appendix section, whereas additional technical details are deferred to
the Supplementary Material.
2
Background and Preliminaries
As a first go, we start with recalling key concepts pertaining to the theory of empirical risk minimization in binary classification, the flagship problem in statistical learning. A few notions related
to survey theory are next described, which will be involved in the subsequent analysis. Throughout
the article, the indicator function of any event E is denoted by I{E}, the Dirac mass at any point a
by δa , the power set of any set E by P(E), the cardinality of any finite set A by #A.
2.1
Binary Classification - Empirical Risk Minimization Theory
The binary classification problem is considered as a running example all along the paper. Because
it can be easily formulated, it is undeniably the most documented statistical learning problem in
the literature and certain results extend to more general frameworks (e.g. multiclass classification,
2
regression, ranking). Let (Ω, A, P) be a probability space and (X, Y) a random pair defined on
(Ω, A, P), taking its values in some measurable product space X × {−1, +1}, with common distribution P(dx, dy): the r.v. X models some observation, hopefully useful for predicting the binary label
Y. The distribution P can also be described by the pair (F, η) where F(dx) denotes the marginal
distribution of the input variable X and η(x) = P{Y = +1 | X = x}, x ∈ X , is the posterior distribution. The objective is to build, based on the training dataset at disposal, a measurable mapping
g : X 7→ {−1, +1}, called a classifier, with minimum risk:
def
L(g) = P{g(X) 6= Y}.
(1)
It is well-known folklore in the probabilistic theory of pattern recognition that the Bayes classifier
g∗ (x) = 2I{η(x) ≥ 1/2} − 1 is a solution of the risk minimization problem inf g L(g), where the
infimum is taken over the collection of all classifiers defined on the input space X . The minimum
risk is denoted by L∗ = L(g∗ ). Since the distribution P of the data is unknown, one substitutes the
true risk with its empirical estimate
X
bn (g) = 1
L
I{g(Xi ) 6= Yi },
n
n
(2)
i=1
based on a sample (X1 , Y1 ), . . . , (Xn , Yn ) of independent copies of the generic random pair (X, Y).
The true risk minimization is then replaced by the empirical risk minimization
bn (g),
min L
(3)
g∈G
where the minimum is taken over a class G of classifier candidates, supposed rich enough to include
bn
the naive Bayes classifier (or a reasonable approximation of the latter). Considering a solution g
of (3), a major problem in statistical learning theory is to establish upper confidence bounds on
the excess of risk L(b
gn ) − L∗ in absence of any distributional assumptions and taking into account
the complexity of the class G (e.g. described by geometric or combinatorial features such as the
VC dimension)
P and some measure of accuracy of approximation of P by its empirical counterpart
Pn = (1/n) ni=1 δ(Xi ,Yi ) over the class G. Indeed, one typically bounds the excess of risk of the
empirical risk minimizers as follows
∗
∗
b
L(b
gn ) − L ≤ 2 sup |Ln (g) − L(g)| + inf L(g) − L .
g∈G
g∈G
The second term on the right hand side is referred to as the bias and depends on the richness of
the class G, while the first term, called the stochastic error, is controlled by means of results in
empirical process theory, see [5].
Remark 1. (On risk surrogates) Although its study is of major interest from a theoretical
perspective, the problem (3) is generally NP-hard. For this reason, the cost function I{−Yg(X) > 0}
is replaced in practice by a nonnegative convex cost function φ(Yg(X)), turning empirical risk
minimization to a tractable convex optimization problem. Typical choices include the exponential
cost φ(u) = exp(u) used in boosting algorithms, the hinge loss φ(u) = (1 + u)+ in the case of
support vector machines and the logit cost φ(u) = log(1 + exp(u)) for Neural Networks, see [2]
and the references therein. Extension of the results established in the present paper to such risk
surrogates are straightforward and left to the reader.
In this paper, we consider the situation where the training data used to compute of the empirical
risk (2) is not an i.i.d. sample but the product of a more general sampling plan of fixed size n ≥ 1.
3
2.2
Sampling Schemes and Horvitz-Thompson Estimation
Let N ≥ 1. In the standard superpopulation framework we consider, (X1 , Y1 ), . . . , (XN , YN ) is a
sample of independent copies of (X, Y) observed on a finite population IN := {1, . . . , N}. We call a
survey sample of (possibly random) size n ≤ N of the population IN , any subset s := {i1 , . . . , in(s) } ∈
P(IN ) with cardinality n =: n(s) less that N. A sampling design without replacement is determined
by a conditional probability distribution RN on the set of all possible samples s ∈ P(IN ) given
the original data DN = {(Xi , Yi ) : i ∈ IN }. For any i ∈ {1, . . . , N}, the first order inclusion
probability, πi = PRN {i ∈ S} is the probability that the unit i belongs to a random sample S drawn
from the conditional distribution RN . We set π = (π1 , . . . , πN ). The second order inclusion
probabilities are denoted by πi,j = PRN {(i, j) ∈ S2 }, for any i 6= j in {1, . . . , N}2 . The information
related to the observed sample S ⊂ {1, . . . , N} is fully enclosed in the r.v. N = (1 , . . . , N ),
where i = I{i ∈ S} for 1 ≤ i ≤ N. The 1-d marginal conditional distributions of the sampling
scheme N given DN are the Bernoulli distributions B(πi ) = πi δ1 + (1 − πi )δ0 , 1 ≤ i ≤ N, and the
covariance matrix ΓN of the r.v. N has entries given by ΓN (i, j) = πi,j − πi πj ,P
with πi,i = πi by
convention, for 1 ≤ i, j ≤ N. Observe that, equipped with the notations above, 1≤i≤N i = n(S).
One may refer to [10] for accounts of survey sampling techniques. Notice also that, in many
applications, the inclusion probabilities are built using some extra information, typically by means
of auxiliary random variables W1 , . . . , WN defined onP(Ω, A, P) and taking their values in some
measurable
space W: ∀i ∈ {1, . . . , N}, πi = nh(Wi )/ 1≤j≤N h(Wj ), where n max1≤i≤n h(Wi ) ≤
P
h(W
i ) almost-surely and h : W →]0, +∞[ is a measurable link function. The (Xi , Yi , Wi )’s
1≤i≤N
are generally supposed to be i.i.d. copies of a generic r.v. (X, Y, W). See [22] for more details. For
simplicity, the πi ’s are supposed to be deterministic in the subsequent analysis, which boils down
to carrying out the study conditionally upon the Wi ’s in the example aforementioned.
Horvitz-Thompson risk. As defined
in [14], the Horvitz-Thompson version of the (not
bN (g) = N−1 P
available) empirical risk L
1≤i≤N I{Yi 6= g(Xi )} of any classifier candidate g based on
the sampled data {(Xi , Yi ) : i ∈ S} with S ∼ RN is given by:
1 X i
1 X 1
I{g(Xi ) 6= Yi } =
I {g(Xi ) 6= Yi }
N
πi
N
πi
N
LN (g) =
i∈S
(4)
i=1
with the convention that 0/0 = 0 and where the subscript N = (1 , . . . , N ) denotes the vector
in correspondence with the sample S. Observe that, conditionally upon the (Xi , Yi )’s, the quantity
(4), that shall be referred to as the empirical Horvitz-Thompson risk (empirical HT risk in short)
bN (g). Its (pointwise) consisthroughout the paper, is an unbiased estimate of the empirical risk L
tency and asymptotic normality are established in [20] and [3] for a variety of sampling schemes.
This article is devoted to investigating the statistical performance of minimizers ḡN of the HT
risk (4) over the class G under adequate assumptions for the sampling scheme RN used to generate
the training dataset. We point out that such an analysis is far from straightforward due to the
possible depence structure of the terms involved in the summation (4): except in the Poisson case
(recalled below), concentration results for empirical processes cannot be directly applied to control
maximal deviations of the type
sup |LN (g) − L(g)|.
g∈G
Conditional Poisson sampling. One of the simplest sampling plan is undeniably the Poisson
survey scheme (without replacement), a generalization of Bernoulli sampling originally proposed
4
in [12] for the case of unequal weights: the i ’s are independent and the sampling distribution is
thus entirely determined by the first order inclusion probabilities pN = (p1 , . . . , pN ) ∈]0, 1[N :
Y Y
∀s ∈ P(IN ), PN (s) =
pi
(1 − pi ).
(5)
i∈S
i∈S
/
Observe in addition that the behavior of the quantity (4) can be then investigated by means of
results established for sums of independent random variables. However, the major drawback of this
sampling plan lies in the random nature of the corresponding sample size, impacting
P significantly
the variability of (4). The variance of the Poisson P
sample size is given by dN = N
i=1 pi (1 − pi ),
n
while the conditional variance of (4) is in this case: i=1 ((1 − pi )/pi )I{g(Xi ) 6= Yi }. For this reason,
rejective sampling, a sampling design RN of fixed size n ≤ N, is often preferred in practice. It
generalizes the simple random sampling without replacement (where all samples with cardinality n
are equally likely to be chosen, with probability (N − n)!/n!, all the corresponding first and second
order probabilities being thus equal to n/N and n(n − 1)/(N(N − 1)) respectively). Denoting by
πN = (π1 , . . . , πN ) its first order inclusion probabilities and by Sn = {s ∈ P(IN ) : #s = n} the
subset of all possible samples of size n, it is defined by:
Y Y
∀s ∈ Sn , RN (s) = C
pi
(1 − pi ),
(6)
i∈s
P
Q
i∈s
/
Q
N
where C = 1/ s∈Sn i∈s pi i∈s
/ (1 − pi ) and the vector pN = (p
P1 , . . . , pN ) ∈]0, 1[ yields first
order inclusion probabilities equal to the πi ’s and is such that i≤N pi = n. Under this latter
additional condition, such a vector pN exists and is unique (see [11]) and the related representation
(6) is then said to be canonical 1 . Comparing (6) and (5) reveals that rejective RN sampling of fixed
size n can be viewed as Poisson sampling given that the sample size is equal to n. It is for this
reason that rejective sampling is usually referred to as conditional Poisson sampling. One must
pay attention not to get the πi ’s and the pi ’s mixed up: the latter are the first order inclusion
probabilities of PN , whereas the former are those of its conditional version RN . However they can
be related by means of the results stated in [13] (see Theorem 5.1 therein): ∀i ∈ {1, . . . , N},
πi (1 − pi ) = pi (1 − πi ) × (1 − (π̃ − πi )/d∗N + o(1/d∗N ),
(7)
pi (1 − πi ) = πi (1 − pi ) × (1 − (p̃ − pi )/dN + o(1/dN ),
(8)
PN
PN 2
P
2
∗
where d∗N = N
i=1 πi (1 − πi ) and p̃ = (1/dN )
i=1 (pi ) (1 − pi ).
i=1 πi (1 − πi ), π̃ = (1/dN )
More examples of sampling schemes with fixed size are given in the Supplementary Material.
of survey theory.
3
Main Results
We first consider the case where statistical learning is based on the observation of a sample drawn
by means of a rejective scheme. As shall be seen below, the main argument underlying the results
obtained relies on the fact that the related scheme form a collection of negatively associated (binary)
random variables, a rather tractable type of dependence structure. This property being shared by
1
0
Notice that any vector pN
∈]0, 1[N such that pi /(1 − pi ) = cpi0 /(1 − pi0 ) for all i ∈ {1, . . . , n} for some constant
c > 0 can be used to write a representation of RN of the same type as (6)
5
many other sampling schemes of deterministic size, the same argument can be thus naturally applied
to carry out a similar rate analysis for training data produced by such plans. Extensions of these
results to more general sampling schemes are also considered by means of a coupling technique.
3.1
Horvitz-Thompson Empirical Risk Minimization in the Rejective Case
For clarity, we first recall the definition of negatively associated random variables, see [16].
Definition 1. Let Z1 , . . . , Zn be random variables defined on the same probability space, valued
in a measurable space (E, E). They are said to be negatively associated iff for any pair of disjoint
subsets A1 and A2 of the index set {1, . . . , n}
Cov (f((Zi )i∈A1 ), g((Zj )j∈A2 )) ≤ 0,
(9)
for any real valued measurable functions f : E#A1 → R and g : E#A2 → R that are both increasing
in each variable.
The theorem stated below reveals that any rejective scheme N forms a collection of negatively
associated r.v.’s. The proof is given in the Appendix section.
Theorem 1. Let N ≥ 1 and N = (1 , . . . , N ) be the vector of indicator variables related to a
rejective plan on IN . Then, the binary random variables 1 , . . . , N are negatively associated.
The result above permits to handle the dependence of the terms involved in the summation
(4). It is the key argument for proving the following proposition, which extends results for training
datasets generated by basic sampling without replacement (i.e. in the case of all equal weights:
πi = n/N for i = 1, . . . , N), refer to [1] (see also [23]).
Proposition 1. Suppose that the sampling scheme N is rejective with first order inclusion probabilities πN and that the class G is of finite VC dimension V < +∞. Set κN = (n/N)/ mini≤N πi .
Then, the following assertions hold true.
(i) For any δ ∈ (0, 1), with probability larger than 1 − δ, we have: ∀n ≤ N,
s
2
2
bN (g)| ≤ 2κN log( δ ) + V log(N + 1) + 2κN log( δ ) + V log(N + 1) .
sup |L̄N (g) − L
3n
n
g∈G
(10)
(ii) For any solution ḡN of the minimization problem inf g∈G LN (g) is such that, for any δ ∈ (0, 1),
with probability at least 1 − δ, we have: ∀N ≥ 1,
s
log( δ4 ) + V log(N + 1)
log( δ4 ) + V log(N + 1)
L(ḡN ) − L∗ ≤ 2 2κN
+ 4κN
n
3n
s
r
2 log( δ2 )
V
+C
+2
+ inf L(g) − L∗ .
g∈G
N
N
6
The factor κN involved in the bounds
P above reflects the influence of the sampling scheme
(notice incidentally that κN ≥ 1 since i≤N πi = n). In the SWOR case, i.e. when πi = n/N
for all i ∈ {1, . . . , N}, it is then minimum, equal to 1. More generally, when n = o(N) as
N → +∞, as soon as the weights cannot
p vanish faster than n/N, the rate achieved by minimizers
of the HT risk is of the order OP ( (log N)/n). Many sampling schemes (e.g. Rao-Sampford
sampling, Pareto sampling, Srinivasan sampling) of fixed size are actually described by random
vectors N with negatively associated components, see [6] or [18]. Hence, a rapid examination of
Proposition 1’s proof shows that the bounds stated above immediately extend to these cases. See
the Supplementary Material for more details and references. Before showing how the rate bounds
established can be extended to even more general sampling schemes, a few remarks are in order.
Remark 2. (Complexity assumptions) We point out that the results stated can be established,
essentially by means of the same argument as that developed in the Appendix, under complexity
assumptions of different nature, involving metric entropy conditions for instance (see e.g. [26]).
Such straightforward extensions are left to the reader.
Remark 3. (Model Selection) A slight modification of the argument involved in Proposition
1 straightforwardly leads to bounds on the expected excess risk E[L(ḡN )] − inf g∈G L(g). Following the Structural Risk Minimization principle (see [27]), such VC bounds can be next used as
complexity regularization terms to penalize additively the HT risk (4) and, for a sequence of
model classes Gk with k ≥ 1 of finite VC dimension, select the classifier among the minimizers
{arg ming∈Gk L̄N (g), k ≥ 1}, which has approximately minimal risk. Due to space limitations,
details are left to the reader.
Remark 4. (Biased HT risk) As recalled in the Supplementary Material, the canonical parameters pN are practically used to build a rejective sampling scheme N rather than its vector of first
order inclusion probabilities (π1 , . . . , πN ), whose explicit computation based on the pi ’s is a difficult task, refer to [9] for dedicated algorithms. For thisPreason, one could be naturally tempted to
e (g) = (1/N)
minimize the alternative risk estimate L
N
i≤N (i /pi )I{Yi 6= g(Xi )}. As proved in the
Supplementary Material, refinements of Eq. (7)-(8) show that
X 1
1
e (g) − L̄ (g)| ≤ 1
sup |L
≤ 6NκN /(ndN ),
−
N
N
N
pi πi
g∈G
N
(11)
i=1
e (g) from bound (ii) in Proposition1.
one may directly derive a rate bound for solutions of inf g∈G L
N
√
In particular, the learning rate achieved by ḡN is preserved when 1/ n = O(mini≤N πi ) as N, n →
+∞.
3.2
Extensions to More General Sampling Schemes
We now extend the rate bound analysis carried out in the previous subsection to more complex sampling schemes (described by a random vector ∗N possibly exhibiting a very complex
dependence structure). In order to give an insight into the arguments which the extension is
based on, additional notations are required. In this section, we consider a general sampling
design R∗N with first order inclusion probabilities π∗N = (π∗1 , . . . , π∗N ) described by the vector
∗N = (∗1 , . . . , ∗N ) and investigate the performance of minimizers ḡ∗N of the HT empirical risk
P
∗
∗
L̄∗N (g) = (1/N) N
i=1 (i /πi )I{Yi 6= g(Xi )} over a class G. We also consider a rejective sampling
7
scheme RN described by the r.v. N , with first order inclusion probabilities πN = (π1 , . . . , πN )
defined on the same probability space, as well as the following quantity:
1 X i
I{Yi 6= g(Xi )}
N
π∗i
N
ĽN (g) =
(12)
i=1
for any classifier g. Observe that (12) differs from the HT empirical risk L̄N (g) related to the rejective sampling scheme N in the weights it involves, the π∗i ’s instead of the πi ’s namely. Equipped
with this notation, the excess of risk of the HT empirical risk minimizer can be bounded as follows:
bN (g) − L̄ (g)
bN (g) + 2 sup L
L(ḡ∗N ) − inf L(g) ≤ 2 sup L(g) − L
N
g∈G
g∈G
g∈G
+ 2 sup L̄N (g) − ĽN (g) + 2 sup ĽN (g) − L̄∗N (g) .
(13)
g∈G
g∈G
Whereas the first term on the right hand side of (13) can be classically controlled using VapnikChervonenkis and McDiarmid inequalities (see e.g. [27]), assertion (i) of Proposition 1 provides a
control of the second term. Following in the footsteps of [13], the third term shall be bounded by
means of a coupling argument, i.e. a specific choice of the joint distribution of (∗N , N ) satisfying
the distributional margin constraints, while the second term is controlled by assumptions related
to the closeness between the first order inclusion probabilities π∗N and πN . More precisely, the
assumptions required in the subsequent analysis involve the total variation distance between the
sampling plans RN and R∗N :
def
dTV (RN , R∗N ) =
1
2
X
|RN (s) − R∗N (s)|.
s∈P(IN )
Theorem 2. Suppose that Proposition 1’s assumptions are fulfilled. Set κ∗N = (n/N) mini≤N π∗i
and κN = (n/N)/ mini≤N πi . Then, there exists a universal constant C < +∞ such that we have,
∀N ≥ 1,
r
V log(N + 1)
V log(N + 1)
∗
E L(ḡN ) − inf L(g) ≤ 2 2κN
+ 4κN
g∈G
n
3n
r
V
+C
+ 2(κ∗N + κN )(N/n)dTV (RN , R∗N ),
(14)
N
where the infimum is taken over the set of rejective schemes RN with first order inclusion probabilities πN = (π1 , . . . , πN ).
The proof is given in the Supplementary Material. The rate bound obtained depends on the
minimum error made when approximating the sampling plan by a rejective sampling plan in terms
of total variation distance. In practice, following in the footsteps of [13] or [3], it can be controlled
by exhibiting a specific coupling (∗N , N ). One may refer to [3] for many coupling results of this
nature, in particular when the approximating scheme N is of rejective type.
8
4
Illustrative Numerical Experiments
In this section we display numerical experiments to illustrate the relevance of HT risk minimization.
We first consider the case where g(X) = sign(k(X)T θ + b), where k is some mapping function, T
denotes the transposition operator, θ, b are some parameters. As mentionned in 1, we consider the
hinge loss as a convex surrogate of the 0 − 1 loss and add some l2 regularization term. This leads
to the ”Weighted SVM” formulation below:
min
θ,b
1 X 1
max(0, 1 − Yi (k(Xi )T θ − b)) + λkθk2 .
N
πi
i∈S
We use the gaussian r.b.f kernel and perform cross validation to appropriately choose the value
of λ. We then consider the task of learning classification trees using the CART algorithm. These
classifiers are trained using the scikit-learn library [19] and, we account for the randomness of our
experiments by shuffling our datasets and repeating the experiments 50 times.
We first generate a two class dataset D in R10 of size 20000 by sampling independent observations
from two multivariate normal distribution. A similar dataset Dtest of size 2000 is generated to test
our classifiers. Denoting by Id the identity matrix in Rd , the positive class has mean (0, . . . , 0) and
covariance matrix equal to I10 , the negative class has mean (1, . . . , 1) and covariance matrix equal
e of size 1100 via a rejective sampling scheme applied to D.
to 10 × I10 . We then build a dataset D
Observations from the negative class being more noisy we assign them first order probability equal
to 0.1, and assign first order probability equal to 0.01 to observation from the positive class. To
b of size 1100 by sampling without replacement
allow for a fair comparison, we also build a dataset D
e and D,
b and display the results in Table??.
within D. We then learn the different classifiers on D
e
Weighted SVM on D
e
Unweighted SVM on D
b
SVM on D
e
Weighted CART on D
e
Unweighted CART on D
b
CART on D
Mean
Std Deviation
0.02
0.18
0.04
0.06
0.11
0.08
0.005
0.02
0.005
0.01
0.03
0.01
Table 1: Average over 50 runs of the prediction error on Dtest and its standard deviation.
Overall, taking into accounts the inclusion probability allows to consider a training set of reduced size and therefore reduce the computationnal complexity of the learning procedure without
damaging the quality of the prediction . Similar experiments on real datasets are displayed in the
Supplementary Material for which similar conclusions hold.
5
Conclusion
Most theoretical studies providing a statistical explanation for the success of learning algorithms
based on the ERM paradigm fully ignore the possible impact of the sampling scheme producing
9
the training data and stipulate that observations are independent replications of a generic r.v.
or are uniformly sampled without replacement in a larger dataset. Through the generalizable
example of rejective sampling, this paper shows that such studies can be extended to situations
where training data are obtained by more general sampling schemes and possibly exhibit a complex
dependence structure, provided that related probablity weights are appropriately incorporated in
the risk functional.
Appendix
Proof of Theorem 1
Considering the usual representation of the distribution of (1 , . . . , N ) as the conditional distribution of a sample of independent Bernoulli variables (∗1 , . . . , ∗N ) conditioned upon the event
PN ∗
i=1 i = n (see subsection 2.2), the result is a consequence of Theorem 2.8 in [16].
Bernstein inequality for sums of negatively associated random variables
For simplicity, we first establish the following tail bound for negatively associated random variables,
which extends the usual Bernstein inequality in the i.i.d. setting, see [4]. Proofs of Proposition 1
and Theorem 2 are then deduced from Theorem 1 and Theorem 3 (see Supplementary Material) .
Theorem 3. Let Z1 , . . . , ZN be negatively associated real valued random variables such that
|Zi | ≤ c < +∞ a.s. E[Zi ] = 0 and E[Z2i ] 6 σ2i for 1 ≤ i ≤ N. Then, for all t > 0, we have: ∀N ≥ 1,
N
!
X
t2
P
Zi ≥ t ≤ exp − 2
PN 2 .
ct
+
2
i=1 σi
3
i=1
Before detailing
observe that a similar bound holds true for the tail probability
P
the proof,
P
N
N
P
i=1 Zi | ≥ t as well, up to a multiplicative factor 2). Refer also
i=1 Zi ≤ −t (and for P |
to Theorem 4 in [15] for a similar result in a more restrictive setting (i.e. for tail bounds related
to sums of negatively associated r.v.’s).
Proof. The proof starts off with the usual Chernoff method: for all λ > 0,
N
h PN
i
X
P
Zi ≥ t ≤ exp −tλ + log E et i=1 Zi .
(15)
i=1
Next, observe that, for all t > 0, we have
h Pn
i
h
Pn−1 i
E et i=1 Zi = E etZn et i=1 Zi
h
i h Pn−1 i
≤ E etZn E et i=1 Zi
≤
n
Y
h
tZi
E e
i
,
i=1
using the property (9) combined with a descending recurrence on i. The proof is finished by
plugging (16) into (15), using an adequate control of the log-Laplace transform of the Zi ’s and
10
optimizing finally the resulting bound w.r.t. λ > 0, just like in the proof of the classic Bernstein
inequality, see [4].
Supplementary - Proof of Proposition 1
i
1 PN
bN (g) as sup
−
1
I {g(Xi ) 6= Yi } |
We start off by writing S := supg∈G L̄N (g) − L
|
g∈G N
i=1 πi
and apply Theorem 1 conditionnaly upon DN to the r.v Zi := N1 πii − 1 I {g(Xi ) 6= Yi } . Indeed,
N
the (πi )N
i=1 and (I {g(Xi ) 6= Yi })i=1 being positive real numbers, Theorem 1 altogether with [16]
κN
1
1
1
implies that (Zi )N
i=1 are negatively associated. Since |Zi | 6 N max(1, πi − 1) 6 Nπi 6 n and
E[Z2i ] 6
1−pii
N2 πi
6
κN
nN
we have :
N
X
P
Zi ≥ t|DN
i=1
nt2
≤ exp − 2
3 κN t + 2κN
!
.
Applying the same method to the r.v (−Zi )N
i=1 and taking the union bound yields :
!
N
X
nt2
.
P |
Zi | ≥ t|DN ≤ 2 exp − 2
κ
t
+
2κ
N
N
3
i=1
By virtue of Sauer’s lemma, since the class G has finite VC-dimension V, we have by taking expectation w.r.t DN :
!
nt2
V
P {S ≥ t} ≤ 2(N + 1) exp − 2
.
3 κN t + 2κN
nt2
V
so
The high probability bound is then easily deduced by choosing δ = 2(N + 1) exp − 2 κ t+2κ
3 N
N
that :
log( δ2 ) + V log(N + 1)
t−
κN
3n
!
!2
log( δ2 ) + V log(N + 1)
=
κN
3n
2 log( δ2 ) + V log(N + 1)
+
κN
n
leading to the following upperbound :
s
2κN log( δ2 ) + V log(N + 1)
log( δ2 ) + V log(N + 1)
t6
+ 2
κN .
3n
n
The second claim of Proposition 1 is established using
L(ḡN ) − L∗
≤
bN (g)
inf L(g) − L∗ + 2 sup L(g) − L
g∈G
g∈G
bN (g) − L̄ (g) , (16)
+ 2 sup L
N
g∈G
bN (g) and a union
altogether with classical results on ERM applied to the term supg∈G L(g) − L
bound.
11
Supplementary - Proof of Theorem2
Starting from (13), we only have to derive bounds for the quantities S1 := supg∈G L̄N (g) − ĽN (g)
and S2 := supg∈G ĽN (g) − L̄∗N (g) . Starting with the first one, we have :
1 X
S1 = sup |
i
g∈G N
N
i=1
1
1
−
π∗i
πi
I {g(Xi ) 6= Yi } |
1 X
1
1
i | ∗ − |
N
πi
πi
N
6
i=1
so that taking expectation w.r.t N conditionned upon DN yields :
1 X π∗i − πi
|
|
N
πi
N
E[S1 |DN ] 6
i=1
N1 X
|πi − π∗i |
nN
N
6 κN
i=1
N
6 κN dTV (RN , R∗N ),
n
taking expectation w.r.t DN gives an upperbound on S1 . We now turn to the analysis of S2 which
is very similar :
1 X ∗i − i
I {g(Xi ) 6= Yi } |
N
π∗i
N
S2 = sup |
g∈G
6
1
N
N
X
i=1
i=1
|∗i − i |
.
π∗i
We then take expectation conditionned upon DN and easily obtain
E[S2 |DN ] 6 κ∗N
N
dTV (RN , R∗N )
n
which conclude the proof.
Supplementary - On biased HT risk minimization
Eq. (11) directly results from the following lemma.
Lemma 1. We have, for p0i s such Suppose that dN ≥ 1. We have, for all i ∈ {1, . . . , N},
|1/πi − 1/pi | ≤
6
× (1 − πi )/πi .
dN
12
Proof. The proof follows from the representation (5.14) on p1509 in [13]. Denote by PN a Poisson
sampling distribution on IN with inclusion probabilities p1 , . . . , pN , the canonical parameters of
RN . For all i ∈ {1, . . . , N}, we have:
−1
X
πi 1 − pi
P(s)
=
pi 1 − πi
s∈P(IN ): i∈IN \{s}
X
×
P(s)
s∈P(IN ): i∈IN \{s}
=
X
X
1 − ph
j∈s (1 − pj ) + (ph − pi )
P
h∈s
−1
PN (s)
s: i∈IN \{s}
×
X
PN (s)
s: i∈IN \{s}
X
P
h∈s
1 − ph
(1
−
p
)
1+
j
j∈s
P(ph −pi )
j∈s (1−pj )
Now recall that for any x ∈] − 1, 1[, we have:
1−x≤
1
≤ 1 − x + x2 .
1+x
It follows that
πi 1 − pi
≤1−
pi 1 − πi
X
−1
P(s)
s: i∈IN \{s}
X
X (1 − ph )(ph − pi )
×
P(s)
P
2
h∈s
s: i∈IN \{s}
j∈s (1 − pj )
−1
X
X
X (1 − ph )(ph − pi )2
+
P(s)
P(s)
P
3
h∈s
(1
−
p
)
s: i∈IN \{s}
s: i∈IN \{s}
j
j∈s
P
Following now line by line the proof on p. 1510 in [13] and noticing that j∈s (1 − pj ) ≥ 1/2dN
(see Lemma 2.2 in [13]), we have
X (1 − ph )(ph − pi )
1
2
≤
P
2 ≤ P
d
N
h∈s
j∈s (1 − pj )
j∈s (1 − pj )
and similarly
X (1 − ph )(ph − pi )2
1
4
P
3 ≤ P
2 ≤ 2 .
dN
h∈s
j∈s (1 − pj )
j∈s (1 − pj )
This yieds: ∀i ∈ {1, . . . , N},
1−
2
πi 1 − pi
2
4
≤
≤1+
+
dN
pi 1 − πi
dN d2N
13
and
pi (1 − πi )(1 −
2
2
4
) ≤ πi (1 − pi ) ≤ pi (1 − πi )(1 +
+ 2 ),
dN
dN dN
leading then to
−
2
2
4
(1 − πi )pi ≤ πi − pi ≤ pi (1 − πi )(
+
)
dN
dN d2N
and finally to
−
(1 − πi ) 2
1
1
(1 − πi ) 2
4
≤
−
≤
(
+ 2 ).
πi
dN
pi πi
πi
dN dN
Since 1/d2N ≤ 1/dN as soon as dN ≥ 1, the lemma is proved.
Supplementary - Further details on the rejective scheme
Let n ≤ N and consider a vector π = (π1 , . . . , πN ) of first order inclusion probabilities. Further
define Sn := {s ∈ P(IN ) : #s = n}, the set of all samples in population IN with cardinality n.
The rejective sampling [13, 3], sometimes called conditional Poisson sampling, exponential design
without replacement or maximum entropy design, is the sampling design RP
N that selects samples
of fixed size n(s) = n so as to maximize the entropy measure H(RN ) = − s∈Sn RN (s) log RN (s),
subject to the constraint that its vector of first order inclusion probabilities coincides with π. It is
easily implemented in two steps:
1. Draw a sample S according to a Poisson plan PN , with properly chosen firstPorder inclusion
probabilities pN = (p1 , . . . , pN ). The representation is called canonical if N
i=1 pi = n. In
that case, relationships between each pi and πi , 1 ≤ i ≤ N, are established in [13].
2. If n(S) 6= n, then reject sample S and go back to step one, otherwise stop.
Vector p must be chosen in a way that the resulting first order inclusion probabilities coincide with
π, by means of a dedicated optimization algorithm [25]. The corresponding probability distribution
Q
Q
PN (s) I{#s=n}
is given for all s ∈ P(IN ) by RN (s) = P
pi i∈s
0) ∝
i∈s
/ (1 − pi ) × I{#s = n}, where ∝
P
(s
0
N
s ∈Sn
denotes the proportionality.
Supplementary - Stratified sampling
A stratified sampling design permits to draw a sample S of fixed size n(S) = n ≤ N within a
population IN that can be partitioned into K ≥ 1 distinct strata IN1 , . . . , INK (known a priori)
of respective sizes N1 , . . . , NK adding up to N. Let n1 , . . . , nK be non-negative integers such that
n1 + · · · + nK = n, then the drawing procedure is implemented in K steps: within each stratum
INk , k ∈ {1, . . . , K}, perform a SWOR of size nk ≤ Nk yielding a sample Sk . The final sample
S
is obtained by assembling these sub-samples: S = Kk=1 Sk . The probability of drawing a specific
−1
PK
Nk
str
sample s by means of this survey design is RN (s) = k=1
. Naturally, first and second
nk
order inclusion
depend on the stratum
which each unit belong: for all i 6= j in UN ,
PK to
PK probabilities
nk (nk −1)
nk
str
2 }.
str
πi (RN ) = k=1 Nk I{i ∈ UNk } and πi,j (RN ) = k=1 Nk (Nk −1) I{(i, j) ∈ UN
k
14
Supplementary - Rao-Sampford sampling
The Rao-Sampford sampling design generates samples s ∈ P(IN ) of fixed size n(s) = n with
RS
respect to some given first order inclusion probabilities πRS := (πRS
1 , . . . , πN ), fulfilling the condition
PN RS
i=1 πi = n, with probability
RRS
N (s) = η
X
πRS
i
i∈s
Y
πRS
j
j∈s
/
1 − πRS
j
.
P
Here, η > 0 is chosen such that s∈P(IN ) RRS
N (s) = 1. In practice, the following algorithm is often
used to implement such a design [3]:
1. select the first unit i with probability πRS
i /n,
RS
2. select the remaining n − 1 units j with drawing probabilities proportional to πRS
j /(1 − πj ),
j = 1, . . . , N,
3. accept the sample if the units drawn are all distinct, otherwise reject it and go back to step
one.
Additional experiments
We consider the following datasets which were obtained via a stratified sampling design. We point
out that this sampling scheme involves negatively associated (binary) random variables so that the
theoretical results obtained in the rejective case extend to this scheme.
incaIndiv
GJB
privacy3
privacy4
N
Number of features
4079
2001
316
301
326
130
95
124
The dataset incaIndiv 2 contain informations on the food consumption of the french population.
The dataset GJB 3 contains questions about job seeking and the internet, workforce automation,
online dating and smartphone use among Americans. The datasets privacy3 4 and privacy4 5 contain
questions about privacy and information sharing. On the datasets incaIndiv and incaCompl we try
to predict whether or not someone is an adult, on the dataset GJB we will try to learn to predict
the gender, and on the datasets privacy3 and privacy4 we will predict an answer to some questions
among 5 possibilities.
We perform our experiments by randomly splitting the datasets incaIndiv, incaCompl,GJB into
a training set (roughly 70 percent of the initial dataset) and a test set. The size of privacy3 and
privacy4 being much smaller we perform 10-fold cross-validation.
2
https://https://www.data.gouv.fr/fr/datasets/
http://www.pewinternet.org/datasets/june-10-july-12-2015-gaming-jobs-and-broadband/
4
http://www.pewinternet.org/datasets/nov-26-2014-jan-3-2015-privacy-panel-3/
5
http://www.pewinternet.org/datasets/jan-27-feb-16-2015-privacy-panel-4/
3
15
Weighted SVM
Unweighted SVM
Weighted CART
Unweighted CART
incaIndiv
GJB
privacy3
privacy4
0.16
0.19
0.04
0.05
0.36
0.43
0.41
0.43
0.46
0.50
0.49
0.52
0.48
0.52
0.54
0.57
Table 2: Average over 50 runs of the prediction error
References
[1] R Bardenet and O.A. Maillard. Concentration inequalities for sampling without replacement.
Bernoulli, 21(3):1361–1385, 2015.
[2] P. Bartlett, M. Jordan, and J. McAuliffe. Convexity, classification and risk bounds. J. of the
A.M.S, 101(473):138–156, 2006.
[3] Y.G. Berger. Rate of convergence to normal distribution for the Horvitz-Thompson estimator.
J. Stat. Plan. Inf, 67(2):209–226, 1998.
[4] S. N. Bernstein. On a modification of chebyshev’s inequality and on the error in laplace
formula. Collected Works, Izd-vo ’Nauka’, Moscow (in Russian), 4:71–80, 1964.
[5] S. Boucheron, O. Bousquet, and G. Lugosi. Theory of Classification: A Survey of Some Recent
Advances. ESAIM: Probability and Statistics, 9:323–375, 2005.
[6] P. Brändén and J. Jonasson. Negative dependence in sampling. Scandinavian Journal of
Statistics, 39(4):830–838, 2012.
[7] N.E. Breslow, T. Lumley, C. Ballantyne, L. Chambless, and M. Kulich. Improved HorvitzThompson estimation of model parameters from two-phase stratified samples: applications in
epidemiology. Stat. Biosc., 1:32–49, 2009.
[8] N.E. Breslow and J.A. Wellner. Weighted likelihood for semiparametric models and two-phase
stratified samples, with application to Cox regression. Scandinavian Journal of Statistics,
35:186–192, 2007.
[9] X.H. Chen, Dempster A.P., and J.S. Liu. Weighted finite population sampling to maximize
entropy. Biometrika, 81(3):457–469, 1994.
[10] J.C. Deville. Réplications d’échantillons, demi-échantillons, Jackknife, bootstrap dans les
sondages. Economica, Ed. Droesbeke, Tassi, Fichet, 1987.
[11] J. Dupacova. A note on rejective sampling. Contribution to Statistics (J. Hajek memorial
volume) Academia Prague, pages 71–78, 1979.
[12] L.A. Goodman. On the estimation of the number of classes in a population. Annals of
Mathematical Statistics, 20:572–579, 1949.
16
[13] J. Hajek. Asymptotic theory of rejective sampling with varying probabilities from a finite
population. The Annals of Mathematical Statistics, 35(4):1491–1523, 1964.
[14] D.G. Horvitz and D.J. Thompson. A generalization of sampling without replacement from a
finite universe. JASA, 47:663–685, 1951.
[15] S. Janson. Large deviation inequalities for sums of indicator variables. 1994.
[16] K. Joag-Dev and F. Proschan. Negative Association of Random Variables with Applications.
The Annals of Statistics, 11(1):286–295, 1983.
[17] V. Koltchinskii. Local Rademacher complexities and oracle inequalities in risk minimization
(with discussion). The Annals of Statistics, 34:2593–2706, 2006.
[18] J.B. Kramer, J. Cutler, and A.J. Radcliffe. Negative dependence and srinivasan’s sampling
process. Combinatorics, Probability and Computing, 20(3):347–361, 2011.
[19] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel,
P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher,
M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine
Learning Research, 12:2825–2830, 2011.
[20] P.M. Robinson. On the convergence of the Horvitz-Thompson estimator. Australian Journal
of Statistics, 24(2):234–238, 1982.
[21] T. Saegusa and J.A. Wellner. Weighted likelihood estimation under two-phase sampling. 2011.
[22] C.E. Särndall and J. Wretman B. Swensson. Model assisted survey sampling. Springer-Verlag,
NY, 2003.
[23] R.J. Serfling. Probability inequalities for the sum in sampling without replacement. The
Annals of Statistics, 2:39–48, 1974.
[24] I. Steinwart, D. Hush, and C. Scovel. Learning from dependent observations. Journal of
Multivariate Analysis, 100(1):175–194, 2009.
[25] Y. Tillé. Sampling algorithms. Springer Series in Statistics, 2006.
[26] A.W. van der Vaart and J.A. Wellner. Weak convergence and empirical processes. Springer,
1996.
[27] V. Vapnik. The Nature of Statistical Learning Theory. Springer-Verlag, New-York, 2001.
17
| 10math.ST
|
Approximate Counting of Matchings in
(3, 3)-Hypergraphs?
arXiv:1402.6190v3 [math.CO] 25 Oct 2017
Andrzej Dudek1?? , Marek Karpinski2? ? ? , Andrzej Ruciński3† , and Edyta
Szymańska3‡
1
Western Michigan University, Kalamazoo, MI, USA, [email protected]
2
Department of Computer Science, University of Bonn, Germany,
[email protected]
3
Faculty of Mathematics and Computer Science, Adam Mickiewicz University,
Poznań, Poland, (rucinski,edka)@amu.edu.pl
Abstract. We design a fully polynomial time approximation scheme
(FPTAS) for counting the number of matchings (packings) in arbitrary
3-uniform hypergraphs of maximum degree three, referred to as (3, 3)hypergraphs. It is the first polynomial time approximation scheme for
that problem, which includes also, as a special case, the 3D Matching
counting problem for 3-partite (3, 3)-hypergraphs. The proof technique
of this paper uses the general correlation decay technique and a new
combinatorial analysis of the underlying structures of the intersection
graphs. The proof method could be also of independent interest.
1
Introduction
The computational status of approximate counting of matchings in hypergraphs
has been open for some time now, contrary to the existence of polynomial time
approximation schemes for graphs. The matching (packing) counting problems
in hypergraphs occur naturally in the higher dimensional free energy problems,
like in the monomer-trimer systems discussed, e.g, by Heilmann [10]. The corresponding optimization versions of hypergraph matching problem relate also to
various allocations problems.
This paper aims at shedding some light on the approximation complexity of
that problem in 3-uniform hypergraphs of maximum vertex degree three (called
(3, 3)-hypergraphs or (3, 3)-graphs for short). This class of hypergraphs includes
?
??
???
†
‡
Part of research of the 3rd and 4th authors done at Emory University, Atlanta and
another part during their visits to the Institut Mittag-Leffler (Djursholm, Sweden).
Research supported by Simons Foundation Grant #244712 and by a grant from
the Faculty Research and Creative Activities Award (FRACAA), Western Michigan
University.
Research supported by DFG grants and the Hausdorff grant EXC59-1.
Research supported by the Polish NSC grant N201 604 940 and the NSF grant
DMS-1102086.
Research supported by the Polish NSC grant N206 565 740.
2
also so-called 3D hypergraphs, that is, (3,3)-graphs that are 3-partite. In [12],
based on a generalization of the canonical path method of Jerrum and Sinclair
[11], we established a fully polynomial time randomized approximation scheme
(FPRAS) for counting matchings in the classes of k-uniform hypergraphs without
structures called 3-combs. However, the status of the problem in arbitrary (3, 3)graphs was left wide open among with other general problems for 3-, 4- and 5uniform hypergraphs (for k ≥ 6 it is known to be hard, see Sec. 2). In particular,
the existence of an FPRAS for counting matchings in (3, 3)-graphs was unknown.
In this paper we design the first fully polynomial time approximation scheme
(FPTAS) for arbitrary (3, 3)-graphs. The method of solution depends on the
general correlation decay technique and some new structural analysis of underlying intersections graphs based on an extension of the classical claw-freeness
notion. The proof method used in the analysis of our algorithm could be also of
independent interest.
The paper is organized as follows. Section 2 contains some basic notions and
preparatory discussions. In Sec. 3 we formulate our main results and provide
the proofs. Finally, Sec. 4 is devoted to the summary and an outlook for future
research.
2
Preliminaries
A hypergraph H = (V, E) is a finite set of vertices V together with a family E
of distinct, nonempty subsets of vertices called edges. In this paper we consider
k-uniform hypergraphs (called further k-graphs) in which, for a fixed k ≥ 2, each
edge is of size k. A matching in a hypergraph is a set (possibly empty) of disjoint
edges.
Counting matchings is a #P-complete problem already for graphs (k = 2)
as proved by Valiant [17]. In view of this hardness barrier, researchers turned to
approximate counting, which initially has been accomplished via probabilistic
techniques.
Given a function C and a random variable Y (defined on some probability space), and given two real numbers , δ > 0, we say that Y is an (, δ)approximation of C if the probability P (|Y (x) − C(x)| ≥ C(x)) ≤ δ. A fully
polynomial randomized approximation scheme (FPRAS) for a function f on
{0, 1}∗ is a randomized algorithm which, for every triple (, δ, x), with > 0, δ >
0, and x ∈ {0, 1}∗ , returns an (, δ)-approximation Y of f (x) and runs in time
polynomial in 1/, log(1/δ), and |x|.
In this paper we investigate the problem of counting the number of matchings
in hypergraphs and try to determine the status of this problem for k-graphs with
bounded degrees.
Let degH (v) be the degree of vertex v in a hypergraph H, that is, the number
of edges of H containing v. We denote by ∆(H) the maximum of degH (v) over
all v in H. We call a k-graph H a (k, r)-graph if ∆(H) ≤ r. Let #M (k, r) be
the problem of counting the number of matchings in (k, r)-graphs.
3
Our inspiration comes from new results (both positive and negative) that
emerged for approximate counting of the number of independent sets in graphs
with bounded degree and shed some light on the problem #M (k, r).
Let #IS(d) [#IS(≤ d)] be the problem of counting the number of all independent sets in d-regular graphs [graphs of maximum degree bounded by d, that
is, (2, d)-graphs]. Luby and Vigoda [14] established an FPRAS for #IS(≤ 4).
This was complemented later by the approximation hardness results for the
higher degree instances by Dyer, Frieze and Jerrum [7]. The subsequent progress
has coincided with the revival of a deterministic technique – the spatial correlation decay method – based on early papers of Dobrushin [5] and Kelly [13]. It
resulted in constructing deterministic approximation schemes for counting independent sets in several classes of graphs with degree (and other) restrictions, as
well as for counting matchings in graphs of bounded degree.
Definition 1. A fully polynomial time approximation scheme (FPTAS) for a
function f on {0, 1}∗ is a deterministic algorithm which for every pair (, x) with
> 0, and x ∈ {0, 1}∗ , returns a number y(x) such that
|y(x) − f (x)| ≤ f (x),
and runs in time polynomial in 1/, and |x|.
In 2007 Weitz [18] found an FPTAS for #IS(≤ 5), while, more recently, Sly
[15] and Sly and Sun [16] complemented Weitz’s result by proving the approximation hardness for #IS(6), that is, proving that unless NP=RP, there exists no
FPRAS (and thus, no FPTAS) for #IS(6). By applying two reductions: from
#IS(6) to #M (6, 2) (taking the dual hypergraph of a 6-regular graph), and
from #M (k, 2) to #IS(k) (taking the intersection graph of a (k, 2)-graph) for
k = 3, 4, 5, we conclude that
(i) (unless NP=RP) there exists no FPRAS for #M (6, 2);
(ii) there is an FPTAS for #M (k, 2) with k ∈ {3, 4, 5}.
Note that the first reduction results, in fact, in a linear (6, 2)-graph, so the
class of hypergraphs in question is even narrower. (A hypergraph is called linear
when no two edges share more than one vertex.) On the other hand, by the same
kind of reduction it follows from a result of Greenhill [9] that exact counting of
matchings is #P-complete already in the class of linear (3, 2)-graphs.
Facts (i) and (ii) above imply that the only interesting cases for the positive
results are those for (k, r)-graphs with k = 3, 4, 5 and r ≥ 3, and thus, the
smallest one among them is that of (3, 3)-graphs. Our main result establishes an
FPTAS for counting the number of matchings in this class of hypergraphs.
3
Main Result and the Proof
The following theorem is the main result of this paper.
4
Theorem 2. The algorithm CountMatchings given in Section
3.2 provides an
FPTAS for #M (3, 3) and runs in time O n2 (n/)log50/49 144 .
The intersection graph of a hypergraph H is the graph G = L(H) with
vertex set V (G) = E(H) and edge set E(G) consisting of all intersecting pairs
of edges of H. When H is a graph, the intersection graph L(H) is called the line
graph of H. Graphs which are line graphs of some graphs are characterized by
9 forbidden induced subgraphs [3], one of which is the claw, an induced copy
of K1,3 . There is no similar characterization for intersection graphs of k-graphs.
Still, it is easy to observe that for any k-graph H, its intersection graph L(H)
does not contain an induced copy of K1,k+1 . We shall call such graphs (k + 1)claw-free.
Our proof of Thm. 2 begins with an obvious observation that counting the
number of matchings in a hypergraph H is equivalent to counting the number of
independent sets in the intersection graph G = L(H). More precisely, let ZM (H)
be the number of matchings in a hypergraph H and, for a graph G, let ZI (G)
be the number of independent sets in G. (Note that both quantities count the
empty set.) Then ZM (H) = ZI (L(H)).
To approximately count the number of independent sets in a graph G =
L(H) for a (3, 3)-graph H, we apply some of the ideas from [2] (the preliminary
version of this paper appeared in [1]) and [8]. In [2] two new instances of FPTAS
were constructed, both based on the spatial correlation decay method. First, for
#M (2, r) with any given r. Then, still in [2], the authors refined their approach
to yield an FPTAS for counting independent sets in claw-free graphs of bounded
clique number which contain so called simplicial cliques. The last restriction has
been removed by an ingenious observation in [8].
Papers [2,8] inspired us to seek adequate methods for (3, 3)-graphs. Indeed,
for every (3, 3)-graph H its intersection graph G = L(H) is 4-claw-free and
has ∆(G) ≤ 6. This turned out to be the right approach, as we deduced our
Thm. 2 from a technical lemma (Lem. 3 below) which constructs an FPTAS
for the number of independent sets in K1,4 -free graphs G with ∆(G) ≤ 6 and
an additional property stemming from their being intersection graphs of (3, 3)graphs.
3.1
Proof of Theorem 2 – Sketch and Preliminaries
We deduce Thm. 2 from a technical lemma. The assumptions of this lemma
reflect some properties of the intersection graphs of (3, 3)-graphs.
Lemma 3. There exists an FPTAS for the problem of counting independent
sets in every 4-claw-free graph with maximum degree at most 6 and such that
the neighborhood of every vertex of degree d ≥ 5 induces a subgraph that spans a
matching of size bd/2c.
Proof (of Thm. 2). Given a (3, 3)-graph H, consider its intersection graph G.
Then G is 4-claw-free, has maximum degree at most 6 and every vertex neighborhood of size d ≥ 5 must span in G a matching of size bd/2c. This means that
5
Lem. 3 applies to G and there is an FPTAS for counting independent sets of G
which is the same as counting matchings in H.
t
u
It remains to prove Lem. 3. We begin with underlining some properties of
4-claw-free graphs which are relevant for our method. First, we introduce the
notion of a simplicial 2-clique which is a generalization of a simplicial clique
introduced in [4] and utilized in [2]. Throughout we assume notation A \ B for
set differences and, for A ⊂ V (G), we write G − A for the graph operation of
deleting from G all vertices belonging to A. In other words, G−A = G[V (G)\A].
Also, for any graph G, we use δ(G) to denote its minimum vertex degree and
α(G) for the size of the largest independent set in G.
Definition 4. A set K ⊆ V (G) is a 2-clique if α(G[K]) ≤ 2. A 2-clique is
simplicial if for every v ∈ K, NG (v) \ K is a 2-clique in G − K.
For us a crucial property of simplicial 2-cliques is that if G is a connected
graph containing a nonempty simplicial 2-clique K then it is easy to find another
simplicial 2-clique in the induced subgraph G − K, and consequently, the whole
vertex set of G can be partitioned into blocks which are simplicial 2-cliques in
suitable nested sequence of induced subgraphs of G (see Claim 8).
However, in the proof of Lem. 3 we shall use a special class of 2-cliques.
Definition 5. A 2-clique K in a graph G is called a block if |K| ≤ 4 and
δ(G[K]) ≥ 1 whenever |K| = 4. A block K is simplicial if for every v ∈ K the
set NG (v) \ K is a block in G − K.
Next, we state a trivial but useful observation which follows straight from
the above definition. (We consider the empty set as a block too.)
Fact 6. If K is a (simplicial) block in G then for every V 0 ⊆ V (G) the set
K ∩ V 0 is a (simplicial) block in the induced subgraph G[V 0 ] of G.
Let a graph G satisfy the assumptions of Lem. 3. The next claim provides a
vital, “self-reproducing” property of blocks in G.
Claim 7. If K is a simplicial block in G, then for every v ∈ K the set NG (v)\K
is a simplicial block in G − K.
Proof. Set Kv := NG (v) \ K for convenience. By definition of K, Kv is a block.
It remains to show that Kv is simplicial. Let u ∈ Kv and let Ku = NG (u) \ (K ∪
NG (v)). Suppose there is an independent set I in G[Ku ] of size |I| = 3. Then
u, v and the vertices of I would form an induced K1,4 in G with u in the center.
As this is a contradiction, we conclude that Ku is a 2-clique.
To show that Ku is indeed a block, note first that, by the assumptions that
∆(G) ≤ 6, we have |Ku | ≤ 5. However, if |Ku | = 5 then v would be an isolated
vertex in G[NG (u)]. But G[NG (u)] spans a matching of size 3 since δ(u) = 6 – a
contradiction. For the same reason, if |Ku | = 4 then regardless of the degree of
u in G (which might be 5 or 6) there can be no isolated vertex in G[Ku ], since
G[Ku ] must span a matching of size 2.
t
u
6
Our next claim asserts that once there is a nonempty block in G, one can
find a suitable partition of V (G) into sets which are blocks in a nested sequence
of induced subgraphs of G defined by deleting these sets one after another.
Claim 8. Let K be a nonempty simplicial block in G. If, in addition, G is
connected then there exists a partition V (G) = K1 ∪ · · · ∪ Km such that K1 = K
and
Si−1for every i = 2, . . . , m, Ki is a nonempty, simplicial block in Gi := G −
j=1 Kj .
Proof. Suppose we have already constructed disjoint sets K1 ∪ · · · ∪ Ks , for some
s ≥ 1, such that K1S= K, for every i = 2, . . . , s, Ki isSa nonempty, simplicial
i−1
s
block in Gi := G − j=1 Kj , and that Rs := V (G) \ i=1 Ki 6= ∅. Since G is
connected, there is an edge between a vertex in Rs and a vertex v ∈ Ki for some
1 ≤ i ≤ s. Since Ki is a simplicial block in Gi , by Fact 6, it is also simplicial in
its subgraph Gi [V 0 ], where V 0 = Ki ∪ Rs , that is the subgraph of Gi obtained
0
by deleting all vertices of Ki+1 ∪ · · · ∪ Ks . Now apply Claim 7 to Gi [V
Ss ], Ki , and
v, to conclude that NG (v) ∩ Rs is a simplicial block in Gs+1 := G − i=1 Ki . t
u
Let K1 , K2 , . . . , Km be as in Claim 8. Then,
ZI (G) =
ZI (Gi )
ZI (Gm )
ZI (G1 ) ZI (G2 )
·
· ... ·
· ... ·
,
ZI (G2 ) ZI (G3 )
ZI (Gi+1 )
ZI (Gm+1 )
(1)
where Gm+1 = ∅ and ZI (Gm+1 ) = 1. Observe that for each i, Gi+1 = Gi − Ki
and the reciprocal of each quotient in (1) is precisely the probability
PGi (Ki ∩ I = ∅) =
ZI (Gi − Ki )
,
ZI (Gi )
(2)
where I is an independent set of Gi chosen uniformly at random. In view of
this, the main step in building an FPTAS for ZI (G) will be to approximate the
probability PG (Ki ∩ I = ∅) within 1 ± n (see Sec. 3.2 and Algorithm 2 therein).
But what if G isSdisconnected or does not contain a simplicial block to start
c
with? First, if G = i=1 Gi consists of c connected components G1 , . . . , Gc , then,
clearly
c
Y
ZI (G) =
ZI (Gi )
(3)
i=1
and the problem reduces to that for connected graphs.
As for the second obstacle, Fadnavis [8] proposed a very clever observation to
cope with it. Let G be a connected graph satisfying the assumptions of Lem. 3
and let v ∈ V (G) be such that G − v is connected. By considering the fate of
vertex v, we obtain the recurrence
ZI (G) = ZI (G − v) + ZI (Gv ),
(4)
Sc
where Gv = G − NG [v] and NG [v] = NG (v) ∪ {v}. Let Gv = i=1 Gvi be the
partition of Gv into its connected components. For each i let ui ∈ NG (v) be
such that NG (ui ) ∩ V (Gvi ) 6= ∅. Owing to the connectedness of G − v, a vertex
ui must exist. Set Ki = NG (ui ) ∩ V (Gvi ).
7
Claim 9. The set Ki is a simplicial block in Gvi .
Proof. The proof is quite similar to that of Claim 7. We first prove that Ki is a
block. Suppose there is an independent set I in G[Ki ] of size |I| = 3. Then ui , v
and the vertices of I would form an induced K1,4 in G with ui in the center.
As this is a contradiction, we conclude that Ki is a 2-clique. To prove that Ki
is, in fact, a block, notice that there is no edge between v and Ki . Thus, we
cannot have |Ki | = 5 because then v would be an isolated vertex in G[N (ui )] –
a contradiction with the assumption on G. If, however, |Ki | = 4 then v is the
(only) isolated vertex in G[N (ui )] and, consequently, δ(G[Ki ]) ≥ 1.
It remains to show that the block Ki is simplicial, that is, for every w ∈ Ki ,
the set NGvi (w) \ Ki is a block in Gvi − Ki . This, however, can be proved mutatis
mutandis as in the proof of Claim 7.
t
u
For the first term of recurrence (4) we apply (4) recursively. In view of Claim
9, to the second term of recurrence (4) one can apply formula (3) and then each
term ZI (Gvi ) can be approximated based on (1) and (2).
3.2
The Remainder of the Proof of Lemma 3
within 1 ± n , where
Hence, it remains to approximate PG (K ∩ I = ∅) = ZIZ(G−K)
I (G)
K is a simplicial block in G. We set Nv := NG (v) and formulate the following
recurrence relation by considering how an independent set may intersect K:
ZI (G) = ZI (G − K) +
X
ZI (G − (Nv ∪ K)) +
v∈K
1
2
X
ZI (G − (Nu ∪ Nv ∪ K))
uv ∈G[K]
/
or equivalently, after dividing sidewise by ZI (G − K),
X ZI (G − (Nv ∪ K)) 1
ZI (G)
= 1+
+
ZI (G − K)
ZI (G − K)
2
v∈K
X
uv ∈G[K]
/
ZI (G − (Nu ∪ Nv ∪ K))
.
ZI (G − K)
Here and throughout the inner summation ranges over all ordered pairs of
distinct vertices of K such that {u, v} ∈
/ G[K]. At this point, in view of symmetry,
it seems redundant to consider ordered pairs (and consequently have the factor
of 12 in front of the sum), but we break the symmetry right now as we further
observe that
ZI (G − (Nu ∪ Nv ∪ K)) ZI (G − (Nv ∪ K))
ZI (G − (Nu ∪ Nv ∪ K))
=
·
.
ZI (G − K)
ZI (G − (Nv ∪ K))
ZI (G − K)
By Claim 7, Nv \ K is a simplicial block in G − K. We need to show that,
similarly, Nu \ (Nv ∪ K) is a simplicial block in G − (Nv ∪ K).
Claim 10. Let K be a simplicial block in G and let u, v ∈ K be such that u 6= v
and uv ∈
/ G[K]. Further, let H := G − (NG (v) ∪ K). Then NH (u) is a simplicial
block in H.
8
Proof. By Claim 7, the set NG (u) \ K is a simplicial block in G − K. Apply Fact
6 to NG (u) \ K and G − K with V 0 = V (H).
t
u
Let
ΠG (K) := P(K ∩ I = ∅) =
ZI (G − K)
,
ZI (G)
where I is a random independent set of G. Finally, setting Kv := Nv \ K and
Kuv := Nu \ (Nv ∪ K), and rewriting G − (Nv ∪ K) = G − K − Kv , we get the
recurrence for the probabilities:
X
X
1
−1
ΠG−K−Kv (Kuv ) .
ΠG
(K) = 1 +
ΠG−K (Kv ) 1 +
2
v∈K
uv ∈G[K]
/
This recurrence, in principle, allows one to compute ΠG (K) exactly, but only
in an exponential number of steps. Instead, we will approximate it by a function
ΦG (K, t), also defined recursively, which “mimics” ΠG (K) but has a built-in
time counter t.
Definition 11. For every graph G, every simplicial block K in G and an integer
t ∈ Z+ , the function ΦG (K, t) is defined recursively as follows: ΦG (K, 0) =
ΦG (K, 1) = 1 as well as ΦG (∅, t) = 1, while for t ≥ 2 and K 6= ∅
X
X
1
Φ−1
ΦG−K (Kv , t − 1) 1 +
ΦG−K−Kv (Kuv , t − 2) .
G (K, t) = 1 +
2
v∈K
uv ∈G[K]
/
Now we are ready to state the algorithm CountMatchings for computing
ZM (H) for any connected (3, 3)-graph H and its subroutine CountIS for computing ZI (G) in a subgraph of G = L(H) containing a simplicial block K.
Algorithm 1 CountMatchings(H, t)
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
G := L(H).
ZM := 1, F := G.
while F 6= ∅ do
Pick v ∈ V (F ) s.t. F − v is connected.
F v := F − NF [v]
If F v =
S ∅ then ZM = ZM + 1 and go to Line 3.
F v = ci=1 Fiv , where Fiv are connected components of F v .
for i := 1 to c do
Find Ki as in Claim 9
end for
Q
ZM := ZM + ci=1 CountIS(Fiv , Ki , t)
F := F − v
end while
Return ZM
9
Algorithm 2 CountIS(G, K, t)
1:
2:
3:
4:
5:
6:
7:
S
Let V (G) = m
i=1 Ki be a partition of V (G) as in Claim 8 with K1 = K.
ZI := 1, F := G
for i = 1 to m do
ZI
ZI := ΦF (K
i ,t)
F := F − Ki
end for
Return ZI
We will show that already for t = Θ(log n), when Φ can be easily computed
in polynomial time, the two functions become close to each other.
Note that both quantities, ΠG (K) and ΦG (K, t), fall into the interval [ 91 , 1].
The lower bound is due to the fact that a block has at most 4 vertices and each
of them has degree at most 2 in Gc , so that the total number of terms in the
denominator is at most nine, five of them do not exceed 1, while eight of them
do not exceed 12 . Our goal is to approximate ΠG (K) by ΦG (K, t), for a suitably
chosen t, within the multiplicative factor of 1 ± /n. In view of the above lower
bound, it suffices to show that |ΠG (K) − ΦG (K, t)| ≤ 9n
.
To achieve this goal, we will use the correlation decay technique which boils
down to establishing a recursive bound on the above difference (cf. [2]). The
success of this method depends on the right choice of a pair of functions g and
h, with g : [0, 1] → <, such that they are inverses of each other, that is, g ◦ h ≡ 1.
Then we define a function fK of |K| + 2e(Gc [K]) variables, one for each vertex
and each (ordered) non-edge of G[K], as follows. Let z = (z1 , . . . , z|K| , zuv : uv ∈
/
G[K]) be a vector of variables of that function. For ease of notation, we denote
the set of all indices of the coordinates of function fK by J, that is, we set
J := K ∪ {(u, v) : {u, v} ∈
/ G[K]}. Then
−1
X
X
1
h(zuv )
fK (z) := f (z) = g 1 +
h(zv ) 1 +
(5)
.
2
v∈K
uv ∈G[K]
/
To understand the reason for this set-up, put x := g(ΠG (K)), xv := g(ΠG−K (Kv )),
xuv := g(ΠG−K−Kv (Kuv )), and, correspondingly,
y := g(ΦG (K, t)) yv := g(ΦG−K (Kv , t − 1))
yuv := g(ΦG−K−Kv (Kuv , t − 2)).
Then, f (x) = x and f (y) = y, and so the difference we are after can be expressed
as |x − y| = |f (x) − f (y)|. Thus, we are in position to apply the Mean Value
Theorem to f and conclude that there exists α ∈ [0, 1] such that, setting zα =
αx + (1 − α)y,
|f (x) − f (y)| = |∇f (zα )(x − y)| ≤ |∇f (zα )| × max |xκ − yκ |.
κ∈J
It remains to bound maxz |∇f (z)| from above, uniformly by a constant γ < 1.
Then, after iterating at most t but at least t/2 times, we will arrive at a triple
10
(G0 , K 0 , t0 ), where G0 is an induced subgraph of G, K 0 is a block in G0 , and
t0 ∈ {0, 1}. At this point, setting µg := |g(1)| + | maxs g(s))|, we will obtain the
ultimate bound
|x − y| ≤ γ t/2 × |g(ΠG0 (K 0 )) − g(1)| ≤ γ t/2 × µg ≤
,
9n
for
t ≥ 2 log((9µg n)/)/ log(1/γ).
(6)
In [2], to estimate |∇f (z)| for a similar function f , the authors chose g(s) =
log s and h(s) = es . This choice, however, does not work for us. Instead, we set
g(s) = s1/4 and h(s) = s4 . Then, µg = 2 and
(
)
P
P
4
3
zv3 + 12
(zv3 zuv
+ zv4 zuv
)
X ∂f (z)
v∈K
uv ∈G[K]
/
=(
|∇f (z)| ≤
!)5/4 .
∂zκ
P 4
P
κ∈J
1
4
1+
zuv
zv 1 + 2
v∈K
uv ∈G[K]
/
Observe that fK depends only on the isomorphism type of G[K], a graph on
up to 4 vertices, with no independent set of size 3, and with no isolated vertex
when |K| = 4. Let us call all these graphs block graphs. One block graph is given
in Figure 1 below.
z1
z2
z13
z14
z23
z24
z4
z3
Fig. 1. The essential block graph.
In a sense we just need to consider this one block graph. Indeed, the complement of every block graph is contained in the complement of the block graph
in Figure 1. Hence, it suffices to maximize |∇f (z)| just for this graph. Our computational task is, therefore, to bound from above
1
1 + z14 + z24 + z34 + z44 +
4
−5/4
1 4 4
4
4
4
z14 z1 + z44 + z13
z14 + z34 + z23
z24 + z34 + z24
z24 + z44
×
2
4
4
4
4
4
4
4
4
2z13 2 + z14
+ z13
+ 2z23 2 + z23
+ z24
+ 2z33 2 + z13
+ z23
+ 2z43 2 + z14
+ z24
+
3
3
3
3
2z14
z14 + z44 + 2z13
z14 + z34 + 2z23
z24 + z34 + 2z24
z24 + z44 .
F (z) = k∇(z)k1 =
11
One can show (using, e.g., Mathematica) that F (z) < 0.971 for 0 ≤ zi ≤ 1
and 0 ≤ zij ≤ 1. Thus, we have (6) with µg = 2 and, say, γ = 0.98 = 49
50 .
Summarizing, the running time of computing ΦG (K, t) in Step 4 of Algorithm 2
is 12t since there at most 12 expressions to compute in each step of the recurrence
relation (see Def. 11). Also, CountIS takes at most |V (Fiv )|12t steps and hence,
Line 11 of CountMatchings takes n12t steps and is invoked at most n times.
Consequently, with t = 2dlog((18n)/)/ log(50/49)e
we get the running time of
our algorithm of order O n2 (n/)log50/49 144 .
Remark 12. With basically the same proof wePcan construct an FPTAS for
calculating the partition function ZM (H, λ) = M λ|M | , where the sum runs
over all matchings in H, for any constant λ ∈ (0, 1.077]. The λ factor will appear
4
in front of each summation in (5), which one can neutralize by setting h(s) = sλ
and g(s) = (λs)1/4 .
4
Summary, Discussion, and Further Research
The main result of this paper (Thm. 2) establishes an FPTAS for the problem
#M (3, 3) of counting the number of matchings in a (3, 3)-graph. A reformulation
of Thm. 2 in terms of graphs yields an FPTAS for the problem of counting
independent sets in every graph which is the intersection graph of a (3, 3)-graph.
As mentioned earlier, every intersection graph of a (3, 3)-graph is 4-claw-free.
Moreover, its maximum degree is at most six. We wonder if there exists an
FPTAS for the problem of counting independent sets in every 4-claw-free graph
with maximum degree at most 6. Lemma 3 falls short of proving that. The
missing part is due to our inability to repeat the above estimates for 2-cliques
of size five.
In an earlier paper [12] three of the authors have found an FPRAS for the
number of matchings in k-graphs without 3-combs. As their intersection graphs
are claw-free, it follows from the above mentioned result on independent sets in
[2,8] that there is also an FPTAS for the number of matchings in (k, r)-graphs
without 3-combs, for any fixed r. In view of this conclusion and Thm. 2, we raise
the question if for all k ≤ 5 and r there is an FPTAS (or at least FPRAS) for the
problem #M (k, r). The first open instance is that of (3, 4)-graphs. For k = 4, 5,
to avoid recurrences of depth k − 1 ≥ 3, as an intermediate step, one could first
consider the restriction of the class of (k, r)-graphs to those without a 4-comb,
that is, to those whose intersection graphs are 4-claw-free. Here, the first open
instance is that of (4, 3)-graphs without 4-combs. In general, it would be also
very interesting to elucidate the status of the problem for arbitrary k-graphs for
k = 3, 4 and 5, or for some generic subclasses of them.
Acknowledgements
We thank Martin Dyer and Mark Jerrum for stimulating discussions on the
subject of this paper and the referees for their valuable comments. We are also
12
very grateful to Michael Simkin who pointed out and fixed an error (cf. Lemma 3
and the proof of Claim 7) in an earlier version of this paper [6].
References
1. Bayati, M., Gamarnik, D., Katz, D., Nair, C., Tetali, P.: Simple deterministic
approximation algorithms for counting matchings. In: STOC’07—Proceedings of
the 39th Annual ACM Symposium on Theory of Computing, pp. 122–127. ACM
(2007)
2. Bayati, M., Gamarnik, D., Katz, D., Nair, C., Tetali, P.: Simple deterministic
approximation algorithms for counting matchings (2008), http://people.math.
gatech.edu/~tetali/PUBLIS/BGKNT_final.pdf
3. Beineke, L.W.: Characterizations of derived graphs. J. Combin. Theory 9, 129–135
(1970)
4. Chudnovsky, M., Seymour, P.: The roots of the independence polynomial of a
clawfree graph. J. Combin. Theory Ser. B 97(3), 350–357 (2007)
5. Dobrushin, R.: Prescribing a system of random variables by conditional distributions. Theor. Probab. Appl. 15, 458–486 (1970)
6. Dudek, A., Karpinski, M., Ruciński, A., Szymańska, E.: Approximate counting
of matchings in (3, 3)-hypergraphs. In: Algorithm theory—SWAT 2014, Lecture
Notes in Comput. Sci., vol. 8503, pp. 380–391. Springer, Cham (2014)
7. Dyer, M., Frieze, A., Jerrum, M.: On counting independent sets in sparse graphs.
SIAM J. Comput. 31(5), 1527–1541 (2002)
8. Fadnavis, S.: Approximating independence polynomials of claw-free graphs (2012),
http://www.math.harvard.edu/~sukhada/IndependencePolynomial.pdf
9. Greenhill, C.: The complexity of counting colourings and independent sets in sparse
graphs and hypergraphs. Comput. Complexity 9(1), 52–72 (2000)
10. Heilmann, O.: Existence of phase transitions in certain lattice gases with repulsive
potential. Lett. Al Nuovo Cimento Series 2 3(3), 95–98 (1972)
11. Jerrum, M., Sinclair, A.: Approximating the permanent. SIAM J. Comput. 18(6),
1149–1178 (1989)
12. Karpiński, M., Ruciński, A., Szymańska, E.: Approximate counting of matchings
in sparse uniform hypergraphs. In: 2013 Proceedings of the Workshop on Analytic
Algorithmics and Combinatorics (ANALCO), pp. 72–79. SIAM (2013)
13. Kelly, F.P.: Stochastic models of computer communication systems. J. Roy. Statist.
Soc. Ser. B 47(3), 379–395, 415–428 (1985)
14. Luby, M., Vigoda, E.: Fast convergence of the Glauber dynamics for sampling
independent sets. Random Structures Algorithms 15(3-4), 229–241 (1999)
15. Sly, A.: Computational transition at the uniqueness threshold. In: 2010 IEEE 51st
Annual Symposium on Foundations of Computer Science FOCS 2010, pp. 287–296
(2010)
16. Sly, A., Sun, N.: The computational hardness of counting in two-spin models on dregular graphs. In: FOCS, pp. 361–369 (2012), http://arxiv.org/abs/1203.2602
17. Valiant, L.G.: The complexity of enumeration and reliability problems. SIAM J.
Comput. 8(3), 410–421 (1979)
18. Weitz, D.: Counting independent sets up to the tree threshold. In: STOC’06: Proceedings of the 38th Annual ACM Symposium on Theory of Computing, pp. 140–
149. ACM (2006)
13
Appendix: Mathematica expressions
First we define F (z) function:
F[z1_, z2_, z3_, z4_, z14_, z13_, z23_, z24_]:=\
1/4(1+z1^4+z2^4+z3^4+z4^4+\
1/2(z14^4(z1^4+z4^4)+z13^4(z1^4+z3^4)+\
z23^4(z2^4+z3^4)+z24^4(z2^4+z4^4)))^(-5/4)\
(2z1^3(2+z14^4+z13^4)+2z2^3(2+z23^4+z24^4)+\
2z3^3(2+z13^4+z23^4)+2z4^3(2+z14^4+z24^4)+\
2z14^3(z1^4+z4^4)+2z13^3(z1^4+z3^4)+\
2z23^3(z2^4+z3^4)+2z24^3(z2^4+z4^4))
Next we find the absolute maximum:
NMaximize[{F[z1, z2, z3, z4, z14, z13, z23, z24],\
0<=z1<=1 && 0<=z2<=1 && 0<=z3<=1 && 0<=z4<=1 &&
0<=z14<=1 && 0<=z13<=1 && 0<=z23<=1 && 0<=z24<=1},\
{z1, z2, z3, z4, z14, z13, z23, z24}]
obtaining that
F (z) ≤ F (ζ, ζ, ζ, ζ, 1, 1, 1, 1) ∼ 0.970247,
where ζ ∼ 0.695347.
| 8cs.DS
|
arXiv:1709.01606v1 [math.AC] 5 Sep 2017
Factoring in the Chicken McNugget monoid1
Scott T. Chapman
Sam Houston State University
Department of Mathematics and Statistics
Box 2206
Huntsville, TX 77341-2206
[email protected]
Chris O’Neill
University of California at Davis
Department of Mathematics
One Shield Ave.
Davis, CA 95616
[email protected]
People just want more of it. - Ray Kroc 2
1
Introduction
Every day, 34 million Chicken McNuggets are sold worldwide.3 At most
McDonalds locations in the United States today, Chicken McNuggets are
sold in packs of 4, 6, 10, 20, 40, and 50 pieces. However, shortly after their
introduction in 1979 they were sold in packs of 6, 9, and 20. The following
problem spawned from the use of these latter three numbers.
The Chicken McNugget Problem. What numbers of Chicken McNuggets
can be ordered using only packs with 6, 9, or 20 pieces?
1
This article is based on a 2013 PURE Mathematics REU Project by Emelie Curl,
Staci Gleen, and Katrina Quinata which was directed by the authors and Roberto Pelayo.
2
https://www.brainyquote.com/quotes/authors/r/ray_kroc.html
3
http://www.answers.com/Q/How_many_nuggets_does_mcdonalds_sell_a_day?
1
Early references to this problem can be
found in [26, 30]. Positive integers satisfying the Chicken McNugget Problem are
now known as McNugget numbers [22]. In
particular, if n is a McNugget number, then
there is an ordered triple (a, b, c) of nonnegative integers such that
6a + 9b + 20c = n.
(1)
We will call (a, b, c) a McNugget expansion
Figure 1: The 6 piece box
of n (again see [22]). Since both (3, 0, 0) and
(0, 2, 0) are McNugget expansions of 18, it is clear that McNugget expansions
are not unique. This phenomenon will be the central focus of the remainder
of this article.
If max{a, b, c} ≥ 8 in (1), then n ≥ 48 and hence determining the numbers x with 0 ≤ x ≤ 48 which are McNugget numbers can be checked either
by hand or your favorite computer algebra system. The only such x’s which
are not McNugget numbers are: 1, 2, 3, 4, 5, 7, 8, 10, 11, 13, 14, 16, 17, 19,
22, 23, 25, 28, 31, 34, 37, and 43. (The non-McNugget numbers are sequence
A065003 in the On-Line Encyclopedia of Integer Sequences4 .) We demonstrate this in Table 1 with a chart that offers the McNugget expansions
(when they exist) of all numbers ≤ 50.
What happens with larger values? Table 1 has already verified that 44,
45, 46, 47, 48, and 49 are McNugget numbers. Hence, we have a sequence
of 6 consecutive McNugget numbers, and by repeatedly adding 6 to these
values, we obtain the following.
Proposition 1.1. Any x > 43 is a McNugget number.
Thus, 43 is the largest number of McNuggets that cannot be ordered with
packs of 6, 9, and 20.
Our aim in this paper is to consider issues related to the multiple occurances of McNugget expansions as seen in Table 1. Such investigations fall
4
http://oeis.org/A065003.
2
#
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
(a, b, c)
(0, 0, 0)
NONE
NONE
NONE
NONE
NONE
(1, 0, 0)
NONE
NONE
(0, 1, 0)
NONE
NONE
(2, 0, 0)
NONE
NONE
(1, 1, 0)
NONE
#
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
(a, b, c)
NONE
(3, 0, 0) (0, 2, 0)
NONE
(0, 0, 1)
(2, 1, 0)
NONE
NONE
(4, 0, 0) (1, 2, 0)
NONE
(1, 0, 1)
(0, 3, 0) (3, 1, 0)
NONE
(0, 1, 1)
(5, 0, 0) (2, 2, 0)
NONE
(2, 0, 1)
(1, 3, 0) (4, 1, 0)
#
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
(a, b, c)
NONE
(1, 1, 1)
(0, 4, 0) (3, 2, 0) (6, 0, 0)
NONE
(0, 2, 1) (3, 0, 1)
(2, 3, 0) (5, 1, 0)
(0, 0, 2)
(2, 1, 1)
(1, 4, 0) (4, 2, 0) (7, 0, 0)
NONE
(1, 2, 1) (4, 0, 1)
(0, 5, 0) (3, 3, 0) (6, 1, 0)
(1, 0, 2)
(0, 3, 1) (3, 1, 1)
(2, 4, 0) (5, 2, 0) (8, 0, 0)
(0, 1, 2)
(2, 2, 1) (5, 0, 1)
Table 1: The McNugget numbers and their expansions from 0 to 50
under the more general purview of the theory of non-unique factorizations
in integral domains and monoids (a good technical reference on this subject
is [20]). Using a general context, we show in Section 2 that the McNugget
numbers form an additive monoid and discuss some properites shared by
the class of additive submonoids of the nonnegative integers. In Section 3,
we define the particular combinatorial characteristics of McNugget expansions which we will consider. Computations of these characteristics for the
McNugget Monoid will appear in Section 4.
By emphasizing results concerning McNugget numbers, we offer the
reader a glimpse into vast literature surrounding non-unque factorizations.
While we stick to the calculation of basic factorization invariants, our results
indicate that such computations involve a fair amount of complexity. Many
of the results we touch on have appeared in papers authored or co-authored
3
by undergraduates in National Science Foundation Sponsored REU Programs. This is an area which remains rich in open problems, and we hope
our discussion here spurs our readers (both young and old) to explore this
rewarding subject more deeply.
2
A brief diversion into generality
As illustrated above, Chicken McNugget numbers fit into a long studied
mathematical concept. Whether called the Postage Stamp Problem [25],
the Coin Problem [15], or the Knapsack Problem [21], the idea is as follows.
Given a set of k objects with predetermined values n1 , n2 , . . . , nk , what
possible values of n can be had from combinations of these objects? Thus, if
a value of n can be obtained, then there is an ordered k-tuple of nonnegative
integers (x1 , . . . , xk ) which satisfy the linear diophatine equation
n = x1 n1 + x2 n2 + · · · + xk nk .
(2)
We view this in a more algebraic manner. Given integers n1 , . . . , nk > 0, set
hn1 , . . . , nk i = {x1 n1 + · · · + xk nk | x1 , . . . , xk ∈ N0 }.
Notice that if s1 and s2 are in hn1 , . . . , nk i, then s1 +s2 is also in hn1 , . . . , nk i
(where + represents regular addition). Since 0 ∈ hn1 , . . . , nk i and + is an
associative operation, the set hn1 , . . . , nk i under + forms a monoid. Monoids
of nonnegative integers under addition, like the one above, are known as
numerical monoids, and n1 , . . . , nk are called generators. We will call the
numerical monoid h6, 9, 20i the Chicken McNugget monoid, and denote it
by .
Since
consists of the same elements as h6, 9, 20, 27i, it is clear that
generating sets are not unique. Using elementary number theory, it is easy
to argue that any numerical monoid hn1 , . . . , nk i does have a unique generating set with minimal cardinality, obtained by eliminating those generators
ni that lie in the numerical monoid generated by {n1 , . . . , nk } − {ni }. In
this way, it is clear that 6, 9, 20 is indeed the minimal generating set of .
4
When dealing with a general numerical monoid hn1 , . . . , nk i, we will assume
without loss of generality that given generating set n1 , . . . , nk is minimal.
In view of this broader setting, the Chicken McNugget Problem can be
generalized as follows.
The Numerical Monoid Problem. If n1 , . . . , nk are positive integers,
then which nonnegative integers lie in hn1 , . . . , nk i?
Example 2.1. We have already determined above exactly which nonnegative integers are McNugget numbers. Suppose the Post Office issues stamps
in denominations of 4 cents, 7 cents, and 10 cents. What values of postage
can be placed on a letter (assuming that as many stamps as necessary can
be placed on the envelope)? In particular, we are looking for the elements
of h4, 7, 10i. We can again use brute force to find all the solutions to
4a + 7b + 10c = n
and conclude that 1, 2, 3, 5, 6, 9, and 13 cannot be obtained. Since 14, 15,
16, and 17 can, all postage values larger than 13 are possible.
2
Figure 2: The 9 piece box
Let’s return to the largest number of
McNuggets that can’t be ordered (namely,
43) and the companion number 13 obtained
in Example 2.1. The existence of these
numbers is no accident. To see this in general, let n1 , . . . , nk be a set of positive integers which are relatively prime. By elementary number theory, there is a set y1 , . . . , yk
of (possibly negative) integers such that
1 = y1 n 1 + · · · + yk n k .
By choosing an element V = x1 n1 +· · ·+xk nk ∈ hn1 , . . . , nk i with sufficiently
large coefficients (for instance, if each xi ≥ n1 |yi |), we see V + 1, . . . , V + n1
all lie in hn1 , . . . , nk i as well. As such, any integers greater than V can be
obtained in hn1 , . . . , nk i by adding copies of n1 .
This motivates the following definition.
5
Definition 2.2. If n1 , . . . , nk are relatively prime positive integers, then
the Frobenius number of hn1 , . . . , nk i, denoted F (hn1 , . . . , nk i), is the largest
positive integer n such that n ∈
/ hn1 , . . . , nk i.
We have already shown that F ( ) = 43
and F (h4, 7, 10i) = 13. A famous result
of Sylvester from 1884 [29] states that if a
and b are relatively prime, then F (ha, bi) =
ab − a − b (a nice proof of this can be
found in [6]). This is where the fun begins, as strictly speaking no formula exists to compute the Frobenius number of
numerical monoids that require 3 or more
generators. While there are fast algorithms
Figure 3: The 20 piece box
which can compute F (hn1 , n2 , n3 i) (see for
instance [17]), at best formulas for F (hn1 , . . . , nk i) exist only in special cases
(you can find one such special case where F ( ) = 43 pops out in [1, p. 14]).
Our purpose is not to compile or expand upon the vast literature behind the
Frobenius number; in fact, we direct the reader to the excellent monograph
of Ramı́rez Alfonsı́n [27] for more background reading on the Diophatine
Frobenius Problem.
3
The McNugget factorization toolkit
We focus now on the multiple McNugget expansions we saw in Table 1.
In particular, notice that there are McNugget numbers which have unique
triples associated to them (6, 9, 12, 15, 20, 21, 26, 29, 32, 35, 40, 41, 46, and
49), some of which have two (18, 24, 27, 30, 33, 35, 39, 44, 47, and 50), and
even some which have three (36, 42, 45, and 48). While the “normal” notion
of factoring occurs in systems where multiplication prevails, notice that the
ordered triples representing McNugget numbers are actually factorizations
of these numbers into “additive” factors of 6, 9, and 20.
Let’s borrow some terminology from abstract algebra ([18] is a good
beginning reference on the topic). Let x and y ∈ hn1 , . . . , nk i. We say
6
that x divides y in hn1 , . . . , nk i if there exists a z ∈ hn1 , . . . , nk i such that
y = x+z. We call a nonzero element x ∈ hn1 , . . . , nk i irreducible if whenever
x = y + z, either y = 0 or z = 0. (Hence, x is irreducible if its only proper
divisors are 0 and itself). Both of these definitions are obtained from the
usual “multiplicative” definition by replacing “·” with “+” and 1 with 0.
We leave the proof of the following to the reader.
Proposition 3.1. If hn1 , . . . , nk i is a numerical monoid, then its irreducible
elements are precisely n1 , . . . , nk .
Related to irreducibility is the notion of prime elements. A nonzero
element x ∈ hn1 , . . . , nk i is prime if whenever x divides a sum y + z, then
either x divides y or x divides z (this definition is again borrowed from the
multiplicative setting). It is easy to check from the definitions that prime
elements are always irreducible, but it turns out that in general irreducible
elements need not be prime. In fact, the irreducible elements n1 , . . . , nk
of a numerical monoid are never prime. To see this, let ni be an irreducible
element and let T be the numerical monoid generated by {n1 , . . . , nk }−{ni }.
Although ni ∈
/ T , some multiple of ni must lie in T (take, for instance, n2 ni ).
P
Let kn = j6=i xj nj (for some k > 1) be the smallest multiple of ni in T .
P
Then n divides
j6=i xj nj over hn1 , . . . , nk i, but by the minimality of k,
n does not divide any proper subsum. Thus ni is not prime.
For our purposes, we restate Proposition 3.1 in terms of .
Corollary 3.2. The irreducible elements of the McNugget monoid are 6, 9,
and 20. There are no prime elements.
3.1
The set of factorizations of an element
We refer once again to the elements in Table 1 with multiple irreducible
factorizations. For each x ∈ , let
Z(x) = {(a, b, c) | 6a + 9b + 20c = x}.
We will refer to Z(x) as the complete set of factorizations x in , and as
such, we could relabel columns 2, 4, and 6 of Table 1 as “Z(x).” While we
7
will not dwell on general structure problems involving Z(x), we do briefly
address one in the next example.
Example 3.3. What elements x in the McNugget monoid are uniquely
factorable (i.e., |Z(x)| = 1)? A quick glance at Table 1 yields 14 such
nonzero elements (namely, 6, 9, 12, 15, 20, 21, 26, 29, 32, 35, 40, 41, 46, 49).
Are there others? We begin by noting in Table 1 that
(3, 0, 0), (0, 2, 0) ∈ Z(18)
and
(10, 0, 0), (0, 0, 3) ∈ Z(60).
This implies that in any factorization in , 3 copies of 6 can be freely replaced
with 2 copies of 9 (this is called a trade). Similarly, 2 copies of 9 can be
traded for 3 copies of 6, and 3 copies of 20 can be traded for 10 copies of 6.
In particular, for n = 6a + 9b + 20c ∈ , if either a ≥ 3, b ≥ 2 or c ≥ 3,
then n has more than one factorization in . As such, if n is to have unique
factorization, then 0 ≤ a ≤ 2, 0 ≤ b ≤ 1, and 0 ≤ c ≤ 2. This leaves
18 possibilities, and a quick check yields that the 3 missing elements are
52 = (2, 0, 2), 55 = (1, 1, 2) and 61 = (2, 1, 2).
2.
The argument in Example 3.3 easily generalizes – every numerical monoid
which requires more than one generator has finitely many elements that factor uniquely – but note that minimal trades need not be as simple as replacing a multiple of one generator with a multiple of another. Indeed, in the
numerical monoid h5, 7, 9, 11i, there is a trade (1, 0, 0, 1), (0, 1, 1, 0) ∈ Z(16),
though 16 is not a multiple of any generator. Determining the “minimal”
trades of a numerical monoid, even computationally, is known to be a very
hard problem in general [28].
3.2
The length set of an element and related invariants
Extracting information from the factorizations of numerical monoid elements
(or even simply writing them all down) can be a tall order. To this end,
combinatorially-flavored factorization invariants are often used, assigning
to each element (or to the monoid as a whole) a value measuring its failure
to admit unique factorization. We devote the remainder of this paper to
examining several factorization invariants, and what they tell us about the
McNugget monoid as compared to more general numerical monoids.
8
We begin by considering a set, derived from the set of factorizations,
that has been the focus of many papers in the mathematical literature over
the past 30 years. If x ∈
and (a, b, c) ∈ Z(x), then the length of the
factorization (a, b, c) is denoted by
|(a, b, c)| = a + b + c.
We have shown earlier that factorizations in may not be unique, and a
quick look at Table 1 shows that their lengths can also differ. For instance,
42 has three different factorizations, with lengths 5, 6 and 7, respectively.
Thus, we denote the set of lengths of x in by
L(x) = {|(a, b, c)| : (a, b, c) ∈ Z(x)}.
In particular, L(42) = {5, 6, 7}. Moreover, set
`(x) = min L(x)
and
L(x) = max L(x).
(In our setting, it is easy to argue that L(x) must be finite, so the maximum
and minimum above are both well-defined.) To give the reader a feel for
these invariants, in Table 2 we list all the McNugget numbers from 1 to 50
and their associated values L(x), `(x), and L(x).
The following recent result describes the functions L(x) and `(x) for
elements x ∈ hn1 , . . . , nk i that are sufficiently large with respect to the
generators. Intuitively, Theorem 3.4 says that for “most” elements x, any
factorization with maximal length is almost entirely comprised of n1 , so
L(x + n1 ) is obtained by taking a maximum length factorization for x and
adding one additional copy of n1 . In general, the “sufficiently large” hypothesis is needed, since, for example, both 41 = 2 · 9 + 1 · 23 and 50 = 5 · 10
are maximum length factorizations in the numerical monoid h9, 10, 23i.
Theorem 3.4 ([4, Theorems 4.2 and 4.3]). Suppose hn1 , . . . , nk i is a numerical monoid. If x > n1 nk , then
L(x + n1 ) = L(x) + 1,
and if x > nk−1 nk , then
`(x + nk ) = `(x) + 1.
9
x
0
6
9
12
15
18
20
21
24
26
L(x) `(x) L(x)
{0}
0
0
{1}
1
1
{1}
1
1
{2}
2
2
{2}
2
2
{2, 3}
2
3
{1}
1
1
{3}
3
3
{3, 4}
3
4
{2}
2
2
x
27
29
30
32
33
35
36
38
39
40
L(x)
{3, 4}
{2}
{4, 5}
{3}
{4, 5}
{3}
{4, 5, 6}
{3, 4}
{5, 6}
{2}
`(x) L(x)
3
4
2
2
4
5
3
3
4
5
3
3
4
6
3
4
5
6
2
2
x
L(x)
41
{4}
42 {5, 6, 7}
44 {4, 5}
45 {5, 6, 7}
46
{3}
47 {4, 5}
48 {6, 7, 8}
49
{3}
50 {5, 6}
`(x) L(x)
4
4
5
7
4
5
5
7
3
3
4
5
6
8
3
3
5
6
Table 2: The McNugget numbers from 0 to 50 with L(x), `(x), and L(x)
We will return to this result in Section 4.1, where we give a closed formula
for L(x) and `(x) that holds for all x ∈ .
Given our definitions to this point, we can now mention perhaps the
most heavily studied invariant in the theory of non-unique factorizations.
For x ∈ hn1 , . . . , nk i, the ratio
ρ(x) =
L(x)
,
`(x)
is called the elasticity of x, and
ρ(hn1 , . . . , nk i) = sup{ρ(x) | x ∈ hn1 , . . . , nk i}
is the elasticity of hn1 , . . . , nk i. The elasticity of an element n ∈ hn1 , . . . , nk i
measures how “spread out” its factorization lengths are; the larger ρ(n) is,
the more spread out L(n) is. To this end, the elasticity ρ(hn1 , . . . , nk i)
encodes the highest such “spread” throughout the entire monoid. For example, if ρ(hn1 , . . . , nk i) = 2, then the maximum factorization length of any
element n ∈ hn1 , . . . , nk i is at most twice its minimum factorization length.
A formula for the elasticity of a general numerical monoid, given below,
was given in [10], and was the result of an undergraduate research project.
10
3
2.5
2
1.5
1
0.5
100
200
300
400
Figure 4: A plot depicting the elasticity function ρ(n) for n ∈ .
Theorem 3.5 ([10], Theorem 2.1 and Corollary 2.3). The elasticity of the
numerical monoid hn1 , . . . , nk i is
ρ(hn1 , . . . , nk i) =
nk
.
n1
Moreover, ρ(n) = nnk1 precisely when n is an integer multiple of the least
common multiple of n1 and nk , and for any rational r < nnk1 , there are only
finitely many elements x ∈ hn1 , . . . , nk i with ρ(x) ≤ r.
The significance of the final statement in Theorem 3.5 is that there are
rationals 1 ≤ q ≤ nnk1 which do not lie in the set {ρ(x) | x ∈ hn1 , . . . , nk i}
and hence {ρ(x) | x ∈ hn1 , . . . , nk i} ( Q ∩ [1, nnk1 ] (to use terminology from
the literature, numerical monoids are not fully elastic). Figure 4 depicts
the elasticities of elements of up to n = 400; indeed, as n increases, the
elasticity ρ(n) appears to converge to 10
3 = ρ( ). In general, the complete
image {ρ(x) | x ∈ hn1 , . . . , nk i} has been determined by Barron, O’Neill, and
Pelayo in another student co-authored paper [4, Corollary 4.5]; we direct the
reader there for a thorough mathematical description of Figure 4.
We close our discussion of elasticity with the following.
11
Corollary 3.6. The elasticity of the McNugget monoid is
ρ( ) =
10
.
3
While a popular invariant to study, the elasticity only tells us about the
largest and smallest elements of L(x). Looking at Table 2, it appears that
the length sets of the first few McNugget numbers are uniformly constructed
(each is of the form [a, b] ∩ N for positive integers a and b). One need not
look too much further to break this pattern; the element 60 ∈ has
Z(60) = {(0, 0, 3), (1, 6, 0), (4, 4, 0), (7, 2, 0), (10, 0, 0)}
and thus
L(60) = {3, 7, 8, 9, 10}.
This behavior motivates the following “finer” factorization invariant. Fix
x ∈ hn1 , . . . , nk i, and let L(x) = {m1 , . . . , mt } with m1 < m2 < · · · < mt .
Define the delta set of x as
∆(x) = {mi − mi−1 | 2 ≤ i ≤ t},
and the delta set of hn1 , . . . , nk i as
∆(hn1 , . . . , nk i) =
[
∆(x).
x∈hn1 ,...,nk i
The study of the Delta sets of numerical monoids (and more generally, of
cancellative commutative monoids) has been an extremely popular topic;
many such papers feature results from REU programs (see, for instance,
[7, 8, 9, 11, 12, 14]).
From Table 1 we see that the McNugget numbers from 1 to 50 all have
Delta set ∅ or {1}, and we have further showed that ∆(60) = {1, 4}. What
is the Delta set of and moreover, what possible subsets of this set occur as
∆(x) for some x ∈ ? We will address those questions is Section 4.2, with
the help of a result from [11], stated below as Theorem 3.7.
One of the primary difficulties in determining the set ∆(hn1 , . . . , nk i) is
that even though each element’s delta set ∆(x) is finite, the definition of
12
∆(hn1 , . . . , nk i) involves the union of infinitely many such sets. The key
turns out to be a description of the sequence {∆(x)}x∈hn1 ,...,nk i for large x
(note that this is a sequence of sets, not integers). Baginski conjectured
during the writing of [7] that this sequence is eventually periodic, and three
years later this was settled in the affirmative, again in an REU project.
Theorem 3.7 ([11, Theorem 1 and Corollary 3]). For x ∈ hn1 , . . . , nk i,
∆(x) = ∆(x + n1 nk )
whenever x > 2kn2 n2k . In particular,
∆(hn1 , . . . , nk i) =
[
∆(x)
x∈D
where D = {x ∈ hn1 , . . . , nk i | x ≤ 2kn2 n2k + n1 nk } is a finite set.
Thus ∆(hn1 , . . . , nk i) can be computed in finite time. The bound given
in Theorem 3.7 is far from optimal; it is drastically improved in [19], albeit
with a much less concise formula. For convenience, we will use the bound
given above in our computation of ∆( ) in Section 4.2.
3.3
Beyond the length set
We remarked earlier that no element of a numerical monoid is prime. Let’s
consider this more closely in . For instance, since 6 is not prime, there is
a sum x + y in such that 6 divides x + y, but 6 does not divide x nor
does 6 divide y (take, for instance, x = y = 9). But note that 6 satisfies the
following slightly weaker property. Suppose that 6 divides a sum x1 +· · ·+xt
where t > 3. Then there is a subsum of at most 3 of the xi ’s that 6 does
divide. To see this, notice that if 6 divides any of the xi ’s, then we are
done. So suppose it does not. If 9 divides both xi and xj , then 6 divides
xi + xj since 6 divides 9 + 9. If no two xi ’s are divisible by 9, then at least 3
xi ’s are divisible by 20, and nearly identical reasoning to the previous case
completes the argument. This value of 3 offers some measure as to how far
6 is from being prime, and motivates the following definition.
13
Definition 3.8. Let hn1 , . . . , nk i be a numerical monoid. For any nonzero
x ∈ hn1 , . . . , nk i, define ω(x) = m if m is the smallest positive integer such
that whenever x divides x1 + · · · + xt , with xi ∈ hn1 , . . . , nk i, then there is
P
a set T ⊂ {1, 2, . . . , t} of indices with |T | ≤ m such that x divides i∈T xi .
Using Definition 3.8, a prime element would have ω-value 1, so ω(x) can
be interpreted as a measure of how far x is from being prime. In , we
argued that ω(6) = 3; a similar argument yields ω(9) = 3 and ω(20) = 10.
Notice that the computation of ω(x) is dependent more on Z(x) than L(x),
and hence encodes much different information than either ρ(x) or ∆(x).
Let us more closely examine the argument that ω(6) = 3. The key is that
6 divides 9 + 9 and 20 + 20 + 20, but does not divide any subsum of either.
Indeed, the latter of these expressions yields a lower bound of ω(6) ≥ 3,
and the given argument implies that equality holds. With this in mind, we
give the following equivalent form of Definition 3.8, which often simplifies
computation of ω(x).
Theorem 3.9 ([24, Proposition 2.10]). Suppose hn1 , . . . , nk i is a numerical
monoid and x ∈ hn1 , . . . , nk i. The following conditions are equivalent.
(a) ω(x) = m.
(b) m is the maximum length of a sum x1 + · · · + xt of irreducible elements
in hn1 , . . . , nk i with the property that (i) x divides x1 + · · · + xt , and
(ii) x does not divide x1 + · · · + xj−1 + xj+1 + · · · + xt for 1 ≤ j ≤ t.
The sum x1 + · · · + xt alluded to in part (b) above is called a bullet for x.
Hence, 20 + 20 + 20 is a bullet for 6 in , and moreover has maximal length.
The benefit of Theorem 3.9 is twofold: (i) each x ∈ hn1 , . . . , nk i has only
finitely many bullets, and (ii) the list of bullets can be computed in a similar
fashion to the set Z(x) of factorizations. We refer the reader to [3, 5], both of
which give explicit algorithms (again resulting from undergraduate research
projects) for computing ω-values.
Our goal is to completely describe the behavior of the ω-function of the
McNugget Monoid. We do so in Section 4.3, using the following result,
which is clearly similar in spirit to Theorems 3.4 and 3.7.
14
Theorem 3.10 ([23, Theorem 3.6]). For x ∈ hn1 , . . . , nk i sufficiently large,
ω(x + n1 ) = ω(x) + 1.
In particular, this holds for
x>
F + n2
n2 /n1 − 1
where F = F (hn1 , . . . , nk i) is the Frobenius number.
The similarity between Theorems 3.4 and 3.10 is not a coincidence.
While L(x) and ω(x) are indeed different functions (for instance, L(6) = 1
while ω(6) = 3), they are closely related; the ω-function can be expressed in
terms of max factorization length that is computed when some collections
of generators are omitted. We direct the interested reader to [5, Section 6],
where an explicit formula of this form for ω(n) is given.
3.4
Computer software for numerical monoids
Many of the computations referenced in this paper can be performed using the numericalsgps package [16] for the computer algebra system GAP.
The brief snippet of sample code below demonstrates how the package is
used to compute various quantities discussed in this paper.
gap> LoadPackage("num");
true
gap> McN:=NumericalSemigroup(6,9,20);
<Numerical semigroup with 3 generators>
gap> FrobeniusNumberOfNumericalSemigroup(McN);
43
gap> 43 in McN;
false
gap> 44 in McN;
true
gap> FactorizationsElementWRTNumericalSemigroup(18,McN);
[ [ 3, 0, 0 ], [ 0, 2, 0 ] ]
15
gap> OmegaPrimalityOfElementInNumericalSemigroup(6,McN);
3
This only scratches the surface of the extensive functionality offered by
the numericalsgps package. We encourage the interested reader to install
and experiment with the package; instructions can be found on the official
webpage, whose URL is included below.
https://www.gap-system.org/Packages/numericalsgps.html
4
Calculations for the Chicken McNugget monoid
In the final section of this paper, we give explicit expressions for L(x), `(x),
∆(x) and ω(x) for every x ∈ . The derivation of each such expression
makes use of a theoretical result in Section 3.
We note that each of the formulas provided in this section could also be
derived in a purely computational manner, using Theorems 3.4, 3.7, and 3.10
and the inductive algorithms introduced in [5] (indeed, these computations
finish in a reasonably short amount of time using the implementation in
the numericalsgps package discussed in Section 3.4). However, several of
the following results identify an interesting phenomenon that distinguish
from more general numerical monoids (see the discussion preceeding Question 4.5), and the arguments that follow give the reader an idea of how
theorems involving factorization in numerical monoids can be proven.
4.1
Calculating factorization lengths
Theorem 3.4 states that L(x + n1 ) = L(x) + 1 and `(x + nk ) = `(x) + 1 for
sufficiently large x ∈ hn1 , . . . , nk i. but, it was observed during the writing
of [4] that for many numerical monoids, the “sufficiently large” requirement
is unecessary. As it turns out, one such example is the McNugget monoid ,
which we detail below.
16
Theorem 4.1. For each x ∈ , L(x + 6) = L(x) + 1. In particular, if we
write x = 6q + r for q, r ∈ N and r < 6, then
q
if r = 0 or 3
q−5
if r = 1
L(x) =
q−2
if r = 2 or 5
q−4
if r = 4
for each x ∈ .
and a factorization (a, b, c) of x. If b > 1, then x has
Proof. Fix x ∈
another factorization (a + 3, b − 2, c) with length a + b + c + 1. Similarly,
if c ≥ 3, then (a + 10, b, c − 3) is also a factorization of x and has length
a + b + c + 7. This implies that if (a, b, c) has maximum length among
factorizations of x, then b ≤ 1 and c ≤ 2. Upon inspecting Table 1, we see
that unless x ∈ {0, 9, 20, 29, 40, 49}, we must have a > 0.
Now, assume (a, b, c) has maximum length among factorizations of x.
We claim (a + 1, b, c) is a factorization of x + 6 with maximum length. From
/ {0, 9, 20, 29, 40, 49},
Table 1, we see that since x ∈ , we must have x + 6 ∈
meaning any maximum length factorization of x + 6 must have the form
(a0 + 1, b0 , c0 ). This yields a factorization (a0 , b0 , c0 ) of x, and since (a, b, c)
has maximum length, we have a + b + c ≥ a0 + b0 + c0 . As such, (a + 1, b, c)
is at least as long as (a0 + 1, b0 , c0 ), and the claim is proved. Thus,
L(x + 6) = a + 1 + b + c = L(x) + 1.
From here, the given formula for L(x) now follows from the first claim
and the values L(0), L(9), L(20), L(29), L(40), and L(49) in Table 3.2.
A similar expression can be obtained for `(x), ableit with 20 cases instead
of 6, this time based on the value of x modulo 20. We encourage the reader
to adapt the argument above for Theorem 4.2.
Theorem 4.2. For each x ∈ , `(x + 20) = `(x) + 1. In particular, if we
17
write x = 20q + r for q, r ∈ N and
q
q+1
q+2
`(x) =
q+3
q+4
q+5
r < 20, then
if
if
if
if
if
if
r
r
r
r
r
r
=0
= 6, 9
= 1, 4, 7, 12, 15, 18
= 2, 5, 10, 13, 16
= 8, 11, 14, 19
= 3, 17
for each x ∈ .
Theorems 4.1 and 4.2 together yield a closed form for ρ(x) that holds for
all x ∈ . Since lcm(6, 20) = 60 cases are required, we leave the construction
of this closed form to the interested reader.
4.2
Calculating delta sets
Unlike maximum and minimum factorization length, ∆(x) is only periodic
for sufficiently large x ∈ . For example, a computer algebra system can
be used to check that ∆(91) = {1} while ∆(211) = {1, 2}. Theorem 3.7
only guarantees ∆(x + 120) = ∆(x) for x > 21600, but some considerable
reductions can be made. In particular, we will reduce the period from 120
down to 20, and will show that equality holds for all x ≥ 92 (that is to say,
91 is the largest value of x for which ∆(x + 20) 6= ∆(x)).
Theorem 4.3. Each x ∈ with x ≥ 92 has ∆(x + 20) = ∆(x). Moreover,
if r = 3, 8, 14, 17
{1}
{1, 2}
if r = 2, 5, 10, 11, 16, 19
∆(x) =
{1, 3}
if r = 1, 4, 7, 12, 13, 18
{1, 4}
if r = 0, 6, 9, 15
where x = 20q + r for q, r ∈ N and r < 20. Hence ∆( ) = {1, 2, 3, 4}.
Proof. We will show that ∆(x+20) = ∆(x) for each x > 103. The remaining
claims can be verified by extending Table 2 using computer software.
18
Suppose x > 103, fix a factorization (a, b, c) for x, and let l = a + b + c.
If c ≥ 3, then x also has factorizations (a + 10, b, c − 3), (a + 7, b + 2, c − 3),
(a + 4, b + 4, c − 3), and (a + 1, b + 6, c − 3), meaning
{l, l + 4, l + 5, l + 6, l + 7} ⊂ L(x).
Alternatively, since x > 103, if c ≤ 2, then 6a + 9b ≥ 63, and thus
l ≥ a + b + 2 ≥ 9 ≥ `(x) + 4.
The above arguments imply (i) any gap in successive lengths in L(x) occurs
between `(x) and `(x) + 4, and (ii) every factorization with length in that
interval has at least one copy of 20. As such, x + 20 has the same gaps
between `(x + 20) and `(x + 20) + 4 as x does between `(x) and `(x) + 4,
which proves ∆(x + 20) = ∆(x) for all x > 103.
With a slightly more refined argument than the one given above, one
can prove without the use of software that ∆(x + 20) = ∆(x) for all x ≥ 92.
We encourage the interested reader to work out such an argument.
4.3
Calculating ω-primality
We conclude our study of with an expression for the ω-primality of x ∈
and show (in some sense) how far a McNugget number is from being prime.
We proceed in a similar fashion to Theorems 4.1 and 4.2, showing that with
only two exceptions, ω(x + n1 ) = ω(x) + 1 for all x ∈ .
Theorem 4.4. With the exception of x = 6 and x = 12, every nonzero
x ∈ satisfies ω(x + 6) = ω(x) + 1. In particular, we have
q
if r = 0
q+5
if r = 1
q+7
if r = 2
ω(x) =
q+2
if r = 3
q+4
if r = 4
q+9
if r = 5
for each x 6= 6, 12, where x = 6q + r for q, r ∈ N and r < 6.
19
Proof. Fix x ∈ . Following the spirit of the proof of Theorem 4.1, we begin
by proving each x > 12 has a maximum length bullet (a, b, c) with a > 0.
Indeed, suppose (0, b, c) is a bullet for x for some b, c ≥ 0. The element x ∈
also has some bullet of the form (a0 , 0, 0), where a0 the smallest integer such
that 6a0 − x ∈ . Notice a0 ≥ 3 since x > 12. We consider several cases.
/ . If b ≤ 3, then a0 ≥ b.
• If c = 0, then 9b − x ∈ but 9b − x − 9 ∈
Otherwise, either 9(b − 1) or 9(b − 2) is a multiple of 6, and since
9(b − 2) − x ∈
/ as well, we see a0 ≥ 23 (b − 2) + 1 ≥ b.
• If b = 0, then there are two possibilities. If c ≤ 3, then a0 ≥ c.
Otherwise, either 20(c − 1), 20(c − 2) or 20(c − 3) is a multiple of 6,
so we conclude a0 ≥ 10
3 (c − 3) + 1 ≥ c.
• If b, c > 0, then 9b + 20c − x − 9, 9b + 20c − x − 20 ∈
/ , so 9b + 20c − x
is either 0, 6, or 12. This means either (3, b − 1, c), (2, b − 1, c), or
(1, b − 1, c) is also a bullet for x, respectively.
In each case, we have constructed a bullet for x at least as long as (0, b, c),
but with positive first coordinate, so we conclude x has a maximal bullet
with nonzero first coordinate.
Now, using a similar argument to that given in the proof of Theorem 4.1,
if (a+1, b, c) is a maximum length bullet for x+6, then (a, b, c) is a maximum
length bullet for x. This implies ω(x + 6) = ω(x) + 1 whenever x + 6 has
a maximum length bullet with positive first coordinate, which by the above
argument holds whenever x > 12. This proves the first claim.
The formula for ω(x) now follows from the first claim, the computations
ω(9) = 3 and ω(20) = 10 from Section 3.3, and analogous computations for
ω(15) = 4, ω(18) = 3, ω(29) = 13, ω(40) = 10, and ω(49) = 13.
Figure 5 plots ω-values of elements of the McNugget monoid . Since
ω(x + 6) = ω(x) + 1 for large x ∈ , most of the plotted points occur on one
of 6 lines with slope 16 . It is also evident in the plot that x = 6 and x = 12
are the only exceptions.
Although ω(x + 6) = ω(x) + 1 does not hold for every x ∈ , there
are some numerical monoids for which the “sufficiently large” hypothesis in
20
20
15
10
5
0
20
40
60
80
100
Figure 5: A plot depicting the ω-primality function ω(n) for n ∈ .
Theorem 3.10 can be dropped (for instance, any numerical monoids with 2
minimal generators has this property). Hence, we conclude with a problem
suitable for attack by undergraduates.
Question 4.5. Determine which numerical monoids hn1 , . . . , nk i satisfy
each of the following conditions for all x (i.e. not just sufficiently large x):
1. L(x + n1 ) = L(x) + 1,
2. `(x + nk ) = `(x) + 1, or
3. ω(x + n1 ) = ω(x) + 1.
References
[1] G. Alon, P. Clark, On the number of representations of an integer by a
linear form, J. Integer Sequences 8(2005) 05.5.2.
[2] D. F. Anderson, S. T. Chapman, On bounding measures of primeness
in integral domains, Inter. J. Algebra Comp. 22(2012) 1250040.
21
[3] D. F. Anderson, S. T. Chapman, N. Kaplan, D. Torkornoo, An algorithm to compute ω-primality in a numerical monoid, Semigroup Forum
82(2011) 96–108.
[4] T. Barron, C. O’Neill, R. Pelayo, On the set of elasticities in numerical
monoids, Semigroup Forum 94 no. 1 (2017) 37–50.
[5] T. Barron, C. O’Neill, R. Pelayo, On dymamic algorithms for factorization invariants in numerical monoids, Mathematics of Computation
86(2017) 2429–2447.
[6] M Beck, How to change coins, M&M’s, or Chicken McNuggets: the linear Diophatine problem of Frobenius, http://math.sfsu.edu/beck/
papers/frobprojects.pdf.
[7] C. Bowles, S. T. Chapman, N. Kaplan, D. Reiser. On delta sets of
numerical monoids. Journal of Algebra and its Applications 5(2006)
695–718.
[8] S. T. Chapman, J. Daigle, R. Hoyer, N. Kaplan, Delta sets of numerical
monoids using nonminimal sets of generators, Comm. Algebra 38(2010)
2622–2634.
[9] S. T. Chapman, P. A. Garcı́a-Sánchez, D. Llena, A. Malyshev, D. Steinberg, On the delta set and the Betti elements of a BF-monoid, Arabian
J. of Mathematics 1(2012) 53–61.
[10] S. T. Chapman, M. T. Holden, T. A. Moore. Full elasticity in atomic
monoids and integral domains, Rocky Mountain J. Math 37(2006)
1437–1455.
[11] S. T. Chapman, R. Hoyer, N. Kaplan, Delta sets of numerical monoids
are eventually periodic, Aequationes Mathematicae 77(2009) 273–279.
[12] S. T. Chapman, N. Kaplan, T. Lemburg, A. Niles, C. Zlogar, Shifts of
generators and delta sets of numerical monoids, Inter. J. Algebra and
Computation 24(2014) 655–669.
22
[13] S. T. Chapman, W. Puckett, K. Shour, On the omega values of generators of embedding dimension-three numerical monoids generated by an
interval, Involve 7(2014) 657–667.
[14] S. Colton, N. Kaplan, The realization problem for delta sets of numerical semigroups, J. Comm. Algebra, to appear.
[15] https://en.wikipedia.org/wiki/Coin_problem.
[16] M. Delgado, P. Garcı́a-Sánchez, J. Morais, NumericalSgps, A package
for numerical semigroups, Version 0.980 dev (2013), (GAP package),
http://www.fc.up.pt/cmup/mdelgado/numericalsgps/.
[17] D. Einstein, D. Lichtblau, A. Strzebonski, S. Wagon, Frobenius numbers by lattice point enumeration, Integers 7(2007) #A15.
[18] J. Gallian, Contemporary Abstract Algebra. Ninth Ed., Cengage Learning, Boston, MA, 2016.
[19] J. I. Garca-Garca, M. A. Moreno-Fras, A. Vigneron-Tenorio, Computation of Delta sets of numerical monoids, Monatshefte für Mathematik
178(2015) 457–472.
[20] A. Geroldinge, F. Halter-Koch, Nonunique Factorizations: Algebraic,
Combinatorial and Analytic Theory, Pure and Applied Mathematics,
vol. 278, Chapman & Hall/CRC, 2006.
[21] https://en.wikipedia.org/wiki/Knapsack_problem.
[22] http://mathworld.wolfram.com/McNuggetNumber.html.
[23] C. O’Neill, R. Pelayo, On the linearity of ω-primality in numerical
monoids, J. Pure Appl. Algebra 218(2014) 1620–1627.
[24] C. O’Neill, R. Pelayo, How do you measure primality?, Amer. Math.
Monthly 122(2015) 121–137.
[25] https://en.wikipedia.org/wiki/Postage_stamp-problem.
23
[26] H. Picciotto, A. Wah, Algebra: Themes, Tools, Concepts, Creative Publications, Mountain View, CA, 1994, p. 186.
[27] J. L. Ramı́rez Alfonsı́n, The Diophantine Frobenius Problem, Oxford
Lecture Series in Mathematics and Its Applications 30, Oxford University Press, 2005, 256 pp.
[28] B. Sturmfels, Gröbner bases and convex polytopes, University Lecture Series, 8. American Mathematical Society, Providence, RI, 1996.
xii+162 pp. ISBN: 0-8218-0487-1.
[29] J. J. Sylvester, Mathematical questions with their solutions, Educational Times 41(1884) 171–178.
[30] I. Vardi, Computational Recreations in Mathematics, Addison-Wesley,
Reading, MA, 1992, pp. 19–20 and 233–234.
24
| 0math.AC
|
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
1
Control of an Unmanned Surface Vehicle with
Uncertain Displacement and Drag
Wilhelm B. Klinger, Ivan R. Bertaska, Karl D. von Ellenrieder and M. R. Dhanak
Abstract— Experimental testing of an unmanned surface vehicle (USV) has been performed to evaluate the performance of two lowlevel controllers when displacement and drag properties are time-varying and uncertain. The USV is a 4.3 meter long, 150 kilogram
wave adaptive modular vessel (WAM-V) with an inflatable twin hull configuration and waterjet propulsion. Open loop maneuvering
tests were conducted to characterize the dynamics of the vehicle. The hydrodynamic coefficients of the vehicle were determined through
system identification of the maneuvering data and were used for simulations during control system development. The resulting
controllers were experimentally field tested on-water. Variable mass and drag tests show that the vehicle is best controlled by a model
reference adaptive backstepping speed and heading controller. The backstepping controller developed by Liao et. al (2010) is modified
to account for an overprediction of necessary control action and motor saturation. It is shown that when an adaptive algorithm is
implemented for the surge control subsystem of the modified backstepping controller, the effects of variable mass and drag are mitigated.
Index Terms—Adaptive Control, Unmanned Vehicles, Parameter Uncertainty
I. INTRODUCTION
A
N understanding of a vehicle’s maneuverability and dynamic performance characteristics is important for the development
of an automatic control system that will command its motion and can also serve as a guide for the future design of similar
vehicles. A potential application of unmanned surface vehicles (USV’s) is the automatic launch and recovery (ALR) of an
autonomous underwater vehicle (AUV) [27], [29], [33], [34]. A good closed loop controller for a USV used in such operations
needs to allow for the changes in displacement and vehicle resistance associated with launch/recovery of the AUV payload. While
gain-scheduling or other methods may be used when an AUV is either completely onboard the USV, or after it has been launched,
the transition time during launch/recovery requires special considerations in developing automatic controllers because the
magnitude and direction of the drag force due to the AUV and any associated docking/recovery system may be unpredictable.
Several alternatives for speed and heading control of a small twin-hulled unmanned surface vehicle with time-varying and
uncertain displacement and drag are presented. Experimental on-water test results are presented for a USV under the variable and
Manuscript received Month day, 2015. This work was supported by the U.S. Office of Naval Research under Grant N000141110926, managed by K. Cooper.
I. R. Bertaska gratefully acknowledges the support of the Link Foundation Ocean Engineering and Instrumentation Ph.D. Fellowship Program. W. Klinger, I.
Bertaska, K. D. von Ellenrieder and M. R. Dhanak are with the Department of Ocean & Mechanical Engineering, Florida Atlantic University, Dania Beach, FL
33004-3023 USA (phone:+1-954-924-7232; fax:+1-954-924-7270; e-mail: [email protected]).
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
2
uncertain displacement and drag conditions expected during an ALR. The main scholarly outcomes of this work are the
development and experimental validation of low level automatic control systems for a USV during an ALR as it transitions from
carrying to towing an AUV, and also after deploying the AUV and operating independently.
The paper is organized as follows. Background for ALR and adaptive control for USVs is presented in Section II. The particulars
of the vehicle, including its design and instrumentation, are described in Section III. In Section IV, we characterize the maneuvering
performance of the vehicle. We develop controls oriented state space models of the vehicle through the system identification of
experimental data in Section V. Three control alternatives for speed and heading control are explored in Section VI. In Section
VII, we compare the controller’s performance for time-varying and uncertain displacement and drag conditions. Finally, in Section
VIII, some concluding remarks are given regarding the implications of our results.
II. RELATED WORK
There have been a substantial number of studies on the nonlinear control of ships, boats, and unmanned surface vehicles,
including [5],[39] ,[15],[10],[21],[17],[24]. However, as pointed out in [5] and [8], there is a lack of comprehensive and practical
control laws that are robust to the uncertainties in vessel dynamics and large disturbances associated with actual on-water
conditions. Additionally, as elaborated upon in [5], few studies attempt experimental implementation on real USV platforms.
Recent work in controls for unmanned surface vehicles incorporate nonlinearities within the equations of motion to produce
nonlinear control algorithms that are able to achieve a wider performance envelope than their linear counterparts [4]. These include
tracking and setpoint control utilizing sliding mode methods, as in [6] and [25], as well as integrator backstepping as proposed in
[19]. Although various algorithms have been put forward, few have had the opportunity to be validated on full-size USVs in the
presence of environmental disturbances and equipped with traditional navigation sensors, rather than lab-based analogs (e.g.
overhead cameras in lieu of compass/GPS/IMU sensor suite) [9]. Some exceptions include [39],[8],[7],[9],[35],[31], [32], and [3].
In [39], the authors used the backstepping method using feedback linearization and Lyapunov’s direct method to create a speedscheduled algorithm that was field-tested on a USV in riverine operations. In [3], a sliding mode controller was developed for an
amphibious surface vehicle navigating the transitional area between the shore and open ocean, where breaking waves cause large
nonlinearities that may not be adequately modeled in situ. Automatically generated behaviors for USVs were presented in [8], [7],
and [40] where three setpoint controllers were proposed and validated in the field. A controller from this paper is incorporated here
in Section VI.
Adaptive control is also an active area of research for unmanned marine vehicles. Skjetne et al. used adaptive control to
determine the added mass hydrodynamic coefficients while performing free-running maneuvering experiments in [36] to create a
nonlinear maneuvering model suited to either tracking or dynamic positioning control. The methods behind backstepping and
adaptive control were combined in [23], using a Nomoto-like steering model that neglected the coupling of yaw rate with sway
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
3
velocity commonly found in USVs. A recent method of handling uncertain displacement and drag terms is presented in [4], where
model predictive control is used to stabilize the vehicle’s course under large mass variations upwards of 50%. It should be noted
that the authors used a linearized model for plant dynamics, which may not fully capture the dynamics of the vehicle at high speeds.
Aguiar and Hespanha proposed an approach utilizing supervisory switching control to handle parametric uncertainties in an
underactuated marine vehicle [1]. Although favorable simulation results were presented for both an underactuated hovercraft and
an AUV, the use of finite number of system models could potentially lead to large errors if parameters’ range is infinite. Aguiar
and Pascoal extended this work in parametric uncertainty in [2], where an adaptive backstepping controller was derived to estimate
poorly modeled hydrodynamic parameters in an AUV with an unconventional hull form. Kragelund et al [22] proposed three
different adaptive controllers – a gain scheduled PID controller, a model reference adaptive controller and an 𝐿1 adaptive controller
– tested on a SeaFox USV to mitigate the effects of the transitional region between displacement and planing modes of operations.
The work presented here extends the results of previous studies to include the effects of the adaptive control on course-keeping
relating to vehicle bearing, as well as using a nonlinear controller derived from a nonlinear maneuvering model.
III. THE WAM-V USV14
A. Overview
The WAM-V USV14 (Fig. 1) is a twin hull pontoon style vessel designed and built by Marine Advanced Research, Inc. of
Berkeley, CA USA. The vessel structure consists of two inflatable pontoons, two motor pods, a payload tray, and two supporting
arches. The WAM-V class of vehicles is designed to provide a stable platform that can mitigate the effects of waves [11], [16],[30].
The vehicle’s physical characteristics are shown in TABLE I. The WAM-V demonstrates appreciable motion in surge, sway, and
yaw during operation, but can only apply control force and torque in surge and yaw. Because the steering scheme can only actuate
in two of the three appreciable degrees of freedom, the USV is underactuated.
Fig. 1: The WAM-V USV14 Unmanned Surface Vehicle underway in the Intracoastal Waterway (ICW).
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
4
TABLE I
PRINCIPLE CHARACTERISTICS OF THE WAM-V14 USV. THE LOCATION OF THE “KEEL” IS TAKEN AS
THE BOTTOM OF THE PONTOONS. W.R.T. IS AN ACRONYM FOR THE PHRASE “WITH RESPECT TO”.
Parameter
Length Overall (L)
Length on the Waterline (LWL)
Draft (aft and mid-length)(lightship)
Beam Overall (BOA)
Beam on the Waterline (BWL)
Depth (keel to pontoon skid top)
Area of the Waterplane (AWP)
Centerline-to-centerline Side Hull Separation (B)
Length to Beam Ratio (L/B)
Volumetric Displacement (∇) (lightship)
Mass (𝑚)
Longitudinal Center of Gravity (LCG) w.r.t.
aft plane of engine pods
Value
4.29 [m]
3.21 [m]
0.127 and 0.105 [m]
2.20 [m]
2.19 [m]
0.39 [m]
1.1 [m2]
1.83 [m]
2.34
0.34 [m3]
150 [kg]
1.27 [m]
B. Waterjet Propulsion System
The waterjet configuration is similar to that found on the stock version of the WAM-V USV12 [41], except that each of the
USV14 motor pods contains an electric-powered motor, rather than a gasoline-powered engine. The motor pods attach aft of the
inflatable pontoons with a hinged connection at the bottom and a flexible O-ring connection at the top, which allows rotation about
the bottom connection. The design is meant to ensure the motor pods’ continuous alignment with the water, despite vehicle pitching
motion, which maintains flow through the waterjets. Each motor is powered by eight lithium polymer batteries, wired in parallel,
that provide a nominal voltage of 22.2 V. The motors are NeuMotors 2230, and are rated for a maximum of 22,000 RPM. The
nozzles of the waterjet have been fixed so that they point directly aft, and the vehicle is steered using differential thrust. The
waterjet reversing buckets are not used, such that thrust is only developed in the forward direction.
C. Guidance, Navigation, and Control Electronics
In order to conduct development and testing of autonomous control for the WAM-V USV14, a guidance, navigation and control
(GNC) system is implemented. The GNC system is housed in a plastic, water-resistant box and contains a motherboard, singleboard computer, inertial measurement unit (IMU) with global positioning system (GPS) capability, tilt-compensated digital
compass, radio frequency (RF) transceiver, and pulse width modulation (PWM) signal generator. A detailed summary of the GNC
system can be found in [7] and [26]. For the purpose of this project, the key components of the GNC system are the sensor suite
(IMU/GPS and digital compass), single-board computer, and RF transceiver. The IMU/GPS is an XSENS MTi-G sensor , which
is used to measure the position and orientation of the USV during operations. This information is relayed to the single board
computer, where the low level control code is implemented, and maneuvering data are logged. The GPS is specified to provide up
to 2.0 meter accuracy, depending on weather and satellite availability. The digital compass is used to record vehicle heading and
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
5
has a resolution of 0.1 degrees. The RC receiver allows the user the option of maneuvering the USV under remote manual control
or by autonomous navigation. A hand-held remote control is used both for operating the thrusters and reversing buckets and also
for switching between RC and autonomous modes.
IV. MANEUVERING CHARACTERISTICS
Maneuvering tests were performed in sheltered, calm water sections of the US Intracoastal Waterway (ICW) in Dania Beach
and Hollywood, FL USA. The vehicle experienced only minor wind-generated waves. Wind speed and direction were collected
during all testing. Wind-generated forces and moments do not appear to have had any effect on the test results.
The total weight (displacement) of the system will affect both the inertial properties and, through changes in hull submergence
(draft), the hydrodynamic resistance of the vehicle. Thus, the maneuvering characteristics are dependent on how heavily laden the
vehicle is. Here, three different loading conditions (total vehicle displacements), which have been termed “slick”, “lightship” and
“full displacement”, were tested:
Slick: The vehicle displacement with only GNC hardware onboard – the minimum loading to run autonomously. This
configuration includes neither the simulated ALR retrieval mechanism, nor the AUV mock-up (see Section VII.A). The
total displacement is approximately 150 kg.
Lightship: The vehicle displacement with both the GNC hardware and the simulated ALR retrieval mechanism aboard,
but not the AUV mock-up. This arrangement would correspond to the loading condition of the USV after a launch and
before a recovery of the AUV. The ALR system has a mass of approximately 70 kg, so that the total displacement of the
vehicle is about 220 kg.
Full Displacement: The loading condition with the GNC hardware, simulated ALR retrieval mechanism, and AUV
mock-up aboard. This corresponds to the weight of the USV before and after the ALR procedure. To simulate the weight
of the AUV, an additional mass of 26 kg is added to the lightship system, such that its total displacement is about 246 kg.
A. Acceleration Tests
Acceleration tests were conducted to obtain the steady state speeds associated with the available range of motor commands. The
vehicle started in a stationary position, with the surge and sway velocities, and the yaw rate as close to zero (as possible on the
water), and was accelerated to a specified throttle command on both motors for 60 seconds. The test was conducted for a throttle
range of 60-100%. Preliminary experimental results for the slick acceleration test are presented in [20]. Fig. 2 shows the results of
two full throttle acceleration tests for the lightship (no AUV in payload) and full displacement (AUV in payload) vehicle conditions.
It can be seen from the data in Fig. 2 that the maximum surge velocity of the vessel decreases as the payload is increased. The
average maximum surge speed for the lightship and full displacement cases is 2.8 m/s and 2.5 m/s, respectively.
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
6
a)
b)
Fig. 2: Vehicle surge velocity for two full throttle acceleration tests using the automatic control box with preprogrammed motor
commands. Vehicle conditions were: a) Lightship and b) Full Displacement.
The results of the acceleration tests provided a means to approximate a forward drag model for each displacement condition of
the vessel. In order to relate motor commands to drag, a bollard-pull test was conducted. The test consisted of tethering the WAMV USV14 to a fixed object (a pier) with a scale in line with the tether. The speed of the motors was slowly increased to full throttle;
the maximum force measured is 𝑇wj_max = 204 Newtons. Repetitions of this test at multiple throttle commands indicated that thrust
scales linearly with motor command when the forward speed of the vessel is zero. The results of the bollard-pull test were then
incorporated into a linear model for thrust output as a function of input motor command. As the vehicle was tethered during these
tests, the effect of surge speed on thrust production is ignored in this model:
𝑇wj =
𝑛
𝑇
.
𝑛max wj_max
(1)
Here 𝑛 is the commanded motor speed in Hz; 𝑛⁄𝑛max can be interpreted as a commanded percentage of the maximum thrust. Using
this linear model, experiments were conducted at steady state forward speeds 𝑢, where it was assumed that the USV was operating
at the “thrust equals drag” condition. With a drag model composed of both quadratic and linear terms, the thrust equals drag 𝐷
condition would give:
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
𝑇wj_max
𝑛
= 𝑋𝑢|𝑢| 𝑢|𝑢| + 𝑋𝑢 𝑢.
𝑛max
7
(2)
𝑋𝑢|𝑢| and 𝑋𝑢 were then determined from curvefits to the experimental data in Fig. 3.
Fig. 3: Surge direction drag model for the WAM-V USV14 in three displacement conditions: a) slick, b) lightship, and c) fulldisplacement. The points represent steady state values of thrust equals drag (𝐷) for a given speed (𝑢).
Because (1) neglects the effect of surge speed on thrust developed, the drag model does not capture the characteristic variation of
drag with surge speed that might be expected, e.g. 𝐷~𝑢2 . However, use of the 𝑋𝑢|𝑢| and 𝑋𝑢 terms found from these experiments
should intrinsically reproduce the surge speed effects when the desired thrust is determined using (1).
1) Alternative Thruster Models
A second possible thruster model, which has been developed as an analog to conventional propeller models, was also explored.
In conventional propellers, the curves of thrust coefficient
𝐾𝑇 ≡
𝑇𝑝
,
𝜌𝑛2 𝑑 4
versus advance ratio
𝐽≡
𝑢
,
𝑛𝑑
are approximately linear such that
𝐾𝑇 (𝐽) = −𝛼1 𝐽 + 𝛼2 .
Here, 𝑇𝑝 is propeller thrust, 𝑑 is propeller diameter, 𝜌 is water density and 𝛼1 , 𝛼2 > 0 are constants. Using this model the thrust
can be rewritten as
𝑇𝑝 = −𝛼1
or
𝑢
𝜌𝑛2 𝑑 4 + 𝛼2 𝜌𝑛2 𝑑 4 ,
𝑛𝑑
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
8
𝑇𝑝 = 𝑎1 (𝜌𝑑 3 )𝑢𝑛 + 𝑎2 (𝜌𝑑 4 )𝑛2 .
Here the thrust developed varies quadratically with the rotational rate of the propeller and linearly with the surge speed of the
vessel. The authors have not come across similar waterjet models in the literature, but it seems that such a model might be
appropriate. Examining a representative plot of waterjet pump head coefficient 𝐶𝐻 vs. capacity coefficient 𝐶𝑄 (Fig. 4), one can see
that an almost linear, relation exists between the pressure head coefficient
𝐶𝐻 ≡
𝑔𝐻
,
(𝑛𝑑)2
(where 𝑔 is gravitational acceleration, 𝐻 is the pressure head across the waterjet, and 𝑑 is the diameter of the waterjet impeller) and
the capacity coefficient
𝐶𝑄 ≡
𝑄𝑓
,
𝑛𝑑 3
(where 𝑄𝑓 is the volumetric flow rate through the waterjet). Note that 𝐶𝐻 and 𝐶𝑄 are often combined with measured waterjet pump
data to estimate the steady state waterjet impeller shaft speed (see [12], for example).
Fig. 4: Example pump diagram showing pump efficiency 𝜂𝑃 and head coefficient 𝐶𝐻 ≡ 𝑔𝐻/(𝑛𝑑)2 as a function of the capacity
coefficient 𝐶𝑄 ≡ 𝑄𝑓 /𝑛𝑑 3 . 𝑛=shaft revolutions per second, 𝑑 = impeller diameter (reproduced from [13]).
Following a similar development as used above for a conventional propeller, we have
𝐶𝐻 (𝐶𝑄 ) = −𝛼1 𝐶𝑄 + 𝛼2 ,
which can be rewritten as
𝜌𝑔𝐻 = −𝛼1
𝜌𝑄𝑓
(𝑛𝑑)2 + 𝛼2 𝜌(𝑛𝑑)2 .
𝑛𝑑 3
Scaling the volumetric flow rate through the waterjet as 𝑢𝑑 2 , and assuming that the thrust generated is proportional to the pressure
head across the waterjet multiplied by the square of the characteristic diameter of the waterjet we get
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
𝑇wj ~𝜌𝑑 2 [−𝛼1
9
𝑢𝑑 2
(𝑛𝑑)2 + 𝛼2 (𝑛𝑑)2 ],
𝑛𝑑 3
which has the same form as the thrust developed on a propeller
𝑇wj ~𝑎1 (𝜌𝑑 3 )𝑢𝑛 + 𝑎2 (𝜌𝑑 4 )𝑛2 .
Curvefits of this relation to the data collected during our field tests indicate that the model works well for the lightship condition,
but does not match as well for the slick or full displacement conditions. This waterjet thrust model may be appropriate in certain
circumstances and appears to be worthy of additional investigation.
B. Maneuvering Tests
A modified on-water zig-zag test was conducted to obtain motion data. The traditional zig-zag test specifies rudder commands
for a certain heading range [18], but for a vehicle steered using differential thrust, the test must be conducted by varying the motor
speeds. Motor commands were alternated between 0 and 100% power, with the motors 180o out of phase. Each motor setting was
held for 7 seconds. Trajectory and yaw rate (𝑟) results for three tests in the lightship condition are shown in Fig. 5. Yaw rate is of
interest when conducting the zig-zag test because it is later used in model development, specifically in determining the maneuvering
coefficients pertaining to how fast the vessel changes heading. This concept is explored further in Section V. Several runs of each
maneuvering test were carried out to show a convergence of the results and to inspire confidence in the results before using them
to develop the dynamic model of the vehicle.
a)
b)
Fig. 5. a) Trajectory and b) yaw rate time histories for three zig-zag tests conducted in the lightship displacement condition.
The maximum yaw rate was 13 deg/s. In a) the starting position is marked by the circle at the origin.
V. SYSTEM IDENTIFICATION
A. Equations of Motion
Data collected from the acceleration and zig-zag tests in Sections IV.A and IV.B, respectively, were used to develop three degree
of freedom state space maneuvering models for the vehicle in full displacement and lightship displacement conditions. Following
the notation developed in [13] and [14], a three degree of freedom (surge, sway, and yaw) vehicle model is used, as presented in
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
10
[20] and [21], where
𝑴𝒗̇ + 𝑪(𝒗)𝒗 + 𝑫(𝒗)𝒗 = 𝝉
(3)
𝑇
𝜼̇ = [𝑥̇ 𝑦̇ 𝜓̇] ,
(4)
and
𝒗 = [𝑢 𝑣 𝑟]𝑇 .
(5)
Here, the 𝜼̇ describes the vehicle’s North (𝑥̇ ) East (𝑦̇ ) linear velocities and the Z-axis angular velocity (𝜓̇) in an inertial reference
frame (𝑥, 𝑦, 𝜓). 𝒗 contains the vehicle surge velocity (𝑢), sway velocity (𝑣) and yaw rate (𝑟) in the body fixed frame. Fig. 6
illustrates the two coordinate systems and the motor forces acting on the vehicle.
Fig. 6. Top view of WAM-V USV14 with body-fixed coordinate system overlaid. 𝑥𝑏 and 𝒚𝒃 denote vessel surge and
sway axes, respectively. 𝑇𝑝𝑜𝑟𝑡 and 𝑇𝑠𝑡𝑏𝑑 denote the port and starboard waterjet thrust.
The transformation matrix for converting from body-fixed to earth-fixed frames is given by:
cos 𝜓
𝑱(𝜼) = [ sin 𝜓
0
− sin 𝜓
cos 𝜓
0
0
0].
1
(6)
𝑴 is an inertia tensor that is the sum of a rigid body mass matrix, 𝑴𝑹𝑩 , and an added mass matrix, 𝑴𝐴 :
𝑚 − 𝑋𝑢̇
𝑴 = 𝑴𝑹𝑩 + 𝑴𝐴 = [ 0
−𝑚𝑦𝐺
0
𝑚 − 𝑌𝑣̇
𝑚𝑥𝐺 − 𝑁𝑣̇
−𝑚𝑦𝐺
𝑚𝑥𝐺 − 𝑌𝑟̇ ],
𝐼𝑧 − 𝑁𝑟̇
(7)
where, 𝑚 denotes the mass of the USV-AUV system, 𝑥𝐺 and 𝑦𝐺 represent the coordinates of the vessel center of mass in the bodyfixed frame, and 𝐼𝑧 denotes moment of inertia about the 𝑧𝑏 -axis. Several of the terms in the mass matrix utilize SNAME (1950)
[38] nomenclature for representing a force or moment created by motion in a specific degree of freedom, where 𝑋 and 𝑌 correspond
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
11
to force in the 𝑥𝑏 and 𝑦𝑏 directions, respectively, and 𝑁 denotes a moment about the 𝑧𝑏 -axis (vertical axis). The subscript on each
coefficient denotes the cause of the force/moment. It is pertinent to note that the added mass in the surge direction is modeled as
16% of the total mass of the vehicle, whereas theoretically the value should be less than 5%. This change is made so that the
simulation outputs a forward acceleration curve that matches the experimental data.
𝑪(𝒗) is a Coriolis matrix, which includes the sum of a rigid body (𝑪𝑅𝐵 ) and added mass (𝑪𝐴 ) matrix:
𝑪(𝒗) = 𝑪𝑅𝐵 + 𝑪𝐴
=
+
0
0
−𝑚(𝑥𝐺 𝑟 + 𝑣)
0
0
−𝑚(𝑦𝐺 𝑟 − 𝑢)]
[
𝑚(𝑥𝐺 𝑟 + 𝑣) 𝑚(𝑦𝐺 𝑟 − 𝑢)
0
𝑌𝑣̇ 𝑣 + (
𝑌𝑟̇ +𝑁𝑣̇
2
0
0
0
0
200
−𝑋𝑢̇ 𝑢
𝑋𝑢̇ 𝑢
0
𝑌𝑟̇ +𝑁𝑣̇
−𝑌𝑣̇ 𝑣 − (
[
200
2
)𝑟
)𝑟
(8)
]
The model designates the body-fixed origin at the center of gravity and assumes horizontal symmetry, making 𝑥𝐺 = 0 and 𝑦𝐺 = 0.
The factor of 1/200 in the upper-rightmost and lower-leftmost matrix positions of 𝑪𝐴 was applied during model development to
better fit the experimental data and does not appear in [13] and [14]. This scale factor was necessary to fix an issue with the
simulation where the USV could not pull out of a turn after one had been initiated, even with maximum torque applied. The original
scale factor was 2 to represent each demihull, and was manually incremented until the turning behavior of the simulation model
was deemed acceptable. 𝑫(𝒗) is the summation of linear and nonlinear drag matrices:
𝑫(𝒗) = 𝑫𝒍 + 𝑫𝑛
(9)
where,
𝑋𝑢
𝑫𝒍 = − [ 0
0
0
𝑌𝑣
𝑁𝑣
0
𝑌𝑟 ]
𝑁𝑟
(10)
and,
𝑋𝑢|𝑢| |𝑢|
0
0
𝑌𝑣|𝑣| |𝑣| + 𝑌𝑣|𝑟| |𝑟| 𝑌𝑟|𝑣| |𝑣| + 𝑌𝑟|𝑟| |𝑟| ].
𝑫𝒏 = − [ 0
0
𝑁𝑣|𝑣| |𝑣| + 𝑁𝑣|𝑟| |𝑟| 𝑁𝑟|𝑣| |𝑣| + 𝑁𝑟|𝑟| |𝑟|
(11)
The surge direction drag terms 𝐷𝑝𝑜𝑟𝑡 and 𝐷𝑠𝑡𝑏𝑑 are modeled using a polynomial curve fit derived from experimental testing. A
coordinate transformation is carried out to obtain the velocities of each individual pontoon hull because they are offset from the
CG. These transformed velocities are used in the drag model below as 𝑢𝑝𝑜𝑟𝑡 and 𝑢𝑠𝑡𝑏𝑑 :
𝐷𝑝𝑜𝑟𝑡 = (
𝑋𝑢|𝑢|
⁄ ) |𝑢𝑝𝑜𝑟𝑡 |𝑢𝑝𝑜𝑟𝑡 + (𝑋𝑢⁄ ) 𝑢𝑝𝑜𝑟𝑡
2
2
(12)
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
𝐷𝑠𝑡𝑏𝑑 = (
12
𝑋𝑢|𝑢|
⁄ ) |𝑢𝑠𝑡𝑏𝑑 |𝑢𝑠𝑡𝑏𝑑 + (𝑋𝑢⁄ ) 𝑢𝑠𝑡𝑏𝑑
2
2
(13)
The values of 𝑋𝑢|𝑢| and 𝑋𝑢 vary based on the displacement condition of the vehicle and are shown in TABLE II. Incorporating the
moment created by the two drag forces 𝐷𝑝𝑜𝑟𝑡 and 𝐷𝑠𝑡𝑏𝑑 , the term
(14)
(𝐷𝑠𝑡𝑏𝑑 − 𝐷𝑝𝑜𝑟𝑡 )𝐵/2
is added to the yaw moment (not modeled in 𝑫).
Finally, 𝝉 is a vector of the forces and moments generated by the port and starboard waterjets:
𝜏𝑥
(𝑇𝑝𝑜𝑟𝑡 + 𝑇𝑠𝑡𝑏𝑑 )
𝜏
𝝉 = ( 𝑦) = [
].
0
𝜏𝑧
(𝑇𝑝𝑜𝑟𝑡 − 𝑇𝑠𝑡𝑏𝑑 )𝐵/2
(15)
As shown in Fig. 6, 𝑇𝑝𝑜𝑟𝑡 and 𝑇𝑠𝑡𝑏𝑑 denote the thrust from the port and starboard waterjets, respectively, and 𝐵/2 is the transverse
distance from the centerline of the vessel to the centerline of each hull. The forces caused by wind and waves are neglected.
TABLE II
SURGE DRAG COEFFICIENTS FOR THE WAM-V USV14.
Displacement Condition
Slick
Lightship
Full Displacement
𝑋𝑢|𝑢|
-5.8722
-6.9627
-2.6693
𝑋𝑢
50.897
55.771
47.341
The amount of thrust that must allocated to each waterjet can be determined by solving (15) to find
𝑇𝑝𝑜𝑟𝑡 = 𝜏𝑥 /2 + 𝜏𝑧 /𝐵 and 𝑇𝑠𝑡𝑏𝑑 = 𝜏𝑥 /2 − 𝜏𝑧 /𝐵.
(16)
The thrust developed on each waterjet is modeled using the linear relation between commanded waterjet motor speed and thrust
output from equation (1).
B. Simulation Results
In order to determine the accuracy of the mathematical model presented above and to determine the hydrodynamic coefficients
in the added mass matrix (7), Coriolis matrix (8), and drag matrices (10), (11), the experimental zig-zag and acceleration data are
overlaid with simulation results for matching motor input. The values of the hydrodynamic coefficients were manually adjusted
until the simulation results qualitatively matched the experimental data.
For the lightship condition, the straight-line acceleration data were compared against simulation results for three separate waterjet
operating speeds: 𝑅𝑃𝑀 = {0.8𝑅𝑃𝑀𝑚𝑎𝑥 , 0.9𝑅𝑃𝑀𝑚𝑎𝑥 , 𝑅𝑃𝑀𝑚𝑎𝑥 }. Fig. 7-Fig. 10 display final results of the surge drag model in the
vehicle’s equations of motion. Specifically, the parameters that were identified to fit the experimental data were: 𝑋𝑢̇ , 𝑋𝑢 , and 𝑋𝑢|𝑢| .
𝑋𝑢 and 𝑋𝑢|𝑢| characterize the steady state surge velocity of the vehicle during the acceleration test (the flat area of the curves
between 20 and 60 seconds on the figures below). Each steady state surge speed for the three test cases was tabulated and the results
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
13
were fitted to a quadratic polynomial, as explained in Section IV.A. The equations displayed in Fig. 3 are of the form:
𝑋𝑑𝑟𝑎𝑔 = 𝐷𝑝𝑜𝑟𝑡 + 𝐷𝑠𝑡𝑏𝑑 = 𝑋𝑢|𝑢| 𝑢|𝑢| + 𝑋𝑢 𝑢,
(17)
where 𝑋𝑑𝑟𝑎𝑔 denotes drag in the vessel’s surge direction. As the submerged hull form and wetted surface area varies with
displacement, 𝑋𝑢 and 𝑋𝑢|𝑢| also vary with displacement. The term 𝑋𝑢̇ affects the transient behavior during acceleration/deceleration,
i.e. between 0-10 seconds/past 60 seconds. 𝑋𝑢̇ affects the slope of the simulation surge velocity curve during those beginning and
end time periods.
Fig. 7. Lightship surge histories for full throttle acceleration tests. Three experimental tests are shown, each at different initial
conditions: 𝑢 = 0, 0.3, 0.9 m/s. Three simulation results are overlaid with initial conditions corresponding to field tests.
Fig. 8. Lightship surge histories for 90% throttle acceleration tests. Three experimental tests are shown, each at different initial
conditions: 𝑢 = 0.13, 0.54, 0.75 m/s. Three simulation results are overlaid with initial conditions corresponding to field tests.
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
14
Fig. 9. Lightship surge histories for 80% throttle acceleration tests. Three experimental tests are shown, each at different initial
conditions: 𝑢 = 0.44, 0.8, 0.8 m/s. Two simulation results are overlaid with initial conditions corresponding to the field tests.
Results for the full displacement acceleration tests are displayed in Fig. 10-Fig. 14. When comparing the results, which cover
the upper 40% of the vessel’s motor operating limits and the two displacement conditions of interest in the ALR mission, one can
conclude that the surge model of the WAM-V USV14 is adequately identified in the case of straight line forward motion.
Fig. 10. Full displacement surge histories for full throttle acceleration tests. Two experimental tests are shown, each at
different initial conditions: 𝑢=0.5, 0.67 m/s. Two simulation results are overlaid with corresponding initial conditions.
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
15
Fig. 11. Full displacement surge histories for 80% throttle acceleration tests. Two experimental tests are shown, each at the
initial condition: 𝑢=0.5 m/s. One simulation result is overlaid with the same initial condition.
Fig. 12. Full displacement surge histories for 60% throttle acceleration tests. Two experimental tests are shown, each at the
initial condition: 𝑢=0.13 m/s. One simulation result is overlaid with the same initial condition.
It can be noted that in Figs. 7-12 the experimental data show the boat accelerating faster when thrust is applied than it decelerates
when thrust is removed. However, the simulations show the boat decelerating about as quickly as it accelerates. The reason for this
is because, when moving, the real vehicle creates a wake of fluid traveling in the same direction as the vehicle. When thrust is
removed and the vehicle starts to decelerate, the inertia of the fluid in the wake causes the fluid to pile up against the stern, pushing
against the vehicle in the forward direction and slowing its deceleration.
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
16
Unfortunately, a simple quadratic drag model cannot capture this wake effect. Instead the relative rates of acceleration and
deceleration are governed by the relative magnitude of the thrust and surge speed prior to acceleration/deceleration. One can see
this from the equation of motion in the surge direction, which can be simplified as:
1
𝑚𝑢̇ = 𝑇 − 𝜌𝑢2 𝐶𝐷 𝐴𝑓 ,
2
where 𝑇 is the constant thrust applied, 𝐶𝐷 is a drag coefficient and 𝐴𝑓 is the frontal area of the hulls. This can be rewritten as
𝑢̇ =
𝑇 𝑏 2
− 𝑢 .
𝑚 𝑚
When the vehicle reaches a steady state speed 𝑢𝑠𝑠 , the acceleration is zero it can be seen that
𝑇
𝑢𝑠𝑠 = √ .
𝑏
When thrust is initially applied, the acceleration is
𝑢̇ =
𝑇 𝑏 2
− 𝑢 ,
𝑚 𝑚 0
where 𝑢0 is the initial velocity. As the surge speed increases the acceleration decreases to zero. After the vehicle has reached a
steady state speed, at the instant that thrust is removed, the deceleration is
𝑢̇ = −
𝑏 2
𝑇
𝑢𝑠𝑠 = − .
𝑚
𝑚
Thus, one can see that, for the model, the magnitude of the initial acceleration is smaller than that of the initial deceleration when
the model has some initial velocity. It may be possible circumvent this issue by including a switching added mass term in the surge
equation of motion to produce the desired result. Since the bow is finer than the transom, one would expect the added mass
associated with acceleration 𝑚𝑎1 to be smaller than that of deceleration 𝑚𝑎2 . One could implement an equation, such as:
𝑚𝑎 =
1 + sgn(𝑢̇ )
1 − sgn(𝑢̇ )
𝑚𝑎1 +
𝑚𝑎2 ,
2
2
so that
𝑚 , 𝑢̇ > 0
𝑚𝑎 = { 𝑎1
𝑚𝑎2 , 𝑢̇ < 0
A plot of acceleration and deceleration shows that this could produce a result similar to what would be expected (Fig. 13).
However, it is conventional to use a single added mass coefficient for both acceleration and deceleration. From a practical
standpoint, it does not appear to be critical for the purposes of controller design, which is the main focus of the current study.
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
17
Fig. 13. Acceleration and deceleration with switching added mass term. The added mass for deceleration was made
significantly larger than that of acceleration.
Now that verification of the model’s straight line dynamics have been presented, the turning dynamics are addressed by
examining the results of simulated zig-zag maneuvers. Fig. 14 shows an example comparison between the measured and simulated
yaw rate for the zig-zag maneuver at 100% throttle. Based on the results of this and other yaw rate experiment/simulation
comparisons, it was concluded that the dynamic model captures the turning dynamics of the vessel sufficiently well to be used for
the design of a speed and heading controller for the ALR mission.
Fig. 14. Yaw rate data for the lightship 100% throttle zig-zag maneuver. Maximum experimental and simulated yaw rate is
observed to be 13.2 deg/s.
Several assumptions are carried forward into the control system design process. First, it is assumed that the speed of the vessel
is accurately predicted during straight line transits (i.e. when heading error has stabilized around zero) and that the surge velocity
will be underestimated when heading error is large and hence the vehicle will turn aggressively to minimize the heading error. The
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
18
second assumption about the model is that heading is accurately determined. Considering that the intent of this project is the
development and implementation of an appropriate controller for the ALR mission, knowing that the speed and heading of the
dynamic model can be trusted, or its errors anticipated, allows the search for a more accurate model to stop. The final values for
the hydrodynamic coefficients used in the model are given in TABLE III. The middle column of this table contains the nondimensional factors that are applied to the dimensional terms derived using strip theory and principles of drag for cylinders
(𝑁𝑟 , 𝑌𝑟, 𝑌𝑣 ). They are the correction factors for the disparity between theoretical vehicle dynamics and experimental data. In TABLE
III: 𝜌 is water density, 𝑇 is the draft, 𝐶𝑑 is the drag coefficient for a lateral cylinder (𝐶𝑑 = 1.1), and 𝐵ℎ𝑢𝑙𝑙 is the beam of the
individual pontoon hulls.
TABLE III
HYDRODYNAMIC COEFFICIENTS FOR THE WAM-V USV14
Coefficient Name
Non-Dimensional Factor
Dimensional Term
𝑁𝑣̇
2.5
𝑁𝑟̇
1.2
[(𝐿 − 𝐿𝐶𝐺)3 + 𝐿𝐶𝐺 3 ]
4.75
𝐵ℎ𝑢𝑙𝑙 4
−{
𝜋𝜌
𝑇 + 𝜋𝜌𝑇 2
}
2
2
3
𝑋𝑢̇
0.075
−𝑚
𝑌𝑟̇
0.2
𝑌𝑣̇
0.9
−𝜋𝜌𝑇 2
−𝜋𝜌𝑇 2
[(𝐿 − 𝐿𝐶𝐺)2 + 𝐿𝐶𝐺 2 ]
2
[(𝐿 − 𝐿𝐶𝐺)2 + 𝐿𝐶𝐺 2 ]
2
−𝜋𝜌𝑇 2 𝐿
𝑋𝑢
See Fig. 3.
𝐿
𝐵ℎ𝑢𝑙𝑙
𝐵ℎ𝑢𝑙𝑙 2 𝜋𝑇𝐿
− 0.1
+ 0.016 (
) ](
)
𝑇
𝑇
𝑇
2
𝑌𝑣
0.5
𝑁𝑟
0.02
−𝜋𝜌√(𝑢2 + 𝑣 2 )𝑇 2 𝐿2
𝑁𝑣
0.06
−𝜋𝜌√(𝑢2 + 𝑣 2 )𝑇 2 𝐿
𝑌𝑟
6
−𝜋𝜌√(𝑢2 + 𝑣 2 )𝑇 2 𝐿
−40𝜌|𝑣| [1.1 + 0.0045
𝑋𝑢|𝑢|
See Fig. 3.
𝑌𝑣|𝑣|
1
𝑌𝑣|𝑟|
1
𝑌𝑟|𝑣|
1
𝑌𝑟|𝑟|
1
−𝜌𝑇𝐶𝑑 𝐿
−𝜌𝑇
1.1
[(𝐿 − 𝐿𝐶𝐺)2 − 𝐿𝐶𝐺 2 ]
2
Same as 𝑌𝑣|𝑟|
−𝜌𝑇
𝐶𝑑
[(𝐿 − 𝐿𝐶𝐺)3 + 𝐿𝐶𝐺 3 ]
3
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
19
𝑁𝑣|𝑣|
1
Same as 𝑌𝑣|𝑟|
𝑁𝑣|𝑟|
1
Same as 𝑌𝑟|𝑟|
𝑁𝑟|𝑣|
1
Same as 𝑌𝑟|𝑟|
𝑁𝑟|𝑟|
1
−𝜌𝑇
𝐶𝑑
[(𝐿 − 𝐿𝐶𝐺)4 + 𝐿𝐶𝐺 4 ]
4
VI. SPEED AND HEADING CONTROL
Due to the underactuation of the USV14, the lowest level of motion control is in surge speed and heading angle. This approach
has worked well with this type of vehicle [3], [7],[8]. Several controllers regulating those motion primitives have been developed
and implemented on the USV14.
A. Backstepping Control
A surge speed and heading controller has been developed using feedback linearization [37]. The controller is the same as that
presented by Liao et al. [24], apart from changes in formulating the controller for tracking control, rather than set point control, and
in the use of a nonlinear drag model. Note that the approach is very similar to that developed by Ghommam et al. [17] for position
and heading control of an underactuated surface vessel. As in both [17] and [24], the procedures developed by Panteley and Loria
[28] can be used to show that the cascaded system is globally uniformly asymptotically stable.
To develop a surge speed controller, the surge equations of motion given by (3), (7)-(11), and (17) were reduced to:
(𝑚 − 𝑋𝑢̇ )𝑢̇ − (𝑚 − 𝑌𝑣̇ )𝑣𝑟 − 𝑋𝑑𝑟𝑎𝑔 = 𝜏𝑥 ,
(18)
where 𝜏𝑥 is the surge component of the forces/moments (15) generated by the waterjet. Input state linearization [37] can be used to
find a nonlinear feedback control law for the surge subsystem. Let 𝜉𝑢 be a linearizing control, related to the commanded value of
𝜏𝑥 from the controller by
𝜏𝑥 = (𝑚 − 𝑋𝑢̇ )𝜉𝑢 − (𝑚 − 𝑌𝑣̇ )𝑣𝑟 − 𝑋𝑑𝑟𝑎𝑔 .
(19)
With this value of 𝜏𝑥 , (19) reduces to 𝑢̇ = 𝜉𝑢 . If the surge tracking error is defined as 𝑒𝑢 = (𝑢 − 𝑢𝑑 ), where 𝑢𝑑 is the desired surge
speed, and a linearizing control of the form 𝜉𝑢 = 𝑢̇ 𝑑 − 𝑘𝑢 𝑒𝑢 is selected, the surge tracking error dynamics will be:
𝑒̇𝑢 + 𝑘𝑢 𝑒𝑢 = 0.
(20)
For positive values of the surge velocity gain 𝑘𝑢 , the error dynamics for 𝑒𝑢 are stable. As in [24], a hyperbolic tangent form can be
selected for the desired surge acceleration as
𝑢̇ 𝑑 = 𝑢̇ 𝑎,𝑚𝑎𝑥 tanh[𝑘𝑎,𝑚𝑎𝑥 (𝑢𝑑,𝑟𝑒𝑓 − 𝑢)/𝑢̇ 𝑎,𝑚𝑎𝑥 ],
(21)
where 𝑢̇ 𝑎,𝑚𝑎𝑥 is the maximum allowed surge acceleration, 𝑘𝑎,𝑚𝑎𝑥 is a positive surge acceleration gain that shapes the approach to
the desired speed, and 𝑢𝑑,𝑟𝑒𝑓 is defined as
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
𝑢𝑑,𝑟𝑒𝑓 ≡ min{[𝑢𝑑,𝑦𝑎𝑤 + (𝑢𝑑 − 𝑢𝑑,𝑦𝑎𝑤 )exp(−5.73|𝑒𝜓 |)], 𝑢𝑑 },
20
(22)
where 𝑒𝜓 = 𝜓 − 𝜓𝑑 is the vehicle heading error, and 𝑢𝑑,𝑦𝑎𝑤 is the desired surge velocity for periods of yaw motion. The function
min{ } denotes taking the minimum of the two values in the function’s argument. The intent of including the 𝑢𝑑,𝑟𝑒𝑓 term is to keep
the surge speed low during turning maneuvers in order to minimize the vehicle’s turning radius.
To develop a heading controller, the yaw rate equations of motion given by (3), and (7)-(11) were simplified as:
(𝐼𝑧 − 𝑁𝑟̇ )𝑟̇ − (−𝑋𝑢̇ + 𝑌𝑣̇ )𝑢𝑣 − 𝑁𝑟 𝑟 = 𝜏𝑧 .
(23)
As above, input state linearization was used to identify a nonlinear feedback control law for the yaw subsystem. Let 𝜉𝜓 be a
linearizing control related to the commanded torque from the controller 𝜏𝑧 by the relation
𝜏𝑧 = (𝐼𝑧 − 𝑁𝑟̇ )ξ𝜓 − (−𝑋𝑢̇ + 𝑌𝑣̇ )𝑢𝑣 − 𝑁𝑟 𝑟.
(24)
Using this in (23) gives 𝑟̇ = 𝜓̈ = 𝜉𝜓 . If the heading tracking error is defined as 𝑒𝜓 = (𝜓 − 𝜓𝑑 ), where 𝜓𝑑 is the desired heading
angle, and an input of the form 𝜉𝜓 = 𝜓̈ = 𝜓̈𝑑 − 𝑘1 𝑒𝜓 − 𝑘2 𝑒̇𝜓 is selected, the error dynamics will be:
𝑒̈𝜓 + 𝑘2 𝑒̇𝜓 + 𝑘1 𝑒𝜓 = 0,
(25)
which are exponentially stable when the control gains are positive: 𝑘1 > 0 and 𝑘2 > 0. As only the desired yaw 𝜓𝑑 , and not also
𝜓̇𝑑 and 𝜓̈𝑑 , is provided by the ALR system, the control law reduces to:
𝜏𝑧 = (𝐼𝑧 − 𝑁𝑟̇ )(−𝑘1 𝑒𝜓 − 𝑘2 𝑟) − (−𝑋𝑢̇ + 𝑌𝑣̇ )𝑢𝑣 − 𝑁𝑟 𝑟.
(26)
As shown in [24], the combined surge speed and heading control laws make the system state (𝑢, 𝑟, 𝜓)𝑇 bounded and globally
uniformly asymptotically convergent to (𝑢𝑑 , 0, 𝜓𝑑 )𝑇 when 𝑡 → ∞. They are also well-defined and bounded for all 𝑡 > 0 when
{𝑘𝑢 , 𝑘1 , 𝑘2 } > 0.
Examining the sway velocity 𝑣 subsystem,
(𝑚 − 𝑌𝑣̇ )𝑣̇ + (𝑚 − 𝑋𝑢̇ )𝑢𝑟 − 𝑌𝑣 𝑣 = 0,
(27)
And using the error dynamics produced in (20) and (25), one can readily see that the sway subsystem in (27) reduces to,
(𝑚 − 𝑌𝑣̇ )𝑣̇ = −(𝑚 − 𝑋𝑢̇ )𝑢𝑑 𝜓̇𝑑 + 𝑌𝑣 𝑣,
(28)
when 𝑡 → ∞, which is exponentially stable for the setpoint case of 𝜓̇𝑑 = 0. A block diagram of the controller is shown in Fig. 15.
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
21
Fig. 15: Structure of the nonlinear backstepping controller.
During initial on-water controller testing it was found that the waterjet motors were exhibiting a chatter-like behavior. An
examination of the thrust model Section IV.A shows that the commanded speed of each waterjet is sensitive to small changes in
desired thrust. To reduce the chattering effect, the commanded surge force (19) is scaled by 0.5 and the commanded yaw moment
(26) is scaled by 0.05. Coupled with the fact that the waterjet pump motor driver has a maximum speed that it can output, the scale
factors basically act like a saturation function for the commanded motor speed to alleviate the chattering issue.
Despite the fact that a reduced set of the equations of motion in equations (3)-(11) were used to develop the backstepping surge
speed and heading controllers, the backstepping controller was found to be robust to unmodeled dynamics throughout the range of
the USV operating conditions tested.
B. Backstepping Yaw Controller with Adaptive Speed Control
Based on the results of simulation and initial on-water testing of the non-adaptive backstepping controller, the need for adaptive
speed control was established. Although the change in turning dynamics between displacement conditions does not appear to
drastically affect heading control performance, the change in drag significantly affects the performance of surge speed control. To
remedy the issue of steady state speed errors fluctuating with mass and drag parameters throughout the ALR mission, an adaptive
surge speed control law was developed and implemented in combination with the backstepping heading controller developed in
Section VI.A above.
By combining (17) and (18), one can write the surge equation of motion as:
ℎ𝑢̇ − (𝑚 − 𝑌𝑣̇ )𝑣𝑟 − 𝑋𝑢 𝑢 − 𝑋𝑢|𝑢| 𝑢|𝑢| = 𝜏𝑥 ,
(29)
where ℎ ≡ (𝑚 − 𝑋𝑢̇ ). A control law for the thrust commanded by the controller of the form
𝜏𝑥 = − (𝑚 − 𝑌𝑣̇ )𝑣𝑟 − 𝑋̂𝑢 𝑢 − 𝑋̂𝑢|𝑢| 𝑢|𝑢| + 𝑎̂𝑑 𝑢𝑑 ,
(30)
where 𝑋̂𝑢 , 𝑋̂𝑢|𝑢| and 𝑎̂𝑑 are estimated parameters, is chosen to decouple the surge dynamics from those of sway and yaw. With this
control law, (29) becomes
ℎ𝑢̇ = (𝑋𝑢 − 𝑋̂𝑢 )𝑢 + (𝑋𝑢|𝑢| − 𝑋̂𝑢|𝑢| )𝑢|𝑢| + 𝑎̂𝑑 𝑢𝑑 .
(31)
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
22
A linear reference model of the form
𝑢̇ 𝑚 = −𝑎𝑚 𝑢𝑚 + 𝑏𝑚 𝑢𝑑 ,
(32)
is chosen and the model error is defined as 𝑒𝑚 ≡ (𝑢 − 𝑢𝑚 ). With this reference model (31) can be reduced to an equation
describing the surge speed error dynamics:
ℎ𝑒̇𝑚 + ℎ𝑎𝑚 𝑒𝑚 = (ℎ𝑎𝑚 + 𝑎̃1 )𝑢 + 𝑎̃2 𝑢|𝑢| − (ℎ𝑏𝑚 − 𝑎̂𝑑 )𝑢𝑑 ,
(33)
where 𝑎̃1 ≡ (𝑋𝑢 − 𝑋̂𝑢 ) and 𝑎̃2 ≡ (𝑋𝑢|𝑢| − 𝑋̂𝑢|𝑢| ) are parameter estimation errors. An adaptation law of the form
𝑑𝑎̃1
𝑑𝑡
𝑑𝑎̃2
𝑑𝑡
𝑑𝑎̂𝑑
𝑑𝑡
=−
=−
𝑑𝑋̂𝑢
=
𝑑𝑡
𝑑𝑋̂𝑢|𝑢|
𝑑𝑡
−𝛾𝑒𝑚 𝑢 sgn(ℎ)
= −𝛾𝑒𝑚 𝑢|𝑢| sgn(ℎ) ,
=
(34)
−𝛾𝑒𝑚 𝑢𝑑 sgn(ℎ) }
is chosen for stability, where sgn(ℎ) is the sign (signum) function and 𝛾 is a positive adaptation gain. The stability of the surge
speed controller can be shown for the Lyapunov function candidate:
1
2
𝑉 = |ℎ|𝑒𝑚
+ [(ℎ𝑎𝑚 + 𝑎̃1 )2 + 𝑎̃22 + (ℎ𝑏𝑚 − 𝑎̂𝑑 )2 ].
𝛾
(35)
Using |ℎ| = ℎ sgn(ℎ) together with (33) and (34), the derivative of the Lyapunov function reduces to
2 |ℎ|,
𝑉̇ = −2𝑎𝑚 𝑒𝑚
(36)
which is negative definite for 𝑎𝑚 > 0. The coefficients 𝑎𝑚 and 𝑏𝑚 of the reference model (32) can be determined by linearizing
the hyperbolic tangent based model acceleration for the nonlinear backstepping controller (21) with 𝑢𝑑,𝑟𝑒𝑓 = 𝑢𝑑 . When the surge
speed 𝑢 achieved by the USV approaches that of the reference model 𝑢𝑚 , the argument of the hyperbolic tangent function becomes
small and the model acceleration can be approximated as:
𝑢̇ 𝑚 ≈
𝑢̇ 𝑎,𝑚𝑎𝑥 𝑘𝑎,𝑚𝑎𝑥
𝑢̇ 𝑎,𝑚𝑎𝑥
(𝑢𝑑 − 𝑢𝑚 ),
such that 𝑎𝑚 = 𝑏𝑚 = 𝑢̇ 𝑎,𝑚𝑎𝑥 𝑘𝑎,𝑚𝑎𝑥 ⁄𝑢̇ 𝑎,𝑚𝑎𝑥 . A block diagram of the backstepping adaptive controller is shown in Fig. 16.
(37)
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
23
Fig. 16: Structure of the adaptive backstepping controller.
As in Section VI.A, the commanded surge force 𝜏𝑥 is multiplied by a factor of 0.5 to prevent chattering. During field trials the
adaptive backstepping controller was found to work well with large, sudden changes in payload mass (see Section VII). In practice,
it was also found that when the parameter estimates 𝑋̂𝑢|𝑢| and 𝑎̂𝑑 were held constant, the adaptive controller still performed well,
because the linear drag term in the control law (30) dominates the other two terms at low speeds (𝑢 ≤ 1.5 m/s).
C. Comparative Set Point Tests
In order to establish a reasonable comparison between the two controllers, a set point test is conducted on each controller. The test
first consists of placing the USV in a predetermined position at rest with a predetermined initial heading. The test is started by
delivering a constant speed command and an initial desired heading command that is the same as the initial heading of the USV.
The USV accelerates towards the desired speed and also attempts to maintain heading. At a specified time (15 seconds), the heading
command is changed to a value 90 degrees from the initial heading and the USV executes a turn towards the new desired heading.
The metrics for evaluating controller performance during this test are steady state speed and heading error before and after the jump
in heading command. The test is conducted a total of three times for each controller and the performance metrics are averaged over
all three tests.
To quantify the difference in performance between two controllers, the difference between the errors is divided by the larger
error, resulting in the percentage amount that the lower error is ‘better’ when compared to the larger. This metric is relative to the
least desirable result, which is the baseline for comparison. For example, consider results 𝐴 and 𝐵 where 𝐴 is more desirable than
𝐵. The difference between 𝐴 and 𝐵 is divided by 𝐵 and the result 𝜆 = (𝐵 − 𝐴)/max(𝐴, 𝐵) × 100% is said to show that “𝐴
performs (𝐵 − 𝐴)/max(𝐴, 𝐵) × 100% better than 𝐵.” Since the metric we are comparing is steady state error, a positive value for
𝜆 denotes an improvement in A with respect to B. Naturally, a negative value would denote a performance decrease. This convention
applies for all subsequent sections where two control results are compared.
For the results presented in this section, the inputs used for the test are 𝜓𝑑 = 𝜓(𝑡 = 0) for 0 ≤ 𝑡 ≤ 15 seconds, 𝜓𝑑 = 𝜓(𝑡 = 0) +
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
24
90 degrees for 𝑡 > 15 seconds, and 𝑢𝑑 = 1.5 m/s. For the non-adaptive backstepping controller the control gains are set to 𝑘1 =
0.1, 𝑘2 = 1.0, 𝑘𝑢 = 8.0, and 𝑘𝑎,𝑚𝑎𝑥 = 1.2. The adaptive backstepping controller has control gains 𝑘1 = 0.1, 𝑘2 = 1.0, 𝑘𝑎,𝑚𝑎𝑥 =
1.2, and 𝛾 = 0.05. The backstepping controller gains were found (through previous testing) to work well on the USV. The WAMV USV14 was in the lightship condition for the tests, a 10-15 knot sustained wind blew from the southeast (approximately 1600),
and a slight tidal current of <0.2 knots flowed Northerly (approximately 3450). The small wind driven waves were taken to be
negligible in affecting the performance of the vehicle.
The results of one set point test when the USV is controlled by the non-adaptive backstepping controller are shown in Fig. 17.
The average speed error at steady state before the jump in heading is 0.25 m/s, and decreases to 0.24 m/s after the jump. The average
heading error at steady state before the jump is 0.74 degrees and decreases to 0.41 degrees afterwards.
Fig. 17. Results for the set point test using the backstepping (BS) and adaptive backstepping (ABS) controllers.
Results for the same set point test using the adaptive backstepping controller are overlaid on those of the backstepping controller
in Fig. 16 above. The average speed error at steady state before the jump in heading is 0.065 m/s, and 0.070 m/s after the jump.
The average heading error at steady state before and after the jump is 1.9 degrees.
The overall results averaged over three runs for each controller are shown in Table IV. The controllers have been annotated as
ABS and BS adaptive backstepping and backstepping, respectively. For brevity, the charts only show the comparisons of the
controllers after the jump in heading command. The results show that the non-adaptive backstepping controller shows 62% better
heading performance over the adaptive controller before the heading change. After the heading change, the non-adaptive
backstepping controller shows a 78% better heading performance than the adaptive controller.
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
25
TABLE IV
COMPARISON OF CONTROLLERS FOR THE SET POINT TEST. USING THE METRIC IDENTIFIED IN SECTION VI.C, THE STEADY STATE SPEED ERROR OF THE ABS
CONTROLLER IS 71% BETTER THAN THE BS CONTROLLER AFTER THE JUMP. HOWEVER, THE STEADY STATE HEADING ERROR OF THE BS CONTROLLER IS 78%
BETTER THAN THE ABS CONTROLLER AFTER THE JUMP.
Controller:
ABS
BS
Steady State Speed Error Before Jump (m/s):
0.065
0.25
Percent Error:
4%
17%
Steady State Speed Error After Jump (m/s):
0.070
0.24
Percent Error:
5%
16%
Steady State Heading Error Before Jump (deg):
Steady State Heading Error After Jump (deg):
1.9
0.74
1.9
0.41
VII. VARIABLE MASS AND DRAG EXPERIMENTS
Mock ALR tests were performed in Dania Beach and Hollywood, Florida USA. In order for the vehicle to operate effectively, it
cannot be used in sea state greater than 1 or winds higher than 15 knots. The maximum sea state and wind conditions for operations
were determined by trial and error.
During the experiments the vehicle was driven under closed loop speed and heading control in the ICW. Vehicle heading,
acceleration, position, speed over ground, and motor command data were recorded, as were the wind and current speed/direction,
and wave characteristics. The WAM-V USV14 was tested over a range of surge speeds that are expected for the ALR mission.
A. Variable Mass
One of the possible scenarios for the launch phase is the AUV being dropped into the water from the USV’s undercarriage by
the ALR system. In order to simulate this scenario, a test case where the USV’s mass properties vary during operation is conducted.
Specifically, the mass change test involves the USV beginning in the full displacement condition and coming to a set desired speed
and heading. The mass and drag properties of the lightship condition are incorporated into the adaptive and non-adaptive
backstepping control laws (creating the uncertain mass and drag situation). Once the USV reaches steady state in both speed and
heading, an AUV-like mass is dropped from the payload and the USV continues along its course. No real-time feedback is provided
to the low-level controller about the actual displacement condition during the test, either before or after the drop. The performance
of each controller across the two displacement conditions is then compared against the other controllers.
For the experimental setup, the AUV-like mass is a Polyvinyl Chloride (PVC) pipe measuring 130 cm in length, 12 cm in
diameter, weighted with lead shot and capped at both ends (Fig. 18a). The overall weight in air of the PVC apparatus is 39 kg
(approximately the same as a REMUS 100 AUV), with the weight evenly distributed along the length. Four buoys (each rated at
13.6 kg buoyancy) are lashed to the PVC apparatus, two at each end. The device is secured to the undercarriage of the USV using
both fore and aft lines. The aft line runs through a loop directly above the aft section of the PVC device and forward to a quick-
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
26
release mechanism, where both lines are clipped in, thus suspending the PVC device under the payload tray of the USV and along
the USV centerline. The quick release mechanism holds the two lines together until a pin is pulled, which releases the connection
and the device falls into the water. A line is tied to the pin, which allows a user to pull the pin from a distance (from a chase boat
in front of the USV during the test).
A view of the USV in the full displacement condition from below as it is lifted with a crane is shown in Fig. 18b. Notice that the
AUV-like object is located along the centerline of the USV, and the two additional PVC pipes provide the weight of the intended
ALR mechanical system (68 kg).
a)
b)
Fig. 18. a) PVC device used to mimic a REMUS 100 AUV. b) Bottom view of the WAM-V USV14 in the full displacement
condition.
The results of the mass change test when the USV is controlled by the backstepping controller tuned to the lightship condition
are shown in Fig. 18. For the time up to the drop of the PVC device, the USV is in the full displacement condition traveling along
a desired heading of 150o at a desired speed of 1 m/s. The quick release mechanism can only be pulled from the front of the USV.
Accordingly, the drop is marked by a spike in surge velocity (and sometimes a perturbation to the heading) which is partly due to
the impulsive force applied to the quick release line. Because the impulsive force is unmeasured, the data corresponding to the time
when the force is applied are not included in the steady state averages. A picture, taken just before the mass drop occurs, is shown
in Fig. 19. The average speed error for the backstepping controller at steady state before the drop is 0.14 m/s, and decreases to 0.13
m/s after the drop. The data are shifted in time so that the results are synchronized with those of the adaptive backstepping controller
presented below where the drop occurs at t = 78 seconds. The average heading error at steady state before the drop is 1.4 degrees
and increases to 2.6 degrees afterwards.
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
27
Fig. 19. Example results of the variable mass test using the backstepping controller and the adaptive backstepping controller.
Data from the different tests are synched according to the time of the mass drop for the ABS controller. The large perturbations
in surge speed and heading in the adaptive backstepping controller test at the time of the mass drop are attributed to a large
impulse delivered through the quick release line.
Fig. 20. Snapshot just before the PVC device is released from the USV undercarriage during the variable mass test.
Results for the variable mass test using the adaptive backstepping controller tuned to the lightship condition are also shown in
Fig. 18 above. The average speed error at steady state before the drop is 0.046 m/s, and decreases to 0.042 m/s after the drop,
which is synched to the backstepping controller such that it occurs at t = 78 seconds. The average heading error at steady state
before the drop is 3.3 degrees and increases to 4.7 degrees afterwards.
A comparison of the speed and heading errors for the three controllers is shown in Table V. For brevity, the full displacement
condition results are excluded from the comparison table. The results show that for each backstepping controller the speed error
decreased when the USV transitioned from the full displacement to the lightship condition. The magnitude of the speed errors,
however, show that the adaptive controller performs 67% better than the backstepping controller when in the full displacement
condition and performs 68% better than the backstepping controller when in the lightship condition. Note that the short period of
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
28
time during which the USV is reacting to the impulse of the quick-release mechanism are excluded from the analysis of the steady
state characteristics of the controlled USV.
TABLE V
COMPARISON OF THE ADAPTIVE BACKSTEPPING AND BACKSTEPPING CONTROLLERS FOR THE VARIABLE MASS TEST. USING THE METRIC IDENTIFIED IN SECTION
VI.C, THE STEADY STATE SPEED ERROR OF THE ABS CONTROLLER IS 68% BETTER THAN THE BS CONTROLLER AFTER THE DROP. HOWEVER, THE STEADY STATE
HEADING ERROR OF THE BS CONTROLLER IS 43% BETTER THAN THE ABS CONTROLLER AFTER THE DROP.
Controller:
ABS
BS
Steady State Speed Error Before Drop (m/s):
0.05
0.14
Percent Error:
4.6%
14%
Steady State Speed Error After Drop (m/s):
0.04
0.13
Percent Error:
Steady State Heading Error Before Drop (deg):
Steady State Heading Error After Drop (deg):
4%
3.3
13%
1.3
3.8
1.4
The results for heading do not show the same trend in performance. The non-adaptive backstepping controller shows 61% better
heading performance over the adaptive controller in the full displacement condition. For the lightship condition, the non-adaptive
backstepping controller shows 43% better heading performance over the adaptive controller.
B. Variable Drag
A second scenario expected for the ALR mission is when the USV is navigating a straight trajectory in the lightship condition
and the AUV (approaching from behind) impulsively docks with a recovery device suspended below the surface. This towing
situation would create an abrupt change in drag experienced by the USV, and also alter the turning dynamics, i.e. a change in both
the heading and surge subsystems of the USV would occur. As in the change of mass testing, it is assumed that no real-time
feedback about the presence of the AUV is provided to the USV low-level controller, which is tuned for the lightship condition.
To create this scenario, the AUV-like PVC device was secured at two tow points on the aft section of the USV payload tray
using a tow line. Each line is 7.3 m long, and attached to the front section of the PVC device. The PVC device is initially carried
aboard a chase boat behind the USV so that the towing lines are slack. Once the USV comes to a steady heading and speed, the
PVC device is dropped into the water from the chase boat and begins to be towed by the USV (Fig. 21). For this set of tests, since
the mass of the USV is not intended to be altered, the PVC device was weighted to 28 kg, which eased the lifting of the device to
and from the chase boat. The shape of the device was unaltered from the other tests; however the buoys sat higher in the water
which decreased the overall drag of the device. In order to determine the amount of drag experienced by the USV during the
variable drag tests, the PVC device is towed behind the chase boat. The tow line is attached to a scale, which is then attached to the
tow point on the chase boat. The control box is used to record the boat’s velocity and values are read from the scale for different
speeds ranging from 0.5 to 1.6 m/s. Fig. 22 shows the drag model developed for the PVC device. Each point on the plot represents
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
29
the average value of force recorded for that speed and the solid line is the curve fit to the points, valid for the range from 0.5 to 1.6
m/s. The data are averaged over a total of 76 force readings from the scale. For the data presented in this section, with a desired
speed of 1.0 m/s, the drag of the PVC device is estimated at 84 N. The same drag test conducted for the REMUS 100 AUV resulted
in a drag value of no more than 23 N at 1.0 m/s. The variable drag tests, therefore, allow for a towed recovery device drag of at
least 61 N.
Fig. 21. Picture of the WAM-V USV14 towing the AUV-like object through the water during the variable drag testing.
Fig. 22. Drag model for the 28 kg PVC device used in the variable drag tests.
Results of the variable drag test using the non-adaptive backstepping controller tuned to the lightship condition are shown in Fig.
23. The average speed error at steady state before the drop is 0.13 m/s, and increases to 0.17 m/s after the drop. The results are
shifted in time so that the drop occurs at t = 42 seconds and are synchronized with those of the adaptive backstepping controller.
The average heading error at steady state before the drop is 0.65 degrees and increases to 0.94 degrees afterwards. The same test
was conducted for a desired speed of 1.2 m/s, however the drag of the AUV-like PVC device was too high and the propulsion
system could not achieve the desired speed; the heading control also degraded significantly. Only the tests where desired speed
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
30
was 1.0 m/s are shown. The quantitative results are averages of the data shown in the figures only, and are similar to the results
from other tests.
Fig. 23. Results for the variable drag test using the backstepping and the adaptive backstepping controller. The two tests are
synched according to the time of the drop.
Results using the adaptive backstepping controller tuned initially to the lightship condition are also shown in Fig. 22. The average
speed error at steady state before the drop is 0.091 m/s and after the drop it is 0.090 m/s. The drop occurs at t = 42 seconds. The
average heading error at steady state before the drop is 1.81 degrees and decreases to 0.35 degrees afterwards.
A comparison of the speed and heading errors for the three controllers is shown in Table VI. For brevity, the results for the
comparison chart before the drop are excluded. The results show that for each controller the speed error increased when the USV
transitioned from the lightship to towing condition; however, the adaptive controller showed a 58% improvement on the
backstepping controller while in the lightship displacement condition. The adaptive controller performs 64% better than the
backstepping controller while in the towing condition.
TABLE VI
COMPARISON OF THE ADAPTIVE BACKSTEPPING AND BACKSTEPPING CONTROLLERS FOR THE VARIABLE DRAG TEST. USING THE METRIC IDENTIFIED IN SECTION
VI.C, THE STEADY STATE SPEED ERROR OF THE ABS CONTROLLER IS 64% BETTER THAN THE BS CONTROLLER AFTER THE DROP. HOWEVER, THE STEADY STATE
HEADING ERROR OF THE BS CONTROLLER IS 66% BETTER THAN THE ABS CONTROLLER AFTER THE DROP.
Controller:
ABS
BS
Steady State Speed Error Before Drop (m/s):
0.06
0.13
Percent Error:
5.5%
13%
Steady State Speed Error After Drop (m/s):
0.06
0.18
Percent Error:
Steady State Heading Error Before Drop (deg):
Steady State Heading Error After Drop (deg):
6%
0.1
18%
0.5
2.6
0.9
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
31
The results for heading are as follows. The adaptive backstepping controller shows 83% better heading performance over the
non-adaptive backstepping controller in the lightship condition. For the towing condition, the non-adaptive backstepping controller
shows 66% better heading performance over the adaptive controller.
The maximum speed at which the WAM-V can maintain heading control while towing a device with 84 N of drag is 1.0 m/s.
This number will decrease if the drag of the recovery device decreases because the initial estimate of the REMUS 100 drag at 1.0
m/s is 23 N – far less than the AUV-like object used for testing (see Fig. 22).
C. Variable Mass and Drag
A third scenario expected for the ALR mission is where the USV is navigating a straight trajectory in the full displacement
condition when the AUV is lowered/dropped into the water and is still attached to the ALR device for some time before being
completely launched from the USV. This carriage to towing transition would create a change in the drag and mass of the USV. The
turning dynamics would also be altered as in the case of the variable drag testing presented in Section VII.B. As in the other testing,
it is assumed that no real-time feedback about the presence of the AUV is provided to the USV low-level controller.
To conduct the variable mass and drag experiment, the setups for the variable mass (Section VII.A) and variable drag test (Section
VII.B) are combined. The PVC device is loaded in the undercarriage of the USV and the tow lines are coiled on the payload tray
and attached at the ends to the rear tow points on the USV. The USV comes to a steady heading and speed in the full displacement
condition, the quick release is pulled, and the device falls into the water. As the USV continues forward the tow lines are brought
into tension and the AUV is towed behind the USV.
Results of the variable mass and drag test using the non-adaptive backstepping controller tuned to the lightship condition are
shown in Fig. 24. In this case, the PVC device separated cleanly from the suspension points when the quick-release was pulled so
the device immediately transitioned to towing at distance. The average speed error at steady state before the drop is 0.14 m/s, and
increases to 0.19 m/s after the drop. As above, the results of the backstepping controller test are synchronized with those of the
adaptive backstepping controller, such that the drop occurs at t = 87 seconds. The average heading error at steady state before the
drop is 0.36 degrees and increases to 2.9 degrees afterwards. Results using the adaptive backstepping controller tuned initially to
the lightship condition are also shown in Fig. 23. The average speed error at steady state before the drop is 0.071 m/s and after the
drop it is 0.066 m/s. The average heading error at steady state before the drop is 3.1 degrees and increases to 3.4 degrees afterwards.
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
32
Fig. 24. Results of the variable mass and drag test using the backstepping and the adaptive backstepping controllers. The data
from the different tests are synched according to the time of the mass drop.
A comparison of the speed and heading errors for the controllers is shown in Table VII. For brevity, the comparison chart excluded
the results before the drop. The results show that for the non-adaptive controllers the speed error increased when the USV
transitioned from the full displacement to towing condition. The magnitude of the speed errors show that the adaptive controller
retains an advantage of 48% on the backstepping controller while in the full displacement condition. The adaptive controller
performs 65% better than the backstepping controller while in the towing condition. Overall, the adaptive controller out-performed
both non-adaptive controllers in in transitioning from full displacement to lightship condition.
The results for heading are as follows. The non-adaptive backstepping controller shows 89% better heading performance over
the adaptive backstepping controller. For the towing condition, the non-adaptive backstepping controller shows 15% better heading
performance over the adaptive controller.
The main findings of the set point, variable mass, variable drag, and variable mass and drag tests are that the surge speed
performance of the adaptive controller is superior to the non-adaptive backstepping controller in the transitioning region. Finally,
the adaptive controller shows less effects to changing mass and drag parameters than the backstepping controller, which is the
ultimate goal of the low-level control for the ALR mission.
VIII. CONCLUDING REMARKS
In this paper, we present the results of experimental testing of an unmanned surface vehicle under closed loop speed and heading
control through lightship, transitional, and full displacement conditions. The tests were performed to evaluate the tracking
performance of backstepping and model reference adaptive controllers under the variable mass and drag conditions expected during
the automatic launch and recovery of an AUV from the USV. System identification of the data recorded from the water-borne
maneuvering trials was used to find approximate values of hydrodynamic coefficients for the vehicle that adequately reproduce its
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
33
motion in zig-zag and acceleration tests.
TABLE VII
COMPARISON OF THE ADAPTIVE BACKSTEPPING AND BACKSTEPPING CONTROLLERS FOR THE VARIABLE MASS AND DRAG TEST. USING THE METRIC IDENTIFIED IN
SECTION VI.C, THE STEADY STATE SPEED ERROR OF THE ABS CONTROLLER IS 65% BETTER THAN THE BS CONTROLLER AFTER THE DROP. HOWEVER, THE STEADY
STATE HEADING ERROR OF THE BS CONTROLLER IS 15% BETTER THAN THE ABS CONTROLLER AFTER THE DROP.
Controller:
ABS
BS
Steady State Speed Error Before Drop (m/s):
0.071
0.14
Percent Error:
7%
14%
Steady State Speed Error After Drop (m/s):
0.066
0.19
Percent Error:
Steady State Heading Error Before Drop (deg):
Steady State Heading Error After Drop (deg):
6%
3.1
19%
0.36
3.4
2.9
We modified the backstepping controller for mono-hull vessels developed by Liao et al. (2010) [24] to account for the twin hull
dynamics and actuator saturation limits of a wave-adaptive modular vessel. The modified controller was then used as the basis for
a direct method model reference adaptive controller [17]. The reference model was derived using a physics-based model of the
vessel that was tuned using system identification test data. We performed a comparative analysis of the steady state tracking
performance between the two controllers (TABLE VII). Results show that the adaptive controller is superior in surge speed tracking
performance – the improvement comes with a slight penalty in the effectiveness of heading control. The variable displacement and
drag experiments show that the vehicle exhibits steady state surge tracking errors when an adaptive control approach is not used.
The variable mass and drag experiment is the most representative of an ALR mission, and clearly demonstrates that the adaptive
backstepping controller outperformed the non-adaptive backstepping controller in terms of speed and exhibited a similar response
as the non-adaptive backstepping case in heading. These same results are illustrative of the separate variable mass and variable
drag test, where the adaptive backstepping controller clearly outperformed in speed, while displaying a common performance in
heading.
REFERENCES
[1] Aguiar, A. P. and Hespanha, J. P. “Logic-based switching control for trajectory-tracking and path-following of underactuated
autonomous vehicles with parametric modeling uncertainty.” In Proc. of the 2004American Control Conference. Vol. 4, pp.
3004-3010, Boston, MA, USA, 2004.
[2] Aguiar, A. P. and Pascoal, A. M. Dynamic positioning and way-point tracking of underactuated AUVs in the presence of ocean
currents. International Journal of Control vol. 80, no. 7, pp.1092-1108, 2007.
[3] Alvarez, J., Bertaska, I., and von Ellenrieder, K. “Nonlinear Control of an Amphibious Vehicle.” Proc. ASME Dynamic
Systems Control Conference, Stanford, CA USA, 2013.
[4] Annamalai, A. S. K., Sutton, R., Yang, C., Culverhouse, P. and Sharma, S. Robust Adaptive Control of an Uninhabited Surface
Vehicle. J. Intelligent & Robotic Systems, May 2014. DOI 10.1007/s10846-014-0057-2.
[5] Ashrafiuon, H. Muske, K. R. and McNinch, L.C., “Review of Nonlinear Tracking and Setpoint Control Approaches for
Autonmous Underactuated Marine Vehicles,” in American Control Conference, Baltimore, MD. USA, 2010.
[6] Ashrafiuoun H. Muske, K. R. McNinch, L. C. and Soltan, R. A. Sliding-mode Tracking Control of Surface Vessels. IEEE
Transactions of Industrial Electronics, vol. 55, no. 11, pp. 4004-4011, 2008.
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
34
[7] Bertaska, I., Alvarez, J., et al., “Experimental Evaluation of Approach Behavior for Autonomous Surface Vehicles.” Proc.
ASME Dynamic Systems Control Conference, Stanford, CA USA, 2013.
[8] Bertaska, I.R. and von Ellenrieder, K.D., " Supervisory Switching Control of an Unmanned Surface Vehicle," Proc. MTS/IEEE
OCEANS’15 Washington DC, USA pp.1-10, 2015.
[9] Bertaska, I.R, et al. “Experimental Evaluation of Automatically Generated Behaviors for USV Operations,” J. Ocean
Engineering¸ vol. 106, pp. 496-514, 2015.
[10] Caccia, M. Bibuli, M. Bono, R. and Bruzzone, G. Basic navigation, guidance and control of an unmanned surface vehicle.
Autonomous Robots, vol. 25, no. 4, pp. 349-365, 2008.
[11] Dhanak, M. R., Ananthakrishnan, P., Frankenfield, J., & von Ellenrieder, K. "Seakeeping Characteristics of a Wave-Adaptive
Modular Unmanned Surface Vehicle." ASME 2013 32nd Intl Conf on Ocean, Offshore and Arctic Engineering. American
Society of Mechanical Engineers, 2013.
[12] Duerr, P. and von Ellenrieder, K.D., Scaling and Numerical Analysis of Nonuniform Waterjet Pump Inflows, IEEE J. Oceanic
Engineering, vol.40, no.3, pp.701-709, 2015.
[13] Faltinsen, O. M. Hydrodynamics of high-speed marine vehicles. Cambridge University Press, 2005.
[14] Fossen, T. I. Guidance and Control of Ocean Vehicles. Chichester: John Wiley and Sons Ltd, 1994.
[15] Fossen, T. I. and Strand, J. P. Tutorial on nonlinear backstepping: applications to ship control. Modeling, Identification and
Control, vol. 20, no. 2, pp. 83-134, 1999.
[16] Fratello, J. and Ahmadian, M. "Multi-body Dynamic Simulation and Analysis of Wave-adaptive Modular Vessels." Proc. 11th
Intl Conf on Fast Sea Transportation FAST 2011, Honolulu, Hawaii, USA, 2011.
[17] Ghommam, J. Mnif, F. Benali, A. and Derbel, N. Asymptotic backstepping stabilization of an underactuated surface
vessel. Control Systems Technology, IEEE Transactions on, vol. 14, no. 3, pp. 1150-1157, 2006.
[18] International Maritime Organization. “Explanatory Notes to the Standards for Ship Manoeuvrability.” December 2002.
London.
[19] Khalil, H.K. Nonlinear Systems, Upper Saddle River, NJ: Prentice Hall, 2002.
[20] Klinger, W. Bertaska, I. Alvarez, J. and von Ellenrieder, Karl, “Controller Design Challenges for Waterjet Propelled Unmanned
Surface Vehicles with Uncertain Drag and Mass Properties.” Proc. MTS/IEEE Oceans’13 San Diego, CA. pp. 23-26. 2013.
[21] Klinger, W. Bertaska and von Ellenrieder, K. D. “Experimental Testing of an Adaptive Controller for USVs with Uncertain
Displacement and Drag.” Proc. MTS/IEEE Oceans’14 St. John’s, NS Canada. 2014.
[22] Kragelund, S, et al. “Adaptive Speed Control for Autonomous Surface Vessels,” in IEEE/MTS Oceans Conference 2013
(Oceans ’13), San Diego, CA, USA. 2013.
[23] Li. J, et al. “Robust Adaptive Backstepping Design for Course-keeping Control of Ship with Parameter Uncertainty and Input
Saturation,” in 2011 International Conference on Soft Computing and Pattern Recognition (SoCPaR), Dalian, Chiana, 2011.
[24] Liao, Y. Pang, Y. and Wan, L. "Combined speed and yaw control of underactuated unmanned surface vehicles," 2010 2nd
International Asia Conference on Informatics in Control, Automation and Robotics (CAR), vol.1, pp. 157-161, March 2010.
[25] Mahini, F. and Ashrafiuon H, “Autonomous Surface Vessel Target Tracking Experiments in Simulated Rough Sea
Conditions,” in ASME Dynamci Systems and Control Conference, Ft. Lauderdale, FL, USA 2012.
[26] Marquardt, J. G. Alvarez, J. and von Ellenrieder, K. D. (2014) Characterization and System Identification of an Unmanned
Amphibious Tracked Vehicle. IEEE J Oceanic Engineering, 39(4):641-661.
[27] Miranda, M., Beaujean, P. P., An, E., & Dhanak, M. "Homing an Unmanned Underwater Vehicle Equipped with a DUSBL
to an Unmanned Surface Platform: A Feasibility Study." Proc. MTS/IEEE Oceans’13 San Diego, CA. pp. 23-26. 2013.
[28] Panteley, E. and Loria, A. Growth rate conditions for uniform asymptotic stability of cascaded time-varying
systems. Automatica, vol. 37, no. 3, 453-460, 2001.
[29] Pearson, D. An, P.-C., Dhanak, M. R. von Ellenrieder, K. D. and Beaujean, P.-P. “High-level fuzzy logic guidance system for
an unmanned surface vehicle (USV) tasked to perform autonomous launch and recovery (ALR) of an autonomous underwater
vehicle (AUV).” Proc. 2014 IEEE/OES Autonomous Underwater Vehicles (AUV) Oxford, MS pp.1-15, 2014.
[30] Peterson, A. Ahmadian, M. Craft, M. and Shen, A. "Simulation and scale-testing to improve the next generation of waveadaptive modular vessels." Proc. 2013 Grand Challenges on Modeling and Simulation Conference, p. 19. Society for Modeling
& Simulation International, 2013.
[31] Qu, H. Sarda, E.I. Bertaska, I.R. and von Ellenrieder, K.D., "Wind feed-forward control of a USV," Proc. MTS/IEEE
OCEANS’15 Genova, Italy pp.1-10, 2015.
[32] Qu, H. Sarda, E.I. Bertaska, I.R. and von Ellenrieder, K.D., "Adaptive Wind Feedforward Control of an Unmanned Surface
Vehicle for Station Keeping," Proc. MTS/IEEE OCEANS’15 Washington DC, USA pp.1-10, 2015.
[33] Sarda, E. and Dhanak, M. R. "Unmanned recovery of an AUV from a surface platform." Proc. MTS/IEEE Oceans’13 San
Diego, CA. pp. 23-26. 2013.
[34] Sarda, E.I. Dhanak, M. R. and von Ellenrieder, K. D. “Concept for a USV–based autonomous launch and recovery system,”
Proc. ASNE Launch & Recovery Symp. 2014, Linthicum, MD USA, November 2014.
[35] Sarda, E.I. Bertaska, I.R. Qu, A. and von Ellenrieder, K.D., "Development of a USV station-keeping controller," Proc.
MTS/IEEE OCEANS’15 Genova, Italy pp.1-10, 2015.
Please cite published version: https://doi.org/10.1109/JOE.2016.2571158
35
[36] Skjetne, R., Smogeli, O.N., and Fossen T.I. “A Nonlinear Ship Manoeuvering Model: Identification and Adaptive Control
With Experiments for a Model Ship,” Modeling, Identification, and Control, vol 25, no. 1, pp 3-27, 2004.
[37] Slotine, J.-J. E. and Li, W.-P. Applied Nonlinear Control. New Jersey: Prentice Hall, 1991.
[38] SNAME “Nomenclature for Treating the Motion of a Submerged Body through a Fluid”. The Society of Naval Architects and
Marine Engineers, Technical and Research Bulletin No. 1-5, pp. 1-15. April 1950.
[39] Sonnenburg, C.R., and Woolsey, A.C. “Modeling, Identification, and Control of an Unmanned Surface Vehicle,” J. Field
Robotics, vol 30, no. 3, pp1-28, 2013.
[40] Svec, Peter, Brual C. Shah, Ivan R. Bertaska, Jose Alvarez, Armando J. Sinisterra, Karl von Ellenrieder, Manhar Dhanak, and
S. K. Gupta. "Dynamics-aware target following for an autonomous surface vehicle operating under COLREGs in civilian
traffic." In Intelligent Robots and Systems (IROS), 2013 IEEE/RSJ International Conference on, pp. 3871-3878. IEEE, 2013.
[41] von Ellenrieder, K. D. Free Running Tests of a Waterjet Propelled Unmanned Surface Vehicle. J. Mar. Engng & Tech 12(1):19, 2013.
Wilhelm Klinger received the B.S. degree in ocean engineering from the U.S. Naval Academy, Annapolis, MD USA in 2012 and M.S.
degree in ocean engineering from Florida Atlantic University, Boca Raton, FL USA in 2014. He worked as a Research Assistant at
Florida Atlantic University’s Seatech Institute for Ocean and Systems Engineering, Dania Beach, FL USA while completing his M.S.
degree and is currently in flight training for the US Navy. Mr. Klinger is a member of the Marine Technology Society and Tau Beta Pi
Engineering Honor Society.
Ivan Bertaska (M’13) received the B.S. and M.S. degree in ocean engineering from Florida Atlantic University, Boca Raton, FL USA in
2012 and 2013, respectively. He is currently a Ph.D. candidate at Florida Atlantic University’s Seatech Institute for Ocean and Systems
Engineering, Dania Beach, FL USA.
Mr. Bertaska is a member of the Tau Beta Pi Engineering Honor Society and the IEEE Oceanic
Engineering Society. His areas of interest include unmanned vehicle systems control, software architecture and navigation.
Karl von Ellenrieder (M’08) received the B.S. degree in aeronautics and astronautics from the Massachusetts Institute of Technology (MIT),
Cambridge, MA USA in 1990 and the M.S. and Ph.D. degrees in aeronautics and astronautics from Stanford University, Stanford, CA USA
in 1992 and 1998, respectively. He is currently Associate Director of The Institute for Ocean and Systems Engineering (SeaTech),and a
Professor in the Department of Ocean & Mechanical Engineering at Florida Atlantic University in Dania Beach, FL. His research interests
include unmanned surface vehicles, marine hydrodynamics and ship design. Prof. von Ellenrieder is a Member of the Society of Naval
Architects and Marine Engineers, the Marine Technology Society and the IEEE Oceanic Engineering Society.
Manhar Dhanak is the Director of The Institute for Ocean and Systems Engineering (SeaTech), Director of Research for Southeast National
Marine Renewable Energy Center and Professor of Ocean Engineering at Florida Atlantic University. He is the past chair (2003-2009) of
the Department of Ocean Engineering at FAU. He is a graduate of Imperial College, University of London and served as a Research Scientist
at Topexpress Ltd., Cambridge, UK and as a Senior Research Associate at University of Cambridge before joining FAU. Dr. Dhanak has
research interests in hydrodynamics, physical oceanography, autonomous underwater vehicles (AUV) and ocean energy.
| 3cs.SY
|
arXiv:cs/0111043v1 [cs.PL] 16 Nov 2001
Prototyping clp(fd) Tracers: a Trace Model
and an Experimental Validation Environment
∗†
Ludovic Langevine, Pierre Deransart, Mireille Ducassé and Erwan Jahier
INRIA/INSA, INRIA, IRISA/INSA, IRISA/IFSIC
{Ludovic.Langevine Pierre.Deransart}@inria.fr, {Mireille.Ducasse Erwan.Jahier}@irisa.fr
Abstract
Developing and maintaining CLP programs requires visualization and explanation
tools. However, existing tools are built in an ad hoc way. Therefore porting tools from
one platform to another is very difficult. We have shown in previous work that, from a
fine-grained execution trace, a number of interesting views about logic program executions
could be generated by trace analysis.
In this article, we propose a trace model for constraint solving by narrowing. This trace
model is the first one proposed for clp(fd) and does not pretend to be the ultimate one.
We also propose an instrumented meta-interpreter in order to experiment with the model.
Furthermore, we show that the proposed trace model contains the necessary information
to build known and useful execution views. This work sets the basis for generic execution
analysis of clp(fd) programs.
[23] is a comprehensive version of this paper.
1
Introduction
Developing and maintaining CLP programs benefits from visualization and explanation tools
such as the ones designed by the DiSCiPl European project [10]. For example, the CHIP
search-tree tool [28] helps users understand the effect of a search procedure on the search space;
and the S-Box model [17] allows users to inspect the store with graphical and hierarchical
representations.
However, existing tools are built in an ad hoc way. Therefore porting tools from one platform to another is very difficult. One has to duplicate the whole design and implementation
effort every time.
We have shown in previous work that, from a fine-grained execution trace, a number of
interesting views about logic program executions could be generated by trace analysis [11, 19].
Here we want to generate a generic trace that can be used by several debugging tools. Each
tool gives a useful view of the clp(fd) execution (e.g. the labelling tree or the domain state
evolution during a propagation stage). Each tool selects the relevant data in the generic
∗
This work is partly supported by OADymPPaC, a French RNTL project [26].
In A. Kusalik (ed), Proceedings of the Eleventh Workshop on Logic Programming Environments
(WLPE’01), December 1, 2001, Paphos, Cyprus. Computer Research Repository (http://www.acm.org/corr/),
cs.PL/0111043; whole proceedings: cs.PL/0111042.
†
1
2
WLPE’01
trace in order to build its execution view. Therefore, the trace is supposed to contain all
potentially useful information. The tools can be developed simultaneously and reused on
different platforms 1 . This approach assumes that there is a generic trace model. In order to
reach this objective, we start with a general model of finite domain constraint solving which
allows to define generic trace events.
So far, no fine-grained tracers exist for constraint solvers. Implementing a formally defined
tracer is a delicate task, if even at all possible. This is especially true in the context of
constraint solving where the solvers are highly optimized. Therefore, before going for a “real”
implementation, it is essential to elaborate trace models and experiment them.
In order to rigorously define the constraint solving trace model, and as we did for Prolog
[20], we use an operational semantics based on [2] and [15]. We define execution events of
interest with respect to this semantics. On one hand, this formal approach prevents the finedgrained trace model to be platform dependent. On the other hand, the consistence between
the model and concrete clp(fd) platforms has to be validated.
In order to develop a first experimentation, the formal model is implemented as an instrumented meta-interpreter 2 which exactly reflects it.
The proposed instrumented meta-interpreter is useful to experimentally validate the model.
It is based on well known Prolog meta-interpretation techniques for the Prolog part and on
the described operational semantic for the constraint part.
The ultimate goal of the meta-interpreter (as in [20]) is to provide an executable specification of traces. The traces generated by the meta-interpreter could then be used to validate
a real (and efficient) tracer. At this stage, efficiency of the meta-interpreter is not a key issue;
it is used as a prototype rather than as an effective implementation.
Such a tracer may generate a large volume of data. The generic trace must be analyzed
and filtered in order to condense this data down to provide the sufficient information to the
debugging tools. To push further the validation we have mimicked a trace analyzer à la
Opium [12] which allow to obtain different kinds of views of the execution. This is a way to
show that the proposed trace model contains the necessary information to reproduce several
existing clp(fd) debugging tools.
In this article, due to lack of space we concentrate on the constraint solving part, a trace
of the logic programming part “à la Byrd” [3] could be integrated in the meta-interpreter
(see for example [13]).
The contribution of this article is first to rigorously define a trace model for constraint
solving and to provide an environment to experimentally validate it. This sets the basis for
generic execution analysis of clp(fd) programs.
In the following, Section 2 formally defines an execution model of clp(fd) narrowing. In
particular, it defines a 8 steps operational semantics of constraint solving. These steps are the
basis for the trace format defined in Section 3. Section 4 explains how to build an instrumented
meta-interpreter in order to build an experimental environment. Section 5 briefly describes
experiments with the proposed trace and our trace analyzer. Finally, Section 6 discusses the
content of the trace with respect to existing debugging tools.
1
Our OADymPPaC partners will use our trace model to build tools in the CHIP [7], CLAIRE [6] and
GNU Prolog [16] platforms.
2
The meta-interpreter approach to trace clp(fd) program executions has already been used in the APT tool
by Carro and Hermenegildo [4]. However, APT does not access the propagation details. We propose a more
informative trace.
Prototyping clp(fd) Tracers
2
3
Operational Semantics of Constraint Programming
In this section, we propose an execution model of (finite domain) constraint programming
which is language independent. The operational semantics of constraint programming results from the combination of two paradigms: control and propagation. The control part
depends on the programming language in which the solver is embedded, and the propagation
corresponds to narrowing. Although the notions introduced here are essentially language
independent, we will illustrate them in the context of clp(fd).
2.1
Basic Notations
In the rest of the paper, P(A) denotes the power set of A; r|w denotes the restriction of the
relation r ⊆ A × B to w ⊆ A: r|w = {(x, y) | x ∈ w ∧ (x, y) ∈ r}.
The following notations are attached to variables and constraints: V is the set of all the
finite domain variables of the problem; D is a finite set containing all possible values for
variables in V; D is a function D : V → P(D), which associates to each variable x its current
domain, denoted by Dx ; minx and maxx are respectively the lower and upper bounds of Dx ;
C is the set of the problem constraints; var is a function C → P(V), which associates to each
constraint C ∈ C the set of variables of the constraint.
2.2
Reduction Operators
Many explanation tools focus on domain reduction (see for example [15, 21]). When constraints are propagated, the evolution of variable domains is a sequence of withdrawals of
inconsistent values. At each step, for a given constraint, a set of inconsistent values is withdrawn from one (and only one) domain. These values can be determined by algorithms such
as “node consistency”, “arc consistency”, “hyper-arc consistency” or “bounds consistency”
described for example by Marriott and Stuckey [24].
Following Ferrand et al [15], we define reduction operators. The application of all the
reduction operators of a constraint gives the narrowing operator introduced by Benhamou [2].
Definition 2.1 (Reduction operator). A reduction operator redxC is a function attached
to a constraint C and a variable x. Given the domains of all the variables used in C, it
returns the domain Dx without the values of x which are inconsistent with the domains of the
other variables. The set of withdrawn values is denoted Wx .
redxC (D|var(C) ) = Dx − Wx .
There are as many reduction operators attached to C as variables in var(C). In general,
for efficiency reasons, a reduction operator does not withdraw all inconsistent values.
A simple example of reduction operator for C ≡ x > n, where n is a given integer, is
redxC (D|{x} ) = Dx − {v | v ∈ Dx ∧ v ≤ n}.
The evolution of the domains can be viewed as a sequence of applications of reduction
operators attached to the constraints of the store. Each operator can be applied several times
until the computation reaches a fix-point [15]. This fix-point is the set of final domain states.
An example of computation with reduction operators is shown in Figure 1. There are
three variables x, y and z and two constraints, x > y and y > z. At the beginning, Dx =
Dy = Dz = {1, 2, 3}, represented by three columns of white squares. Considering the first
constraint, it appears that x cannot take the value “1”, because otherwise there would be
4
WLPE’01
x
3
2
1
y
z
x
red
x
x>y
y
z
y
red x > y
x
y
z
y
red y > z
x
y
z
z
red y > z
x
y
z
x
red x > y
x
y
z
Figure 1: Application of reductions to the system {x > y; y > z}.
no value for y such that x > y; redxx>y withdraws this inconsistent value from Dx . This
withdrawal is marked with a black square. In the same way, redyx>y withdraws the value 3
from the domain of y. Then, considering the constraint y > z, the operators redyy>z and
redzy>z withdraw respectively the sets {1} and {2, 3} from Dy and Dz . Finally, a second
application of redxx>y reduces Dx to the singleton {3}. The fix-point is reached. The final
solution is: {x = 3, y = 2, z = 1}.
2.3
Awakening and Solved Conditions
An essential notion of constraint propagation is the propagation queue. This queue contains all
the constraints whose reduction operators have to be applied. At each step of the propagation,
a constraint is selected in the propagation queue, according to a given strategy depending on
implementation (for example a constraint with more variables first). The reduction operators
of the selected constraint are applied. These applications can make new domain reductions.
When a variable domain is updated, the system puts in the queue all the constraints where this
variable appears. When the queue becomes empty, a fix-point is reached and the propagation
ends. Thus, there are three fundamental operations: selection from the queue, reduction of
variable domains and awakening of constraints.
It is actually not necessary to wake a constraint on each update of its variable domains.
For example, it is irrelevant to wake the constraint x > y at each modification of Dx or Dy .
The reduction operators of this constraint are unable to withdraw a value in any domain if
neither the upper bound of Dx nor the lower bound of Dy is updated. Hence, it is sufficient
to wake the constraint only when one of those particular modifications occurs.
Definition 2.2 (Awakening condition). The awakening condition of a constraint C is a
predicate depending upon the modifications of D|var(C) . This condition holds when a new
value withdrawal can be made by the constraint reduction operators. The condition is optimal
when it holds only when a new value withdrawal can be made.
The awakening condition of C is denoted by awake cond(C).
The actual awakening conditions are often a compromise between the cost of their computation and how many awakenings they prevent.
Another condition type permits irrelevant awakenings to be prevented. Let us consider a
constraint C; if the domains of its variables are such that no future value withdrawals can
invalidate the constraint satisfaction, the constraint is said to be solved. In that case, the
reduction operators of this constraint cannot make any new value withdrawal anymore, unless
the system backtracks to a former point. For example, if the two domains Dx and Dy are
such that Dx ∩ Dy = ∅, it is useless to apply the reduction operators of x 6= y. Thus, it is
useless to wake a solved constraint.
Prototyping clp(fd) Tracers
5
Definition 2.3 (Solved condition). The solved condition of a constraint C is a predicate
depending upon the state of D|var(C) . This condition holds only when the domains are such
that C is solved.
The solved condition of C is denoted by solved cond(C).
We can note that a sufficient condition is that all variables in var(C) are ground to some
values satisfying the constraint.
In the rest of the paper, a primitive constraint is defined by those three characteristics:
its reduction operators, awakening condition and solved condition.
2.4
Structure of the Constraint Store
The constraint store S is the set of all constraints taken into account by the computation at
a given moment. When the computation begins, the store is empty. Then, constraints are
individually added or withdrawn according to the control part. In the store, constraints may
have different status and the store can be partitioned into five subsets denoted A, S, Q, T, R.
• A is the set of active constraints. It is either a singleton3 (only one constraint is active)
or empty (no constraint is active). The reduction operators associated to the active
constraint will perform the reductions of variable domains.
• S is the set constraints which are said suspended, namely they are waiting to be “woken”
and put into Q in the case of domain modifications of some of their variable (none has
empty domain).
• Q is the set of constraints in the propagation queue 4 , it contains the constraints waiting
to be activated. In order to reach the fix-point all reduction operators associated to
these constraints must be considered.
• T is the set of solved constraints (true), namely the constraints which hold whatever
future withdrawals will be made.
• R is the set of rejected constraints, i.e. the constraints for which the domain of at least
one variable is empty. In practice it is empty or a singleton, since as soon as there is
one constraint in R the store is considered as “unsatisfiable” and the computations will
continue according to the control.
2.5
Propagation
The evolution of the store can now be described as state transition functions or “events”
acting in the store, in the style of Guerevitch’s evolving algebras [18]. This is illustrated
in Figure 2. When propagation begins, there is an active constraint. The active constraint
applies its reduction operators as far as possible. Each application of a reduction operator
is a reduce event. It narrows the domain of one and only one variable. If a domain becomes
empty, the constraint is rejected (reject event). A rejected constraint is a sign of failure. If
no failure occurs and no other reduction can be made, the constraint is either solved (true) or
suspended (suspend). A solved constraint will not be woken anymore. A suspended constraint
will be woken as soon as its awakening condition holds. On awakening, the constraint is put
in the propagation queue (wake-up event). When there is no active constraint, one is selected
in the queue (select) and becomes active. If the queue is empty, the propagation ends.
3
This restriction could be alleviated to handle multiple active constraints, for example to handle unification
viewed as equality constraints on Herbrand’s domains.
4
The term queue comes from standard usage. It is in fact a set.
6
WLPE’01
SELECT
Q
Queue
WAKE−UP
SUSPEND
S
A
TELL
Suspended
constraints
TOLD
Active
Constraint
Solved
TRUE
T
REJECT
R
constraints
REDUCE
Rejected
constraints
Store
Figure 2: Events related to the structured store
select
∃C ∈ Q ∧ A = ∅ ∧ R = ∅
Q ← Q − {C}, A ← {C}
reject
∃C ∈ A, ∃x ∈ var(C) · (Dx = ∅)
A ← A − {C}, R ← {C}
wake-up
reduce
true
suspend
∃C ∈ S · awake cond(C) ∧ R = ∅
S ← S − {C}, Q ← Q ∪ {C}
∃C ∈ A, ∃x ∈ var(C) · (Wx 6= ∅) ∧ R = ∅
Dx ← Dx − Wx
∃C ∈ A · solved cond(C) ∧ R = ∅
A ← A − {C}, T ← T ∪ {C}
∃C ∈ A ∧ R = ∅
A ← A − {C}, S ← S ∪ {C}
Figure 3: Propagation events.
tell(C)
told
Wx = Dx − redxC (D|var(C) )
Rule format : Name
A=∅
push h{C}, S, Q, T, R, Di
pop
Figure 4: Control events
Conditions
Def initions
Actions
Prototyping clp(fd) Tracers
7
The propagation is completely defined by the rules given in Figure 3. The rules are
applied in the order in which they are given (from top to bottom). Each rule specifies an
event type. An event modifies the system state: hA, S, Q, T, R, Di. An event occurs when
its pre-conditions hold and no higher-priority event is possible. The rule priority prevents
redundant conditions. For example, a suspend event is made only when no true event is
possible. In the same way, if a reduce event wakes some constraints, all wake-up events are
performed before any other reduce. The rule system still contains some indeterminism: the
choice of C in the rules select and wake-up, and the choice of x in the rule reduce depends on
the solver strategy.
The rules are applied until all the constraints are in T , R and S. No further application
is possible. If R is non empty, the store is “rejected”. Otherwise, it may be considered that a
(set of) solution(s) has been obtained. All constraints in T are already solved. Therefore, any
tuple of values of V which satisfies the constraints in S is a solution. The way the computation
will continue depends of the control.
2.6
Control
The evolution of the store described so far assumes that the set of constraints in the store
is invariant, i.e., only their status is modified. The host constraint programming language
provides the way to build the store and to manage it, with possible interleaving of constraint
management and propagation steps.
In order to remain as independent as possible from the host language, we restrict the
control part to two events: the first one adds a constraint C into the store (tell(C) event) and
the second one restores the store and the domains in some previous state (told event).
These events act on the store as illustrated in Figure 2. The tell event puts a new constraint
in the store as the active one.
In clp(fd) the control part is particularly simple and can be described by a (control) stack of
states of the system 5 . The state contains the current store and the domains. The propagation
phase is always performed in the state on top of the control stack. Each tell event includes
a constraint into the store, push the new state on the stack and is followed by a complete
propagation phase. When the fix-point is reached either a new tell(C ′ ) is performed, or a
told. In the later case the stack is popped, thus the former state is restored. As long as the
control stack is not empty, new tell events can be performed on choice points and correspond
to alternative computations.
This semantics is formalized by the two rules of Figure 4. The push and pop operations
work on the control stack. With a different host language, the same told event could be used
but with a possibly different meaning.
3
Trace Definition
An execution is represented by a trace which is a sequence of events. An event corresponds
to an elementary step of the execution. It is a tuple of attributes. Following the notation of
Prolog traces, the types of events are called ports.
Most attributes are common to all events. For some ports, specific attributes are added.
For example, reduce events have two additional attributes: the withdrawn values Wx , for the
5
This corresponds to the description of the visit of a standard search-tree of Prolog [9].
8
WLPE’01
variable x whose domain is reduced, and the types of updates made, such a modification of
the domain upper bound. The attributes are as follows.
Attributes for all events
• chrono: the event number (starting with 1);
• depth: the depth of the execution, starting with 0, it is incremented at each
tell and decremented at each told;
• port: the event type as presented in Sections 2.5 and 2.6: one of reduce,
wake-up, suspend, true, reject, select, tell and told;
• constraint: the concerned constraint, represented by a quadruple:
• a unique identifier generated at its tell;
• an abstract representation, identical to the source formulation of the
program;
• an concrete representation (e.g. diffN(X, Y, N) for X ## Y + N or
X - N ## Y) ;
• the invocation context, namely the Prolog goal from which the tell is
performed.
• domains: the value of the variable domains before the event occurs;
• store: the content of the constraint store represented by the 5 components
described in section 2.4. Each set of constraints is represented by a list of
pairs (constraint identifier, external representation):
• store A the set of active constraints;
• store S the set of suspended constraints;
• store Q the propagation queue;
• store T the set of solved constraints;
• store R the set of rejected constraints.
Specific attributes for reduce events
• withdrawn: The withdrawn domain
• update: The list of updates, an update is of the form (variable -> type),
where type can be one of ground, any, min, max, see Section 4 for further
explanation;
Specific attribute for wake-up events
• cause: The verified part of the awakening condition
The attributes are numerous and contain large chunks of information. Indeed, they aim
at providing useful information to automatic trace analysis programs. The more contents the
better. In a default display for users, only some attributes would be chosen. Furthermore,
and as in Opium [12] the trace analysis will be mainly done on the fly, only the attributes
relevant to a given analysis will be retrieved, and no trace will be stored. Therefore there is
no a priori restriction on the number and size of attributes.
Figure 5 shows the source code of program sorted(L). The program sorts three numbers
between 1 and 3 in a very naive way. Following the convention of many systems, constraints
operators are prefixed by a “#”. The figure also shows a trace of the execution. All events
are listed but only with a few event attributes: the event number and port, the constraint
concerned by the event and its variable domains. At reduce events, the variable whose domain
is being reduced as well as the withdrawn values are added.
The first two constraints are entered (tell) and suspended without any reduction (events
#1 to #4). The tell of the third one, Y #> Z gives two value withdrawals, ‘1’ from Dy (#6)
Prototyping clp(fd) Tracers
9
sorted([X, Y, Z]):[X, Y, Z] :: 1..3,
% At the beginning, Dx = Dy = Dz = [1..3]
X ## Y, X #>= Y, Y #> Z, % 3 constraints : x 6= y, x ≥ y and y > z
labelling([X, Y, Z]).
% labelling phase, with a “first fail” strategy
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
[1]
[1]
[2]
[2]
[3]
[3]
[3]
[3]
[3]
[3]
[3]
[3]
[4]
[4]
[4]
[4]
[4]
[4]
[4]
[4]
Tell
Suspend
Tell
Suspend
Tell
Reduce
Wake-up
Reduce
Suspend
Select
Reduce
Suspend
Tell
Reduce
Wake-up
Wake-up
True
Select
Reduce
Wake-up
X##Y
X##Y
X#>=Y
X#>=Y
Y#>Z
Y#>Z
X#>=Y
Y#>Z
Y#>Z
X#>=Y
X#>=Y
X#>=Y
X#=2
X#=2
X#>=Y
X##Y
X#=2
X#>=Y
X#>=Y
Y#>Z
X:[1,2,3]
X:[1,2,3]
X:[1,2,3]
X:[1,2,3]
Y:[1,2,3]
Y:[1,2,3]
X:[1,2,3]
Y:[2,3]
Y:[2,3]
X:[1,2,3]
X:[1,2,3]
X:[2,3]
X:[2,3]
X:[2,3]
X:[2]
X:[2]
X:[2]
X:[2]
X:[2]
Y:[2]
Y:[1,2,3]
Y:[1,2,3]
Y:[1,2,3]
Y:[1,2,3]
Z:[1,2,3]
Z:[1,2,3] Y[1]
Y:[2,3]
Z:[1,2,3] Z[3]
Z:[1,2]
Y:[2,3]
Y:[2,3]
X[1]
Y:[2,3]
X[3]
Y:[2,3]
Y:[2,3]
Y:[2,3]
Y:[2,3]
Z:[1,2]
Y[3]
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
[4]
[4]
[4]
[4]
[4]
[4]
[4]
[4]
[4]
[4]
[4]
[4]
[4]
[4]
[4]
[4]
[4]
[3]
[2]
[1]
True
Select
Reduce
Reject
Told
Tell
Reduce
Wake-up
True
Select
Reduce
Wake-up
True
Select
Reduce
True
Told
Told
Told
Told
X#>=Y
X##Y
X##Y
X##Y
X#=2
X#=3
X#=3
X##Y
X#=3
X##Y
X##Y
Y#>Z
X##Y
Y#>Z
Y#>Z
Y#>Z
X#=3
Y#>Z
X#>=Y
X##Y
X:[2]
X:[2]
X:[2]
X:[]
X:[]
X:[2,3]
X:[2,3]
X:[3]
X:[3]
X:[3]
X:[3]
Y:[2]
X:[3]
Y:[2]
Y:[2]
Y:[2]
X:[3]
Y:[2,3]
X:[1,2,3]
X:[1,2,3]
Y:[2]
Y:[2]
Y:[2]
Y:[2]
X[2]
X[2]
Y:[2,3]
Y:[2,3]
Y:[2,3] Y[3]
Z:[1,2]
Y:[2]
Z:[1,2]
Z:[1,2] Z[2]
Z:[1]
Z:[1,2]
Y:[1,2,3]
Y:[1,2,3]
chrono
depth
port
constraint
domains
withdrawn
update
store_A
store_Q
=
=
=
=
=
=
=
=
=
14
4
REDUCE
(4, X#=2, assign(var(1, X), 2), labelling([X, Y, Z]))
[X::[2..3], Y::[2..3], Z::[1..2]]
X::[3]
[X->any, X->ground, X->max]
[(4, X#=2)]
store_S = [(2, X#>=Y), (3, Y#>Z), (1, X##Y)]
[]
store_T = []
store_R = []
chrono
depth
port
constraint
domains
cause
store_A
store_Q
=
=
=
=
=
=
=
=
16
4
WAKE-UP
(1, X##Y, diff(var(1, X), var(2, Y)), sorted([X, Y, Z]))
[X::[2], Y::[2..3], Z::[1..2]]
[X->ground]
[(4, X#=2)]
store_S
= [(3, Y#>Z), (1, X##Y)]
[(2, X#>=Y)]
store_T = []
store_R = []
Figure 5: A trace of the execution of program sorted([X, Y, Z]), all events are present
with attributes (event number, [depth], constraint C, D|var(C) ). Events #14 and #16 are
displayed with all their attributes.
10
WLPE’01
and ‘3’ from Dz (#8). The first reduction modifies the lower bound of Dy and so wakes the
suspended constraint X #>= Y (#7). After those two reductions the constraint is suspended
and the waiting one is selected (#10). At event #12, the domains are Dx = {2, 3}, Dy = {2, 3}
and Dz = {1, 2}. Then the labelling phase begins. With our simple “first fail” strategy, the
first added constraint is X #= 2. X is ground and equal to 2 and this constraint is solved (#17).
Two other constraints are solved during the propagation, but it leads to Dy = {3}, Dz = {1, 2}
and an empty domain for X (#25). Another labelling constraint is tried (#26), X #= 3 and
leads to the unique solution {X:3, Y:2, Z:1}.
4
Deriving a Tracer from the Operational Semantics
In order to experimentally validate the trace defined in Section 3, we derive, from the operational semantics of Section 2, a clp(fd) interpreter in Prolog that we instrument with trace
hooks. The resulting interpreter, which produces traces, is not meant to be an efficient clp(fd)
system, but to be faithful to the semantics of Section 2. The faithfulness comes from the fact
that the translation of the semantic rules into executable Prolog code is syntactical.
In Section 2, we left the primitive constraints undefined. We therefore first propose a
definition for 8 primitive constraints by specifying, for each, its reduction operators, its solved
condition, and its awakening condition. Note that the definitions we propose in Table 1 define
reduction operators that perform a full-arc consistency.
Then, we show how to translate the primitive constraints and the semantic rules into
Prolog. We also show how to interface this Prolog code with the Prolog underlying system.
4.1
Primitive Constraint Definitions
In order to define a primitive constraint, we need to define its reduction operators, its solved
condition, and its awakening condition (see Section 2). We define in Table 1 the 8 primitive
constraints x = y, x 6= y, x = y + n, x 6= y + n, x > y, x ≥ y, x = n, and x 6= n, where x and
y represent two finite domain variables, and n represents an integer constant. The reduction
operator redxC is defined by the set of values Wx it withdraws from the domain of variable x.
The equality constraint between two variables x and y (x = y) withdraws from the domains
of Dx and Dy the values which are not contained in both domains. The constraint is solved
only when the two variables are ground and have the same value v. At each modification of
Dx or Dy , the reduction operators may withdraw new values. The constraints must therefore
be woken at each of their modification.
The difference constraint between two variables x and y (x 6= y) can only reduce the
domains when one of them is ground. However, it is solved as soon as the two domains are
disjoint.
4.2
Data Structures
A constraint variable is represented by a term containing a unique integer, and a string
(its name in the source). A constraint instance is represented by a quadruple containing a
unique constraint number, a string (the constraint as displayed in the source, i.e., its abstract
representation), an internal (concrete) form, a list of constraint variables, and an invocation
context. The invocation context of a constraint is the Prolog goal from which it was invoked.
Prototyping clp(fd) Tracers
Constraint
C
x=y
Wx
Wy
Wx
x 6= y
Wy
x=y+n
x 6= y + n
x>y
x≥y
x=n
x 6= n
Reduction operators
(redxC (D|var(c) ) = Dx − Wx )
= Dx − (Dx ∩ Dy )
=D
y − (Dx ∩ Dy )
{v} ∩ Dx if Dy = {v}
=
otherwise
∅
{v} ∩ Dy si Dx = {v}
=
∅
otherwise
Wx = Dx − (Dx ∩ {v + n, v ∈ Dy })
Wy = Dy − (Dx ∩ {v + n, v ∈ Dy })
{v + n} ∩ Dx if Dy = {v}
Wx =
otherwise
∅
{v − n} ∩ Dy if Dx = {v}
Wy =
∅
otherwise
Wx = {v ∈ Dx , v ≤ miny }
Wy = {v ∈ Dy , v ≥ maxx }
Wx = {v ∈ Dx , v < miny }
Wy = {v ∈ Dy , v > maxx }
Wx = D
x − {n}
{n} if n ∈ Dx
Wx =
∅
otherwise
11
Solved condition
solved cond(C)
Awakening cond.
awake cond(C)
Dx = Dy = {v}
xany ∨ yany
Dx ∩ Dy = ∅
xground ∨ yground
Dx = {vx } ∧
Dy = {vy } ∧
vx = vy + n
xany ∨ yany
∀v ∈ Dy , v+n 6∈ Dx
xground ∨ yground
minx > maxy
xmax ∨ ymin
minx ≥ maxy
xmax ∨ ymin
Dx = {n}
n∈
/ Dx
Table 1: Characteristics of the primitive constraints implemented in the interpreter
The solver state is represented by a sextuplet: (A, S, Q, T, R, D), where A, S, Q, T and R
define the store as described in Section 2; they are lists of constraints. D is a list of domains.
In order to represent awakening conditions and domain narrowing, we define five types of
domain modifications (following what is done in Eclipse [14]). Each type refers to a particular
constraint variable x.
• xmin refers to a modification of the Dx lower bound, e.g., {1, 2, 4} → {2, 4};
• xmax refers to a modification of the Dx upper bound, e.g., {1, 2, 4} → {1, 2};
• xany refers to any modification of Dx , e.g., {1, 2, 4} → {1, 4};
• xground refers to a grounding of x (Dx becomes a singleton), e.g., {1, 2, 4} → {1};
• xempty refers to an emptying of Dx , e.g., {1, 2, 4} → ∅
Those five modification types respectively appear in the code as X->min, X->max, X->any,
X->ground, and X->empty. Awakening conditions are disjunctions of such modification types;
such disjunctions are encoded by lists.
4.3
Translation of the Primitive Constraints
The reduction operators, the solved condition, and the awakening condition defining a primitive constraint are encoded by the following predicates:
• cd reduction(+C, +D, +X, -Wx)6 : takes as input a constraint C, a domain state D,
and a constraint variable x; it succeeds iff the application of redxC (D|var(C) ) withdraws
a non-empty set (bound to Wx). There is one clause per reduction operator of C;
6
As specified in the standard Prolog [9], + denotes input arguments and - denotes outputs arguments.
12
1
2
3
4
5
6
7
8
WLPE’01
cd_reduction(diff(X,Y),D,Y,[Vx]) :is_ground(X,D,Vx),
get_domain(Y,D,Dy),
member(Vx, Dy).
cd_reduction(diff(X,Y),D,X,[Vy]) :is_ground(Y,D,Vy),
get_domain(X,D,Dx),
member(Vy,Dx).
9
10
11
12
13
14
15
cd_awake(diff(X,Y),Cond) :Cond = [X->ground,Y->ground]).
cd_solved(diff(X,Y),D) :get_domain(X,D,Dx),
get_domain(Y,D,Dy),
d_intersection(Dx,Dy,[]).
Figure 6: Definition of the primitive constraint x 6= y: its two reduction operators, its solved
condition, and its awakening condition, as specified in Table 1.
• cd solved(+C, +D): takes as input a constraint C and a domain state D; it succeeds
iff the constraint C is solved in the domain state D;
• cd awake(+C, -Cond): takes as input a constraint C and outputs (in Cond) the list of
awakening conditions of C.
Figure 6 shows the implementation of the primitive constraint x 6= y which is simply a
Prolog encoding of the first entry of Table 1. The functor diff/2 is the internal encoding of
6=. Predicate is ground(+X, +D, -Vx) takes a constraint variable x and a domain D, and
succeeds iff Dx is a singleton (bound to Vx); get domain(+X, +D, -Dx) takes a constraint
variable x and a domain state D, and outputs the domain of x (Dx ); d intersection(+D1,
+D2, -D) computes the intersection of two domains.
4.4
Translation of the Semantic Rules
Figure 7 contains the translation of the semantic rules of Figures 3 and 4. Each rule is encoded
by a predicate with the same name as the rule. The translation is merely syntactical, except
for the tell and told rules, for which it is unnecessary to save and restore the solver states
(push and pop) since this work is done by the Prolog backtracking mechanism.
Before paraphrasing the code for one rule, we give the meaning of all the (simple) predicates that are not defined elsewhere in the paper: choose in queue(+Q0, -C) takes as input
a queue Q and outputs one of the queue constraints ; it succeeds iff Q is not empty. The choice
of the constraint depends of the solver strategy; subtract(+L1, +L2, -L) computes the difference between two lists; get varC(+C, -V) takes as input a constraint C and outputs a
list of the constraint variables that appear in C; trace(+Port, +C, +St0, +O1, +O2, +O3)
takes as input the different event attributes as described in Section 3; it calls the trace analysis
system which can, for example, print a trace line; put end of queue(+C, +Q0, -Q) puts a
constraint at the end of a queue; update domain(+X, +Wx, +D0, -D, -Mod) takes as input
a constraint variable x, a value set Wx (to withdraw), and a domain state D 0 ; it outputs the
state domain D such that Dx = Dx0 − Wx , and the list of modification types xmod1 , ..., xmodn
(where modi ∈ {min, max, ground, any, empty}) that characterizes the Wx value removals;
internal(+C, -Ci) takes as input a constraint and outputs its internal representation.
All the predicates translating rules take as input a solver state St0 and output a new
solver state St. St0 and St respectively denote the state of the solver before and after the
application of a rule. The only exceptions are predicates wake up(+St0, -St, +ModIn) and
reduce(+St0, -St, -ModOut) that respectively inputs and outputs an additional argument:
Prototyping clp(fd) Tracers
13
a list of modification types (xmod1 , ..., xmodn ). This list is computed in reduce/3 and used
in wake up/3 to check the awakening condition.
Predicate select(+St0, -St) translates the select rule. That rule needs to fulfill 3
conditions to be allowed to be applied: ∃C ∈ Q, that is checked line 17 by choose in queue/2
(which fails iff the queue is empty); A = ∅ and R = ∅ that are checked line 15. The 2 actions
to perform when the conditions of the rule hold are Q ← Q − {C}, which is done line 18 by
subtract/3, and A = {C} which is done line 16. The other rules are translated in the same
way.
4.5
Integration with the Underlying Prolog System
The integration of our instrumented constraint solver with the underlying Prolog system is
done by the predicate call constraint/3 which is given in Figure 8. After propagation and
success, it returns the new state of the constraint part (St). If the propagation leads to a
failure, the goal fails.
Predicate prop step/4 performs a propagation step, i.e., it applies one of the 6 propagation rules of Figure 3; it fails if no rule can be applied. The choice of the rule to apply
is done according to the strategy discussed in Section 2.5. Predicate propagation/3 calls
prop step/4 in loop until either a propagation step fails (the fix-point is reached) or the
solver rejects a constraint (the constraint goal is unsatisfiable).
5
Experimentation
We have made some preliminary experiments of trace analysis with our tracer.
A first experiment consists in displaying the labelling (or search-) tree. The analyzed
program solves the well known n-queens problem. Figure 9 shows a search-tree obtained for
the 4-queens problem by trace analysis of a trace generated by our meta-interpreter, and
displayed with dot [22]. Each variable corresponds to a row of the chess-board. The top node
of the displayed tree contains the initial domains, the other nodes contain the reduced domains
if any. The tree shows four failures in square boxes, two solutions and three choice points.
The uppest arc sums up the tell operations for all permanent constraints of the programs.
Each other arc represents a labelling constraint addition.
The trace analysis uses only the tell, told, reduce events. Few attributes were needed: the
port, the concerned constraint and the domains of constraint variables. For reduce events,
the updated variable and the withdrawn set were also needed.
The sequence of tell and told events in the trace corresponds to a depth-first left-toright visit of the search-tree where tell and told events respectively correspond to downward
and upward moves in the search-tree. Reconstructing a tree from its visit (for a fixed visit
strategy) is easy. In Figure 9, we took advantage of reduce events to label the nodes with the
propagation results.
A second experiment is the generation of a 3D variable update view [29]. The evolution
of the domains of the variables during the computation is displayed in three dimensions. It
gives a tool à la TRIFID [5] (here however colors are introduced to display specific events as
in the variable update view of [28]). The trace analyzer makes a VRML file by computing
domain size on each tell and reject event, and when a solution is found. The details of reduce
14
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
WLPE’01
tell(C, St0, St)
St0 = ( [], S,
St = ([C], S,
trace(tell, C,
:Q, T, R, D),
Q, T, R, D),
St0, -, -, -).
told(C, St0) :trace(told, C, St0, -, -, -).
select(St0, St) :St0 = ( [], S, Q0, T, [], D),
St = ([C], S, Q, T, [], D),
choose_in_queue(Q0, C),
subtract(Q0, [C], Q),
trace(select, C, St0, -, -, -).
reject(St0, St) :St0 = ([C], S, Q, T, [], D),
St = ( [], S, Q, T, [C], D),
get_varC(C, VarC),
member(X, VarC),
get_domain(X, D, []),
trace(reject, C, St0, -, -, -).
wake_up(St0, St, ModIn) :St0 = (A, S0, Q0, T, [], D),
St = (A, S,
Q, T, [], D),
member(C, S0),
awake_cond(C, ModIn, True),
subtract(S0, [C], S),
put_end_of_queue(C, Q0, Q),
trace(wake_up, C, St0, True, -, -).
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
reduce(St0, St, ModOut) :St0 = ([C], S, Q, T, [],
St = ([C], S, Q, T, [],
get_varC(C, VarC),
member(X, VarC),
reduction(C, D0, X, Wx),
update_domain(X, Wx, D0,
trace(reduce, C, St0, X,
true(St0, St) :St0 = ([C], S,
St = ( [], S,
solved_cond(C,
T = [C|T0],
trace(true, C,
D0),
D),
D, ModOut),
Wx, ModOut).
Q, T0, [], D),
Q, T, [], D),
D),
St0, -, -, -).
suspend(St0, St) :St0 = ([C], S0, Q, T, [], D),
St = ( [], S, Q, T, [], D),
S = [C|S0],
trace(suspend, C, St0, -, -, -).
reduction(C, D, X, Wx) :internal(C, Ci),
cd_reduction(Ci, D, X, Wx).
awake_cond(C, ModIn, True) :internal(C, Ci),
cd_awake(Ci, Cond),
intersection(Cond, ModIn, True),
not True = [].
solved_cond(C, D) :internal(C, Ci),
cd_solved(Ci, D).
Figure 7: Prolog translation of the semantic rules of Figures 3 and 4
Prototyping clp(fd) Tracers
1
2
3
4
5
6
7
8
9
10
11
12
15
13
14
15
16
17
18
19
20
21
call_constraint(C, St0, St) :tell(C, St0, St1),
propagation(none, St1, St),
( true ; told(C, St), fail ),
St = (_, _, _, _, [], _).
propagation(Mod0, St0, St) :prop_step(St0, St1, Mod0, Mod)
-> ( St1 = (_, _, _, _, [], _)
-> propagation(Mod, St1, St)
; St = St1 )
; St = St0.
prop_step(St0, St, Mod0, Mod)
select(St0, St) -> Mod =
;
reject(St0, St) -> Mod =
;
wake_up(St0, St, Mod0)
-> Mod =
;
reduce(St0, St, Mod)
-> true
;
true(St0, St)
-> Mod =
;
suspend(St0, St) -> Mod =
:none
none
Mod0
none
none.
Figure 8: Integrating the constraint solver with the underlying Prolog system
A::[1..4]
B::[1..4]
C::[1..4]
D::[1..4]
A##B A##C A##D
B##C B##D C##D
A+1##B
A-1##B
A+2##C
A-2##C
A+3##D
A-3##D
B+1##C B-1##C B+2##D
B-2##D C+1##D C-1##D
A#=1
A -> 1
B -> [3..4]
C -> [2, 4]
D -> [2..3]
B#=3
B#=4
D -> 2
B -> 4
C -> 4
D -> 3
B -> []
C -> []
A -> 2
B -> 4
C -> 1
D -> 3
A#=2
A#=3
A -> 3
B -> 1
C -> 4
D -> 2
A#=4
A -> 4
B -> [1..2]
C -> [1, 3]
D -> [2..3]
B#=1
B#=2
B -> 1
B -> 2
C -> 3
D -> 3
D -> []
C -> []
Figure 9: A search-tree of a 4-queens constraint program execution, obtained by trace analysis,
and displayed by dot.
16
WLPE’01
Figure 10: Comparing two search procedures for the 40-queens problem with VRML views
computed by trace analysis.
events allow us to assign color to each kind of domain update (for example minimum or
maximum value removed or domain emptied) as made by Simonis and Aggoun in the Cosytec
Search-Tree Visualizer [28]. The trace analysis is implemented in about 125 lines of Prolog
and generates an intermediate file. A program implemented in 240 lines of C converts this
file into VRML format.
Figure 10 shows the resolution of the 40-queens problem with two different labelling strategies. We have three axes: variables (horizontal axis of the vertical “wall”), domain size (vertical axis) and time. The first strategy is a first-fail selection of the labelled variable and the
first value tried is the minimum of its domain. The second strategy is also a first-fail strategy
but variable list is sorted with the middle variable first and the middle of domain is preferred
to its minimum. This strategy derives from one described by Simonis and Aggoun [28]. This
approach allows to compare the efficiency of these two strategies by manipulating the 3Dmodel. With the first strategy, we get a quick decreasing of the domain size on one side of
the chess-board and a long oscillation of the domain size on the other side. With the second
strategy, the decreasing of domain size is more regular and more symmetrical, the solution
is found faster. In fact, the second strategy, which consists in putting the queens from the
center of the chess-board, benefits more from the symmetrical nature of the problem. The
possibility of moving manually the figure facilitates observation of such property.
6
Discussion and Conclusion
This article is a first attempt to define new ports for tracing finite domain solvers. We do not
claim that the trace model presented here is the ultimate one. On the contrary we propose
a methodology to experiment and to improve it. This methodology is based on the following
steps: definition of an executable formal model of trace, extraction of relevant informations
by a trace analyzer, utilization of the extracted informations in several debugging tools. Each
debugging tool extracts from the same generic trace the information it needs.
The formal approach of trace modeling used here allows to clearly define the ports. Then
the implementation of the model by a meta-interpreter almost written in ISO-Prolog, a logic
programming language with a clear semantics, allows to preserve its correctness wrt the formal
model. This methodology is efficient enough on small examples and therefore is of practical
interest. However to handle large realistic examples will require hard-coded implementation
Prototyping clp(fd) Tracers
17
of the trace model.
Different aspects must be examined to estimate the results. First the formal model itself.
Two ports are directly related to logic programming (tell and told) and correspond to the
well-known ports call and exit of [3], the others correspond to a small number of different
steps of computation of the reduction operator fix-point. On the other side we defined a great
number of (possibly large sized) attributes to ensure that each event carries enough potentially
useful informations. This immediately shows the limits of our model: it is probably general
enough to take into account several finite domain solvers, but tracing the complete behavior
of different solvers will require new or different ports to take into account different kinds
of control, specific steps of computation (e.g. constraint posting, labelling phase, . . . ), or
different algorithms.
The proposed methodology shows the way to progress: defining the trace with a formal
model makes easier to compare different trace models. It is thus easier to see which are
the missing ports or attributes. Some solvers use a propagation queue with events instead of
constraints (in this case new ports or same ports but with different attributes are necessary), or
do not use backtracking (then told is never used). The question is also to find the right balance
between the number of events and the attributes, in such a way that a hard-coded efficient
implementation of (a part of) the trace model is still possible. For example the attribute
withdrawn of the port reduce concerns one variable only. Therefore if several variables are
involved by a single reduction step, there will be several reduce events. Another possibility
would be to have a unique event with a port whose withdrawn attribute includes several
variables at once. Gathering too many informations in one event may slow down the tracer
considerably. The trace production must be as fast as possible in order to keep the best
performances of the solver.
The ports and attributes presented here seem to be a good basis to start the study of
a generic trace for clp(fd). In [8] a sample of additional ports are suggested to cope with
different finite domain solvers.
The second aspect concerns the extraction of tool specific relevant informations. The
generic trace, by definition, contains too many informations if used by a unique debugging
tool only. It is also not intended to be stored in a huge file, but it should be filtered on the
fly and re-formatted for use by some given tool. The methodology we proposed here, à la
opium, is well-known and has been shown to be efficient and general enough to be used in
practice also in hard-coded implementations. It allows to specify the trace analysis in a high
level language (here Prolog) in a way which is independent from the trace production.
For each experimented tool we have presented here there is a specific analyzer which needs
only few lines of code. With this approach, building different views of the same execution
requires only to modify the trace analyzer. Notice that the implementation presented here
assumes that there is only one trace analyzer running in parallel with the solver. The question
of how different analyzers, working simultaneously, could be combined is not considered here.
Finally the third aspect concerns the experimentation. The main challenge in constraint
debugging is performance debugging. Our objective is to facilitate the development of constraint resolution analysis tools in a manner which is as independent as possible from the
solvers platforms. The three steps method (generic tracer/trace analyzer/debugging tool)
is a way to approach such a goal. We experimented it by building several analyzers and
(limited) tools very easily, without having to change the trace format. Of course much more
experiences are still needed.
Another way to estimate the results of this article is to consider some of the existing
18
WLPE’01
debugging tools and to observe that most of the informations relevant for each of these tools
are already present in the ports and their attributes. Here are briefly reviewed some platform independent tools7 , namely debugging tools using graphical interfaces where labelling,
constraints and propagation can be visualized.
The Search-Tree Visualization tool for CHIP, described by Simonis and Aggoun [28] displays search-trees, variables and domain evolution. The whole of this information is present in
the proposed ports. Our experimentation shows that search-tree and constraints can easily be
displayed. The relevant information is also present for the Oz Explorer, a visual programming
tool described by Schulte [27], also centered on the search-tree visualization with user-defined
displays for nodes. It is also the case in the search-tree abstractor described by Aillaud and
Deransart [1] where search-tree are displayed with constraints at the nodes.
In Grace, a constraint program debugger designed by Meier [25], users can get information on domain updates and constraint awakenings. This is available in our trace via our 8
ports. Grace also provides the ability to evaluate expressions using the current domain of the
variables. As the domain of the variables is an attribute of all events, it can be obtained by
any analyzer of our trace. The CHIP tool also provide update views in which, for example,
useless awakenings are visible. In our environment useless awakenings could be detected by a
select not followed by a reduce. The visualization tool by Carro and Hermenegildo [5] traces
the constraint propagation in 3 dimensions according to time, variable, and cardinal of the
domain. The required information is present in our trace. The S-boxes of Goualard and Benhamou [17] structure the propagation in and the display of the constraint store. Structuring
the propagation requires to modify the control in the solver. Displaying the store according
to the clausal structure would however cause no problem.
An important aspect, not presented in this paper, concerns the possibility of interactive
tools. This means that some debugging tools may request to re-execute some already traced
execution. Special mechanisms must be involved in order to re-execute efficiently and with
the same semantics some part of the execution. How such capability may influences the form
of the generic trace is still an open question and will be further investigated.
References
[1] C. Aillaud and P. Deransart. Towards a language for clp choice-tree visualisation. In
Deransart et al. [10], chapter 8.
[2] F. Benhamou. Interval constraint logic programming. In A. Podelski, editor, Constraint
Programming: Basics and Trends, pages 1–21. Springer-Verlag, Lecture Notes in Computer Science 910, 1994.
[3] L. Byrd. Understanding the control flow of Prolog programs. In S.-A. Tärnlund, editor,
Logic Programming Workshop, Debrecen, Hungary, 1980.
[4] M. Carro and M. Hermenegildo. Tools for search-tree visualisation: The apt tool. In
Deransart et al. [10], chapter 9.
[5] M. Carro and M. Hermenegildo. The vifid/trifid tool. In Deransart et al. [10], chapter 10.
7
They are considered as (clp(fd)) platform independent tools in the sense that they are focused on general
properties of finite domain solving: choice-tree, labelling, variables domain evolution.
Prototyping clp(fd) Tracers
19
[6] Y. Caseau, F.-X. Josset, and F. Laburthe. Claire: combining sets, search and rules to
better express algorithms. In D. De Schreye, editor, Proc. of the 15th Int. Conference
on Logic Programming, pages 245–259. MIT Press, 1999.
[7] Cosytec. CHIP++ Version 5.2. documentation volume 6. http://www.cosytec.com, 1998.
[8] Romuald Debruyune, Jean-Daniel Fekete, Narendra Jussien, Mohammad Ghoniem,
Pierre Deransart, Ludovic Langevine, and e al. A proposal of concrete format for fracing
constraint programming (based on a dtd xml), October 2001. Deliverable D2.2.2.1 (in
French).
[9] P. Deransart, A. Ed-Dbali, and L. Cervoni. Prolog, The Standard; Reference Manual.
Springer Verlag, April 1996.
[10] P. Deransart, M. Hermenegildo, and J. Maluszyński, editors. Analysis and Visualisation
Tools for Constraint Programming. Number 1870 in LNCS. Springer Verlag, 2000.
[11] M. Ducassé. Abstract views of Prolog executions with Opium. In P. Brna, B. du Boulay,
and H. Pain, editors, Learning to Build and Comprehend Complex Information Structures: Prolog as a Case Study, Cognitive Science and Technology, chapter 10, pages
223–243. Ablex, 1999.
[12] M. Ducassé. Opium: An extendable trace analyser for Prolog. The Journal of Logic
programming, 39:177–223, 1999. Special issue on Synthesis, Transformation and Analysis
of Logic Programs, A. Bossi and Y. Deville (eds).
[13] M. Ducassé and J. Noyé. Logic programming environments: Dynamic program analysis
and debugging. The Journal of Logic Programming, 19/20:351–384, May/July 1994.
[14] ECLiPSe.
Constraint
icparc.doc.ic.ac.uk/eclipse/.
logic
programming
system.
http://www-
[15] G. Ferrand, W. Lesaint, and A. Tessier. Value withdrawal explanation in CSP. In
M. Ducassé, editor, AADEBUG’00 (Fourth International Workshop on Automated
Debugging), pages 188–201, 2000. The COmputer Research Repository (CORR)
cs.SE/0012005.
[16] GNU-Prolog. A clp(fd) system based on Standard Prolog (ISO) developed by D. Diaz.
http://gprolog.sourceforge.net/ Distributed under the GNU license.
[17] F. Goualard and F. Benhamou. Debugging Constraint Programs by Store Inspection. In
Deransart et al. [10], chapter 11.
[18] Y. Gurevitch. Evolving algebras, a tutorial introduction. Bulletin of the European Association for Theoretical Computer Science, 43:264–284, 1991.
[19] E. Jahier. Collecting graphical abstract views of Mercury program executions. In
M. Ducassé, editor, Proceedings of the International Workshop on Automated Debugging (AADEBUG2000), Munich, August 2000. The COmputer Research Repository
(CORR) cs.SE/0010038.
20
WLPE’01
[20] E. Jahier, M. Ducassé, and O. Ridoux. Specifying Prolog trace models with a continuation semantics. In K.-K. Lau, editor, Proc. of LOgic-based Program Synthesis and
TRansformation, London, July 2000. Springer-Verlag, Lecture Notes in Computer Science 2042.
[21] N. Jussien and V. Barichard. The PaLM system: explanation-based constraint programming. In Proceedings of TRICS: Techniques foR Implementing Constraint programming
Systems, a post-conference workshop of CP 2000, pages 118–133, Singapore, September
2000.
[22] E. Koutsofios and S. North. Drawing graphs with dot. TR 910904-59113-08TM, AT&T
Bell Laboratories, 1991.
[23] Ludovic Langevine, Pierre Deransart, Mireille Ducasse, and Ewan Jahier. Tracing and analyzing execution of clp(fd) programs: a trace model and an experimental validation environment.
Rr, INRIA, Novembre 2001.
Full version of
this paper, Public Deliverable D1.2.1, to appear as INRIA/IRISA Research report.
http://contraintes.inria.fr/OADymPPaC/.
[24] K. Marriott and P.J. Stuckey. Programming with Constraints: An Introduction. The
MIT Press, Cambridge, Massachussets, 1998.
[25] M. Meier. Debugging constraint programs. In U. Montanari and F. Rossi, editors, Proceedings of the First International Conference on Principles and Practice of Constraint
Programming, number 976 in LNCS, pages 204–221. Springer Verlag, 1995.
[26] INRIA-Rocquencourt,
EMN-Nantes,
INSA-Rennes,
University of Orléans,
Cosytec, and ILOG.
Tools for dynamic analysis and development of constraint programs (OADymPPaC), November 2000.
An RNTL French Project.
http://contraintes.inria.fr/OADymPPaC.
[27] C. Schulte. Oz Explorer: A Visual Constraint Programming Tool. In Proceedings of the
Fourteenth International Conference on Logic Programming (Iclp’97), pages 286–300,
Leuven, Belgium, June 1997. The MIT Press.
[28] H. Simonis and A. Aggoun. Search-Tree Visualisation. In Deransart et al. [10], chapter 7.
[29] R. Zoumman. Analyse du comportement du solveur de contraintes basée sur la visualisation de traces (in French). Rapport de stage, INRIA, 2001.
| 6cs.PL
|
EasyTime++: A case study of incremental domain-specific
language development
Iztok Fister Jr.,∗ Tomaž Kosar,† Iztok Fister,‡ and Marjan Mernik§
Abstract
EasyTime is a domain-specific language (DSL) for measuring time during sports competitions.
arXiv:1303.0722v1 [cs.PL] 4 Mar 2013
A distinguishing feature of DSLs is that they are much more amenable to change, and EasyTime
is no exception in this regard. This paper introduces two new EasyTime features: classifications
of competitors into categories, and the inclusion of competitions where the number of laps must
be dynamically determined. It shows how such extensions can be incrementally added into the
base-language reusing most of the language specifications. Two case studies are presented showing
the suitability of this approach.
To cite paper as follows: I. Jr. Fister, T. Kosar, I. Fister, M. Mernik, EasyTime++: A case
study of incremental domain-specific language development. Information technology and control,
42(1), 77–85, 2013.
∗
University of Maribor, Faculty of electrical engineering and computer science Smetanova 17, 2000 Maribor;
Electronic address: [email protected]
†
University of Maribor, Faculty of electrical engineering and computer science Smetanova 17, 2000 Maribor;
Electronic address: [email protected]
‡
University of Maribor, Faculty of electrical engineering and computer science Smetanova 17, 2000 Maribor;
Electronic address: [email protected]
§
University of Maribor, Faculty of electrical engineering and computer science Smetanova 17, 2000 Maribor;
Electronic address: [email protected]
1
I.
INTRODUCTION
Domain-specific languages (DSLs) are languages tailored to specific application domain
[1–4]. They offer substantial gains regarding expressiveness and ease of use compared with
general-purpose languages (GPLs) within their domain of application [5–7]. However, DSLs
are more amenable to changes [1, 8] since stakeholders’ requirements frequently change. In
order to design and implement DSLs more easily, we need to develop fully modular, extensible, and reusable language descriptions, whilst some of the descriptions could even be
inferred from DSL programs [9, 10]. The language designer wants to include new language
features incrementally as the programming language evolves. Ideally, a language designer
would like to build a language simply by reusing different language definition modules (language components), such as modules for expressions, declarations, etc., as well as to extend
previous language specifications. In the case of general software development the use of
object-oriented techniques and concepts like encapsulation and inheritance, greatly improves
incremental software development, whilst reusability is even further enhanced using aspectoriented techniques [11]. The object-oriented, as well as the aspect-oriented techniques and
concepts, have also been integrated into programming language specifications [12, 13] making new features more easily implemented. One of such tools, where object-oriented and
aspect-oriented concepts have been incorporated, is the LISA tool [8, 14]. This paper shows
how LISA is used within the incremental development of EasyTime DSL, which has been
developed recently for measuring time at different sports competitions (e.g., triathlon, cycling) [15, 16]. EasyTime DSL has already proved to be successful when used at real sport
events (e.g., World Championship in the double ultra triathlon in 2009, National (Slovenian) Championship in the time-trials for cycles in 2010), so the requirements are changing
quickly. Recent extensions to EasyTime have included the possibility of classifying competitors into different categories, where the number of laps is different for each category, and the
inclusion of competitions where the number of laps can be dynamically determined during
a competition (e.g., biathlon, where the number of extra laps depends on missed shots).
The objective of this paper is to introduce EasyTime++ DSL, which supports these new
extensions, as well as to show how such an extension can be incrementally developed using
the introduced LISA tool.
The structure of this paper is as follows: in Section II an overview of the language
2
composition is presented. Section III briefly introduces EasyTime DSL, whilst the core of this
paper is Section IV, which describes how the extensions in EasyTime++ have been specified
and implemented. Some examples are presented in Section V. The paper is concluded with
Section VI, where a brief overview and word about future work is described.
II.
RELATED WORK
Several kinds of language composition have been identified in the literature [8, 17–24]. In
their recent paper [17], Erdweg et al., point out that language composition has obtained little attention, that it is still insufficiently understood, and that the terminology is confusing
thus indicating that the research is inadequate, as yet. Erdweg et al. identified the language
composeability not as a property of languages themselves, but as a property of language definition (e.g., how language specifications can be composed together). The following types of
language composition have been distinguished in [17]: language extension (which subsumes
also language restriction), language unification, self-extension, and extension composition.
In language extension the specifications of base language B are extended with a new
language specification fragment E, which typically makes little sense when regarded independently from the base language B. Hence, language B is a dominant language, which can
be a DSL or a GPL, and serves as a base for other languages. Language extension, as a kind
of language composition, is denoted as B / E indicating that the base language B has been
extended with the language E. The LISA tool supports language extensions when single
attribute grammar inheritance [8] is employed. As is shown in Section IV, EasyTime++ is a
language extension over the base language EasyTime (EasyTime / EasyTime++). In language unification the composition of language specifications is not based on the dominance
of one language, but is based on equal terms. The dominance of one language over another
does not exist and both language specifications are complete and standalone (note that in
the case of language extension the language specifications for the extended part makes little
sense alone). Language unification, as a kind of language composition, is denoted as L1 ]g L2 ,
describing the language composition of languages L1 and L2 using a glue code g. Since LISA
supports multiple attribute grammar inheritance [8], language unification is easily achieved
by inheriting both language specifications (from L1 and L2 ), where the glue code is specified
as a new language specification fragment. In self-extension the language specifications do
3
not change. The language itself is powerful enough for new extensions to be implemented
using macros, function composition, and libraries that provide domain-specific constructs.
This form of language composition is called ’pure language embedding’ [3]. Functional languages are these languages particularly suitable for self-extension. Self-extension, as a kind
of language composition, is denoted as H ← E indicating that the host language H has
been self-extended with the embedded language E. The last form of language composition
is extension composition, which describes how language specifications also support the
combination of various language compositions. That is showing how different compositions
can work together. This kind of language composition can also be described as high-order
language composition. Language unification allows for such higher-order composition per
se (e.g., L1 ]g (L2 ]h L3 )). Whilst some other useful examples of higher-order language
composition like (B / E1 ) / E2 , and B / (L1 ]g L2 ) can not always be easily achieved. Extension compositions involving language extension and language unification can also be easily
achieved in the LISA tool.
TA 1
CP 0
RUN
BIKE
SWIM
TA 2
3.8 km
180 km
42 km
4x
0.95
4x
45
8x
5.25
CP 1
CP 2
MP 1
CP 3
CP 4
CP 5
CP 6
MP 3
MP 2
CP 7
CP 8
MP 4
FIG. 1: Measuring time in Ironman triathlon.
In addition to LISA, which has been in existence since 1999, there are also other similar
tools (e.g., Phobos [18], JastAdd [19], Silver [20], XMF [21], Tatoo [22], MontiCore [23],
JAYCO [25], UUAG [26]) that enable various language compositions. Note, that the most
well-known tools for syntax and the semantic specification of programming languages, Lex
and Yacc [27], don’t support language composition per se. For example, language extension
is possible by manually changing base language specifications B by invasively adding the
specification for extended language E. Hence, change is done in a non-disciplined manner,
thus prohibiting further reuse of specifications. On the other hand, language composition
4
can be done on top of Lex and Yacc (e.g., [28]). Here, it is desirable to briefly mention
JastAdd [19] and Silver [20], since both are based on Attribute Grammars, as in the LISA
case. JastAdd [19] is centered around object-oriented representation of the abstract syntax
tree (AST). Non-terminals act as abstract super classes and productions act as specialized
concrete subclasses that specify the syntactic structure, attributes, and semantic rules. All
these elements can be inherited, specialized, and overridden within subclasses. The idea of
aspect-orientation in JastAdd is to define each aspect of the language in a separate class
and then weave them together at appropriate places. The JastAdd system is a class weaver:
it reads all the JastAdd modules and weaves the fields and methods into the appropriate
classes during the generation of the AST classes. Developers have the possibility of combining various language specifications following the separation of different language aspects
amongst different classes. Silver [20] uses a concept called ’forwarding’ to achieve modular
language extensions, where the extension construct is translated into semantically equivalent
constructs within the host language. Hence, forwarding only allows those new constructs
that can be expressed as a combination of existing language constructs. Additional Silver
features like: with-clause, auto-copying of inherited attributes, collection attributes, pattern
matching, and type-safe polymorphic lists, allow for the host language to be extended in a
more flexible manner, although still restrictive.
III.
EASYTIME
EasyTime was developed for measuring time during Double ultra triathlon in 2009. At
that time, the organizers of this competition were confronted with the problem of how
to measure the times of competitors within three disciplines using a limited number of
measuring devices. Besides this limitation, measures needed to be reliable and accurate,
especially, because of its long duration. Although the measuring time for the triathlon was
our first specific task, the goal was to develop a DSL for measuring time for any competition.
A domain analysis was performed [15] using feature diagrams [29] with the aim of identifying
common and variable concepts, their relations, and structure of particular concepts. In the
case of EasyTime, the concept race consists of sub-concepts: events (e.g., swimming, cycling,
and running), control points (starting and finishing lines, the number of laps), the measuring
time (updating time and decrementing laps), optional transition area (difference between the
5
finish and start times), and agents (automatic or manual). In the next step, these concepts
were mapped to the context-free grammar non-terminals of EasyTime. Finally, its whole
syntax and semantics were developed [15].
In order to illustrate the power of EasyTime, lets describe the Ironman triathlon, as
presented in Figure 1. This triathlon consists of: 3.8 km swim, 180 km cycling, and a 42
km run. These disciplines run one after another with two interruptions: In the first, those
competitors who have finished with swimming prepare themselves for the cycling, whilst in
the second, those competitors who have finished the cycling prepare themselves for running.
Both interruptions occur within so-called transition areas. Their times spent within these
areas are added to their swimming, cycling, and running times, in order to obtain the total
times of specific competitors.
Typically, the organizers divide those courses on which they run particular disciplines
into laps because of easier management. As can be seen in Figure 1, competitors need
to accomplish 4 laps of swimming, 4 laps of cycling, and 8 laps of running. These laps
represent another demand for the measuring time in such competitions because, besides
the intermediate times of each lap, decrementing also needs to be performed. In order to
reduce the number of measuring devices, a measuring point (MP in Figure 1) at which the
intermediate time is measured and the number of laps is decremented, can be incorporated.
In other words, when the number of laps is zero the last intermediate time becomes the final
time of a specific discipline.
This characteristic of the triathlon is put to profitable use by EasyTime. In fact, EasyTime is a DSL that enables the organizers of sporting competitions to adapt measuring
systems for various kinds of competitions, reduce the number of measuring devices, and
achieve accuracy and reliability. The EasyTime program runs on a measuring system and
employs a set of agents that control the measuring devices. For measuring time during
Inronman, as illustrated in Figure 1, the EasyTime program presented in Program 1 is used.
At the start of Program 1, two agents are defined: The former describes a measuring
device on which manual measuring time is performed on a portable computer by an operator,
whilst the latter denotes a measuring device that automatically tracks an event caused when
a competitor crossing the measuring place, based on RFID technology [30]. Typically, the
automatic measuring place is implemented as a mat that acts as an antenna having two
functions: Firstly, the antenna induces a passive tag that is born by competitor. Secondly,
6
Program 1 EasyTime program for measuring time in Ironman
1: 1 manual ”man.dat”;
// definition of manual agent
2: 2 auto 192.168.225.100; // definition of automatic agent
3: // definition of variables
4: var ROUND1 := 4;
5: var INTER1 := 0;
6: var SWIM := 0;
7: var TRANS1 :=0;
8: var ROUND2 := 4;
9: var INTER2 :=0;
10: var BIKE := 0;
11: var TRANS2 :=0;
12: var ROUND3 := 8;
13: var INTER3 := 0;
14: var RUN := 0;
15: // definition of measuring place 1
16: mp[1] → agnt[1] {
17: (true) → upd INTER1;
18: (true) → dec ROUND1;
19: (ROUND1 == 0) → upd SWIM;
20: }
21: // definition of measuring place 2
22: mp[2] → agnt[1] {
23: (true) → upd TRANS1;
24: }
25: // definition of measuring place 3
26: mp[3] → agnt[2] {
27: (true) → upd INTER2;
28: (true) → dec ROUND2;
29: (ROUND2 == 0) → upd BIKE;
30: }
31: // definition of measuring place 4
32: mp[4] → agnt[2] {
33: (true) → upd INTER3;
34: (ROUND3 == 8) → upd TRANS2;
35: (true) → dec ROUND3;
36: (ROUND3 == 0) → upd RUN;
37: }
this induced tag acts as a transmitter that transmits its identification code to a measuring
device. The transmitted code is detected by the measuring device as an event. This event
is transmitted to the measuring system and recorded into a database by an agent.
After the agents definition in Program 1, a declaration of variables follows. For each
measuring place, two variables are defined in general: an intermediate time INTERx and a
laps counter ROUNDx. The final achievements of a competitor for specific disciplines are
saved within variables SWIM, BIKE, and RUN.
The EasyTime program is completed by definitions of measuring places mp[i], where i
represents its identification number that must be defined uniquely. The measuring place
represents a physical device that is connected to a measuring device. The measuring device
can support more measuring places simultaneously. Conversely to a measuring place, a
control point (CP in Figure 1) represents an event from a logical point of view and denotes
the specific location on the course, where the referees need to track the time information
7
about competitors. As a matter of fact, the control points in EasyTime are directly mapped
into variables. The identification numbers of each agent responsible for transmitting the
events is assigned to each measuring place. For example, manual agent agnt[1] in line 16 of
Program 1 controls the first measuring place.
Before recording the event into a database, a sequence of statements in curly brackets
are interpreted on an abstract machine (AM). These statements are in forms of (predicate)
→ operation, where operation denotes a sequence of instructions that are executed when
the value of predicate returns value true. Typically, two instructions are employed in
EasyTime++: upd and dec. The former update the value of variable in the database,
whilst the latter decrements its value.
Although, DSLs can be implemented in vastly possible ways [1], an appropriate implementation when the end-users are not also the programmers is, a compiler/interpreter approach
[31]. Hence, EasyTime was implemented using a compiler generator tool called LISA [8, 14].
The LISA specifications include lexical, syntax and semantic specifications. Whilst classical
regular expressions and BNF are used for the first two specifications, the third specifications
are based on Attribute Grammars [32]. One of the distinguishing features of a LISA compiler generator is that specifications (lexical, syntax, and semantics) can be easily reused
and extended. An overall view of LISA specifications is given in Listing I.
IV.
EASYTIME++
EasyTime’s formal description was introduced in [15], whilst the mapping of EasyTime’s
denotational semantics into attribute grammars, as well as its implementation, are presented in [16]. Due to requested extensions of EasyTime the language has evolved into
EasyTime++. This section describes the formal specifications that were necessary for the
change. Due to the space constraints, we are unable to include complete specifications.
Interested readers are further referred to [15, 16].
The first small change was done within the semantic domain Runners, which represents
a database of competitors (Listing II). The additional components are now Gender and
Category. Along with Id, RF ID, LastN ame, and F irstN ame, the Gender and Category
regarding competitors have added to the semantic domain Runners. The second, and the
most important change within the semantic domains is how the State, which is the mapping
8
Listing I: Overall view of LISA specifications
language L1 [ extends L2 , ..., LN ] {
lexicon {
[[P] overrides | [P] extends] R regular expr.
..
.
}
attributes type A 1, ..., A M
..
.
rule [[Y] extends | [Y] overrides] Z {
X ::= X11 X12 ... X1p compute {
semantic functions }
..
.
|
Xr1 Xr2 ... Xrt compute {
semantic functions }
;
}
..
.
method [[N] overrides | [N] extends] M {
operations on semantic domains
}
..
.
}
Listing II: Semantic domains in EasyTime++
Category={0, 1, 2, 3 . . .}
Gender={f emale, male}
Runners=(Id × RF ID × LastN ame × F irstN ame × Gender × Category)∗
State=Var → ((Category → Integer) × Truth-Value)
from variables to values, has been modelled (Listing II). Within EasyTime, the State was
a simple mapping: State=Var→Integer, however in EasyTime++ an initial value of an
attribute depends on categories, and a variable, called ’dynamicvar’, can also be initialized
during a run-time. Hence, the State is now modelled as: State = Var → ((Category →
Integer) × Truth-Value)
Let us describe the State using a simple excerpt from EasyTime++ declarations. Three
9
Listing III: Meaning of declarations in EasyTime++
D : Dec → State
D[[var x := a]]s
D[[var x := {cat1 → a1 , cat2 → a2 }]]s
D[[dynamicvar x]]s
D[[D1 ; D2 ]]s
→
=
=
=
=
State
s[x → ((λcategory.a) × f alse)]
s[x → ({cat1 → a1 , cat2 → a2 } × f alse)]
s[x → ((λcategory. ⊥) × true)]
D[[D2 ]](D[[D1 ]]s)
variables were declared within an EasyTime++ program (Program 2). The first variable,
ROU N D1 specifies that all competitors need to complete 50 laps, hence in a database of
competitors the attribute ROU N D1 is set at 50 for all competitors. The second variable,
ROU N D2, specifies that, in a case where a competitor belongs to category = 1, he/she
needs to complete 20 laps, whilst a competitor within category = 2 only needs 10 laps. In
the database of competitors, the attribute ROU N D2 is initialized according to the category.
For all competitors in the first category this attribute will be initialized to 20, and for all
competitors in the second category to 10. The third variable, P EN ALT Y , is a dynamic
variable and its initial value for each competitor will be set during the run-time.
Program 2 Excerpt from EasyTime++ declarations
1: var ROUND1 := 50;
2: var ROUND2 := { (category==1) → 20,
3:
(category==2) → 10 };
4: dynamicvar PENALTY; // definition of dynamic var.
The State in EasyTime++ is mapping which maps variable names (e.g., ROU N D1,
ROU N D2, P EN ALT Y ) into two components. The first component is itself a mapping from
Category to Integer (e.g., 1→20, 2→10), whilst the second component indicates whether
a variable is dynamic or not. To cope with this new model for variables in EasyTime++,
the following LISA methods are needed (note that the mapping from Category to Integer
can be implemented using a hashtable [16], see Program 3).
Since all changes in EasyTime++ are done in a declaration part the semantic function
D (for full description of EasyTime semantic functions please see [15, 16]), which describes
the meanings of the declarations needs to be changed accordingly (Listing III).
10
Program 3 Implementation of EasyTime++ State in LISA
1: method M Var {
class Var {
2:
3:
String name;
4:
Hashtable values;
5:
boolean isDynamic;
6:
Var (String name, Hashtable values,
7:
boolean isDynamic) {
8:
this.name = name;
9:
this.values = values;
this.isDynamic = isDynamic;
10:
11:
}
12:
// Java methods are omitted
13:
...
}
14:
15: }
16:
17: method VarEnvironment {
18:
import java.util.*;
19:
public Hashtable put (Hashtable env, Var aVar) {
20:
env = (Hashtable)env.clone();
21:
env.put(aVar.getName(), aVar);
22:
return env;
23:
} // java method
24: } // Lisa method
Semantic function D maps the syntactic construct Dec, representing the declarations,
into its meaning State → State, which is a mapping from State to State. Note, how
the first component of State is defined in a case where the categories are unspecified (first
equation in Listing III), and in a case of dynamic variables (third equation in Listing III).
In the first equation, it is stated that variable x is mapped to value a regardless of cat-
11
egory. The mapping function λcategory.a is a constant function. The second equation
states that variable x is mapped to different values (e.g., a1 , a2 ) according to different categories (e.g., cat1 , cat2 ), whilst in the third equation, the variable x is mapped to undefined
value ⊥ regardless of category. In the case of dynamic variables the second component of
(Category → Integer) × Truth-Value is true, otherwise it is false.
The aforementioned changes in formal specifications of EasyTime++ also require changes
in the implementation part. Note that changes are required in the lexical part (new keywords
category and dynamicvar, new separator), syntax part (new syntax rules for declarations),
as well as in the semantic part (new semantics for declarations). All the other parts of
EasyTime (e.g., agents, measuring places, statements) [16] are intact and hence can be completely reused. Since EasyTime is implemented in LISA, which supports attribute grammar
inheritance [8], and where lexical, syntax and semantic specifications can be inherited, it was
natural to extend EasyTime specifications written in LISA for implementing EasyTime++,
thus achieving incremental language development. Program 4 shows the LISA specification
of EasyTime++. Note, how all EasyTime specifications have been reused (’language EasyTime++ extends EasyTime’). In the inherited specifications it was necessary to override rule
Dec, which contained syntactic and semantic specifications for declarations, add some new
grammar productions and their semantics (rule Categories), as well as add new attribute
varvalues of type Hashtable, which were attached to the non-terminal CT GRS, and extend regular definitions for Separator and Keyword. Overall less than 70 lines of LISA
specifications have been newly written to obtain the complete compiler for EasyTime++.
Note that this is an example of language extension where language specifications’ fragment
(Program 4) alone does not make any sense and can not exist without base-language specifications (for complete EasyTime specifications in LISA see [16]). Hence, this kind of language
composition can be denoted as EasyTime / EasyTime++.
V.
EXAMPLES
In order to test EasyTime++ DSL two case-studies were performed:
• cyclo-cross Grand-prix, and
• biathlon.
12
Program 4 LISA specification of EasyTime++
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
38:
39:
40:
41:
42:
43:
44:
language EasyTime++ extends EasyTime {
lexicon {
extends Separator ,
extends Keyword category — dynamicvar
}
attributes Hashtable *.varvalues;
rule extends Start {
compute { }
}
rule overrides Dec {
DEC ::= var #Id := #Int ; compute {
// category is not specified; isDynamic=false
DEC.outState = put(DEC.inState,
new Var(#Id.value(),
put(new Hashtable(), ”0”,
Integer.valueOf(#Int.value()).intValue()), false));
};
DEC ::= dynamicvar #Id ; compute {
// category can not be specified; isDynamic=true
DEC.outState = put(DEC.inState,
new Var(#Id.value(), null, true));
};
DEC ::= var #Id := { CTGRS } ; compute {
// categories are specified and can’t be dynamic
DEC.outState = put(DEC.inState,
new Var(#Id.value(), CTGRS.varvalues, false));
};
}
rule Categories {
CTGRS ::= ( category == #Int ) → #Int , CTGRS
compute {
CTGRS[0].varvalues = put(CTGRS[1].varvalues,
#Int[0].value(),
Integer.valueOf(#Int[1].value()).intValue());
};
CTGRS ::= ( category == #Int ) → #Int compute {
CTGRS.varvalues = put(new Hashtable(),
#Int[0].value(),
Integer.valueOf(#Int[1].value()).intValue());
};
}
...
// LISA methods
}
13
The former was experienced in practice, whilst the latter could be taken as proof of
concept. In the rest of this section, both case-studies are discussed in detail.
A.
Case-study 1: Cyclo-cross Grand-Prix
This case-study tested the introduction of categories in EasyTime++. Cyclo-cross is a
relatively new sport that typically takes place in winter and is dedicated to cycle road-riders
who are preparing for the new season. Races usually consist of several laps of a short course
featuring pavements, wooded trails, grass, steep hills, and obstacles.
In this case-study, one lap of 2.5 km was used (Figure 2). According to the number of
laps, the competitors were divided into three categories, as follows:
• 4 laps: junior men and women up to 15 years old (U-15),
• 6 laps: junior men and women up to 19 years old (U-19), and
• 9 laps: absolute categories (U-23, Elite, Masters).
In order to make the competition more interesting, the organizers allowed all the competitors onto the course simultaneously. Only one measuring device with two measuring
places was needed for measuring this competition because the course passed at one location.
Here, the intermediate times of laps were measured and, thereby, decremented the laps’
counters of specific competitors. When the laps counter reached zero the finish time of the
competitor was reported. However, how many laps to go depended on the category to which
the specific competitor belonged.
The EasyTime++ program for measuring time in this competition can be seen by Program 5. Note that here both measuring places, i.e., mats, were laid so that the whole length
of the finish line was captured. In line with this, a competitor can cross either of both mats.
As a result, the programs for both measuring devices are the same, and work in parallel.
14
FIG. 2: Track layout of the cyclo-cross competition.
Program 5 EasyTime++ program for measuring time in cyclo-cross competition
1: 2 auto 192.168.225.100; // definition of agent
2: var BIKE := 0;
3: var ROUND1 := { (category==1) → 4,
(category==2) → 6, (category==3) → 9 };
4:
5: // definition of measuring place 1
6: mp[1] → agnt[2] {
7:
(true) → dec ROUND1;
8:
(ROUND1 == 0) → upd BIKE;
9: }
10: // definition of measuring place 2
11: mp[2] → agnt[2] {
12:
(true) → dec ROUND1;
13:
(ROUND1 == 0) → upd BIKE;
14: }
In summary, measuring time in cyclo-cross performed well with EasyTime++. Although
the organizers prepared three different lengths of courses, six different lists of results were
15
obtained according to gender. Fortunately, in our case the gender could be handled by a
database system, whilst the EasyTime++ program was unaware of it.
B.
Case-study 2: Biathlon
A biathlon was the second case-study for EasyTime++. Biathlon refers specifically to
the winter sport that combines cross-country skiing and rifle shooting. As can be seen from
Figure 3, competitors start with skiing. Skiing is interrupted by rifle shooting. Typically,
biathlon consists of 4 laps of skiing. The shooting appears close to the end of a lap. Two
positions for competitors are allowed when shooting, i.e., prone and standing. Interestingly,
the number of missed shoots is penalized by the additional number of penalty laps. Note
that the time spent within the penalty laps are added to the total time of the competitor.
The time of the penalty lap is typically taken to be between 20-30 seconds.
The EasyTime++ program for measuring time in biathlon can be seen by Program 6.
Three measuring devices are needed to cover this competition. Each measuring device
realizes one measuring place. Moreover, each measuring place also represents a control
point. In contrast to Ironman, in a biathlon time spent in penalty loops is of no interest in
the preferred race. Here, only the total time of competitor is important.
CP 0
CP 1
CP 2
CP 3
PENALTY
MP 1
MP 2
MP 3
4x
FIG. 3: Measuring time in biathlon competitions.
In summary, the first device represents the special measuring device for counting hits.
The agent assigned to this device puts the number of missed hits into the database variable
PENALTY, dynamically. Note that this device is treated in EasyTime like an ordinary
measuring device. The second measuring device is dealt with by counting the penalty laps,
whilst the third device measures the final time.
16
Program 6 EasyTime++ program for measuring time in biathlon competition
1 auto 192.168.225.110; // definition of agent 1
2 auto 192.168.225.100; // definition of agent 2
var ROUND := 4;
var RUN := 0;
dynamicvar PENALTY; // definition of dynam.variable
// definition of measuring place 1
mp[1] → agnt[1] {
(true) → upd PENALTY;
}
// definition of measuring place 2
mp[2] → agnt[2] {
(true) → dec PENALTY;
}
// definition of measuring place 3
mp[3] → agnt[2] {
(true) → dec ROUND;
(ROUND == 0) → upd RUN;
}
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
VI.
CONCLUSIONS
Easy language composition is still an open-issue within programming language research.
In particular, a new young field of software language engineering is of interest regarding engineering principles when constructing new languages, whether general-purpose or domainspecific. A language designer would like to build a new language simply by composing
different components and/or extending previous components. This paper has presented
EasyTime++ DSL as a language extension of EasyTime, where the base language specifications written in the LISA compiler generator have been extended with new features,
thus enabling the introduction of categories into competitions, and those new competitions
where the number of laps is dynamically determined. The implemented multiple attribute
grammar inheritance in LISA enables easy language composition since lexical, syntax, and
semantic specifications can be reused and extended. In such a manner, an incremental language development using LISA has been demonstrated. The suitability of EasyTime++ was
shown in two case studies: cyclo-cross Grand-prix and a biathlon. More extensive experimental work, which would include other types of language compositions and more DSLs, is
17
also planned in the future.
[1] M. Mernik, J. Heering, A. M. Sloane. When and how to develop domain-specific languages. ACM Computing Surveys, 2005, 37(4), 316–344.
[2] A. van Deursen, P. Klint, J. Visser. Domain-specific languages: an annotated bibliography. ACM SIGPLAN Notices, 2000, 35(6), 26–36.
[3] P. Hudak. Building domain-specific embedded languages. ACM Computing Surveys, 1996,
28(4es).
[4] M. Fowler. Domain Specific Languages. Addison-Wesley Professional, 2010.
[5] M. J. Varanda Pereira, M. Mernik, D. da Cruz, P. R. Henriques. Program comprehension for domain-specific languages. Computer Science and Information Systems, 2008,
5(2), 1–17.
[6] T. Kosar, N. Oliveira, M. Mernik, M. J. Varanda Pereira, M. Črepinšek, D. da
Cruz, P. R. Henriques. Comparing General-Purpose and Domain-Specific Languages: An
Empirical Study. Computer Science and Information Systems, 2010, 7(2), 247–264.
[7] T. Kosar, M. Mernik, J. C. Carver. Program comprehension of domain-specific and
general-purpose languages: comparison using a family of experiments. Empirical software
engineering, 2012, 17(3), 276–304.
[8] M. Mernik, V. Žumer. Incremental programming language development. Computer Languages, Systems and Structures, 2005, 31(1), 1–16.
[9] D. Hrnčič, M. Mernik, B. R. Bryant. Embedding DSLs into GPLs: A Grammatical
Inference Approach. Information Technology and Control, 2011, 40(4), 307–315.
[10] D. Hrnčič, M. Mernik, B. R. Bryant, F. Javed. A memetic grammar inference algorithm
for language learning. Applied Soft Computing, 2012, 12(3), 1006–1020.
[11] G. Kiczales. Aspect-oriented programming. ACM Computing Surveys, 1996, 28(4), Article
No. 154.
[12] P. Klint, T. van der Storm, J. J. Vinju. Term rewriting meets aspect-oriented programming. Processes, Terms and Cycles: Steps on the Road to Infinity. Springer-Verlag, 2005,
88–105.
[13] M. Mernik, D. Rebernak. Aspect-Oriented Attribute Grammars. Electronics and Electrical
18
Engineering, 2011, 10(116), 99–104.
[14] P. R. Henriques, M. J. Varanda Pereira, M. Mernik, M. Lenič, J. Gray, H. Wu.
Automatic generation of language-based tools using LISA. IEE Proceedings - Software Engineering, 2005, 152(2), 54–69.
[15] I. Jr. Fister, I. Fister, M. Mernik, J. Brest. Design and implementation of domainspecific language Easytime. Computer Languages, Systems and Structures, 2011, 37(4), 276–
304.
[16] I. Jr. Fister, M. Mernik, I. Fister, D. Hrnčič. Implementation of EasyTime Formal
Semantics using a LISA Compiler Generator. Computer Science and Information Systems,
2012, 9(3): 1019–1044.
[17] S. Erdweg, P. G. Giarrusso, T. Rendel. Language Composition Untangled. In: Proceedings of Workshop on Language Descriptions, Tools and Applications (LDTA’12). Published online at http://www.informatik.uni-marburg.de/∼seba/publications/languagecomposition. pdf,
2012.
[18] A. Granicz, J. Hickey. Phobos: A front-end approach to extensible compilers. In: Proceedings of the 36th Annual Hawaii International Conference on System Sciences (HICSS36),
2003.
[19] G. Hedin, E. Magnusson. JastAdd: an aspect-oriented compiler construction system. Science of Computer Programming, 2003, 47(1), 37–58.
[20] E. Van Wyk, D. Bodin, J. Gao, L. Krishnan. Silver: An extensible attribute grammar
system. Science of Computer Programming, 2010, 75(1), 39–54.
[21] T. Clark, P. Sammut, J. Willans. Superlanguages: Developing languages and applications
with XMF. Published online at hhtp://bit.ly/HiTOKp, 2008.
[22] J. Cervelle, R. Forax, G. Roussel. A simple implementation of grammar libraries. Computer Science and Information Systems, 2007, 4(2), 65–77.
[23] H. Krahn, B. Rumpe, S. Voelkel. MontiCore: Modular development of textual domain
specific languages. In: Proceedings of the 30th International Conference on Software Engineering (SLE 2008), 2008, 925–926.
[24] A. Stonis, D. Rubliauskas, J. Blonskis. Mapping Syntax Extensions. Information Technology and Control, 2002, 24(3), 35–48.
[25] J. Porubän, M. Forgáč, M. Sabo, M. Bı̀hálek. Annotation Based Parser Generator.
19
Computer Science and Information Systems, 2010, 7(2), 291–307.
[26] M.
Viera,
S.
D.
Swierstra,
A.
Middelkoop.
How to make Attribute Grammars First-Class. In:
Language
Descriptions,
Tools
and
Applications
UUAG
Meets
AspectAG:
Proceedings of Workshop on
(LDTA’12).
Published
online
at
http://www.cs.uu.nl/research/techreps/repo/CS-2011/2011-029.pdf, 2012.
[27] D. Brown, J. Levine, T. Mason. lex & yacc. O’Reilly Media, 2nd Edition, 1992.
[28] X. Wu, B.R. Bryant, J. Gray, M. Mernik. Component-based LR parsing. Computer
Languages, Systems and Structures, 2010, 36(1), 16–33.
[29] V. Štuikys, R. Damaševicius. Measuring Complexity of Domain Models Represented by
Feature Diagrams. Information Technology and Control, 2009, 38(3), 179–187.
[30] K. Finkenzeller. RFID Handbook. John Wiley & Sons, 2010.
[31] T. Kosar, P. E. Martı́nez López, P. A. Barrientos, M. Mernik. A Preliminary Study on
Various Implementation Approaches of Domain-Specific Language. Information and Software
Technology, 2008, 50(5), 390–405.
[32] J. Paakki. Attribute Grammar Paradigms - A High-Level Methodology in Language Implementation. ACM Computing Surveys, 1995, 27(2), 196–255.
Updated 3 March 2013.
20
| 6cs.PL
|
Concave Flow on Small Depth Directed Networks
arXiv:1704.07791v1 [cs.DS] 25 Apr 2017
Tung Mai
Georgia Tech
[email protected]
Richard Peng
Georgia Tech
[email protected]
Anup B. Rao
Georgia Tech
[email protected]
Vijay V. Vazirani
Georgia Tech
[email protected]
April 26, 2017
Abstract
Small depth networks arise in a variety of network related applications, often in the form of
maximum flow and maximum weighted matching. Recent works have generalized such methods
to include costs arising from concave functions. In this paper we give an algorithm that takes
a depth D network and strictly increasing concave weight functions of flows on the edges and
computes a (1−ǫ)-approximation to the maximum weight flow in time mDǫ−1 times an overhead
that is logarithmic in the various numerical parameters related to the magnitudes of gradients
and capacities.
Our approach is based on extending the scaling algorithm for approximate maximum weighted
matchings by [Duan-Pettie JACM‘14] to the setting of small depth networks, and then generalizing it to concave functions. In this more restricted setting of linear weights in the range
[w min , w max ], it produces a (1 − ǫ)-approximation in time O(mDǫ−1 log(w max /w min )). The
algorithm combines a variety of tools and provides a unified approach towards several problems
involving small depth networks.
1
Introduction
Combinatorial problems have traditionally been solved using combinatorial methods, e.g., see
[Sch03]. Aside from providing many state-of-the-art results, these methods also have the advantages of simplicity and interpretability of intermediate solutions. Recently, algorithms that
utilize continuous methods have led to performances similar or better than these combinatorial
methods [CKM+ 11, Mad13, ZO15]. A significant advantage of continuous approaches is that they
extend naturally to general objective functions, whereas in combinatorial algorithms there is much
difference, for example, between even finding a maximum flow and a minimum cost flow.
This paper is motivated by recent works that used accelerated gradient descent to obtain
O(mǫ−1 log n) time routines for computing approximate solutions of positive linear programs [ZO15,
WRM15]. Such linear programs include in particular (integral) maximum weighted matchings on
bipartite graphs, and fractional weighted matchings on general graphs. These algorithms, as well
as earlier works using continuous methods for the maximum weighted matching problems [AG14],
represent a very different approach that obtained bounds similar to the result by Duan and Pettie [DP14]; the latter computed similar approximations using combinatorial methods. We study
this connection in the reverse direction to show that the augmenting path based approaches can be
extend to 1) small depth, acyclic networks, and 2) more general cost objectives. Formally, given a
DAG with source s and sink t, along with capacities c e and differentiable concave functions fe on
each
P edge, we want to find a flow from s to t that obeys capacity c e and maximizes the function
e fe (x e ). Our main result to this direction is:
Theorem 1.1. Given a depth D acyclic network with positive capacities and concave cost functions
whose derivatives are between w min and w max and queryable in O(1) time, ConvexFlow returns
a (1 − ǫ)-approximation to the maximum weighted flow in O Dmǫ−1 log(w max /w min ) log n time.
We note that the logarithmic dependency on w max /w min is common in scaling algorithms
for flows: the case of linear costs corresponds to fe (x e ) = w e x e for scalars w e , and in turn
w min = mine w e and w max = maxe w e . We will discuss the relation between this formulation and
other flow problems, as well this dependency on weights in Section 1.2.
1. Initialize flow x , potentials p, step size δ
2. Repeat until δ is sufficiently small:
(a) Repeat O(1/ǫ) steps:
i. Compute eligible capacities on the edges, forming the eligible graph.
ii. Send a maximal flow in the eligible graph.
iii. Recompute eligible graph, adjust potentials of vertices unreachable from s.
(b) Reduce δ, and adjust the potentials.
Figure 1: Algorithmic Template for Our Routines
The main components of our algorithm are in Figure 1. It involves the construction and flow
routing on an eligiblility graph, which consists of all positive capacitiated edges calculated from a set
of eligibility rules. Note that aside from the blocking-flow computation on this graph in Step 2(a)ii,
1
all other steps are local steps based on values on an edge, potentials of its two end points, and an
overall step size. This algorithm in its fullest generalization is in Figure 8 in Section 4.
The paper is organized as follows. In Section 2 we describe the simplest variant: max weight
flow on unit-capacity, small depth networks where the cost functions on edges are linear. The main
reason for presenting this simple setting is to show how we trade running time with accuracy by
appropriately relaxing eligibility conditions for edges. In Section 3, we extend our algorithm to
situations with widely varying edge weights via a scaling scheme. We then extend the result to
concave functions in Section 4, leading to our main result as stated in Theorem 1.1.
1.1
Our Techniques
Our main technical contribution is a method for combining augmenting paths algorithms with
primal-dual schemes for flows with costs. We combine blocking flows, which are essentially maximal sets of augmenting paths, with the weight functions on edges via eligibility rules that dictate
which edge in the residual graph can admit flow based on the current primal / dual solution. Invariants that follow from these eligibility rules then allow us to maintain approximate complementary
slackness between the primal and dual solutions. In turn, these yield approximation guarantees.
Our eligibility rules are directly motivated by the scaling algorithms for approximate maximum
matching [DP14], which is in turn based on scaling algorithms for maximum weighted matchings [GT89, GT91, Gab17]. However, the fact that augmenting paths can contain edges from
different depth levels pose several challenges. Foremost is the issue that a very heavily weighted
edge may cause flow to go through a lightly weighted edge. As a result, we need to find a way to
work around the key idea from [DP14] of relating the weight of an edge to the first time a flow
passes through it. We do so by using paths as the basic unit by which we analyze our flows, and
analyzing the entire history of paths going through an edge (see Section 3.3 for details).
We will develop our algorithms first in the case where all weights on edges are linear, and then
extend them to concave functions. These extensions are obtained by viewing each edge with nonlinear cost functions as a collection of edges with differing weights and infinitesimal capacities, and
then simulating a continuous version of our linear algorithm through blocking flows.
1.2
Related Works
As our results combine techniques applicable to a wide range of combinatorial problems, our results
have similarities, but also differ, from many previous works. An incomplete list of them is in
Figure 2. In order to save space, we use Õ(f (n)) to hide terms that are polylog(f (n)): this is a
widely used convention for the more numerical variants of these objectives, and here mostly hides
factors of log n.
While the guarnatees of many of the algorithms listed here are stated for inputs with integer
capacities / costs [GT89, GR98, LS14], their running times are equivalent to producing 1 ± ǫapproximates with poly(log(ǫ−1 )) overhead in running time.
Our guarantees, which contain poly(ǫ−1 ) in the running time terms, closest resemble the approximation algorithms for weighted matchings [DP14, ZO15, AG14] and flows on small depth
networks [Coh95]. These running times only outperform the routines with poly(log(ǫ−1 )) depdencies for mild ranges of ǫ. In such cases, the solution can differ significantly point-wise from the
optimum: for example, deleting ǫ-fraction of a maximum cardinality matching still gives a solution
that’s within 1 ± ǫ of the optimum. As a result, the barriers towards producing exact solutions in
2
reference
[GR98]
[Orl13]
[Coh95]
[Pen16]
our work
[LS14]
[Orl93]
[GT87]
our work
[GT89]
[DP14]
[ZO15]
[AG14]
our work
[Hoc07]
[Vég12]
our work
objective
flow amount
flow amount
flow amount
flow amount
flow amount
min linear cost
min linear cost
min linear cost
max linear weight
matching weight
matching weight
matching weight
b-matching weight
b-matching weight
concave/convex
min separable convex cost
max concave weight
graph
any
any
small depth
undirected
small depth
any
any
any
small depth
any
any
bipartite 2
any
bipartite
any
any
small depth
approximation
1 ± ǫ in value
exact
1 ± ǫ in value
1 ± ǫ in value
1 ± ǫ in value
1 ± ǫ in value
exact
exact
1 ± ǫ in value
1 ± ǫ in value
1 ± ǫ in value
1 ± ǫ in value
1 ± ǫ in value
1 ± ǫ in value
1 ± ǫ pointwise
exact
1 ± ǫ in value
runtime 1
Õ(m3/2 log(ǫ−1 ))
Õ(mn)
Õ(mD 2 ǫ−3 )
Õ(mǫ−3 )
Õ(mDǫ−1 )
Õ(m3/2 log(ǫ−1 ))
Õ(m2 )
Õ(mn)
Õ(mDǫ−1 )
Õ(mn1/2 log(ǫ−1 ))
O(mǫ−1 log(ǫ−1 ))
Õ(mǫ−1 )
Õ(mpoly(ǫ−1 ))
Õ(mǫ−1 )
Õ(m2 log(ǫ−1 ))
≥ O(n4 )
Õ(mDǫ−1 ) 3
Figure 2: Results Related to Our Algorithm
strongly polynomial time [Hoc07] are circumvented by these assumptions, especially the bounds on
the minimum and maximum gradients. These bounds are natural generalizations of the bounds on
costs in scaling algorithms for flows with costs [GT89]. In some cases, they can also be removed
via preprocessing steps that find poly(n) crude approximations to the optimum, and trim edge
weights via those values [CKM+ 11, DP14, Pen16]. In particular, if all the capacities are integers,
we can increase the minimum weight, or gradient, in a way analogous to the matching algorithm
from [DP14].
Lemma 1.2. If all capacities are integers and the functions fe are non-negative and increasing,
then we can assume
ǫw max
P
,
w min ≥
O (m2 · e c e )
P
2
or w max /w min ≤ O( mǫ · e c e ) without perturbing the objective by a factor of more than 1 ± ǫ.
Proof. First, note that the assumption about integral capacities means that there exists a feasible
flow that routes at least 1/m2 per edge, and such a flow, x ′ , can be produced using a single
depth-first search in O(m) time.
We will fix the gradients to create functions f whose gradients are all within poly(m, ǫ−1 ) of
ǫ
def
w
b = max f
.
e
m2
1
Many of the approximate algorithms assume integer capacities/weights, and have factors log C and/or log N in
their running times, where C and N are the largest capacity and weight respectively. We omit those factors due to
space constraints: they are discussed as dependencies on the accuracy ǫ in the text before Lemma 1.2.
2
This algorithm also can produce fractional matching solutions on general graphs.
3
There is a hidden factor of logarithm of the ratio between max and min gradients.
3
First, note that OP T ≥ w
b since the functions are non-negative. So we can add Pǫwbc e to the gradient
e
of all the functions without changing the optimum by more than ǫw
b ≤ ǫOP T additively. Formally
the functions:
ǫw
b
def
.
f e (x) = fe (x) + x · P
e′ c e′
with associated maximum OP T satisfies OP T ≤ OP T ≤ (1 + ǫ)OP T .
To fix the maximum gradient, consider the truncated functions
(
f e (x)
if x ≥ mǫ2 , and
def
.
fbe (x) = xm
2
ǫ
if x ≤ mǫ2
ǫ · f e m2
It can be checked that fbe (x) ≤ f e (x) at all points, and that the maximum gradient of fbe (x) is
bounded by
ǫ
xm2
2 −1
· fe
≤
O
m
ǫ
w
b
.
ǫ
m2
[
[
So it remains to show that OP
T , the maximum with the functions fb satisfies OP
T ≥ (1 − ǫ)OP T .
For this, consider the optimum solution to OP T , x , the solution
(1 − ǫ) x + ǫx ′
is feasible because it’s a linear combination of two feasible solutions, has objective at least 1 − ǫ
times the objective of x due to the concavity and monotonicity of f , and has flow at least ǫ/m2
per edge. The result then follows from observing that fbe (x) = f e (x) for all x ≥ ǫ/m2 .
Remark By Lemma 1.2, if all capacities are integer, the running time of our result can be written
as Õ(Dmǫ−1 log C), where C is the largest capacity.
Our assumption on the topology of the graph follows the study of small depth networks
from [Coh95], which used the depth of the network to bound the number of (parallel) steps required by augmenting path finding routines. From an optimization perspective, small depth is also
natural because it represents the number of gradient steps required to pass information from s to t.
Gradient steps, in the absence of preconditioners, only pass information from a node to its neighbor, so diameter is kind of required for any purely gradient based methods The acyclicity requirement is a consequence of our adaptation of tools from approximating weighted matchings [DP14],
and represent a major shortcoming of our result compared to exact algorithms for minimum cost
flow [Orl93] and convex flows [Hoc07, Vég12]. However, as we will discuss in Appendix A, there
are many applications of flows where the formulations naturally lead to small depth networks.
2
Simple Algorithm for Unit Capacities and Linear Weights
In this section, we propose a simple (1 − ǫ) approximation algorithm for the case of fe (·) being
linear functions, and all capacities are unit. This assumption means we can write the cost functions
as:
fe (x e ) = w e x e
for some w e ≥ 0. Under these simplifications, the linear program (LP) that we are solving and its
dual become:
4
max
X
e
∀u,
∀e,
X
vu
x vu =
X
min
w ex e
X
ye
e
∀e,
x uv
uv
0 ≤ xe ≤ 1
pv − pu + y e ≥ w e
ps = pt
∀e,
y e ≥ 0.
We add a dummy directed edge from t to s with infinite capacity to ensure flow conservation
at each vertex. For every edge e we mention in the paper, we implicitly assume that e = uv is an
edge in the original network and e 6= ts, unless otherwise stated. The dual inequality involving y e
is equivalent to:
y e ≥ w e + pu − pv,
The RHS is critical for our algorithm and its analysis. We will define it as the reduced weight of
the edge, w pe :
w pe = w e + p u − p v .
Note that the choice of w pe uniquely determines the optimum choice of y e through y e ← max 0, w pe .
To find an optimum flow on these networks it is sufficient to obtain feasible pairs x and p, y
that satisfy the following complementary slackness conditions:
1. For all e, if x e = 0, then y e = 0, which is equivalent to w pe ≤ 0.
2. For all e, if x e = 1 then y e = w pe , which is equivalent to w pe ≥ 0.
The role of these values of w pe can be further illustrated by the following Lemma, which is
widely used in scaling algorithms.
Lemma 2.1. For any flow x for which the flow conservation is satisfied, and any potential p such
that p t = p s , we have:
X
X
w ex e =
w pe x e .
e
e
Proof. Expanding the summation gives
X
X
X
X
w ex e +
x ep u −
x ep v .
w pe x e =
e
e=uv
e=uv
e
Rearranging the last terms gives
X
u
pu
X
uv
x uv −
X
vu
x vu
!
,
which we know is 0 for every u 6= s, t. We also have that the amount of flow leaving s is same as
the amount of flow entering t, and p s = p t so the total contribution of these terms is 0.
5
Note that if p s = 0 and w pe is close to 0 for each e in a path from s to some u, the total
reduced cost of the path is close to p u . This is the view taken by most primal-dual algorithms for
approximate flows: conditions are imposed on valid paths in order to make p u s emulate max weight
residual paths to the source. These routines then start with p t set to very large, and maintaining
a complementary primal/dual pair while gradually reducing p t . Shortest path algorithms such as
the network simplex method and the Hungarian algorithm fall within this framework [AMO93,
Chapter 12]. Such algorithms produce exact answers, but may take Ω(n) steps to converge even in
the unit weighted case due to very long augmenting paths.
The central idea in our algorithms is to regularize these complementary slackness conditions.
Specifically, we relax the x e = 0 case from w pe ≤ 0 to w pe ≤ δ, for some δ which we eventually set to
ǫw min . We may assume that all the weights are integer multiples of δ since the additive rounding
error of each weight w e is at most ǫw e . We use a set of eligibility rules to define when we can send
flows along edges, or whether an arc is eligible. The modified eligibility conditions are shown in
Figure 3.
1. Forward edge: if x e = 0, can send flow along u → v if w pe = δ.
2. Backward edge: if x e = 1, can send flow back from v → u if w pe = 0.
Figure 3: Eligiblity Rules for Simple Algorithm on Unit Capacity Networks with Linear Weights.
(A1) If x e = 1, then w pe ≥ 0.
(A2) If x e = 0, then w pe ≤ δ.
Figure 4: Relaxed Invariant for Simple Algorithm on Unit Capacity Networks with Linear Weights.
These rules leads our key construct, the eligible graph:
Definition 2.2. Given a graph G and a set of eligibility rules, the eligible graph consists of all
eligible residual arcs with positive capacities, equaling to the amount of flow that can be routed on
them. A vertex v is an eligible vertex if there exists a path from s to v in the eligible graph.
The main idea behind working on this graph is that this approximate invariant can be maintained analogously to exact complementary slackness.
This leads to an instantiation of the algorithmic template from Figure 1. Its pseudocode is
given in Figure 5.
The key property given by the introduction of δ is after sending 1 unit of flow along a maximal
collection of paths, t is no longer reachable from s in the eligible graph. We can show this from
the observation that all edges in a path that we augment along become ineligible. For an eligible
forward edge uv, x uv = 0 and w puv = δ. Therefore, after augmentation, x uv = 1 and vu is ineligible.
A similar argument can be applied for an eligible backward edge vu to show that after augmentation
the residual forward edge uv is also ineligible.
6
x = SimpleFlow (G, s, t, {w e } , w min , w max , ǫ)
Input: Unit capacity directed acyclic graph G from s to t,
Weights w e for each edge e, along with bounds w min = mine w e and w max = maxe w e .
Error tolerance ǫ.
Output: (1 − ǫ) approximate maximum weight flow x .
1. Compute levels lu of the vertices u (length of a longest path from s to u).
Initialize x e = 0, p u = w max · lu , δ = ǫw min .
2. Repeat until p t = 0:
(a) Eligibility labeling: Compute the eligible graph.
(b) Augmentation: Send 1 unit of flow along a maximal collection of edge-disjoint paths
in the eligible graph.
(c) Dual Adjustment: Recompute the eligible graph, for each vertex u unreachable from
s, set p u ← p u − δ.
Figure 5: Simple Algorithm Based on the Eligibility Conditions From Figure 3
This fact plays an important role in our faster convergence rate:
Lemma 2.3. The running time of the algorithm given in Figure 5 is O(Dw max m/ǫw min ) where
D is the depth of the network.
Proof. The algorithm starts with p t = Dw max and in each iteration, p t is decreased by δ.
Therefore, the number of iterations is Dw max /δ. Since in each iteration it takes O(m) to find
a maximal set of disjoint augmenting paths using depth-first search, the total running time is
O(Dw max m/ǫw min ).
We also need to check that our eligibility rules and augmentation steps indeed maintain the
invariants throughout this algorithm.
Lemma 2.4. Throughout the algorithm, the invariants from Figure 4 are maintained.
Proof. It is easy to check that at the beginning of the algorithm, the invariants are satisfied. We
will show by induction that dual adjustments and augmentations do not break them. Assume the
two conditions hold before a dual adjustment step. Notice that a dual adjustment can only change
reduced weights of edges that have exactly one eligible endpoint. For an edge e = uv, we consider
2 cases:
• u is eligible, and v is ineligible: a dual adjustment step decreases p v , and thus increases w pe .
Condition (A1) still holds because the reduced weight increases. Condition (A2) still holds
because reduced weight can only increase until it is equal to δ and both u and v are eligible.
• v is eligible, and u is ineligible: a dual adjustment step decreases p u , and thus decreases
w pe . Condition (A1) still holds because the reduced weight can only decrease until it is equal
7
to 0 and both u and v are eligible. Condition (A2) still holds because the reduced weight
decreases.
Now assume that the two conditions hold before an augmentation step, and notice that we only
send flow along eligible edges. For an edge e with x e = 0 before augmentation, the eligibility
condition is w pe = δ, thus after augmentation x e = 1 and condition (A1) is satisfied. For an edge
e with x e = 1 before augmentation, the eligibility condition is w pe = 0, thus after augmentation
x e = 0 and condition (A2) is satisfied.
It remains to bound the quality of the solution produced. Once again, the invariants from
Figure 4 play crucial roles here.
Lemma 2.5. If all the invariants and feasibility conditions hold when the algorithm terminates
with p s = p t , then the solution produced, x , is feasible and (1 − ǫ)-optimal.
Proof. The flow x is feasible because in each augmentation step we only send one unit of flow along
a path in the residual graph.
Let x ∗ be any other st flow obeying the capacities. Our goal is to show that
X
X
w e x e ≥ (1 − ǫ)
w e x ∗e .
e
e
To do so, we first invoke Lemma 2.1, which allows us consider the weights w pe instead. We
invoke the invariants based on the two cases:
1. If x e = 1, then x ∗e ≤ x e since the edges are unit capacitated. The invariant also gives w pe ≥ 0,
which implies
x e w pe ≥ x ∗e w pe .
2. If x e = 0, then w pe x e = 0. The invariant gives w pe ≤ δ, which combined with x ∗e ≥ 0 gives:
x e w pe = 0 ≥ x ∗e (w pe − δ) .
Furthermore, since δ ≤ ǫw e , this also implies
x e w pe ≥ x ∗e w pe − ǫw e x ∗e .
Summing the implications of these invariants across all edges gives
X
X
X
X
X
x ∗e w e .
x e w pe ≥
x ∗e w pe − ǫ
x ∗e w e = (1 − ǫ)
x ew e =
e
e
e
e
8
e
3
Scaling Algorithm for Linear Weights
In this section, we show how to improve the running time of the algorithm in the previous section
through scaling. The scaling algorithm makes more aggressive changes to vertex potentials p u at
the beginning, and moves to progressively smaller changes. To be more precise, the algorithm starts
with δ0 = ǫw max and ends with δT = ǫw min , and the value of δ is halved each scale. Assume ǫ and
w max /w min are powers of 2, this immediately implies
T = log (w max /w min ) .
The vertex potentials are also initialized in a similar way, i.e p u = w max · lu for all u. Scale
i > 0 starts with
p t = Dw max /2i + 2Dδi
and ends when
p t = Dw max /2i+1 .
The change in step sizes leads to more delicate invariants. Instead of having a sharp cut-off at δ as
in the single-scale routine, we further utilize δi to create a small separation between edges eligible
in the forward and backward directions.
Between scales, we make adjustments to all potentials to maintain the invariants. Pseudocode
of this routine is given in Figure 6.
The rest of this section is a more sophisticated analysis built upon Section 2 and the scaling
algorithms from [DP14], leading to the following main claim:
Theorem 3.1. For any ǫ < 1/10,
ScalingFlow returns x that is an (1 − 8ǫ)-approximation in
−1
O Dmǫ log(w max /w min ) log n time.
A main issue in setting up the invariants, as well as the overall proof, is that errors from the
iterations accumulate. Here we need an additional definition about the last time forward flow as
pushed on an edge.
Definition 3.2. During the course of the algorithm, for each edge e = uv, define scale(e) to be
the the largest scale j such that j is at most i, the current scale, and there is a forward flow from
u to v at scale j. Define scalef (e) to be the final value of scale(e) when the algorithm terminates.
Note that scale(e) resets to i if uv is used in the blocking flow generated on Line 2(a)ii at scale
i. For simplicity, we use δe to denote δscale(e) , δef to denote δscalef (e) , and le to denote lv − lu . The
key eligibility rules and invariants are in Figure 7.
The intricacies in Condition (B1) is closely related to the changes in the dual rescaling step
from Line 2(b)i. In particular, Condition (B2), w pe < 2δi , may be violated if δi is halved. The
dual rescaling on Line 2(b)i is in place precisely to fix this issue by lowering w pe . It on the other
hand causes further problems for lower bounds on w pe , leading to modifications to Condition (B1)
as well. We first verify that these modified invariants are preserved by the dual rescaling steps.
3.1
Dual Rescaling Steps
Lemma 3.3. The dual rescaling step on Line 2(b)i of Figure 7 maintain invariants (B1) and (B2).
9
ScalingFlow (G, s, t, {w e , c e } , w min , w max , ǫ)
Input: Directed acyclic graph G from s to t,
weights w e for each edge e, along with bounds w min = mine w e and w max = maxe w e ,
capacities c e for each edge e,
error tolerance ǫ.
Output: (1 − ǫ) approximate maximum weight flow x .
1. Compute levels lu of the vertices u (length of a longest path from s to u).
Initialize x = 0, p u = w max · lu , δ0 = ǫw max .
max
2. For i from 1 to T = log w
w min :
(a) Repeat until p t =
Dw max
2i+1
if i < T or until p t = 0 if i = T :
i. Eligibility labeling: Compute the eligible graph based on the eligibility rules:
A. uv is eligible in the forward direction if δi ≤ w puv ,
B. a backward edge vu is eligible if w puv ≤ 0.
ii. Augmentation: Send a blocking flow in the eligible graph.
iii. Dual Adjustment: Recompute the eligible graph, for each vertex u unreachable
from s, set p u ← p u − δi .
(b) If i < T :
i. Dual Rescale: p u ← p u + δi lu ,
ii. δi+1 ← δ2i .
Figure 6: Scaling Algorithm Based on the Eligibility Condition From Figure 7.
′
Proof. We start with Condition (B2). If w pe ≤ 2δi , then if we pick p ′ such that w pe ≤ w pe − δi , we
get:
′
w pe ≤ δi = 2δi+1 .
This is done by updating vertex potential p u ← p u + δi lu for each vertex u at the end of scale i.
It remains to check that the condition
w pe ≥ −3le (δe − δi ) − δi
does not break by decreasing p u − p v by le δi :
− 3le (δe − δi ) − δi − le δi
= −3le δe + 2le δi − δi
= −3le δe + 4le δi+1 − 2δi+1
≥ −3le (δe − δi+1 ) − δi+1 .
where the last line follows because le ≥ 1.
10
1. Forward edge: if x e < c e , u → v is eligible if w pe ≥ δi .
2. Backward edge: if x e > 0, v → u is eligible if w pe ≤ 0.
(B1) If x e > 0, then w pe ≥ −3le (δe − δi ) − δi .
(B2) If x e < c e , then w pe ≤ 2δi .
Figure 7: Relaxed Invariant and Corresponding Eligibility Rules for Scaling Algorithm on Non-unit
Capacity Networks with Linear Weights.
3.2
Blocking Flows
Our algorithm also incorporates more general capacities c e . The changes caused by this in the
linear programming formulations are:
1. In the primal, the capacity constraints become 0 ≤ x e ≤ c e ,
P
2. the dual objective becomes e c e y e .
This extension can be incorporated naturally: instead of finding a maximal set of edge-disjoint
augmenting paths, we find a blocking flow from s to t. Blocking flows can also be viewed as
maximal sets of paths when the edges are broken down into multi-edges of infinitesimal capacities.
They can be found in O(m log n) time using dynamic trees [ST83], and therefore offer similar levels
of theoretical algorithmic efficiency.
In the unit capacity case, an augmentation along any path in the residual graph destroys all
eligible edges in the path, and the asymmetry of the eligibility condition means the new reverse edges
are also ineligible. However, when we move to general capacities, an augmentation may destroy just
a single edge on the path. As a result, it’s simpler to work with the following equivalent definition
of blocking flows.
Definition 3.4. A blocking flow on a graph G is a flow such that every path from s to t goes
through a fully saturated edge.
We first verify that this definition also maintains all the invariants both after augmentation and
dual adjustments. These proofs closely mirror those of Lemma 2.4.
Lemma 3.5. An augmentation by a blocking flow in the eligible graph, as on Line 2(a)ii, maintains
the invariant.
Proof. The algorithm only sends flow through eligible edges. Consider an edge uv on a blocking
flow
1. If the augmentation sends flow from u to v, then uv is an eligible forward edge, and w puv ≥ δi .
After augmentation, condition (B1) must hold.
11
2. If the augmentation sends flow from v to u, then vu is an eligible backward edge, and 0 ≥ w puv .
After augmentation, condition (B2) must hold.
Therefore, all edges on a blocking flow satisfy the set of reduced weight conditions after augmentation.
Lemma 3.6. A dual adjustment step, as specified on Line 2(b)i, maintains the invariant.
Proof. Notice that a dual adjustment can only change the effective reduce weights of edges that
have exactly one eligible endpoint. For an edge e = uv in the graph, we consider 2 cases:
• u is eligible, and v is ineligible: a dual adjustment step decreases p v , and thus increases w pe .
– if x e = c e then condition (B1) still holds because w pe increases.
– if x e < c e then we know that uv is in the residual graph and ineligible. Since uv is
ineligible, w pe must be less than δi . Condition (B2) still holds because w pe can only
increase until it is at least δi and e becomes eligible.
• v is eligible, and u is ineligible: a dual adjustment step decreases p u , and thus decreases w pe .
– if x e > 0 then we know that vu is in the residual graph and ineligible. Since vu is
ineligible, w pe must be greater than 0. Condition (B1) still holds because w pe can only
decrease until it is at most 0 and vu becomes eligible.
– if x e = 0 then condition (B2) still holds because w pe decreases.
Note that at the beginning of the algorithm, x e = 0 and w pe = w e − le w max ≤ 0 for each edge
e. Therefore, the invariants (B2) holds. Induction then gives that these invariants hold throughout
the course of the algorithm.
3.3
Bounding Weights of Paths
We can now turn our attention to the approximation factor. The main difficulty in this proof
is due to cancellations of flows pushed at previous scales. Our proof relies on performing a path
decomposition on each blocking flow in the history of the final solution x . We will view the final
flow, x , as a sum of flows from the collection P.
X
x =
xP.
P ∈P
Note here that x P is a quantity that specifies the amount of flow on that path. We also define FP
to be the set of edges going forward on P , and BP to be the set of edges going backward on P . We
use this decomposition to map backward flow to forward flow in an earlier scale. We will also use
scale(P ) to denote the scale at which P was pushed, and δP to denote δscale(P ) .
Note that we will likely push flows on an edge in forward and backward directions many times
throughout the course of the algorithm. Each of these instances belong to a different path, and is
taken care of separately in the summation. We first establish a lower bound on the weight of each
path using the eligibility criteria: we show that the weight of the path routed at phase i can be
lower bounded by terms involving δi , the step size at that phase.
12
Lemma 3.7.
X
e∈FP
we −
X
e∈BP
we ≥
DδP
2ǫ
Proof. Let p be the vertex potentials from when flow was pushed along P . Note that p may not
be the final potentials. By the definition of w pe , we have
X
X
X
X
we −
we =
w pe −
w pe + (p t − p s ) .
e∈FP
e∈BP
e∈FP
e∈BP
For the edges on the path, the eligibility rules give
• w pe ≥ δP ≥ 0 if e ∈ FP , and
• −w pe ≥ 0 if e ∈ BP .
Furthermore, our outer loop condition on Line 2a ensures that at scale i, p t is at least Dw max /2i+1 =
Dδi /2ǫ, while p s is always 0. Therefore the total is at least Dδi /2ǫ.
This allows us to map the final inaccuracies of weights caused by the Condition (B2) of the
invariants to ǫ of the total weights. The following proof hinges on path decompositions
Lemma 3.8. Let x be the final answer. Then we have
X
X
w ex e.
3le δef x e ≤ 6ǫ
e
e
Proof. Let x P be the amount of flow sent through P. The definition of x and summation of flows
gives
X
X
xe =
xP −
xP.
P :e∈FP
P :e∈BP
reversing the ordering of summation, and aggregating per edge gives:
X
X
X
X
3le δef x e =
3le δef
xP −
xP.
e
e
P :e∈FP
P :e∈BP
For each edge e,
δef
X
P :e∈FP
xP −
X
P :e∈BP
xP =
X
P :e∈FP
δef x P −
X
δef x P
(1)
P :e∈BP
Consider the set of paths P1 , . . . Pk sending flow on edge e (in both directions) in the order of index
i for i = 1 . . . k. Whenever a flow is sent from v to u (a backward direction), there must be some
existing flow from u to v. By breaking a path into multiple paths carrying smaller amounts of flow,
we may assume that for each Pi such that e ∈ BPi , there exists Pj such that j < i, e ∈ FPj and
x Pi = x Pj . We define a function that maps i to j. Since j < i, we must have δPj ≥ δPi . It follows
that
δef x Pj − δef x Pi = 0 ≤ δPj x Pj − δPi x Pi
13
We apply the following modifications to the RHS of equation (1). For each pair (i, j) such that i
is mapped to j, we increase the multiplication factor of x Pi from δef to δPi and the multiplication
factor of x Pj from δef to δPj . Finally, for the remaining forward flows that are not canceled, we
increase δef to the value δP of the path containing them. Such modifications can only increase the
value of the RHS of (1). Therefore, for each edge e,
X
P :e∈FP
δef x P −
X
P :e∈BP
δef x P ≤
X
P :e∈FP
δP x P −
X
δP x P .
P :e∈BP
Plugging back to the summation, we have
X
X
X
X
3le δef x e ≤
3le
δP x P −
δP x P .
e
e
P :e∈FP
P :e∈BP
Changing the order of summation and notice that for each path P
X
X
le δP −
le δP = DδP ,
e∈FP
which by Lemma 3.7 is bounded by 2ǫ
then gives the result.
3.4
e∈BP
P
e∈FP
we −
P
e∈BP
w e . Summing over all paths P ∈ P
Putting Things Together
It remains to put these together through the use of reduced costs in a way analogous to Lemma 2.5.
Proof. (of Theorem 3.1) The number of iterations in each scale is D/2ǫ + 2D = O(D/ǫ). The
number of scales is log(w max /w min ). Each iteration takes O(m log n) to find a blocking flow.
Therefore, the total running time is O ((Dm log n/ǫ) log(w max /w min )).
∗
For the approximation guarantees,
any
pP
denote the final potenP consider
P optimal flow
Px . pLet
p
tials. Once again, Lemma 2.1 gives e w e x e = e w e x e and e w e x ∗e = e w e x ∗e .
We will incorporate the invariants from Figure 7 through two cases:
1. If x e < x ∗e , then x e < c e , and the invariant gives w pe ≤ 2δT , which implies
w pe x e ≥ (w pe − 2δT ) x e ≥ (w pe − 2δT ) x ∗e .
Since δT = ǫw min ≤ w e , this implies
x e w pe ≥ x ∗e w pe − 2ǫw e x ∗e .
2. If x e > x ∗e , then x e > 0, and the invariant gives
w pe ≥ −3le δef
which combined with x ∗e ≤ x e gives:
x e w pe + 3le δef ≥ x ∗e w pe + 3le δef ≥ x ∗e w pe
14
Combining these gives
X
e
X
x e w pe + 3le δef
x e w e + 3le δef =
e
≥
X
e
x ∗e w pe − 2ǫ
= (1 − 2ǫ)
X
X
w e x ∗e
e
w e x ∗e
e
and the result then follows from Lemma 3.8.
4
Non-Linear Networks
We now move to the non-linear setting with arbitrary capacities. The primal objective becomes:
X
max
fe (x e )
e
Here we will let w e (x e ) denote the weight of e when x e units are on it, i.e.,
w e (x e ) =
d
fe (x e ).
dx e
Since fe (·) is concave, we know that w e (x e ) is monotonically decreasing. Pseudocode of our
algorithm is in Figure 8, and its eligibility rules are in Figure 9. Note that the eligibility rules in
the pseudocode is a shorter version of the one in Figure 9.
We remark that the cost of finding Le and Ue , the maximum amount the flow can increase/decrease
before the gradient of fe (x e ) changes by ∆, depends on the structures of fe (·). In case of simple
functions such as quadratics, they can be calculated in closed form. For more complex functions,
finding these points can be viewed as analogs of line searches [BV04].
Conceptually, we divide each edge uv into multiple imaginary parallel edges between u and v,
each with a linear weight function. This can be viewed as approximating w e (x e ) by a decreasing
step function. To be precise, for each edge e, we break it into k parallel edges e1 , . . . , ek such that:
1. w e1 = ⌊w e (0)/δT ⌋δT , w ek = ⌊w e (c e )/δT ⌋δT and w ei = w ei−1 − δT for all 1 < i < k.
2. c ei = x i − x i−1 where x i s are such that x 0 = 0, x k = c e and w e (x i ) = w ei for all 0 < i < k.
Let G = (V, E) be the original network, and G = (V, E) be the network with multiedges. We say
that a flow x in G is well-ordered if the following condition holds for every e:
x ei > 0 ⇒ x ej = c ej
∀j < i.
A flow x in G is equivalent with a flow x in G if:
X
x ej = x e .
j
Notice that the well-ordered property ensures that there is a unique reverse mapping from x e to
x e1 . . . x ek as well.
15
x = ConcaveFlow (G, s, t, {c e , fe } , w min , w max , ǫ)
Input: DAG G from s to t,
capacities c e and concave weight functions fe along each edge e,
bounds w min and w max s.t. w min ≤ dxd e fe (x e ) ≤ w max ∀ 0 ≤ x e ≤ c e , error tolerance ǫ.
Output: (1 − ǫ) approximate maximum weight flow x .
1. Compute levels lu of the vertices u (length of a longest path from s to u).
Initialize δ(0) = ǫw max , p u = w max · lu , x e = 0.
max
2. For i from 1 to T = log w
w min :
(a) Repeat until p t =
Dw max
2i+1
if i < T or until p t = 0 if i = T :
i. Form eligibility graph from all arcs with positive capacities given by:
A. Forward (u → v): max ∆ s.t. x e + ∆ ≤ c e and p u − p v + dxd e fe (x e + ∆) ≥ δi .
B. Backward (v → u): max ∆ s.t. ∆ ≤ x e and p u − p v + dxd e fe (x e − ∆) ≤ 0.
ii. Augmentation: Send a blocking flow in the eligible graph.
iii. Dual Adjustment: Recompute the eligible graph, for each vertex u unreachable
from s, set p u ← p u − δi .
(b) If i < T :
i. Dual Rescale: p u ← p u + δi lu ,
ii. δi+1 ← δ2i .
Figure 8: Algorithm for Concave Networks
Lemma 4.1. Let (x , p) be a flow-potential pair at some point in the course of ConcaveFlow on
input G. ScalingFlow on input G can maintain a flow-potential pair (x , p) such that:
1. x is well-ordered.
2. x and x are equivalent.
Proof. It suffices to prove that the above claim holds in an Eligibility Labeling and an Augmentation
step.
Consider an edge uv that is eligible in forward direction in ConcaveFlow algorithm. Since
x is well-ordered and equivalent to x , there are some eligible multiedges between u and v in
ScalingFlow. Moreover, the first condition in Figure 9 guarantees that the amount of flow
allowed on uv is equal to sum of capacities of all eligible parallel edges from u to v in the multiedges
setting. This follows because
1. The value of w e (x e + ∆) such that w pe (x e + ∆) = δi is a multiple of δi , since p u and p v are
also multiples of δi assuming ǫ is a power of 2.
2. A weight of each edge in G is a multiple of δT by construction.
16
1. Forward edge: if x e < c e , uv is eligible if δi < w pe (x e ), and can has forward capacity
Ue (x e , i) := min {c e − x e , ∆ such that w pe (x e + ∆) = δi } .
2. Backward edge: if x e > 0, vu is eligible if 0 > w pe (x e ), with capacity
Le (x e , i) := min {x e , ∆ such that w pe (x e − ∆) = 0} .
Figure 9: Eligiblity Rules for Scaling Algorithm for Concave Weight Functions.
A similar argument can be applied to the backward eligibility case. Specifically, the amount of
flow that can be pushed back on uv in ConcaveFlow is equal to sum of capacities of all eligible
multiedges from v to u in ScalingFlow.
It follows that a blocking flow in the eligible graph in ConcaveFlow is also a blocking flow
in the eligible graph in ScalingFlow. Given a blocking flow z in the single-edge setting, we
can distribute z uv among eligible multiedges between u and v such that the resulting flow x in
ScalingFlow is still well-order, and still equivalent to the resulting flow x in ConcaveFlow.
This mapping then allows us to bound the result of the continuous process, leading to our main
result.
Proof. (Of Theorem 1.1)
Let x ∗ be the optimal solution in the concave flow problem in G, and x be the final solution
returned by ConcaveFlow. Let x ∗ be a solution in G such that x ∗ is well-ordered and equivalent
to x ∗ . By Lemma 4.1, ScalingFlow on input G can produce a solution x in such that x is
well-ordered and equivalent to x . At the end of the algorithm, the invariant in Section 3 must hold
for the solution x . By Theorem 3.1,
X
X
w e x e ≥ (1 − 8ǫ)
w e x ∗e .
e∈G
e∈G
Since x is equivalent to x and the step function defined by {w ei , c ei } is always below the function
w e,
Z xe
X
w e (t)dt ≥
fe (x e ) =
w ex e.
0
e∈G
Notice that the step function defined by {w ei + δT , c ei }, however, is always above the function w e .
Since x ∗ is equivalent to x ∗ ,
Z x ∗e
X
∗
(w e + δT ) x e ≥
w e (t)dt = fe (x ∗e ).
0
e∈G
Since δT = ǫw min ≤ ǫw e for all e ∈ G, we have w e + δT ≤ w e (1 + ǫ). It follows that
X
w e x ∗e ≥ fe (x ∗e ).
(1 + ǫ)
e∈G
17
Putting everything together,
fe (x e ) ≥
≥
X
e∈G
w e x e ≥ (1 − 8ǫ)
X
w e x ∗e
e∈G
1 − 8ǫ
fe (x ∗e ) ≥ (1 − 9ǫ)fe (x ∗e ).
1+ǫ
Finally, by applying a similar argument as in Theorem3.1, it is easy to see that the running time
of ConcaveFlow is O Dmǫ−1 log(w max /w min ) log n .
5
Discussion
We extended the Duan-Pettie framework for approximating weighted matchings [DP14] to flows on
small depth networks [Coh95]. Our approach can be viewed as viewing the length of augmenting
paths as a regularizing parameter on the weights of augmenting paths. For general concave weight
functions, the routine can also be viewed as a dynamical system that adjusts flows based on the
gradient of the current weight functions [BMV12, BBD+ 13, SV16]. Also, we believe our approach
can be parallelized, and extends to non-bipartite graphs: here the challenge is mostly with finding
an approximate extension of blocking flow routines.
Another related application with small depth networks is finding high density subgraphs in
networks [GT15, MPP+ 15, BHNT15]. These problems can be solved using parameteric maximum
flows or binary searching over maximum flows [Gol84]. In both of these cases, the networks have
constant depth, but the reductions also change the notions of approximation. Weights represents
an additional tool in such routines, and we plan to further investigate its applications.
References
[AG14]
Kook Jin Ahn and Sudipto Guha. Near linear time approximation schemes for uncapacitated and capacitated b-matching problems in nonbipartite graphs. In Proceedings
of the Twenty-fifth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ’14,
pages 239–258, 2014. Available at: https://arxiv.org/abs/1307.4355.
[AMO93]
Ravindra K. Ahuja, Thomas L. Magnanti, and James B. Orlin. Network Flows: Theory,
Algorithms, and Applications. Prentice Hall, Englewood Cliffs, NJ, 1993.
[AMOR95] Ravindra K Ahuja, Thomas L Magnanti, James B Orlin, and MR Reddy. Applications of network optimization. Handbooks in Operations Research and Management
Science, 7:1–83, 1995. Available at: http://jorlin.scripts.mit.edu/docs/publications/52applications%20of%20network.pdf.
[BBD+ 13] Luca Becchetti, Vincenzo Bonifaci, Michael Dirnberger, Andreas Karrenbauer,
and Kurt Mehlhorn.
Physarum can compute shortest paths: Convergence
proofs and complexity bounds.
In Automata, Languages, and Programming
- 40th International Colloquium, ICALP 2013, Riga, Latvia, July 8-12, 2013,
Proceedings, Part II, pages 472–483, 2013.
Available at https://people.mpiinf.mpg.de/˜mtd/files/physarum shortest path.pdf.
18
[BHNT15] Sayan Bhattacharya, Monika Henzinger, Danupon Nanongkai, and Charalampos
Tsourakakis. Space-and time-efficient algorithm for maintaining dense subgraphs on
one-pass dynamic streams. In Proceedings of the Forty-Seventh Annual ACM on Symposium on Theory of Computing, pages 173–182. ACM, 2015.
[BMV12]
Vincenzo Bonifaci, Kurt Mehlhorn, and Girish Varma. Physarum can compute shortest
paths. In Proceedings of the Twenty-Third Annual ACM-SIAM Symposium on Discrete
Algorithms, SODA 2012, Kyoto, Japan, January 17-19, 2012, pages 233–240, 2012.
Available at http://arxiv.org/abs/1106.0423.
[BV04]
S. Boyd and L. Vandenberghe. Convex Optimization. Camebridge University Press,
2004. Available at https://web.stanford.edu/˜boyd/cvxbook/bv cvxbook.pdf.
[CKM+ 11] Paul Christiano, Jonathan A. Kelner, Aleksander Madry, Daniel A. Spielman, and
Shang-Hua Teng. Electrical flows, Laplacian systems, and faster approximation of
maximum flow in undirected graphs. In Proceedings of the 43rd ACM Symposium on
Theory of Computing, STOC 2011, San Jose, CA, USA, 6-8 June 2011, pages 273–282,
2011. Available at http://arxiv.org/abs/1010.2921.
[Coh95]
Edith Cohen. Approximate max-flow on small depth networks. SIAM J. Comput.,
24(3):579–597, June 1995.
[DJ12]
N. Devanur and K. Jain. Online matching with concave returns. In Proceedings of the
44th ACM Symposium on Theory of Computing, STOC, 2012.
[DP14]
Ran Duan and Seth Pettie. Linear-time approximation for maximum weight matching.
Journal of the ACM (JACM), 61(1):1, 2014.
[Gab17]
Harold N. Gabow. The weighted matching approach to maximum cardinality matching.
CoRR, abs/1703.03998, 2017. Available at: http://arxiv.org/abs/1703.03998.
[Gol84]
A. V. Goldberg. Finding a maximum density subgraph. Technical report, Berkeley,
CA, USA, 1984.
[GR98]
Andrew V. Goldberg and Satish Rao. Beyond the flow decomposition barrier. J. ACM,
45:783–797, September 1998.
[GT87]
Andrew V. Goldberg and Robert Endre Tarjan. Solving minimum-cost flow problems
by successive approximation. In Proceedings of the nineteenth annual ACM symposium
on Theory of computing, 1987.
[GT89]
Harold N. Gabow and Robert Endre Tarjan. Faster scaling algorithms for network
problems. SIAM J. Comput., 18(5):1013–1036, 1989.
[GT91]
Harold N. Gabow and Robert Endre Tarjan. Faster scaling algorithms for general
graph-matching problems. J. ACM, 38(4):815–853, 1991.
[GT15]
Aristides Gionis and Charalampos E. Tsourakakis. Dense subgraph discovery: Kdd
2015 tutorial. In Proceedings of the 21th ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining, KDD ’15, pages 2313–2314, New York, NY,
USA, 2015. ACM.
19
[Hoc07]
Dorit S Hochbaum. Complexity and algorithms for nonlinear optimization problems.
Annals of Operations Research, 153(1):257–296, 2007.
[Kar72]
Richard M. Karp. Reducibility among combinatorial problems. In R. E. Miller and J. W.
Thatcher, editors, Complexity of Computer Computations, pages 85–103. Plenum Press,
1972.
[KMVV15] Ravi Kumar, Benjamin Moseley, Sergei Vassilvitskii, and Andrea Vattani. Fast greedy
algorithms in mapreduce and streaming. ACM Transactions on Parallel Computing,
2(3):14, 2015.
[LMSV11] Silvio Lattanzi, Benjamin Moseley, Siddharth Suri, and Sergei Vassilvitskii. Filtering:
a method for solving graph problems in mapreduce. In Proceedings of the twenty-third
annual ACM symposium on Parallelism in algorithms and architectures, pages 85–94.
ACM, 2011.
[LS14]
Yin Tat Lee and Aaron Sidford.
Path finding methods for linear programming: Solv√
ing linear programs in Õ( rank) iterations and faster algorithms for maximum flow.
In Foundations of Computer Science (FOCS), 2014 IEEE 55th Annual Symposium
on, pages 424–433. IEEE, 2014. Available at http://arxiv.org/abs/1312.6677 and
http://arxiv.org/abs/1312.6713.
[Mad13]
Aleksander Madry. Navigating central path with electrical flows: From flows to matchings, and back. In Foundations of Computer Science (FOCS), 2013 IEEE 54th Annual
Symposium on, pages 253–262. IEEE, 2013.
[MAG+ 13] Faraz Makari Manshadi, Baruch Awerbuch, Rainer Gemulla, Rohit Khandekar, Julián
Mestre, and Mauro Sozio. A distributed algorithm for large-scale generalized matching.
Proceedings of the VLDB Endowment, 6(9):613–624, 2013.
[MPP+ 15] Michael Mitzenmacher, Jakub Pachocki, Richard Peng, Charalampos Tsourakakis, and
Shen Chen Xu. Scalable large near-clique detection in large-scale networks via sampling. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, pages 815–824. ACM, 2015.
[Orl93]
James B Orlin. A faster strongly polynomial minimum cost flow algorithm. Operations
research, 41(2):338–350, 1993.
[Orl13]
James B Orlin. Max flows in o(nm) time, or better. In Proceedings of the forty-fifth
annual ACM symposium on Theory of computing, pages 765–774. ACM, 2013.
[Pen16]
Richard Peng. Approximate undirected maximum flows in o (m polylog (n)) time. In
Proceedings of the Twenty-Seventh Annual ACM-SIAM Symposium on Discrete Algorithms, pages 1862–1867. SIAM, 2016. Available at http://arxiv.org/abs/1411.7631.
[Sch03]
A. Schrijver. Combinatorial Optimization. Springer-Verlag, 2003.
[ST83]
Daniel D. Sleator and Robert Endre Tarjan. A data structure for dynamic trees. J.
Comput. Syst. Sci., 26(3):362–391, June 1983.
20
[SV16]
Damian Straszak and Nisheeth K. Vishnoi. Natural algorithms for flow problems. In
Proceedings of the Twenty-Seventh Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2016, Arlington, VA, USA, January 10-12, 2016, pages 1868–1883,
2016.
[TTL+ 12] Wenbin Tang, Jie Tang, Tao Lei, Chenhao Tan, Bo Gao, and Tian Li. On optimization
of expertise matching with various constraints. Neurocomputing, 76(1):71–83, 2012.
[Vég12]
László A Végh. Strongly polynomial algorithm for a class of minimum-cost flow problems with separable convex objectives. In Proceedings of the forty-fourth annual ACM
symposium on Theory of computing, pages 27–40. ACM, 2012.
[VVS10]
Erik Vee, Sergei Vassilvitskii, and Jayavel Shanmugasundaram. Optimal online assignment with forecasts. In Proceedings of the 11th ACM conference on Electronic
commerce, pages 109–118. ACM, 2010.
[WRM15] Di Wang, Satish Rao, and Michael W. Mahoney. Unified acceleration method for
packing and covering problems via diameter reduction. CoRR, abs/1508.02439, 2015.
[ZO15]
A
Zeyuan Allen Zhu and Lorenzo Orecchia. Nearly-linear time positive LP solver with
faster convergence rate. In Proceedings of the Forty-Seventh Annual ACM on Symposium on Theory of Computing, STOC 2015, Portland, OR, USA, June 14-17, 2015,
pages 229–236, 2015.
Applications
We now outline some concrete applications of our formulation on small depth networks: many of
them are discussed in more details in the survey by Ahuja et al. [AMOR95].
A.1
Assignment Problems
Our techniques are directly motivated by those developed for approximate max weight matchings [DP14]. This problem, when phrased as the assignment problem, has applications in matching
objects across snapshots as well as direct uses in assignment indicated by its name. Names of some
of these applications, as enumerated in Application 16-18 of [AMOR95], include locating objects
in space, matching moving objects, and rewiring typewriters.
The problem of finding matchings in large scale graphs has received much attention in experimental algorithm design [LMSV11, MAG+ 13, KMVV15]. These instances often involve capacities
on both sides, as well as edges in between. They are related to b-matchings, and have been studied
as generalized matchings. Previous works that incorporated concave costs include:
• [VVS10] studied Adwords matching using concave objective functions, and showed several
stability and robustness properties of these formulations.
• [TTL+ 12] utilized concave networks to perform assignments of experts to queries with additional constraints on groups of experts as well as queries. This leads to a depth 4 network
with 3 layers corresponding to node groups respectively.
21
• [DJ12] considered a generalization of the well-known Adwords problem, which asks for maximizing the sum of budgeted linear functions of agents. In the generalization, the utility of
each agents is a concave function (of her choice) of her total allocation. The objective is to
maximize the sum of budgeted utilities, and an optimal online algorithm is given for this
generalization.
Another instance of concave cost matching is matrix balancing (Application 28 in [AMOR95]):
given a fractional matrix, we would like to find a matrix with a specified row /column sum while
minimizing distance under some norm. Here a bipartite network can be constructed by turning
the rows into left vertices, columns into right vertices. The concave functions can then go onto the
edges as costs.
Our algorithm is able to handle such generalizations, and lead to bounds similar to those from
recent improvements to positive packing linear programs [ZO15].
Generalizations of matching and assignment lead naturally to small depth networks. One such
generalization is 3-Dimensional matching which is NP-hard [Kar72]. Here the generalization of
edges is 3-tuples of the form (x, y, z), with one element each from sets X, Y , and Z, and 3dimensional matching asks for the maximum weighted set of tuples that uses each element in X,
Y , and Z at most once.
A restriction of this problem on the other hand is in P : the set of acceptable tuples are given
by pairs of edges between XY and Y Z. Specifically, a tuple (x, y, z) can be used iff xy ∈ EXY
and yz ∈ EY Z , and its cost is given by the sum of the two edges. This version can be solved by
maximum cost flow on a depth 5 network:
• For each element x, y or z, create an in and out vertex, connected by an edge with unit
capacity.
• Connect Xout to Yin and Yout to Zin with edges corresponding to EXY and EY Z .
• Connect s to Xin and Zout to t.
A.2
Mincost Flow
Minimum cost flow asks to minimize the cost instead of maximizing it. This is perhaps the most
well-known formulation of our problem. However, it doesn’t tolerate approximations very well: if
we stay with non-negative costs, we need to specify the amount of flow. If this is done exactly, it
becomes flow feasiblity on a small depth network. There are known reductions of directed maximum
flows to this instance (e.g. [Mad13], Section 3).
One way to limit the total amount of flow is to associate a reward with each unit sent from s
to t. Specifically, let q e > 0 be the cost of one unit of flow on e, and Q > 0 be the reward of one
unit sent from s to t, the objective function is
X
min
q e x e − Qf
e
where f is the total amount of flow from s to t. In this case, negating the cost brings us back to
the maximum weight flow problem. The approximation factor on the other hand becomes additive:
22
Lemma A.1. Let (x ∗ , f ∗ ) be the optimal solution to the above mincost flow problem, The algorithm
ScalingFlow in Section 3 gives a solution (x , f ) such that:
!
X
X
q e x ∗e − Qf ∗ + 2ǫQf ∗
q e x e − Qf ≤ (1 + ǫ)
e
e
A proof of Lemma A.1 can be found in Appendix B. Applications of minimum cost flows that
lead to small depth networks include:
• Job scheduling, or optimal loading of a hopping airplane (application 13 from [AMOR95]):
When phrased as scheduling, this can be viewed as completing a maximum set of jobs each
taking place between days [si , ti ] and giving gain ci so that at no point more than ui jobs are
being done simultaneously on day i.
This can be solved with a maximum cost flow whose depth is the number of days:
– The days become a path, with capacity ui ,
– For each job, create a node (connected to source with capacity 1) with two edges, one
with cost 0 to ti , and one with cost ci to si .
– Each node on the path has an edge to sink, t, with capacity equalling the number of
jobs ending that day.
This leads to a network whose depth is the number of days, or number of stops. Both of these
quantities are small in real instances.
• Balancing of schools (Application 15 from [AMOR95]): the goal is to assign students from
varying backgrounds to schools, belong to several districts, so that each school as well as each
district has students with balanced backgrounds. Furthermore, the cost corresponds to the
distance the student needs to travel to the schools. This leads to a depth 3 network where
the layers are the students, schools, and districts respectively.
Finally, we believe several applications can benefit from the incorporation of non-linear weight
functions. As an instance, we can solve the following problem using our framework: Let G be a
network with k sources, {si , 1 ≤ P
i ≤ k}, and a sink, t, and let fi be a concave function for each
1 ≤ i ≤ k. We wish to maximize i fi (Fi ), where Fi is the amount of flow sent from si to t.
B
Negative Weight Functions
We consider the case where our network can have negative weights. For simplicity, we only study
linear weight functions. The analysis, however, can be extended to the case of concave function as
in Section 4. Here, we assume that the absolute values of the weights are bounded:
w max ≥ |w e | ≥ w min
∀e.
Lemma B.1. Let x be a solution returned by ScalingFlow, and let x ∗ be the optimal solution
X
X
X
w e x e ≥ (1 − 8ǫ)
w e x ∗e + (1 + 8ǫ)
w e x ∗e .
e
w e >0
w e <0
23
Proof. A key observation is that the Lemmas 3.7 and 3.8 are robust under negative weights. Using
an argument similar to as in Theorem 3.1, we have:
1. If x e < x ∗e , then x e < c e , and the invariant gives w pe ≤ 2δT , which implies
w pe x e ≥ (w pe − 2δT ) x e ≥ (w pe − 2δT ) x ∗e .
Since δ(T ) ≤ ǫw min ≤ ǫ|w e | for all e, we consider 2 cases:
(
w pe x ∗e − 2ǫw e x ∗e
w pe x e ≥ (w pe − 2δT ) x ∗e ≥
w pe x ∗e + 2ǫw e x ∗e
if w e > 0,
if w e < 0.
2. If x e > x ∗e , then x e > 0, and the invariant gives w pe ≥ −3le δef , which combined with x ∗e ≤ x e
gives:
x e w pe + 3le δef ≥ x ∗e w pe + 3le δef ≥ x ∗e w pe
Combining these then gives
X
X
x e w pe + 3le δef
x e w e + 3le δef =
e
e
≥
≥
X
e
w pe x ∗e −
X
e:w e >0
X
2ǫw e x ∗e +
e:w e <0
= (1 − 2ǫ)
w e x ∗e
e:w e >0
2ǫw e x ∗e
e:w e <0
(w e x ∗e − 2ǫw e x ∗e ) +
X
X
X
(w e x ∗e + 2ǫw e x ∗e )
e:w e <0
+ (1 + 2ǫ)
X
w e x ∗e
e:w e <0
and the result then follows from Lemma 3.8.
Proof. (Of Lemma A.1) Let w e = −q e for every e, and add a directed edge from t to a dummy
vertex t′ with w tt′ = Q. By Lemma B.1, a solution returned by ScalingFlow on that instance
has the following guarantee:
X
X
Qf −
q e x e ≥ (1 − ǫ)Qf ∗ − (1 + ǫ)
q e x ∗e .
e
Negating both sides gives
X
e
q e x e − Qf ≤ (1 + ǫ)
= (1 + ǫ)
X
q e x ∗e − (1 − ǫ)Qf ∗
X
24
q e x ∗e − Qf ∗ + 2ǫQf ∗ .
| 8cs.DS
|
Comparing global input-output behavior of
frozen-equivalent LPV state-space models
Ziad Alkhoury ∗,∗∗ Mihály Petreczky ∗∗∗ Guillaume Mercère ∗
∗
arXiv:1703.03679v1 [cs.SY] 10 Mar 2017
University of Poitiers, Laboratoire d’Informatique et d’Automatique
pour les Systèmes, Bâtiment B25, 2 rue Pierre Brousse, TSA 41105,
86073 Poitiers Cedex 9. (e-mails: [email protected],
[email protected])
∗∗
Ecole des Mines de Douai, F-59500 Douai, France.
∗∗∗
CNRS, Centrale Lille, UMR 9189 - CRIStAL-Centre de Recherche
en Informatique, Signal et Automatique de Lille, F-59000 Lille,
France. (e-mail: [email protected])
Abstract: It is known that in general, frozen equivalent (Linear Parameter-Varying) LPV
models, i.e., LPV models which have the same input-output behavior for each constant
scheduling signal, might exhibit different input-output behavior for non-constant scheduling
signals. In this paper, we provide an analytic error bound on the difference between the inputoutput behaviors of two LPV models which are frozen equivalent. This error bound turns out to
be a function of both the speed of the change of the scheduling signal and the discrepancy
between the coherent bases of the two LPV models. In particular, the difference between
the outputs of the two models can be made arbitrarily small by choosing a scheduling signal
which changes slowly enough. An illustrative example is presented to show that the choice of
the scheduling signal can reduce the difference between the input-output behaviors of frozenequivalent LPV models.
Keywords: LPV system identification, local approach, interpolation, modeling error.
1. INTRODUCTION
As shown, e.g., in Tóth (2010); Lopes dos Santos et al.
(2011), existing techniques dedicated to the identification
of Linear Parameter-Varying (LPV) systems can be split
up into two main families: the local and the global approach. On the one hand, the global approach focuses
on a global procedure for which it is assumed that one
global experiment can be performed in which the control
inputs as well as the scheduling variables can be both
excited (see, e.g., Bamieh and Giarré (2002); Felici et al.
(2007)). By construction, these techniques are restricted to
specific systems where the components of the scheduling
variables are totally controllable and excitable (and not
only measurable). On the other hand, the local approach
is based on a multi-step procedure (see, e.g.,Vizer and
Mercère (2014), Vizer et al. (2013b), Lovera and Mercère
(2007)) where first, frozen linear time-invariant models
are estimated for constant values of the the scheduling
variables, then the global LPV model is built from the interpolation of the local LTI models. A substantial amount
of local techniques are available in the literature (see
De Caigny et al. (2011), Luspay et al. (2011)). This is
probably due to the fact that the identification techniques
for LTI systems are more mature than the LPV ones, in
addition to the development of algorithms to optimize the
number of local operating points involved in the procedure
Vizer and Mercère (2014), as well as their strong link
1 This work was partially supported by ESTIREZ project of Region
Nord-Pas de Calais, France.
with the gain scheduling procedure Rugh and Shamma
(2000). However, it is clear that none of these techniques
is consistent when black-box state-space LPV models are
estimated. As pointed out first in Tóth et al. (2007), the
main reasons why the current solutions fail to mimic the
global behavior of the LPV system are due to (i) the
lack of information about the dynamical behavior of the
scheduling variables p in the frozen models when LPV
models depending on p dynamically are handled, (ii) the
difficulty to express the local LTI models w.r.t. a coherent
basis Tóth (2010) before the interpolation. While solutions
have been suggested to bypass these difficulties when graybox state-space LPV models are considered Vizer et al.
(2013b,a, 2015), restricting slow variations of p (w.r.t. the
dynamics of the system) is the only assumption to make in
order to guarantee a reliable global model behavior when
black-box models are handled, i.e., when the only source
of information is restricted to measured data, see Khalate
et al. (2009).
In this paper, we present an analytical error bound on
the difference between the global input-output behavior
of two frozen equivalent LPV models. By global inputoutput behavior we mean the input-output behavior along
all scheduling signals, while by frozen equivalence of two
LPV models we mean that their input-output behavior
is the same along all constant scheduling signals. As it
was discussed above, even if the data were generated by
an LPV model, local methods for system identification
in general result in LPV models which are only frozen
equivalent to the ‘true’ one. Hence, the error bound of
this paper allow us
(1) to evaluate the difference between the input-output
behavior LPV model obtained by local identification
methods, and the ‘true’ LPV model which generated
the data,
(2) to evaluate the difference between the input-output
behavior of LPV models obtained by different local
identification methods.
The error bound includes the speed of the change of the
scheduling signal, the minimal time the scheduling signal
remains constant (can be 1 time step in general) and a
measure of inconsistency of the bases of the frozen LTI
models. Note that this bound is valid only for (quadratically) stable LPV models. It turns out that even if the
frozen LTI models are not in a coherent basis, the error
between the input-output behaviors of the corresponding
LPV models can be kept arbitrary small by using sufficiently slow scheduling signals, or by keeping the scheduling signal constant for long enough. This allows us to
conclude that the problem of choosing the ‘correct’ state
transformation required for the interpolation step becomes
less critical for slowly varying scheduling signals.
While the results of the paper seem to be intuitive, they
have never been presented formally, to the best of our
knowledge. A precise error bound could be of great interest for system identification and control design for LPV
models, because it allows the user to quantify the effect
of the modeling error due to local identification on the
system performance. In turn, this could be incorporated as
an additional constraint in control synthesis, or used as a
minimization criteria in parametric system identification.
Outline of the paper. In Section 2, we present the basic
notions to set up the framework of the paper. Next, in
Section 3, we present the main results. Section 4 presents
an illustrative example. Finally, Section 5 concludes the
paper. The proofs of the results are presented in Appendix.
2. PROBLEM FORMULATION
To set up the framework of the paper, we give some basic
definitions for LTI and LPV models.
Definition 1. (LTI models). A discrete-time Linear TimeInvariant (LTI) model is defined as follows
x(t + 1) = Ax(t) + Bu(t),
L
(1)
y(t)
= Cx(t).
x(t) ∈ X = Rnx is the state at time t, y(t) ∈ Y = Rny is
the output at time t, u(t) ∈ U = Rnu is the input at time
t, and A ∈ Rnx ×nx , B ∈ Rnx ×nu and C ∈ Rny ×nx In the
sequel, we will use the notation
L = (A, B, C)
(2)
for an LTI model of the form (1).
Definition 2. (LPV models). A discrete-time Linear Parameter Varying (LPV) model (Σ) is defined as follows
x(t + 1) = A(p(t))x(t) + B(p(t))u(t),
Σ
(3)
y(t)
= C(p(t))x(t).
x(t) ∈ X = Rnx is the state at time t, y(t) ∈ Y = Rny
is the output at time t, u(t) ∈ U = Rnu is the input at t
and p(t) ∈ P is the value of the scheduling signal at time
t. The matrix functions A : P 7→ Rnx ×nx , B : P 7→ Rnx ×nu
and C : P 7→ Rny ×nx are assumed to be continuous. We
assume that P is compact, i.e., closed and bounded set. In
the sequel, we will use the short notation
Σ = (A(.), B(.), C(.))
(4)
to denote a model of the form (3). Notice that we use
the notation A(.), B(.) and C(.) to emphasize that these
matrices are maps, not constant matrices.
Remark 1. In order to avoid notational ambiguity, we
denote that matrices of the LTI models by bold symbols,
i.e., (A, B, C), while we denote that matrices of the LPV
model by (A(.), B(.), C(.)).
From now on, we will use the following notation: for a set
A, we denote by AN the set of all functions of the form
φ : N → A. An element of AN can be thought of as a signal
in discrete-time. By a solution of L (respectively Σ), we
mean a tuple of trajectories (x, y, u) ∈ (X , Y, U) satisfying
(1), (respectively (x, y, u, p) ∈ (X , Y, U, P) satisfying (3)),
for all k ∈ N, where X = XN , Y = YN , U = UN , P = PN .
Note that P denotes the set of scheduling signals, while
P denotes the set of all possible values of the scheduling
parameters. That is, an element of P is a sequence, while
an element of P is a vector. The same remark holds
for X , Y, U and X, Y, U respectively. Note that in the
sequel, unless stated otherwise, we will use p to denote
a scheduling signal, and we use p to denote a potential
value of the scheduling variable, i.e., p ∈ P is a sequence,
but p ∈ P is a vector of constant values.
Definition 3. (Input-output map of LPV models).
Define the functions
YΣ : U × P 7→ Y,
(5)
for an LPV model Σ, such that for any (x, y, u, p) ∈ (X ×
Y × U × P), y = YΣ (u, p) hold if and only if (x, y, u, p)
satisfy (3), and x(0) = 0. The function YΣ is called the
input-output map of Σ.
Definition 4. (Frozen model of Σ at p). A frozen model
of an LPV Σ = (A(.), B(.), C(.)) at p, denoted by LΣ (p),
is the LTI model defined by
LΣ (p) = (A(p), B(p), C(p)).
(6)
Definition 5. (Frozen-minimal LPV models).
An LPV model Σ = (A(.), B(.), C(.)) is frozen-minimal, if
∀p ∈ P, the LTI LΣ (p) = (A(p), B(p), C(p)) is minimal 2 .
In this paper, all models have the same state, input, output
and scheduling signal dimensions. Therefore, the matrices
and vectors dimensions nx , np , ny and nu will be omitted
when clear from the context.
Definition 6. (Frozen-equivalent LPV models at p).
We say that two LPV models Σ = (A(.), B(.), C(.)) and
Σ̂ = (Â(.), B̂(.), Ĉ(.)) are frozen equivalent, if ∀p ∈ P
the LTI models LΣ (p) and LΣ̂ (p) have the same transfer
function, i.e., for all s ∈ C which is not an eigenvalue of
A(p) or Â(p), C(p)(sInx − A(p))−1 B(p) = Ĉ(p)(sInx −
Â(p))−1 B̂(p).
That is, if Σ and Σ̂ are frozen input-output equivalent
LPV models, then, for any p ∈ P such that p is constant
2
According to the classical definition, i.e., (A(p), B(p)) is controllable and (A(p), C(p)) is observable, (see Rugh (1996)).
(i.e., p(0) = p(k), ∀k ∈ N), and for any u ∈ U, YΣ (u, p) =
YΣ̂ (u, p).
sup kB(p)k = KB ,
(9)
p∈P
sup kC(p)k = KC ,
(10)
p∈P
3. MAIN RESULTS
sup kTΣ̂,Σ (p)k = KT .
In this section, the difference between the outputs of
two frozen-minimal and frozen-equivalent LPV models is
characterized.
It is important to note that KB , KC and KT are finite.
This follows directly from the fact that each of the maps
B, C and TΣ̂,Σ are continuous, and P is a compact set.
3.1 Preliminaries
Now, for every scheduling signal p ∈ P, define
KM (p) = sup kInx − Mp(k),p(k+1) k.
(11)
p∈P
(12)
k∈N
Then,
In order to compare the global behavior of the LPV models
Σ and Σ̂, we need to introduce some concepts. We start
from the fact that there is a unique isomorphism between
any two input-output equivalent minimal LTI models (i.e.,
they have the same transfer function) L and L̂ (see Rugh
(1996)). This means that there is a unique nonsingular
nx × nx matrix T that fulfills
T AT −1 = Â,
T B = B̂,
CT −1 = Ĉ.
(7)
where {A, B, C} and {Â, B̂, Ĉ} are the state-space matrices of L and L̂ respectively. We say that T is an isomorphism from L to L̂.
Now, let us assume that LΣ (p) = (A(p), B(p), C(p)) and
LΣ̂ (p) = (Â(p), B̂(p), Ĉ(p)) are two frozen LTI models
of Σ and Σ̂, respectively, at p ∈ P. We assume that
Σ and Σ̂ are both frozen-equivalent and frozen-minimal
LPV models, then LΣ (p) and LΣ̂ (p) are input-output
equivalent and both are minimal for all p ∈ P. Hence,
there exists a (unique) isomorphism T (p) from LΣ̂ (p) to
LΣ (p) for all p ∈ P. Let us call the map TΣ̂,Σ : p → T (p),
the frozen isomorphism map from Σ̂ to Σ.
Lemma 1. Let Σ, Σ̂ be two frozen-minimal and frozenequivalent LPV models .If A, B, C, Â, B̂ and Ĉ are continuous maps, then TΣ̂,Σ is a continuous map as well.
The proof of Lemma 1 can be found in Appendix A.1
Remark 2. It is important to note that TΣ̂,Σ is not an
LPV isomorphism between the two LPV models. In order
for TΣ̂,Σ to be an LPV isomorphism, A(p(t)) = TΣ̂,Σ (p(t +
−1
1))Â(p(t))TΣ̂,Σ
(p(t)), should hold for any scheduling signal
p ∈ P, see Kulcsár and Tóth (2011); Tóth (2010) for more
details.
In the rest of the paper, we will need the following
notations.
Notation 1. In the sequel, kAk will denote the induced 2Norm of any matrix A, while for a vector p, kpk will denote
the Euclidean vector norm of p, (see Meyer (2000)). We
denote by Inx the nx × nx identity matrix.
To proceed with the characterization of the global behavior
equivalence, let us define
∀p1 , p2 ∈ P : Mp1 ,p2 = TΣ̂,Σ (p1 )TΣ̂,Σ (p2 )−1 .
(8)
Notice that, if T (p1 ) = T (p2 ), then, Mp1 ,p2 = Inx . Define
the coefficients KB , KC and KT as follows,
KM (p) ≤
sup kInx − Mp1 ,p2 k < +∞.
(13)
p1 ,p2 ∈P
Notice that
sup kInx − Mp1 ,p2 k exists because TΣ̂,Σ is
p1 ,p2 ∈P
continuous, and hence P × P 3 (p1 , p2 ) 7→ Mp1 ,p2 is
continuous and defined on a compact set. This means that,
∃KM > 0, such that
∀p ∈ P : KM (p) ≤ KM .
(14)
Let us also assume that Σ and Σ̂, are quadratically stable,
i.e., there exists a positive definite matrix P > 0, such
that for all p ∈ P, A(p)T P A(p) − P < 0 and for Â(p)
similarly.
Lemma 2. If Σ is quadratically stable, then there exists
a nonsingular S ∈ Rnx ×nx , such that the LPV
Σ̄ = (SA(.)S −1 , SB(.), C(.)S −1 )
satisfies
sup kSA(p)S −1 k < α < 1,
p∈P
for some α ∈ (0, 1).
In the sequel, we will assume that
sup kA(p)k ≤ α < 1,
(15)
p∈P
sup kÂ(p)k ≤ α̂ < 1.
(16)
p∈P
If it is not the case, then we can replace Σ by Σ̄. Note that
α in (15) is an upper bound on the Lyapunov exponent of
dynamical system x(t + 1) = A(p(t))x(t); the smaller α is,
the faster the solutions of x(t + 1) = A(p(t))x(t) converge
to zero.
If Σ̂ satisfies (15), then it follows from (Rugh, 1996, Proof
of Lemma 27.4 and Theorem 27.2) and (9) – (10) that
there exists µ1 > 0 such that for any solution (x̂, ŷ, u, p)
of Σ̂ such that x(0) = 0 and supt∈N ku(t)k < +∞,
sup kx̂(t)k ≤ µ1 sup ku(t)k.
(17)
t
t∈N
From now on, when comparing the outputs of two LPV
models, we will assume that they are subject to the same
bounded input from the set
l∞ (U) = {u ∈ U : kukl∞ = sup ku(t)k < +∞}.
t∈N
Moreover, we will consider piecewise-constant scheduling
signals from the following set.
S∆={p ∈ P : ∀k ∈ N,∀i ∈ {0, 1, · · · , ∆ − 1} :
p(k∆ + i)=p(k∆)},
(18)
3.2 General error bound
We are now ready to present our main result on the
global input-output equivalence of frozen-equivalent LPV
models.
Theorem 1. (Difference characterization). Let Σ and Σ̂
be two frozen-minimal, frozen-equivalent and quadratically stable LPV models. Then, for all p ∈ S∆ , ∆ > 0
and u ∈ l∞ (U), t ∈ N,
kYΣ (u, p)(t) − YΣ̂ (u, p)(t)k
≤ g(∆, KM (p), t mod ∆) · kukl∞
≤ g(∆, KM , t mod ∆).kukl∞ .
(19)
where t mod ∆ denotes the remainder of dividing t by ∆,
and for all i = 0, . . . , ∆ − 1,
αi
g(∆, K, i) =
K(αKT µ1 + KB )KC ,
(20)
1 − α∆
and KB , KC , KT , KM (p), α and µ1 are defined in equations (9), (10), (11), (12), (15) and (17) respectively.
The proof of Theorem 1 can be found in Appendix A.3.
Eq. (19) gives a conservative bound on the output error of
two LPV systems whose frozen transfer functions are the
same. It clearly shows that the error is affected by each of
the following:
• The length of the interval ∆, on which the scheduling
signal p is constant.
• The Lyapunov exponent α of the system, which
expresses the degree of stability.
• The distance KM (p) which expresses the degree of
inconsistency of the basis of the frozen models: if
all the frozen models are in a consistent basis, then
KM (p) is zero.
• The amplitude of the input kukl∞ . Note that (19)
remains true if instead of kukl∞ we take kukl2 =
P
1
( k=0 ku(k)k22 ) 2 .
3.3 Special cases
In this subsection, we formally derive and present important results based on Theorem 1, i.e., which aim at
showing the effects of different variables on the difference
of the input-output map presented in Equations (19) and
(20) from Subsection 3.2.
Theorem 2. (Switching interval). Let Σ and Σ̂ be two
frozen-minimal, frozen-equivalent and quadratically stable
LPV models. Then, for all ε > 0 there exists ∆m > 0 such
that for all p ∈ S∆ , ∆ > ∆m , u ∈ l∞ (U), and for all t ∈ N
with t mod ∆ > ∆m ,
kYΣ (u, p)(t) − YΣ̂ (u, p)(t)k < ε.kukl∞ .
(21)
The proof of Theorem 2 can be found in Appendix A.4.
Theorem 2 tells us that if we have a piecewise constant
scheduling signal, and the duration of each constant piece
is large enough, then on each constant piece the outputs
can get arbitrarily close, no matter how bad we choose
the coordinates of the frozen models. In case of switching
sequences we call the length of the interval on which p is
constant the dwell time. For switched systems, the theorem
says that for large enough dwell times, the outputs of the
system can get arbitrary close towards the end of the active
period of each discrete mode.
Theorem 3. (Speed of change of scheduling parameter).
Let Σ and Σ̂ be two frozen-minimal, frozen-equivalent and
quadratically stable LPV models. Then, for every ε > 0,
there exists δm > 0 such that, for all p ∈ P satisfying
∀t ≥ 0 :|p(t + 1) − p(t)| < δm ,
(22)
and for all u ∈ l∞ (U), t ∈ N,
kYΣ (u, p)(t) − YΣ̂ (u, p)(t)k < ε.kukl∞ .
The proof of Theorem 3 can be found in Appendix A.5.
This result is particularly interesting, because the scheduling signal is no longer restricted to be a piece-wise constant
signal. This implies that we can still guarantee the inputoutput ‘approximate’ equivalence using a slowly changing
scheduling signal, and the result holds for any t.
However, if no conditions can be imposed on the scheduling
signal, it is of great interest to evaluate the effect of the
stability of the system on the difference between inputoutput maps of the two frozen-equivalent models. In the
next theorem, we study the mentioned effect.
Theorem 4. (Stability). Let Σ and Σ̂ be two frozenminimal, frozen-equivalent and quadratically stable LPV
models. Then, for all ε > 0, there exists αm > 0 such that,
if Σ satisfies
sup kA(p)k ≤ αm ,
(23)
p∈P
then, for all u ∈ l∞ (U), p ∈ P, t ∈ N,
kYΣ (u, p)(t) − YΣ̂ (u, p)(t)k < ε.kukl∞ .
(24)
The proof of Theorem 4 can be found in Appendix A.6.
3.4 Discussion: basic trade-offs
Notice that, from Eq. (19), we have three factors that play
important role in the error characterization.
(1) The switching interval of the scheduling signal. By
choosing a relatively long switching interval ∆, we
guarantee that the difference between the outputs of
the LPV models becomes small enough towards the
end of each constant interval.
(2) The rapidity of change of the scheduling signal, i.e.,
the difference between two consecutive values of the
scheduling variable, which is |p(t + 1) − p(t)|. This
case is of great interest when ∆ = 1, because the
scheduling variable will not be restricted to a piecewise constant signal any longer, and it will be allowed
to change at each time instant. By slowly changing
the scheduling signal, we can get relatively very
small difference between the input-output behaviors
of frozen-equivalent LPV models.
(3) The stability of the system. The more the system is
stable, the less it responds to the applied input, i.e.,
the less the difference between two LPV models is.
The results above actually show the following. On the
one hand, even if the identified LTI models are in a noncoherent state-space basis, the obtained state-space LPV
Model and Real system output signals
y(t); y^(t)
y(t)
y^(t)
2
0
jjy(t) ! y^(t)jj
0
10
20
30
40
50
60
70
4
2
1
y(t)
y^(t)
2
0
80
t
Output di,erence
0
jjy(t) ! y^(t)jj
y(t); y^(t)
Model and Real system output signals
4
10
20
30
40
50
60
70
80
60
70
80
60
70
80
t
Output di,erence
0.1
0.05
0
0
10
20
30
40
50
60
70
80
0
0
10
20
t
Scheduling signal
30
40
50
t
Scheduling signal
p(t)
0.4
p(t)
0.5
0
0.3
0.2
0
10
20
30
40
50
60
70
80
0
10
20
30
t
Fig. 2. Output difference for p(t) = 0.3 + 0.1sin(t/5).
Model and Real system output signals
y(t); y^(t)
model can be still an accurate approximation of the real
system provided that the scheduling signal is slow enough.
More precisely, using slow changing scheduling signal, or
using a piece-wise constant scheduling signal with long
switching intervals is enough to obtain satisfying results.
On the other hand, if the basis of the frozen state-space
representation is close to a consistent one, then, even for
‘faster’ scheduling signal the approximation error will be
small.
50
t
4
y(t)
y^(t)
2
0
0
jjy(t) ! y^(t)jj
Fig. 1. Output difference for p a piece-wise constant signal.
40
10
20
30
40
50
60
70
80
60
70
80
60
70
80
t
Output di,erence
0.2
0.1
0
0
10
20
30
40
50
t
Scheduling signal
Let us consider the LPV model Σ of the form (3), such
that nx = 2, np = 1, nu = ny = 1, and such that, for any
p ∈ P := [0.1, 0.4],
>
0 0.2p
1
1
A(p)=
, B(p)=
, C(p)=
.
0.2 p
1
1
In order to obtain an LPV model Σ̂, which is frozenequivalent to Σ, we perform an LPV identification process
which includes the following steps:
(1) Input-output data generation at each possible constant value of the scheduling signal.
(2) Black-box LTI system identification using a subspace
method. This step results in multiple LTI models, all
are obtained in the same canonical form.
(3) Linear interpolation of the matrices of the frozen
models with respect to the values of the scheduling
signal. This step gives us the matrices of the LPV
model Σ̂, which is frozen equivalent to the original
one.
For comparing the outputs of the original LPV model and
the identified one, we excite both LPV models Σ and Σ̂
using the same input u(t) = 1 and the same scheduling
signal p, which is different for each of the following figures.
In Fig 1, the scheduling signal is chosen to be a piecewise constant one as is clearly shown in the figure. Notice
that the difference between the outputs is non-zero at
the switching instances of p, i.e., at t = 11, 21, ..., 71.
Notice also that the difference decreases approaching zero
as we go away from the switching instances. This result is
expected based on Theorem 2. We can also see that the
difference between the outputs depends on the jumps in
the switching signal.
p(t)
0.4
4. ILLUSTRATIVE EXAMPLE
0.3
0.2
0
10
20
30
40
50
t
Fig. 3. Output difference for p(t) = 0.3 + 0.1sin(t/2).
In Fig. 2 and 3, the scheduling signal is chosen to be a
sinusoidal one. This allows us to illustrate the results of
Theorem 3. It is clear that the output difference of Σ
and Σ̂ becomes bigger as we increase the frequency of the
scheduling signal in Fig 3, in comparison to Fig 2.
It is important to mention that the interpolation step
is a crucial one, because the resulting LPV model is
assumed to be frozen-equivalent to the original model
(or the real system) at frozen values of p; otherwise
the difference between the output of Σ and Σ̂ does not
approach zero even if the interval on which the scheduling
signal is constant was very long. Instead, the difference
cannot become smaller than a certain value, which is
affected by the distance between the frozen models of
Σ and Σ̂. Studying the effect of having an upper bound
on the distance between the frozen models is left for
future work. In this example, a linear interpolation is
performed, as it is sufficient to give us the required
model. For different systems, higher complexity models
and interpolation methods might be required in order to
avoid getting frozen-non-equivalent LPV models.
5. CONCLUSIONS
In this paper, an analytical expression for the difference
between any two frozen-equivalent LPV models is presented. This difference could be seen as a modeling error of
LPV systems, and the introduced expression allows us to
evaluate the modeling error of LPV systems numerically,
and gives the controller designer a bound on the global
difference between frozen-equivalent LPV models. Future
work will be aimed at using this expression in an optimization criteria during the identification process, more
precisely, during the search for a coherent basis for the
identified frozen models, which helps in minimizing the
modeling error of LPV systems.
REFERENCES
Bamieh, B. and Giarré, L. (2002). Identification of linear
parameter varying models. International Journal of
Robust and Nonlinear Control, 12(9), 841–853.
De Caigny, J., Camino, J., and Swevers, J. (2011).
Interpolation-based modeling of MIMO LPV systems.
IEEE Transactions on Control Systems Technology,
19(1), 46–63.
Felici, F., van Wingerden, J.W., and Verhaegen, M. (2007).
Subspace identification of MIMO LPV systems using a
periodic scheduling sequence. Automatica, 43(10), 1684–
1697.
Khalate, A., Bombois, X., Tóth, R., and Babuska, R.
(2009). Optimal experimental design for LPV identification using a local approach. In Proceedings of the
IFAC Symposium on System Identification. Saint-malo,
France.
Kulcsár, B. and Tóth, R. (2011). On the similarity state
transformation for linear parameter-varying systems. In
Proceedings of the IFAC World Congress. Milano, Italy.
Lopes dos Santos, P., Azevedo Perdicolis, T., Novara, C.,
Ramos, J., and Rivera, D. (2011). Linear ParameterVarying System Identification: new developments and
trends. Advanced Series in Electrical and Computer
Engineering, World Scientific.
Lovera, M. and Mercère, G. (2007). Identification for
gain scheduling: a balanced subspace approach. In
Proceedings of the American Control Conference. New
York, USA.
Luspay, T., Kulcsár, B., van Wingerden, J.and Verhaegen,
M., and Bokor, J. (2011). Linear Parameter Varying
Identification of Freeway Traffic Models. IEEE Transactions on Control Systems Technology, 19(1), 31–45.
Meyer, C. (2000). Matrix analysis and applied linear
algebra. SIAM, Cambridge University Press.
Rugh, W. and Shamma, J. (2000). Research on gain
scheduling. Automatica, 36(10), 1401–1425.
Rugh, W.J. (1996). Linear System Theory (2Nd Ed.).
Prentice-Hall, Inc., Upper Saddle River, NJ, USA.
Tóth, R. (2010). Modeling and identification of linear
parameter-varying systems. Springer-Verlag, Berlin Heidelberg.
Tóth, R., Felici, F., Heuberger, P., and den Hof, P.V.
(2007). Discrete-time LPV I/O and state-space representations. differences of behavior and pitfalls of interpolation. In Proceedings of the European Control
Conference. Kos, Greece.
Vizer, D., Mercère, G., and Laroche, E. (2015). Gray-box
LPV model identification of a 2-DoF surgical robotic
manipulator by using an H∞ -norm-based local approach. In Proceedings of the IFAC Workshop on Linear
Parameter Varying Systems. Grenoble, France.
Vizer, D., Mercère, G., Prot, O., Laroche, E., and Lovera,
M. (2013a). Linear fractional LPV model identification
from local experiments: an H∞ -based optimization technique. In Proceedings of the Conference on Decision and
Control. Florence, Italy.
Vizer, D., Mercère, G., Prot, O., and Ramos, J. (2013b).
A local approach framework for black-box and graybox LPV system identification. In Proceedings of the
European Control Conference. Zurich, Switzerland.
Vizer, D. and Mercère, G. (2014). An H∞ -norm-based
approach for operating point selection and LPV model
identification from local experiments. Periodica Polytechnica Electrical Engineering and Computer Science,
58(3), 121–131.
Appendix A. PROOFS
A.1 Proof of Lemma 1
Note that for all p ∈ P, TΣ̂,Σ (p) isomorphism satisfies
TΣ̂,Σ (p) =R(LΣ (p))R> (LΣ̂ (p))×
(A.1)
[R(LΣ̂ (p))R> (LΣ̂ (p))]−1 ,
where for all p ∈ P,
R(LΣ (p))= B(p) A(p)B(p) . . . A(p)nx −1 B(p) , (A.2)
R(LΣ̂ (p))= B̂(p) Â(p)B̂(p) . . . Â(p)nx −1 B̂(p) (A.3)
are the reachability matrices of LΣ (p) and LΣ̂ (p) respectively. Since A, B, C and Â, B̂, Ĉ are continuous maps,
then from Eq. (A.2) and Eq. (A.3), we can conclude that
the maps P 3 p 7→ R(LΣ (p)) and P 3 p 7→ R(LΣ̂ (p))
are continuous. Hence, from Eq. (A.1), TΣ̂,Σ is also a
continuous map.
A.2 Proof of Lemma 2
Proof. From the definition of quadratic stability of Σ, ∃P
positive definite matrix, such that the LMI
A(p)> P A(p) − P < 0
(A.4)
holds for p ∈ P. Then, let us choose S such that, P = S > S
and define Â(p) := SA(p)S −1 . Then, if follows directly
that, ∀p ∈ P
Â(p)> Â(p) − Inx < 0,
(A.5)
i.e.,
kÂ(p)k < 1.
Since P is compact, then,
sup kÂ(p)k = kÂ(pm )k
(A.6)
(A.7)
p∈P
for some pm ∈ P and the statement of the Lemma follows
directly.
A.3 Proof of Theorem 1
In order to simplify the equations, we define the following
quantities, for each ∆ > 0, k ≥ 0 and i ∈ {0, 1, . . . , ∆ − 1}:
tks = k∆,
tke = k∆ + ∆ − 1,
tki = k∆ + i.
The intuition behind this notation is as follows. Consider
any p ∈ S∆ . Then p is a piece-wise constant scheduling
signal. Then tks , tke and tki denote the starting time instance, the ending time instance and the ith time instance
(k − 1)th interval
tk−1
s
...
From the first equation in Eq. (A.9) and the fact that
−1
(p(tki )), B(p(tki )) =
A(p(tki )) = TΣ̂,Σ (p(tki ))Â(p(tki ))TΣ̂,Σ
k th interval
tk−1
tks . . .
e
tki . . .
TΣ̂,Σ (p(tki ))B̂(p(tki )), it follows that, ∀i ∈ {0, 1, . . . ∆ − 1},
tke
Fig. A.1. Scheduling signal time instances tks , tke and tki .
of the k th interval of the scheduling signal p respectively.
We illustrate these time instances in Figure A.1.
Note that on each set {tks , tks + 1, · · · , tke }, the signal p is
constant, and tks = tk−1
+1, i.e., the sets {tks , tks +1, · · · , tke }
e
are adjacent and cover the whole time axis.
To derive the equations in the statement of the theorem,
which characterize the differences between Σ and Σ̂, let
(x, y, p, u) be a solution of
x(t + 1) = A(p(t))x(t) + B(p(t))u(t),
Σ
(A.8)
y(t)
= C(p(t))x(t),
such that x(0) = 0 and let (x̂, ŷ, p, u), x̂(0) = 0 be a
solution of
x̂(t + 1) = Â(p(t))x̂(t) + B̂(p(t))u(t),
Σ̂
(A.9)
ŷ(t)
= Ĉ(p(t))x̂(t).
In order to be able to compare the states of the two models
Σ and Σ̂, we need to apply a coordinates transformation
to transform the Σ̂ to the same state basis of Σ. Define,
∀i ∈ {0, 1, · · · , ∆ − 1},
x̄(tki ) = TΣ̂,Σ (p(tki ))x̂(tki ),
(A.10)
where TΣ̂,Σ (p(tki )) is the isomorphism from the frozen
model LΣ̂ (p(tki )) to LΣ (p(tki )). Remember that p(tki ) is
constant on the set {tks , tks + 1, · · · , tke }, i.e., p(tki ) = p(tks ),
then x̄(tki ) = TΣ̂,Σ (p(tks ))x̂(tki ) for all tki .
In order to characterize the difference between y(tki ) =
YΣ (u, p)(tki ) and ŷ(tki ) = YΣ̂ (u, p)(tki ) by means of the
model-related coefficients and the input, we break down
the process to three steps for the sake of readability. In
the first step, we relate the difference between x and x̄
inside the set {tks , tks + 1, . . . , tke }, to the difference between
x and x̄ at the beginning of this interval.
In the second step we relate the difference between x and
x̄ at the beginning (tks ) of each interval, to the difference
between x and x̄ at the end (tk−1
) of the preceding interval.
e
Finally, in the third step, we use the results of the
preceding two steps in a recursive manner to relate the
difference between x and x̄ at any time instant t to the
difference between x and x̄ at time instant ∆ − 1.
Next, we implement these three steps mathematically.
x̄(tki + 1) = A(p(tki ))x̄(tki ) + B(p(tki ))u(tki ).
(A.11)
By subtracting above equation from the first equation in
Eq. (A.8) we get,
x(tki + 1) − x̄(tki + 1) = A(p(tki ))(x(tki ) − x̄(tki )). (A.12)
Then, recursively, on the same time interval, we get
x(tki ) − x̄(tki ) = A(p(tks ))i (x(tks ) − x̄(tks )).
(A.13)
Next, by replacing Eq. (15) in Eq. (A.13), we find that,
kx(tki ) − x̄(tki )k ≤ αi kx(tks ) − x̄(tks )k.
(A.14)
2) Step two: relating to the precedent interval. Now, in
order to get a step backward on the time axis, we need to
relate the initial values of the time interval with the last
value of the previous time interval. For this purpose, define
β =A(p(tk−1
))x̄(tk−1
) + B(p(tk−1
))u(tk−1
).
e
e
e
e
Notice that β is a constant value, and one should not
confuse it with a state trajectory.
By taking the norm of β, we get
kβk ≤ αKT µ1 kukl∞ + KB ku(tk−1
)k.
e
Notice, from the definition of the l∞ norm,
∀k ∈ N : kukl∞ ≥ ku(k)k.
(A.15)
Then,
kβk ≤ (αKT µ1 + KB ).kukl∞ .
(A.16)
Now, using the triangle inequality, we can write
kx(tks ) − x̄(tks )k ≤ kx(tks ) − βk + kβ − x̄(tks )k. (A.17)
Notice that,
kx(tks ) − βk = kA(p(tk−1
))(x(tk−1
) − x̄(tk−1
))k
e
e
e
≤ kA(p(tk−1
))k.kx(tk−1
) − x̄(tk−1
)k.
e
e
e
Then,
kx(tks ) − βk ≤ αkx(tk−1
) − x̄(tk−1
)k.
e
e
Remember that
x̄(tks ) =TΣ̂,Σ (p(tks ))x̂(tks )
(A.18)
=TΣ̂,Σ (p(tks ))[Â(p(tk−1
))x̂(tk−1
) + B̂(p(tk−1
))u(tk−1
)]
e
e
e
e
))−1 [A(p(tk−1
))x̄(tk−1
)
=TΣ̂,Σ (p(tks ))TΣ̂,Σ (p(tk−1
e
e
e
+ B(p(tk−1
))u(tk−1
)],
e
e
(A.19)
i.e.,
x̄(tks ) =Mp(tk ),p(tk−1
) β.
e
s
Along with Eq. (A.10), we can directly find that
β − x̄(tks ) = I − Mp(tk ),p(tk−1
β.
)
e
s
(A.20)
(A.21)
Then,
kβ − x̄(tks )k ≤ kI − Mp(tk ),p(tk−1
) k.kβk.
e
s
(A.22)
Then, by replacing (A.16) and (12) in Eq. (A.22),we get
1) Step one: relating to first value of the interval. Now,
we will characterize the difference between the states of
the two LPV models over one interval of the scheduling
signal, i.e., when the scheduling signal is constant. In
this characterization, we expect to see the effect of the
initial states, i.e., the difference at start time instant of
the current interval of the scheduling signal, in addition to
the length of the time interval ∆.
kβ − x̄(tks )k ≤ KM (p)(αKT µ1 + KB )kukl∞ .
(A.23)
Then, combining (A.18) and (A.23), we find that
kx(tks ) − x̄(tks )k ≤αkx(tk−1
) − x̄(tk−1
)k
(A.24)
e
e
+ KM (p)(αKT µ1 + KB )kukl∞ .
In the last step, we recursively find the relation of the
output difference with the initial state of the systems.
3) Step three: recursivity. From Eq. (A.14) and Eq. (A.24), As ϕ is a continuous function defined on a compact set,
then it is uniformly continuous.
we can recursively find that
k
k
i (k−1)∆
kx(ti ) − x̄(ti )k ≤ α α
kx(∆ − 1) − x̄(∆ − 1)k
In particular, since ϕ(p1 , p1 ) = 0 for all p1 ∈ P,
k−1
∀ε
1 > 0 : ∃δ(ε1 ) > 0, such that, if kp1 − p2 k =
X
+ αi
α∆j Ku (p)kukl∞ ,
(A.25) k(p1 − p2 ) − (p1 − p1 )k < δ(ε1 ), then,
j=0
kϕ(p1 , p2 ) − ϕ(p1 , p1 )k = kϕ(p1 , p2 )k < ε1 .
(A.34)
where Ku (p) = KM (p)(αKT µ1 + KB ). Notice that Hence, if p ∈ P is such that kp(t + 1) − p(t)k < δ(ε1 ), then
TΣ̂,Σ (p(0)) is an isomorphism from LΣ̂ (p(0)) to LΣ (p(0)), KM (p) < ε1 .
i.e., the local models of Σ̂ and Σ during the first ∆ time Then, we can take
instances. This yields directly that,
α
−1
ε1 = ε
(αK
µ
+
K
)K
.
(A.35)
T
1
B
C
kx(∆ − 1) − x̄(∆ − 1)k = 0.
1 − α∆
Then,
This means that,
k−1
α
X
KM (p)(αKT µ1 + KB )KC
g(∆, KM (p), 1) =
kx(tki ) − x̄(tki )k ≤ αi Ku (p)kukl∞
α∆j .
(A.26)
1 − α∆
α
j=0
≤ ε1
(αKT µ1 + KB )KC
1 − α∆
Notice that, for α ∈ [0, 1], we have
= ε.
k−1
∞
X
X
1
∆j
∆j
.
(A.27) Then, from Eq (19), with ∆ = 1, ∀ε > 0, ∃δ > 0 such that
α ≤
α =
1 − α∆
i=0
j=0
if kp(t + 1) − p(t)k < δ, then,
Then, replacing in the equation (A.26), we get
kYΣ (u, p)(t) − YΣ̂ (u, p)(t)k ≤ g(∆, KM (p), 1).kukl∞
i
≤ ε.kukl∞ .
α
kx(tki ) − x̄(tki )k ≤
Ku (p)kukl∞ .
(A.28)
∆
1−α
A.6 Proof of Theorem 4
Then, notice that, using Eq (A.10), we can write
C(p(tki ))x̄(tki ) = C(p(tki ))TΣ̂,Σ (p(tki ))x̂(tki )
=
Ĉ(tki )x̂(tki )
=
(A.29)
ŷ(tki ).
C(p(tki )),
Then, from Eq. (A.28), after multiplication by
we get that,
αi
Ku (p) sup kC(p)k.kukl∞ .
ky(tki )− ŷ(tki )k≤
1 − α∆
p
i.e.,
kYΣ (u, p)(tki ) − YΣ̂ (u, p)(tki )k ≤ g(∆, KM (p), i).kukl∞ .
Remember that for all p ∈ P, KM (p) ≤ KM , then,
kYΣ (u, p)(tki ) − YΣ̂ (u, p)(tki )k ≤ g(∆, KM , i).kukl∞ .
A.4 Proof of Theorem 2
Remember that 0 < α < 1, then,
αi
= 0.
lim
i→∞ 1 − α∆
Then, from Eq. (20), we can find that
lim g(∆, KM (p), i) = 0.
i→∞
(A.30)
(A.31)
This means, ∀ε > 0, ∃∆m > 0, such that,
∀i ≥ ∆m : g(∆, KM (p), i) < ε,
(A.32)
Then, from Eq (19), ∀i > ∆m :
kYΣ (u, p)(k∆ + i) − YΣ̂ (u, p)(k∆ + i)k
≤ g(∆m , KM (p), i).kukl∞
≤ ε.kukl∞ .
A.5 Proof of Theorem 3
Consider the map
ϕ : P × P 7→ Rnx ×nx : (p1 , p2 ) → I − Mp1 ,p2 .
−1
TΣ̂,Σ
(A.33)
Since TΣ̂,Σ is continuous, so is
and hence ϕ is
continuous. Remember that P is compact, then so is P × P.
Notice that, from Eq. (20), we can find that,
lim g(∆, KM (p), i) = 0.
α→0
(A.36)
This means that ∀ε > 0, ∃αm > 0, such that, ∀α < αm ,
g(∆, KM (p), i) < ε.
(A.37)
Then, from Eq. (19), in additions to our assumption that
kAk < α, we directly get that, if α < αm , then, ∀t > 0,
kYΣ (u, p)(t) − YΣ̂ (u, p)(t)k ≤ g(∆, KM (p), i).kukl∞
≤ ε.kukl∞ .
| 3cs.SY
|
REGULARITY OF EDGE IDEALS AND THEIR POWERS
arXiv:1712.00887v2 [math.AC] 11 Mar 2018
ARINDAM BANERJEE, SELVI BEYARSLAN, AND HUY TÀI HÀ
Abstract. We survey recent studies on the Castelnuovo-Mumford regularity of edge ideals
of graphs and their powers. Our focus is on bounds and exact values of reg I(G) and the
asymptotic linear function reg I(G)q , for q ≥ 1, in terms of combinatorial data of the given
graph G.
1. Introduction
Monomial ideals are classical objects that live at the crossroad of three areas in mathematics: algebra, combinatorics and topology. Investigating monomial ideals has led to many
important results in these areas. The new construction of edge ideals of (hyper)graphs has
again resurrected much interest in and regenerated a large amount of work on this class of
ideals (cf. [28, 51, 61] and references therein). In this paper, we survey recent works on the
Castelnuovo-Mumford regularity of edge ideals of graphs and their powers.
Castelnuovo-Mumford regularity is an important algebraic invariant which, roughly speaking, measures the complexity of ideals and modules. Restricting to the class of edge ideals
of graphs, our focus is on studies that relate this algebraic invariant to the combinatorial
data of given graphs. Our interest in powers of edge ideals is driven by the well-celebrated
result of Cutkosky, Herzog and Trung [15], and independently, Kodiyalam [42], that for any
homogeneous ideal I in a standard graded k-algebra R, the regularity of I q is asymptotically a linear function in q; that is, there exist constants a and b such that for all q 0,
reg I q = aq + b. Generally, the problem of finding the exact linear form aq + b and the
smallest value q0 such that reg I q = aq + b for all q ≥ q0 has proved to be very difficult. We,
thus, focus our attention also on the problem of understanding the linear form aq + b and
the value q0 for edge ideals via combinatorial data of given graphs.
The on-going research program in which algebraic invariants and properties of edge ideals
and their powers are investigated through combinatorial structures of corresponding graphs
has produced many exciting results and, at the same time, opened many further interesting
questions and conjectures. It is our hope to collect these works together in a systematic
way to give a better overall picture of the problems and the current state of the art of this
research area. For this purpose, we shall state theorems and present sketches of the proofs;
instead of giving full detailed arguments, our aim is to exhibit general ideas behind these
results, the similarities and differences when developing from one theorem to the next.
This paper can be viewed as a complement to the survey done in [27]. While in [27] the
focus was on the regularity of squarefree monomial ideals in general, our attention in this
paper is restricted mostly to edge ideals of graphs, but we enlarge our scope by discussing
1
also the regularity of powers of edge ideals. The later is an important area of study itself,
with a deep motivation from geometry, and has seen a surge of interest in the last few years.
The paper is outlined as follows. In Section 2, we collect notations and terminology used
in the paper. In Section 3, we present necessary tools which were used in works in this area.
Particularly, we shall give Hochster’s and Takayama’s formulas, which relate the graded
Betti numbers of a monomial ideal to the reduced homology groups of certain simplicial
complexes. We shall also describe inductive techniques that have been the backbones of
most of the studies being surveyed. In Section 4, we survey results on the regularity of edge
ideals of graphs. This section is divided into two subsections; the first one focusses on bounds
on the regularity in terms of combinatorial data of the graph, and the second one exhausts
cases where the regularity of an edge ideal can be computed explicitly. Section 5 discusses
the regularity of powers of edge ideals. This section again is divided into two subsections,
which in turn examine bounds for the asymptotic linear function of the regularity of powers
of edge ideals and cases when this asymptotic linear function can be explicitly described.
In Section 6, we recall a number of results extending the study of edge ideals of graphs to
hypergraphs. Since, our focus in this paper is on mostly edge ideals of graphs, our results
in Section 6 will be representative rather than exhaustive. We end the paper with Section
7, in which we state a number of open problems and questions which we would like to see
answered. We hope that these problems and questions would stimulate further studies in
this research area.
Acknowledgement. The authors would like to thank the organizers of SARC (Southern
Regional Algebra Conference) 2017 for their encouragement, which led us to writing this
survey. The last named author is partially supported by the Simons Foundation (grant
#279786) and Louisiana Board of Regent (grant #LEQSF(2017-19)-ENH-TR-25).
2. Preliminaries
In this section we recall preliminary notations and terminology of combinatorics and algebra that will be used throughout this survey. In Subsection 2.1 we give various definitions on
graphs, hypergraphs, and simplicial complexes. In Subsection 2.2 we recall basic homological algebra terminology. Finally, in the last subsection we define various algebraic objects
associated to (hyper)graphs and simplicial complexes.
2.1. Combinatorial preliminaries. For any finite simple graph G, with set of vertices
V (G) and set of edges E(G), we define some graph-theoretic notions as follows.
For any A ⊆ V (G), the induced subgraph on A is the graph whose set of vertices is A
and whose edges are exactly the edges of G that join two elements of A. In the following
example (Figure 1) H 0 is an induced subgraph of H, while H 00 is not an induced subgraph
of H as it misses the edge yw of H.
For any vertex x in a graph G the degree of x, denoted by d(x) is the number of vertices
connected to x. For any vertex x in a graph G the neighborhood of x, denoted by NG [x]
is the set consisting of x and all its neighbors. For the graph H 00 above, d(x) = 2 and
2
y
z
H: x
w
t
y
s
z
H 0: x
y
z
H 00 : x
w
w
s
s
Figure 1. Induced subgraphs
NG [x] = {x, y, w}. For any graph G and any vertex x, by G \ x we denote the induced
subgraph on V (G) \ {x}.
For any graph G, the complement graph, denoted by Gc , is the graph whose vertices are
the vertices of G and whose edges are the non-edges of G, i.e., for a, b ∈ V (G), ab is an edge
in Gc if and only if ab is not an edge in G. For the graph H in Figure 1, the following is H c .
y
z
H c: x
w
t
s
Figure 2. Complement graph
A cycle of length n in a graph G is a closed walk along its edges, x1 x2 , x2 x3 , . . . , xn−1 xn , xn x1 ,
such that xi 6= xj for i 6= j. We denote the cycle on n vertices by Cn . A chord in the cycle
Cn is an edge xi xj where xj 6= xi−1 , xi+1 . A graph is said to be chordal if for any cycle of
length greater than or equal to 4 there is a chord. A graph is said to be co-chordal if the
complement of G is chordal. The graph H in Figure 2 is chordal and H c is co-chordal.
A forest is a graph without any cycles. A tree is a connected forest.
A complete graph (or clique) on n vertices is a graph where for any two vertices there is
an edge joining them. It is denoted by Kn .
3
A bipartite graph is a graph whose vertices can be split into two groups such that there
is no edge between vertices of same group; only edges are between vertices coming from
different groups. It is easy to see that a graph is bipartite if and only if it is without any
cycle of odd length. We use Km,n to denote the complete bipartite graph with m vertices on
one side, and n on the other. We observe that H 00 in Figure 1 has no cycle, so it is a tree.
As a consequence it is also a bipartite graph with bipartition {x, z, s} and {y, w}.
Let G be a graph. We say two disjoint edges uv and xy form a gap in G if G does not have
an edge with one endpoint in {u, v} and the other in {x, y}. A graph without an induced
gap is called gap-free. Equivalently, G is gap-free if and only if Gc contains no induced C4 .
The graph H in Figure 1 is gap-free. In the following graph {x, y} and {z, w} form a gap.
y
z
G: x
w
s
Figure 3. A graph which is not gap-free
A matching in a graph is a set of pairwise disjoint edges. A matching is called an induced
matching if the induced subgraph on the vertices of the edges forming the matching has no
other edge. We observe that {x, y} and {z, w} forms an induced matching for graph G in
Figure 3.
Any graph isomorphic to K1,3 is called a claw. Any graph isomorphic to K1,n is called an
n-claw. If n > 1, the vertex with degree n is called the root in K1,n . A graph without an
induced claw is called claw-free. A graph without an induced n-claw is called n-claw-free.
y
x
w
t
Figure 4. A 3-claw (or simply, a claw)
A graph is called a diamond if it is isomorphic to the graph with vertex set {a, b, c, d} and
edge set {ab, bc, ac, ad, cd}. A graph without an induced diamond is called diamond-free.
4
A graph is said to be planar if as a 1-dimensional topological space it can be embedded
in the complex plane, i.e., if it can be drawn in the plane in such a way that no pair of edges
cross. The following graph in Figure 5 is a diamond. We observe that it is also planar.
y
z
x
t
Figure 5. A diamond
Any graph isomorphic to the graph with set of vertices {w1 , w2 , w3 , w4 , w5 } and set of edges
{w1 w3 , w2 w3 , w3 w4 , w3 w5 , w4 w5 } is called a cricket. A graph without an induced cricket is
called cricket-free. It is easy to see that a claw-free graph is also cricket-free.
y
w
x
t
z
Figure 6. A cricket
An edge in a graph is called a whisker if one of its vertices has degree one. A graph is
called an anticycle if its complement is a cycle.
x1
x4
x7
x5
x2
x3
x8
x9
x6
Figure 7. A hypergraph
A hypergraph is the natural higher degree analogue of graphs, in the sense that we allow
an edge to be a collection of any number of vertices. Formally speaking, a hypergraph
5
H = (V (H), E(H)) consists of the vertices V (H) = {x1 , . . . , xn } and the edges E(H), where
E(H) is a collection of nonempty subsets of the vertices. With this notation, a graph G is a
hypergraph whose edges are subsets of cardinality 2. By abusing notation, we often identify
an edge {xi1 , . . . , xir } ∈ E(H) with the monomial xi1 . . . xir . Figure 7 depicts a hypergraph
with edges {x1 x2 x3 x4 , x3 x4 x5 , x5 x6 x7 , x7 x8 , x9 }. In this example, the vertex x9 can also be
viewed as an isolated vertex.
A clutter is a hypergraph none of whose edges contains any other edge as a subset. The
hypergraph in Figure 7 is a clutter.
A simplicial complex with vertices {x1 , . . . , xn } is subset of the power set of {x1 , . . . , xn }
which is closed under the subset operation. The sets that constitute a simplicial complex are
called its faces. Maximal faces under inclusion are called facets. Figure 8 gives a simpicial
complex with facets {a, b, f }, {c, d, e}, {f, e},{b, c}.
e
f
a
d
b
c
Figure 8. A simplicial complex
An independent set in a graph G is a set of vertices no two of which forms an edge. The
independence complex of a graph G, denoted by ∆(G), is the simplicial complex whose faces
are independent sets in G.
a
e
c
d
b
f
Figure 9. A simple graph whose independence complex is in Figure 8
Let ∆ be a simplicial complex, and let σ ∈ ∆. The deletion of σ in ∆, denoted by del∆ (σ),
is the simplicial complex obtained by removing σ and all faces containing σ from ∆. The
link of σ in ∆, denoted by link∆ (σ), is the simplicial complex whose faces are
{F ∈ ∆ | F ∩ σ = ∅, σ ∪ F ∈ ∆}.
A simplicial complex ∆ is recursively defined to be vertex decomposable if either:
6
(i) ∆ is a simplex; or
(ii) there is a vertex v in ∆ such that both link∆ (v) and del∆ (v) are vertex decomposable,
and all facets of del∆ (v) are facets of ∆.
A vertex satisfying condition (2) is called a shedding vertex, and the recursive choice of
shedding vertices is called a shedding order of ∆.
Recall that a simplicial complex ∆ is said to be shellable if there exists a linear
order of
Sk−1 T
its facets F1 , F2 , . . . , Ft such that for all k = 2, . . . , t, the subcomplex
Fk is pure
i=1 Fi
and of dimension (dim Fk − 1). Here F represents the simplex over the vertices of F . It is a
celebrated fact that pure shellable complexes give rise to Cohen-Macaulay Stanley-Reisner
rings. Recall also that a ring or module is sequentially Cohen-Macaulay if it has a filtration
in which the factors are Cohen-Macaulay and their dimensions are increasing. This property
corresponds to (nonpure) shellability in general. Vertex decomposability can be thought
of as a combinatorial criterion for shellability and sequentially Cohen-Macaulayness. In
particular, for a simplicial complex ∆,
∆ vertex decomposable ⇒ ∆ shellable ⇒ ∆ sequentially Cohen-Macaulay.
2.2. Algebraic preliminaries. Let S = K[x1 , . . . , xn ] be a polynomial ring over a field K.
Let M be a finitely generated Zn -graded S-module. It is known that M can be successively
approximated by free modules. Formally speaking, there exists an exact sequence of minimal
possible length, called a minimal free resolution of M :
dp
d
d
d
2
1
0
0 −→ Fp −→ Fp−1 · · · −→
F1 −→
F0 −→
M −→ 0
L
(∗)
βi,σ
Here, Fi = σ∈Zn S(−σ) , where S(−σ) denotes the free module obtained by shifting the
degrees in S by σ. The numbers βi,σ ’s are positive integers and are called the multigraded
Betti numbers of M . We often identify σ with the monomial whose exponent vector is σ.
For example, over K[x, y], we may write βi,x2 y (M ) instead of βi,(2,1) (M ).
P
For every j ∈ Z, βi,j =
{σ | |σ|=j} βi,σ is called the (i, j)-th standard graded Betti
number of M . Three very important homological invariants that are related to these numbers
are the Castelnuovo-Mumford regularity, or simply regularity, the depth and the projective
dimension, denoted by reg(M ), depth(M ) and pd(M ) respectively:
reg M = max{|σ| − i | βi,σ 6= 0}
depth M = inf{i | Exti (K, M ) 6= 0}
pd M = max{i | there is a σ, βi,σ 6= 0}.
If S is viewed as a standard graded K-algebra and M is a graded S-module, then the
graded Betti numbers of M are also given by βi,j (M ) = dimk Tori (M, K)j , and so we have
reg M = max{j − i | Tori (M, K)j 6= 0}
pd M = max{i | Tori (M, K) 6= 0}
depth M = n − pd M.
In practice, we often work with short exact sequences, so it is worthwhile to mention
that the regularity can also be defined via the vanishings of local cohomology modules with
7
respect to the “irrelevant maximal ideal” m = (x1 . . . , xn ). Particularly, for i ≥ 0, define
(
i
i
max{j | Hm
(M )j 6= 0} if Hm
(M ) 6= 0
ai (M ) :=
−∞
otherwise.
Then, the regularity of M is also given by
reg M = max{ai (M ) + i}.
i
Example 2.1. Let M = Q[x1 , . . . , x5 ]/(x1 x2 , x2 x3 , x3 x4 , x4 x5 , x5 x1 ). Then the minimal free
resolution of M is:
d3
d2
d1
d0
0 −→ F3 −→
F2 −→
F1 −→
F0 −→
M −→ 0
Here:
β0,σ = 1 if σ = 1, and β0,σ = 0 otherwise
β1,σ = 1 if σ = x1 x2 , x2 x3 , x3 x4 , x4 x5 , x5 x1 , and β1,σ = 0 otherwise
β2,σ = 1 if σ = x1 x2 x3 , x2 x3 x4 , x1 x2 x5 , x1 x4 x5 , x3 x4 x5 , and β2,σ = 0 otherwise
β3,σ = 1 if σ = x1 x2 x3 x4 x5 , and β3,σ = 0 otherwise
An ideal I in S is said to be unmixed if all its associated primes are minimal of same height.
We say I (or equivalently S/I) is Cohen-Macaulay if the Krull dimension and depth are
equal. Cohen-Macaulay ideals are always unmixed and known to have many nice geometric
properties. It can be checked that the module M in Example 2.1 is Cohen-Macaulay.
2.3. Algebraic objects with underlying combinatorial structures. For any (hyper)graph
H over the vertex set V (H) = {x1 , . . . , xn }, its edge ideal is defined as follows:
Y
I(H) = ( x e ∈ E(H)) ⊆ S = K[x1 , . . . , xn ].
x∈e
Example 2.2. The edge ideal of the hypergraph, say H, in Figure 7, is
I(H) = (x1 x2 x3 x4 , x3 x4 x5 , x5 x6 x7 , x7 x8 , x9 ).
For any simplicial complex ∆ over the vertex set V = {x1 , . . . , xn }, its Stanley-Reisner
ideal is defined as follows:
Y
I∆ = ( x e is a minimal non-face of ∆) ⊆ S.
x∈e
Example 2.3. The Stanley-Reisner ideal of the simplicial complex in Figure 8 is
I∆ = (ac, ad, ae, bd, be, cf, df ).
As we have seen before, (hyper)graphs are related to simplicial complexes via the notion
of independence complex. The algebraic view of this relation exposes the following equality.
Lemma 2.4. Let G be a simple (hyper)graph and let ∆(G) be its independence complex.
Then
I(G) = I∆(G) .
8
Example 2.5. The edge ideal of the graph G in Figure 9 is the same as the the StanleyReisner ideal of the simplicial complex ∆ in Figure 8. Note that ∆ = ∆(G).
I(G) = (ae, ad, bd, be, bf, ec, ef )
For any simple graph G with V (G) = {x1 , . . . , xn }, we define the t-path ideal as:
It (G) = (xi1 · · · xit | ik 6= il for k 6= l, xik xik+1 an edge of G).
Example 2.6. The 3-path ideal of the five cycle C5 : x1 x2 x3 x4 x5 is
I3 (C5 ) = (x1 x2 x3 , x2 x3 x4 , x3 x4 x5 , x4 x5 x1 , x5 x1 x2 ).
Finally, as a matter of convention, for any (hyper)graph G, by reg G we mean reg I(G).
3. Formulas and inductive approaches
Computing non-vanishings of local cohomology modules or Betti numbers of an ideal
can be quite complicated. As a result, inductive techniques that relate regularity of edge
ideals with simplicial complexes and smaller ideals are employed as a common tool in the
literature. In this context, induced structures such as induced subcomplexes and induced
subgraphs have proven to be significant objects to investigate regularity of an edge ideal and
its powers.
In this section, we focus on the methods that enable us to bound and compute regularity
of an edge ideal and its powers. We start the section by recalling two important formulas.
Then we address a few inductive bounds that are widely used in the literature.
3.1. Hochster’s and Takayama’s Formulas. Hochster’s formula has been a significant
tool in the study of squarefree monomial ideals due to its power to relate the multigraded
Betti numbers of a simplicial complex ∆ to the non-vanishings of the reduced simplicial
homology groups of ∆ and its induced subcomplexes.
Theorem 3.1 (Hochster’s Formula [35]). Let ∆ be a simplicial complex on the vertex set V
and let I(∆) be its Stanley-Reisner ideal. Then
βi,j (I(∆)) =
X
e j−i−2 (∆W ; K)
dimk H
W ⊆V, |W |=j
where ∆W is the restriction of ∆ to the vertex set W.
For an arbitrary monomial ideal, Hochster’s formula can not be employed. Takayama’s
formula given in [58] can perform a similar task as Hochster’s formula for this class of ideals.
Let I ⊆ S = K[x1 , . . . , xn ] be a monomial ideal. Takayama’s formula provides a combinatorial description for the non-vanishings of Zn -graded component Hmi (S/I)a for a ∈ Zn , and
this description is given in terms of certain simplicial complexes ∆a (I) related to I. Note
that S/I is an Nn -graded algebra, and so Hmi (S/I) is a Zn -graded module over S/I.
The simplicial complex ∆a (I) is called the degree complex of I. The construction of ∆a (I)
was first given in [58] and then simplified in [48]. We recall the construction from [48].
9
For a = (a1 , . . . , an ) ∈ Zn , set xa = xa11 · · · xann and Ga := {j ∈ {1, . . . , n} | aj < 0}. For
every subset F ⊆ {1, . . . , n}, let SF = S[x−1
j | j ∈ F ]. Define
∆a (I) = {F \ Ga | Ga ⊆ F, xa ∈
/ ISF }.
Theorem 3.2 (Takayama’s Formula [58]). Let√I ⊆ S be a monomial ideal, and let ∆(I)
denote the simplicial complex corresponding to I. Then
(
e i−|Ga |−1 (∆a (I), K
dimk H
i
dimk Hm (S/I)a =
0
if Ga ∈ ∆(I),
otherwise.
The formula stated in [48] and here is different than the original formula introduced in
[58]. The original formula has additional conditions on a for Hmi (S/I)a = 0. It follows from
the proof in [58] that those conditions can be omitted.
Due to the equality between edge ideals of graphs and Stanley-Reisner ideals of their
independence complexes, we can employ Hochster’s and Takayama’s formulas in the study
of regularity of edge ideals.
In order to deal with arbitrary monomial ideals, polarization is proved to be a powerful
process to obtain a squarefree monomial ideal from a given monomial ideal. For details of
polarization we refer the reader to [32].
Definition 3.3. Let M = xa11 . . . xann be a monomial in S = K[x1 , . . . , xn ]. Then we define
the squarefree monomial P (M ) (polarization of M ) as
P (M ) = x11 . . . x1a1 x21 . . . x2a2 . . . xn1 . . . xnan
in the polynomial ring R = K[xij | 1 ≤ i ≤ n, 1 ≤ j ≤ ai ]. If I = (M1 , . . . , Mq ) is an ideal in
S, then the polarization of I, denoted by I pol , is defined as I pol = (P (M1 ), . . . , P (Mq )).
The regularity is preserved under polarization.
Corollary 3.4. [32, Corollary 1.6.3.d] Let I ⊂ S be a monomial ideal and I pol ⊂ R be its
polarization. Then reg(S/I) = reg(R/I pol ).
3.2. Inductive techniques. We start the section with an easy yet essential consequence
of Hochster’s formula that links the regularity of a graph with regularity of its induced
subgraphs.
Lemma 3.5. Let G be a simple graph. Then reg H ≤ reg G for any induced subgraph H of
G.
For any homogeneous ideal I ⊆ S and any homogeneous element m of degree d, the
following short exact sequences are used as standard tools in commutative algebra and also
proved to be very useful in computing regularity of edge ideals and their powers.
0 −→
R
R
·m R
(−d) −→
−→
−→ 0
I:m
I
I + (m)
10
(3.1)
Let I and J be ideals in S. Another useful exact sequence is:
0 −→
R R
R
R
→ ⊕ −→
−→ 0
I ∩J
I
J
I +J
(3.2)
We can see how regularity changes in a short exact sequence by taking the associated long
exact sequence of local cohomology modules. Particularly, we have the following inductive
bound (the second statement is the content of [17, Lemma 2.10 ]).
Lemma 3.6. Let I ⊆ S be a monomial ideal, and let m be a monomial of degree d. Then
reg I ≤ max{reg(I : m) + d, reg(I, m)}.
Furthermore, if x is a variable appearing in I, then
reg I ∈ {reg(I : x) + 1, reg(I, x)}.
In the statement of Lemma 3.6, the expression x is a variable appearing in I means some
of the minimal generators of I is divisible by x. Note that if x is a variable not appearing
in I, then reg(I, x) = reg I. In case of edge ideals, if x is an isolated vertex in G, we can
drop the vertex x when computing regularity. Thus reg(I(G) : x) = reg I(G \ NG [x]) and
reg(I(G), x) = reg(I(G \ x)) for any vertex x in G. Then Lemma 3.6 can be restated in terms
of edge ideals.
Lemma 3.7. Let x be a vertex in G. Then
reg G ∈ {reg(G \ NG [x]) + 1, reg(G \ x)}.
Kalai and Meshulam [39] proved the following result for squarefree monomial ideals and
Herzog [31] generalized it to any monomial ideal.
Theorem 3.8. Let I1 , . . . , Is be squarefree monomial ideals in S. Then
s
s
X
X
reg S/Ii
reg S
Ii ≤
i=1
i=1
In the case of edge ideals, we have the following bound.
Corollary
3.9. Let G be a simple graph. If G1 , . . . , Gs are subgraphs of G such that E(G) =
Ss
i=1 E(Gi ) then
s
X
reg S/I(G) ≤
reg S/I(Gi )
i=1
If G is the disjoint union of graphs G1 , . . . , Gs then the above equality is achieved by using
Künneth formula in algebraic topology.
Corollary 3.10. Let G be a simple graph. If G can be written as a disjoint union of graphs
G1 , . . . , Gs then
s
X
reg S/I(G) =
reg S/I(Gi )
i=1
11
In the study of powers of edge ideals, Banerjee developed the notion of even-connection
and gave an important inductive inequality in [5].
Theorem 3.11. For any simple graph G and any s ≥ 1, let the set of minimal monomial
generators of I(G)s be {m1 , . . . , mk }. Then
reg I(G)s+1 ≤ max{reg(I(G)s+1 : ml ) + 2s, 1 ≤ l ≤ k, reg I(G)s }.
In particular, if for all s ≥ 1 and for all minimal monomial generators m of I(G)s ,
reg (I(G)s+1 : m) ≤ 2 and reg I(G) ≤ 4, then reg I(G)s+1 = 2s + 2 for all s ≥ 1; as a
consequence I(G)s+1 has a linear minimal free resolution.
Working with the above inequality requires a good understanding of the ideal (I(G)s+1 :
m) and its generators when m is a minimal monomial generator of I(G)s . The even-connection
definition is key to attain this goal. We recall the definition of even-connectedness and its
important properties from [5].
Definition 3.12. Let G = (V, E) be a graph. Two vertices u and v (u may be the same
as v) are said to be even-connected with respect to an s-fold product e1 · · · es where ei ’s are
edges of G, not necessarily distinct, if there is a path p0 p1 · · · p2k+1 , k ≥ 1 in G such that:
(1)
(2)
(3)
(4)
p0 = u, p2k+1 = v.
For all 0 ≤ l ≤ k − 1, p2l+1 p2l+2 = ei for some i.
For all i, | {l ≥ 0 | p2l+1 p2l+2 = ei } | ≤ | {j | ej = ei } |.
For all 0 ≤ r ≤ 2k, pr pr+1 is an edge in G.
Fortunately, it turns out that (I(G)s+1 : m) is generated by monomials in degree 2.
Theorem 3.13. [5, Theorem 6.1 and Theorem 6.7] Let G be a graph with edge ideal I = I(G),
and let s ≥ 1 be an integer. Let m be a minimal generator of I s . Then (I s+1 : m) is minimally
generated by monomials of degree 2, and uv (u and v may be the same) is a minimal generator
of (I s+1 : m) if and only if either {u, v} ∈ E(G) or u and v are even-connected with respect
to m.
After polarization, the ideal (I(G)s+1 : m) can be viewed as the edge ideal of a graph
that is obtained from G by adding even-connected edges with respect to m. Some of the
combinatorial properties of this ideal in relation to G are studied in [37].
4. Regularity of edge ideals
Computing and/or bounding the regularity of edge ideals is the foundation in the study of
regularity of powers of edge ideals. In this section, we identify the combinatorial structures
that are related to regularity of edge ideals. First, we collect the general upper and lower
bounds given for the regularity of edge ideals, then we present the list of known classes of
graphs where the regularity is computed explicitly.
Note that regularity of an edge ideal is bounded below by 2, which is the generating degree
of an edge ideal. Thus, identifying combinatorial structures of a graph with regularity 2
12
can be considered as the base case of results in this section. The following combinatorial
characterization of such graphs is nowadays often referred to as Fröberg’s characterization.
It was, in fact, given first in topological language by Wegner [63] and later, independently,
by Lyubeznik [45] and Fröberg [25] in monomial ideals language.
Theorem 4.1 ([25, Theorem 1]). Let G be a simple graph. Then reg I(G) = 2 if and only
if G is a co-chordal graph.
4.1. Lower and upper bounds. One of the graph-theoretical invariants that can be related
to the regularity is the induced matching number. The first result revealing this relation is
due to Katzman and it provides a general lower bound on the regularity of edge ideals.
Theorem 4.2 ([41, Lemma 2.2]). Let G be a simple graph and ν(G) be the maximum size
of an induced matching in G. Then
reg I(G) ≥ ν(G) + 1.
Sketch of proof. Let {e1 , . . . , er } be an induced matching of maximal size in G. Suppose
H is the induced subgraph of G with E(H) = {e1 , . . . , er }. Note that all the edges in H are
disjoint. Thus reg(H) = r + 1. By Lemma 3.5, reg(G) ≥ ν(G) + 1.
Another graph-theoretical invariant of interest is the matching number and this invariant
actually emerges as a general upper bound for any graph.
Theorem 4.3 ([29, Theorem 6.7], [64, Theorem 11]). Let G be a simple graph. Let β(G) be
the minimum size of a maximal matching in G. Then
reg I(G) ≤ β(G) + 1.
Example 4.4. Let G be the graph given in Figure 10. It is clear that ν(G) = β(G) = 3.
Then reg G = 4 by Theorem 4.2 and Theorem 4.3. In general, if G is a simple graph with
disjoint edges, then the above bounds coincide and become an equality.
Figure 10. A graph with disjoint edges
Comparing the existing lower and upper bounds yields to interesting classifications, particularly when these bounds coincide. For example, Cameron and Walker [16, Theorem 1]
gave the first classification of graphs G with ν(G) = β(G). Then Hibi et al [34] modified
their result slightly and gave a full generalization with some corrections.
Example 4.5. Let G be the graph given in Figure 11. One can easily verify that β(G) = 2.
However reg G = 2 by Fröberg’s characterization. In general, we can find examples of graphs
where β(G) + 1 can be arbitrarily large compared to reg G.
13
Figure 11. A co-chordal graph
The above upper bound is strengthened by making use of co-chordal subgraph covers
of a graph and this bound is proved by Woodroofe. Recall that a graph is co-chordal if
its complement is chordal and co-chord(G), the co-chordal number of G, denotes the least
number of co-chordal subgraphs of G whose union is G. The graph in Figure 11 is an example
of a co-chordal graph.
Let {e1 , . . . , er } be a maximal matching of minimal size in G. For each i, let Gi be the
subgraph of G with edges ei ∪ { edges in G adjacent to ei }. Note that G1 , . . . , Gr forms a
co-chordal subgraph cover of G, thus co-chord(G) ≤ β(G) and the bound in Theorem 4.6
improves the bound in Theorem 4.3.
Theorem 4.6 ([64, Lemma 1]). Let G be a simple graph. Then
reg I(G) ≤ co-chord(G) + 1.
Sketch of proof. Let co-chord(G) = r and let G1 , . . . , Gr be a co-chordal cover of G. It
follows from Fröberg’s characterization of regularity 2 graphs, Theorem 4.1, that reg(Gi ) = 2
for each i ∈ {1, . . . , r}. Then the result follows immediately from Corollary 3.9.
Gap-free graphs have also been of interest in the investigation of regularity. These graphs
arise naturally since their induced matching number is 1 which is the smallest possible
number it could be. However, computing or bounding the regularity of this class of graphs
is not so easy. Furthermore, there are very few examples on how large the regularity of such
edge ideals can be, see [53] by Nevo and Peeva for an example of a gap-free graph G in
12 variables with reg G = 4. Putting an additional condition on the gap-freeness of G may
result with an upper bound and it is indeed achieved in [52] .
Theorem 4.7 ([52, Theorem 1.2], [17, Proposition 19]). If G is gap-free and claw-free, then
reg I(G) ≤ 3.
Sketch of proof. Let x be a vertex in G with the maximum possible degree. By Lemma
3.7, we have reg(G) ≤ max{reg(G \ NG [x]) + 1, reg(G \ x)}. Note that induced subgraphs
G \ NG [x] and G \ x of G are both gap-free and claw-free. It follows from the induction on
the number of vertices that reg(G \ x) ≤ 3. Thus it suffices to show that reg(G \ NG [x]) ≤ 2.
By Fröberg’s characterization, it is enough to prove that (G \ NG [x])c is chordal and it is
proved by contradiction:
(1) Suppose (G \ NG [x])c has an induced cycle on w1 , w2 , . . . , wn (in order) of length at
least 4.
(2) Any vertex of G is distance 2 from x in G by [17, Proposition 3.3], then {x, y} and
{y, w1 } are edges in G for some vertex y.
14
(3) Note that either {y, w2 } or {y, wn } must be an edge in G. Otherwise edges {w2 , wn }
and {x, y} form a gap in G.
(4) Without loss of generality, suppose {y, w2 } is an edge. Then the induced subgraph
on {x, y, w1 , w2 } is a claw in G, a contradiction.
Example 4.8. There are examples of gap-free and claw-free graphs where both values in
Theorem 4.7 can be attained for regularity. For example, if Gc is a tree then reg G = 2 and
if Gc is Cn for n ≥ 5 then reg G = 3.
This result is generalized to n-claw free and gap-free graphs by Banerjee in [5]. The proof
of the general case follows similarly and uses induction on n.
Theorem 4.9 ([5, Theorem 3.5]). If G is gap-free and n-claw-free, then reg I(G) ≤ n.
Another special class of graphs are planar graphs. This class of graphs emerges frequently
in applications since they can be drawn in the plane without edges crossing. It is given in
[64] that even though regularity of a planar graph may be arbitrarily large, regularity of its
complement can be bounded above by 4.
Theorem 4.10 ([64, Theorem 3.4]). If G is a planar graph, then reg I(Gc ) ≤ 4.
4.2. Exact values. Computing the regularity for special classes of graphs has been an
attractive research topic in the recent years. To that extend characterizing edge ideals with
certain regularity has been of interest as well. However, very little is known in the latter
case.
A combinatorial characterization of edge ideals with regularity 3 is still not known. However, a partial result for bipartite graphs is achieved by Fernández-Ramos and Gimenez in
[23]. Recall that a graph G is bipartite if the vertices V can be partitioned into disjoint
subsets V = X ∪ Y such that {x, y} is an edge in G only if x ∈ X and y ∈ Y or vice versa.
The bipartite complement of a bipartite graph G, denoted by Gbc , is the bipartite graph over
the same partition of vertices and {x, y} ∈ Gbc if and only if {x, y} ∈
/ G for x ∈ X, y ∈ Y.
Theorem 4.11 ([23, Theorem 3.1]). Let G be a connected bipartite graph. Then reg I(G) = 3
if and only if Gc has no induced cycles of length ≥ 4 and Gbc has no induced cycles of length
≥ 6.
In most of the known cases, it turns out that regularity can be expressed in terms of the
induced matching number. We first recall the results when the regularity is one more than
the induced matching number.
Theorem 4.12. Let G be a simple graph and ν(G) be the induced matching number of G.
Then
reg I(G) = ν(G) + 1
in the following cases:
15
(a)
(b)
(c)
(d)
(e)
(f )
(g)
(h)
G is a chordal graph (see [29]);
G is a weakly chordal graph (see [64]);
G is sequentially Cohen-Macaulay bipartite graph (see [62]);
G is unmixed bipartite graph (see [43]);
G is very well-covered graph (see [46]);
G is vertex decomposable graph and has no closed circuit of length 5 (see [40]);
G is a (C4 , C5 )-free vertex decomposable graph (see [7]);
G is a unicyclic graph with cycle Cn when
(i) n ≡ 0, 1 (mod 3) (see [2, 10, 36]) or
(ii) ν(G \ Γ(G)) < ν(G) where Γ(G) is the collection of all neighbors of the roots in
the rooted trees attached to Cn (see [2]).
x
w
y
t
z
Figure 12. A vertex decomposable and a (weakly) chordal graph
Remark 4.13. Chordal graphs are vertex decomposable and vertex decomposable graphs are
sequentially Cohen-Macaulay (see [26, 65]). Sequentially Cohen-Macaulay bipartite graphs
are vertex decomposable (see [62]). Also if G is a very well-covered graph then G is unmixed
and if G is chordal, then it is weakly chordal. Thus, (b) implies (a), and (e) implies (d).
Note that bipartite graphs have no odd cycles. Hence, (f ) implies (c).
Example 4.14. Let G be the graph given in Figure 13. Then Γ(G) = {x6 , x7 , x8 } and
ν(G \ Γ(G)) = 2 < ν(G) = 3. Furthermore, this graph does not belong to any of the classes
described in (a)—(g).
x5
x6
x7
x4
x1
x2
x8
x9
x10
x3
Figure 13. A graph satisfying (h) (ii) in Theorem 4.12
16
It is of interest to find different expressions for the regularity of edge ideals. The next
result collects all known cases in which the regularity has a different expression than the
above classes and it is still in terms of induced matching.
Theorem 4.15. Let G be a simple graph and ν(G) be the induced matching number of G.
Then
reg I(G) = ν(G) + 2
in the following cases:
(a) G is an n-cycle Cn when n ≡ 2 (mod 3) (see [10, 36]);
(b) G is a unicyclic graph with cycle Cn when n ≡ 2 (mod 3) and ν(G \ Γ(G)) = ν(G)
where Γ(G) is the collection of all neighbors of the roots in the rooted trees attached
to Cn (see [2]).
5. Regularity of powers of edge ideals
The regularity of powers of an edge ideal is considerably harder to compute than that
of the edge ideal itself. However, in many cases, known bounds and exact formulas for
the regularity of I(G) inspire new bounds and exact formulas for the regularity of I(G)q , for
q ≥ 1. This section is divided into two parts, where in the first subsection we list a number of
lower and upper bounds, and in the second subsection we give the exact values of reg I(G)q ,
for q ≥ 1, for special classes of graphs.
5.1. Lower and upper bounds. Just like with studies on the regularity of edge ideals,
the induced matching number of a graph is ultimately connected to the regularity of powers
of its edge ideal. The following general lower bound generalizes that of Theorem 4.2 for an
edge ideal to its powers.
Theorem 5.1 ([10, Theorem 4.5]). Let G be any graph and let I = I(G) be its edge ideal.
Then for all q ≥ 1, we have
reg I q ≥ 2q + ν(G) − 1.
Sketch of proof. The proof is based on the following observations:
(1) If H is an induced subgraph of G then for any i, j ∈ Z, we have
βi,j (I(H)q ) ≤ βi,j (I(G)q ).
(5.1)
In particular, this gives reg I(H)q ≤ reg I(G)q .
(2) If H is the induced subgraph of G consisting of a maximal induced matching of ν(G)
edges then for any q ≥ 1, we have
reg I(H)q = 2q + ν(G) − 1.
(5.2)
To establish the first observation, recall that the upper-Koszul simplicial complex K α (I)
associated to a monomial ideal I ⊆ S = K[x1 , . . . , xn ] at degree α ∈ Zn consists of faces
n
o
xα
W ⊆ {1, . . . , n} Q
∈I ,
w∈W w
17
and a variation of Hochster’s formula (Theorem 3.1) gives
e i−1 (K α (I); K) for all i ≥ 0.
βi,α (I) = dimk H
The inequality (5.1) then follows by noting that for α ∈ Zn with supp(α) ⊆ VH , K α (I(H)q ) =
K α (I(G)q ).
The second observation is proved by induction, noting that in this case I(H) is a complete
intersection.
Figure 14. A uniyclic graph
Example 5.2. Though there are many classes of graphs for which the lower bound given in
Theorem 5.1 is attained, there are also classes of graphs where the asymptotic linear function
reg I(G)q is strictly bigger than 2q + ν(G) − 1 for all q 0. Let G be the graph depicted in
Figure 14. Then it is easy to see that ν(G) = 2, whereas reg I(G)q = 2q + 2 for q ≥ 1.
A similar general upper bound generalizing that of Theorem 4.6 is, unfortunately, not
available. It is established, by Jayanthan, Narayanan and Selvaraja [37], only for a special
class of graphs — bipartite graphs.
Theorem 5.3 ([37, Theorem 1.1]). Let G be a bipartite graph and let I = I(G) be its edge
ideal. Then for all q ≥ 1, we have
reg I q ≤ 2q + co-chord(G) − 1.
Sketch of proof. The statement is proved by induction utilizing Theorem 3.11. The crucial
step in the proof is to show that for any q ≥ 1 and any collection of edges e1 , . . . , eq of G
(not necessarily distinct), we have
reg(I q+1 : e1 . . . eq ) ≤ co-chord(G) + 1.
(5.3)
Observe that when G is a bipartite graph, by [1, Lemma 3.7],
I q+1 : e1 . . . eq = (((I 2 : e1 )2 : . . . )2 : eq ),
and so (5.3) itself can be obtained by induction. To this end, let G0 be the graph associated
to the polarization of I 2 : e for an edge e in G (which is generated in degree 2 by Theorem
3.13). The proof is completed by establishing the following facts:
(1) reg I(G0 ) ≤ co-chord(G0 ) + 1. This inequality was proved to hold for any graph in
[64].
(2) co-chord(G0 ) ≤ co-chord(G). This is combinatorial statement, which can be shown
by analyzing how G0 is constructed from G.
18
Remark 5.4. Let G be a 4-cycle. We can easily verify that ν(G) = 1 and co-chord(G) = 1.
Since the lower bound in Theorem 5.1 coincides with the upper bound in Theorem 5.3, we
have reg I(G)q = 2q = 2q + ν(G) − 1 = 2q + co-chord(G) − 1 for all q ≥ 1. Classes of graphs
for which the two upper and lower bounds agree were discussed in [37, Corollary 5.1].
On the other hand, the upper bound given in Theorem 5.3 can be strict. For example,
if G is C8 , then ν(G) = 2 and co-chord(G) = 3. By [10, Theorem 5.2], it is known that
reg I(G)q = 2q + 1 < 2q + co-chord(G) − 1 for all q ≥ 1.
For a special class of graphs — gap-free graphs — there is another upper bound that was
proved by Banerjee [5].
Theorem 5.5 ([5, Theorem 6.19]). Let G be a gap-free graph and let I = I(G) be its edge
ideal. Then for all q ≥ 2, we have
reg I q ≤ 2q + reg I − 1.
Remark 5.6. The bound in Theorem 5.5 is slightly weaker than the conjectural bound of
Conjecture 7.11.
5.2. Exact values. In most of the known cases where the asymptotic linear function reg I(G)q
can be computed explicitly, the lower bound in Theorem 5.1 turns out to give the exact formula. In this subsection, we describe those instances.
In [33, Theorem 3.2], the authors showed that I(G) has a linear resolution if and only
if I(G)q has a linear resolution for all q ≥ 1. Combination of their result with Fröberg’s
characterization yields to the exact value of regularity of powers of co-chordal graphs.
Theorem 5.7. Let G be a co-chordal graph and I = I(G) be its edge ideal. Then for all
q ≥ 1, we have
reg I q = 2q.
As often the case when dealing with graphs, one of the first classes to consider is that of
trees and forests.
Theorem 5.8 ([10, Theorem 4.7]). Let G be a forest and let I = I(G) be its edge ideal.
Then for all q ≥ 1, we have
reg I q = 2q + ν(G) − 1.
Sketch of proof. Thanks to the general lower bound in Theorem 5.1, it remains to establish
the upper bound
reg I q ≤ 2q + ν(G) − 1.
In fact, a more general inequality can be proved for induced subgraphs of G. Let H and K
be induced subgraphs of K such that
EH ∪ EK = EG and EH ∩ EK = ∅.
Then the required upper bound follows from the following inequality (by setting H to be the
empty graph):
reg(I(H) + I(K)q ) ≤ 2q + ν(G) − 1.
19
(5.4)
The inequality (5.4) is proved by induction on q + |VK |, making use of the short exact
sequence arising by taking quotient and colon with respect to a leaf xy in K. Note that, by
[50, Lemma 2.10], we have
(I(H) + I(K)q ) : xy = (I(H) : xy) + (I(K)q : xy) = (I(H) : xy) + I(K)q−1 .
The next natural class of graphs to consider is that of cycles and graphs containing exactly
one cycle (i.e., unicyclic graphs).
Theorem 5.9 ([10, Theorem 5.2]). Let Cn denote the n-cycle and let I = I(Cn ) be its edge
ideal. Let ν = b n3 c be the induced matching number of Cn . Then
ν + 1 if n ≡ 0, 1 (mod 3)
reg I =
ν + 2 if n ≡ 2
(mod 3)
and for any q ≥ 2, we have
reg I q = 2q + ν − 1.
Sketch of proof. The first statement was already proved in [36]. To prove the second
statement, again thanks to the general lower bound of Theorem 5.1, it remains to establish
the upper bound
reg I q ≤ 2q + ν − 1.
(5.5)
The inductive method of Theorem 3.11 once again is invoked, and it reduces the problem
to showing that for any collection of edges e1 , . . . , eq of G (not necessarily distinct), we have
reg(I q+1 : e1 . . . eq ) ≤ ν + 1.
To this end, let J be the polarization of I q+1 : e1 . . . eq . It can be seen that
J = I(H) + (xi1 yi1 , . . . , xit yit ),
where H is a graph over the vertices {x1 , . . . , xn } and x2i1 , . . . , x2it are the non-squarefree
generators of I q+1 : e1 . . . eq . By using standard short exact sequences, it can be shown that
reg(J) = reg I(H).
Observe further that H contains Cn as a subgraph, so H has a Hamiltonian cycle. Thus,
the required upper bound now follows from a more general upper bound for the regularity of
a graph admitting either a Hamiltonian path or a Hamiltonian cycle (which is the content
of [10, Theorems 3.1 and 3.2]).
The following theorem was proved in the special case of whiskered cycles by Moghimian,
Seyed Fakhari, and Yassemi [49], and then in more generality for unicyclic graphs by Alilooee,
Beyarslan and Selvaraja [2].
Theorem 5.10 ([2, Theorem 1.2], [49, Proposition 1.1]). Let G be a unicyclic graph and let
I = I(G) be its edge ideal. Then for all q ≥ 1, we have
reg I q = 2q + reg I − 2.
20
Sketch of proof. The proof is based on establishing the conjectural bound of Conjecture
7.11.(2)
reg(I q ) ≤ 2q + reg I − 2.
(5.6)
The inequality (5.6) is proved by using induction on q. It also requires a good understanding of reg(I(Cn )q , f1 , . . . , fk ) where Cn is the cycle in G and f1 , . . . , fk are the edges of G
that are not in Cn . By making use of the general lower bound of Theorem 5.1 and the first
main result of [2], namely,
ν(G) + 2 if n ≡ 2(mod 3) and ν(G \ Γ(G)) = ν(G)
reg I =
ν(G) + 1 otherwise,
equality is achieved for the latter case. (Here, Γ(G) is a well described subset of the vertices
in G.) The proof is completed by showing reg(I(G\Γ(G))q ) = 2q +ν(G) and using inequality
(5.1).
Example 5.11. Let G be the graph depicted in Figure 15. Macaulay 2 computations show
that
reg I(G) = 5, reg I(G)2 = 6, reg I(G)3 = 8, reg I(G)4 = 10, reg I(G)5 = 12.
Thus, the formula given in Theorem 5.10 does not necessarily hold for a graph containing
more than 1 cycle.
Figure 15. A bicyclic graph
A particular interesting class of graph is those for which ν(G) = 1. It is expected that for
such a graph G, powers of its edge ideal should asymptotically have linear resolutions. This
is examined in the next two theorems under various additional conditions.
Theorem 5.12 ([5, Theorem 1.2]). Let G be a gap-free and cricket-free graph, and let
I = I(G) be its edge ideal. Then for all q ≥ 2, we have
reg I q = 2q;
i.e., I q has a linear resolution.
Sketch of proof. Since I q is generated in degree 2q, it remains to show that for q ≥ 2,
reg I q ≤ 2q. By induction, making use of the inductive techniques of Theorem 3.11 (and [5,
Theorem 3.4] which proves that reg I ≤ 3), it suffices to show that for any collection of edges
e1 , . . . , eq in G, reg(I q+1 : e1 . . . eq ) ≤ 2. Let J be the polarization of I q+1 : e1 . . . eq and let
H be the simple graph associated to J. The statement is reduced to showing that reg J = 2,
or equivalently (by Theorem 4.1), that H c is a chordal graph. That is, H does not have any
induced anticycle of length at least 4.
21
By contradiction, suppose that H has an anticycle w1 , . . . , ws of length at least 4. Since
an induced anticycle of length 4 gives a gap, we may assume further that s ≥ 5. Suppose
also that e1 = xy. The proof follows from the following observations:
(1) There must be an edge between {x, y} and {w1 , w3 }; otherwise xy and w1 w3 form a
gap in G.
(2) Suppose that xw1 ∈ EG . Then neither of w2 nor wn can coincide with x.
(3) Neither of w2 nor wn can coincide with y; otherwise the other vertex (among {w2 , wn })
and w1 would be even-connected implying that w1 w2 or w1 wn is an edge in H.
(4) Either w2 or wn must be connected to x; otherwise, xy and w2 wn form a gap in G.
(5) Suppose that xw2 ∈ EG . Now, by the same line of arguments applying to {w3 , wn },
we deduce that either w3 or wn must be a neighbor of x.
(6) Suppose that xw3 ∈ EG . We arrive at a contradiction that {w1 , w3 , x, y, w2 } forms a
cricket in G.
Theorem 5.13 ([21, Theorem 4.9]). Let G be a gap-free and diamond-free graph, and let
I = I(G) be its edge ideal. Then for all q ≥ 2, we have
reg I q = 2q;
i.e., I q has a linear resolution.
Sketch of proof. The proof of this theorem is based on a good understanding of the
combinatorial structures of gap-free and diamond-free graphs. Let ω(G) denote the largest
size of a complete subgraph in G (the clique number of G). If ω(G) < 3 then G is cricketfree, and the assertion follows from Theorem 5.12. If ω(G) ≥ 4, then it is shown that the
complement of G is chordal, and the conclusion follows from Theorem 4.1.
Consider the case where ω(G) = 3. It is shown that if, in addition, G is also C5 -free then
the complement of G is either chordal or C6 , and the assertion again follows from known
results. The essential part of the proof is then to examine the structures of G when G is
gap-free, diamond-free and contains a C5 . This is where novel and interesting combinatorics
happen. It is shown that, in this case, G can be obtained from a list of ten specific graphs
via a so-called process of multiplying vertices. The proof is completed with a careful analysis
of each of these ten graphs to show that reg(I(G)q+1 : e1 . . . eq ) ≤ 2 and to employ Banerjee’s
result, Theorem 3.11.
Another large class of graphs for which the regularity of powers of their edge ideals can be
computed explicitly is that of very well-covered graphs. The following theorem was proved
by Norouzi, Seyed Fakhari, and Yassemi [54] for very well-covered graphs with an additional
condition, and then by Jayanthan and Selvaraja [38] in full generality for any very wellcovered graph.
Theorem 5.14 ([38, Theorem 5.3], [54, Theorem 3.6]). Let G be a very well-covered graph
and let I = I(G) be its edge ideal. Then for all q ≥ 1, we have
reg I q = 2q + ν(G) − 1.
22
Sketch of proof. The proof also uses inductive techniques given by Theorem 3.11 and goes
along the same line as that of previous theorems in this section. The heart of the arguments
is to verify that for any collection of edges e1 , . . . , eq in G, we have
reg(I q+1 : e1 . . . eq ) ≤ ν(G) + 1.
If J = (I q+1 : e1 . . . eq ) is squarefree then this is achieved by letting H be the graph
associated to J and establishing the following facts:
(1) H is also very well-covered;
(2) ν(H) ≤ ν(G);
(3) reg I(H) = ν(H) + 1 and reg I(G) = ν(G) + 1 (this is the content of [46, Theorem
4.12]).
The arguments are much more involved for the case where J is not squarefree. Let H be
the graph corresponding to the polarization of J. The proof is completed by an ingenious
induction on the number of vertices added to G in order to obtain H.
6. Higher Dimension
In this section we discuss the regularity of squarefree monomial ideals generated in degree
more than two. The goal, as in the case of the edge ideals, is to find bounds and/or formulas
for regularity of the ideal in terms of “its combinatorics”. As mentioned in the preliminaries,
one can view these ideals both as Stanley-Reisner ideals or edge ideals of hypergraphs.
If interpreted as Stanley-Reisner ideals, via Hochster’s formula one can potentially get all
Betti numbers of a given ideal in terms of the combinatorics of the underlying complex. The
hypergraph case, however, is far less understood compared to either the edge ideal case or
the Stanley-Reisner ideal interpretation. However, in the last decade, some results have been
proven in that direction. In some cases, general squarefree monomial ideals can be viewed
as path ideals of simple graphs, and there has been some progress for those cases in the last
few years.
Since the main topic of this survey is the edge ideals, we are not trying to be comprehensive
in this section. Our aim is to give the reader an idea of possible approaches to generalize
the results on edge ideals to higher dimensions. We split the section into three subsections;
devoted to Stanley-Reisner ideals, edge ideals of hypergraphs and path ideals respectively.
6.1. Stanley-Reisner Ideals. As mentioned earlier, any squarefree monomial ideal can be
viewed as the Stanley-Reisner ideal of some simplicial complex. This combinatorial interpretation is by far the most studied one among monomial ideals generated in higher degrees.
The first result we mention is a famous theorem by Eagon and Reiner which establishes the
relation between minimal regularity and maximal depth as well as relation between regularity
of a Stanley-Reisner ideal and the structure of a special kind of dual complex, namely, the
Alexander dual. In a sense it is similar to Fröberg’s result; this gives a classification of linear
resolution. The Alexander dual of a simplicial complex ∆ is the simplicial complex whose
faces are the complements of the nonfaces of ∆. If I is a Stanley-Reisner ideal then by I ∨
we denote the Stanley-Reisner ideal of the Alexander dual of the simplicial complex of I.
23
Theorem 6.1 ([18, Theorem 3]). Let I be a Stanley-Reisner ideal in S = K[x1 , . . . , xn ].
Then I has q-linear resolution if and only if depth of S/I ∨ has depth n − q. In particular, I
has linear resolution if and only if the Alexander dual of ∆(I) is Cohen-Macaulay.
The next result bounds regularity with another important algebraic invariant, the arithmetic degree. For a squarefree monomial ideal I, its arithmetic degree, denoted by adeg(I),
is given by number of facets in the corresponding simplicial complex.
Theorem 6.2 ([24, Theorem 3.8]). Suppose I is a Stanley-Reisner ideal with codimension
≥ 2. Then, reg I ≤ adeg(I).
6.2. Hypergraphs. Any squarefree monomial ideal is the edge ideal of a hypergraph. The
corresponding simplicial complex is the independence complex of this hypergraph and computing that is an NP-hard problem in general.
The problem of finding bounds for the regularity of graphs can be extended to hypergraphs. In [13] the authors provided a sufficient condition for a hypergraph to have regularity
≤ 3. For every vertex x of a hypergraph H, let H : x denote the simple hypergraph of all
minimal subsets A ⊂ V \ {x} such that A or A ∪ {x} is an edge of H.
Theorem 6.3 ([13, Theorem 6.4]). Let H be a simple hypergraph such that H : x is a graph
whose complement is chordal for all vertices of H. Then, reg H ≤ 3.
Any hypergraph whose all edges have same cardinality is called uniform. A d-uniform
hypergraph (i.e., a uniform hypergraph whose all edges are represented by monomials of
degree d) is called properly-connected if for any two edges E, E 0 sharing at least one vertex,
the length of the shortest path between E and E 0 is d − |E ∩ E 0 |. If the length of the shortest
path between two edges is ≥ t then they are called t-disjoint (for relevant definitions see
[29]).
Example 6.4. Consider the 4-uniform hypergraph H with edge set:
{x1 x2 x3 x4 , x1 x2 x3 x7 , x1 x2 x6 x7 , x1 x5 x6 x7 , x1 x5 x6 x8 }
There is a proper irredundant chain of length 4 from the edge E = x1 x2 x3 x4 to E0 =
x1 x5 x6 x8 . Furthermore, there is no shorter such chain. But these edges have a nonempty
intersection. So H is not properly-connected since the distance between them is 4 which is
not same as 4 − |E ∩ E0 | = 3. This hypergraph is not properly-connected. On the other
hand every finite simple graph is properly connected.
A d-uniform properly connected hypergraph H is called triangulated if for any subset A
of vertices the induced subgraph H 0 on A has a vertex x such that the induced subgraph on
its neighborhood N (x) ∪ {x} is complete. For a more formal definition and relevant notions
see [29].
Example 6.5. Simple graphs that are triangulated are precisely the chordal graphs due to
Theorem 5.2 of [29].
The following result by Hà and Van-Tuyl gives a lower bound for regularity of properly
connected hypergraphs and the bound becomes an equality if the hypergraph is triangulated.
24
Theorem 6.6 ([29, Theorem 6.5 and 6.8]). Let H be a properly-connected hypergraph. Suppose d is the common cardinality of the edges in H. Let c be the maximal number of pairwise
d + 1 disjoint edges in H. Then, reg H ≥ (d − 1)c + 1, and the equality occurs if H is
triangulated.
A 2-collage for a hypergraph is a subset C of the edges with the property that for every
edge E of the hypergraph, we can delete a vertex v of E so that E \ {v} is contained in
some edge of C. For uniform hypergraphs, the condition for a collection C of the edges to
be a 2-collage is equivalent to requiring that for any edge E not in C, there exists F ∈ C
such that the symmetric difference of E and F consists of exactly two vertices. When H
is a graph, it is straightforward to see that for any minimal 2-collage, there is a maximal
matching of the same or lesser cardinality.
The following theorem gives a formula for regularity of uniform hypergraphs in terms of
its collages.
Theorem 6.7 ([30, Theorem 1.1]). Let H be a simple d-uniform hypergraph and let c be
the minimum size of a 2 collage in H. Then, reg I(H) ≤ (d − 1)c + 1.
The next theorem is a generalization of the previous that work for all simple hypergraph.
Theorem 6.8 ([30, TheoremP1.2]). Let H be a simple hypergraph and let {m1 , ...., mc } be a
2-collage. Then, reg I(H) ≤ ci=1 |mi | − c + 1.
The next theorem gives a formula for regularity of edge ideals of clutters, which are
hypergraphs where no edge contains any other edge as a subset. A monomial ideal I has
linear quotients if the monomials that generate I can be ordered g1 , . . . , gq such that for all
1 ≤ i ≤ q − 1, ((g1 , . . . , gi ) : gi+1 ) is generated by linear forms. For further details see [51].
Theorem 6.9 ([51, Corollary 3.35]). Let H be a clutter such that I(H) has linear quotients.
Then, reg I(H) = max{|E| : E ∈ E(H)}.
Taylor resolutions, introduced in [59], are free resolutions of monomial ideals constructed
in a natural way that makes computations easy to deal with. These are in general not
minimal. The following theorem gives a formula for regularity when Taylor resolutions are
minimal. A hypergraph H is called saturated when the Taylor resolution of I(H) is minimal.
For further details see [44].
Theorem 6.10 ([44, Proposition 4.1]). For a saturated hypergraph H, we have reg I(H =
|X| − |L| + 1. Here L is the minimal monomial generating set of I(H) and X is the number
of variables that divide some minimal monomial generator.
6.3. Path Ideal. Path ideals of finite simple graphs are interesting generalizations of edge
ideals, see Section 2.2 to recall the definition of path ideals. For gap-free graphs their
regularity too tend to behave nicely.
Theorem 6.11 ([11, Theorem 1.1]). Let G be a gap-free and claw-free graph. Then for all
t ≤ 6, It has linear resolution. If further G is induced whiskered-K4 free then for all t, It
has linear resolution.
25
Let Ln denote a line of length n. The regularity of path ideals of lines have been computed
by Alilooee and Faridi in the next theorem.
Theorem 6.12 ([3, Theorem 3.2]). Let n, t, p and d be integers such that n ≥ 2, 2 ≤ t ≤
n, n = (t + 1)p + d, where p ≥ 0, 0 ≤ d ≤ t. Then, reg It (Ln ) is p(t − 1) + 1 for d < t and
(p + 1)(t − 1) + 1 for d = t.
We end this section by a result regarding a somewhat different kind of path ideals. A
tree is a graph in which there exists a unique path between every pair of distinct vertices; a
rooted tree is a tree together with a fixed vertex called the root. In particular, in a rooted
tree there exists a unique path from the root to any given vertices. We can also view a rooted
tree as a directed graph by assigning to each edge the direction that goes “away” from the
root. Let Γ be a rooted tree and It (Γ) be the squarefree monomial ideal generated by all
“directed paths” of length t − 1 in the above sense. The following theorem gives regularity
of such ideals.
In this theorem, we define lt (Γ) to be the number of leaves in Γ whose level is at least
t − 1 and pt (Γ) to be the maximal number of pairwise disjoint paths of length t in Γ (i.e.,
pt (Γ) = max{|D| | D is a set of disjoint paths of length t in Γ}).
Theorem 6.13 ([9, Theorem 3.4]). Let Γ be a rooted tree on n vertices. Then, reg It (Γ) ≤
(t − 1)[lt (Γ) + pt (Γ)] + 1.
7. Open problems and questions
We end the paper by listing a number of open problems and questions in the research
area which we hope to be solved.
Our first problem is inspired by Theorems 4.1 and 4.11. Since graphs with regularity 2
are classified. The next class of graph to examine is that of regularity 3.
Problem 7.1. Characterize graphs G for which reg I(G) = 3.
In various results, for example Theorems 4.7 and 6.3, “local” condition on the regularity
of G : x for all vertices x lead to a “global” statement on the regularity of G. We ask if
similar local conditions on G : x, for all vertices x, would also lead to a statement on the
asymptotic linear function reg I(G)q .
Question 7.2. Let G be a simple graph and let I = I(G) be its edge ideal. Suppose that
for any vertex x in G, we have reg(I : x) ≤ r. Does this imply that for any q ≥ 1,
reg I q ≤ 2q + r − 1?
As noted in Lemma 3.7, for any vertex x, the regularity of I(G) is always equal to either
the regularity of I(G) : x or the regularity of (I(G), x). It would be interesting to know for
which vertex x the equality happens one way or another.
Problem 7.3. Let G be a graph and let I = I(G) be its edge ideal. Find conditions on a
vertex x of G such that
(1) reg I = reg(I, x).
26
(2) reg I = reg(I : x).
The regularity of I(G) has been computed for several special classes of graphs (see Theorem 4.12). A particular class of graphs which is of interest is that of vertex decomposable
graphs. For a vertex decomposable graph G, the statement of Lemma 3.7 can be made
slightly more precise (see [30]), namely, there exists a vertex x such that
reg G = max{reg(G \ NG [x]) + 1, reg(G \ x)}.
For such a graph G, it is also known that the independent complex ∆(G) is shellable and
the quotient ring S/I(G) is sequentially Cohen-Macaulay.
Problem 7.4. Let G be a vertex decomposable graph. Compute reg I(G) via combinatorial
invariant of G.
As noted throughout Sections 4 and 5, the induced matching number of a graph G is
closely related to the regularity of I(G). In fact, ν(G) + 1 gives a lower bound for reg I(G)
and, more generally, 2q +ν(G)−1 gives a lower bound for reg I(G)q for any q ≥ 1. Moreover,
for many special classes of graphs the equality has been shown to hold. Thus, it is desirable
to characterize all graphs for which the equality is attained.
Problem 7.5. Characterize graphs G for which the edge ideals I = I(G) satisfy
(1) reg I = ν(G) + 1.
(2) reg I q = 2q + ν(G) − 1 ∀ q 0.
When ν(G) = 1, the answer to Problem 7.5.(2) is predicted in the following open problem.
Problem 7.6 (Francisco-Hà-Van Tuyl and Nevo-Peeva). Suppose that ν(G) = 1, i.e., Gc
has no induced 4-cycle and let I = I(G).
(1) Prove (or disprove) that reg I q = 2q for all q 0.
(2) Prove (or disprove) that reg I q+1 = reg I q + 2 for all q ≥ reg(I) − 1.
Note that examples exist in which ν(G) = 1 and reg I q 6= 2q for small values of q (cf.
[53]), so in Problem 7.6 it is necessary to consider q 0. A satisfactory solution to Problem
7.1 would be a good starting point to tackle Problem 7.6, since regularity 3 is the first open
case of the problem. In fact, in this case, reg I(G)q is expected to be linear starting at q = 2.
Problem 7.7 (Nevo-Peeva). Suppose that ν(G) = 1 and reg I(G) = 3. Then is it true that
for all q ≥ 2,
reg I(G)q = 2q?
In many known cases where the asymptotic linear function reg I(G)q can be computed, it
happens to be reg I(G)q = 2q + ν(G) − 1. We would like to see if this is the case when the
equality is already known to hold for small values of q. If this is indeed the case then how
far one must go before concluding that the equality holds for all q ≥ 1?
Problem 7.8. Let G be a graph. Find a number N such that if reg I(G)q = 2q + ν(G) − 1
for all 1 ≤ q ≤ N then, for all q ≥ 1, we have
reg I(G)q = 2q + ν(G) − 1.
27
(Computational experiments seem to suggest that N can be taken to be 2.)
A particularly related question is whether for special classes of graphs satisfying reg I(G) =
ν(G) + 1 one would have reg I(G)q = 2q + ν(G) − 1 for all q ≥ 1 (or for all q 0). Inspired
by Theorem 4.12, we raise the following question.
Question 7.9. Let G be a graph and let I = I(G) be its edge ideal. Suppose that G is of
one of the following types:
(1)
(2)
(3)
(4)
(5)
G
G
G
G
G
is
is
is
is
is
chordal;
weakly chordal;
sequentially Cohen-Macaulay bipartite;
vertex decomposable and contains no closed circuit of length 5;
(C4 , C5 )-free vertex decomposable.
Is it true that for all q 0,
reg I q = 2q + ν(G) − 1?
Note that chordal graphs and sequentially Cohen-Macaulay bipartite graphs are vertex
decomposable. Also bipartite graphs contain no closed circuit of length 5. Thus, in Question
7.9, an affirmative answer to part (4) would imply that for part (3).
It is well-known that the resolution of a monomial ideal is dependent on the characteristic
of the ground field. And yet, in all known cases where the regularity of powers of an edge
ideal can be computed, it is characteristic-independent. We would like to see examples where
this is not the case, or a confirmation that this is always the case if the regularity of the edge
ideal itself is characteristic-independent.
Problem 7.10.
(1) Find examples of graphs G for which the asymptotic linear function reg I(G)q , for
q 0, is characteristic-dependent.
(2) Suppose that reg I(G) is independent of the characteristic of the ground field. Is
the asymptotic linear function reg I(G)q , for q 0, necessarily also characteristicindependent?
When an exact formula may not be available, it is of interest to find a lower and an upper
bound. Since the lower bound in Theorem 5.1 holds for any graph, it is desirable to find
an upper bound to couple with this general lower bound. In this direction, one may either
try to prove the bound in Theorem 5.3 for any graph or to relate the regularity of powers of
I(G) to the regularity of I(G) itself.
Conjecture 7.11 (Alilooee, Banerjee, Beyarslan and Hà). Let G be any graph and let
I = I(G) be its edge ideal. Then for all q ≥ 1, we have
(1) reg I q ≤ 2q + co-chord(G) − 1.
(2) reg I q ≤ 2q + reg I − 2.
(3) reg I q+1 ≤ reg I q + 2.
Note that by Theorem 4.6, Conjecture 7.11.(3) ⇒ Conjecture 7.11.(2) ⇒ Conjecture
7.11.(1).
28
Even though, in general, reg I q is asymptotically a linear function, for small values of q,
there are examples for which reg I q > reg I q+1 . We ask if this would not be the case for edge
ideals of graphs.
Question 7.12. Let G be a graph and let I = I(G) be its edge ideal. Is the function reg I q
increasing for all q ≥ 1?
In investigating the asymptotic linear function reg I q , it is also of interest to know the
smallest value q0 starting from which reg I q attains its linear form. In known cases where the
regularity of I(G)q can be computed explicitly for all q ≥ 1, we have q0 ≤ 2. Computational
experiments seem to suggest that this is indeed always the case.
Question 7.13. Let G be a graph and let q0 be the least integer such that reg I(G)q is a
linear function for all q ≥ q0 . Is it true that q0 ≤ 2?
A much weaker question, yet still very interesting, would be whether q0 ≤ reg I(G)? —
This question is still wide open! A significant step toward answering Question 7.13 is to give
a uniform bound for q0 (i.e., a bound that does not depend on G).
Problem 7.14. Find a number N (which may depend on n and m) such that for any graph
G over n vertices and m edges, we have q0 ≤ N .
References
[1] A. Alilooee and A. Banerjee, Powers of edge ideals of regularity three bipartite graphs. J. Commut.
Algebra 9 (2017), no. 4, 441-454.
[2] A. Alilooee, S. Beyarslan and S. Selvaraja, Regularity of powers of unicyclic graphs. Preprint (2017),
arXiv:1702.00916.
[3] A. Alilooee and S. Faridi, On the resolutions of path ideals of cycles, Communications in Algebra 43
(2015), no. 12, 5413-5433.
[4] A. Bagheri, M. Chardin and H.T. Hà, The eventual shape of Betti tables of powers of ideals. Math.
Res. Lett. 20 (2013), no. 6, 1033-1046.
[5] A. Banerjee, The regularity of powers of edge ideals. J. Algebraic Combin. 41 (2015), no. 2, 303-321.
[6] C. Berge, Hypergraphs: combinatorics of finite sets. North-Holland, New York, 1989.
[7] T. Biyikoglu and Y. Civan, Vertex decomposable graphs, codismantlability, Cohen-Macaulayness and
Castelnuovo-Mumford regularity. Electron. J. Combin. 21 (2014), no. 1, Paper 1.1, 17 pp.
[8] Anders Björner, Topological methods, Handbook of combinatorics, Vol. 1, 2, Elsevier, Amsterdam,
1995, pp. 1819-1872
[9] R. Bouchat, H. T. Hà and A. O’Keefe, Path ideals of rooted trees and their graded Betti numbers. J.
Combin. Theory Ser. A 118 (2011), 2411-2425.
[10] S. Beyarslan, H.T. Hà and T.N. Trung, Regularity of powers of forests and cycles. J. Algebraic
Combin. 42 (2015), no. 4, 1077-1095.
[11] A. Banerjee, Regularity of path ideals of gap free graphs, J. Pure and Applied Alg. 221 (2017), no.
10, 2409-2419.
[12] W. Bruns and J. Herzog, Cohen-Macaulay rings. Cambridge Studies in Advanced Mathematics, 39.
Cambridge University Press, Cambridge, 1993.
[13] G. Caviglia, H. T. Hà, J. Herzog, M. Kummini, N.Terai, N. V. Trung, Depth and regularity modulo
a principal ideal. Preprint (2017), arXiv:1706.09675.
[14] M. Chardin, Some results and questions on Castelnuovo-Mumford regularity. Syzygiez and Hilbert
functions. Lect. Notes Pure Appl. Math., 254, 1-40. Chapman & Hall/CRC, 2007.
[15] S.D. Cutkosky, J. Herzog, and N.V. Trung, Asymptotic behaviour of the Castelnuovo-Mumford regularity. Composito Mathematica 118 (1999), 243-261.
29
[16] K. Cameron and T. Walker, The graphs with maximum induced matching and maximum matching
the same size. Discrete Math. 299 (2005), 49-55.
[17] H. Dao, C. Huneke and J. Schweig, Bounds on the regularity and projective dimension of ideals
associated to graphs. J. Algebraic Combin. 38 (2013), no. 1, 37-55.
[18] J.A. Eagon and V. Reiner, Resolutions of Stanley-Reisner rings and Alexander duality, J. Pure and
Appl. Algebra 130 (1998), no. 3, 265-267.
[19] D. Eisenbud, Commutative Algebra: with a View Toward Algebraic Geometry. Springer-Verlag, New
York, 1995.
[20] D. Eisenbud and S. Goto, Linear free resolutions and minimal multiplicity. J. Algebra 88 (1984), no.
1, 89-133.
[21] N. Erey, Powers of edge ideals with linear resolutions. Preprint (2017), arXiv:1703.01561.
[22] S. Faridi, The facet ideal of a simplicial complex. Manuscripta Math. 109 (2002), no. 2, 159-174.
[23] O. Fernàndez-Ramos and P. Gimenez, Regularity 3 in edge ideals associated to bipartite graphs. J.
Algebraic Combin. 39 (2014), no. 4, 919-937.
[24] A. Frühbis-Krüger and N. Terai, Bounds for the regularity of monomial ideals. LE MATEMATICHE
LIII, Supplemento, pp. 83-97, 2007.
[25] R. Fröberg, On Stanley-Reisner rings. Topics in algebra, Part 2 (Warsaw, 1988), 57-70, Banach Center
Publ., 26, Part 2, PWN, Warsaw, 1990.
[26] C. A. Francisco and A. V. Tuyl, Sequentially Cohen-Macaulay edge ideals, Proc. Amer. Math. Soc.
135 (2007), no. 8, 23-37 (electronic).
[27] H.T. Hà, Regularity of squarefree monomial ideals. Connections between algebra, combinatorics, and
geometry, 251-276, Springer Proc. Math. Stat., 76, Springer, New York, 2014.
[28] H.T. Hà and A. Van Tuyl, Resolutions of square-free monomial ideals via facet ideals: a survey. Algebra, geometry and their interactions, 91-117, Contemp. Math., 448, Amer. Math. Soc., Providence,
RI, 2007.
[29] H. T. Hà and A. Van Tuyl, Monomial ideals, edge ideals of hypergraphs, and their minimal graded
free resolutions. J. Algebraic Combin. 27 (2008), no. 2, 215-245.
[30] H.T. Hà and R. Woodroofe, Results on the regularity of square-free monomial ideals. Adv. in Appl.
Math. 58 (2014), 21-36.
[31] J. Herzog, A generalization of the Taylor complex construction. Comm. Algebra 35 (2007), no. 5,
1747-1756.
[32] J. Herzog and T. Hibi, Monomial ideals. GTM 260, Springer-Verlag, 2011.
[33] J. Herzog, T. Hibi, X. Zheng, Monomial ideals whose powers have a linear resolution. Math. Scand.
95 (2004), no. 1, 23-32.
[34] T. Hibi, A. Higashitani, K. Kimura and A. B. O’Keefe, Algebraic study on Cameron-Walker graphs,
J. Algebra 422 (2015), 257-269.
[35] M. Hochster, Cohen-Macaulay rings, combinatorics, and simplicial complexes. Proc. of the 2nd Oklahoma Conf., Lecture Notes in Pure Appl. Math. 26 (1977) Marcel Dekker, 171-224.
[36] S. Jacques, Betti numbers of graph ideals. Ph.D. Thesis, University of Sheffield, 2004.
arXiv:math.AC/0410107.
[37] A.V. Jayanthan, N. Narayanan and S. Selvaraja, Regularity of powers of bipartite graphs. J. Algebraic
Combin. 47 (2018), no.1, 17-38
[38] A.V. Jayanthan, S. Selvaraja, Asymptotic behavior of Castelnuovo-Mumford regularity of edge ideals
of very well-covered graphs. Preprint (2017), arXiv:1708.06883.
[39] G. Kalai and R. Meshulam, Intersections of Leray complexes and regularity of monomial ideals. J.
Combin. Theory Ser. A 113 (2006), no. 7, 1586-1592.
[40] F. Khosh-Ahang and S. Moradi, Regularity and projective dimension of edge ideal of C5 -free vertex
decomposable graphs. Proc. Amer. Math. Soc. 142 (2014), no. 5, 1567-1576.
[41] M. Katzman, Characteristic-independence of Betti numbers of graph ideals. J. Combin. Theory Ser.
A 113 (2006), no. 3, 435-454.
[42] V. Kodiyalam, Asymptotic behaviour of Castelnuovo-Mumford regularity. Proceedings of the American Mathematical Society, 128 (1999), no. 2, 407-411.
30
[43] M. Kummini, Regularity, depth and arithmetic rank of bipartite edge ideals. J. Algebraic Combin.
30 (2009), no. 4, 429-445.
[44] K-N. Lin and J. McCullough, Hypergraphs and Regularity of Square-free Monomial Ideals, Int. J.
Algebra Comput. 23 (2013), 1573-1590.
[45] G. Lyubeznik, The minimal non-Cohen-Macaulay monomial ideals, J. Pure Appl. Algebra 51 (1988),
261-266.
[46] M. Mahmoudi, A. Mousivand, M. Crupi, G. Rinaldo, N. Terai, and S. Yassemi, Vertex decomposability and regularity of very well-covered graphs. J. Pure Appl. Algebra 215 (2011), no. 10,
2473-2480.
[47] E. Miller and B. Sturmfels, Combinatorial Commutative Algebra. GTM 227, Springer-Verlag, 2004.
[48] N.C. Minh and N.V. Trung, Cohen-Macaulayness of monomial ideals and symbolic powers of StanleyReisner ideals. Adv. Math. 226 (2011), 1285-1306.
[49] M. Moghimian, S.A. Seyed Fakhari, S. Yassemi, Regularity of powers of edge ideal of whiskered cycles.
Comm. Algebra 45 (2017), no. 3, 1246-1259.
[50] S. Morey, Depths of powers of the edge ideal of a tree. Comm. Algebra 38 (2010), no. 11, 4042-4055.
[51] S. Morey, R.H. Villarreal, Edge ideals: algebraic and combinatorial properties. Progress in commutative algebra 1, 85-126, de Gruyter, Berlin, 2012.
[52] E. Nevo, Regularity of edge ideals of C4 -free graphs via the topology of the lcm-lattice. J. Combin.
Theory Ser. A 118 (2011), 491-501.
[53] E. Nevo and I. Peeva, C4 -free edge ideals. J. Algebraic Combin. 37 (2013), no. 2, 243-248.
[54] P. Norouzi, S.A. Seyed Fakhari, S. Yassemi, Regularity of powers of edge ideals of very well-covered
graphs. Preprint (2017), arXiv:1707.04874.
[55] I. Peeva, Graded syzygies. Algebra and Applications, 14. Springer-Verlag London, Ltd., London,
2011.
[56] R. Stanley, Combinatorics and Commutative Algebra. Progress in Mathematics, 41. Birkhäuser
Boston, Inc., Boston, MA, 1983.
[57] S. Sullivant, Combinatorial symbolic powers. J. Algebra 319 (2008), no. 1, 115-142.
[58] Y. Takayama, Combinatorial characterizations of generalized Cohen-Macaulay monomial ideals, Bull.
Math. Soc. Sci. Math. Roumanie (N.S.) 48 (2005), 327-344.
[59] D. K. Taylor, Ideals generated by monomials in an R-sequence, ProQuest LLC, Ann Arbor, MI,
Thesis (Ph.D)-The University of Chicago, 1996
[60] N.V. Trung and H. Wang, On the asymptotic behavior of Castelnuovo-Mumford regularity. J. Pure
Appl. Algebra 201 (2005), no. 1-3, 42-48.
[61] R. H. Villarreal, Monomial algebras. Monographs and Textbooks in Pure and Applied Mathematics,
238. Marcel Dekker, Inc., New York, 2001.
[62] A. Van Tuyl, Sequentially Cohen-Macaulay bipartite graphs: vertex decomposability and regularity.
Arch. Math. (Basel) 93 (2009), no. 5, 451-459.
[63] G. Wegner, d-collapsing and nerves of families of convex sets. Arch. Math. (Basel) 26 (1975), 317-321.
[64] R. Woodroofe, Matchings, coverings, and Castelnuovo-Mumford regularity. J. Commut. Algebra 6
(2014), no. 2, 287-304.
[65] Russ Woodroofe, Vertex decomposable graphs and obstructions to shellability. Proc. Amer. Math.
Soc., 137 (2009), no. 10, 3235-3246.
[66] X. Zheng, Resolutions of facet ideals, Comm. Algebra 32 (2004) 2301-2324.
31
Ramakrishna Mission Vivekenanda University, Belur, West Bengal, India
E-mail address: [email protected]
URL: https://http://maths.rkmvu.ac.in/∼arindamb/
University of South Alabama, Department of Mathematics and Statistics, 411 University
Boulevard North, Mobile, AL 36688-0002, USA
E-mail address: [email protected]
URL: https://selvikara.wordpress.com/selvi/
Tulane University, Department of Mathematics, 6823 St. Charles Ave., New Orleans,
LA 70118, USA
E-mail address: [email protected]
URL: http://www.math.tulane.edu/∼tai/
32
| 0math.AC
|
Minimum length RNA folding trajectories
P. Clote∗
A.H. Bayegan
arXiv:1802.06328v1 [cs.DS] 18 Feb 2018
Biology Department, Boston College, Chestnut Hill, MA
Abstract
Background: Existent programs for RNA folding kinetics, such as Kinefold, Kinfold and KFOLD,
implement the Gillespie algorithm to generate stochastic folding trajectories from an initial structure s
to a target structure t, in which each intermediate secondary structure is obtained from its predecessor
by the application of a move from a given move set. The Kinfold move set M S1 [resp. M S2 ] allows the
addition or removal [resp. addition, removal or shift] of a single base pair. Define the M S1 [resp. M S2 ]
distance between secondary structures s and t to be the minimum path length to refold s to t, where
a move from M S1 [resp. M S2 ] is applied in each step. The M S1 distance between s and t is trivially
equal to the cardinality of the symmetric difference of s and t, i.e the number of base pairs belonging to
one structure but not the other; in contrast, the computation of M S2 distance is highly non-trivial.
Results: We describe algorithms to compute the shortest M S2 folding trajectory between any two given
RNA secondary structures. These algorithms include an optimal integer programming (IP) algorithm,
an accurate and efficient near-optimal algorithm, a greedy algorithm, a branch-and-bound algorithm,
and an optimal algorithm if one allows intermediate structures to contain pseudoknots. A 10-fold slower
version of our IP algorithm appeared in WABI 2017; the current version exploits special treatment of
closed 2-cycles.
Our optimal IP [resp. near-optimal IP] algorithm maximizes [resp. approximately maximizes] the
number of shifts and minimizes [resp. approximately minimizes] the number of base pair additions and
removals by applying integer programming to (essentially) solve the minimum feedback vertex set (FVS)
problem for the RNA conflict digraph, then applies topological sort to tether subtrajectories into the
final optimal folding trajectory.
We prove NP-hardness of the problem to determine the minimum barrier energy over all possible
M S2 folding pathways, and conjecture that computing the M S2 distance between arbitrary secondary
structures is NP-hard. Since our optimal IP algorithm relies on the FVS, known to be NP-complete
for arbitrary digraphs, we compare the family of RNA conflict digraphs with the following classes of
digraphs – planar, reducible flow graph, Eulerian, and tournament – for which FVS is known to be either
polynomial time computable or NP-hard.
Conclusion: This paper describes a number of optimal and near-optimal algorithms to compute the
shortest M S2 folding trajectory between any two secondary structures. Source code for our algorithms
is available at http://bioinformatics.bc.edu/clotelab/MS2distance/.
1
Background
RNA secondary structure is known to form a scaffold for tertiary structure formation [5]. Moreover, secondary structure can be efficiently predicted with reasonable accuracy by using either machine learning with
stochastic context-free grammars [23, 35, 39], provided that the training set is sufficiently large and representative, or by using ab initio physics-based models with thermodynamics-based algorithms [27, 25]. Since the
latter approach does not depend on any form of homology modeling, it has been successfully used for synthetic RNA molecular design [47, 9, 16], to predict microRNA binding sites [32], to discover noncoding RNA
genes [44], in simulations to study molecular evolution [3, 42, 36, 15] and in folding kinetics [13, 45, 37, 11].
Software to simulate RNA secondary structure folding kinetics, such as Kinfold and KFOLD, implement the
Gillespie algorithm to simulate the moves from one structure to another, for a particular move set. At the
∗ Correspondence
[email protected]. Research supported in part by National Science Foundation grant DBI-1262439.
1
elementary-step resolution, two move sets have extensively been studied – the move set M S1 which allows
the addition or removal of a single base pair, and the move set M S2 , which allows the addition, removal or
shift of a single base pair, where a shift move modifies only one of the two positions in a base pair, as shown
in Figure 1.
In simulation studies related to RNA secondary structure evolution, the structural distance between
two secondary structures s, t is often measured by the base pair distance, denoted dBP (s, t), defined to
be the cardinality of the symmetric difference, |s 4 t| = |s − t| + |t − s|, i.e. the number of base pairs
belonging to s but not t, plus the number of base pairs belonging to t but not s. In studies concerning RNA
folding kinetics, the fast, near-optimal algorithm RNAtabupath [10] and the much slower, but exact (optimal)
Barriers algorithm [25] can be used to determine M S1 folding trajectories that minimize the barrier energy,
defined as the maximum of the (Turner) free energy difference between an intermediate structure and the
initial structure. Thermodynamics-based software such as Kinfold, RNAtabupath, and KFOLD use the nearest
neighbor free energy model [41] whose energy parameters are inferred from optical melting experiments. In
contrast, the two theorems below concern the Nussinov energy model [29], which assigns −1 per base pair
and ignores entropy. Folding trajectories s = s0 , s1 , . . . , sm = t from s to t may either be direct, whereby
each intermediate structure si is required to contain only base pairs from s ∪ t, or indirect, without this
restriction. Note that indirect pathways may be energetically more favorable, though longer, than direct
pathways, and that the problem of constructing an energetically optimal direct folding pathway is NP-hard.
Indeed, the following theorem is proven in [40].
Theorem 1 (Maňuch et al. [40]).
With respect to the Nussinov energy model, it is NP-hard to determine, for given secondary structures s, t
and integer k, whether there exists a direct M S1 folding trajectory from s to t with energy barrier at most k.
By an easy construction, we can show an analogous result for M S2 folding pathways. First, we define
a direct M S2 folding pathway from secondary structure s to secondary structure t to be a folding pathway
s = s0 , s1 , . . . , sn = t where each intermediate structure si is obtained from si−1 by removing a base pair
that belongs to s, adding a base pair that belongs to t, or shifting a base pair belonging to s into a base pair
belonging to t.
Theorem 2. With respect to the Nussinov energy model, it is NP-hard to determine, for given secondary
structures s, t and integer k, whether there exists a direct M S2 folding trajectory from s to t with energy
barrier at most k.
Proof. Given secondary structures s, t for an RNA sequence a = a1 , . . . , an , without loss of generality we can
assume that s, t share no common base pair (otherwise, a minimum energy folding trajectory for s − (s ∩ t)
and t − (s ∩ t) yields a minimum energy folding trajectory for s, t.) Define the corresponding secondary
structures
s0 = {(2i, 2j) : (i, j) ∈ s}
t0 = {(2i − 1, 2j − 1) : (i, j) ∈ t}
a02i = ai = a02i−1
for each 1 ≤ i ≤ n
a0 = a01 , a02 , . . . , a02n
In other words, the sequence a0 = a1 , a1 , a2 , a2 , . . . , an , an is obtained by duplicating each nucleotide of a,
and placing each copy beside the original nucleotide; s0 [resp. t0 ] is obtained by replacing each base pair
(i, j) ∈ s by the base pair (2i, 2j) ∈ s0 [resp. (2i − 1, 2j − 1) ∈ t0 . Since there are no base-paired positions
that are shared between s0 and t0 , no shift moves are possible, thus any direct M S2 folding pathway from s0
to t0 immediately yields a corresponding direct M S1 folding pathway from s to t. Since the Nussinov energy
of any secondary structure equals −1 times the number of base pairs, it follows that barrier energy of the
direct M S2 pathway from s0 to t0 is identical to that of the corresponding direct M S1 pathway from s to t.
Since M S1 direct barrier energy is an NP-hard problem by Theorem 1, it now follows that the M S2 barrier
energy problem is NP-hard.
Shift moves, depicted in Figure 1, naturally model both helix zippering and defect diffusion, depicted in
Figure 2 and described in [31]. However, shift moves have rarely been considered in the literature, except in
2
the context of folding kinetics [13]. For instance, presumably due to the absence of any method to compute
M S2 distance, Hamming distance is used as a proxy for M S2 distance in the work on molecular evolution of
secondary structures appearing in [36] – see also [43], where Hamming distance is used to quantify structural
diversity in defining phenotypic plasticity.
In this paper, we introduce the first algorithms to compute the M S2 distance between two secondary
structures. Although M S1 distance, also known as base pair distance, is trivial to compute, we conjecture
that M S2 distance is NP-hard, where this problem can be formalized as the problem to determine, for any
given secondary structures s, t and integer m, whether there is an M S2 trajectory s = s0 , s1 , . . . , sm = t
of length ≤ m. We describe an optimal (exact, but possibly exponential time) integer programming (IP)
algorithm, a fast, near-optimal algorithm, an exact branch-and-bound algorithm, and a greedy algorithm.
Since our algorithms involve the feedback vertex set problem for RNA conflict digraphs, we now provide a
bit of background on this problem.
Throughout, we are exclusively interested in directed graphs, or digraphs, so unless otherwise indicated, all
graphs are assumed to be directed. Any undefined graph-theoretic concepts can be found in the monograph
by Bang-Jensen and Gutin [1]. Given a directed graph G = (V, E), a feedback vertex set (FVS) is a subset
V 0 ⊆ V which contains at least one vertex from every directed cycle in G, thus rendering G acyclic. Similarly,
a feedback arc set (FAS) is a subset E 0 ⊆ E which contains at least one directed edge (arc) from every directed
cycle in G. The FVS [resp. FAS] problem is the problem to determine a minimum size feedback vertex set
[resp. feedback arc set] which renders G acyclic. The FVS [resp. FAS] problem can be formulated as a
decision problem as follows. Given an integer k and a digraph G = (V, E), determine whether there exists a
subset V 0 ⊆ V of size ≤ k [resp. E 0 ⊆ E of size ≤ k], such that every directed cycle contains a vertex in V 0
[resp. an edge in E 0 ].
In Proposition 10.3.1 of [1], it is proved that FAS and FVS have the same computational complexity,
within a polynomial factor. In Theorem 10.3.2 of [1], it is proved that the FAS problem is NP-complete –
indeed, this problem appears in the original list of 21 problems shown by R.M. Karp to be NP-complete
[22]. Note that Proposition 10.3.1 and Theorem 10.3.2 imply immediately that the FVS problem is NPcomplete. In Theorem 10.3.3 of [1], it is proved that the FAS problem is NP-complete for tournaments,
where a tournament is a digraph G = (V, E), such that there is a directed edge from x to y, or from y to
x, for every pair of distinct vertices x, y ∈ V . In [30], it is proved that the FAS for Eulerian digraphs is
NP-complete, where an Eulerian digraph is characterized by the property that the in-degree of every vertex
equals its out-degree. In Theorem 10.3.15 of [1], it is proved that FAS can be solved in polynomial time for
planar digraphs, a result originally due to [26]. In [33], a polynomial time algorithm is given for the FAS
for reducible flow graphs, a type of digraph that models programs without any GO TO statements (see [19]
for a characterization of reducible flow graphs). There is a long history of work on the feedback vertex set
and feedback arc set problems, both for directed and undirected graphs, including results on computational
complexity as well as exact and approximation algorithms for several classes of graphs – see the survey [12]
for an overview of such results.
The plan of the paper is now as follows. In Section 2, we present the graph-theoretic framework for
our overall approach and describe a simple, fast algorithm to compute the pseudoknotted M S2 distance,
or pk-M S2 distance, between structures s, t. By this we mean the minimum length of an M S2 folding
trajectory between s and t, if intermediate pseudoknotted structures are allowed. We show that the pk-M S2
distance between s and t, denoted by dpk−M S2 (s, t), is approximately equal to one-half the Hamming distance
dH (s, t) between s and t. This result can be seen as justification, ex post facto, for the use of Hamming
distance in the investigation of RNA molecular evolution [36]. In Section 3, we describe an exact integer
programming (IP) algorithm which enumerates all directed cycles, then solves the feedback vertex problem
for the collection of RNA conflict digraphs, as described in Section 3.1. Our IP algorithm is not a simple
reduction to the feedback vertex set (FVS) problem; however, since the complexity of FVS/FAS is known
for certain classes of digraphs, we take initial steps towards the characterization of RNA conflict digraphs.
Our optimal IP algorithm is much faster than a branch-and-bound algorithm, but it can be too slow to be
practical to determine M S2 distance between the minimum free energy (MFE) secondary structure and a
(Zuker) suboptimal secondary structure for some sequences from the Rfam database [28]. For this reason,
in Section 4 we present a fast, near-optimal algorithm, and in Section 5, we present benchmarking results to
compare various algorithms of the paper.
Since we believe that further study of RNA conflict digraphs may lead to a solution of the question
3
whether M S2 distance is NP-hard, in Appendix A, all types of directed edge that are possible in an RNA
conflict digraph are depicted. Appendix B presents details on minimum length pseudoknotted M S2 folding
pathways, used to provide a lower bound in the branch-and-bound algorithm of Appendix C. Appendix D
presents pseudocode for a greedy algorithm.
All algorithms described in this paper have been implemented in Python, and are publicly available
at bioinformatics.bc.edu/clotelab/MS2distance. Our software uses the function simple cycles(G)
from the software NetworkX https://networkx.github.io/documentation/networkx-1.9/reference/
generated/networkx.algorithms.cycles.simple_cycles.html, and the integer programming (IP) solver
Gurobi Optimizer version 6.0 http://www.gurobi.com,2014.
2
M S2 distance between possibly pseudoknotted structures
In this section, we describe a straightforward algorithm to determine the M S2 -distance dpk−M S2 (s, t) between
any two structures s, t of a given RNA sequence a1 , . . . , an , where dpk−M S2 (s, t) is defined to be length of a
minimal length trajectory s = s0 , s1 , . . . , sm = t, where intermediate structures si may contain pseudoknots,
but do not contain any base triples. This variant is called pk-M S2 distance. Clearly, the pk-M S2 distance
is less than or equal to the M S2 distance. The purpose of this section is primarily to introduce some of the
main concepts used in the remainder of the paper. Although the notion of secondary structure is well-known,
we give three distinct but equivalent definitions, that will allow us to overload secondary structure notation
to simplify presentation of our algorithms.
Definition 3 (Secondary structure as set of ordered base pairs). Let [1, n] denote the set {1, 2, . . . , n}. A
secondary structure for a given RNA sequence a1 , . . . , an of length n is defined to be a set s of ordered pairs
(i, j), with 1 ≤ i < j ≤ n, such that the following conditions are satisfied.
1. Watson-Crick and wobble pairs: If (i, j) ∈ s, then ai aj ∈ {GC, CG, AU, U A, GU, U G}.
2. No base triples: If (i, j) and (i, k) belong to s, then j = k; if (i, j) and (k, j) belong to s, then i = k.
3. Nonexistence of pseudoknots: If (i, j) and (k, `) belong to s, then it is not the case that i < k < j < `.
4. Threshold requirement for hairpins: If (i, j) belongs to s, then j − i > θ, for a fixed value θ ≥ 0; i.e.
there must be at least θ unpaired bases in a hairpin loop. Following standard convention, we set θ = 3
for steric constraints.
Without risk of confusion, it will be convenient to overload the concept of secondary structure s with
two alternative, equivalent notations, for which context will determine the intended meaning.
Definition 4 (Secondary structure as set of unordered base pairs). A secondary structure s for the RNA
sequence a1 , . . . , an is a set of unordered pairs {i, j}, with 1 ≤ i, j ≤ n, such that the corresponding set of
ordered pairs
def
{i, j}< = (min(i, j), max(i, j))
(1)
satisfies Definition 3.
Definition 5 (Secondary structure as an integer-valued
function). A secondary
n
o structure s for a1 , . . . , an is
a function s : [1, . . . , n] → [0, . . . , n], such that {i, s[i]}< : 1 ≤ i ≤ n, s[i] 6= 0 satisfies Definition 3; i.e.
s[i] =
0
j
if i is unpaired in s
if (i, j) ∈ s or (j, i) ∈ s
(2)
Definition 6 (Secondary structure distance measures). Let s, t be secondary structures of length n. Base
pair distance is defined by equation (3) below, and Hamming distance is defined by equation (4) below.
dBP (s, t) = |{(x, y) : ((x, y) ∈ s ∧ (x, y) 6∈ t) ∨ ((x, y) ∈ t ∧ (x, y) 6∈ s)}|
dH (s, t) = |{i ∈ [1, n] : s[i] 6= t[i]}|
4
(3)
(4)
Throughout this section, the term pseudoknotted structure is taken to mean a set of ordered pairs [resp.
unordered pairs resp. function], which satisfies conditions 1,2,4 (but not necessarily 3) of Definition 3. Given
structure s on RNA sequence {a1 , . . . , an }, we say that a position x ∈ [1, n] is touched by s if x belongs to
a base pair of s, or equivalently s[x] 6= 0. For possibly pseudoknotted structures s, t on {a1 , . . . , an }, we
partition the set [1, n] into disjoint sets A,B,C,D as follows. Let A be the set of positions that are touched
by both s and t, yet do not belong to the same base pair in s and t, so
A = {i ∈ [1, n] : s[i] 6= 0, t(i) 6= 0, s[i] 6= t[i]}
(5)
Let B be the set of positions that are touched by either s or t, but not by both, so
B = {i ∈ [1, n] : (s[i] 6= 0, t[i] = 0) ∨ (s[i] = 0, t[i] 6= 0)}
(6)
Let C be the set of positions touched by neither s nor t, so
C = {i ∈ [1, n] : s[i] = 0 = t[i]}
(7)
Let D be the set of positions that belong to the same base pair in both s and t, so
D = {i ∈ [1, n] : s[i] 6= 0, t[i] 6= 0, s[i] = t[i]}
(8)
We further partition A ∪ B into a set of maximal paths and cycles, in the following manner. Define an
undirected, vertex-colored and edge-colored graph G = (V, E), whose vertex set V is equal to the set A ∪ B
of positions that are touched by either s or t, but not by a common base pair in (s ∩ t), and whose edge set
E = (s − t) ∪ (t − s) = (s ∪ t) − (s ∩ t) consists of undirected edges between positions that are base-paired
together. Color edge {x, y} green if the base pair (x, y) ∈ s − t and red if (x, y) ∈ t − s. Color vertex x yellow
if x is incident to both a red and green edge, green if x is incident to a green edge, but not to any red edge,
red if x is incident to a red edge, but not to any green edge. The connected components of G can be classified
into 4 types of (maximal) paths and one type of cycle (also called path of type 5): type 1 paths have two
green end nodes, type 2 paths have a green end node x and a red end node y, where x < y, type 3 paths
have a red end node x and a green end node y, where x < y, type 4 paths have two red end nodes, and type
5 paths (cycles) have no end nodes. These are illustrated in Figure 3. Note that all nodes of a cycle and
interior nodes of paths of type 1-4 are yellow, while end nodes (incident to only one edge) are either green
or red. If X is a connected component of G, then define the restriction of s [resp. t] to X, denoted by s X
[resp. t X], to be the set of base pairs (i, j) in s [resp. t] such that i, j ∈ X. With this description, most
readers will be able to determine a minimum length pseudoknotted folding pathway from s X to t X,
where X is a connected component of G. For instance, if X is a path of type 2 or 3, then a sequence of shift
moves transforms s X into t X, beginning with a shift involving the terminal green node. Further details
can be found in Appendix B. The formal definitions given below are necessary to provide a careful proof of
the relation between Hamming distance and pseudoknotted M S2 distance, also found in Appendix B.
Definition 7. Let s, t be (possibly pseudoknotted) structures on the RNA sequence a1 , . . . , an . For i, j ∈
[1, n], define i ∼ j if s[i] = j or t[i] = j, and let ≡ be the reflexive, transitive closure of ∼. Thus i ≡ j if
i = j, or i = i1 ∼ i2 ∼ · · · ∼ im = j for any m ≥ 1. For i ∈ [1, n], let [i] denote the equivalence class of i,
i.e. [i] = {j ∈ [1, n] : i ≡ j}.
It follows that i ≡ j if and only if i is connected to j by an alternating red/green path or cycle. The
equivalence classes X with respect to ≡ are maximal length paths and cycles, as depicted in Figure 3.
Moreover, it is easy to see that elements of A either belong to cycles or are found at interior nodes of paths,
while elements of B are found exclusively at the left or right terminal nodes of paths.
Note that odd-length cycles cannot exist, due to the fact that a structure cannot contain base triples –
see condition 2 of Definition 3. Moreover, even-length cycles can indeed exist – consider, for instance, the
structure s, whose only base pairs are (1, 15) and (5, 10), and the structure t, whose only base pairs are (1, 5)
and (10, 15). Then we have the red/green cycle 1 → 5 → 10 → 15 → 1, consisting of red edge 1 → 5, since
(1, 5) ∈ t, green edge 5 → 10, since (5, 10) ∈ s, red edge 10 → 15, since (10, 15) ∈ t, and green edge 15 → 1,
since (1, 15) ∈ s.
5
From the discussion before Definition 7, it follows that A in equation (5) consists of the nodes of every cycle
together with all interior (yellow) nodes of paths of type 1-4. Moreover, we can think of B in equation (6)
as consisting of all path end nodes, i.e. those that have only one incident edge. Let B1 ⊆ B [resp. B2 ⊆ B]
denote the set of elements of B that belong to type 1 paths [resp. type 4 paths] of length 1, i.e. positions
incident to isolated green [resp. red] edges that correspond to base pairs (i, j) ∈ s where i, j are not touched
by t [resp. (i, j) ∈ t where i, j are not touched by s]. Let B0 = B − B1 − B2 be the set of end nodes of a
path of length 2 or more. Letting BP1 [resp. BP2 ] denote the set of base pairs (i, j) that belong to s and
are not touched by t [resp. belong to t and are not touched by s], we can formalize the previous definitions
as follows.
B1 = {i ∈ [1, n] : ∃j [{i, j} ∈ s, t(i) = 0 = t(j)]
(9)
B2 = {i ∈ [1, n] : ∃j [{i, j} ∈ t, s(i) = 0 = s(j)]
(10)
B0 = B − (B1 ∪ B2 )
(11)
BP1 = {(i, j) ∈ s : t[i] = 0 = t[j]}
(12)
BP2 = {(i, j) ∈ t : s[i] = 0 = s[j]}
(13)
In Appendix B, it is proved that pk-M S2 distance between s X and t X for any maximal path X is
c; in contrast, pk-M S2 distance between s X and t X for any
equal to Hamming distance b dH (sX,tX)
2
cycle X is equal to b dH (sX,tX)
c
+
1.
It
follows that
2
dpk−M S2 (s, t) = b
dH (s, t)
c
2
(14)
if and only if there are no type 5 paths (i.e. cycles). This result justifies ex post facto the use of Hamming
distance in the investigation of RNA molecular evolution [36, 43]. We also have the following.
Lemma 8. Let s, t be two arbitrary (possibly pseudoknotted) structures for the RNA sequence a1 , . . . , an ,
and let X1 , . . . , Xm be the equivalence classes with respect to equivalence relation ≡ on A ∪ B. Then the
pk-M S2 distance between s and t is equal to
dpk−M S2 (s, t) =
m
X
max |s Xi |, |t Xi |
i=1
This lemma is useful, since the pk-M S2 distance provides a lower bound for the M S2 distance between
any two secondary structures, and hence allows a straightforward, but slow (exponential time) branch-andbound algorithm to be implemented for the exact M S2 distance – pseudocode for the branch-and-bound
algorithm is given in Section C of the Appendix. To compute pk-M S2 distance, we remove those base pairs
in s − t that are not touched by t, compute the equivalence classes (connected components) X on the set of
positions belonging to the remaining base pairs (provided that the position does not belong to a common
base pair of both s and t), then determine for each X a minimum length pk-M S2 folding pathway from
s X to t X. The formal pseudocode follows.
Algorithm 9 (pk-M S2 distance). M S2 -path length between two possibly pseudoknotted structures s, t.
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
remove from s all base pairs of BP1
numM oves = |BP1 |
Q = A ∪ B0
while Q 6= ∅ {
x0 = min(Q); X = [x0 ] //X is equivalence class of x0
determine path type of X
compute minimum length folding pathway from s X to t X
numM oves = numM oves + max(|s X|, |t X|)
add to s all base pairs in BP2
numM oves = numM oves + |BP2 |
6
return numM oves
11.
Straightforward details of how to implement line 7 are given in the Appendix. The principle underlying
the reason that Algorithm 9 produces a minimum length (pseudoknotted) M S2 folding trajectory from s
to t is that we maximize the number of shift moves, since a single shift move from {x, y} ∈ s to {y, z} ∈ t
corresponds to the simultaneous removal of {x, y} and addition of {y, z}. We apply this principle in the next
section to determine the minimum length (non-pseudoknotted) M S2 folding trajectory from s to t.
3
M S2 distance between secondary structures
In this section, we present an integer programming (IP) algorithm to compute the M S2 distance between
any two secondary structures s, t, i.e. the minimum length of an M S2 trajectory from s to t. Our algorithm
has been cross-checked with the exhaustive branch-and-bound algorithm mentioned at the end of the last
section.
As in the previous section, our goal is to maximize the number of shift operations in the M S2 trajectory,
formalized in the following simple theorem, whose proof is clear.
Theorem 10. Suppose that the M S2 distance between secondary structures s, t is k, i.e. base pair distance
dBP (s, t) = |s − t| + |t − s| = k. Suppose that ` is the number of shift moves occurring in a minimum length
M S2 refolding trajectory s = s0 , s1 , . . . , sm = t from s to t. Then the M S2 distance between s and t equals
dM S2 (s, t) = ` + (k − 2`) = k − `
(15)
Our strategy will now be to use a graph-theoretic approach to maximize the number of shift moves.
3.1
RNA conflict digraph
Throughout this section, we take s, t to be two arbitrary, distinct, but fixed secondary structures of the RNA
sequence a1 , . . . , an . Recall the definitions of A, B, C, D in equations (5–8), so that A is the set of positions
x ∈ [1, n] that are base-paired in both s and t, but the base pairs in s and t are not identical; B is the set
of positions x ∈ [1, n] that are base-paired in one of s or t, but not both; C is the set of positions x ∈ [1, n]
that are base-paired in neither s nor t, and D is the set of positions x ∈ [1, n] that are base-paired to the
same partner in both s and t.
To determine a minimum length M S2 folding trajectory from secondary structure s to secondary structure
t is to maximize the number of shift moves and minimize the number of base pair additions and removals.
To that end, note that the base pairs in s that do not touch any base pair of t must be removed in any M S2
path from s to t, since there is no shift of such base pairs to a base pair of t – such base pairs are exactly
those in BP1 , defined in equation (12). Similarly, note that the base pairs in t that do not touch any base
pair of s must occur must be added, in the transformation of s to t, since there is no shift of any base pair
from s to obtain such base pairs of t – such base pairs are exactly those in BP2 , defined in equation (13). We
now focus on the remaining base pairs of s, all of which touch a base pair of t, and hence could theoretically
allow a shift move in transforming s to t, provided that there is no base triple or pseudoknot introduced by
performing such a shift move. Examples of all six possible types of shift move are illustrated in Figure 4. To
handle such cases, we define the notion of RNA conflict digraph, solve the feedback vertex set (FVS) problem
[22] by integer programming (IP), apply topological sorting [7] to the acyclic digraph obtained by removing
a minimum set of vertices occurring in feedback loops, then apply shift moves in topologically sorted order.
We now formalize this argument.
Define the digraph G = (V, E), whose vertices (or nodes) n ∈ V are defined in the following Definition 11
and whose directed edges are defined in Definition 12.
Definition 11 (Vertex in an RNA conflict digraph).
If s, t are distinct secondary structures for the RNA sequence a1 , . . . , an , then a vertex in the RNA conflict
digraph G = G(s, t) is a triplet node, or more simply, node v = (x, y, z) consisting of integers x, y, z, such that
the base pair {x, y}< = (min(x, y), max(x, y)) belongs to t, and the base pair {y, z}< = (min(y, z), max(y, z))
belongs to s. Let v.t [resp. v.s] denote the base pair {x, y}< [resp. {y, z}< ] belonging to t [resp. s]. The
7
middle integer y of node v = (x, y, z) is called the pivot position, since it is common to both s and t. Nodes
are ordered by the integer ordering of their pivot positions: (x, y, z) (x0 , y 0 , z 0 ) if and only if y ≤ y 0 (or
y = y 0 and x < x0 , or y = y 0 , x = x0 , and z < z 0 ). If v = (x, y, z) is a node, then f latten(v) is defined to be
the set {x, y, z} of its coordinates.
Nodes are representations of a potential shift move, and can be categorized into six types, as shown in
Figure 4.
Definition 12 (Directed edge in an RNA conflict digraph).
Base pairs {a, b}< and {c, d}< are said to touch if |{a, b} ∩ {c, d}| = 1; in other words, base pairs touch if
they form a base triple. Base pairs {a, b}< and {c, d}< are said to cross if either min(a, b) < min(c, d) <
max(a, b) < max(c, d) or min(c, d) < min(a, b) < max(c, d) < max(a, b); in other words, base pairs cross
if they form a pseudoknot. There is a directed edge from node n1 = (x1 , y1 , z1 ) to node n2 = (x2 , y2 , z2 ),
denoted by (n1 , n2 ) ∈ E or equivalently by n1 → n2 , if (1) |f latten(n1 ) ∩ f latten(n2 )| ≤ 1, or in other words
if n1 and n2 overlap in at most one position, and (2) the base pair {y1 , z1 }< ∈ s from n1 either touches or
crosses the base pair {x2 , y2 }< ∈ t from n2 .
Note that if the base pair {y1 , z1 }< ∈ s from n1 touches the base pair {x2 , y2 }< ∈ t from n2 , then it
must be that z1 = x2 ; indeed, since each pivot node y1 [resp. y2 ] belongs to a base pair of both s and t,
it cannot be that z1 = y2 (because then {y1 , z1 }< ∈ s and {y2 , z2 }< ∈ s would form a base triple in s at
z1 = y2 ), nor can it be that y1 = x2 (because then {x1 , y1 }< ∈ t and {x2 , y2 }< ∈ t would form a base
triple in t at y1 = x2 ). Note as well that if n1 = (x1 , y1 , z1 ) and n2 = (x2 , y2 , z2 ) are triplet nodes, then
|f latten(n1 ) ∩ f latten(n2 )| = 1 implies that either n1 → n2 or n2 → n1 . Indeed, if there is a common
element shared by n1 and n2 , then it cannot be a pivot element, since s and t cannot have a base triple. For
the same reason, the common element cannot belong to the base pairs {x1 , y1 } ∈ t of n1 and {x2 , y2 } ∈ t
of n2 (otherwise t would contain a base triple), nor can the common element belong to the base pairs
{y1 , z1 } ∈ s of n1 and {y2 , z2 } ∈ s of n2 (otherwise s would contain a base triple). It follows that either
{x1 , y1 } ∩ {y2 , z2 } =
6 ∅, or {x2 , y2 } ∩ {y1 , z1 } =
6 ∅. From the assumption that |f latten(n1 ) ∩ f latten(n2 )| = 1,
this implies that either n2 → n1 or that n1 → n2 , but not both. Finally, note that if n1 = (x1 , y1 , z1 ),
n2 = (x2 , y2 , z2 ) and |f latten(n1 ) ∩ f latten(n2 )| = 2, then there are exactly three possibilities, all of which
can be realized:
1. n1 .t = n2 .t, so that {x1 , y1 } = {x2 , y2 }, as in the example (1, 5) ∈ s, (10, 15) ∈ s, (5, 10) ∈ t,
n1 = (10, 5, 1), n2 = (5, 10, 15);
2. n1 .s = n2 .s, so that {y1 , z1 } = {y2 , z2 }, as in the example (1, 5) ∈ t, (10, 15) ∈ t, (5, 10) ∈ s,
n1 = (1, 5, 10), n2 = (15, 10, 5);
3. {x1 , z1 } = {x2 , z2 }, as shown in Figure 5. This latter example will be called a closed 2-cycle.
These considerations produce the equivalent but sharper following definition.
Definition 13 (Conflict digraph G = (V, E)). Let s, t be distinct secondary structures for the RNA sequence
a1 , . . . , an . The RNA confict digraph G(s, t) = (V (s, t), E(s, t)), or G = (V, E) when s, t are clear from
context, is defined by
V = {(x, y, z) : x, y, z ∈ [1, n] ∧ {x, y} ∈ t ∧ {y, z} ∈ s}
(16)
n
E = (n1 , n2 ) : n1 = (x1 , y1 , z1 ) ∈ V ∧ n2 = (x2 , y2 , z2 ) ∈ V ∧
|f latten(n1 ) ∩ f latten(n2 )| ≤ 1 ∧ z1 = x2 ∨
[min(y1 , z1 ) < min(x2 , y2 ) < max(y1 , z1 ) < max(x2 , y2 )] ∨
o
[min(x2 , y2 ) < min(y1 , z1 ) < max(x2 , y2 ) < max(y1 , z1 )]
(17)
8
The set of directed edges of conflict digraph G = (V, E), as defined in Definition 13, establishes a partial
ordering on vertices of V with the property that n1 → n2 holds for vertices n1 = (x, y, z), n2 = (u, v, w) if
and only if (1) n1 and n2 overlap in at most one position, and (2) when shift move n2 is applied, shifting
{v, w} ∈ s to {u, v} ∈ t, the base pair {u, v} either touches or crosses the base pair {y, z} ∈ s in n1 . It
follows that if n1 → n2 , then the shift move in which {y, z} ∈ s shifts to {x, y} ∈ t must be performed before
the shift move where {v, w} ∈ s shifts to {u, v} ∈ t – indeed, if shifts are performed in the opposite order,
then after shifting {v, w} ∈ s to {u, v} ∈ t and before shifting {y, z} ∈ s to {x, y} ∈ t, we would create either
a base triple or a pseudoknot.
Strictly speaking, the overlap condition (1) is not a necessary requirement, and in our first IP algorithm
to compute M S2 distance [2], we considered a somewhat more general edge relation without condition (1).
If distinct vertices n1 , n2 violate condition (1), then |f latten(n1 ) ∩ f latten(n2 )| = 2, and the constraint (‡)
in line 7 of Algorithm 14 would ensure that at most one of n1 , n2 are selected by the IP solver for the set V ,
in the resulting acyclic digraph G = (V , E). Nevertheless, the run time of our algorithm depends heavily on
the number of simple directed cycles in the initial conflict digraph G = (V, E). Without condition (1), nodes
n1 , n2 in a closed 2-cycle (see Figure 5) satisfy n1 → n2 and n2 → n1 . Since it is possible that n1 belong to
other cycles that do not contain n2 , this can (and does) lead to a greatly increased number of directed cycles,
hence much longer run time – indeed, our algorithm in [2] runs 10 times slower than the current algorithm.
Before proceeding with the description of the algorithm, we must explain how to treat closed 2-cycles, as
shown in Figure 5, for which there exist four integers a1 < a2 < a3 < a4 , such that either Case A or Case B
holds.
Case A: Base pairs (a1 , a2 ) and (a3 , a4 ) belong to t, while base pairs (a1 , a4 ) and (a2 , a3 ) belong to s, as
shown in Figure 5a.
In this case, the conflict digraph G = (V, E) contains the following 4 vertices v1 = (a1 , a2 , a3 ) of type 1,
v2 = (a3 , a4 , a1 ) of type 5, v3 = (a2 , a1 , a4 ) of type 4, and v4 = (a4 , a3 , a2 ) of type 2. The overlap of any two
distinct vertices has size 2, so by Definition 13, there can be no directed edge between any vertices. There
are four optimal trajectories of size 3; for specificity we select the following optimal trajectory:
remove (a1 , a4 ) from s
(18)
Case B: Base pairs (a1 , a2 ) and (a3 , a4 ) belong to s, while base pairs (a1 , a4 ) and (a2 , a3 ) belong to t, as
shown in Figure 5b.
In this case, the conflict digraph G = (V, E) contains the following 4 vertices v1 = (a1 , a4 , a3 ) of type 6,
v2 = (a4 , a1 , a2 ) of type 3, v3 = (a2 , a3 , a4 ) of type 1, and v4 = (a3 , a2 , a1 ) of type 2. The overlap of any two
distinct vertices has size 2, so by Definition 13, there can be no directed edge between any vertices. There
are four optimal trajectories of size 3; for specificity we select the following optimal trajectory:
remove (a1 , a2 ) from s
(19)
In Algorithm 14 below, it is necessary to list all closed 2-cycles, as depicted in Figure 5. This can
be done in linear time O(n), for RNA sequence a = a1 , . . . , an and secondary structures s, t by computing
equivalence classes as defined in Definition 7, then inspecting all size 4 equivalence classes X = {a1 , a2 , a3 , a4 }
to determine whether Case A or Case B applies. For each such closed 2-cycle, Algorithm 14 computes the
partial trajectory (18) or (19) appropriately, then the vertices v1 , v2 , v3 , v4 are deleted. No edges need to be
deleted, since there are no edges between vi and vj for 1 ≤ i, j ≤ 4. In creating the partial trajectories, the
variable numMoves must be updated.
By special treatment of closed 2-cycles, we obtain a 10-fold speed-up in the exact IP Algorithm 14
over that of the precursor Algorithm 10 in our WABI 2017 proceedings paper [2] – compare run times of
Figure 14 with those from Figure 5 of [2]. Except for the special case of closed 2-cycles that must be handled
before general treatment, note that Definition 13 establishes a partial ordering on vertices of the conflict
digraph G = (V, E), in that edges determine the order in which shift moves should be performed. Indeed,
if n1 = {x, y, z}, n2 = {u, v, z} and (n1 , n2 ) ∈ E, which we denote from now on by n1 → n2 , then the shift
move in which {y, z} ∈ s shifts to {x, y} ∈ t must be performed before the shift move where {v, w} ∈ s
shifts to {u, v} ∈ t – indeed, if shifts are performed in the opposite order, then after shifting {v, w} ∈ s to
{u, v} ∈ t and before shifting {y, z} ∈ s to {x, y} ∈ t, we would create either a base triple or a pseudoknot.
9
Our strategy to efficiently compute the M S2 distance between secondary structures s and t will be to (1)
enumerate all simple cycles in the conflict digraph G = (V, E) and to (2) apply an integer programming
(IP) solver to solve the minimum feedback arc set problem V 0 ⊂ V . Noticing that the induced digraph
G = (V , E), where V = V − V 0 and E = E ∩ (V × V ), is acyclic, we then (3) topologically sort G, and (4)
perform shift moves from V in topologically sorted order.
Algorithm 14 (M S2 distance from s to t).
Input: Secondary structures s, t for RNA sequence a1 , . . . , an
Output: Folding trajectory s = s0 , s1 , . . . , sm = t, where s0 , . . . , sm are secondary structures, m is the
minimum possible value for which si is obtained from si−1 by a single base pair addition, removal or shift
for each i = 1, . . . , m.
First, initialize the variable numMoves to 0, and the list moveSequence to the empty list [ ]. Recall that
BP2 = {(x, y) : (x, y) ∈ t, (s − t)[x] = 0, (s − t)[y] = 0}. Bear in mind that s is constantly being updated, so
actions performed on s depend on its current value.
//remove base pairs from s that are untouched by t
1. BP1 = {(x, y) : (x, y) ∈ s, (t − s)[x] = 0, (t − s)[y] = 0}
2. for (x, y) ∈ BP1
3.
remove (x, y) from s; numMoves = numMoves+1
//define conflict digraph G = (V, E) on updated s and unchanged t
4. define V by equation (16)
5. define E by equation (17)
6. define conflict digraph G = (V, E)
//IP solution
Pof minimum feedback arc set problem
7. maximize
v∈V xv where xv ∈ {0, 1}, subject to constraints (†) and (‡)
//constraint
to
remove vertex from each simple cycle of G
P
xv < |C| for each simple directed cycle C of G
(†)
v∈C
//constraint to ensure shift moves cannot be applied if they share same base pair from s or t
(‡) xv + xv0 ≤ 1, for all pairs of vertices v = (x, y, z) and v 0 = (x0 , y 0 , z 0 ) with |{x, y, z} ∩ {x0 , y 0 , z 0 }| = 2
//define IP solution acyclic digraph G = (V , E)
8. V = {v ∈ V : xv = 1}; V 0 = {v ∈ V : xv = 0}
9. E = {(v, v 0 ) : v, v 0 ∈ V ∧ (v, v 0 ) ∈ E}
10. G = (V , E)
//handle special, closed 2-cycles
11. for each closed 2-cycle [x] = {a1 , a2 , a3 , a4 } as depicted in Figure 5
12.
if [x] is of type A as depicted in Figure 5a
13.
remove base pair from s by equation (18)
14.
if [x] is of type B as depicted in Figure 5b
15.
remove base pair from s by equation (19)
//remove base pairs from s that are not involved in a shift move
16. V .s = {(x, y) : ∃v ∈ V (v.s = (x, y))}
17. for (x, y) ∈ s − t
18.
if (x, y) 6∈ V .s
19.
remove (x, y) from s; numMoves = numMoves+1
//topological sort for IP solution G = (V , E)
20. topological sort of G using DFS [7] to obtain total ordering ≺ on V
21. for v = (x, y, z) ∈ V in topologically sorted order ≺
22.
shift {y, z} to {x, y} in s; numMoves = numMoves+1
//add remaining base pairs from t − s, e.g. from BP2 and type 4,5 paths in Figure 3
23. for (x, y) ∈ t − s
24.
add (x, y) to s; numMoves = numMoves+1
25. return folding trajectory, numMoves
We now illustrate the definitions and the execution of the algorithm for a tiny example where s =
{(1, 5), (10, 15), (20, 25)} and t = {(5, 10), (15, 20)}. From Definition 7, there is only one equivalence class
X = {1, 5, 10, 15, 20, 25} and it is a path of type 1, as illustrated in Figure 3, where b1 = 1, a1 = 5, b2 = 10,
10
a2 = 15, b3 = 20, a3 = 25. From Definition 11, there are 4 vertices in the conflict digraph G = (V, E),
where v1 = (10, 5, 1), v2 = (5, 10, 15), v3 = (20, 15, 10), v4 = (15, 20, 25) – recall the convention from that
definition that vertex v = (x, y, z) means that base pair {y, z} ∈ s and base pair {x, y} ∈ t, so that the
pivot position y is shared by base pairs from both s and t. From Definition 12, there are only two directed
edges, v1 → v3 since v1 .s touches v2 .t, and v2 → v4 since v2 .s touches v4 .t. Note there is no edge from v1 to
v2 , or from v2 to v3 , or from v3 to v4 , since their overlap has size 2 – for instance f latten(v1 ) = {1, 5, 10},
f latten(v2 ) = {5, 10, 15}, and f latten(v1 )∩f latten(v2 ) = {5, 10} of size 2. There is no cycle, so the constraint
(†) in line 7 of Algorithm 14 is not applied; however the constraint (‡) does apply, so that xv1 + xv2 ≤ 1,
xv2 + xv3 ≤ 1, xv3 + xv4 ≤ 1. It follows that there are three possible IP solutions for the vertex set V .
Case 1: V = {v1 , v3 }
Then v1 .s = (1, 5), v3 .s = (10, 15) so V .s = {(1, 5), (10, 15)} and by lines 11-14 we remove base pair (20, 25)
from s. Now G = (V , E), where E = {v1 → v3 }, so topological sort is trivial and we complete the trajectory
by applying shift v1 and then shift v3 . Trajectory length is 5.
Case 2: V = {v1 , v4 }
Then v1 .s = (1, 5), v4 .s = (20, 25) so V .s = {(1, 5), (20, 25)} and by lines 11-14 we remove base pair (10, 15)
from s. Now G = (V , E), where E = ∅, so topological sort is trivial and we complete the trajectory by
applying shift v1 and then shift v4 , or by applying shift v4 and then shift v1 . Trajectory length is 5.
Case 3: V = {v2 , v4 }
Then v2 .s = (10, 15), v4 .s = (20, 25) so V .s = {(10, 15), (20, 25)} and by lines 11-14 we remove base pair
(1, 5) from s. Now G = (V , E), where E = {v2 → v4 }, so topological sort is trivial and we complete the
trajectory by applying shift v2 and then shift v4 . Trajectory length is 5.
3.2
Examples to illustrate Algorithm 14
We illustrate concepts defined so far with three examples: a toy 20 nt RNA sequence, a 25 nt bistable switch,
and the 56 nt spliced leader RNA from L. collosoma.
3.2.1
Toy 20 nt sequence
For the toy 20 nt sequence GGGAAAUUUC CCCAAAGGGG with initial structure s shown in Figure 6a,
and target structure t shown in Figure 6b, the corresponding conflict digraph is shown in Figure 6c. This is
a toy example, since the empty structure is energetically more favorable than either structure: free energy of
s is +0.70 kcal/mol, while that for t is +3.30 kcal/mol. The conflict digraph contains 6 vertices, 10 directed
edges, and 3 simple cycles: a first cycle {(8, 20, 10), (9, 19, 11), (18, 10, 20), (19, 9, 1)} of size 4, a second cycle
{[(8, 20, 10), (19, 9, 1)} of size 2, and a third cycle {(18, 10, 20), (9, 19, 11)} of size 2.
3.2.2
Bistable switch
Figure 7 depicts the secondary structure for the metastable and the MFE structures, as well as the corresponding conflict digraphs for the 25 nt bistable switch, with sequence UGUACCGGAA GGUGCGAAUC
UUCCG, taken from Figure 1(b).1 of [20], in which the authors report structural probing by comparative
imino proton NMR spectroscopy. The minimum free energy (MFE) structure has -10.20 kcal/mol, while
the next metastable structure has -7.40 kcal/mol. Two lower energy structures exist, having -9.00 kcal/mol
resp. -7.60 kcal/mol; however, each is a minor variant of the MFE structure. Figures 7a and 7b depict
respectively the metastable and the MFE secondary structures for this 25 nt RNA, while Figures 7c and 7d
depict respectively the MFE conflict digraph and the metastable conflict digraph.
For this 25 nt bistable switch, let s denote the metastable structure and t denote the MFE structure. We
11
determine the following. , then we have the following.
s = [(1, 16), (2, 15), (3, 14), (4, 13), (5, 12), (6, 11)] with 6 base pairs
t = [(6, 25), (7, 24), (8, 23), (9, 22), (10, 21), (11, 20), (12, 19), (13, 18)] with 8 base pairs
A = {6, 11, 12, 13}
B = {1, 2, 3, 4, 5, 7, 8, 9, 10, 14, 15, 16, 18, 19, 20, 21, 22, 23, 24, 25}
C = {17}
D=∅
BP1 = {(1, 16), (2, 15), (3, 14)} with 3 base pairs
BP2 = {(7, 24), (8, 23), (9, 22), (10, 21)} with 4 base pairs
B0 = {4, 5, 18, 19, 20, 25}
B1 = {1, 2, 3, 14, 15, 16}
B2 = {7, 8, 9, 10, 21, 22, 23, 24}
and there are three equivalence classes: X1 = {4, 13, 18} of type 2, X2 = {5, 12, 19} of type 2, and X3 =
{6, 11, 20, 25} of type 4.
Figure 7c depicts the MFE conflict digraph, where s denotes the metastable structure and t denotes
the MFE structure. In the MFE conflict digraph G = (V, E), vertices are triplet nodes (x, y, z), where
(unordered) base pair {y, z} ∈ s belongs to the metastable [resp. MFE] structure, and (unordered) base
pair {x, y} ∈ t belongs to the MFE [resp. metastable] structure. A direct edge (x, y, z) → (u, v, w) occurs if {y, z} ∈ s touches or crosses {u, v} ∈ t. Both the MFE and the metastable conflict digraphs
are acyclic. Although there are no cycles, the IP solver is nevertheless invoked in line 7 with constraint
(‡), resulting in either a first solution V = {(18, 13, 4), (19, 12, 5), (20, 11, 6)} or a second solution V =
{(18, 13, 4), (19, 12, 5), (25, 6, 11)}. Indeed, the overlap of vertices (20, 11, 6) and (25, 6, 11) has size 2, so one
of these vertices must be excluded from V in 8 of Algorithm 14. Assume that the first solution is returned
by the IP solver. Then we obtain the following minimum length M S2 folding trajectory from metastable s
to MFE t.
Vertex and edge set of G = (V, E) are given by the following.
V = {(18, 13, 4), (19, 12, 5), (20, 11, 6), (25, 6, 11)}
E = {(18, 13, 4) → (19, 12, 5), (18, 13, 4) → (20, 11, 6), (18, 13, 4) → (25, 6, 11),
(19, 12, 5) → (20, 11, 6), (19, 12, 5) → (25, 6, 11)}
One of two minimum length M S2 folding trajectories is given by the following.
1.
2.
UGUACCGGAAGGUGCGAAUCUUCCG
1234567890123456789012345
0.
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
((((((....)))))).........
.(((((....)))))..........
..((((....))))...........
...(((....)))............
....((....))(....).......
.....(....)((....))......
..........(((....))).....
......(...(((....)))...).
......((..(((....)))..)).
......(((.(((....))).))).
......(((((((....))))))).
.....((((((((....))))))))
metastable s
remove (1,16)
remove (2,15)
remove (3,14)
shift (4,13) to (13,18)
shift (5,12) to (12,19)
shift (6,11) to (11,20)
add (7,24)
add (8,23)
add (9,22)
add (10,21)
add (6,25)
Algorithm 14 executes the following steps: (1) Remove base pairs in BP1 from s. (2) Compute conflict
digraph G = (V, E). (3) Apply IP solver to determine maximum size V ⊆ V , subject to removing a vertex
12
from each cycle (†) and not allowing any two vertices in V to have overlap of size 2. (4) Topologically sorting
the induced digraph G = (V , E). (5) Execute shifts according to total ordering ≺ given by topological sort.
(6) Add remaining base pairs from t − s. Note that in trajectory steps 7-10, the base pair added comes from
BP2 , while that in step 11 is a base pair from t that is “leftover”, due to the fact that triplet node (shift
move) (25, 6, 11) does not belong to IP solution V .
3.2.3
Spliced leader from L. collosoma
For the 56 nt L. collosoma spliced leader RNA, whose switching properties were investigated in [18] by
stopped-flow rapid-mixing and temperature-jump measurements, the MFE and metastable structures are
shown in Figure 8, along with the conflict digraph for M S2 folding from the metastable structure to the MFE
structure. This RNA has sequence AACUAAAACA AUUUUUGAAG AACAGUUUCU GUACUUCAUU
GGUAUGUAGA GACUUC, an MFE structure having -9.40 kcal/mol, and an alternate metastable structure
having -9.20 kcal/mol. Figure 8 displays the MFE and metastable structures for L. collosoma spliced leader
RNA, along with the conflict digraph for M S2 folding from the metastable to the MFE structure.
For L. collosoma spliced leader RNA, if we let s denote the metastable structure and t denote the
MFE structure, then there are seven equivalence classes: X1 = {10, 45, 31, 23} of type 4; X2 = {11, 43, 33}
of type 3; X3 = {12, 42, 34, 20} of type 4, X4 = {13, 41, 35, 19} of type 4, X5 = {22, 32, 44} of type 3,
X6 = {24, 54, 30, 48} of type 1, and X7 = {25, 53, 29, 49} of type 1. As in the case with the 25 nt bistable
switch, the equivalence classes for the situation where s and t are interchanged are identical, although type 1
paths become type 4 paths (and vice versa), and type 2 paths become type 3 paths (and vice versa). Output
from our (optimal) IP algorithm is as follows.
AACUAAAACAAUUUUUGAAGAACAGUUUCUGUACUUCAUUGGUAUGUAGAGACUUC
12345678901234567890123456789012345678901234567890123456
Number of Nodes: 12
Number of edges: 71
Number of cycles: 5
s: .......................((((((((((((.....)))))..)))))))..
t: .......((((((..(((((.((((...)))).)))))..))).))).........
-9.20 kcal/mol
-9.40 kcal/mol
0. .......................((((((((((((.....)))))..))))))).. metastable s
1. .......................((.(((((((((.....)))))..)))).)).. remove (26,52)
2. .......................((..((((((((.....)))))..)))..)).. remove (27,51)
3. .......................((...(((((((.....)))))..))...)).. remove (28,50)
4. .......................((....((((((.....)))))..)....)).. remove (29,49)
5. .......................((.....(((((.....))))).......)).. remove (30,48)
6. .......................((...).(((((.....)))))........).. (25,53) -> (25,29)
7. .......................((...))(((((.....)))))........... (24,54) -> (24,30)
8. .........(.............((...)).((((.....)))))........... (31,45) -> (10,45)
9. .........(...........(.((...)).)(((.....))).)........... (32,44) -> (22,32)
10. .........((..........(.((...)).).((.....))).)........... (33,43) -> (11,43)
11. .........(((.........(.((...)).)..(.....))).)........... (34,42) -> (12,42)
12. .........(((......(..(.((...)).)..)......)).)........... (35,41) -> (19,35)
13. .......(.(((......(..(.((...)).)..)......)).).)......... add (8,47)
14. .......(((((......(..(.((...)).)..)......)).)))......... add (9,46)
15. .......(((((...(..(..(.((...)).)..)..)...)).)))......... add (16,38)
16. .......(((((...((.(..(.((...)).)..).))...)).)))......... add (17,37)
17. .......(((((...((((..(.((...)).)..))))...)).)))......... add (18,36)
18. .......(((((...(((((.(.((...)).).)))))...)).)))......... add (20,34)
19 .......(((((...(((((.((((...)))).)))))...)).)))......... add (23,31)
20. .......((((((..(((((.((((...)))).)))))..))).)))......... add (13,41)
Number of base pair removals: 5
Number of base pair additions: 8
Number of base pair shifts: 7
13
MS2 Distance:
20
Figure 8a depicts the initial structure s, and Figure 8b depicts the target minimum free energy structure
t for spliced leader RNA from L. collosoma. The conflict digraph for the refolding from s to t is shown
in Figure 8c. Figure 9a displays the rainbow diagram for spliced leader RNA from L. collosoma, in which
the base pairs for the initial structure s (Figure 8a) are shown below the line in red, while those for the
target structure t (Figure 8b) are shown above the line in blue. Figure 9c displays the Arrhenius tree,
where leaf index 2 represents the initial metastable structure s with free energy -9.20 kcal/mol as shown
in Figure 8a, while leaf index 1 represents the target MFE structure t with free energy -9.40 kcal/mol as
shown in Figure 8b. In Figure 9b, the dotted blue line depicts the free energies of structures in the shortest
M S2 folding trajectory for spliced leader, as computed by Algorithm 14, while the solid red line depicts
the free energies of the energy-optimal folding trajectory as computed by the programs RNAsubopt [46] and
barriers [14].
3.2.4
xpt riboswitch from B. subtilis
In this section, we describe the shortest M S2 folding trajectory from the initial gene ON structure s to
the target gene OFF structure t for the 156 nt xanthine phosphoribosyltransferase (xpt) riboswitch from B.
subtilis, where the sequence and secondary structures are taken from Figure 1A of [38]. The gene ON [resp.
OFF] structures for the 156 nt xpt RNA sequence AGGAACACUC AUAUAAUCGC GUGGAUAUGG
CACGCAAGUU UCUACCGGGC ACCGUAAAUG UCCGACUAUG GGUGAGCAAU GGAACCGCAC
GUGUACGGUU UUUUGUGAUA UCAGCAUUGC UUGCUCUUUA UUUGAGCGGG CAAUGCUUUU
UUUAUU are displayed in Figure 10a [resp. 10b], while Figure 10c shows the rainbow diagram, where lower
red arcs [resp. upper blue arcs] indicate the base pairs of the initial gene ON [resp. target gene OFF]
structure. The default structure for the xpt riboswitch in B. subtilis is the gene ON structure; however, the
binding of a guanine nucleoside ligand to cytidine in position 66 triggers a conformational change to the
gene OFF structure. Figure 10d depicts the conflict digraph G = (V, E) containing 18 vertices, 113 directed
edges, and 1806 directed cycles, which is used to compute the shortest M S2 folding trajectory from the gene
ON to the gene OFF structure. Figures 11a and 11b show an enlargement of the initial gene ON structure
s and target gene OFF structure t, which allows us to follow the moves in a shortest M S2 trajectory that is
displayed in Figure 11c.
4
An algorithm for near-optimal M S2 distance
Since the exact IP Algorithm 14 could not compute the shortest M S2 folding trajectories between the minimum free energy (MFE) structure and Zuker suboptimal structures for some Rfam sequences of even modest
size (≈ 100 nt), we designed a near-optimal IP algorithm (presented in this section), and a greedy algorithm
(presented in Section D of the Appendix). The exact branch-and-bound algorithm from Appendix C was
used to debug and cross-check all algorithms.
The run time complexity of both the exact IP Algorithm 14 and the greedy algorithm is due to the
possibly exponentially large set of directed simple cycles in the RNA conflict digraph. By designing a 2-step
process, in which the feedback arc set (FAS) problem is first solved for a coarse-grained digraph defined below,
and subsequently the feedback vertex set (FVS) problem is solved for each equivalence class, we obtain a
much faster algorithm to compute a near-optimal M S2 folding trajectory between secondary structures s and
t for the RNA sequence {a1 , . . . , an }. In the first step, we use IP to solve the feedback arc set (FAS) problem
for a particular coarse-grained digraph defined below, whose vertices are the equivalence classes as defined in
Definition 7. The number of cycles for this coarse-grained digraph is quite manageable, even for large RNAs,
hence the FAS can be efficiently solved. After removal of an arc from each directed cycle, topological sorting
is applied to determine a total ordering according to which each individual equivalence class is processed. In
the second step, Algorithm 14 is applied to each equivalence class in topologically sorted order, whereby the
feedback vertex set (FVS) problem is solved for the equivalence class under consideration. In the remainder
of this section, we fill in the details for this overview, and then present pseudocode for the near-optimal
Algorithm 15.
14
Given secondary structures s and t for the RNA sequence {a1 , . . . , an }, we partition the set [1, n] into
disjoint sets A, B, C, D as in Section 2 by following equations (5), (6), (7), (8). The union A∪B is subsequently
partitioned into the equivalence classes X1 , . . . , Xm , defined in Definition 7. Define the coarse-grain, conflict
digraph G = (V, E), whose vertices are the indices of equivalence classes X1 , . . . , Xm , and whose directed
edges i → j are defined if there exists a base pair (x, y) ∈ s, x, y ∈ Xi which crosses a base pair (u, v) ∈ t,
u, v ∈ Xj . Although there may be many such base pairs (x, y) ∈ s and (u, v) ∈ t, there is only one edge
between i and j; i.e. G is a directed graph, not a directed multi-graph. If i → j is an edge, then we define
Ni,j to be the set of all base pairs (u, v) ∈ s, u, v ∈ Xi that cross some base pair (u, v) ∈ t, u, v ∈ Xj , and
let ni,j the number of base pairs in Ni,j . Formally, given equivalence classes X1 , . . . , Xm , the coarse-grain,
conflict digraph G = (V, E) is defined by
V = {1, . . . , m}
n
o
E = i → j : ∃(x, y) ∈ s∃(u, v) ∈ t x, y ∈ Xi ∧ u, v ∈ Xj
(20)
(21)
A directed edge from i to j may be denoted either by i → j ∈ E or by (i, j) ∈ E. For each edge i → j, we
formally define Ni,j and ni,j by the following.
n
o
Ni,j = (x, y) ∈ s : x, y ∈ Xi ∧ ∃(u, v) ∈ t u, v ∈ Xj ∧ (x, y) crosses (u, v)
(22)
ni,j = Ni,j
(23)
We now solve the feedback arc set (FAS) problem, rather than the feedback vertex set (FVS) problem,
for digraph G, by applying an IP solver to solve the following optimization problem:
1.
P
maximize
(i,j)∈E ni,j · xi,j subject to constraint (]):
P
(])
xi,j < |C|
(i,j)∈E
i,j∈C
for every directed cycle C = (i1 , i2 , i3 , . . . , ik−1 , ik )
This IP problem can be quickly solved, since there is usually only a modest number of directed cycles for
the coarse-grained digraph. For each directed edge or arc i → j that is to be removed from a directed
cycle, we remove all base pair (x, y) ∈ from structure s that cross some base pair (u, v) ∈ t for which
u, v ∈ Xj . Removal of certain base pairs from s can disconnect some previous equivalence classes into two
0
or more connected components, hence equivalence classes X10 , . . . , Xm
0 must be recomputed for the updated
0
structure s and (unchanged) structure t. The conflict, digraph G = (V 0 , E 0 ) is then defined by equations
(20) and (21) for (updated) s and (unchanged) t. Since G0 is now acyclic, it can be topologically sorted,
0
which determines an ordering σ(1), . . . , σ(m0 ) for processing equivalence classes X10 , . . . , Xm
0 . To process
0
an equivalence class X , we restrict the exact Algorithm 14 to each equivalence class. Indeed, to process
equivalence class X 0 , we define a (local) conflict digraph G(X 0 ) = (V (X 0 ), E(X 0 )) defined as follows.
V (X 0 ) = {(x, y, z) : x, y, z ∈ X 0 ∧ {x, y} ∈ t ∧ {y, z} ∈ s
0
0
(24)
0
E(X ) = {(x, y, z) → (u, v, w) : (x, y, z) ∈ V (X ) ∧ (u, v, w) ∈ V (X )∧
(25)
{x, y} touches or crosses {v, w} }
Algorithm 15 (Near-optimal M S2 distance from s to t).
Input: Secondary structures s, t for RNA sequence a1 , . . . , an
Output: s = s0 , s1 , . . . , sm = t, where s0 , . . . , sm are secondary structures, m is a near-optimal value for
which si is obtained from si−1 by a single base pair addition, removal or shift for each i = 1, . . . , m.
First, initialize the variable numMoves to 0, and the list moveSequence to the empty list [ ]. Define
BP1 = {(x, y) : (x, y) ∈ t, (t − s)[x] = 0, (t − s)[y] = 0}; i.e. BP1 consists of those base pairs in s which are
15
not touched by any base pair in t. Define BP2 = {(x, y) : (x, y) ∈ t, (s − t)[x] = 0, (s − t)[y] = 0}; i.e. BP2
consists of those base pairs in t which are not touched by any base pair in s.
//remove base pairs from s that are untouched by t
1. for (x, y) ∈ BP1
2.
s = s − {(x, y)}
3.
numMoves = numMoves + 1
//define equivalence classes on updated s, t
4. [1, n] = A ∪ B ∪ C ∪ D
5. determine equivalence classes X1 , . . . , Xm with union A ∪ B
//define conflict digraph G = (V, E) on collection of equivalence classes
6. define V = {1, . . . , m}
7. define E = {(i, j) : 1 ≤ i, j ≤ m} by equation (21)
8. define coarse-grain, conflict digraph G = (V, E)
//IP solution
Pof feedback arc set problem (not feedback vertex set problem)
9. maximize
(i,j)∈E ni,j · xi,j subject to constraint (]):
//removeParc from each simple cycle whereni,j defined in equation (23)
xi,j < |C|
(])
(i,j)∈E
i→j∈C
for every directed cycle C = (i1 , i2 , . . . , ik−1 , ik )
e = {(i, j) : xi,j = 0} //E
e is set of edges that must be removed
10. E
e
//process the IP solution E
e
11. for (i, j) ∈ E
12.
for (x, y) ∈ Ni,j //Ni,j defined in Definition 22
13.
s = s − {(x, y)} //remove base pair from s belonging to feedback arc
14.
numMoves = numMoves + 1
//determine equivalence classes X 0 1 , . . . , X 0 m0 for updated s and (unchanged) t
15. [1, n] = A0 ∪ B 0 ∪ C 0 ∪ D0
16. define X 0 1 , . . . , X 0 m0 whose union is (A0 ∪ B 0 )
17. define V 0 = {1, . . . , m0 }
18. define E 0 = {(i, j) : 1 ≤ i, j ≤ m0 } by equation (21)
19. define G0 = (V 0 , E 0 ) //note that G0 is an acyclic multigraph
20. let σ ∈ Sm0 be a topological sort of V 0 //Sm0 denotes set of all permutations on [1, m0 ]
//process shifts in Xσ(i) in topological order by adapting part of Algorithm 14
21. for i = 1 to m0
22.
define V (Xσ(i) ) by equation (24)
23.
define E(Xσ(i) ) by equation (25)
24.
define G(Xσ(i) ) = (V (Xσ(i) ), E(Xσ(i) ))
//IP solution
Pof minimum feedback vertex set problem
25.
maximize
v∈V (Xσ(i) ) xv where xv ∈ {0, 1}, subject to constraints (†i ) and (‡i )
//first P
constraint removes vertex from each simple cycle of G(Xσ(i) )
(†i )
xv < |C| for each simple directed cycle C of G(Xσ(i) )
v∈C
//ensure shift moves cannot be applied if they share same base pair from s or t
(‡i ) xv + xv0 ≤ 1, for distinct vertices v = (x, y, z), v 0 = (x0 , y 0 , z 0 ) with |{x, y, z} ∩ {x0 , y 0 , z 0 }| = 2
//define the induced, acyclic digraph G(Xσ(i) )
26.
V (Xσ(i) ) = {v ∈ V (Xσ(i) ) : xv = 1}
27.
E(Xσ(i) ) = {(v, v 0 ) : v, v 0 ∈ V ∧ (v, v 0 ) ∈ E(Xσ(i) )}
28.
let G(Xσ(i) ) = (V (Xσ(i) ), E(Xσ(i) ))
//handle special, closed 2-cycles
29.
for each closed 2-cycle [x] = {a1 , a2 , a3 , a4 } as depicted in Figure 5
30.
if [x] is of type A as depicted in Figure 5a
31.
remove base pair from s by equation (18)
32.
if [x] is of type B as depicted in Figure 5b
33.
remove base pair from s by equation (19)
//remove base pairs from s that are not involved in a shift move
34.
V (Xσ(i) ).s = {(x, y) : ∃v ∈ V (Xσ(i) )(v.s = (x, y))}
35.
for (x, y) ∈ s − t
36.
if (x, y) 6∈ V (Xσ(i) ).s
16
remove (x, y) from s; numMoves = numMoves+1
//topological sort of IP solution V (Xσ(i) )
38.
topological sort of G(Xσ(i) ) using DFS to obtain total ordering ≺ on V (Xσ(i) )
38.
for v = (x, y, z) ∈ V in topologically sorted order ≺
//check if shift would create a base triple, as in type 1,5 paths from Figure 3
40.
if s[x] = 1 //i.e. {u, x} ∈ s for some u ∈ [1, n]
41.
remove {u, x} from s; numMoves = numMoves+1
42.
shift {y, z} to {x, y} in s; numMoves = numMoves+1
//remove any remaining base pairs from s that have not been shifted
43.
for (y, z) ∈ s − t which satisfy y, z ∈ Xσ(i)
44.
if (x, y, z) ∈ V (Xσ(i) ) where x = t[y]
45.
shift base pair {y, z} ∈ s to {x, y} ∈ t; numMoves = numMoves+1
46.
else //{y, z} is a remaining base pair of s but cannot be applied in a shift
47.
remove (x, y) from s; numMoves = numMoves+1
//add remaining base pairs from t to s
48. for (x, y) ∈ t − s
49.
s = s ∪ {(x, y)}
50.
numMoves = numMoves + 1
51. return moveSequence, numMoves
37.
5
5.1
Benchmarking results
Random sequences
Given a random RNA sequence a = a1 , . . . an of length n, we generate a list L of all possible base pairs,
then choose with uniform probability a base pairs (x, y) from L, add (x, y) to the secondary structure s
being constructed, then remove all base pairs (x0 , y 0 ) from L that either touch or cross (x, y), and repeat
these last three steps until we have constructed a secondary structure having the desired number (n/5) of
base pairs. If the list L is empty before completion of the construction of secondary structure s, then reject
s and start over. The following pseudocode describes how we generated the benchmarking data set, where
for each sequence length n = 10, 15, 20, · · · , 150 nt, twenty-five random RNA sequences were generated of
length n, with probability of 1/4 for each nucleotide, in which twenty secondary structures s, t were uniformly
randomly generated for each sequence so that 40% of the nucleotides are base-paired.
1.
2.
3.
4.
5.
6.
for n = 10 to 150 with step size 10
for numSeq = 1 to 25
generate random RNA sequence a = a1 , . . . , an of length n
generate 20 random secondary structures of a
for all 20
= 190 pairs of structures s, t of a
2
compute optimal and near-optimal M S2 folding trajectories from s to t
The number of computations per sequence length is thus 25 · 190 = 4750, so the size of the benchmarking
set is 15 · 4750 = 71, 250. This benchmarking set is used in Figures 12 – 16. Figure 12 compares various distance measures discussed in this paper: M S2 distance computed by the optimal IP Algorithm 14,
approximate M S2 distance computed by the near-optimal Algorithm 15, pk − M S2 distance that allows
pseudoknotted intermediate structures, M S1 distance, and Hamming distance divided by 2. Additionally,
this figure distinguishes the number of base pair additions/removals and shifts in the M S2 distance.
Figure 13a shows the scatter plots and Pearson correlation coefficients all pairs of the distance measures:
M S2 distance, near-optimal M S2 distance, pk − M S2 distance, Hamming distance divided by 2, M S1
distance. In contrast to Figure 13, the second panel Figure 13b shows the length-normalized values. It is
unclear why M S2 distance has a slightly higher length-normalized correlation with both Hamming distance
divided by 2 and M S1 distance, than that with approximate M S2 distance, as computed by Algorithm 15
– despite the fact that the latter algorithm approximates M S2 distance much better than either Hamming
distance divided by 2 or M S1 distance. Figure 14 shows that run-time of Algorithms 14 and 15, where the
17
former is broken down into time to generate the set of directed cycles and the time for the IP solver. Note that
there is a 10-fold speed-up in Algorithm 14 from this paper, compared with a precursor of this algorithm that
appeared in the proceedings of the Workshop in Bioinformatics (WABI 2017). Since Algorithm 15 applies
Algorithm 14 to each equivalence class, there is a corresponding, but less striking speed-up in the nearoptimal algorithm. Since run-time depends heavily on the number of directed cycles in the conflict digraphs,
Figure 15a shows the size of vertex and edge sets of the conflict digraphs for the benchmarking data, and
Figure 15b depicts the cycle length distribution for benchmarking data of length 150; for different lengths,
there are similar distributions (data not shown). Finally, Figure 15c showns the (presumably) exponential
increase in the number of directed cycles, as a function of sequence length. Since Algorithm 15 does not
compute the collection of all directed cycles (but only those for each equivalence class), the run time of
Algorithm 15 appears to be linear in sequence length, compared to the (presumably) exponential run time
of Algorithm 14.
5.2
Rfam sequences
In this section, we use data from the Rfam 12.0 database [28] for analogous computations as those from
the previous benchmarking section. For each Rfam family having average sequence length less than 100 nt,
one sequence is randomly selected, provided that the base pair distance between its MFE structure and its
Rfam consensus structure is a minimum. For each such sequence a, the target structure t was taken to be
the secondary structure having minimum free energy among all structures of a that are compatible with
the Rfam consensus structure, as computed by RNAfold -C [25] constrained with the consensus structure
of a. The corresponding initial structure s for sequence a was selected from a Zuker-suboptimal structure,
obtained by RNAsubopt -z [25], with the property that |dBP (s, t) − dH (s, t)| < 0.2 · dBP (s, t). Since we
know from Figure 14 that run time of the optimal IP Algorithm 14 depends on the number of cycles in the
corresponding RNA conflict digraph, the last criterion is likely to result in a less than astronomical number of
cycles. The resulting dataset consisted of 1333 sequences, some of whose lengths exceed 100 nt. Nevertheless,
the number of cycles in the RNA constraint digraph of 22 of the 1333 sequences exceeded 50 million (an
upper bound set for our program), so all figures described in this section are based on 1311 sequences from
Rfam.
Figure 17 depicts the moving averages in centered windows [x−2, x+2] of the following distance measures
for the 1311 sequences extracted from Rfam 12.0 as described. Distance measures include (1) optimal M S2 distance computed by the exact IP (optimal) Algorithm 14 (where the number of base pair additions (+)
or removals (−) is indicated, along with the number of shifts), (2) near-optimal M S2 -distance computed by
near-optimal Algorithm 15, (3) Hamming distance divided by 2, (4) M S1 distance aka base pair distance, (5)
pseudoknotted M S2 distance (pk-M S2 ) computed from Algorithm 9, (6) optimal local M S2 with parameter
d = 10, and (7) optimal local M S2 with parameter d = 20. The latter values were computed by a variant of
the exact IP Algorithm 14 with locality parameter d, defined to allow base pair shifts of the form (x, y) → (x, z)
or (y, x) → (z, x) only when |y −z| ≤ d. This data suggests that Hamming distance over 2 (dH (s, t)/2) closely
approximates the distance computed by near-optimal Algorithm 15, while pk-M S2 distance (dpk-M S2 (s, t))
is a better approximation to M S2 distance than is Hamming distance over 2. Figure 18 presents scatter plots
and Pearson correlation values when comparing various distance measures using the Rfam data. Figure 18a
[resp. Figure 18b] presents Pearson correlation [resp. normalized Pearson correlation] values computed,
where by normalized, we mean that for each of the 1311 extracted Rfam sequences a with corresponding
initial structure s and target structure t, the length-normalized distance measures d(s, t)/|a| are correlated.
Figure 19 depicts the moving average run times as a function of sequence length, where for given value x the
run times are averaged for sequences having length in [x − 2, x + 2]. Finally, Figure 20 depicts the number
of sequences of various lengths used in the Rfam benchmarking set of 1311 sequences.
6
Conclusion
In this paper, we have introduced the first optimal and near-optimal algorithms to compute the shortest
RNA secondary structure folding trajectories in which each intermediate structure is obtained from its
predecessor by the addition, removal or shift of a base pair; i.e. the shortest M S2 trajectories. Since helix
zippering and defect diffusion employ shift moves, one might argue that it is better to include shift moves
18
when physical modeling RNA folding, and indeed the RNA folding kinetics simulation program Kinfold [13]
uses the M S2 move set by default. Using the novel notion of RNA conflict directed graph, we describe an
optimal and near-optimal algorithm to compute the shortest M S2 folding trajectory. Such trajectories pass
through substantially higher energy barriers than trajectories produced by Kinfold, which uses Gillespie’s
algorithm [17] (a version of event-driven Monte Carlo simulation) to generate physically realistic M S2 folding
trajectories. We have shown in Theorem 2 that it is NP-hard to compute the M S2 folding trajectory having
minimum energy barrier, and have presented anecdotal evidence that suggests that it may also NP-hard to
compute the shortest M S2 folding trajectory. For this reason, and because of the exponentially increasing
number of cycles (see Figure 16) and subsequent time requirements of our optimal IP Algorithm 14, it is
unlikely that (exact) M S2 distance prove to be of much use in molecular evolution studies such as [3, 42, 15].
Nevertheless, Figures 12 and 17 suggest that either pk-M S2 distance and/or near-optimal M S2 distance
may be a better approximation to (exact) M S2 distance than using Hamming distance, as done in [36, 43].
However, given the high correlations between these measures, it is unlikely to make much difference in
molecular evolution studies.
Our graph-theoretic formulation involving RNA conflict digraphs raises some interesting mathematical
questions partially addressed in this paper; in particular, it would be very interesting to characterize the
class of digraphs that can be represented by RNA conflict digraphs, and to determine whether computing
the shortest M S2 folding trajectory is N P -hard. We hope that the results presented in this paper may lead
to resolution of these questions.
References
[1] J. Bang-Jensen and G. Gutin. Digraphs : theory, algorithms, and applications. Springer monographs in
mathematics. Springer, London, New York, 2001. Deuxième impression avec corrections.
[2] A.H. Bayegan and P. Clote. An IP algorithm for RNA folding trajectories. In K. Reiner and R. Schwartz,
editors, Algorithms in Bioinformatics: 17th International Workshop, WABI 2017, Boston MA, USA,
August 21-23, 2017. Springer, 2017.
[3] E. Borenstein and E. Ruppin. Direct evolution of genetic robustness in microRNA. Proc. Natl. Acad.
Sci. U.S.A., 103(17):6593–6598, April 2006.
[4] Pierre Charbit, Stéphan Thomassé, and Anders Yeo. The minimum feedback arc set problem is np-hard
for tournaments. Combinatorics, Probability & Computing, 16(1):1–4, 2007.
[5] S. S. Cho, D. L. Pincus, and D. Thirumalai. Assembly mechanisms of RNA pseudoknots are determined
by the stabilities of constituent secondary structures. Proc. Natl. Acad. Sci. U.S.A., 106(41):17349–
17354, October 2009.
[6] P. Clote and A. Bayegan. Network Properties of the Ensemble of RNA Structures. PLoS. One.,
10(10):e0139476, 2015.
[7] T.H. Cormen, C.E. Leiserson, and R.L. Rivest. Algorithms. McGraw-Hill, 1990. 1028 pages.
[8] K. Darty, A. Denise, and Y. Ponty. VARNA: Interactive drawing and editing of the RNA secondary
structure. Bioinformatics, 25(15):1974–1975, August 2009.
[9] I. Dotu, J. A. Garcia-Martin, B. L. Slinger, V. Mechery, M. M. Meyer, and P. Clote. Complete
RNA inverse folding: computational design of functional hammerhead ribozymes. Nucleic. Acids. Res.,
42(18):11752–11762, February 2015.
[10] I. Dotu, W. A. Lorenz, P. VAN Hentenryck, and P. Clote. Computing folding pathways between RNA
secondary structures. Nucleic. Acids. Res., 38(5):1711–1722, 2010.
[11] E. C. Dykeman. An implementation of the Gillespie algorithm for RNA kinetics with logarithmic time
update. Nucleic. Acids. Res., 43(12):5708–5715, July 2015.
19
[12] P. Festa, P. Pardalos, and M.G.C. Resende. Feedback set problems. In C. Floudas and P. Pardalos,
editors, Encyclopedia of Optimization, pages 1005–1016. Springer US, Boston, MA, 2009. Second edition.
[13] C. Flamm, W. Fontana, I.L. Hofacker, and P. Schuster. RNA folding at elementary step resolution.
RNA, 6:325–338, 2000.
[14] C. Flamm, I.L. Hofacker, P.F. Stadler, and M. Wolfinger. Barrier trees of degenerate landscapes. Z.
Phys. Chem., 216:155–173, 2002.
[15] J. A. Garcia-Martin, A. H. Bayegan, I. Dotu, and P. Clote. RNAdualPF: software to compute the
dual partition function with sample applications in molecular evolution theory. BMC. Bioinformatics,
17(1):424, October 2016.
[16] J. A. Garcia-Martin, I. Dotu, and P. Clote. RNAiFold 2.0: a web server and software to design custom
and Rfam-based RNA molecules. Nucleic. Acids. Res., 43(W1):W513–W521, July 2015.
[17] D.T. Gillespie. A general method for numerically simulating the stochastic time evolution of coupled
chemical reactions. J Comp Phys, 22(403):403–434, 1976.
[18] K.A. Harris and D.M. Crothers. The Leptomonas collosoma spliced leader RNA can switch between
two alternate structural forms. Biochemistry, 32(20):5301–5311, 1993.
[19] Matthew S. Hecht and Jeffrey D. Ullman. Flow graph reducibility. SIAM J. Comput., 1(2):188–202,
1972.
[20] C. Hobartner and R. Micura. Bistable secondary structures of small RNAs and their structural probing
by comparative imino proton NMR spectroscopy. J. Mol. Biol., 325(3):421–431, January 2003.
[21] D.B. Johnson. Finding all the elementary circuits of a directed graph. SIAM J. Comput., 4:77–84, 1975.
[22] Richard M. Karp. Reducibility among combinatorial problems. In Proceedings of a symposium on
the Complexity of Computer Computations, held March 20-22, 1972, at the IBM Thomas J. Watson
Research Center, Yorktown Heights, New York., pages 85–103, 1972.
[23] B. Knudsen and J. Hein. Pfold: RNA secondary structure prediction using stochastic context-free
grammars. Nucleic. Acids. Res., 31(13):3423–3428, July 2003.
[24] Kazimierz Kuratowski. Sur le problème des gauches en topologie. Fundamenta Mathematicae, 15:271–
283, 1930.
[25] R. Lorenz, S. H. Bernhart, C. Höner zu Siederdissen, H. Tafer, C. Flamm, P. F. Stadler, and I. L.
Hofacker. Viennarna Package 2.0. Algorithms. Mol. Biol., 6:26, 2011.
[26] C.L. Lucchesi and D.H. Younger. A minimax arc theorem for directed graphs. J. London Math. Soc,
17:369–374, 1978.
[27] D.H. Mathews, J. Sabina, M. Zuker, and H. Turner. Expanded sequence dependence of thermodynamic
parameters provides robust prediction of RNA secondary structure. J. Mol. Biol., 288:911–940, 1999.
[28] E. P. Nawrocki, S. W. Burge, A. Bateman, J. Daub, R. Y. Eberhardt, S. R. Eddy, E. W. Floden, P. P.
Gardner, T. A. Jones, J. Tate, and R. D. Finn. Rfam 12.0: updates to the RNA families database.
Nucleic. Acids. Res., 43(Database):D130–D137, January 2015.
[29] R. Nussinov and A. B. Jacobson. Fast algorithm for predicting the secondary structure of single stranded
RNA. Proceedings of the National Academy of Sciences, USA, 77(11):6309–6313, 1980.
[30] K. Perrot and T.V. Pham. Np-hardness of minimum feedback arc set problem on Eulerian digraphs and
minimum recurrent configuration problem of chip-firing game. CoRR - Computing Research Repository
- arXiv, abs/1303.3708, 2013.
20
[31] D. Pörschke. Model calculations on the kinetics of oligonucleotide double-helix coil transitions: Evidence
for a fast chain sliding reaction. Biophys Chem, 2(2):83–96, August 1974.
[32] N. Rajewsky. microrna target predictions in animals. Nat. Genet., 38:S8–S13, June 2006.
[33] Vijaya Ramachandran. A minimax arc theorem for reducible flow graphs. SIAM J. Discrete Math.,
3(4):554–560, 1990.
[34] Vijaya Ramachandran. A minimax arc theorem for reducible flow graphs. SIAM J Disc. Math., 3(4):554–
560, 1990.
[35] P. Schattner, A. N. Brooks, and T. M. Lowe. The tRNAscan-SE, snoscan and snoGPS web servers for
the detection of tRNAs and snoRNAs. Nucleic. Acids. Res., 33(Web):W686–W689, July 2005.
[36] P. Schuster and P.F. Stadler. Modeling conformational flexibility and evolution of structure: RNA as an
example. In U. Bastille, M. Roman, and M. Vendruscolo, editors, Structural Approaches to SequenceEvolution, page 3–36. Springer, Heidelberg, 2007.
[37] E. Senter, I. Dotu, and P. Clote. RNA folding pathways and kinetics using 2D energy landscapes. J
Math Biol, 2014.
[38] A. Serganov, Y.R. Yuan, O. Pikovskaya, A. Polonskaia, L. Malinina, A.T. Phan, C. Hobartner, R. Micura, R.R. Breaker, and D.J. Patel. Structural basis for discriminative regulation of gene expression by
adenine- and guanine-sensing mRNAs. Chem. Biol., 11(12):1729–1741, 2004.
[39] Z. Sukosd, B. Knudsen, M. Vaerum, J. Kjems, and E. S. Andersen. Multithreaded comparative RNA
secondary structure prediction using stochastic context-free grammars. BMC. Bioinformatics, 12:103,
2011.
[40] C. Thachuk, J. Maňuch, L. Stacho, and A. Condon. NP-completeness of the direct energy barrier height
problem. Natural Computing, 10(1):391–405, 2011.
[41] D. H. Turner and D. H. Mathews. NNDB: the nearest neighbor parameter database for predicting
stability of nucleic acid secondary structure. Nucleic. Acids. Res., 38(Database):D280–D282, January
2010.
[42] A. Wagner. Robustness and evolvability: a paradox resolved. Proc. Biol Sci., 275(1630):91–100, January
2008.
[43] A. Wagner. Mutational robustness accelerates the origin of novel RNA phenotypes through phenotypic
plasticity. Biophys. J., 106(4):955–965, February 2014.
[44] S. Washietl and I. L. Hofacker. Identifying structural noncoding RNAs using RNAz. Curr Protoc
Bioinformatics, 0(O):O, September 2007.
[45] Michael T. Wolfinger, W. Andreas Svrcek-Seiler, Christoph Flamm, Ivo L. Hofacker, and Peter F.
Stadler. Efficient folding dynamics of RNA secondary structures. J. Phys. A: Math. Gen., 37:4731–
4741, 2004.
[46] S. Wuchty, W. Fontana, I.L. Hofacker, and P. Schuster. Complete suboptimal folding of RNA and the
stability of secondary structures. Biopolymers, 49:145–164, 1999.
[47] J. N. Zadeh, B. R. Wolfe, and N. A. Pierce. Nucleic acid sequence design via efficient ensemble defect
optimization. J. Comput. Chem., 32(3):439–452, February 2011.
21
Figure 1: Illustration of shift moves, taken from [6].
Figure 2: Defect diffusion [31], where a bulge migrates stepwise to become absorbed in an hairpin loop. The
move from structure (a) to structure (b) is possible by the shift (1, 12) → (1, 13), the move from (b) to (c) by
shift (2, 11) → (2, 12), etc. Our algorithm properly accounts for such moves with respect to energy models
A,B,C. Image taken from [6].
7
Figures
22
Figure 3: All possible maximal length red-green paths and cycles. Each equivalence class X, as defined
in Definition 7, can be depicted as a maximal length path or cycle, consisting of those positions x ∈ [1, n]
that are connected by alternating base pairs drawn from secondary structures s (green) and t (red). Each
path or cycle X is depicted in a fashion that the leftmost [resp. rightmost position] satisfies the following
properties, where End(s, X) [resp. End(t, X)] denotes the set of elements of X that are untouched by t
[resp. s]: for paths of type 1, |X| is even, |End(s, X)] = 2, b1 = min(End(s, X)), t[am ] = 0; for paths of
type 2, |X| is odd, |End(s, X)] = 1 = |End(t, X)|, b0 = min(End(s, X)), s[bm ] = 0; for paths of type 3,
|X| is odd, |End(s, X)] = 1 = |End(t, X)|, a0 = min(End(t, X)), t[am ] = 0; for paths of type 4, |X| is
even, |End(t, X)] = 2, a1 = min(End(t, X)), s[bm ] = 0; for paths of type 5, or cycles, |X| must be even,
and a1 = min(X). Note that the appearance of positions in left-to-right order does not necessarily respect
integer ordering, so the leftmost position is not necessarily the minimum min(X), nor is the rightmost
position necessarily the maximum max(X).
23
Figure 4: All six possible shift moves, in which a base pairs of s (teal) that touches a base pairs of t (red) is
shifted, thus reducing the base pair distance dBP (s, t) by 2. Each such shift move can uniquely be designated
by the triple (x, y, z), where y is the pivot position (common position to a base pair in both s and t), x is
the remaining position in the base pair in t, and z is the remaining position in the base pair in s.
Figure 5: Two special types of closed 2-cycles. (a) RNA conflict digraph G = (V, E) for secondary structures
t and s, where a1 < a2 < a3 < a4 and t = {(a1 , a2 ), (a3 , a4 )}, and s = {(a1 , a4 ), (a2 , a3 )}. Nodes of
V = {v1 , v2 , v3 , v4 } are the following: v1 = (a1 , a2 , a3 ) of type 1, v2 = (a3 , a4 , a1 ) of type 5, v3 = (a2 , a1 , a4 )
of type 4, and v4 = (a4 , a3 , a2 ) of type 2. (b) RNA conflict digraph G = (V, E) for secondary structures
t and s, where a1 < a2 < a3 < a4 and t = {(a1 , a4 ), (a2 , a3 )} and s = {(a1 , a2 ), (a3 , a4 )}. Nodes of
V = {v1 , v2 , v3 , v4 } are the following: v1 = (a1 , a4 , a3 ) of type 6, v2 = (a4 , a1 , a2 ) of type 3, v3 = (a2 , a3 , a4 )
of type 1, v4 = (a3 , a2 , a1 ) of type 2. Since the overlap between any two distinct vertices in (a) and (b) is
2, there are no edges in E for the conflict digraphs of (a) and (b). An optimal trajectory from s to t is
constructed by removing a base pair from s, performing a shift, and adding the remaining base pair from
t. In each case there are 2 choices for the base pair to remove and two choices for the shift, so 4 optimal
trajectories for each of (a) and (b).
24
Figure 6: Conflict digraph for a 20 nt toy example with sequence GGGAAAUUUC CCCAAAGGGG,
with initial structure s whose free energy is +0.70 kcal/mol, and target structure t whose
free energy is + 3.30 kcal/mol.
The conflict digraph contains 3 simple cycles: a first cycle
{(8, 20, 10), (9, 19, 11), (18, 10, 20), (19, 9, 1)} of size 4, a second cycle {[(8, 20, 10), (19, 9, 1)} of size 2, and
a third cycle {(18, 10, 20), (9, 19, 11)} of size 2.
Figure 7: Conflict digraphs for the 25 nt bistable switch with sequence UGUACCGGAA GGUGCGAAUC
UUCCG taken from Figure 1(b).1 of [20], in which the authors performed structural probing by comparative
imino proton NMR spectroscopy. (a) Minimum free energy (MFE) structure having -10.20 kcal/mol. (b)
Alternate metastable structure having next lowest free free energy of -7.40 kcal/mol. Two lower energy
structures exist, having -9.00 kcal/mol resp. -7.60 kcal/mol; however, each is a minor variant of the MFE
structure. (c) RNA conflict digraph G = (V, E), having directed edges (x, y, z) → (u, v, w) if the (unordered)
base pair {y, z} ∈ s touches or crosses the (unordered) base pair {u, v} ∈ t. Here, s is in the metastable
structure shown in (b) having -7.40 kcal/mol, while t is the MFE structure shown in (a) having -10.20
kcal/mol. The conflict digraph represents a necessary order of application of shift moves, in order to avoid the
creation of base triples or pseudoknots. Note that the digraph G is acyclic, but the IP solver must nevertheless
be invoked with constraint (‡) that precludes both vertices (20, 11, 6) and (25, 6, 11) from belonging to the
solution V . (d) RNA conflict digraph G0 = (V 0 , E 0 ), having similar definition in which roles of s and t are
reversed – i.e. M S2 folding pathways from the MFE structure to the (higher energy) metastable structure.
25
Figure 8: Conflict digraph for the 56 nt spliced leader RNA from L. collosoma with sequence AACUAAAACA
AUUUUUGAAG AACAGUUUCU GUACUUCAUU GGUAUGUAGA GACUUC. (a) Metastable structure
having free energy of -9.20 kcal/mol. (b) Minimum free energy (MFE) structure having free energy of -9.40
kcal/mol. (c) The RNA conflict digraph for refolding from metastable s to MFE t contains 12 vertices, 61
directed edges and no directed cycles. Minimum free energy (MFE) and metastable structures in (a) and
(b) computed by Vienna RNA Package [25]. Secondary structure images generated using VARNA [8].
26
Figure 9: (a) Rainbow diagram for spliced leader RNA from L. collosoma, in which the base pairs for
the initial structure s (Figure 8a) are shown below the line in red, while those for the target structure t
(Figure 8b) are shown above the line in blue. (b) Free energies of structures in the shortest M S2 folding
trajectory for spliced leader are shown by the dotted blue line, while those for the energy-optimal M S2
folding trajectory are shown in the solid red line. Algorithm 14 was used to compute the shortest M S2
trajectory, while the programs RNAsubopt [46] and barriers [14] were used to compute the energy-optimal
folding trajectory.
27
Figure 10: Gene ON and gene OFF structures and the RNA conflict digraph for the 156 nt xanthine
phosphoribosyltransferase (xpt) riboswitch from B. subtilis – structures consistent with in-line probing data
taken from Figure 1A of [38]. (a) Gene ON structure (default) in absence of free guanine, having (computed)
free energy of -33.11 kcal/mol. (b) Gene OFF structure when guanine binds cytidine in position 66, having
(computed) free energy of -56.20 kcal/mol (guanine not shown). (c) Rainbow diagram with red gene ON
structure below line and blue gene OFF structure above line. Rainbow diagrams allow one to determine by
visual inspection when base pairs touch or cross. (d) Conflict digraph G = (V, E), containing 18 vertices,
113 directed edges, and 1806 directed cycles. Minimum free energy (MFE) and metastable structures in (a)
and (b) computed by Vienna RNA Package [25]. Free energy computations using Turner energy model [41]
computed with Vienna RNA Package [25]. Secondary structure images generated using VARNA [8].
28
Figure 11: (a) Initial gene ON structure (default) in absence of free guanine (enlarged image). (b) Target
gene OFF structure when guanine binds cytidine in position 66 (enlarged image). (c) Shortest M S2 folding
trajectory from the gene ON structure s to the gene OFF structure t for the 156 nt xpt riboswitch from
B. subtilis, described in the caption to Figure 10. Note the initial elongation of the P1 helix by the first
shift, followed by the stepwise removal of the anti-terminator and construction of the terminator loops by
shift 2-10, followed by base pair additions to lengthen the terminator loop. Although (a,b) are duplicated
from the previous figure, this is necessary to follow the sequence of moves in the M S2 trajectory. Secondary
structure images generated using VARNA [8]. (d) Free energies of structures in the shortest M S2 folding
trajectory for xpt are shown by the dotted blue line, while those for the an energy near-optimal M S1 folding
trajectory are shown in the solid red line. Algorithm 14 was used to compute the shortest M S2 trajectory,
while the program RNAtabuPath [10] was used to compute the energy near-optimal M S1 folding trajectory.
The size of the 156 nt xpt riboswitch and the fact that the program RNAsubopt would need to generate all
secndary structures within 30 kcal/mol of the minimum free energy -54.1 kcal/mol preclude any possibility
that the optimal M S2 trajectory can be computed by application of the program barriers [14].
29
Figure 12: Benchmarking statistics for optimal and near-optimal algorithm to compute minimum length M S2
folding trajectories between random secondary structures s, t of random RNA sequences of variable lengths.
For each sequence length n = 10, 15, 20, · · · , 150 nt, twenty-five random RNA sequences were generated of
length n, with probability of 1/4 for each nucleotide. For each RNA sequence, twenty secondary structures
s, t were uniformly randomly generated so that 40% of the nucleotides are base-paired. Thus the number of
computations per sequence length is thus 25 · 190 = 4750, so the size of the benchmarking set is 15 · 4750 =
71, 250. Using this dataset, the average M S2 distance was computed for both the exact IP Algorithm 14
and the near-optimal Algorithm 15. In addition to M S2 distance computed by the exact IP and the nearoptimal algorithm, the figure displays pk − M S2 distance (allowing pseudoknots in intermediate structures)
as computed by Algorithm 9, the M S1 distance (also known as base pair distance), Hamming distance over
2, and provides a breakdown of the M S1 distance in terms of the number of base pair addition/removal
moves “num base pair +/- (optimal)” and the shift moves “num shift moves (optimal)”.
30
Figure 13: Pairwise correlations for optimal M S2 distance, pk − M S2 distance, near-optimal M S2 distance,
Hamming distance divided by 2, and M S1 distance (also called base pair distance). For each two measures,
scatter plots were created for the 71, 250 many data points from the benchmarking set described in Figure 12.
Pearson correlation and normalized Pearson correlation values computed, where by normalized, we mean that
for each of the 71, 250 data points, we consider the length-normalized distance (distance divided by sequence
length). These correlations are statistically significant – each Pearson correlation values has a p-value less
than 10−8 .
31
Figure 14: Run time for the exact IP (optimal) algorithm 14 and the near-optimal algorithm 15 to compute
minimum length M S2 folding trajectories for the same data set from previous Figure 12. Each data point
represents the average µ ± σ where error bars indicate one standard deviation, taken over 71, 250 many
sequence/structure pairs. Run time of the optimal algorithm depends on time to perform topological sort,
time to enumerate all directed cycles using our Python implementation of Johnson’s algorithm [21], and time
for the Gurobi IP solver (ordered here by increasing time demands).
32
Figure 15: (Left) Average size of vertex sets V and of directed edge sets E for RNA conflict digraphs G =
(V, E) for the data set of described in Figure 12. Error bars represent ±1 standard deviation. Clearly the size
of a conflict digraph grows linearly in the length n of random RNAs a = a1 , . . . , an , given random secondary
structures s, t having n/5 base pairs. (Right) Cycle length distribution for random RNAs a = a1 , . . . , an of
length n = 150, with randomly chosen secondary structures s, t having n/5 base pairs, using data extracted
from the data set described in Figure 12. For values of n = 50, . . . , 150, the cycle length distribution appears
approximately normal, although this is not the case for n ≤ 40 (data not shown).
Figure 16: Average number of directed cycles as a function of sequence length for the data set described in
Figure 12. For each random RNA sequence a = a1 , . . . , an of length n, and for each pair of random secondary
structures s, t of a having n/5 base pairs, we computed the total number of directed cycles in the conflict
digraph G(a, s, t). The figure suggests that starting at a threshold sequence length n, there is an exponential
growth in the number of directed cycles in the conflict digraph of random sequences of length n.
33
Figure 17: Moving averages of distance measures graphed as a function of sequence length for the 1311
many sequences extracted from Rfam 12.0 (see text). Distance measures included optimal M S2 -distance
computed by the exact IP (optimal) algorithm 14 (where the number of base pair additions (+) or removals
(−) is indicated, along with the number of shifts),, near-optimal M S2 -distance computed by near-optimal
algorithm 15, Hamming distance divided by 2, M S1 distance aka base pair distance, pseudoknotted M S2
distance (pk-M S2 ) computed from Lemma 8, optimal local M S2 with parameter d = 10, and optimal local
M S2 with parameter d = 20. The latter values were computed by a variant of the exact IP algorithm 14 where
shift moves were restricted to be local with parameter d, whereby base pair shifts of the form (x, y) → (x, z)
or (y, x) → (z, x) were allowed only when |y − z| ≤ d. All moving averages were computed over symmetric
windows of size 9, i.e. [i − 4, i + 4]. From smallest to largest value, the measures are: number of shifts in
optimal M S2 trajectory < number of base pair additions or deletions (+/−) in optimal M S2 trajectory <
pk-M S2 < M S2 distance < Hamming distance over 2 ≈ near-optimal M S2 < M S2 with locality parameter
d = 20 < M S2 with locality parameter d = 10 < M S1 .
34
Figure 18: Pairwise correlations using Rfam benchmarking data (see text) for optimal M S2 distance, pk −
M S2 distance, near-optimal M S2 distance, Hamming distance divided by 2, and M S1 distance (also called
base pair distance). For each two measures, scatter plots were created for the 1311 many data points. Pearson
correlation and normalized Pearson correlation values computed, where by normalized, we mean that for each
of the 1311 data points, we consider the length-normalized distance (distance divided by sequence length).
These correlations are statistically significant – each Pearson correlation values has a p-value less than
2.2 · 10−16 .
Figure 19: Moving averages of run time as a function of sequence length for the 1311 sequences extracted from
Rfam 12.0 (see text). Distance measures considered are the exact M S2 distance computed by the optimal
IP Algorithm 14, an approximation to the M S2 distance computed by the near-optimal IP Algorithm 15,
the pk − M S2 distance (allowing pseudoknots in intermediate structures) as computed by Algorithm 9, and
two variants of exact M S2 distance, where shifts are restricted by locality parameter d = 10, 20. These
latter values were computed by the exact IP Algorithm 14 modified to allow base pair shifts of the form
(x, y) → (x, z) or (y, x) → (z, x) only when |y − z| ≤ d.
35
Figure 20: Number of sequences as a function of sequence length for the 1311 sequences extracted from Rfam
12.0 and used in benchmarking tests (see text).
Figure 21: RNA conflict digraph G = (V, E) for secondary structures t and s, where t =
{(a2 , a1 ), (b1 , b2 ), (c1 , c2 ), (d1 , d2 )}, and s = {(a2 , a3 ), (b2 , b3 ), (c2 , c3 ), (d2 , d3 )}. The triplet nodes of V =
{va , vb , vc , vd } are the following: va = (a1 , a2 , a3 ) of type 3, shift vb = (b1 , b2 , b3 ) of type 1, shift
vc = (c1 , c2 , c3 ) of type 1, and shift vd = (d1 , d2 , d3 ) of type 1. The edges in E are the following: va → vb ,
vb → vc , vc → vd , vd → va . The conflict digraph G = (V, E) is order-isomorphic to the digraph G0 = (V 0 , E 0 ),
where V 0 = {1, 2, 3, 4} and edges are as follows: 1 → 2 → 3 → 4 → 1.
36
Figure 22: (a) Complete bipartite graph K3,3 . A finite graph is planar if and only if it does not contain
the forbidden graph K3,3 or the complete graph K5 [24]. (b) Directed graph realized by the RNA conflict
digraph in Figure 23. It follows that RNA conflict digraphs are not necessarily planar. (c) Directed graph
realized by the RNA conflict digraph in Figure 24. A flow graph is reducible if and only if it does not contain
such a forbidden flow graph, where edges between nodes may be replaced by arc-disjoint directed paths [19].
It follows that RNA conflict digraphs are not necessarily reducible flow graphs.
Figure 23: RNA conflict digraph that realizes the digraph K3,3 depicted in Figure 22b, whose undirected
red edges represent the undirected graph K3,3 depicted in Figure 22b. The nonplanar complete, bipartite
digraph K3,3 , with shift moves c, b, a, d, e, f , all of type 1, in order from left to right – i.e. order of positions
along the x-axis is given by: c1 , b1 , a1 , a2 , b2 , c2 , d1 , e1 , f1 , c3 , b3 , a3 , f2 , e2 , d2 , d3 , e3 , f3 . Notice that as crosses
bt , ct , dt , et and ft so c ← a, b ← a and a → d, a → e, a → f ; bs crosses ct , dt , et and ft so c ← b and b → d,
b → e, b → f ; cs crosses dt , et and ft so c → d, c → e, c → f .
Figure 24: Forbidden flow graph with nodes i = (i1 , i2 , i3 ), a = (a1 , a2 , a3 ), b = (b1 , b2 , b3 ), c = (c1 , c2 , c3 ),
where nodes i, a, c are of type 1 and node b is of type 3. Notice that is crosses at so i → a; as crosses bt so
a → b; as crosses ct so a → c; bs crosses ct so b → c; cs crosses bt so c ← b.
37
Figure 25: (a) Digraph of an ordered 4-cycle, which is representable by an RNA conflict digraph, as shown
in Figure 21. (b) Digraph of an ordered 4-cycle, which is not representable by an RNA conflict digraph.
Note that digraph (b) is Eulerian, with the property that the in-degree of each vertex equals its out-degree.
(c) Digraph of a tournament on 4 vertices, which is not representable by an RNA conflict digraph. Digraph
(a) is isomorphic with digraph (b), thus showing that representability is not preserved under isomorphism.
3
Since it is not difficult to show that all 2(2) = 8 tournaments on 3 nodes are representable by RNA conflict
digraphs (data not shown), it follows that digraph (c) is a minimum sized non-representable tournament,
n
which we verified by constraint programming. In general there are 2( 2 ) many tournaments on n.
Figure 26: Example of a 4-node digraph in (a) that is not representable by an RNA conflict digraph. However,
by adding an edge to digraph (a), we obtain a representable digraph in (b). Note that digraph (a) is neither
order-isomorphic to digraph (b), nor is there an order-preserving embedding of digraph (a) into digraph (b).
38
A
Classification of edges in RNA constraint digraphs
In this section, we describe the collection of all possible directed edges v → v 0 in which v.s crosses v 0 .t, that
can appear in an RNA conflict digraph, classified as forward, backward or 2-cycles and according to each
type of vertex (see Figure 4 for the six types of vertices). It is straightforward for the reader to imaging
additional directed edges v → v 0 in which v.s touches v 0 .t, so these are not shown. In addition, we provide
the pseudocode for a slow branch-and-bound algorithm to determine a shortest M S2 trajectory.
Given two secondary structures s, t for the RNA sequence a1 , . . . , an , recall that notation for a shift move
from the (unordered) base pair {x, y} ∈ s to the (unordered) base pair {y, z} ∈ t is given by the triple
(z, y, x), where the middle coordinate y is the pivot position, common to both base pairs {x, y} ∈ s and
{y, z} ∈ t, while the first [resp. last] coordinate z [resp. x] is the remaining position from the base pair
{y, z} ∈ t [resp. {x, y} ∈ s]. A directed edge is given from shift move (x, y, z) to shift move (u, v, w) if the
base pair {y, z} ∈ s from the first shift move crosses with the base pair {u, v} ∈ t from the second shift move;
i.e. min(u, v) < min(y, z) < max(u, v) < max(y, z) or min(y, z) < min(u, v) < max(y, z) < max(u, v). The
reason for the directed edge is that if the second shift (u, v, w) is applied before the first shift (x, y, z), then
a pseudoknot (crossing) would be created; it follows that the first shift must be applied before the second
shift.
Edges may be forward (left-to-right) or backward (right-to-left), depending on whether the pivot position
of the first shift is (strictly) less than or (strictly) greater than the pivot position of the second shift. This
section does not list similar examples, where the (unordered) base pair {y, z} ∈ s from the first shift move
touches the (unordered) base pair {u, v} ∈ t from the second shift move, as such examples are clear from
Figure 3 of the main text.
A.1
1
Forward Edges
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
forward edge of type 5 → 5
forward edge of type 5 → 1
forward edge of type 2 → 5
forward edge of type 5 → 6
(4,5,2) → (3,6,1)
(3,4,1) → (2,5,6)
(5,4,2) → (3,6,1)
(3,4,1) → (2,6,5)
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
forward edge of type 1 → 1
forward edge of type 1 → 6
forward edge of type 2 → 1
forward edge of type 1 → 5
(2,3,6) → (1,4,5)
(2,3,6) → (1,5,4)
(4,3,1) → (2,5,6)
(2,3,5) → (4,6,1)
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
forward edge of type 2 → 6
forward edge of type 1 → 3
forward edge of type 4 → 1
forward edge of type 1 → 2
(4,3,1) → (2,6,5)
(1,2,5) → (6,3,4)
(3,2,6) → (1,4,5)
(1,2,5) → (6,4,3)
39
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
forward edge of type 1 → 1
forward edge of type 4 → 6
forward edge of type 4 → 5
forward edge of type 1 → 6
(1,2,4) → (3,5,6)
(3,2,6) → (1,5,4)
(3,2,5) → (4,6,1)
(1,2,4) → (3,6,5)
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
forward edge of type 3 → 3
forward edge of type 4 → 3
forward edge of type 3 → 2
forward edge of type 3 → 1
(6,1,4) → (5,2,3)
(2,1,5) → (6,3,4)
(6,1,4) → (5,3,2)
(6,1,3) → (2,4,5)
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
forward edge of type 4 → 2
forward edge of type 4 → 1
forward edge of type 3 → 6
forward edge of type 4 → 6
(2,1,5) → (6,4,3)
(2,1,4) → (3,5,6)
(6,1,3) → (2,5,4)
(2,1,4) → (3,6,5)
A.2
1
Backward Edges
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
backward edge of type 6 ← 6
backward edge of type 6 ← 2
backward edge of type 1 ← 6
backward edge of type 6 ← 5
(2,5,4) ← (1,6,3)
(1,4,3) ← (6,5,2)
(2,4,5) ← (1,6,3)
(1,4,3) ← (5,6,2)
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
backward edge of type 2 ← 2
backward edge of type 1 ← 2
backward edge of type 2 ← 5
backward edge of type 1 ← 5
(6,3,2) ← (5,4,1)
(1,3,4) ← (6,5,2)
(6,3,2) ← (4,5,1)
(1,3,4) ← (5,6,2)
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
backward edge of type 2 ← 6
backward edge of type 2 ← 4
backward edge of type 3 ← 2
backward edge of type 2 ← 1
(5,3,2) ← (1,6,4)
(5,2,1) ← (4,3,6)
(6,2,3) ← (5,4,1)
(5,2,1) ← (3,4,6)
40
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
backward edge of type 3 ← 5
backward edge of type 2 ← 2
backward edge of type 3 ← 6
backward edge of type 2 ← 5
(6,2,3) ← (4,5,1)
(4,2,1) ← (6,5,3)
(5,2,3) ← (1,6,4)
(4,2,1) ← (5,6,3)
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
backward edge of type 4 ← 4
backward edge of type 3 ← 4
backward edge of type 4 ← 1
backward edge of type 3 ← 1
(4,1,6) ← (3,2,5)
(5,1,2) ← (4,3,6)
(4,1,6) ← (2,3,5)
(5,1,2) ← (3,4,6)
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
backward edge of type 4 ← 2
backward edge of type 3 ← 2
backward edge of type 4 ← 5
backward edge of type 3 ← 5
(3,1,6) ← (5,4,2)
(4,1,2) ← (6,5,3)
(3,1,6) ← (4,5,2)
(4,1,2) ← (5,6,3)
A.3
1
2-Cycles
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
6
1
2
3
4
5
2-cycle of type 1 ↔ 5
2-cycle of type 2 ↔ 6
2-cycle of type 3 ↔ 5
2-cycle of type 4 ↔ 6
(1,3,5) ↔ (4,6,2)
(5,3,1) ↔ (2,6,4)
(6,2,4) ↔ (3,5,1)
(4,2,6) ↔ (1,5,3)
1
2
3
4
5
6
1
2
3
4
5
2-cycle of type 3 ↔ 1
2-cycle of type 4 ↔ 2
(5,1,3) ↔ (2,4,6)
(3,1,5) ↔ (6,4,2)
A.4
6
Summary tables of shift moves edges
Table 1 presents a count of all 12 possible bidirectional edges, while Table 2 [resp. Table 3] presents a count
of all 34 possible forward [resp. back] directed edges. Here, by bidirectional edge between nodes x and y, we
mean the existence of directed edges x → y and y → x. Figures in Sections A.3, A.1 and A.2 depict all of
these these directed edges.
41
6
1↔5
2↔6
3↔5
4↔6
3↔1
4↔2
Table 1: All 6 possible bidirectional edges, or 2-cycles. Note that 1 ↔ 5 is distinct from 5 ↔ 1, since the pivot
point from the left node must be less than that from the right node in our notation. Here, by bidirectional
edge between nodes x and y, we mean the existence of directed edges x → y and y → x.
edge
1→1
1→2
1→3
1→4
1→5
1→6
1 ⇒∗
num
2
1
1
0
1
2
7
edge
2→1
2→2
2→3
2→4
2→5
2→6
2 ⇒∗
num
1
0
0
0
1
1
3
edge
3→1
3→2
3→3
3→4
3→5
3→6
3 ⇒∗
num
1
1
1
0
0
1
4
edge
4→1
4→2
4→3
4→4
4→5
4→6
4 ⇒∗
num
2
1
1
0
1
2
7
edge
5→1
5→2
5→3
5→4
5→5
5→6
5 ⇒∗
num
1
0
0
0
1
1
3
edge
6→1
6→2
6→3
6→4
6→5
6→6
6 ⇒∗
num
0
0
0
0
0
0
0
Table 2: All 24 possible forward edges and their number. Here only shift moves of the form (x, y, z) →
(u, v, w) are considered, where the (unordered) base pair {y, z} ∈ s crosses the (unordered) base pair {u, v} ∈
t, where y < v.
edge
1←1
1←2
1←3
1←4
1←5
1←6
1 ⇐∗
num
0
1
0
0
1
1
3
edge
2←1
2←2
2←3
2←4
2←5
2←6
2 ⇐∗
num
1
2
0
1
2
1
7
edge
3←1
3←2
3←3
3←4
3←5
3←6
3 ⇐∗
num
1
2
0
1
2
1
7
edge
4←1
4←2
4←3
4←4
4←5
4←6
4 ⇐∗
num
1
1
0
1
1
0
4
edge
5←1
5←2
5←3
5←4
5←5
5←6
5 ⇐∗
num
0
0
0
0
0
0
0
edge
6←1
6←2
6←3
6←4
6←5
6←6
6 ⇐∗
num
0
1
0
0
1
1
3
Table 3: All 24 possible backward edges and their number. Here only shift moves of the form (x, y, z) ←
(u, v, w) are considered, where the (unordered) base pair {v, w} ∈ s crosses the (unordered) base pair
{x, y} ∈ t, where y < v.
42
B
Minimal length pk-M S2 folding pathways
This section provides details on the simple algorithms for pk-M S2 minimum length folding pathways for
each of the five types of paths depicted in Figure 3. If s and t are (possibly pseudoknotted) structures
on [1, n], and X ⊆ [1, n] is an equivalence class, then define the restriction of s [resp. t] to X, denoted
by s X [resp. t X], to be the set of base pairs (i, j) in s [resp. t] such that i, j ∈ X. Each path or
cycle in A ∪ B can be subdivided into the following five cases. Each equivalence class can be classified as
one of five types of paths, depicted in Figure 3 described below. For this classification, we need to define
End(s, X) = {x ∈ X : t[x] = 0} and End(t, X) = {x ∈ X : s[x] = 0} – i.e. End(s, X) [resp. End(t, X)]
is the set of elements x of X that belong to a base pair in s [resp. t], but the path cannot be extended
because x is not touched by a base pair from t [resp. s]. For each type of path X, we present a (trivial)
algorithm that returns the shortest M S2 folding trajectory from s X to t X. Additionally, we determine
the relation between the pseudoknotted M S2 distance between s X and t X, denoted dX
pk−M S2 (s, t), as
X
well as the Hamming distance, denoted dH (s, t).
An equivalence class X of size m is defined to be a path of type 1, if m is even, so path length is odd,
and |End(s, X)| = 2. Let b0 = min(End(s, X)) and for 1 ≤ i < m/2, define ai+1 = s[bi ] and bi = t[ai ], as
shown in Figure 3a. A minimum length sequence of M S2 moves to transform s X into t X is given by
the following:
Path 1 subroutine
1. remove {bm/2 , am/2 } from s
2. for (m/2) − 1 down to 1
3.
shift base pair (bi , ai ) to (ai , bi+1 )
An alternate procedure would be to remove the first base pair {b1 , a1 } and perform shifts from left to right.
Notice that if m = |X| = 2, then a path of type 1 is simply a base pair with the property that neither i
dX
H (s,t)
. The Hamming distance dX
H (s, t) =
2
X
Moreover, dpk−M S2 (s, t) = max(|s X|, |t X|).
nor j is touched by t. For arbitrary m, dX
pk−M S2 (s, t) =
dX
pk−M S2 (s, t)
dX
pk−M S2 (s, t)
dX (s,t)
b H 2 c.
m, and
= m/2, so
=
An equivalence class X of size m is defined to be a path of type 2, if m is odd, so path length is even,
and |End(s, X)| = 1 = |End(t, X)|, and min(End(s, X)) < min(End(t, X)). Let b0 = min(End(s, X)) and
for 1 ≤ i ≤ bm/2c, define ai+1 = s[bi ] and bi = t[ai ], as shown in Figure 3b. A minimum length sequence of
M S2 moves to transform s X into t X is given by the following:
Path 2 subroutine
1. for i = bm/2c down to 1
2.
shift base pair {bi−1 , ai } to {ai , bi }
dX (s,t)
X
X
H
The Hamming distance dX
c. Moreover,
H (s, t) = m, and dpk−M S2 (s, t) = bm/2c, so dpk−M S2 (s, t) = b
2
X
dpk−M S2 (s, t) = max(|s X|, |t X|).
An equivalence class X of size m is defined to be a path of type 3, if m is odd, so path length is even,
and |End(s, X)| = 1 = |End(t, X)|, and min(End(t, X)) < min(End(s, X)). Let a0 = min(End(t, X)) and
for 1 ≤ i ≤ bm/2c, define bi = t[ai−1 ] and ai = s[bi ], as shown in Figure 3c. A minimum length sequence of
M S2 moves to transform s X into t X is given by the following:
Path 3 subroutine
1. for i = 1 to bm/2c
2.
shift base pair {bi , ai } to {ai−1 , bi }
X
X
The Hamming distance dX
H (s, t) = m, and pk-M S2 distance dpk−M S2 (s, t) = bm/2c, so dpk−M S2 (s, t) =
dX
H (s,t)
c.
2
Moreover, dX
pk−M S2 (s, t) = max(|s X|, |t X|).
An equivalence class X of size m is defined to be a path of type 4, if m is even, so path length is odd, and
|End(t, X)| = 2. Let a1 = min(End(t, X)) and for 2 ≤ i < m/2, define ai+1 = s[bi ] and for 1 ≤ i ≤ m/2,
define bi = t[ai ], as shown in Figure 3d. A minimum length sequence of M S2 moves to transform s X into
t X is given by the following:
b
Path 4 subroutine
1. for i = 1 to m/2 − 1
43
2.
shift base pair {bi , ai+1 } to {ai , bi }
3. add base pair {am/2 , bm/2 }
Notice that if m = 2, then a path of type 4 is simply a base pair (i, j) ∈ t, with the property that neither i nor
dX (s,t)
X
X
H
j is touched by s. The Hamming distance dX
.
H (s, t) = m, and dpk−M S2 (s, t) = m/2, so dpk−M S2 (s, t) =
2
X
Moreover, dpk−M S2 (s, t) = max(|s X|, |t X|).
An equivalence class X of size m is defined to be a path of type 5, if it is a cycle, i.e. each element x ∈ X
is touched by both s and t. Since base triples are not allowed due to condition 2 of Definition 3, cycles have
only even length, and so |X| is also even. Let a1 = min(X), and for 1 ≤ i ≤ m/2, define bi = t[ai ], and
for 2 ≤ i ≤ m/2, define ai = s[bi−1 ], as shown in Figure 3e. A minimum length sequence of M S2 moves to
transform s X into t X is given by the following:
Path 5 subroutine
1. remove base pair {bm/2 , a1 }
2. for i = 1 to m/2 − 1
3.
shift base pair {bi , ai+1 } to {ai , bi }
4. add base pair {am/2 , bm/2 }
dX (s,t)
X
X
H
c + 1.
The Hamming distance dX
H (s, t) = m, and dpk−M S2 (s, t) = m/2 + 1, so dpk−M S2 (s, t) = b
2
X
Moreover, dpk−M S2 (s, t) = max(|s X|, |t X|). Note that any base pair could have initially been removed
from s, and by relabeling the remaining positions, the same algorithm would apply.
In summary, pk-M S2 distance between s X and t X for any maximal path (equivalence class) X is
equal to Hamming distance b dH (sX,tX)
c; in contrast, pk-M S2 distance between s X and t X for any
2
dH (sX,tX)
cycle X is equal to b
c + 1. It follows that dpk−M S2 (s, t) = b dH (s,t)
c if and only if there are no
2
2
type 5 paths, thus establishing equation (14.
Now let B1 [resp. B2 ] denote the set of positions of all type 1 paths [resp. type 4 paths] of length 1 –
i.e. positions incident to isolated green [resp. red] edges that correspond to base pairs (i, j) ∈ s where i, j
are not touched by t [resp. (i, j) ∈ t where i, j are not touched by s]. As well, let B0 designate the set of
positions in B not in either B1 or B2 . Note that B1 ⊆ B and B2 ⊆ B, and that formally
B0 = B − (B1 ∪ B2 )
(26)
B1 = {i ∈ [1, n] : ∃j [{i, j} ∈ s, t(i) = 0 = t(j)]
(27)
B2 = {i ∈ [1, n] : ∃j [{i, j} ∈ t, s(i) = 0 = s(j)]
(28)
Note that B1 and B2 have an even number of elements, and that all elements of B − B1 − B2 are incident
to a terminal edge of a path of length 2 or more. Correspondingly, define BP1 and BP2 as follows:
BP1 = {(i, j) ∈ s : t[i] = 0 = t[j]}
(29)
BP2 = {(i, j) ∈ t : s[i] = 0 = s[j]}
(30)
Note that |BP1 | = |B1 |/2 and |BP2 | = |B2 |/2. The following is a restatement of Lemma 8.
Lemma 16. Let s, t be two arbitrary pseudoknotted structures for the RNA sequence a1 , . . . , an , and let
X1 , . . . , Xm be the equivalence classes with respect to equivalence relation ≡ on A ∪ B0 = [1, n] − B1 − B2 −
C − D. Then the pk-M S2 distance between s and t is equal to
|BP1 | + |BP2 | +
m
X
max |s Xi |, |t Xi |
i=1
Alternatively, if X1 , . . . , Xm are the equivalence classes on A ∪ B = [1, n] − C − D, then
dpk−M S2 (s, t) =
m
X
max |s Xi |, |t Xi |
i=1
44
C
Branch-and-bound algorithm
In this section, we present pseudocode for an exact and exhaustive branch-and-bound search strategy [7]
to determine a shortest M S2 folding trajectory between two given secondary structures of a given RNA
sequence.
First base pairs in BP1 and BP3 are removed from s to obtain the root structure. Starting from the
root we perform removal and shift of base pairs until the target or an empty structure is achieved. For each
state Si , we have the distance dist from the root to Si . Furthermore, we compute a lower bound lb for
the M S2 distance from Si to the target structure by calculating the M S2 distance allowing pseudoknots.
The optimistic M S2 distance from the source to target is estimated as dist + lb. If the optimistic distance
is higher than the incumbent M S2 distance computed so far, the search is stopped and a different state
is considered. In order to prune the search tree more effectively, base pairs that cause more conflicts are
considered first. Finally, if an empty structure is achieved we compute the list of remaining moves for s.
In algorithm 1 we describe a branch and bound algorithm for computing the M S2 distance between two
RNA secondary stuctures s and t. First base pairs in BP1 and BP3 are removed from s to obtain the root
structure. Starting from the root we perform removal and shift of base pairs until the target or an empty
structure is achieved. For each state Si , we have the distance dist from the root to Si . Furthermore, we
compute a lower bound lb for the M S2 distance from Si to the target structure by calculating the M S2
distance allowing pseudoknots. The optimistic M S2 distance from the source to target is estimated as
dist + lb. If the optimistic distance is higher than the incumbent M S2 distance computed so far, the search
is stopped and a different state is considered. In order to prune the search tree more effectively, base pairs
that cause more conflicts are considered first. Finally, if an empty structure is achieved we compute the list
of remaining moves for s.
D
Greedy algorithm
For a digraph G = (V, E), in this section, we present the pseudocode for a staightforward greedy algorithm
to determine a (possibly non-maximal) vertex subset V ⊂ V such that the induced subgraph H = (V , E)
contains no directed cycles, where E = E ∩ (V × V ). Nevertheless, in the following greedy algorithm, it is
necessary to first generate a list of all (possibly exponentially many) directed cycles. This computational
overhead is sidestepped by the near-optimal algorithm in the next section.
Algorithm 17 (Greedy approximation of M S2 distance from s to t).
Input: Secondary structures s, t for RNA sequence a1 , . . . , an
Output: Greedy M S2 folding trajectory s = s0 , s1 , . . . , sm = t, where s0 , . . . , sm are secondary structures,
m is the minimum possible value for which si is obtained from si−1 by a single base pair addition, removal
or shift for each i = 1, . . . , m.
First, initialize the variable numMoves to 0, and the list moveSequence to the empty list [ ]. Define
BP1 = {(x, y) : (x, y) ∈ t, (t − s)[x] = 0, (t − s)[y] = 0}; i.e. BP1 consists of those base pairs in s which
are not touched by any base pair in t. Define BP2 = {(x, y) : (x, y) ∈ t, (s − t)[x] = 0, (s − t)[y] = 0}; i.e.
BP2 consists of those base pairs in t which are not touched by any base pair in s. Bear in mind that s is
constantly being updated, so actions performed on s depend on its current value.
//remove base pairs from s that are untouched by t
1. for (x, y) ∈ BP1
2.
remove (x, y) from s; numMoves = numMoves+1
//define conflict digraph G = (V, E) on updated s and unchanged t
3. define V by equation (16)
4. define E by equation (17)
5. define conflict digraph G = (V, E)
6. C = {C1 , . . . , Cm } //list of all simple directed cycles in G
//determine set V0 of vertices to remove so that restriction of G to V − V0 is acyclic
7. V0 = ∅ //V0 is set of vertices to be removed from V
8. for v ∈ V
9.
Cv = {C ∈ C : v ∈ C}
45
Algorithm 1 Branch and bound algorithm for M S2 distance
Input: Two RNA secondary structures s and t with length n
Output: Minimum number of M S2 moves in the path from s to t
1: BP1 is the set of base pairs (i, j) ∈ s such that neither i nor j is touched by any base pair of t
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
38:
39:
40:
41:
42:
43:
BP2 is the set of base pairs (i, j) ∈ t such that neither i nor j is touched by any base pair of s
BP3 is the set of base pairs (i, j) ∈ t, s
V = {(x, y, z) : 1 ≤ x, y, z ≤ n; t[b] = a; s[b] = c}
remove base pairs in BP1 ∪ BP3 from s
remove base pairs in BP2 ∪ BP3 from t
for v ∈ V compute nv the the number of crossings between v.s and all base pairs in t
sort V in decreasing order by nv
best = base pair distance between s and t
. in worst case only M S1 moves are used
curdist = 0
. distance between s and current state cs
define a data structure state = {s, t, dist, lb, rm, ad, sh} representing a node in the search tree.
s and t represent the structures in the current state; dist is the M S2 distance from the root;lb is the lower bound
for the M S2 distance from s to t for all paths passing through the current state cs; rm, ad and sh are respectively
the lists of removals, additions and shifts performed from the root to obtain the current state.
root = state(s, s, best, 0, [ ], [ ], [ ])
define priority queue Q containing states S1 , ..., Sm , ordered by Si .lb
Q = {root}
while (Q is not empty) do
cs = Q.pop()
. state with smallest lower bound cs.lb will be poped
if (cs.lb < best) then
if (s has no base pairs) then
. current state is a leaf
R = set of base pairs in t
cs.ad = cs.ad ∪ R
cs.dist = cs.dist + |R|
if (cs.dist < best) then
Sol = cs
best = cs.dist
else if (t has no base pairs) then
. current state is a leaf
R = set of base pairs in s
cs.rm = cs.rm ∪ R
cs.dist = cs.dist + |R|
if (cs.dist < best) then
Sol = cs
best = cs.dist
else
. current state is an internal node
let M be the list of possible shift moves and removals that can be applied to current state cs
for m ∈ M do
apply move m to s
l = optimal number of moves to go from s to t passing through cs + m and allowing pseudoknots
if (l < best) then
if (m is a shift) then
nt = t − m.t
. remove the resolved base pair from t
ns = s − m.s
. remove the resolved base pair from s
newState = state(ns, nt, cs.dist + 1, l, cs.rm, cs.ad, cs.sh + m)
else if (m is a removal) then
ns = s − m
. remove the resolved base pair from s
newState = state(ns, cs.t, cs.dist + 1, l, cs.rm + m, cs.ad, cs.sh)
Q = Q ∪ newState
f inalP ath = BP1 + Sol.rm + Sol.sh + Sol.add + BP2
46
10. while C =
6 ∅
11.
v0 = argmaxv |Cv | //v0 belongs to largest number of cycles
12.
V0 = V0 ∪ {v0 }
13.
V = V − {v0 }
14.
E = E − {(x, y) : x = v0 ∨ y = v0 }
15.
G = (V, E) //induced subgraph obtained by removing v0
16.
C = C − Cv0 // remove all cycles containing v0
17.
v0 = (x, y, z) //unpack v0 to obtain base pairs {x, y}< ∈ t, {y, z}< ∈ s
18.
s = s − {(min(y, z), max(y, z))}
//topological sort of the now acyclic digraph G = (V, E) for updated V, E
19. topological sort of G using DFS [7] to obtain total ordering ≺ on V
20. for v = (x, y, z) ∈ V in topologically sorted order ≺
//check if shift would create a base triple, as in type 1,5 paths from Figure 3 of main text
21.
if s[x] = 1 //i.e. {u, x} ∈ s for some u ∈ [1, n]
22.
remove {u, x} from s; numMoves = numMoves+1
23.
shift {y, z} to {x, y} in s; numMoves = numMoves+1
//remove any remaining base pairs from s that have not been shifted
24. for (x, y) ∈ s − t
25.
remove (x, y) from s; numMoves = numMoves+1
//add remaining base pairs from t−s, e.g. from BP2 and type 4,5 paths in Figure 3 of main text
26. for (x, y) ∈ t − s
27.
add (x, y) to s; numMoves = numMoves+1
28. return folding trajectory, numMoves
We now analyze the time and space complexity of the greedy algorithm. In line 6, Johnson’s algorithm
[21] is used to enumerate all simple directed cycles, resulting in run time O((|V | + |E|) · (|C| + 1)), where
|V | [resp. |E|] denotes the number of vertices [resp. edges] of the initial conflict digraph G, and |C| denotes
the number of directed cycles of G. Let M = |C| denote the number of directed cycles in C, and let
N = O(|V | · M ) denote the total number of vertices (counting duplicates) in the set of all simple directed
cycles C = {C1 , . . . , CM }. Lines 7 through 28 require O(N ) time and space, provided that one introduces
the data structures A1 , A2 , A3 , A4 , defined by as follows:
A1 [v] = |{C ∈ C : v ∈ C}|
A2 [v] = {k ∈ {1, . . . , |C|} : Ck ∈ C ∧ v ∈ Ck }
A3 [k] = {v ∈ V : v ∈ Ck }
1 if Ck ∈ C
A4 [k] =
0 else
In other words, A1 is a linked list of size |V |, where A1 [v] equals the (current) number of cycles to which v
belongs (in line 13, the node A1 [v] is deleted from the linked list); A2 is an array of size |V |, where A2 [v] is
a linked list of indices k of cycles Ck that contain vertex v (note that the size of linked list A2 [i] is A1 [i]);
A3 is an array of size the number |C| of cycles, where A3 [k] is a linked list of vertices v that belong to Ck ;
A4 is an array of size the number |C| of cycles, where A4 [k] is a a boolean value (true/false), depending on
whether the cycle Ck currently belongs to C (used to implement line 16). Details are left to the reader, or
can be gleaned from reading our publicly available source code. It follows that the run time complexity of
Algorithm 17 is O((|V | + |E|) · (|C| + 1)) with space complexity of O(|V | · (|C| + 1) + |E|).
E
E.1
Graph theoretical properties
Representable digraphs
Recall that digraph G = (V, E) is isomorphic to digraph G0 = (V 0 , E 0 ) if there is a bijective function (i.e.
one-one and onto) Φ : V → V 0 , such that for all u, v ∈ V , (u, v) ∈ E if and only if (Φ(u), Φ(v)) ∈ E 0 . Since
RNA conflict digraphs have a natural ordering of vertices defined in Definition 11, we now define digraph
order-isomorphism.
47
Definition 18 (Order-isomorphism). Let G = (V, E, ) [resp. G0 = (V 0 , E 0 , 0 )] be a digraph, whose vertex
set V [resp. V 0 ] is totally ordered by [resp. 0 ]. We say that G is order-isomorphic to G0 if there exists
an order-preserving bijective function Φ : V → V 0 (i.e. one-one and onto) such that (1) for u, v ∈ V , x y
if and only if Φ(u) 0 Φ(v), (2) for u, v ∈ V , (u, v) ∈ E) if and only if (Φ(u), Φ(v)) ∈ E 0 . If Φ is an injective
function (one-one, but not necessarily onto), then G is said to have an order-preserving embedding in G0 .
We say that a digraph G = (V, E) is representable if it is order-isomorphic to an RNA conflict digraph,
formally defined as follows.
Definition 19 (Representable digraph).
Let V = {1, . . . , n} be a set of vertices and E a set of directed edges on V . The digraph G = (V, E) is said
to be representable if there exist secondary structures s, t of some RNA sequence a1 , . . . , am , an integer N ,
and an order-preserving function Φ : [1, n] → [1, N ]3 such that (1) for v, v 0 ∈ [1, n], x < y if and only if
Φ(v) < Φ(v 0 ), (2) for each v ∈ [1, n], Φ(v) = (x, y, z) where x, y, z are distinct, {x, y}< ∈ t, {y, z}< ∈ s, (3)
there is an edge u → v in E if and only if Φ(u).s = {y, z}< ∈ s touches or crosses Φ(v).t = {x, y}< ∈ t.
As just defined, the notion of representability depends on the nucleotide sequence a1 , . . . , an . In a
mathematical investigation to determine which digraphs are representable, it is more natural to reinterpret
the notion of secondary structure to satisfy requirements 2-4 of Definition 3, but not necessarily requirement
1.
The requirement that mapping Φ be order-preserving is important. Consider the RNA conflict digraph
G in Figure 21, equivalent to the ordered digraph in Figure 25a, having edges 1 → 2 → 3 → 4 → 1.
Clearly G is isomorphic to the digraph G0 in Figure 25b, although there is no order-isomorphism between
G and G0 . Indeed, by writing a program to exhaustively enumerate all representable digraphs having
a vertex set of size 4, we know that G0 is not order-isomorphic to any RNA conflict digraph. It is a
3
straightforward exercise to show that each of the 2(2) = 8 many tounaments on 3 nodes is representable
4
(data not shown); however, not all 2(2) = 64 many tournaments on 4 nodes are representable, as shown
in Figure 25c. Although representability is not invariant under isomorphism, it clearly is invariant under
order-isomorphism. Moreover, we have the following.
Theorem 20. Suppose that Φ is an order-preserving embedding of digraph G = (V, E) into digraph G0 =
(V 0 , E 0 ). If G is not representable, then G0 is not realizable.
The theorem is immediate, since if G0 were order-isomorphic to an RNA conflict digraph, then the
induced subgraph Φ(G) of G0 must be representable, and hence G must be representable. Figure 26a depicts
a nonrepresentable digraph having 4 vertices and 4 edges. By adding an edge to that figure, we obtain the
digraph in Figure 26b, which is not representable.
Recall that an automorphism of a directed graph G = (V, E) is the set of permutations σ on n letters,
for V = {1, . . . , n}, such that G and σ(G) are isomorphic. Using a small program that we wrote to compute
the automorphism group Aut(G) for any connected, directed graph G = (V, E), we found that the digraphs
in Figures 25c and 26b both have the trivial automorphism group consisting only of the identity permutation on 4 letters. Since the former is not representable and the latter is representable, it follows that the
automorphism group of a digraph implies nothing about whether the digraph is representable.
E.1.1
Example of RNA conflict digraphs
Computation time for the IP algorithm 14 is dominated by the time to generate a list of all simple cycles and
the time to obtain an IP solution satisfying the FVS problem (†) as well the constraints (‡) that ensure that
shift moves cannot be applied if they share same base pair from s or t. This raises the question whether the
FVS problem is polynomial time solvable for RNA conflict digraphs. We cannot settle this open question,
but provide examples of RNA conflict digraphs that indicate the there is no reduction to known cases for
which the FVS problem has been resolved to be NP-complete or polynomial time computable. Figure 21
depicts a 4-cycle digraph that is order-isomorphic to the digraph with edges 1 → 2 → 3 → 4 → 1. Figure 22a
shows the complete bipartite graph K3,3 . Recall that in [24], Kuratowski proved that a graph is planar if
and only if it does not contain a subgraph that is a subdivision of K3,3 or the complete graph K5 ; i.e. does
not contain an embedded copy of one of the forbidden graphs K3,3 or K5 . Figure 22b depicts a copy of
48
K3,3 with directed edges, which is realized in the RNA conflict digraph shown in Figure 23. It follows from
Kuratowski’s theorem that RNA conflict digraphs are not planar in general. Figure 22c depicts a forbidden
digraph, with the property that a flow graph (graphical representation of code in a programming language
with possible goto-statements) is reducible if and only if the flow graph does not contain a copy of this
forbidden digraph (or related digraph where the edges of Figure 22c may be replaced by arc-disjoint paths)
[19]. Figure 24 shows an RNA conflict digraph which represents the forbidden digraph of Figure 22c. It
follows that RNA conflict digraphs are not reducible flow graphs in general. Two classes of digraphs for
which FVS is known to be polynomial time computable are: (1) planar digraphs [26], and (2) reducible
flow graphs [34]. On the other hand, FVS is NP-complete for general digraphs [22], for tournaments [4, 1],
3
and for Eulerian digraphs [30]. It is a simple exercise, left to the reader, to show that each of the 2(2) = 8
tournaments on 3 vertices can be represented by an RNA conflict digraph.
Proposition 21. All tournaments on a graph having 3 vertices can be represented as RNA conflict digraphs.
n
There are 2( 2 ) many tournaments on n vertices. By applying constraint programming to each of the
6
2 = 64 tournaments on 4 vertices, we found a number of tournaments that are not representable as conflict
digraphs. Figure 25a depicts a 4-cycle which is representable by the RNA conflict digraph represented in
Figure 21. Figure 25b depicts a simple, connected digraph on 4 nodes (a cycle) which is not representable
as a conflict digraph, and Figure 25c shows a tournament on 4 nodes which is not representable.
The FVS problem is known to be NP-complete for all tournaments, all Eulerian digraphs and for general
digraphs, and there are polynomial time algorithms for FVS for planar digraphs and for flow graphs. Examples we have provided suggest that there is no straightforward application of known results to settle the
question whether the FVS problem is NP-complete for the class of RNA conflict digraphs.
Question 22. Is the FVS problem polynomial time computable or NP-complete for the collection of RNA
conflict digraphs?
Using constraint programming, we determined which connected digraphs on 4 nodes could be represented
as RNA conflict digraphs. This was done by considering all partition of [1, 12] into four classes, each class
corresponding to a triplet node (x, y, z) and determining the resulting edge relations defined by whether a
base pair from s belonging to node v crosses a base pair from t belonging to node v 0 . Is there a better
approach?
Question 23 (Representation of arbitrary G = (V, E) as conflict digraph). Is there an efficient algorithm
to determine whether the labeled digraph G = (V, E) can be realized as an RNA conflict digraph, where
G = (V, E) is given as input.
The following is perhaps tractable.
Question 24. Is there an efficient algorithm to determine whether a given labeled digraph G = (V, E) can
be realized as an RNA conflict digraph, the vertices of G are totally ordered by v1 < · · · < vn , and each vertex
is labeled by one of the node types 1,. . . ,6.
Even if the preceding problem has a polynomial time solution, it is not clear whether the same is true
for the following slight generalizations.
Question 25. Is there an efficient algorithm to determine whether the labeled digraph G = (V, E) can be
realized as an RNA conflict digraph, where vertices of D are totally ordered by v1 < · · · < vn .
Question 26. Is the feedback arc set (FAS) problem complete for RNA conflict digraphs? Is the problem of
computing M S2 distance between arbitrary secondary structures s, t NP-complete?
Since a digraph G is planar if and only if it contains neither the complete graph K5 on 5 vertices nor the
complete bipartite graph K3,3 , so Theorem 20 suggests the following question.
Question 27. Is there a finite set of prohibited digraphs H1 , . . . , Hr such that a digraph G is representable
by an RNA conflict digraph if and only if there is no order-preserving embedding of Hi into G for any
i = 1, . . . , r.
49
| 8cs.DS
|
Myopic Policy Bounds for Information Acquisition POMDPs
0.4
0.6
0.2
0.4
0.6
p(s1 )
0.8
0
3)
p( s
0.6
0.4
0.4
0.4
0.6
p(s1 )
0.8
0.8
0.6
0.4
0.4
0.6
0.2
0.2
0.2
a3
0.8
0
0
a2
0.8
0.6
0.2
0.8
0
3)
0.6
0.4
p( s
3)
0.2
2)
p( s
0.8
a1
p( s
M. Lauri and R. Ritala are with Department of Automation Science and
Engineering, Tampere University of Technology, P.O. Box 692, FI-33101,
Tampere, Finland, {mikko.lauri, risto.ritala}@tut.fi.
N. Atanasov and G. Pappas are with GRASP Lab, University
of Pennsylvania, Philadelphia, PA 19104, USA, {atanasov,
pappasg}@seas.upenn.edu. This work was supported by
TerraSwarm, one of six centers of STARnet, a Semiconductor Research
Corporation program sponsored by MARCO and DARPA.
0.2
Optimal policy
0
2)
The current proliferation of sensors and robots has potential
to transform fields as diverse as environmental monitoring,
security and surveillance, localization and mapping, and structure inspection. One of the technical challenges in these scenarios is to control the sensors and robots to extract accurate
information about various physical phenomena autonomously.
Robotic information acquisition may be thought of as an
optimization problem: given constraints on robot motion and
the available sensing resources, find an optimal control policy
for the robot resulting in the greatest amout of collected information. Assuming Markovian system dynamics and observations conditionally independent given the system state, the
optimization problem is formalized as a partially observable
Markov decision process (POMDP) [1]. A POMDP consists
of the state, action and observation spaces, stochastic system
dynamics and observation models and a real-valued reward
function. As the true state of the robot and environment are
unknown, a probability density function (pdf) known as a
belief state is maintained to represent information regarding
the state. The optimal control policy is a mapping from
belief states to actions such that the expected total reward
accumulated over a given optimization horizon is maximized.
While POMDPs offer a principled framework handling
probabilistic uncertainty in both control actions and sensing
Upper bound policy
0
p( s
I. I NTRODUCTION
Lower bound policy
2)
Abstract—This paper addresses the problem of optimal control
of robotic sensing systems aimed at autonomous information
gathering in scenarios such as environmental monitoring, search
and rescue, and surveillance and reconnaissance. The information gathering problem is formulated as a partially observable
Markov decision process (POMDP) with a reward function that
captures uncertainty reduction. Unlike the classical POMDP
formulation, the resulting reward structure is nonlinear in the
belief state and the traditional approaches do not apply directly.
Instead of developing a new approximation algorithm, we show
that if attention is restricted to a class of problems with certain
structural properties, one can derive (often tight) upper and
lower bounds on the optimal policy via an efficient myopic
computation. These policy bounds can be applied in conjunction
with an online branch-and-bound algorithm to accelerate the
computation of the optimal policy. We obtain informative lower
and upper policy bounds with low computational effort in a target
tracking domain. The performance of branch-and-bounding is
demonstrated and compared with exact value iteration.
p( s
arXiv:1601.07279v1 [cs.SY] 27 Jan 2016
Mikko Lauri, Nikolay Atanasov, George J. Pappas, and Risto Ritala
0.2
0.8
0
0
0.2
0.4
0.6
p(s1 )
0.8
0
Fig. 1. The optimal policy µ∗ in a POMDP (shown on the right) can be
lower and upper bounded by myopic lower and upper bound policies µ∗1 and
µ∗1 (shown on the left and middle, respectively). The ternary plots cover the
belief simplex B in the three-state case with state space S = {s1 , s2 , s3 },
while the dots indicate the action selected by the policy for a particular belief
point. The inequalities µ∗1 (b) ≤ µ∗ (b) ≤ µ∗1 hold for all b ∈ B. The optimal
policy is fully determined if µ∗1 (b) = µ∗1 (b). Even if the bounds do not agree
they may provide information regarding the optimal action.
outcomes, obtaining optimal policies for them is computationally hard [2]. A wide range of algorithms for computing approximately optimal policies for POMDPs have been
proposed in the literature, see e.g. [3], [4], [5] for reviews.
Most research, however, focuses on a standard formulation in
which the reward is a function of the hidden system state, and
hence necessarily linear in the belief state. When it comes to
information gathering applications in robotics (see e.g. [6], [7],
[8], [9]), reward functions that capture uncertainty reduction
(e.g. negative Shannon entropy [10]) or information gain (e.g.,
mutual information [10]) are often more interesting. These
reward function are nonlinear in the belief state, and hence,
many of the existing approximation techniques for the standard
POMDP formulation do not apply. Approximation methods
applicable for such information acquisition POMDPs include
some online algorithms [11] and specialized value iteration
techniques, see e.g. [12]. Sampling-based online planning
algorithms, such as those based on variants of Monte Carlo
tree search (see e.g. [13], [14]), are also applicable.
Besides approximation algorithms, an alternative route to
handling the inherent complexity of POMDPs with nonlinear
belief-dependent rewards is to make structural statements
about the optimal value function. In the case of state-dependent
rewards, earlier work [15], [16], [17] shows that, if we restrict
attention to a class of POMDPs that satisfy certain assumptions
on the state and observation processes, we can bound the
optimal policy from both below and above by an easily
computable myopic (greedy) policy (see Fig. 1 for details).
The specific case of POMDP multi-armed bandits was studied
in [18].
The goal of this work is to extend these structural results to
POMDPs with belief-dependent rewards. Krishnamurthy and
Djonin [19] determined conditions under which the optimal
policy has a threshold structure with respect to the monotone
likelihood ratio (MLR) order, a partial order on the belief
states. Such partial orders on the belief states have been used
in the related active hypothesis testing field [20] to determine
when a given sensing action dominates all other actions. For
example, Naghshvar and Javidi [21] used Blackwell ordering
of the observation models to reduce an active hypothesis test
with K actions to a passive test with a single action. Compared to active hypothesis testing, information acquisition in
POMDPs is even more challenging because the true underlying
state evolves during the sensing process.
Our first contribution is to extend the structural results of
[17] for POMDPs with state-dependent rewards to POMDPs
with nonlinear belief-dependent rewards. This allows us to
generate upper and lower bounds on the optimal policy (which
agree in large portions of the belief state space and hence
determine the optimal action) via a very efficient myopic
policy computation. Our second contribution is to apply the
myopic policy bounds in conjunction with an online branchand-bound pruning algorithm to accelerate computation of the
optimal policy. Our approach allows anytime computation with
a bounded error from the optimal solution, unlike for instance
sampling-based planning which is only asymptotically optimal.
The remainder of the paper is organized as follows. Section II formulates the information acquisition problem as a
POMDP, and discusses suitable choices of reward functions.
Section III briefly reviews existing structural results. Section IV introduces our extension to belief-dependent rewards
and defines the myopic lower and upper bound policies.
In Section V, we show how to apply the myopic policy
bounds in a branch-and-bound pruning algorithm. Section VI
presents empirical evaluation in target tracking domains, and
Section VII concludes the paper.
the set of all possible pdfs over the system state. Finally, ρ :
B × A → R is the expected reward, which is a (nonlinear)
function of the belief state b.
Our definition subsumes the traditional POMDP definition
with a state-and-action-dependent reward
P function r : S×A →
R. This is seen defining ρ(b, a) := s∈S r(s, a)b(s).
The evolution of the belief state b in a POMDP is conditional on the actions and observations. When action a ∈ A
is executed, the belief state evolves according to the state
transition model T. The evolution can be tracked by Bayesian
filtering, which consists of iterating prediction and update
steps. The prediction step revises the current belief state b ∈ B
to the predicted belief ba ∈ B according to
X
ba (s0 ) =
T(s0 , s, a)b(s).
(1)
II. I NFORMATION ACQUISITION IN POMDP S
Definition 1 (Uncertainty function and information gain [22]).
An uncertainty function is a non-negative, concave function
f : B → R+ . The information gain If (b, a) of an action
a ∈ A in belief state b ∈ B is the expected reduction in the
uncertainty function f :
To model information acquisition problems in robotic applications we define a POMDP with a belief-dependent reward
function. In detail, a POMDP is a tuple hT , S, A, Z, T, O, ρi,
where T is a set of decision epochs and S = {1, 2, . . . , S},
A = {1, 2, . . . , A}, and Z = {1, 2, . . . , Z} are the finite
state, action, and observation spaces, respectively. The function
T : S × S × A → [0, 1] is a stochastic state transition
model such that T(s0 , s, a) is the probability of reaching state
s0 ∈ S from state s ∈ S after executing action a ∈ A and
O : Z × S × A → [0, 1] is a stochastic observation model such
that O(z 0 , s0 , a) is the conditional probability of observing
z 0 ∈ Z when state s0 ∈ S was reached after executing action
a ∈ A. To simplify notation, let Ta := [T(j, i, a)]Si,j=1 be
the matrix of state transition probabilities for a given action
a ∈ A, i.e. previous states are column-wise and next states
row-wise. Similarly, let Oa := [O(z, j, a)] for z ∈ Z, j ∈ S,
and a ∈ A.
As information about the system state is incomplete, it is
modeled by a probability density function (pdf) b(s), s ∈ S
over the system state. The set B, called the belief space, is
s∈S
After the state transition, an observation z 0 ∈ Z is perceived
and the information it provides is incorporated via the update
step:
b0 (s0 ) = τ (b, a, z 0 ) :=
1
· O(z 0 , s0 , a)ba (s0 ),
η(z 0 | b, a)
(2)
P
where η(z 0 | b, a) := s∈S O(z 0 , s, a)ba (s) is a normalization
factor equal to the conditional probability of observing z 0 .
A. Reward functions for information acquisition
The goal in information acquisition is to reduce the uncertainty in the unobservable state of interest s ∈ S. It is
therefore desirable to reach “peaked” belief states that have
the majority of their probability mass on a single underlying
state. In other words, actions that lead the belief state towards
the vertices of the probability simplex B should be rewarded.
Suitable measures of uncertainty are concave functions of the
belief state, called uncertainty functions [22, Sec. 14.16].
If (b, a) := f (ba ) − EZ [f (τ (b, a, Z))].
(3)
Examples of uncertainty functions include Shannon entropy,
the more-general Rényi quadratic entropy, and variants of the
value of information, such as the probability of error in hypothesis testing [10]. For example, the information gain associated
with Shannon entropy is known as mutual information.
For information acquisition tasks, either information gain
or a negative uncertainty function may be used as the reward
function in the POMDP formulation. However, as the following example shows, using an uncertainty function might
be more appropriate when the magnitude of the predicted
uncertainty f (ba ) is significantly affected by the action choice.
Example 1. Consider an active localization problem [23] in
which a mobile robot needs to choose an appropriate action
a to reduce the uncertainty in the distribution b of its current
position s. Suppose that the entropy f (b) of its current position
distribution is n bits. The robot may choose between a risky
high-velocity motion ar leading to a predicted entropy of
f (bar ) = n + m bits or a safe low-velocity motion as leading
to f (bas ) = n + h bits, where h < m. Suppose that in both
cases, after moving, the robot receives the same amount of
information, k bits, from its sensor measurement z 0 . In this
example, the negative uncertainty function correctly predicts
that action ar is risky:
−f (τ (b, ar , z 0 )) = −(n + m − k)
< −(n + h − k) = −f (τ (b, as , z))
but, perhaps surprisingly, the information gain prefers the
risky action:
If (b, ar ) = m − k > h − k = If (b, as ).
In the rest of the paper, we restrict attention to beliefdependent reward functions, specific to the task of information
gathering, which have the following form:
X
ρ(b, a) :=
r(s, a)b(s) − wa f (b),
(4)
s∈S
where f is an uncertainty function (preferred to information
gain due to the observations in Example 1), r(s, a) is any
standard state-dependent reward function, and wa are userspecified weights that trade-off measurement rewards and state
uncertainty.
B. Value functions and optimal policies
the value iteration must be solved is uncountable, the basic
scheme above rarely results in practical methods for computing
optimal policies.
III. M YOPIC POLICY BOUNDS FOR POMDP S
In this section, we review existing results on myopic policy
bounds for POMDPs with state-depend rewards. The idea of
a myopic policy bound is that the optimal policy µ∗k for any
k can be either lower or upper bounded by the myopic policy
µ∗1 (Eq. 8). The results exploit concepts of stochastic partial
orders, summarized in the following definition.
Definition 2 (Stochastic partial orders [25]). Let b1 , b2 ∈ B.
1) b1 first-order stochastically domainates b2 , denoted b1 ≥s
S
S
P
P
b2 , if
b1 (i) ≥
b2 (i) ∀j ∈ S.
i=j
i=j
2) Equivalently, b1 ≥s b2 iff
S
P
i=1
b1 (i)g(i) ≥
S
P
b2 (i)g(i)
i=1
for all increasing functions g : S → R.
3) b1 is greater than b2 in the monotone likelihood ratio (MLR) order, denoted b1 ≥r b2 , if b1 (i)b2 (j) ≤
b2 (i)b1 (j), for all i < j, j ∈ S.
4) A function g : B → R is MLR increasing (decreasing) if
b1 ≥r b2 ⇒ g(b1 ) ≥ g(b2 ) (g(b1 ) ≤ g(b2 )).
5) MLR dominance implies first-order stochastic dominance,
b1 ≥r b2 ⇒ b1 ≥s b2 .
Given conditions on the state transition and observation
models, it was shown in [15], [16] that the optimal value
function is MLR increasing and that the optimal policy may
be lower bounded by µ∗1 . The required conditions, however,
were thought to be too restrictive to be of practical significance
until a recent extension [17] proposed the following revised
assumptions.
Let H ∈ N ∪ {∞} denote a planning horizon so that
T := {1, 2, . . . , H}, and let 0 ≤ γ ≤ 11 denote a discount
factor determining the relative value of immediate and future
rewards. The goal in information acquisition is to choose a
policy µ∗k : B → A for each k ∈ T such that the expected Assumption 1 (Sufficient conditions for existence of myopic
sum of rewards over the decisions epochs is maximized. The policy bounds [17]).
sequence of optimal policies µ∗k for k ∈ T remaining decisions (A1) Ta and Oa are totally positive or order 2 (TP2) ∀a ∈ A,
(A2) For all j ∈ {1, 2, . . . , S − 1}, a ∈ {1, 2, . . . , A − 1}, and
can be computed via value iteration [24] according to:
j,a,z
S×S
z ∈ Z, the matrices Dj,a,z := [dj,a,z
m,n + dn,m ] ∈ R
a+1
a+1
Q∗1 (b, a) = ρ(b, a)
(5)
j,a,z
a
a
are copositive, where dm,n := Oz,j Oz,j+1 Tm,j Tn,j+1 −
a+1
a
Vk∗ (b) = max Q∗k (b, a),
(6)
Oaz,j+1 Oa+1
z,j Tm,j+1 Tn,j , and
a∈A
P
P
a+1
X
(A3)
Tai,j Oaz,j − Ta+1
≥ 0, ∀i ∈ S, ∀z̄ ∈ Z,
∗
i,j Oz,j
Q∗k (b, a) = ρ(b, a) + γ
η(z 0 | b, a)Vk−1
(τ (b, a, z 0 )), (7)
z≤z̄ j∈S
z 0 ∈Z
∀a ∈ {1, 2, . . . , A − 1}.
µ∗k (b) = argmax Q∗k (b, a),
(8)
The TP2 property for a matrix requires that all its seconda∈A
order minors are nonnegative [26]. Examples of TP2 obser∗
where Vk : B → R is the optimal value function. For infinitevation models include exponential family distributions such
horizon problems with H → ∞ and bounded rewards, value
as Gaussian, exponential, gamma, and binomial pdfs [26];
∗
iteration converges to a unique fixed point V that satisfies
more examples may be found e.g. in [27]. While (A2) may
the Bellman equation:
be checked directly [28], it is much more efficient (but also
X
V ∗ (b) = ρ(b, a) + γ
η(z 0 | b, a)V ∗ (τ (b, a, z 0 )), (9) more restrictive) to check the following:
j,a,z
(A2’) dj,a,z
z 0 ∈Z
m,n + dn,m ≥ 0 for all m, n ∈ S, j ∈ {1, . . . , S − 1},
a ∈ {1, 2, . . . , A − 1}, and z ∈ Z.
and the corresponding optimal infinite-horizon policy µ∗ is
Assumption
(A3) is a revised version of the original assumpstationary [24]. As the set B of belief states over which
tion in [15], [16]. It states that for every state i ∈ S, the prior
1 Note that γ = 1 is only valid for finite H.
pdf of perceiving z ∈ Z after ending in state j ∈ S after
executing action a + 1 first order stochastically dominates that
of action a. The assumptions lead to the following lemma.
Lemma 1 ([17]). Consider a POMDP hT , S, A, Z, T, O, ρi.
Then for all b ∈ B, and a0 , a ∈ A, a0 ≥ a,
(L1) if (A3) holds, η(z 0 | b, a0 ) ≥s η(z 0 | b, a), and
(L2) if (A2) holds, τ (b, a0 , z 0 ) ≥r τ (b, a, z 0 ) for all z 0 ∈ Z.
The following theorem for POMDPs with rewards linear in
the belief state is due to [15, Prop. 2] and [17, Theorem 1]:
Theorem 1. Consider a POMDP hT , S, A, Z, T, O, ρi
satisfying
Assumptions (A1)-(A3), with ρ(b, a)
:=
P
r(s,
a)b(s).
If
r(s,
a)
is
increasing
(decreasing)
s∈S
in S for all a ∈ A, then the optimal value function Vk∗ is
MLR increasing (decreasing) and the optimal policy µ∗k (b) for
any k ∈ T satisfies µ∗1 (b) ≤ µ∗k (b) (µ∗1 (b) ≥ µ∗k (b)) ∀b ∈ B.
IV. T RANSFORMED MLR MONOTONE REWARDS FOR
INFORMATION ACQUISITION POMDP S
In this section, we extend Theorem 1 to apply to reward
functions of the form in (4), which are of interest for information acquisition tasks. To proceed, we need the following
minor extension of Theorem 1.
Corollary 1. Consider a POMDP hT , S, A, Z, T, O, ρi satisfying Assumptions (A1)-(A3) and suppose that ρ is MLR
increasing (decreasing) in B for all a ∈ A. Then, the
optimal policy µ∗k (b) for any k ∈ T satisfies µ∗1 (b) ≤ µ∗k (b)
(µ∗1 (b) ≥ µ∗k (b)) ∀b ∈ B.
Proof. By [15, Prop. 2], it suffices to show Vk∗ is MLR
increasing. We proceed by induction. For the base case k = 1,
V1∗ (b) = max ρ(b, a) is clearly MLR increasing. For the
a∈A
∗
induction step, suppose the claim holds for Vk−1
. Consider
the sum part of the value iteration in Eq. (7). Let b1 , b2 ∈ B
s.t. b1 ≥r b2 . Now
Z
X
≥
z 0 =1
Z
X
∗
η(z 0 | b1 , a)Vk−1
(τ (b1 , a, z 0 ))
∗
η(z 0 | b2 , a)Vk−1
(τ (b1 , a, z 0 ))
(10)
z 0 =1
≥
Z
X
∗
η(z 0 | b2 , a)Vk−1
(τ (b2 , a, z 0 )).
z 0 =1
The first inequality follows by Lemma (L1), [15,
Lemma 1.2(2),1.3(1)] which state τ (b, a, z 0 ) is MLR
increasing in z 0 , and the the induction hypothesis. The
second inequality follows by Lemma (L2) and the induction
hypothesis. The proof for the MLR decreasing part is similar
and omitted.
Following the observation in Corollary 1, the main idea
is to transform the reward function ρ(b, a) (4) to one that
is MLR monotone but leaves the optimal policy unchanged.
Accordingly, we define two transformed reward functions:
ρ(b, a) = ρ(b, a) + [(I − γ(Ta )T )g]T b
ρ(b, a) = ρ(b, a) + [(I − γ(Ta )T )h]T b,
(11)
where g, h ∈ RS are parameter vectors chosen so that ρ(b, a)
is MLR increasing and ρ(b, a) is MLR decreasing in B.
Note that, by construction the transformed rewards in (11)
leave the optimal policy unaffected. This can be verified
by plugging either one into the Bellman equation (9). The
corresponding infinite horizon value functions are V ∗ (b) =
∗
V ∗ (b) + g T b, V (b) = V ∗ (b) + hT b, and µ∗ ≡ µ∗ ≡ µ∗ [17].
The following theorem gives necessary conditions for the
existence of vectors g and h at a single belief state.
Theorem 2. Let b0 ∈ B. If there exists a g ∈ RS such that
the set of linear constraints given by
∂ρ(b, a)
∂b
K + [(I − γ(Ta )T )g]T K ≥ 0
∀a ∈ A (12)
b=b0
where K is an S-by-S matrix with entries
(
1
if i = j
,
kij =
−1 if j = i + 1
(13)
are satisfied, then ρ(b, a) as defined in Eq. (11) is MLR
increasing at b0 .
Proof. As in the supplementary material of [17], let ∆ = {δ ∈
RS : 1 = δ(1) ≥ δ(2) ≥ . . . ≥ δ(S)}. Any belief state b ∈ B
may be represented as δ = Kb ∈ ∆. Let b1 = Kδ1 , b2 =
Kδ2 such that b1 ≥r b2 . By Definition 2, b1 ≥s b2 , and the
equivalent partial order on ∆ ⊂ RS is the component-wise
partial order between δ1 and δ2 . Then, the function ρ(Kδ, a)
is increasing on ∆ if its partial derivatives are non-negative:
∂ρ(Kδ, a)
∂ρ(Kδ, a) ∂[(I − γ(Ta )T )g]T Kδ
=
+
∂δ
∂δ
∂δ
(14)
∂ρ(b, a)
a T
=
K + [(I − γ(T ) )g]T K ≥ 0
∂b
and the claim follows by evaluating the derivative at b0 .
The proof for ρ(b, a) being MLR decreasing is obtained
similarly, by swapping the required sign for the partial derivatives. Practically, the condition of the theorem may be checked
by solving a feasibility linear program (LP) with (S − 1) · A
constraints. If a single g and h may be chosen such that the
transformed rewards are MLR monotone for all b ∈ B, then
the myopic policies
µ∗1 (b) = argmax ρ(b, a)
a∈A
µ∗1 (b) = argmax ρ(b, a)
(15)
a∈A
satisfy µ∗1 (b) ≤ µ∗ (b) ≤ µ∗1 (b) for all b ∈ B. This follows from
Corollary 1, and the fact that the optimal stationary policy is
unaffected by applying the transformed rewards. It is imporant
to note that the preceding inequalities do not hold anymore for
finite horizon optimal policies.
Consider the following two examples of nonlinear beliefdependent reward functions obtained by applying different
uncertainty functions in Eq. (4). In both examples, we denote
by ra the column vector formed from the elements of r(s, a)
for all s ∈ S.
Original reward
Example 2 (Shannon Entropy). Let the uncertainty function
in (4) be the Shannon entropy:
f (b) := −
S
X
Transformed reward
ρ(b, a1 )
ρ(b, a2 )
b(i) log b(i).
i=1
The derivative of ρ(b, a) for the constraints in Thm. 2 is:
∂ρ(b, a)
= raT + wa 1 + log(b(1)), . . . 1 + log(b(S)) .
∂b
As log(0) is not well defined, we consider the inner belief
simplex B := {b ∈ B | ∀i : b(i) ≥ } for ∈ (0, 1). Note that
log() ≤ log(b(i + 1)) − log(b(i)) ≤ − log().
For i = 1, . . . , S −1 and ∀a ∈ A, the ith constraint in Eq. (12)
can then be bounded by a belief-independent quantity:
φai − φai+1 ≥ wa (log(b(i + 1) − log(b(i))) ≥ −wa log(),
where φa := ra + (I − γ(Ta )T )g.
Example 3 (Rényi Quadratic Entropy). Let the uncertainty
function in (4) be the Rényi quadratic entropy:
f (b) := − log
S
X
b2 (i).
i=1
The derivative of ρ(b, a) for the constraints in Thm. 2 is:
2wa
∂ρ(b, a)
= raT + PS
bT .
2 (i)
∂b
b
i=1
S
and
0.2 0.4 0.6 0.8 1 0
p(s = 1)
0.2 0.4 0.6 0.8 1
p(s = 1)
Fig. 2. Expected original (left axes) and transformed (right axes) rewards as
function of the belief state according to Example 3. Rewards for action a1 are
shown by a solid line and for action a2 by a dashed line. The increasing blue
curves in the right axes depict the MLR increasing rewards ρ(b, a), and the
decreasing red curves depict the MLR decreasing rewards ρ(b, a). The two
regions on the right axes indicated by lines starting and ending with circle
markers are the parts of the belief space where the optimal policy is fully
determined by the myopic policy bounds. The dashed region corresponds to
µ∗1 (b) = µ∗1 (b) = a2 and the solid region to µ∗1 (b) = µ∗1 (b) = a1 .
is equivalent to the familiar notion of monotonicity on R. On
the bottom of the right axes, two regions in B are indicated
by a dashed and a solid line. These are regions where the
myopic policy bounds agree and the optimal policy is thus
fully determined.
V. M YOPIC BOUNDS FOR BRANCH - AND - BOUND PRUNING
Note that:
−1 ≤ b(i + 1) − b(i) ≤ 1
0
X
1
b2 (i) ≤ 1,
≤
S
i=1
which can be used to obtain belief-independent bounds in (12).
In particular, the ith constraint in (12) may be bounded by a
belief-independent quantity:
2wa
φai − φai+1 ≥ PS
(b(i + 1) − b(i)) ≥ −2wa ,
2
j=1 b (j)
where φa := ra + (I − γ(Ta )T )g and i = 1, . . . , S − 1.
In both examples, belief independent bounds for finding
h are obtained with straightforward changes. For Shannon
entropy, one ideally wants to choose as small as possible to
cover the greatest possible subset of the belief space. On the
other hand, the smaller is, the more restrictive the constraints
become, making it more unlikely to find feasible solutions. The
issue is resolved by considering a different quantification of
uncertainty, such as the Rényi quadratic entropy.
Figure 2 shows an example of the transformation of Rényi
quadratic entropy rewards and the associated monotonicity
properties as discussed in Example 3. The left axes show
the original expected reward as function of the belief state,
and the right axes show the transformed MLR increasing and
decreasing rewards, respectively. For S = 2, the belief space
B can be represented by [0, 1] ⊂ R and MLR monotonicity
The significance of Theorems 1 and 2 is that they give a
prescription for constructing policy bounds µ∗1 (b) ≤ µ∗ (b) ≤
µ∗1 (b) for all b ∈ B. In this section, we show that the
bounds can be used in an online branch-and-bound scheme to
accelerate the computation of the optimal policy for POMDPs
that satisfy Assumption 1.
Branch-and-bound pruning for POMDPs is based on a tree
search over the set of belief states reachable by finite-length
action and observation sequences, see e.g. [11]. The search is
started from the current belief state b0 at the root of the search
tree. The tree branches due to action choices and observation
possibilities, so that each node corresponds to a belief state
b reachable from b0 via the prediction and update iteration
in (1), (2). The standard branch-and-bound algorithm works
with lower and upper bounds on the optimal action values,
i.e., Qk (b, a) ≤ Q∗k (b, a) ≤ Qk (b, a). Starting from node b,
the subtree corresponding to action a may be pruned if it is
known to be suboptimal (i.e, the upper bound Qk (b, a) is lower
than the lower bound Qk (b, â) for some other action â), thus
saving computational resources.
Following the intuition of the branch-and-bound scheme
above, we design a new algorithm (Alg. 1), which employs
our policy bounds, instead of bounds on the optimal action
values. Given a belief state b and a search depth d > 0, the
algorithm constructs a search tree over reachable belief states,
storing the estimate of the value Q̂(b, a) for each visited belief
state. Only actions within the myopic lower and upper bound
policies are considered.
Algorithm 1 Branch-and-bound with myopic policy bounds.
Input: Belief state b, search depth d
Output: Estimate of the value of the best action at b
1: function S EARCH (b, d)
2:
if d = 1 then return max ρ(b, a)
a
3:
end if
do
4:
for all a ∈ A ∩ [µ∗1 (b), µ∗1 (b)]
P
5:
Q̂(b, a) ← ρ(b, a) + γ
η(z 0 | b, a)
z 0 ∈Z
6:
×S EARCH(τ (b, a, z 0 ), d − 1)
end for
return max Q̂(b, a)
a
9: end function
7:
8:
for each action a ∈ A
q
(1 − q)/2
Oaz,j =
(1 − q)
0
has elements
if z = j
if j 6= 1, i 6= S and |z − j| = 1
, (16)
if j = 1, i = S and |z − j| = 1
otherwise
i.e. with probability q < 1 the observation identifies the true
state, with symmetric error probability in either direction.
This model corresponds to a situation where e.g. entering a
dangerous state is perceived e.g. by the robot being damaged
by hazardous environmental conditions. This class of target
tracking domains can be checked to satisfy Assumptions 1.
A. Minimizing the uncertainty of the target state
The next proposition follows from [11, Theorem 3.1].
Proposition 1. After Alg. 1 terminates, its action recommendation â is guaranteed to coincide with the optimal action
a∗ := µ∗ (b) if the myopic lower and upper bounds agree at
b, i.e., if µ∗1 (b) = µ∗1 (b). Otherwise, the suboptimality of â is
of order γ d , i.e., |Q∗ (b, â) − Q∗ (b, a∗ )| ∼ o(γ d ).
VI. A PPLICATION TO TARGET T RACKING
We evaluate the performance of Algorithm 1 in a target
tracking application. Let s ∈ S model the state of a physical
process of interest such as the position of an adversarial
aircraft or certain environment conditions (e.g., temperature,
chemical concentration, air pollution, etc.). Let s = 1 correspond to an innocuous target state (e.g., the aircraft poses no
danger or the environment levels are normal) and, at the other
extreme, let s = S correspond to a red-alert state in which
there is imminent danger. We aim to design a control policy
for an autonomous robotic platform to track the target state s.
Let the available robot actions A = {1, 2, . . . , A} model the
priority invested into tracking the target, with 1 corresponding
to greatest priority (e.g, all available resources being dedicated
to tracking the target) and A corresponding to lowest effort,
i.e. the target is not being tracked.
The state transition model is such that the lower the priority
invested into tracking, the more likely it is to transition to
a state that is considered dangerous. Conversely, the higher
the priority given to tracking, the less likely it is to accidentally enter a dangerous state. In other words, supposing
the current state is i, the likelihood of transitioning to state
j > i should increase as function of a. State transition
matrices Ta modeling this can be defined e.g. as follows.
Let {xi }Si=1 and {yj }SA
j=1 be two increasing sequences of real
numbers, and define a matrix Aij = exp(xi yj ), normalizing
it such that each column sums to one, and define Ta as the
submatrix of A containing all rows and columns j for which
1 + (a − 1) · S ≤ j ≤ a · S. Each Ta also satisfies the required
TP2 condition [26].
The robot has access to Z = S observations providing
information about the target state. The observation matrix Oa
We first consider target tracking with a penalty for uncertain
tracker state information. We set S = A = Z = 3 and defined
the sequences {xi } = {1, 2, 3}, {yj } = {−9, −8, . . . , −1}.
The observation model was as defined in Eq. (16). The reward
was as in Eq. (4), with
2 2.5 1
r(s, a) = 1.1 1.2 0.5 ,
(17)
0.3 2 0.2
wa = 1.1 1.6 1 , and f equal to Rényi quadratic entropy.
The discount factor was γ = 0.99, and the observation
accuracy parameter was q = 0.7.
We computed the lower and upper myopic policy bounds for
the problem. The bounds are visualized on the belief simplex
B in Figure 1. The figure shows on the left and middle the
lower and upper myopic policies, and on the right the optimal
policy determined when the two bounds agree.
We defined a larger problem with S = Z = 8 and A = 3,
and in it compared Algorithm 1 against an exhaustive tree
search. There are Nd = (AZ)d reachable belief states after d
decisions,
Pdand the number of belief states in a complete search
tree is i=0 Ni .
TABLE I
AVERAGE PERCENTAGE OF BELIEF STATES PRUNED AS FUNCTION OF THE
SEARCH DEPTH d APPLYING THE MYOPIC POLICY BOUNDS .
Belief states pruned
d=2
d=3
d=4
d=5
d=6
26.3%
36.3%
44.9%
52.3%
58.8%
Table I shows the average percentage of belief state pruned
from the search tree for 100 randomly sampled initial belief
states. Computing the myopic policy bound requires solving
a feasibility LP with (S − 1) · A constraints offline, and
then during the search computing for each belief encountered
the action maximizing the immediate expected transformed
reward. Empirically, we found that for d ≥ 3 branch-andbounding was faster than the exhaustive search measured by
computation time.
B. Comparison with existing approaches
1
s
1
S−s+1
if s ∈ Sp
if s ∈ Sd
otherwise
(18)
where Sp and Sd are the regions of the state space corresponding to poor and dangerous tracking, respectively, and cp and
cd are the respective penalty costs. If the state is not in Sl or
Sh , a tracking reward proportional to kr is received.
1
In our experiments, we set ca = 2A
, cp = 1, cd = 0.1, and
2
kr = S . Furthermore, we set Sl = {s ∈ S | s ≤ S/5} and
Sh = {s ∈ S | s ≥ 9S/10}, and γ = 0.99. The observation
accuracy parameter was q = 0.8.
Table II shows the minimum, average and maximum percentage of actions that could be pruned by applying the myopic
policy bounds as function of the domain size. A value of
0% would indicate that the bounds were completely noninformative, and a value of 100% that the optimal policy is
fully determined by the bounds. The results were obtained by
evaluating the bounds with 500 randomly sampled reachable
belief states in each problem. We note that pruning efficiency
tends to decrease both as function of the state space and action
space size, but in all cases at least one third pruning rate was
achieved even in the worst case. For S = 256 and A = 64, a
vector satisfying the constraints of Thm. 2 could not be found.
As we computed the optimal solution, we could examine
how close the bounds are to the optimal policy. Figure 3 shows
an example of the looseness of the upper bound measured
by the average of µ∗1 (b) − µ∗ (b) over the 500 belief states.
The greater the difference is, the further away the myopic
upper bound policy is from the optimal policy, and the less
informative it is. We note that even as the number of states
and actions increases, the upper bound remains at a distance
of less than 10 actions from the optimal.
2 We applied the implementation from the pomdp-solve package of A.
Cassandra, see http://pomdp.org
µ∗1 (b) − µ∗ (b)
Next, we compare the performance of our approach to an
optimal incremental pruning algorithm2 in a target tracking
scenario with a state space size S between 4 and 256, and
action space size A between 4 and 64. Since the incremental
pruning algorithm (as well as other existing approaches)
can handle only state-dependent rewards, we do not add an
uncertainty function to the reward, i.e., f ≡ 0 in (4). The
state-dependent reward r(s, a) is designed to 1) incur a lower
reward for investing higher priority in target tracking and to 2)
penalize for both tracking the target poorly or too dangerously.
The penalty on dangerous tracking models a situation where
attempting to track the state too aggressively may make the
robot vulnerable to environmental hazards. The reward is set
as r(s, a) = −ca · (A − a + 1) + t(s, a), where ca is the cost
of expending one unit of effort in tracking, and t(s, a) is a
tracking performance reward, defined
−cp ·
t(s, a) = −cd ·
kr · s
Upper bound looseness
15
A = 16
A = 32
A = 64
10
5
0
2
4
8
16
32
64
128
256
S (log scale)
Fig. 3. The average distance between the upper bound and optimal policy as
function of state space size S for various action space sizes A.
VII. C ONCLUSION
We examined information acquisition in POMDPs with a
reward function nonlinear in the belief state. We showed
that if the POMDP fulfills certain structural properties, the
optimal infinite horizon stationary policy may be lower and
upper bounded by myopic (greedy) lower and upper bound
policies. Based on the bounds, we designed a branch-andbound pruning algorithm for online planning in POMDPs, and
demonstrated its effectiveness in a target tracking application.
The main advantage of our approach is that, if the structural
properties are satisified in the POMDP model, the resulting
branch-and-bound algorithm saves orders of magnitude in
computation compared to the existing approaches. The main
drawback is that the requred structural properties can be quite
restrictive in some applications. For example, a natural order
in the state and action spaces is required.
Future work will focus on relaxing the structural requirements while keeping track of the effect on the policy bounds.
This has potential to widen the scope of the possible applications of our apporach significantly.
R EFERENCES
[1] L. Kaelbling, M. Littman, and A. Cassandra, “Planning and acting
in partially observable stochastic domains,” Artificial Intelligence,
vol. 101, no. 1-2, pp. 99–134, 1998. [Online]. Available: http:
//linkinghub.elsevier.com/retrieve/pii/S000437029800023X
[2] O. Madani, S. Hanks, and A. Condon, “On the Undecidability
of Probabilistic Planning and Related Stochastic Optimization
Problems,” Artificial Intelligence, vol. 147, no. 1–2, pp. 5–34,
2003. [Online]. Available: http://www.sciencedirect.com/science/article/
pii/S0004370202003788
[3] W. Lovejoy, “A Survey of Algorithmic Methods for Partially Observed
Markov Decision Processes,” Annals of Operations Research, vol. 28,
no. 1, pp. 47–65, 1991.
[4] M. Hauskrecht, “Value-function Approximations for Partially Observable Markov Decision Processes,” Journal of Artificial Intelligence
Research, vol. 13, no. 1, pp. 33–94, 2000.
[5] G. Shani, J. Pineau, and R. Kaplow, “A Survey of Point-based POMDP
Solvers,” Autonomous Agents and Multi-Agent Systems, vol. 27, no. 1,
pp. 1–51, 2013.
[6] C. Stachniss, G. Grisetti, and W. Burgard, “Information Gain-based
Exploration using Rao-Blackwellized Particle Filters,” in Proc. Robotics:
Science and Systems (RSS), Cambridge, MA, USA, Jun. 2005.
[7] B. Charrow, V. Kumar, and N. Michael, “Approximate Representations
for Multi-robot Control Policies that Maximize Mutual Information,”
Autonomous Robots, vol. 37, no. 4, pp. 383–400, Aug. 2014.
TABLE II
T HE ( MINIMUM , AVERAGE , MAXIMUM ) PERCENTAGE OF ACTIONS PRUNED IN TARGET TRACKING DOMAINS AS FUNCTION OF STATE SPACE SIZE S AND
ACTION SPACE SIZE A.
S
A=4
A=8
A = 16
A = 32
A = 64
4
8
(75.0, 94.9, 100)
(50.0, 75.0, 75)
(75.0, 83.3, 87.5)
(37.5, 59.5, 62.5)
(75.0, 78.7, 87.5)
(50.0, 58.0, 75.0)
(71.9, 74.7, 75.0)
(59.4, 64.9, 71.9)
(64.1, 72.0, 75.0)
(48.4, 57.6, 67.2)
16
32
(50.0, 72.7, 75)
(50.0, 73.9, 75)
(50, 65.2, 75.0)
(37.5, 58.5, 62.5)
(43.8, 61.3, 68.8)
(37.8, 54.9, 62.5)
(46.9, 55.0, 62.5)
(37.5, 49.9, 59.4)
(46.9, 55.5, 62.5)
(43.8, 52.6, 59.4)
64
128
(50.0, 71.9, 75)
(50.0, 73.2, 75)
(37.5, 60.4, 62.5)
(50, 61.1, 62.5)
(43.8, 54.0, 62.5)
(43.8, 54.7, 56.3)
(37.5, 49.6, 59.4)
(40.6, 49.3, 53.1)
(40.6, 47.9, 53.1)
(42.2, 48.3, 53.1)
256
(50.0, 72.8, 75)
(50, 60.4, 62.5)
(43.8, 50.9, 56.3)
(43.8, 49.1, 53.1)
(n/a)
[8] N. Atanasov, J. Le Ny, K. Daniilidis, and G. J. Pappas, “Information
Acquisition with Sensing Robots: Algorithms and Error Bounds,” in
Proc. IEEE International Conference on Robotics and Automation
(ICRA), Hong Kong, China, Jun. 2014, pp. 6447–6454.
[9] M. Lauri and R. Ritala, “Optimal Sensing via Multi-armed Bandit Relaxations in Mixed Observability Domains,” in Proc. IEEE International
Conference on Robotics and Automation (ICRA), Seattle, WA, May
2015, pp. 4807–4812.
[10] T. Cover and J. Thomas, Elements of Information Theory. John Wiley
& Sons, 2012.
[11] S. Ross, J. Pineau, S. Paquet, and B. Chaib-draa, “Online planning
algorithms for POMDPs,” Journal of Artificial Intelligence Research,
vol. 32, no. 1, pp. 663–704, 2008. [Online]. Available: http:
//www.aaai.org/Papers/JAIR/Vol32/JAIR-3217.pdf
[12] M. Araya-López, O. Buffet, V. Thomas, and F. Charpillet, “A POMDP
Extension with Belief-dependent Rewards,” in Advances in Neural
Information Processing Systems 23, J. Lafferty, C. Williams, J. ShaweTaylor, R. Zemel, and A. Culotta, Eds., Vancouver, Canada, Dec. 2010,
pp. 64–72.
[13] D. Silver and J. Veness, “Monte-Carlo Planning in Large POMDPs,”
in Advances in Neural Information Processing Systems 23, J. Lafferty,
C. Williams, J. Shawe-Taylor, R. Zemel, and A. Culotta, Eds., Vancouver, Canada, Dec. 2010, pp. 2164–2172.
[14] A. Somani, N. Ye, D. Hsu, and W. S. Lee, “DESPOT: Online POMDP
Planning with Regularization,” in Advances in Neural Information
Processing Systems 26, C. J. C. Burges, L. Bottou, M. Welling,
Z. Ghahramani, and K. Q. Weinberger, Eds. Curran Associates, Inc.,
2013, pp. 1772–1780. [Online]. Available: http://papers.nips.cc/paper/
5189-despot-online-pomdp-planning-with-regularization.pdf
[15] W. S. Lovejoy, “Some Monotonicity Results for Partially Observed
Markov Decision Processes,” pp. 736–743, 1987.
[16] U. Rieder, “Structural results for partially observed control models,”
ZOR Zeitschrift für Operations Research Methods and Models of
Operations Research, vol. 35, no. 6, pp. 473–490, Nov. 1991. [Online].
Available: http://link.springer.com/10.1007/BF01415990
[17] V. Krishnamurthy and U. Pareek, “Myopic Bounds for Optimal Policy
of POMDPs: An Extension of Lovejoys Structural Results,” Operations
Research, vol. 63, no. 2, pp. 428–434, Apr. 2015. [Online]. Available:
http://pubsonline.informs.org/doi/10.1287/opre.2014.1332
[18] V. Krishnamurthy and B. Wahlberg, “Partially Observed Markov Decision Process Multiarmed Bandits–Structural Results,” Mathematics of
Operations Research, vol. 34, no. 2, pp. 287–302, 2009.
[19] V. Krishnamurthy and D. V. Djonin, “Structured threshold policies
for dynamic sensor scheduling - A partially observed Markov
decision process approach,” IEEE Transactions on Signal Processing,
vol. 55, no. 10, pp. 4938–4957, Oct. 2007. [Online]. Available: http:
//ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=4305446
[20] M. Naghshvar and T. Javidi, “Active sequential hypothesis testing,”
The Annals of Statistics, vol. 41, no. 6, pp. 2703–2738, Dec. 2013.
[Online]. Available: http://projecteuclid.org/euclid.aos/1387313387
[21] ——, “Active M-ary sequential hypothesis testing,” in 2010 IEEE
International Symposium on Information Theory, vol. 1, no. 1. IEEE,
Jun. 2010, pp. 1623–1627. [Online]. Available: http://ieeexplore.ieee.
org/lpdocs/epic03/wrapper.htm?arnumber=5513381
[22] M. H. DeGroot, Optimal Statistical Decisions. Hoboken, New Jersey:
John Wiley & Sons, Inc., 2004, Wiley Classics Library edition.
[23] S. Thrun, W. Burgard, and D. Fox, Probabilistic Robotics. Cambrdige,
MA: The MIT Press, 2006.
[24] D. Bertsekas and S. Shreve, Stochastic Optimal Control: The DiscreteTime Case. Athena Scientific, 1996.
[25] M. Shaked and J. G. Shanthikumar, Stochastic Orders. New
York, NY: Springer New York, 2007. [Online]. Available: http:
//link.springer.com/10.1007/978-0-387-34675-5
[26] S. M. Fallat and C. R. Johnson, Totally Nonnegative Matrices. Princeton, NJ: Princeton University Press, 2011.
[27] S. Karlin, Total Positivity. Stanford University Press, 1968.
[28] W. Kaplan, “A Test for Copositive Matrices,” Linear Algebra and its
Applications, vol. 313, no. 1-3, pp. 203–206, 2000.
| 3cs.SY
|
arXiv:1511.04203v2 [physics.plasm-ph] 25 May 2016
GPEC, A REAL-TIME CAPABLE TOKAMAK EQUILIBRIUM
CODE
M. Rampp1 , R. Preuss2 , R. Fischer2 and the ASDEX Upgrade Team2
1
Max Planck Computing and Data Facility
Gießenbachstraße 2, 85748 Garching, Germany
2
Max-Planck-Institut für Plasmaphysik,
Boltzmannstraße 2, 85748 Garching, Germany
May 26, 2016
Abstract
A new parallel equilibrium reconstruction code for tokamak plasmas is presented.
GPEC allows to compute equilibrium flux distributions sufficiently accurate to derive parameters for plasma control within 1 ms of runtime which enables real-time
applications at the ASDEX Upgrade experiment (AUG) and other machines with a
control cycle of at least this size. The underlying algorithms are based on the wellestablished offline-analysis code CLISTE, following the classical concept of iteratively
solving the Grad-Shafranov equation and feeding in diagnostic signals from the experiment. The new code adopts a hybrid parallelization scheme for computing the
equilibrium flux distribution and extends the fast, shared-memory-parallel Poisson
solver which we have described previously by a distributed computation of the individual Poisson problems corresponding to different basis functions. The code is based
entirely on open-source software components and runs on standard server hardware
and software environments. The real-time capability of GPEC is demonstrated by
performing an offline-computation of a sequence of 1000 flux distributions which are
taken from one second of operation of a typical AUG discharge and deriving the relevant control parameters with a time resolution of a millisecond. On current server
hardware the new code allows employing a grid size of 32 × 64 zones for the spatial
discretization and up to 15 basis functions. It takes into account about 90 diagnostic signals while using up to 4 equilibrium iterations and computing more than 20
plasma-control parameters, including the computationally expensive safety-factor q
on at least 4 different levels of the normalized flux.
1
I
INTRODUCTION
or other simplifying assumptions like function
parametrization 11 are adopted e.g. 12 . Lately,
three codes based on fast Grad-Shafranov
solvers have been able to demonstrate real-time
capability: P-EFIT 13 , a GPU-based variant of
the EFIT equilibrium-reconstruction code 9,7 , is
able to compute one iteration within 0.22 ms on
a 65 × 65 grid with 3 basis functions. With a
similar resolution the JANET software 14,4 has
reached runtimes of about 0.5 ms per iteration
on a few CPU cores within the LabVIEW-based
system at AUG. Using a single CPU core
overclocked at 5 GHz, the LIUQE code 15 which
is deployed at the Swiss TCV experiment,
achieves a cycle time of about 0.2 ms with
a spatial resolution of 28 × 65 points. In all
these cases, however, only a single iteration is
performed and only a very small number of basis
functions can be afforded for the parametrization of the source term of the Grad-Shafranov
equation. While such restrictions may well be
justifiable under specific circumstances 16 the
motivation for our work is to substantially push
the limits of numerical accuracy in terms of
spatial resolution, number of basis functions,
and number of iterations and thus the quality of
the equilibrium solution. This is expected to aid
the generality and robustness of the application,
in particular with respect to variations in the
plasma conditions.
In addition to the reliability and accuracy of
real-time equilibrium reconstructions, the sustainability and maintainability of the codes is
considered an important aspect. Typical implementations of analysis tools in fusion science
have a rather long life time which often exceeds
life times of commercial hardware and software
product cycles. The long usability required for
analysis tools poses special requirements on the
software chosen and the maintenance capabilities
Reconstruction of the plasma equilibrium shape
is a key requirement for the operation of current
and forthcoming tokamak experiments e.g. 1,2,3,4 .
The most commonly employed method numerically reconstructs the plasma equilibrium by
iteratively solving the two-dimensional GradShafranov equation 5,6 for the poloidal flux function (ψ), and using diagnostic signals from the
experiment as constraints 7,8 . In short, the
source term of the Grad-Shafranov equation is
expanded into a linear combination of basis functions which allows formulating a linear regression problem for the expansion coefficients, using
the actual diagnostic signals and their forwardmodelling based on the numerical solution for ψ.
The problem can be tackled numerically by solving a number of independent Grad-Shafranovtype equations — one for each basis function
of the expansion — and employing a Picarditeration procedure that allows evaluating the
source term using the solution from the previous iteration step.
A variety of equilibrium-reconstruction
codes based on this strategy 8 or similar approaches 9,7,10 exist and have been routinely used
for diagnostics and data analysis at the various
tokamak experiments. The codes EFIT 9,7 and
CLISTE 8 , for example, have been employed at
DIII-D, EAST and ASDEX Upgrade (AUG),
respectively, a class of medium-sized tokamaks
which are characterized by a control cycle
on the order of a millisecond. Until recently,
however, computing times in the millisecond
range were inaccessible to such ”first-principles”
codes, and hence their applicability for real-time
diagnostic analysis and plasma-control is very
limited, unless the numerical resolution or the
number of iteration steps is drastically reduced
2
The new code enables computing the equilibrium flux distribution and the derived diagnostics and control parameters within 1 ms of runtime, given a grid size of 32 × 64 zones with up
to 15 spline basis functions for the discretization,
and about 90 diagnostic signals. With a runtime
of less than 0.2 ms for a single equilibrium iteration and about 0.2 ms required for computing a
variety of more than 20 relevant control parameters, one can afford 4 iterations for converging
the equilibrium solution. The latter was checked
to be sufficient for reaching accuracies of better
than a percent for the relevant quantities. As a
proof of principle we shall demonstrate the real
time capability of the new code by performing
an offline analysis using data from an AUG discharge. To our knowledge the new GPEC code
to date is one of the fastest (at a given numerical
accuracy) and most accurate (at a given runtime
constraint) of its kind.
necessary to adapt to changing environments.
Open-source software together with off-the-shelf
computer hardware is thought of being perfectly suited for these demands 17 . This in particular applies to all non-standardized components like, e.g., the implementation of complex,
tailored, and evolving numerical algorithms for
equilibrium reconstruction. Commercial, closedsource software (like, e.g., compilers or numerical libraries) and hardware components may well
comply with such demands, provided that they
can be modularly interchanged using standardized interfaces.
To this end we have implemented a new,
parallel equilibrium-reconstruction code, GPEC
(Garching Parallel Equilibrium Code), which
builds on the fast, shared-memory-parallel GradShafranov solver we have developed previously 18
and a two-level hybrid parallelization scheme
which was pointed out in the same paper. The
basic numerical model and functionality of the
new code originate from the well-established and
validated algorithms implemented by the equilibrium codes CLISTE 8 and specifically its descendant IDE 19 , both of which are being used for
comprehensive offline diagnostics and data analysis at the ASDEX Upgrade experiment. GPEC
is based entirely on open-source software components, runs on standard server hardware and
uses the same code base and computer architecture as employed for performing offline analyses
with the IDE code. Such a strategy is considered a major advantage, in particular concerning verification and validation of the code 7 and
also its evolution: On the one hand, algorithmic
and functional innovations in the offline physics
modeling can gradually be taken over by the realtime version. On the other hand, high-resolution
offline analysis can directly take advantage of optimizations achieved for the real-time variant.
The paper is organized as follows: in Section II
we recall the basic equations and describe our
new, hybrid-parallel implementation of the equilibrium solver and its verification and validation.
Its computational performance and in particular real-time capability is demonstrated in Section III on an example application using real
AUG data. Section IV provides a summary and
conclusions.
II
II.A
ALGORITHMS AND IMPLEMENTATION
Equilibrium Reconstruction
Basic equations The Grad-Shafranov equation 5,6 describing ideal magneto-hydrodynamic
equilibrium in two-dimensional tokamak geome3
try reads
∆∗ ψ(r, z) = −2πµ0 r jφ
,
the right-hand side. The updated flux distribution is then given by
(1)
with cylindrical coordinates (r, z), the elliptic
differential operator
∂ 1 ∂
∂2
+ 2
∂r r ∂r ∂z
ψ=
N
X
ck ψk
,
(7)
k=1
(2) which follows from Eqs. (1,4) and the linearity of the operator ∆∗ . The unknown coeffiand the poloidal flux function (in units of Vs), cients ck are determined by experimental data:
Using the N distributions ψk obtained from
ψ(r, z). The toroidal current density profile,
Eqs.
(5,6) the so-called response matrix B :=
∂p(r, ψ) F (ψ) dF (ψ)
+
,
(3) {bl (ψk )}l=1,...,M,k=1,...N , consisting of predictions
jφ := 2π r
∂ψ
µ0 r dψ
for a set of M diagnostic signals {ml }l=1...,M is
consists of two terms, where p(r, ψ) ≡ p(ψ) is calculated. The prediction bl (ψ) is some (linear)
the plasma pressure (isotropic case) and F (ψ) = function employing the flux distribution ψ to
produce the forward-modelled signal. All mearBφ = (µ0 /2π)Ipol is proportional to the total
surements are located within the poloidal flux
poloidal current, Ipol .
grid such that the poloidal flux outside the grid
is not needed. Linear regression of the response
Numerical solution The classical stratmatrix with the measured diagnostic signals fiegy 8,20,16 for numerically solving Eq. (1) is based
nally yields the coefficients ck for Eq. (7). This
on an expansion of the current density profile jφ
procedure is iterated until certain convergence
on the right-hand-side into a linear combination
criteria are fulfilled.
of a number N = Np + NF of basis functions,
The private flux region close to the divertor
is treated differently, although the same poloidal
jφ (r, z) =
flux as inside the plasma occurs. Since there
Np
NF
X
X
1
are no measurements of the current distribution
cNp +k ϕk (ψ) .
ck πk (ψ) +
2π r
µ0 r
in the private flux region and the chosen funck=1
k=1
tional form for the current decay was proven to
(4)
be of minor importance for the equilibrium reIn each cycle of a Picard-iteration scheme, a construction, the current is chosen to decay approximately exponentially with a decay length
number of N Poisson-type equations,
of about 5 mm, starting at the last closed flux
∆∗ ψk = −4π 2 µ0 r 2 πk (ψ) (k = 1 . . . Np ), (5) surface.
∆∗ := r
,
and
The algorithm can be summarized as follows:
∗
2
∆ ψNp +k = −4π ϕk (ψ)
(k = 1 . . . NF ),
(6)
1. evaluate right-hand sides gk (r, z) :=
4π 2 µ0 r 2 πk (ψ i−1 ) and gNp +k (r, z)
:=
2
i−1
4π ϕk (ψ ) of Eqs. (5,6), using the solu-
are solved individually, where the solution ψ
from the last iteration step is used for evaluating
4
tion, ψ i−1 , from the previous iteration step measurements are used in the IDE code in the
(or initialization).
off-line mode, but are not subject of the present
work due to the lack of reliable on-line availabil2. solve the Poisson-type equations ∆∗ ψk = ity. The relatively large number of basis func−gk (r, z) individually for each k = 1, . . . , N . tions (and hence fit parameters), which is moAn additional Poisson problem arises from tivated by the need to allow for equilibria sufa convergence-stabilization procedure which ficiently flexible to address all occurring plasma
introduces the vertical plasma position as an scenarios, requires application of regularization
auxiliary free parameter 8,15 .
constraints. While in the IDE code additional
3. evaluate {bl (ψk )}l=1,...,M,k=1,...,N to con- smoothness (curvature and amplitude) requireare applied to the source profiles p′ and
struct the response matrix B.
Ad- ments
′
ditional columns {bl,N +1 }l=1,...,M and F F (which effectively adds additional columns
{bl,N +1...N +1+Next }l=1,...,M arise from the to the response matrix), GPEC currently adopts
21
abovementioned convergence-stabilization a simpler ridge-regression procedure .
procedure and from deviations in the
currents measured at a number of Next
Implementation and parallelization The
external field coils to account for wall
implementation of GPEC starts out from the seshielding and plasma-induced wall currents,
rial offline-analysis code IDE 19 , thus closely folrespectively.
lowing the concepts of the well-known CLISTE
4. perform linear regression on B · c = m to code 8 . Specifically, for each of the two sets of
basis functions, {πk }k=1...NP and {ϕk }k=1...NF ,
obtain the set of coefficients {ck }k=1,...,N .
GPEC employs a basis of cubic spline polyPN
5. evaluate new solution ψ i =
k=1 ck ψk + nomials which are defined by a number of Np
ψ ext , adding contributions from the mea- and NF knots at positions xi , respectively, with
sured and fitted deviating currents in ex- πk (xi ) = δki , ϕk (xi ) = δki and natural boundary
ternal coils, ψ ext .
conditions. Bicubic Hermite interpolation is employed for evaluating the flux distribution ψk and
6. go back to step 1.
until convergence
its gradients (which are required for computing
is reached, e.g. in terms of the maxithe magnetic field) at the positions of the magmum norm evaluated over all grid points,
netic probes and the flux loops in order to calcukψ i − ψ i−1 k∞ .
late the corresponding forward-modelled signals
The response matrix B is already prepared for for the response matrix. Thus, besides achieving
additional measurements of the motional Stark higher-order interpolation accuracy, smoothness
effect (MSE), the Faraday rotation, the pressure of the magnetic field is guaranteed by construc22
profile (electron, ion and fast particle pressure), tion .
GPEC utilizes the fast, thread-parallel Poisthe q-profile (e.g. from MHD modes), the divertor tile currents constraining F (ψ) on open son solver we have developed previously 18 . It
flux surfaces, the measurements of loop voltage is based on the two-step scheme for solving
and of iso-flux constraints 19 . These additional the Poisson equation with Dirichlet boundary
5
conditions 23 and replaces the cyclic-reduction
scheme which is traditionally employed in the
codes GEC and CLISTE 24,8 by a parallelizable
Fourier-analysis method for decoupling the linear system into tridiagonal blocks 22 . For details
on the implementation and computational performance we refer to refs. 25,18 .
As briefly sketched in ref. 18 the main idea of
the new code is to exploit an additional level of
parallelism by distributing the individual Poisson problems for determining each of the ψk
to different MPIa -processes. The algorithm, together with notes on the parallelization (relevant
MPI communication routines are noted in brackets) is summarized as follows.
5. performs linear regression on m = B · c
thread-parallelization of linear algebra
routines
6. gathers all ck′ ψk′ computed by the other
processes Pk′6=k and computes ψ =
PN
k′=1 ck′ ψk′
collective communication (MPI Allreduce)
The distribution of the Poisson problems to
different processes comes at the price of collective communication (steps 4 and 6), in which
all processes combine their data with all others, using MPI routines MPI Allgather and
MPI Allreduce, respectively.
Each process Pk=1,...,N is assigned to a single
II.B Computation of Plasma-Control
basis functionb , and process PN +1 is dedicated to
Parameters
the additional Poisson problem that arises from
the introduction of the vertical shift parameter 8 . Given the solution for the poloidal flux function
Accordingly, each process Pk :
ψ(r, z), GPEC computes more than 20 derived
1. computes gk (r, z) using the solution ψ(r, z) quantities which are relevant for plasma confrom the previous iteration step (or initial- trol. Two of them, the (r, z)-coordinates of the
magnetic axis, Rmag and Zmag can be obtained
ization)
thread-parallelization over (r, z)-grid immediately by searching the maximum of the
equilibrium flux. For the rest of the quantities
2. employs fast Poisson solver 18 to solve four major computational tasks (labelled A–D
∆∗ ψk = −gk (r, z)
in the following compilation) have to be carried
thread-parallelization 18 out. Their results allow evaluating the individ3. computes column b(ψk ) of the response ma- ual control parameters (summarized under the
bullet points) in a numerically straightforward
trix
thread-parallelization over (r, z)-grid and inexpensive way. For details on the physical definition and derivation of these quantities,
4. gathers columns b(ψk′ ) computed by the see, e.g. Ref. 27 . Implementation details are given
other processes Pk′6=k and assembles re- further below.
sponse matrix B
collective communication (MPI Allgather) (A) determination of grid points which are ena
closed by the separatrix
The ”Message Passing Interface” standard 26 .
b
In general the implementation supports an even distribution of the N + 1 Poisson problems to a number of
N ′ processes, provided N ′ divides N + 1.
• coordinates (Rgeo , Zgeo ) of the geometric axis, defined as the area6
weighted
integral within the separatrix (D) computation of the pressure distribution
R
(r, z)dS/S.
p(r, z) and the poloidal magnetic field
• the horizontal and vertical minor plasma qradius is given by
R
2
ahor
= 2
r 2 dS/S − Rgeo
and
qR
2 , respecbver = 2
z 2 dS/S − Zgeo
tively, which determines the elongation
k = bver /ahor .
• the total energy content of
R the plasma
is given by WMHD = 3/2 pdV .
• the same integration can be used
for computing theR poloidal beta parameter, βpol =
pdV /Zβ , employing the normalization constant Zβ =
2 /(2 · l2 ) (as used in
µ0 · Vplasma · Itor
pol
CLISTE) with lpol and Itor as specified
under (B) and (C), respectively.
• the plasma
R 2self-inductivity li =
−1
(2µ0 Zβ )
Bpol dV is given by an integral over the squared poloidal mag2 = (2πr)−2 · ((∂ ψ)2 +
netic field Bpol
r
(∂z ψ)2 ).
• dRXP , the difference of the poloidal
flux at the two x-points divided by
(2πRaus ) and by the poloidal magnetic field at (Raus , Zmag ): (ψXP1 −
ψXP2 )/|∇ψ|(Raus ,Zmag ) .
(B) identification of contour lines (r(t), z(t))
with ψ(r(t), z(t)) = ψc for a number of particular values of the normalized flux ψc
• from the separatrix curve (with ψc
taken as the value at the innermost
x-point which defines also the lastclosed flux surface LCFS) a number
of geometric properties can be derived
straightforwardly such as the (r, z)coordinates of the uppermost and lowermost point (in z-direction) on the
plasma surface, or Rin (Raus ) as the
R-coordinate of the innermost (outerImplementation and parallelization The
most) point on the plasma surface.
• the length of the poloidal perimeter lpol algorithms for computing the plasma-control paof the plasma is computed by integra- rameters and their basic implementation are
taken directly from the IDE code. We emphasize
tion over the LCFS.
that unlike other real-time codes 15 GPEC makes
• the safety factor q is computed by inte- no algorithmic simplifications compared with the
gration over contours defined by levels offline variant of the code. In particular, GPEC
ψc ∈ [0.25, 0.50, 0.75, 0.95].
uses the same spatial grid resolution as adopted
(C) computation
of the toroidal plasma current for equilibrium reconstruction and employs the
R
high-order interpolation schemes (cubic splines,
Itor = jφ dS (Eq. 3)
bicubic Hermite interpolation) of IDE. In the fol• the indicators for the current-center lowing we briefly summarize the main concepts
are given by the following current- adopted for the computationally most expensive
weighted integrals of
qR the current routines:
R
r 2 jφ dS/Itor ,
density Rsquad =
(A) for performing the integrations . . . dS/S,
R
Zsquad = zjφ dS/Itor .
grid points located outside the separatrix
7
are masked out according to a comparison
with the value of ψ at the x-point.
lation polynomials for intra-grid interpolation, cf. Sect. II.A).
(B) a custom contour-finding algorithm is used
which employs quadratic interpolation to
obtain the values at the intersections of
the spatial grid with the contour curve.
Quantities on the LCFS, like lpol , are determined by quadratic extrapolation from
three equidistant contour levels close to the
plasma surface.
As shall be shown in Sect. III.B the computing time is dominated by finding and integrating along contours as well as by evaluating the
pressure distribution. Accordingly, the computations are grouped into the following mutually
independent tasks which can be assigned to different MPI tasks, with only a few scalar quantities to be communicated for evaluating the final
results:
(C) for the evaluation
R of the toroidal plasma
current Itor = jφ dS the current density
• computation of the pressure distribution
jφ (r, z) is computed according to Eq. (4).
p(r, z) and the poloidal magnetic field
The summands ck πk (ψ) and cNp +k ϕk (ψ) in
principle would be available from the last
• handling of 3 contour levels for extrapolaequilibrium iteration but are distributed
tion to the last closed flux surface
among processors (cf. Sect. II.A). It turns
• handling of 4 contour levels corresponding
out that locally recomputing πk (ψ) and
to the set of safety factors q25 , q50 , q75 , q95
ϕk (ψ) by using the converged equilibrium
solution ψ (a copy of which is available on
• determination of the x-point and handling
each processor) is faster than collecting the
of 1 contour level corresponding to the sepindividual summands by MPI communicaaratrix curve
tion.
Within each MPI task, OpenMP threads are
(D) the computation of the pressure distriused, e.g. for parallelizing the computation of
bution p(r, z) is based on the identity
the pressure over the individual points of the
P Np
∂ψ p(ψ(r, z)) =
k=1 ck πk (ψ) (cf. Eq. 4). (r, z)-grid or for parallelizing over different conFor fast evaluation of p(r, z) on the poloidal tour levels.
mesh, an equidistant grid of ψl , l = 1, . . . , 12
points is constructed. For each ψl the corII.C Verification and Validation
responding pressure contribution πk (ψl ) is
computed by numerical integration starting The codes CLISTE and IDE have been extenat the plasma boundary. A cubic spline in- sively validated and verified by application to
terpolation is used to evaluate πk (ψ(r, z)) on ASDEX Upgrade data and by means of code
this grid. The gradients for computing the comparisons. Thus, only those parts of the new
poloidal magnetic field are obtained using GPEC code need to be verified which are treated
centered finite differences of the neighboring differently from the IDE code. The procedure is
grid points (consistent with the derivatives greatly facilitated by the fact that both codes
used to define the bicubic Hermite interpo- implement the same algorithms and share the
8
same code base. The only differences in the III APPLICATION PERFORGPEC implementation with respect to IDE are
MANCE
1) the prescription of a fixed number of equilibrium iterations and 2) the use of only 12 instead III.A Demonstration of Real-Time
of 30 integration points for computing the presCapability
sure distribution p(r, z) from its gradient ∂ψ p(ψ)
In order to demonstrate the real-time capabil(see item D above) on every grid point (r, z).
ity of the new code a post-processing run is perFigure 1 shows the sensitivity of a number of formed using data from a typical AUG discharge.
selected parameters to the number of equilibrium We chose AUG shot number #23221, and focus
iterations for an arbitrarily chosen time point on a time window between t1 = 3.014 s and
(t = 3.015 ms) of AUG discharge #23221. By t1000 = 4.013 s with 1000 time points, correusing the solution ψ from the last time point sponding to a time resolution of exactly 1 ms.
as an initial value, the plasma geometry (pa- As shall be shown below this simulation of 1000
rameters Rin , Raus , and the total volume of the time points can be performed on standard server
plasma, Vplasma ) is already very well determined. hardware within less than a second of computing
Within only a few iteration steps accuracies of time.
The GPEC code computes a fixed number of
better than a percent compared with the converged solution (obtained with 200 iterations) four equilibrium iteration steps for every time
are obtained also for the quantities which de- point and uses the equilibrium solution ψ(tn ) at
pend on the pressure distribution (βpol , WMHD ), time point tn as the initial value for the next
or on the toroidal current (Itor ). For the real- time point tn+1 . A standard spatial resolution
time demonstration presented in the subsequent of (Nr , Nz ) = (32, 64) grid points is employed,
section we shall fix the number of iterations to and N = 7 basis functions are used. The refour and show that the accuracy level of 1 % is sponse matrix B contains signals from 61 magmaintained over a time sequence of 1 s with 1000 netic probes and 18 flux loops and in addition
time points.
takes into account 10 external coils. Figure 3
provides an overview of the time evolution of a
Figure 2 shows that a number of 12 integration number of characteristic parameters, such as sepoints for the numerical integration of ∂ψ p(ψ) lected plasma-shape parameters, energy content
is more than sufficient in order to reach sub- and safety factor q95 (black, solid lines). For refpercent accuracies for the control parameters erence, the figure shows an extended time winβpolR and WMHD , both of which are proportional dow between t = 0.4 s and t = 4.4 s, which
to pdV (see item D above). The time for com- includes the dynamic start-up phase of the disputing p(r, z) is proportional to the number of charge with significant variations in the plasma
integration points and scales almost ideally with parameters, and also shows results from a run
the number of threads which motivates our spe- which converges the solution ψ until the maxcific choice of 12 points as a multiple of 6 threads imum changes on the grid, kψ i − ψ i−1 k∞ , are
that shall be employed for the majority of com- below 10−4 (blue, dashed lines). Compared with
putations (cf. Sect. III.B).
the converged reference run one notices good
9
Fig. 1. Sensitivity of selected parameters to the number of equilibrium iterations for computing ψ(r, z).
For each quantity the relative deviation from its converged value is shown as a function of the number
of iterations. Note that the deviations for the geometric quantities, Rin , Raus , and the total volume of
the plasma, Vplasma are multiplied by a factor of 10. Data was taken from AUG shot #23221 at time
t = 3.015 ms.
overall agreement of the ”real-time” simulation
with a fixed number of 4 equilibrium iterations
throughout the extended time interval, even during the initial start-up phase which lasts until
approx. 1.5 s. It is hence justified to confine the
analysis on the — arbitrarily chosen — time interval, [3.014 s, 4.013 s] in the ”plateau” phase
of the discharge. Focusing on this time window,
Figure 4 shows that the relative errors for most
quantities are at smaller than one percent. The
comparably large variations of Zmag are due to
its small absolute value. The absolute scatter of
Zmag is in the order of mm. Note that the cycle
time of ∆t = 1 ms is much shorter than the
10
Fig. 2. Sensitivity of βpol and of the total energy content WMHD to the number of integration points, Npsi ,
spent for computing the pressure from its gradient on every grid point (r, z). The relative deviation from
the converged value is shown as a function of the number of integration points, Npsi . Data was taken from
AUG shot #23221 at time t = 3.015 ms.
timescale of the changes of plasma conditions
(cf. the increase of WMHD , βpol at t ≈ 3.5 s).
Thus the equilibrium solution, even if not fully
converged, can easily follow such secular trends,
which is reflected by the fact that there is no
visible change in the magnitudes of the relative
error at t ≈ 3.5 s.
The employed spatial resolution of (Nr , Nz ) =
(32, 64) grid points is a common choice
for real-time analysis in medium-sized tokamaks 9,7,14,4,15 . For the chosen AUG discharge
a comparison with a run using a twofold finer
spatial grid spacing, (Nr , Nz ) = (64, 128), shows
good agreement for the majority of quantities,
with q95 and Router exhibiting the largest sensitivity to the resolution.
We conclude that by performing four iterations per time point and using (Nr , Nz ) =
(32, 64) spatial grid points, sufficiently accurate
equilibrium solutions for deriving real-time control parameters are obtained, at least in the sense
of a proof-of-principle using the chosen example
data from AUG shot #23221. A systematic analysis of the accuracy and a comprehensive validation of the code under various tokamak operational scenarios is beyond the scope of this
paper.
III.B
Computational Performance
Overview The computational performance of
GPEC is assessed on a standard compute cluster with x86 64 CPUs and InfiniBand (FDR 14)
interconnect. The individual compute nodes
are equipped with two Intel Xeon E5-2680v3
”Haswell” CPUs (2.5 GHz, 2x12 cores). We use a
11
resolution
(Nr , Nz , N+1)
(32, 64, 6)
(32, 64, 8)
(32, 64, 8)
(32, 64, 12)
(32, 64, 16)
(32, 64, 16)
(32, 64, 16)
(64, 128, 6)
(64, 128, 8)
(64, 128, 12)
(64, 128, 16)
Ttot
[ms]
0.85
1.08
0.88
0.97
1.55
1.26
0.99
2.07
1.88
2.02
2.04
Teq
[ms]
0.62
0.87
0.67
0.74
1.32
1.02
0.73
1.55
1.28
1.39
1.40
Tctl
[ms]
0.18
0.17
0.17
0.19
0.18
0.18
0.21
0.39
0.48
0.49
0.51
cores
(nodes)
24 (1)
24 (1)
48 (2)
72 (3)
24 (1)
48 (2)
96 (4)
48 (2)
48 (2)
72 (3)
96 (4)
MPI
tasks
6
4
8
12
4
8
16
6
8
12
16
threads
per task
4
6
6
6
6
6
6
8
6
6
6
TABLE I. Overview of the total runtime Ttot per time point (2nd column) for different combinations of
the numerical resolution (1st column) and compute resources. The latter are given in terms of the total
number of cores (nodes) (5th column), of MPI tasks (6th column) and of OpenMP threads per MPI task
(7th column). Ttot is the sum of the runtime Teq (3rd column) for computing the equilibrium distribution
ψ with 4 iteration steps (cf. Sect. II.A), and Tctl (4th column) for deriving plasma-control parameters (cf.
Sect. II.B). The benchmarks were performed on Xeon E5-2680v3 ”Haswell” CPUs (2.5 GHz, 24 cores per
node) and used data from AUG shot #23221.
standard Intel software stack (FORTRAN compiler v14.0, MPI v5.0) on top of the Linux operating system (SLES11 SP3). For the required
linear algebra functionality the OpenBLAS 28 library (version 0.2.13) is employed using the defacto-standard interfaces from BLAS 29,30 and
LAPACK 31,30 , and the FFTW 32 library (version 3.3.4) is used for the discrete sine transforms in the Poisson solver. Both libraries are
open-source software released under the BSD or
the GPL license, respectively. If desired, opensource alternatives to the Intel compilers (e.g.
GCC 33 ) and MPI libraries (e.g. OpenMPI 34 ,
MVAPICH 35 ) could be readily utilized, albeit
possibly with a certain performance impact.
Table I shows an overview of the total runtime per time point and of the individual contributions from computing the equilibrium solution
and from deriving the set of plasma-control parameters. We chose combinations of different nu-
merical resolutions (in terms of the spatial grid
resolution Nr × Nz and number of basis functions N ) and computing resources (in terms of
the number of CPU cores).
With two compute nodes (each with 24 CPU
cores) and using a resolution of (Nr , Nz , N+1) =
(32, 64, 8) a runtime well below 1 ms and hence
real-time capability is reached. When doubling
the number of basis functions (N +1 = 16) realtime calculation is still possible on four compute
nodes, indicating overall good weak scalability of
the code.
Roughly three quarters of the runtime is required for computing the equilibrium solution
and the rest goes into evaluating the set of
plasma control parameters. Hence, with a reduction of the number of equilibrium iterations
to only one or two 13,14,4,15 runtimes of 0.5 ms are
reached on just a single compute server.
With a higher spatial resolution of (Nr , Nz ) =
12
Fig. 3. Time evolution of selected parameters for AUG shot #23221 as computed with the new real-time
code during an extended time window between t = 0.4 s and t = 4.4 s. The time interval [3.014 s, 4.014 s]
which is chosen for performance analysis is indicated by vertical lines. The individual plots show the
coordinates (Rmag , Zmag ) and (RX , ZX ) of the magnetic axis, and of the x-point, respectively, the horizontal
and vertical minor plasma radius, ahor and bver , the plasma self-inductivity, li , the poloidal beta parameter,
13 the safety factor q95 (see Sect. II.B for the definition
βpol , the total energy content of the plasma, WMHD , and
of the quantities). The solid, black lines correspond to a ”real-time” application which is restricted to 4
equilibrium iterations. For comparison, values derived from converged equilibrium solutions (convergence
criterion kψ i − ψ i−1 k∞ < 10−4 ) are shown as dashed, blue lines. A standard spatial resolution of (32, 64)
grid points and 7 basis functions were employed.
Fig. 4. Evolution of the relative deviation (denoted by square brackets) of the quantities which were
computed after 4 equilibrium iterations, from the values computed with converged equilibrium solutions
(convergence criterion kψ i − ψ i−1 k∞ < 10−4 ) during the time interval chosen for performance analysis (as
indicated by vertical lines in Fig. 3). A standard spatial resolution of (32, 64) grid points and 7 basis functions
were employed.
14
(64, 128) a cycle time of 2 ms is possible with 4
iterations, although from the linear algorithmic
complexity of the algorithm 18 , and the amount
of data which is communicated between the processes one would expect the runtime to increase
by a factor of four compared to (Nr , Nz ) =
(32, 64). The reason is a higher parallel efficiency of the Poisson solver which is limited by
OpenMP overhead at lower resolutions 18 as well
as lower communication overhead due to larger
message sizes in the expensive MPI Allreduce
operation which collects the distributed
P fields
ψk (r, z) and computes the sum ψ = N
k=1 ck ψk
(see the rows labelled ”Poisson solver” and ”ψ
sum” in Table II).
tions would be somewhat faster in this case the
resultant doubling of the effective runtime for the
Poisson solver cannot be compensated for.
Plasma-control parameters For the example of (Nr , Nz , N + 1) = (32, 64, 8) computed
with 48 CPU cores, Figure 5 shows how the runtime of Tctl = 0.17 ms (cf. Table I) is composed and illustrates the benefits of exploiting
task parallelism. Due to the final MPI Allreduce
operation in the equilibrium construction algorithm, which comes at little extra effective runtime cost as compared to a single MPI Reduce,
all MPI tasks hold a copy of the equilibrium
flux distribution ψ(r, z). Hence, independent
parts of the time-consuming computations for
the plasma-control parameters like, e.g. evaluating the pressure distribution or determining the
contours for the safety factor q, can be handled
in parallel by different MPI tasks and only scalar
quantities need to be communicated (within a
few microseconds) between the MPI tasks at the
end. If necessary, the remaining load-imbalance
indicated by Figure 5, namely task 0 and task
3 becoming idle after about 0.1 ms, and tasks
4, . . . , 7 not computing anything at all at this
stage, could be exploited for computing additional quantities and/or for further reducing the
effective runtime. Without task-parallelization,
on the contrary, the runtimes would add up to
more than 0.4 ms despite the thorough OpenMP
parallelization (e.g. across different contour levels) within the task.
Equilibrium reconstruction Table II further shows that the runtime for the equilibrium
iterations is dominated
P by computing the distributed sum, ψ = N
k=1 ck ψk , which is also an
implicit synchronization point for all processes,
and by the Poisson solver. Due to the use of
high-order interpolation schemes (cf. Sect.II.A)
a significant fraction of about 25% is spent on the
evaluation of the right-hand sides of Eqs. (5,6)
(”RHS update”) and of the response matrix B
(”RM evaluation”).
In general, the chosen products of number
of MPI tasks times OpenMP threads per task
turned out as the most efficient ones for this
hardware platform with 24 cores per node: using
more than 6 threads (or 8 threads in the case of
(Nr , Nz ) = (64, 128)) results only in a modest
speedup of the Poisson solver 18 but doubles the
required number of cores and network resources. IV CONCLUSIONS
Keeping those constant and reducing the number of MPI tasks instead by a factor of two would With the motivation of achieving subrequire each MPI task to handle two basis func- millisecond runtimes a new, parallel equilibriumtions. Although the collective MPI communica- reconstruction code, GPEC, was presented
15
resolution
(Nr , Nz , N+1)
cores (MPI tasks
· threads/task)
RHS update
Poisson solver
RM evaluation
RM gather∗
lin. regression
ψ sum∗∗
other
total
(32, 64, 8)
(32, 64, 16)
(64, 128, 8)
(64, 128, 16)
48(8 · 6)
96(16 · 6)
48(8 · 6)
96(16 · 6)
T [ms]
0.022
0.036
0.024
0.015
0.013
0.048
0.008
0.167
%
13.2
21.6
14.5
8.9
7.9
28.9
5.0
100.0
T [ms]
0.021
0.036
0.024
0.021
0.017
0.056
0.008
0.182
%
11.7
19.5
13.3
11.4
9.5
30.5
4.2
100.0
T [ms]
0.050
0.085
0.025
0.019
0.014
0.110
0.011
0.319
%
15.7
26.6
7.8
5.9
4.4
36.2
3.4
100.0
T [ms]
0.053
0.086
0.025
0.029
0.018
0.128
0.014
0.353
%
15.0
24.5
7.1
8.1
5.2
36.2
3.9
100.0
TABLE II. Runtime breakdown of a single Picard equilibrium-iteration step (average over 4 iteration
steps times 1000 time points) for different combinations (subset of Table I) of numerical resolution and
computational resources. The individual algorithmic steps from top to bottom correspond to the sequence
in Sect. II.A. Routines labelled with asterisks are dominated by MPI communication (∗ MPI Allgather,
∗∗
MPI Allreduce).
which is suitable for real-time applications in
medium-sized tokamaks like ASDEX Upgrade
(AUG). GPEC implements the classical concept of iteratively solving the Grad-Shafranov
equation and feeding in diagnostic signals
from the experiment 7,8 . Specifically, GPEC is
implemented as a variant of the IDE code 19 ,
which is a descendant of CLISTE 8 . Compared
with these well-established and validated offlineanalysis codes no algorithmic simplifications
are necessary for achieving the desired cycle
times of less than a millisecond, besides limiting
the number of equilibrium iterations to four.
In addition to real-time applications the new
code enables fast and highly accurate offline
analyses soon after the tokamak discharge:
Here, tolerable runtimes are in the range of
0.1–1 s per cycle which allows computing fully
converged equilibria employing highly resolved
spatial and basis functions grids.
a fast shared-memory-parallel Grad-Shafranov
solver 18 together with the MPI-distributed solution of the individual Poisson-type problems and
a thorough parallelization of the post-processing
algorithms for computing the relevant plasmacontrol parameters from the equilibrium flux distribution.
Using data from a typical AUG discharge the
real-time capability of the new code was demonstrated by the offline computation of a sequence
of 1000 time points within less than a second of
runtime. The relative accuracy was ascertained
by comparing the relevant plasma parameters
with a converged run. By allowing four iterations for computing the equilibrium solution the
majority of control parameters can be computed
with an accuracy of a percent or better.
The adopted two-level, hybrid parallelization
scheme allow efficient utilization available compute resources (in terms of the numbers of
The parallelization of the new code builds on nodes, of CPU sockets per node and of cores
16
Fig. 5. Runtime of the major groups (A-D) of computations for the plasma-control parameters (cf.
Sect. II.B) and their assignment to four MPI tasks which execute in parallel. Within each task the individual parts of the algorithm (indicated by different colours) execute in the order from bottom to top.
Accordingly, the dashed horizontal line marks the effective runtime in the application. The data corresponds
to AUG shot #23221 computed with a resolution of (Nr , Nz , N +1) = (32, 64, 8) using in total 8 MPI tasks,
each with 6 OpenMP threads (cf. Table I).
per socket) for a given numerical resolution (in
terms of the spatial grid and number of basis
functions). Moreover, foreseeable advances in
computer technology, in particular the increasing number of CPU cores per compute node, are
expected to push the limits of real-time applications with GPEC towards even higher numerical
accuracies (in terms of affordable resolution and
number of equilibrium iterations). We note that
the benchmarks figures reported in this work can
be considered rather conservative as the computations were performed on a standard com-
pute cluster which comprises hundreds of nodes.
For energy-budget reasons such clusters are typically configured not with CPUs of the highest
clock-frequency. In our case, for example, CPUs
with 2.5 GHz and only 24 cores per node were
available. For deployment in the control system of a tokamak experiment such as AUG, by
contrast, we envision dedicated server hardware
with higher clock frequencies and at least four
CPU sockets, both of which is expected to further boost the computational performance of the
real-time application. Thus it should be rather
17
straightforward to save enough time for the communication of the code with the control system,
which, depending of the specifics of the system,
requires another few hundred microseconds per
cycle.
Finally, it is worth mentioning that GPEC
is based entirely on open-source software components, relies on established industry (FORTRAN, MPI, OpenMP) or de-facto software
standards (BLAS, LAPACK, FFTW), runs on
standard server hardware and software environments and hence can be released to the community and utilized without legal or commercial
restrictions.
ACKNOWLEDGEMENT
We acknowledge stimulating discussions with K.
Lackner and W. Treutterer. Thanks to P. Martin who has provided a subversion of the CLISTE
code with which this study started and to S. Gori
who has developed an initial FORTRAN 90version.
This work has been carried out within the framework of the EUROfusion Consortium and has
received funding from the Euratom research
and training programme 2014-2018 under grant
agreement No 633053. The views and opinions
expressed herein do not necessarily reflect those
of the European Commission.
18
REFERENCES
[1] C. GORMEZANO, A. SIPS, T. LUCE,
S. IDE, A. BECOULET, X. LITAUDON,
A. ISAYAMA, J. HOBIRK, M. WADE,
T. OIKAWA, R. PRATER, A. ZVONKOV,
B. LLOYD, T. SUZUKI, E. BARBATO, P. BONOLI, C. PHILLIPS,
V. VDOVIN, E. JOFFRIN, T. CASPER,
J. FERRON, D. MAZON, D. MOREAU,
R. BUNDY, C. KESSEL, A. FUKUYAMA,
N. HAYASHI, F. IMBEAUX, M. MURAKAMI, A. POLEVOI, & H. S. JOHN.
Chapter 6: Steady state operation. Nuclear
Fusion, 47(6), S285 (2007).
[2] Y. GRIBOV, D. HUMPHREYS, K. KAJIWARA, E. LAZARUS, J. LISTER,
T. OZEKI, A. PORTONE, M. SHIMADA,
A. SIPS, & J. WESLEY.
Chapter 8:
Plasma operation and control. Nuclear Fusion, 47(6), S385 (2007).
[3] M. MARASCHEK, G. GANTENBEIN,
T. P. GOODMAN, S. GÜNTER, D. F.
HOWELL, F. LEUTERER, A. MÜCK,
O. SAUTER, & H. ZOHM. Active control
of MHD instabilities by ECCD in ASDEX
upgrade. Nuclear Fusion, 45, 1369 (2005).
[4] L.
GIANNONE,
M.
REICH,
M.
MARASCHEK,
C.
RAPSON,
W. TREUTTERER, L. BARRERA,
E. POLI, A. BOCK, G. CONWAY,
F. FISCHER, J. FUCHS, K. LACKNER, P. MCCARTHY, A. MLYNEK,
19
R. PREUSS, K.-H. SCHUHBECK, J. STOBER, M. RAMPP, R. MCDERMOTT,
Q. RUAN, L. WENZEL, & ASDEX
UPGRADE TEAM. A data acquisition
system for real-time magnetic equilibrium
reconstruction on ASDEX Upgrade and its
application to NTM stabilization experiments. Fusion Engineering and Design, 88,
3299 (2013).
[5] V. D. SHAFRANOV. About MHD equilibrium configurations. Zh. Eksp. Teor. Fiz.,
33, 710 (1957).
[6] R. LÜST & A. SCHLÜTER. Axisymmetrische magnetohydrodynamische Gleichgewichtskonfigurationen.
Z. Naturforsch., 129, 850 (1957).
[7] L. L. LAO, J. R. FERRON, R. J. GROEBNER, W. HOWL, H. ST. JOHN, E. J.
STRAIT, & T. S. TAYLOR. Equilibrium
analysis of current profiles in tokamaks. Nuclear Fusion, 30, 1035 (1990).
[8] P. J. MCCARTHY, P. MARTIN, &
W. SCHNEIDER. The CLISTE interpretive equilibrium code.
Tech. Rep.
IPP 5/85, Max-Planck-Institut für Plasmaphysik (1999).
[9] L. LAO, H. S. JOHN, R. STAMBAUGH,
A. KELLMAN, & W. PFEIFFER. Reconstruction of current profile parameters and
plasma shapes in tokamaks. Nuclear Fusion, [14] A. BARP, M. CERNA, S. CONCEZZI,
25(11), 1611 (1985).
L. GIANNONE, G. MORROW, Q. RUAN,
A. VEERAMANI, & L. WENZEL. A
[10] J. BLUM, C. BOULBE, & B. FAUGERAS.
real-time Grad-Shafranov PDE solver and
Reconstruction of the equilibrium of the
MIMO controller. Fusion Engineering and
plasma in a tokamak and identification of
Design, 87(12), 2112 (2012). Proceedings
the current density profile in real time.
of the 8th IAEA Technical Meeting on ConJournal of Computational Physics, 231(3),
trol, Data Acquisition, and Remote Partic960 (2012).
ipation for Fusion Research.
[11] B. BRAAMS, W. JILGE, & K. LACKNER. [15] J.-M. MORET, B. DUVAL, H. LE,
Fast determination of plasma parameters
S. CODA, F. FELICI, & H. REIMERDES.
through function parametrization. Nuclear
Tokamak equilibrium reconstruction code
Fusion, 26(6), 699 (1986).
LIUQE and its real time implementation.
Fusion Engineering and Design, 91, 1
[12] L. GIANNONE, M. CERNA, R. COLE,
(2015).
M. FITZEK, A. KALLENBACH, K. LDDECKE, P. MCCARTHY, A. SCARA- [16] J. FERRON, M. WALKER, L. LAO, H. S.
BOSIO, W. SCHNEIDER, A. SIPS,
JOHN, D. HUMPHREYS, & J. LEUER.
W. TREUTTERER, A. VRANCIC,
Real time equilibrium reconstruction for
L. WENZEL, H. YI, K. BEHLER,
tokamak discharge control. Nuclear Fusion,
T. EICH, H. EIXENBERGER, J. FUCHS,
38, 1055 (1998).
G. HAAS, G. LEXA, M. MARQUARDT,
A. MLYNEK, G. NEU, G. RAUPP, M. RE- [17] I. YONEKAWA, A. V. FERNANDEZ, J.ICH, J. SACHTLEBEN, K. SCHUHBECK,
M. FOURNERON, J.-Y. JOURNEAUX,
T. ZEHETBAUER, S. CONCEZZI, T. DEW.-D. KLOTZ, A. WALLANDER, & COBELLE, B. MARKER, M. MUNROE,
DAC TEAM. ITER instrumentation and
N. PETERSEN, & D. SCHMIDT. Data
control system towards long pulse operaacquisition and real-time signal processing
tion. Plasma Fusion Research, 7, 2505047
of plasma diagnostics on ASDEX upgrade
(2012).
using labview RT. Fusion Engineering and
Design, 85(34), 303 (2010). Proceedings [18] M. RAMPP, R. PREUSS, R. FISCHER,
K. HALLATSCHEK, & L. GIANNONE. A
of the 7th IAEA Technical Meeting on
parallel Grad-Shafranov solver for real-time
Control, Data Acquisition, and Remote
control of tokamak plasmas. Fusion Science
Participation for Fusion Research.
and Technology, 62(3), 409 (2012).
[13] X. N. YUE, B. J. XIAO, Z. P. LUO, &
Y. GUO. Fast equilibrium reconstruction [19] R. FISCHER, J. HOBIRK, L. B. ORTE,
A. BOCK, A. BURCKHART, I. CLASSEN,
for tokamak discharge control based on
M. DUNNE, J. FUCHS, L. GIANGPU. Plasma Physics and Controlled FuNONE, K. LACKNER, P. MCCARTHY,
sion, 55(8), 085016 (2013).
20
R. PREUSS, M. RAMPP, S. RATHGEBER, M. REICH, B. SIEGLIN, W. SUTTROP, E. WOLFRUM, & ASDEX
UPGRADE TEAM. Magnetic equilibrium
reconstruction using geometric information
from temperature measurements at asdex
upgrade. In V. NAULIN, C. ANGIONI,
M. BORGHESI, S. RATYNSKAIA,
S. POEDTS, & T. DONN (eds.), 40th EPS
Conference on Plasma Physics, vol. 37D
of Europhysics conference abstracts, page
P2.139. European Physical Society, Geneva
(2013).
[25] R. PREUSS, R. FISCHER, M. RAMPP,
K. HALLATSCHEK, U. VON TOUSSAINT, L. GIANNONE, & P. MCCARTHY. Parallel equilibrium algorithm
for real-time control of tokamak plasmas.
Tech. Rep. IPP R/47, Max-Planck-Institut
für Plasmaphysik (2012).
[26] http://www.mpi-forum.org/.
[27] W. M. STACEY. Fusion Plasma Physics.
Wiley, 2 edn. (2012).
[28] http://www.openblas.net.
[20] P. J. MCCARTHY & ASDEX UPGRADE [29] BLAST FORUM. Basic linear algebra subprograms technical forum standard. InTEAM. Identification of edge-localized moternational Journal of High Performance
ments of the current density profile in a
Applications and Supercomputing, 16(1)
tokamak equilibrium from external mag(2002).
netic measurements. Plasma Physics and
Controlled Fusion, 54(1), 015010 (2012).
[30] http://www.netlib.org.
[21] A. E. HOERL & R. W. KENNARD. [31] E. ANDERSON, Z. BAI, C. BISCHOF,
Ridge regression: Biased estimation for
S. BLACKFORD, J. DEMMEL, J. DONnonorthogonal problems. Technometrics,
GARRA, J. DU CROZ, A. GREENBAUM,
12(1), 55 (1970).
S. HAMMARLING, A. MCKENNEY, &
D. SORENSEN. LAPACK Users’ Guide.
[22] W. H. PRESS, S. A. TEUKOLSKY, W. T.
Society for Industrial and Applied MatheVETTERLING, & B. P. FLANNERY. Numatics, Philadelphia, PA, third edn. (1999).
merical Recipes: The Art of Scientific ComISBN 0-89871-447-8 (paperback).
puting. Cambridge University Press, 3rd
edn. (2007).
[32] M. FRIGO & S. G. JOHNSON. The design
[23] K. V. HAGENOW & K. LACKNER. Computation of axisymmetric MHD equilibria.
In Proc. 7th Conf. on the Numerical Simulation of Plasmas, page 140. New York, US
(1975).
[33]
and implementation of FFTW3. Proceedings of the IEEE, 93(2), 216 (2005). Special
issue on “Program Generation, Optimization, and Platform Adaptation”.
https://gcc.gnu.org/.
[24] K. LACKNER. Computation of ideal MHD [34] http://www.open-mpi.de/.
equilibria. Comp. Phys. Comm., 12, 33
[35] http://mvapich.cse.ohio-state.edu/.
(1976).
21
| 5cs.CE
|
arXiv:1612.06749v1 [cs.DC] 20 Dec 2016
GuStL – An Experimental Guarded States
Language
Oskar Schirmer
Göttingen, 2016-12-20
Abstract
Programming a parallel computing system that consists of several thousands
or even up to a million message passing processing units may ask for a language that supports waiting for and sending messages over hardware channels.
As programs are looked upon as state machines, the language provides
syntax to implement a main event driven loop.
The language presented herewith surely will not serve as a generic programming language for any arbitrary task. Its main purpose is to allow for
a prototypical implementation of a dynamic software system as a proof of
concept.
1
Contents
1 Introduction
3
2 Syntax
3
3 Semantics
5
4 Process Creation
8
5 Segmented Program Development
9
6 Compiler
10
7 Discussion
10
2
1
Introduction
The Guarded States Language essentially does combine well known concepts
of existing languages, while any concept not needed for its main purpose is
avoided. Its basic concept is imperative sequential operations. Syntactically
it is inspired by Algol ([1963nb]) and Pascal ([1975jw]). Main program
execution is organized as a state machine as with SDL ([2007ir]). States
are subject to guards, similar to those introduced in [1975ed]. For channel
operations, input and output, the notation follows [1978ch].
A software system may consist of various processes running in parallel,
each of which is a program executed sequentially, one instruction at a time.
Channels may be set up dynamically to allow transmission of data, i.e. messages, from one process to the other.
Depending on the current process state, different guards may be evaluated,
causing some statement sequence to be executed as soon as the condition of
the corresponding guard is met.
Statements serve to implement traditional imperative operations, such as
assigments, loops, conditional execution and subroutine invocation.
The language does only provide features restricted to the processing unit
the program is executed on. As a consequence, and as dynamic process creation does imply activity of various system processes – especially in the case
of creating a remote process –, the language does not provide syntactic means
to describe instantiation of a new process.
2
Syntax
A program is made up of a sequence of symbols. A symbol may be one of:
special symbol, reserved word, identifier, number, or string.
Symbols may be separated by white space, i.e. blank, tabulator, or newline characters. Any character sequence starting with a left brace up to and
including the next right brace is considered white space, too, as long as it
does not start within a string. The latter allows insertion of comments. Note,
that nested comments are not valid.
For the special symbols and reserved words, see the formal syntax definition
below. Reserved words are to be written in lower case.
An identifier is composed of a sequence of upper and lower case letters,
digits, and the underline character. It must not start with a digit. Identifiers
are case sensitive.
Numbers are given in either decimal, hexadecimal, or character notation.
For decimal notation a number is written as a sequence of digits, for hexadec3
imal notation, it is preceded by 0x. A number given in character notation
is the characters implicit textual representation1 enclosed by single quotes.
Scientific notation is not supported.
A string is a series of characters – i.e. their implicit textual representations
– enclosed by double quotes. To include a double quote in a string, it is
duplicated. A string is encoded as an array at one word per character.
A valid sentence for a program must match the following partially formalised syntax description.2
process ::= { constdef | wordsdecl | proceduredecl | f unctiondecl }
“process” ident!process “(” ident!port [ “,” ident!const ] “)”
[ statesdecl ] [ portsdecl ]
{ constdef | wordsdecl | proceduredecl | f unctiondecl }
“start” { statement } { guardedstate } “stop”
statesdecl ::= “state” ident!state { “,” ident!state }
portsdecl ::= “port” ident!port { “,” ident!port }
constdef ::= “const” ( ident!const “=” expression | ident!constarray
“[” [ expression ] “]” “=” constarray { “,” constarray } )
constarray ::= expression | “"” chars “"”
wordsdecl ::= “word” worddecl { “,” worddecl }
worddecl ::= ident!word | ident!array “[” expression “]”
proceduredecl ::= “procedure” ident!procedure “(” f ormalparameters “)”
{ wordsdecl } “do” { statement } “return”
f unctiondecl ::= “function” ident!f unction “(” f ormalparameters “)”
{ wordsdecl } “do” { statement } “return” expression
f ormalparameters ::= [ f ormalparameter { “,” f ormalparameter } ]
f ormalparameter ::= “const” ident!constarray “[” [ expression ] “]” |
ident!array “[” [ expression ] “]” | ident!word | “port” ident!port
guardedstate ::= “on” statelist [ “\” guard ] “:” { statement }
statelist ::= ident?state { “,” ident?state }
guard ::= ident?port “!” | reception | expiration
statement ::= conditional | repetition | transition | assignment |
transmission | procedurecall | inlineasm
1
It is assumed that the usual implicit textual representation for characters is its Unicode code point as defined in [2009uc]
2
For reasons of simplicity, it is given in a somewhat modified version of the notation
inspired by Pān.ini ([1840ob]) and initially used by Backus ([1959jb]). To achieve a
fully formalised definition, including semantics, one would need to describe the language
by using, e.g., an extended affix grammar ([1997dw]). The scope of identifiers is hinted
with a subscript !scope , where identifiers are declared, and with a subscript ?scope , where
they are assumed to be already declared in the scope given
4
conditional ::= “if” expression “then” { statement }
{ “elseif” expression “then” { statement } }
[ “else” { statement } ] “done”
repetition ::= [ “while” expression ] “repeat” expression “times”
{ statement } ( “done” | “until” expression )
transition ::= “next” ident?state
assignment ::= variable “:=” ( expression | inlineasm )
transmission ::= ident?port “!” ( expression | “end” | “pause” )
reception ::= ident?port “?” ( variable | “end” )
variable ::= ident?word | ident?array “[” expression “]” | ident?port
expression ::= simpleexpr
[ ( “=” | “<>” | ( “<” | “>” | “<=” | “>=” ) [ “$” ] ) simpleexpr ]
simpleexpr ::= term { ( “+” | “-” | “|” | “^” | “or” ) term }
term ::= f actor
{ ( “*” | “&” | “and” | “<<” | ( “/” | “%” | “>>” ) [ “$” ] ) f actor }
f actor ::= [ unaryop ] ( variable | constant | f unctioncall | “now”
| reception | “(” expression “)” )
unaryop ::= “-” | “~” | “not”
constant ::= number | “’” char “’” | “#” ( ident?array | ident?constarray ) |
ident?const | ident?constarray “[” expression “]”
procedurecall ::= ident?procedure “(” actualparameters “)”
f unctioncall ::= ident?f unction “(” actualparameters “)”
actualparameters ::= [ actualparameter { “,” actualparameter } ]
actualparameter ::= expression |
( ident?array | ident?constarray ) [ “[” expression “]” ]
expiration ::= “after” expression
inlineasm ::= “asm” singleasm { “,” singleasm }
singleasm ::= [ unaryop ] constant | “(” expression “)”
3
Semantics
The process is a state machine, based upon an arbitrary number of states.
As a special case, the number of states may be zero. In any other case, all
state identifiers must be declared next to the process header.
To allow external communication, a process may want to transmit or receive data over channels. To set up a channel at runtime, it must be dynamically assigned to a port, which in turn is static. Any port is referred to by
an identifier declared before being used.
Constant identifiers may be declared to stand for constant words or con5
stant arrays. The value of a constant array is given as a list of numbers,
where all or parts of the list may be replaced by a string.
The basic unit of data to operate on is a word. It represents a natural
or integer number, which depends on which operation is performed on it.
Different basic types to restrict specific operations or to determine details
of an operation – like signedness – do not exist.3 A variables identifier may
either be declared to refer to a single word, or, when declared as an array, to
refer to a sequence of words. Variables must be declared prior to being used.
Subroutines – both functions and procedures – may be declared to implement sequences of statements for multiple use in the program. Upon termination, functions provide a word return value to be used in the invocating
expression, while procedures do not return any value. Subroutines may be
declared to take a list of parameters, which may be words or arrays or ports.
Furthermore, subroutines allow declaring local single word variables, which
are instantiated at subroutine invocation and cease at the end of the subroutines execution. Subroutines must be declared prior to being used.
Statements in the program or in a subroutine are executed sequentially,
one by one. For a conditional statement, a series of expressions is evaluated, until one is met – i.e. the expression evaluates to non-zero – and the
corresponding list of statements is executed. If no condition is met and an
alternative branch is given with else, that alternative branches statements
are executed. A repetition statement is executed for a maximum number of
iterations. Furthermore, it is not iterated or not executed at all, when the
initial condition is not met, and it is no longer iterated when the terminating
condition is met.
The right side of an assignment is evaluated and the result is assigned to
the word or port on the left side. A value assigned to a port takes it for a
destination port number and prepares a channel to be set up from the local
to the destination port.
For channel data transmission, the port to communicate over is given on
the left side of the statement. The data to transmit is given on the right side,
this may either be an arbitrary value, or a special token to end or pause a
message.
Channel data reception is not a statement, but a factor in an expression.
As with transmission, the port to receive data from is given on the left side,
while on the right side is either a variable to store data into or the end
symbol. Both variants consume an incoming end token when it is available,
3
Various programming languages define signedness as a property of each single data
unit. This way, wherever data units of different type are subject to an operation, implicit
conversion rules need to be applied, an everlasting source of confusion. To avoid this mess,
GuStL makes signedness be a property of the operation itself.
6
but if data is pending, it is only consumed in the first case and stored into the
variable. Same as with the assignment, storing the newly received word to
a port sets the ports destination. The factor evaluates to non-zero whenever
the expected token is received, which is data for the first variant and an end
token for the second. Otherwise, it evaluates to zero.
Calling a subroutine implies evaluation of the list of actual parameters,
which must match the list of formal parameters given in the subroutines
declaration. For formal word parameters, the corresponding expression is
evaluated and its result used in the subroutine (call by value). For formal
array parameters, the reference to the corresponding variable is determined
and used in the subroutine (call by reference). For formal port parameters,
the given port identification is used in the subroutine (call by reference). For
a subroutine with more than one parameter the order of parameter evaluation
is not defined.
The transition statement is used to change the machines state. It must
not be used in a subroutine. Additionally, every execution path in the main
program must end in a transition statement.
The main program starts with a list of statements for initialisation purposes, and ends with a list of statements to handle proper process termination. In between, an arbitrary number of guarded statements is provided.
Whenever the initialising list of statements or another guarded statement of
the program has finished executing the next guarded statement may be selected for execution. Only those guarded statements, that match the current
state, are considered. From these, one random guarded statement is selected,
for which the condition of its guard is met. When no condition of any guard
is met at all, the process is stalled until at least one condition is met.
A guard does either check a port for transmission readiness, for availability of data to receive, for an end token received on a port, or for a given
time being expired. When data is available and the corresponding guarded
statement is selected for execution, the data received is actually stored into
the variable given. If for a state a guard is declared to check for availability
of input data, another guard to check for an end token on the same port
is obligatory for the same state. For one state several different guards may
be declared, provided they do not check for the same condition on the same
port, and that no two guards check for expiration.
By default, when evaluating an expression, all operations are performed as
unsigned. For various operators, however, appending a $ sign to the operator
itself may be used to ask for signed operation. Signed division is defined to
be Euclidean (see e.g. [2001dl]). Any expression is composed from one or
two simple expressions, simple expressions are composed from terms, terms
are composed from factors. A factor may be a single word or a word from
7
an array, a constant number or an element from a constant array, the size
of an array, the call of a function – using its return value, the current time
(now), the result of checking an input port for having received an end token,
or another expression. Any factor may be prepended by a unary operator.
The order in which parts of an expression are evaluated is not defined.
To allow machine specific operations, inline assembly is allowed both as a
statement and as the right side of an assignment. Inline assembly is given as
a list of opcodes, but may include also ordinary expressions to be evaluated.
Opcodes depend on the implementation, see section 6. An expression included
in inline assembly will leave the corresponding result on the data stack.
4
Process Creation
For traditional single node computers according to the von-Neumann architecture ([1945jn]), the basic resource to be allocated by processes at runtime
is memory, taken from a uniform pool of contiguous memory words.
For machines made up of large amounts of parallel processing units, each
with its own local memory, the allocatable quantity is a single processing
unit. Thus, allocating resource does no longer mean to reserve a portion of
memory, but to start a process on a previously idle processing unit: process
creation.
The program to run on the allocated processing unit has to be identified,
e.g. by name, which in turn may be variable, and does not refer to data held
inside the current program, so for the language to natively support process
creation a hypothetical rule could look like the following – where the identifier would refer to a string denoting the new processes name:
creation ::= “new” ident?array “(” expression “)”
However, depending on the hardware, starting a process may highly depend
on supportive software, i.e. the operating system. The compiler would need
to produce operating system dependent code for process creation. To avoid
complexity and increase flexibility, means for process creation shall instead
be provided by the operating system as an ordinary subroutine. An outline
on how it could be declared is given hereafter.
Once the new process is being executed, the process that started it may
need to communicate with it. To do so, it needs to know at least one port
of the new process to attach to, the control port. The subroutine for process
creation will return its global identification, so it must be a function. For
the new process, the port is declared as the first parameter to the process
8
declaration header.
As the processing unit and its local memory are – once allocated – an
indivisible unit, the amount of memory available during lifetime of the process
cannot be changed. Only at the time of creation of the process, the amount
of memory given to it may be chosen. For this purpose, the subroutine to
create the new process not only needs to know its name, but also a number,
which will be available inside the new process as a constant, the dimension.
To make it usable in a running program, it is declared to be a constant as
the second parameter to the process header. It is optional and defaults to
zero.
Hardware driver processes may need to access specific hardware locations,
so it is crucial to start such a process on a given processing unit. For other
reasons, it may be desirable to choose the processing unit from a restricted
subset of all units available, so some flags might possibly be helpful. An
additional parameter to the function starting the process is used to commit
the information needed. The function may be declared as follows:
function new(name[], dimension, extra)
5
Segmented Program Development
When writing a new program, parts of it may be similar or even identical
to portions of a previously implemented program. The wish to reuse these
portions and avoid repeated work should represent one of the basic desires
of every programmer.
The most simple solution would be to simply copy and paste the portion
in question to the new program. While in earlier days this was frowned upon
for the reason of wasting storage, today the untraceable propagation of programming mistakes is a major concern.
Separate compilation units had the advantage to circumvent the need to
recompile all of the code used in a program by keeping the program divided
into portions even in its compiled state. This was useful in times of slow
compiling machines, but bears the risk of binary incompatibilities when interfaces are changed in source files. Further, an additional linker program or
a linking loader is needed to enable program execution.
Assembly of the program at compile time may be achieved using include
files, a method to splice a set of source files.4 But it may also lead to confusion
of which include file is used where, especially when using nested inclusions.
4
Using C ([1978kr]) usually involves a combination of both linkable program portions,
called libraries, and include files for interface declaration
9
To avoid the latter problem, source file composition shall be restricted to
simple concatenation. In fact, neither the language nor the compiler need to
take any provisions for it, as the concatenation may well be performed by the
invoking entity before feeding the complete source program into the compiler
itself.
The fact that insertion of one source file into another is impossible makes it
necessary to allow declarations before the process declaration of a program,
as generic declarations need to be prepended to the main program.
6
Compiler
A first compiler for GuStL has been implemented to produce code for the
experimental machine NOP ([2016os]). The compiler itself is implemented
in C, running in a Unix environment. Its single pass design follows the basic
concept of [1977nw]. It compiles source from standard input to binary code
on standard output. Matching the target machine, words are implemented as
32 bit wide, least significant byte first, and signed numbers are represented
in two’s complement. The implicit textual representation for characters is
its Unicode code point as defined in [2009uc], encoded using UTF-8 in the
source file (see also [2003fy]).
The binary code itself is preceded by a six word header to allow direct use
in an operating system environment:
magic
flags
d1
d0
code
entry
7
0x85cf80cf
0
data memory size: ld1 -ld0 = d1 · dimension + d0 , for
dimension see page 9, see also [2016os]
code and constants memory size: lc1 -lc0 , see [2016os]
word offset to initial instruction, relative to lc0 , see [2016os]
Discussion
The main purpose of the programming language GuStL is to allow for a
prototypical implementation of a dynamic software system, and its design is
restricted to features that are needed to serve this purpose.
There is only very limited support for string handling, namely definition
of constant arrays. It might be useful to provide more support for string
handling, e.g. transmission and reception of sequences instead of a single
word. However, the latter might reduce program flow determinism, because
time behavior for the processing of a sequence potentially is more complex
10
than for a single word, scattering non-determinism from controlling guards
into imperative program flow.
Direct recursion in a subroutine is supported, but recursion involving several subroutines is not, as there is no forward declaration. With local memory
being quite limited, it is arguable whether system abstraction is adequate to
justify support of massive recursion.
Likewise, there is no specific support for floating point calculations, list
processing, bit field computation or set arithmetics5 , data pointers or complex
data structures. Where this is perceived as a deficiency, it might be worth
considering the choice or design of a different language that better fits the
particular project context.
Execution of code in parallel may be handled on different levels of a system. GuStL is designed to implement a dynamic software system, and thus
it provides parallel execution on process level, where every process start requires to dynamically determine a free resource, the loading of code, and
setup of control communication. For static software systems, e.g. embedded
control, things are different in that all parts of the software needed at runtime are known prior to system start, i.e. at compile time. Consequently,
dynamically loading code is not necessary and even communication layout
between processes may be arranged at compile time. There are examples
that allow much more efficient instantiation of parallelity in a static software
system (e.g. [2010hh]), but the question of how to apply these techniques
to a dynamic software system remains open.
Directly addressing memory is not needed even in the context of an operating system, as the intended target architecture provides peripheral access
through dedicated operations instead of memory mapped control registers,
see [2016os].
Literature
[1840ob] Otto Böhtlingk (Ed.): “Pān.ini’s Grammatik”, Bonn 1840, Delhi
2001, ISBN 81-208-1025-2
[1945jn] John von Neumann: “First Draft of a Report on the EDVAC”,
Moore School of Electrical Engineering, University of Pennsylvania,
June 30, 1945
[1959jb] J. W. Backus: “The syntax and semantics of the proposed international algebraic language of the Zürich ACM-GAMM conference”,
Proc. Internat. Conf. Inf. Proc., UNESCO, Paris, June 1959
5
as defined for Pascal
11
[1963nb] J. W. Backus, F. L. Bauer, J. Green, C. Katz, J. McCarthy, A.
J. Perlis, H. Rutishauser, K. Samelson, B. Vauquois, J. H. Wegstein,
A. van Wijngaarden, M. Woodger, P. Naur (Ed.): “Revised Report
on the algorithmic language ALGOL 60”, Numerische Mathematik 4,
p. 420-453 (1963)
[1975ed] Edsger W. Dijkstra: “Guarded commands, nondeterminacy, and
formal derivation of programs”, Comm. ACM 18, 8 (Aug. 1975), p. 453457, EWD472
[1975jw] Kathleen Jensen, Niklaus Wirth: “Pascal: user manual and report”, Springer-Verlag, 1975
[1977nw] Niklaus Wirth: “Compilerbau”, Teubner, Stuttgart, ISBN 3-51902338-5, 1977
[1978ch] C. A. R. Hoare: “Communicating sequential processes”, The
Queen’s University, Belfast, Northern Ireland Commun. ACM 21, 8
(Aug. 1978), p. 666-677
[1978kr] Brian W. Kernighan, Dennis M. Ritchie: “The C Programming
Language (1st ed.)”, February 1978, Englewood Cliffs, NJ: Prentice
Hall, ISBN 0-13-110163-3
[1997dw] Jochen Demuth, Stephan Weber, Sönke Kannapinn, Mario Kröplin:
“Echte Compilergenerierung – Effiziente Implementierung einer abgeschlossenen Theorie”, Forschungsbericht des Fachbereichs Informatik
1997/6, Technische Universität Berlin
[2001dl] Daan Leijen: “Division and Modulus for Computer Scientists”,
University of Utrecht, Dept. of Computer Science, December 3, 2001
[2003fy] F. Yergeau: “Request for Comments: 3629”, Network Working
Group, Alis Technologies, November 2003
[2007ir] ITU-T Recommendation Z.100: “Specification and Description
Language (SDL)”, International Telecommunication Union, 11/2007
[2009uc] The Unicode Consortium: “The Unicode Standard, Version 5.2.0”,
Mountain View, CA, 2009, ISBN 978-1-936213-00-9
[2010hh] James Hanlon, Simon J. Hollis: “Dynamic generation of parallel
computations”, Department of Computer Science, University of Bristol, UK, in proc. UK Electronics Forum 2010, (pp. 7-17), 2010
[2016os] Oskar Schirmer: “NOP – A Simple Experimental Processor for
Parallel Deployment”, Göttingen, 2016
12
| 6cs.PL
|
Mining a Sub-Matrix of Maximal Sum
Vincent Branders, Pierre Schaus, and Pierre Dupont
arXiv:1709.08461v1 [stat.ML] 25 Sep 2017
Université catholique de Louvain - ICTEAM/INGI - Machine Learning Group
Place Sainte Barbe 2, 1348 Louvain-la-Neuve - Belgium
[email protected]
Abstract. Biclustering techniques have been widely used to identify
homogeneous subgroups within large data matrices, such as subsets of
genes similarly expressed across subsets of patients. Mining a max-sum
sub-matrix is a related but distinct problem for which one looks for a
(non-necessarily contiguous) rectangular sub-matrix with a maximal sum
of its entries. Le Van et al. [6] already illustrated its applicability to gene
expression analysis and addressed it with a constraint programming (CP)
approach combined with large neighborhood search (CP-LNS). In this
work, we exhibit some key properties of this N P-hard problem and define
a bounding function such that larger problems can be solved in reasonable time. Two different algorithms are proposed in order to exploit the
highlighted characteristics of the problem: a CP approach with a global
constraint (CPGC) and mixed integer linear programming (MILP). Practical experiments conducted both on synthetic and real gene expression
data exhibit the characteristics of these approaches and their relative
benefits over the original CP-LNS method. Overall, the CPGC approach
tends to be the fastest to produce a good solution. Yet, the MILP formulation is arguably the easiest to formulate and can also be competitive.
1
Introduction
Gene expression data is typically represented as a large matrix of gene expression levels across various samples. The study of such data is a valuable tool to
improve the understanding of the underlying biological processes. For example,
biomarker discovery aims at finding indicators of a disease or the physiological state of patients. This problem can be addressed with clustering techniques
which perform a grouping of one dimension, either the rows or the columns of
the original matrix. Yet it is known that breast cancer, for example, exhibits distinct subtypes [11,12]. In other words, specific genes exhibit activation patterns
only in a specific group of patients. Biclustering techniques, or co-clustering,
identify specific subsets of rows and of columns which jointly form homogeneous
entries [8].
In the present work, we focus on a related but different mining task. One
looks in particular for subsets of rows and of columns with globally high values.
In the context of gene expression analysis, the objective is to find a subset of
genes which are relatively highly expressed among a subset of patients, even
though some entries might depart from this pattern. Formally, one looks for a
Presented at NFMCP 2017, The 6th International Workshop on New Frontiers in Mining Complex
Patterns, Skopje, Macedonia, Sep 22, 2017.
rectangular, and non necessarily contiguous, sub-matrix of a large matrix with
a maximal sum of the selected entries.
This max-sum sub-matrix problem is closely related to the maximal ranked
tile mining problem studied by Le Van et al. [6]. In the later case, prior to
the search of a sub-matrix, each matrix entry is replaced by its rank across its
particular row. In other words, the maximal ranked tile mining is equivalent to
the max-sum sub-matrix for which the matrix entries are discrete ranks along the
rows. Since the combinatorial optimization algorithms to solve such problems are
actually not specific to discrete entries we address here the slightly more general
setting of continuous entries.
Our main contributions are 1) the study of the max-sum sub-matrix problem
while exhibiting some of its key properties and the definition of an upper bound
easy to compute in order to speed up the search for a solution; 2) the implementation of two additional algorithms making use of these properties : a CP
approach with a global constraint (CPGC) and mixed integer linear programming (MILP); 3) practical experiments conducted both on synthetic and real
gene expression data showing that the original CP-LNS method can be largely
outperformed; 4) the study of the relative benefits of the proposed methods
across various problem instances.
2
Problem
2.1
Statement
Definition 1 (The Max-Sum Sub-Matrix Problem). Given a matrix M ∈
IRm×n consisting of m rows and n columns, let R = {1, . . . , m} and C =
{1, . . . , n} be index sets for rows and for columns respectively, find the max-sum
sub-matrix (I ∗ , J ∗ ) , with I ∗ ⊆ R and J ∗ ⊆ C, such that:
(I ∗ , J ∗ ) = argmax f (I, J) = argmax
I⊆R,J⊆C
I⊆R,J⊆C
X
i∈I,j∈J
Mi,j .
(1)
The objective function f (I, J) is the sum of the entries of a sub-matrix (I, J).
The maximization term rewards, respectively penalizes, matrix entries with positive, respectively negative, values. The problem is interesting only for matrices
with positive and negative entries. Otherwise the optimal solution is trivially
identified.
The max-sum sub-matrix problem is N P-hard from a simple reduction1 of
the Maximum edge Weight Biclique Problem (MWBP) [3].
1
Essentially by considering the rows and columns of the matrix as the two sets of
nodes of a bipartite graph.
2.2
Properties
For a defined
P subset of columns J, the objective function can be formulated as
f (I, J) = i∈I ri with ri being the contribution of the row i:
X
ri =
Mi,j .
(2)
j∈J
This is important as the search can be limited to one dimension through
independent computation of the contributions along the other dimension. Indeed,
for any of the two dimensions being fixed, optimization along the other dimension
is trivial since it amounts to select only the subset of entries whose contribution
is positive. For a fixed subset of columns J ⊆ C, the subset of rows IJ∗ ⊆ R that
maximizes the objective value is greedily identified as:
X
IJ∗ = argmax
Mi,j = {i ∈ R | ri ≥ 0} .
(3)
I⊆R
i∈I,j∈J
In the gene expression analysis context with order(s) of magnitude more
rows (the genes) than columns (the samples), one typically search for a subset
of columns and greedily select the associated optimal subset of rows.
The search space of the max-sum sub-matrix problem contains 2|R| × 2|C|
feasible solutions that are rectangular sub-matrices (I, J) of the original matrix,
(I ⊆ R, J ⊆ C). Thanks to the independent contribution along one dimension,
the number of feasible solutions to be explicitly evaluated is thus reduced to
2|C| × |R|.
2.3
Related Work
Biclustering techniques address the problem of finding sub-matrices that satisfy
some definition of homogeneity since entries grouped together into biclusters
typically have similar values. A comprehensive review can be found in [8]. There
is no assumption of homogeneity in the max-sum sub-matrix problem but rather
one looks for a rectangular sub-matrix with an overall maximal sum.
The maximum (contiguous) subarray problem introduced in [2] identifies a subarray of maximal sum from an array. For a one-dimensional array, this problem
can be solved in linear time by Kadane’s algorithm [2]. Cubic and sub-cubic time
complexity algorithms have been proposed in the two-dimensional case [2,14,15].
This problem is however simpler than the max-sum sub-matrix since the selected sub-matrix is constrained to be formed of contiguous rows and contiguous
columns from the original matrix.
The maximum ranked tile mining problem has been introduced in [6]. As discussed in Sect. 1, this is a special case of the max-sum sub-matrix problem for
which the matrix entries are discrete ranks, corresponding to a permutation of
column indices on each row. Both problems are examples of subgroup discovery
which aims at finding subsets of a dataset that show an interesting behavior
with respect to a certain property of interest [1].
3
3.1
Optimization Approaches
Boolean Decision Vectors
The max-sum sub-matrix problem can be modeled with two vectors of boolean
decision variables: T = (T1 , . . . , Tm ) for the rows and U = (U1 , . . . , Un ) for the
columns with Ti ∈ {0, 1} and Uj ∈ {0, 1}. A sub-matrix (I, J) is defined by
I = {i ∈ R | Ti = 1} and J = {j ∈ C | Uj = 1}. The problem consists in
assigning a value to each variable of T and U .
Let us denote by U 1 = {j ∈ C | Uj = 1} the selected columns, U 0 = {j ∈
C | Uj = 0} the unselected ones and by U ? = {j ∈ C | Uj = {0, 1}} the undecided
ones.
3.2
Constraint Programming
Constraint Programming (CP) is a flexible programing paradigm that is capable
of solving optimization problems. As a declarative approach, it only requires to
model the problem and, by using existing solvers, let it search and find solutions. A model is defined as a constraint satisfaction problem CSP = (V, D, C)
where V is the set of variables, D is their respective domains, and C is a set of
constraints defined over the variables. A feasible solution is an assignment of the
variables to values of their domains such that all constraints are satisfied.
Constraints are exploited to iteratively reduce the domains of variables. Such
constraint propagation reduces the number of variable assignments to consider.
Once all unfeasible values are removed from the domains of variables, the fixpoint of the propagation is reached. Then the solver selects a variable X ∈ V
that is unbound and recursively calls the solver while assigning a value to this
variable. Through exploration of a depth-first-search tree (DFS), the solver either
reaches a solution or backtrack when the domain of variables becomes empty.
Efficient backtracking is achieved through trailing, which is a state management strategy that facilitates the restoration of the computation state to an
earlier version, effectively undoing changes that were imposed since then. Knuth
[5] attributes to Floyd [4] the first formulation of this mechanism and its usage
in a backtracking algorithm. The trail exposes two methods: pushState and
popState to respectively time-stamp the current state and restore it. Its implementation is captured in terms of two simple stacks. The first stack holds entries
to undo, the second one holds the frame sizes stacked between two consecutive
call to pushState. Trailing enables the design of reversible objects defined on
the trail.
CP-LNS. The max-sum sub-matrix problem can be modeled as:
X
maximize
Ti × ri , ∀i ∈ R : Ti = 1 ⇔ ri ≥ 0 .
(4)
i∈R
The objectiveP
is to maximize the sum of rows contributions which are formalized as ri =
j∈C Uj × Mi,j . Each row with positive (respectively negative)
contribution is constrained to be selected (respectively unselected). This provide
strong filtering during the search.
A speed up of the search, in the form of a large neighborhood search (LNS),
has also been proposed by Le Van et al. [6]. Through modification of the assignment of a large number of variables, LNS avoids being stuck in a region of the
search tree by restarting the exploration in other regions. This comes at the cost
of losing the possibility to prove optimality. In the present problem, half of the
column variables are uniformly selected for the search while leaving the others
unchanged.
CP Global Constraint. We propose to improve the model by Le Van et al. [6]
by designing an efficient algorithm encapsulated inside a global constraint that
captures the whole problem. The pseudo-code is given in Algorithm 1. The call
to the bounding and filtering procedure has been made explicit. In practice, the
lines 11 to 14 would be encapsulated in the global constraint triggered by the
fix-point algorithm. The key ingredients of our approach are:
– A feasible solution at each node of the search tree (evaluate()).
– An efficient bounding procedure (upperBound()).
– An efficient procedure to filter the domains (filter()).
Feasible solution at each node of the search tree. CP usually updates its feasible solution and best so far lower bound in the leaf-node of the search tree,
that is when every variable of the problem is bound. One can observe that for
the max-sum sub-matrix problem, any partial assignment of the variables can
be extended implicitly as a complete solution for which the unbound variables
would be set to 0 (i.e. U ? variables are considered unselected). There is thus
no need to wait that every variable is bound to evaluate the solution and possibly update the best so far lower bound. The value of the objective function
of the feasible solution is computed in the evaluate() method as the sum of
the positive rows contributions
of the partial solution (see eq. (2) and (3) where
P
U 1 = J): f (IU∗ 1 , U 1 ) = i∈R max(0, ri ).
An efficient bounding procedure. CP uses a branch and bound depth-first-search
to avoid the exploration of proven sub-optimal solution. The branch and bound
performances depend on the strength and efficiency of the procedure to compute
the upper bound. We design simple yet efficient bounding procedure for the maxsum sub-matrix problem. Intuitively one computes an upper bound on the row
contribution of each row and sum up all the positive bounds on the rows. The
upper bound on the contribution of a row is the sum of the matrix entries in the
selected columns plus the sum of the positive entries in the unbound columns.
One simply computes the upper bound as:
X
X
g(P ) = g(U 1 , U 0 , U ? ) =
max(0, ri +
max(0, Mi,j )) .
(5)
i∈R
j∈U ?
The bound is admissible but not tight as it may optimistically evaluate the
objective (g(P ) ≥ f (P )). Indeed, it relies on a per-line relaxation of the problem,
each selecting a possibly different set of columns.
The upperBound() method is an implementation of the proposed upper
bound using reversible double to store the incremental modifications of the partial contribution of the rows. Using a reversible sparse-set T for the row variables allows an efficient exclusion or inclusion of the rows through descent or
backtrack [13]. Indeed, as soon as the bound of the row contribution becomes
negative it should not be considered in the subsequent nodes of the search tree.
The number of rows to consider goes from θ(|R|) to O(|R|).
An efficient filtering procedure. The filter() method evaluates the upper
bound result for two one-step look-ahead scenarios: if column j would be selected, this look-ahead upper bound is denoted ub∈
j , or if j would not be selected,
∈
/
∈
denoted ubj . Then, any column j with ubj ≤ best is discarded as inclusion of
the column can only lead to worst solution than the incumbent. Similarly, any
/
column j with ub∈
j ≤ best is included. The time complexity for computing all
the look-ahead upper bounds is in O(|T | × |U ? |) by taking advantage of the fact
that the look-ahead bound of each line for each column can be obtained in O(1)
from riub .
3.3
Mixed Integer Linear Programming
Mixed Integer Linear Programming (MILP) [9] involves the optimization of a
linear objective function, subject to linear constraints. Some or all of the variables
are required to be integer. A MILP solver explores a branch and bound tree using
linear-programming (LP) bounds at each node of the search tree.
As for CP, there are advantages to using MILP. It represent a natural alternative to constraint programming as the problem is modeled with integer decision
variables.
Model. The max-sum sub-matrix problem can be formulated as a linear model
using “big M ” constants. Equations (7) and (8) linearize ri+ = max(0, ri ) with
ri being the sum of the selected entries of row i.
X
maximize
ri+ ,
(6)
i∈R
s.t. ri+ ≤ Ti × M,
ri+
∀i ∈ R ,
≤ ri + (1 − Ti ) × M, ∀i ∈ R ,
X
ri =
Mi,j × Uj , ∀i ∈ R .
(7)
(8)
(9)
j∈C
In order to avoid roundingPerrors and ill conditioned
Pmatrices, the big M
constants can be replaced by j∈C max(0, Mi,j ) and − j∈C min(0, Mi,j ), respectively in equations (7) and (8).
Algorithm 1: CP Global Constraint
best ← −∞ // best so far objective
trail
r: array[m] rev-double // rows partial sums
T : rev-set ← {1, . . . , n} // candidate rows
24
28
15
Method dfs()
if U ? 6= φ then
j ← select j ∈ U ?
foreach v ∈ {0, 1} do
trail.pushState()
D(Uj ) ← v
best ← max(best, evaluate())
ub ← upperBound()
if ub > best then
filter()
dfs()
16
trail.popState()
36
1
2
3
4
5
6
7
8
9
10
11
12
13
14
25
26
27
29
30
31
32
33
34
35
37
17
18
19
20
21
22
23
Method evaluate(): double
// evaluate objective
for j ∈ U ? do
if D(Uj ) 6= {0, 1} then
U? ← U? \ j
if D(Uj ) = 1 then
ri ← ri + Mi,j , ∀i ∈ T
return
P
i∈T
max(0, ri )
38
i
else T ← T \ i
return ub
40
41
42
43
45
46
i
// ri always ≤ 0
Method filter(): double
// remove impossible values
?
ub∈
j ← 0 ∀j ∈ U
/
?
ub∈
j ← 0 ∀j ∈ U
for i ∈ T do
for j ∈ U ? do
if Mi,j > 0 then
if riub − Mi,j > 0 then
/
∈
/
ub − M
ub∈
i,j
j ← ubj + ri
∈
ub
ub∈
j ← ubj + ri
39
44
4
Method upperBound(): double
ub ← 0
for i ∈ T do
P
riub ← ri + j∈U ? max(0, Mi,j )
ub
if r
> 0 then ub ← ub + r ub
else
/
∈
/
ub
ub∈
j ← ubj + ri
if riub + Mi,j > 0 then
∈
ub + M
ub∈
i,j
j ← ubj + ri
for j ∈ U ? do
if ub∈
j ≤ best then D(Uj ) ← 0
/
if ub∈
j ≤ best then D(Uj ) ← 1
Experiments
This section describes experiments conducted to assess the relative performances
of three algorithms to solve the max-sum sub-matrix problem. CP-LNS denotes
the method proposed by Le Van [6]. The other algorithms are specific methods
proposed in the present work: a constraint programming with a global constraint
(CPGC) and a mixed integer programming (MILP) solution.
These algorithms are first compared on data matrices which are generated
in a controlled setting. Experiments on the breast cancer gene expression data
used in [6] are reported next. The main criterion to assess the performance of the
various methods is the computing time to solve a particular problem instance.
This is technically assessed through an any-time profile defined below.
All algorithms have been implemented in the Scala programming language
(2.11.4). Each run is executed with a single thread on a MacBook Pro (OS
version 10.10.5) laptop (Intel i7-2720 CPU @ 2.20-3.30GHz, 4GB RAM per
run). Constraint programming implementations are based on the latest version of OscaR [10] and MILP is based on the latest version of Gurobi (7.0.2).
The code and datasets are available at https://bitbucket.org/vbranders/
maxsumsubmatriximplementation.
4.1
Synthetic Data
We follow a similar protocol as in [6]. Synthetic data are generated by implanting
a sub-matrix (I, J) of interest in a larger matrix M = (R, C) made of m rows
and n columns. The implanted sub-matrix (I, J) forms a specific selection of
rows and columns chosen at random. For each row index (from 1 to m) and each
column index (from 1 to n) of M, a binary variable is sampled from a Bernoulli
distribution B(p) and the associated row or column is included in the sub-matrix
(I, J) if B(p) = 1. Hence, I = {i ∈ R | B(p) = 1} and J = {j ∈ C | B(p) = 1}.
Next, the full matrix M is generated according to two normal distributions,
N (1, 1) whenever the particular entry belongs to the implanted sub-matrix, and
N (−3, 1) otherwise.
(
∼ N (1, 1)
if i ∈ I and j ∈ J ,
∀i ∈ R, ∀j ∈ C : Mi,j =
(10)
∼ N (−3, 1) otherwise .
Such a generation protocol favors the occurrence of higher values in the implanted sub-matrix and lower values elsewhere. Yet, given the standard deviations chosen equal to 1, both ranges of values may overlap. We note that, as
in [6], the implanted sub-matrix is not guaranteed to be an optimal solution
to the max-sum sub-matrix problem. This generation protocol looks however
realistic to define a rectangular (and not necessarily contiguous) sub-matrix of
interest in a larger matrix.
Problem instances are generated for various matrix sizes (m, n) and a varying
parameter p. As p increases, the size of the implanted sub-matrix is expected
to increase as well. In the gene expression analysis context, m can easily be two
orders of magnitude larger than n and the sub-matrix of interest is typically
small as compared to the full matrix. Such cases are included in the controlled
experiments reported below but a larger spectrum of problem instances is also
considered.
4.2
Gene Expression Data
The proposed case study concerns biomarker discovery for breast cancer subtypes using heterogeneous molecular data types. For a biological analysis and
interpretation of the results, the interested reader is redirected to the work of [6].
The pre-processed data provided by [6] consists of a matrix of m = 2, 211 rows
and n = 94 columns. In a first step, the matrix entries are transformed to discrete ranks along each row. A given threshold θ × n is then subtracted from each
matrix entry.
As θ increases, the proportion of positive entries decreases and, consequently,
a smaller sub-matrix of interest is expected to be found. Hence, the control
parameter θ plays a similar role as the parameter p (from Sect. 4.1) but in an
opposite way.
4.3
Evaluation
One could assess algorithms performances through runtime or number of feasible
solutions. While the former may depend on implementation details, the latter
strongly depends on the time spent in each node. We preferred the runtime
comparisons as it is a more common approach and we made sure to implement
the algorithms in the most comparable fashion.
Any-Time Profile. In practice, an important criterion for the user is the time
required to solve an instance and the ability to find the best solution within a
given budget of time. Using any-time profiles, one can summarize these characteristics. The idea behind any-time profiles is that an algorithm should produce
as high quality solution as possible at any moment of its running time [7].
In the max-sum sub-matrix problem, a high solution quality corresponds to
a sub-matrix of large sum. Performances of each algorithm are aggregated on
multiple instances. The solution quality for each algorithm is scaled on a [0, 1]
interval and the reported result is the average per-instance solution quality as
a function of the time. For each instance, runs not completed in a maximum
budget of time tmax are interrupted.
Definition 2 (Max-Sum Sub-Matrix Any-Time Profile). Let f (algo, inst, t)
be the objective value of the best solution found so far by an algorithm algo for
an instance inst at time t. Let tmax be the maximum running time before interrupting an algorithm. The any-time profile of an algorithm is the solution quality
Qalgo (t) computed on all instances as a function of the time:
f (algo, inst, t)
1 X
,
Qalgo (t) =
|inst|
f (algo∗inst , inst, tmax )
inst
(11)
with algo∗inst = argmax f (algo, inst, tmax ).
algo
4.4
Results
Figure 1 presents the any-time profile on 50 synthetic data with 10, 000 rows
and p = {0.05, 0.3, 0.7} for 100 columns (column 1) or 1, 000 columns (column
2) and the any-time profile on breast cancer gene expression data with 2, 211
rows, 94 columns and variable choices of θ (columns 3 and 4).
Synthetic Data. The existing CP-LNS method from [6] is clearly outperformed
by other methods. It can barely produce any solution in the allocated time budget. The best approach is CPGC followed by MILP. The reported curves are
stopped whenever the proof of optimality is obtained or else the maximal running time is reached. Hence, CPGC also exhibits best results whenever proving
optimality is possible in the allocated running time.
8
12
16
20
0
40
80
120
200
1
0.6
0
0.2
0.4
Solution quality
0.8
1
0.6
0
0.2
0.4
Solution quality
0.8
1
0.8
0.6
0.2
0
4
θ = 0.9
0
200
600
1000
0
200
600
10 000 x 100, p = 0.3
10 000 x 1 000, p = 0.3
θ = 0.7
θ = 0.5
4
8
12
16
20
0
40
80
120
200
0.8
0.6
0
0.2
0.4
Solution quality
0.8
0.6
0
0.2
0.4
Solution quality
0.8
0.6
0
0.2
0.4
Solution quality
0.8
0.6
0.4
0
0
0
200
600
1000
0
200
600
10 000 x 100, p = 0.7
10 000 x 1 000, p = 0.7
θ = 0.3
θ = 0.1
8
12
16
20
0
40
80
120
200
0.8
0.6
0.2
0
0
Time (seconds)
CPGC
0.4
Solution quality
0.8
0.6
0
0.2
0.4
Solution quality
0.8
0.6
0
0.2
0.4
Solution quality
0.8
0.6
0.4
0.2
0
4
Time (seconds)
1000
1
Time (seconds)
1
Time (seconds)
1
Time (seconds)
1
Time (seconds)
0
1000
1
Time (seconds)
1
Time (seconds)
1
Time (seconds)
1
Time (seconds)
0.2
Solution quality
0.4
Solution quality
0.8
0.6
0.4
Solution quality
0.2
0
0
Solution quality
θ = 0.95
10 000 x 1 000, p = 0.05
1
10 000 x 100, p = 0.05
200
600
1000
Time (seconds)
CP−LNS
0
200
600
1000
Time (seconds)
MILP
Fig. 1. Any-time profiles of constraint programming with a global constraint (CPGC),
the method proposed by Le Van et al. [6] (CP-LNS) and mixed integer linear programming (MILP). Columns 1 and 2: Curves corresponds to the average solution quality
over all instances as a function of time (in seconds). Results are computed on 50 instances with 10, 000 rows, 100 (column 1) or 1, 000 (column 2) columns, a variable p
and a maximum running time of 20 (column 1) or 200 (column 2) seconds. Columns 3
and 4: Curves corresponds to the solution quality over each problem instance obtained
for a specific θ as a function of the time (in seconds). Results computed on breast
cancer gene expression data with 2, 211 rows, 94 columns and various θ values for a
maximum waiting time of 1, 000 seconds.
Gene Expression Data. Each curve corresponds here to the performance of
an algorithm on a single instance, the one obtained for a specific choice of θ. On
the whole spectrum of instances considered, the clear winner is CPGC. The most
interesting instances are those for which θ ≥ 0.9 since such settings correspond
to small sub-matrices which are more likely to illustrate an interesting biological
pattern. In such cases, the best approaches are CPGC and CP-LNS.
5
Conclusions and Perspectives
We introduce the max-sum sub-matrix problem which consists in finding a (non
necessarily contiguous) rectangular sub-matrix in a large matrix whose sum is
maximal. This problem is originally motivated, in the context of gene expression
analysis, by the search of a subset of highly expressed genes in a specific subset,
to be found, of relevant samples exhibiting such a pattern. A close variant of
this problem, known as maximal ranked tile mining problem, has already been
studied and tackled with constrained programming (CP) combined with large
neighborhood search (CP-LNS) [6].
We present here key properties of the max-sum sub-matrix problem to speed
up the search for a solution and we propose two additional algorithms to solve it.
Experiments reported both on synthetic data and the original gene expression
data used in [6] illustrate the benefits of our proposed methods. In particular,
a CP approach with a global constraint (CPGC) is the most effective one in a
large spectrum of problem instances. Overall, the CPGC method is also best at
proving optimality when such proof can be obtained within the allocated budget
of computing time.
The second approach proposed here relies on mixed integer linear programming (MILP). It is arguably the simplest to formulate and to address with a
standard solver for such problems. It is competitive with the other methods and
largely outperforms CP-LNS as well in our controlled experiments. It exhibits
however some performance degradation on some instances from gene expression
data, most likely as a consequence of the specific relaxation it is based on.
Related biclustering techniques identify multiple sub-matrices. The max-sum
sub-matrix problem could be extended to such a setting by iteratively selecting
max-sum sub-matrices penalizing the previous discovered rows and columns to
avoid the selection of already identified sub-matrices.
The max-sum sub-matrix mining problem could be extended to a supervised
classification setting. For example, in gene expression analysis, one typically
wants to find genes (rows) that allows to discriminate between two conditions. In
other words, the columns could be a priori labeled according to two conditions.
The objective can then be to identify a subset of rows that are maximally relevant
to discriminate between subsets of samples from different conditions. This could
be encoded in a larger matrix for which columns represent pairs of columns in
either conditions from the original matrix and the value stored is interpreted as
a distance value for a particular gene across both conditions.
The max-sum sub-matrix problem could also be applied to outlier detection
and/or biclustering. For example, using an appropriate data transformation,
entries that are close to the mean or to the median could be mapped to relatively
large positive entries. Similarly, entries far away from the mean would be mapped
to low values. Consequently a sub-matrix of maximal sum would correspond to
subsets of rows and of columns exhibiting similar entries. Explicit comparisons
to existing biclustering algorithms could be considered in such a setting.
References
1. Atzmueller, M.: Subgroup discovery. Wiley Interdisciplinary Reviews: Data Mining
and Knowledge Discovery 5(1), 35–49 (2015)
2. Bentley, J.: Programming pearls: algorithm design techniques. Communications of
the ACM 27(9), 865–873 (1984)
3. Dawande, M., Keskinocak, P., Tayur, S.: On the biclique problem in bipartite
graphs (1996)
4. Floyd, R.W.: Nondeterministic algorithms. Journal of the ACM (JACM) 14(4),
636–644 (1967)
5. Knuth, D.E.: The art of computer programming: Volume 4B, Combinatorial Algorithms: Part 2, Backtrack Programming, vol. 4B. Adison-Wesley (2016)
6. Le Van, T., Van Leeuwen, M., Nijssen, S., Fierro, A.C., Marchal, K., De Raedt, L.:
Ranked tiling. In: Joint European Conference on Machine Learning and Knowledge
Discovery in Databases. pp. 98–113. Springer (2014)
7. López-Ibánez, M., Stützle, T.: Automatically improving the anytime behaviour of
optimisation algorithms. European Journal of Operational Research 235(3), 569–
582 (2014)
8. Madeira, S.C., Oliveira, A.L.: Biclustering algorithms for biological data analysis:
a survey. IEEE/ACM Transactions on Computational Biology and Bioinformatics
(TCBB) 1(1), 24–45 (2004)
9. Nemhauser, G.L., Wolsey, L.A.: Integer programming and combinatorial optimization. Wiley, Chichester. GL Nemhauser, MWP Savelsbergh, GS Sigismondi (1992).
Constraint Classification for Mixed Integer Programming Formulations. COAL
Bulletin 20, 8–12 (1988)
10. OscaR
Team:
OscaR:
Scala
in
OR
(2012),
available
from
https://bitbucket.org/oscarlib/oscar
11. Parker, J.S., Mullins, M., Cheang, M.C., Leung, S., Voduc, D., Vickery, T., Davies,
S., Fauron, C., He, X., Hu, Z., et al.: Supervised risk predictor of breast cancer
based on intrinsic subtypes. Journal of clinical oncology 27(8), 1160–1167 (2009)
12. Perou, C.M., Sørlie, T., Eisen, M.B., van de Rijn, M., Jeffrey, S.S., Rees, C.A.,
Pollack, J.R., Ross, D.T., Johnsen, H., Akslen, L.A., et al.: Molecular portraits of
human breast tumours. Nature 406(6797), 747–752 (2000)
13. de Saint-Marcq, V.l.C., Schaus, P., Solnon, C., Lecoutre, C.: Sparse-sets for domain
implementation. In: CP workshop on Techniques foR Implementing Constraint
programming Systems (TRICS). pp. 1–10 (2013)
14. Takaoka, T.: Efficient algorithms for the maximum subarray problem by distance
matrix multiplication. Electronic Notes in Theoretical Computer Science 61, 191–
200 (2002)
15. Tamaki, H., Tokuyama, T.: Algorithms for the maximum subarray problem based
on matrix multiplication. In: SODA. vol. 1998, pp. 446–452 (1998)
| 2cs.AI
|
The Capacity of Private Information Retrieval with
Partially Known Private Side Information∗
arXiv:1710.00809v2 [cs.IT] 27 Nov 2017
Yi-Peng Wei
Karim Banawan
Sennur Ulukus
Department of Electrical and Computer Engineering
University of Maryland, College Park, MD 20742
[email protected]
[email protected]
[email protected]
November 28, 2017
Abstract
We consider the problem of private information retrieval (PIR) of a single message
out of K messages from N replicated and non-colluding databases where a cacheenabled user (retriever) of cache-size M possesses side information in the form of full
messages that are partially known to the databases. In this model, the user and the
databases engage in a two-phase scheme, namely, the prefetching phase where the user
acquires side information and the retrieval phase where the user downloads desired
information. In the prefetching phase, the user receives mn full messages from the nth
P
database, under the cache memory size constraint N
n=1 mn ≤ M . In the retrieval
phase, the user wishes to retrieve a message such that no individual database learns
anything about the identity of the desired message. In addition, the identities of the
side information messages that the user did not prefetch from a database must remain
private against that database. Since the side information provided by each database
in the prefetching phase is known by the providing database and the side information
must be kept private against the remaining databases, we coin this model as partially
known private side information. We characterize the capacity of the PIR with partially
−1
1− 1
1
= 1−( 1 )NK−M .
known private side information to be C = 1 + N1 + · · · + N K−M
−1
N
Interestingly, this result is the same if none of the databases knows any of the prefetched
side information, i.e., when the side information is obtained externally, a problem posed
by Kadhe et al. and settled by Chen-Wang-Jafar recently. Thus, our result implies that
there is no loss in using the same databases for both prefetching and retrieval phases.
This work was supported by NSF Grants CNS 13-14733, CCF 14-22111, CNS 15-26608 and CCF 1713977.
∗
1
1
Introduction
The private information retrieval (PIR) problem is a canonical problem to study privacy
issues that arise when information is downloaded (retrieved) from public databases. Since
its first formulation by Chor et al. in [1], the PIR problem has become a central research
topic in the computer science literature, see e.g., [2–5]. In the classical setting of PIR in [1],
a user wishes to retrieve a single message (or a file) out of K messages replicated across
N non-communicating databases without leaking any information about the identity of the
retrieved message. To that end, the user submits a query to each database. Each database
responds truthfully with an answer string. The user reconstructs the desired message from
the collected answer strings. Trivially, the user can download the entire database and incur a
linear (in number of messages) download cost, but this retrieval strategy is highly inefficient.
The efficiency of a PIR scheme is measured by the normalized download cost, which is the
cost of privately downloading one bit of the desired message. The goal of the PIR problem is
to devise the most efficient retrieval strategy under the privacy and decodability constraints.
The PIR problem has received attention in recent years in the information and coding
theory literatures, see e.g., [6–11]. In the leading work of Sun-Jafar [12], the classical PIR
problem is re-formulated to conform with the conventional information-theoretic arguments,
and the notion of PIR capacity is introduced, which is defined as the supremum of retrieval
rates over all achievable retrieval schemes. Reference [12] characterizes the capacity of the
−1
1
using a greedy achievable scheme
classical PIR model to be C = 1 + N1 + · · · + N K−1
that is closely related to blind interference alignment [13] and an induction-based converse
argument. Following the work of Sun-Jafar [12], the capacity of many interesting variants of
the classical PIR model have been investigated, such as, PIR from colluding databases, robust
PIR, symmetric PIR, PIR from MDS-coded databases, PIR for arbitrary message lengths,
multi-round PIR, multi-message PIR, PIR from Byzantine databases, secure symmetric PIR
with adversaries, and their several combinations [14–28].
In this paper, we consider the problem of PIR with partially known private side information. Our work is most closely related to [29–32]1 . These works investigate the PIR problem
when the user (retriever) possesses some form of side information about the contents of the
databases. However, the models of [29–32] differ in three important aspects, namely, 1) the
structure of the side information, 2) the presence or absence of privacy constraints on the
side information, and 3) the databases’ awareness of the side information at its initial acquisition. Here, structure of the side information refers to whether the side information is in the
form of full messages or parts of messages or whether messages are mixed through functions
(coded/uncoded); privacy of the side information refers to whether the user further aims to
keep the side information private from the databases; and databases’ awareness of the side
information refers to whether the databases knew the initially prefetched side information.
1
A parallel line of work that studies privacy issues of requests and side information in index coding based
broadcast systems can be found in [33, 34].
2
Specifically, reference [29] studies the capacity of the cache-aided PIR where the user
caches rLK bits in the form of any arbitrary function of the K messages, where L is the
message size, and 0 ≤ r ≤ 1 is the caching ratio. Reference [29] assumes that the cache
content is perfectly known by all the databases, and hence there is no need to protect the
privacy of the cached content. Reference [29] determines the optimal download cost for this
1
1
= (1 −r) 1 + N1 + · · · + N K−1
using a memory-sharing achievable
model to be D ∗ (r) = C(r)
scheme and a converse that utilizes Han’s inequality. This conclusion is somewhat pessimistic
in that the user cannot exploit the cached content as useful side information during PIR to
reduce the download cost, since the databases are fully aware of it; the optimum D ∗ (r)
formula indicates that the user should download the uncached part of the content, i.e.,
(1 − r), via the optimum PIR scheme in [12]. This result motivates [30, 31] to study the
other extreme when the databases are completely unaware of the side information at its
initial acquisition. References [30] and [31] differ in terms of the structure of the cached
content: [30] considers the case where rK full messages are cached, and [31] considers the
case where a random r fraction of the symbols of each of K messages is cached. In this
case, [31] shows a significant reduction in the download cost over [29], as the user can now
leverage the cached bits as side information, since they are unknown to the databases. In [31],
there is no privacy constraint on the cached content.
Reference [30] further introduces another model where the cached content (in the form
of full messages) which is unknown to the databases at the time of initial prefetching, must
remain unknown throughout the PIR, i.e., the queries of the user should not leak any information about the cached content to the databases. The exact capacity for this problem
−1
1
. The optimal achievable scheme in
is settled in [32] to be C = 1 + N1 + · · · + N K−M
−1
this case starts from the traditional achievable scheme without side information in [12] and
reduces the download cost by utilizing the reconstruction property of MDS codes.
In this paper, we take a deeper look at the issue of awareness or otherwise unawareness
of the databases about the cached content at its initial acquisition. We first note that it is
practically challenging to make the side information completely unknown to the databases
at its initial acquisition as assumed in [30–32]. One way to do this could be to employ
one of the databases for prefetching the side information and exclude it from the retrieval
process. Therefore, for the remaining N − 1 databases, the side information is completely
unknown. This solution is strictly sub-optimal as the capacity expression in [32] (shown
as C in the previous paragraph) is monotonically decreasing in N. An alternative solution
could be to devise a refreshing mechanism that ensures that the cached content is essentially
random from the perspective of each database [29], which may be challenging to implement.
We also note that the other extreme of the problem, where the databases are fully aware
of the cached content [29], is discouraging as the user cannot benefit from the cached side
information. Therefore, a natural model is to use the databases for both prefetching and
retrieval phases, such that the databases gain partial knowledge about the side information
3
available to the user, which makes it possible for the user to exploit the remaining side
information that is unknown to each individual database to reduce the download cost during
the retrieval process. This poses the following questions: Can we propose efficient joint
prefetching-retrieval strategies that exploit the limited knowledge of each database to drive
down the download cost? How much is the loss from the fully unknown case in [30, 32]?
In this paper, we investigate the PIR problem when the user and the databases engage in
a two-phase scheme, namely, prefetching phase and retrieval phase. In the prefetching phase,
the user caches mn full messages out of the K messages from the nth database under a total
P
cache memory size constraint N
n=1 mn ≤ M. Hence, each database has a partial knowledge
about the side information possessed by the user, namely, the part of the side information
that this database has provided during the prefetching phase. In the retrieval phase, the
user wants to retrieve a message (which is not present in its memory) without leaking any
information to any individual database about the desired message or the remaining side
information messages that are unknown to each database. The goal of this work is to design
a joint prefetching-retrieval scheme that minimizes the download cost in the retrieval phase.
To that end, we first derive a general lower bound for the normalized download cost that
is independent of the prefetching strategy. Then, we prove that this bound is attainable using
two achievable schemes. The first achievable scheme, which is proposed in [32] for completely
unknown side information, is a valid achievable scheme for our problem with partially known
side information for any prefetching strategy.2 We provide a second achievable scheme for
∈ N, which requires smaller sub-packetization
the case of uniform prefetching, i.e., mn = M
N
and smaller field size for realizing MDS codes. While the first achievable scheme [32] requires
a message size of L = N K , the second achievable scheme proposed here requires a message
M
M
size of L = N K− N , which scales down the message size by an exponential factor N N , which
in turn simplifies the achievable scheme and minimizes the total number of downloaded bits
without sacrificing from the capacity. We prove that the exact capacity of this problem is
−1
1
C = 1 + N1 + · · · + N K−M
. Surprisingly, this is the same capacity expression for the
−1
PIR problem when the databases are completely unaware of the side information possessed
by the user as found in [32] recently. Therefore, our result implies that there is no loss in
the capacity if the same databases are employed in both prefetching and retrieval phases.
2
System Model
We consider a classic PIR problem with K independent messages W1 , . . . , WK , where each
message consists of L symbols,
H(W1 ) = · · · = H(WK ) = L,
2
H(W1 , . . . , WK ) = H(W1 ) + · · · + H(WK ).
We thank Dr. Hua Sun for pointing this out.
4
(1)
There are N non-communicating databases, and each database stores all the K messages.
The user (retriever) has a local cache memory which can store up to M messages.
There are two phases: a prefetching phase and a retrieval phase. In the prefetching phase,
∀n ∈ [N], where [N] = {1, 2, . . . , N}, the user caches mn out of total K messages from the
nth database. We denote the indices of the cached messages from the nth database as Hn .
Therefore, |Hn | = mn . We denote the indices of all cached messages as H,
H=
N
[
(2)
Hn ,
n=1
where Hn1 ∩ Hn2 = ∅, if n1 6= n2 . Due to the cache memory size constraint, we require
|H| =
N
X
mn ≤ M.
(3)
n=1
Since the user caches mn messages from the nth database, Hn is known to the nth database.
Since the databases do not communicate with each other, Hn is unknown to the other
databases. We use m = (m1 , . . . , mN ) to represent the prefetching phase. After the prefetching phase, the user learns |H| messages, denoted as WH = {Wi1 , . . . , Wi|H| }. We refer to WH
as partially known private side information.
In the retrieval phase, the user privately generates a desired message index θ ∈ [K] \ H,
and wishes to retrieve message Wθ such that no database knows which message is retrieved.
Since the desired message index θ and cached message indices H are independent of the
message contents, for random variables θ, H, and W1 , . . . , WK , we have
H (θ, H, W1 , . . . , WK ) = H (θ, H) + H(W1 ) + · · · + H(WK ).
[θ,H]
(4)
[θ,H]
In order to retrieve Wθ , the user sends N queries Q1 , . . . , QN to the N databases,
[θ,H]
where Qn is the query sent to the nth database for message Wθ given the user has partially
known private side information WH . The queries are generated according to H, which is
independent of the realizations of the K messages. Therefore, we have
[θ,H]
I(W1 , . . . , WK ; Q1
[θ,H]
, . . . , QN ) = 0.
(5)
To ensure that individual databases do not know which message is retrieved and also
do not know the cached messages from other databases, i.e., to guarantee the privacy of
(θ, H \ Hn ), we need to satisfy the following privacy constraint, ∀n ∈ [N], ∀H, H′ such that
|H| = |H′ | ≤ M, Hn ⊂ H, Hn ⊂ H′ , and ∀θ ∈ [K] \ H, ∀θ′ ∈ [K] \ H′ ,
′
′
′
′
,H ]
, An[θ ,H ] , W1 , . . . , WK , Hn ),
, W1 , . . . , WK , Hn ) ∼ (Q[θ
, A[θ,H]
(Q[θ,H]
n
n
n
5
(6)
where A ∼ B means that A and B are identically distributed.
[θ,H]
[θ,H]
Upon receiving the query Qn , the nth database replies with an answering string An ,
[θ,H]
which is a function of Qn and all the K messages. Therefore, ∀θ ∈ [K] \ H, ∀n ∈ [N],
H(A[θ,H]
|Q[θ,H]
, W1 , . . . , WK ) = 0.
n
n
[θ,H]
(7)
[θ,H]
After receiving the answering strings A1 , . . . , AN from all the N databases, the user
needs to decode the desired message Wθ reliably. By using Fano’s inequality, we have the
following reliability constraint
[θ,H]
[θ,H]
[θ,H]
[θ,H]
= o(L),
H Wθ |WH , H, Q1 , . . . , QN , A1 , . . . , AN
(8)
→ 0 as L → ∞.
where o(L) denotes a function such that o(L)
L
For fixed N, K, and pretching scheme m = (m1 , . . . , mN ), a pair (D(m), L(m)) is achievable if there exists a PIR scheme for messages of size L(m) symbols long with partially known
private side information satisfying the privacy constraint (6) and the reliability constraint (8),
where D(m) represents the expected number of downloaded symbols (over all the queries)
[θ,H]
[θ,H]
[θ,H]
[θ,H]
from the N databases via the answering strings A1:N , where A1:N = (A1 , . . . , AN ), i.e.,
D(m) =
N
X
n=1
.
H A[θ,H]
n
(9)
In this work, for fixed N, K, and M, we aim to characterize the optimal normalized download
cost D ∗ , where
∗
D = inf
m:(3)
3
D(m)
: (D(m), L(m)) is achievable .
L(m)
(10)
Main Results
We characterize the exact normalized download cost for the PIR problem with partially
known private side information as shown in the following theorem.
Theorem 1 In the PIR problem with partially known private side information under the
cache memory size constraint |H| ≤ M, the optimal normalized download cost is
1
1
+ · · · + K−M −1
N
N
1 K−M
1 − (N )
.
=
1 − N1
D∗ = 1 +
(11)
(12)
The converse proof for Theorem 1 is given in Section 4, and the achievability proof for
Theorem 1 is given in Section 5. Theorem 1 does not assume any particular property for
6
the prefetching strategy, i.e., m is arbitrary except for satisfying the memory size constraint.
We have a few remarks.
1−
1
Remark 1 Theorem 1 implies that C = D1∗ = 1−( 1 )NK−M . Surprisingly, this capacity expresN
sion is exactly the same as the capacity for the PIR problem with completely unknown private
side information in [32]. This implies that there is no loss in capacity due to employing the
same databases for both prefetching and retrieval phases. The reason for this phenomenon
is that although each database has a partial knowledge about some of the cached messages at
the user, the privacy constraint on this known side information is relaxed.
Remark 2 The normalized download cost in Theorem 1 is the same as the normalized download cost for the classical PIR problem [12] if the number of messages is K − M. That is, a
cache of size M messages effectively reduces the total number of messages by M. Noting that
the download cost in [12] monotonically increases in the number of messages, the effective
reduction in the number of messages by the cache size results in a significant reduction in the
download cost due to the presence of side information at the user even though it is partially
known by the databases and it needs to be kept private against other databases.
Remark 3 The optimal prefetching strategy exploits the entire cache memory of the user as
the capacity expression is monotonically increasing in M.
Remark 4 In Section 5, we present the capacity achieving schemes for the partially known
private side information. We note that, in general the PIR scheme in [32] is a valid achievable scheme for our problem as well. Nevertheless, in the special case of uniform prefetching,
= m ∈ N, we provide a different achievable scheme that exploits the prefetchi.e., mn = M
N
M
ing uniformity to work with message size L = N K−m = N K− N in contrast to L = N K
needed for the scheme in [32], i.e., the message size is decreased by an exponential factor
M
N N . Furthermore, we note that although both schemes need an MDS code to reduce the
number of downloaded equations, we note that the field size needed to realize this MDS code
∈ N) compared with the field size needed in
is significantly smaller with our scheme (if M
N
the scheme in [32]. This implies that although uniform prefetching does not affect the PIR
capacity, it significantly simplifies the achievable scheme.
4
Converse Proof
In this section, we derive a general lower bound for the normalized download cost D ∗ given
in (10). We extend the techniques presented in [12, 32] to the PIR problem with partially
known private side information.
For the prefetching vector m = (m1 , . . . , mN ) satisfying (3), we note that satisfying the
memory size constraint with equality leads to a valid lower bound on (10). Consequently, we
P
first consider the case N
n=1 mn = M̃ ≤ M, i.e., we study the case when the user learns M̃
7
messages after the prefetching phase. Since we do not specify the prefetching strategy m in
P
advance, the following lower bound is valid for all m such that N
n=1 mn = M̃ . Without loss
of generality, we relabel the M̃ cached messages as W1 , W2 , . . . , WM̃ , i.e., H = {1, 2, . . . , M̃}
and WH = W1:M̃ . We first need the following lemma, which characterizes a lower bound on
the length of the undesired portion of the answering strings as a consequence of the privacy
constraint on the retrieved message.
Lemma 1 (Interference lower bound) For the PIR with partially known private side
information, the interference from undesired messages within the answering strings, D − L,
is lower bounded by,
[M̃+1,H]
[M̃+1,H]
D − L + o(L) ≥ I WM̃ +2:K ; H, Q1:N
, A1:N
|WH , WM̃ +1 .
(13)
If the privacy constraint is absent, the user downloads only L symbols for the desired
message, however, when the privacy constraint is present, it should download D symbols.
The difference between D and L, i.e., D − L, corresponds to the undesired portion of the
answering strings. Note that Lemma 1 is an extension of [12, Lemma 5] if M̃ = 0, i.e., the
user has no partially known private side information. Lemma 1 differs from its counterpart
in [31, Lemma 1] in two aspects, namely, the left hand side is D(r) − L(1 − r) in [31] as the
user requests to download the uncached bits only, and the bound in [31, Lemma 1] constructs
K − 1 distinct lower bounds by changing k in contrast to one bound here as it always starts
from WM̃ +2 . Finally, we note that a similar argument to Lemma 1 can be implied from [32].
Proof: We start with the right hand side of (13),
[M̃+1,H]
[M̃+1,H]
I WM̃ +2:K ; H, Q1:N
, A1:N
|WH , WM̃ +1
[M̃ +1,H]
[M̃+1,H]
= I WM̃ +2:K ; H, Q1:N
, A1:N
, WM̃ +1 |WH − I WM̃ +2:K ; WM̃+1 |WH .
(14)
For the first term on the right hand side of (14), we have
[M̃ +1,H]
[M̃+1,H]
I WM̃ +2:K ; H, Q1:N
, A1:N
, WM̃ +1 |WH
[M̃+1,H]
[M̃+1,H]
[M̃+1,H]
[M̃ +1,H]
= I WM̃ +2:K ; H, Q1:N
, A1:N
|WH + I WM̃+2:K ; WM̃+1 |H, Q1:N
, A1:N
, WH
(15)
(8)
[M̃ +1,H]
[M̃+1,H]
= I WM̃ +2:K ; H, Q1:N
, A1:N
|WH + o(L)
(16)
(4),(5)
[M̃ +1,H]
[M̃+1,H]
+ o(L)
(17)
= I WM̃ +2:K ; A1:N
|WH , H, Q1:N
[M̃ +1,H]
[M̃ +1,H]
[M̃+1,H]
[M̃ +1,H]
= H A1:N
|WH , H, Q1:N
− H A1:N
|WH , H, Q1:N
, WM̃ +2:K + o(L)
(18)
(8)
[M̃ +1,H]
[M̃+1,H]
[M̃+1,H]
[M̃ +1,H]
− H A1:N
, WM̃+1 |WH , H, Q1:N
, WM̃ +2:K + o(L)
= H A1:N
|WH , H, Q1:N
(19)
8
[M̃ +1,H]
− H WM̃ +1 |WH , H, Q1:N
, WM̃ +2:K + o(L)
≤H
(4),(5)
[M̃ +1,H]
[M̃+1,H]
= H A1:N
|WH , H, Q1:N
− H WM̃ +1 |WH , WM̃+2:K + o(L)
[M̃ +1,H]
[M̃ +1,H]
− L + o(L)
= H A1:N
|WH , H, Q1:N
[M̃ +1,H]
≤ H A1:N
− L + o(L)
[M̃ +1,H]
[M̃+1,H]
A1:N
|WH , H, Q1:N
≤ D − L + o(L),
(20)
(21)
(22)
(23)
(24)
[M̃ +1,H]
[M̃+1,H]
H, Q1:N
, A1:N
, WH
where (16), (19) follow from the decodability of WM̃ +1 given
,
[M̃ +1,H]
, (21) follows from the
(17) follows from the independence of WM̃ +2:K and H, Q1:N
[M̃ +1,H]
independence of WM̃ +1 and H, Q1:N
, and (24) follows from the independence bound.
For the second term on the right hand side of (14), we have
I WM̃ +2:K ; WM̃ +1 |WH = H WM̃ +1 |WH − H WM̃ +1 |WH , WM̃ +2:K
= L − L = 0.
(25)
(26)
Combining (14), (24), and (26) yields (13).
In the following lemma, we prove an inductive relation for the mutual information term
on the right hand side of (13).
Lemma 2 (Induction lemma) For all k ∈ {M̃ + 2, . . . , K}, the mutual information term
in Lemma 1 can be inductively lower bounded as,
[k−1,H]
[k−1,H]
I Wk:K ; H, Q1:N , A1:N |WH , WM̃ +1:k−1
L − o(L)
1
[k,H]
[k,H]
.
≥ I Wk+1:K ; H, Q1:N , A1:N |WH , WM̃+1:k +
N
N
(27)
Lemma 2 is a generalization of [12, Lemma 6] to our setting. The main difference between
Lemma 2 and [32] is that in order to apply the partial privacy constraint, the random variable
H should be used in its local form Hn as it corresponds to the partial knowledge of the nth
database.
Proof: We start with the left hand side of (27),
I
[k−1,H]
[k−1,H]
Wk:K ; H, Q1:N , A1:N |WH , WM̃+1:k−1
1
[k−1,H]
[k−1,H]
× N × I Wk:K ; H, Q1:N , A1:N |WH , WM̃+1:k−1
N
N
1 X
I Wk:K ; Hn , Q[k−1,H]
, A[k−1,H]
|WH , WM̃ +1:k−1
≥
n
n
N n=1
=
N
1 X
≥
I Wk:K ; Q[k−1,H]
, A[k−1,H]
|WH , WM̃ +1:k−1 , Hn
n
n
N n=1
9
(28)
(29)
(30)
N
1 X
=
I Wk:K ; Q[k,H]
, A[k,H]
|WH , WM̃+1:k−1 , Hn
n
n
N n=1
(6)
N
1 X
[k,H]
I Wk:K ; A[k,H]
|W
,
W
,
H
,
Q
=
H
n
n
n
M̃ +1:k−1
N n=1
(4),(5)
N
1 X
=
H A[k,H]
|WH , WM̃ +1:k−1, Hn , Q[k,H]
n
n
N n=1
(7)
N
1 X [k,H]
[k,H]
[k,H]
H An |WH , WM̃ +1:k−1, H, Q1:N , A1:n−1
≥
N n=1
(31)
(32)
(33)
(34)
N
1 X
[k,H]
[k,H]
=
(35)
I Wk:K ; A[k,H]
|W
,
W
,
H,
Q
,
A
H
n
1:n−1
M̃ +1:k−1
1:N
N n=1
1
[k,H]
[k,H]
(36)
= I Wk:K ; A1:N |WH , WM̃ +1:k−1, H, Q1:N
N
(4),(5) 1
[k,H]
[k,H]
=
(37)
I Wk:K ; H, Q1:N , A1:N |WH , WM̃ +1:k−1
N
o(L)
(8) 1
[k,H]
[k,H]
(38)
= I Wk:K ; Wk , H, Q1:N , A1:N |WH , WM̃ +1:k−1 −
N
N
o(L)
1
1
[k,H]
[k,H]
= I Wk:K ; Wk |WH , WM̃+1:k−1 + I Wk:K ; H, Q1:N , A1:N |WH , WM̃+1:k −
N
N
N
(39)
L − o(L)
1
[k,H]
[k,H]
= I Wk+1:K ; H, Q1:N , A1:N |WH , WM̃ +1:k +
,
(40)
N
N
(7)
where (29) follows from the non-negativity of mutual information, (31) follows from the
privacy constraint, (32) follows from the independence of the messages and the queries, (33),
(35) follow from the fact that answer strings are deterministic functions of the messages and
the queries, (34) follows from
the fact that
conditioning reduces entropy, (37) follows from
[k,H]
the independence of Wk:K and H, Q1:N , (38) follows from the reliability constraint on Wk ,
and (40) follows from the independence of Wk and (WH , WM̃+1:k−1 ).
Now, we are ready to derive the lower bound for arbitrary K, N, and M̃ . This can be
obtained by applying Lemma 1 and Lemma 2 successively.
Lemma 3 For fixed N, K, and M̃ ≤ M, we have
1
1
− o(L).
D ≥L 1+
+···+
N
N K−M̃ −1
(41)
Proof: We have
(13)
[M̃ +1,H]
[M̃+1,H]
D ≥ L + I WM̃ +2:K ; H, Q1:N
, A1:N
|WH , WM̃ +1 − o(L)
(27)
L
1
[M̃ +2,H]
[M̃+2,H]
≥ L+
+ I WM̃ +3:K ; H, Q1:N
, A1:N
|WH , WM̃ +1:M̃ +2 − o(L)
N
N
10
(42)
(43)
(27)
≥ L+
(27)
L
1
L
[M̃ +3,H]
[M̃+3,H]
+ 2 + I WM̃ +4:K ; H, Q1:N
, A1:N
|WH , WM̃ +1:M̃ +3 − o(L)
N
N
N
≥ ...
(27)
1
1
≥ L 1+
+···+
− o(L),
N
N K−M̃−1
(44)
(45)
(46)
where (42) follows from Lemma 1, (43)-(46) follow from applying Lemma 2 starting from
k = M̃ + 2 to k = K, which differs from [12] in terms of the starting point of the induction.
We conclude the converse proof by dividing by L and taking L → ∞ in (41), to have
D∗ ≥ 1 +
1
1
+···+
.
K−
N
N M̃−1
(47)
Finally, we note that the right hand side of (47) is monotonically decreasing in M̃. Since
M̃ ≤ M, the lowest lower bound is obtained by taking M̃ = M, which yields the final
converse bound,
D∗ ≥ 1 +
1
1
+ · · · + K−M −1 .
N
N
(48)
P
Remark 5 We note that if (48) is tight, any prefetching strategy m such that N
n=1 mn < M
is strictly suboptimal. Furthermore, the lower bound in (48) is the same for all prefetching
P
strategies m satisfying N
n=1 mn = M. In Section 5, we show that this lower bound is tight.
5
Achievability Proof
We first note that the achievability scheme proposed in [32] for the PIR problem with completely unknown private side information also works for the PIR problem with partially
known private side information here. The PIR scheme in [32] is based on MDS codes and
consists of two stages. The first stage determines the systematic part of the MDS code according to the queries generated in [12], which protects the privacy of the desired message,
i.e., in the first stage, the user designs the queries such that no information is leaked about
which message out of the K messages is the desired one. In the second stage, the user reduces
the number of the downloaded equations by downloading the parity part of the MDS code
only. For the case of partially known private side information here, two privacy constraints
should be satisfied: the desired message privacy constraint and the side information privacy
constraint. For the desired message, we note that the user should guarantee that the queries
designed to retrieve any of the K − mn messages should be indistinguishable at the nth
database (i.e., with the exception of the mn messages that the nth database has provided).
Due to the first stage, the privacy of the desired message holds as it was designed to protect
11
the privacy of all K messages, which is more restricted. Furthermore, the PIR scheme in [32]
also protects the privacy of the side information. The scheme in [32] ensures that the queries
do not reveal the identity of the M messages that are possessed by the user as side information. In our model, we note that we need to protect the privacy of M − mn messages from
the nth database, as the remaining mn messages are known to the nth database. Since the
privacy constraint imposed on the side information in our model is less restricted than [32],
using the scheme in [32] satisfies the privacy constraint of the side information in our case as
well. That is, the nth database cannot infer which other M − mn messages the user holds.
The PIR scheme in [32] achieves the normalized download cost in Theorem 1. The PIR
scheme in [32] requires a message size of N K symbols. In the following, we propose another
M
∈ N. Thus, this
achievability scheme which requires a message size of N K− N , if mn = M
N
scheme requires smaller sub-packetization and smaller field size for the MDS code.
Our PIR scheme for partially known private side information is based on the PIR schemes
in [12, 32]. To protect the privacy of the partially known private side information and the
privacy of the desired message, similar to [12], we apply the following three principles recursively: 1) database symmetry, 2) message symmetry within each database, and 3) exploiting
undesired messages as side information. We reduce the download cost by utilizing the reconstruction property of MDS codes by exploiting partially known private side information
as in [32]. The side information enables the user to request reduced number of equations as
a consequence of the user’s knowledge of M messages from the prefetching phase. Nevertheless, to protect the privacy of the side information, the user actually queries MDS coded
symbols which is mixture of K −mn messages. The main difference between our achievability
scheme and that in [12,32] is that since the nth database knows that the user has prefetched
mn messages, the user does not need to protect the privacy for these mn messages from the
nth database. This effectively reduces the number of messages that the scheme in [32] needs
to operate on to K −mn messages in contrast to K in [32]. When M
∈ N, we show that if the
N
, for all n, then
user caches the same number of messages from each database, i.e., mn = M
N
the lower bound in (11) is achievable by this scheme. This scheme reduces the message size
M
requirement from L = N K in [32] to L = N K− N here, simplifying the achievable scheme.
5.1
5.1.1
Motivating Examples
N = 2 Databases, K = 4 Messages, and M = 2 Cached Messages
Assume that each message is of size 8 symbols. We use ai , bi , ci and di , for i = 1, . . . , 8,
to denote the symbols of messages W1 , W2 , W3 and W4 , respectively. In this example, in
the prefetching phase, the user caches message W3 from database 1, and message W4 from
database 2; and in the retrieval phase, the user wishes to retrieve message W1 privately.
The user first generates the query table in Table 1. In Table 1, the user queries 7 symbols.
Since the user knows d1 from the cached message W4 , in order to use the partially known
12
private side information, the user can in fact reduce the number of queries to 6 equations per
database by ignoring d1 . However, if the user simply does not download d1 , it compromises
the privacy of W4 at database 1. Alternatively, the user queries the MDS coded version of
the 7 symbols. By using these 7 symbols as the systematic part, we can use a (13, 7) MDS
code. By downloading the 6 parity symbols, the user can reconstruct the whole 7 symbols
utilizing the knowledge of d1 . Therefore, the normalized download cost for our achievability
= 23 , which matches the lower bound in (11) for this case.
scheme is 6+6
8
For database 1, the query table in Table 1 induces the same distribution on the messages
W1 , W2 and W4 . Therefore, we guarantee the privacy of the desired message. The reliability
constraint can also be verified. Note that b2 is downloaded from database 2, and d2 is
downloaded in the prefetching phase. Therefore, a3 and a4 are decodable. By getting b4 + c3
from database 2, the user can get b4 due to the private side information W3 . Therefore, the
user can decode a7 from a7 + b4 + d4 . Similar arguments follow for database 2.
Table 1: Query table for K = 4, N = 2, M = 2.
DB1
DB2
a1
a2
b1
b2
d1
c1
a3 + b2
a5 + b1
a4 + d2
a6 + c2
b3 + d3
b4 + c3
a7 + b4 + d4 a8 + b3 + c4
WH1 ={W3 }
5.1.2
WH2 ={W4 }
N = 2 Databases, K = 5 Messages, and M = 2 Cached Messages
Assume that each message is of size 16 symbols. We use ai , bi , ci , di and ei , for i = 1, . . . , 16,
to denote the symbols of messages W1 , W2 , W3 , W4 , and W5 , respectively. In this example,
in the prefetching phase, the user caches message W4 from database 1, and message W5 from
database 2; and in the retrieval phase, the user wishes to retrieve message W1 privately. The
user first generates the query table in Table 2. In Table 2, the user queries 15 symbols. Since
the user knows e1 from the cached message W5 , in order to use the partially known private
side information, the user in fact queries the MDS coded version of the 15 symbols. By using
these 15 symbols as the systematic part, we can use a (29, 15) MDS code. By downloading the
14 parity symbols, the user can reconstruct the whole 15 symbols. Therefore, the normalized
download cost for our achievability scheme is 14+14
= 47 , which matches the lower bound in
16
(11) for this case.
For database 1, the query table in Table 2 induces the same distribution on the messages
W1 , W2 , W3 and W5 . Therefore, we guarantee the privacy of the desired message. The
13
reliability constraint can also be verified. Note that b2 , c2 are downloaded from database 2,
and e2 is downloaded in the prefetching phase. Therefore, a3 , a4 and a5 are decodable. By
getting b6 + d3 from database 2, the user can get b6 due to the private side information W4 .
Similarly, c6 is also decodable. Therefore, the user can decode a10 from a10 + b6 + e5 and a11
from a11 + c6 + e6 . By getting b8 + c8 + d7 from database 2, the user can get b8 + c8 due to
the private side information W4 . Therefore, the user can decode a15 from a15 + b8 + c8 + e8 .
Similar arguments follow for database 2.
Table 2: Query table for K = 5, N = 2, M = 2.
DB1
DB2
a1
a2
b1
b2
c1
c2
e1
d1
a3 + b2
a6 + b1
a4 + c2
a7 + c1
a5 + e2
a8 + d2
b3 + c3
b5 + c5
b4 + e3
b6 + d3
c4 + e4
c6 + d 4
a9 + b5 + c5
a12 + b3 + c3
a10 + b6 + e5
a13 + b4 + d5
a11 + c6 + e6
a14 + c4 + d6
b7 + c7 + e7
b8 + c8 + d7
a15 + b8 + c8 + e8 a16 + b7 + c7 + d8
WH1 ={W4 }
5.2
WH2 ={W5 }
General Achievable Scheme for
M
N
∈N
Let M
= m. In the prefetching phase, the user caches m messages from each database. To
N
achieve the lower bound shown in (11), in the retrieval phase, we choose the message size as
L = N K−m symbols. The details of the achievable scheme are as follows:
1. Initialization: The user permutes each message randomly and independently. After the
random permutation, we use Ui (j) to denote the jth symbol of the permuted message
Wi . Suppose the user wishes to retrieve Wθ privately. We then prepare the query table
by first querying Uθ (1) from database 1. Set the round index to r = 1.
2. Symmetry across databases: The user queries the same number of equations with the
same structure as database 1 from the remaining databases.
3. Message symmetry: For each database, to satisfy the privacy constraint, the user
should query equal amount of symbols from all other K − m messages. Since the user
14
has cached m messages from each database in the prefetching phase, the user does not
need to protect the privacy for these m messages. For the rth round, the user queries
sums of every r combinations of the K − m messages.
4. Exploiting side information: For database 1, the user exploits the side information
equations obtained from the other (N −1) databases to query sum of r+1 combinations
of the K − m messages, where sum of r combinations is the side information. If the r
combinations contain the cached message from database 1, we replace the overlapping
symbols through the symbols cached from other databases.
5. Repeat steps 2, 3, 4 after setting r = r + 1 until r = K − m + 1.
6. Shuffling the order of queries: By shuffling the order of queries uniformly, all possible
queries can be made equally likely regardless of the message index. This guarantees
the privacy of the desired message.
7. Downloading MDS parity parts: Now, the query table is finished. For each database,
let p be the number of queried symbols in the query table, and let q be the number of
queried symbols which are determined by the side information the user cached in the
prefetching phase. Apply a (2p − q, p) MDS code to the queried symbols by letting the
p symbols to be the systematic part. Finally, the user downloads the parity parts of
the MDS-coded answering strings which are p − q symbols for each database.
5.3
Normalized Download Cost
We now calculate the total number of downloaded symbols. We first calculate p, which is
the number of queried symbols in the query table for each database,
K−m
K −m
K−m
(N − 1)K−m−1
(49)
(N − 1) + · · · +
+
p=
K−m
2
1
K−m
K −m
K −m
1
K−m
2
(50)
(N − 1)
(N − 1) + · · · +
(N − 1) +
=
K−m
2
1
N −1
1
=
N K−m − 1 ,
(51)
N −1
where K−m
in (49) corresponds to the queries of sums of every r combinations of the K −m
r
messages, and (N − 1)r−1 corresponds to the number of sets of the available side information
from other (N − 1) databases.
We then calculate q, which is the number of queried symbols which are determined by
the side information the user cached in the prefetching phase,
q=
(N − 1)m
(N − 1)m
(N − 1)m
(N − 1)(N −1)m−1
(N − 1) + · · · +
+
(N − 1)m
2
1
15
(52)
(N − 1)m
(N − 1)m
1
(N −1)m
(N − 1)
(N − 1) + · · · +
=
(N − 1)m
1
N −1
1
=
N (N −1)m − 1 ,
N −1
(53)
(54)
where (N −1)m
in (52) corresponds to the queries which can be determined by the partially
r
known private side information, and (N − 1)r−1 corresponds to the number of sets of queries
consisting of r combinations.
Next, we calculate the number of symbols for the desired message,
L=N
K −m−1
K −m−1
K −m−1
K−m−1
(N − 1)
(N − 1) + · · · +
+
K −m−1
1
0
(55)
= N × N K−m−1 = N K−m ,
(56)
where K−m−1
in (55) corresponds to the queries containing the desired message and (N −
r−1
1)r−1 corresponds to the number of sets of queries consisting of r combinations.
Therefore, the normalized download cost becomes,
D
N(p − q)
=
L
L
N
N
K−m
(N −1)m
N
−
1
−
N
−
1
N −1
= N −1
N K−m
N K−m − N (N −1)m
N
×
=
N −1 "
N K−m
K−M #
1
1
,
=
1 × 1−
N
1− N
(57)
(58)
(59)
(60)
which matches the lower bound in (11).
Remark 6 Note that although our achievable scheme and the scheme in [32] are both using
MDS coding to exploit the available side information, the field size requirements for realizing
the MDS codes are different. For the scheme of [32], a (2p̃ − q̃, p̃) MDS code is used, where
p̃ = N 1−1 (N K − 1) and q̃ = N 1−1 (N M − 1). This requires larger field size than the (2p − q, p)
∈ N), since 2p̃ − q̃ > (2p − q).
MDS code used in our scheme (if M
N
6
Conclusion
In this paper, we have introduced a new PIR model, namely, PIR with partially known
private side information as a natural model for studying practical PIR problems with cached
side information. In this model, the user and the databases engage in a caching/PIR scenario which consists of two phases, namely, prefetching phase and retrieval phase. The nth
16
database provides the user with mn side information messages in the prefetching phase such
P
that N
n=1 mn ≤ M, hence, each database has partial knowledge about the side information
in contrast to full knowledge in [29] and no knowledge in [30–32]. Based on this side information, the user designs a retrieval scheme that does not reveal the identity of the desired
message or the identities of the remaining M − mn messages to the nth database. For this
1− 1
model, we determined the exact capacity to be C = 1−( 1 )NK−M . The capacity is attained
N
for any prefetching strategy that satisfies the cache memory size constraint with equality.
The achievable scheme in [32] can also be used for this model. We further proposed another
PIR scheme which requires smaller sub-packetization and field size for the case of uniform
prefetching. Uniform prefetching, when feasible, is optimal. Interestingly, the capacity expression we derive for this problem is exactly the same as the capacity expression for the
PIR problem with completely unknown side information [32]. Therefore, our result implies
that there is no loss in employing the same databases for prefetching and retrieval purposes.
References
[1] B. Chor, E. Kushilevitz, O. Goldreich, and M. Sudan. Private information retrieval.
Journal of the ACM, 45(6):965–981, 1998.
[2] W. Gasarch. A survey on private information retrieval. In Bulletin of the EATCS, 2004.
[3] C. Cachin, S. Micali, and M. Stadler. Computationally private information retrieval
with polylogarithmic communication. In International Conference on the Theory and
Applications of Cryptographic Techniques. Springer, 1999.
[4] R. Ostrovsky and W. Skeith III. A survey of single-database private information retrieval: Techniques and applications. In International Workshop on Public Key Cryptography, pages 393–411. Springer, 2007.
[5] S. Yekhanin. Private information retrieval. Communications of the ACM, 53(4):68–73,
2010.
[6] N. B. Shah, K. V. Rashmi, and K. Ramchandran. One extra bit of download ensures
perfectly private information retrieval. In IEEE ISIT, June 2014.
[7] G. Fanti and K. Ramchandran. Efficient private information retrieval over unsynchronized databases. IEEE Journal of Selected Topics in Signal Processing, 9(7):1229–1239,
October 2015.
[8] T. Chan, S. Ho, and H. Yamamoto. Private information retrieval for coded storage. In
IEEE ISIT, June 2015.
17
[9] A. Fazeli, A. Vardy, and E. Yaakobi. Codes for distributed pir with low storage overhead.
In IEEE ISIT, June 2015.
[10] R. Tajeddine and S. El Rouayheb. Private information retrieval from MDS coded data
in distributed storage systems. In IEEE ISIT, July 2016.
[11] H. Sun and S. A. Jafar. The capacity of private information retrieval. In IEEE Globecom,
December 2016.
[12] H. Sun and S. A. Jafar. The capacity of private information retrieval. IEEE Trans. on
Info. Theory, 63(7):4075–4088, July 2017.
[13] H. Sun and S. Jafar. Blind interference alignment for private information retrieval. 2016.
Available at arXiv:1601.07885.
[14] H. Sun and S. Jafar. The capacity of robust private information retrieval with colluding
databases. 2016. Available at arXiv:1605.00635.
[15] R. Tajeddine, O. W. Gnilke, D. Karpuk, R. Freij-Hollanti, C. Hollanti, and S. El Rouayheb. Private information retrieval schemes for coded data with arbitrary collusion patterns. 2017. Available at arXiv:1701.07636.
[16] H. Sun and S. Jafar. The capacity of symmetric private information retrieval. 2016.
Available at arXiv:1606.08828.
[17] K. Banawan and S. Ulukus. The capacity of private information retrieval from coded
databases. IEEE Trans. on Info. Theory. Submitted September 2016. Also available at
arXiv:1609.08138.
[18] H. Sun and S. Jafar. Optimal download cost of private information retrieval for arbitrary
message length. 2016. Available at arXiv:1610.03048.
[19] H. Sun and S. Jafar. Multiround private information retrieval: Capacity and storage
overhead. 2016. Available at arXiv:1611.02257.
[20] K. Banawan and S. Ulukus. Multi-message private information retrieval: Capacity
results and near-optimal schemes. IEEE Trans. on Info. Theory. Submitted February
2017. Also available at arXiv:1702.01739.
[21] K. Banawan and S. Ulukus. The capacity of private information retrieval from Byzantine
and colluding databases. IEEE Trans. on Info. Theory. Submitted June 2017. Also
available at arXiv:1706.01442.
[22] Q. Wang and M. Skoglund. Symmetric private information retrieval for MDS coded
distributed storage. 2016. Available at arXiv:1610.04530.
18
[23] R. Freij-Hollanti, O. Gnilke, C. Hollanti, and D. Karpuk. Private information retrieval
from coded databases with colluding servers. 2016. Available at arXiv:1611.02062.
[24] H. Sun and S. Jafar. Private information retrieval from MDS coded data with colluding servers: Settling a conjecture by Freij-Hollanti et al. 2017. Available at arXiv:
1701.07807.
[25] Y. Zhang and G. Ge. A general private information retrieval scheme for MDS coded
databases with colluding servers. 2017. Available at arXiv: 1704.06785.
[26] Y. Zhang and G. Ge. Multi-file private information retrieval from MDS coded databases
with colluding servers. 2017. Available at arXiv: 1705.03186.
[27] Q. Wang and M. Skoglund. Linear symmetric private information retrieval for MDS
coded distributed storage with colluding servers. 2017. Available at arXiv:1708.05673.
[28] Q. Wang and M. Skoglund. Secure symmetric private information retrieval from colluding databases with adversaries. 2017. Available at arXiv:1707.02152.
[29] R. Tandon. The capacity of cache aided private information retrieval. 2017. Available
at arXiv: 1706.07035.
[30] S. Kadhe, B. Garcia, A. Heidarzadeh, S. El Rouayheb, and A. Sprintson. Private
information retrieval with side information. 2017. Available at arXiv:1709.00112.
[31] Y.-P. Wei, K. Banawan, and S. Ulukus. Fundamental limits of cache-aided private
information retrieval with unknown and uncoded prefetching. 2017. Available at
arXiv:1709.01056.
[32] Z. Chen, Z. Wang, and S. Jafar. The capacity of private information retrieval with
private side information. 2017. Available at arXiv:1709.03022.
[33] M. Karmoose, L. Song, M. Cardone, and C. Fragouli. Private broadcasting: an index
coding approach. 2017. Available at arXiv: 1701.04958.
[34] M. Karmoose, L. Song, M. Cardone, and C. Fragouli. Preserving privacy while broadcasting: k-limited-access schemes. 2017. Available at arXiv:1705.08437.
19
| 7cs.IT
|
M ONOMIAL IDEALS VIA SQUARE - FREE MONOMIAL IDEALS
arXiv:math/0507238v2 [math.AC] 10 Mar 2017
Sara Faridi∗
2005
1 Introduction
In this paper we study monomial ideals using the operation “polarization” to first turn them into
square-free monomial ideals. Various forms of polarization appear throughout the literature and
have been used for different purposes in algebra and algebraic combinatorics (for example, Weyman [17], Fröberg [8], Schwartau [13], or Rota and Stein [11]). One of the most useful features
of polarization is that the chain of substitutions that turn a given monomial ideal into a square-free
one can be described in terms of a regular sequence (Fröberg [8]). This fact allows many properties of a monomial ideal to transfer to its polarization. Conversely, to study a given monomial
ideal, one could examine its polarization. The advantage of this latter approach is that there are
many combinatorial tools dealing with square-free monomial ideals. One of these tools is StanleyReisner theory: Schwartau’s thesis [13] and the book by Stückrad and Vogel [15] discuss how the
Stanley-Reisner theory of square-free monomial ideals produces results about general monomial
ideals using polarization. Another tool for studying square-free monomial ideals, which will be our
focus here, is facet ideal theory, developed by the author in [5], [6] and [7].
This paper is organized as follows. In Section 2 we define polarization and introduce some
of its basic properties. In Section 3 we introduce facet ideals and its features that are relevant to
this paper. In particular, we introduce simplicial trees, which correspond to square-free monomial
ideals with exceptionally strong algebraic properties. Section 4 extends the results of facet ideal
theory to general monomial ideals. Here we study a monomial ideal I whose polarization is a
tree, and show that many of the properties of simplicial trees hold for such ideals. This includes
Cohen-Macaulayness of the Rees ring of I (Corollary 4.8), I being sequentially Cohen-Macaulay
(Corollary 4.12), and several inductive tools for studying such ideals, such as localization (see
Section 4.1).
Appendix A is an independent study of primary decomposition in a sequentially Cohen-Macaulay
module. We demonstrate how in a sequentially Cohen-Macaulay module M , every submodule
appearing in the filtration of M can be described in terms of the primary decomposition of the
0-submodule of M . This is used to prove Proposition 4.11.
∗
This research was supported by NSERC, and by Laboratoire de combinatoire et d’informatique mathématique at
UQÀM, Montreal.
1
2 Polarization
Definition 2.1. Let R = k[x1 , . . . , xn ] be a polynomial ring over a field k. Suppose M =
x1 a1 . . . xn an is a monomial in R. Then we define the polarization of M to be the square-free
monomial
P(M ) = x1,1 x1,2 . . . x1,a1 x2,1 . . . x2,a2 . . . xn,1 . . . xn,an
in the polynomial ring S = k[xi,j | 1 ≤ i ≤ n, 1 ≤ j ≤ ai ].
If I is an ideal of R generated by monomials M1 , . . . , Mq , then the polarization of I is defined
as:
P(I) = P(M1 ), . . . , P(Mq )
which is a square-free monomial ideal in a polynomial ring S.
Here is an example of how polarization works.
Example 2.2. Let J = (x1 2 , x1 x2 , x2 3 ) ⊆ R = k[x1 , x2 ]. Then
P(J) = (x1,1 x1,2 , x1,1 x2,1 , x2,1 x2,2 x2,3 )
is the polarization of J in the polynomial ring
S = k[x1,1 , x1,2 , x2,1 , x2,2 , x2,3 ]
Note that by identifying each xi with xi,1 , one can consider S as a polynomial extension of R.
Exactly how many variables S has will always depend on what we polarize. Therefore, as long as
we are interested in the polarizations of finitely many monomials and ideals, S remains a finitely
generated algebra.
Below we describe some basic properties of polarization, some of which appear (without proof)
in [15]. Here we record the proofs where appropriate.
Proposition 2.3 (basic properties of polarization). Suppose that R = k[x1 , . . . , xn ] is a polynomial
ring over a field k, and I and J are two monomial ideals of R.
1. P(I + J) = P(I) + P(J);
2. For two monomials M and N in R, M | N if and only if P(M ) | P(N );
3. P(I ∩ J) = P(I) ∩ P(J);
4. If p = (xi1 , . . . , xir ) is a (minimal) prime containing I, then P(p) is a (minimal) prime
containing P(I);
5. If p′ = (xi1 ,e1 , . . . , xir ,er ) is a prime over P(I), then p = (xi1 , . . . , xir ) is a prime over I.
Moreover, if p′ has minimal height (among all primes containing P(I)), then p must have
minimal height as well (among all primes containing I);
2
6. height I = height P(I);
1. Follows directly from Definition 2.1.
Proof.
2. Suppose that M = x1 b1 . . . xn bn and N = x1 c1 . . . xn cn , and suppose that
P(M ) = x1,1 . . . x1,b1 . . . xn,1 . . . xn,bn
and
P(N ) = x1,1 . . . x1,c1 . . . xn,1 . . . xn,cn .
If M | N , then bi ≤ ci for all i, which implies that P(M ) | P(N ). The converse is also clear
using the same argument.
3. Suppose that I = (M1 , . . . , Mq ) and J = (N1 , . . . , Ns ) where the generators are all monomials. If U = x1 b1 . . . xn bn is a monomial in I ∩ J, then for some generator Mi of I and
Nj and J, we have Mi | U and Nj | U , hence by part 2, P(Mi ) | P(U ) and P(Nj ) | P(U ),
which implies that P(U ) ∈ P(I) ∩ P(J).
Conversely, if U ′ is a monomial in P(I) ∩ P(J), then for some generator Mi = x1 b1 . . . xn bn
of I and Nj = x1 c1 . . . xn cn and J we have P(Mi ) | U ′ and P(Nj ) | U ′ . This means that
lcm(P(Mi ), P(Nj )) | U ′ . It is easy to see (by an argument similar to the one in part 2) that
lcm(P(Mi ), P(Nj )) = P(lcm(Mi , Nj )). Since lcm(Mi , Nj ) is one of the generators of
I ∩ J, it follows that P(lcm(Mi , Nj )) is a generator of P(I ∩ J) and hence U ′ ∈ P(I ∩ J).
4. If p = (xi1 , . . . , xir ) is a minimal prime over I = (M1 , . . . , Mq ), then for each of the xij
there is a Mt such that xij | Mt , and no other generator of p divides Mt . The same holds for the
polarization of the two ideals: P(p) = (xi1 ,1 , . . . , xir ,1 ) and P(I) = (P(M1 ), . . . , P(Mt )),
and so P(p) is minimal over P(I).
5. Suppose that p′ = (xi1 ,e1 , . . . , xir ,er ) is a prime lying over P(I). Then for every generator
Mt of I, there is a xij ,ej in p′ such that xij ,ej | P(Mt ). But this implies that xij | Mt , and
therefore I ⊆ p = (xi1 , . . . , xir ).
Now suppose that p′ has minimal height r over P(I), and there is a prime ideal q over I with
height q < r. This implies (from part 4) that P(q), which is a prime of height less than r,
contains P(I), which is a contradiction.
6. This follows from parts 4 and 5.
Example 2.4. It is not true that every minimal prime of P(I) comes from a minimal prime of I.
For example, let I = (x21 , x1 x22 ). Then
P(I) = (x1,1 x1,2 , x1,1 x2,1 x2,2 ).
The ideal (x1,2 , x2,1 ) is a minimal prime over P(I), but the corresponding prime (x1 , x2 ) is not a
minimal prime of I (however, if we had taken any minimal prime of minimal height of P(I), e.g.
(x1,1 ), then the corresponding prime over I would have been minimal; this is part 5 above).
3
For a monomial ideal I in a polynomial ring R = k[x1 , . . . , xn ] as above, there is a unique
irredundant irreducible decomposition of the form
I = q1 ∩ . . . ∩ qm
where each qi is a primary ideal generated by powers of the variables x1 , . . . , xn (see [16, Theorem 5.1.17]).
Proposition 2.5 (polarization and primary decomposition). Let I be a monomial ideal in a polynomial ring R = k[x1 , . . . , xn ], and let P(I) be the polarization of I in S = k[xi,j ] as described in
Definition 2.1.
1. If I = (xi1 a1 , . . . , xir ar ) where the aj are positive integers, then
\
P(I) =
(xi1 ,c1 , . . . , xir ,cr )
1≤cj ≤aj
1≤j≤r
2. If I = (xi1 , . . . , xir )m , where 1 ≤ i1 , . . . , ir ≤ n and m is a positive integer, then P(I) has
the following irredundant irreducible primary decomposition:
\
P(I) =
(xi1 ,c1 , . . . , xir ,cr )
1≤cj ≤m
Σcj ≤m+r−1
3. Suppose that I = q1 ∩ . . . ∩ qm is the unique irredundant irreducible primary decomposition
of I, such that for each i = 1, . . . , m,
i
i
qi = (x1 a1 , . . . , xn an ),
i
where the aij are nonnegative integers, and if aij = 0 we assume that xj aj = 0.
Then P(I) has the following irreducible primary decomposition (some primes might be repeated).
\
\
(x1,c1 , . . . , xn,cn )
P(I) =
1≤i≤m 1≤cj ≤ai
j
1≤j≤n
where when aij = 0, we assume that cj = xj,0 = 0.
Proof.
1. We know that
P(I) = (xi1 ,1 . . . xi1 ,a1 , . . . , xi1 ,1 . . . xi1 ,ar ).
Clearly the minimal primes of P(I) are (xi1 ,c1 , . . . , xir ,cr ) for all cj ≤ aj . This settles the
claim.
4
2. Assume, without loss of generality, that I = (x1 , . . . , xr )m . So we can write
I = x1 b1 . . . xr br | 0 ≤ bi ≤ m, b1 + · · · + br = m
so that
P(I) = x1,1 . . . x1,b1 . . . xr,1 . . . xr,br | 0 ≤ bi ≤ m, b1 + · · · + br = m .
We first show that P(I) is contained in the intersection of the ideals of the form (x1,c1 , . . . , xr,cr )
described above. It is enough to show this for each generator of P(I). So we show that
U = x1,1 . . . x1,b1 . . . xr,1 . . . xr,br ∈ I = (x1,c1 , . . . , xr,cr )
where 0 ≤ bi ≤ m, b1 + · · · + br = m, 1 ≤ cj ≤ m and c1 + · · · + cr ≤ m + r − 1.
If for any i, bi ≥ ci , then it would be clear that U ∈ I.
Assume bi ≤ ci − 1 for i = 1, . . . , r − 1. It follows that
m − br = b1 + · · · + br−1
≤ c1 + · · · + cr−1 − (r − 1)
≤ m + r − 1 − cr − (r − 1)
= m − cr
which implies that br ≥ cr , hence U ∈ I.
So far we have shown one direction of the inclusion.
To show the opposite direction, take any monomial
\
U ∈ (x1,c1 , . . . , xr,cr )
where 1 ≤ cj ≤ m and c1 + · · · + cr ≤ m + r − 1.
Notice that for some i ≤ r, xi,1 | U ; this is because U ∈ (x1,1 , . . . , xr,1 ).
We write U as
U = x1,1 . . . x1,b1 . . . xr,1 . . . xr,br U ′
where U ′ is a monomial, and the bi are nonnegative integers such that for each j < bi , xi,j | U
(if xi,1 6 | U then set bi = 0). We need to show that it is possible to find such bi so that
b1 + · · · + br = m.
Suppose b1 + · · · + br ≤ m − 1, and xi,bi +1 6 | U for 1 ≤ i ≤ r. Then
b1 + · · · + br + r ≤ m + r − 1,
hence
U ∈ (x1,b1 +1 , . . . , xr,br +1 )
implying that xi,bi +1 | U for some i, which is a contradiction.
Therefore b1 , . . . , br can be picked so that they add up to m, and hence U ∈ P(I); this settles
the opposite inclusion.
5
3. This follows from part 1 and Proposition 2.3 part 3.
Corollary 2.6 (polarization and associated primes). Let I be a monomial ideal in a polynomial ring
R = k[x1 , . . . , xn ], and let P(I) be its polarization in S = k[xi,j ] as described in Definition 2.1.
Then (xi1 , . . . , xir ) ∈ AssR (R/I) if and only if (xi1 ,c1 , . . . , xir ,cr ) ∈ AssS (S/P(I)) for some
positive integers c1 , . . . , cr .
Example 2.7. Consider the primary decomposition of J = (x1 2 , x2 3 , x1 x2 ):
J = (x1 , x2 3 ) ∩ (x1 2 , x2 ).
By Proposition 2.5, P(J) = (x1,1 x1,2 , x2,1 x2,2 x2,3 , x1,1 x2,1 ) will have primary decomposition
P(J) = (x1,1 , x2,1 ) ∩ (x1,1 , x2,2 ) ∩ (x1,1 , x2,3 ) ∩ (x1,2 , x2,1 ).
A very useful property of polarization is that the final polarized ideal is related to the original
ideal via a regular sequence. The proposition below, which looks slightly different here than the
original statement in [8], states this fact.
Proposition 2.8 (Fröberg [8]). Let k be a field and
R = k[x1 , . . . , xn ]/(M1 , . . . , Mq ),
where M1 , . . . , Mq are monomials in the variables x1 , . . . , xn , and let
N1 = P(M1 ), . . . , Nq = P(Mq )
be a set of square-free monomials in the polynomial ring
S = k[xi,j | 1 ≤ i ≤ n, 1 ≤ j ≤ ai ]
such that for each i, the variable xi,ai appears in at least one of the monomials N1 , . . . , Nq . Then
the sequence of elements
xi,1 − xi,j where 1 ≤ i ≤ n and 1 < j ≤ ai
forms a regular sequence in the quotient ring
R′ = S/(N1 , . . . , Nq )
and if J is the ideal of R′ generated by the elements in (1), then
R = R′ /J.
Moreover, R is Cohen-Macaulay (Gorenstein) if and only if R′ is.
Example 2.9. Let J and R be as in Example 2.2. According to Proposition 2.8, the sequence
x1,1 − x1,2 , x2,1 − x2,2 , x2,1 − x2,3
is a regular sequence in S/P(J), and
R/J = S/ P(J) + (x1,1 − x1,2 , x2,1 − x2,2 , x2,1 − x2,3 ) .
6
(1)
3 Square-free monomial ideals as facet ideals
Now that we have introduced polarization as a method of transforming a monomial ideal into a
square-free one, we can focus on square-free monomial ideals. In particular, here we are interested
in properties of square-free monomial ideals that come as a result of them being considered as facet
ideals of simplicial complexes. Below we review the basic definitions and notations in facet ideal
theory, as well as some of the basic concepts of Stanley-Reisner theory. We refer the reader to [2],
[5], [6], [7], and [14] for more details and proofs in each of these topics.
Definition 3.1 (simplicial complex, facet, subcollection and more). A simplicial complex ∆ over a
set of vertices V = {v1 , . . . , vn } is a collection of subsets of V , with the property that {vi } ∈ ∆ for
all i, and if F ∈ ∆ then all subsets of F are also in ∆ (including the empty set). An element of ∆ is
called a face of ∆, and the dimension of a face F of ∆ is defined as |F |− 1, where |F | is the number
of vertices of F . The faces of dimensions 0 and 1 are called vertices and edges, respectively, and
dim ∅ = −1. The maximal faces of ∆ under inclusion are called facets of ∆. The dimension of the
simplicial complex ∆ is the maximal dimension of its facets.
We denote the simplicial complex ∆ with facets F1 , . . . , Fq by
∆ = hF1 , . . . , Fq i
and we call {F1 , . . . , Fq } the facet set of ∆. A simplicial complex with only one facet is called a
simplex. By a subcollection of ∆ we mean a simplicial complex whose facet set is a subset of the
facet set of ∆.
Definition 3.2 (connected simplicial complex). A simplicial complex ∆ = hF1 , . . . , Fq i is connected if for every pair i, j, 1 ≤ i < j ≤ q, there exists a sequence of facets Ft1 , . . . , Ftr of ∆ such
that Ft1 = Fi , Ftr = Fj and Fts ∩ Fts+1 6= ∅ for s = 1, . . . , r − 1.
Definition 3.3 (facet/non-face ideals and complexes). Consider a polynomial ring R = k[x1 , . . . , xn ]
over a field k and a set of indeterminates x1 , . . . , xn . Let I = (M1 , . . . , Mq ) be an ideal in R, where
M1 , . . . , Mq are square-free monomials that form a minimal set of generators for I.
• The facet complex of I, denoted by δF (I), is the simplicial complex over a set of vertices
v1 , . . . , vn with facets F1 , . . . , Fq , where for each i, Fi = {vj | xj |Mi , 1 ≤ j ≤ n}. The
non-face complex or the Stanley-Reisner complex of I, denoted by δN (I) is be the simplicial
complex over a set of vertices v1 , . . . , vn , where {vi1 , . . . , vis } is a face of δN (I) if and only
if xi1 . . . xis ∈
/ I.
• Conversely, if ∆ is a simplicial complex over n vertices labeled v1 , . . . , vn , we define the
facet ideal of ∆, denoted by F(∆), to be the ideal of R generated by square-free monomials xi1 . . . xis , where {vi1 , . . . , vis } is a facet of ∆. The non-face ideal or the StanleyReisner ideal of ∆, denoted by N (∆), is the ideal of R generated by square-free monomials
xi1 . . . xis , where {vi1 , . . . , vis } is not a face of ∆.
7
Throughout this paper we often use a letter x to denote both a vertex of ∆ and the corresponding
variable appearing in F(∆), and xi1 . . . xir to denote a facet of ∆ as well as a monomial generator
of F(∆).
Example 3.4. If ∆ is the simplicial complex hxyz, yu, uvwi drawn below,
y
x
u
z
w
v
then F(∆) = (xyz, yu, uvw) and N (∆) = (xu, xv, xw, yv, yw, zu, zv, zw) are its facet ideal and
nonface (Stanley-Reisner) ideal, respectively.
Facet ideals give a one-to-one correspondence between simplicial complexes and square-free
monomial ideals.
Next we define the notion of a vertex cover. The combinatorial idea here comes from graph
theory. In algebra, it corresponds to prime ideals lying over the facet ideal of a given simplicial
complex.
Definition 3.5 (vertex covering, independence, unmixed). Let ∆ be a simplicial complex with vertex set V . A vertex cover for ∆ is a subset A of V that intersects every facet of ∆. If A is a minimal
element (under inclusion) of the set of vertex covers of ∆, it is called a minimal vertex cover. The
smallest of the cardinalities of the vertex covers of ∆ is called the vertex covering number of ∆ and
is denoted by α(∆). A simplicial complex ∆ is unmixed if all of its minimal vertex covers have the
same cardinality.
A set {F1 , . . . , Fu } of facets of ∆ is called an independent set if Fi ∩ Fj = ∅ whenever i 6= j.
The maximum possible cardinality of an independent set of facets in ∆, denoted by β(∆), is called
the independence number of ∆. An independent set of facets which is not a proper subset of any
other independent set is called a maximal independent set of facets.
Example 3.6. If ∆ is the simplicial complex in Example 3.4, then the vertex covers of ∆ are:
{x, u}, {y, u}, {y, v}, {y, w}, {z, u}, {x, y, u}, {x, z, u}, {x, y, v}, . . . .
The first five vertex covers above (highlighted in bold), are the minimal vertex covers of ∆. It
follows that α(∆) = 2, and ∆ is unmixed. On the other hand, {xyz, uvw} is the largest maximal
independent set of facets that ∆ contains, and so β(∆) = 2.
Definition 3.7 (Alexander dual). Let I be a square-free monomial ideal in the polynomial ring k[V ]
with V = {x1 , . . . , xn }, and let ∆N be the non-face complex of I (i.e. ∆N = δN (I)). Then the
Alexander dual of ∆N is the simplicial complex
∆N ∨ = {F ⊂ V | F c ∈
/ ∆N }
where F c is the complement of the face F in V .
We call the nonface ideal of ∆N ∨ the Alexander dual of I and denote it by I ∨ .
8
3.1 Simplicial Trees
Considering simplicial complexes as higher dimensional graphs, one can define the notion of a tree
by extending the same concept from graph theory. Before we define a tree, we determine what
“removing a facet” from a simplicial complex means. We define this idea so that it corresponds to
dropping a generator from the facet ideal of the complex.
Definition 3.8 (facet removal). Suppose ∆ is a simplicial complex with facets F1 , . . . , Fq and
F(∆) = (M1 , . . . , Mq ) its facet ideal in R = k[x1 , . . . , xn ]. The simplicial complex obtained by
removing the facet Fi from ∆ is the simplicial complex
∆ \ hFi i = hF1 , . . . , F̂i , . . . , Fq i
and F(∆ \ hFi i) = (M1 , . . . , M̂i , . . . , Mq ).
The definition that we give below for a simplicial tree is one generalized from graph theory. See
[5] and [6] for more on this concept.
Definition 3.9 (leaf, joint). A facet F of a simplicial complex is called a leaf if either F is the only
facet of ∆, or for some facet G ∈ ∆ \ hF i we have
F ∩ (∆ \ hF i) ⊆ G.
If F ∩ G 6= ∅, the facet G above is called a joint of the leaf F .
Equivalently, a facet F is a leaf of ∆ if F ∩ (∆ \ hF i) is a face of ∆ \ hF i.
Example 3.10. Let I = (xyz, yzu, zuv). Then F = xyz is a leaf, but H = yzu is not, as one can
see in the picture below.
x
y
F ∩ (∆ \ hF i) =
y
∩
z
y
H ∩ (∆ \ hHi) =
z
z
∩
y
=
u
x
u
z
v
y
v
z
y
u
=
z
u
Definition 3.11 (tree, forest). A connected simplicial complex ∆ is a tree if every nonempty subcollection of ∆ has a leaf. If ∆ is not necessarily connected, but every subcollection has a leaf, then
∆ is called a forest.
Example 3.12. The simplicial complexes in examples 3.4 and 3.10 are both trees, but the one below
is not because it has no leaves. It is an easy exercise to see that a leaf must contain a free vertex,
where a vertex is free if it belongs to only one facet.
9
One of the most powerful properties of simplicial trees from the point of view of algebra is that
they behave well under localization. This property makes it easy to use induction on the number of
vertices of a tree for proving its various properties.
Lemma 3.13 (Localization of a tree is a forest). Let I ⊆ k[x1 , . . . , xn ] be the facet ideal of a
simplicial tree, where k is a field. Then for any prime ideal p of k[x1 , . . . , xn ], δF (Ip ) is a forest.
Proof. See [6, Lemma 4.5].
4 properties of monomial ideals via polarization
For the purpose of all discussions in this section, unless otherwise stated, let I be a monomial ideal
in the polynomial ring R = k[x1 , . . . , xn ] over a field k, whose polarization is the square-free
monomial ideal P(I) in the polynomial ring
S = k[xi,j | 1 ≤ i ≤ n, 1 ≤ j ≤ ai ].
We assume that the polarizing sequence (as described in (1) in Proposition 2.8) is
ν = ν1 , . . . , νv
which is a regular sequence in S/P(I) and
R/I = S/ P(I) + (ν) .
4.1 Monomial ideals whose polarization is a simplicial tree
A natural question, and one that this paper is mainly concerned with, is what properties of facet
ideals of simplicial trees can be extended to general (non-square-free) monomial ideals using polarization? In other words, if for a monomial ideal I in a polynomial ring P(I) is the facet ideal of a
tree (Definition 3.11), then what properties of P(I) are inherited by I?
The strongest tool when dealing with square-free monomial ideals is induction– either on the
number of generators, or the number of variables in the ambient polynomial ring. This is particularly
the case when the facet complex of the ideal is a tree, or in some cases when it just has a leaf. In
this section we show that via polarization, one can extend these tools to monomial ideals in general.
For a given monomial ideal I, we show that if P(I) is the facet ideal of a tree, and p is a prime
ideal containing I, then P(Ip ) is the facet ideal of a forest (Theorem 4.1); this allows induction on
number of variables. Similarly, Theorem 4.3 provides us with a way to use induction on number of
generators of I.
Theorem 4.1 (localization and polarization). If P(I) is the facet ideal of a tree, and p is a prime
ideal of R containing I, then P(Ip ) is the facet ideal of a forest.
10
Proof. The first step is to show that it is enough to prove this for prime ideals of R generated by a
subset of {x1 , . . . , xn }. To see this, assume that p is a prime ideal of R containing I, and that p′
is another prime of R generated by all xi ∈ {x1 , . . . , xn } such that xi ∈ p (recall that the minimal
primes of I are generated by subsets of {x1 , . . . , xn }; see [16, Corollary 5.1.5]). So p′ ⊆ p. If
I = (M1 , . . . , Mq ), then
Ip′ = (M1 ′ , . . . , Mq ′ )
where for each i, Mi ′ is the image of Mi in Ip′ . In other words, Mi ′ is obtained by dividing
Mi = x1 a1 . . . xn an by the product of all the xj aj such that xj ∈
/ p′ . But xj ∈
/ p′ implies that
′
xj ∈
/ p, and so it follows that Mi ∈ Ip . Therefore Ip′ ⊆ Ip . On the other hand since p′ ⊆ p,
Ip ⊆ Ip′ , which implies that Ip′ = Ip (the equality and inclusions of the ideals here mean equality
and inclusion of their generating sets).
Now suppose I = (M1 , . . . , Mq ), and p = (x1 , . . . , xr ) is a prime containing I. Suppose that
for each i, we write Mi = Mi′ .Mi′′ , where
Mi′ ∈ k[x1 , . . . , xr ] and Mi′′ ∈ k[xr+1 , . . . , xn ]
so that
Ip = (M1′ , . . . , Mt′ ),
where without loss of generality M1′ , . . . , Mt′ is a minimal generating set for Ip .
We would like to show that the facet complex ∆ of P(Ip ) is a forest. Suppose that, again without
loss of generality,
I ′ = (P(M1′ ), . . . , P(Ms′ ))
is the facet ideal of a subcollection ∆′ of ∆. We need to show that ∆′ has a leaf.
If s = 1, then there is nothing to prove. Otherwise, suppose that P(M1 ) represents a leaf of the
tree δF (P(I)), and P(M2 ) is a joint of P(M1 ). Then we have
P(M1 ) ∩ P(Mi ) ⊆ P(M2 ) f or all i ∈ {2, . . . , s}.
Now let xe,f be in P(M1′ ) ∩ P(Mi′ ) for some i ∈ {2, . . . , s}. This implies that
(i) xe,f ∈ P(M1 ) ∩ P(Mi ) ⊆ P(M2 ), and
(ii) e ∈ {1 . . . , r}
From (i) and (ii) we can conclude that xe,f ∈ P(M2′ ), which proves that P(M1′ ) is a leaf for ∆′ .
Remark 4.2. It is not true in general that if p is a (minimal) prime of I, then P(Ip ) = P(I)P(p) .
For example, if I = (x1 3 , x1 2 x2 ) and p = (x1 ), then Ip = (x1 2 ) so P(Ip ) = (x1,1 x1,2 ), but
P(I)P(p) = (x1,1 ).
Another feature of simplicial trees is that they satisfy a generalization of König’s theorem ([6,
Theorem 5.3]). Below we explain how this property, and another property of trees that is very useful
for induction, behave under polarization.
11
Recall that for a simplicial complex ∆, α(∆) and β(∆) are the vertex covering number and
the independence number of ∆, respectively (Definition 3.5). For simplicity of notation, if I =
(M1 , . . . , Mq ) is a monomial ideal, we let β(I) denote the maximum cardinality of a subset of
{M1 , . . . , Mq } consisting of pairwise coprime elements (so β(∆) = β(F(∆)) for any simplicial
complex ∆).
Theorem 4.3 (joint removal and polarization). Suppose M1 , . . . , Mq are monomials that form a
minimal generating set for I, and P(I) is the facet ideal of a simplicial complex ∆. Assume that ∆
has a leaf, whose joint corresponds to P(M1 ). Then, if we let I ′ = (M2 , . . . , Mq ), we have
height I = height I ′ .
Proof. If G is the joint of ∆ corresponding to P(M1 ), then P(I ′ ) = F(∆ \ hGi). From [6,
Lemma 5.1] it follows that α(∆) = α(∆ \ hGi), so that height P(I) = height P(I ′ ), and therefore
height I = height I ′ .
Theorem 4.4. Suppose M1 , . . . , Mq are monomials that form a minimal generating set for I, and
P(I) is the facet ideal of a simplicial tree ∆. Then height I = β(I).
Proof. We already know that height I = height P(I) = α(∆). It is also clear that β(I) =
β(P(I)), since the monomials in a subset {Mi1 , . . . , Mir } of the generating set of I are pairwise
coprime if and only if the monomials in {P(Mi1 ), . . . , P(Mir )} are pairwise coprime. On the other
hand, from [6, Theorem 5.3] we know that α(∆) = β(∆). Our claim follows immediately.
We demonstrate how to apply these theorems via an example.
Example 4.5. Suppose I = (x1 3 , x1 2 x2 x3 , x3 2 , x2 3 x3 ). Then
P(I) = (x1,1 x1,2 x1,3 , x1,1 x1,2 x2,1 x3,1 , x3,1 x3,2 , x2,1 x2,2 x2,3 x3,1 )
is the facet ideal of the following simplicial complex (tree) ∆.
x 3,1
x 3,2
x 1,3
x 2,3
x 1,1
x 1,2
x 2,1
x 2,2
Now α(∆) = height I = 2 because the prime of minimal height over I is (x1 , x3 ). From
Theorem 4.4 it follows that β(I) = 2. This means that you can find a set of two monomials in the
generating set of I that have no common variables: for example {x1 3 , x3 2 } is such a set.
Since the monomials x1 2 x2 x3 and x2 3 x3 polarize into joints of ∆, by Theorem 4.3 the ideals
I, (x1 3 , x3 2 , x2 3 x3 ), (x1 3 , x1 2 x2 x3 , x3 2 ), and (x1 3 , x3 2 )
all have the same height.
12
We now focus on the Cohen-Macaulay property. In [6] we showed that for a simplicial tree ∆,
F(∆) is a Cohen-Macaulay ideal if and only if ∆ is an unmixed simplicial complex. The condition
unmixed for ∆ is equivalent to all minimal primes of the ideal F(∆) (which in this case are all
the associated primes of F(∆)) having the same height. In general, an ideal all whose associated
primes have the same height (equal to the height of the ideal) is called an unmixed ideal.
It now follows that
Theorem 4.6 (Cohen-Macaulay criterion for trees). Let P(I) be the facet ideal of a simplicial tree
∆. Then R/I is Cohen-Macaulay if and only if I is unmixed.
Proof. From Proposition 2.8, R/I is Cohen-Macaulay if and only if S/P(I) is Cohen-Macaulay.
By [6, Corollary 8.3], this is happens if and only if P(I) is unmixed. Corollary 2.6 now proves the
claim.
If R is a ring and J is an ideal of R, then the Rees ring of R along J is defined as
R[Jt] = ⊕n∈N J n tn .
Rees rings come up in the algebraic process of “blowing up” ideals. One reason that trees were
defined as they are, is that their facet ideals produce normal and Cohen-Macaulay Rees rings ([5]).
Proposition 4.7. If S[P(I)t] is Cohen-Macaulay, then so is R[It]. Conversely, if we assume that R
and S are localized at their irrelevant maximal ideals, then R[It] being Cohen-Macaulay implies
that S[P(I)t] is Cohen-Macaulay.
Proof. Suppose that ν1 , . . . , νv is the polarizing sequence as described before. For i = 1, . . . , v − 1
let
Ri = S/(ν1 , . . . , νi ), Ii = P(I)/(ν1 , . . . , νi ), Rv = R and Iv = I.
Notice that S[P(I)t] and R[It] are both domains. Also note that for each i,
S[P(I)t]/(ν1 , . . . , νi ) = Ri [Ii t]
is the Rees ring of the monomial ideal Ii in the polynomial ring Ri , and is therefore also a domain.
Therefore νi+1 is a regular element in the ring S[P(I)t]/(ν1 , . . . , νi ), which means that ν1 , . . . , νv
is a regular sequence in S[P(I)t].
Similarly, we see that
R[It] = S[P(I)t]/(ν1 , . . . , νv ).
[2, Theorem 2.1.3] now implies that if S[P(I)t] is Cohen-Macaulay, then so is R[It]. The
converse follows from [2, Exercise 2.1.28].
Corollary 4.8 (Rees ring of a tree is Cohen-Macaulay). Suppose that P(I) is the facet ideal of a
simplicial tree. Then the Rees ring R[It] of I is Cohen-Macaulay.
Proof. This follows from the Proposition 4.7, and from [5, Corollary 4], which states that the Rees
ring of the facet ideal of a simplicial tree is Cohen-Macaulay.
13
4.2 Polarization of sequentially Cohen-Macaulay ideals
The main result of this section is that if the polarization of a monomial ideal I is the facet ideal of
a tree, then I is a sequentially Cohen-Macaulay ideal. The theorem that implies this fact (Proposition 4.11) is interesting in its own right. For a square-free monomial ideal J, Eagon and Reiner
[4] proved that J is Cohen-Macaulay if and only if its Alexander dual J ∨ has a linear resolution.
Herzog and Hibi [9] then defined componentwise linear ideals and generalized their result, so that a
square-free monomial ideal J is sequentially Cohen-Macaulay if and only if J ∨ is componentwise
linear (see [9] or [7]). But even though Alexander duality has been generalized to all monomial
ideals from square-free ones, the criterion for sequential Cohen-Macaulayness does not generalize:
it is not true that if I is any monomial ideal, then I is sequentially Cohen-Macaulay if and only
of I ∨ is a componentwise linear ideal; see Miller [10]. We show that the statement is true if I ∨ is
replaced by P(I)∨ .
Definition 4.9 ([14, Chapter III, Definition 2.9]). Let M be a finitely generated Z-graded module
over a finitely generated N-graded k-algebra, with R0 = k. We say that M is sequentially CohenMacaulay if there exists a finite filtration
0 = M0 ⊂ M1 ⊂ . . . ⊂ Mr = M
of M by graded submodules Mi satisfying the following two conditions.
(a) Each quotient Mi /Mi−1 is Cohen-Macaulay.
(b) dim (M1 /M0 ) < dim (M2 /M1 ) < . . . < dim (Mr /Mr−1 ), where dim denotes Krull
dimension.
We define a componentwise linear ideal in the square-free case using [9, Proposition 1.5].
Definition 4.10 (componentwise linear). Let I be a square-free monomial ideal in a polynomial
ring R. For a positive integer k, the k-th square-free homogeneous component of I, denoted by
I[k] is the ideal generated by all square-free monomials in I of degree k. The ideal I above is said
to be componentwise linear if for all k, the square-free homogeneous component I[k] has a linear
resolution.
For our monomial ideal I, let Q(I) denote the set of primary ideals appearing in a reduced
primary decomposition of I. Suppose that h = height I and s = max{height q | q ∈ Q(I)}, and
set
\
Ii =
q.
q∈Q(I)
height q≤s−i
So we have the following filtration for R/I (we assume that all inclusions in the filtration are
proper; if there is an equality anywhere, we just drop all but one of the equal ideals).
0 = I = I0 ⊂ I1 ⊂ . . . ⊂ Is−h ⊂ R/I.
14
(2)
If R/I is sequentially Cohen-Macaulay, then by Theorem A.4, (2) is the appropriate filtration
that satisfies Conditions (a) and (b) in Definition 4.9.
For the square-free monomial ideal J = P(I), we similarly define
\
Q(J) = {minimal primes over J} and Ji =
p
p∈Q(J)
height p≤s−i
where the numbers h and s are the same as for I because of Proposition 2.5. It follows from
Proposition 2.5 and Corollary 2.6 that for each i, P(Ii ) = Ji and the polarization sequence that
transforms Ii into Ji is a subsequence of ν = ν1 , . . . , νv .
What we have done so far is to translate, via polarization, the filtration (2) of the quotient ring
R/I into one of S/J:
0 = J = J0 ⊂ J1 ⊂ . . . ⊂ Js−h ⊂ S/J.
(3)
Now note that for a given i, the sequence ν is a Ji+1 /Ji -regular sequence in S, as ν is a regular
sequence in S/Ji , which contains Ji+1 /Ji . Also note that
R/Ii ≃ S/(Ji + ν).
It follows that Ji+1 /Ji is Cohen-Macaulay if and only if Ii+1 /Ii is (see [2, Exercise 2.1.27(c),
Exercise 2.1.28, and Theorem 2.1.3]).
Proposition 4.11. The monomial ideal I is sequentially Cohen-Macaulay if and only if P(I) is
sequentially Cohen-Macaulay, or equivalently, P(I)∨ is a componentwise linear ideal.
Proof. By [7, Proposition 4.5], P(I)∨ is componentwise linear if and only if P(I) is sequentially Cohen-Macaulay, which by the discussion above is equivalent to I being sequentially CohenMacaulay.
Corollary 4.12 (Trees are sequentially Cohen-Macaulay). Let P(I) be the facet ideal of a simplicial
tree. Then I is sequentially Cohen-Macaulay.
Proof. This follows from Proposition 4.11 and by [7, Corollary 5.5], which states that P(I)∨ is a
componentwise linear ideal.
5 Further examples and remarks
To use the main results of this paper for computations on a given monomial ideal, there are two
steps. One is to compute the polarization of the ideal, which as can be seen from the definition, is a
quick and simple procedure. This has already been implemented in Macaulay2. The second step is
to determine whether the polarization is the facet ideal of a tree, or has a leaf. Algorithms that serve
this purpose are under construction [3].
15
Remark 5.1. Let I = (M1 , . . . , Mq ) be a monomial ideal in a polynomial ring R. If P(I) is the
facet ideal of a tree, then by Corollary 4.8, R[It] is Cohen-Macaulay. But more is true: if you drop
any generator of I, for example if you consider I ′ = (M1 , . . . , M̂i , . . . , Mq ), then R[I ′ t] is still
Cohen-Macaulay. This is because P(I ′ ) corresponds to the facet ideal of a forest, so one can apply
the same result.
A natural question is whether one can say the same with the property “Cohen-Macaulay” replaced by “normal”. If I is square-free, this is indeed the case. But in general, polarization does not
preserve normality of ideals.
Example 5.2 (normality and polarization). A valid question is whether Proposition 4.7 holds if the
word “Cohen-Macaulay” is replaced with “normal”, given that simplicial trees have normal facet
ideals ([5])?
The answer is negative. Here is an example.
Let I = (x1 3 , x1 2 x2 , x2 3 ) be an ideal of k[x1 , x2 ]. Then I is not normal; this is because I is not
even integrally closed: x1 x2 2 ∈ I as (x1 x2 2 )3 − x1 3 x2 6 = 0, but x1 x2 2 6∈ I. Now
P(I) = (x1,1 x1,2 x1,3 , x1,1 x1,2 x2,1 , x2,1 x2,2 x2,3 )
is the facet ideal of the tree
x 1,3
x 1,2
x 1,1
x 2,2
x 2,1
x 2,3
which is normal by [5].
The reason that normality (or integral closure in general) does not pass through polarization is
much more basic: polarization does not respect multiplication of ideals, or monomials. Take, for
example, two monomials M and N and two monomial ideals I and J, such that M N ∈ IJ. It is
not necessarily true that P(M )P(N ) ∈ P(I)P(J).
Indeed, let I = J = (x1 x2 ) and M = x1 2 and N = x2 2 . Then M N = x1 2 x2 2 ∈ IJ. But
P(M ) = x1,1 x1,2 , P(N ) = x2,1 x2,2 , and P(I) = P(J) = (x1,1 x2,1 )
and clearly P(M )P(N ) ∈
/ P(I)P(J) = (x1,1 2 x2,1 2 ).
Remark 5.3. It is useful to think of polarization as a chain of substitutions. This way, as a monomial
ideal I gets polarized, the ambient ring extends one variable at a time. All the in-between ideals
before we hit the final square-free ideal P(I) have the same polarization.
For example let J = (x2 , xy, y 3 ) ⊆ k[x, y]. We use a diagram to demonstrate the process
described in the previous paragraph. Each linear form a − b stands for “replacing the variable b with
a”, or vice versa, depending on which direction we are going.
v−y
u−x
J = (x2 , xy, y 3 ) −→ J1 = (xu, uy, y 3 ) −→ J2 = (xu, uv, y 2 v)
16
w−y
−→ J3 = (xu, uv, yvw) .
This final square-free monomial ideal J3 is the polarization of J, and is isomorphic to P(J) as
we defined it in Definition 2.1. Note that all ideals J, J1 , J2 and J3 have the same (isomorphic)
polarization. We can classify monomial ideals according to their polarizations. An interesting
question is to see what properties do ideals in the same polarization class have. A more difficult
question is how far can one “depolarize” a square-free monomial ideal I, where by depolarizing I
we mean finding monomial ideals whose polarization is equal to I, or equivalently, traveling the
opposite direction on the above diagram.
A
Appendix: Primary decomposition in a sequentially
Cohen-Macaulay module
The purpose of this appendix is to study, using basic facts about primary decomposition of modules, the structure of the submodules appearing in the (unique) filtration of a sequentially CohenMacaulay module M . The main result (Theorem A.4) states that each submodule appearing in the
filtration of M is the intersection of all primary submodules whose associated primes have a certain height and appear in an irredundant primary decomposition of the 0-submodule of M . Similar
results, stated in a different language, appear in [12]; the author thanks Jürgen Herzog for pointing
this out.
We first record two basic lemmas that we shall use later (the second one is an exercise in Bourbaki [1]). Throughout the discussions below, we assume that R is a finitely generated algebra over
a field, and M is a finite module over R.
Lemma A.1. Let Q 1 , . . . , Q t , P all be primary submodules of an R-module M , such that Ass(M/Q i ) =
{qi } and Ass(M/P) = {p}. If Q 1 ∩ . . . ∩ Q t ⊆ P and Q i 6⊆ P for some i, then there is a j 6= i
such that qj ⊆ p.
m
Proof. Let x ∈ Q i \P. For each j not equal to i, pick the positive integer mj such that qj j x ⊆ Q j .
So we have that
m
m
m
m
m1
mt
mt
i+1
i−1
i+1
i−1
1
qm
1 . . . qi−1 qi+1 . . . qt x ⊆ Q 1 ∩ . . . ∩ Q t ⊆ P =⇒ q1 . . . qi−1 qi+1 . . . qt ⊆ p
where the second inclusion is because x ∈
/ P. Hence for some j 6= i, qj ⊆ p.
Lemma A.2. Let M be an R-module and N be a submodule of M . Then for every p ∈ Ass(M/N ),
if p 6⊇ Ann(N ), then p ∈ Ass(M ).
Proof. Since p ∈ Ass(M/N ), there exists x ∈ M \ N such that p = Ann(x); in other words
px ⊆ N. Suppose Ann(N ) 6⊆ p, and let y ∈ Ann(N ) \ p. Now ypx = 0, and so p ⊆ Ann(yx) in
M . On the other hand, if z ∈ Ann(yx), then zyx = 0 ⊆ N and so zy ∈ p. But y ∈
/ p, so z ∈ p.
Therefore p ∈ Ass(M ).
17
Suppose M is a sequentially Cohen-Macaulay module with filtration as in Definition 4.9. We
adopt the following notation. For a given integer j, we let
Ass(M )j = {p ∈ Ass(M ) | height p = j}.
Suppose that all the j where Ass(M )j 6= ∅ form the sequence of integers
0 ≤ h1 < . . . < hc ≤ dim R
so that Ass(M ) =
S
1≤j≤c Ass(M )hj .
Proposition A.3. For all i = 0, . . . , r − 1, we have
1. Ass(Mi+1 /Mi ) ∩ Ass(M ) 6= ∅;
2. Ass(M )hr−i ⊆ Ass(Mi+1 /Mi ) and c = r;
3. If p ∈ Ass(Mi+1 ), then height p ≥ hr−i ;
4. If p ∈ Ass(Mi+1 /Mi ), then Ann(Mi ) 6⊆ p;
5. Ass(Mi+1 /Mi ) ⊆ Ass(M );
6. Ass(Mi+1 /Mi ) = Ass(M )hr−i ;
7. Ass(M/Mi ) = Ass(M )≤hr−i ;
8. Ass(Mi+1 ) = Ass(M )≥hr−i .
Proof.
1. We use induction on the length r of the filtration of M . The case r = 1 is clear, as
we have a filtration 0 ⊂ M , and the assertion follows. Now suppose the statement holds for
sequentially Cohen-Macaulay modules with filtrations of length less than r. Notice that Mr−1
that appears in the filtration of M in Definition 4.9 is also sequentially Cohen-Macaulay, and
so by the induction hypothesis, we have
Ass(Mi+1 /Mi ) ∩ Ass(Mr−1 ) 6= ∅ for i = 0, . . . , r − 2
and since Ass(Mr−1 ) ⊆ Ass(M ) it follows that
Ass(Mi+1 /Mi ) ∩ Ass(M ) 6= ∅ for i = 0, . . . , r − 2.
It remains to show that Ass(M/Mr−1 ) ∩ Ass(M ) 6= ∅.
For each i, Mi−1 ⊂ Mi , so we have ([1] Chapter IV)
Ass(M1 ) ⊆ Ass(M2 ) ⊆ Ass(M1 ) ∪ Ass(M2 /M1 )
The inclusion M2 ⊆ M3 along with the inclusions in (4) imply that
Ass(M2 ) ⊆ Ass(M3 ) ⊆ Ass(M2 ) ∪ Ass(M3 /M2 )
⊆ Ass(M1 ) ∪ Ass(M2 /M1 ) ∪ Ass(M3 /M2 ).
18
(4)
If we continue this process inductively, at the i-th stage we have
Ass(Mi ) ⊆ Ass(Mi−1 ) ∪ Ass(Mi /Mi−1 )
⊆ Ass(M1 ) ∪ Ass(M2 /M1 ) ∪ Ass(M3 /M2 ) ∪ . . . ∪ Ass(Mi /Mi−1 )
and finally, when i = r it gives
Ass(M ) ⊆ Ass(M1 ) ∪ Ass(M2 /M1 ) ∪ . . . ∪ Ass(M/Mr−1 ).
(5)
Because of Condition (b) in Definition 4.9, and the fact that each Mi+1 /Mi is Cohen-Macaulay
(and hence all its associated primes have the same height; see [2] Chapter 2), if for every i we
pick pi ∈ Ass(Mi+1 /Mi ), then
hc ≥ height p0 > height p1 > . . . > height pr−1 .
where the left-hand-side inequality comes from the fact that Ass(M1 ) ⊆ Ass(M ). By our
induction hypothesis, Ass(M ) intersects Ass(Mi+1 /Mi ) for all i ≤ r − 2, and so because of
(5) we conclude that
height pi = hc−i , and Ass(M )hc−i ⊆ Ass(Mi+1 /Mi ) for 0 ≤ i ≤ r − 2.
And now Ass(M )h0 has no choice but to be included in Ass(M/Mr−1 ), which settles our
claim. It also follows that c = r.
2. See the proof for part 1.
3. We use induction. The case i = 0 is clear, since for every p ∈ Ass(M1 ) = Ass(M1 /M0 ) we
know from part 2 that height p = hr . Suppose the statement holds for all indices up to i − 1.
Consider the inclusion
Ass(Mi ) ⊆ Ass(Mi+1 ) ⊆ Ass(Mi ) ∪ Ass(Mi+1 /Mi ).
From part 2 and the induction hypothesis it follows that if p ∈ Ass(Mi+1 ) then height p ≥
hr−i .
p
T
T
4. Suppose Ann(Mi ) ⊆ p. Since Ann(Mi ) = p′ ∈Ass(Mi ) p′ , it follows that p′ ∈Ass(Mi ) p′ ⊆
p, so there is a p′ ∈ Ass(Mi ) such that p′ ⊆ p. But from parts 2 and 3 above it follows that
height p′ ≥ hr−i+1 and height p = hr−i , which is a contradiction.
5. From part 4 and Lemma A.2, it follows that Ass(Mi+1 /Mi ) ⊆ Ass(Mi+1 ) ⊆ Ass(M ).
6. This follows from parts 2 and 5, and the fact that Mi+1 /Mi is Cohen-Macaulay, and hence
all associated primes have the same height.
19
7. We show this by induction on e = r − i. The case e = 1 (or i = r − 1) is clear, because by
part 6 we have Ass(M/Mr−1 ) = Ass(M )h1 = Ass(M )≤h1 .
Now suppose the equation holds for all integers up to e − 1 (namely i = r − e + 1), and we
would like to prove the statement for e (or i = r − e). Since Mi+1 /Mi ⊆ M/Mi , we have
Ass(Mi+1 /Mi ) ⊆ Ass(M/Mi ) ⊆ Ass(Mi+1 /Mi ) ∪ Ass(M/Mi+1 )
(6)
By the induction hypothesis and part 6 we know that
Ass(M/Mi+1 ) = Ass(M )≤hr−i−1 and Ass(Mi+1 /Mi ) = Ass(M )hr−i ,
which put together with (6) implies that
Ass(M )hr−i ⊆ Ass(M/Mi ) ⊆ Ass(M )≤hr−i
We still have to show that Ass(M/Mi ) ⊇ Ass(M )≤hr−i−1 .
Let
p ∈ Ass(M )≤hr−i−1 = Ass(M/Mi+1 ) = Ass((M/Mi )/(Mi+1 /Mi )).
If p ⊇ Ann(Mi+1 /Mi ), then (by part 6)
\
p⊇
q =⇒ p ⊇ q for some q ∈ Ass(M )hr−i
q∈Ass(M )hr−i
which is a contradiction, as height p ≤ hr−i−1 < height q. It follows from Lemma A.2 that
p ∈ Ass(M/Mi ).
8. The argument is based on induction, and exactly the same as the one in part 4, using more
information; from the inclusions
Ass(Mi ) ⊆ Ass(Mi+1 ) ⊆ Ass(Mi ) ∪ Ass(Mi+1 /Mi ),
the induction hypothesis, and part 6 we deduce that
Ass(M )≥hr−i+1 ⊆ Ass(Mi+1 ) ⊆ Ass(M )≥hr−i+1 ∪ Ass(M )hr−i ,
which put together with part 4, along with Lemma A.2 produces the equality.
Now suppose that as a submodule of M , M0 = 0 has an irredundant primary decomposition of
the form:
\
h
h
M0 = 0 =
Q 1j ∩ . . . ∩ Q sjj
(7)
1≤j≤r
h
h
h
where for a fixed j ≤ r and e ≤ sj , Q ej is a primary submodule of M with Ass(M/Q ej ) = {pe j }
h
h
and Ass(M )hj = {p1 j , . . . , psjj }.
20
Theorem A.4. Let M be a sequentially Cohen-Macaulay module with filtration as in Definition 4.9,
and suppose that M0 = 0 has a primary decomposition as in (7). Then for each i = 0, . . . , r − 1,
Mi has the following primary decomposition
\
h
h
Mi =
Q 1j ∩ . . . ∩ Q sjj .
(8)
1≤j≤r−i
Proof. We prove this by induction on r (length of the filtration). The case r = 1 is clear, as the
filtration is of the form 0 = M0 ⊂ M . Now consider M with filtration
0 = M0 ⊂ M1 ⊂ . . . ⊂ Mr = M.
Since Mr−1 is a sequentially Cohen-Macaulay module of length r − 1, it satisfies the statement of
the theorem. We first show that Mr−1 has a primary decomposition as described in (8). From part 7
of Proposition A.3 it follows that
Ass(M/Mr−1 ) = Ass(M )h1
and so for some phe 1 -primary submodules Peh1 of M (1 ≤ e ≤ sj ), we have
Mr−1 = P1h1 ∩ . . . ∩ Psh11 .
(9)
We would like to show that Qhe1 = Peh1 for e = 1, . . . , s1 .
Fix e = 1 and assume Qh11 6⊂ P1h1 . From the inclusion M0 ⊂ P1h1 and Lemma A.1 it follows
h
that for some e and j (with e 6= 1 if j = 1), we have pe j ⊆ ph1 1 . Because of the difference in heights
h
of these ideals the only conclusion is pe j = ph1 1 , which is not possible. With a similar argument we
deduce that Qhe1 ⊂ Peh1 , for e = 1, . . . , s1 .
h
Now fix j ∈ {1, . . . , r} and e ∈ {1, . . . , sj }. If Mr−1 = Q ej we are done. Otherwise, note that
h
h
for every j and pe j -primary submodule Q ej of M ,
h
Q ej ∩ Mr−1
h
h
is a pe j -primary submodule of Mr−1 (as ∅ =
6 Ass(Mr−1 /(Q ej ∩ Mr−1 )) = Ass((Mr−1 +
h
h
h
h
Q ej )/Q ej ) ⊆ Ass(M/Q ej ) = {pe j }). So M0 = 0 as a submodule of Mr−1 has a primary
decomposition
\
h
h
M0 ∩ Mr−1 = 0 =
(Q 1j ∩ Mr−1 ) ∩ . . . ∩ (Q sjj ∩ Mr−1 ).
1≤j≤r
From Proposition A.3 part 8 it follows that
Ass(Mr−1 ) = Ass(M )≥h2
so the components Qht1 ∩ Mr−1 are redundant for t = 1, . . . , s1 , so for each such t we have
\
h
(Q 1j ∩ Mr−1 ) ⊆ Qht1 ∩ Mr−1 .
Q
hj
h1
e 6=Q t
21
h
h
If Q ej ∩ Mr−1 6⊆ Qht1 ∩ Mr−1 for some e and j (with Q ej 6= Qht1 ), then by Lemma A.1 for some
h
such e and j we have pe j ⊆ pht 1 , which is a contradiction (because of the difference of heights).
h
Therefore, for each t (1 ≤ t ≤ s1 ), there exists indices e and j (with Q ej 6= Qht1 ) such that
h
Q ej ∩ Mr−1 ⊆ Qht1 ∩ Mr−1 .
It follows now, from the primary decomposition of Mr−1 in (9) that for a fixed t
h
P1h1 ∩ . . . ∩ Psh11 ∩ Q ej ⊆ Qht1 .
Assume Pth1 6⊆ Qht1 . Applying Lemma A.1 again, we deduce that
h
pe j ⊆ pht 1 , or there is t′ 6= t such that pht′1 ⊆ pht 1 .
Neither of these is possible, so Pth1 ⊆ Qht1 for all t.
We have therefore proved that
Mr−1 = Qh11 ∩ . . . ∩ Qhs11 .
By the induction hypothesis, for each i ≤ r − 2, Mi has the following primary decomposition
\
\
h
h
h
h
Mi =
(Q 1j ∩ Mr−1 ) ∩ . . . ∩ (Q sjj ∩ Mr−1 ) =
Q 1j ∩ . . . ∩ Q sjj
2≤j≤r−i
1≤j≤r−i
which proves the theorem.
References
[1]
N. Bourbaki, Commutative algebra, Chapters 1–7, Elements of Mathematics (Berlin),
Springer-Verlag, Berlin, 1989.
[2]
W. Bruns and J. Herzog, Cohen-Macaulay rings, Cambridge Studies in Advanced Mathematics 39, revised edition, 1998.
[3]
M. Caboara, S. Faridi and P. Selinger, Simplicial tree computations, in preparation.
[4]
J.A. Eagon and V. Reiner, Resolution of Stanley-Reisner rings and Alexander duality, J.
Pure Appl. Algebra 130 (1998), 265–275.
[5]
S. Faridi, The facet ideal of a simplicial complex, manuscripta math. 109 (2002), 159–174.
[6]
S. Faridi, Cohen-Macaulay properties of square-free monomial ideals, preprint.
[7]
S. Faridi, Simplicial trees are sequentially Cohen-Macaulay, J. Pure Appl. Algebra 190
(2004), 121–136.
22
[8]
R. Fröberg, A study of graded extremal rings and of monomial rings, Math. Scand. 51
(1982), 22–34.
[9]
J. Herzog and T. Hibi, Componentwise linear ideals, Nagoya Math. J. 153 (1999), 141–153.
[10]
E. Miller, Alexander duality for monomial ideals and their resolutions, math.AG/9812095.
[11]
G. Rota and J. Stein, Standard basis in supersymplectic algebras, Proc. Nat. Acad. Sci.
U.S.A. 86 (1989), no. 8, 2521–2524.
[12]
P. Schenzel, On the dimension filtration and Cohen-Macaulay filtered modules, Commutative algebra and algebraic geometry (Ferrara), 245–264, Lecture Notes in Pure and Appl.
Math., 206, Dekker, New York, 1999.
[13]
P. Schwartau, Liaison addition and monomial ideals, Ph.D. Thesis, Brandeis University,
1982.
[14]
R.P. Stanley, Combinatorics and commutative algebra, second edition, Progress in Mathematics 41, Birkhäuser Boston, Inc., Boston, MA, 1996.
[15]
J. Stückrad and W. Vogel, Buchsbaum rings and applications, Springer-Verlag, Berlin,
1986.
[16]
R. Villarreal, Monomial algebras, Monographs and Textbooks in Pure and Applied Mathematics 238, Marcel Dekker, Inc., New York, 2001.
[17]
J. Weyman, Ideals generated by monomials, Masters Thesis, N. Copernicus University,
Torun, Poland in 1977.
23
| 0math.AC
|
arXiv:1512.01351v1 [math.RA] 4 Dec 2015
CLASSICAL INVARIANT THEORY
FOR FREE METABELIAN LIE ALGEBRAS
VESSELIN DRENSKY, ŞEHMUS FINDIK
Abstract. Let KXd be a vector space with basis Xd = {x1 , . . . , xd } over a
field K of characteristic 0. One of the main topics of classical invariant theory
is the study of the algebra of invariants K[Xd ]SL2 (K) , where KXd is a module
of the special linear group SL2 (K) isomorphic to a direct sum Vk1 ⊕ · · · ⊕ Vkr
and Vk is the SL2 (K)-module of binary forms of degree k. Noncommutative
invariant theory deals with the algebra of invariants Fd (V)G of the group G <
GLd (K) acting on the relatively free algebra Fd (V) of a variety of K-algebras
V. In this paper we consider the free metabelian Lie algebra Fd (A2 ) which
is the relatively free algebra in the variety A2 of metabelian (solvable of class
2) Lie algebras. We study the algebra Fd (A2 )SL2 (K) of SL2 (K)-invariants of
Fd (A2 ). We describe the cases when this algebra is finitely generated. This
happens if and only if KXd ∼
= V2 as an SL2 (K)= V1 ⊕ V0 ⊕ · · · ⊕ V0 or KXd ∼
module (and in the trivial case KXd ∼
= V0 ⊕· · ·⊕V0 ). For small d we give a list
of generators even when Fd (A2 )SL2 (K) is not finitely generated. The methods
for establishing that the algebra Fd (A2 )SL2 (K) is not finitely generated work
also for other relatively free algebras Fd (V) and for other groups G.
1. Introduction
One of the most intensively studied objects in classical invariant theory is the
algebra of invariants of the special linear group SL2 (C) acting on binary forms.
One of the possible noncommutative generalizations is the following. We fix a field
K of characteristic 0 and a vector space KXd with basis Xd = {x1 , . . . , xd }, d ≥ 2.
We assume that KXd is a module of the special linear group SL2 (K) isomorphic
to a direct sum Vk1 ⊕ · · · ⊕ Vkr , where Vk is the SL2 (K)-module of binary forms
of degree k (and the action of SL2 (K) on KXd is not trivial, i.e., at least one ki
is positive). Then we extend diagonally the action of SL2 (K) on the relatively
free algebra Fd (V) freely generated by Xd in a variety of (associative, Lie, Jordan,
etc.) K-algebras V. The algebra Fd (V)SL2 (K) is a noncommutative analogue of
the “classical” object K[Xd ]SL2 (K) of polynomial SL2 (K)-invariants.
In the general case when G is an arbitrary subgroup of GLd (K) and V is an arbitrary variety of algebras not too much is known about the algebra of G-invariants
2010 Mathematics Subject Classification. 17B01, 17B30, 13A50, 15A72, 17B63.
Key words and phrases. free metabelian Lie algebras, classical invariant theory, noncommutative invariant theory.
The research of the first named author was partially supported by Grant Ukraine 01/0007 of
the Bulgarian Science Fund for Bilateral Scientific Cooperation between Bulgaria and Ukraine.
The research of the second named author was partially supported by the Council of Higher
Education (YÖK) in Turkey.
The research of both authors was partially supported by Grant I02/18 of the Bulgarian Science
Fund.
1
2
VESSELIN DRENSKY, ŞEHMUS FINDIK
Fd (V)G in Fd (V). The picture is more or less clear for varieties of associative algebras. For finite groups G see the surveys [13, 9, 17] and for G reductive – the papers
[26, 7]. In particular, Fd (V)G is finitely generated for all reductive groups G acting
rationally on KXd if and only if V satisfies the Engel identity [[[x2 , x1 ], . . .], x1 ] = 0.
It is well known that for varieties of Lie algebras there is a dichotomy: The
variety V either contains the metabelian variety A2 or satisfies the Engel identity.
Recall that the variety A2 of metabelian (solvable of class 2) Lie algebras is defined
by the polynomial identity
[[x1 , x2 ], [x3 , x4 ]] = 0
and the free metabelian Lie algebra Fd (A2 ) is isomorphic to the factor algebra
Ld /L′′d of the free d-generated Lie algebra Ld modulo the second term L′′d =
[[Ld , Ld ], [Ld , Ld ]] of its derived series. If V satisfies the Engel identity, then the
celebrated result of Zel’manov [28] gives that V is nilpotent. (The fact that the Engel identity implies that Fd (V) is nilpotent with the class of nilpotency depending
on d follows from the famous paper by Kostrikin [18] on the Burnside problem.)
When V is nilpotent, the algebra Fd (V) is finite dimensional. Hence Fd (V)G is also
finite dimensional and therefore finitely generated as a Lie algebra for all groups
G. If V contains A2 , then the canonical homomorphism Fd (V) → Fd (A2 ) maps
Fd (V)G onto Fd (A2 )G . Hence, concerning the finite generation of Fd (V)G , the free
metabelian algebra is the key object. If the group G is finite and V contains A2 ,
then Fd (V)G is never finitely generated by [4, 8]. To the best of our knowledge, not
too much information is available for Fd (V)G when the group G is not finite. The
algebra of invariants Fd (V)G of the group G = U T2 (K) of unipotent 2 × 2 matrices
acting unipotently on KXd was studied in [10, 5]. In particular, when V = A2 , the
algebra Fd (A2 )UT2 (K) is finitely generated if and only if KXd = K 2 ⊕ K d−2 and
U T2 (K) acts canonically on K 2 and trivially on K d−2 .
In the present paper we study the algebra Fd (A2 )SL2 (K) of SL2 (K)-invariants
of Fd (A2 ). We show that Fd (A2 )SL2 (K) is finitely generated if and only if KXd ∼
=
V1 ⊕ V0 ⊕ · · · ⊕ V0 or KXd ∼
= V2 as an SL2 (K)-module (and in the trivial case
KXd ∼
= V0 ⊕ · · · ⊕ V0 ).
One of the main tools in our considerations is the Shmel’kin embedding theorem
[23] which allows to consider Fd (A2 ) as a subalgebra of the abelian wreath product
KAd wrKYd of two d-dimensional abelian Lie algebras KAd and KYd with bases
Ad = {a1 , . . . , ad } and Yd = {y1 , . . . , yd }, respectively.
To see whether Fd (A2 )SL2 (K) is finitely generated, the most difficult part of the
proof is to find whether the commutator ideal Fd (A2 )′ contains nonzero invariants.
For small d it can be established calculating the Hilbert (or Poincaré) series of
Fd (A2 )SL2 (K) . But it has turned out that in the general case we have to embed
KAd wrKYd into a Poisson algebra and then to use classical results for the existence
of sufficiently many invariants of SL2 (K).
The commutator ideal Fd (A2 )′ has a natural structure of a K[Xd ]-module.
Well known results give that (Fd (A2 )′ )SL2 (K) is a finitely generated K[Xd ]SL2 (K) module. For small d we give a list of generators of the K[Xd ]SL2 (K) -module
(Fd (A2 )′ )SL2 (K) . This allows to find explicit sets of generators of the Lie algebra
Fd (A2 )SL2 (K) even if is not finitely generated. In this part our paper may be considered as a continuation of [5] where similar problems were solved for Fd (A2 )UT2 (K) .
CLASSICAL INVARIANT THEORY FOR FREE METABELIAN LIE ALGEBRAS
3
Finally, we want to mention that the methods for establishing that the algebra
Fd (A2 )SL2 (K) is not finitely generated work also for other relatively free algebras
Fd (V) and for other groups G.
2. Preliminaries
In the sequel we fix the base field K of characteristic 0. All vector spaces,
algebras, tensor products, etc., will be over K. If R is any algebra generated by
the elements r1 , . . . , rk and I is an ideal of R, we shall denote the generators of the
factor algebra R/I with the same symbols r1 , . . . , rk .
2.1. Free metabelian Lie algebras and abelian wreath products. For a
background on varieties of Lie algebras we refer to [2]. We denote by Fd the free
metabelian Lie algebra Fd (A2 ) freely generated by Xd = {x1 , . . . , xd }. We assume
that all Lie commutators are left normed, e.g.,
[x1 , x2 , x3 ] = [[x1 , x2 ], x3 ] = [x1 , x2 ]adx3 .
The commutator ideal Fd′ has a basis consisting of all
[xj1 , xj2 , xj3 , . . . , xjk ],
1 ≤ ji ≤ d,
j1 > j2 ≤ j3 ≤ · · · ≤ jk .
The algebra Fd has a grading
Fd = (Fd )1 ⊕ (Fd )2 ⊕ (Fd )3 ⊕ · · ·
where (Fd )n is the vector subspace of all homogeneous elements of degree n. Similarly, Fd is Zd -graded (or multigraded), with grading which counts the degree of
each variable xi , i = 1, . . . , d. The Hilbert series of Fd is the formal power series
defined by
X
H(Fd , Zd ) = H(Fd , z1 , . . . , zd ) =
dim(Fd )n Zdn ,
ni ≥0
where (Fd )n is the homogeneous component of degree n = (n1 , . . . , nd ) and Zdn =
z1n1 · · · zdnd . It is well known, see e.g., [8], that
H(Fd , Zd ) = 1 + (z1 + · · · + zd ) + (z1 + · · · + zd − 1)
d
Y
1
.
1 − zj
j=1
The consequence of the metabelian identity
[xj1 , xj2 , xjσ(3) , . . . , xjσ(k) ] = [xj1 , xj2 , xj3 , . . . , xjk ],
where σ is an arbitrary permutation of 3, . . . , k, allows to define an action of the
polynomial algebra K[Xd ] on Fd′ by the rule
uf (x1 , . . . , xd ) = uf (adx1 , . . . , adxd ),
u ∈ Fd′ ,
f (Xd ) ∈ K[Xd ].
One of the main tools in our considerations is the Shmel’kin embedding theorem [23] which allows to consider Fd (A2 ) as a subalgebra of the abelian wreath
product KAd wrKYd of two d-dimensional abelian Lie algebras KAd and KYd .
Recall the construction of abelian wreath products due to Shmel’kin [23]. Let
KAd and KYd be the abelian Lie algebras with bases Ad = {a1 , . . . , ad } and
Yd = {y1 , . . . , yd }, respectively. Let Cd be the free right K[Yd ]-module with free
4
VESSELIN DRENSKY, ŞEHMUS FINDIK
generators a1 , . . . , ad . We give it the structure of a Lie algebra with trivial multiplication. Then KAd wrKYd is equal to the semidirect sum Cd ⋋ KYd . The elements
of KAd wrKYd are of the form
d
X
aj fj (Yd ) +
d
X
βj y j ,
fj ∈ K[Yd ],
βj ∈ K,
j = 1, . . . , d.
j=1
j=1
The multiplication in KAd wrKYd is defined by
[Cd , Cd ] = [Yd , Yd ] = 0,
[aj fj (Yd ), yi ] = aj fj (Yd )yi ,
i, j = 1, . . . , d.
The embedding theorem of Shmel’kin [23] gives that the free metabelian Lie algebra
Fd is isomorphic to the subalgebra of KAd wrKYd generated by aj +yj , j = 1, . . . , d.
In the sequel we identify Fd with its copy in KAd wrKYd and assume that
xj = aj + yj ,
j = 1, . . . , d.
If u belongs to Fd′ ,
X
X
[xi , xj ]fij (adx1 , . . . , adxd ),
[xi , xj ]fij (Xd ) =
u=
fij (Xd ) ∈ K[Xd ],
i>j
i>j
then in KAd wrKYd
u=
The element
d
X
X
(ai yj − aj yi )fij (Yd ).
ai fi (Xd ) ∈ KAd wrKYd
i=1
belongs to Fd′ if and only if
d
X
yi fi (Yd ) = 0.
i=1
2.2. Poisson algebras. Given two algebraically independent SL2 (K)-invariants
f1 (Xd ), f2 (Xd ) ∈ K[Xd ] we shall construct a nonzero SL2 (K)-invariant π(f1 , f2 )
in Fd′ . For this purpose we need to embed KAd wrKYd into a Poisson algebra. Recall
that the commutative-associative algebra P with multiplication · is a (commutative)
Poisson algebra if it has a Poisson bracket [·, ·] such that (P, [·, ·]) is a Lie algebra
and [·, ·] satisfies the Leibniz rule
[f1 · f2 , f3 ] = [f1 , f3 ] · f2 + f1 · [f2 , f3 ],
f1 , f2 , f3 ∈ P.
The free Poisson algebra P (Xd ) freely generated by Xd can be obtained in the
following way. Fix a basis Bd = {b1 , b2 , . . .} of the free Lie algebra Ld . Then, as a
commutative algebra, P (Xd ) is isomorphic to the symmetric algebra S(Ld ), i.e., to
the polynomial algebra K[Bd ]. The Poisson bracket in P (Xd ) is defined extending
the Lie bracket in Ld using the Leibniz rule. In the sequel we shall omit the · in
the multiplication in Poisson algebras.
Lemma 2.1. Let P (Ad , Yd ) be the free Poisson algebra freely generated by Ad ∪ Yd
and let
Rd = P (Ad , Yd )/(A2d , [Ad , Ad ], [Yd , Yd ], [ai , yj ] − ai yj | i, j = 1, . . . , d)
CLASSICAL INVARIANT THEORY FOR FREE METABELIAN LIE ALGEBRAS
5
be the factor algebra of P (Ad , Yd ) modulo the ideal generated by all elements
ai aj , [ai , aj ], [yi , yj ], [ai , yj ] − ai yj ,
i, j = 1, . . . , d.
(i) As a vector space Rd has a basis
{Ydq = y1q1 · · · ydqd ,
ai Ydq = ai y1q1 · · · ydqd | qj ≥ 0,
i = 1, . . . , d}.
(ii) As a commutative algebra Rd is isomorphic to the factor algebra K[Ad , Yd ]/(A2d )
modulo the ideal generated by all products ai aj , i, j = 1, . . . , d.
(iii) The Poisson bracket in Rd satisfies
[ai y1p1 · · · ydpd , y1q1 · · · ydqd ] = (q1 + · · · + qd )ai y1p1 +q1 · · · ydpd +qd ,
[y1p1 · · · ydpd , y1q1 · · · ydqd ] = [ai y1p1 · · · ydpd , aj y1q1 · · · ydqd ] = 0.
(iv) The Lie algebra KAd wrKYd is isomorphic to the Lie subalgebra of Rd generated by Ad and Yd .
Proof. The defining relations
[ai , aj ] = [yi , yj ] = 0,
[ai , yj ] = ai yj
and the Leibniz rule imply that the commutators of the generators Ad and Yd
can be expressed as monomials of the polynomial algebra K[Ad , Yd ]. The defining
relations ai aj = 0 give that as a vector space Rd is spanned by the elements from
the statement (i) of the lemma. Hence we shall prove (i), (ii), and (iii), if we
show that the commutative algebra K[Ad , Yd ]/(A2d ) which has the desired basis, is
a Poisson algebra with respect to the anticommutative bracket defined in (iii). In
order to verify the Jacobi identity it is sufficient to check the case
u1 = ai Ydp ,
p = (p1 , . . . , pd ),
u2 = Ydq ,
q = (q1 , . . . , qd ),
u3 = Ydr ,
r = (r1 , . . . , rd ).
Direct verification gives that
[u1 , u2 , u3 ] + [u2 , u3 , u1 ] + [u3 , u1 , u2 ]
= (q1 + · · · + qd )[ai Ydp+q , Ydr ] − (r1 + · · · + rd )[ai Ydp+r , Ydq ]
= ((q1 +· · ·+qd )+(r1 +· · ·+rd ))ai Ydp+q+r −((r1 +· · ·+rd )+(q1 +· · ·+qd ))ai Ydp+q+r = 0.
Similarly, for the Leibniz rule it is sufficient to check the case [u1 , u2 u3 ] for the same
elements u1 , u2 , u3 :
[u1 , u2 u3 ] = ((q1 + r1 ) + · · · + (qd + rd ))ai Ydp+q+r
= (q1 + · · · + qd )ai Ydp+q Ydr + (r1 + · · · + rd )Ydq ai Ydp+r = [u1 , u2 ]u3 + u2 [u1 , u3 ].
This proves (i), (ii), and (iii). Finally, for the proof of (iv) we see that the commutator ideal of the Lie subalgebra of Rd generated by Ad and Yd has a basis consisting
of all
ai Ydp = ai (adp1 y1 ) · · · (adpd yq ), p = (p1 , . . . , pd ) 6= (0, . . . , 0),
the same as the commutator ideal of KAd wrKYd . The multiplication table is also
the same. Hence both Lie algebras are isomorphic.
6
VESSELIN DRENSKY, ŞEHMUS FINDIK
2.3. Algebraically independent elements in Poisson algebras. The well known
Jacobian criterion [16] from 1841 gives that the polynomials f1 , . . . , fk ∈ K[Xd ],
k ≤ d, are algebraically dependent if and only if
∂f1 /∂xj1
∂f1 /∂xj2
..
.
∂f2 /∂xj1
∂f2 /∂xj2
..
.
···
···
..
.
∂fk /∂xj1
∂fk /∂xj2
=0
..
.
∂f1 /∂xjk
∂f2 /∂xjk
···
∂fk /∂xjk
for all determinants with 1 ≤ j1 < j2 < · · · < jk ≤ d. A simple proof for k = d
can be found e.g., in [11] or [22]. We shall use the Jacobian criterion for k = 2 only
(and for any d ≥ 2). For self-completeness we present an elementary proof of the
part we need.
Lemma 2.2. Let f1 (Xd ), f2 (Xd ) ∈ K[Xd ] and let
Jij (f1 , f2 ) =
∂f1 /∂xi
∂f1 /∂xj
∂f2 /∂xi
=0
∂f2 /∂xj
for all 1 ≤ i < j ≤ d. Then f1 (Xd ) and f2 (Xd ) are algebraically dependent.
Proof. Let f1 (Xd ) and f2 (Xd ) be algebraically independent. We may assume that
they depend essentially on all variables xi , i.e., (∂f1 /∂xi , ∂f2 /∂xi ) 6= (0, 0) for all
i = 1, . . . , d. Since all determinants Jij (f1 , f2 ) are equal to 0, this means that there
exist nonzero polynomials aij , bij ∈ K[Xd ] such that
aij ∂f1 ∂f2
∂f1 ∂f2
=
, 1 ≤ i < j ≤ d.
,
,
∂xi ∂xi
bij ∂xj ∂xj
Since every d + 1 polynomials in K[Xd ] are algebraically dependent, there exists a
nonzero polynomial p(t1 , t2 , u1 , . . . , uk ) such that k ≤ d − 1 and
p(f1 , f2 , x1 , . . . , xk ) = p(f1 , f2 , Xk ) = 0.
We choose k minimal with this property, and among all such p we choose this of
minimal degree with respect to uk . Since f1 , f2 are algebraically independent, we
have that k > 0. We take the partial derivatives of p(f1 , f2 , x1 , . . . , xk ) with respect
to xk and xk+1 :
0=
∂p(f1 , f2 , Xk )
∂p
∂f1
∂p
∂f2
∂p
=
(f1 , f2 , Xk )
+
(f1 , f2 , Xk )
+
(f1 , f2 , Xk ),
∂xk
∂t1
∂xk
∂t2
∂xk
∂uk
∂p(f1 , f2 , Xk )
∂p
∂f1
∂p
∂f2
=
(f1 , f2 , Xk )
+
(f1 , f2 , Xk )
.
∂xk+1
∂t1
∂xk+1
∂t2
∂xk+1
Since ∂f1 /∂xk and ∂f2 /∂xk are equal, respectively, to (ak,k+1 /bk,k+1 )∂f1 /∂xk+1
and (ak,k+1 /bk,k+1 )∂f2 /∂xk+1 , the above equations give that (∂p/∂uk )(f1 , f2 , Xk ) =
0. Hence f1 , f2 , Xk annulate the polynomial (∂p/∂uk )(t1 , t2 , u1 , . . . , uk ) which is of
lower degree in xk than p(t1 , t2 , u1 , . . . , uk ) and this is a contradiction. Hence k = 0,
i.e., f1 and f2 are algebraically dependent.
0=
If f1 (Xd ), f2 (Xd ) are polynomials in K[Xd ] considered as elements of the free
Poisson algebra P (Xd ), then, see e.g., [22],
X
∂f1 ∂f2
∂f1 ∂f2
.
−
[xi , xj ]
[f1 , f2 ] =
∂xi ∂xj
∂xj ∂xi
i<j
CLASSICAL INVARIANT THEORY FOR FREE METABELIAN LIE ALGEBRAS
7
Hence f1 (Xd ), f2 (Xd ) ∈ K[Xd ] ⊂ P (Xd ) are algebraically independent in K[Xd ] if
and only if [f1 , f2 ] 6= 0 in P (Xd ). An analogue for any f1 , f2 ∈ P (Xd ) was established in [21]. We cannot apply directly this result for the algebra Rd defined in the
previous subsection because it is a homomorphic image of the free Poisson algebra.
We believe that the following proposition is of independent interest because it allows, starting from two polynomials in commuting variables, to construct elements
with prescribed properties in the free metabelian Lie algebra.
Proposition 2.3. Let the homogeneous polynomials f1 (Xd ) and f2 (Xd ) be algebraically independent in K[Xd ]. Then in the Poisson algebra Rd the commutator
π(f1 , f2 ) = [f1 (Ad + Yd ), f2 (Ad + Yd )]
= [f1 (a1 + y1 , . . . , ad + yd ), f2 (a1 + y1 , . . . , ad + yd )]
X
∂f1 (Yd ) ∂f2 (Yd ) ∂f1 (Yd ) ∂f2 (Yd )
=
(ai yj − aj yi )
−
∂yi
∂yj
∂yj
∂yi
i<j
is different from zero and belongs to the commutator ideal Fd′ of the free metabelian
Lie algebra Fd embedded in Rd .
Proof. If P is any Poisson algebra, Wd = {w1 , . . . , wd } ⊂ P , and
Wdq = w1q1 · · · wdqd ,
Wdr = w1r1 · · · wdrd ,
then the Leibniz rule implies that
[Wdq , Wdr ] =
d
X
q +rj −1
[wi , wj ]qi rj w1q1 +r1 · · · wiqi +ri −1 · · · wj j
· · · wdqd +rd
i,j=1
=
X
[wi , wj ]
i<j
∂Wdq ∂Wdr
∂Wdq ∂Wdr
−
∂wi ∂wj
∂wj ∂wi
.
Hence, as in the case of the free Poisson algebra, for any polynomials f1 (Xd ), f2 (Xd ) ∈
K[Xd ],
X
∂f1 (Wd ) ∂f2 (Wd ) ∂f1 (Wd ) ∂f2 (Wd )
−
[wi , wj ]
[f1 (Wd ), f2 (Wd )] =
∂wi
∂wj
∂wj
∂wi
i<j
=
X
[wi , wj ]
i<j
∂f1 (Wd )/∂wi
∂f1 (Wd )/∂wj
X
∂f2 (Wd )/∂wi
[wi , wj ]Jij (f1 (Wd ), f2 (Wd )).
=
∂f2 (Wd )/∂wj
i<j
We apply this equality for
Wd = Ad + Yd = {a1 + y1 , . . . , ad + yd } ⊂ Rd .
Since [ai + yi , aj + yj ] = ai yj − aj yi and ak al = 0, we obtain
X
∂f1 (Yd )/∂yi
(ai yj − aj yi )
[f1 (Ad + Yd ), f2 (Ad + Yd )] =
∂f1 (Yd )/∂yj
i<j
=
X
(ai yj − aj yi )
i<j
∂f2 (Yd )/∂yi
∂f2 (Yd )/∂yj
∂f1 (Yd ) ∂f2 (Yd ) ∂f1 (Yd ) ∂f2 (Yd )
−
∂yi
∂yj
∂yj
∂yi
d
d
X
X
∂f
(Y
)
∂f
(Y
)
∂f
(Y
)
∂f
(Y
)
2 d
1 d
2 d
1 d
yj
−
yj
.
ai
=
∂y
∂y
∂y
∂y
i
j
j
i
j=1
j=1
i=1
d
X
8
VESSELIN DRENSKY, ŞEHMUS FINDIK
Since f2 (Yd ) is homogeneous, it is easy to see that
d
X
j=1
yj
∂f2 (Yd )
= deg(f2 )f2 (Yd ).
∂yj
If [f1 (Ad + Yd ), f2 (Ad + Yd )] = 0, then the coordinates of ai are equal to 0 and we
obtain a homogeneous linear system
d
X
∂f1 (Yd ) ∂f2 (Yd )
∂f1 (Yd )
deg(f2 )f2 (Yd ) −
yj
=0
∂yi
∂yj
∂yi
j=1
of d equations, for i = 1, . . . , d, with unknowns ∂f1 (Yd )/∂yj , j = 1, . . . , d. The
matrix of this system is
deg(f2 )f2 − y1 ∂f2 /∂y1
−y2 ∂f2 /∂y1
···
−yd ∂f2 /∂y1
−y1 ∂f2 /∂y2
deg(f2 )f2 − y2 ∂f2 /∂y2 · · ·
−yd ∂f2 /∂y2
.
..
..
.
..
..
.
.
.
−y1 ∂f2 /∂yd
−y2 ∂f2 /∂yd
· · · deg(f2 )f2 − yd ∂f2 /∂yd
Since f2 (Yd ) 6= 0, at least one partial derivative if nonzero. Let, for example,
∂f2 /∂yd 6= 0. Subtracting the last row of the system multiplied by (∂f2 /∂yi )/(∂f2 /∂yd )
from the i-th row, we obtain an equivalent system with a matrix
deg(f2 )f2
−0
···
0
∗
0
deg(f2 )f2
···
0
∗
..
..
..
..
..
.
.
.
.
.
.
0
0
···
deg(f2 )f2
∗
−y1 ∂f2 /∂yd −y2 ∂f2 /∂yd · · · −yd−1 ∂f2 /∂yd deg(f2 )f2 − yd ∂f2 /∂yd
Since f2 (Yd ) 6= 0, the rank of the matrix is ≥ d − 1. Since the system has a nonzero
solution ∂f2 /∂yj = ∂f1 /∂yj , j = 1, . . . , d, the rank is d − 1. Hence all solutions of
the system are
∂f2
p(Yd ) ∂f1
=
, j = 1, . . . , d,
∂yj
q(Yd ) ∂yj
for some polynomials p(Yd ), q(Yd ) ∈ K[Yd ]. Therefore,
Jij (f1 (Yd ), f2 (Yd )) = 0,
i, j = 1, . . . , d,
which contradicts with the algebraic independence of f1 (Y ) and f2 (Y ). Clearly,
the form of the element
X
(ai yj − aj yi )fij (Yd ), fij (Yd ) ∈ K[Yd ],
[f1 (Ad + Yd ), f2 (Ad + Yd )] =
i<j
shows that [f1 (Ad + Yd ), f2 (Ad + Yd )] is a Lie element and belongs to Fd′ .
2.4. Invariant theory of the special linear group. All necessary information
on representation theory of the special linear group SL2 (K) and the general linear
group GL2 (K) and on invariant theory of SL2 (K) can be found in many books,
see e.g., [24, 27]. For a background on symmetric functions see, e.g., [19]. In our
considerations we combine ideas of De Concini, Eisenbud, Procesi [6] and for the
noncommutative generalizations – of Almkvist, Dicks, Formanek [1], see [3] for
details. Every rational representation ρ : SL2 (K) → GLd (K) of SL2 (K) is a direct
CLASSICAL INVARIANT THEORY FOR FREE METABELIAN LIE ALGEBRAS
9
sum of irreducible representations. For each nonnegative integer k there exists a
unique irreducible rational (k + 1)-dimensional SL2 (K)-representation
ρk : SL2 (K) → GLk+1 (K) = GL(Vk ),
where Vk is the vector space of the
Ξk = {ξ0 , ξ1 , . . . , ξk } of Vk and define
g
g = 11
g21
bilinear forms of degree k. We fix a basis
the action of GL2 (K) as follows. Let
g12
∈ GL2 (K).
g22
If K 2 is the vector space with basis {e1 , e2 } and the canonical action of GL2 (K)
g(e1 ) = g11 e1 + g21 e2 ,
g(e2 ) = g12 e1 + g22 e2 ,
then the action of g on the polynomial functions f : K 2 → K is
g(f )(v) = f (g −1 (v)),
v = u 1 e1 + u2 e2 ∈ K 2 .
In particular, the action on the coordinate functions is
1
1
u′1 = g(u1 ) =
(g22 u1 − g12 u2 ), u′2 = g(u2 ) =
(−g21 u1 + g11 u2 ).
det(g)
det(g)
Hence
u1 = g11 u′1 + g12 u′2 ,
If the binary form of degree k is
u2 = g21 u′1 + g22 u′2 .
k
X
k
ξj u1k−j uj2
j
j=0
k
k
k
= ξ0 uk1 +
ξ1 u1k−1 u2 +
ξ2 u1k−2 u22 + · · · +
ξk−1 u1 u2k−1 + ξk uk2 ,
1
2
k−1
then g acts on f (u1 , u2 ) by the rule
f (u1 , u2 ) =
g : f (u1 , u2 ) → f (g(u1 ), g(u2 )) = f (g11 u′1 + g12 u′2 , g21 u′1 + g22 u′2 )
k
X
k ′ ′ k−j ′ j
ξ (u ) (u2 )
= f ′ (u′1 , u′2 ) =
j j 1
j=0
and we define
g(ξj ) = ξj′ ,
j = 0, 1, . . . , k.
Example 2.4. We shall describe the action of the matrix
1 1
g1 =
∈ SL2 (K)
0 1
on Vk . Since
g1 (u1 ) = u′1 + u′2 ,
g1 (u2 ) = u′2 ,
we have
k
k
X
X
k
k
k−j j
ξj (u′1 + u′2 )k−j (u′2 )j
ξj u1 u2 =
f (u1 , u2 ) =
j
j
j=0
j=0
=
k−j
k X
X
k
k−j
ξj (u′1 )k−j−i (u′2 )i+j
j
i
j=0 i=0
10
VESSELIN DRENSKY, ŞEHMUS FINDIK
=
k−j
k X
X
k
i+j
ξj (u′1 )k−(j+i) (u′2 )i+j .
i
+
j
j
j=0 i=0
In the above expression the summation is the same as for l = i + j = 0, 1, . . . , k,
j = 0, 1, . . . , i + j = l, and we obtain
k X
l
k
X
X
k
l
k ′ ′ k−l ′ l
f (u1 , u2 ) =
ξj (u′1 )k−l (u′2 )l =
ξ (u ) (u2 ) .
l
j
l l 1
j=0
l=0
l=0
Hence the action of g1 on ξl is defined by
l
X
l
ξj ,
g1 (ξl ) = ξl′ =
j
j=0
l = 0, 1, . . . , k.
Similarly, the action of
g2 =
1 0
1 1
on Vk is
g2 (ξk−l ) =
l
X
l
j=0
j
ξk−j ,
l = 0, 1, . . . , k.
If KXd is isomorphic to a direct sum Vk1 ⊕ · · · ⊕ Vkr , we may change linearly
the variables and assume that the bases of the SL2 (K)-submodules Vki of KXd are
subsets of Xd . If KXk+1 = Vk we shall identify x1 , x2 , . . . , xk+1 with ξ0 , ξ1 , . . . , ξk ,
respectively.
The following statement is a part of classical invariant theory, see [15, p. 318 in
the 1893 original]. A modern proof can be found, e.g., in [24, p. 65].
Proposition 2.5. Let KXk+1 = Vk be the SL2 (K)-module of the binary forms of
degree k. Then the transcendence degree on the algebra of invariants K[Xk+1 ]SL2 (K)
is
0, if k = 1;
SL2 (K)
transc.deg(K[Xk+1 ]
) = 1, if k = 2;
k − 2, if k > 2.
Example 2.6. The following cases are well known for more than 100 years, see [14].
They can be seen directly computing the Hilbert series of the algebra of invariants
and checking that the given elements are SL2 (K)-invariants (see the comments
below).
(i) If KX2 = V1 , then K[X2 ]SL2 (K) = K.
(ii) Let KX3 = V2 , then K[X3 ]SL2 (K) = K[f ], where f = x22 − x1 x3 is the
discriminant.
(iii) Let KX4 = V1 ⊕ V1 , where the two copies of V1 are generated by x1 , x2 and
x3 , x4 , respectively. Then K[X4 ]SL2 (K) = K[f ], where f = x1 x4 − x2 x3 .
We use the same notation ρk and Vk , respectively, for the polynomial representation of GL2 (K) and the corresponding GL2 (K)-module of the binary forms of
degree k. As in the case of rational representations of SL2 (K), every polynomial
representation of GL2 (K) is a direct sum of irreducible representations. Every irreducible representation is of the form detl ⊗ ρk , where det : GL2 (K) → K is defined
by det : g → det(g), g ∈ GL2 (K). Clearly, the representation detl ⊗ ρk is the
CLASSICAL INVARIANT THEORY FOR FREE METABELIAN LIE ALGEBRAS
11
same as the representation indexed by the partition (k + l, l). If the representation
ρ : SL2 (K) → GLd (K) = GL(KXd ) is of the form
ρk1 ⊕ · · · ⊕ ρkr ,
i.e., KXd = Vk1 ⊕ · · · ⊕ Vkr , we extend the action of SL2 (K) to the action of
GL2 (K). Then we extend this action diagonally on the polynomial algebra K[Xd ]:
g(f (Xd )) = g(f (x1 , . . . , xd )) = f (g(x1 ), . . . , g(xd )),
g ∈ SL2 (K),
f (Xd ) ∈ K[Xd ].
The homogeneous component K[Xd ]n of degree n is a direct sum of GL2 (K)modules
X
K[Xd ]n =
mn (k, l)(detl ⊗ Vk ),
l,k≥0
where the nonnegative integer mn (k, l) is the multiplicity of detl ⊗Vk in K[Xd ]n .
The following well known lemma, see e.g., [6] describes the SL2 (K)- and U T2 (K)invariants of polynomial GL2 (K)-modules.
Lemma 2.7. Let the (finite dimensional) polynomial GL2 (K)-module V decompose
as
X
V =
m(k, l)(detl ⊗ Vk ).
l,k≥0
SL2 (K)
(i) The subspace V
of the SL2 (K)-invariants in V is spanned by the onedimensional GL2 (K)-submodules detl ⊗ V0 .
(ii) Each summand m(k, l)(detl ⊗ Vk ) of V has m(k, l) linearly independent elei
ments wkl
, i = 1, 2, . . . , m(k, l), (one in each irreducible component detl ⊗ Vk ) such
that the vector space V UT2 (K) has a basis
i
{wkl
| i = 1, 2, . . . , m(k, l), k, l ≥ 0}.
SL2 (K)
The above lemma implies that the homogeneous component K[Xd ]n
degree n of the algebra of SL2 (K)-invariants K[Xd ]SL2 (K) is
X
K[Xd ]nSL2 (K) =
mn (0, l)(detl ⊗ V0 )
of
l≥0
and the Hilbert series of K[Xd ]SL2 (K) is
H(K[Xd ]SL2 (K) , z) =
XX
mn (0, l)z n .
n≥0 l≥0
Let D2 (K) be the diagonal subgroup of GL2 (K). Then detl ⊗ Vk has a basis
{v0 , v1 , . . . , vk } of eigenvectors of D2 (K) and
α1 0
: vj → αk+l−j
αl+j
0 6= α1 , α2 ∈ K.
1
2 vj ,
0 α2
The element v0 is equal up to a multiplicative constant to the element wkl ∈ detl ⊗Vk
from Lemma 2.7. The corresponding character of D2 (K) is
α1 0
→ S(k+l,l) (α1 , α2 ),
χ(detl ⊗ Vk ) :
0 α2
where
S(k+l,l) (t1 , t2 ) = (t1 t2 )l (tk1 + t1k−1 t2 + · · · + t1 t2k−1 + tk2 )
12
VESSELIN DRENSKY, ŞEHMUS FINDIK
is the Schur function. This defines a Z2 -grading (or bigrading) on detl ⊗ Vk and
hence on K[Xd ]. The Hilbert series of K[Xd ] which counts both this bigrading and
the usual grading is
X
HGL2 (K) (K[Xd ], t1 , t2 , z) =
mn (k, l)S(k+l,l) (t1 , t2 )z n .
l,k,n≥0
The Hilbert series
H(K[Xd ], Zd ) = H(K[Xd ], z1 , . . . , zd ) =
d
Y
i=1
1
1 − zi
of the Zd -graded algebra K[Xd ] and HGL2 (K) (K[Xd ], t1 , t2 , z) are related in the
following way. If {xj , xj+1 , . . . , xj+ki } is the basis of Vki , we replace the variables
zj , zj+1 , . . . , zj+ki , respectively, with tk1i z, t1ki −1 t2 z, . . . , tk2i z. Then
HGL2 (K) (K[Xd ], t1 , t2 , z) = H(K[Xd ], tk11 z, t1k1 −1 t2 z, . . . , tk2r z).
There are many methods to compute the Hilbert series of K[Xd ]SL2 (K) . We have
chosen the following. We define the multiplicity series of HGL2 (K) (K[Xd ], t1 , t2 , z)
X
MGL2 (K) (K[Xd ], t1 , t2 , z) =
mn (k, l)tk1 (t1 t2 )l z n ,
l,k,n≥0
which is the generating function of the multiplicities mn (k, l) of the irreducible
GL2 (K)-components in K[Xd ]n . We introduce new variables t = t1 , u = t1 t2 and
rewrite MGL2 (K) (K[Xd ], t1 , t2 , z) in the form
X
′
(K[X
],
t,
u,
z)
=
MGL
mn (k, l)tk ul z n .
d
(K)
2
l,k,n≥0
′
MGL
(K[Xd ], 0, u, z)
2 (K)
Then
does not depend on t and
′
is the component of MGL
(K[Xd ], t, u, z) which
2 (K)
′
H(K[Xd ]SL2 (K) , z) = MGL
(K[Xd ], 0, 1, z).
2 (K)
′
For the computing of MGL2 (K) (K[Xd ], t1 , t2 , z) and MGL
(K[Xd ], t, u, z) we have
2 (K)
used the methods in [3] which origin from classical work of Elliott [12] and MacMahon [20]. But it is much easier to verify that the formal power series f (t1 , t2 , z) is
equal to MGL2 (K) (K[Xd ], t1 , t2 , z). It is sufficient to see whether
1
(t1 f (t1 , t2 , z) − t2 f (t2 , t1 , z)).
t1 − t2
In the sequel we shall also need partial cases of the Young rule and rules for symmetric and skew-symmetric tensor squares. The proofs can be found, respectively,
in [19] and [25].
HGL2 (K) (K[Xd ], t1 , t2 , z) =
Lemma 2.8. (i) The tensor product of the GL2 (K)-modules Vk and Vm , k ≥ m,
decomposes as
m
M
Vk ⊗ Vm =
detn ⊗ Vk+m−2n .
n=0
(ii) The symmetric tensor square of the GL2 (K)-module Vk which identifies v ′ ⊗
′′
v and v ′′ ⊗ v ′ decomposes as
m
M
V2m ⊗sym V2m =
det2n ⊗ V4(m−n) ,
n=0
CLASSICAL INVARIANT THEORY FOR FREE METABELIAN LIE ALGEBRAS
13
if k = 2m is even and
V2m+1 ⊗sym V2m+1 =
m
M
det2n ⊗ V4(m−n)+2 ,
n=0
if k = 2m + 1 is odd.
(iii) The skew-symmetric tensor square of the GL2 (K)-module Vk which identifies
v ′ ⊗ v ′′ and −v ′′ ⊗ v ′ decomposes as
m
M
V2m ⊗skew V2m =
det2n−1 ⊗ V4(m−n)+2 ,
n=1
if k = 2m is even and
V2m+1 ⊗skew V2m+1 =
m
M
det2n+1 ⊗ V4(m−n) ,
n=0
if k = 2m + 1 is odd.
Remark 2.9. Let v be an element of the polynomial GL2 (K)-module V . In order
to check that v is SL2 (K)-invariant it is not necessary to check that g(v) = v for
all g ∈ SL2 (K). It is well known that it is enough to see this only for the matrices
g1 and g2 from Example 2.4. The matrices g1 and g2 are unitriangular. Hence the
matrices gi − 1, i = 1, 2, act as nilpotent operators on V and the logarithm
X
(gi − 1)n
, i = 1, 2,
δi = log(gi ) =
(−1)n−1
n
n≥1
is well defined. In the polynomial case, one may extend the action of δi on V to
a derivation of K[V ] by the rule δ(uv) = δ(u)v + uδ(v), u, v ∈ K[V ], and then by
linearity on the whole K[V ]. Then f ∈ K[V ] is an SL2 (K)-invariant if and only
if δi (f ) = 0, i = 1, 2. This verification is much simpler than the checking that
gi (f ) = f , i = 1, 2. Similar arguments work for the tensor algebra on V (which is
isomorphic to the free associative algebra KhV i), for the free Lie algebra on V , etc.
Example 2.10. Consider the element
k
X
k
(−1)i ξi ⊗ ξk−i ∈ Vk ⊗ Vk .
w=
i
i=0
This element is bihomogeneous of degree (k, k). Direct verification shows that
g1 (w) = g2 (w) = w (or, equivalently, w belongs to the kernel of the derivations
log(g1 ) and log(g2 ) of the tensor algebra on Vk ). Hence w belongs to (Vk ⊗Vk )SL2 (K)
and, in the notation of Lemma 2.8 (i), generates the irreducible GL2 (K)-component
detk ⊗V0 . By Lemma 2.8 (ii) and (iii), when k = 2m is even, w 6= 0 in the symmetric
square V2m ⊗sym V2m . When k = 2m + 1 is odd, w 6= 0 in V2m+1 ⊗skew V2m+1 .
We identify Vk and KXk+1 . Then the symmetric tensor square V2m ⊗sym V2m
may be identifies with the GL2 (K)-module of the homogeneous quadratic polynomials in K[X2m+1 ]. Similarly, we may consider the elements of the skew-symmetric
tensor square V2m+1 ⊗skew V2m+1 as homogeneous elements of second degree in the
metabelian Lie algebra F2m+2 . In this way we obtain the SL2 (K)-invariants
2m+1
X 2m
w2m+1 =
(−1)i−1 xi x2m+2−i
i
−
1
i=1
14
VESSELIN DRENSKY, ŞEHMUS FINDIK
m
X
2m
2m
i−1
(−1) xi x2m+2−i +
(−1)m x2m+1 ∈ K[X2m+1 ]SL2 (K) ,
=2
i
−
1
m
i=1
2m+2
X 2m + 1
(−1)i−1 [xi , x2m+3−i ]
w2m+2 =
i
−
1
i=1
m+1
X 2m + 1
′
(−1)i−1 [xi , x2m+3−i ] ∈ (F2m+2
)SL2 (K) .
=2
i
−
1
i=1
2.5. Noncommutative invariant theory for arbitrary groups. Let V be a
variety of K-algebras. We shall assume that V is a variety of Lie algebras. Since
the base field K is of characteristic 0, and hence infinite, the relatively free algebra
Fd (V) is Zd -graded. The action of the group GLd (K) = GL(KXd ) on KXd is
extended diagonally on Fd (V):
g(f (Xd )) = g(f (x1 , . . . , xd )) = f (g(x1 ), . . . , g(xd )),
g ∈ GLd (K), f (Xd ) ∈ Fd (V).
Let G be a subgroup of GLd (K) = GL(KXd ). Then the algebra of G-invariants is
Fd (V)G = {f (Xd ) ∈ Fd (V) | g(f (Xd )) = f (Xd ) for all g ∈ G}.
The following lemma is well known, see e.g., [10, Proposition 4.2] for the case of
varieties of associative algebras.
Lemma 2.11. Let V and W be two varieties of algebras and V ⊂ W. Let G be
a subgroup of GLd (K) = GL(KXd ). Then the natural homomorphism Fd (W) →
Fd (V) maps Fd (W)G onto Fd (V)G . Hence, if the algebra Fd (V)G is not finitely
generated, then the same is Fd (W)G .
As a result of the dichotomy of the varieties of Lie algebras (either V contains
A2 or V satisfies the Engel identity and is nilpotent), the G-invariants of the free
metabelian Lie algebra are the key object in the study of the finite generation
of Fd (V)G . We give a couple of cases when the algebra Fd (A2 )G is not finitely
generated. We shall assume that GLd (K) acts in the same way on the vector
spaces KXd , KAd , and KYd . If g ∈ GLd (K), then
g(xj ) =
d
X
i=1
αij xi ,
g(aj ) =
d
X
αij ai ,
i=1
g(yj ) =
d
X
αij yi ,
j = 1, . . . , d,
i=1
with the same αij for Xd , Ad , and Yd .
Lemma 2.12. Let the subgroup G of GLd (K) = GL(KXd ) be such that the vector
space KXd has no G-fixed elements (i.e., (KXd )G = 0) and the algebras K[Xd ]G
and Fd (A2 )G are nontrivial (i.e., K[Xd ]G 6= K and Fd (A2 )G 6= 0). Then the Lie
algebra Fd (A2 )G is not finitely generated.
Proof. Since the algebra FdG = Fd (A2 )G is graded, the condition (KXd )G = 0 gives
that FdG ⊂ Fd′ . Hence the algebra FdG is abelian and it is sufficient to show that
it is infinite dimensional. Working in the Poisson algebra Rd from Subsection 2.2,
if u ∈ Rd′ = [Rd , Rd ] and f ∈ K[Yd ] ⊂ Rd are homogeneous nonzero elements,
then the element uf is also nonzero. Starting with a nonzero homogeneous u ∈
(Fd′ )G and with a homogeneous polynomial of positive degree f ∈ K[Xd ]G , we
obtain an infinite sequence of elements uf n ∈ Fd′ , n = 0, 1, 2, . . . which are linearly
independent in Rd because are of different degrees. Since g(uf n ) = g(u)g(f )n =
uf n for all g ∈ G, we conclude that uf n ∈ (Fd′ )G and complete the proof.
CLASSICAL INVARIANT THEORY FOR FREE METABELIAN LIE ALGEBRAS
15
Proposition 2.13. Let (KXd )G = 0 for the subgroup G of GLd (K) = GL(KXd )
and let the transcendence degree of K[Xd ]G be greater than 1. Then the Lie algebra
Fd (A2 )G is not finitely generated.
Proof. Since transc.deg(K[Xd ]G ) > 1, the algebra contains two algebraically independent homogeneous invariants f1 and f2 . Then the element
π(f1 , f2 ) = [f1 (Ad + Yd ), f2 (Ad + Yd )] ∈ Rd
Fd′ .
belongs to
Also, if g ∈ G, then
g(f1 (Ad + Yd )) = f1 (Ad + Yd ),
g(f2 (Ad + Yd )) = f2 (Ad + Yd ).
Hence
g(π(f1 , f2 )) = [g(f1 (Ad + Yd )), g(f2 (Ad + Yd ))] = [f1 (Ad + Yd ), f2 (Ad + Yd )]
and π(f1 , f2 ) ∈ (Fd′ )G . Now the proof follows from Lemma 2.12.
Finally, computing the Hilbert series of the SL2 (K)-invariants of the free metabelian
Lie algebra Fd , we may apply exactly the same methods as for the SL2 (K)invariants of K[Xd ]. Starting with the decomposition of the SL2 (K)-module KXd
and the Hilbert series H(Fd , Zd ) of Fd we obtain the Hilbert series
X
HGL2 (K) (Fd , t1 , t2 , z) =
mn,k,l (Fd )S(k+l,l) (t1 , t2 )z n ,
k,l,n≥0
the multiplicity series
MGL2 (K) (Fd , t1 , t2 , z) =
X
mn,k,l (Fd )tk1 (t1 t2 )l z n ,
k,l,n≥0
′
MGL
(Fd , t, u, z) =
2 (K)
X
mn,k,l (Fd )tk ul z n ,
k,l,n≥0
and the Hilbert series
SL2 (K)
H(Fd
′
, z) = MGL
(Fd , 0, 1, z)
2 (K)
SL (K)
of Fd 2 . See [5] where such computations were performed for the multiplicity
series of Fd .
3. Infinite generation of the metabelian invariants
of the special linear group
In this section we assume that SL2 (K) acts rationally on the vector space KXd
and
KXd ∼
= Vk1 ⊕ · · · ⊕ Vkr ,
where Vk is the SL2 (K)-module of binary forms of degree k. When necessary we
extend the SL2 (K)-action to GL2 (K). If V ′ is an SL2 (K)-submodule of KXd we
assume that it is spanned by a subset of Xd . We denote by Fd the free metabelian
Lie algebra generated by Xd .
Following our agreement, if KXd has an SL2 (K)-invariant subspace we may assume that it has a basis {xk+1 , . . . , xd } and its complement has a basis {x1 , . . . , xk }.
Then the subalgebra of SL2 (K)-invariants in K[Xd ] has the form
K[Xd ]SL2 (K) = (K[Xk ]SL2 (K) )[xk+1 , . . . , xd ].
16
VESSELIN DRENSKY, ŞEHMUS FINDIK
SL (K)
We shall derive an analogue of this fact for the algebra Fd 2
which shows how
SL2 (K)
SL (K)
to obtain the generators of Fd
if we know the generators of Fd−12 . The
proof is based on a similar statement for the algebra of U T2 (K)-invariants from [5].
We shall work in the Poisson algebra
Rd = P (Ad , Yd )/(A2d , [Ad , Ad ], [Yd , Yd ], [ai , yj ] − ai yj | i, j = 1, . . . , d)
from Lemma 2.1. The lemma gives that the Lie algebra KAd wrKYd is isomorphic
to the Lie subalgebra of Rd generated by Ad and Yd . Hence the free metabelian Lie
algebra Fd is embedded in Rd as the Lie subalgebra generated by
Xd = Ad + Yd = {xi = ai + yi | i = 1, . . . , d}.
Clearly, the unitary associative subalgebra of Rd generated by Xd is isomorphic to
the polynomial algebra K[Xd ]. By Proposition 2.3, if f1 (Xd ) and f2 (Xd ) are two
polynomials in K[Xd ] ⊂ Rd , then the commutator
π(f1 (Xd ), f2 (Xd )) = [f1 (Ad + Yd ), f2 (Ad + Yd )]
X
∂f1 (Yd ) ∂f2 (Yd ) ∂f1 (Yd ) ∂f2 (Yd )
−
(ai yj − aj yi )
=
∂yi
∂yj
∂yj
∂yi
i<j
X
∂f1 (Xd ) ∂f2 (Xd ) ∂f1 (Xd ) ∂f2 (Xd )
[xi , xj ]
=
−
∂xi
∂xj
∂xj
∂xi
i<j
belongs to the commutator ideal Fd′ of Fd ⊂ Rd .
Proposition 3.1. Let xd ∈ KXd be SL2 (K)-invariant and let KXd = KXd−1 ⊕
Kxd as SL2 (K)-modules. Let {vi | i ∈ I} and {uj | j ∈ J} be, respectively, homo′
geneous bases of (Fd−1
)SL2 (K) and ω(K[Xd−1 ])SL2 (K) with respect to the usual Zgrading, where ω(K[Xd−1 ]) is the augmentation ideal of K[Xd−1 ]. Then (Fd′ )SL2 (K) ⊂
Rd has a basis consisting of
vi ydn = vi adn xd ,
π(xd , uj )ydn =
d−1
X
[xd , xk ]
k=1
i ∈ I,
n ≥ 0,
∂uj
(adXd−1 )adn xd ,
∂xk
j ∈ J,
n ≥ 0.
Proof. We consider the group U T2 (K) as a subgroup of SL2 (K), both groups
acting on KXd . Hence U T2 (K) acts unitriangularly on the irreducible SL2 (K)components of KXd . Then the algebras of U T2 (K)-invariants K[Xd ]UT2 (K) and
UT (K)
Fd 2
coincide, respectively, with the algebras of constants (i.e., the kernels)
δ1
K[Xd ] and Fdδ1 of the Weitzenböck (i.e., locally nilpotent linear) derivation δ1 =
log(g1 ). A basis of the vector space (Fd′ )UT2 (K) is given in [5, Theorem 4.3]. We
translate it in the language of our paper. Let {vi | i ∈ I} and {uj | j ∈ J} be, re′
spectively, homogeneous bases of (Fd−1
)UT2 (K) and ω(K[Xd−1 ])UT2 (K) with respect
2
′ UT2 (K)
to both Z- and Z -gradings. Then (Fd )
as a subspace of KAd wrKYd ⊂ Rd
has a basis
{vi ydn , π(xd , uj )ydn | i ∈ I, j ∈ J, n ≥ 0}.
UT (K)
A bihomogeneous element in K[Xd ]UT2 (K) or in Fd 2
of bidegree (p, q) is SL2 (K)invariant if and only if p = q. Since the elements xd , ad , yd are of bidegree (0, 0),
the vector space (Fd′ )SL2 (K) has a basis consisting of the basis elements vi ydn and
π(xd , uj )ydn of (Fd′ )UT2 (K) with the property that the bidegrees of vi and uj are
CLASSICAL INVARIANT THEORY FOR FREE METABELIAN LIE ALGEBRAS
17
equal to (pi , pi ) and (qj , qj ), respectively. But this condition simply means that
′
vi ∈ (Fd−1
)SL2 (K) and uj ∈ ω(K[Xd−1 ])SL2 (K) . To complete the proof it is sufficient to see that
vi ydn = vi adn xd , i ∈ I, n ≥ 0,
π(xd , uj )xnd =
d−1
X
k=1
[xd , xk ]
∂uj
(adXd−1 )adn xd
∂xk
j ∈ J,
n ≥ 0.
Lemma 3.2. Let KXd = V ′ ⊕ V ′′ , where V ′ = KXe = K{x1 , . . . , xe } and V ′′
are SL2 (K)-submodules of KXd and let Fe be the Lie subalgebra of Fd generated
by Xe . Then the natural projection ν : Fd → Fe which send xe+1 , . . . , xd to 0 maps
SL (K)
SL (K)
Fd 2
onto Fe 2 .
Proof. As a vector space Fd = Fe ⊕ ker(ν). Since both Fe and ker(ν) are SL2 (K)modules, it is clear that every SL2 (K)-invariant f ∈ Fd is a sum of an SL2 (K)invariant f ′ ∈ Fe (which depends on the variables Xe only) and an SL2 (K)invariant f ′′ ∈ ker(ν) (and f ′′ is a linear combination of commutators which depend
SL (K)
SL (K)
essentially on the other variables xe+1 , . . . , xd ). Hence ν(Fd 2 ) = Fe 2 .
SL (K)
Our strategy to determine whether the algebra Fd 2
is finitely generated
SL (K)
will be the following. We shall describe the cases when Fk+12
is not finitely
V
is
an
irreducible
SL
(K)-module.
We
shall
show that
generated and KXk+1 ∼
= k
2
this holds if and only if k > 2. In view of the above Lemma 3.2, this will imply
SL (K)
that Fd 2
is never finitely generated if KXd contains a submodule Vk , k > 2.
By similar arguments we shall eliminate the cases when KXd contains a submodule
isomorphic to V2 ⊕ Vk , k = 0, 1, 2, and V1 ⊕ V1 . In all other cases, namely when
SL (K)
d = 3 and KXd = V2 or KXd = V1 ⊕ V0 ⊕ · · · ⊕ V0 , we shall show that Fd 2
is
finitely generated.
SL (K)
is
Lemma 3.3. Let d = k + 1 and KXk+1 ∼
= Vk , k ≥ 3. Then the algebra Fk+12
not finitely generated.
Proof. The group SL2 (K) does not fix any nonzero element of Vk because k ≥ 3,
If k ≥ 4 we use Proposition 2.5 which gives that the transcendence degree of
SL (K)
K[Xk+1 ]SL2 (K) is equal to k − 2 ≥ 2 and by Proposition 2.13 the algebra Fk+12
is
not finitely generated. If k = 3 we shall apply Lemma 2.12. Since the transcendence
degree of K[X4 ]SL2 (K) 6= K is equal to 1 and K[X4 ]SL2 (K) 6= K, it is sufficient
SL (K)
SL (K)
to show that F4 2
6= 0. A nonzero element in F4 2
is given in Example
2.10.
Remark 3.4. We may give a constructive proof of Lemma 3.3. Example 2.10
SL (K)
provides a nonzero element in Fk+12
for k ≥ 3 odd. In order to apply Lemma 2.12
we need an SL2 (K)-invariant polynomial of positive degree. For this purpose we
may take the discriminant. Recall that if θ1 , . . . , θk are the zeros of the polynomial
f (t) = ξ0 tk + ξ1 tk−1 + · · · + ξk−1 t + ξk , then the discriminant
Y
ξ02k−2
(θi − θj )2
1≤i<j≤k
18
VESSELIN DRENSKY, ŞEHMUS FINDIK
of f (t) is an SL2 (K)-invariant polynomial of degree 2(k − 1) in K[Vk ] = K[Xk+1 ].
For k > 3 even we may choose one more explicit SL2 (K)-invariant of second degree
in K[Xk+1 ], namely the element Example 2.10. Then we may produce a nonzero
SL (K)
element in Fk+12
applying Proposition 2.3.
Lemma 3.5. Let
∼ V1 ⊕ V0 ⊕ · · · ⊕ V0 ,
KXd =
where V1 and V0 ⊕ · · · ⊕ V0 have bases X2 and {x3 , . . . , xd }, respectively. Then the
SL (K)
algebra Fd 2
is generated by [x2 , x1 ] and {x3 , . . . , xd }.
Proof. We proceed by induction on d applying Proposition 3.1 in each step. For
d = 2 we have that K[X2 ]SL2 (K) = K[V1 ]SL2 (K) = K. Since as a GL2 (K)-module
F2′ decomposes as a direct sum of det ⊗Vn , n ≥ 0, we obtain that F2SL2 is the
one-dimensional vector space spanned by the commutator [x2 , x1 ]. By induction,
SL (K)
we assume that the Lie algebra Fd−12
is generated by [x2 , x1 ], x3 , . . . , xd−1 and
has a basis consisting of all commutators
[x2 , x1 ]adn3 x3 · · · adnd−1 xd−1 ,
ni ≥ 0,
i = 3, . . . , d − 1,
containing [x2 , x1 ] and all commutators
[xi1 , xi2 , . . . , xin ],
i1 > i2 ≤ · · · ≤ in ,
ij = 3, . . . , d − 1,
n = 1, 2, . . . .
The vector space of the SL2 (K)-invariants in the augmentation ideal of K[Xd−1 ] has
a basis consisting of all monomials in x3 , . . . , xd−1 . Now Proposition 3.1 implies that
the SL2 (K)-invariants in Fd′ are spanned by the elements [x2 , x1 ]adn3 x3 · · · adnd xd
and the commutators of length ≥ 2 in x3 , . . . , xd . Since x3 , . . . , xd belong to
SL (K)
SL (K)
Fd 2 , this implies that Fd 2
is generated by [x2 , x1 ] and {x3 , . . . , xd }.
SL (K)
= 0.
Lemma 3.6. Let d = 3 and KX3 ∼
= V2 , Then F3 2
Proof. In [5, Example 3.2] we have computed the Hilbert series of the algebra
UT (K)
F3 2 :
t31 t2 z 2
UT (K)
HGL2 (F3 2 , t1 , t2 , z) = t21 z +
(1 − t21 z)(1 − t1 t2 z)
= t21 z +
tu2 z 2
t1 (t1 t2 )2 z 2
= t2 z +
,
2
2
(1 − t1 z)(1 − t1 t2 z)
(1 − t z)(1 − uz)
SL (K)
where t = t1 , u = t1 t2 . Since the Hilbert series of F3 2
is obtained from
UT2 (K)
the Hilbert series of F3
substituting t = 0 and u = 1, we obtain that
SL (K)
SL (K)
H(F3 2 , z) = 0, and hence F3 2
= 0.
SL (K)
is
Lemma 3.7. Let KXd ∼
= V1 ⊕ V1 , V2 ⊕ V0 , V2 ⊕ V1 or V2 ⊕ V2 . Then Fd 2
not finitely generated.
Proof. If KXd ∼
= Vi ⊕ Vj , we assume that Vi and Vj have bases {x1 , . . . , xi+1 }
and {xi+2 , . . . , xi+j+2 }, respectively. First, we shall construct nonzero elements in
(Fd′ )SL2 (K) .
(i) Let d = 4 and KX4 ∼
= V1 ⊕ V1 . Consider the element
v11 = [x1 , x4 ] − [x2 , x3 ] ∈ F4 .
SL2 (K)
Applying Remark 2.9 we see that g1 (v11 ) = g2 (v11 ) = v11 . Hence v11 ∈ F4
.
CLASSICAL INVARIANT THEORY FOR FREE METABELIAN LIE ALGEBRAS
19
(ii) Let d = 4 and KX4 ∼
= V2 ⊕ V0 . By Example 2.6 the element f = x22 − x1 x3
SL2 (K)
belongs to K[X3 ]
= K[V2 ]SL2 (K) . Since x4 is an SL2 (K)-invariant, applying
Proposition 3.1 we obtain the nonzero element
v20 = π(x4 , f ) =
3
X
[x4 , xk ]
k=1
∂f
(adX3 )
∂xk
= 2[x4 , x2 , x2 ] − [x4 , x1 , x3 ] − [x4 , x3 , x1 ] ∈ (F4′ )SL2 (K) .
(iii) Let d = 5 and KX5 ∼
= V2 ⊕ V1 . Then the simplest nonzero SL2 (K)-invariant
is
SL (K)
v21 = [x4 , x5 ] ∈ F5 2 .
(iv) Let d = 6 and KX6 ∼
= V2 ⊕ V2 . Using the element
2
X 2
(−1)i ξi ⊗ ξ2−i = ξ0 ⊗ ξ2 − 2ξ1 ⊗ ξ1 + ξ2 ⊗ ξ0 ∈ V2 ⊗ V2
w=
i
i=0
from Example 2.10 we can obtain the nonzero SL2 (K)-invariant
SL2 (K)
v22 = [x1 , x6 ] − 2[x2 , x5 ] + [x3 , x4 ] ∈ F6
.
By Example 2.6 in all four cases we have SL2 (K)-invariants of positive degree
in K[Xd ]. In (i) this is f = x1 x4 − x2 x3 and in (ii)–(iv) – f = x22 − x1 x3 . Since
SL (K)
KXd 2
= 0 in the cases (i), (iii) and (iv), by Lemma 2.12 we conclude that the
SL2 (K)
algebra Fd
is not finitely generated.
SL (K)
Let us assume that in the case (ii) the algebra F4 2
is finitely generated.
SL2 (K)
SL2 (K)
Since KX4
= Kx4 , this implies that F4
is generated by x4 and a finite
number of elements u1 , . . . , uk ∈ (F4′ )SL2 (K) . Hence (F4′ )SL2 (K) is spanned by
ui adn x4 ,
But
(F4′ )SL2 (K)
i = 1, . . . , k,
n ≥ 0.
contains the infinite sequence of elements
v22 (ad2 x2 − adx1 adx3 )m ,
m = 0, 1, 2, . . . ,
and for sufficiently large m these elements cannot be expressed as linear combinaSL (K)
tions of ui adn x4 . Again, F4 2
is not finitely generated.
The following theorem is one of the main results of the paper.
Theorem 3.8. Let KXd be isomorphic as an SL2 (K)-module to the direct sum
Vk1 ⊕· · ·⊕Vkr , where (k1 , . . . , kr ) 6= (0, . . . , 0) and let Fd = Fd (A2 ) be the d-generated
SL (K)
free metabelian Lie algebra. Then the algebra Fd 2
of SL2 (K)-invariants is
finitely generated if and only if KXd ∼
= V2 .
= V1 ⊕ V0 ⊕ · · · ⊕ V0 or KXd ∼
Proof. Let KXd ∼
= Vk1 ⊕ · · · ⊕ Vkr . We may assume that k1 ≥ · · · ≥ kr . By
SL (K)
Lemmas 3.5 and 3.6 the algebra Fd 2
is finitely generated in the cases KXd ∼
=
∼
V1 ⊕ V0 ⊕ · · · ⊕ V0 and KXd = V2 . Hence we have to show that in the cases (i)
SL (K)
k1 ≥ 3; (ii) k1 = 2, r ≥ 2; (iii) k1 = k2 = 1 the algebra Fd 2
is not finitely
generated. Applying Lemma 3.2 we reduce the considerations to the following three
cases:
SL (K)
(i′ ) KXd ∼
is not finitely
= Vk , k ≥ 3. By Lemma 3.3 the algebra Fd 2
generated.
(ii′ ) KXd ∼
= V2 ⊕ Vi , i = 0, 1, 2. Now we apply Lemma 3.7.
20
VESSELIN DRENSKY, ŞEHMUS FINDIK
(iii′ ) KXd ∼
= V1 ⊕ V1 . We apply Lemma 3.7 again and complete the proof of the
theorem.
4. Explicit invariants in small dimensions
In this section we compute the explicit generators of the finitely generated
K[Xd ]SL2 (K) -module (Fd′ )SL2 (K) for d ≤ 5 and for d = 6 when KX6 ∼
= V2 ⊕ V2 and
KX6 ∼
= V1 ⊕ V1 ⊕ V1 . We omit the cases containing V0 as a component of KXd by
Proposition 3.1. In what follows, the Hilbert series are obtained from [5, Example
3.2] and the generators and the relations between them from [5, Section 5], taking
into account the homogeneous components of bidegree (n, n) only.
Example 4.1. (i) Let d = 3 and KX3 ∼
= V2 . In this case we have from Lemma 3.6
that
H((F3′ )SL2 (K) , z) = 0,
and the algebra (F3′ )SL2 (K) does not have nonzero elements. On the other hand
the only SL2 (K)-invariant of polynomial algebra is f = x1 x3 − x22 , and
1
.
H(K[X3 ]SL2 (K) , z) =
1 − z2
(ii) Let d = 4 and KX4 ∼
= V3 . Then
1
z2
and H(K[X4 ]SL2 (K) , z) =
.
1 − z4
1 − z4
The K[X4 ]SL2 (K) -module (F4′ )SL2 (K) has one generator of degree 2. It is
H((F4′ )SL2 (K) , z) =
v = [x4 , x1 ] − 3[x3 , x2 ],
and the single generator of K[X4 ]SL2 (K) is
f = x21 x24 − 6x1 x2 x3 x4 + 4x1 x33 + 4x32 x4 − 3x22 x23 .
∼ V1 ⊕ V1 . We have a nonzero SL2 (K)-invariant in
(iii) Let d = 4 and KX4 =
(F4′ )SL2 (K) by Lemma 3.7. Now let us find all of the generators of finitely generated
K[X4 ]SL2 (K) -module (F4′ )SL2 (K) . By computations using the suggested method we
have
3z 2
1
H((F4′ )SL2 (K) , z) =
and H(K[X4 ]SL2 (K) , z) =
.
1 − z2
1 − z2
The free generators of the K[X4 ]SL2 (K) -module (F4′ )SL2 (K) are as follows:
v1 = [x4 , x1 ] − [x3 , x2 ],
v2 = [x2 , x1 ] and v3 = [x4 , x3 ].
The single generator of K[X4 ]SL2 (K) is f = x1 x4 − x2 x3 .
(iv) Let d = 5 and KX5 ∼
= V4 . Then
H((F5′ )SL2 (K) , z) =
z5
,
(1 − z 2 )(1 − z 3 )
and
H(K[X5 ]SL2 (K) , z) =
1
.
(1 −
− z3)
The only generator of the K[X5 ]SL2 (K) -module (F5′ )SL2 (K) is
z 2 )(1
v = [x2 , x1 ]·(12x34 +4x2 x25 −16x3 x4 x5 )+[x3 , x1 ]·(−2x1 x25 +18x23 x5 −18x3 x24 +4x2 x4 x5 )
+[x3 , x2 ] · (16x2 x24 − 18x2 x3 x5 ) + [x4 , x1 ] · (−4x2 x24 + 12x23 x4 + 4x1 x4 x5 − 16x2 x3 x5 )
+[x4 , x2 ] · (−8x2 x3 x4 + 12x22 x5 ) + [x5 , x1 ] · (x1 x3 x5 − 9x33 − 3x1 x24 − x22 x5 + 14x2 x3 x4 )
CLASSICAL INVARIANT THEORY FOR FREE METABELIAN LIE ALGEBRAS
21
+[x5 , x2 ] · (−8x22 x4 + 6x2 x23 ),
where w · p(Xn ) = wp(adXn ) for w ∈ F5′ and p(Xn ) ∈ K[Xn ]. The generators of
K[X5 ]SL2 (K) are
f1 = x1 x5 − 4x2 x4 + 3x23 ,
and
f2 = −x1 x3 x5 − 2x2 x3 x4 + x33 + x1 x24 + x22 x5 .
(v) Let d = 5 and KX5 ∼
= V2 ⊕ V1 . Then
H((F5′ )SL2 (K) , z) =
z2 + z3 + z4 + z5
,
(1 − z 2 )(1 − z 3 )
and
H(K[X5 ]SL2 (K) , z) =
1
.
(1 − z 2 )(1 − z 3 )
The four generators of the K[X5 ]SL2 (K) -module (F5′ )SL2 (K) are
v1 = [x5 , x4 ],
v2 = [x4 , x2 ] · x5 − [x4 , x3 ] · x4 − [x5 , x1 ] · x5 + [x5 , x2 ] · x4 ,
v3 = [x4 , x1 ]·x3 x5 −[x4 , x2 ]·x3 x4 +[x4 , x3 ]·(x2 x4 −x1 x5 )−[x5 , x1 ]·x2 x5 +[x5 , x2 ]·x1 x5 ,
v4 = −[x4 , x1 ]·x23 x4 +2[x4 , x2 ]·x2 x3 x4 −[x4 , x3 ]·x1 x3 x4 +[x5 , x1 ]·(2x2 x3 x4 −x1 x3 x5 )
+[x5 , x2 ] · (2x1 x2 x5 − 4x22 x4 ) + [x5 , x3 ] · (2x1 x2 x4 − x21 x5 ) + [x5 , x4 ] · (2x32 − 2x1 x2 x3 ).
Note that the first generator was suggested in Lemma 3.7. The generators of
K[X5 ]SL2 (K) are
f1 = x22 − x1 x3 and f2 = x1 x25 − 2x2 x4 x5 + x3 x24 .
(vi) Let d = 6 and KX6 ∼
= V1 ⊕ V1 ⊕ V1 . Then
H((F6′ )SL2 (K) , z) =
6z 2 − z 6
,
(1 − z 2 )3
and
H(K[X5 ]SL2 (K) , z) =
1
.
(1 − z 2 )3
The generators of the K[X6 ]SL2 (K) -module (F6′ )SL2 (K) are
v1 = [x2 , x1 ],
v2 = [x4 , x3 ],
v5 = [x6 , x1 ] − [x5 , x2 ],
The generators of K[X6 ]
SL2 (K)
f1 = x1 x4 − x2 x3 ,
v3 = [x6 , x5 ],
v4 = [x4 , x1 ] − [x3 , x2 ],
and v6 = [x6 , x3 ] − [x5 , x4 ].
are
f2 = x1 x6 − x2 x5 ,
and f3 = x3 x6 − x4 x5 .
There is a relation of degree 6 between the generators:
v1 · f32 + v2 · f22 + v3 · f12 − v4 · f2 f3 + v5 · f1 f3 − v6 · f1 f2 = 0.
(vii) Let d = 6 and KX6 ∼
= V2 ⊕ V2 . Then
H((F6′ )SL2 (K) , z) =
z 2 + 2z 3 + 3z 4 − z 6
,
(1 − z 2 )3
and
H(K[X5 ]SL2 (K) , z) =
1
.
(1 − z 2 )3
22
VESSELIN DRENSKY, ŞEHMUS FINDIK
The generators of the K[X6 ]SL2 (K) -module (F6′ )SL2 (K) , the first one was obtained
in Lemma 3.7, are
v1 = [x6 , x1 ] − 2[x5 , x2 ] + [x4 , x3 ],
v2 = [x2 , x1 ] · x6 − [x3 , x1 ] · x5 + [x3 , x2 ] · x4 ,
v3 = [x6 , x1 ] · x5 − [x5 , x1 ] · x6 + [x4 , x2 ] · x6 − [x4 , x3 ] · x5 − [x6 , x2 ] · x4 + [x5 , x3 ] · x4 ,
v4 = [x3 , x1 ] · x25 + [x6 , x1 ] · (2x2 x5 − x1 x6 ) − [x4 , x3 ] · x3 x4 − 2[x2 , x1 ] · x5 x6
+[x5 , x1 ] · (2x2 x6 − 2x3 x5 ) + [x4 , x2 ] · (2x3 x5 − x2 x6 )
+[x5 , x2 ] · (2x3 x4 − 2x2 x5 ) − 2[x3 , x2 ] · x4 x5 − [x6 , x2 ] · x2 x4 ,
v5 = [x3 , x1 ] · (x1 x6 − 2x2 x5 ) + [x6 , x1 ] · (x22 − x1 x3 ) + [x2 , x1 ] · (2x3 x5 − 2x2 x6 )
+2[x5 , x1 ] · x2 x3 + [x4 , x2 ] · x2 x3 − 2[x5 , x2 ] · x22 + [x3 , x2 ] · x2 x4 − [x4 , x1 ] · x23 ,
and
v6 = [x6 , x1 ] · x4 x6 + [x4 , x3 ] · x4 x6 − 2[x5 , x1 ] · x5 x6 − 2[x4 , x2 ] · x5 x6 + 4[x5 , x2 ] · x25
−2[x6 , x2 ] · x4 x5 − 2[x5 , x3 ] · x4 x5 + [x4 , x1 ] · x26 + [x6 , x3 ] · x24 ,
while the generators of K[X6 ]SL2 (K) are
f1 = x1 x3 − x22 ,
f2 = x1 x6 − 2x2 x5 + x3 x4 ,
and f3 = x4 x6 − x25 ,
with a relation between the generators:
v4 · f2 + v5 · f3 − v6 · f1 + v1 · (f22 + f1 f3 ) = 0.
Now we explain one of the cases in detail.
Example 4.2. Let d = 4 and KX4 ∼
= V3 . Then we have the Hilbert series of the
algebras (F4′ )UT2 (K) and K[X4 ]UT2 (K) from [5, Example 5.2] as follows:
HGL2 ((F4′ )UT2 (K) , t1 , t2 , z) =
=
t31 t2 z 2 (t21 + t22 + t41 t42 z 2 + t51 t62 z 3 − t81 t62 z 4 )
(1 − t31 z)(1 − t21 t2 z)(1 − t61 t62 z 4 )
t4 uz 2 + u3 z 2 + t2 u5 z 4 + tu7 z 5 − t4 u7 z 6
(1 − t3 z)(1 − tuz)(1 − u6 z 4 )
and
HGL2 (K[X4 ]UT2 (K) , t1 , t2 , z) =
(1 − t31 z)(1
3 3 3
1 + t61 t32 z 3
− t41 t22 z 2 )(1 − t61 t62 z 4 )
1+t u z
(1 −
− t2 u2 z 2 )(1 − u6 z 4 )
where t = t1 , u = t1 t2 . We know that Hilbert series of (F4′ )SL2 (K) and of
K[X4 ]SL2 (K) are obtained from the Hilbert series above, respectively, substituting t = 0 and u = 1. Note that the degree of u gives the information of symmetric
bidegrees of the homogeneous generators. Thus we have
=
t3 z)(1
1
z2
and H(K[X4 ]SL2 (K) , z) =
.
4
1−z
1 − z4
The Hilbert series suggest that the K[X4 ]SL2 (K) -module (F4′ )SL2 (K) has one generator of bidegree (3, 3) and degree 2, and K[X4 ]SL2 (K) is generated by a generator
of bidegree (6, 6) and degree 4. A candidate for a generator of K[X4 ]SL2 (K) is of
the form
f = c1 x21 x24 + c2 x1 x2 x3 x4 + c3 x1 x33 + c4 x32 x4 + c5 x22 x23 .
Since it is SL2 (K)-invariant, then it is preserved under the actions of g1 and g2 .
Easy computations give that (c1 , c2 , c3 , c4 , c5 ) = (1, −6, 4, 4, −3). We follow the
H((F4′ )SL2 (K) , z) =
CLASSICAL INVARIANT THEORY FOR FREE METABELIAN LIE ALGEBRAS
23
similar argument as in the previous case to obtain the generator of the K[X4 ]SL2 (K) module (F4′ )SL2 (K) of the form
v = d1 [x4 , x1 ] + d2 [x3 , x2 ],
and find that (d1 , d2 ) = (1, −3).
On the other hand, the Hilbert series of K[X4 ]SL2 (K) -submodule of (F4′ )SL2 (K)
generated by v = [x4 , x1 ] − 3[x3 , x2 ] is equal to the Hilbert series of whole algebra.
Thus (F4′ )SL2 (K) is a free cyclic K[X4 ]SL2 (K) -module generated by v. As a vector
space (F4′ )SL2 (K) is spanned by the elements v · f n , n ≥ 0, forming an infinite set
of generators of (F4′ )SL2 (K) as an algebra.
Acknowledgements
The second named author is very thankful to the Institute of Mathematics and
Informatics of the Bulgarian Academy of Sciences for the creative atmosphere and
the warm hospitality during his visit as a post-doctoral fellow when part of this
project was carried out. The first named author is grateful to Leonid Bedratyuk
for the useful discussions on classical invariant theory.
References
[1] G. Almkvist, W. Dicks, E. Formanek, Hilbert series of fixed free algebras and noncommutative
classical invariant theory, J. Algebra 93 (1985), 189-214.
[2] Yu. A. Bahturin, Identical Relations in Lie Algebras (Russian), Nauka, Moscow, 1985. Translation: VNU Science Press, Utrecht, 1987.
[3] F. Benanti, S. Boumova, V. Drensky, G. K. Genov, P. Koev, Computing with rational symmetric functions and applications to invariant theory and PI-algebras, Serdica Math. J. 38
(2012), 137-188. http://www.math.bas.bg/serdica/2012/2012-137-188.pdf.
[4] R. M. Bryant, On the fixed points of a finite group acting on a free Lie algebra, J. London
Math. Soc. (2) 43 (1991), 215-224.
[5] R. Dangovski, V. Drensky, Ş. Fındık, Weitzenböck derivations of free metabelian Lie algebras,
Linear Algebra Appl. 439 (2013), No. 10, 3279-3296.
[6] C. De Concini, D. Eisenbud, C. Procesi, Young diagrams and determinantal varieties, Invent.
Math. 56 (1980), 129-165.
[7] M. Domokos, V. Drensky, A Hilbert-Nagata theorem in noncommutative invariant theory,
Trans. Amer. Math. Soc. 350 (1998), 2797-2811.
[8] V. Drensky, Fixed algebras of residually nilpotent Lie algebras, Proc. Amer. Math. Soc. 120
(1994), 1021-1028.
[9] V. Drensky, Commutative and noncommutative invariant theory, Math. and Education in
Math., Proc. of the 24-th Spring Conf. of the Union of Bulgar. Mathematicians, Svishtov,
April 4-7, 1995, Sofia, 1995, 14-50.
[10] V. Drensky, C. K. Gupta, Constants of Weitzenböck derivations and invariants of unipotent
transformations acting on relatively free algebras, J. Algebra 292 (2005), 393-428.
[11] R. Ehrenborg, G.-C. Rota, Apolarity and canonical forms for homogeneous polynomials, Eur.
J. Comb. 14 (1993), 157-181.
[12] E. B. Elliott, On linear homogeneous diophantine equations, Quart. J. Pure Appl. Math. 34
(1903), 348-377.
[13] E. Formanek, Noncommutative invariant theory, Contemp. Math. 43 (1985), 87-119.
[14] P. Gordan, Vorlesungen über Invariantentheorie, Herausgegeben von G. Kerschensteiner,
Zweiter Band: Binare Formen, Teubner, Leipzig, 1887. Reprinted by Chelsea Publishing
Company, New York, 1987.
[15] D. Hilbert, Über die vollen Invariantensysteme, Math. Ann. 42 (1893), 313-373. Reprinted
in Gesammelte Abhandlungen, Band II: Algebra. Invariantentheorie. Geometrie, 282-344,
Springer-Verlag, Berlin-Heidelberg-New York, 1970.
24
VESSELIN DRENSKY, ŞEHMUS FINDIK
[16] C. G. J. Jacobi, De determinantibus functionalibus, J. Reine Angew. Math. 22 (1841), 319359. Translation: Ueber die Functionaldeterminanten, Ostwald’s Klassiker d. exacten Wissensch. 78, Leipzig, W. Engelmann, 1896.
[17] O. G. Kharlampovich, M. V. Sapir, Algorithmic problems in varieties, Intern. J. Algebra and
Computation 5 (1995), 379-602.
[18] A. I. Kostrikin, The Burnside problem (Russian), Izv. Akad. Nauk SSSR, Ser. Mat. 23 (1959),
3-34. Translation: Amer. Math. Soc. Transl. (2) 36 (1964), 63-99.
[19] I. G. Macdonald, Symmetric Functions and Hall Polynomials, Oxford Univ. Press (Clarendon), Oxford, 1979, Second Edition, 1995.
[20] P. A. MacMahon, Combinatory Analysis, vols. 1 and 2, Cambridge Univ. Press. 1915, 1916.
Reprinted in one volume: Chelsea, New York, 1960.
[21] L. Makar-Limanov, U. Umirbaev, The Freiheitssatz for Poisson algebras, J. Algebra 328
(2011), 495-503.
[22] I. P. Shestakov, U. U. Umirbaev, Poisson brackets and two-generated subalgebras of rings of
polynomials, J. Amer. Math. Soc. 17 (2004), 181-196.
[23] A. L. Shmel’kin, Wreath products of Lie algebras and their application in the theory of groups
(Russian), Trudy Moskov. Mat. Obshch. 29 (1973), 247-260. Translation: Trans. Moscow
Math. Soc. 29 (1973), 239-252.
[24] T. A. Springer, Invariant Theory, Lecture Notes in Mathematics 585, Springer-Verlag, BerlinHeidelberg-New York, 1977.
[25] R. M. Thrall, On symmetrized Kronecker powers and the structure of the free Lie ring, Amer.
J. Math. 64 (1942), 371-388.
[26] N. Vonessen, Actions of Linearly Reductive Groups on Affine PI-Algebras, Mem. Amer.
Math. Soc. 414, 1989.
[27] H. Weyl, The Classical Groups, Their Invariants and Representations, Princeton Mathematical Series 1, Princeton University Press, Princeton, NJ, 1946. Reprint in Princeton
Landmarks in Mathematics, 1997.
[28] E. I. Zel’manov, On Engel Lie algebras (Russian), Sibirsk. Mat. Zh. 29 (1988), No. 5, 112-117.
Translation: Siberian Math. J. 29 (1988), 777-781.
Institute of Mathematics and Informatics, Bulgarian Academy of Sciences, 1113
Sofia, Bulgaria
E-mail address: [email protected]
Department of Mathematics, Çukurova University, 01330 Balcalı, Adana, Turkey
E-mail address: [email protected]
| 0math.AC
|
THE ARTIN-SPRINGER THEOREM FOR QUADRATIC FORMS
OVER SEMI-LOCAL RINGS WITH FINITE RESIDUE FIELDS
arXiv:1602.07739v2 [math.AC] 29 Feb 2016
STEPHEN SCULLY
Abstract. Let R be a commutative and unital semi-local ring in which 2 is
invertible. In this note, we show that anisotropic quadratic spaces over R remain
anisotropic after base change to any odd-degree finite étale extension of R. This
generalization of the classical Artin-Springer theorem (concerning the situation
where R is a field) was previously established in the case where all residue fields of
R are infinite by I. Panin and U. Rehmann. The more general result presented here
permits to extend a fundamental isotropy criterion of I. Panin and K. Pimenov for
quadratic spaces over regular semi-local domains containing a field of characteristic
6= 2 to the case where the ring has at least one residue field which is finite.
1. Introduction
Let R be a commutative and unital ring in which 2 is invertible. By a (free)
quadratic space over R, we mean a pair (V, q) consisting of a free R-module V of
finite rank and a non-degenerate quadratic form q on V (where, by non-degenerate,
we mean that the polar bilinear form bq of q (see §2.A) has trivial radical, i.e., that
the R-linear map V → V ∗ given by v 7→ bq (v, −) is bijective). Given an extension
of (commutative, unital) rings R ֒→ S, there is an induced base-change procedure
transforming quadratic spaces over R to quadratic spaces over S; namely, to any
quadratic space (V, q) over R, one associates the quadratic space (VS , qS ) over S
obtained by setting VS = V ⊗R S and taking qS to be the unique (non-degenerate)
quadratic form on VS which restricts to q on V (via the canonical inclusion V ֒→ VS ).
A quadratic space (V, q) over R is said to be (strictly) isotropic if there exists a
unimodular vector v ∈ V for which q(v) = 0. If no such unimodular vector exists,
then we say that (V, q) is (strictly) anisotropic. In other words, (V, q) is anisotropic
if the projective quadric {q = 0} ⊆ P(V ) over R admits no R-valued points.
A basic, yet fundamental problem for the theory of quadratic spaces over rings is
that of understanding the isotropy behaviour of quadratic spaces after base change
to another ring of interest. In [P09], I. Panin proved the following important result
in this direction: If (V, q) is an anisotropic quadratic space over a regular semi-local
domain R containing a field of characteristic 0, then it remains anisotropic after
base change to the fraction field of R. Later, I. Panin and K. Pimenov ([PP10])
showed that the statement also holds if R is a regular semi-local domain of finite
characteristic (different from 2), provided that all residue fields of R are infinite.
Date: February 5, 2018.
2010 Mathematics Subject Classification. Primary: 11E81; Secondary: 11E08.
Key words and phrases. Quadratic forms, semi-local rings, Artin-Springer theorem.
The author was supported by a PIMS postdoctoral fellowship held at the University of Alberta.
1
2
STEPHEN SCULLY
The significance of these results is that they permit to reduce a variety of problems
concerning quadratic spaces over sufficiently nice semi-local domains to the case of
fields, where a rich and extensive theory is already available. Recent applications
to the study of Witt groups and Milnor-Witt K-groups of regular equicharacteristic
semi-local domains may be found in [G15] and [GSZ16].
One of the main ingredients used in the aforementioned work of Panin and
Panin-Pimenov is the following generalization of the classical Artin-Springer theorem ([Sp52]) concerning the existence of odd-degree points on quadrics: If (V, q) is
an anisotropic quadratic space over a semi-local ring R with infinite residue fields
(of characteristic 6= 2), then (V, q) remains anisotropic after base change to any
odd-degree finite étale extension of R (the Artin-Springer theorem being the special
case where R is a field ). This statement, which was established by Panin and U.
Rehmann in [PR07], was used by Panin and Pimenov to prove the following more
precise version of their result (cf. [PP10, Proof of Thm. 1.1]): If R is a regular
semi-local domain containing a field of characteristic 6= 2, and the Artin-Springer
theorem holds for quadratic spaces over R, then anisotropic quadratic spaces over
R remain anisotropic after base change to the fraction field of R.
The purpose of this note is to show that the Artin-Springer theorem holds for
quadratic spaces over any semi-local ring in which 2 is invertible (see Theorem 4.1
below). In light of the above discussion, this makes the Panin-Pimenov isotropy
criterion available for quadratic spaces over any regular semi-local domain of finite
characteristic 6= 2, which was our original objective. The stated extension of the
Artin-Springer theorem is achieved by means of an elementary reduction to the case
already treated in [PR07]. The key tool used to carry out this reduction is Proposition 3.1 below, which is perhaps of some interest in its own right. Some additional
consequences of our extension results are also briefly discussed in §5.
Terminology. In this note, all rings are assumed to be commutative and unital.
By a variety, we mean an integral scheme of finite type over a field.
Acknowledgements. I would like to thank Stefan Gille for bringing the problem
discussed in this text to my attention. I would also like to thank Olivier Haution
and Ivan Panin for helpful discussions.
2. Preliminaries
In this section we recall some facts and fix some notation concerning quadratic
forms over rings (in which 2 is invertible). The main reference here is [Ba78], but we
do not assume much more than a basic knowledge of the theory of quadratic forms
over fields. Note that we only consider quadratic forms defined on free modules of
finite rank, as opposed to quadratic forms defined on arbitrary projective modules.
In particular, the Witt rings considered below are Witt rings of free quadratic spaces.
2.A. Quadratic spaces over rings. Let R be a ring in which 2 is invertible and
let V be an R-module. By a quadratic form on V , we mean a map q : V → R
such that (i) q(λv) = λ2 q(v) for all (λ, v) ∈ R × V , and (ii) the symmetric form
bq : V × V → R given by bq (v, w) = q(v + w) − q(v) − q(w) is R-bilinear. If W is
THE ARTIN-SPINGER THEOREM FOR QUADRATIC FORMS OVER SEMI-LOCAL RINGS 3
an R-submodule of V , then the orthogonal complement of W in V (with respect
to q) is defined as the R-submodule W ⊥ consisting of all vectors v ∈ V such that
bq (v, w) = 0 for all w ∈ W . If the R-linear map W → W ∗ given by w 7→ bq (w, −) is
bijective, then we say that W is non-degenerate (with respect to q). In this case, V
decomposes as the internal direct sum of W and W ⊥ . If V itself is non-degenerate,
then we say that the quadratic form q is non-degenerate. In this note, we will be
exclusively concerned with non-degenerate quadratic forms on free modules of finite
rank. To this end, it is worth noting the following:
Lemma 2.1. Let (V, q) be a pair consisting of an R-module V and a quadratic form
q on V , let W be an R-submodule of V , and let n be a positive integer. Then W is
free and non-degenerate of rank n if and only if there exist n R-module generators
w1 , . . . , wn of W such that the matrix bq (wi , wj ) is invertible over R.
Proof. Theonly point worthy of remark is that the invertibility of the matrix M :=
bq (wi , wj ) implies the freeness of Rw1 + · · · + Rwn ⊆ V . Observe, however, that
if r1 w1 + · · · + rn wn = 0 for some r1 , . . . , rn ∈ R, then the vector (r1 , . . . , rn ) lies in
the kernel of M, whence the claim.
By a (free) quadratic space over R, we mean a pair (V, q) consisting of a free
R-module V of finite rank and a non-degenerate quadratic form q on V . The rank
of (V, q) is defined as the rank of the free R-module V . By a subspace of (V, q) we
mean a quadratic space (W, p) such that W is a free non-degenerate submodule of
V and p = q|W (i.e., p is the restriction of q to W ).
Given an R-algebra S, we can define a quadratic space (VS , qS ) over S by setting
VS = V ⊗R S and taking qS to be the unique (non-degenerate) quadratic form on
VS which restricts to q on V (we suppress from the notation the dependency of this
construction on the fixed R-algebra structure on S). In the other direction, if S is a
free R-algebra of finite rank and we are given an R-linear map s : S → R such that
the quadratic form S → R, x 7→ s(x2 ) is non-degenerate, then, given any quadratic
space (W, p) over S, the pair (R W, s ◦ p) is a quadratic space over R (where by R W
we mean W regarded as an R-module via the given R-algebra structure on S).
If (U, p) is another quadratic space over R, then an isometry
from (U, p) to (V, q)
is an R-linear isomorphism φ : U → V such that q φ(u) = p(u) for all u ∈ U. If
such an isometry exists, then we say that (U, p) and (V, q) are isometric. The set
O(V, q) of all isometries from a fixed quadratic space (V, q) to itself is equipped with
a natural group structure, and is called the orthogonal group of (V, q). If w
∈ V is
−1
such that q(w) is invertible in R, then the map v 7→ v − q(w) bq (v, w) w is an
element of O(V, q). Self-isometries of this kind are called reflections.
The orthogonal sum (U ⊕V, p ⊥ q) and tensor product (U ⊗R V, p ⊗q) of quadratic
spaces (U, p) and (V, q) over R are defined in the obvious way. These operations
descend to the level of isometry classes, thus equipping the set of all isometry classes
of quadratic spaces over R with the structure
of a commutative unital semi-ring
2
with unit given by the class of (R, x 7→ x ) . The Grothendieck completion of this
semi-ring is called the Grothendieck-Witt ring of (free) quadratic spaces over R, and
is denoted GW(R) (see [Ba78, §I.2] for more details).
4
STEPHEN SCULLY
Given an invertible element a in R, we write hai for the rank-1 quadratic space
(R, x 7→ ax2 ) over R. Given n invertible elements a1 , . . . , an in R, we then write
ha1 , . . . , an i for the n-fold orthgonal sum ha1 i ⊥ · · · ⊥ han i. A quadratic space
isometric to one of this kind is said to be diagonalizable. Note that every rank-1
quadratic space is diagonalizable. If a quadratic space (V, q) over R admits a free
non-degenerate rank-1 R-submodule L, then we have the orthogonal decomposition
(V, q) = (L, q|L ) ⊥ (L⊥ , q|L⊥ ). In light of the preceding remarks, this implies:
Lemma 2.2. If every quadratic space over R admits a free non-degenerate rank-1
R-submodule, then every quadratic space over R is diagonalizable.
If (V, q) is a quadratic space over R, then a vector v ∈ V is said to be (strictly)
isotropic if v is unimodular (i.e., v generates a free rank-1 R-submodule of V ) and
q(v) = 0. A quadratic space which admits an isotropic vector is said to be (strictly)
isotropic, and (strictly) anisotropic otherwise. The most important example of an
isotropic quadratic space over R is the hyperbolic plane H = (R ⊕ R, (x, y) 7→ xy).
A quadratic space is said to be hyperbolic if it is isometric to an orthogonal sum of
hyperbolic planes. The set of isometry classes of hyperbolic spaces over R forms an
ideal in the Grothendieck-Witt ring GW(R). The quotient of GW(R) by this ideal
is called the Witt ring of (free) quadratic spaces over R, and is denoted by W(R).
If f : R → S is a homomorphism of rings, then the assignment (V, q) 7→ (VS , qS )
descends to a ring homomorphism f ∗ : W(R) → W(S) called restriction along f . If f
equips S with the structure of a free R-module of finite rank, and s : S → R is an Rlinear map such that the quadratic form S → R, x 7→ s(x2 ) is non-degenerate, then
the assignment (W, p) 7→ (R W, s ◦ p) gives rise to a ring homomorphism s∗ : W(S) →
W(R), called (Scharlau) transfer along s. The pair (f ∗ , s∗ ) then satisfies the obvious
projection formula; in particular, the composition
f∗
s
∗
W(R) −→ W(S) −
→
W(R)
coincides with multiplication by the element s∗ (h1i) (see [Ba78, I.2.12]).
2.B. Quadratic spaces over finite direct products of rings. Let R1 , . . . , Rs
be rings in which 2 is invertible, and let R = R1 × · · · × Rs be their direct product.
Given a quadratic space (V, q) over R, we obtain, for each 1 ≤ i ≤ s, an induced
quadratic space (Vi , qi ) := (VRi , qRi ) over Ri by restriction along the natural projection R → Ri . Note that the s quadratic spaces (V1 , q1 ), . . . , (Vs , qs ) completely
determine (V, q). Indeed, for each i, the map Ri ֒→ R given by x 7→ x · 1 gives rise
to an inclusion Vi ⊆ V under which q|Vi = qi . Via these inclusions, an arbitrary
vector v ∈ V decomposes uniquely as a sum v = v1 + · · · + vs with vi ∈ Vi for each
i. The quadratic form q then acts on v by the formula q(v) = q1 (v1 ) + · · · + qs (vs ).
Note, furthermore, that the subsets Vi ⊆ V are pairwise mutually orthogonal with
respect to the quadratic form q. This readily implies the following statements:
Lemma 2.3. Let R and (V, q) be as above.
(1) If v ∈ V , then q(v) = 0 if and only if qi (vi ) = 0 for all 1 ≤ i ≤ s.
(2) Vectors v1 , . . . , vn ∈ V generate a free non-degenerate rank-n R-submodule of V
if and only if, for each 1 ≤ i ≤ s, (v1 )i , . . . , (vn )i generate a free non-degenerate
rank-n Ri -submodule of Vi .
THE ARTIN-SPINGER THEOREM FOR QUADRATIC FORMS OVER SEMI-LOCAL RINGS 5
2.C. Quadratic spaces over semi-local rings. The basic strategy for studying
quadratic forms over general rings is that of reduction to the case of fields, where
a rich and extensive theory has been developing since its initiation by E. Witt in
[Wi37]. As expounded in the book [Ba78], this strategy can be applied successfully
to the study of quadratic forms over semi-local rings using reduction modulo the
Jacobson radical and the remarks of the previous subsection. For later reference,
we now point out some basic results which emerge from this philosophy. For the
remainder of this subsection, we let R be a semi-local ring in which 2 is invertible
and set R = R/Jac(R), where Jac(R) denotes the Jacobson radical of R. Then
R decomposes (essentially uniquely) as a finite direct product of fields, say R ≃
k1 × · · · × ks . If V is an R-module, then we write V for the R-module obtained by
restricting V along the quotient map R 7→ R. Similarly, for each 1 ≤ i ≤ s, we write
V i for the ki -vector space obtained from V by restriction along R → ki . The image
of a vector v ∈ V under the canonical projection V → V will be denoted by v. In
the same way, for each 1 ≤ i ≤ s, we write v i for the image of v under the projection
V → V i . Note that we have v = v 1 + · · · + vs . Furthermore, v is unimodular if and
only if v i 6= 0 for all 1 ≤ i ≤ s. If q is a quadratic form on V , then the induced
quadratic form on V will be denoted by q. Similarly, for each 1 ≤ i ≤ s, the induced
quadratic form on the ki -vector space V i will be denoted by q i . In particular, given
a quadratic space (V, q) over R, we obtain in this way a quadratic space (V , q) over
R and, for each 1 ≤ i ≤ s, a quadratic space (V i , qi ) over ki .
Lemma 2.4. Let (V, q) be a quadratic space over R. Then v1 , . . . , vn ∈ V generate
a free non-degenerate rank-n R-submodule of V if and only if, for each 1 ≤ i ≤ s,
(v1 )i , . . . , (vn )i generate a non-degenerate rank-n ki -linear subspace of V i .
Proof. In view of Lemma 2.3 (2), it suffices to show that v1 , . . . , vn generate a free
non-degenerate rank-n R-submodule of V if and only if v1 , . . . , vn generate a free
non-degenerate rank-n R-submodule of V . This follows from Lemma
2.1, since an
element of R (in this case the determinant of the matrix bq (vi , vj ) ) is invertible in
R if and only if its image under the projection R → R is invertible in R.
Any quadratic space over a field of characteristic 6= 2 admits a non-degenerate
subspace of rank-1 (see [EKM08, Prop. 7.29]). By the preceding lemma, it follows
that every quadratic space over the semi-local ring R admits a free non-degenerate
submodule of rank 1. Lemma 2.2 therefore implies:
Lemma 2.5. Every quadratic space over R is diagonalizable.
A slightly more subtle observation is the following important lemma:
Lemma 2.6. Let (V, q) be a quadratic space over R, and let v, w ∈ V be unimodular
vectors such that qi (wi ) = q i (v i ) for all 1 ≤ i ≤ s. Then there exists u ∈ V such
that q(u) = q(v) and ui = wi for all 1 ≤ i ≤ s.
Remark 2.7. The condition that q i (wi ) = q i (vi ) for all 1 ≤ i ≤ s just means that
q(w) = q(v). Similarly, the assertion that ui = w i for all 1 ≤ i ≤ s simply means
that u = w. We choose to state the lemma this way for later reference.
6
STEPHEN SCULLY
Proof. The rank-1 case is trivial, so we can assume that (V, q) has rank at least 2.
Since v and w are unimodular, we have vi 6= 0 6= wi for all 1 ≤ i ≤ s. Thus, by
an observation of E. Witt (see [EKM08, §II.8]), there exists, for each 1 ≤ i ≤ s,
an isometry τi ∈ O(V i , q i ) which is a product of (at most 2) reflections and which
satisfies τi (vi ) = w i . Since rank(V ) ≥ 2, every reflection in O(V i , q i ) extends to
a reflection in O(V , q) which fixes vj for all j 6= i. It follows that there exists
τ ∈ O(V , q) which is a product of reflections and which sends v to w. Since an
element of R is invertible if and only if its image in R is invertible, every reflection in
O(V , q) lifts to a reflection in O(V, q). As a result, τ lifts to an isometry φ ∈ O(V, q).
The vector u := φ(v) then has the desired properties.
Remark 2.8. Using the Cartan-Dieudonné theorem on the generation of the orthogonal group of a quadratic space over a field by reflections (see [Sc85, Theorem I.5.4]),
one can show by similar means that the natural projection SO(V, q) → SO(V , q) on
special orthogonal groups is surjective. This observation is due to M. Knebusch (see
[Kn69]). If R is a local ring, then even O(V, q) → O(V , q) is surjective.
It is worth stating explicitly the following special case of Lemma 2.6 which will
be needed later:
Lemma 2.9. Let (V, q) be an isotropic quadratic space over R. Suppose that, for
each 1 ≤ i ≤ s, we are given a non-zero vector vi ∈ V i such that q i (vi ) = 0. Then
there exists a (strictly) isotropic vector v ∈ V such that vi = vi for all 1 ≤ i ≤ s.
Now, using Lemma 2.6 (see also Remark 2.8), one can readily show that Witt’s
cancellation and decomposition theorems hold for (free) quadratic spaces over R
(see [Ba78, Ch. III]). As a result, the standard characterizations of isotropy for
quadratic spaces over fields are all valid over the semi-local ring R:
Lemma 2.10. Let (V, q) be a quadratic space over R and let Q ⊆ P(V ) be the
projective R-quadric defined by the vanishing of q. Then the following are equivalent:
(1) (V, q) is isotropic.
(2) (V, q) ≃ H ⊥ (W, p) for some quadratic space (W, p) over R.
(3) There exists a quadratic space (W, p) over R with rank(W ) < rank(V ) such that
(V, q) = (W, p) in W(R).
(4) Q has an R-valued point.
(5) Q is rational over R.
(6) The affine quadric {q = 0} ⊆ A(V ) is rational over R.
Proof. If (2) holds, then we can choose coordinates (x : y : z1 : · · · : zn ) on V ∗
so that q(x, y, z1 , . . . , zn ) = xy + p(z1 , . . . , zn ). The open complement in Q of the
principal divisor {x = 0} is then isomorphic to AnR , and so (5) holds. The remaining
implications are either trivial or straightforward consequences of Lemma 2.6 and
standard facts concerning quadratic forms over fields; for the sake of brevity we
omit the details and refer the reader instead to [Ba78, Ch. III].
THE ARTIN-SPINGER THEOREM FOR QUADRATIC FORMS OVER SEMI-LOCAL RINGS 7
3. Key proposition
The key ingredient needed to extend the Artin-Springer theorem to quadratic
spaces over arbitrary semi-local rings in which 2 is invertible is the (n = 3 case of
the) following proposition, which is of some interest in its own right:
Proposition 3.1. Let R be a semi-local ring in which 2 is invertible, and let S be
a degree-n finite étale extension of R for some (not necessarily odd) integer n ≥ 2.
If (V, q) is a quadratic space of rank > n over R such that (VS , qS ) is isotropic, then
(V, q) contains a rank-n subspace which becomes isotropic over S.
The case where n = 2 is well known and can be made rather more precise (see
[Ba78, Thm. V.4.2]). The general case of Proposition 3.1 follows immediately from
the more specific Lemma 3.2 below. Before proceeding, let us pause to fix some
notation and terminology which will also be used in the next section:
Notation and Terminology. Let R and S be as in the statement of Proposition
3.1. By [EGAIV, 18.4.5], we can write S = R[t]/ f (t) , where f is a monic separable
polynomial of degree n over R. By monic, we mean that the leading coefficient of f
is invertible in R; by separable, we mean that the discriminant of f is invertible in
R. Both conditions are local in the following sense: a polynomial over R is monic
(resp. separable) if and only if it is monic (resp. separable) when restricted to every
residue field of R. The image of the variable t in S will be denoted by θ. Note that
S is a free rank-n R-module with basis 1, θ, . . . , θn−1 .
Our refinement of Proposition 3.1 may now be stated as follows:
Lemma 3.2. In the situation of Proposition 3.1, there exist vectors v0 , v1 , . . . , vn−1 ∈
V such that
(1) v0 + θv1 + · · · + θn−1 vn−1 ∈ VS is a (strictly) isotropic vector for qS .
(2) v0 , v1 , . . . , vn−1 generate a free non-degenerate rank-n submodule of V .
Proof. Let us first remark that the unimodularity of the vector in (1) is guaranteed
by condition (2), and so we will suppress this point in the discussion that follows.
We begin the proof with some reductions. Firstly, since S is a finite étale extension
of R, its Jacobson radical is generated (as an ideal of S) by the Jacobson radical of R.
Thus, by Lemmas 2.4 and 2.9, the problem reduces immediately to the case where
R is a field, which we will now choose to denote by k. Before proceeding, we make
one further reduction to the case where the u-invariant of k is ≤ 2 (that is, every
quadratic space of rank ≥ 3 over k is isotropic). If k is finite, then it has u-invariant
2, and so no reduction is necessary (see [EKM08, Ex. 36.2 (3)]). To treat the case
where k is infinite, consider the affine quadric XqS = {qS = 0} ⊆ A(VS ). Since
qS admits a (strictly) isotropic vector, XqS is a rational S-scheme (Lemma 2.10).
The Weil restriction Y := RS/k (XqS ) is therefore a rational k-variety. Note that
Y is a closed subvariety of A(V ⊕n ): if A is a k-algebra, then
an n-tuple of vectors
Pn−1
⊕n
(v0 , v1 , . . . , vn−1 ) ∈ (VA ) lies in Y (A) if and only if qS⊗k A ( i=0 θi vi ) = 0. Consider
now the k-morphism g : A(V ⊕n ) → A1k given by (v0 , v1 , . . . , vn−1 ) 7→ det bq (vi , vj ) .
By Lemma 1.1, n-vectors v0 , v1 , . . . , vn−1 in V generate a non-degenerate rank-n
8
STEPHEN SCULLY
subspace of V if and only if g(v0, v1 , . . . , vn−1 ) 6= 0. Thus, in light of the above
remarks, the lemma holds if and only if U(k) 6= ∅, where U is the open subvariety
of Y given by U = Y ∩ g −1(A1k \ {0}). Now Y is a rational k-variety, and so if k is
infinite, then U(k) 6= ∅ if and only if U is non-empty as a scheme. In particular, in
order to prove the lemma in this case, we may assume that k is algebraically closed.
Since any quadratic space of rank 2 or more over an algebraically closed field is
evidently isotropic, we have made the necessary reduction.
Working under the assumption that the u-invariant of k is ≤ 2, we will now show
that there exist vectors v0 , v1 , . . . , vn−1 ∈ V satisfying conditions (1) and (2) in the
statement of the lemma. Let s = [ n2 ]. Note that (V, q) has rank strictly greater than
2s by assumption. Since every rank-3 quadratic space over k is isotropic, it follows
that V admits a non-degenerate subspace W of dimension 2s + 1 such that
(W, q|W ) ≃ H
· · ⊥ H} ⊥ hai
| ⊥ ·{z
s times
∗
for some a ∈ k (see Lemma 2.10). More explicitly, this means that we can find
vectors e1 , f1 , e2 , f2 , . . . , es , fs and w in V such that
• q(ei ) = q(fi ) = 0 for all i,
• q(w) = a,
• bq (ei , ej ) = bq (fi , fj ) = 0 for all i 6= j,
• bq (ei , fj ) = δij for all i, j (where δ is the Kronecker delta), and
• bq (ei , w) = bq (fi , w) = 0 for all i.
(the (ei , fi ) are hyperbolic pairs in the sense of [EKM08, p. 40]) To finish the proof,
let us now separate the case where n is even from that where n is odd:
Case 1. n = 2s. Suppose first that s = 1, and let α, β ∈ k be such that θ2 = α + βθ
in S. A direct calculation shows that if we set
β2
β
v0 = (−α − )ae1 + f1 − w
and
v1 = w,
4
2
then qS (v0 + θv1 ) = 0, i.e., condition (1) of the statement holds for this choice
of v0
and v1 . At the same time, the determinant of the Gram matrix bq (vi , vj ) is in this
case equal to −a2 (β 2 + 4α), which is non-zero because S/k is étale. Thus, condition
(2) also holds for this choice of v0 and v1 .
P
Suppose now that s > 1. Let λ1 , . . . , λs ∈ k ∗ be such that si=1 λi = 0, and set
(
λi+1 ei+1 if 0 ≤ i ≤ s − 1
vi =
fn−i
if s ≤ i ≤ n − 1
Then, by construction, we have
X
X
n−1
n−1
s
X
X
i
qS
θ vi =
q(vi )θ2i +
bq (vi , vj )θi+j =
λi θn−1 = 0,
i=0
i=0
i<j
i=1
and so the vi satisfy the first condition of the statement. Condition (2) is also satisfied, since the restriction of q to the subspace spanned by the vi is a hyperbolic form.
THE ARTIN-SPINGER THEOREM FOR QUADRATIC FORMS OVER SEMI-LOCAL RINGS 9
Case 2. n = 2s + 1. Let λ1 , . . . , λs ∈ k ∗
λi+1 ei+1
vi = w
f
n−i
be such that
Ps
i=1
λi = −a, and set
if 0 ≤ i ≤ s − 1
if i = s
if s + 1 ≤ i ≤ n − 1
Then, by construction, we have
X
X
s
n−1
n−1
X
X
i
qS
θ vi =
q(vi )θ2i +
bq (vi , vj )θi+j = q(w)θn−1+
λi θn−1 = (a−a)θn−1 = 0,
i=0
i=0
i<j
i=1
and so the vi satisfy condition (1) of the statement. Again, condition (2) is also
satisfied, since the subspace of V spanned by the vi is nothing else but W .
4. The Artin-Springer theorem
We are now ready to prove our main result:
Theorem 4.1. Let R be a semi-local ring in which 2 is invertible. If (V, q) is an
anisotropic quadratic space over R, then (VS , qS ) is anisotropic for any odd-degree
finite étale extension S of R.
As discussed in §1, the particular case where R is a field is due (independently) to
E. Artin (unpublished) and T.A. Springer ([Sp52]), while the case where all residue
fields of R are infinite was proved in [PR07] by I. Panin and U. Rehmann. Our
purpose is to treat the case where at least one residue field of R is finite. We will
argue by reduction to the case where all residue fields are “large enough”. The key
ingredient here is (the n = 3 case of) Proposition 3.1 above. We continue with the
notation of §3, but, as per the statement of Theorem 4.1, we now assume that n is
odd. We present the proof of Theorem 4.1 in a series of short steps:
4.A. The hyperbolicity analogue of Theorem 4.1. Let i denote the natural
inclusion of R into S. Below, we will make use of the following hyperbolicity analogue
of Theorem 4.1, which is well known and straightforward to prove (see [OP99, §3]):
Proposition 4.2. The natural restriction homomorphism i∗ : W(R) → W(S) is
injective. In other words, if (V, q) is a quadratic space over R such that (VS , qS ) is
hyperbolic, then (V, q) is hyperbolic.
Proof. Let s : S → R be the R-linear map defined by setting s(θn−1 ) = 1 and
s(θi ) = 0 for all 0 ≤ i ≤ n − 2. A straightforward calculation shows that the
quadratic form S → R, x 7→ s(x2 ) is isometric to h1i ⊥ n−1
· H (here n−1
· H denotes
2
2
copies
of
H).
In
particular,
it
is
non-degenerate,
and so
the orthogonal sum of n−1
2
we can consider the Scharlau transfer s∗ : W(S) → W(R). Recall from §2.1 above
i∗
s∗
that the composition W(R) −
→ W(S) −
→ W(R) coincides with multiplication by the
element s∗ (h1i). However, by the above remarks, we have s∗ (h1i) = h1i in W(R),
and so we see that s∗ is a section of i∗ . Hence i∗ is injective, as claimed.
10
STEPHEN SCULLY
4.B. Proof of Theorem 4.1 for quadratic spaces of rank ≤ 3. Using Proposition 4.2, we now show that Theorem 4.1 holds for quadratic spaces of rank ≤ 3.
Proposition 4.3. Theorem 4.1 holds when (V, q) has rank ≤ 3.
Proof. For quadratic spaces of rank 1, the statement is trivial. Next, an isotropic
space of rank 2 over a semi-local ring is hyperbolic (see Lemma 2.10), and so the
rank-2 case of the statement follows from Proposition 4.2. Assume now that (V, q)
has rank 3. For ease of notation, we simply denote the pair (V, q) by q. Now,
by Lemma 2.5, we can write q ≃ ha, b, ci for some a, b, c ∈ R∗ . Then the form
π = q ⊥ habci = ha, b, c, abci is similar to (i.e., isometric to a unit multiple of) the
2-fold Pfister form h1, abi⊗h1, aci. By the main property of Pfister forms (see [Ba78,
Cor. IV.3.2]), πS is either anisotropic or hyperbolic. Suppose that πS is hyperbolic.
Then, by Proposition 4.2, π is already hyperbolic, and so q = −habci in the Witt
ring W(R). But, by Lemma 2.10, this implies that q is isotropic, thus contradicting
our initial hypothesis. We can therefore conclude that πS is anisotropic. Since qS is
a restriction of πS , it must be anisotropic as well.
4.C. Proof of Theorem 4.1 for extensions of degree 3. Next, we use Proposition 4.3 to show that Theorem 4.1 holds in the case where S has degree 3 (with
no restriction on the rank of (V, q)):
Proposition 4.4. Theorem 4.1 holds when S is a degree-3 extension of R.
Proof. The case where (V, q) has rank ≤ 3 was treated in Proposition 4.3 above.
If (V, q) has rank strictly greater than 3, then Proposition 4.3 at least implies that
every rank-3 subspace of (V, q) remains anisotropic over S. In view of the n = 3
case of Proposition 3.1, however, this is all that is needed to make the desired
conclusion.
4.D. Proof of Theorem 3.1 in the general case. We now conclude the proof of
Theorem 3.1. The main ingredient, together with Proposition 4.4, is the following
statement due (essentially) to I. Panin and U. Rehmann:
Lemma 4.5 (Panin-Rehmann). There exists a natural number Mn for which the
following holds: Assume that all residue fields of R contain at least Mn elements.
Then, for any quadratic space (V, q) over R such that (VS , qS ) is isotropic, there
exist v0 , v1 , . . . , vn−1 ∈ V with the following properties:
(1) v0 + θv1 + · · · + θn−1 vn−1 ∈ VS is a (strictly) isotropic vector for qS .
(2) qR[t] (v0 + tv1 + · · · + tn−1 vn−1 ) ∈ R[t] is monic and separable of degree 2n − 2
(here R[t] denotes the polynomial ring in a single variable t over R).
Proof. This is a mild extension of [PR07, Prop. 1.1]. In fact, the statement is proved
in [PR07] under the stronger hypothesis that R is local with infinite residue field.
Note, however, that since S is a finite étale extension of R, its Jacobson radical is
generated (as an ideal of S) by that of R. Thus, in view of Lemma 2.9 and the fact
that condition (2) can be checked locally (see the remarks at the beginning of this
section), everything reduces immediately to the case where R is a field (regardless
of whether R is local or not). Moreover, under the assumption that R is a field, it
was shown in [PR07, §3] that there exists a non-empty rational R-variety U with
THE ARTIN-SPINGER THEOREM FOR QUADRATIC FORMS OVER SEMI-LOCAL RINGS 11
property that U(R) 6= ∅ if and only if the conclusion of the lemma holds. This
conclusion is therefore valid provide that R contains ”enough” elements; the term
“enough” is made formal here by the integer Mn in the statement of the lemma (in
particular, it is not necessary for R to be infinite).
We now give the proof of Theorem 4.1:
Proof of Theorem 4.1. Suppose first that at least one residue field of R is finite. Let
h(x) ∈ R[x] be a monic separable polynomial of degree 3 in a single variable x
over R which is irreducible when restricted to any finite residue field of R. Then
e := R[x]/ h(x) is a degree-3 étale extension of R such that the minimal cardinality
R
e is strictly greater than that of the finite residue fields
of the finite residue fields of R
of R. Now, in order to show that (VS , qS ) is anisotropic, it is sufficient to show that
e By Proposition 4.4, however, (V e , q e )
(VSe, qSe) is anisotropic, where Se = S ⊗R R.
R R
e and, by repeating this process
is anisotropic. We can therefore replace R by R,
sufficiently many times, we ultimately reduce to the case where all residue fields of
R contain at least Mn elements (where Mn is the natural number in the statement of
Lemma 4.5). Having made this reduction, the proof proceeds as in [PR07]. For the
reader’s convenience, we give the details. Thus, suppose for the sake of contradiction
that (VS , qS ) is isotropic, and let v0 , v1 , . . . , vn be as in Lemma 4.5 so that
(1) v0 + θv1 + · · · + θn−1 vn−1 ∈ VS is a (strictly) isotropic vector for qS .
(2) qR[t] (v0 + tv1 + . . . tn−1 vn−1 ) ∈ R[t] is monic and separable of degree 2n − 2.
Let v(t) = v0 + tv1 + . . . tn−1 vn−1 ∈ V ⊗R R[t]. By (1), the polynomial qR[t] v(t) ∈
R[t] is divisible by f (t), say
qR[t] v(t) = f (t)g(t)
with g(t) ∈ R[t].
By (2), g(t) is monic and separable of degree n − 2 over R. Let
S ′ = S[t]/ g(t) . Then S ′ is a degree-(n − 2) étale extension of R. We claim that
(VS ′ , qS ′ ) is isotropic. Note first that if θ′ denotes the image of t in S ′ , then we have
qS ′ v(θ′ ) = f (θ′ )g(θ′ ) = 0
in S ′ . To prove the claim, it therefore suffices to check that v(θ′ ) ∈ VS ′ is unimodular.
We claim that this follows from the separability part of (2). Since unimodularity
of vectors and separability of polynomials can be checked locally (see §2.3 and the
remarks at the beginning of §3), we are, to this end, free to assume that R is a field.
Then g(t) decomposes into a product of pairwise coprime irreducible polynomials
over R, say g(t) = g1 (t) · · · gk (t). If v(θ′ ) were not unimodular, then v(t) would be
divisible by some gi (t), whence qR[t] would be divisible by gi (t)2 in R[t]. As this
would contradict the separability part of (2), the claim follows. Thus, (VS ′ , qS ′ ) is
isotropic. Now S ′ is an odd-degree étale extension of R which has strictly smaller
degree than S. By repeating the above argument sufficiently many times, we come
to the conclusion that (V, q) is isotropic, thus contradicting our original hypothesis.
This completes the argument.
12
STEPHEN SCULLY
5. The Panin-Pimenov isotropy criterion and applications
5.A. The Panin-Pimenov isotropy criterion. As an immediate application of
our extension of the Artin-Springer theorem (Theorem 4.1 above), we obtain the
following extension of an important result due to I. Panin and K. Pimenov:
Theorem 5.1. Let R be a regular semi-local domain containing a field of characteristic 6= 2. If (V, q) is an anisotropic quadratic space over R, then (VK , qK ) is
anisotropic, where K denotes the fraction field of R.
This theorem was originally proved by I. Panin in [P09] under the stronger assumption that R contains a field of characteristic 0. In [PP10], I. Panin and K.
Pimenov treated the case where R has finite characteristic different from 2, but under the assumption that all residue fields of R are infinite. In order to extend the
statement to the case where some residue field of R is finite (which was our original goal), the arguments presented in [PP10] showed that it would suffice to prove
that the Artin-Springer theorem holds for quadratic spaces over R. This extension
therefore follows from our Theorem 4.1 above. In more detail:
Proof of Theorem 4.1. Let (V, q) be a quadratic space over R. If (VK , qK ) is isotropic,
then it is shown in [PP10, Proof of Thm. 1.1] that (V, q) becomes isotropic over an
odd-degree finite étale extension of R (in fact, over R ⊗k k ′ , where k denotes the
prime subfield of R and k ′ is some finite extension of k of degree prime to 2 and the
characteristic of k). By Theorem 4.1, this implies that (V, q) is isotropic.
5.B. Applications. The Panin-Pimenov isotropy criterion is of fundamental importance for the study of quadratic forms over (nice) semi-local rings, since it permits to
reduce many problems to the case of fields, where an extensive literature is already
available (see, e.g., [EKM08]). Let R be a regular semi-local domain containing a
field of characteristic 6= 2. To conclude this note, we now mention (without proof)
some consequences of our extension of the Panin-Pimenov criterion to the case where
at least one residue field of R is finite (Theorem 5.1). These statements were known
previously in the case where the residue fields of R are all infinite.
• All results proved in [G15, §3] are valid for quadratic spaces over R; these include
the linkage theorem (analogous to the original result of R. Elman and T. Y.
Lam concerning quadratic spaces over fields) as well as certain other purity-type
statements for Pfister forms.
• If, moreover, R local, then the purity theorem of [PP10, Cor. 1] holds for quadratic
spaces over R, i.e., unramified quadratic spaces over K := Frac(R) are defined
over R (on the level of quadratic spaces themselves, and not only Witt groups).
• If R is local, then the main results of [CP07] and [CP13] (concerning purity statements for G2 -torsors and certain F4 -torsors, respectively) are valid over R.
References
[Ba78] R. Baeza. Quadratic forms over semilocal rings. Lecture Notes in Mathematics, Vol. 655, Springer-Verlag, Berlin-New York, 1978, vi+199 pp.
[CP07] V. Chernousov and I. Panin. Purity of G2 -torsors. C. R. Math. Acad. Sci.
Paris 345 (2007), no.6, 307–312.
THE ARTIN-SPINGER THEOREM FOR QUADRATIC FORMS OVER SEMI-LOCAL RINGS 13
[CP13] V. Chernousov and I. Panin. Purity for Pfister forms and F4 -torsors with
trivial g3 invariant. J. Reine Angew. Math. 685 (2013), 99–104.
[EGAIV] A. Grothendieck. Eléments de géometrie algébrique. IV.. Inst. Hautes
Études Sci. Publ. Math.
[EKM08] R. Elman, N. Karpenko, and A. Merkurjev. The algebraic and geometric
theory of quadratic forms. AMS Colloquium Publications, 56, American
Mathematical Society, 2008.
[G15] S. Gille. On quadratic forms over semilocal rings. Preprint, 2015.
[GSZ16] S. Gille, S. Scully and C. Zhong. Milnor-Witt K-groups of local rings.
Adv. Math. 286 (2016), 729–753.
[Kn69] M. Knebusch. Isometrien über semilokalen Ringen. (German). Math. Z.
108, 1969, 255–268.
[OP99] M. Ojanguren and I. Panin. A purity theorem for the Witt group. Ann.
Sci. École Norm. Sup. (4) 32 (1999), no. 1, 71–86.
[P09] I. Panin. Rationally isotropic quadratic spaces are locally isotropic. Invent.
Math. 176 (2009), no. 2, 397–403.
[PP10] I. Panin and K. Pimenov. Rationally isotropic quadratic spaces are locally
isotropic: II. Doc. Math. 2010, Extra volume: Andrei A. Suslin sixtieth
birthday, 515–523.
[PR07] I. Panin and U. Rehmann. A variant of a theorem by Springer. Algebra
i Analiz 19 (2007), no. 6, 117–125; translation in St. Petersburg Math. J.
19 (2008), no. 6, 953–959.
[Sc85] W. Scharlau. Quadratic and hermitian forms. Springer-Verlag, Berlin,
1985.
[Sp52] T.A. Springer. Sur les formes quadratiques d’indice zéro. C. R. Acad. Sci.
Paris 234, (1952). 1517-1519.
[Wi37] E. Witt. Theorie der quadratischen Formen in beliebigen Körpern. J.
Reine Angew. Math. 176 (1937), 31–44.
Department of Mathematical and Statistical Sciences, University of Alberta,
Edmonton AB T6G 2G1, Canada
E-mail address: [email protected]
| 0math.AC
|
Non-Parametric Transformation Networks
Dipan K. Pal 1 Marios Savvides 1
arXiv:1801.04520v3 [cs.CV] 14 Feb 2018
Abstract
ConvNets, through their architecture, only enforce invariance to translation. In this paper, we
introduce a new class of deep convolutional architectures called Non-Parametric Transformation Networks (NPTNs) which can learn general
invariances and symmetries directly from data.
NPTNs are a natural generalization of ConvNets
and can be optimized directly using gradient descent. Unlike almost all previous works in deep
architectures, they make no assumption regarding
the structure of the invariances present in the data
and in that aspect are flexible and powerful. We
also model ConvNets and NPTNs under a unified framework called Transformation Networks
(TN), which yields a better understanding of the
connection between the two. We demonstrate the
efficacy of NPTNs on data such as MNIST and
CIFAR10 where they outperform ConvNet baselines with the same number of parameters. We
show it is more effective than ConvNets in modelling symmetries from data, without the explicit
knowledge of the added arbitrary nuisance transformations. Finally, we replace ConvNets with
NPTNs within Capsule Networks and show that
this enables Capsule Nets to perform even better.
1. Introduction
The Fundamental Problem. One of the central problems
of deep learning, and machine learning in general, has been
supervised classification. An instantiation of which, in vision, is object classification. A core challenge towards these
problems is the encoding or learning of invariances and
symmetries that exist in the training data. Any general classification problem would require invariance to the withinclass transformations and symmetries (nuisance transformations) while being selective to between-class transformations. There also has been evidence that the sample complexity of a model is inversely proportional to the amount of
invariance it can invoke towards nuisance transformations
(Anselmi et al., 2013). Indeed, methods which incorporate
some known invariances or promote learning of more powerful invariances for a learning problem perform better in
Figure 1. Operation performed by a single Non-Parametric Transformation Network (NPTN) node (single channel input and single
channel output). NPTNs are a generalization of ConvNets towards
learning general invariances and symmetries. The node has two
main components (a) Convolution and (b) Transformation Pooling. The dot-product between the input patch (x) and a set of |G|
number of filters gw (green) is computed (this results in convolution when implemented with spatially replicated nodes). Here
|G| = 6 (different shades of green indicate transformed templates).
g indicates the transformation applied to the template or filter w
. The resultant six output scalars (red) are then max-pooled over
to produce the final output s (black). The pooling operation here
is not spatially (as in vanilla ConvNets) but rather across the |G|
channels which encode non-parametric transformations. The output s is now invariant to the transformation encoded by the set of
filters G. Each plane indicates a single feature map/filter.
the target task given a certain amount of data. A number
of ways exist to achieve this. One can present transformed
versions of the training data (Burges & Schölkopf, 1997;
Niyogi et al., 1998), minimize auxiliary objectives promoting invariance during training (Schroff et al., 2015; Hadsell
et al., 2006) or pool over transformed versions of the representation itself (Liao et al., 2013; Pal et al., 2016; 2017).
Convolutional Networks and Beyond. Towards this goal,
ideas proposed in (LeCun et al., 1998) with the introduction of convolutional neural networks have proved to be
very useful. Weight sharing implemented as convolutions,
helped to regularize the network and vastly reduce the number of parameters learned. Additionally, it resulted in the
hard encoding of translation invariances (and symmetries)
in the network, making it one of the first applications of
modelling invariance through a network’s architecture itself.
Such a mechanism resulted in greater sample efficiency and
in regularization in the form of a structural or inductive
bias in the network. With this motivation in mind, it is almost natural to ask whether networks which model more
complicated invariances and symmetries perform better?
Non-Parametric Transformation Networks
Investigating architectures which invoke invariances not implicitly through the model’s functional map but explicitly
through an architectural property seems important.
New Dimensions in Network Architecture. Over the
years, deep convolutional networks (ConvNets) have enjoyed a wide array of improvements in architecture. It was
observed early on that a larger number of filters (width) in
ConvNets led to improved performance, though with diminishing returns. Another significant milestone was the development and maturity of residual connections and dense skip
connections (He et al., 2016; Huang et al., 2016). Though
there have been more advances in network architecture,
many of the improvements have been derivatives of these
two ideas (Zagoruyko & Komodakis, 2016; Chen et al.,
2017; Hu et al., 2017). Recently however, Capsule Nets
were introduced (Sabour et al., 2017) which presented another potentially fundamental idea of encoding properties of
an entity or an object in an activity vector rather than a scalar.
With the goal of designing more powerful networks, ideas
for modelling general invariances in the same framework
as ConvNets, open up a new and potentially key dimension
for architecture development.
Primary Contribution. In this work, we explore one such
architecture class, called Transformation Networks (TN).
We introduce a new layer which can form a class of networks
called Non-Parametric Transformation Networks (NPTNs).
These networks have the ability to learn invariances to general transformations present in the data which are (potentially) non-parametric in nature. NPTNs are named so for
their explicit handling of transformation invariances and
symmetries in the data. They can be easily implemented
using standard off-the-shelf deep learning frameworks and
libraries. Further, they can be optimized using vanilla gradient descent methods such as SGD. Unlike other methods
that enforce additional invariances in convolutional architectures, NPTNs do not need to transform the input or the
filters at any stage of the learning/testing process. They enjoy benefits of a standard convolutional architecture such as
speed and memory efficiency while being more powerful in
modelling invariances and being elegant in their operation.
When forced to ignore any learnable transformation invariances in data, they gracefully reduce to vanilla ConvNets in
theory and practice. However, when allowed to do so, they
outperform ConvNets by capture more general invariances.
Some properties of NPTNs. The architecture itself of an
NPTN allows it to be able to learn powerful invariances from
data (a single node is illustrated in Fig. 1). This offer better sample complexity and shortens the generalization gap1 .
Learning invariances from data is different and more pow1
Indeed, empirically we find that three or more layered NPTNs
have a smaller generalization gap, i.e. have higher training losses
than vanilla ConvNets, but lower testing losses.
erful than enforcing known and specific invariances such
as rotation symmetry in networks. Such networks which
enforce predefined symmetries (including vanilla ConvNets)
force the same invariances at all layers which is a strong
prior. More complex invariances are left for the network
to learn using the implicit functional map as opposed to
the explicit architecture. The proposed NPTNs have the
ability to learn different and independent invariances for
different layers and in fact for different channels themselves.
Standard convolution architectures enforce translation invariance through the convolution operation followed by a
aggregation operation (either pooling or a second convolution layer). In this aspect, each node has a predefined
invariance (translation) and only needs to learn the filter
instantiation. However, an NPTN node needs to learn two
independent entities. First, the instantiation of the filter and
second, the transformation that the particular node is invariant towards. Each node learns these entities independently
of each other which allows for a more flexible invariance
model as opposed to architectures which replicate invariances across the network.
2. Prior Art
There has been considerable interest in the past, in developing methods which incorporate prior knowledge of invariances or symmetries in data. Although the applications
previously tackled were specific and relatively narrow, development of such methods offers a better understanding of
the importance of modelling symmetries in data. In these
cases, the architectures explicitly enforce such structure
through different approaches. Though in this work we focus
on deep architectures, it is interesting to note a number of
works on modifications of Markov Random Fields and Restricted Boltzman Machines to achieve rotational invariance
(Schmidt & Roth, 2012; Kivinen & Williams, 2011; Sohn
& Lee, 2012).
Incorporating known invariances using deep networks.
Convolutional architectures have also seen efforts to produce
rotation invariant representations. (Fasel & Gatica-Perez,
2006) and (Dieleman et al., 2015) rotate the input itself before feeding it into stacks of CNNs and generating rotation
invariant representations through gradual pooling or parameter sharing across orientations. (Teney & Hebert, 2016;
Wu et al., 2015; Li et al., 2017) rotate the convolution filters
instead of the input. This is followed by a pooling operation
to invoke invariance. Having the filters transform instead
alleviates the need to transform the inputs, which is more
expensive. Nonetheless, this requirement also remains considerably expensive during training. A similar approach was
explored for scale by (Xu et al., 2014). (Clark & Storkey,
2015) propose a network to play Go incorporating reflective
symmetry within the filters through weight tying. An inter-
Non-Parametric Transformation Networks
esting direction of research was explored by (Sifre & Mallat,
2013) where the filters were fixed and non-trainable. The
properties of the filter structure allowed them to be rotation,
scale and translation invariant. All these methods however
make an critical assumptions about the invariances present
in each of the applications. Although this is useful, it is very
restrictive when dealing with general applications or cases
where the transformations present are unknown and complex in nature such as general vision. (Cohen & Welling,
2016a) presented a method to incorporate invariances towards parametric groups. Recently, (Henriques & Vedaldi,
2017) proposed an interesting warped convolution based
layer which can implement more general (but parameterized) equivariance. The transformations are known apriori
and the sample grids are generated offline. Also, (Cohen &
Welling, 2016b) introduced steerable filters into the convolutional framework. They require generating the filters offline
and apriori, and hence have limited capability in learning
arbitrary and adaptive transformations. NPTNs need no
such apriori knowledge, can learn arbitrary non-parametric
transformations and finally are simpler and more elegant in
theory and in implementation.
variance through complex weight sharing. Optimization
was carried out through Topographic ICA and only carried
out layer wise for deep networks. A separate approach
towards modelling invariances was also developed where
a normalizing transformation is applied to every input independently. This approach was applied to transforming
auto encoders (Hinton et al., 2011) and Spatial Transformer
Networks (Jaderberg et al., 2015).
Learning unknown invariances from data. To address
the previous shortcoming, a few studies with deep networks
have been conducted towards learning more general transformations. Indeed, in most real world problems, nuisance
transformations present in data are unknown or too complicated to be parameterized by some function. (Anselmi
et al., 2013) proposed a theory of group invariances called
I-theory and explored its connection to general classification
problems and deep networks. Based off the core idea of
measuring moments of a group invariant distribution, multiple works had demonstrated efficacy of the ideas in more
challenging real-world problems such as face recognition,
though not in a neural network setting (Liao et al., 2013; Pal
et al., 2016; 2017).
Each TN node (single input channel and single output channel) internally consists of two operations 1) (convolution)
the convolution operation with a bank of filters and 2) (transformation pooling) a max pooling operation across the set
of the resultant convolution feature maps from the single
input channel. The pooling here is across only transformed
versions of the single input channel. This is in contrast to the
single convolution operation of the vanilla convolution node.
Note that the pooling operation is not the same superficially
to the standard pooling in a ConvNet, where the pooling is
spatial in nature. Here the pooling is not spatial but rather
across channels originating from the same input channel.
Fig. 1 illustrates the operation of single TN node with a
single input channel for a single patch. The single channel illustrated in the figure takes in a single input feature
map and convolves it with a bank of |G| filters. Here |G|
is the cardinality (or size) of the set of transformations that
the TN node is invariant towards, with G being the actual
set itself. Next, the transformation max pooling operation
simply max pools across the |G| feature values to obtain a
single TN activation value. When this node is replicated
spatially, standard convolution layers can be utilized. A TN
layer with M input channels and N output channels will
simply have M N of these TN nodes with each of the N
output channel having a TN node connected to each of the
M input channels. Fig. 2 illustrates a multi-channel TN.
Learning unknown invariances from data using deep
networks. (Gens & Domingos, 2014) introduced Symnets
which was one of the first to model general invariances in
deep networks with back propagation. They utilize kernel
based interpolation to tie weights model general symmetries.
Consistent with the connection between sample complexity
and invariance modelling, they find that Symnets perform
better with fewer samples compared to vanilla ConvNets.
Nonetheless, the approach is complicated and difficult to
scale. To the best of our knowledge, this was the only approach which learned invariances from data in a neural
network setting, albeit with a more complicated approach.
(Anselmi et al., 2017) provide sufficient conditions to enforce the learned representation to have symmetries learned
from data. (Kavukcuoglu et al., 2009) modelled local invariances using pooling over sparse coefficients of a dictionary
of basis functions. (Ngiam et al., 2010) achieved local in-
3. Transformation Networks
The Transformation Network (TN) is a feed forward network with its architecture designed to enforce invariance to
some class of transformations. At the core of the framework
is the TN node, whose structure itself enforces desirable
invariance properties. A TN network consists of multiple
such nodes stacked in layers. A TN node is analogous to
a single channel output with a single channel input in a
ConvNet. More specifically, if a vanilla ConvNet layer has
M input channels and N output channels, the TN version of
the layer would have M N TN nodes arranged in the same
fashion.
Formally, a TN node denoted by Υ acting on a 2D image
patch vectorized as x ∈ Rd can be defined as follows.
Υ(x) = max(hx, gwi)
g∈G
(1)
Here, G is formally defined as a unitary group, i.e. a finite
Non-Parametric Transformation Networks
(a) Convolution layer
(b) NPTN layer
Figure 2. Comparison between (a) a standard Convolution layer
and (b) a TN layer with |G| = 3. Each layer depicted has 2
input (shades of grey) and 2 output channels (shades of blue). The
convolution layer has therefore, 2 × 2 = 4 filters, whereas the TN
layer has 2 × 2 × 3 = 12 filters. NPTN (later introduced) has the
same structure as the TN layer. The different shades of filters in
the TN layer denote transformed versions of the same filter which
are max pooled over (support denoted by inverted curly bracket).
The + operation denotes channel addition. In our experiments, we
adjust the input/output channels of the NPTN layer to have the
same number of filters as the ConvNet baselines.
set obeying group axioms with each element being unitary.
w ∈ Rd is the weight or template instantiation of the set
of transformed weights {gw | g ∈ G}2 . Therefore, the
convolution kernel weights of a TN node are simply the
transformed versions of w as transformed by the unitary
group G. In a TN node, both the entities (w, G) are learned,
i.e. a TN node is tasked with learning both the template
instantiation w, and the set of transformations G to which
the node is to be invariant towards. This is in sharp contrast
with the vanilla convolutional node in which only the template instantiation w is learned, where G is hard coded to
be the translation group. Theoretically, the TN node has to
transform weight template w according to G to generate the
rest of the filters to be pooled over during the trasformation
pooling stage. In practice however, these are simply stored
as a set of templates or filters which only implicitly encodes
G. Gradient descent updates each filter differently over time
facilitating this encoding. Thus, during any forward pass, no
generation of transformed filters is necessary which significantly reduces computational complexity comapred to some
previous works (Teney & Hebert, 2016; Wu et al., 2015).
Invariances in a TN node. Invariance in the TN node arises
in theory due to the symmetry of the unitary group structure of the filters. The max operation simply measures the
infinite moment of an invariant distribution which leads to
invariance. We demonstrate this in the form of the following
result.
Theorem 3.1. (Invariance Property) Given vectors x, w ∈
Rd , a unitary group G and Υ(x) = maxg∈G (hx, gwi), for
any fixed g 0 ∈ G, the following is true
Υ(x) = Υ(g 0 x)
2
We denote the action of a group element g on w with the
notation gw to reduce clutter.
Proof. Consider the distribution of elements of the set
Sg0 = {hg 0 x, gwi} over all g ∈ G and for any particular g 0 ∈ G. This 1-D distribution characterizes the vector
g 0 x through the projections onto gw. Due to unitarity of G,
and that g 0 ∈ G, we have hg 0 x, gwi = hx, g 0−1 gwi. Now,
since G is a group, we have for any g 0 ∈ G, g 0−1 g ∈ G
due to the closure property. The set of elements in Sg0 contains all elements of G and hence must also contain g 0−1 g.
This implies that the action of g 0−1 on the group G results
in just a reordering of the group, leaving the distribution
unchanged. Thus, the set Sg0 is unchanged. More specifically, Sg0 = {hg 0 x, gwi} = {hx, gwi} = Se , where e is the
identity element of G. Thus, the two sets invoke the exact
same distribution, which results in their moments being the
same. This includes the infinite moment, which implies
Υ(g 0 x) = maxg∈G Sg0 = maxg∈G Se = Υ(x).
Theorem 3 shows that for any input x, the node output is
invariant to the transformation group G. This is interesting,
since one does not need to observe any transformed version of x during training which reduces sample complexity.
Invariance is invoked for any arbitrary input x during test
thereby demonstrating good generalization properties. It is
worthwhile to note that vanilla Convnets with their pooling
layers pool over perfect unitary groups since translation is a
unitary operation and pooling structure is enforced over a
finite group.
General Non-group Structure in a TN node. In practice,
a TN node does not explicitly enforce any group structure
to the set of templates or convolution filters. Although such
a structure is required for theoretically generating invariance, sufficient approximate invariance has been observed
in empirical studies on real-world data utilizing non-group
structures as found in (Pal et al., 2016; Liao et al., 2013).
In our experiments, we observe that the TN architecture
networks are able to perform better by learning invariance
towards both group structured transformations such as translation and rotation, and also towards general non-parametric
transformations.
ConvNets are a kind of Parametric Transformation Networks (PTNs). The vanilla convolution layer simply performs the convolution operation of the filters onto the image.
This by itself does not produce any invariance to any transformation. However, when followed by a spatial pooling
operation (or even a second convolution layer), the resultant
feature is explicitly invariant to only translation. This convolution and pooling operation can be modelled by the TN
node when the group G is defined to be a finite translation
group. Following the TN node operation, the translation
group G acts on the filter template w and transformed versions of it are computed on-the-fly for a ConvNet, which (for
a single patch) are then dot-producted with the input patch
to generates features. These features are then max-pooled
Non-Parametric Transformation Networks
0.5
0.8
0.78
0.45
0.76
Conv (48)
NPTN (9 5)
NPTN (12 4)
NPTN ch (16 3)
NPTN (24 2)
NPTN (48 1)
0.74
0.4
0.35
Loss
0.72
Loss
over by the second operation of the TN node resulting in
translation invariance. It is straight-forward to observe that
a generalization of the vanilla convolution node would be
where the group G is parameterized to model more complicated transformations. Networks with this more general
type of parametric TN node are called Parametric Transformation Networks (PTN). Thereby, a vanilla convolution
network (ConvNet) with its pooling operation is a kind of
PTN.
0.7
0.68
0.25
0.66
0.2
0.64
0.15
0.62
4. Non-Parametric Transformation Networks
We now introduce our main practical contribution, which
are Transformation Networks that can learn general nonparametric symmetries from data. A Non-Parametric Transformation Network (NPTN) is a network of TN nodes that
lack any parametric model for the transformation which
describes the set of templates/filters in the nodes. They are
able to explicitly learn arbitrary invariances and symmetries
from data itself better than a vanilla CNN. Compared to
other approaches to model general invariances, their architecture is a natural generalization of CNNs and is elegantly
simple. Setting |G| = 1 reduces a NPTN to a standard CNN
in theory and in practice.
NPTN layer structure. NPTNs can be implemented using
standard deep learning libraries and convolutional routines.
They can be optimized using standard gradient descent methods such as SGD. They can replace any convolution layer
in any architecture making them versatile. We describe one
way to implement a NPTN layer with M input channels, N
output channels and that models a set of transformations
with cardinality |G| (as illustrated in Fig. 2). Note that
each of the M × N channels (analogous to the input ×
output number of channels in ConvNets) models |G| filters
independently. For every input channel, |G| filters must
be learned which would encode the pre-transformed filter
set. This continuously maintained set of filters bypasses
the need to transform the filters on-the-fly during a forward
pass or whenever the weights change due to an update. The
gradients itself encode the invariance updates. Once the
input is convolved with these M × |G| filters, the M sets
each with |G| feature maps each are max pooled across.
More specifically, each |G| number of feature maps from
a single input channel results in one intermediate feature
map after max pooling (across the |G| channels). After this
operation there are M intermediate feature maps which are
transformation invariant. Now, the sum (alternatively the
mean) of these M feature maps results in one output feature
map or channel. Each of the N output channels performs
the same operation with independent of the other channels.
The total number of filters learned for a NPTN layer with
M input channels, N output channels and that models |G|
number of transformations is M N |G|. Importantly, every
0.3
0.6
150
200
250
300
Epoch
0.1
150
Conv (48)
Conv (72)
Conv (89)
Conv (104)
Conv (118)
Conv (130)
Conv (141)
NPTN (48 1)
NPTN (48 2)
NPTN (48 3)
NPTN (48 4)
NPTN (48 5)
NPTN (48 6)
NPTN (48 7)
200
250
300
Epoch
(a) Training loss for a 2 layer net-(b) Training loss for a 3 layer network
work
Figure 3. Training loss on CIFAR10. (a) 2 layered network. Each
network has the same number of parameters in the second layer
and lower number of parameters for the first layer for the NPTN
variants. (b) 3 layered network. Networks with the same color have
the same number of parameters. All non-trivial NPTN variants
(|G| > 1) have higher training error but lower test error, thereby
shortening the generalization gap.
connection from input to output of a layer independently
models invariance using separate |G| number of filters.
Non-unitary structure in a NPTN node. NPTNs do not
explicitly enforce a unitary structure. Preserving a unitary
structure would restrict the complexity of transformations
that can be modelled. Nonetheless, such a restriction might
also bring about regularization benefits which is a question
that is left for future exploration. NPTNs in this aspect are
an approximation whereas ConvNets exactly maintain the
unitary and group structure explicitly through hard coded
operations. Despite this fact, we find in our experiments
that NPTNs perform better, leading to the hypothesis that
modelling more complex transformations is more important
than exactly preserving the unitary group structure.
Relation to Maxout Networks. NPTNs deviate significantly in motivation and architecture from Maxout Networks (Goodfellow et al., 2013). Maxout Networks were
introduced as a more general activation function which also
applied max pooling across channels. However, each of
the set of channels pooled over has a support over all input channels. There is no relation to invariance modelling
of a single input feature map. NPTNs on the other hand
only max pool across channels which take in only a single
channel as input. The filters of the pooled over channels get
applied to a single input. Pooling across these responses
results in an invariant description of the input.
5. Empirical Evaluation
5.1. Implementation.
We implement NPTNs using PyTorch with its standard convolution routines. Our NPTN implementation consisted of
Non-Parametric Transformation Networks
5.2. Benchmarking against ConvNets
In our first set of experiments, we benchmark and characterize the behavior of NPTNs against the standard ConvNets
augmented with Batch Normalization (Ioffe & Szegedy,
2015). The goal of this set of experiments is to observe
whether learning non-parametric transformation invariance
from complex visual data itself helps with object classification. For this experiment, we utilize the CIFAR10 dataset3 .
The networks we experiment with are not designed to compete on this data but rather throw light into the behavior of
NPTNs. We therefore utilize shallow networks, namely a
two and three layered network for these experiments. Each
layer block of the baseline ConvNets consist of the convolution layer, followed by batch normalization and the
non-linearity (PReLU) and finally by a 2 by 2 spatial max
pooling layer. Each corresponding NPTN network replaces
only the convolution layer with the NPTN layer. Thus,
NPTN is allowed to model non-parametric invariance in
addition to the typically enforced translation invariance due
to spatial max pooling.
Two layered NPTN. Our first pilot experiment works with
a two layered network with the baseline ConvNet having
channels [48, 16] with a total of 3 × 48 + 48 × 16 = 912 filters. The NPTN variants in this experiment keep the number
of filters in the second layer constant with 48 channels with
|G| = 1 denoted by (48 1), 24 channels with |G| = 2 denoted by (24 2), and so on up until 9 channels with |G| = 5
3
With standard data augmentation of random cropping after a
4 pixel pad, and random horizontal flipping. Training was for 300
epochs with the learning rate being 0.1 and decreased at epoch
150, and 225 by a factor of 10.
ConvNet
NPTN
0.64
Test loss
four operations performed in sequence, namely convolution,
volumettric max pooling, channel reordering and finally volumetric mean pooling. For M inputs, N outputs and |G|
transformations, the first function is the standard convolution function with M inputs and M N |G| outputs with the
groups option set to M . Hence a total of M N |G| filters are
learned. The second function is the volumetric max pooling
to pool only across channels with a kernel size (|G|, 1, 1)
(where the kernel size is (channels, pool h, pool w)). The
kernel size along height and width can be increased from
1 if spatial max pooling is desired as well. Note that the
output of the convolution function will have N channels in
sequence from the same input channel replicated M times.
However, we want the alternate configuration before the
volumetric mean pooling layer. We instead require one feature map from every input channel to be together in order
(i.e. from each of the M input channels) replicated N times.
This is solved by a channel reordering operation which is
the third operation. The final operation is the volumetric
mean pooling with a kernel of size (M, 1, 1) resulting N
output channels as required.
0.62
0.6
(48, 1)
(24, 2)
(16, 3)
(12, 4)
(9, 5)
(Channel, |G|)
Figure 4. Test losses on CIFAR10 for the two layered network.
Each network listed has the same number of filters. ConvNet
denoted had 48 channels.
Method
ConvNet (48)
NPTN (48, 1)
ConvNet (72)
NPTN (48, 2)
ConvNet (89)
NPTN (48, 3)
ConvNet (104)
NPTN (48, 4)
ConvNet (118)
NPTN (48, 5)
ConvNet (130)
NPTN (48, 6)
ConvNet (141)
NPTN (48, 7)
(k = 3)
0.555
0.548
0.506
0.498
0.485
0.476
0.474
0.479
0.464
0.478
0.453
0.483
0.456
0.477
(k = 5)
0.467
0.472
0.440
0.431
0.440
0.407
0.441
0.418
0.431
0.412
0.438
0.422
0.435
0.421
(k = 7)
0.465
0.456
0.481
0.427
0.467
0.438
0.485
0.448
0.482
0.461
0.478
0.457
0.473
0.455
Table 1. Test loss on CIFAR10 for the three layered network. Each
NPTN is paired with its corresponding ConvNet baseline which
has approximately the same number of parameters.
(9 5). Fig. 3(a) and Fig. 4 show the training and testing
losses. Each network experimented with has the same number of parameters. For the two layered network we find that
the (48 1) performs slightly worse than the ConvNet baseline in training. However, all NPTN variants which learn
a non-trivial set of transformations (|G| > 1) have lower
training loss except (9 5). NPTNs (24 2), (16 3), (12 4)
perform significantly better. Fig. 3(a) shows that all NPTN
variants have much lower test losses with (16 3) performing
the best.
Three layered NPTN. For the second comparison experiment, we explore the behavior of a three layered network
with different kernel sizes. To maintain a fair comparison,
we have an approximately equal number of filters between
the corresponding baseline ConvNet (by slightly increasing
the number of channels) and the NPTN variant while keeping the number of channels constant for the NPTNs at 48
but instead increasing cardinality of the transformation set
(|G|). Thus, the transformation modelling capacity of the
Non-Parametric Transformation Networks
Rotations
ConvNet (36)
NPTN (36, 1)
NPTN (18, 2)
NPTN (12, 3)
NPTN (9, 4)
Translations
ConvNet (36)
NPTN (36, 1)
NPTN (18, 2)
NPTN (12, 3)
NPTN (9, 4)
0◦
0.022
0.021
0.020
0.019
0.018
0 pixels
0.019
0.018
0.019
0.019
0.019
30◦
0.037
0.041
0.034
0.033
0.036
4 pixels
0.030
0.029
0.023
0.022
0.023
60◦
0.066
0.064
0.053
0.055
0.057
8 pixels
0.063
0.061
0.054
0.051
0.051
90◦
0.106
0.108
0.092
0.087
0.095
12 pixels
0.218
0.224
0.190
0.189
0.195
Table 2. Test loss on progressively transformed MNIST with (a)
random rotations and (b) random pixel shifts. NPTNs can learn
invariances to arbitrary transformations from the data itself without
any apriori knowledge.
network is slowly increased. We train/test all networks on
CIFAR10 and report results in Fig. 3(b) and Table 1. We
find that a three layered NPTN has higher training loss as
compared to its ConvNet baseline but demonstrates lower
test loss. This indicates that NPTN have a structural bias
which further shortens the generalization gap. This behavior
in training is in contrast to the two layered network but in
line with the behavior during test time with decreased loss.
We find that NPTNs with roughly the same computational
complexity and parameters generalize better than ConvNets.
Note that however, we do not observe such benefits for a
kernel size of 3, in fact in many cases NPTNs perform worse.
We hypothesize that, this is because a 3 × 3 filter size is
too small an area of an activation map to exhibit meaningful structured spatial transformations that can be learned.
Better generalization is consistently observed for the larger
kernel sizes of 5 and 7 which do contain more structured
transformations due to larger receptive field sizes.
Effect of training with larger |G|. We find that performance peaks at |G| being around 3 and going higher offers
less performance gains. We believe that this is because the
forward pass of the NPTN selects a single transformation
channel (per input channel) to update due to the max operation. Every back-propagation therefore updates only a
single filter out of the |G| filters (per input channel) leading
to each filter being updated on an average by the total num1
ber of iterations times a factor of |G|
. This results in the
filters being less optimized than standard ConvNet filters
resulting in a trade off against modelling invariance. The
effect seems more pronounced for higher |G|. A standard
ConvNet does not face this problem as every filter is updated
for any given backprop iteration (|G| = 1).
5.3. Learning Transformation Invariances
Efficient learning of unknown invariances from data.
We demonstrate the ability of NPTN networks to learn invariances directly from data without any apriori knowledge.
For this experiment, we augment MNIST with a) random
rotations b) random translations, both in training and testing data thereby increasing the complexity of the learning
problem itself. For each sample, a random instantiation
of the transformation was applied. For rotation, the angular range was increased, whereas for translations it was
the pixel shift range. Table 2 presents these results. All
networks in the table are two layered and have the exact
same number of parameters. As expected, NPTNs match
the performance of vanilla ConvNets when there were no
additional transformations added (0◦ and 0 pixels)4 . However, as the transformation intensity (range) is increased,
NPTNs perform significantly better than ConvNets. Trends
consistent with previous experiments were observed with
the highest performance observed with NPTN (|G| = 3).
This highlights the main feature of NPTNs, i.e. their ability
to model arbitrary transformations observed in data without
any apriori information and without changes in architecture
whatsoever. They exhibit better performance in settings
where both rotation invariance and stronger translation invariance is required (even though ConvNets are designed
specifically to handle translations). This ability is something that previous deep architectures did not posses nor
demonstrate.
Efficacy with depth. As an exploratory study, we examine
how the relative depth of the NPTN layer assists in learning
transformation invariance on MNIST. For this, we train
three networks with number of layers being 5, 6 and 8
with 48 channels per layer expect the last layer with 16
channels. We replace layers at different depth with a single
NPTN layer (|G| = 3) starting at depth 2 through to the
last but one layer and observe the impact on performance.
We do this while keeping the number of parameters and the
computation complexity exactly the same for all networks of
a particular depth. To enforce this, the number of channels
in the ConvNet baselines were slightly increased to maintain
the same number of parameters as the NPTN versions. Fig. 5
shows the performance of these networks on CIFAR10. The
term layer ratio signifies the relative depth of the NPTN
layer replacement (layer ratio is the layer of replacement
divided by the total number of layers). We find that, as a
general trend, NPTNs offer most performance gains towards
the lower layers where it can learn more complicated low
level transformations than translations. Performance gains
decrease towards a layer ratio of 0.7. It is not clear at this
NPTNs perform slightly better than ConvNets for 0◦ rotations
because for all rotation experiments, small translations up to 2
pixels were applied only in training.
4
Non-Parametric Transformation Networks
2
Conv 5 Layers
Conv 6 Layers
Conv 8 Layers
NPTN 5 layers
NPTN 6 layers
NPTN 8 layers
0.7
0.6
0.5
Test Error %
Test Loss
0.8
1.5
1
(128)
(64, 2)
(42, 3)
(32, 4)
(Channel, |G|)
0.2
0.4
0.6
0.8
1
1.2
Layer Ratio
Figure 5. Performance of NPTN layer replacement (|G| = 3)
w.r.t the relative depth of the replacement on MNIST. Layer ratio
signifies how close to the top layer is the NPTN replacement.
point why this effect takes place. Nonetheless, the trend of
better performance at lower layer replacements is clear.
5.4. NPTNs with Capsule Networks
Capsule Networks with dynamic routing were recently introduced as an extension of standard neural networks (Sabour
et al., 2017). The main motivation behind the use of capsules (group of neurons) to represent entities is to allow
capsules to encode different properties of the object. However, since the architecture is implemented using vanilla
convolution layers, invariance properties of the networks
are limited. Our goal for this final experiment is to augment Capsule Nets with NPTNs. We do this by replacing
the convolution layers in the Primary Capsule layer of the
published architecture with NPTN layers while maintaining
the same number of parameters (by reducing number of
channels and increasing |G|). In effect, the convolutional
capsules were replaced by NPTN capsules. Our baseline is
the proposed CapsuleNet with 3 layers using a third party
implementation PyTorch5 . The number of output channels
in the first convolution layer was kept at 48. The baseline
convolution capsule layer had 128 output channels. The
NPTN variants progressively decreased the number of channels as |G| was increased. All other hyperparameters were
preserved. The networks were trained on the 2-pixel shifted
MNIST for 50 epochs with a learning rate of 10−3 . They
were tested on the test set with no shift. The performance
statistics of 5 runs are reported in Fig. 6. We find that for
roughly the same number of kernel filters (and parameters),
Capsule Nets have much to gain from the use of NPTN
layers (a significant loss decrease from 1.90 to 0.78 for 13 of
the baseline number of channels and |G| = 3). The learning
of invariances within each capsule significantly increases
efficacy and performance of the overall architecture.
5
https://github.com/dragen1860/CapsNet-Pytorch.git
Figure 6. Test errors on MNIST for Capsule Nets augmented with
NPTNs. (128) denotes a Capsule Network with a vanilla ConvNet.
Other labels are NPTNs with (channels, |G|). The number of
filters from left to right is {4224, 4160, 4074, 4128}. NPTNs
significantly outperform ConvNets in Capsule Nets with fewer
filters.
6. Discussion
It is clear that the success of ConvNets is not the whole story
towards solving perception. Studies into different aspects of
network design will prove to be paramount in addressing the
complex problem of not just visual but general perception.
The development of NPTNs offer one such design aspect,
i.e. modelling non-parametric invariances and symmetries
simultaneously from data. Through our experiments, we
found that NPTNs can indeed effectively learn general invariances without any apriori information. Further, they are
effective and improve upon vanilla ConvNets even when
applied to general vision data as presented in CIFAR10 with
complex unknown symmetries. This seems to be a critical
requirement for any system that is aimed at taking a step
towards general perception. Assuming knowledge of symmetries in real-world data (not just visual) is impractical and
succesful models would need to adapt accordingly.
In all of our experiments, NPTNs were compared to vanilla
ConvNet baselines with the same number of filters (and
thereby more channels). Interestingly, the superior performance of NPTNs with fewer channels than ConvNet
baselines indicates that better modelling of invariances is
a useful goal to pursue during design. Explicit and efficient modelling of invariances has the potential to improve
many existing architectures. In our experiments, we also
find that Capsule Networks which utilized NPTNs instead
of vanilla ConvNets performed much better. This motivates
and will justify more attention towards architectures and
other solutions that efficiently model general invariances
in deep networks. Such an endeavour might not only produce networks performing better in practice, it promises to
deepen our understanding of deep networks and perception
in general.
Non-Parametric Transformation Networks
References
Anselmi, Fabio, Leibo, Joel Z, Rosasco, Lorenzo, Mutch,
Jim, Tacchetti, Andrea, and Poggio, Tomaso. Unsupervised learning of invariant representations in hierarchical
architectures. arXiv preprint arXiv:1311.4158, 2013.
Anselmi, Fabio, Evangelopoulos, Georgios, Rosasco,
Lorenzo, and Poggio, Tomaso. Symmetry regularization.
Technical report, Center for Brains, Minds and Machines
(CBMM), 2017.
Burges, Christopher JC and Schölkopf, Bernhard. Improving the accuracy and speed of support vector machines.
In Advances in neural information processing systems,
pp. 375–381, 1997.
Chen, Yunpeng, Li, Jianan, Xiao, Huaxin, Jin, Xiaojie, Yan,
Shuicheng, and Feng, Jiashi. Dual path networks. In
Advances in Neural Information Processing Systems, pp.
4470–4478, 2017.
Clark, Christopher and Storkey, Amos. Training deep convolutional neural networks to play go. In International
Conference on Machine Learning, pp. 1766–1774, 2015.
Cohen, Taco and Welling, Max. Group equivariant convolutional networks. In International Conference on Machine
Learning, pp. 2990–2999, 2016a.
Cohen, Taco S and Welling, Max. Steerable cnns. arXiv
preprint arXiv:1612.08498, 2016b.
Dieleman, Sander, Willett, Kyle W, and Dambre, Joni.
Rotation-invariant convolutional neural networks for
galaxy morphology prediction. Monthly notices of the
royal astronomical society, 450(2):1441–1459, 2015.
Fasel, Beat and Gatica-Perez, Daniel. Rotation-invariant
neoperceptron. In Pattern Recognition, ICPR. 18th International Conference on, volume 3, pp. 336–339. IEEE,
2006.
Gens, Robert and Domingos, Pedro M. Deep symmetry
networks. In Advances in neural information processing
systems, pp. 2537–2545, 2014.
Goodfellow, Ian, Warde-Farley, David, Mirza, Mehdi,
Courville, Aaron, and Bengio, Yoshua. Maxout networks.
In International Conference on Machine Learning, pp.
1319–1327, 2013.
Hadsell, Raia, Chopra, Sumit, and LeCun, Yann. Dimensionality reduction by learning an invariant mapping. In
Computer vision and pattern recognition, 2006 IEEE computer society conference on, volume 2, pp. 1735–1742.
IEEE, 2006.
He, Kaiming, Zhang, Xiangyu, Ren, Shaoqing, and Sun,
Jian. Deep residual learning for image recognition. In
Proceedings of the IEEE conference on computer vision
and pattern recognition, pp. 770–778, 2016.
Henriques, João F and Vedaldi, Andrea. Warped convolutions: Efficient invariance to spatial transformations. In
International Conference on Machine Learning, 2017.
Hinton, Geoffrey E, Krizhevsky, Alex, and Wang, Sida D.
Transforming auto-encoders. In International Conference
on Artificial Neural Networks, pp. 44–51. Springer, 2011.
Hu, Jie, Shen, Li, and Sun, Gang. Squeeze-and-excitation
networks. arXiv preprint arXiv:1709.01507, 2017.
Huang, Gao, Liu, Zhuang, Weinberger, Kilian Q, and
van der Maaten, Laurens. Densely connected convolutional networks. arXiv preprint arXiv:1608.06993, 2016.
Ioffe, Sergey and Szegedy, Christian. Batch normalization:
Accelerating deep network training by reducing internal
covariate shift. In International conference on machine
learning, pp. 448–456, 2015.
Jaderberg, Max, Simonyan, Karen, Zisserman, Andrew, et al.
Spatial transformer networks. In Advances in Neural
Information Processing Systems, pp. 2017–2025, 2015.
Kavukcuoglu, Koray, Fergus, Rob, LeCun, Yann, et al.
Learning invariant features through topographic filter
maps. In Computer Vision and Pattern Recognition, 2009.
CVPR 2009. IEEE Conference on, pp. 1605–1612. IEEE,
2009.
Kivinen, Jyri J and Williams, Christopher KI. Transformation equivariant boltzmann machines. In International Conference on Artificial Neural Networks, pp. 1–9.
Springer, 2011.
LeCun, Yann, Bottou, Léon, Bengio, Yoshua, and Haffner,
Patrick. Gradient-based learning applied to document
recognition. Proceedings of the IEEE, 86(11):2278–2324,
1998.
Li, Junying, Yang, Zichen, Liu, Haifeng, and Cai, Deng.
Deep rotation equivariant network. arXiv preprint
arXiv:1705.08623, 2017.
Liao, Qianli, Leibo, Joel Z, and Poggio, Tomaso. Learning
invariant representations and applications to face verification. In Advances in Neural Information Processing
Systems, pp. 3057–3065, 2013.
Ngiam, Jiquan, Chen, Zhenghao, Chia, Daniel, Koh,
Pang W, Le, Quoc V, and Ng, Andrew Y. Tiled convolutional neural networks. In Advances in neural information
processing systems, pp. 1279–1287, 2010.
Non-Parametric Transformation Networks
Niyogi, Partha, Girosi, Federico, and Poggio, Tomaso. Incorporating prior information in machine learning by creating virtual examples. Proceedings of the IEEE, 86(11):
2196–2209, 1998.
Pal, Dipan, Kannan, Ashwin, Arakalgud, Gautam, and Savvides, Marios. Max-margin invariant features from transformed unlabelled data. In Advances in Neural Information Processing Systems 30, pp. 1438–1446, 2017.
Pal, Dipan K, Juefei-Xu, Felix, and Savvides, Marios. Discriminative invariant kernel features: a bells-and-whistlesfree approach to unsupervised face recognition and pose
estimation. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, pp. 5590–
5599, 2016.
Sabour, Sara, Frosst, Nicholas, and Hinton, Geoffrey E. Dynamic routing between capsules. In Advances in Neural
Information Processing Systems, pp. 3859–3869, 2017.
Schmidt, Uwe and Roth, Stefan. Learning rotation-aware
features: From invariant priors to equivariant descriptors.
In Computer Vision and Pattern Recognition (CVPR),
2012 IEEE Conference on, pp. 2050–2057. IEEE, 2012.
Schroff, Florian, Kalenichenko, Dmitry, and Philbin, James.
Facenet: A unified embedding for face recognition and
clustering. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pp. 815–823,
2015.
Sifre, Laurent and Mallat, Stéphane. Rotation, scaling and
deformation invariant scattering for texture discrimination. In Proceedings of the IEEE conference on computer
vision and pattern recognition, pp. 1233–1240, 2013.
Sohn, Kihyuk and Lee, Honglak. Learning invariant representations with local transformations. arXiv preprint
arXiv:1206.6418, 2012.
Teney, Damien and Hebert, Martial. Learning to extract
motion from videos in convolutional neural networks.
In Asian Conference on Computer Vision, pp. 412–428.
Springer, 2016.
Wu, Fa, Hu, Peijun, and Kong, Dexing. Flip-rotatepooling convolution and split dropout on convolution
neural networks for image classification. arXiv preprint
arXiv:1507.08754, 2015.
Xu, Yichong, Xiao, Tianjun, Zhang, Jiaxing, Yang, Kuiyuan,
and Zhang, Zheng. Scale-invariant convolutional neural
networks. arXiv preprint arXiv:1411.6369, 2014.
Zagoruyko, Sergey and Komodakis, Nikos. Wide residual
networks. arXiv preprint arXiv:1605.07146, 2016.
| 2cs.AI
|
A distributed primal-dual algorithm for computation of generalized Nash equilibria with
shared affine coupling constraints via operator splitting methods
Peng Yi, Lacra Pavela,
a Department
of Electrical and Computer Engineering, University of Toronto
arXiv:1703.05388v1 [math.OC] 15 Mar 2017
Abstract
In this paper, we propose a distributed primal-dual algorithm for computation of a generalized Nash equilibrium (GNE) in
noncooperative games over network systems. In the considered game, not only each player’s local objective function depends
on other players’ decisions, but also the feasible decision sets of all the players are coupled together with a globally shared affine
inequality constraint. Adopting the variational GNE, that is the solution of a variational inequality, as a refinement of GNE, we
introduce a primal-dual algorithm that players can use to seek it in a distributed manner. Each player only needs to know its local
objective function, local feasible set, and a local block of the affine constraint. Meanwhile, each player only needs to observe the
decisions on which its local objective function explicitly depends through the interference graph and share information related to
multipliers with its neighbors through a multiplier graph. Through a primal-dual analysis and an augmentation of variables, we
reformulate the problem as finding the zeros of a sum of monotone operators. Our distributed primal-dual algorithm is based on
forward-backward operator splitting methods. We prove its convergence to the variational GNE for fixed step-sizes under some
mild assumptions. Then a distributed algorithm with inertia is also introduced and analyzed for variational GNE seeking. Finally,
numerical simulations for network Cournot competition are given to illustrate the algorithm efficiency and performance.
Keywords: Network system; generalized Nash equilibrium; multi-agent systems; distributed algorithm; operator splitting;
1. Introduction
Engineering network systems, like power grids,
communication networks, transportation networks and
sensor networks, play a foundation role in modern society.
The efficient and secure operation of various network systems
relies on efficiently solving decision and control problems
arising in those large scale network systems. In many decision
problems, the nodes can be regarded as agents that need to
make local decisions possibly limited by the shared network
resources within local feasible sets. Meanwhile, each agent
has a local cost/utility function to be optimized, which depends
on the decisions of other agents. The traditional manner
for solving such decision problems over networks is the
centralized optimization approach, which relies on a control
center to gather the data of the problem and to optimize the
social welfare (usually taking the form of the sum of local
objective functions) within the local and global constraints.
The centralized optimization approach may not be suitable for
decision problems over large scale networks, since it needs
bidirectional communication between all the network nodes
and the control center, it is not robust to the failure of the
center node, and the computational burden for the center is
unbearable. It is also not preferable because the privacy of
each agent might be compromised when the data is transferred
I This
work was supported by NSERC Discovery Grant (261764).
Email address: [email protected],
[email protected] (Peng Yi, Lacra Pavel)
Preprint submitted to Elsevier
to the center. Recently, a distributed optimization approach is
proposed as an alternative methodology for solving decision
problems in network systems (Yi, Hong & Liu (2016), Shi,
Ling, Wu and Yin (2015) and Zeng, Yi, Hong & Xie (2016)).
In the distributed optimization approach, the data is distributed
throughout the network nodes and there is no control center,
and each agent in the network can just utilize its local data
and share information with its neighbour agents to compute its
local decision that corresponds to the optimal solution of the
social welfare optimization problem. Therefore, the distributed
optimization approach overcomes the drawbacks of the
centralized optimization approach by decomposing the data,
computation and communication to each agent. Moreover,
each agent has the authority and autonomy to formulate its own
objective function without worrying about privacy leaking out.
However, both approaches adopt the same solution concept,
that is the optimal social welfare solution with local and global
constraints, as the solution criterion of decision problems in
network systems.
However, optimal solutions of social welfare may not be
proper solution concepts in many applications. In fact, with the
deregulation and liberalization of markets, there is no guarantee
that the agents will not deliberately deviate from their local
optimal solutions to increase (decrease) own local utility (cost),
possibly by deceiving to utilize more network resources. In
this paper, we consider the game theoretic approach where
each agent in the network has its own local autonomy and
rationality. In such a setup of multiple interacting rational
March 17, 2017
players making decisions in a noncooperative environment,
Nash Equilibrium (NE) is a more reasonable solution. In
an NE, no player can increase (decrease) its local utility
(cost) by unilaterally changing its local decision, therefore,
no agent has the incentive to deviate from it. In other
words, NE is a self-enforceable solution in the sense that
once NE is computed all the agents will execute that NE.
Recently, there has been increasing interest in adopting game
theory and NE as the modeling framework and solution
concept for various network decision problems, like wireless
communication systems (Scutari,Palomar, Facchinei, and Pang,
(2010) and Menache & Ozdaglar (2011)), network flow
control (Alpcan & Basar (2005)), optical networks (Pan &
Pavel (2009)) and smart grids (Ye & Hu (2016)).
Moreover, in engineering network systems, not only the
local objective function of each agent depends on other
players’ decisions, but also the feasible set of each local
decision could depend on other agents’ decisions, because
the agents may compete for the utilization of some shared
or common network resources like bandwidth, spectrum or
power. This type of network decision problems can be
modeled as noncooperative games with coupling constraints,
and generalized Nash equilibrium (GNE) can be adopted as its
solution. The study of GNE dates back to the social equilibrium
concept proposed by Debreu (1952), and flourished in the
last two decades with applications in practical problems like
environment pollution games (Krawczyk & Uryasev (2000)),
power market design (Contreras, Klusch, Krawczyk (2004)),
optical networks (Pavel (2007)), wireless communication
(Facchinei & Pang (2010)). Interested readers can refer to
Facchinei & Kanzow (2010) for a historical review of GNE,
and refer to Fischer, Herrich & Schonefeld (2014) for recent
developments, and to Facchinei & Pang (2010) for a technical
treatment.
Even though NE or GNE is a reasonable and expectable
solution as a result of multiple rational agents making decisions
in a noncooperative manner, how to arrive at an NE (GNE) is
by no way a self-evident task. Each player needs to know the
complete game information, including objective functions and
feasible decision sets of all the other agents, in order to compute
NE in an introspective manner. It gets much more complicated
for computing GNE, because the agents also have to consider
the coupling in the feasible decision sets. In fact, as of yet there
is no universal manner to efficiently compute GNE in games
with coupling constraints (Harker (1991) and Fischer, Herrich
& Schonefeld (2014)), except for games with shared coupling
constraints (Facchinei, Fischer & Piccialli (2007)). Moreover,
for games in large scale network systems, it is quite unrealistic
and undesirable to assume that each agent could have the
complete information of the whole network, because this
implies prohibitive communication and computation burden
and no privacy protection. Therefore, each player (agent)
should compute its local decision corresponding with an NE
or GNE in a distributed manner, somehow resembling the
distributed optimization approach. In other words, each agent
should only utilize its local objective function, local feasible
set and possible local data related to coupling constraints, and
should only share information with its neighbouring agents to
compute its local decision in the NE (GNE). This turns out to be
an emerging research topic and gets studied in Salehisadaghiani
& Pavel (2016a), Koshal,Nedić & Shanbhag (2016), Parise,
Gentile, Grammatico & Lygeros (2015), Ye & Hu (2016) and
Swenson, Kar & Xavier (2015), etc.
Motivated by the above, in this work we consider a
distributed algorithm for iterative computation of GNE in
noncooperative games with shared affine coupling constraints
over network systems. The considered noncooperative game
has each agent’s local objective function depending on other
agents’ decisions as specified by an interference graph, and
has also an affine constraint shared by all agents, coupling all
players’ feasible decision sets. The considered game model
covers many practical problems, like the power market model
in Contreras, Klusch, Krawczyk (2004), environment pollution
game in Krawczyk & Uryasev (2000), power allocation
game in communication systems in Yin, Shanbhag & Mehta
(2011). A (centralized) numerical algorithm was recently
studied in Schiro, Pang & Shanbhag (2013) for quadratic
objective functions and in Dreves & Sudermann (2016) for
linear objective functions. Generally speaking, the GNE of the
considered game may not be unique. In this work, we adopt
the variational GNE, that corresponds with the solution of a
variational inequality proposed in Facchinei, Fischer & Piccialli
(2007), to be a refinement GNE solution. The variational GNE
is a particular type of the normalized equilibrium proposed in
Rosen (1965), and enjoys a nice economical interpretation that
all the agents have the same shadow price for shared network
resources without any discrimination as pointed in Kulkarni &
Shanbhag (2012). Furthermore, the variational GNE enjoys
a sensitivity and stability property (Facchinei & Pang (2010)
and Facchinei & Kanzow (2010)), hence we adopt it as the
desirable solution.
We propose a new type of distributed algorithm that agents
can use to compute the variational GNE by only manipulating
their local data and communicating with neighbouring agents.
Observing that the KKT condition of the corresponding
variational inequality requires all agents to reach consensus
on the multiplier of the shared affine constraint, we introduce
a local copy of the multiplier and an auxiliary variable for
each player. To enforce the consensus of local multipliers,
we use a reformulation that incorporates the Laplacian matrix
of a connected graph. Motivated by the forward-backward
operator splitting method for finding zeros of a sum of
monotone operators (refer to Bauschke & Combettes (2011))
and the recent primal-dual algorithm proposed in Condat
(2013) for optimization problems with linear composition
terms, we propose a novel distributed algorithm for iterative
computation of GNE. The main idea is to introduce a suitable
metric matrix and to split the equivalent reformulation into
two monotone operators. An operator splitting method has
been adopted for NE computation in a centralized manner in
Briceno-Arias & Combettes (2013). A different splitting idea
is adopted here appropriate for distributed GNE computation.
Moreover, a distributed algorithm with inertia is also proposed
and investigated, motivated by the acceleration algorithms in
2
Alvarez & Attouch (2001), Attouch, Chbani, Peypouquet, and
Redont (2016), Iutzeler & Hendrickx (2016) and Lorenz &
Pock (2015), most of which only focused on optimization
problems. The convergence of the proposed algorithms is
verified under suitable fixed step-size choice and some mild
assumptions on the objective functions and communication
graphs.
The recent works of Zhu & Frazzoli (2016), Yu, van
der Schaar & Sayed (2016), Liang, Yi and Hong (2016)
and Paccagnan, Gentile, Parise, Kamgarpour & Lygeros
(2016) are closely related with this work since all of them
are concerned with the distributed algorithm for seeking GNE
of noncooperative games with coupling constraints. Zhu
& Frazzoli (2016) address the GNE seeking for the case
where each player has non-shared local coupling constraints.
Assuming that each player can observe other players’ decisions
on which its local objective function and local constraint
functions depend through the interference graph, Zhu &
Frazzoli (2016) propose a distributed primal-dual GNE seeking
algorithm based on variational inequality methods, and show
algorithm convergence under diminishing step-sizes. Yu,
van der Schaar & Sayed (2016) investigate the distributed
GNE seeking under stochastic data observations. The authors
assume that the coupling constraints have a locally shared
property that if one player has its one of local constraints
dependent on the decision of another player, then this constraint
must be shared between those two players. Their algorithm
design is based on a penalty-type gradient method. Under
the assumption that each player can observe the decisions
on which its local objective function and constraint functions
depend, Yu, van der Schaar & Sayed (2016) utilize a gradient
type algorithm to seek the pure NE of the game derived by
penalizing the coupling and local constraints. They show that
their algorithm can reach a region near the pure penalized NE
with a constant step-size, which will approach a GNE if the
penalizing parameter goes to infinity. Both Liang, Yi and
Hong (2016) and Paccagnan, Gentile, Parise, Kamgarpour &
Lygeros (2016) consider the distributed algorithm for seeking a
variational GNE of the aggregative game with globally shared
affine coupling constraints. This represents a particular type of
game where the players’ local objective functions depend on
some aggregative variables of all agents’ decisions. Liang, Yi
and Hong (2016) assume that each player has local copies of
both the aggregative variables and the multipliers, and combine
a finite-time convergent continuous-time consensus dynamics
and a projected gradient flow to derive their distributed
GNE seeking dynamics. Meanwhile, Paccagnan, Gentile,
Parise, Kamgarpour & Lygeros (2016) adopt the asymmetric
projection algorithm for variational inequalities to design their
variational GNE seeking algorithm. However Paccagnan,
Gentile, Parise, Kamgarpour & Lygeros (2016) assume that
there is an additional central node for the update of the common
multiplier, and only address quadratic objective functions.
Compared with these works, our paper has following
contributions,
(i): The considered noncooperative game model is
completely general, thus a generalization of the aggregative
game in Liang, Yi and Hong (2016) and Paccagnan, Gentile,
Parise, Kamgarpour & Lygeros (2016). We further assume
that the shared affine coupling constraint is also decomposed
such that each player only knows its local contribution to the
global constraint, that is only a sub-block matrix of the whole
constraint matrix. In this sense, no player knows exactly the
shared constraints, hence, our problem model is also different
from the ones in Yu, van der Schaar & Sayed (2016) and
Zhu & Frazzoli (2016). The decomposition of the coupling
constraints, together with the localization of player’s local
objective function and local feasibility set, is quite appealing for
iterative computation of GNE in large-scale network systems
because this reduces the data transmission and computation
burden, and protects the players’ privacies.
(ii):The proposed distributed algorithms can compute the
variational GNE iteratively under a more localized data
structure and information observing structure compared to
previous ones. Firstly, each player only utilizes the local
objective function and local feasible set, and its local sub-block
matrix of the affine constraints. Secondly, we assume the
players have two (different) information observing graphs, i.e.,
interference graph and multiplier graph. Each player only
needs to observe the decisions that its local objective function
directly depends on through the interference graph. This type
of information observation assumption has also been adopted in
Yu, van der Schaar & Sayed (2016) and Zhu & Frazzoli (2016).
Meanwhile, each player only needs to share information related
to multipliers with its neighbouring agents through another
multiplier graph. Here it is not required that each player
should know the decisions that coupling constraints depend
on, as assumed in Yu, van der Schaar & Sayed (2016) and
Zhu & Frazzoli (2016). Therefore, our information sharing
(observing) structure is more localized and sparse.
(iii): The algorithm development and convergence analysis
is motivated by the operator splitting method (Bauschke &
Combettes (2011)), different from the penalized method
adopted in Yu, van der Schaar & Sayed (2016) and the
variational inequality approach in Liang, Yi and Hong (2016),
Zhu & Frazzoli (2016) and Paccagnan, Gentile, Parise,
Kamgarpour & Lygeros (2016). Based on this operator
splitting approach, we prove the algorithm converges to the
variational GNE under fixed step-sizes. Note that neither
convergence nor non-bias estimation is achieved in Yu, van
der Schaar & Sayed (2016) under fixed step-sizes, while
Zhu & Frazzoli (2016) achieve convergence with diminishing
step-sizes. On the other hand, compared with Briceno-Arias &
Combettes (2013), this paper addresses the GNE seeking under
coupling constraints, adopts a different splitting technique,
and achieves fully distributed computations. The operator
splitting method is powerful and provides additional insights.
Moreover, a distributed algorithm with inertia is proposed
and analyzed, resembling the acceleration algorithms in
optimization (Nesterov (2013) and Iutzeler & Hendrickx
(2016)). The algorithm performance is illustrated via numerical
experiments of network Cournot competitions with bounded
market capacities.
The paper is organized as follows. Section 2 gives the
3
domRA = Rm if A is maximally monotone 1 . For a proper lower
semi-continuous convex (l.s.c.) function f , its sub-differential
m
is a set-valued operator ∂ f : dom f → 2R and
notations and preliminary background. Section 3 formulates
the noncooperative game and gives the distributed algorithm
for iterative computation of a GNE. Section 4 shows how the
operator splitting method motivates the algorithm development,
and Section 5 presents the algorithm convergence analysis.
Then a distributed GNE seeking algorithm with inertia is
proposed and analyzed in Section 6. Finally, a network Cournot
competition with bounded market capacities is formulated with
numerical studies in Section 7, while concluding remarks are
given in Section 8.
∂ f : x 7→ {g ∈ Rm | f (y) ≥ f (x) + hg, y − xi, ∀y ∈ dom f }. (1)
∂ f is a maximally monotone operator 2 . Then Prox f = R∂ f :
Rm → dom f is called the proximal operator of f 3 , i.e.
1
Prox f : x 7→ arg min f (u) + ||u − x||22 .
(2)
u∈dom f
2
0, x ∈ Ω;
Define the indicator function of set Ω as ιΩ (x) =
∞, x < Ω.
For a closed convex set Ω, ιΩ is a proper l.s.c. function. ∂ιΩ is
just the normal cone operator of set Ω, that is ∂ιΩ (x) = NΩ (x) 4
and
∅
x<Ω
(3)
NΩ (x) =
{v|hv,
y
−
xi
≤
0,
∀y
∈
Ω}
x
∈ bd(Ω)
0
x ∈ int(Ω)
2. Notations and preliminary background
In this section, we review the notations and some preliminary
notions in monotone operators and graph theory.
Notations: In the following, Rm (Rm
+ ) denotes the
m−dimesional (nonnegative) Euclidean space. For a column
vector x ∈ Rm (matrix A ∈ Rm×n ), xT (AT ) denotes its transpose.
√
xT y = hx, yi denotes the inner product of x, y, and ||x||2 = xT x
denotes the norm induced by inner product h·, ·i. Given a
symmetric positive definite matrix G, denote the G-induced
inner product hx, yiG = hGx,
√ yi. The G-matrix induced norm,
|| · ||G , is defined as ||x||G = hGx, xi. Denote by || · || any matrix
induced norm in the Euclidean space. Denote 1m = (1, ..., 1)T ∈
Rm and 0m = (0, ..., 0)T ∈ Rm . For column vectors x, y, x ≥ (>)y
is understood componentwise. diag{A1 , ..., AN } represents the
block diagonal matrix with matrices A1 , ..., AN on its main
diagonal. Null(A) and Range(A) denote the null space and
range space of matrix A, respectively. Denote col(x1 , ...., xN )
as the column vector stacked with column vectors x1 , ..., xN . In
denotes the identity matrix in Rn×n . For a matrix A = [ai j ], ai j or
[A]i j stands for the matrix entry in the ith row and jth column
of A. We also use [x]k to denote the k−th element in column
Q
vector x. Denote ×i=1,...,n Ωi or ni=1 Ωi as the Cartesian product
of the sets Ωi , i = 1, ..., n. Denote int(Ω) as the interior of Ω, and
bd(Ω) as the boundary set of Ω. Define the projection of x onto
a set Ω by PΩ (x) = arg miny∈Ω ||x − y||2 . A set Ω is a convex set
if λx + (1 − λ)y ∈ Ω, ∀λ ∈ [0, 1], ∀x, y ∈ Ω. An extended
value proper function f : Rm → R is a convex function if
f (λx+(1−λ)y) ≤ λ f (x)+(1−λ) f (y), ∀λ ∈ [0, 1], ∀x, y ∈ dom f .
In this case, we also have 5
ProxιΩ (x) = RNΩ (x) = PΩ (x).
(4)
For a single-valued operator T : Rm → Rm , a point x ∈ Rm
is a fixed point of T if T x = x, and the set of fixed points of
T is denoted as FixT . The composition of operators A and
B, denoted by A ◦ B, is defined via its graph graA ◦ B =
{(x, z)|∃y ∈ ranB, (x, y) ∈ graB, (y, z) ∈ graA}. We also use AB
to denote the composition A ◦ B when they are single-valued.
Similarly, their sum A + B is defined as gra(A + B) = {(x, y +
z)|(x, y) ∈ graA, (x, z) ∈ graB}. Suppose operators A and B
are maximally monotone and 0 ∈ int(domA − domB), then
A + B is also maximally monotone6 . Further suppose that
A is single-valued, then zer(A + B) = FixRB ◦ (Id − A)7 ,
which helps to formulate the basic forward-backward operator
splitting algorithm for finding zeros of a sum of monotone
operators.
2.2. Graph theory
The following concepts are reviewed from Mesbahi &
Egerstedt (2010). The information sharing or exchanging
among the agents is described by graph G = (N, E). N =
{1, · · · , N} is the set of agents, and the edge set E ⊂ N × N
contains all the information interactions. If agent i can get
information from agent j, then ( j, i) ∈ E and agent j belongs
to agent i’s neighbor set Ni = { j|( j, i) ∈ E}, and i < Ni . G is
said to be undirected when (i, j) ∈ E if and only if ( j, i) ∈ E. A
2.1. Monotone operators
The following concepts are reviewed from Bauschke &
m
Combettes (2011). Let A : Rm → 2R be a set-valued operator.
Denote Id as the identity operator, i.e, Id(x) = x. The domain
of A is domA = {x ∈ Rm |Ax , ∅} where ∅ stands for the empty
set, and the range of A is ranA = {y ∈ Rm |∃x ∈ Rm , y ∈ Ax}.
The graph of A is graA = {(x, u) ∈ Rm × Rm |u ∈ Ax}, then
the inverse of A is defined through its graph as graA−1 =
{(u, x) ∈ Rm × Rm |(x, u) ∈ graA}. The zero set of operator
A is zerA = {x ∈ Rm |0 ∈ Ax}. Define the resolvent of operator
A as RA = (Id + A)−1 . An operator A is called monotone if
∀(x, u), ∀(y, v) ∈ graA, we have hx − y, u − vi ≥ 0. Moreover, it
is maximally monotone if graA is not strictly contained in the
graph of any other monotone operator. RA is single-valued and
1
Proposition 23.7 in Bauschke & Combettes (2011)
Theorem 20.40 in Bauschke & Combettes (2011)
3 Proposition 16.34 in Bauschke & Combettes (2011)
4 Example 16.12 of Bauschke & Combettes (2011)
5 Example 23.4 of Bauschke & Combettes (2011)
6 Corollary 24.4 in Bauschke & Combettes (2011)
7 Proposition 25.1 in Bauschke & Combettes (2011)
2
4
path of graph G is a sequence of distinct agents in N such that
any consecutive agents in the sequence correspond to an edge
of graph G. Agent j is said to be connected to agent i if there is
a path from j to i. G is said to be connected if any two agents
are connected.
Define the weighted adjacency matrix W = [wi j ] ∈ RN×N of
G with wi j > 0 if j ∈ Ni and wi j = 0 otherwise. Assume W =
W T for undirected graphs. Define the weighted degree matrix
P
P
Deg = diag{d1 , · · · , dN } = diag{ Nj=1 w1 j , ..., Nj=1 wN j }. Then
the weighted Laplacian of graph G is L = Deg−W. When graph
G is a connected and undirected graph, 0 is a simple eigenvalue
of Laplacian L with the eigenspace {α1N |α ∈ R}, and L1N = 0N ,
1TN L = 0TN , while all other eigenvalues are positive. Denote the
eigenvalues of L in an ascending order as 0 < s2 ≤ · · · ≤ sN ,
then by Courant-Fischer Theorem,
min xT Lx = s2 ||x||22 ,
x,0,
1T x=0
max xT Lx = sN ||x||22 .
x,0
by globally shared affine constraints. Denote
X ⊂ Rm :=
where Ωi ⊂ R is a private feasible decision set of player i, and
A = [A1 , · · · , AN ] ∈ Rm×n with Ai ∈ Rm×ni , and b ∈ Rm . Denote
QN
Ω = i=1
Ωi . Given the globally shared set X (which may not
be known by any agents), the following set-valued map gives
the feasible decision set map of agent i: Xi (x−i ) := {xi ∈ Rni :
(xi , x−i ) ∈ X}, or in other words:
X
A j x j }.
Xi (x−i ) := {xi ∈ Ωi |Ai xi ≥ b −
j,i, j∈N
Hence, each agent has a local feasible constraint xi ∈ Ωi ,
and there exists a coupling constraint shared by all agents with
sub-matrix Ai characterizing how agent i is involved in the
coupling constraint (shares the global resource). Notice that
agent i may only know its local Ai , in which case the globally
shared affine constraint couples the agents’ feasible decision
sets, but is not known by any agents.
(5)
Remark 3.1. We consider affine coupling constraints for
various reasons. Even though not as general as the nonlinear
constraints considered in Pavel (2007) and Zhu & Frazzoli
(2016), this setup does enjoy quite strong modeling flexibility.
As pointed out in page 191 of Facchinei & Kanzow (2010), “
However, it should be noted that the jointly convex assumption
on the constraints · · · practically is likely to be satisfied only
when the joint constraints gµ = g, µ = 1, ..., N are linear, i.e. of
the form Ax ≤ b for some suitable matrix A and vector b.” In
fact, many existing generalized Nash game models adopt affine
coupling constraints, as well documented in Schiro, Pang &
Shanbhag (2013) and Dreves & Sudermann (2016).
(6)
3.1. Game formulation
Consider a group of agents (players) N = {1, · · · , N}
that seek the generalized Nash equilibrium (GNE) of a
noncooperative game with coupling constraints defined as
follows. Each player i ∈ N controls its local decision
(strategy or action) xi ∈ Rni . Denote x = col(x1 , · · · , xN ) ∈
Rn as the decision profile, i.e., the stacked vector of all
PN
ni = n. Denote x−i =
the agents’ decisions where i=1
col(x1 , · · · , xi−1 , xi+1 , · · · , xN ) as the decision profile stacked
vector of the agents’ decisions except player i. Agent i aims to
optimize its local objective function within its feasible decision
set. The local objective function for agent i is fi (xi , x−i ) : Rn →
R. Notice that the local objective function of agent i is coupled
with other players’ decisions (however, may not be explicitly
coupled with all other players’ decisions). Moreover, the
feasible decision set of player i also depends on the decisions
n
of the other players with Xi (x−i ) : Rn−ni → 2R i denoting a
set-valued map that maps x−i to the feasible decision set of
agent i. The aim of agent i is to find the best-response strategy
set given the other players’ decision x−i ,
xi ∈R
i
s.t., xi ∈ Xi (x−i ).
Assumption 1. For player i, fi (xi , x−i ) is a differentiable
convex function with respect to xi given any fixed x−i , and Ωi
is a closed convex set. X has nonempty interior point (Slater’s
condition).
Suppose x∗ is a GNE of game (7), then for agent i, xi∗ is the
optimal solution to the following convex optimization problem:
X
min fi (xi , x∗−i ), s.t. xi ∈ Ωi , Ai xi ≥ b −
A j x∗j .
(9)
xi
∀i ∈ N.
j,i, j∈N
Define a local Lagrangian function for agent i with multiplier
λi ∈ Rm
+ as
Li (xi , λi ; x−i ) = fi (xi , x−i ) + λTi (b − Ax).
(7)
λ∗i
xi∗
(10)
Rm
+
When is an optimal solution to (9), there exists ∈
such
that the following optimality conditions (KKT) are satisfied:
The GNE x∗ = col(x1∗ , · · · , x∗N ) of the game in (7) is obtained
at the intersection of all the players’ best-response sets, and is
defined as:
xi∗ ∈ arg min fi (xi , x∗−i ), s.t. xi ∈ Xi (x∗−i ),
\
{x ∈ Rn |Ax ≥ b},
i=1
3. Problem formulation and distributed algorithm
minn fi (xi , x−i )
Ωi
ni
Denote d∗ = max{d1 , · · · , dN } as the maximal weighted degree
of graph G, then we have the following estimation,
d∗ ≤ sN ≤ 2d∗ .
N
Y
0 ∈ ∇ xi Li (xi∗ , λ∗i ; x∗−i ) + NΩi (xi∗ ), xi∗ ∈ Ωi
hλ∗i , b − Ax∗ i = 0, b − Ax∗ ≤ 0, λ∗i ≥ 0,
(11)
These can be equivalently written in the following form by
using (10) and the definition of the normal cone operator in (3)
(8)
0 ∈ ∇ xi fi (xi∗ , x∗−i ) − ATi λ∗i + NΩi (xi∗ )
0 ∈ (Ax∗ − b) + NRm+ (λ∗i )
Here we consider the GNE seeking in noncooperative games
where the couplings between players’ feasible sets are specified
5
(12)
In fact, since NΩ (x) = ∅ when x < Ω, it must hold that xi∗ ∈
Ωi and λ∗i ∈ Rm
=
+ when (12) is satisfied. Furthermore, NRm
+
Qm
∗
∗
N
.
If
[λ
]
=
0,
then
N
([λ
]
)
=
−R
,
and
hence
R+
+
j=1 R+
i k
i k
[b − Ax∗ ]k ≤ 0; and if [λ∗i ]k > 0, we have NR+ ([λ∗i ]k ) = 0, hence
[b − Ax∗ ]k = 0. Therefore, b − Ax∗ ≤ 0 and hλ∗i , b − Ax∗ i = 0.
Denote λ̄ = col(λ1 , · · · , λN ). Therefore, by Theorem 4.6 in
Facchinei & Kanzow (2010) when (x∗ , λ̄∗ ) satisfies KKT (12)
for i = 1, ..., N, x∗ is a GNE of the game in (7)
According to the above discussions, given x∗ as a GNE of
game in (7), its corresponding Lagrangian multipliers for the
globally shared affine coupling constraint may be different for
the agents, i.e., λ∗1 , λ∗2 ,, · · · , , λ∗N . In this work, we aim
to seek a GNE with the same Lagrangian multiplier for all the
agents, which has a nice interpretation from the viewpoint of
variational inequality.
Define
F(x) = col(∇ x1 f1 (x1 , x−1 ), · · · , ∇ xN fN (xN , x−N )),
the coupling constraint is met and a variational GNE of the
game is found.
n
m
Define two operators A and B, both from Ω × Rm
+ to R × R
as follows,
!
!
x
F(x)
A:
7→
λ !
−b
!
!
!
(17)
0 −AT
x
x
NΩ (x)
+
B:
7→
A
0
λ
λ
NRm+ (λ))
By the definition of F(x) in (13), the KKT conditions (16)
can be equivalently written as 0 ∈ (A + B)col(x∗ , λ∗ ). Notice
that B is a maximally monotone operator (similar arguments
for this can be found in Lemma 5.4). Hence, if F(x) has some
additional properties, then solving V I(F, X) can be converted to
the problem of finding zeros of a sum of monotone operators.
Assumption 2. F(x) defined in (13) is strongly monotone with
parameter η over Ω: hF(x) − F(y), x − yi ≥ η||x − y||22 , ∀x, y ∈ Ω,
and θ− Lipschitz continuous over Ω: ||F(x) − F(y)||2 ≤ θ||x −
y||2 , ∀x, y ∈ Ω.
(13)
which is usually called the pseudo-gradient. The variational
inequality (VI) approach to find a GNE of game (7) is to find
the solution of the following V I(F, X):
Find x∗ ∈ X, hF(x∗ ), x − x∗ i ≥ 0, ∀x ∈ X.
Remark 3.3. Assumption 2 has also been adopted in Yu, van
der Schaar & Sayed (2016), Paccagnan, Gentile, Parise,
Kamgarpour & Lygeros (2016) and Zhu & Frazzoli (2016).
Assumption 2 guarantees that there exists a unique solution
to V I(F, X) in (14) 9 , thus guarantees the existence of a GNE
for game in (7). However, the GNE of (7) may not be unique.
The algorithm for computing all the GNE of noncooperative
games with coupling constraints is still an opening research
topic, and interested readers can refer to Nabetani, Tseng, and
Fukushima (2011). This work aims to provide a distributed
algorithm for iterative computation of a variational GNE of the
considered game, which enjoys nice economical interpretations
and stability properties.
(14)
Let us check the KKT condition for V I(F, X) in (14). In fact, x∗
is a solution to V I(F, X) in (14) if and only if x∗ is the optimal
solution to the following optimization problem:
minn hF(x∗ ), xi,
x∈R
s.t.,
x ∈ Ω, Ax ≥ b
(15)
According to the optimization formulation of V I(F, X) in (15),
if x∗ solves V I(F, X), there exists λ∗ ∈ Rm such that the
following optimality conditions (KKT) are satisfied:
0 ∈ ∇ xi fi (xi∗ , x∗−i ) − ATi λ∗ + NΩi (xi∗ ), i = 1, · · · , N
0 ∈ (Ax∗ − b) + NRm+ (λ∗ )
(16)
3.2. Distributed algorithm
By comparing the two sets of KKT conditions in (12) and
(16) we obtain,
In practice, each player only knows its private information
in game (7), especially when the players interact over large
scale networks. It is quite natural that each player can only
know its local objective function fi (xi , x−i ) and local feasible set
Ωi , which cannot be shared with other players, because those
data contain its local private information, such as cost function,
preference and action ability. Moreover, matrix Ai specifies
how player i participates in the resource allocation or market
behavior, hence also contains private information, and b can be
PN
decomposed as b = i=1
bi . Thus matrix Ai can be regarded as a
map from decision space Rni to resource space Rm , while vector
bi can be regarded as a local contribution or observation for
the global resource. For example, if there are total m markets,
and each player i produces a kind of product with amount of
T
xi ∈ R+ , then Ai ∈ Rm
+ satisfying 1m Ai = 1 and 0 ≤ Ai ≤ 1
just specifies how each player allocates its production to each
market. In this case, if b̃i ∈ Rm
+ is a local observation of market
Theorem 3.2. 8 Suppose Assumption 1 holds. Every solution
x∗ of V I(F, X) in (14) is a GNE of game in (7). Furthermore,
if x∗ together with λ∗ satisfies the KKT conditions for the
V I(F, X), i.e., (16), then x∗ together with λ∗1 =, · · · , = λ∗N = λ∗
satisfies the KKT conditions for the GNE, i.e., (12).
The solution x∗ of V I(F, X) in (14) is termed as a variational
GNE or normalized equilibrium of the game with coupling
constraints in (7). A variational GNE enjoys an economical
interpretations of no price discrimination and has better stability
and sensitivity properties, therefore, can be regarded as a
refinement of GNE (refer to Kulkarni & Shanbhag (2012) for
more discussions). This paper will propose a novel distributed
algorithm for the agents to find a solution of V I(F, X) in (14),
thus provides a distributed coordination mechanism such that
8 Theorem
9 Theorem
2.1 of Facchinei, Fischer & Piccialli (2007)
6
2.3.3 of Facchinei and Pang (2007)
coupling constraint: Ax ≥ b. To ensure that the globally shared
coupling constraint is satisfied and all the agents have the same
local multipliers, all players must coordinate which necessarily
requires that multiplier graph Gλ must be connected. Therefore,
G f and Gλ could be two different information observation or
information sharing graphs because they serve for different
purposes.
capacity vector, then the true market capacities can be taken as
PN
PN 1
b = i=1
bi = i=1
N b̃i .
Therefore, we assume that player i only knows its local
fi (xi , x−i ), Ωi , and matrix Ai and bi with A = [A1 , · · · , AN ] and
PN
b = i=1
bi . In other words, player i has a local first-order
oracle of fi (xi , x−i ) which returns ∇ xi fi (xi , x−i ) given (xi , x−i ),
meanwhile player i can manipulate Ωi , Ai and bi for its local
computation.
We are ready to present the main distributed algorithm
after the introduction of algorithm notations. Agent (player)
i controls its local decision variable xi ∈ Rni and local
Lagragian multiplier λi ∈ Rm . Meanwhile, we also assume
that player i has a local auxiliary variable zi ∈ Rm for the
coordinations needed to satisfy the affine coupling constraint
and to reach consensus of the local multipliers λi . As indicated
before, player i can compute ∇ xi fi (xi , x−i ) by observing the
adversary players’ decisions that its local objective function
fi (xi , x−i ) directly depends on, that is the decisions of its
interference neighbors in Nif . On the other hand, player i can
also share information related to the local multiplier λi and
local auxiliary variable zi with its multiplier neighbours in Niλ
through multiplier network Gλ .
Next we show the distributed algorithm for agent i.
The players need to find the solution to V I(F, X) in (14)
in a distributed manner by local information observation and
sharing, hence find a variational GNE of the game in (7) without
any coordinator. To facilitate the local coordination between
agents, here we specify two graphs, G f and Gλ , related to the
local information observations or exchanging between players.
Graph G f , termed as interference graph, is defined according
to the dependence relationships between the agents’ objective
functions and the other players’ decisions, which is also called
graphical model for games in computer science (refer to
Kearns, Littman, & Singh (2001)). For graph G f = (N, E f ),
( j, i) ∈ E f if the objective function of agent i, fi (xi , x−i )
explicitly depends on the decision of player j. We define
Nif = { j|( j, i) ∈ E f } as the set of interference neighbors
whose decisions directly influence the objective function of
player i. Therefore, the objective function of player i can also
be written as fi (xi , {x j } j∈N f ), and the local oracle of player i
i
returns ∇ xi fi (xi , {x j } j∈N f ), i.e., ∇ xi fi (xi , x−i ), given {x j } j∈N f . The
i
i
local oracle might compute ∇ xi fi (xi , {x j } j∈N f ) by approximating
i
∇ xi fi (xi , x−i ) with local objective function value observations
(taking the simultaneous perturbation stochastic approximation
in Spall (1992) as an example), or by utilizing the estimation
techniques developed in Salehisadaghiani & Pavel (2016b).
On the other hand, for the coordination of the feasibility of
action sets and the consensus of local multipliers λ∗1 =, · · · , =
λ∗N = λ∗ in Theorem 3.2, we also assume that the agents can
exchange certain local information through a multiplier graph
Gλ = (N, Eλ ). ( j, i) ∈ Eλ if player i can receive certain
information from player j, while the information to be shared
through Gλ will be specified later. Thereby, player i has its
multiplier neighbors Niλ = { j|( j, i) ∈ Eλ }. W = [wi j ] is the
weighted adjacency matrix associated with multiplier graph Gλ ,
and L is the corresponding weighted Laplacian matrix.
Algorithm 3.5.
xi,k+1 = PΩi xi,k − τi (∇ xi fi (xi,k , {x j,k } j∈N f ) − ATi λi,k )
i
X
zi,k+1 = zi,k + νi
wi j (λi,k − λ j,k )
(18)
(19)
j∈Niλ
X
n
λi,k+1 = PRm+ λi,k − σi 2Ai xi,k+1 − Ai xi,k − bi +
wi j (λi,k − λ j,k )
j∈Niλ
+2
X
j∈Niλ
wi j (zi,k+1 − z j,k+1 ) −
X
wi j (zi,k − z j,k )
o
(20)
j∈Niλ
τi , νi , σi > 0 are fixed constant step-sizes of player i, and
W = [wi j ] is the weighted adjacency matrix of Gλ .
Algorithm 3.5 runs sequentially as follows. At the iteration
time k, player i gets ∇ xi fi (xi,k , x−i,k ) by observing x j,k , j ∈ Nif
through interference graph G f , and updates xi,k with (18);
meanwhile, player i gets λ j,k , j ∈ Niλ through multiplier graph
Gλ , and updates zi,k by (19). Then player i gets z j,k+1 , j ∈ Niλ
through multiplier graph Gλ and updates λi,k with (20) that also
employs the most recent local information xi,k+1 .
Intuitively speaking, (18) employs the projected gradient
descent of the local Lagrangian function in (10). (19) can be
regarded as the discrete-time intergration for the consensual
errors of local copies of multipliers, which will ensure the
consensus of λi eventually. In fact, a similar dynamics has
been employed in distributed optimization in Lei, Chen & Fang
(2016). Finally, (20) updates local multiplier by a combination
of the projected gradient ascent of local Lagrangian function
(10) and a proportional-integral dynamics for consensual errors
of multipliers. Section 4 will give a detailed algorithm
development from the viewpoint of operator splitting methods.
Assumption 3. Gλ = {N, Eλ } is undirected and connected.
W = WT .
Remark 3.4. We assume that each agent can observe the
decisions which its local objective function directly depends
on through interference graph G f . Therefore, player i can
get its local gradient ∇ xi fi (xi , {x j } j∈N f ). This type of local
i
information observation model has also been adopted in Yu,
van der Schaar & Sayed (2016) and Zhu & Frazzoli (2016).
On the other hand, player i’s feasible decision set may depend
on any other player k’s decision even if fi (xi , x−i ) does not
explicitly depend on the decision of player k, i.e. k < Nif .
In fact, player i’s feasible decision set implicitly depends on
all other players’ decisions through the globally shared affine
7
Algorithm 4.1.
Algorithm 3.5 is a totally distributed algorithm and has
following key features:
i). The full data decomposition and privacy protection
is achieved since each player only needs to know its local
objective function fi (xi , x−i ) and local feasible set Ωi .
ii). The matrix A is decomposed and each block Ai is
kept by player i, hence the privacy of each player is protected
because Ai describes how player i is involved in the market or
competition.
iii). Each player only needs to observe the decisions which
its local objective function directly depends on, and only needs
to share information with its multiplier neighbours, through
G f and Gλ , respectively. Both graphs usually have sparse
edge connections, therefore, the observation or communication
burden is relieved. Furthermore, the information observation
related with decisions and the information sharing related
with multipliers is decoupled and accomplished with different
graphs G f and Gλ , respectively.
Therefore, those two
information sharing processes can work in a parallel manner,
and can be designed independently.
iv). The algorithm converges with fixed step-sizes under
some mild conditions, and works in a Gauss-Seidel manner
that utilizes the most recent information when updating λi .
Moreover, (18) and (19) can even be computed in parallel for
player i.
xk+1 = PΩ xk − τ̄(F(xk ) − ΛT λ̄k )
z̄k+1 = z̄k + ν̄L̄λ̄k
(22)
n
o
λ̄k+1 = PRmN
λ̄k − σ̄ Λ(2xk+1 − xk ) − b̄ + L̄λ̄k + L̄(2z̄k+1 − z̄k )
+
(23)
Using the fact that PΩ (x) = ProxιΩ (x) = RNΩ (x) in (4) and
the definition of resolvent operator as RNΩ (x) = (Id + NΩ )−1 ,
equation (21) be be written as xk+1 = (Id + NΩ )−1 [xk − τ̄(F(xk ) −
ΛT λ̄k )], or equivalently,
xk − τ̄(F(xk ) − ΛT λ̄k ) ∈ xk+1 + NΩ (xk+1 ).
(24)
diag{ τ11 In1 , · · ·
−1
, τ1N InN }, ν̄−1
=
1
1
diag{ σ1 Im , · · · , σN Im }.
QN
= i=1
NΩi (xi ), hence
Notice that τ̄
=
1
1
diag{ ν1 Im , · · · , νN Im } and σ̄
=
Furthermore, NΩ is a cone and NΩ (x)
τ̄−1 NΩ (x) = NΩ (x). Therefore, (24) can be written as
−1
− F(xk ) ∈ NΩ (xk+1 ) − ΛT λ̄k+1 + τ̄−1 (xk+1 − xk ) + ΛT (λ̄k+1 − λ̄k ).
(25)
QN
m (λ ), and
mN (λ̄) =
Moreover, NRmN
is
a
cone,
N
N
R
i
R+
i=1
+
+
σ̄−1 NRmN
= NRmN
. Then with similar arguments, equation (23)
+
+
can be written as:
h
i
λ̄k − σ̄ Λ(2xk+1 −xk )− b̄+ L̄λ̄k + L̄(2z̄k+1 − z̄k ) ∈ λ̄k+1 +NRmN
(λ̄k+1 )
+
(26)
or equivalently,
4. Algorithm development
In this section, we first show how Algorithm 3.5 is developed
and provide the motivations behind the algorithm’s convergence
analysis. Then we verify that the limiting point of Algorithm
3.5 solves the V I(F, X) in (14), and thus finds a variational GNE
of the game in (7).
Algorithm 3.5 is inspired by the forward-backward splitting
methods for finding zeros of the sum of monotone operators
(Bauschke & Combettes (2011)) and the primal-dual algorithm
for optimization with linear composition terms by Condat
(2013). The key difference are the specific operator splitting
form and the augmentation of variables to achieve distributed
computations. Next, we systematically show how to write
Algorithm 3.5 in the form of a forward-backward operator
splitting algorithm.
Let us define some notations to write Algorithm 3.5 in a
compact form. Denote xk = col(x1,k , · · · , xN,k ) ∈ Rn , λ̄k =
col(λ1,k , · · · , λN,k ) ∈ RmN , z̄k = col(z1,k , · · · , zN,k ) ∈ RmN , b̄ =
col(b1 , · · · , bN ) ∈ RmN , τ̄ = diag{τ1 In1 , · · · , τN InN } ∈ Rn×n , ν̄ =
diag{ν1 Im , · · · , νN Im } ∈ RmN×mN , σ̄ = diag{σ1 Im , · · · , σN Im } ∈
RmN×mN , Λ = diag{A1 , · · · , AN } ∈ RmN×n and ΛT =
diag{AT1 , · · · , ATN } ∈ Rn×mN . L̄ = L ⊗ Im where L ∈ RN×N is
the weighted Laplacian matrix of multiplier graph Gλ .
Using these notations, the definition of pseudo-gradient
N
F(x)
in
(13)
and
PQi=1
=
Ωi (col(x1 , · · · , xN ))
10
col(PΩ1 (x1 ), · · · , PΩN (xN )) , Algorithm 3.5 can be written in a
compact form as:
10 Proposition
(21)
−[L̄λ̄k − b̄] ∈
+
NRmN
(λ̄k+1 ) + Λxk+1 + L̄z̄k+1
+
Λ(xk+1 − xk ) + L̄(z̄k+1 − z̄k ) + σ̄−1 (λ̄k+1 − λ̄k )
(27)
Therefore, equation(22) together with (25) and (27) can be
written in a compact form as:
NΩ (xk+1 ) − ΛT λ̄k+1
F(xk )
−
L̄
λ̄
0
−
∈
k+1
¯
NRmN
(
λ̄
)
+
Λx
+
L̄z̄
L̄λk − b̄
k+1
k+1
k+1
+
−1
τ̄
0
ΛT xk+1 − xk
L̄ z̄k+1 − z̄k
+ 0 ν̄−1
(28)
Λ
L̄ σ̄−1
λ̄k+1 − λ̄k
Denote
−1
τ̄
Φ = 0
Λ
0
ν̄−1
L̄
ΛT
L̄
σ̄−1
(29)
Notice that matrix Φ ∈ R(n+2mN)×(n+2mN) is symmetric due to
L = LT and L̄ = L ⊗ Im .
Denote Ω̄ = Ω × RmN × RmN
+ . Define the operators Ā : Ω̄ →
n+2mN
and B̄ : Ω̄ → 2R
as follows,
R
x
F(x)
,
0
A := z̄ 7→
λ̄
L̄λ̄ − b̄
x
NΩ (x) 0 0 −ΛT x
+ 0 0 −L̄ z̄ (30)
0
B := z̄ 7→
NRmN
(λ̄)
λ̄
Λ L̄
0
λ̄
+
n+2mN
23.16 of Bauschke & Combettes (2011)
8
Remark 4.2. Operators Ā and B̄ in (30) can be regarded as an
extension of operators A and B in (17) by augmenting λ ∈ Rm
to λ̄ ∈ RmN and introducing auxiliary variables z̄ ∈ RmN .
Moreover, operators in (30) utilize L1N = 0N of Laplacian
matrix L to ensure the consensus of λi , and utilize 1TN L = 0TN
to ensure the feasibility of affine coupling constraints.
The next result shows that any limiting point of Algorithm
3.5, that is, any zero point of operator Ā + B̄, is a variational
GNE of game (7).
Theorem 4.5. Suppose that Assumptions 1-3 hold. Consider
operators Ā and B̄ defined in (30), and operators A and B
defined in (17). Then the following statements hold:
(i): Given any col(x∗ , z̄∗ , λ̄∗ ) ∈ zer(Ā + B̄), then x∗ solves the
V I(F, X) in (14), hence x∗ is a variational GNE of game in (7).
Moreover λ̄∗ = 1N ⊗ λ∗ , and the multiplier λ∗ together with x∗
satisfy the KKT condition in (16), i.e., col(x∗ , λ∗ ) ∈ zer(A + B).
(ii): zer(A + B) , ∅ and zer(Ā + B̄) , ∅.
The next result shows that Algorithm 3.5, or equivalently
Algorithm 4.1, can be regarded as a forward-backward operator
splitting method for finding zeros of a sum of operators, or
an iterative computation of fixed points of a composition of
operators.
Lemma 4.3. Suppose Φ in (29) is positive definite, and
operators Ā and B̄ in (30) are maximally monotone. Denote
T 1 := Id − Φ−1 Ā and T 2 := (Id + Φ−1 B̄)−1 . Then any limiting
point of Algorithm 3.5, i.e., col(x∗ , z̄∗ , λ̄∗ ), is a zero of Ā + B̄ and
is a fixed point of T 2 ◦ T 1 .
Proof: (i): By the definition of operators Ā, B̄ in (30), we
have,
Ā + B̄ : col(x, z̄, λ̄) 7→
0 0 −ΛT x 0 F(x) + NΩ (x)
0
0 0 −L̄ z̄ + 0 +
NRmN
(λ̄)
−b̄
λ̄
Λ L̄
L̄
+
Proof: Denote $ = col(x, z̄, λ̄), then using (28), (29) and (30),
Algorithm 3.5 can written in a compact form as follows:
− Ā($k ) ∈ B̄($k+1 ) + Φ($k+1 − $k ).
(34)
Suppose that col(x∗ , z̄∗ , λ̄∗ ) ∈ zer(Ā + B̄). From the second
line of (34),
−L̄λ̄∗ = −L ⊗ Im λ̄∗ = 0mN .
(31)
Since Φ is symmetric and positive definite, we can write
equation (31) as $k − Φ−1 Ā($k ) ∈ $k+1 + Φ−1 B̄($k+1 ), or
equivalently,
(Id − Φ−1 Ā)($k ) ∈ (Id + Φ−1 B̄)($k+1 ).
It follows that λ̄∗ = 1N ⊗ λ∗ , λ∗ ∈ Rm since L is the weighted
Laplacian of multiplier graph Gλ and Gλ is connected due to
Assumption 3.
Then by the first line of (34), combined with ΛT =
diag{AT1 , · · · , ATN } and λ∗1 = λ∗2 =, · · · , = λ∗N = λ∗ , we have
(32)
Since Φ−1 B̄ is maximally monotone (refer to Lemma 5.6), (Id+
Φ−1 B̄)−1 is single-valued. Then by the definition of the inverse
of a set-valued operator, Algorithm 3.5 is written as
$k+1 = (Id + Φ B̄) (Id − Φ Ā)$k := T 2 ◦ T 1 $k .
−1
−1
−1
0n ∈ −ΛT 1N ⊗ λ∗ + F(x∗ ) + NΩ (x∗ ),
(33)
(35)
or equivalently,
Suppose that Algorithm 3.5, or equivalently (33), converges
to a limiting point $∗ . Then by the continuity of the right
hand of Algorithm 3.5 (In fact, the right hand of Algorithm
3.5 is Lipschitz continuous due to Assumption 2 and the
nonexpansive property of projection operator), $∗ = T 2 T 1 $∗ .
Therefore, any limiting point of Algorithm 3.5 is a fixed point
of the composition T 2 ◦ T 1 , and Algorithm 3.5 can be regarded
as an iterative computation of fixed points of T 2 ◦ T 1 .
By Theorem 25.8 of Bauschke & Combettes (2011), (33) is
also the forward-backward splitting algorithm for finding zeros
of a sum of monotone operators, hence $∗ ∈ zer(Φ−1 Ā+Φ−1 B̄)
for any limiting point $∗ . Since Φ is positive definite, any
limiting point $∗ = col(x∗ , z̄∗ , λ̄∗ ), i.e., $∗ = T 2 T 1 $∗ , also
belongs to zer(Ā + B̄). In fact,
0ni ∈ −ATi λ∗ + ∇ xi fi (xi∗ , x∗−i ) + NΩi (xi∗ ), i = 1, · · · , N.
(36)
By the third line of (34) and using L̄λ̄∗ = 0, it follows that
0mN ∈ Λx∗ − b̄ + L̄z̄∗ + NRmN
(λ̄∗ ).
+
This implies that there exist v1 , v2 , · · · , vN ∈ NRm+ (λ∗ ), such that
0mN = Λx∗ − b̄ + L ⊗ Im z̄∗ + col(v1 , · · · , vN ).
Multiplying both sides of above equation with 1TN ⊗ Im and
combining with 1T L = 0T , we have
0m = (1TN ⊗ Im )(Λx∗ − b̄ + L ⊗ Im z̄∗ + col(v1 , · · · , vN ))
=
$∗ = (Id + Φ−1 B̄)−1 (Id − Φ−1 Ā)$∗
⇔ (Id − Φ−1 Ā)$∗ ∈ (Id + Φ−1 B̄)$∗
⇔ 0 ∈ Φ−1 (Ā + B̄)($∗ ) ⇔ 0 ∈ (Ā + B̄)($∗ ).
N
X
i=1
Ai xi∗ −
N
X
i=1
bi +
N
X
vi
i=1
PN
TN
11
N
Due to the fact that NTi=1
,
Ωi =
i=1 NΩi if
i=1 int(Ωi ) , ∅
we have
PN
PN
PN
0m ∈
Ai xi∗ − i=1
bi + i=1
NRm+ (λ∗ )
Pi=1
P
N
N
∗
T
N
∈
Ai xi − i=1 bi + N i=1 Rm+ (λ∗ )
(37)
Pi=1
PN
N
∗
∗
m (λ ).
∈
A
x
−
b
+
N
i
i
R
i=1
i=1
+
i
2
Remark 4.4. The iteration $k+1 = T 2 T 1 $k is also known as
Picard iteration for iteratively approximating fixed points of
T 2 T 1 (refer to Berinde (2007)). Lemma 5.5 will give a sufficient
condition for Φ to be positive definite. Lemma 5.4 will give the
condition that ensures Ā and B̄ to be maximally monotone.
11 Corollary
9
16.39 of Bauschke & Combettes (2011)
By (36) and (37), for any col(x∗ , λ̄∗ , z̄∗ ) ∈ zer(Ā + B̄), the
KKT condition for V I(F, X) in (14), i.e. (16), is satisfied for
x∗ , λ∗ . We conclude that x∗ solves V I(F, X) in (14), and is a
variational GNE of game (7) by Theorem 3.2. It also follows
that λ∗ together with x∗ satisfy the KKT condition in (16). This
also implies col(x∗ , λ∗ ) ∈ zer(A + B).
analysis on the contractive property of underlying operators.
However, the contractivity assumption on operators is usually
quite restrictive. Herein we resort to the theory of averaged
operators and firmly nonexpansive operators for convergence
analysis. Firstly we give some basic definitions and properties
of averaged operators and firmly nonexpansive operators13 . All
the following results are valid in Hilbert spaces, thus they hold
in Euclidean spaces with any G−matrix induced norm || · ||G ,
given G as a symmetric positive definite matrix. Denote by
|| · || an arbitrary matrix induced norm in a finite dimensional
Euclidean space.
An operator T : Ω ⊂ Rm → Rm is nonexpansive if it
is 1−Lipschitzian, i.e., ||T (x) − T (y)|| ≤ ||x − y||, ∀x, y ∈ Ω.
An operator T is α−averaged if there exists a nonexpansive
0
0
operator T such that T = (1 − α)Id + αT . Denote the class
of α−averaged operators as A(α). If T ∈ A( 12 ), then T is also
called firmly nonexpansive operator.
(ii) Under Assumptions 1 and 2, the considered game in (7)
has a unique variational GNE x∗ , and there exists λ∗ ∈ Rm
such that the KKT condition (16) is satisfied, i.e. col(x∗ , λ∗ ) ∈
zer(A + B). Therefore zer(A + B) , ∅.
Then we need to show that there exists col(x∗ , λ̄∗ , z̄∗ ) such
that col(x∗ , λ̄∗ , z̄∗ ) ∈ zer(Ā + B̄).
Take λ̄∗ = 1N ⊗ λ∗ . Because L1N = 0, the second line of (34)
is satisfied.
Since col(x∗ , λ∗ ) ∈ zer(A + B), 0 ∈ F(x∗ ) − AT λ∗ +
NΩ (x∗ ). Using λ∗1 = λ∗2 =, · · · , = λ∗N = λ∗ , ΛT λ̄∗ =
col(AT1 λ∗ , · · · , ATN λ∗ ) = AT λ∗ . Therefore, the first line of (34) is
satisfied with x∗ , 1N ⊗ λ∗ .
Moreover, with col(x∗ , λ∗ ) ∈ zer(A + B), 0m ∈ Ax∗ −
b + NRm+ (λ∗ ). Then we need to show that there exists z̄∗ =
col(z∗1 , · · · , z∗N ) ∈ RmN , such that the third line of (34) is
satisfied. Take v∗ ∈ NRm+ (λ∗ ) such that 0 = Ax∗ − b + v∗ . Since
QN
λ̄∗ = 1N ⊗ λ∗ and NRmN
NRm+ (λ∗ ), take v∗1 = v∗2 =
(λ̄∗ ) = i=1
+
, · · · , = v∗N = N1 v∗ ∈ NRm+ (λ∗ ). Then (1TN ⊗ Im )(Λx∗ − b̄ + L̄λ̄∗ +
PN
PN
bi + v∗ = Ax∗ − b + v∗ = 0m .
Ai xi∗ − i=1
col(v∗1 , · · · , v∗N )) = i=1
∗
∗
∗
That is Λx − b̄ + L̄λ̄ + col(v1 , · · · , v∗N ) ∈ Null(1TN ⊗ Im ). By
the fundamental theorem of linear algebra12 , Null(1TN ⊗ Im ) =
Range(1N ⊗ Im )⊥ and Range(L̄) = Null(L̄)⊥ since L̄ is also
symmetric. Notice that Null(L̄) = Range(1N ⊗ Im ), hence
Λx∗ − b̄ + L̄λ̄∗ + col(v∗1 , · · · , v∗N ) ∈ Range(L̄). Noticing that
col(v∗1 , · · · , v∗N ) ∈ NRmN
(λ̄∗ ), there exists z̄∗ ∈ RmN such that the
+
third line of (34) is satisfied with x∗ , z̄∗ , 1N ⊗ λ∗ . Therefore,
zer(Ā + B̄) , ∅.
2
Lemma 5.1. 14 Given an operator T : Ω ⊂ Rm → Rm and
α ∈ (0, 1), the following three statements are equivalent:
(i): T ∈ A(α);
2
(ii): ||T x−T y||2 ≤ ||x−y||2 − 1−α
α ||(x−y)−(T x−T y)|| , ∀x, y ∈ Ω
2
2
(iii): ||T x − T y|| + (1 − 2α)||x − y|| ≤ 2(1 − α)hx − y, T x −
T yi, ∀x, y ∈ Ω.
By (iii) of Lemma (5.1), T ∈ A( 12 ) if and only if
||T (x) − T (y)||2 ≤ hx − y, T (x) − T (y)i, ∀x, y ∈ Ω.
The operator T is called β−cocoercive (or β−inverse strongly
monotone) if βT is firmly nonexpansive, i.e.,
β||T (x) − T (y)||2 ≤ hx − y, T (x) − T (y)i, ∀x, y ∈ Ω.
(39)
Lemma 5.2. 15 For a convex differentiable function f with ϑ−
Lipschitzian gradient, we have ∇ f to be ϑ1 −cocoercive, i.e.,
1
||∇ f (x) − ∇ f (y)||22 ≤ hx − y, ∇ f (x) − ∇ f (y)i.
(40)
ϑ
Lemma 5.2 is known as Baillon-Haddad theorem, and one
elementary proof can be found in Theorem 2.1.5 of Nesterov
(2013).
5. Convergence Analysis
In this section, we prove the convergence of Algorithm 3.5
by giving a sufficient step-size choice condition. The analysis
is based on the compact reformulation (33). We will first
show that all the prerequisites in Lemma 4.3 can be satisfied
under suitable step-sizes. Then (33), i.e., Algorithm 3.5,
can be regarded as a forward-backward splitting algorithm for
finding zeros of a sum of monotone operators, or equivalently,
an iterative computation of fixed points of a composition of
operators.
In fact, some existing NE (GNE) algorithms can also be
regarded as a type of iterative computation of fixed points
of operators, such as the best-response learning dynamics
(Parise, Gentile, Grammatico & Lygeros (2015)), relaxation
algorithms based on Nikaido-Isoda function (Krawczyk &
Uryasev (2000) and Contreras, Klusch, Krawczyk (2004))
and the proximal-best response algorithm in Facchinei & Pang
(2010). Most of above works built their algorithm convergence
12 Page
(38)
Lemma 5.3. 16 If operator ∆ is maximally monotone, then
T = R∆ = (Id + ∆)−1 is firmly nonexpasive and 2R∆ − Id is
nonexpansive.
Hence, the projection operator onto a closed convex set is firmly
nonexpansive since PΩ = ProxιΩ = R∂ιΩ = RNΩ and NΩ is
maximally monotone17 .
In the following, we analyze the maximal monotonicity of
operators Ā, B̄ in (30), the positive definite property of matrix
Φ, and the properties of operators T 1 and T 2 defined in Lemma
4.3 by giving sufficient step-sizes choice conditions, which are
shown in Lemma 5.4, 5.5 and 5.6, respectively.
13 Chapter
4 and Chapter 20 of Bauschke & Combettes (2011)
4.25 of Bauschke & Combettes (2011)
15 Theorem 18.15 of Bauschke & Combettes (2011)
16 Proposition 23.7 of Bauschke & Combettes (2011)
17 Example 20.41 and Proposition 4.8 of Bauschke & Combettes (2011)
14 Proposition
405 of Meyer (2000)
10
Obviously, domB1 = Rn+2mN , hence B̄ = B1 + B2 is also
maximally monotone21 .
2
Lemma 5.4. Suppose Assumptions 1-3 hold.
Given an
Euclidean space with norm || · ||2 , then
(i): Operator Ā in (30) is β−cocoercive with 0 < β ≤
min{ 2d1 ∗ , θη2 } where d∗ is the maximal weighted degree of
P
P
multiplier graph Gλ , i.e., d∗ = max{ Nj=1 w1 j , · · · , Nj=1 wN j },
and η, θ are parameters in Assumption 2;
(ii): Operator B̄ in (30) is maximally monotone.
Lemma 5.5. Given any δ > 0, if each player i takes τi > 0,
νi > 0 and σi > 0 as its local fixed step-sizes in Algorithm 3.5
that satisfy:
Proof: (i): According to the definition of Ā in (30) and the
definition of β−cocercive in (39), we need to prove that
F(x1 ) F(x2 ) x1 x2
0
0
h
−
, z̄1 − z̄2 i ≥
λ̄2
L̄λ̄1 − b̄ L̄λ̄2 − b̄ λ̄1
x1 x2
F(x1 ) F(x2 )
2
0
0
β||
||2 , ∀ z̄1 , z̄2 ∈ Ω̄
−
λ̄2
λ̄1
L̄λ̄2 − b̄
L̄λ̄1 − b̄
(41)
Notice that L̄λ̄− b̄ is the gradient of function f˜(λ̄) := 21 λ̄T L̄λ̄−
b̄T λ̄. Moveover, f˜(λ̄) is a convex function since ∇2 f˜(λ̄) = L̄ is
positive semi-definite due to Assumption 3 18 . It can easily be
verified that L̄λ̄− b̄ is ||L||2 −Lipschitz continuous (notice that the
eigenvalues of L̄ are just the elements in col(0, s2 · · · , sN ) ⊗ 1m ),
therefore, by Lemma 5.2
hL̄λ̄1 − b̄ − (L̄λ̄2 − b̄), λ̄1 − λ̄2 i ≥
1
||L̄λ̄1 − L̄λ̄2 ||22 .
||L||2
τi
≤
νi
σi
≤
≤
P1
,
T
max j=1,...ni { m
k=1 |[Ai ] jk |}+δ
1
2di +δ
1
Pn i
max j=1,...m { k=1
|[Ai ] jk |}+2di +δ
(44)
then matrix Φ defined in (29) is positive definite, and Φ−δIn+2mN
is positive semi-definite.
Proof: It is sufficient to show that Φ − δIn+2mN is positive
semi-definite.
−1
τ̄ − δIn
0
ΛT
−1
0
ν̄ − δImN
L̄
Φ − δIn+2mN =
,
−1
Λ
L̄
σ̄ − δImN
(45)
where τ̄−1 − δIn = diag{( τ11 − δ)In1 , · · · , ( τ1N − δ)InN }, ν̄−1 −
δImN = diag{( ν11 − δ)Im , · · · , ( ν1N − δ)Im }, and σ̄−1 − δImN =
diag{( σ11 − δ)Im , · · · , ( σ1N − δ)Im }. One sufficient condition
for matrix Φ − δIn+2mN to be positive semi-definite is that it
is diagonally dominant with nonnegative diagonally elements,
that is for every row of the matrix the diagonal entry is larger
than or equal to the sum of the magnitudes of all the other
(non-diagonal) entries in that row. This is equivalent to require
that,
P
1
T
≥ max j=1,...ni { m
k=1 |[Ai ] jk |}
τi − δ
P
m
1
≥
2di
j=1 |Li j | =P
νi − δ
P
(46)
ni
1
−
δ
≥
max
{
|[Ai ] jk |} + mj=1 |Li j |
j=1,...m
σi
Pk=1
i
= max j=1,...m { nk=1
|[Ai ] jk |} + 2di
(42)
Since ||L||2 ≤ sN and d∗ ≤ sN ≤ 2d∗ by (6) where d∗ =
max{d1 , · · · , dN } is the maximal weighted degree of multiplier
1
graph Gλ , we have ||L||
≥ 2d1 ∗ .
2
Meanwhile by Assumption 2, F(x) is η−strongly monotone
and θ− Lipschitz continuous over Ω. By ||F(x1 ) − F(x2 )||22 ≤
θ2 ||x1 − x2 ||22 , ∀ x1 , x2 ∈ Ω we have
η
||F(x1 ) − F(x2 )||22 .
θ2
(43)
Taking 0 < β ≤ min{ 2d1 ∗ , θη2 } and adding (42) and (43) yields
(41). Thus operator Ā is β−cocoercive. By the definition
of β−cocoercive in (39), operator Ā is also monotone. Since
operator Ā is also single-valued, it is also maximally monotone.
(ii): The operator B̄ in (30) can be written as:
0 0 −ΛT x NΩ (x)
0 0 −L̄ z̄ + 0 := B1 + B2 .
NRmN
Λ L̄
0
λ̄
+
hF(x1 ) − F(x2 ), x1 − x2 i ≥ η||x1 − x2 ||22 ≥
It can easily be verified that if each agent chooses its local
step-sizes satisfying (44), then (46) is satisfied.
2
Given a globally known parameter δ, each agent can
independently choose its local step sizes τi ,νi , and σi with the
rule given in (44).
Suppose that the step-sizes τi , µi , σi in Algorithm 3.5 are
chosen such that Φ in (29) is positive definite. Thus we can
define a norm
√ induced by√matrix Φ for a finite Euclidean space
as ||x||Φ = hx, xiΦ = hΦx, xi. The next result investigates
the properties of operators Φ−1 Ā, Φ−1 B̄ and T 1 , T 2 defined in
Lemma 4.3 under Φ−induced norm || · ||Φ .
Since L̄ is symmetric, B1 is a skew-symmetric matrix, i.e.,
BT1 = −B1 . Hence, B1 is maximally monotone19 .
B2 can be written as the direct sum of NΩ × 0mN × NRmN
.
+
Both NΩ and NRmN
are
maximally
monotone
as
normal
cones
+
of closed convex sets. Obviously, 0mN is also maximally
monotone as a single-valued operator. Furthermore, the direct
sum of maximally monotone operators is also maximally
monotone20 , hence B2 is also maximally monotone.
Lemma 5.6. Suppose Assumptions 1-3 hold. Take 0 < β ≤
min{ 2d1 ∗ , θη2 } where d∗ is the maximal weighted degree of
multiplier graph Gλ , and η, θ are parameters in Assumption
1
2. Take δ > 2β
. Suppose that the step-sizes τi , νi , σi in
Algorithm 3.5 are chosen to satisfy (44). Then the operators
Φ−1 Ā and Φ−1 B̄ with Φ in (29) and Ā, B̄ in (30), and operators
18 Proposition
17.10 of Bauschke & Combettes (2011)
20.30 of Bauschke & Combettes (2011)
20 Proposition 20.23 of Bauschke & Combettes (2011)
19 Example
21
11
Corollary 24.4 of Bauschke & Combettes (2011)
T 1 = Id − Φ−1 Ā, T 2 = RΦ−1 B̄ = (Id + Φ−1 B̄)−1 defined as Lemma
4.3 satisfy the following properties under the Φ−induced norm
|| · ||Φ :
1
(i). Φ−1 Ā is βδ−cocoercive, and T 1 ∈ A( 2δβ
).
−1
(ii). Φ B̄ is maximally monotone, and T 2 ∈ A( 12 ).
1
A( 2δβ
), and T 2 := (Id + Φ−1 B̄)−1 ∈ A( 12 ) in the Euclidean
space with Φ−matrix induced norm || · ||Φ . The next result
shows the convergence of Algorithm 3.5 based on its compact
reformulation (33) and the properties of T 1 and T 2 .
Proof: (i): By the definition of cocoercivity in (39), we need to
prove hΦ−1 Ā(x) − Φ−1 Ā(y), x − yiΦ ≥ βδ||Φ−1 Ā(x) − Φ−1 Ā(y)||2Φ ,
∀x, y ∈ Ω̄. Noticing that by the choice of parameters τi , νi , σi ,
we have that matrix Φ − δIn+2mN is positive semi-definite from
Lemma 5.5. Denote smax (Φ) and smin (Φ) as the maximal and
minimal eigenvalues of matrix Φ. It must hold that smax (Φ) ≥
smin (Φ) ≥ δ. Furthermore, ||Φ||2 = smax (Φ) ≥ smin (Φ) =
1
≥ δ 22 , therefore, we also have ||Φ−1 ||2 ≤ 1δ . Notice that
||Φ−1 ||2
the operator Ā is single-valued and Φ−1 is also nonsingular, so
that for any x, y ∈ Ω̄,
Theorem 5.7. Suppose Assumptions 1-3 hold. Take 0 <
β ≤ min{ 2d1 ∗ , θη2 } where d∗ is the maximal weighted degree of
multiplier graph Gλ , and η, θ are parameters in Assumption
1
. The step-sizes τi , νi , σi in Algorithm 3.5
2. Take δ > 2β
are chosen to satisfy (44). Then with Algorithm 3.5, each
player has its local strategy xi,k converging to its corresponding
component in the variational GNE of game (7), and the local
Lagrangian multipliers λi,k of all the agents converge to the
same Lagrangian multiplier corresponding with KKT condition
(16), i.e.,
||Φ−1 Ā(x) − Φ−1 Ā(y)||2Φ
= hΦΦ−1 (Ā(x) − Ā(y)), Φ−1 (Ā(x) − Ā(y))i
= ||Ā(x) − Ā(y)||2Φ−1 ≤ 1δ ||Ā(x) − Ā(y)||22
lim xi,k → xi∗ ,
k→∞
By the β−cocoercive property of Ā in Lemma 5.4 and the above
inequality,
Firstly, by (i) and (ii) of Lemma 5.1 and the fact that T 1 , T 2
are averaged operators due to Lemma 5.6, T 1 , T 2 are also
nonexpansive operators under the Φ−matrix induced norm ||·||Φ .
Take any $∗ ∈ zer(Ā + B̄), or equivalently any fixed point of
T 2 ◦ T 1 , i.e., $∗ = T 2 T 1 $∗ , and then by Lemma 4.3 and (33),
1
1
1
)Id +
(−T̆ ) ∈ A(
)
2βδ
2βδ
2βδ
||$k+1 − $∗ ||Φ = ||T 2 (T 1 $k ) − T 2 (T 1 $∗ )||Φ
≤ ||T 1 $k − T 1 $∗ || ≤ ||$k − $∗ ||Φ
1
since 1 < 2βδ by the assumption that δ > 2β
and −T̆ is also
nonexpansive.
(ii). Φ is symmetric positive definite and nonsingular. For
any (x, u) ∈ graΦ−1 B̄ and (y, v) ∈ graΦ−1 B̄, Φu ∈ ΦΦ−1 B̄(x) ∈
B̄(x) and Φv ∈ ΦΦ−1 B̄(y) ∈ B̄(y). Then hx − y, u − viΦ =
hx − y, Φ(u − v)i ≥ 0, ∀x, y since B̄ is monotone by Lemma
5.4. Therefore, Φ−1 B̄ is monotone under the Φ−matrix induced
product h·, ·iΦ .
Furthermore, take (y, v) ∈ Ω̄ × Rn+2mN , and hx − y, u − viΦ ≥
0, for any other (x, u) ∈ gra(Φ−1 B̄). For any (x, ũ) ∈ graB̄,
we have (x, Φ−1 ũ) ∈ gra(Φ−1 B̄). hx − y, Φ(Φ−1 ũ − v)i ≥ 0,
or equivalently, hx − y, ũ − Φv)i ≥ 0. Since B̄ is maximally
monotone, then (y, Φv) ∈ graB̄. We conclude that v ∈ Φ−1 B̄(y)
which implies that Φ−1 B̄ is maximally monotone.
Therefore, by Lemma 5.3 T 2 = (Id + Φ−1 B̄)−1 is firmly
nonexpansive under the Φ−matrix induced norm || · ||Φ .
2
(49)
Hence the sequence {||$k − $∗ ||Φ } is non-increasing and
bounded from below. By the monotonic convergence theorem,
{||$k − $∗ ||} is bounded and converges for every $∗ ∈ zer(A +
B).
1
) and T 2 ∈ A( 12 ). Denote ξ =
By Lemma 5.6, T 1 ∈ A( 2δβ
1
2δβ ∈ (0, 1). Then with (ii) of Lemma 5.1 and (33) we have,
||$k+1 − $∗ ||2Φ = ||T 2 (T 1 $k ) − T 2 (T 1 $∗ )||2Φ
≤ ||T 1 $k − T 1 $∗ ||2Φ
−||(T 1 $k − T 1 $∗ ) − (T 2 T 1 $k − T 2 T 1 $∗ )||2Φ
≤ ||$k − $∗ ||2Φ − ||(T 1 $k − T 1 $∗ ) − (T 2 T 1 $k − T 2 T 1 $∗ )||2Φ
∗
∗ 2
− 1−ξ
ξ ||$k − $ − (T 1 $k − T 1 $ )||Φ
(50)
where the first inequality follows by T 2 ∈ A( 21 ) and the second
1
inequality follows by T 1 ∈ A( 2δβ
), both utilizing (ii) of Lemma
5.1. Notice that
Summarizing the above results, take 0 < β ≤ min{ 2d1 ∗ , θη2 },
1
δ > 2β
, and τi , νi , σi satisfying (44), then T 1 := Id − Φ−1 Ā ∈
22 Proposition
(48)
Proof: With Lemma 5.4 and 5.5, Algorithm 3.5 can be written
in a compact form (33) according to Lemma 4.3, i.e., $k+1 =
T 2 T 1 $k . The convergence analysis will be conducted via the
analysis of this iterative computation of fixed points of T 2 ◦ T 1 .
hΦ−1 Ā(x) − Φ−1 Ā(y), x − yiΦ = hĀ(x) − Ā(y), x − yi
≥ β||Ā(x) − Ā(y)||22 ≥ βδ||Φ−1 Ā(x) − Φ−1 Ā(y)||2Φ .
(47)
Therefore, the operator Φ−1 Ā is βδ−cocoercive under the
Φ−induced norm || · ||Φ .
Moreover, βδΦ−1 Ā is firmly nonexpansive by the definition
of cocoercive operator. This implies that there exists a
nonexpansive operator T̆ such that βδΦ−1 Ā = 12 T̆ + 21 Id. Then
T 1 = Id − Φ−1 Ā = (1 −
lim λi,k → λ∗ , ∀i = 1, · · · , N.
k→∞
α||x||2 + (1 − α)||y||2 = ||αx + (1 − α)y||2 + α(1 − α)||x − y||2 . (51)
5.2.7 and 5.2.8 in Meyer (2000)
12
For the second and third terms on the right hand side of (50),
the current operator output with previous iterate, leads to
the well-known Krasnosel’skiĭ-Mann type of fixed point
iteration23 , and has been utilized in (generalized) Nash
equilibrium computation in Contreras, Klusch, Krawczyk
(2004) and Krawczyk & Uryasev (2000). Meanwhile, inertial
algorithms in operator splitting methods have received attention
in recent years, such as Alvarez & Attouch (2001), Attouch,
Chbani, Peypouquet, and Redont (2016), Lorenz & Pock
(2015) and Rosasco, Villa & Vũ (2016). These efforts
are partially motivated by the heavy ball method in Polyak
(1987) and Nesterov’s acceleration algorithm in Nesterov
(2013)) for optimization problems and their recent success
in machine learning applications (refer to Wibisono, Wilson,
and Jordan, (2016)). In particular, Nesterov’s acceleration
algorithm is proved to enjoy an optimal convergence speed with
a specific step-size choice. Thereby, in this work we consider a
distributed algorithm with inertia for variational GNE seeking
given as below:
||(T 1 $k − T 1 $∗ ) − (T 2 T 1 $k − T 2 T 1 $∗ )||2Φ
∗
∗ 2
+ 1−ξ
ξ ||$k − $ − (T 1 $k − T 1 $ )||Φ
1
∗
= ξ [ξ||(T 1 $k − T 1 $ ) − (T 2 T 1 $k − T 2 T 1 $∗ )||2Φ
+(1 − ξ)||(T 1 $k − T 1 $∗ ) − ($k − $∗ )||2Φ ]
1
= ξ ||(T 1 $k − T 1 $∗ ) − ξ(T 2 T 1 $k − T 2 T 1 $∗ )
−(1 − ξ)($k − $∗ )||2Φ
1
+ ξ ξ(1 − ξ)||(T 1 $k − T 1 $∗ ) − (T 2 T 1 $k − T 2 T 1 $∗ )
−(T 1 $k − T 1 $∗ ) + ($k − $∗ )||2Φ
≥ (1 − ξ)||($k − $∗ ) − (T 2 T 1 $k − T 2 T 1 $∗ )||2Φ
= (1 − ξ)||$k − T 2 T 1 $k ||2Φ
= (1 − ξ)||$k − $k+1 ||2Φ
(52)
where the second equality follows from (51) by setting α = ξ,
x = (T 1 $k − T 1 $∗ ) − (T 2 T 1 $k − T 2 T 1 $∗ ) and y = (T 1 $k −
T 1 $∗ ) − ($k − $∗ ).
Combining (50) and (52) yields ∀k ≥ 0,
||$k+1 − $∗ ||2Φ ≤ ||$k − $∗ ||2Φ − (1 − ξ)||$k − $k+1 ||2Φ .
Algorithm 6.1.
(53)
Using (53) from 0 to k and adding all k + 1 inequalities yields
||$k+1 − $∗ ||2Φ ≤ ||$0 − $∗ ||2Φ − (1 − ξ)
k
X
Acceleration phase
||$l − $l+1 ||2Φ .
l=0
P
Taking limit as k → ∞ we have, (1 − ξ) ∞
||$ − $k+1 ||2Φ ≤
k=1
P∞ k
∗ 2
||$0 − $ ||Φ . Since 1 − ξ > 0, it follows that k=1 ||$k − $k+1 ||2Φ
converges and limk→∞ $k − $k+1 = 0.
Since {||$k − $∗ ||} is bounded and converges, {$k } is a
bounded sequence. There exists a subsequence{$nk } that
converges to $̃∗ . Notice that the composition T 2 ◦ T 1 is
(Lipschitz) continuous and single-valued, because (33) is just
an equivalent expression of Algorithm 3.5, and obviously the
right hand side of Algorithm 3.5 is continuous. $nk +1 =
T 2 T 1 $nk . Since T 2 T 1 is continuous, and limnk →∞ $nk − $nk +1 =
0, passing to limiting point, we have $̃∗ = T 2 T 1 $̃∗ . Therefore,
the limiting point $̃∗ is a fixed point of T 2 T 1 , or equivalently,
$̃∗ ∈ zer(Ā + B̄).
Setting $∗ = $̃∗ in (49), we have {||$k − $̃∗ ||} is bounded
and converges. Since there exists a subsequence {$nk } that
converges to $̃∗ , it follows that {||$k − $̃∗ ||} converges to zero.
Therefore, limk→∞ $k → $̃∗ . By Theorem 4.5, this just implies
(48).
2
x̃i,k = xi,k + α(xi,k − xi,k−1 )
(54)
z̃i,k = zi,k + α(zi,k − zi,k−1 )
(55)
λ̃i,k = λi,k + α(λi,k − λi,k−1 )
(56)
U pdate phase
xi,k+1 = PΩi x̃i,k − τi (∇ xi fi ( x̃i,k , { x̃ j,k } j∈N f ) − ATi λ̃i,k )
i
X
zi,k+1 = z̃i,k + νi
wi j (λ̃i,k − λ̃ j,k )
(57)
(58)
j∈Niλ
X
n
λi,k+1 = PRm+ λ̃i,k − σi 2Ai xi,k+1 − Ai x̃i,k − bi +
wi j (λ̃i,k − λ̃ j,k )
j∈Niλ
+2
X
j∈Niλ
wi j (zi,k+1 − z j,k+1 ) −
X
wi j (z̃i,k − z̃ j,k )
o
(59)
j∈Niλ
α > 0 is a fixed step-size in the acceleration phase, and
τi > 0, νi > 0, σi > 0 are fixed step-sizes of player i, and
W = [wi j ] is the weighted adjacency matrix of multiplier graph
Gλ .
Compared with Algorithm 3.5, Algorithm 6.1 has two
phases. In the acceleration phase, each player uses the local
state information of the last two steps to get predictive variables
by a simple linear extrapolation. In the update phase, the
players just feed the predictive variables to Algorithm 3.5 to
get the next iterates. Hence compared with Algorithm 3.5,
Algorithm 6.1 has only an additional simple local computation
burden. Obviously, Algorithm 6.1 is also totally distributed,
and shares all the features of Algorithm 3.5. However, there is
an additional need to choose a proper step-size α.
In the following two subsections, we will first give some
intuitive interpretation of Algorithm 6.1 from the viewpoint of a
6. Distributed algorithm with inertia
In this section, we propose a distributed algorithm with
inertia for variational GNE seeking, which possibly accelerates
the convergence under some mild additional computation
burden.
There are various modifications of Picard fixed point iteration
to achieve the possible acceleration of convergence speed, and
most of them fall into the domains of relaxation algorithm
and inertial algorithm (Refer to Iutzeler & Hendrickx
(2016) for reviews and numerical comparisons for optimization
problems). The relaxation algorithm that simply combines
23 Chapter
13
5 of Bauschke & Combettes (2011)
where $̃(kh) is an interpolation point to be determined later.
Denote $k = $(kh) and take h = 1, then (66) can be written as
discretization of continuous-time dynamical systems, and then
prove its convergence.
$k +(1−α̃)($k −$k−1 )−Φ−1 Ā($̃k ) ∈ $k+1 +Φ−1 B̄($k+1 ). (67)
6.1. Interpretations from viewpoints of dynamical systems
Denote α = 1 − α̃ and take $̃k = $k + α($k − $k−1 ), then (67)
can be written as
The interpretation of inertial (acceleration) algorithms from
a continuous-time dynamical system viewpoint can be found
in Polyak (1987) and most recently in Wibisono, Wilson,
and Jordan, (2016) for optimization problems and in Attouch,
Chbani, Peypouquet, and Redont (2016) for proximal point
algorithms. Here we give a comparative development of
Algorithm 3.5 and Algorithm 6.1 just for illustrations of the
differences behind the algorithms.
Firstly, let us show that Algorithm 3.5, or equivalently its
compact reformulation (33) in Lemma 4.3, can be interpreted
as the discretization of the following dynamical system:
$̇ ∈ −Φ−1 Ā($) − Φ−1 B̄($).
$̃k = $k + α($k − $k−1 ),
$̃k − Φ−1 Ā($̃k ) ∈ $k+1 + Φ−1 B̄($k+1 ).
(68) leads to equations (63)-(64), or equivalently Algorithm 6.1.
Remark 6.2. Compared with (60), (65) is a second order
dynamical system with an additional inertial term α$̇, hence
(65) enjoys better convergence properties than (60). Therefore,
it is expected that Algorithm 6.1, as a discretization of (65),
would have better convergence properties than Algorithm 3.5.
(60)
6.2. Convergence analysis
The following result proves the convergence of Algorithm
6.1 by providing sufficient step-size choices for α as well as
τi , νi , σi . The sufficient choice condition for α can be ensured
by solving a simple algebra inequality. The proof idea of the
following result is motivated by inertial algorithms works for
optimization and operator splitting such as Alvarez & Attouch
(2001), Attouch, Chbani, Peypouquet, and Redont (2016),
Rosasco, Villa & Vũ (2016), Iutzeler & Hendrickx (2016),
and especially Lorenz & Pock (2015). However, since this
work considers a noncooperative game setup and adopts a fixed
step-size in the distributed algorithm, Theorem 6.3’s proof is
also provided for completeness.
In fact, for differential inclusion (60), we have the following
implicit/explicit discretization with step-size of h,
$(kh + h) − $(kh)
∈ −Φ−1 Ā($(kh))−Φ−1 B̄($(kh+h)). (61)
h
Denote $k = $(kh) and take h = 1, then (61) can be written as
$k − Φ−1 Ā($k ) ∈ $k+1 + Φ−1 B̄($k+1 ).
(62)
Therefore, the implicit/explicit discretization of (60) is exactly
(32) that leads to (33), or equivalently Algorithm 3.5.
Moreover, the explicit discretization corresponds with the
forward step, and the implicit discretization corresponds with
the backward step. That’s the reason why Algorithm 3.5 is
called a forward-backward splitting algorithm.
Adopt similar compact notations as in Section 4, and
denote x̃ = col( x̃1 , · · · , x̃N ), λ̃¯ = col(λ̃1 , · · · , λ̃N ), and z̃¯ =
col(z̃1 , · · · , z̃N ). And further denote $̃ = col(x̃, λ̃¯ , z̃¯). Then by
similar arguments as in Section 4 and using operators Ā and
B̄ defined in (30) and Φ defined in (29), Algorithm 6.1 can be
written in a compact form (assume B̄ is maximally monotone),
$̃k = $k + α($k − $k−1 )
−1
Theorem 6.3. Suppose Assumptions 1-3 hold. Take 0 <
β ≤ min{ 2d1 ∗ , θη2 } where d∗ is the maximal weighted degree of
multiplier graph Gλ , and η, θ are parameters in Assumption 2.
1
and 0 < α < 1
Given a sufficient small 0 < < 1, take δ > 2β
in Algorithm 6.1 such that 2βδ(1 − 3α − ) ≥ (1 − α)2 . Suppose
that player i chooses its step-sizes τi , νi , σi in Algorithm 6.1
satisfying (44). Then with Algorithm 6.1, players’ local
strategies converge to the variational GNE of game in (7), and
the local multipliers λi,k of all the agents converge to the same
multiplier corresponding with KKT condition (16), i.e.,
(63)
$k+1 = (Id + Φ B̄) (Id − Φ Ā)$̃k
−1
−1
(64)
lim xi,k → xi∗ ,
k→∞
where $k is defined as in Section 4.
We can show that Algorithm 6.1, or equivalently (63)-(64),
can be interpreted as the discretization of the following
second-order continuous-time dynamical system,
$̈ + α̃$̇ ∈ −Φ−1 Ā($) − Φ−1 B̄($).
(68)
lim λi,k → λ∗ , ∀i = 1, · · · , N.
k→∞
(69)
Proof: By the choice of β, operator Ā is β−cocoercive and
operator B̄ is maximally monotone due to Lemma 5.4. By the
choice of δ, τi , νi and σi , Φ − δIn+2mN is positive semi-definite
due to Lemma 5.5, and Φ−1 Ā and Φ−1 B̄ satisfy the properties
in Lemma 5.6. Therefore, Algorithm 6.1 can be exactly written
in the compact form of (63)-(64) with similar arguments as in
Lemma 4.3.
Resorting to Theorem 4.5, we only need to show that
Algorithm 6.1 converges and its limiting point belongs to
zer(Ā + B̄). In fact, any limiting point of Algorithm 6.1 satisfies
$∗ ∈ zer(Ā + B̄) as shown next. Suppose limk→∞ $k → $∗ ,
then $̃k → $∗ and $∗ = (Id+Φ−1 B̄)−1 (Id−Φ−1 Ā)$∗ using the
(65)
In fact, for differential inclusion (65) consider the following
type of implicit/explicit discretization,
$(kh) − $(kh − h)
$(kh + h) − 2$(kh) + $(kh − h)
+ α̃
∈
2
h
h
− Φ−1 Ā($̃(kh)) − Φ−1 B̄($(kh + h)),
(66)
14
continuity of the right hand of (64). Therefore, $∗ ∈ zer(Ā+ B̄)
because Φ is a positive definite matrix.
The following relationship (similar with the cosine rule) will
be heavily utilized in the convergence analysis.
||a −
c||2Q
− ||b −
c||2Q
= 2ha − b, a − ciQ − ||a −
b||2Q ,
Using (76) for the second term on the right hand of (72)
yields
||$k − $∗ ||2Φ − ||$k+1 − $∗ ||2Φ
≥ ||$k − $k+1 ||2Φ + 2hĀ($̃k ) − Ā($∗ ), $k+1 − $∗ i
−2αh$k − $k−1 , $k+1 − $∗ iΦ
(70)
By Lemma 5.4, Ā is β−cocoercive. For the second term on
the right hand of (77), we have
which can be verified by directly expanding with ||a + b||2Q =
||a||2Q + 2ha, biQ + ||b||2Q .
The proof is divided into three parts:
• Part 1: Given any $∗ ∈ zer(Ā + B̄), ||$k+1 − $∗ ||2Φ − ||$k −
$∗ ||2Φ follows a recursive inequality as follows,
||$k+1 − $∗ ||2Φ − ||$k − $∗ ||2Φ
≤ α(||$k − $∗ ||2Φ − ||$k−1 − $∗ ||2Φ ) + (α + α2 )||$k − $k−1 ||2Φ
(71)
P
2
• Part 2: Given any $∗ ∈ zer(Ā+ B̄), ∞
k ||$k+1 −$k ||Φ < ∞
and limk→∞ $k+1 − $k = 0.
or equivalently we have
||$k+1 − $∗ ||2Φ − ||$k − $∗ ||2Φ ≤ −||$k − $k+1 ||2Φ
1
||$k+1 − $̃k ||22 + 2αh$k − $k−1 , $k+1 − $∗ iΦ .
+ 2β
Part 1: Given any point $∗ ∈ zer(Ā + B̄), we first prove a
recursive inequality (71) for ||$k+1 − $∗ ||2Φ − ||$k − $∗ ||2Φ .
Using (70) to expand the left hand of (71) yields,
α(||$k − $∗ ||2Φ − ||$k−1 − $∗ ||2Φ )
= α(2h$k − $k−1 , $k − $∗ iΦ − ||$k − $k−1 ||2Φ )
(80)
Combining (79) and (80), we have,
||$k+1 − $∗ ||2Φ − ||$k − $∗ ||2Φ
−α(||$k − $∗ ||2Φ − ||$k−1 − $∗ ||2Φ )
1
≤ −||$k − $k+1 ||2Φ + 2β
||$k+1 − $̃k ||22
+2αh$k − $k−1 , $k+1 − $∗ iΦ
−α(2h$k − $k−1 , $k − $∗ iΦ − ||$k − $k−1 ||2Φ )
1
≤ −||$k − $k+1 ||2Φ + 2β
||$k+1 − $̃k ||22
+2αh$k − $k−1 , $k+1 − $k iΦ + α||$k − $k−1 ||2Φ
(81)
Next using (63) and (70) for the first and third terms on the
right hand of (81) yields
(73)
Because $∗ ∈ zer(Ā + B̄), we also have
2αh$k − $k−1 , $k+1 − $k iΦ − ||$k − $k+1 ||2Φ
= 2αh α1 ($̃k − $k ), $k+1 − $k iΦ − ||$k − $k+1 ||2Φ
= 2h$̃k − $k , $k+1 − $k iΦ − ||$k − $k+1 ||2Φ
= ||$k − $̃k ||2Φ − ||$k+1 − $̃k ||Φ
= α2 ||$k − $k−1 ||2Φ − ||$k+1 − $̃k ||2Φ
(74)
Due to the maximal monotonicity of B̄ proved in Lemma 5.4,
hu − v, $k+1 − $∗ i ≥ 0, ∀ u ∈ B̄($k+1 ), v ∈ B̄($∗ )
(79)
Utilizing the equality (70) we also have,
||$k − $∗ ||2Φ − ||$k+1 − $∗ ||2Φ
= 2h$k − $k+1 , $k − $∗ iΦ − ||$k − $k+1 ||2Φ
= 2h$k − $k+1 , $k − $k+1 + $k+1 − $∗ iΦ − ||$k − $k+1 ||2Φ
= ||$k − $k+1 ||2Φ + 2h$k − $k+1 , $k+1 − $∗ iΦ
= ||$k − $k+1 ||2Φ + 2h$̃k − $k+1 , $k+1 − $∗ iΦ
−2αh$k − $k−1 , $k+1 − $∗ iΦ
(72)
where the last step is derived by incorporating (63).
To tackle the second term on the right hand of (72), we
proceed as follows. By (64) we have, $̃k − Φ−1 Ā($̃k ) ∈
$k+1 + Φ−1 B̄($k+1 ), or equivalently,
− Ā($∗ ) ∈ B̄($∗ ).
hĀ($̃k ) − Ā($∗ ), $k+1 − $∗ i
= hĀ($̃k ) − Ā($∗ ), $k+1 − √
$̃k + $̃k − $∗ i
≥ β||Ā($̃k ) − Ā($∗ )||22 + 2h β(Ā($̃k ) − Ā($∗ )), 2 √1 β ($k+1 − $̃k )i
1
≥ β||Ā($̃k ) − Ā($∗ )||22 − β||Ā($̃k ) − Ā($∗ )||22 − 4β
||$k+1 − $̃k ||22
1
2
= − 4β ||$k+1 − $̃k ||2
(78)
where the first inequality is obtained by the cocoercive property
(39) and the second inequality is obtained by 2ha, bi ≥ −||a||22 −
||b||22 .
Combining (77) with (78) we have
||$k − $∗ ||2Φ − ||$k+1 − $∗ ||2Φ ≥ ||$k − $k+1 ||2Φ
1
||$k+1 − $̃k ||22 − 2αh$k − $k−1 , $k+1 − $∗ iΦ ,
− 2β
• Part 3: We first show the convergence of {||$k − $∗ ||2Φ }
given any $∗ ∈ zer(Ā+ B̄), and then show the convergence
of Algorithm 6.1.
Φ($̃k − $k+1 ) − Ā($̃k ) ∈ B̄($k+1 ).
(77)
(75)
(82)
1
For the second term on the right hand of (81), 2β
||$k+1 −
1
2
2
$̃k ||2 ≤ 2βδ ||$k+1 − $̃k ||Φ , since Φ > δIn+2mN . Incorporating
this and (82) into (81)
By incorporating (73) and (74) into (75), we have
hΦ($̃k − $k+1 ) − Ā($̃k ) + Ā($∗ ), $k+1 − $∗ i ≥ 0
||$k+1 − $∗ ||2Φ − ||$k − $∗ ||2Φ
−α(||$k − $∗ ||2Φ − ||$k−1 − $∗ ||2Φ )
1
≤ −||$k+1 − $̃k ||2Φ + 2β
||$k+1 − $̃k ||22 + (α + α2 )||$k − $k−1 ||Φ
1
≤ −(1 − 2δβ
)||$k+1 − $̃k ||2Φ + (α + α2 )||$k − $k−1 ||2Φ
(83)
or equivalently,
h$̃k − $k+1 , $k+1 − $∗ iΦ − hĀ($̃k ) − Ā($∗ ), $k+1 − $∗ i ≥ 0.
(76)
15
Since δ >
1
2β ,
we derive (71).
P
2
Part 2: In this step, we will prove ∞
k ||$k+1 − $k ||Φ < ∞ and
limk→∞ $k+1 − $k = 0.
1
Denote S = Φ − 2β
In+2mN . Then S is symmetric and positive
1
definite since Φ ≥ δIn+2mN and δ > 2β
. The first inequality of
(83) can also be written as:
||$k+1 − $∗ ||2Φ − ||$k − $∗ ||2Φ − α(||$k − $∗ ||2Φ − ||$k−1 − $∗ ||2Φ )
≤ −||$k+1 − $̃k ||2S + (α + α2 )||$k − $k−1 ||2Φ .
(84)
For the first term on the right hand of (84),
−||$k+1 − $̃k ||2S
= 2h$k − $k+1 , $k − $̃k iS − ||$k − $k+1 ||2S − ||$k − $̃k ||2S
= −2h$k − $k+1 , α($k − $k−1 )iS
−||$k − $k+1 ||2S − α2 ||$k − $k−1 ||2S
≤ α||$k − $k+1 ||2S + α||$k − $k−1 ||2S
−||$k − $k+1 ||2S − α2 ||$k − $k−1 ||2S
= (α − 1)||$k − $k+1 ||2S + (α − α2 )||$k − $k−1 ||2S .
(85)
where the first equality follows from (70), the second equality
follows from (63), and the third inequality follows from
−2hx, yi ≤ ||x||2 + ||y||2 .
Denote Q = 2Φ − 1−α
2β In+2mN . Then Q is also symmetric and
positive definite, since α < 1 and Φ ≥ δIn+2mN . Combining (84)
with (85),
k
Therefore, limk→∞ $k+1 − $k = 0, and
$k−1 ||2Φ < ∞.
P∞
k=1 (α
+ α2 )||$k −
Part 3: In this part, we first show the convergence of
{||$k − $∗ ||2Φ } given any $∗ ∈ zer(Ā + B̄), and then show the
convergence of Algorithm 6.1.
Denote φk = max{0, ||$k − $∗ ||2Φ − ||$k−1 − $∗ ||2Φ } and ψk =
(α + α2 )||$k − $k−1 ||2Φ , and recall (71), we have φk+1 ≤ αφk + ψk .
Apply this relationship recursively,
φk+1 ≤ αk φ1 +
k−1
X
αi ψk−i .
(89)
i=0
Summing (89) from k = 1 to k = J,
P J Pk−1 i
PJ
1−α J
k=1 φk+1 ≤ φ1 1−α + k=1 i=0 α ψk−i
P J−1 i P J
J
≤ φ1 1−α
k=i+1 ψk−i
1−α + i=0 α
||$k+1 − $∗ ||2Φ − ||$k − $∗ ||2Φ
−α(||$k − $∗ ||2Φ − ||$k−1 − $∗ ||2Φ )
≤ (α − 1)||$k − $k+1 ||2S + (α − α2 )||$k − $k−1 ||2S
+(α + α2 )||$k − $k−1 ||2Φ
= (α − 1)||$k − $k+1 ||2S
1
+αh[(1 − α)(Φ − 2β
I) + (1 + α)Φ]$k − $k−1 , $k − $k−1 i
= (α − 1)||$k − $k+1 ||2S + α||$k − $k−1 ||2Q .
(86)
Denote µk = ||$k − $∗ ||2Φ − α||$k−1 − $∗ ||2Φ + α||$k − $k−1 ||2Q ,
then
µk+1 − µk = ||$k+1 − $∗ ||2Φ − α||$k − $∗ ||2Φ + α||$k+1 − $k ||2Q
−||$k − $∗ ||2Φ + α||$k−1 − $∗ ||2Φ − α||$k − $k−1 ||2Q
= ||$k+1 − $∗ ||2Φ − ||$k − $∗ ||2Φ − α(||$k − $∗ ||2Φ
−||$k−1 − $∗ ||2Φ ) + α||$k+1 − $k ||2Q − α||$k − $k−1 ||2Q
≤ (α − 1)||$k − $k+1 ||2S + α||$k − $k−1 ||2Q
+α||$k+1 − $k ||2Q − α||$k − $k−1 ||2Q
= (α − 1)||$k − $k+1 ||2S + α||$k+1 − $k ||2Q
1
= h[(α − 1)(Φ − 2β
I) + α(2Φ − 1−α
2β I)]$k+1 − $k ,
$k+1 − $k i
(87)
where the third inequality follows by (86).
Given a sufficient small 0 < < 1, choose 0 < α < 1 and
1
δ > 2β
such that 2βδ(1 − 3α − ) ≥ (1 − α)2 , then
1
1−α
2β In+2mN ) − α(2Φ − 2β In+2mN )
2
− (1−α)
2β In+2mN ≥ Φ
−(α − 1)(Φ −
= (1 − 3α)Φ
α||$k−1 −$∗ ||2Φ ≤ ||$k −$∗ ||2Φ −α||$k−1 −$∗ ||2Φ +α||$k −$k−1 ||2Q ≤
µ1 . Therefore, ||$k − $∗ ||2Φ ≤ α||$k−1 − $∗ ||2Φ + µ1 . By Lemma
1 at Page 44 of Polyak (1987), ||$k − $∗ ||2Φ ≤ αk (||$1 − $∗ ||2Φ −
µ1
µ1
1−a ) + 1−α , and {$k } is bounded sequence.
P
We also have µk+1 − µ1 ≤ − ki=1 ||$i+1 − $i ||2Φ . Then
P
ki=1 ||$i+1 − $i ||2Φ ≤ µ1 − µk+1 ≤ µ1 + α||$k − $∗ ||2Φ ≤
αµ1
µ1
) + 1−α
. Let k goes to infinity,
µ1 + αk+1 (||$1 − $∗ ||2Φ − 1−a
we have,
∞
X
||$k+1 − $k ||2Φ < ∞
(88)
Therefore, (87) yields µk+1 − µk ≤ −||$k+1 − $k ||2Φ .
Therefore, µk+1 ≤ µk ≤ µ1 . By the definition of µk , ||$k −$∗ ||2Φ −
(90)
Let J → ∞, then since 0 < α < 1,
P∞
P
φ1
i P∞
+ ∞
≤ 1−α
i=0 α Pk=i+1 ψk−i
k=1 φk
P
φ1
i ∞
(91)
+ ∞
≤ 1−α
t=1 ψt
i=0 α
φ1
1 P∞
≤ 1−α + 1−α t=1 ψt
P
P∞
2
2
Noticing that ∞
t=1 ψt =
k=1 (α + α )||$k − $k−1 ||Φ < ∞,
P∞
Pk
k=1 φk < ∞, and hence the sequence { i=1 φi }, being a
nonnegative and non-decreasing sequence, converges and is
bounded.
P
Consider another sequence {||$k − $∗ ||2Φ − ki=1 φi }. Since
P
||$k − $∗ ||2Φ is nonnegative and { ki=1 φi } is bounded, {||$k −
P
$∗ ||2Φ − ki=1 φi } is bounded from below. Furthermore, {||$k −
P
$∗ ||2Φ − ki=1 φi } is a non-increasing sequence. In fact,
P
||$k+1 − $∗ ||2Φ − k+1
i=1 φiP
= ||$k+1 − $∗ ||2Φ − φk+1 − ki=1 φi
P
≤ ||$k+1 − $∗ ||2Φ − ||$k+1 − $∗ ||2Φ + ||$k − $∗ ||2Φ − ki=1 φi
P
= ||$k − $∗ ||2Φ − ki=1 φi
where the second inequality follows from the definition of φk ,
P
φk+1 ≥ ||$k+1 − $∗ ||2Φ − ||$k − $∗ ||2Φ . As {||$k − $∗ ||2Φ − ki=1 φi }
is a non-increasing sequence and bounded from below, {||$k −
P
$∗ ||2Φ − ki=1 φi } converges.
Therefore, {||$k − $∗ ||2Φ }, being the sum of two convergent
P
P
sequences {||$k − $∗ ||2Φ − ki=1 φi } and { ki=1 φi }, also converges.
We are ready to show the convergence of Algorithm 6.1 using
the results in Part 1 and Part 2. Since {$k } is a bounded
16
sequence, it has a convergent subsequence {$nk } that converges
to $̃∗ . Because limk→∞ $k+1 − $k = 0 by Part 2, we have
limk→∞ $nk −1 − $nk = 0 and limk→∞ $nk +1 − $nk = 0. Pass
to limiting point of {$nk }, then we have $̃∗ = T 1 T 2 $̃∗ because
the righthand side of (63)-(64) is continuous. Hence, $̃∗ ∈
zer(Ā + B̄). Taking $∗ = $̃∗ in (71) of Part 1, we also have
{||$k − $̃∗ ||2Φ } converges by Part 3. Because a subsequence
{||$nk − $̃∗ ||2Φ } converges to zero, the whole sequence {||$k −
$̃∗ ||2Φ } converges to zero. Therefore, the whole sequence of
{$k } converges to $̃∗ ∈ zer(Ā + B̄). Resorting to Theorem
4.5 gives the desired result.
2
total product supply to all the markets given the action profile x
of all the companies. Market M j has a maximal capacity of
r j > 0, therefore, it should be satisfied that Ax ≤ r where
r = col(r1 , · · · , rm ) ∈ Rm . Suppose that P : Rm → Rm
is a price vector function that maps the total supply of each
market to the corresponding market’s price. Each company
has also a local production cost function ci (xi ) : Ωi → R.
Then the local objective function of company (player) Fi is
fi (xi , x−i ) : ci (xi ) − PT (Ax)Ai xi .
Overall, in this network Cournot game, each company needs
to solve the following optimization problem given the other
companies’ profile x−i ,
Remark 6.4. A sufficient and simple choice of parameters to
ensure the conditions in Theorem (6.3) is δ = β1 , = α and
√
1
0 < α < 10 − 3. In fact, δ > 2β
implies that 2βδ could be
be any real number % > 1. If we take = ςα, ς > 0, then the
quadratic inequality becomes α2 − (2 − 3% − ς%)α + 1 − % < 0.
Since 1 − % < 0, α2 − (2 − 3% − ς%)α + 1 − % takes value strictly
less than zero when α takes 0. By the continuity of quadratic
equation, there always exists 0 < α < 1 that ensures the above
quadratic inequality given any % > 1 and ς > 0.
min xi ∈Ωi
s.t.
ci (xi ) − PT (Ax)Ai xi
P
Ai xi ≤ r − Nj=1, j,i A j x j
(92)
Obviously, the above network Cournot game in (92) is a
particular problem of game in (7). Some practical decision
problems in engineering networks can be well described by the
network Cournot game in (92), such as the rate control game in
communication network (Yin, Shanbhag & Mehta (2011)) and
the demand response game in smart grids (Ye & Hu (2016)).
Example 7.1 (Rate control game). Consider a group of
source-destination pairs (nodes) in a communication
network, that is {S 1 , ..., S N }, to decide their data rates in
a non-cooperative setting. The data is transferred through a
group of communication links (channels), that is {L1 , ..., Lm },
and each link L j has a maximal data rate capacity of c j > 0.
Assume that an additional layer has decided the routine table
for each source-destination pair S i , which is encoded by
Ai ∈ Rm×ni . Each column of Ai has only one element being 1
and all the other elements being zero, and the k-th element of
column j is 1 if S i utilizes the link Lk and transfers data rate
[xi ] j on link Lk . The local decision variable of S i is the data
rate on each link that it utilizes, denoted by xi ∈ Rni . xi also
PN
has a local feasibility constraint xi ∈ Ωi . Denote n = i=1
ni ,
x = col(x1 , · · · , xN ) ∈ Rn , A = [A1 , · · · , AN ] ∈ Rm×n and
c = col(c1 , · · · , cm ) ∈ Rm . The total data rate on each
link should be less than the capacity of that link: Ax ≤ c.
Given the data rate profile of all the nodes x, the payoff
function of S i , Ji (xi , x−i ) : Rn → R, takes the form as
Ji (xi , x−i ) = −ui (xi ) + DT (Ax)Ai xi , where ui (xi ) : Ωi → R is the
utility of source S i , and D : Rm → Rm is a delay function that
maps the total data rate on each link to the unit delay of that
link. Thereby, the data rate control game can be well described
by the network Cournot game in (92).
7. Network Cournot game and simulation studies
There are various practical problems that can be well
modeled by the game in (7), such as the river basin pollution
game in Krawczyk & Uryasev (2000), the power market
competition in Contreras, Klusch, Krawczyk (2004), plug-in
electric vehicles charging management in Paccagnan, Gentile,
Parise, Kamgarpour & Lygeros (2016), and communication
network congestion game in Yin, Shanbhag & Mehta (2011).
All above examples can be regarded as the type of the network
Cournot game described below, which is a generalization of the
network Cournot competition in Bimpikis, Ehsani, & Ilkilic
(2014) by introducing additional market capacity constraints or
equivalently globally shared coupling affine constraints. This
type of network Cournot game with affine coupling constraints
also appeared in the numerical studies of Yu, van der Schaar &
Sayed (2016).
7.1. Network Cournot game
Suppose that there are N companies (players) with labels
F1 , · · · , F N and m markets with labels M1 , · · · , Mm . Company
Fi decides its strategy to participate in the competition in ni
markets by producing and delivering xi ∈ Rni amounts of
products to the markets it connects with. The production
limitation of company Fi is xi ∈ Ωi ⊂ Rni . Company Fi has
a local matrix Ai ∈ Rm×ni that specifies which market it will
participate in. The j-th column of Ai , that is [Ai ]: j , has only
one element being 1 and all other elements being 0, and [Ai ]: j
has its k-th element being 1 if and only if player Fi delivers
[xi ] j amount of production to the market Mk . Therefore,
matrices A1 , · · · , AN can be used to specify a bipartite graph
that represents the connections between the companies and the
PN
markets. Denote n = i=1
ni , x = col(x1 , · · · , xN ) ∈ Rn , and
PN
m×n
A = [A1 , · · · , AN ] ∈ R . Then Ax ∈ Rm = i=1
Ai xi is just the
Example 7.2 (Demand response game). Given a distribution
network in power grids, suppose that there are T time periods,
and each period has a desirable minimal total load shedding
di > 0. Suppose that there are N load managers (energy
management units or players) in the network, and each load
manager i can decide a local vector xi ∈ Rti as its local
load shedding vector in some specific time periods. Each load
manager i also has a local matrix Ai ∈ RT ×ti that specifies
which time period player i will participate. For j−th column
of Ai , it has one element being 1 while all other elements
17
F5
being zero. The k−th element of the j−th column of Ai is 1
if load manager i decides to decrease its load by [xi ] j at time k.
PN
Denote x = col(x1 , · · · , xN ), and A = [A1 , · · · , AN ] ∈ RT × i=1 ti ,
d = col(d1 , · · · , dT ). Naturally, it is required that the total load
shedding of all the load managers should meet the minimal
value, Ax ≥ d. Each player has a local feasible constraint
xi ∈ Ωi ⊂ Rti , and a cost function ci (xi ) : Ωi → R due
to local load shedding. P : RT → RT is the payment price
vector function that maps total load shedding of each period to
the payment price vector, therefore, PT (Ax)Ai xi is the payment
awards of player i for its load shedding. The disutility function
of player i is Ji (xi , x−i ) = ci (xi ) − PT (Ax)Ai xi given all the
players’ action profile x. All in all, the demand response
management game is well described by the network Cournot
game model in (92).
ci (xi ) − (P̄ − DAx)T Ai xi
P
ci (xi ) − P̄T Ai xi + ( Nj=1 DA j x j )T Ai xi
ci (xi ) + xiT ATi DAi xi − ATi P̄T xi
P
+ Nj=1, j,i xTj ATj DAi xi
AT1 DA2
2AT2 DA2
···
ATN DA2
···
···
···
···
AT1 DAN
AT2 DAN
···
2ATN DAN
F20
M5 F12
F8
M3
F13 F16
F9
F4
Figure 1: Network cournot game: An edge from Fi to M j on this graph implies
that Fi participates in the competition in M j .
F5
F14
F10
F1
F17
F18
F2
F15
F7
F6
F3
F4
F19
F12
F8
F9
F11
F20
F16
F13
Figure 2: Interference graph G f : An undirected edge from Fi to F j on this
graph means that the objective function of Fi depends on x j , and vice versa.
Company Fi is assumed to be able to observe the decisions of its neighbors on
this graph .
(93)
F(x) is strongly monotone24 and Lipschitz continuous, and
Assumptions 1 and 2 are satisfied.
Remark 7.3. Notice that in the network Cournot game of (92),
each player i’s local objective function ci (xi ) − PT (Ax)Ai xi
only depends on the decisions of the players that participate
the same markets as player i. Denote Z[a, b] as the set of
integers from a to b. Mathematically, j ∈ Nif if and only if
∃k ∈ Z[1, m], p ∈ Z[1, ni ] and q ∈ Z[1, n j ] such that [Ai ]kp = 1
and [A j ]kq = 1. Since A = [A1 , · · · , AN ] is usually a sparse
matrix, the interference graph G f of network Cournot game
also has sparse edge connections.
,
(94)
7.2. Simulation studies
In the studies, we adopt a similar simulation setting as
Yu, van der Schaar & Sayed (2016) without considering the
stochastic factors. Consider 20 companies and 7 markets, and
the connection relationship between the companies and the
markets is depicted in Figure 1. If there is an edge from Fi to
M j in Figure 1, then company Fi participates the competition in
Notice that Q can be written as
Q = diag{AT1 DA1 , · · · , ATN DAN } + S T S
F19
M7
M4
F3
F18
M6
F15
F7
F6
then
F(x) = ∇c(x) + Qx − AT P̄.
F11
M2
Denote F(x) = col(∇ x1 J1 (x1 , x−1 ), · · · , ∇ xN JN (xN , x−N )),
∇c(x) = col(∇c1 (x1 ), ∇c2 (x2 ), · · · , ∇cN (xN )) and Q ∈ Rn×n , and
2AT1 DA1
AT DA
1
Q = 2
···
ATN DA1
F14 F
17
M1
F2
Moreover, the Assumptions 1 and 2 for Algorithm 3.5 and
6.1 can easily be satisfied for many practical cost functions and
price functions. For example, take company Fi ’s production
cost function to be a strongly convex function with Lipschitz
continuous gradients (A quadratic function ci (xi ) = xiT Qi xi +
bTi xi with Qi ∈ Rni ×ni being a symmetric and positive definite
matrix and bi ∈ Rni is one possible choice). The price of
market M j is taken as the linear function of the total supplying
p j (x) = P̄ j − d j [Ax] j (known as a linear inverse demand
function in economics) with P̄ j > 0, d j > 0. Denote P =
col(p1 , · · · , pm ) : Rn → Rm , P̄ = col(P̄1 , · · · , P̄m ) ∈ Rm ,
D = diag{d1 , · · · , dm } ∈ Rm×m . Then P = P̄ − DAx is the
vector price function. The payments of company Fi by selling
product xi to the markets that it connects with is just PT Ai xi .
Therefore, the objective function of company Fi is,
Ji (xi , x−i ) =
=
=
F10
F1
(95)
where
S is
√ a block matrix
√ defined√ as S √ =
√
[ DA1 , · · · , DAN ] ∈ Rm×n and D = diag{ d1 , · · · , dm }.
Therefore,
Q
is
positive
semi-definite
matrix.
Hence, the Jacobian matric of F(x), JF(x)
=
diag{∇2 c1 (x1 ), ∇2 c2 (x2 ), · · · , ∇2 cN (xN )} + Q is postive definite
since the cost functions ci (xi ) are strongly convex. Therefore,
24 Proposition
18
2.3.2 of Facchinei and Pang (2007)
F1
F2
F3
F4
F5
F6
F7
F20
F8
F19
F9
F18
F10
F17
F16
F15
F14
F13
F12
F11
Figure 3: Multiplier graph Gλ : Company Fi and F j are able to exchange their
local {λi , zi } and {λ j , z j } if there exists an edge between them on this graph.
||x −x∗ ||
||$ −$∗ ||
2
2
k
k
× 100% and ||$
× 100%
Figure 5:
The trajectories of ||x
∗ ||
∗ ||
2
2
generated with Algorithm 3.5 and 6.1: This also shows the convergence of
both algorithms, and the superior convergence speed of Algorithm 6.1.
Figure 4: The trajectories of ||xk+1 − xk ||2 and ||$k+1 − $k ||2 generated with
Algorithm 3.5 and 6.1: This shows the convergence of both algorithms, and the
superior convergence speed of the algorithm with inertia 6.1.
market M j by producing and delivering products to market M j .
Each company Fi has a local constraint as 0 < xi < Θi where
each component of Θi is randomly drawn from (10, 25). Each
market M j has a maximal capacity of r j > 0, j = 1, · · · , 7 and r j
is randomly drawn from (20, 80). The local objective function
is taken as (93). The local cost function of company i is ci (xi ) =
Pi
πi ( nj=1
[xi ] j )2 + bTi xi which is a strongly convex function with
Lipschitz continuous gradient. Here πi is randomly drawn from
(1, 8), and each component of bi is randomly drawn from (1, 4).
The price function is taken as the linear function P = P̄ − DAx,
and P̄ j and d j are randomly drawn from (250, 500) and (1, 5),
respectively.
With Figure 1 and the definition of objective function in (93),
the interference graph G f can be easily obtained and is depicted
in Figure 2. Meanwhile, we adopte the multiplier graph Gλ
shown in Figure 3. The weighted adjacency matrix W = [wi j ]
of multiplier graph Gλ has all its nonzero elements to be 1.
With Figure 1 and the definition of objective function in (93),
the interference graph G f can be easily obtained and is depicted
in Figure 2. Meanwhile, we adopte the multiplier graph Gλ
shown in Figure 3. The weighted adjacency matrix W = [wi j ]
of multiplier graph Gλ has all its nonzero elements to be 1.
Set the step-sizes in Algorithm 3.5 as τi = 0.03, νi = 0.2, σ =
0.02 for all companies, and for Algorithm 6.1 set α = 0.12
while other step-sizes are the same as Algorithm 3.5. The initial
starting points xi,0 , λi,0 and zi,0 of both algorithms are set to be
zeros.
Figure 6: The trajectory of ||(L ⊗ I7 )λ̄k ||2 and [(1T20 ⊗ I7 )λ̄k ]T (Axk − r): The
left sub-figure shows that λ̄k converges to the null space of L ⊗ I7 , hence the
local multipliers of all companies reach consensus asymptotically. (1T20 ⊗ I7 )λ̄k
P
is the averaged vector 20
i=1 λi,k , therefore, the right sub-figure shows that the
complementary condition hλ∗ , Ax∗ − ri = 0 is asymptotically satisfied.
Figure 7: The trajectories of local decisions xi,k of F1 , F6 , F10 and F16 . The
solid red lines are generated with Algorithm 3.5 and dashed lines are generated
with Algorithm 6.1
19
extended for stochastic GNE seeking with noisy gradient
observations and noisy information sharing by resorting to the
stochastic forward-backward splitting algorithm in Rosasco,
Villa & Vũ (2016). The strong monotonicity assumption on the
pseudo-gradient might be relaxed to monotonicity assumption
by utilizing the forward-backward-forward splitting method in
Briceno-Arias & Combettes (2013).
Reference
References
Figure 8: The trajectories of some components of local multiplier [λi,k ] j where
solid red lines are generated with Algorithm 3.5 and dashed lines are generated
with Algorithm 6.1. [λi,k ] j is the j−th component of local multiplier λi at
iteration time k. Here we select local multipliers of F1 , F6 , F10 and F16
Alpcan, T. and Basar, T., 2005. Distributed algorithms for Nash equilibria
of flow control games. In Advances in dynamic games (pp. 473-498).
Birkhauser Boston.
Alvarez, F. and Attouch, H., 2001. An inertial proximal method for maximal
monotone operators via discretization of a nonlinear oscillator with
damping. Set-Valued Analysis, 9(1-2), pp.3-11.
Attouch, H., Chbani, Z., Peypouquet, J. and Redont, P., 2016. Fast convergence
of inertial dynamics and algorithms with asymptotic vanishing viscosity.
Mathematical Programming, pp.1-53.
Bauschke, H.H. and Combettes, P.L., 2011. Convex analysis and monotone
operator theory in Hilbert spaces. Springer Science & Business Media.
Berinde, V., 2007. Iterative approximation of fixed points. Berlin, Germany:
Springer.
Bimpikis, K., Ehsani, S. and Ilkilic, R., 2014. Cournot competition in
networked markets. In EC (p. 733).
Briceno-Arias, L.M. and Combettes, P.L., 2013. Monotone operator methods
for Nash equilibria in non-potential games. In Computational and Analytical
Mathematics (pp. 143-159). Springer New York.
Combettes, P.L. and Vũ, B.C., 2014. Variable metric forward-backward
splitting with applications to monotone inclusions in duality. Optimization,
63(9), pp.1289-1318.
Contreras, J., Klusch, M. and Krawczyk, J.B., 2004. Numerical solutions to
Nash-Cournot equilibria in coupled constraint electricity markets. IEEE
Transactions on Power Systems, 19(1), pp.195-206.
Condat, L., 2013. A primal-dual splitting method for convex optimization
involving Lipschitzian, proximable and linear composite terms. Journal of
Optimization Theory and Applications, 158(2), pp.460-479.
Debreu, G., 1952. A social equilibrium existence theorem. Proceedings of the
National Academy of Sciences, 38(10), pp.886-893.
Dreves, A. and Sudermann-Merx, N., 2016. Solving linear generalized Nash
equilibrium problems numerically. Optimization Methods and Software,
pp.1-28.
Facchinei, F. and Pang, J.S., 2007. Finite-dimensional variational inequalities
and complementarity problems. Springer Science & Business Media.
Facchinei, F. and Pang, J.S., 2010. Nash equilibria: the variational approach.
Convex optimization in signal processing and communications, p.443.
Facchinei, F. and Kanzow, C., 2010. Generalized Nash equilibrium problems.
Annals of Operations Research, 175(1), pp.177-211.
Facchinei, F., Fischer, A. and Piccialli, V., 2007. On generalized Nash games
and variational inequalities. Operations Research Letters, 35(2), pp.159-164.
Fischer, A., Herrich, M. and Schonefeld, K., 2014. Generalized Nash
equilibrium problems-recent advances and challenges. Pesquisa
Operacional, 34(3), pp.521-558.
Harker, P.T., 1991. Generalized Nash games and quasi-variational inequalities.
European journal of Operational research, 54(1), pp.81-94.
Iutzeler, F. and Hendrickx, J.M., 2016. A Generic Linear Rate Acceleration
of Optimization algorithms via Relaxation and Inertia. arXiv preprint
arXiv:1603.05398.
Kearns, M., Littman, M.L. and Singh, S., 2001. Graphical models for game
theory. In Proceedings of the Seventeenth conference on Uncertainty in
artificial intelligence (pp. 253-260). Morgan Kaufmann Publishers Inc..
Koshal, J., Nedić, A. and Shanbhag, U.V., 2016. Distributed Algorithms for
Aggregative Games on Graphs. Operations Research, 64(3), pp.680-704.
Krawczyk, J.B. and Uryasev, S., 2000. Relaxation algorithms to find
Nash equilibria with economic applications. Environmental Modeling &
Assessment, 5(1), pp.63-73.
The trajectories of selected algorithm performance
indexes,
∗
∗
k −$ ||
k −x ||
× 100%, ||$||$
×
including ||xk+1 − xk ||2 , ||$k+1 − $k ||2 , ||x||x
∗ ||
∗ ||
100%, ||L ⊗ I7 λ̄k ||2 and [(1T20 ⊗ I7 )λ̄k ]T (Axk − r), are shown in
Figures 4, 5 and 6. The trajectories of the local decisions xi,k of
some companies are shown in Figure 7, and the trajectories of
[λi,k ] j , which stands for the j-th component of local Lagrangian
multiplier λi,k , are shown in Figure 8.
8. Conclusions
In this paper, we proposed a primal-dual distributed
algorithm based on operator splitting methods for iterative
computation of a variational GNE in noncooperative games
with globally shared affine coupling constraints. The algorithm
is motivated by the forward-backward operator splitting method
for finding zeros of a sum of monotone operators. Each player
only needs to knows its local information, especially a block
of the affine coupling constraints. The proposed algorithm
is proved to converge with fixed step-sizes under some
mild assumptions by exploiting properties of composition of
averaged operators. Furthermore, a distributed algorithm with
inertia is also proposed and analyzed for possible acceleration
of convergence speed. Numerical simulation studies for
a network Cournot game demonstrate the efficiency of the
proposed algorithms and the superior convergence speed of the
inertial algorithm.
Many challenging and exciting topics are still open for
distributed NE/GNE seeking. Here we only list some problems
with probable solution hints. Finding all the generalized
Nash equilibria has its only interests, and this could be
partially solved by combining the design in this paper with
the parameterized variational inequality method in Nabetani,
Tseng, and Fukushima (2011). The algorithm requires that
each player is able to observe all its neighbors’ decisions
through the interference graph G f . This assumption could
be relaxed by adopting the local consensus dynamics in
Salehisadaghiani & Pavel (2016b), and then it could only
be required that the players were able to observe parts of its
neighbors’ decisions through a maximal triangle-free spanning
subgraph of G f . The methodology of this paper could be
20
Kulkarni, A.A. and Shanbhag, U.V., 2012. On the variational equilibrium
as a refinement of the generalized Nash equilibrium. Automatica, 48(1),
pp.45-55.
Lei, J., Chen, H.F. and Fang, H.T., 2016. Primal-Dual Algorithm for Distributed
Constrained Optimization. Systems & Control Letters, 96, 110-117.
Liang, S., Yi, P. and Hong, Y., 2016. Distributed Nash equilibrium
seeking for aggregative games with coupled constraints. arXiv preprint
arXiv:1609.02253.
Lorenz, D.A. and Pock, T., 2015. An inertial forward-backward algorithm for
monotone inclusions. Journal of Mathematical Imaging and Vision, 51(2),
pp.311-325.
Menache, I. and Ozdaglar, A., 2011. Network games: Theory, models, and
dynamics. Synthesis Lectures on Communication Networks, 4(1), pp.1-159.
Mesbahi, M. and Egerstedt, M. (2010). Graph Theoretic Methods for
Multiagent Networks, Princeton: Princeton University Press.
Meyer, C.D., 2000. Matrix analysis and applied linear algebra (Vol. 2). SIAM.
Nabetani, K., Tseng, P. and Fukushima, M., 2011. Parametrized variational
inequality approaches to generalized Nash equilibrium problems with
shared constraints. Computational Optimization and Applications, 48(3),
pp.423-452.
Nesterov, Y., 2013. Introductory lectures on convex optimization: A basic
course (Vol. 87). Springer Science & Business Media.
Paccagnan D., Gentile B., Parise F.,Kamgarpour M. and Lygeros J. 2016.
Distributed computation of generalized Nash equilibria in quadratic
aggregative games with affine coupling constraints. In 2016 55th IEEE
Conference on Decision and Control (CDC), (pp.6123-6128), IEEE.
Pan, Y. and Pavel, L., 2009. Games with coupled propagated constraints in
optical networks with multi-link topologies. Automatica, 45(4), pp.871-880.
Parise, F., Gentile, B., Grammatico, S. and Lygeros, J., 2015, December.
Network aggregative games: Distributed convergence to Nash equilibria.
In 2015 54th IEEE Conference on Decision and Control (CDC) (pp.
2295-2300). IEEE.
Pavel, L., 2007. An extension of duality to a game-theoretic framework.
Automatica, 43(2), pp.226-237.
Poljak B. T., 1987. Introduction to optimization. Optimization Software.
Rosasco, L., Villa, S. and Vũ, B.C., 2016. A stochastic inertial
forward-backward splitting algorithm for multivariate monotone inclusions.
Optimization, 65(6), pp.1293-1314.
Rosen, J.B., 1965. Existence and uniqueness of equilibrium points for concave
n-person games. Econometrica: Journal of the Econometric Society,
pp.520-534.
Salehisadaghiani, F. and Pavel, L., 2016a. Distributed Nash equilibrium
seeking: A gossip-based algorithm. Automatica, 72, pp.209-216.
Salehisadaghiani, F. and Pavel, L., 2016b. Distributed Nash Equilibrium
Seeking By Gossip in Games on Graphs. arXiv preprint arXiv:1610.01896.
Schiro, D.A., Pang, J.S. and Shanbhag, U.V., 2013. On the solution of affine
generalized Nash equilibrium problems with shared constraints by Lemkes
method. Mathematical Programming, 142(1-2), pp.1-46.
Scutari, G., Palomar, D.P., Facchinei, F. and Pang, J.S., 2010. Convex
optimization, game theory, and variational inequality theory. IEEE Signal
Processing Magazine, 27(3), pp.35-49.
Shi, W., Ling, Q., Wu, G. and Yin, W., 2015. A proximal gradient algorithm
for decentralized composite optimization. IEEE Transactions on Signal
Processing, 63(22), pp.6013-6023.
Spall, J.C., 1992. Multivariate stochastic approximation using a simultaneous
perturbation gradient approximation. IEEE transactions on Automatic
Control, 37(3), pp.332-341.
Swenson, B., Kar, S. and Xavier, J., 2015. Empirical centroid fictitious play: An
approach for distributed learning in multi-agent games. IEEE Transactions
on Signal Processing, 63(15), pp.3888-3901.
Wibisono, A., Wilson, A.C. and Jordan, M.I., 2016. A Variational Perspective
on Accelerated Methods in Optimization. Proceedings of the National
Academy of Sciences, 133, doi:10.1073/pnas.1614734113
Ye, M. and Hu, G., 2016. Game design and analysis for price-based demand
response: An aggregate game approach. IEEE Transactions on Cybernetics,
to appear.
Yi P, Hong Y, Liu F. 2016. Initialization-free Distributed Algorithms
for Optimal Resource Allocation with Feasibility Constraints and its
Application to Economic Dispatch of Power Systems. Automatica 74,
259-269.
Yin, H., Shanbhag, U.V. and Mehta, P.G., 2011. Nash equilibrium problems
with scaled congestion costs and shared constraints. IEEE Transactions on
Automatic Control, 56(7), pp.1702-1708.
Yu, C. K., van der Schaar, M. and Sayed, A. H., 2016. Distributed Learning
for Stochastic Generalized Nash Equilibrium Problems. arXiv preprint
arXiv:1608.00039.
Zeng X, Yi P, Hong Y, Xie L. 2016. Continuous-Time Distributed
Algorithms for Extended Monotropic Optimization Problems, arXiv
preprint arXiv:1608.01167, 2016
Zhu, M. and Frazzoli, E., 2016. Distributed robust adaptive equilibrium
computation for generalized convex games. Automatica, 63, pp.82-91.
21
| 3cs.SY
|
1
MIMO UWB Radar System with Compressive
Sensing
arXiv:1604.04291v1 [cs.IT] 14 Apr 2016
Xia Li, Student Member, IEEE, Zhen Hu, Robert C. Qiu, Fellow, IEEE,
Abstract—A multiple input multiple output ultra-wideband
cognitive radar based on compressive sensing is presented in this
letter. For traditional UWB radar, high sampling rate analog to
digital converter at the receiver is required to meet Shannon
theorem, which increases hardware complexity. In order to
bypass the bottleneck of ADC or further increase the radar
bandwidth using the latest wideband ADC, we propose to exploit
CS for signal reconstruction at the receiver of UWB radar for
the sparse targets in the surveillance area. Besides, the function
of narrowband interference cancellation is integrated into the
proposed MIMO UWB radar. The field demonstration proves
the feasibility and reliability of the proposed algorithm.
Target 2
Target 1
AWG
RX antenna
RX antenna
TX antenna
TX antenna
Index Terms—MIMO, UWB, radar, CS, NBI, testbed.
Transmitter
based on
FPGA
Trigger signal
Receiver
I. I NTRODUCTION
USB 2.0
In the radar system, the unprecedented radio bandwidth
provides advantages such as the high precision of range
estimation. However, the extremely high sampling rate of
analog to digital converter (ADC) becomes a major challenge
in the ultra-wideband (UWB) radar systems. According to
Nyquist sampling theorem, the sampling rate should be at least
twice the bandwidth of the signal. For better system performance, oversampling is often required and results even higher
sampling rate. Compressed sensing (CS) gives an opportunity
to overcome this challenge. CS [1] theory specifies a new
signal acquisition approach, in a high probability, allowing
the acquisition of signals at a much lower data rate than the
Nyquist sampling rate.
Some previous work of our group [2] applied CS to the
communication application. In the communication system, the
received waveforms are synchronized by some mechanism
like preamble. However, in the MIMO radar application, the
received signal cannot be considered as synchronized since
the arrival time of signal depends on the over the air delay,
which is determined by the target distance to the radar. The
unknown delay leads to the randomness of the measurement
matrix construction. A cyclic prefix (CP) is used in waveform
design to make sure the measurement matrix is constant and
unrelated to the time delay. Therefore, each receiver antenna
can construct the basis matrix locally. Based on the analysis
of the system, the architecture of the testbed is proposed. This
testbed is based on our previous results of the UWB time reversal experiment [3] and cognitive UWB radar experiment [4]–
[6]. Both of them are on the full sampling rate. However, the
new system satisfies the specialty of the compressed sensing
X. Li, Z. Hu, and R. C. Qiu are with the Department of Electrical and Computer Engineering, Center for Manufacturing Research, Tennessee Technological University, Cookeville, TN, 38505, e-mail: [email protected];
[email protected]; [email protected].
Gigabyte
Ethernet
MATLAB
computing
engine
Fig. 1.
Block diagram of the system architecture
structure. The sampling rate at receiver can be much lower
than the Nyquist rate used in the two previous systems.
The experiment results based on measured data are provided
to demonstrate the effectiveness of the algorithm. The performance curves of the algorithm under different NBI and SNR
conditions are provided in simulation.
II. BACKGROUND AND P ROBLEM F ORMULATION
Since the transmitter and receiver of the our system are
synchronized by a trigger signal as shown in Fig. 1, ToAs
between transmitter antennas and receiver antennas can be
estimated from the recovered channel information, which are
used to estimate the target positions [4].
In traditional channel estimation theory, the channel vector
h ∈ Cn can be estimated from the observations of the linear
convolution, Xh = y ∈ Cn ., where the observation matrix
X ∈ Cn×n is a partial Toeplitz matrix.
By using CS theory, a sparse, high-dimensional vector
h ∈ Cn can be accurately recovered from a small number of
linear observations, RΩ Xh = ȳ ∈ Cm . Consider an arbitrary
index set Ω ⊂ {0, 1, . . . , n − 1} whose cardinality |Ω| = m.
We define the operator RΩ ∈ Cm×n that restricts a vector to
the entries listed in Ω. In our testbed, the ȳ can be viewed as
the sub-sampling of y. A delay in ȳ means the change of RΩ ,
which means the measurement matrix RΩ X is different. Since
the delay is unknown to the receiver, channel information
cannot be recovered at receiver locally.
2
A cyclic extension of time length TG is used to change
the linear convolution to circulant convolution. If the channel
delay is within TG , the system can be described in matrix as
y = Xh + n, where X is a circulant matrix as
xn−1 xn−2 · · · x1
x0
x0
xn−1 . . . x2
x1
(1)
X= .
.
.
.
.. .
..
..
..
..
.
xn−2 xn−3 · · · x0 xn−1
The delay in dictionary nd can be represented by the
channel latency. This is proved by (2). It holds when the delay
nd is smaller than the length of CP.
yc [n − nd ]
=
=
=
=
N
−1
X
m=0
N
−1
X
of the received waveform with full sampling rate.
h11
h21
h12
h22
ȳ1 ȳ2 = X1 X2
| {z } | {z }
h1
h2
+ n̄1 n̄2
(5)
The corresponding partial circulant matrix is defined as X̄ =
RΩ X. The partial random circulant matrix is generated by a
Rademacher sequence. It is proved in [8], the sth restricted
isometry constant δs has
( 3
)
r
3
s
s2
E [δs ] 6 C1 max
log 2 n,
log s log n ,
(6)
m
m
h [m] x [hn − (nd + m)iN ]
where C1 > 0 is a universal constant.
(6) means that for a given δ ∈ (0, 1), we have E [δs ] 6 δ
provided by
o
n
3
3
(7)
m > C2 max δ −1 s 2 log 2 n, δ −2 slog2 nlog2 s ,
h [m0 − nd ] x [hn − m0 iN ]
where C2 > 0 is another universal constant.
For 0 6 λ 6 1 and C3 > 0, we can have
h [m] x [hn − nd − miN ]
m=0
N
−1
X
m0 =0
N
−1
X
h [m − nd ] x [hn − miN ]
P (δs > E [δs ] + λ) 6 e
(2)
m=0
The vector h is the sampled signal of the channel impulse
response h (t). h (t) P
can be treated as a time limited pulse train
of the form h (t) = αm δ (t − τm Ts ), where the amplitudes
2
m
If the delay τm is an integer, then all the energy from αm is
mapped to tap hτm . However, if τm is not an integer, its energy
will leak to all taps of hk . However, most of the energy is kept
in the neighborhood of the original pulse locations. When the
target number is small, the number of wireless paths is also
small. In this case, h can be considered as a sparse signal.
For a 2 × 2 MIMO system shown in Fig. 1, the system
equation can be described as
h11
h21
h12
h22
y1 y2 = X1 X2
| {z } | {z }
h1
h2
+ n1 n2 ,
(4)
where y1 and y2 are the received waveforms from receiver
antennas 1 and 2. For the transmitted waveforms x1 and x2 ,
X1 and X2 are the respective Toeplitz matrices. n1 and n2
are the i.i.d. complex Gaussian variables for the two receiver
antennas. hji is the impulse response of the channel between
transmitter antenna i to receiver antenna j.
For the CS-based radar system, the sub-sampled received
waveform can be considered as the result of down-sampling
,
(8)
s
C3 m
log2 slog2 n.
where σ =
In radar systems, the number of multipath is equivalent to
the number of targets. The channel information is sparse to
the most cases. The channel information can be recovered by
using the following convex optimization program
min (kh1 k1 + kh2 k1 )
Xh1 − ȳ1
subject to
Xh2 − ȳ2
αm of channel taps are complex valued. The samples of h are
modeled as [7]
π
sin (πτm )
1 X
. (3)
hk = √
αm e−j N (k+(N −1)τm )
π
n m
sin N
(τm − k)
−λ2
σ2
2
2
6 ε1 ,
6 ε2
(9)
where k·k1 is the `1 norm and k·k2 is the `2 norm. εj is
chosen such that it bounds the amount of noise in the measurement from jth receiver antenna. The convex optimization
problem (9) is a second-order cone problem and can be solved
efficiently by utilizing the Basis Pursuit Denoising (BPDN)
algorithm [9].
III. NBI CANCELLATION USING CS THEORY
The system is modified to achieve the robustness under
NBI. In (4), NBI signal is not taken into account in algorithm,
NBI is treated as noise. However, the sparsity of the NBI can
be used to recover its information from sub-sampled received
waveform.
ȳ1
ȳ2
=X
h1
h2
+
n̄1
n̄2
=X
h1
h2
+
n̄1
n̄2
+
j̄1
+
FH jf 1
j̄2
FH j f 2
(10)
where FH is a randomly chosen rows of the discrete Fourier
matrix and can be written as FH = RΩ FH . jj is a vector of
the NBI signal received at antenna j. jf j is a vector of the
NBI in frequency domain. A recent research [10] has proved
3
that this measurement matrix has small restricted isometry
constants with very high probability.
For the receiver antenna j, we can have
hj1
ȳj = X1 X2
+ F−1 jf j + n̄j
hj2
i hj1
h
(11)
=
FH hj2 +n̄j .
X1 X2
|
{z
}
jf j
| {z }
Ā
sj
Because both the channel information and the NBI interference information are sparse, the interference signal and
channel information can be recovered simultaneously by using
the following convex optimization program
min (ks1 k1 + ks2 k1 )
As1 − ȳ1
subject to
As2 − ȳ2
=
|
X1
X2
X1
X2
{z
X̄
hj1
hj2
2
2
6
6
,
(12)
+ FH jf j + n̄j − FH j̃fj
Local Oscillator 4.0 GHz
I
D/A Conveter
Dual- core
1Gsps/14- bit
Xilinx Virtex-5 FPGA
Q
Power
Amplifier
Quadrature Modulator
Block diagram of the system architecture
Quadrature Demodulator
400~6000 MHz
I
A/D Converter
DDR2
Memory
Low Noise RF Bandpass
Amplifier
filter
A/D Converter
Q
FPGA
Array
Local Oscillator
4.0 GHz
Host Computer
Xilinx
Chipscope Pro
Analyzer
I
DDR2
Memory
A/D Converter
Low Noise RF Bandpass
Amplifier
filter
Q
Quadrature Demodulator
400~6000 MHz
Fig. 3.
A/D Converter
4 Channels 3Gsps Digitizer with FPGA Array In NI PXIe Chassis
Block diagram of the system architecture
traditional system. Due to the CS algorithm, the sampling rate
of the received signal can be reduced to much lower than
the Nyquist rate. As an example, we set the receiver ADC
sampling rate 13 of the full signal bandwidth, which is 166.66
Msps, as shown in Fig. 3. The reduced sampling rate in the
receiver can greatly reduce the cost and power consumption of
the receiver. The analog RF down-converter is the same as the
traditional system. A bandpass filter is used to remove noise
outside of signal bandwidth.
V. E XPERIMENT R ESULTS AND P ERFORMANCE A NALYSIS
(13)
hj
min (kh1 k1 + kh2 k1 )
Xh1 − ȳ10
subject to
Xh2 − ȳ20
where
Power Splitter
1:2
Reference Clock 20 MHz
For Local Oscillator
Baseband
Processing
Fig. 2.
Power
Amplifier
Q
JTAG
The channel information can be recovered by using the
following convex optimization program
ε00j
1Gsps/8-bit
2 Cores both with I&Q
Host
computer
Quadrature Modulator
hj1
+w̄j + n̄j
} hj2
| {z }
Digital Arbitrary
Waveform Generators
Power Splitter
1:2
ε01
ε02
where s1 contains both the channel information and interference position in frequency domain. ε0j bounds the energy of
n̄j .
NBI signal is always sparse in frequency domain due to
its narrow band feature. The number of taps in channel
information equals the number of targets in the surveillance
area and suffers the influence of the environment like clutters.
It is more likely NBI has better sparsity structure than the
channel information. The precision of the recovery algorithm
relies on the signal sparsity. The NBI signal recovery should be
more accurate. Thus, the estimated interference signal j̃f can
be used to cancel the interference in the received waveform.
By subtracting the interference signal jf using the estimated
interference signal j̃f , the remaining signal w̄ is obtained.
ȳj0 =
I
D/A Conveter
Dual- core
1Gsps/14- bit
2
2
6 ε001 ,
6 ε002
(14)
bounds the energy of noise signal w̄j + n̄j .
IV. CS- BASED MIMO RADAR SYSTEM
Based on the analysis on the previous section, a CS-based
cognitive MIMO radar system is proposed. The transmitter
architecture is the same as our previous work [3], as shown
in Fig. 2. In the experiment, the carrier frequency is set to
4 GHz. The bandwidth of the radar waveform is 500 MHz,
making the system an ultra-wideband system.
To recover the original signal, the sampling rate of the
radar received waveform must be higher than 500 MHz in
With one target in surveillance area, the experiment result is
shown in Fig. 4. The starting time of recording received signal
is the starting time of waveform transmitting. The position of
radar waveform in the received signal can be detected by using
the correlation relationship of CP. The channel information
should have only one path in this situation. However, in real
system, the channel delay is not an integer in model (3). Fig. 4
(b) is the result by solving (14). Most energy is mapped to
tap 3 while the neighborhood of the tap holds the residue.
Since sparsity is considered, the values of all other taps are
close to zero. Fig. 4 (c) Interference information is obtained
by solving (11). All of the energy is mapped to tap 45.
In order to analyze the performance of the proposed algorithms, several simulations are conducted. For each point,
1 × 104 simulations are performed. Fig. 5 shows the effect of
single tone NBI to compressed sensing system when SNR =
20 dB. Rhombused curve is obtained by (9) while stared curve
is obtained by recovering channel information using (12). The
vertical axis is the probability of detection, which means the
probability when correct channel tap location is obtained. The
result shows when SIR goes below 15 dB, system performance
would be severely influenced by NBI.
4
20
100
0
50
0
1
0
0.9
-10
0.8
-20
0.7
-30
-40
0.6
Pd
Amplitude (dB)
Amplitude (dB)
Amplitude
-20
10
-40
0.5
-50
0.4
-60
With second step, single tone NBI
With only one step, single tone NBI
With second step, dual tone NBI
With only first step, dual tone NBI
0.3
-60
-50
-70
0.2
-80
0.1
-20
-80
-90
-100
0
4
2
4 x 10
Time index
-100
-10
SNR (dB)
-5
0
-100
0
20
40
Channel tap index
0
20
40
60
Frequency bin
Fig. 4. Experiment data: (a) Received sub-sampled baseband signal; (b)
Recovered channel information; (c) Recovered interference information
1
0.9
Fig. 6.
Performance curve of simulation
algorithm in real-world situation. Simulation results show that
the proposed algorithms are able to estimate the locations
of the jamming interference and provide a performance gain
compared to the system without NBI cancellation.
0.8
ACKNOWLEDGMENT
0.7
This work is funded by National Science Foundation
through two grants (ECCS-0901420 and ECCS-0821658), and
Office of Naval Research through two grants (N00010-10-10810 and N00014-11-1-0006). This work is partly funded by
an AFOSR subcontract though the prime contractor (RNET
Technologies, Inc.) on the CIRE (Center for Innovative Radar
Engineering) contract FA8650- 10-D-1750 with AFRL/RY.
Pd
0.6
0.5
0.4
0.3
Without NBI cancellation
With NBI cancellation
0.2
0.1
12
Fig. 5.
-15
12.5
13
13.5
14
SIR (dB)
14.5
15
15.5
16
The effect of NBI cancellation
With a strong interference (SIR = 0 dB), Fig. 6 gives system
performance under difference additive white Gaussian noise
(AWGN) levels. Three targets situation is considered. Both
single tone and dual tone NBIs results are shown. Two stared
curves are for (14). Two rhombused curves are for (12). With
the second step of channel estimation, the accuracy becomes
much higher. In the case of dual-tone, the NBI signal becomes
less sparse than the single tone case. Because the CS algorithm
highly relies on the sparsity of the signal, the performance is
worse than the single tone case.
VI. C ONCLUSION
In this letter, we have proposed a CS-based MIMO UWB
cognitive radar. An algorithm is designed to get the ToA information from the sub-sampled echoed radar waveform. The
algorithm is based on circulant convolution. It can estimate
the channel information. Based on the equations of the whole
system, a hardware architecture has been proposed to fit into
the special structure of the CS system. NBI signal in the
surveillance area is also considered. An algorithm to estimate
and cancel the NBI interference is proposed. The hardware
testbed is used to verify the effectiveness of the proposed
R EFERENCES
[1] E. Candès, J. Romberg, and T. Tao, “Robust uncertainty principles: Exact
signal reconstruction from highly incomplete frequency information,”
Information Theory, IEEE Transactions on, vol. 52, no. 2, pp. 489–509,
2006.
[2] P. Zhang, Z. Hu, R. Qiu, and B. Sadler, “A compressed sensing based
ultra-wideband communication system,” in Communications, 2009.
ICC’09. IEEE International Conference on. IEEE, 2009, pp. 1–5.
[3] Z. Chen, C. Zhang, F. Lin, J. Yu, X. Li, Y. Song, R. Ranganathan,
N. Guo, and R. C. Qiu, “Towards a large-scale cognitive radio network:
Testbed, intensive computing, frequency agility and security,” in Computing, Networking and Communications (ICNC), 2012 International
Conference on. IEEE, 2012, pp. 556–562.
[4] X. Li, Z. Hu, R. Qiu, N. Guo, G. Scalzi, and M. Wicks, “Waveform
diversity for closed-loop wide-band cognitive sensing system,” in IEEE
Waveform Diversity and Design Conference, vol. 2012, 2012.
[5] X. Li, Z. Hu, R. Qiu, Z. Wu, J. Browning, and M. Wicks, “Demonstration
of cognitive radar for target localization under interference,” Aerospace
and Electronic Systems, IEEE Transactions on, vol. 50, no. 4, pp. 2440–
2455, 2014.
[6] X. Li, Z. Hu, R. C. Qiu, and M. Wicks, “Experimental demonstration
of cognitive radar for target localization under strong interference,”
in MILITARY COMMUNICATIONS CONFERENCE, 2012-MILCOM
2012. IEEE, 2012, pp. 1–6.
[7] J. Van de Beek, O. Edfors, M. Sandell, S. Wilson, and P. Borjesson,
“On channel estimation in ofdm systems,” in Vehicular Technology
Conference, 1995 IEEE 45th, vol. 2. IEEE, 1995, pp. 815–819.
[8] H. Rauhut, J. Romberg, and J. A. Tropp, “Restricted isometries for partial random circulant matrices,” Applied and Computational Harmonic
Analysis, vol. 32, no. 2, pp. 242–254, 2012.
[9] S. S. Chen, D. L. Donoho, and M. A. Saunders, “Atomic decomposition
by basis pursuit,” SIAM review, vol. 43, no. 1, pp. 129–159, 2001.
[10] E. J. Candes, Y. C. Eldar, D. Needell, and P. Randall, “Compressed
sensing with coherent and redundant dictionaries,” Applied and Computational Harmonic Analysis, vol. 31, no. 1, pp. 59–73, 2011.
| 3cs.SY
|
Capacity Planning for Cluster Tools in the
Semiconductor Industry
Martin Romaucha,∗, Richard F. Hartla
a Department of Business Administration
University of Vienna, Oskar-Morgenstern-Platz 1, 1090 Wien
Abstract
This paper proposes a new model for Cluster-tools with two load locks. Clustertools are widely used to automate single wafer processing in semiconductor
industry. The load locks are the entry points into the vacuum of the Clustertool’s mainframe. Usually there are two of them available. Each lot being
processed, is dedicated to a single load-lock. Therefore at most two different
lots (with possibly different processing times and qualification) can be processed
simultaneously. This restriction is one of the major potential bottlenecks.
Capacity planning is one of the possible applications for the proposed model
and the paper demonstrates the integration into a more general framework that
considers different tool types and different operational modes.
The paper also generalizes an earlier model that is limited to three processing
chambers. The proposed modeling approach is based on makespan reductions
by parallel processing. It turns out that the performance of the new approach
is similar, when compared to the generalized model for three chambers, but the
new approach outperforms the generalized model for four and more chambers.
Keywords: Cluster-Tool , Capacity Planning, Semiconductor Industry, Linear
Programming , Duality
1. Introduction
This paper considers Cluster-Tools, a tool type that can be found in semiconductor fabrication front-end sites, also called wafer fabrication facilities. The
wafer is the substrate used for the fabrication of integrated circuits for semicon∗ Corresponding
author
Email addresses: [email protected] (Martin Romauch),
[email protected] (Richard F. Hartl)
Preprint submitted to —————–
3rd May 2016
2
Design,
Maskmaking
reticles
Wafermaking
wafer
starts
Epitaxy
Front-End
Process:
------Oxidation,
Deposition,
Diffusion
Photolitography
Photoresist
Exposure
Planarization
Etch
semiconductor
device
Test &
back-end
wafer
finishes
Cleaning
Inspection & Measurement
raw
silicon
Implantation
FRONT-END (WAFER FABRICATION)
Figure 1: front end fabrication as a re-entrant flow shop (based on Sorenson [24] and Mönch
et al. [19])
ductor devices. A wafer is a disc, usually a slice of mono-crystalline silicon with
a width between 150 and 300 mm, and a thickness between 0.5 and 1 mm.
The input of front-end facilities are containers of wafers (also called lots or
FOUPs, Front Opening Unified Pods). A lot usually consists of 20 or more
wafers that require the same processing. The fabrication is characterized by a
large number loops or layers (compare Hutcheson [15]). Each loop consists of
one or more processing steps; one of them is a lithographic process that allows
to cover parts of the wafer (photo resist) from the process that follows. Figure
1 gives a schematic view on wafer fabrication.
The flow diagram in Figure 1, is a representation of the fabrication process
as a re-entrant flow shop. According to Thiesse and Fleisch [25] the traversed
path of a single lot can be several kilometers long.
In semiconductor industry most of the equipment is automated and characterized by load/unload operations, robot handling, testing, alignment, cooling
and much more. Cluster Tools (compare Franssila [6]) for instance, have one or
more load locks and multiple processing chambers. Cluster-tools are used to automate several process types. The internal software defines the behavior of the
system. Cluster tools allow an automatic transfer of wafers between load ports
and process chambers with possibly different processes inside a vacuum. The
system can also be used for parallel processing to increase throughput and productivity. The wafers are transferred between chambers under a vacuum using a
robotic arm to prevent exposure to air to prevent oxidation and contamination.
Since a Cluster-tool may operate on different lots at the same time, the combination of lots is important and the Cluster-tool may show different operation
cycle times for the same type of lot. According to Mönch et al. [19] several researchers studied the optimization of internal scheduling, but the corresponding
models are not suitable to consider several Cluster-tools at the same time. One
3
L1
3
LL1
2
1
A
3
B
A
LL1
2
L1
B
1
L2
3
2
L2
1
3
2
LL2
LL2
C
(i) t = 0
3
LL1
A
1
C
(ii) t = 5
A
2
A
B
1
LL1
2
L1
B
L1
1
1
LL2
2
C
(iii) t = 6
L2
3
3
B
2
1
C
1
2
3
3
L2
3
LL1
1
LL2
2
C
(iv) t = 10
LL2
t = 1 2 3 4 5 6 7 8 9 10 11 12
Figure 2: A Cluster-tool with two load locks that is processing wafers in parallel. On left
hand side: four snapshots of the Cluster-tool for t = 0, 5, 6, 10 can be found. The thickness of
the outline of the wafers, indicates the processing progress. The corresponding Gantt-chart
can be found on the right hand side of the Figure.
of the major bottlenecks of a Cluster-tool are the load locks (see Christopher
[3]). The load locks are the entry point into the vacuum of the Cluster tool.
In order to describe a Cluster-Tool with respect to load locks and chambers, a
simplified schedule for a Cluster-tool that is working in parallel mode is discussed
in Example 1. The example is based on Figure 2.
Example 1. On the left hand side of Figure 2, a Cluster-tool that is working in
parallel mode with two load locks LL1 and LL2 is illustrated. In this example
the chambers are supposed to be identical. Later in the example, one of the
chambers will be less flexible than the others. In the beginning (t = 0), each
load lock is occupied by one lot (LL1 with lot L1 and LL2 with lot L2 ), and
each lot consists of three wafers. Each wafer needs to be processed in one of the
chambers. The load lock LL1 is occupied with the lot L1 that needs the longer
processing time. More precisely, each wafers from L1 need a processing time of
six time units. For L2 the processing time is five time units per wafer. In this
example the wafer handling time is neglected, therefore it is possible to start all
processes in each chamber at the same time.
In the first step (i) with t = 0, two wafers from lot L1 are assigned to
chambers A and B; and one wafer from L2 is assigned to chamber C. At t = 5
the wafer in chamber C is finished and it is passed back to L2 . Then, the second
wafer from L2 is assigned to chamber C. At t = 6 the wafers in A and B are
handed back to L1 and chamber A will be occupied by the last wafer of L1 ;
chamber B will be occupied by the last wafer of L2 .
The corresponding Gant-chart can be found on the right hand side of Figure
1
2. The make-span is 12 with an idle time percentage of 12
on average. If wafers
from L1 are excluded from chamber C then at least two wafers from L1 need to
be assigned to the same chamber (A or B), hence the schedule is optimal with
4
respect to the make-span.
Now, additionally suppose that the problem is scaled by thirty, and each lot
counts 90 wafers. Then the wafers can be distributed in such a way that the
idle time percentage vanishes. If this is possible, then the total processing time
is 90 · (6 + 5) = 990. Therefore it is sufficient to find a distribution where each
chamber finishes after 330 time units. Note that 330 is divisible by five and six:
A
B
C
}|
{ z }| {
z }| { z
330 = 55 · 6 = 35 · 6 + 24 · 5 = 66 · 5
Therefore 55 wafers from L1 can be assigned to chamber A and 66 wafers
of L2 can be assigned to chamber C. The remaining wafers can be assigned to
chamber B.
This example shows that on a larger scale where larger quantities are considered, a continuous approximation - where it is allowed to computationally
“split” wafers - leads to reasonable results.
LP based Cluster-tool models can be directly integrated in capacity models
or in master planning like proposed in Ponsignon and Mönch [22] and Romauch
and Klemmt [23]. The combination of simulation and optimization (LP) as
discussed in Almeder et al. [1] Gansterer et al. [8] and Juan et al. [16] is also
suitable for the integration of LP based Cluster-Tools models.
Finally, real-time dispatching that integrates LP approaches can be found
in Doleschal et al. [4] and Ham et al. [11] which can be extended to LP based
Cluster-tool models. According to Duemmler and Wohlleben [5] there are also
various alignments necessary to achieve an effective WIP Flow Management
that assures that the planned capacity consumption (utilization) is coherent
with the dispatching reality. Therefore, improvements of LP based Cluster-tool
models are important to several areas.
The focus of this paper is on static capacity planning, but it is important to
note that the LP models introduced in this paper may also be useful to other
applications, like master planning, simulation and dispatching.
In the context of capacity planning for front-end facilities, a concise prediction of the equipment utilization is one of the major tasks. For given demands
- that is wafer starts for given product routes - a calculation needs to confirm
that the capacity restrictions are met. Furthermore it is necessary to identify
hard/soft bottlenecks that are limiting the throughput.
Depending on the purpose and time frame, there are several methods to
tackle this capacity planning problem - sometimes elementary MRP calculations are sufficient, but the large majority is covered by simulation and linear
programming. This paper is dedicated to linear programming formulations and
proposes a Cluster Tools model for multiple chambers and two load locks that
is performant, flexible and easy to analyze.
The contribution of this paper is threefold. First, we present a new model
that considers more than three chambers. Second we derive the model presented
in Ortner [21] in a more general way that also allows to handle than three chambers. Third, computational experiments are used to compare the performance
of the proposed methods.
5
2. Capacity Planning in Semiconductor Industry
In order to show the applicability of the proposed Cluster-Tool model, the integration into a basic capacity planning model will be presented. The discussed
model is static, the processing sequence is not considered and the quantities of
wafers are continuous. Aspects like multiple time-frames can be understood as
an elementary extension (cf. Romauch and Klemmt [23]), therefore the discussion (and the computational experiments) is restricted to a basic model to prove
the concept and to avoid a cumbersome notation.
The discussion of input and output of the basic capacity planning model
starts with an example that is depicted in Figure 3. The input for the capacity
planning model is the weekly going rate dp for each product p ∈ P on the basis
of lots for a given fixed lot size. That means, dp wafers of product product
p need to be produced in the following week. Each product corresponds to a
product route that consists of processing steps. Some of the processing steps
can be considered equivalent for different product routes and these equivalent
processing steps constitute the so called job classes (j ∈ J). The capacity
consumption (e.g. service times) and flexibility for members of the same job
class is the same. Analogous to Figure 3 the primary demand dp on basis
products infers a secondary demand λj on basis job classes. For instance job
class 1 has a demand λ1 = 30 because this job class occurs once in the product
routes for product one and product two (λ1 = d1 + d2 ).
The wafers λj are distributed among the available machines, i.e. yji wafers
from job class j are assigned to machine
i. In total, for each job class j the
P
demand λj needs to be satisfied, i.e. i∈{a,b,c,d} yij = λj . The distribution that
corresponds to yji infers a load on the corresponding machines that depends on
P6
the service time tji , i.e. ui = j=1 yji tji .
For the static capacity model the arrival rate λj on basis job class is distributed over the machines and results in a utilization profile, that corresponds
to the capacity usage. A proper distribution or resource allocation is therefore
a center piece of capacity planning, it will be used to judge if the demand can
be satisfied and to identify the bottlenecks. The capacity model selected for
the investigations is based on the idea of minimizing the maximum utilization
- that means, to minimize utilization on the bottleneck. This approach can be
extended to a lexicographic mini-max objective like in Ogryczak [20] to identify all levels of equally utilized equipment (cf. resource pool concept in Gold
[9, 10]). A solution that is optimal with respect to the lexicographic mini-max
objective has the property that only machines of the same level may share the
load of a common job class.
The LP defined in (1-4) is called the basic model, where the maximum load is
minimized. According to (3) ρ is an upper limit to the utilization ui for all tools
i ∈ I and the decision variable xji corresponds to the time invested in processing
lots from job class j on machine i. The service rate µji = t1ji corresponds to
number of lots that can be processed in one time unit. Therefore the number
of processed units of job class j on machine j is yji = xji µji . In order to satisfy
the demand, constraint (2) has to be satisfied.
6
product mix
(pP)
dp
d1=20
40
ua
utilization
job classes
j J 1
lj l1=30
a
wafer starts
per week
(pP)
d2=10
product routes
expected utilization
for each machine
demand for each
job class (given)
demand for each
product (given)
10
demand for each product
2
3
l2=80 l3=20
d3=40
4
5
l4=60 l5=10
ub
uc
b
c
6
l6=40
ud
d
(iI)
machines
Figure 3: A simplified capacity planning process - the Figure is adapted from Romauch and
Klemmt [23]. A given demand on the basis of products is decomposed into a demand for
each job class. Finally a distribution of the workload results in a forecast for the expected
utilization.
min ρ
s.t.
P
i∈I xji µji = λj
P
ui = j∈J xji ≤ ρ
xji ≥ 0.
(1)
(∀j ∈ J),
(2)
(∀i ∈ I),
(3)
(4)
3. Capacity Planning for Cluster Tools
The LP defined in (1-4) is a valid model for parallel single servers in heavy
traffic (see e.g. Harrison and Lopez [12]), where each machine is represented by
a queue with a given service rate.
The models presented in this paper are based on these assumptions and
extend the model to Cluster-Tools. According to Lopez and Wood [18] the
main operating modes are the serial and the parallel mode. A Cluster-Tool is
operated in serial mode if each wafer visits all chambers in a sequence and it is
working in parallel mode if each wafer visits only one chamber.
The serial mode can be approximated (lower bound) by separating the chambers of the Cluster-Tool from the mainframe. For instance, a Cluster-tool i with
three chambers {A, B, C} is represented by the mainframe i, and the chambers
(i, A), (i, B) and (i, C). An activity j calls chambers and mainframe at the same
time. The corresponding service rates for the chambers are µj,(i,A) , µj,(i,B) and
7
µj,(i,C) . Since the slowest process defines the throughput of the tool, hence
the service rate for the mainframe is defined by µj,i = minu∈A,B,C {µj,(i,v) }.
Therefore it is sufficient to add the following constraint to (1-4):
P
ui,v =
xji µji
j∈J µj,(i,v)
≤ ρ (∀i ∈ I, v ∈ {A, B, C})
(5)
The objective of this paper is to propose a similar approximation for ClusterTools in parallel mode. The corresponding models need a bit more explanation.
For Cluster-Tools in parallel mode the service rate depends on the selected
subset of chambers r∈ R ={A, B, C, AB, ...} that are selected to process a given
type of lot. The subset of chambers is also called a recipe. The service time µj,i,r
depends on the job class j, the Cluster tool i and the recipe r. The demand is
represented by the arrival rate λj .
The distribution of lots is determined by xj,i,r , which is the time invested
in activity (j, i) with recipe r. In Ortner [21] a parallel Cluster-Tool model for
two Load locks and three processing chambers is discussed. The basic idea is
the identification of categories for optimal schedules (see Figure 4) where the
makespan can be represented by a simple formula. Each formula is the
Pweighted
sums of the times
invested
in
a
given
recipes
amd
can
be
written
as
r wi,r xi,r
P
where xi,r = j xj,i,r .
For instance, if all possible processes are running in parallel which means that
x
+x
+x
+xi,A +xi,B +xi,C
the load locks are the limiting resource, then xi,ABC + i,AB i,AC i,BC
2
is the corresponding make-span. If this is not the case, then a single chamber
recipe exists that is too short or too long to run in parallel, which means that
the make-span must be larger than that.
It was shown, that the make-span can be represented by the maximum value
of five formulas. The corresponding proof in Ortner [21] is constructive and
generates feasible schedules for each case identified.
The LP (6-9) is an abstraction of the LP found in Ortner [21], where the
matrix B is defined in Table 1.
In (6-9) we present an extension of (1-4) that can be used to represent the
approach presented in Ortner [21]. For the three chamber case, the matrix B
represents the formulas given in Figure 4 and the corresponding values can be
found in Table 1.
One of the contributions of the paper is to derive matrices for more than
three chambers.
min ρ
P
(6)
xj,i,r µj,i,r = λj
(∀j ∈ J)
(7)
Bk,r xj,i,r ≤ ρ
(∀i ∈ I, k ∈ K)
(8)
(j, i, r) ∈ C
(9)
i ∈ I, r ∈ R :
(j, i, r) ∈ C
P
j∈J,r∈R
xj,i,r ≥ 0
8
xi,ABC + xi,AB + xi,AC + xi,A
A
xi,ABC
xi,AB
xi,AC
B
xi,ABC
xi,AB
xi,B
C
xi,ABC
xi,C
xi,ABC +
xi,A
xi,BC
xi,B
xi,AC xi,BC
A
xi,ABC
xi,AB
B
xi,ABC
xi,AB
C
xi,ABC
xi,C
xi,ABC + xi,AB + xi,BC + xi,B
A
xi,ABC
xi,AB
B
xi,ABC
xi,AB
xi,BC
C
xi,ABC
xi,C
xi,BC
xi,AB +xi,AC +xi,BC +xi,A +xi,B +xi,C
2
xi,A
xi,B
xi,B
xi,C
xi,ABC + xi,AC + xi,BC + xi,C
xi,A
xi,B
A
xi,ABC
xi,AC
xi,A
B
xi,ABC
xi,B
xi,BC xi,B
C
xi,ABC
xi,AC
xi,BC
xi,C
xi,ABC + xi,AB + xi,BC + xi,AC
A
xi,ABC
xi,AB
B
xi,ABC
xi,AB
C
xi,ABC
xi,C
xi,AC
xi,A
xi,B
xi,BC
xi,AC
xi,BC
Figure 4: Illuistration of different categories of optimal schedules. Five examples are given,
one for each representative schedule. Each Figure ia a Gantt-chart (x-axis: time) on basis
chambers. The formula in the header contains the formula to calculate the makespan, it
represents this category. Two-chamber-recipes are colored to highlight the joint usage of two
chambers.
Table 1: Matrix B = (Bk,r ) for the dual based model (6-9) when considering three chambers.
k=1
2
3
4
5
r=A
0.5
1
B
0.5
C
0.5
AC
0.5
1
1
1
1
1
BC
0.5
1
1
1
AB
0.5
1
1
1
ABC
1
1
1
1
1
3.1
An alternative modeling approach that exploits parallel processing
BC
AC
xi,BC
9
xi,AC
ξi,AC,B
A
B
ξi,A,BC
xi,A
ξi,A,B
xi,ABC
ABC
ξi,A,C
C
xi,B
ξi,B,C
xi,C
ξi,AB,C
AB
xi,AB
Figure 5: Parallel processing for three chambers, graph representation G(R, E). The vertices
R represent the recipes and the edges E represent possible parallelizations. The Figure on
the right hand side represents the parallelization opportunities for a given Cluster-tool i. For
instance, the variable ξi,AB,C defines the time invested in parallel processing of recipe AB
and recipe C.
Before tackling this problem, an alternative modeling approach that considers more than three chambers is presented.
3.1. An alternative modeling approach that exploits parallel processing
For a single equipment i, where the time invested in certain recipes xj,i,r
is given and P
no parallel processing is allowed (one load lock) the resulting
makespan is j,r xj,i,r . In case two load locks are available it is possible to
exploit parallel processing. Since xj,i,r cannot be changed, the time spent in
parallel processing is the only way to reduce the makespan. This parallelization
problem can be seed as a sub problem and will be explained in this chapter. A
similar idea was used in Tian et al. [26] for parallel processor scheduling.
In Figure 5 the graph representation G(R, E) of the parallelization problem
for three chambers can be found. The vertices R represent the recipes and the
edges E represent possible parallelizations. On the right hand side the parallelization opportunities for a given Cluster-tool i are illustrated. For instance,
when considering the parallelizations with recipe C, then there are three possible candidates {A, B, AB}. Each of these options is reflected by a variable
that defines the time used for this recipe combination, i.e. ξi,A,C , ξi,B,C and
ξi,AB,C . Also note that the time invested in parallel processing is limited by the
given values xj,i,r . For instance when considering ξi,AB,C two restrictions can
be found: ξi,AB,C ≤ xi,AB and ξi,AB,CP+ ξi,A,C + ξi,B,C ≤ xi,C . According to
that, a parallelization that maximizes (r1 ,r2 )∈E ξi,r1 ,r2 solves the sub-problem.
The prallelization graph that considers four chambers is depicted in Figure
6 and the number of vertices is already twice as large. The graph is also not
planar anymore, since it contains a K3,3 given by {A, B, AB} × {C, D, CD}.
The general formulation of the corresponding sub problem for optimal parallelization is represented by the following LP:
3.1
An alternative modeling approach that exploits parallel processing
xi,BC
10
xi,AD
xi,ABCD
xi,AC
xi,ACD
xi,ABC
xi,B
xi,D
xi,CD
xi,AB
xi,A
xi,C
xi,BCD
xi,ABD
xi,BD
Figure 6: Parallel processing for four chambers, graph representation G(R, E).
max
P
r2 :(r,r2 )∈E
P
(r1 ,r2 )∈E
ξi,r,r2 +
ξi,r1 ,r2
P
r1 :(r1 ,r)∈E
ξi,r1 ,r ≤ xi,r
ξi,r1 ,r2 ≥ 0
(10)
(∀r ∈ R)
(11)
(∀(r1 , r2 ) ∈ E)
(12)
The sub-problem (10-12) be integrated in the master LP as follows:
min ρ
P
(13)
j∈J
(14)
i ∈ I, r ∈ R
(15)
i ∈ I, r ∈ R
(16)
i∈I
(17)
xj,i,r ≥ 0
(j, i, r) ∈ C
(18)
xi,r ≥ 0
i ∈ I, r ∈ R
(19)
i ∈ I, (r1 , r2 ) ∈ E
(20)
xj,i,r µj,i,r = λj
i ∈ I, r ∈ R :
(j, i, r) ∈ C
P
xi,r = j xj,i,r
P
P
r2 :(r,r2 )∈E ξi,r,r2 +
r :(r ,r)∈E ξi,r1 ,r ≤ xi,r
P
P 1 1
r∈R xi,r −
(r1 ,r2 )∈E ξi,r1 ,r2 ≤ ρ
ξi,r1 ,r2 ≥ 0
In order to count the number of variables xj,i,r , xi,r and ξi,r1 ,r2 it is sufficient
to count the vertices and edges in the parallelization graph. Suppose that n
chambers
are available, then |R| = 2n − 1 recipes are available. There are
n
k possibilities to choose a set of k chambers and each partition defines nonk
conflicting recipe pair, that makes 2 2−2 (= 2k−1 − 1) edges. Therefore, the
Pn
n
number of edges is |E| = k=0 (2k−1 − 1) = 3 2−1 − (2n − 1) in total. According
n
to that the coefficient matrix of (13-20) has 1+|C| + |I| 3 2−1 columns, |J| +
|I|(1 + 2|R|) rows and 2|C| + |I|(1 + 3|R| + 2|E|) nonzeros.
3.2
Analysis of the parallelization problem
11
3.2. Analysis of the parallelization problem
The second approach is based on investigating the dual of the sub-problem
(10-12). The problem (10-12) can be represented by a max-flow problem which
gives some insights about the structure of basic feasible solutions.
The corresponding max-flow problem is defined as follows, the set of vertices
e ∪ {s, t}, where R
e
and edges and the capacity limits are defined by: V = R ∪ R
e
is a copy of R. The arcs between R and R are a bipartite representation of E,
together with the links to the source, and the links to the sink, the set of arcs
e i.e Ee ={(r, re) ∈ R × R
e : (r, re) ∈ E} ∪ {(s, r) : r∈ R} ∪ {(e
is defined by E,
r, t) :
e The arc capacities are infinite for all arcs except the links to source
re ∈ R}.
and the links to the sink. In Figure 7 the max-flow version for three chambers
e ∪ {s, t}, E)
e is illustrated. The meaning
is depicted where the graph G(R ∪ R
of the flows ηi,r1 ,r˜2 and ηi,r2 ,r˜1 is directly linked to ξi,r1 ,r2 since the maximum
flow formulation can be understood as a duplication of the original problem.
The details will be elaborated in a transformation of the problem. Before that,
the LP of the corresponding max-flow problem will be defined in (21-26). The
objective (22) is to maximize the flow into the sink t. The constraints (22)
and (23) represent the flow balance equations. The capacities on the arcs are
formulated in (24) and (25).
P
e ηi,re,t
(e
r ,t)∈E
P
= ηi,s,r
e ηi,r,e
r
e
r : (r,e
r )∈E
P
= ηi,e
e ηi,r,e
r
r ,t
r: (r,e
r )∈E
xi,r
ηi,s,r ≤ 2
max
ηi,e
≤
r ,t
xi,r
2
ηi,e ≥ 0
(21)
(∀r ∈ R)
(22)
e
(∀e
r ∈ R)
(23)
(∀r ∈ R)
(24)
(∀r ∈ R)
(25)
e
(∀e ∈ E)
(26)
To see that the max-flow version is a suitable representation of the sub
problem, a transformation of solutions will be established.
If η is a feasible solution for (21-26), then ξ defined by ξi,r1 ,r2 ← ηi,r1 ,re2 +
ηi,r2 ,re1 is a feasible solution for (10-12) and the objective value is the same.
In order to prove this statement it is necessary to provePthat (11) is satisfied. The left hand side of (11) has the following form:
r2 :(r,r2 )∈E (ηi,r,re2 +
| {z }
(i.1)
(ii.1)
(ii.2)
P
z }| {
z }| {
ηi,r2 ,e
) + r1 :(r1 ,r)∈E (ηi,r1 ,e
+ ηi,r,re1 ). Note that (r, r2 ) ∈ E is represented by
r
r
| {z }
(i.2)
e Therefore the components (i.1) and (i.2)
two arcs (r, r˜2 ) ∈ Ee and (r2 , r̃) ∈ E,
can
by
P be combined andPbecause of (22,24) the
P corresponding part is bounded
xi,r
η
+
η
=
η
=
η
≤
.
Bei,s,r
r2 :(r,r2 )∈E i,r,re2
r1 :(r1 ,r)∈E i,r,re1
2
e i,r,e
r
e
r :(r,e
r )∈E
cause of (23,25) the part that corresponds to (ii.1) and (ii.2) is bounded by
3.2
Analysis of the parallelization problem
≤
s
12
ABC
]
ABC
AB
g
AB
xi,AB
2
≤
ηAB,Ce
ηC,AB
f
AC
g
AC
BC
g
BC
C
f
C
B
f
B
A
xi,AB
2
f
A
R
t
R̃
e ∪ {s, t}, Ee) of the maximum flow formulation of the
Figure 7: Graph representation G(R ∪ R
sub problem for three chambers. The flow can be interpreted as the time used for parallel
processing: ξAB,C ← η
+η
.
AB,C
C,AB
e
f
P
P
P
ηi,r2 ,e
+ r1 :(r1 ,r)∈E ηi,r1 ,e
= r:(r,e
= ηi,e
≤
r
r
e ηi,r,e
r
r ,t
r )∈E
ξ is feasible for (10-12).
r2 :(r,r2 )∈E
xi,r
2 .
Hence,
The dual is therefore a min-cut problem and for the solution space it is
sufficient to consider selecting arcs that are connected to the sink or to the
source - these cuts are called basic cuts. In other words, for every cut it is
possible to construct a basic cut that has a cut value that is smaller or equal.
The formulation of the dual of the maximum flow formulation (21-26) can be
found in (27-30). The dual variables are called πi,r and π
ei,r . The constraint
(28) reveals the vertex covering aspect of the problem.
min
P
r∈R
xi,r
πi,r +e
πi,r
2
πi,r1 + π
ei,r2 ≥ 1
(27)
∀(r1 , re2 ) ∈ Ee
(28)
πi,r ≥ 0
∀r ∈ R
(29)
π
ei,r ≥ 0
∀r ∈ R
(30)
This problem is totally uni-modular, because the columns can be partitioned
according to the criterion proposed in Heller and Tompkins [13].
For the search space, it is sufficient to consider basic cuts. A basic cut B can
e and
be represented by the corresponding nodes that are selected, B ⊂ R ∪ R
e
each arc (r1 , re2 ) between R and R is covered by B to satisfy (28). Furthermore,
3.2
Analysis of the parallelization problem
13
it is sufficient to consider basic cut that cannot be reduced by eliminating a
node, that means that by definition, removing one node of a minimal basic cut
leads to an infeasible solution.
e \ B) of a minimal basic cut is a maxTherefore the complement (B c = R ∪ R
imum independent set. Efficient algorithm for listing all maximum independent
sets are known (see Kashiwabara et al. [17]) and can be used to find all minimal
basic cuts.
k
k
Suppose that a listing {(πi,r
,π
ei,r
)}k∈K that represents all minimal basic cuts
is given, then the optimal value for (27-30) corresponds to the cut value with
the smallest value in the listing of all minimal basic cuts.
k
k
)}k∈K are given, then the
,π
ei,r
Suppose that all minimal basic cuts {(πi,r
k
k
P
πi,r
+e
πi,r
smallest value r∈R xi,r
defines the optimum.
2
In order to remove
symmetric solutions,
the solutions will be identified by
k
k
πi,r
+e
πi,r
, hence the optimum is defined by
vectors (Πi,r,k ) where Πi,r,k ←
2
P
mink r∈R xi,r Πi,r,k . The minimum of a set of numbers is defined as the largest
lower bound of these numbers, therefore the optimum can be represented by the
LP formulated in (31-32), where constraint (32) ensures that χ is a lower bound
of the cut values and the objective (31) defines that χ is the largest lower bound:
max χ
χ≤
P
r∈R
xi,r Πi,r,k
(31)
∀k ∈ K
(32)
Therefore the shortest possible makespan
for given xi,r can be obtained by
P
changing
the
the
objective
(32)
to
min
x
r∈R i,r − χ. A substitution of χ with
P
ρ ← r∈R xi,r − χ in (33-32) leads to the following equivalent LP:
ρ
min
P
r∈R
(33)
(1 − Πi,r,k ) xi,r ≤ ρ ∀k ∈ K
(34)
The new formulation of the sub problem will be integrated in the master problem
(13-20) by replacing the constraints (17,16) with (34). The resulting master LP
is formulated in (3540):
min ρ
P
(35)
j∈J
(36)
i ∈ I, r ∈ R
(37)
− Πi,r,k )xi,r ≤ ρ
i ∈ I, k ∈ K
(38)
xj,i ≥ 0
i ∈ I, r ∈ R
(39)
xj,i,r ≥ 0
(j, i, r) ∈ C
(40)
xj,i,r µj,i,r = λj
i ∈ I, r ∈ R :
(j, i, r) ∈ C
xi,r =
P
r∈R (1
P
j
xj,i,r
14
The problem (27-30) can be interpreted as a relaxation of the vertex cover
problem, therefore the basic solutions are half integral (compare Hochbaum [14],
type 2var).
If the number of chambers is fixed, the coefficients (1 − Πi,r,k ) only depend
on the number of chambers, therefore a representation by a R × K matrix
1−Π = (1−Πr,k ) can be used. In order to reduce the size of the matrix, all rows
of 1 − Π that are redundant will be removed. The details to the corresponding
Algorithm can be found in AppendixA.
For three chambers the model (3540) coincides with the model presented in
Ortner [21] (compare (6-9)) and the matrix (1 − Π) is equivalent to the matrix
presented in Table 1. Therefore, the presented methodology replaces (6-9) and
provides an alternative to the elementary proof given in Ortner [21].
More importantly, the approach reveals a generalization of the model in
Ortner [21] to more than three chambers. For instance, the matrix for four
chambers is given in Table 2. Note that line 1 corresponds to complete parallelization, while line 8 corresponds to the sum of all recipes that contain A and
line 12 corresponds to chamber B, line 18 represents the utilization for chamber
C and line 23 represents the utilization for chamber D. Note that the coeffin
cient matrix of (3540) has 1+|C| + |I| 3 2−1 columns, |J| + |I|(|R| + |K|) rows and
2|C| + |I|(|R| + 1 + |K| + |1 − Π|) nonzeros, where |1 − Π| denotes the number
of nonzeros in (1 − Πi,r,k ). Known values for |K| and |1 − Π| can be found in
Table 4. According to that, the generalized model for five chamber is already
getting fairly large.
In the next section a computational experiment will be used to compare the
new approach and the generalized model.
4. Computational Experiments
In this Section the computational results for the generalized formulation
(35-40) and the alternative model (13-20) for the a large set of benchmark
instances (http://www.univie.ac.at/prolog/research/Clustertool/) is presented.
The instances were solved with Gurobi 6.0 on a desktop computer with a [email protected] GHz processor.
The structure of the generalized formulation (13-20) seems to be much simpler than the alternative model (35-40), but especially for more than three
chambers the coefficient matrix of the generalized formulation tends to grow
faster and therefore it is interesting to examine which model performs better.
The size of the problems is indicated in Table 5 and Table 7, where the number
of nonzero entries in the coefficient matrix of the alternative model and the
scaling factor for the generalized model are documented (all chamber released).
The number of nonzeros of the coefficient matrix for the generalized model is
nonzerogen = 2|C|+|I|(|R|+1+|K|+|Π|) and the for the alternative formulation
it is nonzeroalt = 2|C| + |I|(1 + 3|R| + 2|E|) . Therefore the difference of the
δ =nonzerogen − nonzeroalt can be calculated as follows:
δ = |I|(|K| + |Π| − 2(|R| + |E|)) = |I|(1 + |K| + |Π| − 3n )
15
Table 2: Matrix B for the dual based model (6-9) when considering four chambers.
A
AB
ABC
ABCD
ABD
AC
ACD
AD
B
BC
BCD
BD
C
CD
D
1
0.5
0.5
0.5
1
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
2
0.5
0.5
1
1
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
3
0.5
0.5
0.5
1
1
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
4
0.5
0.5
1
1
1
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
5
0.5
0.5
0.5
1
0.5
0.5
1
0.5
0.5
0.5
0.5
0.5
0.5
6
0.5
0.5
1
1
0.5
0.5
1
0.5
0.5
0.5
0.5
0.5
0.5
7
0.5
0.5
0.5
1
1
0.5
1
0.5
0.5
0.5
0.5
8
1
1
1
1
1
1
1
1
9
0.5
0.5
1
0.5
0.5
0.5
0.5
0.5
0.5
1
0.5
10
0.5
1
1
0.5
0.5
0.5
0.5
0.5
0.5
1
0.5
11
0.5
0.5
1
1
0.5
0.5
0.5
0.5
0.5
1
0.5
12
1
1
1
1
1
1
1
1
13
0.5
0.5
1
0.5
0.5
1
0.5
1
0.5
14
1
1
1
1
1
1
1
1
15
1
1
1
1
1
1
16
1
1
1
1
1
17
1
1
1
1
1
18
1
1
19
1
1
20
1
21
22
23
0.5
1
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
Table 4: Known values for |K| and |1 − Π|
|K|
|1 − Π|
n=1
1
1
2
2
4
3
5
22
4
23
245
5
590
13740
Table 5: Number of nonzero entries in the coefficient matrix of the alternative model; unit:
thousands.
three chambers
four chambers
five chambers
density shape sizecat=0 1
2
3
0
1
2
3
0
1
2
3
1:1
3.4 11.3 40.9 156.8 8.0 25.7 90.6 341.7 18.9 57.7 196.3 724.4
1:4
2.2 8.7 32.9 123.7 5.1 19.4 71.8 267.9 11.6 42.4 153.0 562.8
1
16:1
6.2 17.2 54.1 188.2 16.1 42.5 127.3 425.8 42.5 106.0 299.5 953.0
4:1
4.4 13.8 46.7 171.7 10.8 32.3 105.8 379.2 26.8 75.9 236.9 820.2
1:1
4.3 15.3 59.2 232.5 9.9 34.2 129.8 504.0 22.7 75.3 277.3 1059.7
1:4
3.0 10.7 42.2 165.5 6.9 23.6 91.9 357.5 15.3 51.1 194.5 747.9
2
16:1
7.0 21.0 74.0 271.4 17.9 50.6 169.9 604.1 46.0 122.8 387.6 1321.4
4:1
5.3 18.4 66.6 256.1 12.7 42.2 148.3 560.0 30.7 96.3 324.8 1193.8
1:1
4.5 17.4 68.7 271.7 10.4 38.7 149.9 587.8 23.8 84.6 319.0 1233.0
1:4
3.2 11.5 46.0 181.4 7.2 25.4 99.9 391.4 16.1 54.8 211.0 818.1
3
16:1
7.1 23.9 85.4 318.7 18.0 56.8 194.4 705.4 46.4 135.6 438.2 1530.8
4:1
5.8 20.3 78.4 302.7 13.8 46.3 173.6 659.9 33.0 104.8 376.9 1400.2
16
Table 7: Scaling factor for nonzero entries in the coefficient matrix of the generalized model
relative to alternative model. An entry of 2 means that the number of nonzero elements is
twice as large.
three chambers
four chambers
five chambers
density shape sizecat=0 1
2
3
0
1
2
3
0
1
2
3
1:1
1.01
1.00 1.00 1.00 1.47 1.29 1.17 1.09 15.93 10.76 6.74
4.11
1:4
1.00
1.00 1.00 1.00 1.38 1.22 1.12 1.06 13.44 8.49 5.06
3.13
1
16:1
1.00
1.00 1.00 1.00 1.36 1.19 1.10 1.05 12.85 7.66 4.53
2.83
4:1
1.00
1.00 1.00 1.00 1.37 1.19 1.10 1.06 13.17 7.64 4.68
3.00
1:1
1.00
1.00 1.00 1.00 1.27 1.16 1.08 1.04 10.21 6.52 3.90
2.51
1:4
1.00
1.00 1.00 1.00 1.26 1.15 1.08 1.04 9.75 6.14 3.67
2.38
2
16:1
1.01
1.01 1.01 1.00 1.93 1.71 1.47 1.28 27.55 22.26 16.05 10.46
4:1
1.01
1.01 1.00 1.00 1.84 1.59 1.35 1.20 25.48 19.35 12.63 7.82
1:1
1.01
1.01 1.00 1.00 1.83 1.53 1.31 1.17 25.31 17.63 11.29 6.89
1:4
1.01
1.01 1.00 1.00 1.70 1.47 1.28 1.16 22.00 15.86 10.51 6.50
3
16:1
1.01
1.00 1.00 1.00 1.59 1.36 1.20 1.11 19.33 12.70 7.94
4.78
4:1
1.01
1.00 1.00 1.00 1.55 1.32 1.17 1.09 18.09 11.76 6.98
4.22
Table 9: Auxiliary parameters δn and γn for calculating δ = nonzerosgen − nonzerosalt =
|I|δn and nonzeroalt = 2|C| + |I|γn .
n=1
2
3
4
5
γn =
δn
3n+1 −2n+1 +1
2
0
-2
1
188
14088
3
10
33
106
333
17
In Table 9 some values for
δ
|I|
can be found.
The instances that were used in the computational experiment vary in size
and structure and the following four factors were controlling the construction:
• sizecat {0, 1, 2, 3}
where, |I| · |J| = 100 · 4sizecat = 400,1600, 6400, 25600
• shape {1 : 4, 1 : 1, 4 : 1, 8 : 1}
where |I| : |J| = shape, the proportion of tools and jobs
• locked {0, 3, 6, 9}
is the probability that a chamber is not available.
locked chambers: locked
10
• density {1,2,3}
1=sparse, 2=medium, 3=dense
For the three chamber model, the corresponding instances have up to 300.000
nonzeros and for four chamber up to 700.000 nonzeros for the case where all
chambers are available. According to Table 10 the alternative model is significantly faster for four chambers, but in average slower for three chambers.
18
Table 10: Speedup factor for alternative model versus generalized model (detailed results).
The bold entries represent cases were the generalized model is faster.
three chambers
four chambers
sizecat shape density locked=0 3
6
9
avg locked=0 3
6
9
avg
1
0.84 0.82 0.88 0.8 0.84
1.25
1.12 0.83 1.11 1.08
1:4
2
1.07
0.81 0.61 0.93 0.86
1.2
1.15 1.33 1.18 1.22
3
0.83 0.85 0.91 0.93 0.88
1.2
1.21 1.27 1.22 1.23
1
0.97 0.83 0.8 0.86 0.87
0.92 0.82 0.71 1.22 0.92
1:1
2
0.79 0.97 0.88 0.9 0.89
1.11
1.12 1.23 1.14 1.15
3
0.86
0.9 0.92 0.82 0.88
0.93
1.1 1.09 1.16 1.07
0
1
0.67 0.65 0.81 0.68 0.7
0.66 0.79 0.82 0.8 0.77
4:1
2
0.96 0.73 0.77 0.76 0.81
0.77 0.82 0.82 0.93 0.84
3
0.83 0.74 0.77 0.71 0.76
0.82
0.8 0.83 0.9 0.84
1
0.72
0.6 0.63 0.51 0.62
0.74
0.6 0.58 0.59 0.63
16:1
2
0.63 0.58 0.65 0.68 0.64
0.78 0.72 0.6 0.67 0.69
3
0.6
0.58 0.59 0.52 0.57
0.81 0.71 0.66 0.69 0.72
1
0.98 0.84 0.83 0.78 0.86
1.63
1.33 1.37 1.48 1.45
1:4
2
0.89 0.91 0.96 0.96 0.93
2.29
1.73 1.62 1.61 1.81
3
0.76 0.75 0.94 0.97 0.86
1.89
1.78 1.52 1.6 1.7
1
1.08
0.92 1.02 0.85 0.97
1.43
1.43 1.42 1.34 1.41
1:1
2
1.11
0.82 1.09 1.01 1.01
1.53
1.48 1.44 1.43 1.47
3
1.11
1.07 1.1 1.73 1.25
1.83
1.43 1.46 1.42 1.54
1
1
0.8
0.7 0.84 0.62 0.74
1.22
1.07 0.98 1.11 1.1
4:1
2
1.21
0.54 1.15 0.86 0.94
1.52
1.41 1.21 1.22 1.34
3
1.18
0.8 0.86 0.79 0.91
1.49
1.49 1.24 1.38 1.4
1
0.66 0.57 0.71 0.68 0.66
0.99
1.04 0.83 0.84 0.96
16:1
2
0.68
0.6 0.48 0.72 0.62
1.29
1.27 1.12 0.94 1.16
3
0.74 0.51 0.7 0.76 0.68
1.26
1.27 1.14 1.07 1.19
1
0.91 0.97 0.74 1.02 0.91
2.05
2.32 1.92 1.82 2.03
1:4
2
1.08
1.06 0.98 0.96 1.02
2.22
2.02 2.03 2.05 2.08
3
1.06
1.09 1.06 1.09 1.08
2.04
2.02 1.87 1.98 1.98
1
1.3
1.13 0.95 1.29 1.17
2.35
1.68 1.81 1.62 1.87
1:1
2
1.19
1.09 1.22 1.05 1.14
2.08
2.35 1.94 1.8 2.04
3
1.26
1.4 1.11 1.05 1.21
1.87
1.9 1.84 1.94 1.89
2
1
0.95 0.89 0.88 0.92 0.91
1.69
1.78 1.66 1.53 1.67
4:1
2
1
0.91 0.89 0.93 0.93
1.68
1.7 1.83 1.7 1.73
3
2.43
0.99 0.94 1
1.34
1.82
1.76 1.78 1.69 1.76
1
0.74 0.78 0.75 0.98 0.81
1.44
1.72 1.41 1.27 1.46
16:1
2
0.84 0.82 0.83 0.85 0.84
1.5
1.71 1.63 1.51 1.59
3
0.85
0.8 0.89 0.96 0.88
1.48
1.67 1.68 1.53 1.59
1
1.41
1.02 1.19 1.19 1.34
2.5
2.44 2.17
2
2.28
1:4
2
1.24
1.16 1.1 0.94 1.11
2.41
2.42 2.74 2.23 2.45
3
1.22
1.1 0.99 1.2 1.13
2.33
2.2 2.93 2.1 2.39
1
1.77
1.32 1.05 1.07 1.3
1.82
2.26 2.04 1.94 2.02
1:1
2
1.28
1.48 1.33 1.05 1.29
2.94
2.24 2.13 2.11 2.36
3
1.33
1.31 1.32 1.09 1.26
3.09
3.52 2.21 2.11 2.73
3
1
1.08
0.92 0.95 1.01 0.99
1.7
1.81 1.87 1.82 1.8
4:1
2
1.04
0.92 1
1.03
1
1.85
2
1.91 1.87 1.91
3
2.7
1.11
1
1.1 1.48
2.03
1.91 1.82 1.96 1.93
1
1.03
0.83 0.89 0.92 0.92
1.63
1.91 1.88 1.62 1.76
16:1
2
1.02
0.94 0.92 1.01 0.97
1.7
1.7 1.94 1.74 1.77
3
1.16
0.91 0.95 0.99 1
1.63
1.64 1.88 1.79 1.74
avg
1.11
0.9 0.91 0.93 0.97
1.55
1.59 1.52 1.47 1.54
19
Table 12: CPU time in milliseconds.
three
sizecat shape density locked=0 3
1
25
28
1:4
2
30
36
3
29
33
1
36
36
1:1
2
39
36
3
37
42
0
1
48
49
4:1
2
47
49
3
58
53
1
64
84
16:1
2
63
71
3
70
73
1
60
68
1:4
2
76
79
3
96
101
1
73
95
1:1
2
84
132
3
98
118
1
1
106
124
4:1
2
122
219
3
129
147
1
144
184
16:1
2
155
199
3
170
284
1
230
234
1:4
2
237
269
3
268
316
1
237
271
1:1
2
349
380
3
366
454
2
1
306
325
4:1
2
414
462
3
482
504
1
423
417
16:1
2
516
532
3
600
636
1
844.5 920
1:4
2
893
1171
3
1000 1270
1
783
1106
1:1
2
1498 1533
3
1749 2017
3
1
1097 1339
4:1
2
1751 2012
3
1823 2184
1
1216 1450
16:1
2
1820 2002
3
1998 2465
chambers
6
9
25
25
41
27
32
30
35
35
42
39
38
40
47
47
48
50
53
56
65
83
68
65
70
77
70
86
74
72
89
79
88
88
99 100
114 135
111 130
127 130
139 158
158 155
244 164
194 176
301 219
284 286
297 282
275 263
366 353
438 418
346 343
490 437
519 487
435 377
527 502
579 568
799 798
1048 1194
1318 1110
1010 932
1567 1470
1888 1705
1170 1077
1744 1632
2162 1816
1348 1229
1976 1715
2271 2057
four chambers
avg locked=0 3
6
9
25.75
48
41
48
35
33.50
60
60
55
55
31.00
70
66
59
58
35.50
72
85
95
67
39.00
97
82
80
80
39.25
109
94
90
83
47.75
129
138 107 98
48.50
149
144 131 115
55.00
164
158 145 127
74.00
219
233 203 169
66.75
230
249 238 192
72.50
241
263 245 190
71.00
166
175 153 128
75.25
175
200 181 159
91.25
190
213 202 174
86.00
221
225 206 185
103.75
267
285 262 243
116.25
273
319 313 275
117.75
301
329 333 282
149.50
372
376 378 343
143.25
395
381 418 407
160.25
615
513 507 459
190.50
538
549 552 533
206.00
583
606 586 542
246.00
498
491 518 468
269.00
643
717 663 585
290.75
727
778 734 620
261.50
581
801 693 627
362.00 1015 1010 940 849
419.00 1213 1226 1098 969
330.00
904
907 888 802
450.75 1428 1408 1189 1063
498.00 1460 1639 1398 1214
413.00 1467 1347 1242 1111
519.25 2049 1731 1540 1342
595.75 2373 2030 1731 1556
843.23 2574 2330 1912 1655
1076.50 3771 3387 2693 2244
1174.50 4066 3881 2970 2481
957.75 4401 3155 2750 2228
1517.00 6157 5217 3992 3243
1839.75 6818 5885 4755 3793
1170.75 5851 4490 3302 2538
1784.75 8029 6460 4972 3733
1996.25 8498 7921 6246 4475
1310.75 6815 4901 3794 3274
1878.25 10745 8655 5332 4430
2197.75 12061 10248 6514 5102
avg
43.00
57.50
63.25
79.75
84.75
94.00
118.00
134.75
148.50
206.00
227.25
234.75
155.50
178.75
194.75
209.25
264.25
295.00
311.25
367.25
400.25
574.33
543.00
579.25
493.75
652.00
714.75
675.50
953.50
1126.50
875.25
1272.00
1427.75
1291.75
1665.50
1922.50
2117.75
3023.75
3349.50
3133.50
4652.25
5312.75
4045.25
5798.50
6785.00
4696.00
7290.50
8481.25
20
5. Summary
The paper proposes a new model for representing Cluster-Tools with two
load locks and provides generalization of an existing Cluster-Tool model Ortner
[21] that was restricted to three chambers. A computational study indicates
that on average the generalized model for three chambers is a little faster than
the new model proposed in this paper, but for several instances the proposed
alternative model seems to have small but significant benefits.
For four chambers, the new model performs significantly better than the
generalized model. Except for very small instances, the new model out-performs
the generalized one.
Our future research will consider Cluster-tools with more than two load
locks and we will investigate the impact of conflicting recipes. We will also
work on sensitivity analysis and robustness with respect to recipe combinations
and availability.
References
[1] Almeder, C., Preusser, M., Hartl, R. F., 2009. Simulation and optimization
of supply chains: alternative or complementary approaches? OR spectrum
31 (1), 95–119.
[2] Chazelle, B., 1993. An optimal convex hull algorithm in any fixed dimension. Discrete & Computational Geometry 10 (4), 377–409.
[3] Christopher, J., 2008. Study of optimal load lock dedication for cluster
tools. In: Proceedings of the 40th Conference on Winter Simulation. Winter
Simulation Conference, pp. 2136–2140.
[4] Doleschal, D., Weigert, G., Klemmt, A., Lehmann, F., 2013. Advanced
secondary resource control in semiconductor lithography areas: from theory
to practice. In: Proceedings of the 2013 Winter Simulation Conference:
Simulation: Making Decisions in a Complex World. IEEE Press, pp. 3879–
3890.
[5] Duemmler, M., Wohlleben, J., 2012. A framework for effective wip flow
management in semiconductor frontend fabs. In: Simulation Conference
(WSC), Proceedings of the 2012 Winter. IEEE, pp. 1–6.
[6] Franssila, S., 2010. Introduction to microfabrication. John Wiley & Sons.
[7] Gale, D., 1989. The theory of linear economic models. University of Chicago
press.
[8] Gansterer, M., Almeder, C., Hartl, R. F., 2014. Simulation-based optimization methods for setting production planning parameters. International
Journal of Production Economics 151, 206–213.
21
[9] Gold, H., 2004. Dynamic optimization of routing in a semiconductor manufacturing plant. In: Hein Fleuren, Dick den Hertog, P. K. (Ed.), Operations
research proceedings 2004: selected papers of the annual international conference of the German Operations Research Society (GOR); jointly organized with the Netherlands Society for Operations Research (NGB). German Operations Research Society (GOR) and the Netherlands Society for
Operations Research (NGB), Tilburg, The Netherlands, pp. 76–83.
[10] Gold, H., 2008. Lastverbund, european patent ep1567921.
[11] Ham, M., Lee, Y. H., Fowler, J. W., 2009. Integer programming-based
real-time scheduler in semiconductor manufacturing. In: Winter Simulation
Conference. Winter Simulation Conference, pp. 1657–1666.
[12] Harrison, J. M., Lopez, M. J., 1999. Heavy traffic resource pooling in
parallel-server systems. Queueing Systems 33, 339–368.
[13] Heller, I., Tompkins, C., 1956. An extension of a theorem of Dantzig’s.
Linear inequalities and related systems 38, 247–254.
[14] Hochbaum, D. S., 1996. Approximating covering and packing problems:
set cover, vertex cover, independent set, and related problems. In: Approximation algorithms for NP-hard problems. PWS Publishing Co., pp.
94–143.
[15] Hutcheson, J. D., 2000. Introduction to semiconductor equipment. Handbook of Semiconductor Manufacturing Technology, 23–33.
[16] Juan, A. A., Faulin, J., Grasman, S. E., Rabe, M., Figueira, G., 2015. A
review of simheuristics: Extending metaheuristics to deal with stochastic
combinatorial optimization problems. Operations Research Perspectives 2,
62–72.
[17] Kashiwabara, T., Masuda, S., Nakajima, K., Fujisawa, T., 1992. Generation of maximum independent sets of a bipartite graph and maximum
cliques of a circular-arc graph. Journal of Algorithms 13 (1), 161–174.
[18] Lopez, M. J., Wood, S. C., 2003. Systems of multiple cluster tools: Configuration, reliability, and performance: Single-wafer manufacturing in the
nanochip era. IEEE transactions on semiconductor manufacturing 16 (2),
170–178.
[19] Mönch, L., Fowler, J. W., Dauzère-Pérès, S., Mason, S. J., Rose, O., 2011.
A survey of problems, solution techniques, and future challenges in scheduling semiconductor manufacturing operations. Journal of Scheduling 14 (6),
583–599.
[20] Ogryczak, W., 1997. On the lexicographic minimax approach to location
problems. European Journal of Operational Research 100 (3), 566–585.
22
[21] Ortner, A. M., 2008. capacity planning for cluster tools - a mathematical
model that reflects scheduling for parallel processing with two loadlocks.
Master thesis, Alpen-Adria-Universität Klagenfurt, Klangenfurt, Austria.
[22] Ponsignon, T., Mönch, L., 2012. Heuristic approaches for master planning
in semiconductor manufacturing. Computers & Operations Research 39 (3),
479–491.
[23] Romauch, M., Klemmt, A., 2015. Product mix optimization for a semiconductor fab: Modeling approaches and decomposition techniques. Computers & Operations Research 53, 338 – 352.
URL http://www.sciencedirect.com/science/article/pii/
S0305054814001087
[24] Sorenson, C. T., 1999. Basic principles of semiconductor manufacturing.
Microelectronics Manufacturing and the Environment Course Materials,
http://www.erc.arizona.edu/education/mme%20course%
20materials/MME%20Modules/Manufacturing%20Module/
Manufac%20Overview.ppt, accessed: 2016-04-07.
[25] Thiesse, F., Fleisch, E., 2008. On the value of location information to lot
scheduling in complex manufacturing processes. International Journal of
Production Economics 112 (2), 532–547.
[26] Tian, K., Jiang, Y., Shen, X., 2009. A study on optimally co-scheduling
jobs of different lengths on chip multiprocessors. In: Proceedings of the 6th
ACM Conference on Computing Frontiers. CF ’09. ACM, New York, NY,
USA, pp. 41–50.
URL http://doi.acm.org/10.1145/1531743.1531752
AppendixA. Algorithm to find a minimal representation for the dual
based model
With respect to (6-9) it is important to keep the set K as small as possible.
Therefore, the concept of redundant vectors (redundant minimal basic cuts) will
be introduced and different criteria for redundancy will be stated and proven.
According to (38), vectors Πk := (Πr,k ) ∈ RR are used as a lower bound on ρ,
i.e. h1 − Πk , xi ≤ ρ. Suppose for two vectors Πk , Πk0 the following condition
true: h1 − Πk , xi ≤ h1 − Πk0 , xi ≤ ρ. Then Πk0 dominates Πk and Πk can be
removed. The condition can be reformulated to hΠ, xi ≥ hΠk0 , xi and leads to
the following more general definition of redundancy:
Definition 2. Assuming that a set A of m vectors, A = {a1 , a2 . . . , am } with
ai ∈ Rn and ai ≥ 0 is given. Then a vector b is called redundant if the following
is true:
∀x ≥ 0 : hb, xi ≥ mini hai , xi
23
Or, in other words b is called redundant if ∀x ≥ 0∃i : hb, xi ≥ hai , xi. Note,
that it is sufficient to consider vectors x of unit length, therefore b is called
redundant if:
∀x ≥ 0, kxk = 1∃i : hb, xi ≥ hai , xi ⇔
∀x ≥ 0, kxk = 1∃i : h1 − b, xi ≤ h1 − ai , xi ⇔
∀x ≥ 0, kxk = 1∃i : kh1 − b, xi xk ≤ kh1 − ai , xi xk
Geometrically, h1 − b, xi x can be interpreted as the projection of 1 − b on x.
Also note that
i
i
h1 − b, xi x − 1−b
= 1−b
and h1 − ai , xi x − 1−a
= 1−a
for kxk =
2
2
2
2
1, therefore the statement is equivalent to:
S
i
U +1−b ( 1−b
) ⊂ i U +1−ai ( 1−a
)
2
k 2 k
k 2 k 2
where U+ (b) = {x ∈ Rn : x ≥ 0, kx − bk ≤ }.
Figure A.8 depicts an example in three dimensions. If the sphere U +1−b ( 1−b
)
k 2 k 2
S +
i
is contained in i U 1−ai ( 1−a
) then b is redundant. The geometric interprek 2 k 2
tation can be used to compute if a vector b is redundant or not:
) defines
If b is not redundant, then some part of the boundary of U +1−b ( 1−b
k 2 k 2
S +
i
the boundary of the union of spheres i U 1−ai ( 1−a
). The boundary of the
k 2 k 2
S +
1−ai
union of spheres i U 1−ai ( 2 ) can be described by transformed facets like
k 2 k
in Figure A.8 where the white region is touching three neighboring regions. If
it is possible to find two facets that belong to two spheres (i, j) that are neighboring to the sphere U +1−b ( 1−b
), then a hyperplane Hij exists that contains
k 2 k 2
the intersection of the boundaries of the spheres (in 3d it contains a circle).
Furthermore, if b is not redundant this hyperplane has a nonempty intersection
with the facet that corresponds to b. Therefore, a vector can be found that is on
the hyperplane and outside the union. If no such two spheres can be found then
the dominating part of the region intersects with at least one of the hyperplanes
that are defined by xi = 0.
Therefore, b is not redundant if:
S
i
U +1−b ( 1−b
)|Hij 6⊂ i U +1−ai ( 1−a
)|
or
2
k 2 k
k 2 k 2 Hij
S
i
U +1−b ( 1−b
)|
6⊂ i U +1−ai ( 1−a
)|
k 2 k 2 xi =0
k 2 k 2 xi =0
The dimension of the corresponding sub problems is n − 1, the method can
be reapplied until the dimension of all subproblems is one. Note that if b
i
is not redundant then it is possible to identify one or more spheres U +1−ai ( 1−a
)
k 2 k 2
that contain b, the set of the corresponding indices is called M , and M 6= ∅,
otherwise it is trivial that b is not redundant. Also suppose that U +1−b ( 1−b
) is
k 2 k 2
i
not fully contained in one of the U +1−ai ( 1−a
) (i ∈ M ), otherwise b is clearly
k 2 k 2
dominated. For the procedure it is not necessary to consider all possible Hij ,
it is sufficient to assume that i ∈ M . The corresponding procedure is not very
effective and the worst case time complexity is larger than O(mn ).
24
In the following, an alternative method, that is based on convex hulls is
presented:
Proposition 3. criterion for redundancy: A vector b is redundant with respect
to a set of vectors {ai }, if and only if it is contained in the convex hull h{ai }i.
Proof. the negation redundancy is the following:
∃x ≥ 0
∀i :
hb, xi > hai , xi
∃x ≥ 0
∀i :
hb − ai , xi > 0
(A.1)
According to Gale [7] (Theorem 2.10, p. 49, which is a corollary of Farkas’
Lemma) the corresponding elimination criterion has the form:
∃λ ≥ 0(λ 6= 0)
It is also possible to assume that
∃λi > 0
P
i6=j
P
i6=j
b≤
λi (b − ai ) ≤ 0
λi = 1 and therefore:
P
i6=j
λi ai
This condition states that b is contained in the convex hull of {ai }i6=j . Therefore the problem is solvable by calculating the convex hull; and according to
n
Chazelle [2], the convex hull is computable in O(m log(m) + m 2 ) time.
A further alternative that avoids the calculation of the convex hull ( see
Chazelle [2]) is presented in the following proposition. Among the presented
methods it is the most efficient one.
Proposition 4. criterion for redundancy: A vector b is redundant if and only
if the following LP is infeasible.
min
1t x
(b − ai )t x ≥ 1
∀i
(A.2)
x≥0
Proof. ⇒: A feasible solution x of the LP also satisfies (A.1) and therefore the
vector b is redundant - the objective is irrelevant.
⇒: On the other hand, b is redundant and satisfies (A.1), therefore a vector
x0 and a positive number > 0 exist such that (b − ai )t x > . Hence, x is a
feasible solution of the LP and the proof is complete. The complexity is defined
by the complexity of linear programming algorithms. Algorithms with O(n3 L)
are known, where L is the bit length of the input data.
25
Figure A.8: A three dimensional example with four vectors, where the corresponding set of
vectors that is non-dominating is represented by the union of spheres.
cut
A
e B,
e AB}
g
Π1 :
{A, B, AB, A,
1
e B,
e C}
e
Π2 :
{A, B, C, A,
1
e
e
Π3 :
{A, B, C, AB, A, B}
1
1
(Π
+
Π
)
:
1
1
2
2
B
1
1
1
1
C
AB
1
AC
BC
1
1
2
1
2
1
2
1
2
Table A.14: minimal basic cuts and redundancy
Definition 5. Note, that a set of vectors A = {ai ∈ [0, ∞)n } is called minimal
if none of its members is redundant with respect to A \ {ai }, more precisely:
∀j∃x ≥ 0 ∀i 6= j : aj t x < ati x.
Finally, to show that not all minimal basic cuts are needed in the LP formulation, a simple example that considers three chambers is presented:
Example 6. For three chamber, Table A.14 lists three of the minimal basic cuts,
named Π1 , Π2 , Π3 . Obviously, Π3 = 21 Π1 + 12 Π2 , therefore the criterion formulated in Proposition 3 shows that Π3 is redundant with respect to {Π1 , Π2 }.
| 8cs.DS
|
Hierarchical Actor-Critic
Andrew Levy 1 Robert Platt 2 Kate Saenko 1
arXiv:1712.00948v3 [cs.AI] 28 Feb 2018
Abstract
The ability to learn at different resolutions in time
may help overcome one of the main challenges in
deep reinforcement learning — sample efficiency.
Hierarchical agents that operate at different levels of temporal abstraction can learn tasks more
quickly because they can divide the work of learning behaviors among multiple policies and can
also explore the environment at a higher level. In
this paper, we present a novel approach to hierarchical reinforcement learning called Hierarchical
Actor-Critic (HAC) that enables agents to learn to
break down problems involving continuous action
spaces into simpler subproblems belonging to different time scales. HAC has two key advantages
over most existing hierarchical learning methods:
(i) the potential for faster learning as agents learn
short policies at each level of the hierarchy and
(ii) an end-to-end approach. We demonstrate that
HAC significantly accelerates learning in a series
of tasks that require behavior over a relatively
long time horizon and involve sparse rewards.
1. Introduction
Despite major successes in both simulated and real-world
tasks, a key problem with many deep reinforcement learning
(RL) algorithms is that they are slow. Learning is particularly slow when the rewards granted to agents are sparse.
One major reason for reinforcement learning’s poor sample
efficiency is that many existing algorithms force agents to
learn at the lowest level of temporal abstraction. For instance, if a simulated robot agent is given a task involving
locomotion, the agent will need to learn the entire sequence
of joint torques to accomplish the task (Lillicrap et al., 2015)
instead of trying to break the problem down at a higher level.
Learning exclusively at low levels of abstraction slows down
learning for two key reasons. First, agents must learn longer
sequences of actions in order to achieve the desired behavior. This is problematic because policies involving longer
1
Department of Computer Science, Boston University, Boston,
MA, USA 2 College of Information and Computer Science, Northeastern University, Boston, MA, USA.
sequences of actions are more difficult to learn, particularly
when rewards are sparse. The process of propagating back
Q-values from the actions that produce the sparse reward to
the preceding actions takes longer. The issue of long-term
credit assignment also becomes more severe as the actionvalue function needs to learn Q-values for a larger portion
of the state-action space. Second, learning at the lowest
level restrains exploration. Agents that can propose higher
level subgoals can more quickly determine the distant states
that are helpful in achieving certain behavior goals. A faster
exploration of the state space of the environment may speed
up the process of learning a robust policy.
Yet most existing hierarchical RL methods do not provide
an approach for breaking down tasks involving continuous
action spaces that guarantees shorter policies at each level
of abstraction and is end-to-end. Most current hierarchical
approaches only enable agents to learn at higher levels if the
action space is discrete (Dayan & Hinton, 1993) (Vezhnevets
et al., 2017). Further, many existing hierarchical learning
approaches do not implement hierarchical agents that equitably divide up the work of learning a behavior among the
agent’s multiple policies. For instance, many approaches
choose to decompose problems into smaller state spaces
rather than into smaller time scales (Dayan & Hinton, 1993).
This can be problematic in continuous action space environments as a policy that acts within a small region of the state
space may need a lengthy sequence of actions to escape that
region. Most existing hierarchical approaches also require
non-trivial manual work including designing non-sparse reward functions, preselecting the set of possible higher level
subgoals rather than learning them from experience, and
staggering the training of different policies (Sutton et al.,
1999) (Kulkarni et al., 2016).
In this paper, we introduce a novel approach to hierarchical reinforcement learning called Hierarchical Actor-Critic
(HAC). The algorithm enables agents to learn to divide tasks
involving continuous state and action spaces into simpler
problems belonging to different time scales. HAC achieves
this objective by implementing agents that learn multiple
policies in parallel. Each successive policy in the hierarchy
is responsible for learning how to break down problems
into subproblems with increasingly fine time resolutions.
Figure 1 should provide some intuition on how HAC agents
learn at different time scales. The figure shows an agent
Hierarchical Actor-Critic
Start
Goal
High-Level
Policy
Goal
Mid-Level
Policy
action. Each goal-based actor network learns limited length
policies that operate at different time resolutions due to a
critical feature of the algorithm — time limits. Each actor
network has only a certain number of actions to achieve its
higher level input goal. Section 3 explains how time limits
enable each actor network to specialize in a different time
scale.
Another key advantage of HAC is that it provides an end-toend hierarchical learning approach. HAC learns to separate
goals into subgoals using just the agent’s experience and
the algorithm only requires sparse reward functions. The
hierarchical policies are also learned in parallel and do not
need to be learned in different phases.
Goal
Low-Level
Policy
Time
Figure 1. Example HAC Hierarchy. Agent in figure uses three
policies to learn a behavior. Each policy specializes in breaking
down problems into subproblems with finer time resolutions.
that uses three policies to accomplish some behavior. The
solid vertical lines represent the time resolutions of subgoals output by each policy. The more distance there is
between consecutive vertical lines, the more time the agent
has to achieve each subgoal. The low-level policy outputs
actual agent actions so the vertical lines for the low-level
policy can be interpreted as subgoals requiring one action.
In the figure, the high-level policy breaks down the end goal
into three subgoals with relatively large time resolutions.
The mid-level policy specializes in dividing each subgoal
from the high-level policy into three subgoals belonging to
shorter time scales. Finally, the low-level policy specializes
in decomposing each subgoal from the mid-level policy into
three agent actions, which represent the smallest time resolution. A crucial benefit of having each policy specialize in
breaking down goals of particular time scale into subgoals
of a certain smaller time scale is that the policies that are
learned are limited in length. This is beneficial because
shorter policies can be learned more quickly than longer
ones. Further, having multiple policies that operate at different levels of temporal abstraction is helpful because it
enables high-level exploration, which can also accelerate
learning.
Hierarchical Actor-Critic helps agents learn a hierarchy of
policies similar to Figure 1 using a set of actor-critic networks. Each actor-critic network is responsible for learning
one of the policies within the hierarchy. The policies or
actor networks that are learned are goal-based, meaning that
they take as input the current state and a goal and output an
For this paper, we ran a series of experiments comparing the performance of agents that did and did not use
the Hierarchical Actor-Critic algorithm. The tasks examined include pendulum, reacher, cartpole, and pick-andplace environments. In each task, agents that used Hierarchical Actor-Critic significantly outperformed those that
did not. In some tasks, the use of Hierarchical ActorCritic appears to be the difference between consistently
solving a task and rarely solving a task. A video showing the results of our experiments is available at https:
//www.youtube.com/watch?v=m3EYeBpGepo.
2. Background
Hierarchical Actor-Critic builds off three techniques from
the reinforcement learning literature: (i) the Deep Deterministic Policy Gradient (DDPG) learning algorithm (Lillicrap
et al., 2015), (ii) Universal Value Function Approximators
(UVFA) (Schaul et al., 2015), and (iii) Hindsight Experience
Replay (HER) (Andrychowicz et al., 2017).
DDPG serves as the key learning infrastructure within Hierarchical Actor-Critic. DDPG is an actor-critic algorithm and
thus uses two neural networks to enable agents to learn from
experience. The actor network learns a deterministic policy
that maps from states to actions π : S → A. The critic
network approximates the Q-function or the action-value
function of the current policy Qπ (st , at ) = E[Rt |st , at ],
in
which Rt is the discounted sum of future rewards
P∞
i−t
ri . Thus, the critic network maps from (state,
i=t γ
action) pairs to expected long-term reward Q : S × A → R.
In order to learn a near-optimal policy that results in large
expected long-term reward, DDPG follows a cyclical process composed of two steps: (i) policy evaluation and (ii)
policy improvement. In the policy evaluation phase, the
agent first interacts with the environment for a period of
time using a noisy policy π(s) + N (0, 1), in which N (·)
is some normal distribution. The transitions experienced
are stored as (st , at , rt , st+1 ) tuples in a replay buffer. The
agent then updates its approximation of the Q-function of
Hierarchical Actor-Critic
the current policy by performing mini-batch gradient descent on the loss function L = (Q(st , at ) − yt )2 , in which
the target yt is the Bellman estimate of the Q-function
yt = rt + γQ(st+1 , π(st+1 )). In the policy improvement
phase, the agent modifies its policy based on the updated
approximation of the action-value function. The actor function is trained by moving its parameters in the direction of
the gradient of Q w.r.t. the actors parameters.
Universal Value Function Approximators is a second idea
that is critical to HAC. UVFA extends the action-value function to incorporate goals. The Q-function now represents
the expected long-term reward of taking an action given
the current state and goal Qπ (st , at , gt ) = E[Rt |st , at , gt ].
Each goal has its own reward function rg (st , at , st+1 ) and
discount function γg (s). γg (s) = 0 when the agent is in a
state that achieves the prescribed goal as the current state
can be viewed as a terminating one. Goals are critical to
HAC because goals are often hierarchical and can be broken
down into subgoals. Goals are also useful because they can
be used easily with sparse and binary reward functions.
Hindsight Experience Replay is another component from
the reinforcement learning literature that is integral to Hierarchical Actor-Critic. HER helps agents learn goal-based
policies more quickly when sparse reward functions are
used. The idea behind HER is that even though an agent
may have failed to achieve its given goal in an episode, the
agent did learn a sequence of actions to achieve a different objective in hindsight — the state in which the agent
finished. Learning how to achieve different goals in the
goal space should help the agent better determine how to
achieve the original goal. Hindsight Experience Replay is
implemented by creating a separate copy of the transitions
(st , at , rt , st+1 , g) that occurred in an episode and replacing
(i) the original goal with the goal achieved in hindsight and
(ii) the original reward with the appropriate value given the
new goal.
3. Hierarchical Actor-Critic
We introduce a new hierarchical RL approach called Hierarchical Actor-Critic. The algorithm helps agents learn long
time horizon tasks involving continuous action spaces and
sparse rewards more quickly by enabling agents to learn to
break down those tasks into easier subtasks belonging to
different time scales. HAC directly addresses the issue of
lengthy policies that hinder many existing non-hierarchical
and hierarchical approaches as HAC agents learn limited
policies at each level of temporal abstraction. The approach
is also end-to-end as it learns subgoal policies at different
levels of temporal abstraction on its own and in parallel and
only requires sparse reward functions.
Action
Neural
Network
State
Low-Level
Actor
Subgoal
Neural
Network
State
High-Level
Actor
Goal
Figure 2. Hierarchical policy with 1 subgoal layer
3.1. Architecture
The objective of the algorithm is to learn a hierarchical
policy like the one shown in Figure 2. The hierarchical
policy is composed of multiple goal-based policies or actor
networks. Each actor network takes as input the current
state and higher level goal and outputs an action belonging
to a particular time scale. For the subgoal actor networks,
such as the bottom network in Figure 2, this action is a
proposed subgoal. The proposed subgoal is a desired future
state or set of future states for the agent. For the actor
network operating at the lowest level of abstraction, such
as the top network in Figure 2, the action is the agent’s
actual output. In our experiments, we trained agents that
used hierarchical policies composed of two and three actor
networks. Additional layers can be easily added.
Each actor network has its own critic network and replay
buffer to learn a near-optimal policy. The actor networks
from Figure 2 are shown connected to their respective critic
networks in Figure 3. Each critic network approximates
the Q-function for its associated policy Qπi (st , at , gt ) =
E[Rt |st , at , gt ] using the Bellman equation as a target yt =
rg + γg Qπi (st+1 , πi (st+1 ), gt ). rg is sparse and binary and
is granted when the agent has reached a state within a certain
distance of the goal. As described in (Schaul et al., 2015),
γg = 0 when the prescribed goal has been achieved as a
terminating state has been reached.
3.2. Temporal Abstraction via Limited Policies
Each actor network within the hierarchical policy learns a
limited length policy as a result of time limits. Actor net-
Hierarchical Actor-Critic
Q-Value
Neural
Network
State
Subgoal
Low-Level
Actor-Critic
Action
Neural
Network
State
Subgoal
Q-Value
Neural
Network
State
Goal
High-Level
Actor-Critic
Subgoal
Neural
Network
State
Goal
Figure 3. Actor-Critic networks for hierarchical policy with 1 subgoal layer.
works can only take a certain number of actions to achieve
their higher level goal. For the lowest level actor network,
this means it can only execute a certain number of actual
agent actions to achieve a subgoal. For higher level subgoal
actor networks, the policy limit means the network must
achieve its higher level goal within a maximum number
of subgoals. The policy length limit for actor network i is
controlled by a hyper parameter Ti . In our experiments, we
generally used the same value for Ti for each actor network
within the hierarchical policy. This limit thus restricts each
actor network to only learning how to achieve goals that can
be accomplished within a certain number of actions, which
shortens the goal-based policy that is learned.
The combination of time limits and Hindsight Experience
Replay enables each actor network to specialize in different
time resolutions. Actor networks learn to operate at different time scales largely as a result of hindsight learning, in
which the agent learns how to achieve the goal states the
agent actually reached during an episode with the help of
HER. The following example should provide some clarity.
Consider an agent using a two layer hierarchical policy with
T0 = T1 = 10, meaning that the agent must achieve each
subgoal in no more than 10 agent actions and achieve the
end goal in no more than 10 subgoals. Even in the worst
case scenario in which the agent fails after 100 agent actions
to achieve any of the 10 subgoals and the end goal, the agent
will still be able to learn how it could have divided up the
task of achieving the final state into problems belonging
to different time scales. In order to achieve the last state
reached on the 100th action, the agent could have chosen
every 10th state to be a subgoal state. This is a valid breakdown because the end goal is reached in no more than 10
subgoals and each subgoal is achieved in no more than 10
actions by the agent. From this breakdown, the higher level
network’s replay buffer receives a sequence of 10 transitions
showing how it can use subgoals belonging to larger time
resolutions to achieve the hindsight end goal. Similarly,
the low-level network’s replay buffer receives 10 sequences
of 10 transitions each showing how it can use actions belonging to the smallest time scale to achieve each of the
10 subgoals. Over the course of many episodes, each actor
network learns to achieve goals with actions belonging to
its respective time scale. Thus, time limits are critical for
helping each actor network specialize in a different time
resolution because they provide a simple way to divide up a
sequence of actions that achieved some goal into multiple
sequences belonging to different time scales.
Learning limited policies at different time resolutions
presents significant benefits as well as new challenges for
the agent. The key benefit is that it should be easier to
learn multiple shorter policies in parallel than one long policy. Credit assignment is less of a problem as the critic
function for each policy only needs to learn Q-values for a
more limited region of the state-action-goal space. Learning
is faster because reinforcement learning agents essentially
learn from end to beginning when sparse rewards are used.
This backwards learning process occurs more quickly if the
policy requires fewer actions. However, the use of limited
policies also results in a significant new dilemma — subgoal actor networks now have conflicting missions. Subgoal
actor networks need to learn a policy that can simultaneously (i) achieve its higher level goal in as few actions (i.e.,
subgoals) as possible but also (ii) output subgoals that can
be achieved by the lower-level actor network in a limited
number of steps. Producing subgoals that are too ambitious
may not be achievable by lower level layers as they specialize in limited policies. Overly ambitious subgoals could
thus result in the failure of the higher level actor network
to achieve its own goal. To simultaneously solve both of
its conflicting objectives, the upper and lower level layers
need to coordinate as the upper level needs to understand
the types of subgoals the lower level can accomplish.
We take two steps to incentivize subgoal actor network i to
output subgoals that can be achieved by actor network i − 1
in no more than Ti−1 actions. First, as in the example discussed above, all experience transitions passed to the replay
buffers of subgoal actor networks contain actions (i.e. subgoals) that were actually achieved by the succeeding, lower
level actor network within the maximum number of actions.
Second, HAC penalizes proposed subgoals that were not
achieved. A certain percentage of the time defined by a hy-
Hierarchical Actor-Critic
Algorithm 1 Hierarchical Actor-Critic
Initialize Actor-Critic networks [(π0 , Q0 ), . . . , (πn , Qn )]
Initialize Replay Buffers (R0 , . . . , Rn )
for episode = 1 to M do
Sample actual goal, G, and initial state, s0
for tn = 1 to Tn do
Testing Boolean B ← {0, 1}
Sample layer n subgoal from πn
gtn ← πn (s(tn ), G) + B · N (0, 1)
...
for t0 = 1 to T0 do
Sample action from π0
at0 ← π0 (s(t0 ), g1 ) + B · N (0, 1)
ProcessTrans(st0 ,at0 ,rt0 +1 ,st0 +1 ,gt1 )
if gt1 , ..., G achieved or t0 = T0 then
Perform HER on layer 0 transitions
break
end if
end for
...
ProcessTrans(stn ,gtn ,rtn +1 ,stn +1 ,G)
if G achieved or tn = Tn then
Perform HER on layer n transitions
break
end if
end for
Update Actor-Critic Networks
end for
per parameter, agents will test subgoals by not adding noise
to the subgoals and actions prescribed by its hierarchical
policy. Noise needs to be removed when testing subgoals
because an agent may miss a subgoal due to noise added to
lower level actions. Subgoals that cannot be achieved with
the agent’s current, noise-free policy will be penalized. In
our experiments, if layer i proposed a subgoal that was not
achieved, layer i received a reward of −Ti . In addition, we
set γ g = 0 when a subgoal is missed during testing as the
Q-value of a missed proposed subgoal should not depend on
the Q-value of a different state. If γ g = 0, then the reward
for proposing the subgoal must be −Ti or less. Otherwise,
proposing the missed subgoal may have a larger Q-value
than some subgoals that do efficiently guide the agent to
the goal, and the subgoal actor network may not update
its policy as a result. We observed that the extra penalty
for proposing a missed subgoal works as intended as the
subgoal layer would adjust to proposing more conservative
subgoals that the lower-level layer can achieve.
3.3. Algorithm
The main part of the HAC algorithm, shown in Algorithm 1,
is concerned with ensuring the correct experience transitions
Algorithm 2 Transition Processing
Function ProcessTrans(sti ,ati ,rti +1 ,sti +1 ,gti+1 )
if subgoal layer and ati not achieved then
Ri ← (sti , gt0 i , rti +1 , sti +1 , gti+1 )
if Testing then
Ri ← (sti , ati , −Ti , sti +1 , gti+1 )
end if
else
Ri ← (sti , ati , rti +1 , sti +1 , gti+1 )
if subgoal layer then
Store HER trans (sti , gt0 i , T BD, sti +1 , T BD)
else
Store HER trans (sti , ati , T BD, sti +1 , T BD)
end if
end function
are passed to each actor network’s replay buffer. In order
to learn policies of a certain length, actor networks need
to be passed sequences of transitions of the same length.
Also, as discussed above, in order for each subgoal actor
network to learn to output effective subgoals that belong to
its respective time resolution, certain transitions need to be
passed to each actor network.
HAC proceeds as follows. The initial state and end goal
are passed to the actor network n, which then triggers the
passing of a series of subgoals from higher level actor networks to lower level actor networks. If the agent is in a
subgoal testing phase, noise will not be added to these subgoals. The cascade of subgoals terminates when the last
subgoal is passed to the lowest level actor function, which
is located in the innermost for loop. The lowest level actor
then has T0 attempts to try to achieve the provided subgoal. After each action by the lowest level actor network,
two copies of transitions are created as a result of the call
to the ProcessTrans function shown in Algorithm 2. The
first transition (st0 , at0 , rt0 +1 , st0 +1 , gt1 ) is placed in replay
buffer R0 . This transition indicates whether the action at0
taken in state st0 was able to achieve goal gt1 . The second transition (st0 , at0 , T BD, st0 +1 , T BD) awaits further
processing during Hindsight Experience Replay. Once the
lowest level actor has run out of T0 attempts or achieves
any of the higher-level goals gt1 ,...,G, Hindsight Experience
Replay is executed. During HER, a few random transitions
are selected to serve as hindsight goals. As a result, for the
selected transition and all preceding transitions, the TBD
goal component is replaced with gt0 1 = f (st1 +1 ), in which
f (·) is a function that transforms an array from the state
space to the the goal space. gt0 1 can thus be interpreted as
the subgoal achieved in hindsight. The TBD reward component is then replaced with its appropriate value given the
updated goal. The purpose of these HER transitions is thus
to help the agents learn the sequences of actions that were
Hierarchical Actor-Critic
used to achieve hindsight goals, even though these hindsight
goals may be different than the original target goals. After
HER is executed for the lowest level actor network, control
is then passed to the preceding higher level actor network.
were thereby only using DDPG and HER. A video showing
the results of our experiments is available at https://
www.youtube.com/watch?v=m3EYeBpGepo.
The completion of the innermost for loop consists of 1
subgoal action for actor network 1. Different transitions
will be created depending on (i) whether its original proposed subgoal gt1 was achieved by actor network 0 and
(ii) whether the agent is currently in a testing phase. If
the proposed subgoal was not achieved, the transition
(st1 , gt0 1 , rt1 +1 , st1 +1 , gt2 ) is passed to replay buffer R1 .
This transition indicates whether the subgoal achieved in
hindsight, gt0 1 , was able to achieve the higher level goal
gt2 . If the proposed subgoal was missed and the agent is
currently testing subgoals, the actor network 1 is penalized
with the transition (st1 , gt1 , −T1 , st1 +1 , gt2 ). On the other
hand, if subgoal gt1 was achieved by actor network 0, the
transition (st1 , gt1 , rt1 +1 , st1 +1 , gt2 ) is passed to R1 . Finally, an additional transition (st1 , gt0 1 , T BD, sti +1 , T BD)
is created for Hindsight Experience Replay. Actor network 1
then proposes a second subgoal for actor network 0 and the
process begins again. The key insight here is that the second
component or the ”action” component in every transition
passed to the subgoal actor networks identifies a subgoal
that has been achieved by the lower level actor network in
no more than Ti−1 steps, except for the transitions that penalize bad subgoals. As a result, each subgoal actor network
learns to output actions that both belong to a a particular
time resolution and are helpful towards achieving the actor
network’s higher level goal.
4.1. Environments
After interacting with the environment for one or more
episodes and filling its replay buffers with transitions, the
agent now learns from the experience by updating its actorcritic functions. Mini-batch gradient descent is performed
on the critic network to push the Q-function closer to the
Bellman estimates. Next, the parameters of the actor function θi are moved in the direction of the gradient of the
Q-function w.r.t θi . For the lowest level actor network, updating its actor-critic network should enable the actor to find
more efficient routes to achieving its shorter term goals. For
each subgoal actor-critic layer, updating its actor-critic network means finding a better way to balance its conflicting
objectives of finding the sequence of subgoals that can (i)
most quickly solve the higher level goal and (ii) be achieved
in no more than Ti−1 steps by the succeeding actor layer.
4. Experiments
We evaluated the Hierarchical Actor-Critic approach on a
total of 5 tasks: Pendulum-Climb, Pendulum-Balance, UR5
Reacher, Cartpole, and 1-Object Pick-and-Place. In each
environment, we compared the performance of agents using
0, 1, and 2 subgoal layers. Agents using 0 subgoal layers
We assessed HAC on the following 5 tasks. Each of these
simulations was built using the Mujoco physics engine
(Todorov et al., 2012).
1. Pendulum - Climb
The goal of this task is for the agent to swing the pendulum to its maximum height, marked by a yellow
cube. The agent only needs to touch the yellow sphere
located at the peak and does not need to try to balance
the pole. We found that an efficient policy could solve
this task in around 100 low-level actions.
2. Pendulum - Balance
The goal for this environment is to balance the pendulum at its peak near the yellow sphere. Thus, to
achieve the goal the pole must be located near the peak
and have angular velocity near 0. Figure 4 shows a
few frames from a successful episode. We found that
an efficient policy could solve this task in around 150
low-level actions.
3. UR5 Reacher
The goal of this task is for the agent to learn to move to
a randomly designated point, marked by a yellow cube.
The agent in this task is a simulated UR5, a 6 DOF
robotic arm. To make the task require a longer time
horizon, the goal location is always in the quadrant in
front and opposite the starting location of the gripper.
We found that an efficient policy could solve this task
in around 60 individual actions.
4. Cartpole Swingup
The goal for this task is to swing the pole up to the
yellow cube. In order to achieve the goal, the angular
velocity of the pole must also be near 0 and the position
of the cart must be below the yellow cube. We found
that an efficient policy could solve this task in around
170 low-level actions.
5. 1-Object Pick-and-Place
The idea for this task was to assess how Hierarchical
Actor-Critic would perform in a task with natural hierarchy. The objective in this task is to pick up the blue
rod and move it to the yellow rod. The agent is a 2
joint robot worm. Our more efficient agents can solve
this task in around 110 steps.
States, Actions, Rewards
The state space in all environments include joints positions
and joint velocities. The actions are joint torques. Further,
Hierarchical Actor-Critic
Figure 4. Three frames from a successful episode of the Pendulum - Balance task. The agent learns to reach the goal (touch the yellow
sphere) by breaking down the task into subgoals (touch the purple spheres).
Table 1. Environment subgoal and goal descriptions
TASK
P ENDULUM -C LIMB
P ENDULUM -BALANCE
UR5 R EACHER
C ARTPOLE
P ICK - AND -P LACE
S UBGOAL
G OAL
θxy
θxy , θ̇
θxy
θxy , θ̇
θef f ectorxyz
θxy , θ̇
θxy
θef f ectorxyz
θxy , θ̇
θxy
each network within the stack has its own sparse reward
function. In our experiment, a reward of -1 was granted if
an actor network took an action that did not complete its
goal and a reward of 0 if the action was successful. Also,
a larger negative reward was issued for subgoal networks
that proposed a subgoal that the lower level layers could not
achieve. Rewards are granted by a layer if the agent has
moved within a certain distance of the layer’s goal. Further,
the rewards granted by each network are independent of
each other. For instance, an agent may miss the low-level
subgoal but achieve a high-level subgoal. In this case, the
transition passed to the low-level replay buffer will still list
the negative reward for taking that action given the current
state and subgoal because the action was unsuccessful.
Goals and Subgoals
Table 1 shows the subgoals and end goals for each task. θxy
and θ̇ represent Cartesian joint positions and joint velocities,
respectively. The subgoals in our experiments were essentially desired future states that the agent learns to achieve
along the way to reaching the end goal. For some of our
experiments, we used a lower dimensional version of the
state as the subgoal.
The end goal shown in the last column of Table 1 depends
on the behavior the user would like the agent to learn. The
more specific the desired behavior, the more similar the end
goal should be to the state. For instance, in our Pendulum
- Climb task, the agent just needed to learn to swing up to
the peak in as few actions as possible. The agent did not
need to try to balance the pole. Thus, for this task the end
goal was 2-dimensional and includes the (x,y) coordinates
of maximum height of the pendulum. On the other hand,
in the Pendulum - Balance task, the agent needed to both
swing the pole to its peak and maintain a near 0 angular
velocity when the pole was upright. Thus, the goal for this
task was 3-dimensional and included the (x,y) coordinates
of maximum height of the pendulum and the desired angular
velocity of 0.
4.2. Results
The accompanying video and Figure 4.1 show the results
of our experiments. In Figure 4.1, each chart plots the
percentage of goals achieved by agents using 0, 1, and 2
subgoal layers in each testing period. Testing periods are
separated by about 300 episodes and each testing period
consists of 64 episodes. Each plot represents the average
performance over 7-10 runs of each task. Agents learn each
task from scratch and thus begin with no pre-training. In
episode 0, agents are thereby using random policies.
The key result from our work is that the use of Hierarchical Actor-Critic resulted in a substantial improvement in
performance. In all 5 environments, agents that used HAC
learned a robust policy significantly faster. Indeed, for many
tasks, agents that used no subgoal layers were not able to
consistently solve the task.
The video and the frames in Figure 4 also demonstrate that
the algorithm is working as intended – the agents are learning how to break problems down into easier subproblems.
In the Pendulum - Balance environment, the subgoals represented by purple spheres show that the agent has learned
that if it wants to balance the pole upright it first needs to
swing back and forth. In the UR5 Reacher environment,
if the goal location marked by the yellow cube is far away
from an agent using 2 subgoal layers, the agent will often
put the high-level subgoal marker, the green cube, about
halfway to the end goal. The low-level subgoals, marked
by the purple cubes, then appear to guide agent toward the
high-level subgoal. This indicates the agent has learned how
to effectively separate high-level goals into easier low-level
goals. Similar hierarchical behavior is evident in the pickand-place environments. In the 1-Object Pick-and-Place,
the high-level subgoal, represented by the green sphere and
Hierarchical Actor-Critic
Figure 5. The above plots compare the performance of agents that use the baseline DDPG+HER method with 0 subgoal layers (blue), and
our HAC method with 1 (red) and 2 (yellow) subgoal layers, for each of the five tasks. Performance is measured by the percentage of end
goals achieved in each testing period, which consists of 64 episodes.
rod, will often appear in the vicinity of the blue rod when
the blue rod has not yet been picked up. Once the blue rod
has been attached, the high-level subgoal moves to the yellow rod. The low-level subgoal, represented by the purple
sphere and rod, again guide the agent to the green high-level
subgoal.
We also observed that for most trials in each task except
for the pick-and-place task, agents using 2 subgoal layers
learned significantly faster than agents using 1 subgoal layer.
Agent using 2 subgoal layers learn an extra policy but each
of the three policies specialize in learning shorter policies
than the 2 policies learned by agents using 1 subgoal layer.
This result further supports the main premise of the algorithm — hierarchical agents that learn more shorter policies
in parallel can outperform agents learning fewer longer policies.
5. Related Work
Hierarchical RL is a topic of ongoing research (Sutton et al.,
1999), (Dayan & Hinton, 1993), (Vezhnevets et al., 2017),
(Dietterich, 1998). One popular hierarchical reinforcement
learning approach is feudal reinforcement learning (Dayan
& Hinton, 1993). In feudal reinforcement learning, the
state space is divided into increasingly small regions at each
level of abstraction. Dayan and Hinton (Dayan & Hinton,
1993) present a grid world example, in which a maze is
continually divided into quarters at each level. Each level
has a set of managers that are in charge of providing goals
and rewards to the 4 sub-managers below. Sub-managers
need to learn to complete their goals by learning how to give
tasks to their own sub-managers. Dayan and Hinton (Dayan
& Hinton, 1993) show that a feudal structure outperforms
a non-hierarchical Q-learning approach. A key difference
between HAC and feudal reinforcement learning is that the
latter breaks down problems along the spatial dimension
instead of the temporal dimension. This is problematic
for two reasons. First, it is unclear how the state space
would be divided for high-dimensional and continuous state
spaces. Second, even if there was a way to divide a highdimensional continuous state space, the feudal approach
does not guarantee the hierarchical policies learned will
each be short. There may be some small region of the
continuous state space that is difficult to maneuver and may
require many actions from a manager. HAC, on the other
hand, motivates its actor networks to learn shorter policies,
which can accelerate learning.
Another popular framework in hierarchical reinforcement
learning is the options framework (Sutton et al., 1999). This
approach generally uses a hierarchy of two layers to enable
agents to break problems down. The low-level layer consists
of multiple options, each of which is a policy that can solve a
specific task. The high-level layer is responsible for learning
the sequence of these specific policies that can achieve a
task. HAC uses a different approach to breaking problems
down. Instead of having the high-level policy select one
of many specific low-level policies, the high-level network
provides a subgoal to a single low-level network, which
is trained to achieve a variety of subgoals as it learns a
goal-based policy. Using one low-level goal-based policy
Hierarchical Actor-Critic
network instead of several non-goal-based policies should
provide some efficiency advantages because learning how
to achieve one subgoal will often help in learning how to
achieve different subgoals. For instance, in a pick-and-place
task, learning how pick up and drop off an object in a certain
location should help the agent learn to pick up and drop off
the object in a different target location.
Kulkarni et al.(Kulkarni et al., 2016) proposed an approach
with some similarities to both the options framework and
HAC. The algorithm, named hierarchical-DQN (h-DQN),
aims to help agents solve tasks in environments with discrete
action spaces. Agents implemented with h-DQN break
down tasks using two value functions. The high-level layer
attempts to learn a sequence of subgoals that can accomplish
a task. The low-level layer attempts to learn a sequence
of individual actions that can achieve the provided subgoal.
The low-level layer thus learns a goal-based policy and value
functions similar to HAC. However, unlike the Hierarchical
Actor-Critic method, h-DQN does not enable agents to learn
the sequence of high-level subgoals from scratch while using
only sparse reward functions. In the papers Montezumas
Revenge example, the agent was provided with the set of the
possible subgoals, which included objects in the game such
as doors, ladders, and keys. The agent was then responsible
for learning the order these items needed to be reached. An
external reward function was also used to help the agent
more quickly find the order of these subgoals. One key
reason Hierarchical Actor-Critic does not need aids like sets
of subgoals or manually-engineered reward functions is the
use of Hindsight Experience Replay. With HER, as long as
the agent can occasionally achieve goals that are nearby the
intended goal, the agent should have a chance to learn the
desired behavior.
6. Conclusion
We introduced a new technique called Hierarchical ActorCritic that uses temporal abstraction to break down complex
problems into easier subproblems. Our results indicate that
only using one policy to learn a challenging behavior in an
environment with sparse rewards can be problematic. A
better approach may be to learn a set of policies operating at
different time resolutions that work together to learn some
behavior.
Acknowledgements
This research was supported by NSF award IIS-1724237
and by DARPA.
References
Andrychowicz, Marcin, Crow, Dwight, Ray, Alex, Schneider, Jonas, Fong, Rachel, Welinder, Peter, McGrew, Bob, Tobin, Josh, Pieter Abbeel, OpenAI,
and Zaremba, Wojciech. Hindsight experience replay. In Guyon, I., Luxburg, U. V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., and Garnett,
R. (eds.), Advances in Neural Information Processing
Systems 30, pp. 5055–5065. Curran Associates, Inc.,
2017. URL http://papers.nips.cc/paper/
7090-hindsight-experience-replay.pdf.
Dayan, Peter and Hinton, Geoffrey E. Feudal reinforcement learning. In Hanson, S. J., Cowan, J. D., and
Giles, C. L. (eds.), Advances in Neural Information
Processing Systems 5, pp. 271–278. Morgan-Kaufmann,
1993. URL http://papers.nips.cc/paper/
714-feudal-reinforcement-learning.
pdf.
Dietterich, Thomas G. The maxq method for hierarchical
reinforcement learning. In In Proceedings of the Fifteenth
International Conference on Machine Learning, pp. 118–
126. Morgan Kaufmann, 1998.
Kulkarni, Tejas D, Narasimhan, Karthik, Saeedi, Ardavan,
and Tenenbaum, Josh. Hierarchical deep reinforcement
learning: Integrating temporal abstraction and intrinsic
motivation. In Lee, D. D., Sugiyama, M., Luxburg, U. V.,
Guyon, I., and Garnett, R. (eds.), Advances in Neural Information Processing Systems 29, pp. 3675–3683. Curran
Associates, Inc., 2016.
Lillicrap, Timothy P., Hunt, Jonathan J., Pritzel, Alexander,
Heess, Nicolas, Erez, Tom, Tassa, Yuval, Silver, David,
and Wierstra, Daan. Continuous control with deep reinforcement learning. CoRR, abs/1509.02971, 2015. URL
http://arxiv.org/abs/1509.02971.
Schaul, Tom, Horgan, Daniel, Gregor, Karol, and Silver,
David. Universal value function approximators. In
Bach, Francis and Blei, David (eds.), Proceedings of
the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning
Research, pp. 1312–1320, Lille, France, 07–09 Jul
2015. PMLR. URL http://proceedings.mlr.
press/v37/schaul15.html.
Sutton, R. S., Precup, D., and Singh, S. Between MDPs
and semi-MDPs: a framework for temporal abstraction
in reinforcement learning. Artificial Intelligence Journal,
112:181–211, 1999.
Todorov, Emanuel, Erez, Tom, and Tassa, Yuval. Mujoco: A
physics engine for model-based control. 2012 IEEE/RSJ
International Conference on Intelligent Robots and Systems, pp. 5026–5033, 2012.
Hierarchical Actor-Critic
Vezhnevets, Alexander Sasha, Osindero, Simon, Schaul,
Tom, Heess, Nicolas, Jaderberg, Max, Silver, David, and
Kavukcuoglu, Koray. Feudal networks for hierarchical
reinforcement learning. CoRR, abs/1703.01161, 2017.
URL http://arxiv.org/abs/1703.01161.
| 9cs.NE
|
1
On the Restricted Isometry of the Columnwise
Khatri-Rao Product
Saurabh Khanna, Student Member, IEEE and Chandra R. Murthy, Senior Member, IEEE
arXiv:1709.05789v2 [cs.IT] 10 Oct 2017
Dept. of ECE, Indian Institute of Science, Bangalore, India
{saurabh, cmurthy}@iisc.ac.in
Abstract—The columnwise Khatri-Rao product of two matrices
is an important matrix type, reprising its role as a structured
sensing matrix in many fundamental linear inverse problems.
Robust signal recovery in such inverse problems is often contingent on proving the restricted isometry property (RIP) of a
certain system matrix expressible as a Khatri-Rao product of
two matrices. In this work, we analyze the RIP of a generic
columnwise Khatri-Rao product by deriving two upper bounds
for its kth order Restricted Isometry Constant (k-RIC) for
different values of k. The first RIC bound is computed in terms
of the individual RICs of the input matrices participating in the
Khatri-Rao product. The second RIC bound is probabilistic in
nature, and is specified in terms of the input matrix dimensions.
We show that the Khatri-Rao product of a pair of m × n
sized random matrices comprising independent and identically
distributed subgaussian entries satisfies k-RIP
with arbitrarily
√
high probability, provided m exceeds O( k log3/2 n). This is
a substantially milder condition compared to O(k log n) rows
needed to guarantee k-RIP of the input subgaussian random
matrices participating in the Khatri-Rao product. Our results
confirm that the Khatri-Rao product exhibits stronger restricted
isometry compared to its constituent matrices for the same
RIP order. The proposed RIC bounds are potentially useful in
obtaining improved performance guarantees in several sparse
signal recovery and tensor decomposition problems.
examples include compressive sensing [4], [5], covariance
matrix estimation [6], [7], direction of arrival estimation [8]
and tensor decomposition [9]. In each of these examples,
the KR product A
B, for certain m × n sized system
matrices A and B, plays the role of the sensing matrix used to
generate linear measurements y of an unknown signal vector
x according to
y = (A B) x + w,
(2)
Index Terms—Khatri-Rao product, Kronecker product, compressive sensing, Restricted isometry property, covariance matrix
estimation, multiple measurement vectors, PARAFAC, CANDECOMP, tensor decomposition, direction of arrival estimation.
The smallest constant δk (Φ) for which (3) holds for all ksparse z is called the k th order restricted isometry constant or
the k-RIC of Φ. Matrices with small k-RICs are good encoders
for storing/sketching high dimensional vectors with k or fewer
nonzero entries [14]. For example, δk (A B) < 0.307 is a
sufficient condition for a unique k-sparse solution to (2) in the
noiseless case, and its perfect recovery via the `1 minimization
technique [15]. As pointed out earlier, in many structured
signal recovery problems, the main sensing matrix can be
expressed as a columnwise Khatri-Rao product between two
matrices. Thus, from a practitioner’s viewpoint, it is pertinent
to study the restricted isometry property of a columnwise
Khatri-Rao product matrix, which is the focus of this work.
I. I NTRODUCTION
The Khatri-Rao product, denoted by the symbol , is
a columnwise Kronecker product, which was originally introduced by Khatri and Rao in [1]. For any two matrices
A = [a1 , a2 . . . , ap ] and B = [b1 , b2 . . . , bp ] of sizes m × p
and n × p, respectively, the columnwise Khatri-Rao product
A B is a matrix of dimension mn × p defined as
A
B = [a1 ⊗ b1 a2 ⊗ b2
...
ap ⊗ bp ] ,
(1)
where a ⊗ b denotes the Kronecker product [2] between
vectors a and b. That is, each column of A
B is the
Kronecker product between the respective columns of the
two input matrices A and B. In this article, we shall refer
to the columnwise Khatri-Rao product as simply the KhatriRao product or the KR product. Since the Kronecker product
A⊗B comprises all pairwise Kronecker product combinations
of the columns of the input matrices, it can be shown that
A B = (A ⊗ B)J, where J is a p2 × p selection matrix
2
with columns as a subset of the standard basis in Rp [3].
Khatri-Rao product matrices are encountered in several
linear inverse problems of fundamental importance. Recent
where w represents the additive measurement noise. It is now
well established in the sparse signal recovery literature [10]–
[12] that, if the signal of interest, x, is a k-sparse1 vector in
Rn , it can be stably recovered from its noisy underdetermined
2
linear observations y ∈ Rm (m2 < n) in a computationally
efficient manner provided that the sensing matrix (here, A B)
satisfies the restricted isometry property defined next.
A matrix Φ ∈ Rm×n is said to satisfy the Restricted
Isometry Property (RIP) [13] of order k, if there exists a
constant δk (Φ) ∈ (0, 1), such that for all k-sparse vectors
z ∈ Rn ,
(1 − δk (Φ))||z||22 ≤ ||Φz||22 ≤ (1 + δk (Φ))||z||22 .
(3)
A. Applications involving Khatri-Rao matrices
We briefly describe some examples where it is required to
show the restricted isometry property of a KR product matrix.
1) Support recovery of joint sparse vectors from underdetermined linear measurements: Suppose x1 , x2 , . . . , xL are
unknown joint sparse signals in Rn with a common k-sized
support denoted by an index set S. A canonical problem in
multi-sensor signal processing is concerned with the recovery
of the common support S of the unknown signals from their
1A
vector is said to be k-sparse if at most k of its entries are nonzero.
2
noisy underdetermined linear measurements y1 , y2 , . . . , yL ∈
Rm generated according to
yj = Axj + wj ,
1 ≤ j ≤ L,
(4)
m×n
where A ∈ R
(m < n) is a known measurement matrix,
and wj ∈ Rn models the noise in the measurements. This
problem arises in many practical applications such as MIMO
channel estimation, cooperative wideband spectrum sensing in
cognitive radio networks, target localization, and direction of
arrival estimation. In [16], the support set S is recovered as
the support of γ̂, the solution to the Co-LASSO problem:
Co-LASSO: min vec(R̂yy ) − (A
γ0
2
A)γ
2
+λ ||γ||1 , (5)
PL
where R̂yy , L1 j=1 yj yjT . From compressive sensing
theory [11], the RIP of A A (also called the self Khatri-Rao
product of A) determines the stability of the sparse solution in
the Co-LASSO problem. In M-SBL [17], a different support
recovery algorithm, A A satisfying 2k-RIP can guarantee
exact recovery of S from multiple measurements [18].
2) Vandermonde decomposition of Toeplitz matrices:
According to a classical result by Carathéodory and Fejér [19],
any n×n positive semidefinite Toeplitz matrix T of rank r < n
admits the following decomposition
T = APAT ,
(6)
where P is an n × n positive semidefinite diagonal matrix
with an r-sparse diagonal, and A is an n × n Vandermonde
matrix with uniformly sampled complex sinusoids of different frequencies as its columns. This Toeplitz decomposition
underpins subspace based spectrum estimation methods such
as MUltiple SIgnal Classification (MuSiC) and EStimation
of Parameters by Rotationally Invariant Techniques (ESPRIT)
[20], [21]. By replacing T with a data covariance matrix,
the r-sparse support of diag(P) in (6) corresponds to the
r-dimensional signal subspace of the data. Estimation of
p , diag(P) is tantamount to finding an r-sparse solution
to the vectorized form of (6), i.e., vec(T) = (A A)p. Here
again, the recovery of a unique r-sparse solution for p can be
guaranteed if A A satisfies the RIP of order 2r.
3) PARAFAC model for low-rank three-way arrays:
Consider an I × J × K tensor X of rank r. We can
express
PrX as the sum of r rank-one three way arrays as
X = i=1 ai ◦ bi ◦ ci , where ai , bi , ci are loading vectors of
dimension I, J, K, respectively, and ◦ denotes the vector outer
product. The tensor X itself can be arranged into a matrix
as X = [vec(X1 ), vec(X2 ), . . . , vec(XK )]. In the parallel
factor analysis (PARAFAC) model [22], the matrix X can be
approximated as
X ≈ (A B)CT ,
(7)
where A, B and C are the loading matrices with columns
as the loading vectors ai , bi and ci , respectively. In many
problems such as direction of arrival estimation using a 2Dantenna array, the loading matrix C turns out to be rowsparse matrix [23]. In such cases, the uniqueness of the
PARAFAC model shown in (7) depends on the restricted
isometry property of the Khatri-Rao product A B.
Finding the exact kth order RIC of any matrix entails
searching for the smallest and largest eigenvalues among all
possible k-column submatrices of the matrix, which is, in
general, an NP hard task [24]. In this work, we follow an
alternative approach to analyzing the RIP of a KR product
matrix. We seek to derive tight upper bounds for its RICs.
B. Related Work
Perhaps the most straightforward way to analyze the RICs
of the KR product matrix is to use the eigenvalue interlacing
theorem [25], which relates the singular values of any kcolumn submatrix of the KR product to the singular values of
the Kronecker product. This is possible because any k columns
of the KR product can together be interpreted as a submatrix
of the Kronecker product. However, barring the maximum and
minimum singular values of the Kronecker product, there is
no explicit characterization of its non-extremal singular values
available, that can be used to obtain tight bounds the k-RIC of
the KR product. Bounding the RIC using the extreme singular
values of the Kronecker product matrix turns out to be too
loose to be useful. In this context, we note that an upper bound
for the k-RIC of the Kronecker product matrix is derived in
terms of the k-RICs of the input matrices in [4], [26]. However,
the k-RIC of the KR product matrix is yet to be analyzed.
Recently, [27], [28] gave probabilistic lower bounds for
the minimum singular value of the columnwise KR product
between two or more matrices. These bounds are limited to
randomly constructed input matrices, and are polynomial in
the matrix size. In [29], it is shown that for any two matrices
A and B, the Kruskal-rank2 of A B has a lower bound in
terms of K-rank(A) and K-rank(B). In fact, Krank(A B)
is at least as high as max (K-rank(A), K-rank(B)), thereby
suggesting that A B satisfies a stronger restricted isometry
property than both A and B. The RIC bounds presented in
this work ratify this fact.
A closely related yet weaker notion of restricted isometry
constant is the τ -robust K-rank, denoted by K-rankτ . For a
given matrix Φ, the K-rankτ (Φ) is defined as the largest k for
which every n × k submatrix of Φ has its smallest singular
value larger than 1/τ . In [27], it is shown that the τ -robust Krank is super-additive, implying that the K-rankτ of the KhatriRao product is strictly larger than individual K-rankτ s of the
input matrices. We show a similar result for the restricted
isometry constants of the KR product matrix.
Our work is perhaps closest to [7], which provides a
polynomially tight probabilistic upper bound for the k-RIC
(defined using `1 -norm in (3)) of the Khatri-Rao product
A B, when the input matrices A and B are the adjacency
matrices of two independent uniformly random δ-left regular
bipartite graphs. This work instead assumes A and B to be
random matrices with independent subgaussian elements.
C. Our Contributions
We derive two upper bounds on the k-RIC of the columnwise KR product of two m × n sized matrices A and B. The
2 The Kruskal rank of any matrix A is the largest integer r such that any
r columns of A are linearly independent.
3
bounds are listed below.
1) A deterministic upper bound for the k-RIC of A B in
terms of the k-RICs of the input matrices A and B. The
bound is valid for k ≤ m, and for input matrices with
unit `2 -norm columns.
2a) A probabilistic upper bound for the k-RIC of A B
in terms of k and the input matrix dimensions (m, n),
for A, B as random matrices with i.i.d. subgaussian
elements. The probabilistic bound is polynomially tight
with respect to the input matrix dimension n. The bound
is valid for k ≤ n.
2b) A probabilistic upper bound for the k-RIC of the self
KR product A A in terms of m, n, and k, for A as a
random matrix with i.i.d. subgaussian elements. Although
the RIC bounds for the self KR product and the general
KR product with distinct input matrices are of similar
form, the derivation of former RIC bound is slightly more
intricate as it involves showing sharp concentration for
functions of dependent random variables.
A key idea in the RIC analysis is to use the fact (stated
formally as Proposition 12) that for any two matrices A and
B, the Gram matrix of their KR product (A B)T (A B) can
be interpreted as the Hadamard product (element wise multiplication) between AT A and BT B. The Hadamard product
form is more analytically tractable than columnwise Kronecker
product form of the KR matrix.
Lately, in several machine learning problems, the necessary
and sufficient conditions for successful signal recovery have
been reported in terms of the RICs of a certain Khatri-Rao
product matrix serving as a pseudo sensing matrix [8], [16].
In light of this, our proposed RIC bounds are quite timely,
and pave the way towards obtaining order-wise tight sample
complexity bounds in several fundamental learning problems.
The rest of this article is organized as follows. In Secs. II
and III, we present our main results: deterministic and probabilistic RIC bounds, respectively, for a generic columnwise
KR product matrix. Sec. III also discusses about the RIP of the
self Khatri-Rao product of a matrix with itself, an important
matrix type encountered in the sparse diagonal covariance
matrix estimation problem. In Secs. IV and VI, we provide
some background concepts needed in proving the proposed
RIC bounds. Secs. V and VII provide the detailed proofs of
the deterministic and probabilistic RIC bounds, respectively.
Final conclusions are presented in Sec. VIII.
Notation: In this work, bold lowercase letters are used for
representing both scalar random variables as well as vectors.
Bold uppercase letters are reserved for matrices. The `2 -norm
of vector x is denoted by ||x||2 . For an m × n matrix A,
||Ax||
||A|| denotes its operator norm, ||A|| = supx∈Rn ,x6=0 ||x|| 2 .
2
The Hilbert-Schmidt
(or Frobenius) norm of A is defined
Pm P
n
2
as ||A||HS =
i=1
j=1 |Ai,j | . The symbol [n] denotes
the index set {1, 2, . . . , n}. For any index set S ⊆ [n],
AS denotes the submatrix compromising the columns of A
indexed by S. The matrices A ⊗ B, A ◦ B and A B denote
the Kronecker product, Hadamard product and columnwise
Khatri-Rao product of A and B, respectively. A ≤ B implies
that B − A is a positive semidefinite matrix. AT , AH , A−1 ,
and A† denote the transpose, conjugate-transpose, inverse and
generalized matrix inverse operations, respectively.
II. D ETERMINISTIC k-RIC B OUND
In this section, we present our first upper bound on the kRIC of a generic columnwise KR product A B, for any two
similar sized matrices A and B with normalized columns. The
bound is given in terms of the k-RICs of A and B.
Theorem 1. Let A and B be m × n sized real-valued
matrices with unit `2 -norm columns and satisfying the k th
order restricted isometry property with constants δkA and
δkB , respectively. Then, their columnwise Khatri-Rao product
A B satisfies the restricted isometry
property with k-RIC at
most δ 2 , where δ , max δkA , δkB , i.e.,
(1 − δ 2 )||z||22 ≤ ||(A
B)z||22 ≤ (1 + δ 2 )||z||22
(8)
n
holds for all k-sparse vectors z ∈ R .
Proof. The proof is provided in Section V.
Remark 1: The RIC bound for A B in Theorem 1 is
relevant only when δk (A) and δk (B) lie in (0, 1), which
is true only for k ≤ m. In other words, the above k-RIC
characterization for A B requires the input matrices A and
B to be k-RIP compliant.
Remark 2: Since the input matrices A and B satisfy kRIP with δk (A), δk (B) ∈ (0, 1), it follows from Theorem 1
that δk (A B) is strictly smaller than max (δk (A), δk (B)).
If B = A, the special case of self Khatri-Rao product A A
arises, for which
δk (A
A) < δk2 (A).
(9)
Above implies that the self Khatri-Rao product A A is a better restricted isometry compared to A itself. This observation
is in alignment with the expanding Kruskal rank and shrinking
mutual coherence of the self Khatri-Rao product reported in
[16]. In fact, for k = 2, the 2-RIC bound (9) exactly matches
the mutual coherence bound shown in [16].
For k ∈ (m, m2 ], using ( [30], Theorem 1), one can
√
show that δk (A
B) ≤
k + 1 δ√k , where δ√k =
max δ√k (A), δ√k (B) . This bound, however, loses its tightness and quickly becomes unattractive for larger values of k.
Finding a tighter k-RIC upper bound for the k > m case
remains an open problem.
To gauge the tightness of the proposed k-RIC bound for
A
B, we present its simulation-based quantification for
the case when the input matrices A and B are random
Gaussian matrices with i.i.d. N (0, 1/m) entries. Fig. 1 plots
δk (A), δk (B), δk (A B) and the upper bound δk (A B) =
2
(max (δk (A), δk (B))) for a range of input matrix dimension
m. The aspect ratio m/n of the input matrices is fixed to
0.5.3 For computational tractability, we restrict our analysis
to the cases k = 2 and 3. The RICs: δk (A), δk (B) and
3 While the m×n matrices A and B may represent highly underdetermined
linear systems (when m n), their m2 ×n sized Khatri-Rao product A B
can become an overdetermined system. In fact, many covariance matching
based sparse support recovery algorithms [16], [17], [31] exploit this fact to
offer significantly better support reconstruction performance.
4
The minimum such K1 is called the subgaussian or ψ2 norm
of the random variable x, i.e.,
1
δ2A
δ2B
δ2A⊙B
0.8
p
Bound for δ2A⊙Bin Theorem-1
(11)
p≥1
Given a pair of random input matrices with i.i.d. subgaussian entries, Theorem 2 presents a new upper bound on the
k-RIC of their columnwise KR product.
δ2
0.6
0.4
0.2
0
100
1
||x||ψ2 = sup p−1/2 (E |x| ) p .
300
500
700
900
n
1.2
1
Theorem 2. Suppose A and B are m × n random matrices
with real i.i.d. subgaussian entries, such that EAij = 0 and
EA2ij = 1, and ||Aij ||ψ2 ≤ K, and similarly for B. Let Ko ,
max (K, 1). Then, the k th order restricted isometry constant
of √Am √Bm , denoted by δk , satisfies δk ≤ δ with probability
at least 1 − 8en−2(c2 γ−1) for all γ > 1/c2 , and c2 a universal
positive constant, provided that
!
√
16 2πγ 3/2 Ko2 √
3/2
k (log n) .
m≥
δ
δ3
0.8
0.6
0.4
0.2
0
50
100
150
200
250
300
n
Fig. 1: Variation of k-RICs of A, B, A B and the proposed upper bound
with increasing input matrix dimensions. The top and the bottom plots are
for k = 2 and 3, respectively. Each data point is averaged over 10 trials.
δk (A B) are computed by exhaustively searching for the
worst conditioned submatrix comprising k columns of A, B
and A B, respectively. From Fig. 1, we observe that the
proposed k-RIC upper bound becomes tighter as the input
matrices grow in size.
III. P ROBABILISTIC k-RIC B OUND
The deterministic RIC bound for the columnwise KR product discussed in the previous section is only applicable when
k ≤ m. It also fails to explain the empirical observation that
the KR product A B often satisfies k-RIP in spite of the
input matrices A and B failing to do so. These concerns are
addressed by our second RIC bound for the columnwise KR
product. This second RIC bound is probabilistic in nature and
is applicable to the KR product of random input matrices
with i.i.d. subgaussian entries. Below, we define a subgaussian
random variable and state some of its properties.
Definition 1. (Subgaussian Random Variable): A random variable x is called subgaussian, if its tail probability is dominated
by that of a Gaussian random variable. In other words, there
2
2
exists a constant K > 0 such that P(|x| ≥ t) ≤ e−t /K .
Gaussian, Bernoulli and all bounded random variables are
subgaussian random variables. For a subgaussian random
variable, its pth order moment grows only as fast as O(pp/2 )
[32]. In other words, there exists K1 > 0 such that
√
p 1
(E |x| ) p ≤ K1 p, p ≥ 1.
(10)
Proof. The proof is provided in Section VII.
√
The normalization constant m used while computing
√B
the KR product √Am
ensures that the columns of
m
A
B
√
√
the input matrices m , m have unit average energy, i.e.
√
√
E ||ai / m||2 = E ||bi / m||2 = 1 for 1 ≤ i ≤ n. Column
normalization is a key assumption towards correct modelling
of the isotropic, norm-preserving nature of the effective sens1
ing matrix m
(A B), an attribute found in most sensing
matrices employed in practice.
Theorem 2 implies that
!
√
3/2
√
A
B
k (log n)
3/2 2
√
δk √
≤ 16 2πγ Ko
m
m
m
(12)
with probability exceeding 1 − 8en−2(c2 γ−1) . Thus, the above
k-RIC bound decreases as m increases, which is intuitively
appealing. Interestingly, for fixed k and n, the above k√B decays as O( 1 ). This is
RIC upper bound for √Am
m
m
a significant improvement over the O( √1m ) decay rate [11]
already known for the individual k-RICs of the input subgaussian matrices √Am and √Bm . Thus, for any m, the Khatri√B exhibits stronger restricted isometry
Rao product √Am
m
property, with significantly smaller k-RICs in comparison to
the k-RICs for the input matrices.
In some applications, the effective sensing matrix can be
1
(A A) of a
expressed as the self-Khatri Rao product m
A
certain column normalized system matrix √m with itself [16].
In Theorem 3 below, we present the k-RIC bound for the
special case of self-Khatri-Rao product matrices.
Theorem 3. Let A be an m × n random matrix with real
i.i.d. subgaussian entries, such that EAij = 0 and EA2ij = 1,
and ||Aij ||ψ2 ≤ K. Then, the k th order restricted isometry
constant of the column normalized self Khatri-Rao product
1
A) satisfies δk ≤ δ with probability at least 1 −
m (A
4en−2(c2 γ−1) for any γ ≥ 1/c2 , provided
!
√
16 2πγ 3/2 Ko2 √
3/2
m≥
k (log n) .
δ
5
Here, Ko , max (K, 1) and c2 > 0 is a universal constant.
Proof. A proof sketch is provided in Appendix F.
Theorem 3 implies that
√
A
A
√
≤ 16 2πγ 3/2 Ko2
δk √
m
m
√
3/2
k (log n)
m
!
(13)
with probability exceeding 1−4en
. The above k-RIC
bound for the self Khatri-Rao product scales with m, n, and
k in a similar fashion as the asymmetric Khatri-Rao product.
Remark 3: From [11], for an mq× n matrix A with i.i.d.
2Ck log(eN/k)
with high
subgaussian entries, δk ( √Am ) ≤
m
probability. Now, even though the columns of A do not exactly
have unit norm, directly using this in Theorem 1 for the sake of
2Ck log(eN/k)
√A ) ≤
, when
comparison results in δk ( √Am
m
m
k ≤ m. Our probabilistic bound in (13) is therefore tighter
than the deterministic
bound in Theorem 1 by a multiplicative
√
factor of O( k). Note that, the deterministic bound is valid
for any pair of input matrices with normalized columns, while
the probabilistic bound holds when the input matrices have
i.i.d. subgaussian entries.
Remark 4: In covariance matching based signal support
recovery algorithms like Co-LASSO [16] and M-SBL [17],
given a system matrix √Am , δ2k ( √Am √Am ) ≤ 1 is one of the
sufficient conditions for exact recovery of the true k-sparse
support of the unknown signal of interest [18]. This sufficiency
condition is met with arbitrarily high probability according
to Theorem
√ 3, when the
subgaussian matrix A has at least
m = O
k log3/2 n rows. This is a significantly milder
condition when compared with O(k log n) rows required for
exact k-sparse support recovery by conventional compressive
sensing algorithms such as SOMP [33]. In fact, to the best
of our knowledge, our k-RIC bound for the self Khatri-Rao
product provides the first ever theoretical confirmation for
the empirically observed performance of covariance matching
based sparse support recovery algorithms, i.e.,
√ the number of
measurements m need to scale as only O( k) rather than
O(k) in the case of conventional support recovery algorithms.
−2(c2 γ−1)
IV. P RELIMINARIES FOR D ETERMINISTIC k-RIC B OUND
In this section, we present some preliminary concepts and
results which are necessary for the derivation of the deterministic k-RIC bound in Theorem 1. For the sake of brevity, we
provide proofs only for claims that have not been explicitly
shown in their cited sources.
A. Properties of the Kronecker and Hadamard product
For any two matrices A and B of dimensions m × n and
p × q, the kronecker product A ⊗ B is the mp × nq matrix
a11 B a12 B · · · a1n B
a21 B a22 B · · · a2n B
A⊗B= .
(14)
..
.. .
.
.
.
.
.
.
.
am1 B am2 B · · ·
amn B
The following Proposition relates the spectral properties of
the Kronecker product and its constituent matrices.
Proposition 4 (7.1.10 in [2]). Let A ∈ Rn×n and B ∈ Rp×p
admit eigenvalue decompositions UA ΛA UTA and UB ΛB UTB ,
respectively. Then,
(UA ⊗ UB )(ΛA ⊗ ΛB )(UA ⊗ UB )T
yields the eigenvalue decomposition for A ⊗ B.
For any two matrices of matching dimensions, say m × n,
their Hadamard product A ◦ B is obtained by elementwise
multiplication of the entries of the input matrices, i.e.,
[A ◦ B]i,j = aij bij
for i ∈ [m], j ∈ [n].
(15)
The Hadamard product A ◦ B is a principal submatrix of the
Kronecker product A ⊗ B [3], [34]. For n × n sized square
matrices A and B, one can write,
A ◦ B = JT (A ⊗ B)J,
(16)
where J is an n2 ×n sized selection matrix constructed entirely
from 0’s and 1’s which satisfies JT J = In .
In Proposition 5, we present an upper bound on the spectral
radius of a generic Hadamard product.
Proposition 5. For every A, B ∈ Rm×n , we have
σmax (A ◦ B) ≤ rmax (A)cmax (B)
(17)
where σmax (·), rmax (·) and cmax (·) are the largest singular
value, the largest row norm and the largest column norm of
the input matrix, respectively.
Proof. See Theorem 5.5.3 in [35].
We now state an important result about the Hadamard
product of two positive semidefinite matrices.
Proposition 6 (Mond and Pec̆arić [36]). Let A and B be
positive semidefinite n × n Hermitian matrices and let r and
s be two nonzero integers such that s > r. Then,
1/s
(As ◦ Bs )
1/r
≥ (Ar ◦ Br )
.
(18)
In Propositions 7 and 8, we state some spectral properties of
correlation matrices and their Hadamard products. Correlation
matrices are symmetric positive semidefinite matrices with
diagonal entries equal to one. Later on, we will exploit the
fact that the singular values of the columnwise KR product
are related to the singular values of the Hadamard product of
certain correlation matrices.
Proposition 7. If A is an n×n correlation matrix, then A1/2 ◦
A1/2 is a doubly stochastic matrix.
Proof. See Appendix A.
Proposition 8 (Werner [37]). For any correlation matrices A
and B of the same size, we have A1/2 ◦ B1/2 ≤ I, where
A1/2 and B1/2 are the positive square roots of A and B,
respectively.
Proof. Since A is a correlation matrix, from Proposition 7,
it follows that A1/2 ◦ A1/2 is doubly stochastic. Since the
rows and columns of A1/2 ◦ A1/2 sum to unity, we have
rmax (A1/2 ) = cmax (A1/2 ) = 1. Similarly, rmax (B1/2 ) =
cmax (B1/2 ) = 1. Then, from Proposition 5, it follows that
the largest eigenvalue of A1/2 ◦ B1/2 is at most unity.
6
B. Matrix Kantorovich Inequalities
Matrix Kantorovich inequalities relate positive definite matrices by inequalities in the sense of the Löwner partial order.4
These inequalities can be used to extend the Löwner partial
order to the Hadamard product of positive definite matrices.
Our proposed RIC bound relies on the tightness of these
Kantorovich inequalities and their extensions.
A matrix version of the Kantorovich inequality was first
proposed by Marshall and Olkin in [38]. It is stated below as
Proposition 9.
Proposition 9 (Marshall and Olkin [38]). Let A be an n × n
positive definite matrix. Let A admit the Schur decomposition
A = UΛUT with unitary U and Λ = diag(λ1 , λ2 , . . . , λn )
such that λi ∈ [m, M ]. Then, we have
A2 ≤ (M + m)A − mM In .
(19)
The above inequality (19) is the starting point for obtaining
a variety of forward and reverse Kantorovich-type matrix
inequalities for positive definite matrices. In Propositions 10
and 11, we state specific forward and reverse inequalities,
respectively, which are relevant to us.
Proposition 10 (Liu and Neudecker [39]). Let A be an n × n
positive definite Hermitian matrix, with eigenvalues in [m, M ].
Let V be an n × n matrix such that VT V = I. Then,
2
1
(20)
VT A2 V − VT A2 V ≤ (M − m)2 I.
4
Proposition 11 (Liu and Neudecker [40]). Let A and B be
n×n positive definite matrices. Let m and M be the minimum
and maximum eigenvalues of B1/2 A−1 B1/2 . Let X be an
n × p matrix with rank q (n > p ≥ q). Then, we have
4mM
(XT BX)(XT AX)† (XT BX) ≥
XT BA−1 BX.
(M + m)2
(21)
Proposition 10 can be proved using (19) by pre- and
post-multiplying by VT and V, respectively, followed by
completion of squares for the right hand side terms. The proof
of Proposition 11 is given in [40].
C. Kantorovich Matrix Inequalities for the Hadamard Products of Positive Definite Matrices
Lemmas 1 and 2 stated below extend the Kantorovich inequalities from the previous subsection to Hadamard products.
Lemma 1 (Liu and Neudecker [39]). Let A and B be
n × n positive definite matrices, with m and M denoting the
minimum and maximum eigenvalues of A ⊗ B. Then, we have
1
A2 ◦ B2 ≤ (A ◦ B)2 + (M − m)2 In .
(22)
4
Proof. Let J be the selection matrix such that JT J = I and
A ◦ B = JT (A ⊗ B)J. Then, by applying Proposition 10 with
A replaced with A ⊗ B, and V replaced with J, we obtain
1
JT (A ⊗ B)2 J − (JT (A ⊗ B)J)2 ≤ (M − m)2 I.
4
4 The Löwner partial order here refers to the relation “≤”. For positive
definite matrices A and B, A ≤ B if and only if B − A is a positive
semi-definite matrix.
Using Fact 8.21.29 in [2], i.e., (A ⊗ B)2 = A2 ⊗ B2 , we get
JT (A2 ⊗ B2 )J − (A ◦ B)2 ≤
1
(M − m)2 I,
4
Finally, by observing that JT (A2 ⊗ B2 )J = A2 ◦ B2 , we
obtain the desired result.
Lemma 2 (Liu [41]). Let A, B be n × n positive definite
correlation matrices. Then,
A1/2 ◦ B1/2 ≥
2cd
In
+ d2
c2
(23)
where the eigenvalues of A1/2 and B1/2 lie inside [c, d].
Lemma 2 follows from Proposition 11, by replacing A with
In ⊗ B, B with A1/2 ⊗ B1/2 , and X with J, where J is
the n2 × n binary selection matrix such that JT J = I and
A ◦ B = JT (A ⊗ B)J.
V. P ROOF OF THE D ETERMINISTIC k-RIC B OUND
(T HEOREM 1)
The key idea used in bounding the k-RIC of the columnwise
KR product A B is the observation that the Gram matrix
of A B can be interpreted as a Hadamard product between
the two correlation matrices AT A and BT B, as mentioned
in the following Proposition.
Proposition 12 (Rao and Rao [42]). For A, B ∈ Rm×n ,
(A
B)T (A
B) = (AT A) ◦ (BT B)
(24)
Proof. See Proposition 6.4.2 in [42].
Then, by using the forward and reverse Kantorovich matrix
inequalities, we obtain the proposed upper bound for k-RIC
of A B as explained in the following arguments.
Without loss of generality, let S ⊂ [n] be an arbitrary index
set representing the nonzero support of z in (8), with |S| ≤ k.
Let AS denote the m × |S| submatrix of A, constituting |S|
columns of A indexed by the set S. Let BS be constructed
similarly. Since δk (A), δk (B) < 1, both AS , BS have full
rank, and consequently the associated Gram matrices ATS AS ,
BTS BS are positive definite. Further, since A and B have
unit norm columns, both ATS AS and BTS BS are correlation
matrices with unit diagonal entries. Using Proposition 12, we
can write
(AS
BS )T (AS
BS ) = ATS AS ◦ BTS BS .
(25)
Next, for k ≤ m, by applying Lemma 1 to the positive
definite matrices (ATS AS )1/2 and (BTS BS )1/2 , we get,
1
1
1 2
ATS AS ◦ BTS BS ≤ (ATS AS ) 2 ◦(BTS BS ) 2 + (M −m)2 Ik
4
1
2
≤ Ik + (M − m) Ik ,
(26)
4
where the second inequality is a consequence of the unity
bound on the spectral radius of the Hadamard product between
correlation matrices, shown in Proposition 8. In (26), M and
m are upper and lower bounds for the maximum and minimum
eigenvalues of (ATS AS )1/2 ⊗ (BTS BS )1/2 , respectively. From
the restricted isometry of A and B, and by application of
7
Proposition 4, the minimum and maximum eigenvalues of
T
1/2
(A
⊗ (BTS BS )1/2qare lower and upper bounded by
q S AS )
almost surely. Then, for all t > 0,
By introducing δ , max(δkA , δkB ), it is easy to check that
the eigenvalues of (ATS AS )1/2 ⊗ (BTS BS )1/2 also lie inside
the interval [1 − δ, 1 + δ]. Plugging m = 1 − δ and M = 1 + δ
in (26), and by using (25), we get
(AS BS )T (AS BS ) ≤ 1 + δ 2 Ik .
(27)
Proof. See Theorem 6.7 in [45].
(1 −
δkA )(1
−
δkB )
(1 +
and
δkA )(1
+
δkB ),
respectively.
Similarly,
by applying√Lemma 2 to ATS AS and BTS BS with
√
c = 1 − δ and d = 1 + δ, we obtain
p
1 − δ 2 Ik .
(ATS AS )1/2 ◦ (BTS BS )1/2 ≥
ATS AS
BTS BS
From Proposition 6, we have
◦
2
(ATS AS )1/2 ◦ (BTS BS )1/2 . Therefore, we can write
ATS AS ◦ BTS BS ≥ 1 − δ 2 Ik .
≥
(AS
BS ) (AS
BS ) ≥ 1 − δ
2
Ik .
2
/2ν
.
Theorem 14 can be used to show a subgaussian tail for the
maximal linear combination of bounded nonnegative random
variables, stated below.
Lemma 3. Let x1 , x2 , . . . , xn be i.i.d. nonnegative, bounded
random variables
Pn satisfying xi ≤ b. Then, y ,
supz∈Rn ,||z||2 ≤1 i=1 zi xi is a nonnegative random variable
which satisfies the following tail inequality:
P (y − Ey ≥ t) ≤ e−t
2
/2b2
.
Proof. See Appendix B.
The following proposition bounds the expectation of a
nonnegative subgaussian random variable.
Further, using (25), we get
T
P (z − Ez ≥ t) ≤ e−t
(28)
Finally, Theorem 1’s statement follows from (27) and (28).
VI. P RELIMINARIES FOR P ROBABILISTIC k-RIC B OUND
Proposition 15. Let x be a nonnegative random variable with
a subgaussian tail, i.e., for t ≥ 0,
P(z − Ez ≥ t) ≤ e−t
√
for some ν > 0. Then, Ez ≤ 2πν.
2
/2ν
In this section, we briefly discuss some concentration results
on certain functions of subgaussian random variables which
will appear in the proofs of Theorems 2 and 3.
Proof. See Appendix C.
A. A Tail Probability for Subgaussian Vectors
The proof of Theorem
2 starts
with a variational definition
of the k-RIC, δk √Am √Bm given below.
The Theorem below presents the Hanson-Wright inequality
[43], [44], a tail probability for a quadratic form constructed
using independent subgaussian random variables.
Theorem 13 (Rudelson and Vershynin [44]). Let x =
(x1 , x2 , . . . , xn ) ∈ Rn be a random vector with independent
components xi satisfying Exi = 0 and ||xi ||ψ2 ≤ K. Let A
be an n × n matrix. Then, for every t ≥ 0,
P xT Ax − ExT Ax > t
"
!#
t2
t
≤ 2 exp −c min
2 ,
K 4 ||A||HS K 2 ||A||
where c is a universal positive constant.
B. Concentration of Supremum of Linear Combination of
Bounded, Nonnegative Independent Random Variables
The following theorem bounds the subgaussian tail of a
function of several independent random variables satisfying
a relaxed bounded difference property.
Theorem 14. Consider a general real-valued function of n
independent random variables z = f (x1 , x2 , . . . , xn ) and
zi denotes an xi -measurable random variable defined by
zi = inf x0i f (x1 , x2 , . . . , x0i , . . . , xn ). Assume that the random
variable z is such that there exists a constant ν > 0, for which,
n
X
i=1
2
(z − zi ) ≤ ν
VII. P ROOF OF THE P ROBABILISTIC k-RIC B OUND
(T HEOREM 2)
δk
A
√
m
B
√
m
=
sup
z∈Rn ,
||z||2 =1,||z||0 ≤k
A
√
m
B
√
m
2
−1 .
z
2
(29)
In order to find a probabilistic upper bound for δk , we intend
to find a constant δ ∈ (0, 1) such that P(δk √Am √Bm ≥ δ)
is arbitrarily close to zero. We therefore consider the tail event
2
B
A
√
√
E,
sup
z −1 ≥δ ,
m
m
z∈Rn ,
2
||z||2 =1,||z||0 ≤k
(30)
and show that for m sufficiently large, P(E) can be driven
arbitrarily close to zero. In other words,
δ serves
the constant
B
A
√
√
.
as a probabilistic upper bound for δk
m
m
Let Uk be the set of all k or less sparse unit norm vectors
in Rn . Then, using Proposition 12, the tail event in (30) can
be rewritten as
T
P(E) = P sup zT (A B) (A B) z − m2 ≥ δm2
z∈U
k
= P sup zT AT A ◦ BT B z − m2 ≥ δm2
z∈U
k
n X
n
X
= P sup
zi zj aTi aj bTi bj − m2 ≥ δm2
z∈Uk i=1 j=1
8
where ai and bi denote the ith column of A and B, respectively. Further, by applying the triangle inequality and the
union bound, we get
!
n
X
2
2
2
2
2
P(E) ≤ P sup
zi ||ai ||2 ||bi ||2 − m ≥ αδm +
z∈Uk i=1
n
n
X
X
zi zj aTi aj bTi bj
z∈Uk i=1
j=1,j6=i
P sup
≥ (1 − α)δm2 . (31)
In the above, α ∈ (0, 1) is a variational union bound parameter
which can be optimized at a later stage. We now proceed to
find separate upper bounds for each of the two probability
terms in (31).
The first probability term in (31) admits the following
sequence of relaxations.
!
n
X
2
2
2
2
2
P sup
zi ||ai ||2 ||bi ||2 − m ≥ αδm
In step (d), β ∈ (0, 1) is a variational union bound parameter.
Step (g) is once again the union bound and exploits the i.i.d.
nature of the columns of A and B. Step (h) is obtained
by setting β = αδ/4. Lastly, step (i) is the Hanson-Wright
inequality (Theorem 13) applied to the subgaussian vector a1 .
Next, we turn our attention to the second probability term
in (31). It can be upper bounded as follows:
sup
n
n
X
X
z∈Uk i=1
j=1,j6=i
≤ sup
≤
n
X
z∈Uk i=1
(b)
≤
sup
P
≤
P
2
2
(c)
=
P
(d)
n
X
[ n
2
2
||ai ||2 ||bi ||2 − m2
P
2
2
i=1
(e)
=
(f )
≤
(g)
≤
(h)
≤
(i)
≤
nP
2
2
||a1 ||2 ||b1 ||2 − m2 ≥ αδm2 .
2
2
||a1 ||2 − m ||b1 ||2 − m ≥ αβδm2
α(1 − β)δm
2
+2nP ||a1 ||2 − m ≥
2
p
2
2nP ||a1 ||2 − m ≥ αβδm
α(1 − β)δm
2
+2nP ||a1 ||2 − m ≥
.
2
αδm
αδ
2
4nP ||a1 ||2 − m ≥
1−
2
4
nP
4ne
2 2
−cm α4Kδ4
o
.
|zj | |uij vij |
!
|zi | ||ai ||2 ||bi ||2
n
X
|zj0 | |uij vij | ,
(33)
j=1,j6=i
z∈Uk i=1
j=1,j6=i
≤ P max sup
1≤i≤n z∈Uk
n
X
√
zj uij vij ≥ 4 2πKo2 (γ log n)3/2
j=1,j6=i
n
X
(1 − α)δm2
|zi | ||ai ||2 ||bi ||2 ≥ √
+ P sup
4 2πKo2 (γ log n)3/2
z∈Uk i=1
!
(.34)
The following Lemmas provide upper bounds for the two tail
probability terms in (34).
2
(1−αδ/4)
where uij , h ||aaii|| , aj i and vij , h ||bbii|| , bj i. From
2
2
the approximate rotational invariance of subgaussian random
variables [32], it can be shown that the inner products uij and
vij are also subgaussian with ||uij ||ψ2 = ||vij ||ψ2 ≤ c3 K,
with c3 > 0 being a universal numerical constant.
Using (33), and by applying the union bound, the second
probability term in (31) can be written as the sum of following
two tail probabilities.
n
n
X
X
zi zj aTi aj bTi bj ≥ (1 − α)δm2
P sup
o
≥ αδm2
||ai ||2 ||bi ||2 − m2 ≥ αδm2
n
X
sup
n
X
j=1,j6=i
1≤i≤n z0 ∈Uk
1≤i≤n
≤
|zi | ||ai ||2 ||bi ||2
max sup
1≤i≤n
2
zi2 ||ai ||2 ||bi ||2 − m2 ≥ αδm2
max ||ai ||2 ||bi ||2 − m2 ≥ αδm2
n
X
z∈Uk i=1
!
2
zi zj aTi aj bTi bj
z∈Uk i=1
z∈Uk i=1
(a)
2
2
+m ||a||2 − m + m ||b||2 − m .
(32)
In the above, step (a) follows from the triangle inequality
combined with the fact that zi2 ’s sum to one. The inequality
in step (b) is a consequence of the fact that any nonnegative
convex combination of n arbitrary numbers is at most the
maximum among the n numbers. Step (c) is obtained by
simply rewriting the tail event for the maximum of n random
variables as the union of tail events for the individual random
variables. Step (d) is obtained by using the union bound over
values of index i ∈ [n]. Step (e) follows because the columns
of A and B are i.i.d.. Step (f ) is the union bound combined
with the fact that for any two vectors a, b ∈ Rm , the following
triangle inequality holds:
2
2
2
2
||a||2 ||b||2 − m2 ≤
||a||2 − m ||b||2 − m
Lemma 4. Let A and B be m × n random matrices as
specified in Theorem 2. Let uij , h ||aaii|| , aj i and vij ,
2
h ||bbii|| , bj i, where ai and bi denote the ith column of A and
2
B, respectively. Then,
n
X
√
P max sup
|zj ||uij vij | ≥ 4 2πKo2 (γ log n)3/2
1≤i≤n z∈Uk
j=1,j6=i
≤
2e
n2(c2 γ−1)
,
(35)
where Ko , max(K, 1) and c2 > 0 is a universal constant.
Proof. See Appendix D.
9
Lemma 5. Let A and B be m × n random matrices with independent zero mean, unit variance, i.i.d. subgaussian entries
satisfying ||Aij ||ψ2 = ||Bij ||ψ2 ≤ K. Then,
!
n
X
(1 − α)δm2
P sup
|zi | ||ai ||2 ||bi ||2 ≥ √
4 2πKo2 (γ log n)3/2
z∈Uk i=1
≤ 2ne
cm
−K
4
o
,
(36)
√
8 2πKo2 γ 3/2 √
k log3/2
(1−α)δ
n. Here, Ko = max (K, 1)
provided m ≥
and c > 0 is a universal constant.
Proof. See Appendix E.
By using Lemmas 4 and 5 together in (34), we get
n
n
X
X
P sup
zi zj aTi aj bTi bj ≥ (1 − α)δm2
z∈Uk i=1
j=1,j6=i
≤
2e
cm
−K
4
to the dimensions of the input matrices, we had to resort to the
randomness in their entries. Removing this randomness aspect
of our results could be an interesting direction for future work.
A PPENDIX
A. Proof of Proposition 7
Proof. Since A is a correlation matrix, it admits the Schur
decomposition, A = UΛUT , with unitary U and eigenvalue
matrix Λ = diag(λ1 , λ2 , . . . , λn ). Since A is positive semidefinite, its nonnegative square-root exists and is given by
A1/2 = UΛ1/2 UT . Consider
n
n
X
X
1/2
1/2
A1/2 ◦ A1/2 =
λi ui uTi ◦
λj uj uTj
i=1
=
4e
+ 2ne o ≤ 2(c γ−1) ,
(37)
n2(c2 γ−1)
n 2
√
3
8 2πKo2 γ 3/2 √
(2c2 γ−1)Ko4 log n
2 n,
provided m ≥ max
k
log
.
(1−α)δ
c
Using (32) and (37) together in (31), we get
=
n X
n
X
i=1 j=1
n X
n
X
j=1
1/2 1/2
λi λj
1/2 1/2
λi λj
ui uTi ◦ uj uTj
T
(ui ◦ uj ) (ui ◦ uj ) .(39)
i=1 j=1
Above, the second equality follows from the distributive
property of the Hadamard product and the last step follows
4e
8e
4
[2]. Using (39), we can show that the rows
+
≤ 2(c γ−1) , (38) from Fact 7.6.2 in 1/2
P(E) ≤ 2 2
2(c2 γ−1)
2
cα δ (1−αδ/4)2 m
n
n
and
columns
of
A
◦ A1/2 sum to one, as follows:
−1
4
4Ko log n
n
n X
n
X
√
1/2 1/2
T
1/2
1/2
3
8 2πKo2 γ 3/2 √
(2c2 γ−1)ξKo4 log n
(A
◦
A
)1
=
λi λj (ui ◦ uj ) (ui ◦ uj ) 1
2
provided m ≥ max
k log n,
,
(1−α)δ
c
i=1 j=1
4
n X
n
where ξ = max 1, α2 δ2 (1−αδ/4)
.
2
X
1/2 1/2
T
=
λi λj (ui ◦ uj ) (ui ◦ 1) uj
Lastly, by choosing α = 1/2, we get P(δk ≥ δ) = P(E) ≤
i=1 j=1
8en−2(c2 γ−1) , provided that
n
n X
!
X
√
1/2 1/2
2 3/2√
4
3
λi λj (ui ◦ uj ) uTi uj
=
(2c2 γ −1)ξKo
16 2πKo γ
k log 2 n,
m ≥ max
log n ,
i=1 j=1
δ
c
n
X
λi (ui ◦ ui ) = d (say).
=
16
where ξ = max 1, γδ2 (1−δ/8)2 . Note that, in terms of k and
i=1
√
n, the first term in the inequality for m scales as k log3/2 n;
The above arguments follow from the orthonormality of
it dominates the second term, which scales as log n. This ends
the columns of U, and repeated application
in
Pn of Fact 7.6.1
our proof.
2
[2]. Note that for k ∈ [n], d(k) =
λ(i)
(u
(k))
=
i
i=1
[UΛUT ]kk = Akk = 1. Thus, we have shown that (A1/2 ◦
VIII. C ONCLUSIONS
A1/2 )1 = 1. Likewise, it can be shown that 1T (A1/2 ◦
In this work, we have analyzed the restricted isometry A1/2 ) = 1T . Thus, A1/2 ◦ A1/2 is doubly stochastic.
property of the columnwise Khatri-Rao product matrix in
terms of its restricted isometry constants. We gave two upper B. Proof of Lemma 3
bounds for the k-RIC of a generic columnwise Khatri-Rao
Define y(i) , inf0 sup (z1 x1 + · · · + zi x0i + · · · + zn xn ).
product matrix. The first k-RIC bound, a deterministic bound,
xi z∈Rn
||z||≤1
is valid for the Khatri-Rao product of an arbitrary pair of
Pn
∗
sup
input matrices of the same size with normalized columns. Further, let z , arg
j=1 zj xj . Since xi are nonn
z∈R ,||z||2 ≤1
It is conveniently computed in terms of the k-RICs of the negative, clearly z∗ is a nonnegative,
unit norm vector. Then,
input matrices. We also gave a probabilistic RIC bound for for i ∈ [n],
the columnwise KR product of a pair of random matrices with
i.i.d. subgaussian entries. The probabilistic RIC bound is one
n
n
X
X 0
of the key components needed for computing tight sample y − y(i) = sup
zj xj − inf0 sup
zj xj + zi0 x0i
xi z0 ∈Rn ,
complexity bounds for several machine learning algorithms.
z∈Rn , j=1
j=1,
0
||z||2 ≤1
||z ||2 ≤1 j6=i
The analysis of the RIP of Khatri-Rao product matrices in
this article can be extended in multiple ways. The current RIC
n
n
X
X
bounds can be extended to the Khatri-Rao product of three or
≤
zj∗ xj − inf0
zj∗ xj + zi∗ x0i
xi
more matrices. More importantly, in order to relate the RICs
j=1
j=1,j6=i
10
= zi∗ xi − inf0 x0i = zi∗ xi .
a,b
m×n
× Rm×n : x1 = a, y1 = b}
(40) Let E2 , E2 ∩{(X, Y) ∈ R
denote the set of matrix tuples (X, Y) with their first columns
Pn
fixed as a and b ∈ Rm , respectively. Then, conditioned
∗
(i) 2
Since
≤
i ≤ b and ||z ||2 ≤ 1, we have
i=1 y − y
Pn x∗2
on E2a,b , the random variables |u1j v1j |, 2 ≤ j ≤ n are
2
2
z
b
≤
b
.
Finally,
by
invoking
Theorem
14
with
ν
=
i=1 i
i.i.d., bounded and nonnegative random variables. By applying
b2 , we obtain the desired tail bound for y.
Lemma 3, it follows that
n
n
X
X
C. Proof of Proposition 15
P sup
|zj ||uij vij | − E sup
|zj ||u1j v1j | ≥ t E2a,b
z∈Uk j=2
z∈Uk j=2
The said upper bound for Ez can be derived as shown below.
Z ∞
Z ∞
2
4 2
2
≤ e−t /8K γ log n . (43)
P(z ≥ r + Ez)dr
P(z ≥ t)dt =
Ez =
Z −Ez
Z0 ∞
It is to be noted here that the above tail bound in the RHS
∞
2
√
2
− r2ν
−r /2ν
does not depend on a and b. This leads to an interesting
e
e
dr ≤
≤
dr ≤ 2πν.
−∞
−Ez
consequence explained in the following discourse.
a,b
Let T be an event such that P(T
≤ η, where η is
S |E2 ) a,b
independent of a, b. Since E2 = a,b∈Rn E2 , we have
D. Proof of Lemma 4
a,b
n (T ∩ E
P
∪
a,b∈R
2
P(T ∩ E2 )
=
P(T |E2 ) =
Let Q ⊆ (Rm×n × Rm×n ) be the collection of matrix
P(E
)
P(E
)
2
2
tuples (X, Y) satisfying |uij vij | < 2K 2 γ log n for all 1 ≤
R
R
a,b
a,b
a,b
xT
P(T
∩
E
P(T
|E
)
i xj
2
2 )P(E2 )
a,b
a,b
i, j ≤ n, j 6= i, where uij and vij are evaluated as h ||xi || i and
=
=
2
P(E2 )
P(E2 )
yT y
h ||yi i ||j i, respectively. Then, for subgaussian random matrices
R
a,b
2
P(E2 )
P(E2 )
A, B as defined in Lemma 4,
a,b
=η
= η.
(44)
≤η
P(E2 )
P(E2 )
[
Coming back to our proof, using similar arguments as
|uij vij | ≥ 2K 2 γ log n
P ((A, B) ∈
/ Q) = P
above5 and on account of the tail bound in the RHS in (43)
1≤i,j≤n
being independent of a and b, it follows that
j6=i
X
(a)
n
n
X
X
≤
P |uij vij | ≥ 2K 2 γ log n
|zj ||u1j v1j | ≥ t E2
|zj ||uij vij | − E sup
P sup
xi
(b)
≤
1≤i,j≤n
j6=i
2
n
P |u12 v12 | ≥ 2K 2 γ log n
2
√ p
n2 P |u12 | ≥ 2K γ log n
z∈Uk j=2
z∈Uk j=2
2
≤ e−t
/8K 4 γ 2 log2 n
. (45)
Further, using Proposition 15, the probability term in the
LHS of the inequality in (45) can be replaced by a smaller
≤
probability term as shown below.
(d)
e
2 −2c1 γ log n
≤ en e
= 2(c γ−1) .
(41)
n
X
√
n 1
|zj ||u1j v1j | ≥ t + 2 2πK 2 γ log n E2
P sup
In the above, steps (a), (b), and (c) are obtain by applying
z∈Uk j=2
the union bound over n2 combinations of distinct (i, j) pairs
2
2 2
2
≤ e−t /8K γ log n .
and exploiting the identical subgaussian tails of the random
variables uij and vij , i 6= j. Step (d) is simply the standard
Setting t = 4K(γ log n)3/2 , we get
subgaussian tail bound, with c1 > 0 a universal constant. Thus,
n
for γ > 1/c1 , the random variable |uij vij |, i 6= j, is uniformly
X
√
3
1
Psup
|zj ||u1j v1j | ≥ 4 2πKo2 (γ log n) 2 E2 ≤ 2γ .
bounded by 2K 2 γ log n with arbitrarily high probability, for
n
z∈Uk j=2
sufficiently large n.
(46)
Let E2 denote the event that the matrix tuple (A, B) ∈ Q.
Finally,
by
combining
(41),
(42)
and
(46),
we
obtain
the
Then, by the law of total probability, we have
desired tail bound:
n
X
n
X
√
P max sup
|zj ||uij vij | ≥ t
P max sup
|zj ||uij vij | ≥ 4 2πKo2 (γ log n)3/2
1≤i≤n z∈Uk
j=1,j6=i
1≤i≤n z∈Uk
j=1,j6=i
n
X
1
e
2e
≤ 2γ−1 + 2(c γ−1) ≤ 2(c γ−1) ,
(47)
≤ P max sup
|zj ||uij vij | ≥ t E2 + P(E2c )
1
n
n
n 2
1≤i≤n z∈Uk
j=1,j6=i
where c2 = min (c1 , 1) is a universal positive constant.
n
X
5 The tail bound in (45) can also be obtained by using the law of total
≤ nP sup
|zj ||u1j v1j | ≥ t E2 + P(E2c ).
(42)
(c)
z∈Uk j=2
probability extended to conditional probabilities.
11
E. Proof of Lemma 5
Consider the tail event
)
(
n
X
(1 − α)δm2
.
E3 = sup
|zi | ||ai ||2 ||bi ||2 ≥ √
4 2πKo2 (γ log n)3/2
z∈Uk i=1
(48)
While bounding the tail probability P(E3 ), the supremum with
respect to z can be circumvented as shown below.
!
n
X
(1 − α)δm2
|zi |
√ ||ai ||2 ||bi ||2 ≥ √
P(E3 ) = P sup
3
4 2πkKo2 (γ log n) 2
z∈Uk i=1 k
(1 − α)δm2
. (49)
≤ P max ||ai ||2 ||bi ||2 ≥ √
1≤i≤n
4 2πkKo2 (γ log n)3/2
The above inequality is obtained by observing that
since z has unit P`2 -norm, by Cauchy-Schwarz inequaln
ity, we have √1k i=1 |zi | ≤ 1. Consequently, the linPn
zi
√
||ai ||2 ||bi ||2 can be at most
ear combination
i=1
k
max1≤i≤n ||ai ||2 ||bi ||2 , a quantity independent of z.
Since ai and bi , the column vectors of A and B, respectively, are i.i.d., the following series of union bounds apply.
(1 − α)δm2
√
P(E3 ) ≤ nP ||a1 ||2 ||b1 ||2 ≥
4 2πkKo2 (γ log n)3/2
s
!
(1 − α)δm2
√
≤ 2nP ||a1 ||2 ≥
4 2πkKo2 (γ log n)3/2
(1 − α)δm
2
≤ 2nP ||a1 ||2 −m ≥ m √
−1 .(50)
4 2πkKo2 (γ log n)3/2
≤ P
(1−α)δm
√
√
2 γ 3/2 k(log n)3/2
4 2πKo
−1
+P sup
o
and show that for sufficiently large m, P(E1 ) can be driven
arbitrarily close to zero, thereby√implying√that δ is a probabilistic upper bound for δk ((A/ m A/ m)). Now, P(E1 )
admits the following union bound:
T
P(E1 ) = P sup zT (A A) (A A) z − m2 ≥ δm2
z∈U
k
= P sup zT AT A ◦ AT A z − m2 ≥ δm2
z∈U
k
n X
n
X
2
= P sup
zi zj aTi aj − m2 ≥ δm2
z∈Uk i=1 j=1
2
||ai ||4
2
2
≥ αδm
−m
n
n
X
X
aTi aj
2
≥ (1 − α)δm2 .(53)
In the above, the second identity follows from Proposition 12.
The last inequality uses the triangle inequality followed by the
union bound, with α ∈ (0, 1) being a variational parameter to
be optimized later. Similar to the proof of Theorem 2, we now
derive separate upper bounds for each of the two probability
terms in (53).
The first term in (53) admits the following series of relaxations.
!
n
X
4
2
2
2
P sup
zi ||ai ||2 − m ≥ αδm
z∈Uk i=1
(a)
≤
sup
P
n
X
z∈Uk i=1
!
4
zi2 ||ai ||2 − m2 ≥ αδm2
(b)
4
max ||ai ||2
1≤i≤n
2
P
(c)
4
||a1 ||2 − m2 ≥ αδm2 .
2
2
2
nP ||a1 ||2 − m ≥ αβδm
α(1 − β)δm
2
+2nP ||a1 ||2 − m ≥
2
p
2
nP ||a1 ||2 − m ≥ αβδm
α(1 − β)δm
2
+2nP ||a1 ||2 − m ≥
.
2
αδm
αδ
2
3nP ||a1 ||2 − m ≥
1−
2
4
(d)
≤
(e)
≤
(f )
≤
(g)
≤
nP
−m
2
≤
≤
4
F. Proof Sketch for Theorem 3
The proof of Theorem 3 is along similar lines as that of
Theorem 2; the sketch here focuses on the differences in the
steps. We consider the tail event
(
)
2
A
A
√
√
E1 , sup
z −1 ≥δ ,
(52)
m
m
z∈Uk
2
!
zi2
zi zj
z∈Uk i=1
j=1,j6=i
≤ 2ne−cm/Ko .
(51)
√
8 2πKo2 γ 3/2 √
3/2
k(log n) , and with c > 0
provided m ≥
(1−α)δ
being a universal constant.
P(E3 ) ≤ 2ne
z∈Uk i=1
Since a1 is a subgaussian vector with independent entries, by
the Hanson-Wright concentration inequality in Theorem 13,
the desired tail probability can be obtained as follows:
cm
−K
4
sup
n
X
≥ αδm
2 2
−cm α4Kδ4 (1−αδ/4)2
3ne
o
.
(54)
In the above, step (a) is the triangle inequality. The inequality
in step (b) follows from the fact that nonnegative convex
combination of n arbitrary numbers is at most the maximum
among the n numbers. Step (c) is a union bound. Step (d)
is also a union bounding argument with β ∈ (0, 1) as a
variational parameter, combined with the fact that for any vec4
2
tor a, the triangle inequality ||a||2 − m2 ≤ ||a||2 − m
2
||a||2
2
+
2m
− m is always true. Step (f ) is obtained by choosing the union bound parameter β = αδ/4. Finally, step (g)
is the Hanson-Wright inequality (Theorem 13) applied to the
subgaussian vector a1 .
We now derive an upper bound for the second probability
term in (53). Note that
sup
n
n
X
X
z∈Uk i=1
j=1,j6=i
≤ sup
n
X
z∈Uk i=1
zi zj aTi aj
2
|zi | ||ai ||2
2
n
X
j=1,j6=i
|zj |u2ij
12
≤
n
X
sup
z∈Uk i=1
!
n
X
2
|zi | ||ai ||2 max sup
1≤i≤n
z0 ∈U
k
|zj0 |u2ij ,(55)
j=1, j6=i
where uij , h ||aaii|| , aj i. The inner product uij is a subgaus2
sian random variable satisfying ||uij ||ψ2 ≤ c4 K, where c4 is
a universal numerical constant.
Using (55) in (53) yields the following union bound.
n
n
X
X
2
≥ (1 − α)δm2
P sup
zi zj aTi aj
z∈Uk i=1
j=1,j6=i
≤
P
max sup
1≤i≤n z0 ∈Uk
+P
sup
n
X
z∈Uk i=1
n
X
j=1
j6=i
√
|zj0 |u2ij ≥ 4 2πKo2 (γ log n)3/2
2
|zi | ||ai ||2
(1 − α)δm2
≥ √
4 2πKo2 (γ log n)3/2
!
. (56)
The following Lemmas 6 and 7 provide upper bounds for the
two tail probability terms occurring in the RHS of (56).
provided m ≥ max
where ξ = max 1,
Proof. See Appendix G.
Lemma 7. Let A be an m×n random matrix with zero mean,
unit variance, i.i.d. subgaussian entries satisfying ||Aij ||ψ2 ≤
K. Then,
!
n
X
(1 − α)δm2
− cm
2
|zi | ||ai ||2 ≥ √
P sup
≤ ne Ko4 ,
2
3/2
4 2πKo (γ log n)
z∈Uk i=1
(57)
√
8 2πKo2 γ 3/2 √
3/2
provided m ≥
k
log
n.
Here,
c
>
0
is
a
(1−α)δ
universal constant.
Proof. The proof is similar to that of Lemma 5, which is given
in Appendix E.
By using the tail probability bounds in Lemmas 6 and 7
together in (56), we obtain the following upper bound.
n
n
X
X
2
P sup
zi zj aTi aj
≥ (1 − α)δm2
z∈Uk i=1
j=1,j6=i
e
≤
e
−cm/Ko4
1
+ ne
= 2(c γ−1) + ( cm −1)
n2(c2 γ−1)
n 2
n Ko4 log n
2e
≤
.
(58)
n2(c2 γ−1)
The √
above
inequality
is
true
for
m
≥
3
8 2πKo2 γ 3/2 √
(2c2 γ−1)Ko4 log n
2
max
k log n,
. By combining
(1−α)δ
c
(54), (56) and (58), we obtain the following tail bound.
P(E) ≤
3
n
cα2 δ 2 (1−αδ/4)2 m
−1
4 log n
4Ko
+
4
α2 δ 2 (1−αδ/4)2
(2c2 γ−1)ξKo4
log n
c
n,
,
. Finally, by choosing
−2(c2 γ−1)
α = 1/2, we have
P(E) ≤ 4en
, provided that
√
16 2πKo2 γ 3/2 √
(2c2 γ−1)ξKo4
3/2
m ≥ max
k
log
n,
log
n
,
δ
c
16
with ξ = max 1, δ2 (1−δ/8)2 . Note that, in terms of k and
√
n, the first term in the inequality for m scales as k log3/2 n;
it dominates the second term, which scales as log n.
G. Proof of Lemma 6
Let W ⊆ Rm×n be the collection of all m × n sized
matrices X with columns xi , i ∈ [n], and satisfying |uij | ,
√
|xT
i xj |
||xi ||2 < K 2γ log n for all 1 ≤ i, j ≤ n, j 6= i. Then, for a
subgaussian random matrix A,
n
o
[
p
P (A ∈
/ W) = P
|uij | ≥ K 2γ log n
1≤i,j≤n, j6=i
p
P |uij | ≥ K 2γ log n
X
≤
Lemma 6. Let A be an m × n random matrix as specified in
Theorem 3. Let uij , h ||aaii|| , aj i. Then,
2
n
X
√
3
e
Pmax sup
|zj |u2ij ≥ 4 2πKo2 (γ log n) 2≤ 2(c γ−1) ,
1≤i≤n z∈Uk
n 2
j=1,j6=i
where Ko , max (K, 1) and c2 > 0 is a universal constant.
√
3
8 2πKo2 γ 3/2√
k log 2
(1−α)δ
1≤i,j≤n, j6=i
2
p
n
P |u12 | ≥ K 2γ log n
2
en2 −2c1 γ log n
e
e
= 2(c γ−1) ,
2
2n 1
≤
≤
(60)
where c1 > 0 is a universal constant. In the above, the first
and second inequalities are due to the union bound and the
identically distributed nature of uij , j 6= i. The last inequality
is simply the subgaussian tail
√ bound. Thus, for γ ≥ 1/c1 , and
for sufficiently large n, K 2γ log n acts as a uniform upper
bound for the random variables |uij |, j 6= i with arbitrarily
high probability.
Let E3 denote the event that the matrix A ∈ W. Then, by
law of total probability,
n
X
P max sup
|zj ||uij |2 ≥ t
1≤i≤n z∈Uk
j=1,j6=i
≤ P max sup
1≤i≤n z∈Uk
≤ nP sup
n
X
z∈Uk j=2
n
X
|zj ||uij |2 ≥ t E3 + P(E3c )
j=1,j6=i
|zj ||u1j |2 ≥ t E3 + P(E3c ).
(61)
Conditioned on event E3 and the column a1 assuming
some fixed value, the random variables |u1j |2 are bounded,
nonnegative and i.i.d. random variables. By Lemma 3, it
follows that
n
n
X
X
P sup
|zj ||u1j |2 − E sup
|zj ||u1j |2 ≥ t E3 , a1
z∈Uk j=2
z∈Uk j=2
2
≤ e−t
/8K 4 γ 2 log2 n
. (62)
2e
4e
≤ 2(c γ−1) , (59) Since the above tail probability bound is independent of
n2(c2 γ−1)
n 2
a1 , it remains unchanged after marginalizing the conditional
13
probability in the LHS over all values of a1 , i.e.,
n
n
X
X
P sup
|zj ||u1j |2 − E sup
|zj ||u1j |2 ≥ t E3
z∈Uk j=2
z∈Uk j=2
≤ e−t
2
/8K 4 γ 2 log2 n
.
(63)
Further, using Proposition 15, the probability term in the LHS
of the inequality in (63) can be replaced by a smaller tail
probability as shown below.
n
X
√
P sup
|zj ||uij |2 ≥ t + 2 2πK 2 γ log n E3
z∈Uk j=2
2
≤ e−t
/8K 2 γ 2 log2 n
.
3/2
Setting t = 4K(γ log n) , we get
n
X
√
1
P sup
|zj ||u1j |2 ≥ 4 2πKo2 (γ log n)3/2 E3 ≤ 2γ .
n
z∈Uk j=2
(64)
Finally, by combining (60), (61) and (64), we obtain the
desired tail bound.
n
X
√
P max sup
|zj ||uij |2 ≥ 4 2πKo2 (γ log n)3/2
1≤i≤n z∈Uk
j=1,j6=i
1
e
e
+ 2(c γ−1) ≤ 2(c γ−1) ,
n2γ−1
2n 1
n 2
where c2 = min (1, c1 ) is a universal positive constant.
≤
R EFERENCES
[1] C. G. Khatri and C. R. Rao, “Solutions to some functional equations
and their applications to characterization of probability distributions,”
Sankhya: The Indian Journal of Statistics, Series A (1961-2002), vol. 30,
no. 2, pp. 167–180, 1968.
[2] D. S. Bernstein, Matrix Mathematics: Theory, Facts, and Formulas
(Second Edition). Princeton University Press, 2009.
[3] S. Liu and G. Trenkler, “Hadamard, Khatri-Rao, Kronecker and other
matrix products,” International Journal of Information and System
Sciences, vol. 4, no. 1, pp. 160–177, 2007.
[4] M. F. Duarte and R. G. Baraniuk, “Kronecker compressive sensing,”
IEEE Trans. Image Process., vol. 21, no. 2, pp. 494–504, Feb. 2012.
[5] A. Koochakzadeh and P. Pal, “Sparse source localization in presence of
co-array perturbations,” in International Conference on Sampling Theory
and Applications (SampTA), May 2015, pp. 563–567.
[6] D. Romero, D. D. Ariananda, Z. Tian, and G. Leus, “Compressive covariance sensing: Structure-based compressive sensing beyond sparsity,”
IEEE Signal Process. Mag., vol. 33, no. 1, pp. 78–93, Jan. 2016.
[7] G. Dasarathy, P. Shah, B. N. Bhaskar, and R. D. Nowak, “Sketching
sparse matrices, covariances, and graphs via tensor products,” IEEE
Trans. Inf. Theory, vol. 61, no. 3, pp. 1373–1388, Mar. 2015.
[8] W. K. Ma, T. H. Hsieh, and C. Y. Chi, “DOA estimation of quasistationary signals with less sensors than sources and unknown spatial
noise covariance: A Khatri-Rao subspace approach,” IEEE Trans. Signal
Process., vol. 58, no. 4, pp. 2168–2180, Apr. 2010.
[9] N. D. Sidiropoulos and A. Kyrillidis, “Multi-way compressed sensing
for sparse low-rank tensors,” IEEE Signal Process. Lett., vol. 19, no. 11,
pp. 757–760, Nov. 2012.
[10] D. L. Donoho, M. Elad, and V. N. Temlyakov, “Stable recovery of sparse
overcomplete representations in the presence of noise,” IEEE Trans. Inf.
Theory, vol. 52, no. 1, pp. 6–18, Jan. 2006.
[11] S. Foucart and H. Rauhut, A Mathematical Introduction to Compressive
Sensing. Birkhäuser Basel, 2013.
[12] E. J. Candès, J. K. Romberg, and T. Tao, “Stable signal recovery from
incomplete and inaccurate measurements,” Communications on Pure and
Applied Mathematics, vol. 59, no. 8, pp. 1207–1223, 2006.
[13] E. J. Candes and T. Tao, “Decoding by linear programming,” IEEE
Trans. Inf. Theory, vol. 51, no. 12, pp. 4203–4215, Dec 2005.
[14] R. Baraniuk, M. Davenport, R. DeVore, and M. Wakin, “A simple proof
of the restricted isometry property for random matrices,” Constructive
Approximation, vol. 28, no. 3, pp. 253–263, 2008.
[15] T. T. Cai, L. Wang, and G. Xu, “New bounds for restricted isometry
constants,” IEEE Trans. Inf. Theory, vol. 56, no. 9, pp. 4388–4394, Sept
2010.
[16] P. Pal and P. P. Vaidyanathan, “Pushing the limits of sparse support
recovery using correlation information,” IEEE Trans. Signal Process.,
vol. 63, no. 3, pp. 711–726, Feb. 2015.
[17] D. P. Wipf and B. D. Rao, “An empirical Bayesian strategy for solving
the simultaneous sparse approximation problem,” IEEE Transactions on
Signal Processing, vol. 55, no. 7, pp. 3704–3716, Jul. 2007.
[18] S. Khanna and C. R. Murthy, “On the support recovery of jointly
sparse Gaussian sources using sparse Bayesian learning,” CoRR, vol.
abs/1703.04930. [Online]. Available: http://arxiv.org/abs/1703.04930
[19] C. Carathéodory and L. Fejér, “Uber den zusammenhang der extremen
von harmonischen funktionen mit ihren koeffizienten un uber den picardlandauschen satz,” Rendiconti del Circolo Matematico di Palermo,
vol. 32, no. 1, pp. 218–239, 1911.
[20] P. Stoica and R. Moses, Spectral Analysis of Signals. Pearson Prentice
Hall, 2005.
[21] Z. Yang, L. Xie, and P. Stoica, “Vandermonde decomposition of multilevel Toeplitz matrices with application to multidimensional superresolution,” IEEE Trans. Inf. Theory, vol. 62, no. 6, pp. 3685–3701,
Jun. 2016.
[22] N. D. Sidiropoulos and A. Kyrillidis, “Multi-way compressed sensing
for sparse low-rank tensors,” IEEE Signal Processing Letters, vol. 19,
no. 11, pp. 757–760, Nov. 2012.
[23] D. Nion and N. D. Sidiropoulos, “A PARAFAC-based technique for
detection and localization of multiple targets in a MIMO radar system,”
in Proc. ICASSP, 2009, pp. 2077–2080.
[24] A. M. Tillmann and M. E. Pfetsch, “The computational complexity of the
restricted isometry property, the nullspace property, and related concepts
in compressed sensing,” IEEE Trans. Inf. Theory, vol. 60, no. 2, pp.
1248–1259, Feb. 2014.
[25] R. A. Horn and C. R. Johnson, Eds., Matrix Analysis. Cambridge
University Press, 1986.
[26] S. Jokar, “Sparse recovery and Kronecker products,” in 44th Annual
Conference on Information Sciences and Systems, Mar. 2010, pp. 1–4.
[27] A. Bhaskara, M. Charikar, A. Moitra, and A. Vijayaraghavan, “Smoothed
analysis of tensor decompositions,” in Proceedings of the 46th Annual
ACM Symposium on Theory of Computing, 2014, pp. 594–603.
[28] J. Anderson, M. Belkin, N. Goyal, L. Rademacher, and J. R. Voss, “The
more, the merrier: the blessing of dimensionality for learning large
Gaussian mixtures,” in 27th Annual Conference on Learning Theory,
Jun. 2014, pp. 1135–1164.
[29] N. D. Sidiropoulos and R. Bro, “On the uniqueness of multilinear
decomposition of N-way arrays,” Journal of Chemometrics, vol. 14,
no. 3, pp. 229–239, 2000.
[30] P. Koiran and A. Zouzias, “Hidden cliques and the certification of the
restricted isometry property,” IEEE Trans. Inf. Theory, vol. 60, no. 8,
pp. 4999–5006, Aug. 2014.
[31] S. Khanna and C. R. Murthy, “Rényi divergence based covariance
matching pursuit of joint sparse support,” in Sig. Process. Advances
in Wireless Commun. (SPAWC), Jul. 2017, pp. 1 – 5.
[32] R. Vershynin, “Introduction to the non-asymptotic analysis of random
matrices,” 2010.
[33] J. A. Tropp, “Greed is good: algorithmic results for sparse approximation,” IEEE Trans. Inf. Theory, vol. 50, no. 10, pp. 2231–2242, Oct.
2004.
[34] G. Visick, “A quantitative version of the observation that the Hadamard
product is a principal submatrix of the Kronecker product,” Linear
Algebra and its Applications, vol. 304, no. 13, pp. 45 – 68, 2000.
[35] R. A. Horn and C. R. Johnson, Topics in Matrix Analysis. Cambridge;
New York: Cambridge University Press, 1994.
[36] B. Mond and J. E. Pec̆arić, “Inequalities for the Hadamard product of
matrices,” SIAM J. Matrix Anal. Appl., vol. 19, no. 1, pp. 66–70, Jan.
1998.
[37] H. J. Werner, “Hadamard product of square roots of correlation matrices,” Image, vol. 26, pp. 1–32, Apr. 2001.
[38] A. W. Marshall and I. Olkin, “Reversal of the Lyapunov, Hölder,
and Minkowski inequalities and other extensions of the Kantorovich
inequality,” Journal of Mathematical Analysis and Applications, vol. 8,
no. 3, pp. 503 – 514, 1964.
14
[39] S. Liu and H. Neudecker, “Several matrix Kantorovich-type inequalities,” Journal of Mathematical Analysis and Applications, vol. 197, no. 1,
pp. 23 – 26, 1996.
[40] ——, “Kantorovich inequalities and efficiency comparisons for several
classes of estimators in linear models,” Statistica Neerlandica, vol. 51,
no. 3, pp. 345–355, 1997.
[41] S. Liu, “On the Hadamard product of square roots of correlation
matrices,” Econometric Theory, vol. 18, no. 4, p. 1007, Aug. 2002.
[42] C. R. Rao and M. B. Rao, Matrix Algebra and Its Applications to
Statistics and Econometrics. Singapore: World Scientific, 1998.
[43] R. Adamczak, “A note on the Hanson-Wright inequality for random
vectors with dependencies,” Electron. Commun. Probab., vol. 20, pp.
72–84, 2015.
[44] M. Rudelson and R. Vershynin, “Hanson-Wright inequality and subgaussian concentration,” Electron. Commun. Probab., vol. 18, pp. 82–91,
2013.
[45] S. Boucheron, G. Lugosi, and P. Massart, Concentration Inequalities:
A Nonasymptotic Theory of Independence. Oxford University Press,
2013.
| 7cs.IT
|
Approximating Min-Cost Chain-Constrained Spanning Trees: A
Reduction from Weighted to Unweighted Problems∗
arXiv:1605.03203v1 [cs.DS] 10 May 2016
André Linhares†
Chaitanya Swamy†
Abstract
We study the min-cost chain-constrained spanning-tree (abbreviated MCCST) problem: find a mincost spanning tree in a graph subject to degree constraints on a nested family of node sets. We devise
the first polytime algorithm that finds a spanning tree that (i) violates the degree constraints by at most
a constant factor and (ii) whose cost is within a constant factor of the optimum. Previously, only an
algorithm for unweighted CCST was known [14], which satisfied (i) but did not yield any cost bounds.
This also yields the first result that obtains an O(1)-factor for both the cost approximation and violation
of degree constraints for any spanning-tree problem with general degree bounds on node sets, where an
edge participates in a super-constant number of degree constraints.
A notable feature of our algorithm is that we reduce MCCST to unweighted CCST (and then utilize [14]) via a novel application of Lagrangian duality to simplify the cost structure of the underlying
problem and obtain a decomposition into certain uniform-cost subproblems.
We show that this Lagrangian-relaxation based idea is in fact applicable more generally and, for any
cost-minimization problem with packing side-constraints, yields a reduction from the weighted to the
unweighted problem. We believe that this reduction is of independent interest. As another application
of our technique, we consider the k-budgeted matroid basis problem, where we build upon a recent
1.5
rounding algorithm of [4] to obtain an improved nO(k /ǫ) -time algorithm that returns a solution that
satisfies (any) one of the budget constraints exactly and incurs a (1 + ǫ)-violation of the other budget
constraints.
1 Introduction
Constrained spanning-tree problems, where one seeks a minimum-cost spanning tree satisfying additional
({0, 1}-coefficient) packing constraints, constitute an important and widely-studied class of problems. In
particular, when the packing constraints correspond to node-degree bounds, we obtain the classical mincost bounded-degree spanning tree (MBDST) problem, which has a rich history of study [7, 11, 12, 5, 8, 16]
culminating in the work of [16] that yielded an optimal result for MBDST. Such degree-constrained networkdesign problems arise in diverse areas including VLSI design, vehicle routing and communication networks
(see, e.g., the references in [15]), and their study has led to the development of powerful techniques in
approximation algorithms.
Whereas the iterative rounding and relaxation technique introduced in [16] (which extends the iterativerounding framework of [10]) yields a versatile technique for handling node-degree constraints (even for
more-general network-design problems), we have a rather limited understanding of spanning-tree problems
with more-general degree constraints, such as constraints |T ∩ δ(S)| ≤ bS for sets S in some (structured)
∗
A preliminary version [13] will appear in the Proceedings of the 18th IPCO, 2016.
{alinhare,cswamy}@uwaterloo.ca. Dept. of Combinatorics and Optimization, University of Waterloo, Waterloo,
ON N2L 3G1. Research supported partially by the second author’s NSERC grant 327620-09, NSERC Discovery Accelerator
Supplement Award, and Ontario Early Researcher Award.
†
1
family S of node sets.1 A fundamental impediment here is our inability to leverage the techniques in [8, 16].
The few known results yield: (a) (sub-) optimal cost, but a super-constant additive- or multiplicative- factor
violation of the degree bounds [3, 1, 6, 2]; or (b) a multiplicative O(1)-factor violation of the degree bounds
(when S is a nested family), but no cost guarantee [14]. In particular, in stark contrast to the results known
for node-degree-bounded network-design problems, there is no known algorithm that yields an O(1)-factor
cost approximation and an (additive or multiplicative) O(1)-factor violation of the degree bounds. (Such
guarantees are only known when each edge participates in O(1) degree constraints [2]; see however [17] for
an exception.)
We consider the min-cost chain-constrained spanning-tree (MCCST) problem introduced by [14], which
is perhaps the most-basic setting involving general degree bounds where there is a significant gap in our
understanding vis-a-vis node-degree bounded problems. In MCCST, we are given an undirected connected
graph G = (V, E), nonnegative edge costs {ce }, a nested family S (or chain) of node sets S1 ( S2 ( · · · (
Sℓ ( V , and integer degree bounds {bS }S∈S . The goal is to find a minimum-cost spanning tree T such
that |δT (S)| ≤ bS for all S ∈ S, where δT (S) := T ∩ δ(S). Olver and Zenklusen [14] give an algorithm
for unweighted CCST that returns a tree T such that |δT (S)| = O(bS ) (i.e., there is no bound on c(T )),
and show that, for some ρ > 0, it is NP-complete to obtain an additive ρ · logloglog|V|V| | violation of the degree
bounds. We therefore focus on bicriteria (α, β)-guarantees for MCCST, where the tree T returned satisfies
c(T ) ≤ α · OPT and |δT (S)| ≤ β · bS for all S ∈ S.
Our contributions. Our main result is the first O(1), O(1) -approximation algorithm for MCCST. Given
λ
·OPT and |δT (S)| ≤ 9λ·bS for all S ∈ S, using
any λ > 1, our algorithm returns a tree T with c(T ) ≤ λ−1
the algorithm of [14] for unweighted CCST, denoted AOZ , as a black box (Theorem 3.3). As noted above,
this is also the first algorithm that achieves an O(1), O(1) -approximation for any spanning-tree problem
with general degree constraints where an edge belongs to a super-constant number of degree constraints.
We show in Section 4 that our techniques are applicable more generally. We give a reduction showing that for any cost-minimization problem with packing side-constraints, if we have an algorithm for the
unweighted problem that returns a solution with an O(1)-factor violation of the packing constraints and
satisfies a certain property, then one can utilize it to obtain an O(1), O(1) -approximation for the costminimization problem. Furthermore, we show that if the algorithm
for the unweighted counterpart satisfies
a stronger property, then we can utilize it to obtain a 1, O(1) -approximation (Theorem 5.1).
We believe that our reductions are of independent interest and will be useful in other settings as well.
Demonstrating this, we show an application to the k-budgeted matroid basis problem, wherein we seek to
2
find a basis satisfying k budget constraints. Grandoni et al. [9] devised an nO(k /ǫ) -time algorithm that
returned a (1, 1 + ǫ, . . . , 1 + ǫ)-solution: i.e., the solution satisfies (any) one budget constraint exactly and
violates the other budget constraints by a (1+ǫ)-factor (if the problem is feasible). Very recently, Bansal and
1.5
Nagarajan [4] improved the running time to nO(k /ǫ) but return only a (1+ǫ, . . . , 1+ǫ)-solution. Applying
our reduction (to the algorithm in [4]), we obtain the best of both worlds: we return a (1, 1 + ǫ, . . . , 1 + ǫ)1.5
solution in nO(k /ǫ) -time (Theorem 5.7).
The chief novelty in our algorithm and analysis, and the key underlying idea, is an unorthodox use of
Lagrangian duality. Whereas typically Lagrangian relaxation is used to drop complicating constraints and
thereby simplify the constraint structure of the underlying problem, in contrast, we use Lagrangian duality
to simplify the cost structure of the underlying problem by equalizing edge costs in certain subproblems. To
elaborate (see Section 3.1), the algorithm in [14] for unweighted CCST can be viewed as taking a solution x
to the natural linear-programming (LP) relaxation for MCCST, converting it to another feasible solution x′
satisfying a certain structural property, and exploiting this property to round x′ to a spanning tree. The main
1
Such general degree constraints arise in the context of finding thin trees [1], where S consists of all node sets, which turn out
to be a very useful tool in devising approximation algorithms for asymmetric TSP.
2
bottleneck here in handling costs (as also noted in [14]) is that c⊺ x′ could be much larger than c⊺ x since the
conversion ignores the ce s and works with an alternate “potential” function.
∗
Our crucial insight is that we can exploit Lagrangian duality to obtain perturbed edge costs {cye }
such that the change in perturbed cost due to the conversion process is ∗bounded.
Loosely speaking, if
y
y∗
the conversion process shifts weight from xf to xe , then we ensure that ce = cf (see Lemma 3.5); thus,
∗
∗
(cy )⊺ x = (cy )⊺ x′ ! The perturbation also ensures that applying AOZ to x′ yields a tree whose perturbed cost
∗
∗
∗
is equal to (cy )⊺ x′ = (cy )⊺ x. Finally, we show that for an optimal LP solution x∗ , the “error” (cy − c)⊺ x∗
∗
incurred in working with the cy -cost is O(OPT ); this yields the O(1), O(1) -approximation.
We extend the above idea to an arbitrary cost-minimization problem with packing side-constraints as
follows. Let x∗ be an optimal solution to the LP-relaxation, and P be the polytope obtained by dropping the
packing constraints. We observe that the same Lagrangian-duality based perturbation ensures that all points
on the minimal face of P containing x∗ have the same perturbed cost. Therefore, if we have an algorithm
for the unweighted problem that rounds x∗ to a point x̂ on this minimal face, then we again obtain that
∗
∗
(cy )⊺ x̂ = (cy )⊺ x∗ , which then leads to an O(1), O(1) -approximation (as in the case of MCCST).
Related work. Whereas node-degree-bounded spanning-tree problems have been widely studied, relatively few results are known for spanning-tree problems with general degree constraints for a family S of
node-sets. With the exception of the result of [14] for unweighted CCST, these other results [3, 1, 6, 2]
all yield a tree of cost at most the optimum with an ω(1) additive- or multiplicative- factor violation of the
degree bounds. Both [3] and [2] obtain additive factors via iterative rounding and relaxation. The factor
in [3] is (r − 1) for an arbitrary S, where r is the maximum number of degree constraints involving an edge
(which could be |V | even when S is a chain), while [2] yields an O(log |V |) factor when S is a laminar
family (the factor does not improve when S is a chain). The dependent-rounding techniques in [1, 6] yield a
tree T satisfying |δT (S)| ≤ min O logloglog|S||S| bS , (1 + ǫ)bS + O logǫ|S| for all S ∈ S, for any family S.
For MBDST, Goemans [8] obtained the first O(1), O(1) -approximation; his result yields a tree of cost
at most the optimum and at most +2 violation of the degree bounds. This was subsequently improved to an
(optimal) additive +1 violation by [16]. Zenklusen [17] considers an orthogonal generalization of MBDST,
where there is a matroid-independence constraint on the edges incident to each node, and obtains a tree of
cost at most the optimum and “additive” O(1) violation (defined appropriately) of the matroid constraints.
To our knowledge, this is the only prior work that obtains an O(1)-approximation to both the cost and
packing constraints for a constrained spanning-tree problem where an edge participates in ω(1) packing
constraints (albeit this problem is quite different from spanning tree with general degree constraints).
Finally, we note that our Lagrangian-relaxation based technique is somewhat similar to its use in [11].
However, whereas [11] uses this to reduce uniform-degree MBDST to the problem of finding an MST of
minimum maximum degree, which is another weighted problem, we utilize Lagrangian relaxation in a more
refined fashion to reduce the weighted problem to its unweighted counterpart.
2 An LP-relaxation for MCCST and preliminaries
We consider the following natural LP-relaxation for MCCST. Throughout, we use e to index the edges of
the underlying graph G = (V, E). For a set S ⊆ V , let E(S) denote {uv ∈ E : u, v ∈ S}, and δ(S) denote
3
the edges on the boundary of S. For a vector z ∈ RE and an edge-set F , we use z(F ) to denote
min
X
ce xe
P
e∈F
ze .
(P)
e
s.t.
x E(S) ≤ |S| − 1
x(E) = |V | − 1
x δ(S) ≤ bS
x ≥ 0.
∀∅ =
6 S(V
∀S ∈ S
(1)
(2)
(3)
(4)
For any x ∈ RE
+ , let supp(x) := {e : xe > 0} denote the support of x. It is well known that the polytope,
PST (G), defined by (1), (2), and (4) is the convex hull of spanning trees of G. We call points in PST (G)
fractional spanning trees. We refer to (1), (2) as the spanning-tree
constraints. We will also utilize (Pλ ),
the modified version of (P) where we replace (3) with x δ(S) ≤ λbS for all S ∈ S, where λ ≥ 1. Let
OPT (λ) denote the optimal value of (Pλ ), and let OPT := OPT (1).
Preliminaries. A family L ⊆ 2V of sets is a laminar family if for all A, B ∈ L, we have A ⊆ B or B ⊆ A
or A ∩ B = ∅. As is standard, we say that A ∈ L is a child of L ∈ L if L is the minimal
set of L such
L , E L ) be the graph obtained from L, E(L) by contracting the
that A ( L. For each L ∈ L, let GL
=
(V
L
L
L
children of L in L; we drop the superscript L when L is clear from the context.
Given x ∈ PST (G), define a laminar decomposition L of x to be a (inclusion-wise)
maximal laminar
family of sets whose spanning-tree constraints are tight at x, so x E(A) = |A| − 1 for all A ∈ L. Note
that V ∈ L and {v} ∈ L for all v ∈ V . A laminar decomposition can be constructed in polytime (using
network-flow techniques). For any L ∈ L, let xL
L , or simply xL if L is clear from context, denote x restricted
to EL . Observe that xL is a fractional spanning tree of GL .
3 An LP-rounding approximation algorithm
3.1 An overview
We first give a high-level overview. Clearly, if (P) is infeasible, there is no spanning tree satisfying the
degree constraints, so in the sequel, we assume that (P) is feasible. We seek to obtain a spanning tree T
of cost c(T ) = O(OPT ) such that |δT (S)| = O(bS ) for all S ∈ S, where δT (S) is the set of edges of T
crossing S.
In order to explain the key ideas leading to our algorithm, we first briefly discuss the approach of Olver
and Zenklusen [14] for unweighted CCST. Their approach ignores the edge costs {ce } and instead starts
with a feasible solution x to (P) that minimizes a suitable (linear) potential function. They use this potential
function to argue that if L is a laminar decomposition of x, then (x, L) satisfies a key structural property
called rainbow freeness. Exploiting this, they give a rounding algorithm, hereby referred to as AOZ , that for
every L ∈ L, rounds xL to a spanning tree TL of GL such that |δTL (S)| ≈ O xL (δ(S)) for all S ∈ S, so
that concatenating the TL s yields a spanning tree T of G satisfying |δT (S)| = O x(δ(S)) = O(bS ) for all
S ∈ S (Theorem 3.2). However, as already noted in [14], a fundamental obstacle towards generalizing their
approach to handle the weighted version (i.e., MCCST) is that in order to achieve rainbow freeness, which
is crucial for their rounding algorithm, one needs to abandon the cost function c and work with an alternate
potential function.
We circumvent this difficulty as follows. First, we note that the algorithm in [14] can be equivalently
viewed as rounding an arbitrary solution x to (P) as follows. Let L be a laminar decomposition of x.
Using the same potential-function idea, we can convert x to another solution x′ to (P) that admits a laminar
4
decomposition L′ refining L such that (x′ , L′ ) satisfies rainbow freeness (see Lemma 3.1), and then round
x′ using AOZ . Of course, the difficulty noted above remains, since the move to rainbow freeness (which
again ignores c and uses a potential function) does not yield any bounds on the cost c⊺ x′ relative to c⊺ x.
We observe however that there is one simple property (*) under which one can guarantee that c⊺ x′ = c⊺ x,
namely, if for every L ∈ L, all edges in supp(x) ∩ EL have the same cost. However, it is unclear how to
utilize this observation since there is no reason to expect our instance to have this rather special property:
for instance, all edges of G could have very different costs!
Now let x∗ be an optimal solution to (P) (we will later modify this somewhat) and L be a laminar
decomposition of x∗ . The crucial insight that allows us to leverage property (*), and a key notable aspect
of our algorithm and analysis, is that one can leverage Lagrangian duality to suitably perturb the edge
costs so that the perturbed costs satisfy property (*). More precisely, letting y ∗ ∈ RS+ denote the optimal
values∗ of the dualPvariables corresponding to constraints (3), if we define the perturbed cost of edge e to
∗
be cye := ce + S∈S:e∈δ(S) yS∗ , then the cy -cost of all edges in supp(x∗ ) ∩ EL are indeed equal, for
every L ∈ L (Lemma 3.5). In essence, this holds because for any
e′ ∈ supp(x∗ ), by complementary
P
slackness, we have ce′ = (dual contribution to e′ from (1),(2)) − S∈S:e′ ∈δ(S) yS∗ . Since any two edges
e, f ∈ supp(x∗ ) ∩ EL appear in the same sets of L, one can argue that the dual contributions to e and f
∗
∗
from (1), (2) are equal, and thus, cye = cyf .
∗
Now since (x∗ , L∗ ) satisfies (*) with the perturbed costs cy , we can convert (x∗ , L∗ ) to (x′ , L′ ) satisfying rainbow freeness without altering the perturbed cost, and then round x′ to a spanning tree T using AOZ . This immediately yields |δT (S)| = O(bS ) for all S ∈ S. To bound the cost, we argue that
P
P ∗
∗
c(T ) ≤ cy (T ) = e cye x∗e = c⊺ x∗ + S∈S bS yS∗ (Lemma 3.6), where the last equality followsPfrom complementary slackness. (Note that the perturbed costs are used only in the analysis.) However, S∈S bS yS∗
need not be bounded in terms of c⊺ x∗ . To fix this, we modify our starting solution x∗ : we solve (Pλ ) (which
recall is (P) with inflated degree bounds {λbS }), where λ > 1, to obtain x∗ , and use this x∗ in our algorithm.
Now, letting y ∗ be the optimal dual values corresponding to the inflated degree constraints, a simple duality
P
(λ)
argument shows that S∈S bS yS∗ ≤ OPT (1)−OPT
(Lemma 3.7), which yields c(T ) = O(OPT ) (see
λ−1
Theorem 3.3).
A noteworthy feature of our algorithm is the rather unconventional use of Lagrangian relaxation, where
we use duality to simplify the cost structure (as opposed to the constraint-structure) of the underlying problem by equalizing edge costs in certain subproblems. This turns out to be the crucial ingredient that allows us
to utilize the algorithm of [14] for unweighted CCST as a black box without worrying about the difficulties
posed by (the move to) rainbow freeness. In fact, as we show in Sections 4 and 5, this Lagrangian-relaxation
idea is applicable more generally, and yields a novel reduction from weighted problems to their unweighted
counterparts. We believe that this reduction is of independent interest and will find use in other settings as
well; indeed, we demonstrate another application of our ideas in Section 5.2.
3.2 Algorithm details and analysis
To specify our algorithm formally, we first define the rainbow-freeness property and state the main result
of [14] (which we utilize as a black box) precisely.
For an edge e, define Se := {S ∈ S : e ∈ δ(S)}. Note that Se could be empty. We say that two edges
e, f ∈ E form a rainbow if Se ⊆ Sf or Sf ⊆ Se . (This definition is slightly different from the one in [14], in
that we allow Se = Sf .) We say that (x, L) is a rainbow-free decomposition if L is a laminar decomposition
of x and for every set L ∈ L, no two edges of supp(x) ∩ EL form a rainbow. (Recall that GL = (VL , EL )
denotes the graph obtained from L, E(L) by contracting the children of L.) The following lemma shows
that one can convert an arbitrary decomposition (x, L) to a rainbow-free one; we defer the proof to the
Appendix. (As noted earlier, this lemma is used to equivalently view the algorithm in [14] as a rounding
algorithm that rounds an arbitrary solution x to (P).)
5
Lemma 3.1. Let x ∈ PST (G) and L be a laminar decomposition of x. We can compute in polytime a
fractional spanning tree x′ ∈ PST (G) and a rainbow-free decomposition (x′ , L′ ) such that: (i) supp(x′ ) ⊆
supp(x); (ii) L ⊆ L′ ; and (iii) x′ (δ(S)) ≤ x(δ(S)) for all S ∈ S.
Theorem 3.2 ([14]). There is a polytime algorithm, AOZ , that given a fractional spanning tree x′ ∈ PST (G)
and a rainbow-free decomposition (x′ , L′ ), returns a spanning tree TL ⊆ supp(x′ ) of GL for everyL ∈ L′
such that the concatenation T of the TL s is a spanning tree of G satisfying |δT (S)| ≤ 9x′ δ(S) for all
S ∈ S.
We can now describe our algorithm quite compactly. Let λ > 1 be a parameter.
1. Compute an optimal solution x∗ to (Pλ ), a laminar decomposition L of x∗ .
2. Apply Lemma 3.1 to (x∗ , L) to obtain a rainbow-free decomposition (x′ , L′ ).
′
′
′
3. Apply AOZ to the input (x′ , L′ ) to obtain spanning trees TLL of GL
L for every L ∈ L . Return the
′
concatenation T of all the TLL s.
Analysis.
We show that the above algorithm satisfies the following guarantee.
Theorem 3.3. The above algorithm run with parameter λ > 1 returns a spanning tree T satisfying c(T ) ≤
λ
λ−1 · OPT and |δT (S)| ≤ 9λbS for all S ∈ S.
The bound on |δT (S)| follows immediately from Lemma 3.1 and Theorem 3.2 since x∗ , and hence x′
obtained in step 2, is a feasible solution to (Pλ ). So we focus on bounding c(T ). This will follow from
∗
three things. First, we define the perturbed cy -cost precisely. Next, Lemma 3.5 proves the key result
that for every L ∈ L, all edges in supp(x∗ ) ∩ EL have the same perturbed cost. Using this it is easy to
P
P ∗
∗
show that c(T ) ≤ cy (T ) = e cye x∗e = OPT (λ) + λ S∈S bS yS∗ (Lemma 3.6). Finally, we show that
P
OPT −OPT (λ)
∗
(Lemma 3.7), which yields the bound stated in Theorem 3.3.
S∈S bS yS ≤
λ−1
To define the perturbed costs, we consider the Lagrangian dual of (Pλ ) obtained by dualizing the (inflated) degree constraints x δ(S) ≤ λbS for all S ∈ S:
X
X
max
x(δ(S)) − λbS )yS .
(LDλ )
ce xe +
gλ (y) :=
min
y∈RS
+
x∈PST (G)
e
S∈S
P
P
For y ∈ RS , let Gλ,y (x) := e ce xe + S∈S x(δ(S)) − λbS ySP= e cye xe − λ S∈S bS yS denote the
objective function of the LP that defines gλ (y), where cye := ce + S∈S:e∈δ(S) yS .
P
P
∗
Let y ∗ be an optimal solution to (LDλ ). Our perturbed costs are {cye }. We prove the following preliminary lemma, then proceed to show that the perturbed costs satisfy property (*).
Lemma 3.4. We have gλ (y ∗ ) = Gλ,y∗ (x∗ ) = OPT (λ).
P
Proof. For any y ∈ RS+ , we have gλ (y) + λ S∈S bS yS =
X
min
cye xe s.t. x E(S) ≤ |S| − 1 ∀∅ =
6 S ( V,
|
e
X
max −
(|S| − 1)µS
∅6=S⊆V
|
{z
x(E) = |V | − 1,
(Pλ,y )
s.t.
−
X
∅6=S⊆V :
e∈E(S)
µS ≤ cye ∀e ∈ E,
{z
(Dλ,y )
6
x≥0 =
}
µS ≥ 0 ∀∅ =
6 S(V
}
where the second equality follows since (Dλ,y ) is the dual of (Pλ,y ). It follows that
gλ (y ∗ ) = max gλ (y) = max −
y∈RS
+
X
∅6=S⊆V
(|S| − 1)µS − λ
−
s.t.
X
∅6=S⊆V :
e∈E(S)
µS −
X
bS y S
(Dλ )
S∈S
X
S∈S:
e∈δ(S)
yS ≤ ce
y ≥ 0, µS ≥ 0
∀e ∈ E
∀∅ =
6 S ( V.
Notice that (Dλ ) is the dual of (Pλ ), hence, gλ (y ∗ ) = OPT (λ). Moreover, it also follows that ŷ is an optimal
solution to (LDλ ) iff there exists µ̂ = (µ̂S )∅6=S⊆V such that (µ̂, ŷ) is an optimal solution to (Dλ ).
So let µ∗ be such that (µ∗ , y ∗ ) is an optimal solution to (Dλ ). It follows
that x∗ and (µ∗ , y ∗ ) satisfy
complementary slackness. So we have that if µ∗S > 0 then x∗ E(S) = |S| − 1, and if x∗e > 0 then
P
P
P
∗
− ∅6=S⊆V :e∈E(S) µ∗S − S∈S:e∈δ(S) yS∗ = ce , or equivalently cye = − ∅6=S⊆V :e∈E(S) µ∗S . Therefore,
Gλ,y∗ (x∗ ) =
X
e
=−
=−
∗
cye x∗e − λ
X
∅6=S⊆V
X
∅6=S⊆V
X
bS yS∗ =
X
−
e
S∈S
X
∅6=S⊆V :e∈E(S)
X
µ∗S x∗ E(S) − λ
bS yS∗
X
bS yS∗
µ∗S x∗e − λ
S∈S
S∈S
(|S| −
1)µ∗S
−λ
X
bS yS∗ = gλ (y ∗ ).
S∈S
∗
Lemma 3.5. For any L ∈ L, all edges of supp(x∗ ) ∩ EL have the same cy -cost.
y∗
Proof. Consider any two edges e, f ∈ supp(x∗ ) ∩ EL . Suppose for a contradiction that cy∗
e < cf . Obtain
/ {e, f }). Using the same
x̂ from x∗ by increasing x∗e by ǫ and decreasing x∗f by ǫ (so x̂e′ = x∗e′ for all e′ ∈
argument as in the proof of Lemma 3.1, one can show that x̂ ∈ PST (G) for a sufficiently small ǫ > 0. Since
∗
∗
cye < cyf , we have gλ (y ∗ ) ≤ Gλ,y∗ (x̂) < Gλ,y∗ (x∗ ) = gλ (y ∗ ), where the last equality follows from Lemma
3.4, and we obtain a contradiction.
P ∗
P
P
Lemma 3.6. We have c(T ) ≤ e cye x∗e = e ce x∗e + λ S∈S bS yS∗ .
∗
Proof. Observe that c(T ) ≤ cy (T ) since ce ≤ cye for all e ∈ E as y ∗ ≥ 0. We now bound cy (T ). To
∗ L
∗
keep notation simple, we use GL = (VL , EL ) and x∗L to denote GL
L and (x′ )L (which recall is x restricted
′
′
∗ L
to ELL ) respectively, and P
G′L = (VL′ , EL′ ) and x∗L to denote GL
L and (x )L respectively.
∗
∗
y
y
We have c (T ) =
L∈L c (T ∩ EL ) since the sets {EL }L∈L partition E. Fix L ∈ L. Note that
x∗L is a fractional spanning tree of GL = (VL , EL ) since for any ∅ =
6 Q ⊆ VL , if R is the subset of V
corresponding to Q, and A1 , . . . , Ak areP
the children of L whose corresponding contracted nodes lie in Q,
we have x∗L EL (Q) = x∗ E(R) − ki=1 x∗ E(Ai ) ≤ |R \ (A1 ∪ . . . ∪ Ak )| + k − 1 = |Q| − 1
with equality holding when Q = VL . Note that T ∩ EL is a spanning tree of GL since T is obtained by
concatenating spanning trees for the graphs {G′L′ }L′ ∈L′ , and L′ refines L. Also, all edges of supp(x∗ ) ∩ EL
P
∗
∗
∗
have the same cy -cost by Lemma 3.5. So we have cy (T ∩ EL ) = e∈EL cye x∗e . It follows that
∗
∗
cy (T ) =
X
∗
cye x∗e =
e
=
X
e
X
ce x∗e +
e
ce x∗e
+
∗
X
S∈S
yS∗ x∗
X
yS∗ x∗e
S∈S:e∈δ(S)
X
X
δ(S) =
ce x∗e + λ
bS yS∗ .
e
7
S∈S
Lemma 3.7. We have
P
∗
S∈S bS yS
≤
OPT (1)−OPT (λ)
.
λ−1
Proof. By Lemma 3.4, we have that
OPT (λ) = gλ (y ∗ ) = Gλ,y∗ (x∗ ).
Using Lemma 3.4 again, now with λ = 1, and since y ∗ is a feasible solution to (LD1 ), we obtain that
OPT (1) = maxy∈RS g1 (y) ≥ g1 (y ∗ ). Note that the objective functions of the LPs defining g1 (y ∗ ) and
+
P
gλ (y ∗ ) differ by a constant: G1,y∗ (x) − Gλ,y∗ (x) = (λ − 1) S∈S bS yS∗ for all x ∈ PST (G). Since x∗ is an
optimal solution to minx∈PST (G) Gλ,y∗ (x), it is also an optimal solution to minx∈PST (G) G1,y∗ (x). It follows
that
OPT (1) ≥ g1 (y ∗ ) = G1,y∗ (x∗ ) .
P
Therefore, OPT (1) − OPT (λ) ≥ G1,y∗ (x∗ ) − Gλ,y∗ (x∗ ) = (λ − 1) S∈S bS yS∗ .
Proof of Theorem 3.3. As noted earlier, the bounds on δT (S)
from Lemma 3.1 and
follow immediately
′ δ(S) ≤ 9x∗ δ(S) ≤ 9λb . The bound on c(T )
Theorem 3.2: for any S ∈ S, we have |δ
(S)|
≤
9x
S
PT
follows from Lemmas 3.6 and 3.7 since e ce x∗e = OPT (λ).
4 A reduction from weighted to unweighted problems
We now show that our ideas are applicable more generally, and yield bicriteria approximation algorithms
for any cost-minimization problem with packing side-constraints, provided we have a suitable approximation algorithm for the unweighted counterpart. Thus, our technique yields a reduction from weighted to
unweighted problems, which we believe is of independent interest.
To demonstrate this, we first isolate the key properties of the rounding algorithm B used above for
unweighted CCST that enable us to use it as a black box to obtain our result for MCCST; this will yield
an alternate, illuminating explanation of Theorem 3.3. Note that B is obtained by combining the procedure
in Lemma 3.1 and AOZ (Theorem 3.2). First, we of course utilize that B is an approximation
algorithm
∗
for unweighted CCST, so it returns a spanning tree T such that |δT (S)| = O x (δ(S)) for all S ∈ S.
Second, we exploit the fact that B returns a tree T that preserves tightness of all spanning-tree constraints
that are tight at x∗ . This is the crucial property that allows us to bound c(T ), since this implies (as we
P y∗ ∗
∗
explain below) that cy (T ) =
e ce xe , which then yields the bound on c(T ) as before. The equality
follows because the set of optimal solutions to the LP minx∈PST (G) Gλ,y∗ (x) is a face of PST (G); thus all
points on the minimal face of PST (G) containing x∗ are optimal solutions to this LP, and the stated property
implies that the characteristic vector of T lies on this minimal face. In other words, while AOZ proceeds
by exploiting the notions of rainbow freeness and laminar decomposition, these notions are not essential to
obtaining our result; any rounding algorithm for unweighted CCST satisfying the above two properties can
be utilized to obtain our guarantee for MCCST.
We now formalize the above two properties for an arbitrary cost-minimization problem with packing
side-constraints, and prove that they suffice to yield a bicriteria guarantee. Consider the following abstract
, and b ∈ Rm
problem, where P ⊆ Rn+ is a fixed polytope: given c ∈ Rn+ , A ∈ Rm×n
+ , find
+
min c⊺ x
s.t.
x is an extreme point of P,
Ax ≤ b.
(QP )
Observe that we can cast MCCST as a special case of (QP ), by taking P = PST (G) (whose extreme points
are spanning trees of G), c to be the edge costs, and Ax ≤ b to be the degree constraints. Moreover,
by taking P to be the convex hull of a bounded set {x ∈ Zn+ : Cx ≤ d} we can use (QP ) to encode a
discrete-optimization problem.
8
We say that B is a face-preserving rounding algorithm (FPRA) for unweighted (QP ) if given any point
x ∈ P, B finds in polytime an extreme point x̂ of P such that:
(P1) x̂ belongs to the minimal face of P that contains x.
We say that B is a β-approximation FPRA (where β ≥ 1) if we also have:
(P2) Ax̂ ≤ βAx.
⊺
P
P
Let (RP
λ ) denote the LP min c x : x ∈ P, Ax ≤ λb ; note that (R1 ) is the LP-relaxation of (Q ).
Let opt(λ) denote the optimal value of (RP
λ ), and let opt := opt(1). We say that an algorithm is a (ρ1 , ρ2 )approximation algorithm for (QP ) if it finds in polytime an extreme point x̂ of P such that c⊺ x̂ ≤ ρ1 opt and
Ax̂ ≤ ρ2 b.
Theorem 4.1. LetB be a β-approximation FPRA for unweighted (QP ). Then, given any λ > 1, one can
λ
obtain a λ−1
, βλ -approximation algorithm for (QP ) using a single call to B.
Proof sketch. We dovetail the algorithm for MCCST and its analysis. We simply compute an optimal solution x∗ to (RP
λ ) and round it to an extreme point x̂ of P using B. By property (P2), it is clear that
Ax̂ ≤ β(Ax∗ ) ≤ βλb.
y
⊺
For y ∈ Rm
+ , define c := c + A y. To bound the cost, as before, we consider the Lagrangian dual of
(RP
λ ) obtained by dualizing the side-constraints Ax ≤ λb.
h
(y)
:=
min
H
(x)
, where Hλ,y (x) := (cy )⊺ x − λy ⊺ b.
max
λ
λ,y
m
y∈R+
x∈P
Let y ∗ = argmaxy∈Rm
hλ (y). We can mimic the proof of Lemma 3.4 to show that x∗ is an optimal solution
+
to minx∈P Hλ,y∗ (x). The set of optimal solutions to this LP is a face of P. So all points on the minimal face
of P containing x∗ are optimal solutions to this LP. By property (P1), x̂ belongs to this minimal face and so
∗
∗
is an optimal solution to this LP. So (cy )⊺ x̂ = (cy )⊺ x∗ = c⊺ x∗ + (y ∗ )⊺ Ax∗ = opt(λ) + λ(y ∗ )⊺ b, where
the last equality follows by complementary slackness. Also, by the same arguments as in Lemma 3.7, we
∗
∗
λ
. Since c ≤ cy , we have c⊺ x̂ ≤ (cy )⊺ x̂ ≤ λ−1
· opt.
have (y ∗ )⊺ b ≤ opt(1)−opt(λ)
λ−1
5 Towards a 1, O(1) -approximation algorithm for (QP )
A natural question that emerges from Theorems 3.3 and 4.1 is whether one can obtain a 1, O(1) -approximation,
i.e., obtain a solution of cost at most opt that violates the packing side-constraints by an (multiplicative)
O(1)-factor. Such results are known for degree-bounded spanning tree problems with various kinds of
degree constraints [8, 16, 3, 17], so, in particular, it is natural to ask whether such a result also holds
for MCCST. (Note that for MCCST, the dependent-rounding techniques in [1, 6] yield a tree T with
for all S ∈ S.) We show
c(T ) ≤ OPT and |δT (S)| ≤ min O logloglog|S||S| bS , (1 + ǫ)bS + O logǫ|S|
that our approach is versatile enough to yield such a guarantee provided we assume a stronger property from
the rounding algorithm B for unweighted (QP ).
Let Ai denote the i-th row of A, for i = 1, . . . , m. We say that B is an (α, β)-approximation FPRA for
unweighted (QP ) if in addition to properties (P1) and (P2), it satisfies:
(P3)
⊺
it rounds a feasible solution x to (RP
α ) to an extreme point x̂ of P satisfying Ai x̂ ≥
such that A⊺i x = αbi .
A⊺i x
α
for every i
(For MCCST, property (P3) requires that |δT (S)| ≥ bS for every set S ∈ S whose degree constraint (in
(Pα )) is tight at the fractional spanning tree x.)
Theorem 5.1. Let B be an (α, β)-approximation FPRA for unweighted (QP ). Then, one can obtain a
(1, αβ)-approximation algorithm for (QP ) using a single call to B.
9
Proof. We show that applying the algorithm from Theorem 4.1 with λ = α yields the claimed result. It
is clear that the extreme point x̂ returned satisfies Ax̂ ≤ αβb. As in the proof of Theorem 4.1, let y ∗
be an optimal solution to maxy∈Rm
hλ (y) (where λ = α). In Lemma 3.6 and the proof of Theorem 4.1,
+
∗ ⊺
⊺
y
we use the weak bound c x̂ ≤ (c ) x̂. We tighten this to obtain the improved bound on c⊺ x̂. We have
∗
c⊺ x̂ = (cy )⊺ x̂ − (y ∗ )⊺ Ax̂, and
(y ∗ )⊺ Ax̂ =
X
i:A⊺i x∗ =λbi
yi∗ (A⊺i x̂) ≥
X
i:A⊺i x∗ =λbi
yi∗ A⊺i x∗
=
α
X
yi∗ bi = (y ∗ )⊺ b.
i:A⊺i x∗ =λbi
The first and last equalities above follow because yi∗ > 0 implies that A⊺i x∗ = λbi . The inequality follows
from property (P3). Thus, following the rest of the arguments in the proof of Theorem 4.1, we obtain that
∗
c⊺ x̂ ≤ (cy )⊺ x̂ − (y ∗ )⊺ b = c⊺ x∗ + (λ − 1)(y ∗ )⊺ b ≤ opt(1).
5.1 Obtaining an additive approximation for (QP ) and cost at most opt via an FPRA with
two-sided additive guarantees
We now present a variant of Theorem 5.1 that shows that we can achieve cost at most opt and additive
approximation for the packing side constraints using an FPRA with two-sided additive guarantees. We
give an application of this result in Section 5.2, where we utilize it to obtain improved guarantees for the
k-budgeted matroid basis problem.
Theorem 5.2. Let B be an FPRA for unweighted (QP ) that given x ∈ P returns an extreme point x̂ of P
such that Ax − ∆ ≤ Ax̂ ≤ Ax + ∆, where ∆ ∈ Rm
+ may depend on A and c (but not on b). Using a single
call to B, we can obtain an extreme point x̃ of P such that c⊺ x̃ ≤ opt and Ax̃ ≤ b + 2∆.
The above result is obtained via essentially the
⊺same arguments as those in Theorems 4.1 and 5.1. For
P
a vector ∆ ∈ Rm
,
let
(W
)
denote
the
LP
min
c x : x ∈ P, Ax ≤ b + ∆ . Let ~0 denote the all-zeros
+
∆
vector, and note that (W~P
)
is
the
LP-relaxation
of
(QP ). Let opt(∆) denote the optimum value of (WP
∆ ),
0
P
and let opt := opt(~0). The Lagrangian dual of (W∆ ) obtained by dualizing the side-constraints Ax ≤ b+ ∆
is
ϕ
(y)
:=
min
Φ
(x)
,
(LD∆ )
max
∆
∆,y
m
x∈P
y∈R+
where Φ∆,y (x) := (cy )⊺ x − y ⊺ (b + ∆). (Recall that cy := c + A⊺ y.) Let x∗ be an optimal solution to (WP
∆)
∗
ϕ∆ (y). We have the following variants of Lemmas 3.4 and 3.7.
and y = argmaxy∈Rm
+
Lemma 5.3. We have ϕ∆ (y ∗ ) = Φ∆,y∗ (x∗ ) = opt(∆).
Proof. This follows by mimicking the arguments used in the proof of Lemma 3.4.
Lemma 5.4. We have (y ∗ )⊺ ∆ ≤ opt(~0) − opt(∆).
Proof. We mimic the proof of Lemma 3.7. By Lemma 5.3, we have that
opt(∆) = ϕ∆ (y ∗ ) = Φ∆,y∗ (x∗ )
and opt(~0) = maxy∈RS ϕ~0 (y) ≥ ϕ~0 (y ∗ ). Note that Φ~0,y∗ (x) − Φ∆,y∗ (x) = (y ∗ )⊺ ∆, which is inde+
pendent of x. So since x∗ is an optimal solution to minx∈P Φ∆,y∗ (x), it is also an optimal solution to
minx∈P Φ~0,y∗ (x). It follows that
opt(~0) ≥ ϕ~0 (y ∗ ) = Φ~0,y∗ (x∗ ).
Hence, opt(~0) − opt(∆) ≥ Φ~0,y∗ (x∗ ) − Φ∆,y∗ (x∗ ) = (y ∗ )⊺ ∆.
10
Proof of Theorem 5.2. The algorithm simply computes an optimal solution x∗ to (WP
∆ ), and rounds it to an
extreme point x̃ of P using algorithm B.
It is clear that Ax̃ ≤ Ax∗ + ∆ ≤ (b + ∆) + ∆ = b + 2∆. Next we argue that c⊺ x̃ ≤ opt. We have
∗
c⊺ x̃ = (cy )⊺ x̃ − (y ∗ )⊺ Ax̃, and
X
X
(y ∗ )⊺ Ax̃ =
yi∗ (A⊺i x̃) ≥
yi∗ (A⊺i x∗ − ∆i )
i:A⊺i x∗ =bi +∆i
=
X
i:A⊺i x∗ =bi +∆i
yi∗ bi = (y ∗ )⊺ b.
i:A⊺i x∗ =bi +∆i
By Lemma 5.3, x∗ is an optimal solution to minx∈P Ψ∆,y∗ (x). So all points on the minimal face of P
containing x∗ are optimal solutions to this LP. In particular, since x̃ belongs to this minimal face (by property
(P1)), x̃ is an optimal solution to this LP. This observation, along with the inequality above, yields c⊺ x̃ ≤
∗
(cy )⊺ x∗ − (y ∗ )⊺ b = opt(∆) + (y ∗ )⊺ ∆. Finally, using Lemma 5.4 yields c⊺ x̃ ≤ opt(~0) as required.
5.2 Application to k-budgeted matroid basis
Here, we seek toP
find a basis S of a matroid M = (U, I) satisfying k budget constraints {di (S) ≤ Bi }1≤i≤k ,
where di (S) := e∈S di (e). Note that this can be cast a special case of (QP ), where P = P(M ) is the basis
polytope of M , the objective function encodes a chosen budget constraint (say the k-th budget constraint),
and Ax ≤ b encodes the remaining budget constraints. We show that our techniques, combined with a recent
1.5
randomized algorithm of [4], yields a (randomized) algorithm that, for any ǫ > 0, returns in nO(k /ǫ) time
a basis that (exactly) satisfies the chosen budget constraint, and violates the other budget constraints by
(at most) a (1 + ǫ)-factor, where n := |U | is the size of the ground-set of M . This matches the currentbest approximation guarantee of [9] (who give a deterministic algorithm) and the current-best running time
of [4].
Theorem 5.5 ([4]). For some constant ν > 0, there exists a randomized FPRA, BBN , √
for unweighted
(QP(M )√
) that rounds any x ∈ P(M ) to an extreme point x̂ of P(M ) such that Ax − ν k∆ ≤ Ax̂ ≤
Ax + ν k∆, where ∆ = (max1≤j≤n aij )1≤i≤k−1 = (maxe∈U di (e))1≤i≤k−1 .
Lemma 5.6. There exists
√ a polytime randomized algorithm that finds a basis S of M such that dk (S) ≤ Bk ,
and di (S) ≤ Bi + 2ν k maxe∈U di (e) for all 1 ≤ i ≤ k − 1, or determines that the instance is infeasible.
Proof. As explained above, we cast the problem as a special case of (QP ) by using the k-th budget constraint
as the objective function, and the remaining budget constraints as packing side-constraints. If the LPrelaxation of (QP ) is infeasible, then the budgeted-matroid-basis instance is infeasible. Otherwise, the
above guarantee follows by applying Theorem 5.2 with the algorithm B=BBN .
Using ideas from [4], we combine the algorithm from Lemma 5.6 with a partial enumeration step as
follows. We say an element e ∈ U is heavy if the inequality di (e) > 2νǫ√k Bi holds for at least one index
i ∈ {1, . . . , k}. Let H denote the set of all heavy elements. We state our algorithm below. Let ǫ > 0 be a
parameter.
e ⊆ H of size |H|
e ≤ 2νk1.5 , we do the following.
1. For every set H
ǫ
e and deleting the
(a) Let M ′ be the matroid obtained from M by contracting the elements of H
e
elements of H \ H.
e for i ∈ {1, . . . , k}.
(b) Compute residual budgets Bi′ := Bi − di (H),
(c) Run the algorithm from Lemma 5.6 on matroid M ′ with budgets {Bi′ }1≤i≤k .
11
(d) If the algorithm succeeds (that is, if the LP that it attempts to solve is feasible), then let T be
e ∪ T . If S is a basis of M , dk (S) ≤ Bk , and
the set of elements returned, and let S := H
di (S) ≤ (1 + ǫ)Bi for all 1 ≤ i ≤ k − 1, then return S.
2. If step 1 does not return any set S, then return that the instance is infeasible.
1.5
Theorem 5.7. The algorithm above, run with parameter ǫ > 0, finds in nO(k /ǫ) time a basis S of M such
that dk (S) ≤ Bk and di (S) ≤ (1 + ǫ)Bi for all 1 ≤ i ≤ k − 1, or determines that the instance is infeasible.
2νk1.5
1.5
= nO(k /ǫ) . Since steps 1(a)–1(d) run in
Proof. Note that the number of iterations is at most n ǫ
1.5 /ǫ)
O(k
poly(n) time, the overall running time is n
as claimed.
If the instance is infeasible, then any outcome of the algorithm (infeasible, or a basis S) is correct.
(Note that due to the verification done at the end of step 1(d), any set S returned must have the required
properties.) So assume that the instance is feasible, and let S ∗ be a basis of M that exactly satisfies all
the budget constraints. We argue that in this case the algorithm does indeed return a basis with the desired
properties. Let H ∗ := S ∗ ∩H be the set of heavy elements that S ∗ contains. Note that since a heavy element
uses up at least one budget to an extent greater than 2νǫ√k , and since S ∗ satisfies all the k budget constraints,
1.5
e = H ∗ (if the algorithm
we must have |H ∗ | ≤ kǫ = 2νk . Note that at the iteration corresponding to H
2ν
√
k
ǫ
reaches it), the set S ∗ \ H ∗ is feasible for the residual problem (with a matroid M ′ and residual budgets
{Bi′ } defined in steps 1(a) and 1(b)). Further, note that this set also certifies that the resulting set S satisfies
dk (S) = dk (H ∗ ) + dk (T ) ≤ dk (H ∗ ) + dk (S ∗ \ H ∗ ) = dk (S ∗ ) ≤ Bk . Finally, for every i ∈ {1, . . . , k − 1},
we have
√
di (S) = di (H ∗ ) + di (T ) ≤ di (H ∗ ) + Bi′ + 2ν k max di (e)
√
ǫ
≤ Bi + 2ν k √ Bi = (1 + ǫ)Bi ,
2ν k
e∈U \H
and so the set S will pass the verification done at step 1(d) and will be returned by the algorithm.
References
[1] A. Asadpour, M. Goemans, A. Madry, S. Oveis Gharan, and A. Saberi. An O(log n/ log log n)approximation algorithm for the asymmetric traveling salesman problem. In Proceedings of the 20th
SODA, pages 379–389, 2010.
[2] N. Bansal, R. Khandekar, J. Könemann, V. Nagarajan, and B. Peis. On generalizations of network
design problems with degree bounds. Mathematical Programming, 141(1-2):479–506, 2013.
[3] N. Bansal, R. Khandekar, and V. Nagarajan. Additive guarantees for degree-bounded directed network
design. SICOMP, 39(4):1413–1431, 2009.
[4] N. Bansal and V. Nagarajan. Approximation-friendly discrepancy rounding. To appear in Proc., IPCO
2016. Also appears as arXiv:1512.02254, 2015.
[5] K. Chaudhuri, S. Rao, S. Riesenfeld, and K. Talwar. What would Edmonds do? Augmenting paths and
witnesses for degree-bounded MSTs. Algorithmica 55:157–189, 2009.
[6] C. Chekuri, J. Vondrak, and R. Zenklusen. Dependent randomized rounding via exchange properties
of combinatorial structures. In 51st FOCS, 2010.
12
[7] M. Fürer and B. Raghavachari. Approximating the minimum-degree Steiner Tree to within one of
optimal. Journal of Algorithms 17(3):409–423, 1994.
[8] M. Goemans. Minimum bounded degree spanning trees. In 47th FOCS, 2006.
[9] F. Grandoni, R. Ravi, M. Singh, and R. Zenklusen. New approaches to multi-objective optimization.
Mathematical Programming, 146(1-2):525–554, 2014.
[10] K. Jain. A factor 2 approximation algorithm for the generalized Steiner network problem. Combinatorica, 21:39–60, 2001.
[11] J. Könemann and R. Ravi. A matter of degree: improved approximation algorithms for degree-bounded
minimum spanning trees. SICOMP, 31:1783–1793, 2002. SIAM Journal on Computing, 31:1783–
1793, 2002.
[12] J. Könemann and R. Ravi. Primal-dual meets local search: approximating MST’s with nonuniform
degree bounds. In Proceedings of the 35th STOC, 389–395 2003.
[13] A. Linhares and C. Swamy. Approximating min-cost chain-constrained spanning trees: a reduction
from weighted to unweighted problems. To appear in Proceedings of the 18th IPCO, 2016.
[14] N. Olver and R. Zenklusen. Chain-constrained spanning trees. In Proceedings of the 16th IPCO, pages
324–335, 2013.
[15] R. Ravi, M. Marathe, S. Ravi, D. Rosenkrantz, and H. Hunt III. Approximation algorithms for degreeconstrained minimum-cost network-design problems. Algorithmica 31(1):58–78, 2001.
[16] M. Singh and L. Lau. Approximating minimum bounded degree spanning trees to within one of
optimal. In Proceedings of the 39th STOC, pages 661–670. 2007.
[17] R. Zenklusen. Matroidal degree-bounded minimum spanning trees. In Proceedings of the 23rd SODA,
pages 1512–1521, 2012.
A
Proof of Lemma 3.1
This follows from essentially the same potential-function argument as used in [14] to obtain a rainbow-free
solution. Sort the edges of supp(x) in increasing order of |Se | breaking ties arbitrarily. Let e1 , e2 , . . . , ek
denote this ordering. Let w ∈ RE be any weight function such that we1 < we2 < · · · < wek (e.g., wei = i
for all i). Let x′ be an optimal solution to the following LP. (Note that the LP has variables {ze }e∈E , and
that the {xe }e∈E values are fixed.)
X
min
we ze
(P’)
e
s.t.
z ∈ PST (G), ze = 0
z δ(S) ≤ x δ(S)
z E(L) = |L| − 1
∀e ∈
/ supp(x)
∀S ∈ S
∀L ∈ L.
Properties (i) and (iii) hold by construction. Since we force the spanning-tree constraints corresponding to
sets in L to be tight, we can start with L and extend it to obtain a laminar decomposition L′ of x′ that refines
L, so (ii) holds.
13
It remains to show that (x′ , L′ ) is a rainbow-free decomposition. Consider any set L ∈ L′ and any two
′
edges e, f ∈ supp(x′ )∩ELL , and suppose that e, f form a rainbow. Let we < wf , so we must have Se ⊆ Sf .
Now perturb x′ by adding ǫ to x′e (the argument below will show that x′e < 1) and subtracting ǫ from x′f ,
where ǫ > 0 is chosen to be suitably small; let x′′ be this perturbed vector. Clearly, wT x′′ < wT x′ , so if
we show that x′′ is feasible to (P’),
then we obtain a contradiction. Clearly,
supp(x′′ ) ⊆
supp(x). Since
Se ⊆ Sf it follows that x′′ δ(S) ≤ x δ(S) for all S ∈ S. Also, x′′ E(L) = x′ E(L) = |L| − 1 for all
L ∈ L.
Finally, we show
that x′′ ∈ PST (G) for a sufficiently
small ǫ > 0. (Hence, x′e < x′′e ≤ 1.) For A ⊆ V
such thatx′ E(A) < |A| − 1, we obtain x′′ E(A) ≤ |A| − 1 by taking ǫ > 0 suitably small; for A with
x′ E(A) = |A| − 1, we obtain x′′ E(A) = |A| − 1 since the spanning-tree constraints for all L ∈ L′ are
tight at (x′ and) x′′ and these span all other tight spanning-tree constraints.
14
| 8cs.DS
|
Networked Intelligence
Towards Autonomous Cyber Physical Systems
André Karpištšenko
[email protected]
Abstract
Developing intelligent systems requires the combination of results from both industry and academia. In this
report you ind an overview of relevant research ields and industrially applicable technologies for building
large scale cyber physical systems. A conceptual architecture is used to illustrate how existing pieces may it
together, and the maturity of the subsystems is estimated.
The goal is to structure developments and the challenge of machine intelligence for Consumer and Industrial
Internet technologists, cyber physical systems researchers and people interested in the convergence of data
and the Internet of Things (IoT). This report will be useful for planning developments of intelligent systems.
Acknowledgements: I would like to thank Ando Saabas for help in inding the focus, Lauri Koobas and Erki
Suurjaak for pushing for brevity and clarity, Meelis Kull, Tambet Matiisen, Anna Leontjeva, Taivo Pungas and
Kairit Sirts for improvements. I’d also like to thank Machine Learning Estonia presenters for providing the
information on the latest state in machine learning, and Skype ML and AI reading group for updates in the
ield. I am grateful to the early readers as well.
________
Technical Report, 2016
Technical Report
Networked Intelligence, 2016
Executive Summary
The bene its of intelligent systems are demonstrated by many arti icial narrow intelligence (ANI) applications
used by businesses, governments and society. Some are highlighted later in the Applications section. In
developing intelligent systems there is a positive feedback loop ‐ as the products, services and processes get
better, the increase in pro itability is reinvested into further developments. There are many more innovations
one can imagine. Only what exists on the market today is highlighted here.
Cyber physical systems integrate computational and physical capabilities to interact with humans, machines,
and the environment. This report brings together the progress in developments of such systems to develop a
vision of a concept architecture for the machine intelligence challenge. To demonstrate the feasibility of the
concept we give an overview of relevant existing research results and technologies necessary for building it.
The report is structured as follows. In the Research Progress the focus is mainly on progress made in
computer science, while highlighting some of the adjacent ields. The Technology Action provides an overview
of trends in data and IoT. Core infrastructure such as data & IoT platforms and advanced analytics platforms
are maturing with potential high‐value commercial applications. Relevant to the intelligent systems, the
highest maturity is in data gathering and preprocessing, followed by advanced analytics.
In the illustrative section How the Pieces Fit Together, a hierarchical system of systems approach is used,
consisting of parts related to data preprocessing, models and self‐improvement. Higher‐order functions
related to inding solutions with limited resources1 have been left out, while the aspects where the
involvement of people is necessary are highlighted. This vision is not intended as a technical blueprint. I
estimated that the concept illustrated here is achievable within a 5‐10 year time frame.
Bringing together research and technology developments is necessary for engineering intelligent systems.
Hopefully this report will spark developments and conversations that speed up the progress.
The report is intended for Consumer and Industrial Internet technologists, cyber physical systems researchers
and people interested in the convergence of data and the IoT. It will be invaluable for planning the
development of intelligent systems, identifying bottlenecks, estimating resources and selecting technologies.
How to read guide: The irst sections give the context for the vision. If you are practically minded, start from
Technology Action. If you know the market and seek ideas for your system, start from the section How the
Pieces Fit Together.
1
R. Kurzweil, How to Create a Mind: The Secret of Human Thought Revealed, Penguin Books, 2012
Technical Report
Networked Intelligence, 2016
Research Progress
The aim of this section is to describe the research ields directly relevant to developing and designing large scale
autonomous cyber physical systems2.
One of the foundations for developing large scale systems is the system of systems concept. That is, large scale
integrated systems which are diverse and independently operable on their own, but are networked together
for a common goal.3
The ield of arti icial general intelligence (AGI)4 is the forefront of related research ields. The approach here is
pragmatic and business driven. Given the possible long road to practical AGI, there is no focus here on
providing a complete overview. In the industry, combinations of ANIs are used to solve complex situations.
The Figure 1 below illustrates how different research ields it together to form networked systems. It is
assumed that creating intelligent systems is mainly a software and systems engineering challenge, thus ields
like mechanical and electrical engineering, which focus on physical objects and their applications, are
excluded. However, some of the relevant results from these ields are listed in the Appendix.
Figure 1. Illustration of relevant research ields needed the for design and development of networks of systems
In Figure 1, research is grouped into one of four categories depending on the key aspect they address. In the
following sections you will ind a brief overview and highlights with links to technical material.
2
3
4
Partially inspired by IARPA Research Programs
M. Jamshidi, Systems of Systems Engineering: Principles and Applications, CRC Press, 2008
P. Wang, B. Goertzel, Theoretical Foundations of Arti icial General Intelligence, Atlantis Press, 2012
Technical Report
Networked Intelligence, 2016
Input Interfaces
The ields most relevant to the systems’ ability to take input from external systems and to observe context and
environment are information theory, information fusion5 and knowledge representation6.
Information theory de ines the limits of systems coping with exponentially increasing7 amount of incoming
data in processing and storage infrastructure. Information fusion deals with the art of merging information
from heterogeneous sources. For cyber physical systems, a relevant subset of the ield is sensor fusion8. It is
becoming commonplace that information fusion is a stream processing activity that operates with continuous
data lows. In modern data processing and analysis systems, multimodal information integration is the most
time consuming activity. This is due to high variety of data formats, versioning of data and variety in content.
Successful fusion relies on data curation and domain knowledge that can be partially automated. Knowledge
representation is about information required to understand incoming data. The extra structural, semantic and
meta information allows machines to process raw data into something meaningful.
There are different types of data that cyber physical systems transform into a knowledge‐base: textual data,
usually in natural language, but also in formal languages; image data, for example photographs, satellite
imagery and other; video data; audio data; numeric data and other binary signals. At large scales, a key aspect
is to provide means to cope with large datasets which may also contain errors.
Internal Processing
The following research ields aim at making sense of the incoming and historic information and internal
system state to ind patterns, make predictions and take decisions to deliver solutions for applications.
The basis is to make stored data, information and knowledge available in machine‐readable and
machine‐interpretable formats. Indexing and probabilistic retrieval methods decide what data can be
ef iciently accessed and retrieved within query results. Information extraction and retrieval9 is a ield that has
evolved rapidly after the onset of web era and is highly relevant to working with text, audio transcripts and
image & video captions. Knowledge organization and validation related functionalities are further layers on
top of related technologies.
Probability theory that form the foundations for many aspects in intelligent systems, such as information
retrieval, machine learning (ML) and knowledge discovery, requires highlighting. Probabilistic programming
languages10 focus on simplifying software engineering that deals with probabilities, so that less code needs to
be written. Bayesian inference and subjective Bayesian probability form an important basis for designing
models capable of dynamic problem solving.
5
V. Torra, Information Fusion in Data Mining, Springer, 2003
R. Brachman, H. Levesque, Knowledge Representation and Reasoning, Morgan Kaufmann 2004
7
EMC Digital Universe with Research & Analysis by IDC
8
Mitchell, H. B. Multi‐sensor Data Fusion, Springer‐Verlag, 2007
9
C. Manning, P. Raghavan, H. Schütze, Introduction to Information Retrieval, Cambridge University Press, 2008
10
http://probabilistic‐programming.org/
6
Technical Report
Networked Intelligence, 2016
Machine Learning and Knowledge Discovery
Machine Learning11 is a rapidly advancing ield with a range of different statistical learning methods12 which
can be used for supervised, partially supervised and unsupervised learning. A method quickly gaining traction
is deep learning13 which is about representation learning and automated feature engineering. The branch is
used in many industrial applications. Other popular methods are random forests and gradient boosting.
An increasing trend is one of hybrid intelligent systems where multiple methods are combined to bene it from
the strengths of each14. The approach combines neural networks, expert systems, fuzzy logic, symbolic
systems, genetic algorithms and case‐based reasoning.
A common aspect of machine learning is its dependence on data. There exists a subset of intelligently solvable
problems that require less data, like those based on chemical and physical processes. Such models are based
on systems of differential equations and laws of nature. There is a trend of replacing some classical algorithms
with neural network based approximations to improve performance15.
In developing intelligent systems one must choose between models as well as create ensemble models. To
ease the process of communication, models should be interpretable. See the example of random forests16.
Computational Learning and Process Automation
Computational learning theory deals with the design and analysis of machine learning algorithms, mainly
concerned with proving algorithms. Here a broader approach is taken. Self‐improvement processes17 and the
applications of machine learning to guide model development are not new. For example, neuroevolution uses
evolutionary algorithms for training18 and online machine learning deals with continuous sequential data,
updating the model at each step. As more models are produced, model curation and deployment automation is
necessary.
Development of new models involves data preparation steps such as normalization, semi‐automated feature
extraction and dimensionality reduction. To perform data exploration for useful pattern discovery in
high‐dimensional data, methods such as topological data analysis19 can be used. Some progress has been
made in automating feature relationship identi ication20.
Other Fields
There are other important ields, such as data mining, that are not covered here. In building cyber physical
systems there are more aspects than just working with the data and models. For example to decide on actions,
statistical decision theory facilitates identi ication of decision input uncertainties and inding a solution. To
11
K. Murphy, Machine Learning: A Probabilistic Perspective, MIT Press, 2012
T. Hastie, R. Tibshirani, J. Friedman, The Elements of Statistical Learning: Data Mining, Inference, and Prediction, Second Edition,
Springer, 2009
13
I. Goodfellow, Y. Bengio and A. Courville, Deep Learning, MIT Press (in preparation), 2016
14
L. Medsker, Hybrid Intelligent Systems, Springer, 2013
12
15
K. Abhishek, M.P. Singh, S. Ghosh, A. Anand, Weather Forecasting Model using Arti icial Neural Network, C3IT, 2012
A. Saabas, Interpreting random forests, Diving into data blog, 2014
17
J. Dunn, Introducing FBLearner Flow: Facebook's AI backbone, Facebook Code, 2016
18
D. Floreano, P. Dürr, C. Mattiussi, Neuroevolution, Evolutionary Intelligence, 2008
19
G. Carlsson, Topology and Data, Bulletin of the American Mathematical Society, 2009
20
D. Reshef, Y. Reshef, H. Finucane, S. Grossman, G. McVean, P. Turnbaugh, E. Lander, M. Mitzenmacher, P. Sabeti, Detecting novel
associations in large datasets, Science, 2011
16
Technical Report
Networked Intelligence, 2016
deal with optimally unsolvable situations,21 heuristics can be used. Attention based methods invest resources
based on areas of interest22.
In industries such as inance, where systems are in co‐opetition, means to think strategically are built using
game theory23. While this aspect is not relevant for inding a solution in a one off application, it becomes
relevant when operating continuously. Finding solutions for useful situations requires explicit accounting for
internal and external resource constraints. Optimization theory deals with selecting the best element from a
set of available alternatives. In machine learning, regularization deals with over itting.
For autonomous systems, situation awareness that originated in human factors research has recently
advanced to hierarchical situation awareness for cyber physical systems24. This is relevant to perceiving
context and comprehending it.
Neuroscience25, especially computational neuroscience, provides inspiration for ways of structuring
intelligent systems. Progress in neuroscience26 can be applicable in creating autonomous cyber physical
systems. Initial progress has been made with robots27.
Output Interfaces
For interacting with external systems and the environment, general system theory has broad applicability. The
most relevant class is cyber physical systems28 that refers to systems with integrated computational and
physical capabilities that interact with humans and machines.
Systems have to deal with dynamic behaviour and control theory provides guidance for process control in
order to steer a system towards a desired state. Many AI control challenge concepts and behaviours are
manifested in simple setups29. Control interfaces can take many forms, especially for human‐machine
interaction, for example natural language interfaces. Human‐machine interaction is an essential part of an
intelligent system. Extended intelligence30 is an emerging ield studying intelligence as a fundamentally
distributed phenomenon.
Networked Processing
System of systems forms a dynamically changing interactive network of processes that encode or decode
information from one form to another. Network theory provides a means to understand the form and function
of the system that the network represents31. Utilizing network theory, system designers can start to reason
about its structure, network resilience and resistance to perturbations. Advancements in control of networks
32
provide an initial framework for controlling complex self‐organized systems. More speci ically, one can
21
C. Moore, S. Mertens, The Nature of Computation, Oxford University Press, 2011
A. Almahairi, N. Ballas, T. Cooijmans, Y. Zheng, H. Larochelle, A. C., Dynamic Capacity Networks, ICLR 2016
23
S. Hargreaves Heap, Y. Varoufakis, Game Theory: A Critical Introduction, Routledge, 1995
24
J. Preden, Enhancing Situation‐Awareness, Cognition and Reasoning of Ad‐Hoc Network Agents, Tallinn University of Technology, 2010
25
Human Brain Project, EU FET Flagship Initiative
26
B. Baars, Global workspace theory of consciousness: toward a cognitive neuroscience of human experience, Progress in Brain Research,
2005
27
S. Bringsjord, J. Licato, N. Govindarajulu, R. Ghosh, A. Sen, Real Robots that Pass Human Tests of Self‐Consciousness, IEEE Ro‐Man, 2015
28
R. Baheti, H. Gill, Cyber‐physical Systems, The Impact of Control Theory, IEEE Control Systems Society, 2011
29
J. Tallinn, Toy Model of the Control Problem, 2015, http://www.slideshare.net/AndrKarpitenko/ai‐control
30
http://www.pubpub.org/pub/extended‐intelligence
31
M. Newman, Networks: An Introduction, Oxford University Press, 2010
32
Y. Liu, J. Slotine, A. Barabasi, Controllability of Complex Networks, Nature, 2011
22
Technical Report
Networked Intelligence, 2016
identify the set of driver nodes with time‐dependent control to guide the system’s entire dynamics. However,
sparse inhomogeneous networks, which emerge in many real complex systems, are the most dif icult to
control. In addition networked systems may be represented as multilayer structures33, even evolving as a
function of time34. Methodologies dealing with such representations sets the contemporary directions in
network science with straightforward applications in system design, IoT, and interdependent systems.
Knowledge about the network can be classi ied and visualized for designers and operators35.
Drawing the parallel between a dedicated system and an agent, a traditional multi‐agent systems approach
provides concepts for structuring organizations of agents that interact with the environment. Using
organizational theory as inspiration for organizing a cyber physical system of systems (for example the ields
of self‐organization, scienti ic management and organizational processes) can aid designers in choosing the
right protocols for evolving their systems.
Risks emerge when a system interacts with people and other systems. Cybersecurity is about managing these
risks with appropriate measures. For example cryptographic methods can be used for exchanging sensitive
information. Information security is a cross‐cutting concern and an important aspect of any system, especially
so for intelligent systems interacting with people and nature. Intelligent system design should factor in threat
intelligence and modeling to account for the risks of its actions and operations.
Technology Action
An overview of state of the art in technologies relevant to large scale autonomous cyber physical systems36.
These are fun times for anyone in software and systems engineering. The systems are becoming intelligent,
boundaries between ields are blurring and the pace of development is increasing. Mature trends like mobile
and new exponential trends like Industrial IoT, Consumer IoT, natural interfaces and data science are creating
new commercial and technological opportunities. The list of technologies and companies is long, here a
summary of some promising developments is provided.
33
Kivelä, M., Arenas, A., Barthelemy, M., Gleeson, J.P., Moreno, Y. and Porter, M.A., 2014. Multilayer networks. Journal of complex networks,
2(3), pp.203‐271.
34
Holme, P. and Saramäki, J., 2012. Temporal networks. Physics reports, 519(3), pp.97‐125.
K. Börner, Atlas of Science: Visualizing What We Know, MIT Press, 2010
36
Inspired by market research done by Shivon Zilis and FirstMark Capital
35
Technical Report
Networked Intelligence, 2016
Figure 2. Technology areas relevant to designing and building intelligent systems
The technologies relevant to building cyber physical systems are grouped into three categories: Infrastructure
Layer, Processing Layer and Usage Layer as illustrated in Figure 2. In the following sections you will ind an
overview of the categories37.
Infrastructure Layer
The infrastructure offers basic building blocks for creating commercially viable cyber physical systems. The
technologies de ine what is achievable in meeting non‐functional requirements.
Details on the overview of the Infrastructure Layer are explained in the Appendix. The base infrastructure
exists for development of autonomous cyber physical systems and is advancing fast. The Data & IoT Platforms,
as well as Data Exchanges address the need to adapt to the increasing amount of devices, machines and
protocols.
Processing Layer
The processing layer delivers the basis for building the core functionality of intelligent systems.
37
For paper readers, link to companies and technologies are provided, use the links to access the highlights directly.
Technical Report
Networked Intelligence, 2016
Data & IoT Platforms
Data and devices are the cornerstones of cyber physical systems. The available datasets and data streams
de ine the scope of applications. Processing, storing, indexing and querying the data (e.g. Drill) are the most
fundamental functions.
Data platforms address working with data, companies like Cloudera (recently launched Kudu), MapR and
others provide open source and proprietary technologies that scale to very large data sizes. Company
products like HortonWorks HDF further extend these technologies with IoT ready solutions that integrate
open‐source technologies from organizations like LinkedIn (Ka ka), Twitter (Storm) and NSA (Ni i). Further
progress is made by companies like DataBricks (Spark), Data Artisans (Flink) and others who commercialize
open‐source research results of thought leaders in data systems like UC Berkeley (AMPLab), MIT (DBg), CMU
(DB Group) and other institutions. A driving trend is the move from general purpose data warehouses and
relational databases to purpose built databases for small‐data, columnar, time‐series, array, key‐value and
other types data. Relevant to intelligent systems, probabilistic platforms are emerging. Cloud platform
providers commoditize the most widely used data solutions for petabyte scale needs.
While the solutions for big data are maturing and converging, the early phase of the IoT trend has seen a
response by many different companies like Cisco Jasper and PTC ThingWorx who are competing for the
leading position in different markets. Driven by the data and IoT trends, traditional companies like GE are
rede ining themselves as software‐driven service companies and are strategically investing into partnerships
38
. Companies such as Pivotal are recognizing this major shift and deliver software for those enterprises.
Younger companies for example Helium and others, are addressing IoT development needs to speed up the
time it takes for new connected devices to reach mass markets. Many IoT development platforms, like
Hologram, focus on the unique needs that come with dealing with many sensors, devices and data streams.
Data and IoT trends converge in platforms like Planet OS that combine open‐source, proprietary and cloud
technologies to deliver real value in demanding industries like energy. Reactive platforms like Lightbend form
a good basis for building such platforms. Building upon such endeavours brings the bene it of machine and
sensor independent data formats and protocols.
Data Preprocessing and Exchanges
Finding the right data, cleaning, curating and enriching it is one of the most time‐consuming activities in data
science. An intelligent system depends on the quality of input directly. Thus it makes sense to rely on
dedicated technologies like Tamr, Trifacta and CrowdFlower in these activities. For data integration,
platforms like snapLogic and others exist.
Preprocessing endeavours are commonly too large for a single entity. A solution is delivered by data
syndication and publishing services, such as Xignite, Planet OS, Enigma, Qlik and potentially Microsoft Azure
DataMarket, that provide high quality datasets and data streams via well‐de ined APIs. Building upon Data
Exchanges and preprocessing related technologies is a good use of resources.
38
https://www.oracle.com/corporate/pressrelease/ge‐digital‐042116.html
Technical Report
Networked Intelligence, 2016
Advanced Analytics Platforms
Advanced analytics is the part of Technology Action dealing directly with putting intelligence into systems.
Given high quality input, model building, execution, optimization and postprocessing are the major steps in
analytics. The major drivers are the need for increasing the speed of learning39 and the nature of diminishing
returns in model improvements40. All analytical methods have limits to their precision and accuracy.
The natural starting point for inding solutions to new situations is to use existing models that have solved
similar situations before. Deep learning frameworks like Caffe provide such facilities. For more complex
applications, training new models is possible via software libraries and services like deep learning focused
Theano and DL4J, that simplify the creation and execution of distributed models. Advanced libraries like
TensorFlow enable ef icient design of deep learning model architectures. Efforts similar to Keras, modularize
and integrate these lower‐level libraries into easily usable solutions. Machine learning libraries such as
Photon ML build upon existing data platforms. Investments by companies into products like Nervana Systems
Neon and Preferred Networks Chainer focus on optimizing the use of GPUs. Widely used community efforts
like scikit‐learn serve as the growing space for new ideas. Foundational building blocks are delivered by
company products like Nvidia Deep Learning and Microsoft DMTK. Advanced Analytics technologies are
mostly based on Java, Scala, Python, Lua and C++. Experimental and small scale model building frequently
happens in R.
When performing large scale model building, the move to machine learning platforms makes sense. There are
many companies pushing the performance and capabilities of machine learning. ML platforms are developed
by Continuum Analytics, H2O, Dato, Skytree and others. There are specialized companies like manufacturing
and Industrial IoT focused Uptake and others.
From a data science process perspective, generic solution companies like DataRobot are streamlining the
entire process of model building, selection and deployment for organizations doing large scale model building.
Some technologies and companies are specializing further in a particular step in model building like model
optimization by Sigopt and deployment by Yhat.
From a different angle, Wolfram is pushing mathematics‐based computational knowledge, and is gaining
traction in R&D intensive industries like biotechnology. Interesting IoT venues are opened by cloud HPC
simulation platform Rescale that provides scalable model execution for variety of models like those by Ansys.
From knowledge creation, an interesting approach is taken by SparkBeyond that is automating research as
well as the search for complex patterns in data. Relevant to knowledge analytics, semantic analytics
technologies are developed by companies like Cycorp. Topological analysis for knowledge discovery is
developed by Ayasdi. The frontier of AGI is developed by companies like Vicarious and stealth startups, with a
focus on applications in analyzing images and video.
As the saying goes, build what you must, buy what you can.
39
G‐B. Huang, Z. Bai, L. Lekamalage, C. Kasun, C. M. Vong, Local Receptive Fields Based Extreme Learning Machine, IEEE Computational
Intelligence Magazine, 2015
40
D. Ferrucci, E. Brown, J. Chu‐Carroll, J. Fan, D. Gondek, A. Kalyanpur, A. Lally, J. Murdock, E. Nyberg, J. Prager, N. Schlaefer, C. Welty,
Building Watson: An Overview of the DeepQA Project, AI Magazine, 2010
Technical Report
Networked Intelligence, 2016
Data Speci ic Analytics
Language, audio and vision are the most important means of communication and perceiving the world for
humans. For machines to be able to interact seamlessly with people, the frontier of enabling technologies is
developed by IBM Watson, Google and Microsoft, among others.
Besides bigger players, there are younger companies specializing in speci ic types of data analysis. For
example Dextro is focused on making videos discoverable, searchable and curatable. Vision.ai is pushing
vision services to the cloud. In satellite imagery, Planet Labs and others are focused on analysing the very
large datasets. Clarifai Vision uses deep learning for detecting duplicates and doing visual searches. Similarly,
in audio analysis there are companies like Nuance Communications providing speech and imaging
applications for businesses.
Given the very large knowledge base that the written Internet represents, there are a lot of commercial
activities to extract knowledge from it, see for example import.io. For building your own, high performing text
parsers like SyntaxNet and industrial spaCy tokenizer provide the basis for ef icient natural language
processing (NLP) in English language. State of the art, research focused, NLP libraries like Carmel, NLTK, Gate
provide advanced features like semantic reasoning and information extraction. For automated statistical
machine translation, academic software like Moses exist. Applications of deep convolutional neural networks
are the most successful in dealing with perceptual data41.
Overall, the ield of Data Speci ic Analytics is advancing fast and considerable progress has been made in
making language, audio and vision understandable to machines.
Usage Layer
Technologies deliver value through Applications. Here are some examples of how organizations are applying
technologies in the data and IoT trend and how people can interact with the technologies.
Applications
I explicitly list existing applications rather than communicate speculative promises. Refer to angel and venture
capital investments42 and venues like KDD to see where the market is heading:
● A frequent and most basic generic application is event and outlier detection.
● In the automotive industry, route optimization based on traf ic, as demonstrated by Google Waze and
Zubie, is relevant for transportation and logistics, like Uber and Didi Chuxing.
● In synthetic biology and organism engineering, Zymergen is demonstrating the use of automation,
data architecture and machine learning.
● In investment inance, Alphasense and Bloomberg are leading data application frontier.
● In retail inance, Lendup, Af irm, Mirador Lending, Inventure and Earnest are applying data for better
servicing their customers.
● Energy intelligence is developed by EnerNOC.
● In internal intelligence, Palantir is leading the way.
● In market intelligence look for Funderbeam, Quid, Mattermark and CB Insights.
● Bayes Impact is applying data science for non‐pro it initiatives.
41
42
http://www.computervisionblog.com/2016/06/deep‐learning‐trends‐iclr‐2016.html
For example Data Collective, Founders Fund, DFJ, Khosla Ventures, Andreessen Horowitz, Y Combinator.
Technical Report
●
●
●
●
●
●
●
●
Networked Intelligence, 2016
In personal health, companies like Jawbone, Mis it and Garmin help people track their itness goals.
In agriculture, the Climate Corporation has found a way to protect farmers against the climate change.
John Deere is providing APIs for precision agriculture.
In enterprise sales, Salesforce is leading the way with acquisitions like MetaMind, Implisit, MinHash
and RelateIQ as well as pipeline management by InsideSales and Pipedrive. Gainsight is pushing with
customer success software.
In enterprise customer service, wise.io and NICE Systems lead to data driven decisions.
In enterprise security and fraud detection, Feedzai and Sift Science are the frontier.
In cloud infrastructure monitoring, Datadog.
New generation of company wide business intelligence by companies like Looker and Domo.
In adtech, Rocketfuel. In enterprise legal, Ravel.
There are valuable uses of data in many other areas demonstrated by many companies.
Human Interfaces
Intelligent systems must have processes of communication and control that interface with people43. The
report’s focus is on user endpoints, rather than on data visualization and communication44. Machines have to
adapt to the needs of people, so that they can interact naturally. User and customer experience design has
become an integral part of developing applications. The most recent transformative technologies in user
interfaces are virtual reality and augmented reality technologies, such as developed by HTC Vive, Magic Leap,
Facebook Oculus and Meta.
In mobile technology, user interfaces like Facebook M and wit.ai, Microsoft Cortana, Google Voice Recognition,
Apple Siri, Turing Robot and api.ai make interaction with devices more intuitive. Automated assistants for
interaction with corporate processes, systems and people, are reducing the burden of high volume
communication brought by the email, chat and mobile apps era. By 2020 it is estimated that at least 50% of all
searches are going to be either through images or speech45. Some of the companies to follow are x.ai and
ClaraBridge. Assistants in combination with solutions like IFTTT can automate and connect Internet services
so that they work together for the bene it of people and make their lives feel more natural. A system moving in
this direction is Viv, based on a dynamically evolving cognitive architecture extending intent of the user.
Limitations
There are many detailed, technical limitations and API designs of existing technologies that must be taken into
account in systems architecture. Here are the upper bounds for some of the factors:
● Financial cost and latency numbers46 of storage, network and processing services.
● Geographical locations of data centers and CDNs. Physical iber networks layout, with bandwidths
realistically in the range of 10GBps on planetary scale. Coverage and quality of mobile networks.
These mainly follow the trends of population and the Internet adoption.
● Linear scalability and high availability of data‐center scale OS proven on 10,000s of nodes.
43
T. Mikolov, A. Joulin, M. Baroni, A Roadmap towards Machine Intelligence, arXiv, 2015
C. Viau, Try Datavis Now, Github Books (in preparation), 2016
45
http://www.fastcompany.com/3035721/
46
http://people.eecs.berkeley.edu/~rcs/research/interactive_latency.html
44
Technical Report
●
●
●
●
●
Networked Intelligence, 2016
Automatic scalability of Data & IoT platforms to billions of devices47, proven scalability to trillions of
messages per day48.
Proven scalability of storage services to a few exabytes49. Some organizations think in zettabytes and
above. For benchmarks on data platforms see TPC50 related materials.
Limitations of Data Exchanges focused on a few industries like inancial and geospatial.
Resource intensity of machine learning training processes addressed with GPU and FPGA.
Lack of good solutions for highly multidimensional data visualization and communication.
How the Pieces Fit Together
Illustration with a possible top down structure of a very large scale, autonomous cyber physical system concept.
It is just one way of clustering and labeling a network of processes that form such a system.
To bring different research and commercial results together I illustrate a hierarchical system of systems
concept where different independent subsystems deal with different levels of abstractions. In designing the
concept the approach of systems thinking is used. Systems thinking includes the idea of layered structure.
Architecture is an abstract description of the entities of a system and the relationships between those entities
51
. The allocation of physical or informational elements of function into elements of form is one of the earliest
and most important decisions. This vision is not intended as a technical blueprint52, it is an illustration.
The actual systems architecture would bene it from principles like reactive systems53, microservices54 and
patterns like the blackboard architecture to achieve non‐functional requirements and provide programmable
APIs to subsystems that can be used independently and for system management.
Figure 3. A concept architecture of an intelligent system. The colors represent components with varying degree of
maturity based on Technology Action: blue stands for mature, grey for emerging and white for nascent
commercial grade technologies.
47
https://aws.amazon.com/iot/how‐it‐works/
http://www.con luent.io/blog/apache‐ka ka‐hits‐1.1‐trillion‐messages‐per‐day‐joins‐the‐4‐comma‐club
49
http://www.slideshare.net/AmazonWebServices/stg306‐efs‐how‐to‐store‐8‐exabytes‐look‐good‐doing‐it
50
http://www.tpc.org/
51
E. Crawley, ESD.34 Systems Architecting — Lecture Notes. MIT Engineering Systems Division, IAP, 2007
48
52
J. Spolsky, Don't Let Architecture Astronauts Scare You, Joel on Software, 2001
http://www.reactivemanifesto.org/
54
http://microservices.io/
53
Technical Report
Networked Intelligence, 2016
The structure is based on a direct mapping from a more detailed Research Progress and Technology Actions
analysis. Systems can be built on top of platforms such as those in the Infrastructure Layer and Processing
Layer sections. There can be multiple instantiations of subsystems or the entire system.
The lowest abstraction level of the system is in Preprocessing Data low, that has to cope with diverse input
data from other systems and its own subsystem processes. With well‐prepared data, high quality features and
input information, the next level of abstractions is addressed by modeling and advanced analytics focused
Core Data low. These subsystems form the basis for intelligent systems. The system has to have a connection
to the environment through well‐de ined Output Interfaces and Core Human Interfaces. In large systems,
explicit attention to system management is necessary to help designers, developers and operators decide how
to impact the overall system to evolve in a certain direction. It makes sense to observe and analyze its
behaviour and interactions with explicit threat intelligence and modeling functions.
Preprocessing Data low
This part of the illustration deals with lower abstraction levels of data, information and knowledge produced
by both external and internal systems.
Figure 4. Illustration of the subsystems dealing with low‐abstraction levels of data, information and knowledge. *
stands for a mature feature, / stands for emerging feature
Input Data, Info Merger and Knowledge Manager
To observe context and environments, the system has to have well‐de ined input APIs. Preprocessing is one of
the most human resource intensive activities, so aiming for high levels of automation in this part of the system
has a high return of investment.
To accelerate the progress further, the preprocessing can build upon Data Exchanges available in speci ic
industries with high quality, preprocessed data streams and data batches. Given that the incoming information
may be in a format and structure not directly usable by the system, decoding of information is necessary. As
Technical Report
Networked Intelligence, 2016
the input data is growing exponentially, the system needs measures to manage the growth rate of information.
For achieving this, compression, archival and in some cases iltering and aggregation may be used.
The types of information relevant to adaptive systems vary a lot in the Input Data. To understand language,
the system should be able to adapt to changes in the languages55. For effective processing of images,
segmentation, captioning and classi ication is necessary. For processing video, video analytics conducting
captioning and extraction of timed events from the content, is necessary. Processing audio with speech
recognition can be transformed into textual representation that can be fed into next processing steps.
Furthermore, both internal and external model output can provide high‐dimensional input that should be
classi ied and clustered. Vectors and multidimensional arrays are the output of this initial preprocessing step.
These can be fused together based on structure, semantics and meta information, such as location, time,
source, channel and destination. As shown by Data Exchanges. As a result, the Core Data low receives a set of
uni ied data streams and batches for conducting advanced analytics.
For the system’s ability to remember history, ef icient methods of storing both structured and unstructured
data, information and knowledge are necessary. The data must be indexed according to the expected queries.
Ef iciency in indexing is mandatory as it can be larger than the data itself. Depending on the nature of the data,
there could be means for probabilistic retrieval.
Data Curator, Data Preparer & Explorer
For ensuring the quality of input data and enabling knowledge discovery, further subsystems are required in
the preprocessing part of very large scale intelligent systems.
Namely, a Data Curator has to ind data through involving experts and browsing Data Exchanges. The
functions for managing the data, based on its use frequency, must be implemented so that the contents in the
Knowledge Manager can be grouped into hot, warm and cold data, all using different technologies with
different retrieval latencies. The overall quality management function in the Data Curator should offer means
for assigning a quality rank to the data as well as operate automated quality engineering processes on the
data.
Input Data and the contents of the Knowledge Manager can rarely be used directly by the processes
conducting advanced analytics. A Data Preparer takes care of processes that clean, normalize and do overall
data enrichment before the modeling step. The functions have to deal with errors and their distributions,
interpreting which requires a high degree of domain knowledge possessed by dedicated people.
As the content is high‐dimensional, guidance for machines is needed to indicate what are the best features for
solving a situation at hand. Feature manufacturing is a key process56. The Data Explorer targets automated
feature extraction, feature relationship identi ication and dimensionality reduction processes to narrow down
the sets of possible input variables for models. Given the state of technologies and the need for domain
knowledge, human involvement is necessary. They can utilize new methods like topological data analysis to
discover the most important relationships in the data for the models.
55
http://blog.stephenwolfram.com/2015/11/how‐should‐we‐talk‐to‐ais/
56
http://www.kdnuggets.com/2015/12/harasymiv‐lessons‐kaggle‐machine‐learning.html
Technical Report
Networked Intelligence, 2016
Core Data low
This part of the illustration is responsible for taking high quality inputs from the Preprocessing Data low and
the Knowledge Manager and turning those into metrics, estimates, predictions, simulations and projections.
The core parts of the subsystem are the Model Engine, Interpreter and Self‐improvement subsystems. The
Model Engine subsystems focus on providing a wide range of useful models with effective training, execution
and combining. The Interpreter takes the output of the models and processes the results along with models,
so that these can be effectively shared with other systems and people. The Self‐improvement is about
automating the processes of data science related to models.
Figure 5. Illustration of the subsystem functions in the Core Data low responsible for the base intelligence
Base Library, Executor and Selector
The Base Library consists of models as well as metadata pointers to other intelligent system APIs. The goal of
the models is to ind patterns, simulate speci ic scenarios for speci ic conditions, make predictions, create
insights and enable decision‐making. The machine learning methods and models in the Base Library can be
supervised, partially‐supervised and unsupervised, which means that the Self‐improvement subsystem has to
have means to produce and manage content in the library. Some models, especially those related to physical
objects, can be models based on systems of differential equations. These require different execution mode
compared to ML.
The Executor schedules and allocates resources for running selected models from a prioritized processing
backlog. The distribution of the execution of training, simulation, operational and online learning models is a
must‐have. The Executor may break the processes into intermediate steps that may produce intermediate
datasets. These can be reused for speeding up the use of models in recurring situations. Looking for patterns
Technical Report
Networked Intelligence, 2016
in the model execution steps will improve the contents of the Base Library with new models that deal with
different levels of abstractions.
Successful use of models is usually a combination of methods and algorithms. The Selector builds the
knowledge base of model performance on speci ic input combinations and uses the content for facilitating
decisions when a hybrid intelligence or an ensemble model should be used. The content is managed by the
Self‐improvement subsystem. The Selector also holds information about the empirical resource use of models
to build a library of heuristics that can serve as input for model selection.
Self‐improvement and Interpreter
Building applicable models is a process that involves people with data science and domain expertise. The
processes have to be automated to a great extent to enable fast adaption to new situations and changing
conditions. Model performance improvements demonstrate diminishing returns over time, thus it is
necessary to be able to monitor these changes to make decisions about the maturity and applicability of
solutions to different situations.
Similar to Data Curation, there should be means for model curation, as some types of methods like online
learning can produce hundreds of thousands of models. Situations that are mission critical or frequently
encountered by a cyber physical system, must be addressed quickly, while non‐mission‐critical situations can
be addressed with a certain latency. This means that models can also be classi ied as hot, warm and cold in the
libraries. This information can be used by the Executor in allocating resources and doing model deployments
to runtime environment.
The output of the models, as well as the way a model produced the output, are important bits of information
for sound decision‐making and collaboration. For evaluating different forms of bias and uncertainties, the
Interpreter collects information from the Model Engine and monitors for important occurrences of events.
Since the model output, model description and metrics can be very large, this data needs to be compressed
and encoded for ef icient communication. This is the most speculative piece of the illustration.
Output Interfaces and Core Human Interfaces
The events, decisions and insights are communicated through Output APIs. De ining these early in the
development of a system is a must‐have activity. It guides the design and provides means to de ine baseline
subsystem implementation that can be improved over time. As with any interface that is used in real
Applications, in‐depth versioning and planning is required.
The Core Human Interfaces are the most important control measures next to the system management.
De ining these early in the system architecture is important for success. These are also measures for safety of
the entire system. For further measures of safety, there should be a chain of command built throughout the
subsystems impacting system actions directly. This chain can introduce interruptibility57 into the system.
Given the complexity of the interfaces to an Intelligent System, two main functions are proposed. One is
visualization that translates data into understandable form for human decision making. The other is
interactive bots with which text, audio, images and video can be exchanged. The communication with
57
http://intelligence.org/ iles/Interruptibility.pdf
Technical Report
Networked Intelligence, 2016
interactive bots is a high priority input channel for the Preprocessing Data low, introducing a direct feedback
loop into the system for training models.
Notes on the Architecture
Architectural principles for systems of systems have been long explored in both industry and academia. In
short, the elements must be able to usefully operate independently, have managerial independence and
support evolutionary development. In communication, the interfaces have to be standardized between
different layers, which allows for dedicated developments in each system. For successful software
modularisation and architecture, one should assume most components will be replaced during the lifecycle of
the software.
Unable to centrally analyse the data in its entirety, a system of systems remains distributed, with speci ic
groups addressing speci ic aspects of the global networked system. Data protection gap needs to be explicitly
addressed in the systems architecture58. Continuous deployment, testing and monitoring enable development
speed and meeting non‐functional requirements.
The traditional 7 layers of OSI model are insuf icient to describe the richness of new systems. Examples are
reactive systems based on the concept of Actors. Actors represent a system of completely decoupled
components that interact only using (a)synchronous messages. While it can be argued that it is a part of the
“applications layer”, it unnecessarily bloats the layer. For future systems, such communication could be better
represented by a new layer not yet present; call it the “systems layer” if you will. This layer would present
systems for dynamic analysis of interaction networks to infer properties of the whole system.
Intelligent System Initiatives
The realization of the illustrated concept would be a bottom‐up endeavour by numerous dedicated parties59.
Minimum Viable Products would be in order. Independent parties develop technologies, products and
research according to their own agenda. In short, there are three types of actions relevant to such systems:
●
●
●
58
59
60
Incremental improvement with immediate commercial impact:
● Preprocessing Data low: Info Merger, Knowledge Manager
● Core Data low: Executor
Innovation and applied research with impact in 3‐5 years60:
● Preprocessing Data low: Input Data, Data Curator and Preparer & Explorer
● Core Data low: Selector, Base Library and Self‐improvement
● Output Interfaces, Core Human Interfaces
Fundamental research with tangible commercial impact in 10+ years61:
● Core Data low: Interpreter
http://www.gigya.com/resource/whitepaper/the‐2015‐state‐of‐consumer‐privacy‐personalization/
S. Berkun, The Myths of Innovation, O'Reilly, 2010
Based on a list of public investment sizes and rounds in companies related to the ield
The timeframe estimate is based on the observation that the research is emerging and ACM SigSoft Impact Project conclusion that
software engineering breakthroughs take 10‐15 years to reach wide scale practice.
61
Technical Report
Networked Intelligence, 2016
Leading organizations who are pushing the boundaries of intelligent system developments are listed below.
Smaller companies tend to keep a lower pro ile and are not listed. Keeping an eye on the publications, public
releases and acquisitions by these companies can serve as an indicator of directions and progress in the ield:
●
●
●
●
●
●
●
●
●
Google
Microsoft
Apple
Facebook
AWS
IBM
Baidu
IARPA
OpenAI
It is important to release software early in order to understand the dynamics and to ind improvement areas
in design and implementation. Iterations and increments are essential in software engineering. Building real
cyber physical systems architecture has to be based on actual requirements.
Conclusion
The bene its of intelligent systems are demonstrated by many ANI applications used by businesses,
governments and society. Research relevant to the ield offers good basis for tackling complex scenarios and
advancing the ield. The core infrastructure like data & IoT platforms and advanced analytics platforms are
maturing with potential high‐value commercial applications. Relevant to the intelligent systems, the highest
maturity is in data gathering and preprocessing, followed by advanced analytics.
The hierarchical system of systems concept, consisting of parts related to data preprocessing, models and
self‐improvement, shows how the different pieces may it together. This vision is not intended as a technical
blueprint. I estimate that the concept illustrated here is achievable within a 5‐10 year timeframe.
Bringing together research and technology developments is necessary for engineering intelligent systems.
Hopefully this writing sparks developments and conversations that speed up the progress.
Technical Report
Networked Intelligence, 2016
Appendix
Background on the Infrastructure Layer that impacts building of cyber physical systems.
Building Blocks
The base components that drive advancements of systems are processors, memories, drives and chargers.
These de ine the basis for non‐functional requirements, especially performance.
Processors by companies like Intel62, Qualcomm63 and others are targeting the new IoT and data era. While
the era of Moore’s Law is coming to an end64, new types of processors like Tensor Processing Units65 and
increasingly ield‐programmable gate arrays (FPGA) and potentially HP Nanostores66 and adiabatic quantum
optimization processors by DWave67 68, are pushing the boundaries of what is possible with data focused
hardware. Visual computing technologies by companies like Nvidia offer step jump increases in performance
of ML, as GPUs increase the bandwidth, reduce latency and communication costs with built in support for
matrix algebra69.
Mature drive technologies like Solid State Drives productized by companies like Nimble Storage, offer
signi icant read performance improvements compared to traditional hard disk drives.
Wireless charging using new technologies by companies like uBeam and WiTricity offer new means to charge
independent devices without the hassle of cables. This unlocks the potential of smart, autonomous devices in
many applications.
Devices and Machines
Devices are becoming open and extensible platforms that are capable of complex tasks independently from
central control, the so‐called Fog70 and Mist Computing71 that complement Cloud Computing. This creates new
means for cyber physical systems to interact with the world.
Incumbent electronics, appliances and mobile device manufacturers like LG are continuously innovating in
new types of consumer devices. At the same time new Industrial Internet machines and software are
developed by giants like Siemens. Companies like NXP Semiconductors and Psikick are innovating in the
foundations of circuit design suitable for IoT era. Very small electronic devices (MEMS) by companies such as
mCube offer new means to interact and control devices. At the same time prototyping platforms like Arduino,
Raspberry Pi and Xilinx reduce the speed and cost of innovation in creation of devices. New applications can
be developed quickly.
62
Intel Xeon Processor E7 Family for data analytics
Qualcomm Snapdragon for mobile devices
64
https://www.technologyreview.com/s/601102/intel‐puts‐the‐brakes‐on‐moores‐law/
65
N. Jouppi, Google supercharges machine learning tasks with TPU custom chip, Google Cloud Platform Blog, 2016
66
Ranganathan, From Microprocessors to Nanostores: Rethinking Data‐Centric Systems, IEEE Computer, 2011
67
D‐Wave Quantum Hardware
68
https://plus.google.com/+QuantumAILab/posts/DymNo8DzAYi
69
http://www.nvidia.com/object/machine‐learning.html
70
http://www.cisco.com/c/en/us/solutions/internet‐of‐things/iot‐fog‐computing.html
71
http://www.thinnect.com/mist‐computing/
63
Technical Report
Networked Intelligence, 2016
Overall commoditization and maturity of electronic components has led to major innovations of devices
operating on land, air, sea and space. For example on land, products like Google Self Driving Car and Tesla are
creating autonomous cars. Furthermore, incumbent companies like Ford, are turning their products into open
platforms. In air, companies like DJI are producing commercial and recreational unmanned aerial systems.
These devices are getting information platforms for developing and operating them, for example Airware and
Dreamhammer. On the sea, companies like Liquid Robotics are developing autonomous ocean robots capable
of covering vast distances. In space, new entrants like SpaceX and BlueOrigin are creating rockets for orbital
light. At the same time companies like Spire and Planet Labs are driving bringing down the costs of space
with nanosatellites for applications like satellite imagery.
A similar transformation is happening in home automation, where home hubs like Google Nest and Amazon
Echo are changing how we connect devices like thermostats and appliances at home. Companies like Tado and
View Dynamic Glass are changing the historically plain devices, such as air conditioners and windows, into
smart devices. Home robotics like social robots by Jibo and cleaning robots by iRobot are changing household
chores and social interaction. Incumbent companies like Samsung, Honeywell and Philips are also investing
heavily into home automation.
Industrial automation is evolving at a similar scale and pace. Amazon Robotics, Starship and other companies
are creating mobile robotic ful illment systems. Industrial Internet machine producers like Bosch and ABB are
producing heavy‐industry electronics and machinery. Younger companies like Rethink Robotics and Kuka
Robotics are creating industrial robots and factory automation solutions so humans can focus on other tasks.
Increasingly more devices are attached to people with the wearables trend. Devices like Apple Watch (3
GFlops) and Pebble bring new human interfaces and ease of use to the people. Big companies like Japanese
Rakuten and South Korean Samsung, and smaller ones like HumanCondition and Ringly are turning clothes,
jewelry and accessories into smart endpoints.
Overall, these trends are a big enabler for builders of cyber physical systems. Innovative devices and machines
enable rapid adaptation not only in software, but also on hardware side. Moravec’s paradox72 seems to be
becoming less of a challenge thanks to working real‐life applications.
Operating Systems
For dealing with the diversity of devices, there is also diversity in operating systems (OS). Universal OS like
Windows, Ubuntu, Debian and FreeBSD will continue to unify this layer of technologies, but the more
specialized operating systems like OS X, Android, iOS, Micrium and TinyOS will have a continued role to play.
This means that at least when developing applications, multiple platform support requires dedicated
attention of software engineers. For universal OS, technologies like Docker simplify deployment.
When planning systems of systems, combining computations into a dynamic network of data processing
quickly becomes a complex task of managing clusters of devices. Thus, new ways of expressing computation
are developed by abstracting the device layer into programmable concepts, which can be combined into
computational process lows. An example is the data center OS Mesosphere and Stratoscale. In designing very
large scale systems, it makes sense to utilize this abstraction level to increase the focus on the core
functionality.
72
V. Rotenberg, Moravec's Paradox: Consideration in the context of two brain hemisphere functions, Activitas Nervosa Superior, 2013
Technical Report
Networked Intelligence, 2016
Protocols
Different intents of communicating systems call for different protocols. This is clearly observable in the
evolution of protocols in the IoT era. Most common ixed and Wireless protocols like WiFi, Bluetooth, RFID
and LTE are seeing numerous additions of new industry speci ic protocols like Zigbee, LoRa, MQTT, NFC,
6LoWPAN, DDS, LWM2M, and application speci ic protocols like BitX and M‐Bus. The important takeaway is
that these protocols address speci ic needs like speed, power consumption, resource constraint, reliability,
machine‐to‐machine communication and other aspects relevant to building cyber physical systems. Products
like Eero, simplify the complexity for dedicated needs, like use at homes. The same can be expected for of ices
and factories.
In designing the input and output interfaces of systems we must ind ways to deal with the diversity of
protocols. The Data & IoT Platforms, as well as Data Exchange address this.
Networks and Connectivity
Networks form an integral part of systems as they enable communication. The infrastructure is advancing
towards open, programmable environments, capable of operating at 1GBit/s with ms latencies. Companies
like Cisco, Thinnect and others continue to innovate in this arena with concepts like software de ined
networks and network as a service.
With the increase of machine‐to machine communication, companies like Texas Instruments, Atmel and
others, traditionally not focused on networks, are pushing the solution space of M2M communications with
components. Younger companies like goTenna and others design and market wireless connectivity products
for decentralized communications. There are also dedicated connectivity companies like Sigfox, Sierra and
others pulling together different technologies to target speci ic needs of IoT, for example city‐scale across the
globe.
Large scale networking equipment and solutions are operated by telecom operators who share the strategic
vision of IoT and data intensity. Thus we see companies like T‐Mobile and Huawei focus on data centers. The
bigger players like Verizon, AT&T, China Mobile, SoftBank, Bharti Airtel, Orange Telecom, Telefonica, Vodafone
and others are all investing into their capabilities to meet the future needs of IoT and data era, where the
number of devices communicating over the network is very large. The largest underdeveloped populated area
of the planet is currently Africa, but with leadership by companies like Millicom and MTN Group, the network
quality will increase in time. In more remote and less populated areas there is still a long way to go, but there
are initiatives targeting this need as well73.
Networked systems pose several challenges, such as variable time delays, failures, recon igurations, not only
at the lower layers of the OSI model, but at the application layer. Given that telecom operators have moved into
providing applications in multimedia, we will see continued improvement in guarantees provided by the
infrastructure, developed in partnership with content delivery networks (CDN) like Akamai and cloud service
providers. Thus it is assumed, that in designing cyber physical systems, the designers and developers have to
mainly focus on the quality of the edge communications required by the new protocols and devices.
73
https://info.internet.org/
Technical Report
Networked Intelligence, 2016
Cloud Platforms
Cloud is the most mature concept compared to Fog and Mist Computing. Networked computation is being
pushed to the edges of networks. This creates a dynamically changing infrastructure environment where
functionality is continuously moving between the edges and centralized systems. Cloud services offer ways to
meet many of the non‐functional requirements, such as security and scalability. Initiatives like OWASP de ine
requirements and ways for systems to advance from the base offer to more secure solutions. Companies like
Ionic Security offer data protection & control platforms.
The main service providers are Google Cloud, Microsoft Azure and Amazon Web Services (AWS) who all target
data and machine learning74 75 76 related functionalities at planetary scale. Other providers like Salesforce
Heroku can be used for fast innovation in specialized applications.
Active Companies
A list of companies mentioned in the report. There are many more pushing the boundaries of possibilities.
Category
Mentioned Companies
Infrastructure Layer
Building Blocks
Intel, Qualcomm, Google, HP, DWave, Nvidia, Nimble Storage, uBeam,
WiTricity
Devices and Machines
LG, Siemens, NXP Semiconductors, Psikick, mCube, Arduino, Xilinx,
Google, Tesla, Ford, DJI, Airware, Dreamhammer, Liquid Robotics,
SpaceX, BlueOrigin, Spire, Planet Labs, Google, Amazon, Tado, View,
iRobot, Samsung, Honeywell, Philips, Bosch, ABB, Rethink Robotics,
Kuka Robotics, Apple, Pebble, Rakuten, HumanCondition, Ringly
Operating Systems
Microsoft, Apple, Canonical, Micrium, Docker, Mesosphere, Stratoscale
Networks and Connectivity
Cisco, Thinnect, Texas Instruments, Atmel, goTenna, Sigfox, Sierra,
T‐Mobile, Huawei, Verizon, AT&T, China Mobile, SoftBank, Bharti Airtel,
Orange Telecom, Telefonica, Vodafone, Millicom, MTN Group,
Internet.org, Akamai
Cloud Platforms
Google Cloud, Microsoft Azure, Amazon Web Services, Salesforce
Heroku
Processing Layer
Data & IoT Platforms
74
Cloudera, MapR, HortonWorks, LinkedIn, Twitter, DataBricks, Data
Artisans, HP, Kx, Paradigm4, Facebook, Cisco Jasper, PTC Thingworx,
GE, Pivotal, Helium, Hologram, Planet OS
https://cloud.google.com/products/machine‐learning/
https://aws.amazon.com/machine‐learning/
76
https://azure.microsoft.com/en‐us/services/machine‐learning/
75
Technical Report
Networked Intelligence, 2016
Data Preprocessing and
Exchanges
Tamr, Trifacta, CrowdFlower, snapLogic, Xignite, Planet OS, Enigma,
Qlik, Microsoft
Advanced Analytics Platforms
Google, Nervana Systems, Preferred Networks, Nvidia, Microsoft,
Continuum Analytics, H2O, Dato, Skytree, Uptake, DataRobot, Yhat,
Sigopt, Wolfram, Rescale, SparkBeyond, Cycorp, Ayasdi, Vicarious
Data Speci ic Analytics
IBM Watson, Google, Microsoft, Dextro, Vision.ai, Planet Labs, Clarifai
Vision, Nuance Communications, import.io
Usage Layer
Applications
Google, Zymergen, Alphasense, Bloomberg, Lendup, Af irm, Mirador
Lending, Inventure, Earnest, EnerNOC, Palantir, Funderbeam, Quid,
Mattermark, CB Insights, Bayes Impact, Jawbone, Mis it, Garmin,
Climate Corp, John Deere, Salesforce, InsideSales, Pipedrive, Gainsight,
wise.io, NICE Systems, Feedzai, Sift Science, Datadog, Looker, Domo,
Rocketfuel, Ravel
Human Interfaces
HTC, Magic Leap, Facebook Oculus, Meta, wit.ai, Microsoft, Apple,
Turing Robot, api.ai, x.ai, ClaraBridge, IFTTT, Viv
| 3cs.SY
|
arXiv:1711.10331v1 [cs.LG] 25 Nov 2017
Complex Structure Leads to Overfitting:
A Structure Regularization Decoding Method for Natural
Language Processing
Xu Suna , Weiwei Suna,b , Shuming Maa , Xuancheng Rena , Yi Zhanga , Wenjie Lic ,
Houfeng Wanga
a MOE
Key Laboratory of Computational Linguistics, Peking University
of Computer Science and Technology, Peking University
c Department of Computing, The Hong Kong Polytechnic University
b Institute
Abstract
Recent systems on structured prediction focus on increasing the level of structural dependencies within the model. However, our study suggests that complex structures
entail high overfitting risks. To control the structure-based overfitting, we propose to
conduct structure regularization decoding (SR decoding). The decoding of the complex structure model is regularized by the additionally trained simple structure model.
We theoretically analyze the quantitative relations between the structural complexity
and the overfitting risk. The analysis shows that complex structure models are prone to
the structure-based overfitting. Empirical evaluations show that the proposed method
improves the performance of the complex structure models by reducing the structurebased overfitting. On the sequence labeling tasks, the proposed method substantially
improves the performance of the complex neural network models. The maximum F1
error rate reduction is 36.4% for the third-order model. The proposed method also
works for the parsing task. The maximum UAS improvement is 5.5% for the tri-sibling
model. The results are competitive with or better than the state-of-the-art results. 1
Keywords: Structural complexity regularization, Structured prediction, Overfitting
risk reduction, Linear structure, Tree structure, Deep Learning
1. Introduction
Structured prediction models are often used to solve the structure dependent problems in a wide range of application domains including natural language processing,
bioinformatics, speech recognition, and computer vision. To solve the structure dependent problems, many structured prediction methods have been developed. Among them
Email addresses: [email protected] (Xu Sun), [email protected] (Weiwei Sun),
[email protected] (Shuming Ma), [email protected] (Xuancheng Ren),
[email protected] (Yi Zhang), [email protected] (Wenjie Li), [email protected]
(Houfeng Wang)
1 This work is a substantial extension of a conference paper presented at NIPS 2014 [1].
Preprint submitted to Artificial Intelligence
November 29, 2017
the representative models are conditional random fields (CRFs), deep neural networks,
and structured perceptron models. In order to capture the structural information more
accurately, some recent studies emphasize on intensifying structural dependencies in
structured prediction by applying long range dependencies among tags, developing
long distance features or global features, and so on.
From the probabilistic perspective, complex structural dependencies may lead to
better modeling power. However, this is not the case for most of the structured prediction problems. It has been noticed that some recent work that tries to intensify the
structural dependencies does not really benefit as expected, especially for neural network models. For example, in sequence labeling tasks, a natural way to increase the
complexity of the structural dependencies is to make the model predict two or more
consecutive tags for a position. The new label for a word now becomes a concatenation of several consecutive tags. To correctly predict the new label, the model can be
forced to learn the complex structural dependencies involved in the transition of the
new label. Nonetheless, the experiments contradict the hypothesis. With the increasing
number of the tags to be predicted for a position, the performance of the model deteriorates. In the majority of the tasks we tested, the performance decreases substantially.
We show the results in Section 4.1.1.
We argue that over-emphasis on intensive structural dependencies could be misleading. Our study suggests that complex structures are actually harmful to model
accuracy. Indeed, while it is obvious that intensive structural dependencies can effectively incorporate the structural information, it is less obvious that intensive structural
dependencies have a drawback of increasing the generalization risk. Increasing the
generalization risk means that the trained models tend to overfit the training data. The
more complex the structures are, the more instable the training is. Thus, the training is
more likely to be affected by the noise in the data, which leads to overfitting. Formally,
our theoretical analysis reveals why and with what degree the structure complexity
lowers the generalization ability of the trained models. Since this type of overfitting
is caused by the structural complexity, it can hardly be solved by ordinary regularization methods, e.g., the weight regularization methods, such as L2 and L1 regularization
schemes, which are used only for controlling the weight complexity.
To deal with this problem, we propose a simple structural complexity regularization
solution based on structure regularization decoding. The proposed method trains both
the complex structure model and the simple structure model. In decoding, the simpler
structure model is used to regularize the complex structure model, deriving a model
with better generalization power.
We show both theoretically and empirically that the proposed method can reduce
the overfitting risk. In theory, the structural complexity has the effect of reducing the
empirical risk, but increasing the overfitting risk. By regularizing the complex structure
with the simple structure, a balance between the empirical risk and the overfitting risk
can be achieved. We apply the proposed method to multiple sequence labeling tasks,
and a parsing task. The formers involve linear-chain models, i.e., LSTM [2] models,
and the latter involves hierarchical models, i.e., structured perceptron [3] models. Experiments demonstrate that the proposed method can easily surpass the performance of
both the simple structure model and the complex structure model. Moreover, the results
are competitive with the state-of-the-art results or better than the state-of-the-arts.
2
To the best of our knowledge, this is the first theoretical effort on quantifying the
relation between the structural complexity and the generalization risk in structured prediction. This is also the first proposal on structural complexity regularization via regularizing the decoding of the complex structure model by the simple structure model.
The contributions of this work are two-fold:
• On the methodology side, we propose a general purpose structural complexity
regularization framework for structured prediction. We show both theoretically
and empirically that the proposed method can effectively reduce the overfitting
risk in structured prediction. The theory reveals the quantitative relation between
the structural complexity and the generalization risk. The theory shows that
the structure-based overfitting risk increases with the structural complexity. By
regularizing the structural complexity, the balance between the empirical risk
and the overfitting risk can be maintained. The proposed method regularizes the
decoding of the complex structure model by the simple structure model. Hence,
the structured-based overfitting can be alleviated.
• On the application side, we derive structure regularization decoding algorithms
for several important natural language processing tasks, including the sequence
labeling tasks, such as chunking and name entity recognition, and the parsing
task, i.e., joint empty category detection and dependency parsing. Experiments
demonstrate that our structure regularization decoding method can effectively
reduce the overfitting risk of the complex structure models. The performance
of the proposed method easily surpasses the performance of both the simple
structure model and the complex structure model. The results are competitive
with the state-of-the-arts or even better.
The structure of the paper is organized as the following. We first introduce the
proposed method in Section 2 (including its implementation on linear-chain models and
hierarchical models). Then, we give theoretical analysis of the problem in Section 3.
The experimental results are presented in Section 4. Finally, we summarize the related
work in Section 5, and draw our conclusions in Section 6.
2. Structure Regularization Decoding
Some recent work focuses on intensifying the structural dependencies. However,
the improvements fail to meet the expectations, and the results are even worse sometimes. Our theoretical study shows that the reason is that although the complex structure results in the low empirical risk, it causes the high structure-based overfitting risk.
The theoretical analysis is presented in Section 3.
According to the theoretical analysis, the key to reduce the overall overfitting risk is
to use a complexity-balanced structure. However, such kind of structure is hard to define in practice. Instead, we propose to conduct joint decoding of the complex structure
model and the simple structure model, which we call Structure Regularization Decoding (SR Decoding). In SR decoding, the simple structure model acts as a regularizer
that balances the structural complexity.
3
As the structures vary with the tasks, the implementation of structure regularization
decoding also varies. In the following, we first introduce the general framework, and
then show two specific algorithms for different structures. One is for the linear-chain
structure models on the sequence labeling tasks, and the other is for the hierarchical
structure models on the joint empty category detection and dependency parsing task.
2.1. General Structure Regularization Decoding
Before describing the method formally, we first define the input and the output of
the model. Suppose we are given a training example (xx, y ), where x ∈ Xn is the input
and y ∈ Yn is the output. Here, X stands for the feature space, and Y stands for the
tag space, which includes the possible substructures regarding to each input position.
In preprocessing, the raw text input is transformed into a sequence of corresponding
features, and the output structure is transformed into a sequence of input position related tags. The reason for such definition is that the output structure varies with tasks.
For simplicity and flexibility, we do not directly model the output structure. Instead,
we regard the output structure as a structural combination of the output tags of each
position, which is the input position related substructure.
Different modeling of the structured output, i.e., different output tag space, will lead
to different complexity of the model. For instance, in the sequence labeling tasks, Y
could be the space of unigram tags, and then y is a sequence of y i . Y could also be the
space of bigram tags. The bigram tag means the output regrading to the input feature
x i involves two consecutive tags, e.g. the tag at position i, and the tag at position i + 1.
Then, y is the combination of the bigram tags y i . This also requires proper handling
of the tags at overlapping positions. It is obvious that the structural complexity of the
bigram tag space is higher than that of the unigram tag space.
Point-wise classification is donated by g(xx, k, y), where g is the model that assigns
the scores to each possible output tag y at the position k. For simplicity, we denote
g(xx, k) = argmaxy∈Y g(xx, k, y), so that g : Xn × R 7→ Y. Given a point-wise cost
function c : Y × Y 7→ R, which scores the predicted output based on the gold standard
output, the model can be learned as:
g = argmin
g
|xx|
X X
c(g(xx, i), y i )
(xx,yy)∈S i=1
For the same structured prediction task, suppose we could learn two models g1 and
g2 of different structural complexity, where g1 : Xn × R 7→ Y is the simple structure
model, and g2 : Xn × R 7→ Z is the complex structure model. Given the corresponding point-wise cost function c1 and c2 , the proposed method first learns the models
separately:
g1 = argmin
g
g2 = argmin
g
|xx|
X X
c1 (g(xx, i), y i )
(xx,yy)∈S 1 i=1
|xx|
X X
(xx,zz)∈S 2 i=1
4
c2 (g(xx, i), z i )
Suppose there is a mapping T from Z to {y|y ∈ Y}, that is, the complex structure
can be decomposed into simple structures. In testing, prediction is done by structure
regularization decoding of the two models, based on the complex structure model:
z = argmax
z ∈Z|xx|
|xx|
X
(g2 (xx, i, z i ) +
X
g1 (xx, i, y))
(1)
y∈T (zzi )
i=1
In the decoding of the complex structure model, the simple structure model acts as a
regularizer that balances the structural complexity.
We try to keep the description of the method as straight-forward as possible without
loss of generality. However, we do make some assumptions. For example, the general
algorithm combines the scores of the complex model and the simple model at each
position by addition. However, the combination method should be adjusted to the task
and the simple structure model used. For example, if the model g is a probabilistic
model, multiplication should be used instead of the addition. Besides, the number of
the models is not limited in theory, as long as (1) is changed accordingly. Moreover,
if the joint training of the models is affordable, the models are not necessarily to be
trained independently. These examples are intended to demonstrate the flexibility of
the structure regularization decoding algorithms. The detailed implementation should
be considered with respect to the task and the model.
In what follows, we show how structure regularization decoding is implemented
on two typical structures in natural language processing, i.e. the sequence labeling
tasks, which involve linear-chain structures, and the dependency parsing task, which
involves hierarchical structures. We focus on the differences and the considerations
when deriving structure regularization decoding algorithms. It needs to be reminded
that the implementation of the structure regularization decoding method can be adapted
to more kinds of structures. The implementation is not limited to the structures or the
settings that we use.
2.2. SR Decoding for Sequence Labeling Tasks
We first describe the model, and then explain how the framework in Section 2.1 can
be implemented for the sequence labeling tasks.
Sequence labeling tasks involve linear-chain structures. For a sequence labeling
task, a reasonable model is to find a label sequence with the maximum probability conditioned on the sequence of observations, i.e. words. Given a sequence of observations
x = x1 , x2 , · · · , xT , and a sequence of labels, y = y1 , y2 , ..., yT , where T denotes the
sentence length, we want to estimate the joint probability of the labels conditioned on
the observations as follows:
p(yy|xx) = p(y1 , y2 , · · · , yT |x1 , x2 , · · · , xT )
If we model the preceding joint probability directly, the number of parameters that
need to be estimated is extremely large, which makes the problem intractable. Most
existing studies make Markov assumption to reduce the parameters. We also make
an order-n Markov assumption. Different from the typical existing work, we decompose the original joint probability into a few localized order-n joint probabilities. The
5
multiplication of these localized order-n joint probabilities is used to approximate the
original joint probability. Furthermore, we decompose each localized order-n joint
probability to the stacked probabilities from order-1 to order-n, such that we can efficiently combine the multi-order information.
By using different orders of the Markov assumptions, we can get different structural
complexity of the model. For example, if we use the Markov assumption of order-1,
we obtain the simplest model in terms of the structural complexity:
p(yy|xx) = p(y1 , y2 , · · · , yT |x1 , x2 , · · · , xT )
, p(y1 |xx)p(y2 |xx) · · · p(yT |xx)
=
T
Y
(p (yt |xx))
t=1
If we use the Markov assumption with order-2, we estimate the original joint probability as follows:
p(yy|xx) = p(y1 , y2 , · · · , yT |x1 , x2 , · · · , xT )
, p(y1 , y2 |xx)p(y2 , y3 |xx) · · · p(yT −1 , yT |xx)
=
T −1
Y
(p (yt , yt+1 |xx))
t=1
This formula models the bigram tags with respect to the input. The search space expands, and entails more complex structural dependencies. To learn such models, we
need to estimate the conditional probabilities. In order to make the problem tractable,
a feature mapping is often introduced to extract features from conditions, to avoid a
large parameter space.
In this paper, we use BLSTM to estimate the conditional probabilities, which has
the advantage that feature engineering is reduced to the minimum. Moreover, the conditional probabilities of higher order models can be converted to the joint probabilities
of the output labels conditioned on the input labels. When using neural networks,
learning of n-order models can be conducted by just extending the tag set from unigram labels to n-gram labels. In training, this only affects the computational cost of the
output layer, which is linear to the size of the tag set. The models can be trained very
efficiently with a affordable training cost. The strategy is showed in Figure 1.
To decode with structure regularization, we need to connect the models of different
complexity. Fortunately, the complex model can be decomposed into a simple model
and another complex model. Notice that:
p(yt , yt+1 |xx) = p(yt+1 |xx, yt )p(yt |xx)
6
𝑦
𝑦
SR decoding
decoding
BIO
B
BI
BLSTM
BLSTM
𝑥
𝑥
BIO
Figure 1: Illustration of SR-decoding on sequence labeling task (Left: third-order BLSTM. Right: thirdorder BLSTM with SR-decoding). BLSTM decodes with third-order tags directly, while SR-decoding jointly
decodes with first-order, second-order, and third-order tags.
The original joint probability can be rewritten as:
p(yy|xx) = p(y1 , y2 , · · · , yT |x1 , x2 , · · · , xT )
, (y1 , y2 |xx)p(y2 , y3 |xx) · · · p(yT −1 , yT |xx)
= p(y2 |xx, y1 )p(y1 |xx)p(y3 |xx, y2 )p(y2 |xx) · · · p(yT |xx, yT −1 )p(yT −1 |xx)
=
T −1
Y
(p (yt+1 |yt , x ) p (yt |xx))
t=1
In the preceding equation, the complex model is p(yt+1 |yt , x ). It predicts the next label based on the current label and the input. p(yt |xx) is the simplest model. In practice, we estimate p(yt+1 |yt , x ) also by BLSTMs, and the computation is the same with
p(yt , yt+1 |xx). The derivation can also be generalized to an order-n case, which consists
of the models predicting the length-1 to length-n label sequence for a position. Moreover, the equation explicitly shows a reasonable way of SR decoding, and how the
simple structure model can be used to regularize the complex structure model. Figure
1 illustrates the method.
However, considering the sequence of length T and an order-n model, decoding is
not scalable, as the computational complexity of the algorithm is O(nT ). To accelerate
SR decoding, we prune the tags of a position in the complex structure model by the topmost possible tags of the simplest structure model. For example, if the output tags of
the complex structure are bigram labels, i.e. Z = Y×Y, the available tags for a position
in the complex structure model are the combination of the most probable unigram tags
of the position, and the position before. In addition, the tag set of the complex model is
also pruned so that it only contains the tags appearing in the training set. The detailed
algorithm with pruning, which we name scalable multi-order decoding, is given in
Appendix B.
2.3. SR Decoding for Dependency Parsing
We first give a brief introduction to the task. Then, we introduce the model, and
finally we show the structure regularization decoding algorithm for this task.
7
root
But it ∅1 is n’t clear how long GM would be willing ∅2 to fight ... ∅3
Figure 2: An example of the dependency parsing analysis augmented with empty elements from PTB. The
dependency structure is according to Stanford Dependency. “∅” denotes an empty element. “∅1 ” indicates
an expletive construction; “∅2 ” indicates that the subject for fight, i.e. GM, is located in another place; “∅3 ”
indicates a wh-movement.
The task in question is the parsing task, specifically, joint empty category detection
and dependency parsing, which involves hierarchical structures. In many versions of
Transformational Generative Grammars, e.g., the Government and Binding [4] theory,
empty category is the key concept bridging S-Structure and D-Structure, due to its
possible contribution to trace movements. Following the linguistic insights, a traditional
dependency analysis can be augmented with empty elements, viz. covert elements [5].
Figure 2 shows an example of the dependency parsing analysis augmented with empty
elements. The new representations leverages hierarchical tree structures to encode not
only surface but also deep syntactic information. The goal of empty category detection
is to find out all empty elements, and the goal of dependency parsing thus includes
predicting not only the dependencies among normal words but also the dependencies
between a normal word and an empty element.
In this paper, we are concerned with how to employ the structural complexity regularization framework to improve the performance of empty category augmented dependency analysis, which is a complex structured prediction problem compared to the
regular dependency analysis.
A traditional dependency graph G = (V, A) is a directed graph, such that for sentence x = w1 , . . . , wn the following holds:
1. V = {0, 1, 2, . . . , n},
2. A ⊆ V × V.
The vertex set V consists of n + 1 nodes, each of which is represented by a single integer. Especially, 0 represents a virtual root node w0 , while all others corresponded to
words in x. The arc set A represents the unlabeled dependency relations of the particular analysis G. Specifically, an arc (i, j) ∈ A represents a dependency from head wi
to dependent w j . A dependency graph G is thus a set of unlabeled dependency relations between the root and the words of x. To represent an empty category augmented
dependency tree, we extend the vertex set and define a directed graph as usual.
To define a parsing model, we denote the index set of all possible dependencies
as I = {(i, j)|i ∈ {0, · · · , n}, j ∈ {1, · · · , n}, i , j}. A dependency parse can then be
represented as a vector
y = {y(i, j) : (i, j) ∈ I}
where y(i, j) = 1 if there is an arc (i, j) in the graph, and 0 otherwise. For a sentence x,
8
we define dependency parsing as a search for the highest-scoring analysis of x:
y∗ (x) = arg max Score(x, y)
y∈Y(x)
(2)
Here, Y(x) is the set of all trees compatible with x and Score(x, y) evaluates the event
that tree y is the analysis of sentence s. In brief, given a sentence x, we compute its
parse y∗ (x) by searching for the highest-scored dependency parse in the set of compatible trees Y(x). The scores are assigned by Score. In this paper, we evaluate structured
perceptron and define Score(x, y) as w> Φ(x, y) ,where Φ(x, y) is a feature-vector mapping and w is the corresponding parameter vector.
In general, performing a direct maximization over the set Y(x) is infeasible. The
common solution used in many parsing approaches is to introduce a part-wise factorization:
X
φ(x, p)
Φ(x, y) =
p∈Part(y)
Above, we have assumed that the dependency parse y can be factored into a set of
parts p, each of which represents a small substructure of y. For example, y might be
factored into the set of its component dependencies. A number of dynamic programming (DP) algorithms have been designed for first- [6], second- [7, 8], third- [9] and
fourth-order [10] factorization.
Parsing for joint empty category detection and dependency parsing can be defined
in a similar way. We use another index set I0 = {(i, j)|i, j ∈ {1, · · · , n + (n + 1)2 }},
where i > n indicates an empty node. Then a dependency parse with empty nodes can
be represented as a vector similar to y:
z = {z(i, j) : (i, j) ∈ I0 }.
Let Z(x) denote the set of all possible z for sentence x. We then define joint empty
category detection and dependency parsing as a search for the highest-scoring analysis
of x:
X
w> φ(x, z)
(3)
z∗ (x) = arg max
z∈Z(x)
p∈Part(z)
When the output of the factorization function, namely Part(x), is defined as the
collection of all sibling or tri-sibling dependencies, decoding for the above two optimization problems, namely (2) and (3), can be resolved in low-degree polynomial time
with respective to the number of words contained in x [7, 9, 11]. In particular, the
decoding algorithms proposed by Zhang et al. [11] are extensions of the algorithms
introduced respectively by McDonald and Pereira [7] and Koo and Collins [9].
To perform structure regularization decoding, we need to combine the two models.
In this problem, as the models are linear and do not involve probability, they can be
easily combined together. Assume that f : Y → R and g : Z → R assign scores to
parse trees without and with empty elements, respectively. In particular, the training
data for estimating f are sub-structures of the training data for estimating g. Therefore,
the training data for f can be viewed as the mini-samples of the training data for g.
9
A reasonable model to integrate f and g is to find the optimal parse by solving the
following optimization problem:
max.
s.t.
λ f (y) + (1 − λ)g(z)
y ∈ Y, z ∈ Z
y(i, j) = z(i, j), ∀(i, j) ∈ I
(4)
where λ is a weight for score combination.
In this paper, we employ dual decomposition to resolve the optimization problem
(4). We sketch the solution as follows.
The Lagrangian of (4) is
X
L(y, z; u) = f (y) + g(z) +
u(i, j)(y(i, j) − z(i, j))
(i, j)∈I
where u is the Lagrangian multiplier. Then the dual is
L(u)
=
max L(y, z; u)
X
X
= max( f (y) +
u(i, j)y(i, j)) + max(g(z) −
u(i, j)z(i, j))
y∈Y,z∈Z
y∈Y
z∈Z
(i, j)∈I
(i, j)∈I
We instead try to find the solution for
min L(u).
u
By using a subgradient method to calculate maxu L(u), we have another SR decoding
algorithm. Notice that, there is no need to train the simple model and the complex
model separately.
Algorithm 1 SR Decoding Algorithm for Dependency Parsing
u(0) ← 0
for k = 0 → K do
P
y ← arg maxy∈Y ( f (y) + i, j u(i, j)y(i, j))
P
z ← arg maxz∈Z (g(z) − i, j u(i, j)z(i, j))
if ∀(i, j) ∈ I, y(i, j) = z(i, j) then
return z
else
u(k+1) ← u(k) − α(k) (y − z)
end if
end for
return z
3. Theoretical Analysis: Structure Complexity vs. Overfitting Risk
We first describe the settings for the theoretical analysis, and give the necessary
definitions (Section 3.1). We then introduce the proposed method with the proper annotations for clearance of the analysis (Section 3.2). Finally, we give the theoretical
10
results on analyzing the generalization risk regarding to the structure complexity based
on stability (Section 3.3). The general idea behind the theoretical analysis is that the
overfitting risk increases with the complexity of the structure, because more complex
structures are less stable in training. If some examples are taken out of the training set,
the impact on the complex structure models is much severer compared to the simple
structure models. The detailed relations among the factors are shown by the analysis.
3.1. Problem Settings of Theoretical Analysis
In this section, we give the preliminary definitions necessary for the analysis, including the learning algorithm, the data, and the cost functions, and especially the definition of structural complexity. We also describe the properties and the assumptions
we make to facilitate the theoretical analysis.
A graph of observations (even with arbitrary structures) can be indexed and be denoted by an indexed sequence of observations O = {o1 , . . . , on }. We use the term sample to denote O = {o1 , . . . , on }. For example, in natural language processing, a sample
may correspond to a sentence of n words with dependencies of linear chain structures
(e.g., in part-of-speech tagging) or tree structures (e.g., in syntactic parsing). In signal
processing, a sample may correspond to a sequence of n signals with dependencies of
arbitrary structures. For simplicity in analysis, we assume all samples have n observations (thus n tags). In the analysis, we define structural complexity as the scope of the
structural dependency. For example, a dependency scope of two tags is considered less
complex than a dependency scope of three tags. In particular, the dependency scope
of n tags is considered the full dependency scope which is of the highest structural
complexity.
A sample is converted to an indexed sequence of feature vectors x = {xx(1) , . . . , x (n) },
where x (k) ∈ X is of the dimension d and corresponds to the local features extracted
from the position/index k. We can use an n × d matrix to represent x ∈ Xn . In other
words, we use X to denote the input space on a position, so that x is sampled from Xn .
Let Yn ⊂ Rn be structured output space, so that the structured output y are sampled
from Yn . Let Z = (Xn , Yn ) be a unified denotation of structured input and output
space. Let z = (xx, y ), which is sampled from Z, be a unified denotation of a (xx, y ) pair
in the training data.
Suppose a training set is
S = {zz1 = (xx1 , y 1 ), . . . , z m = (xxm , y m )},
with size m, and the samples are drawn i.i.d. from a distribution D which is unknown. A
learning algorithm is a function G : Zm 7→ F with the function space F ⊂ {Xn 7→ Yn },
i.e., G maps a training set S to a function GS : Xn 7→ Yn . We suppose G is symmetric
with respect to S , so that G is independent on the order of S .
Structural dependencies among tags are the major difference between structured
prediction and non-structured classification. For the latter case, a local classification
of g based on a position k can be expressed as g(xx(k−a) , . . . , x (k+a) ), where the term
{xx(k−a) , . . . , x (k+a) } represents a local window. However, for structured prediction, a
local classification on a position depends on the whole input x = {xx(1) , . . . , x (n) } rather
than a local window, due to the nature of structural dependencies among tags (e.g.,
11
graphical models like CRFs). Thus, in structured prediction a local classification on k
should be denoted as g(xx(1) , . . . , x (n) , k). To simplify the notation, we define
g(xx, k) , g(xx(1) , . . . , x (n) , k)
Given a training set S of size m, we define S \i as a modified training set, which
removes the i’th training sample:
S \i = {zz1 , . . . , z i−1 , z i+1 , . . . , z m },
and we define S i as another modified training set, which replaces the i’th training sample with a new sample ẑ i drawn from D:
S i = {zz1 , . . . , zi−1 , ẑi , zi+1 , . . . , zm },
We define the point-wise cost function c : Y × Y 7→ R+ as c[GS (xx, k), y (k) ], which
measures the cost on a position k by comparing GS (xx, k) and the gold-standard tag y (k) .
We introduce the point-wise loss as
`(GS , z , k) , c[GS (xx, k), y (k) ]
Then, we define the sample-wise cost function C : Yn × Yn 7→ R+ , which is the
cost function with respect to a whole sample. We introduce the sample-wise loss as
L(GS , z ) , C[GS (xx), y ] =
n
X
`(GS , z , k) =
k=1
n
X
c[GS (xx, k), y (k) ]
k=1
Given G and a training set S , what we are most interested in is the generalization
risk in structured prediction (i.e., the expected average loss) [12, 13]:
R(GS ) = Ez
h L(GS , z ) i
n
Unless specifically indicated in the context, the probabilities and expectations over
random variables, including Ez (.), ES (.), Pz (.), and PS (.), are based on the unknown
distribution D.
Since the distribution D is unknown, we have to estimate R(GS ) from S by using
the empirical risk:
m
Re (GS ) =
m
n
1 X
1 XX
L(GS , z i ) =
`(GS , z i , k)
mn i=1
mn i=1 k=1
In what follows, sometimes we will use the simplified notations, R and Re , to denote
R(GS ) and Re (GS ).
To state our theoretical results, we must describe several quantities and assumptions which are important in structured prediction. We follow some notations and assumptions on non-structured classification [14, 15]. We assume a simple real-valued
structured prediction scheme such that the class predicted on position k of x is the sign
12
of GS (xx, k) ∈ D. In practice, many popular structured prediction models have a realvalued cost function. Also, we assume the point-wise cost function cτ is convex and
τ-smooth such that ∀y1 , y2 ∈ D, ∀y∗ ∈ Y
|cτ (y1 , y∗ ) − cτ (y2 , y∗ )| ≤ τ|y1 − y2 |
(5)
While many structured learning models have convex objective function (e.g., CRFs),
some other models have non-convex objective function (e.g., deep neural networks).
It is well-known that the theoretical analysis on the non-convex cases are quite difficult. Our theoretical analysis is focused on the convex situations and hopefully it can
provide some insight for the more difficult non-convex cases. In fact, we will conduct
experiments on neural network models with non-convex objective functions, such as
LSTM. Experimental results demonstrate that the proposed structural complexity regularization method also works in the non-convex situations, in spite of the difficulty of
the theoretical analysis.
Then, τ-smooth versions of the loss and the cost function can be derived according
to their prior definitions:
Lτ (GS , z ) = Cτ [GS (xx), y ] =
n
X
`τ (GS , z , k) =
k=1
n
X
cτ [GS (xx, k), y (k) ]
k=1
Also, we use a value ρ to quantify the bound of |GS (xx, k)−GS \i (xx, k)| while changing
a single sample (with size n0 ≤ n) in the training set with respect to the structured input
x . This ρ-admissible assumption can be formulated as ∀k,
|GS (xx, k) − GS \i (xx, k)| ≤ ρ||GS − GS \i ||2 · ||xx||2
(6)
where ρ ∈ R+ is a value related to the design of algorithm G.
3.2. Structural Complexity Regularization
Base on the problem settings, we give definitions for the common weight regularization and the proposed structural complexity regularization. In the definition, the
proposed structural complexity regularization decomposes the dependency scope of the
training samples into smaller localized dependency scopes. The smaller localized dependency scopes form mini-samples for the learning algorithms. It is assumed that the
smaller localized dependency scopes are not overlapped. Hence, the analysis is for a
simplified version of structural complexity regularization. We are aware that in implementation, the constraint can be hard to guarantee. From an empirical side, structural
complexity works well without this constraint.
Most existing regularization techniques are proposed to regularize model weights/parameters,
e.g., a representative regularizer is the Gaussian regularizer or so called L2 regularizer.
We call such regularization techniques as weight regularization.
Definition 1 (Weight regularization). Let Nλ : F 7→ R+ be a weight regularization
function on F with regularization strength λ, the structured classification based objective function with general weight regularization is as follows:
Rλ (GS ) , Re (GS ) + Nλ (GS )
13
(7)
While weight regularization normalizes model weights, the proposed structural
complexity regularization method normalizes the structural complexity of the training samples. Our analysis is based on the different dependency scope (i.e., the scope of
the structural dependency), such that, for example, a tag depending on two tags in context is considered to have less structural complexity than a tag depending on four tags
in context. The structural complexity regularization is defined to make the dependency
scope smaller. To simplify the analysis, we suppose a baseline case that a sample z
has full dependency scope n, such that all tags in z have dependencies. Then, we introduce a factor α such that a sample z has localized dependency scope n/α. In this
case, α represents the reduction magnitude of the dependency scope. To simplify the
analysis without losing generality, we assume the localized dependency scopes do not
overlap with each other. Since the dependency scope is localized and non-overlapping,
we can split the original sample of the dependency scope n into α mini-samples of
the dependency scope of n/α. What we want to show is that, the learning with small
and non-overlapping dependency scope has less overfitting risk than the learning with
large dependency scope. Real-world tasks may have an overlapping dependency scope.
Hence, our theoretical analysis is for a simplified “essential” problem distilled from the
real-world tasks.
In what follows, we also directly call the dependency scope of a sample as the
structure complexity of the sample. Then, a simplified version of structural complexity regularization, specifically for our theoretical analysis, can be formally defined as
follows:
Definition 2 (Simplified structural complexity regularization for analysis). Let Nα :
F 7→ F be a structural complexity regularization function on F with regularization
strength α with 1 ≤ α ≤ n, the structured classification based objective function with
structural complexity regularization is as follows2 :
m
Rα (GS ) , Re [G Nα (S ) ] =
α
m
α
n/α
1 XXX
1 XX
L[GS 0 , z (i, j) ] =
`[GS 0 , z (i, j) , k] (8)
mn i=1 j=1
mn i=1 j=1 k=1
where Nα (zzi ) splits z i into α mini-samples {zz(i,1) , . . . , z (i,α) }, so that the mini-samples
have a dependency scope of n0 = n/α. Thus, we get
S 0 = {zz(1,1) , z(1,2) , . . . , z(1,α) , . . . , z(m,1) , z(m,2) , . . . , z(m,α) }
|
{z
}
|
{z
}
α
(9)
α
with mα mini-samples of expected structure complexity n/α. We can denote S 0 more
compactly as S 0 = {zz01 , z 02 , . . . , z 0mα } and Rα (GS ) can be simplified as
mα n/α
mα
Rα (GS ) ,
1 XX
1 X
L(GS 0 , z 0i ) =
`[GS 0 , z 0i , k]
mn i=1
mn i=1 k=1
(10)
2 The notation N is overloaded here. For clarity throughout, N with subscript λ refers to weight regularization function, and N with subscript α refers to structural complexity regularization function.
14
Note that, when the structural complexity regularization strength α = 1, we have
S 0 = S and Rα = Re .
Now, we have given a formal definition of structural complexity regularization,
by comparing it with the traditional weight regularization. Below, we show that the
structural complexity regularization can improve the stability of learned models, and
can finally reduce the overfitting risk of the learned models.
3.3. Stability of Structured Prediction
Because the generalization of a learning algorithm is positively correlated with the
stability of the learning algorithm [14], to analyze the generalization of the proposed
method, we instead examine the stability of the structured prediction. Here, stability
describes the extent to which the resulting learning function changes, when a sample
in the training set is removed. We prove that by decomposing the dependency scopes,
i.e, by regularizing the structural complexity, the stability of the learning algorithm can
be improved.
We first give the formal definitions of the stability with respect to the learning
algorithm, i.e., function stability.
Definition 3 (Function stability). A real-valued structured classification algorithm G
has “function value based stability” (“function stability” for short) ∆ if the following
holds: ∀zz = (xx, y ) ∈ Z, ∀S ∈ Zm , ∀i ∈ {1, . . . , m}, ∀k ∈ {1, . . . , n},
|GS (xx, k) − GS \i (xx, k)| ≤ ∆
The stability with respect to the cost function can be similarly defined.
Definition 4 (Loss stability). A structured classification algorithm G has “uniform
loss-based stability” (“loss stability” for short) ∆l if the following holds: ∀zz ∈ Z, ∀S ∈
Zm , ∀i ∈ {1, . . . , m}, ∀k ∈ {1, . . . , n},
|`(GS , z , k) − `(GS \i , z , k)| ≤ ∆l
G has “sample-wise uniform loss-based stability” (“sample loss stability” for short)
∆ s with respect to the loss function L if the following holds: ∀zz ∈ Z, ∀S ∈ Zm , ∀i ∈
{1, . . . , m},
|L(GS , z) − L(GS \i , z)| ≤ ∆ s
It is clear that the upper bounds of loss stability and function stability are linearly
correlated under the problem settings.
Lemma 5 (Loss stability vs. function stability). If a real-valued structured classification algorithm G has function stability ∆ with respect to loss function `τ , then G has
loss stability
∆l ≤ τ∆
and sample loss stability
∆ s ≤ nτ∆.
15
The proof is provided in Appendix A.
Here, we show that lower structural complexity has lower bound of stability, and
is more stable for the learning algorithm. The proposed method improves stability by
regularizing the structural complexity of training samples.
Theorem 6 (Stability vs. structural complexity regularization). With a training set
S of size m, let the learning algorithm G have the minimizer f based on commonly
used L2 weight regularization:
f = argmin Rα,λ (g) = argmin
g∈F
g∈F
mα
1 X
λ
Lτ (g, z 0j ) + ||g||22
mn j=1
2
(11)
where α denotes the structural complexity regularization strength with 1 ≤ α ≤ n.
Also, we have
1 X
λ
0
\i0
f \i = argmin Rα,λ
(g) = argmin
Lτ (g, z 0j ) + ||g||22
(12)
mn j,i0
2
g∈F
g∈F
where j , i0 means j ∈ {1, . . . , i0 − 1, i0 + 1, . . . , mα}.3 Assume Lτ is convex and
differentiable, and f (xx, k) is ρ-admissible. Let a local feature value is bounded by v
such that x (k,q) ≤ v for q ∈ {1, . . . , d}.4 Let ∆ denote the function stability of f compared
0
with f \i for ∀zz ∈ Z with |zz| = n. Then, ∆ is bounded by
∆≤
dτρ2 v2 n2
,
mλα2
(13)
and the corresponding loss stability is bounded by
∆l ≤
dτ2 ρ2 v2 n2
,
mλα2
and the corresponding sample loss stability is bounded by
∆s ≤
dτ2 ρ2 v2 n3
.
mλα2
The proof is given in Appendix A.
We can see that increasing the size of training set m results in linear improvement
of ∆, and increasing the strength of structural complexity regularization α results in
quadratic improvement of ∆.
0
The function stability ∆ is based on comparing f and f \i , i.e., the stability is based
on removing a mini-sample. Moreover, we can extend the analysis to the function
stability based on comparing f and f \i , i.e., the stability is based on removing a fullsize sample.
3 Note that, in some cases the notation i is ambiguous. For example, f \i can either denote the removing
of a sample in S or denote the removing of a mini-sample in S 0 . Thus, when the case is ambiguous, we use
different index symbols for S and S 0 , with i for indexing S and i0 for indexing S 0 , respectively.
4 Recall that d is the dimension of local feature vectors defined in Section 3.1.
16
Corollary 7 (Stability by removing a full sample). With a training set S of size m,
let the learning algorithm G have the minimizer f as defined before. Also, we have
f \i = argmin R\i
α,λ (g) = argmin
g∈F
g∈F
1 X
λ
Lτ (g, z 0j ) + ||g||22
mn j<i
2
(14)
where j < i means j ∈ {1, . . . , (i − 1)α, iα + 1, . . . , mα}, i.e., all the mini-samples derived
from the sample z i are removed. Assume Lτ is convex and differentiable, and f (xx, k)
is ρ-admissible. Let a local feature value be bounded by v such that x (k,q) ≤ v for
q ∈ {1, . . . , d}. Let ∆¯ denote the function stability of f comparing with f \i for ∀zz ∈ Z
with |zz| = n. Then, ∆¯ is bounded by
dτρ2 v2 n2
= α sup(∆),
∆¯ ≤
mλα
(15)
0
where ∆ is the function stability of f comparing with f \i , and sup(∆) =
described in Eq. (13). Similarly, we have
dτρ2 v2 n2
,
mλα2
as
dτ2 ρ2 v2 n2
∆¯ l ≤
= α sup(∆l ),
mλα
and
dτ2 ρ2 v2 n3
= α sup(∆ s ).
∆¯ s ≤
mλα
The proof is presented in Appendix A.
In the case that a full sample is removed, increasing the strength of structural complexity regularization α results in linear improvement of ∆.
3.4. Reduction of Generalization Risk
In this section, we formally describe the relation between the generalization and
the stability, and summarize the relationship between the proposed method and the
generalization. Finally, we draw our conclusions from the theoretical analysis.
Now, we analyze the relationship between the generalization and the stability.
Theorem 8 (Generalization vs. stability). Let G be a real-valued structured classification algorithm with a point-wise loss function `τ such that ∀k, 0 ≤ `τ (GS , z , k) ≤ γ.
Let f , ∆, and ∆¯ be defined before. Let R( f ) be the generalization risk of f based on
the expected sample z ∈ Z with size n, as defined before. Let Re ( f ) be the empirical
risk of f based on S , as defined like before. Then, for any δ ∈ (0, 1), with probability
at least 1 − δ over the random draw of the training set S , the generalization risk R( f )
is bounded by
r
ln δ−1
¯
¯
(16)
R( f ) ≤ Re ( f ) + 2τ∆ + (4m − 2)τ∆ + γ
2m
17
The proof is in Appendix A.
The upper bound of the generalization risk contains the loss stability, which is
rewritten as the function stability. We can see that better stability leads to lower bound
of the generalization risk.
By substituting the function stability with the formula we get from the structural
complexity regularization, we get the relation between the generalization and the structural complexity regularization.
Theorem 9 (Generalization vs. structural complexity regularization). Let the structured prediction objective function of G be penalized by structural complexity regularization with factor α ∈ [1, n] and L2 weight regularization with factor λ. The penalized
function has a minimizer f :
f = argmin Rα,λ (g) = argmin
g∈F
g∈F
mα
1 X
λ
Lτ (g, z 0j ) + ||g||22
mn j=1
2
(17)
Assume the point-wise loss `τ is convex and differentiable, and is bounded by `τ ( f, z , k) ≤
γ. Assume f (xx, k) is ρ-admissible. Let a local feature value be bounded by v such that
x (k,q) ≤ v for q ∈ {1, . . . , d}. Then, for any δ ∈ (0, 1), with probability at least 1 − δ over
the random draw of the training set S , the generalization risk R( f ) is bounded by
r
ln δ−1
2dτ2 ρ2 v2 n2 (4m − 2)dτ2 ρ2 v2 n2
R( f ) ≤ Re ( f ) +
+
+γ
(18)
mλα
mλα
2m
The proof is in Appendix A.
q −1
2 2 2 2
2 2 2 2
ρ v n
ρ v n
ln δ
We call the term 2dτmλα
+ (4m−2)dτ
+γ
mλα
2m in (18) as the “overfit-bound”.
Reducing the overfit-bound is crucial for reducing the generalization risk bound. Most
importantly, we can see from the overfit-bound that the structural complexity regularization factor α always stays together with the weight regularization factor λ, working
together to reduce the overfit-bound. This indicates that the structural complexity regularization is as important as the weight regularization for reducing the generalization
risk for structured prediction.
Moreover, since τ, ρ, and v are typically small compared with other variables, especially m, (18) can be approximated as follows by ignoring the small terms:
√
dn2 ln δ−1
R( f ) ≤ Re ( f ) + O
(19)
√
λα m
First, (19) suggests that structure complexity n can increase the overfit-bound on a magnitude of O(n2 ), and applying weight regularization can reduce the overfit-bound by
O(λ). Importantly, applying structural complexity regularization further (over weight
regularization) can additionally reduce the overfit-bound by a magnitude of O(α). When
α = 1, which means “no structural complexity regularization”, we have the worst
overfit-bound. Also, (19) suggests that increasing the size of training set can reduce
the overfit-bound on a square root level.
Theorem 9 also indicates that too simple structures may overkill the overfit-bound
but with a dominating empirical risk, while too complex structures may overkill the
18
empirical risk but with a dominating overfit-bound. Thus, to achieve the best prediction
accuracy, a balanced complexity of structures should be used for training the model.
By regularizing the complex structure with the simple structure, a balance between
the empirical risk and the overfitting risk can be achieved. In the proposed method,
the model of the complex structure and the simple structure are both used in decoding.
In essence, the decoding is based on the complex model, for the purpose of keeping
the empirical risk down. The simple model is used to regularize the structure of the
output, which means the structural complexity of the complex model is compromised.
Therefore, the overfitting risk is reduced.
To summarize, the proposed method decomposes the dependency scopes, that is,
regularizes the structural complexity. It leads to better stability of the model, which
means the generalization risk is lower. Under the problem settings, increasing the
regularization strength α can bring linear reduction of the overfit-bound. However,
too simple structure may cause a dominating empirical risk. To achieve a balanced
structural complexity, we could regularize the complex structure model with the simple
structure model. The complex structure model has low empirical risk, while the simple
structure model has low structural risk. The proposed method takes the advantages
of both the simple structure model and the complex structure model. As a result, the
overall overfitting risk can be reduced.
4. Experiments
We conduct experiments on natural language processing tasks. We are concerned
with two types of structures: linear-chain structures, e.g. word sequences, and hierarchical structures, e.g. phrase-structure trees and dependency trees. The natural language processing tasks concerning linear-chain structures include (1) text chunking,
(2) English named entity recognition, and (3) Dutch named entity recognition. We also
conduct experiments on a natural language processing task that involves hierarchical
structures, i.e. (4) dependency parsing with empty category detection.
4.1. Experiments on Sequence Labeling Tasks
Text Chunking (Chunking): The chunking data is from the CoNLL-2000 shared
task [16]. The training set consists of 8,936 sentences, and the test set consists of 2,012
sentences. Since there is no development data provided, we randomly sampled 5%
of the training data as development set for tuning hyper-parameters. The evaluation
metric is F1-score.
English Named Entity Recognition (English-NER): The English NER data is
from the CoNLL-2003 shared task [17]. There are four types of entities to be recognized: PERSON, LOCATION, ORGANIZATION, and MISC. This data has 14,987
sentences for training, 3,466 sentences for development, and 3,684 sentences for testing. The evaluation metric is F1-score.
Dutch Named Entity Recognition (Dutch-NER): We use the D-NER dataset [18]
from the shared task of CoNLL-2002. The dataset contains four types of named entities: PERSON, LOCATION, ORGANIZATION, and MISC. It has 15,806 sentences
for training, 2,895 sentences for development, and 5,195 sentences for testing. The
evaluation metric is F1-score.
19
Test score
BLSTM order1
BLSTM order2
BLSTM order2 + SR
BLSTM order3
BLSTM order3 + SR
Chunking
93.97
93.24
94.81 (+1.57)
92.50
95.23 (+2.73)
English-NER
87.65
87.59
89.72 (+2.13)
87.16
90.59 (+3.43)
Dutch-NER
76.04
76.33
80.51 (+4.18)
76.57
81.62 (+5.05)
Table 1: Comparing SR decoding with the baseline BLSTM models. As we can see, when the order of the
model increases, the performance of BLSTM deteriorates most of the time. With SR decoding, the structural
complexity is controlled, so that the performance of the higher order model is substantially improved. It is
also interesting that the improvement is larger when the order is higher.
Chunking
96
91.0
F−score (%)
95
82.0
90.0
80.0
89.0
94
78.0
88.0
93
76.0
87.0
BLSTM
BLSTM+SR
92
1
2
Order
(a) Chunking
3
86.0
74.0
order-1
BLSTM
order-2
order-3
BLSTM-STAD
(b) English-NER
order-1
BLSTM
order-2
order-3
BLSTM-STAD
(c) Dutch-NER
Figure 3: Comparing SR decoding with baseline LSTM models. SR decoding can substantially improve the
performance of the complex structure model. Moreover, it is clear that SR decoding improves the order-3
models more than the order-2 models.
Since LSTM [19] is a popular implementation of recurrent neural networks, we
highlight experiment results on LSTM. In this work, we use the bi-directional LSTM
(BLSTM) as the implementation of LSTM, considering it has better accuracy in practice. For BLSTM, we set the dimension of input layer to 200 and the dimension of
hidden layer to 300 for all the tasks.
The experiments on BLSTM are based on the Adam learning method [20]. Since
we find the default hyper parameters work satisfactorily on those tasks, following
Kingma and Ba [20] we use the default hyper parameters as follows: β1 = 0.9, β2 =
0.99, = 1 × 10−5 .
For the tasks with BLSTM, we find there is almost no difference by adding L2 regularization or not. Hence, we do not add L2 regularization for BLSTM. All weight
matrices, except for bias vectors and word embeddings, are diagonal matrices and randomly initialized by normal distribution.
We implement our code with the python package Tensorflow.
4.1.1. Results
First, we apply the proposed scalable multi-order decoding method on BLSTM
(BLSTM-SR). Table 1 compares the scores of BLSTM-SR and BLSTM on standard
test data. As we can see, when the order of the model is increased, the baseline model
worsens. The exception is the result of the Dutch-NER task. When the order of the
20
Model
Kudoh and Matsumoto [22]
Kudo and Matsumoto [23]
Sha and Pereira [24]
McDonald et al. [25]
Sun et al. [26]
Collobert et al. [27]
Sun [1]
Huang et al. [28]
This paper
F1
93.48
93.91
94.30
94.29
94.34
94.32
94.52
94.46
95.23
Table 2: Text Chunking: comparing with previous work. Shen and Sarkar [21] also achieve 95.23%, but
their result is based on noun phrase chunking. However, our result is based on all phrase chunking, which
has more tags to predict and is more difficult.
model is increased, the model is slightly improved. It demonstrates that, in practice,
although complex structure models have lower empirical risks, the structural risks are
more dominant.
The proposed method easily surpasses the baseline. For Chunking, the F1 error
rate reduction is 23.2% and 36.4% for the second-order model and the third-order
model, respectively. For English-NER, the proposed method reduces the F1 error rate
by 17.2% and 26.7% for the second-order model and the third-order model, respectively. For Dutch-NER, the F1 error rate reduction of 17.7% and 21.6% is achieved
respectively for the second-order model and the third-order model. It is clear that the
improvement is significant. We suppose the reason is that the proposed method can
combine both low-order and high order information. It helps to reduce the overfitting
risk. Thus, the F1 score is improved.
Moreover, the reduction is larger when the order is higher, i.e., the improvement
of order-3 models is better than that of order-2 models. This confirms the theoretical
results that higher structural complexity leads to higher structural risks. This also suggests the proposed method can alleviate the structural risks and keep the empirical risks
low. The phenomenon is better illustrated in Figure 3.
Table 2 shows the results on Chunking compared to previous work. We achieve the
state-of-the-art in all-phrase chunking. Shen and Sarkar [21] achieve the same score as
ours. However, they conduct experiments in noun phrase chunking (NP-chunking). All
phrase chunking contains much more tags than NP-chunking, which is more difficult.
SR decoding also achieves better results on English NER and Dutch NER than
existing methods. Huang et al. [28] employ a BLSTM-CRF model in the English
NER task and achieve F1 score of 90.10%. The score is lower than our best F1 score.
Chiu and Nichols [29] present a hybrid BLSTM with F1 score of 90.77%. The model
slightly outperforms our method, which may be due to the external CNNs they used to
extract word features. Gillick et al. [30] keep the best result of Dutch NER. However,
the model is trained on corpora of multilingual languages. Their model trained with
a single language gets 78.08% on F1 score and performs worse than ours. Nothman
et al. [18] reach 78.6% F1 with a semi-supervised approach in Dutch NER. Our model
21
English
Chinese
train
test
train
test
#{Sent}
38667
2336
8605
941
#{Overt}
909114
54242
193417
21797
#{Covert}
57342
3447
13518
1520
Table 3: Numbers of sentences, overt and covert elements in training and test sets.
still outperforms the method.
4.2. Experiments on Parsing
Joint Empty Category Detection and Dependency Parsing For joint empty category detection and dependency parsing, we conduct experiments on both English and
Chinese treebanks. In particular, English Penn TreeBank (PTB) [31] and Chinese TreeBank (CTB) [32] are used . Because PTB and CTB are phrase-structure treebanks, we
need to convert them into dependency annotations. To do so, we use the tool provided
by Stanford CoreNLP to process PTB, and the tool introduced by Xue and Yang [5]
to process CTB 5.0. We use gold-standard POS to derive features for disambiguation. To simplify our experiments, we preprocess the obtained dependency trees in the
following way.
1. We combine successive empty elements with identical head into one new empty
node that is still linked to the common head word.
2. Because the high-order algorithm is very expensive with respect to the computational cost, we only use relatively short sentence. Here we only keep the
sentences that are less than 64 tokens.
3. We focus on unlabeled parsing.
The statistics of the data after cleaning are shown in Table 3. We use the standard
training, validation, and test splits to facilitate comparisons. Accuracy is measured
with unlabeled attachment score for all overt words (UASo ): the percentage of the overt
words with the correct head. We are also concerned with the prediction accuracy for
empty elements. To evaluate performance on empty nodes, we consider the correctness
of empty edges. We report the percentage of the empty words in the right slot with
correct head. The i-th slot in the sentence means that the position immediately after the
i-th concrete word. So if we have a sentence with length n, we get n + 1 slots.
4.2.1. Results
Table 4 lists the accuracy of individual models coupled with different decoding
algorithms on the test sets. We focus on the prediction for overt words only. When
we take into account empty categories, more information is available. However, the
increased structural complexity affects the algorithms. From the table, we can see that
the complex sibling factorization works worse than the simple sibling factorization
in English, but works better in Chinese. The results of the tri-sibling factorization
are exactly the opposite. The complex tri-sibling factorization works better than the
22
Factorization
Sibling
Sibling (complex)
Tri-sibling
Tri-sibling (complex)
Empty Element
No
Yes
No
Yes
English
91.73
91.70
92.23
92.41
Chinese
89.16
89.20
90.00
89.82
Table 4: UASo of different individual models on test data. The upper and bottom blocks present results
obtained by sibling and tri-sibling models respectively. We cannot draw a conclusion whether the complex
structure models performs better or not, because the results vary with the structural complexity and the data.
This is expected, because structural complexity affects both the empirical risks and the structural risks. As
the data (English vs. Chinese) and the structure (Sibling vs. Tri-sibling) are different, the relations between
empirical risks and the structural risks are also different. Hence, the varied results.
Method
Sibling (complex)
Sibling (complex) + SR
Tri-sibling (complex)
Tri-sibling (complex) + SR
English
91.70
91.96 (+0.26)
92.41
92.71 (+0.30)
Chinese
89.20
89.53 (+0.33)
89.82
90.38 (+0.56)
Table 5: UASo of different SR decoding models on test data. As we can see, SR decoding consistently
improves the complex structure models. All improvements are statistically significant.
simple tri-sibling factorization in English, but works worse in Chinese. The results can
be explained by our theoretical analysis. While the structural complexity is positively
correlated with the overfitting risk, it is negatively correlated with the empirical risk. In
this task, although the overfitting risk is increased when using the complex structure,
the empirical risk is decreased more sometimes. Hence, the results vary both with the
structural complexity and the data.
Table 5 lists the accuracy of different SR decoding models on the test sets. We can
see that the SR decoding framework is effective to deal with the structure-based overfitting. This time, the accuracy of analysis for overt words is consistently improved.
For the second-order model, SR decoding reduces the error rate by 3.1% for English,
and by 3.0% for Chinese. For the third-order model, the error rate reduction of 4.0%
for English, and 5.5% for Chinese is achieved by the proposed method. Similar to the
sequence labeling tasks, the third-order model is improved more. We suppose the consistent improvements come from the ability of reducing the structural risk of the SR
decoding algorithm. Although in this task, the complex structure is sometimes helpful
to the accuracy of the parsing, the structural risk still increases. By regularizing the
structural complexity, further improvements can be achieved, on top of the decreased
empirical risk brought by the complex structure.
We use the Hypothesis Tests method [33] to evaluate the improvements. When the
p-value is set to 0.05, all improvements in Table 5 are statistically significant.
5. Related Work
The term structural regularization has been used in prior work for regularizing
structures of features. For (typically non-structured) classification problems, there
23
are considerable studies on structure-related regularization. Argyriou et al. [34] apply
spectral regularization for modeling feature structures in multi-task learning, where the
shared structure of the multiple tasks is summarized by a spectral function over the
tasks’ covariance matrix, and then is used to regularize the learning of each individual
task. Xue et al. [35] regularize feature structures for structural large margin binary classifiers, where data points belonging to the same class are clustered into subclasses so
that the features for the data points in the same subclass can be regularized. While those
methods focus on the regularization approaches, many recent studies focus on exploiting the structured sparsity. Structure sparsity is studied for a variety of non-structured
classification models [36, 37] and structured prediction scenarios [38, 39], via adopting
mixed norm regularization [40], Group Lasso [41], posterior regularization [42], and a
string of variations [43, 44, 45].
Compared with those pieces of prior work, the proposed method works on a substantially different basis. This is because the term structure in all of the aforementioned
work refers to structures of feature space, which is substantially different compared
with our proposal on regularizing tag structures (interactions among tags).
There are other related studies, including the studies of Sutton and McCallum [46]
and Samdani and Roth [47] on piecewise/decomposed training methods, and the study
of Tsuruoka et al. [48] on a “lookahead” learning method. They both try to reduce the
computational cost of the model by reducing the structure involved. Sutton and McCallum [46] try to simply the structural dependencies of the graphic probabilistic models,
so that the model can be efficiently trained. Samdani and Roth [47] try to simply the
output space in structured SVM by decomposing the structure into sub-structures, so
that the search space is reduced, and the training is tractable. Tsuruoka et al. [48] try to
train a localized structured perceptron, where the local output is searched by stepping
into the future instead of directly using the result of the classifier.
Our work differs from Sutton and McCallum [46], Samdani and Roth [47], Tsuruoka et al. [48], because our work is built on a regularization framework, with arguments and justifications on reducing generalization risk and for better accuracy, although it has the effect that the decoding space of the complex model is reduced by the
simple model. Also, the theoretical results can fit general graphical models, and the
detailed algorithm is quite different.
On generalization risk analysis, related studies include Tsuruoka et al. [49], Bousquet and Elisseeff [14], Shalev-Shwartz et al. [15] on non-structured classification and
Taskar et al. [12], London et al. [13, 50] on structured classification. This work targets
the theoretical analysis of the relations between the structural complexity of structured
classification problems and the generalization risk, which is a new perspective compared with those studies.
6. Conclusions
We propose a structural complexity regularization framework, called structure regularization decoding. In the proposed method, we train the complex structure model.
In addition, we also train the simple structure model. The simple structure model is
used to regularize the decoding of the complex structure model. The resulting model
24
embodies a balanced structural complexity, which reduces the structure-based overfitting risk. We derive the structure regularization decoding algorithms for linear-chain
models on sequence labeling tasks, and for hierarchical models on parsing tasks.
Our theoretical analysis shows that the proposed method can effectively reduce the
generalization risk, and the analysis is suitable for graphic models. In theory, higher
structural complexity leads to higher structure-based overfitting risk, but lower empirical risk. To achieve better performance, a balanced structural complexity should
be maintained. By regularizing the structural complexity, that is, decomposing the
structural dependencies but also keeping the original structural dependencies, structurebased overfitting risk can be alleviated and empirical risk can be kept low as well.
Experimental results demonstrate that the proposed method easily surpasses the
performance of the complex structure models. Especially, the proposed method is also
suitable for deep learning models. On the sequence labeling tasks, the proposed method
substantially improves the performance of the complex structure models, with the maximum F1 error rate reduction of 23.2% for the second-order models, and 36.4% for the
third-order models. On the parsing task, the maximum UAS improvement of 5.5% on
Chinese tri-sibling factorization is achieved by the proposed method. The results are
competitive with or even better than the state-of-the-art results.
Acknowledgments
This work was supported in part by National Natural Science Foundation of China
(No. 61300063), and Doctoral Fund of Ministry of Education of China (No. 20130001120004).
This work is a substantial extension of a conference paper presented at NIPS 2014 [1].
References
References
[1] X. Sun, Structure regularization for structured prediction, in: Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing
Systems 2014, Montreal, Quebec, Canada, pp. 2402–2410.
[2] S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural Computation 9 (1997)
1735–1780.
[3] M. Collins, Discriminative training methods for hidden markov models: Theory and experiments with perceptron algorithms, in: Proceedings of EMNLP’02, pp. 1–8.
[4] N. Chomsky, Lectures on Government and Binding, Foris Publications, Dordecht, 1981.
[5] N. Xue, Y. Yang, Dependency-based empty category detection via phrase structure trees,
in: Proceedings of the 2013 Conference of the North American Chapter of the Association
for Computational Linguistics: Human Language Technologies, Association for Computational Linguistics, Atlanta, Georgia, 2013, pp. 1051–1060.
[6] J. M. Eisner, Three new probabilistic models for dependency parsing: an exploration, in:
Proceedings of the 16th conference on Computational linguistics - Volume 1, Association
for Computational Linguistics, Stroudsburg, PA, USA, 1996, pp. 340–345.
[7] R. McDonald, F. Pereira, Online learning of approximate dependency parsing algorithms,
in: Proceedings of 11th Conference of the European Chapter of the Association for Computational Linguistics (EACL-2006)), volume 6, pp. 81–88.
25
[8] X. Carreras, Experiments with a higher-order projective dependency parser, in: J. Eisner (Ed.), EMNLP-CoNLL 2007, Proceedings of the 2007 Joint Conference on Empirical
Methods in Natural Language Processing and Computational Natural Language Learning,
ACL, Prague, Czech Republic, 2007, pp. 957–961.
[9] T. Koo, M. Collins, Efficient third-order dependency parsers, in: Proceedings of the
48th Annual Meeting of the Association for Computational Linguistics, Association for
Computational Linguistics, Uppsala, Sweden, 2010, pp. 1–11.
[10] X. Ma, H. Zhao, Fourth-order dependency parsing, in: Proceedings of COLING 2012:
Posters, The COLING 2012 Organizing Committee, Mumbai, India, 2012, pp. 785–796.
[11] X. Zhang, W. Sun, X. Wan, The covert helps parse the overt, in: Proceedings of the 21st
Conference on Computational Natural Language Learning (CoNLL 2017), Association for
Computational Linguistics, Vancouver, Canada, 2017, pp. 343–353.
[12] B. Taskar, C. Guestrin, D. Koller, Max-margin markov networks, in: S. Thrun, L. K.
Saul, B. Schölkopf (Eds.), Advances in Neural Information Processing Systems 16 [Neural
Information Processing Systems, NIPS 2003], MIT Press, Vancouver and Whistler, British
Columbia, Canada, 2003, pp. 25–32.
[13] B. London, B. Huang, B. Taskar, L. Getoor., Pac-bayes generalization bounds for randomized structured prediction, in: NIPS Workshop on Perturbation, Optimization and
Statistics.
[14] O. Bousquet, A. Elisseeff, Stability and generalization., Journal of Machine Learning
Research 2 (2002) 499–526.
[15] S. Shalev-Shwartz, O. Shamir, N. Srebro, K. Sridharan, Learnability and stability in the
general learning setting, in: COLT 2009 - The 22nd Conference on Learning Theory,
Montreal, Quebec, Canada, June 18-21, 2009.
[16] E. T. K. Sang, S. Buchholz, Introduction to the CoNLL-2000 shared task: Chunking, in:
Proceedings of CoNLL’00, pp. 127–132.
[17] E. F. Sang, F. D. Meulder, Introduction to the CoNLL-2003 Shared Task: LanguageIndependent Named Entity Recognition, in: Proceedings of CoNLL-2003, pp. 142–147.
[18] J. Nothman, N. Ringland, W. Radford, T. Murphy, J. R. Curran, Learning multilingual
named entity recognition from wikipedia, Artificial Intelligence 194 (2013) 151–175.
[19] J. Hammerton, Named entity recognition with long short-term memory, in: CoNLL, ACL,
2003, pp. 172–175.
[20] D. P. Kingma, J. Ba, Adam: A method for stochastic optimization, CoRR abs/1412.6980
(2014).
[21] H. Shen, A. Sarkar, Voting between multiple data representations for text chunking, in:
Advances in Artificial Intelligence, 18th Conference of the Canadian Society for Computational Studies of Intelligence, Canadian AI 2005, Victoria, Canada, May 9-11, 2005,
Proceedings, pp. 389–400.
[22] T. Kudoh, Y. Matsumoto, Use of support vector learning for chunk identification, in: Fourth
Conference on Computational Natural Language Learning, CoNLL 2000, and the Second
Learning Language in Logic Workshop, LLL 2000, Held in cooperation with ICGI-2000,
Lisbon, Portugal, September 13-14, 2000, pp. 142–144.
[23] T. Kudo, Y. Matsumoto, Chunking with support vector machines, in: Proceedings of
NAACL’01, pp. 1–8.
[24] F. Sha, F. Pereira, Shallow parsing with conditional random fields, in: NAACL ’03:
Proceedings of the 2003 Conference of the North American Chapter of the Association for
Computational Linguistics on Human Language Technology, pp. 134–141.
[25] R. T. McDonald, K. Crammer, F. Pereira, Flexible text segmentation with structured multilabel classification, in: HLT/EMNLP 2005, Human Language Technology Conference
and Conference on Empirical Methods in Natural Language Processing, Proceedings of
26
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
the Conference, 6-8 October 2005, Vancouver, British Columbia, Canada, The Association
for Computational Linguistics, 2005, pp. 987–994.
X. Sun, L. Morency, D. Okanohara, Y. Tsuruoka, J. Tsujii, Modeling latent-dynamic in
shallow parsing: A latent conditional model with imrpoved inference, in: COLING 2008,
22nd International Conference on Computational Linguistics, Proceedings of the Conference, 18-22 August 2008, Manchester, UK, pp. 841–848.
R. Collobert, J. Weston, L. Bottou, M. Karlen, K. Kavukcuoglu, P. P. Kuksa, Natural
language processing (almost) from scratch, Journal of Machine Learning Research 12
(2011) 2493–2537.
Z. Huang, W. Xu, K. Yu, Bidirectional lstm-crf models for sequence tagging, arXiv preprint
arXiv:1508.01991 (2015).
J. P. C. Chiu, E. Nichols, Named entity recognition with bidirectional lstm-cnns., TACL 4
(2016) 357–370.
D. Gillick, C. Brunk, O. Vinyals, A. Subramanya, Multilingual language processing from
bytes, CoRR abs/1512.00103 (2015).
N. Xue, F. Xia, F.-d. Chiou, M. Palmer, The penn Chinese treebank: Phrase structure
annotation of a large corpus, Natural Language Engineering 11 (2005) 207–238.
M. P. Marcus, M. A. Marcinkiewicz, B. Santorini, Building a large annotated corpus of
english: the penn treebank, Computational Linguistics 19 (1993) 313–330.
T. Berg-Kirkpatrick, D. Burkett, D. Klein, An empirical investigation of statistical significance in nlp, in: Proceedings of the 2012 Joint Conference on Empirical Methods in
Natural Language Processing and Computational Natural Language Learning, Association
for Computational Linguistics, Jeju Island, Korea, 2012, pp. 995–1005.
A. Argyriou, C. A. Micchelli, M. Pontil, Y. Ying, A spectral regularization framework for
multi-task structure learning, in: J. C. Platt, D. Koller, Y. Singer, S. T. Roweis (Eds.), Advances in Neural Information Processing Systems 20, Proceedings of the Twenty-First Annual Conference on Neural Information Processing Systems, Vancouver, British Columbia,
Canada, December 3-6, 2007, Curran Associates, Inc., 2007, pp. 25–32.
H. Xue, S. Chen, Q. Yang, Structural regularized support vector machine: A framework
for structural large margin classifier., IEEE Transactions on Neural Networks 22 (2011)
573–587.
C. A. Micchelli, J. Morales, M. Pontil, A family of penalty functions for structured sparsity., in: Proceedings of NIPS’10, Curran Associates, Inc., 2010, pp. 1612–1623.
J. C. Duchi, Y. Singer, Boosting with structural sparsity, in: A. P. Danyluk, L. Bottou,
M. L. Littman (Eds.), Proceedings of the 26th Annual International Conference on Machine
Learning, ICML 2009, Montreal, Quebec, Canada, June 14-18, 2009, volume 382 of ACM
International Conference Proceeding Series, ACM, 2009, pp. 297–304.
M. W. Schmidt, K. P. Murphy, Convex structure learning in log-linear models: Beyond
pairwise potentials., in: Proceedings of AISTATS’10, volume 9 of JMLR Proceedings, pp.
709–716.
A. F. T. Martins, N. A. Smith, M. A. T. Figueiredo, P. M. Q. Aguiar, Structured sparsity in
structured prediction., in: Proceedings of EMNLP’11, pp. 1500–1511.
A. Quattoni, X. Carreras, M. Collins, T. Darrell, An efficient projection for l1,infinity
regularization., in: Proceedings of ICML’09, p. 108.
M. Yuan, Y. Lin, Model selection and estimation in regression with grouped variables,
Journal of the Royal Statistical Society, Series B 68 (2006) 49–67.
J. Graça, K. Ganchev, B. Taskar, F. Pereira, Posterior vs parameter sparsity in latent variable
models, in: Proceedings of NIPS’09, pp. 664–672.
F. Bach, R. Jenatton, J. Mairal, G. Obozinski, Structured sparsity through convex optimization, CoRR abs/1109.2397 (2011).
27
[44] G. Obozinski, B. Taskar, M. I. Jordan, Joint covariate selection and joint subspace selection
for multiple classification problems., Statistics and Computing 20 (2010) 231–252.
[45] J. Huang, T. Zhang, D. N. Metaxas, Learning with structured sparsity., Journal of Machine
Learning Research 12 (2011) 3371–3412.
[46] C. A. Sutton, A. McCallum, Piecewise pseudolikelihood for efficient training of conditional
random fields., in: ICML’07, ACM, 2007, pp. 863–870.
[47] R. Samdani, D. Roth, Efficient decomposed learning for structured prediction, in: Proceedings of the 29th International Conference on Machine Learning, ICML 2012, Edinburgh,
Scotland, UK, June 26 - July 1, 2012, icml.cc / Omnipress, 2012.
[48] Y. Tsuruoka, Y. Miyao, J. Kazama, Learning with lookahead: Can history-based models
rival globally optimized models?, in: S. Goldwater, C. D. Manning (Eds.), Proceedings
of the Fifteenth Conference on Computational Natural Language Learning, CoNLL 2011,
Portland, Oregon, USA, June 23-24, 2011, ACL, 2011, pp. 238–246.
[49] Y. Tsuruoka, J. Tsujii, S. Ananiadou, Stochastic gradient descent training for l1-regularized
log-linear models with cumulative penalty, in: Proceedings of ACL’09, Suntec, Singapore,
pp. 477–485.
[50] B. London, B. Huang, B. Taskar, L. Getoor, Collective stability in structured prediction:
Generalization from one example, in: Proceedings of the 30th International Conference on
Machine Learning (ICML-13), pp. 828–836.
[51] X. Sun, W. Li, H. Wang, Q. Lu, Feature-frequency-adaptive on-line training for fast and
accurate natural language processing, Computational Linguistics 40 (2014) 563–586.
[52] H. Noji, Y. Miyao, M. Johnson, Using left-corner parsing to encode universal structural
constraints in grammar induction, in: Proceedings of the EMNLP’16, pp. 33–43.
[53] Y. Miyao, R. Sætre, K. Sagae, T. Matsuzaki, J. Tsujii, Task-oriented evaluation of syntactic
parsers and their representations, in: Proceedings of ACL’08, pp. 46–54.
[54] X. Sun, T. Matsuzaki, D. Okanohara, J. Tsujii, Latent variable perceptron algorithm for
structured classification, in: Proceedings of the 21st International Joint Conference on
Artificial Intelligence (IJCAI 2009), pp. 1236–1242.
[55] T. Matsuzaki, Y. Miyao, J. Tsujii, Probabilistic CFG with latent annotations, in: Proceedings of ACL’05, pp. 75–82.
[56] T. Matsuzaki, J. Tsujii, Comparative parser performance analysis across grammar frameworks through automatic tree conversion using synchronous grammars, in: Proceedings of
COLING’08, pp. 545–552.
[57] J. Nivre, J. Hall, J. Nilsson, A. Chanev, G. Eryigit, S. Kübler, S. Marinov, E. Marsi, Maltparser: A language-independent system for data-driven dependency parsing, Natural Language Engineering 13 (2007) 95–135.
[58] J. Nivre, Algorithms for deterministic incremental dependency parsing, Computational
Linguistics 34 (2008) 513–553.
[59] Y. Zhang, S. Clark, A tale of two parsers: Investigating and combining graph-based and
transition-based dependency parsing, in: Proceedings of EMNLP’08, pp. 562–571.
[60] Y. Zhang, S. Clark, Syntactic processing using the generalized perceptron and beam search,
Computational Linguistics 37 (2011) 105–151.
[61] O. Vinyals, L. Kaiser, T. Koo, S. Petrov, I. Sutskever, G. E. Hinton, Grammar as a foreign
language, in: NIPS’15, pp. 2773–2781.
[62] H. He, X. Sun, A unified model for cross-domain and semi-supervised named entity recognition in chinese social media, in: Proceedings of the Thirty-First AAAI Conference on
Artificial Intelligence, February 4-9, 2017, San Francisco, California, USA., pp. 3216–
3222.
[63] H. He, X. Sun, F-score driven max margin neural network for named entity recognition
in chinese social media, in: Proceedings of the 15th Conference of the European Chapter
28
of the Association for Computational Linguistics, EACL 2017, Valencia, Spain, April 3-7,
2017, Volume 2: Short Papers, pp. 713–718.
[64] X. Sun, Structure regularization for structured prediction: Theories and experiments,
CoRR abs/1411.6243 (2014).
Appendix A. Proof
Our analysis sometimes needs to use McDiarmid’s inequality.
Theorem 10 (McDiarmid, 1989). Let S = {q1 , . . . , qm } be independent random variables taking values in the space Qm . Moreover, let g : Qm 7→ R be a function of S that
satisfies ∀i, ∀S ∈ Qm , ∀q̂i ∈ Q,
|g(S ) − g(S i )| ≤ ci .
Then ∀ > 0,
−2 2
PS [g(S ) − ES [g(S )] ≥ ] ≤ exp Pm 2 .
i=1 ci
Lemma 11 (Symmetric learning). For any symmetric (i.e., order-free) learning algorithm G, ∀i ∈ {1, . . . , m}, we have
ES [R(GS ) − Re (GS )] =
1
ES ,ẑẑi [L(GS , ẑi ) − L(GS i , ẑi )]
n
Proof
m
1
1X
ES [R(GS ) − Re (GS )] = ES Ez (L(GS , z)) −
L(GS , z j )
n
m j=1
m
=
1X
1
ES L(GS , z j )
ES ,ẑẑi L(GS , ẑ i ) −
n
m j=1
1
ES ,ẑẑi L(GS , ẑ i ) − ES L(GS , z i )
n
1
= ES ,ẑẑi L(GS , ẑ i ) − ES i L(GS i , ẑ i )
n
1
= ES ,ẑẑi L(GS , ẑ i ) − L(GS i , ẑ i )
n
=
where the 3rd step is based on ES L(GS , z i ) = ES L(GS , z j ) for ∀zzi ∈ S and ∀zz j ∈ S ,
given that G is symmetric.
t
u
Appendix A.1. Proofs
Proof of Lemma 5
29
According to (5), we have ∀i, ∀S , ∀zz, ∀k
|`τ (GS , z, k) − `τ (GS \i , z, k)| = |cτ [GS (xx, k), y(k) ] − cτ [GS \i (xx, k), y(k) ]|
≤ τ|GS (xx, k) − GS \i (xx, k)|
≤ τ∆
This gives the bound of loss stability.
Also, we have ∀i, ∀S , ∀zz
|Lτ (GS , z) − Lτ (GS \i , z)| =
≤
n
X
k=1
n
X
cτ [GS (xx, k), y(k) ] −
n
X
cτ [GS \i (xx, k), y(k) ]
k=1
cτ [GS (xx, k), y (k) ] − cτ [GS \i (xx, k), y (k) ]
k=1
n
X
≤τ
|GS (xx, k) − GS \i (xx, k)|
k=1
≤ nτ∆
t
u
This derives the bound of sample loss stability.
Proof of Theorem 6
When a convex and differentiable function g has a minimum f in space F , its
Bregman divergence has the following property for ∀ f 0 ∈ F :
dg ( f 0 , f ) = g( f 0 ) − g( f )
With this property, we have
0
0
0
0
0
0
\i
\i
dRα,λ ( f \i , f ) + dR\i0 ( f, f \i ) = Rα,λ ( f \i ) − Rα,λ ( f ) + R\i
α,λ ( f ) − Rα,λ ( f )
α,λ
0
0
\i0
\i0
= Rα,λ ( f \i ) − R\i
α,λ ( f ) − Rα,λ ( f ) − Rα,λ ( f )
1
1
0
Lτ ( f \i , z 0i0 ) −
Lτ ( f, z 0i0 )
=
mn
mn
30
(A.1)
Then, based on the property of Bregman divergence that dg+g0 = dg + dg0 , we have
0
0
0
0
dNλ ( f, f \i ) + dNλ ( f \i , f ) = d(R\i0 −R\iα 0 ) ( f, f \i ) + d(Rα,λ −Rα ) ( f \i , f )
α,λ
0
0
0
0
= dRα,λ ( f \i , f ) + dR\i0 ( f, f \i ) − dRα ( f \i , f ) − dR\iα 0 ( f, f \i )
α,λ
(based on non-negativity of Bregman divergence)
0
0
≤ dRα,λ ( f \i , f ) + dR\i0 ( f, f \i )
α,λ
(using (A.1))
1
0
=
Lτ ( f \i , z 0i0 ) − Lτ ( f, z 0i0 )
mn
n/α
1 X
0
`τ ( f \i , z 0i0 , k) − `τ ( f, z 0i0 , k)
=
mn k=1
(A.2)
n/α
1 X \i0 0
cτ f (xxi0 , k), y 0i0 (k) − cτ f (xx0i0 , k), y 0i0 (k)
≤
mn k=1
n/α
≤
τ X \i0 0
f (xxi0 , k) − f (xx0i0 , k)
mn k=1
(using (6))
ρτ
0
|| f − f \i ||2 · ||xx0i0 ||2
≤
mα
Moreover, Nλ (g) =
gence satisfies:
λ
2
2 ||g||2
=
λ
2 hg, gi
is a convex function and its Bregman diver-
λ
hg, gi − hg0 , g0 i − h2g0 , g − g0 i
2
λ
= ||g − g0 ||22
2
dNλ (g, g0 ) =
(A.3)
Combining (A.2) and (A.3) gives
0
λ|| f − f \i ||22 ≤
ρτ
0
|| f − f \i ||2 · ||xx0i0 ||2
mα
(A.4)
which further gives
ρτ 0
||xx 0 ||2
(A.5)
mλα i
Given ρ-admissibility, we derive the bound of function stability ∆( f ) based on sample z with size n. We have ∀zz = (xx, y ), ∀k,
0
|| f − f \i ||2 ≤
0
0
| f (xx, k) − f \i (xx, k)| ≤ ρ|| f − f \i ||2 · ||xx||2
(using (A.5))
τρ2 0
≤
||xx 0 ||2 · ||xx||2
mλα i
31
(A.6)
With the feature dimension d and x (k,q) ≤ v for q ∈ {1, . . . , d} , we have
||xx||2 = ||
n
X
x (k) ||2
k=1
≤ ||hnv, . . . , nvi||2
| {z }
(A.7)
d
√
= dn2 v2
√
= nv d
√
Similarly, we have ||xx0i0 ||2 ≤ nvα d because x 0i0 is with the size n/α.
Inserting the bounds of ||xx||2 and ||xx0i0 ||2 into (A.6), it goes to
0
| f (xx, k) − f \i (xx, k)| ≤
dτρ2 v2 n2
mλα2
which gives (13). Further, using Lemma 5 derives the loss stability bound of
and the sample loss stability bound of
dτ2 ρ2 v2 n3
mλα2
on the minimizer f .
(A.8)
dτ2 ρ2 v2 n2
,
mλα2
t
u
Proof of Corollary 7
The proof is similar to the proof of Theorem 6. First, we have
\i
\i
dRα,λ ( f \i , f ) + dR\i ( f, f \i ) = Rα,λ ( f \i ) − Rα,λ ( f ) + R\i
α,λ ( f ) − Rα,λ ( f )
α,λ
\i
\i
= Rα,λ ( f \i ) − R\i
α,λ ( f ) − Rα,λ ( f ) − Rα,λ ( f )
α
α
1 X
1 X
Lτ ( f \i , z (i, j) ) −
Lτ ( f, z (i, j) )
=
mn j=1
mn j=1
32
(A.9)
Then, we have
dNλ ( f, f \i ) + dNλ ( f \i , f ) = d(R\i
\i
α,λ −Rα )
( f, f \i ) + d(Rα,λ −Rα ) ( f \i , f )
= dRα,λ ( f \i , f ) + dR\i ( f, f \i ) − dRα ( f \i , f ) − dR\iα ( f, f \i )
α,λ
(based on non-negativity of Bregman divergence)
≤ dRα,λ ( f \i , f ) + dR\i ( f, f \i )
α,λ
(using (A.9))
α
α
1 X
1 X
=
Lτ ( f \i , z (i, j) ) −
Lτ ( f, z (i, j) )
mn j=1
mn j=1
=
!
n/α
n/α
α
X
1 X X
`τ ( f \i , z (i, j) , k) −
`τ ( f, z (i, j) , k)
mn j=1 k=1
k=1
≤
1 XX
`τ ( f \i , z (i, j) , k) − `τ ( f, z (i, j) , k)
mn j=1 k=1
≤
τ X X \i
f (xx(i, j) , k) − f (xx(i, j) , k)
mn j=1 k=1
α
n/α
α
n/α
(A.10)
(using (6), and define ||xx(i,max) ||2 = max ||xx(i, j) ||)2 )
ρτ
≤ || f − f \i ||2 · ||xx(i,max) ||2
m
This gives
λ|| f − f \i ||22 ≤
∀j
ρτ
|| f − f \i ||2 · ||xx(i,max) ||2
m
(A.11)
and thus
ρτ
||xx(i,max) ||2
(A.12)
mλ
Then, we derive the bound of function stability ∆( f ) based on sample z with size n,
and based on \i rather than \i0 . We have ∀zz = (xx, y ), ∀k,
|| f − f \i ||2 ≤
| f (xx, k) − f \i (xx, k)| ≤ ρ|| f − f \i ||2 · ||xx||2
(using (A.12))
τρ2
||xx(i,max) ||2 · ||xx||2
mλ
√
√
τρ2 nv d
≤
·
· nv d
mλ
α
dτρ2 v2 n2
=
mλα
(using (13))
≤
(A.13)
= α sup(∆)
t
u
33
Proof of Theorem 8
Let f \i be defined like before. Similar to the definition of f \i based on removing a
sample from S , we define f i based on replacing a sample from S . Let R( f )\i denote
[R( f )]\i = R\i ( f \i ).
First, we derive a bound for |R( f ) − R\i ( f )|:
1
|Ez Lτ ( f, z ) − Ez Lτ ( f \i , z )|
n
n
n
X
1 X
z
= |Ez
`τ ( f, , k) − Ez
`τ ( f \i , z, k)|
n
k=1
k=1
|R( f ) − R( f )\i | =
≤
n
n
X
1 X
`τ ( f \i , z , k)|
`τ ( f, z , k) −
Ez |
n
k=1
k=1
≤
1 X
Ez
|`τ ( f, z , k) − `τ ( f \i , z , k)|
n k=1
(A.14)
n
¯
(based on Lemma 5 and the definition of ∆)
¯
≤ τ∆
Then, we derive a bound for |R( f ) − R( f )i |:
|R( f ) − R( f )i | = |R( f ) − R( f )\i + R( f )\i − R( f )i |
≤ |R( f ) − R( f )\i | + |R( f )\i − R( f )i |
(based on (A.14))
≤ τ∆¯ + τ∆¯
= 2τ∆¯
Moreover, we derive a bound for |Re ( f ) − Re ( f )i |. Let ẑ i denote the full-size sample
34
(with size n and indexed by i) which replaces the sample z i , it goes to:
m
1 X
1
1 X
|Re ( f ) − Re ( f ) | =
Lτ ( f i , z j ) −
Lτ ( f, z j ) −
Lτ ( f i , ẑi )
mn j=1
mn j,i
mn
1 X
1
|Lτ ( f, z i ) − Lτ ( f i , ẑ i )|
≤
|Lτ ( f, z j ) − Lτ ( f i , z j )| +
mn j,i
mn
i
n
≤
1 X
1 X
|`τ ( f, z i , k) − `τ ( f i , ẑ i , k)|
|Lτ ( f, z j ) − Lτ ( f i , z j )| +
mn j,i
mn k=1
(based on 0 ≤ `τ (GS , z , k) ≤ γ)
γ
1 X
|Lτ ( f, z j ) − Lτ ( f i , z j )| +
≤
mn j,i
m
γ
1 X
|Lτ ( f, z j ) − Lτ ( f \i , z j )| + |Lτ ( f \i , z j ) − Lτ ( f i , z j )| +
≤
mn j,i
m
(based on Lemma 5, and ∆( f i , f \i ) = ∆( f, f \i ) from the definition of stability)
γ
1 X ¯
≤
nτ∆ + nτ∆¯ +
mn j,i
m
=
2(m − 1)τ∆¯ + γ
m
(A.15)
Based on the bounds of |R( f )−R( f )i | and |Re ( f )−Re ( f )i |, we show that R( f )−Re ( f )
¯
∆+γ
satisfies the conditions of McDiarmid Inequality (Theorem 10) with ci = (4m−2)τ
:
m
|[R( f ) − Re ( f )] − [R( f ) − Re ( f )]i | = |[R( f ) − R( f )i ] − [Re ( f ) − Re ( f )i ]|
≤ |R( f ) − R( f )i | + |Re ( f ) − Re ( f )i |
2(m − 1)τ∆¯ + γ
≤ 2τ∆¯ +
m
(4m − 2)τ∆¯ + γ
=
m
35
(A.16)
Also, following the proof of Lemma 11, we can get a bound for ES [R( f ) − Re ( f )]:
ES [R( f ) − Re ( f )] =
=
m
1
1X
ES Ez (L( f, z )) −
L( f, z j )
n
m j=1
m
1X
1
ES ,ẑẑi L( f, ẑ i ) −
ES L( f, z j )
n
m j=1
1
ES ,ẑẑi L( f, ẑ i ) − ES L( f, z i )
n
1
= ES ,ẑẑi L( f, ẑ i ) − ES i L( f i , ẑ i )
n
(A.17)
1
= ES ,ẑẑi L( f, ẑ i ) − L( f i , ẑ i )
n
1
≤ ES ,ẑẑi |L( f, ẑ i ) − L( f i , ẑ i )|
n
1
1
≤ ES ,ẑẑi |L( f, ẑ i ) − L( f \i , ẑ i )| + ES ,ẑẑi |L( f \i , ẑ i ) − L( f i , ẑ i )|
n
n
(based on Lemma 5 and the ∆¯ defined in (15))
≤ τ∆¯ + τ∆¯
=
= 2τ∆¯
Now, we can apply McDiarmid Inequality (Theorem 10):
−2 2
PS [R( f ) − Re ( f )] − ES [R( f ) − Re ( f )] ≥ ≤ exp Pm 2
i=1 ci
(A.18)
Based on (A.16) and (A.17), it goes to
PS R( f ) − Re ( f ) ≥ 2τ∆¯ + ≤ exp
Let δ = exp
−2m 2
2
¯
(4m−2)τ∆+γ
−2m 2
(4m − 2)τ∆¯ + γ 2
!
(A.19)
, we have
r
= (4m − 2)τ∆¯ + γ
ln δ−1
2m
(A.20)
Based on (A.19) and (A.20), there is a probability no more than δ such that
R( f ) − Re ( f ) ≥ 2τ∆¯ +
r
ln δ−1
2m
r
ln δ−1
2m
= 2τ∆¯ + (4m − 2)τ∆¯ + γ
Then, there is a probability at least 1 − δ such that
R( f ) ≤ Re ( f ) + 2τ∆¯ + (4m − 2)τ∆¯ + γ
36
(A.21)
t
u
which gives (16).
Proof of Theorem 9
2 2 2
v n
According to (15), we have ∆¯ ≤ dτρmλα
.
Inserting this bound into (16) gives (18).
t
u
Appendix B. Scalable Decoding with Pruning in Sequence Labeling Tasks
Train time (overall)
BLSTM order1 + SR
BLSTM order2 + SR
BLSTM order3 + SR
Test time (overall)
BLSTM order1 + SR
BLSTM order2 + SR
BLSTM order3 + SR
Chunking
443.13
448.65
459.75
Chunking
10.71
13.64
44.81
English-NER
677.16
705.23
726.58
English-NER
10.08
13.13
20.43
Dutch-NER
484.74
511.08
520.85
Dutch-NER
15.89
26.60
28.66
Table B.6: Timing Results on varying the orders.
SR decoding is implemented by extending the Viterbi decoding algorithm, and
multi-order dependencies are jointly considered. Originally, we should consider all
possible transition states for every position, which means the search space is very large
because there are often too many high order tags. However, in the complete search
space, we may compute many tag-transitions that are almost impossible in the best
output tag sequence. Thus, it is crucial to adopt good pruning strategies to reduce
the search space in decoding to avoid the unnecessary computation of the impossible
state transitions. For scalability, we adopt two pruning techniques to greatly reduce the
search space.
First, we use low order information to prune the search space of high order information. There can be different implementations for this idea. In practice, we find a simple
pruning strategy already works well. We simply use order-1 probability to prune the
tag candidates at each position, such that only top-k candidates at each position are
used to generate the search space for higher order dependencies. In practice we find
top-5 pruning gives no loss on accuracy at all.
Second, we prune the search space according to training set, such that only the
tag dependencies appeared in the training data will be considered as probable tag dependencies in the decoding. We collect a dictionary of the tag dependencies from the
training set, and the first pruning technique is based on this dictionary.
With this implementation, our scalable multi-order decoding performs efficiently
on various real-world NLP tasks and keeps good scalability. The overall training time
and the overall test time on the tasks are shown in the Table B.6.5
5 The “overall” time of BLSTM-SR means that the time of all related orders are added together already.
For example, BLSTM-SR-order3’s overall time already includes the time of order-1 and order-2 models.
37
| 2cs.AI
|
arXiv:1710.02067v1 [cs.IT] 5 Oct 2017
Codes Endowed With the Rank Metric
Elisa Gorla
Institut de Mathématiques
Université de Neuchâtel, Switzerland
[email protected]
Alberto Ravagnani
Department of Electrical and Computer Engineering
University of Toronto, Canada
[email protected]
Abstract
We review the main results of the theory of rank-metric codes, with emphasis on their combinatorial properties. We study their duality theory and MacWilliams identities, comparing
in particular rank-metric codes in vector and matrix representation. We then investigate the
combinatorial structure of MRD codes and optimal anticodes in the rank metric, describing
how they relate to each other.
Introduction
A q-ary rank-metric code is a set of matrices over Fq equipped with the rank distance, which
measures the rank of the difference of a pair of matrices. Rank-metric codes were first studied
in [3] by Delsarte for combinatorial interest.
More recently, codes endowed with the rank metric have been re-discovered for error correction in the context of linear network coding, and featured prominently in the coding theory
literature.
In linear network coding, a source attempts to transmit information packets to multiple
destinations via a network of intermediate nodes. The nodes compute and forward in the
direction of the sinks linear functions of the received packets, rather than simply routing them.
In [1, 9] it was shown that linear network coding achieves the optimal multicast throughput
over sufficiently large alphabets.
Rank-metric codes were proposed in [8, 18] for end-to-end error correction in noisy and
adversarial networks. In this context, as shown in [17], the correction capability of a rankmetric code is measured by a fundamental parameter, called the minimum rank distance of
the code.
1
In this work we survey the main results of the mathematical theory of rank-metric codes,
with emphasis on their combinatorial structure.
In Section 1 we introduce the most important parameters of a rank-metric code, namely,
the minimum distance, the weight distribution, and the distance distribution. We then define
the trace-dual of a linear rank-metric code, and compare the duality theories of codes in matrix
and vector representation. In particular, we show that the former generalizes the latter.
Section 2 is devoted to the duality theory of codes endowed with the rank metric. We
study how combinatorial properties of a linear code relate to combinatorial properties of the
dual code. In particular, we show that the weight distribution of a linear code and the weight
distribution of its dual code determine each other via a MacWilliams-type transformation. We
also show an application of the MacWilliams identities for the rank metric to an enumerative
combinatorics problem.
In Section 3 we study codes that have the largest possible cardinality for their parameters.
These are called Maximum Rank Distance codes (MRD in short) and have very remarkable
properties. We first show the existence of linear MRD codes for all choices of the parameters
and of the field size. Then we prove that the dual of a linear MRD code is MRD. Finally, we
show that the distance distribution of a (possibly non-linear) rank-metric code is completely
determined by its parameters.
Section 4 is devoted to rank-metric anticodes, i.e., sets of matrices where the distance
between any two of them is bounded from above by a given integer. We study how codes and
anticodes relate to each other, deriving in particular an upper bound for the cardinality of any
anticode of given parameters. We conclude the section showing that the dual of an optimal
linear anticode is an optimal anticode.
1
Rank-metric codes
In the sequel q denotes a fixed prime power, and Fq the finite field with q elements. Moreover,
k and m denote positive integers with k ≤ m without loss of generality, and Fqk×m is the space
of k × m matrices over Fq . Finally, for given integers a, b ∈ N we denote by
a
b q
the q-ary binomial coefficient of a and b, which counts the number of b-dimensional subspaces
of an a-dimensional space over Fq . See e.g. [19, Section 1.7] for details.
Definition 1. The rank distance is the function d : Fqk×m ×Fqk×m → N defined by d(M, N ) =
rk(M − N ) for all M, N ∈ Fqk×m .
It is easy to check that d is indeed a distance function on Fqk×m .
Definition 2. A (rank-metric) code over Fq is a non-empty subset C ⊆ Fqk×m . When
|C| ≥ 2, the minimum distance of C is the positive integer
d(C) = min{d(M, N ) | M, N ∈ C, M 6= N }.
A code C is linear if it is an Fq -linear subspace of Fqk×m . In this case its dual code is defined
as
C ⊥ = {N ∈ Fqk×m | Tr(M N t ) = 0 for all M ∈ C} ⊆ Fqk×m ,
where Tr(·) denotes the trace of a square k × k matrix.
The map (M, N ) → Tr(M N t ) ∈ Fq is a scalar product on Fqk×m , i.e., it is symmetric,
bilinear and non-degenerate. In particular, the dual of a linear code is a linear code of
dimension
dim(C ⊥ ) = km − dim(C).
Other fundamental parameters of a rank-metric code are the following.
2
Definition 3. The weight distribution and the distance distribution of a code C are
the collections {Wi (C) | i ∈ N} and {Di (C) | i ∈ N} respectively, where
Di (C) = 1/|C| · |{(M, N ) ∈ C 2 | d(M, N ) = i}|
Wi (C) = |{M ∈ C | rk(M ) = i}|,
for all i ∈ N.
If C is a linear code, then for all P ∈ C there are precisely |C| pairs (M, N ) ∈ C 2 such that
M − N = P . Therefore
X
Di (C) = 1/|C| ·
|{(M, N ) ∈ C 2 | M − N = P }| = Wi (C)
P ∈C
rk(P )=i
for all i ∈ N. Moreover, if |C| ≥ 2 then d(C) = min{rk(M ) | M ∈ C, M 6= 0}.
In [5], Gabidulin proposed independently a different notion of rank-metric code, in which
the codewords are vectors with entries from an extension field Fqm rather than matrices over
Fq .
Definition 4. The rank of a vector v = (v1 , ..., vk ) ∈ Fkqm is the dimension of the linear
spaces generated over Fq by its entries, i.e., rkG (v) = dimFq hv1 , ..., vk i. The rank distance
between vectors v, w ∈ Fkqm is dG (v, w) = rkG (v − w).
One can check that dG is a distance function on Fkqm .
Definition 5. A vector rank-metric code over Fqm is a non-empty subset C ⊆ Fkqm . When
|C| ≥ 2, the minimum distance of C is the positive integer
dG (C) = min{dG (v, w) | v, w ∈ C, v 6= w}.
The code C is linear if it is an Fqm -linear subspace of Fkqm . In this case the dual of C is
defined as
)
(
k
X
⊥
k
vi wi = 0 for all v ∈ C ⊆ Fkqm .
C = w ∈ Fq m |
i=1
The map (v, w) 7→ vi wi is an Fqm -scalar product on Fkqm . Therefore for all linear vector
rank-metric codes C ⊆ Fkqm we have
P
dimFqm (C ⊥ ) = k − dimFqm (C).
Definition 6. The weight distribution and the distance distribution of a vector rankmetric code C are the integer vectors (Wi (C) | i ∈ N) and (Di (C) | i ∈ N) respectively,
where
Wi (C) = |{v ∈ C | rkG (v) = i}|,
Di (C) = 1/|C| · |{(v, w) ∈ C 2 | dG (v, w) = i}|
for all i ∈ N.
There exists a natural way to associate to a vector rank-metric code a code in matrix
representation with the same cardinality and metric properties.
Definition 7. Let Γ = {γ1 , ..., γm } be a basis of Fqm over Fq . The matrix associated to a
vector v ∈ Fkqm with respect to Γ is the k × m matrix Γ(v) with entries in Fq defined by
vi =
m
X
Γ(v)ij γj
for all i = 1, ..., k.
j=1
The rank-metric code associated to a vector rank-metric code C ⊆ Fkqm with respect to Γ is
Γ(C) = {Γ(v) | v ∈ C} ⊆ Fqk×m .
3
Notice that in the previous definition the i-th row of Γ(v) is the expansion of the entry vi
over the basis Γ.
The proof of the following result is standard and left to the reader.
Proposition 8. For every Fq -basis Γ of Fqm the map v 7→ Γ(v) is an Fq -linear bijective
isometry (Fkqm , dG ) → (Fqk×m , d).
In particular, if C ⊆ Fkqm is a vector rank-metric code, then Γ(C) has the same cardinality,
rank distribution and distance distribution as C. Moreover, if |C| ≥ 2 then dG (C) = d(Γ(C)).
In the remainder of the section we compare the duality theories of matrix and vector rankmetric codes, showing that the former generalizes the latter. The following results appear
in [12].
Given an Fqm -linear vector rank-metric code C ⊆ Fkqm and a basis Γ of Fqm over Fq , it is
natural to ask whether the codes Γ(C ⊥ ) and Γ(C)⊥ coincide or not. The answer is negative
in general, as we show in the following example.
Example 9. Let q = 3, k = m = 2 and F32 = F3 [η], where η is a root of the irreducible
primitive polynomial x2 + 2x + 2 ∈ F3 [x]. Let ξ = η 2 , so that ξ 2 + 1 = 0. Set α = (ξ, 2),
and let C ⊆ F232 be the 1-dimensional vector rank-metric code generated by α over F32 . Take
Γ = {1, ξ} as basis of F32 over F3 . One can check that Γ(C) is generated over F3 by the two
matrices
0 1
−1 0
Γ(α) =
,
Γ(ξα) =
.
2 0
0 2
Let β = (ξ, 1) ∈ F232 . We have α1 β1 + α2 β2 = 1 6= 0, and so β ∈
/ C ⊥ . It follows Γ(β) ∈
/ Γ(C ⊥ ).
On the other hand,
0 1
Γ(β) =
,
1 0
and it is easy to see that Γ(β) is trace-orthogonal to both Γ(α) and Γ(ξα). Therefore Γ(β) ∈
Γ(C)⊥ , hence Γ(C)⊥ 6= Γ(C ⊥ ).
Although the duality notions for matrix and vector rank-metric codes do not coincide,
there is a simple relation between them via orthogonal bases of finite fields.
m−1
for all
Let Trace : Fqm → Fq be the map defined by Trace(α) = α + αq + · · · + αq
′
} of Fqm over Fq are called orthogonal
α ∈ Fqm . Bases Γ = {γ1 , ..., γm } and Γ′ = {γ1′ , ..., γm
if Trace(γi′ γj ) = δij for all i, j ∈ {1, ..., m}. It is well-known that every basis Γ of Fqm over Fq
has a unique orthogonal basis Γ′ (see [10], page 54).
Theorem 10. Let C ⊆ Fkqm be an Fqm -linear vector rank-metric code, and let Γ, Γ′ be
orthogonal bases of Fqm over Fq . We have
Γ′ (C ⊥ ) = Γ(C)⊥ .
In particular, C has the same weight distribution as Γ(C), and C ⊥ has the same weight
distribution as Γ(C)⊥ .
′
}. Let M ∈ Γ′ (C ⊥ ) and N ∈ Γ(C). There
Proof. Write Γ = {γ1 , ..., γm } and Γ′ = {γ1′ , ..., γm
⊥
′
exist α ∈ C and β ∈ C such that M = Γ (α) and N = Γ(β). By Definition 7 we have
0 =
k
X
αi βi =
i=1
m
k X
X
i=1 j=1
Mij γj′
m
X
Nit γt =
t=1
m
m X
k X
X
Mij Nit γj′ γt .
i=1 j=1 t=1
Applying the function Trace : Fqm → Fq to both sides of equation (10) we obtain
k X
m X
m
k X
m X
m
X
X
0 = Trace
Mij Nit Trace(γj′ γt ) = Tr(M N t ).
Mij Nit γj′ γt =
i=1 j=1 t=1
i=1 j=1 t=1
4
(1)
Therefore Γ′ (C ⊥ ) ⊆ Γ(C)⊥ . Proposition 8 implies that Γ′ (C ⊥ ) and Γ(C)⊥ have the same
dimension over Fq . Hence the two codes are equal. The second part of the statement follows
from Proposition 8.
Theorem 10 shows that the duality theory of Fq -linear rank-metric codes in matrix representation can be regarded as a generalization of the duality theory of Fqm -linear vector
rank-metric codes. For this reason, in the sequel we only treat rank-metric codes in matrix
representation.
2
MacWilliams identities for the rank metric
This section is devoted to the duality theory of codes endowed with the rank metric. We
concentrate on linear rank-metric codes, and show that the weight distributions of a code
C and its dual code C ⊥ determine each other via a MacWilliams-type transformation. This
result was established by Delsarte in [3, Theorem 3.3] using the machinery of association
schemes, and may be regarded as the rank-analogue of a celebrated theorem by MacWilliams
on the weight distribution of linear codes endowed with the Hamming metric (see [11]). In
this section we present a lattice-theoretic proof inspired by [13, Theorem 27].
Notation 11. We denote by colsp(M ) ⊆ Fkq the Fq -space generated by the columns of a
matrix M ∈ Fqk×m . Given a code C ⊆ Fqk×m and an Fq -subspace U ⊆ Fkq , we let
C(U ) = {M ∈ C | colsp(M ) ⊆ U } ⊆ Fqk×m
be the set of matrices in C whose columnspace is contained in U .
Note that for all M, N ∈ Fqk×m we have colsp(M + N ) ⊆ colsp(M ) + colsp(N ). As a
consequence, if U ⊆ Fkq is an Fq -linear subspace and C ⊆ Fqk×m is a linear code, then C(U ) is
a linear code as well.
We start with a series of preliminary results. In the sequel we denote by U ⊥ the orthogonal
(or dual) of an Fq -vector space U ⊆ Fkq with respect to the standard inner product of Fkq . It
will be clear from context if by “⊥” we denote the trace-dual in Fqk×m or the standard dual
in Fkq .
Lemma 12. Let U ⊆ Fkq be a subspace. The following hold.
1. dim(Fqk×m (U )) = m · dim(U ).
2. Fqk×m (U )⊥ = Fqk×m (U ⊥ ).
Proof.
1. Let s = dim(U ) and V = {(x1 , ..., xk ) ∈ Fkq | xi = 0 for i > s} ⊆ Fkq . There exists
an Fq -isomorphism g : Fkq → Fkq that maps U to V . Let G ∈ Fqk×k be the invertible
matrix associated to g with respect to the canonical basis {e1 , ..., ek } of Fkq , i.e.,
g(ej ) =
k
X
Gij ei
for all j = 1, ..., k.
i=1
The map M 7→ GM is an Fq -isomorphism Fqk×m (U ) → Fqk×m (V ). Property 1 of the
lemma now directly follows from the definition of Fqk×m (V ).
2. Let N ∈ Fqk×m (U ⊥ ) and M ∈ Fqk×m (U ). Using the definition of trace-product one
Pm
sees that Tr(M N t ) = i=1 hMi , Ni i, where h·, ·i is the standard inner product of Fkq ,
and Mi , Ni denote the i-th column of M and N (respectively). Each column of N
belongs to U ⊥ , and each column of M belongs to U . Therefore Tr(M N t ) = 0, hence
Fqk×m (U ⊥ ) ⊆ Fqk×m (U )⊥ . By property 1, the two spaces Fqk×m (U ⊥ ) and Fqk×m (U )⊥ have
the same dimension over Fq . Therefore they are equal.
5
The following result is [12, Lemma 28].
Proposition 13. Let C ⊆ Fqk×m be a linear code, and let U ⊆ Fkq be a subspace of dimension
u over Fq . Then
|C|
|C(U )| = m(k−u) |C ⊥ (U ⊥ )|.
q
Proof. We have C(U )⊥ = (C ∩ Fqk×m (U ))⊥ = C ⊥ + Fqk×m (U )⊥ = C ⊥ + Fqk×m (U ⊥ ), where the
last equality follows from part 2 of Lemma 12. Therefore
|C(U )| · |C ⊥ + Fqk×m (U ⊥ )| = q km .
(2)
On the other hand, part 1 of Lemma 12 gives
dim(C ⊥ + Fqk×m (U ⊥ )) = dim(C ⊥ ) + m · dim(U ⊥ ) − dim(C ⊥ (U ⊥ )).
As a consequence,
q km · q m(k−u)
.
|C| · |C ⊥ (U ⊥ )|
Combining equations (2) and (3) one obtains the proposition.
|C ⊥ + Fqk×m (U ⊥ )| =
(3)
We will also need the following preliminary lemma, which is an explicit version of the
Möbius inversion formula for the lattice of subspaces of Fkq . We include a short proof for
completeness. See [19, Sections 3.7 – 3.10] for details.
Lemma 14. Let P(Fkq ) be the set of all Fq -subspaces of Fkq , and let f : P(Fkq ) → Z be
P
any function. Define g : P(Fkq ) → Z by g(V ) = U⊆V f (U ) for all V ⊆ Fkq . Then for all
i ∈ {0, ..., k} and for any subspace V ∈ P(Fkq ) with dim(V ) = i we have
f (V ) =
i
X
i−u
2
(−1)i−u q (
)
u=0
X
g(U ).
U⊆V
dim(U)=u
Proof. Fix an integer i ∈ {0, ..., k} and a vector space V ∈ P(Fkq ) with dim(V ) = i. We
inductivelyPdefine a function µ : {U ∈ P(Fkq ) | U ⊆ V } → Z by µ(U ) = 1 if U = V , and
µ(U ) = − U(S⊆V µ(S) if U ( V . By definition of g we have
X
X
X
X
X
µ(U )g(U ) =
µ(U )
f (S) =
f (S)
µ(U ) = f (V ),
U⊆V
U⊆V
S⊆U
S⊆V
S⊆U⊆V
where the last equality immediately follows from the definition of µ. Therefore it suffices to
show that for all U ⊆ V we have
µ(U ) = (−1)i−u q (
i−j
2
),
(4)
where u = dim(U ). We proceed by induction on i − u. If i = u then equation (4) is trivial.
Now assume i > u. By definition of µ and the induction hypothesis we have
i
X
X
i−j
i−s i−s
µ(U ) = −
µ(S) = −
(−1) q ( 2 )
s−u q
s=u+1
U(S⊆V
=
−
i
X
i−s
(−1)
s=u+1
=
−
i−u
X
s=0
=
i−s
2
q(
) i−u
i−s q
s
i−u
i−u i−u
)
(
2
(−1) q
+ (−1) q ( 2 )
s q
i−u
(−1)
s
i−u
2
q(
),
where the last equality follows from the q-Binomial Theorem (see [19], page 74).
6
We can now prove the main result of this section, first established by Delsarte in [3,
Theorem 3.3]. A proof for the special case of Fqm -linear vector rank-metric codes using
different techniques can be found in [6].
Theorem 15 (MacWilliams identities for the rank metric). Let C ⊆ Fqk×m be an linear
rank-metric code. For all i ∈ {0, ..., k} we have
Wi (C ⊥ ) =
k
k
X
i−u
1 X
k−u
k−j
i−u
Wj (C)
(−1) q mu+( 2 )
.
k−i q
u q
|C| j=0
u=0
Proof. For all subspaces V ⊆ Fkq define
f (V ) = |{M ∈ C ⊥ | colsp(M ) = V }|,
g(V ) =
X
f (U ) = |C ⊥ (V )|.
U⊆V
By Lemma 14, for any i ∈ {0, ..., k} and for any vector space V ⊆ Fkq of dimension i we have
f (V ) =
i
X
(−1)i−u q (
i−u
2
)
u=0
=
i
X
|C ⊥ (U )|
U⊆V
dim(U)=u
i−u
(−1)
q(
)
i−u
2
u=0
=
X
X
|C ⊥ (T ⊥ )|
T ⊆Fk
q
T ⊇V ⊥
dim(T )=k−u
i
i−u
1 X
i−u
(−1) q mu+( 2 )
|C| u=0
X
|C(T )|,
T ⊆Fk
q
⊥
T ⊇V
dim(T )=k−u
where the last equality follows from Proposition 13. Now observe that
X
f (V )
Wi (C ⊥ ) =
V ⊆Fk
q
dim(V )=i
=
=
=
i
i−u
1 X
(−1)i−u q mu+( 2 )
|C| u=0
i
i−u
1 X
i−u
(−1) q mu+( 2 )
|C| u=0
X
X
⊆Fk
q
V
dim(V )=i
|C(T )|
T ⊆Fk
q
⊥
T ⊇V
dim(T )=k−u
X
T ⊆Fk
q
dim(T )=k−u
i
i−u
1 X
k−u
i−u
(−1) q mu+( 2 )
i−u q
|C| u=0
X
|C(T )|
⊆Fk
q
⊥
V
V ⊇T
dim(V )=i
X
|C(T )|.
T ⊆Fk
q
dim(T )=k−u
On the other hand,
X
|C(T )| =
X
T ⊆Fk
q
T ⊆Fk
q
dim(T )=k−u
dim(T )=k−u
k−u
X
j=0
7
X
S⊆T
dim(S)=j
|{M ∈ C | colsp(M ) = S}|
(5)
k−u
X
=
X
j=0
|{M ∈ C | colsp(M ) = S}|
S⊆Fk
q
T ⊆Fk
q
dim(S)=j
T ⊇S
dim(T )=k−u
k−u
X
=
X
j=0
k−j
u
Wj (C).
(6)
q
Combining equations (5) and (6) one obtains the desired result.
Example 16. Let q = 2, k = 2, m = 3. Let
generated over F5 ∼
= Z/5Z by the matrices
1 0 2
,
0 2 4
C ⊆ Fqk×m be the 2-dimensional linear code
2
1
3 0
.
4 0
We have W0 (C) = 1, W1 (C) = 8 and W2 (C) = 16. Applying Theorem 15 one can easily
compute W0 (C ⊥ ) = 1, W1 (C ⊥ ) = 65 and W2 (C) = 560. Observe that C ⊥ has dimension
6 − 2 = 4, and that 1 + 64 + 560 = 625 = 54 , as expected.
We now present a different formulation of the MacWilliams identities for the rank metric.
The following result is [12, Theorem 31].
Theorem 17. Let C ⊆ Fqk×m be a linear code. For all 0 ≤ ν ≤ k we have
k−ν
X
i=0
ν
|C| X
k−j
⊥
Wj (C )
.
ν−j q
q mν j=0
k−i
Wi (C)
=
ν q
Proof. Proposition 13 gives
X
|C(U )|
|C|
q mν
=
U⊆Fk
q
|C(U )| =
|C ⊥ (U )|.
(7)
U⊆Fk
q
dim(U)=ν
dim(U)=k−ν
Observe that
X
X
|{(U, M ) | U ⊆ Fkq , dim(U ) = k − ν, M ∈ C, colsp(M ) ⊆ U }|
U⊆Fk
q
dim(U)=k−ν
=
X
|{U ⊆ Fkq , dim(U ) = k − ν, colsp(M ) ⊆ U }|
M∈C
=
k
X
X
|{U ⊆ Fkq , dim(U ) = k − ν, colsp(M ) ⊆ U }|
rk(M)=i
k
X
X
i=0
=
i=0
M∈C
M∈C
rk(M)=i
k−ν
X
k−i
k−i
Wi (C)
.
=
ν q
k−ν −i q
(8)
i=0
Using the same argument with C ⊥ and k − ν one shows that
X
U⊆Fk
q
dim(U)=ν
⊥
|C (U )| =
k−ν
X
j=0
k−j
Wj (C )
ν −j
⊥
The result now follows combining equations (7), (8) and (9).
8
q
.
(9)
Remark 18. The two formulations of the MacWilliams identities for the rank metric given in
Theorems 15 and 17 are equivalent. See [6, Corollary 1 and Proposition 3] and [12, Theorem
61] for details.
The next theorem is [2, Theorem 27], and shows that the weight distribution of a linear
code is determined by its parameters, together with the number of codewords of small weight.
We state it without proof. An application of this result will be given in Section 3 (see
Corollary 30).
Theorem 19. Let C ⊆ Fqk×m be a linear code with 1 ≤ dim(C) ≤ km − 1, minimum distance
d = d(C), and dual minimum distance d⊥ = d(C ⊥ ). Let ε = 1 if C is MRD, and ε = 0
otherwise. For all 1 ≤ i ≤ d⊥ we have
k−d
X u u − d⊥ + i − 1
i
Wk−d⊥ +i (C) = (−1)i q (2)
Wk−u (C)
d⊥ − i q
i−1
q
⊥
u=d
+
k
d⊥ − i
i−1−ε
X
q u=0
u
⊥
k − d⊥ + i
)
(
2
q dim(C)−m(d −i+u) − 1 .
(−1) q
u
q
u
In particular, k, m, t, d, d⊥ and Wd (C), . . . , Wk−d⊥ (C) completely determine the weight
distribution of C.
We conclude this section showing how MacWilliams identities for the rank metric can be
employed to solve certain enumerative problems of matrices over finite fields. The following
result is [13, Corollary 52].
Corollary 20. Let I ⊆ {(i, j) ∈ {1, ..., k} × {1, ..., m} | i = j} be a set of diagonal entries.
For all 0 ≤ r ≤ k the number of k × m matrices M over Fq having rank r and Mij = 0 for all
(i, j) ∈ I is
q −|I|
k
k
X
X
r−u
|I|
k−u
k−t
(q − 1)t
(−1)r−u q mu+( 2 )
.
k−r q
u q
t
u=0
t=0
Proof. Define the linear code C = {M ∈ Fqk×m | Mij = 0 for all (i, j) ∈
/ I} ⊆ Fqk×m . Then
dim(C) = |I|, Wt (C) = 0 for |I| < t ≤ k, and
|I|
t
Wt (C) =
(q − 1)
t
for 0 ≤ t ≤ |I|. Moreover, C ⊥ = {M ∈ Fqk×m | Mij = 0 for all (i, j) ∈ I}. Therefore the
number of matrices M ∈ Fqk×m having rank r and Mij = 0 for all (i, j) ∈ I is precisely
Wr (C ⊥ ). The corollary now follows from Theorem 15.
3
MRD codes
In this section we study rank-metric codes that have the largest possible cardinality for their
parameters. We start with a Singleton-type bound for the cardinality of a rank-metric code
of given minimum distance. A code is called MRD if it attains the bound. We then show
that for any admissible choice of the parameters there exists a linear MRD code with those
parameters.
In the second part of the section we study general structural properties of MRD codes.
We first prove in Theorem 26 that the dual of a linear MRD code is MRD. Then we show in
Theorem 28 that the weight distribution of a possibly non-linear MRD code C ⊆ Fqk×m with
0 ∈ C is determined by k, m and d(C). As a corollary, we prove that these three parameters
completely determine the distance distribution of any MRD code. Our proofs are inspired by
the lattice-theory approach to the weight functions of coding theory proposed in [14] and [13].
9
Theorem 21 (Singleton-like bound). Let C ⊆ Fqk×m be a rank-metric code with |C| ≥ 2 and
minimum distance d. Then |C| ≤ q m(k−d+1) .
(k−d+1)×m
denote the projection on the last k − d + 1 rows. Since C has
Proof. Let π : C → Fq
minimum distance d, the map π is injective. Therefore
|C| = |π(C)| ≤ q m(k−d+1) .
A code is MRD if its parameters attain the Singleton-like bound.
Definition 22. We say that C ⊆ Fqk×m is an MRD code if |C| = 1, or |C| ≥ 2 and |C| =
q m(k−d+1) , where d = d(C).
We now prove that for any choice of q, k, m and d there exists a linear rank-metric code
C ⊆ Fqk×m that attains the bound of Theorem 21. This result was first shown by Delsarte in
[3], and rediscovered independently by Gabidulin in [5] and by Kötter and Kschischang in [8]
in the context of linear network coding.
Theorem 23. For all 1 ≤ d ≤ k there exists an Fqm -linear vector rank-metric code C ⊆ Fkqm
with dG (C) = d and dimFqm (C) = k − d + 1. In particular, there exists a linear MRD code
C ⊆ Fqk×m with d(C) = d.
We include an elegant proof for Theorem 23 from [8]. Recall that a linearized polynomial p over Fqm is a polynomial of the form
s
2
p(x) = α0 x + α1 xq + α2 xq + · · · + αs xq ,
αi ∈ Fqm , i = 0, ..., s.
The degree of p, denoted by deg(p), is the largest integer i ≥ 0 such that αi 6= 0. The Fqm vector space of linearized polynomials over Fqm of degree at most s is denoted by Linq (m, s).
It is easy to see that dimFqm (Linq (m, s)) = s + 1.
Remark 24. The roots of a linearized polynomial p over Fqm form an Fq -vector subspace of
Fqm (see [10], Theorem 3.50), which we denote by V (p) ⊆ Fqm in the sequel. Clearly, for any
non-zero linearized polynomial p we have dimFq V (p) ≤ deg(p) by the Fundamental Theorem
of Algebra.
Proof of Theorem 23. Let E = {β1 , ..., βk } ⊆ Fqm be a set of Fq -independent elements. These
elements exist as k ≤ m by assumption. Define the Fqm -linear map
evE : Linq (m, k − d) → Fkqm ,
evE (p) = (p(β1 ), ..., p(βk )) for p ∈ Linq (m, k − d).
We claim that C = evE (Linq (m, k − d)) ⊆ Fkqm is a vector rank-metric code with the desired
properties.
Clearly, C is Fqm -linear. Now let p ∈ Linq (m, k − d) be a non-zero linearized polynomial,
and let W ⊆ Fqm denote the space generated over Fq by the evaluations p(β1 ), ..., p(βk ).
The polynomial p induces an Fq -linear evaluation map p : hβ1 , ..., βk iFq → Fqm . The image
of p is W , and therefore by the rank-nullity theorem we have dimFq (W ) = k − dimFq V (p).
By Remark 24 we conclude dimFq (W ) ≥ k − (k − d) = d. This shows that dG (C) ≥ d. In
particular, as d ≥ 1, the map evE is injective, and the dimension of C is dimFqm (C) = k−d+1.
Combining Proposition 8 and Theorem 21 we obtain dG (C) = d.
The second part of the theorem immediately follows from Proposition 8.
The MRD code construction in the proof of Theorem 23 was later generalized by Sheekey
in [15], introducing a new class of MRD codes.
The reminder of the section is devoted to the structural properties of MRD codes. We
start with a preliminary result from [14, Chapter 7].
10
Lemma 25. Let C ⊆ Fqk×m be an MRD code with |C| ≥ 2 and minimum distance d. For all
subspaces U ⊆ Fkq with u = dim(U ) ≥ d − 1 we have
|C(U )| = q m(u−d+1) .
Proof. As in Lemma 12, define the space V = {(x1 , ..., xk ) ∈ Fkq | xi = 0 for i > u} ⊆ Fkq .
Let g : Fkq → Fkq be an Fq -isomorphism with f (U ) = V . Denote by G ∈ Fqk×k the matrix
associated to g with respect to the canonical basis of Fkq . Define the rank-metric code D =
GC = {GM | M ∈ C}. Clearly, D has the same dimension and minimum distance as C. In
particular, it is MRD. Observe moreover that C(U ) = D(V ).
Now consider the maps
π
π
2
1
,
−→
F(k−u)×m
F(k−d+1)×m
D −→
q
q
where π1 is the projection on the last k − d + 1 coordinates, and π2 is the projection on
the last k − u coordinates. Since d(D) = d, π1 is injective. Since D is MRD, we have
logq (|D|) = m(k − d + 1). Therefore π1 is bijective. The map π2 is Fq -linear and surjective.
Therefore
(k−u)×m
|π2−1 (0)| = |π2−1 (M )| = q m(u−d+1) for all M ∈ Fq
.
Since π1 is bijective and π2 is surjective, the map π = π2 ◦ π1 is surjective. Moreover,
(k−u)×m
|π −1 (0)| = |π −1 (M )| = q m(u−d+1) for all M ∈ Fq
.
The lemma now follows from the identity C(U ) = D(V ) = π −1 (0).
We can now show that the dual of a linear MRD code is MRD. The next fundamental
result is [13, Theorem 5.5].
Theorem 26. Let C ⊆ Fqk×m be a linear MRD code. Then C ⊥ is MRD.
Proof. The result is immediate if dim(C) ∈ {0, km}. Assume 1 ≤ dim(C) ≤ km − 1, and let
d = d(C), d⊥ = d(C ⊥ ). Applying Theorem 21 to C and C ⊥ we obtain
dim(C ⊥ ) ≤ m(k − d⊥ + 1).
dim(C) ≤ m(k − d + 1),
Therefore km = dim(C) + dim(C ⊥ ) ≤ 2mk − m(d + d⊥ ) + 2m, i.e.,
d + d⊥ ≤ k + 2.
(10)
Let U ⊆ Fkq be any Fq -subspace with dim(U ) = k − d + 1. By Proposition 13 we have
|C ⊥ (U )| =
|C ⊥ |
q m(d−1)
|C(U ⊥ )|.
(11)
Since dim(U ⊥ ) = d − 1, by Lemma 25 we have |C(U ⊥ )| = |C|/q m(k−d+1) = 1, where the last
equality follows from the fact that C is MRD. Therefore (11) becomes
|C ⊥ (U )| =
|C ⊥ |
q m(d−1)
=
q km /q m(d−1)
= 1.
q m(d−1)
Since U is arbitrary with dim(U ) = k − d + 1, this shows d⊥ ≥ k − d + 2. Using (10) we
conclude d⊥ = k − d + 2. The theorem now follows from
dim(C ⊥ ) = km − dim(C) = km − m(k − d + 1) = m(k − d⊥ + 1).
11
The proof of Theorem 26 also shows the following useful characterization of linear MRD
codes in terms of their minimum distance and dual minimum distance.
Proposition 27. Let C ⊆ Fqk×m be a linear code with 1 ≤ dim(C) ≤ km − 1. The following
are equivalent.
1. C is MRD,
2. C ⊥ is MRD,
3. d(C) + d(C ⊥ ) = k + 2.
In the remainder of the section we concentrate on the weight and distance distributions of
(possibly non-linear) MRD codes. We start with a result on the weight distribution of MRD
codes containing the zero vector (see [14, Theorem 7.46]).
Theorem 28. Let C be an MRD code with |C| ≥ 2 and 0 ∈ C. Let d = d(C). Then W0 (C) = 1,
Wi (C) = 0 for 1 ≤ i ≤ d − 1, and
Wi (C) =
d−1
X
i−u
(−1)
u=0
i
X
i−u
k
i
i
i−u (i−u
+m(u−d+1) k
(
)
)
2
2
q
+
(−1) q
i q u q
i q u q
u=d
for d ≤ i ≤ k.
Proof. Since 0 ∈ C, we have W0 (C) = 1 and Wi (C) = 0 for 1 ≤ i ≤ d − 1. For all subspaces
V ⊆ Fkq define
f (V ) = |{M ∈ C | colsp(M ) = V }|,
g(V ) =
X
f (U ) = |C(V )|.
U⊆V
Fix 0 ≤ i ≤ k and a vector space V ⊆ Fkq of dimension i. By Lemma 14 we have
f (V ) =
i
X
i−u
(−1)
q(
i−u
2
)
u=0
X
g(U ).
U⊆V
dim(U)=u
Using Lemma 25 and the fact that C is MRD with 0 ∈ C we obtain
1
if 0 ≤ dim(U ) ≤ d − 1,
g(U ) =
q m(u−d+1) if d ≤ dim(U ) ≤ k.
Therefore
f (V ) =
d−1
X
u=0
i−u
(−1)
i
X
i−u
i
i−u (i−u
+m(u−d+1) i
)
(
)
2
2
+
(−1) q
.
q
u q
u q
u=d
The result now follows from the identity
Wi (C) =
X
f (V ).
⊆Fk
q
V
dim(V )=i
Different formulas for the weight distribution of linear MRD codes were obtained in [4]
using elementary methods.
Theorem 28 implies the following [3, Theorem 5.6], which states that the distance distribution of any MRD code is determined by its parameters.
12
Corollary 29. Let C ⊆ Fqk×m be an MRD code with |C| ≥ 2 and minimum distance d. We
have D0 (C) = 1, Di (C) = 0 for 1 ≤ i ≤ d − 1, and
Di (C) =
d−1
X
(−1)i−u q (
u=0
i
X
i−u
i
k
i
) k
+
(−1)i−u q ( 2 )+m(u−d+1)
i q u q
i q u q
i−u
2
u=d
for d ≤ i ≤ k.
Proof. Fix an i with d ≤ i ≤ k. For N ∈ C define C − N = {M − N | M ∈ C}. By definition
of distance distribution we have
X
|C| · Di (C) = |{(M, N ) ∈ C 2 | rk(M − N ) = i}| =
Wi (C − N ).
N ∈C
For all N ∈ C the code C − N is MRD. Moreover, 0 ∈ C − N . The result now easily follows
from Theorem 28.
Corollary 29 shows in particular that the weight distribution of a linear MRD code is
determined by k, m and d(C). Recall from Proposition 27 that an MRD code C ⊆ Fqk×m is
characterized by the property d(C)+d(C ⊥ ) = k +2. We now prove that the weight distribution
of a linear code C with d(C) + d(C ⊥ ) = k + 1 is determined by k, m and dim(C). The following
result is [2, Corollary 28].
Corollary 30. Let C ⊆ Fqk×m be a linear rank-metric code with 1 ≤ dim(C) ≤ km − 1 and
d(C) + d(C ⊥ ) = k + 1. Then
dim(C) 6≡ 0
mod m
d(C) = k − ⌈dim(C)/m⌉ + 1.
and
Moreover, for all d ≤ i ≤ k we have
k
Wi (C) =
i q
i−d(C)
X
u=0
u
i
)
(
2
(−1) q
q dim(C)−m(k+u−i) − 1 .
u q
u
Proof. Assume by contradiction that dim(C) = αm for some α. Applying Theorem 21 to C
and C ⊥ we obtain
d(C) ≤ k − α + 1,
d(C ⊥ ) ≤ α + 1.
(12)
By Proposition 27, the two inequalities in (12) are either both equalities, or both strict inequalities. Since d(C) + d(C ⊥ ) = k + 1 by assumption, they must be both strict inequalities.
Therefore
d(C) ≤ k − α,
d(C ⊥ ) ≤ α,
hence d(C) + d(C ⊥ ) ≤ k, a contradiction. This shows that dim(C) 6≡ 0 mod m.
Now write dim(C) = αm + β with 1 ≤ β ≤ m − 1. Applying again Theorem 21 to C and
C ⊥ one finds
km − αm − β
αm + β
⊥
+ 1 = k − α,
d(C ) ≤ k −
= α + 1.
d(C) ≤ k −
m
m
Since d(C) + d(C ⊥ ) = k + 1, we must have
dim(C)
αm + β
+1=k−
+ 1,
d(C) = k −
m
m
as claimed. The last part of the statement follows from Theorem 19.
13
4
Rank-metric anticodes
This section is devoted to rank-metric anticodes, i.e., rank-metric codes in which the distance
between any two matrices is bounded from above by a given integer δ.
In Theorem 33 we give a bound for the cardinality of a (possibly non-linear) anticode,
using a code-anticode-type bound. We also characterize optimal anticodes in terms of MRD
codes. Then we show that the dual of an optimal linear anticode is an optimal linear anticode.
The main results of this section appear in [12] and [14].
Definition 31. Let 0 ≤ δ ≤ k be an integer. A (rank-metric) δ-anticode is a non-empty
subset A ⊆ Fqk×m such that d(M, N ) ≤ δ for all M, N ∈ A. We say that A is linear if it is
an Fq -linear subspace of Fqk×m .
Example 32. Any A ⊆ Fqk×m with |A| = 1 is a 0-anticode. The ambient space Fqk×m is a
k-anticode. The vector space of k × m matrices over Fq whose last k − δ rows are zero is a
linear δ-anticode of dimension mδ.
In the sequel we work with a fixed integer 0 ≤ δ ≤ k. Moreover, for A, C ⊆ Fqk×m we set
A + C = {M + N | M ∈ A, N ∈ C}.
Theorem 33. Let A ⊆ Fqk×m be a δ-anticode. Then |A| ≤ q mδ . Moreover, if δ ≤ k − 1 then
the following are equivalent.
1. |A| = q mδ .
2. A + C = Fqk×m for some MRD code C with d(C) = δ + 1.
3. A + C = Fqk×m for all MRD codes C with d(C) = δ + 1.
Proof. Let C ⊆ Fqk×m be any MRD code with d(C) = δ + 1. Such a code exists by Theorem 23.
For all M ∈ A let [M ] = M + C = {M + N | N ∈ C}. Then [M ] ∩ [M ′ ] = ∅ for all M, M ′ ∈ A
with M 6= M ′ . Moreover, by definition of MRD code we have |[M ]| = |C| = q m(k−δ) for all
M ∈ A, hence
|Fqk×m | ≥
[
M∈A
[M ] =
X
|[M ]| = |A| · |C| = |A| · q m(k−δ) .
M∈A
Therefore |A| ≤ q mδ , and equality holds if and only if
[
Fqk×m =
[M ] = A + C.
M∈A
A similar argument shows that properties 1, 2 and 3 are equivalent.
Definition 34. We say that a δ-anticode A is (cardinality)-optimal if it attains the bound
of Theorem 33.
Remark 35. Example 32 shows the existence of optimal linear δ-anticodes for all choices of
the parameter δ.
In the remainder of the section we prove that the dual of an optimal linear δ-anticode
is an optimal (k − δ)-anticode. The result may be regarded as the analogue of Theorem 26
in the context of rank-metric anticodes. We start with a preliminary result on the weight
distribution of MRD codes.
Lemma 36. Let C ⊆ Fqk×m be an MRD code with 0 ∈ C, |C| ≥ 2 and d(C) = d. Then
Wd+ℓ (C) > 0 for all 0 ≤ ℓ ≤ k − d.
14
Proof. By Theorem 28, we shall prove the lemma for a given MRD code C ⊆ Fqk×m of our
choice with |C| ≥ 2, minimum distance d, and 0 ∈ C. We will first produce a convenient MRD
code with the prescribed properties.
Let C ⊆ Fkqm be the vector rank-metric code constructed in the proof of Theorem 23, with
evaluation set E = {β1 , ..., βk } and evaluation map evE . Let Γ be a basis of Fqm over Fq . By
Proposition 8, the set C = Γ(C) ⊆ Fqk×m is a linear code with dim(C) = m(k − d + 1) and the
same weight distribution as C. In particular, C is a non-zero linear MRD code of minimum
distance d.
Now we prove the lemma for the MRD code C constructed above. Fix ℓ with 0 ≤ ℓ ≤ k − d.
Define t = k − d − ℓ, and let U ⊆ Fqm be the Fq -subspace generated by {β1 , ..., βt }. If t = 0
we set U to be the zero space. By [10], Theorem 3.52,
Y
pU =
(x − γ)
γ∈U
is a linearized polynomial over Fqm of degree t = k − d − ℓ ≤ k − d, i.e., pU ∈ Linq (n, k − d).
Therefore by Proposition 8 it suffices to prove that evE (pU ) = (pU (β1 ), ..., pU (βk )) has rank d+
ℓ = k −t. Clearly, V (pU ) = U . In particular we have evE (pU ) = (0, ..., 0, pU (βt+1 ), ..., pU (βk )).
We will show that pU (βt+1 ), ..., pU (βk ) are linearly independent overFq . Assumethat there
Pk
Pk
exist at+1 , ..., ak ∈ Fq with i=t+1 ai pU (βi ) = 0. Then we have pU
i=t+1 ai βi = 0, i.e.,
Pk
Pt
i=t+1 ai βi ∈ V (pU ) = U . It follows that there exist a1 , ..., at ∈ Fq such that
i=1 ai βi =
Pk
Pt
Pk
i=t+1 ai βi , i.e.,
i=1 ai βi −
i=t+1 ai βi = 0. Since β1 , ..., βk are independent over Fq ,
we have ai = 0 for all i = 1, ..., k. In particular ai = 0 for i = t + 1, ..., k. Hence
pU (βt+1 ), ..., pU (βk ) are linearly independent over Fq , as claimed.
The following proposition characterizes optimal linear anticodes in terms of their intersection with linear MRD codes.
Proposition 37. Assume 0 ≤ δ ≤ k−1, and let A ⊆ Fqk×m be a linear code with dim(C) = mδ.
The following are equivalent.
1. A is an optimal δ-anticode.
2. A ∩ C = {0} for all non-zero MRD linear codes C ⊆ Fqk×m with d(C) = δ + 1.
Proof. By Theorem 33, it suffices to show that if A ∩ C = {0} for all non-zero MRD linear
codes C ⊆ Fqk×m with d(C) = δ + 1, then A is a δ-anticode.
By contradiction, assume that A is not a δ-anticode. Since A is linear, by definition of
δ-anticode there exists N ∈ C with rk(N ) ≥ δ + 1. Let D be a non-zero linear MRD code with
d(D) = δ + 1 (see Theorem 23 for the existence of such a code). By Lemma 36 there exists
M ∈ D with rk(M ) = rk(N ). There exist invertible matrices A and B of size k × k and m× m,
resp., such that N = AM B. Define C = ADB = {AP B | P ∈ D}. Then C ⊆ Fqk×m is a nonzero linear MRD code with d(C) = δ + 1 and such that N ∈ A ∩ C. Since rk(N ) ≥ δ + 1 ≥ 1,
N is not the zero matrix. Therefore A ∩ C 6= {0}, a contradiction.
We conclude the section showing that the dual of an optimal linear anticode is an optimal
linear anticode.
Theorem 38. Let A ⊆ Fqk×m be an optimal linear δ-anticode. Then A⊥ is an optimal linear
(k − δ)-anticode.
Proof. Let A ⊆ Fqk×m be an optimal linear δ-anticode. If δ = k then the result is trivial.
From now on we assume 0 ≤ δ ≤ k − 1. By Definition 34 we have dim(A) = mδ, hence
dim(A⊥ ) = m(k − δ). Therefore by Proposition 37 it suffices to show that A⊥ ∩ C = {0} for
all non-zero linear MRD codes C ⊆ Fqk×m with d(C) = k − δ + 1. Let C be such a code. Then
dim(C) = m(k − (k − δ + 1) + 1) = mδ < mk.
15
Combining Theorem 26 and Proposition 27 one shows that C ⊥ is a linear MRD code with
d(C ⊥ ) = k − (k − δ + 1) + 2 = δ + 1. By Proposition 37 we have A ∩ C ⊥ = {0}. Since
dim(A) + dim(C ⊥ ) = mδ + m(k − (δ + 1) + 1) = mk, we have A ⊕ C ⊥ = Fqk×m . Therefore
{0} = (Fqk×m )⊥ = (A ⊕ C ⊥ )⊥ = A⊥ ∩ C. This shows the theorem.
References
[1] R. Ahlswede, N. Cai, S.-Y.R. Li, R.W. Yeung, Network information flow. IEEE Transactions on Information Theory 46 (2000), 4, pp. 1204 – 1216.
[2] J. De la Cruz, E. Gorla, H. H. López, A. Ravagnani, Rank distribution of Delsarte codes.
Designs, Codes and Cryptography (to appear).
[3] P. Delsarte, Bilinear forms over a finite field, with applications to coding theory. Journal
of Combinatorial Theory A, 25 (1978), 3, pp. 226 – 241.
[4] J. G. Dumas, R. Gow, G. McGuire, J. Sheekey, Subspaces of matrices with special rank
properties. Linear Algebra and Its Applications, 433 (2010), pp. 191 – 202.
[5] E. Gabidulin Theory of codes with maximum rank distance. Problems of Information
Transmission, 1 (1985), 2, pp. 1 – 12.
[6] M. Gadouleau, Z. Yan MacWilliams Identities for Codes with the Rank Metric.
EURASIP Journal on Wireless Communications and Networking, 2008.
[7] E. Gorla, A. Ravagnani, Subspace codes from Ferrers diagrams. Journal of Algebra and
Its Applications, 16 (2017), 7.
[8] R. Kötter, F. R. Kschischang, Coding for Errors and Erasures in Random Network
Coding. IEEE Transactions on Information Theory, 54 (2008), 8, pp. 3579 – 3591.
[9] S.-Y.R. Li, R.W. Yeung, N. Cai, Linear network coding. IEEE Transactions on Information Theory, 49 (2003), 2, pp. 371 – 381.
[10] R. Lidl, H. Niederreiter, Finite Fields. Addison-Wesley Publishing Company (1983).
[11] F. J. MacWilliams, A Theorem on the Distribution of Weights in a Systematic Code.
Bell System Technical Journal, 42 (1963), 1, pp. 79 – 94.
[12] A. Ravagnani, Rank-metric codes and their duality theory. Designs, Codes and Cryptography, 80 (2016), 1, pp. 197 – 216.
[13] A. Ravagnani, Duality of codes supported on regular lattices, with an
application
to
enumerative
combinatorics.
Submitted.
Online
preprint:
https://arxiv.org/abs/1510.02383.
[14] A. Ravagnani, Properties and Constructions of Codes with the Rank and the Subspace
Metric. PhD thesis, Université de Neuchâtel, 2016.
[15] J. Sheekey, A New Family of MRD Codes. Advances in Mathematics of Communications,
10 (2016), 3.
[16] D. Silva, F. R. Kschishang, On metrics for error correction in network coding. IEEE
Transactions on Information Theory, 55 (2009), 12, pp. 5479 – 5490.
[17] D. Silva, F. R. Kschischang, On metrics for error correction in network coding. IEEE
Transactions on Information Theory, 55 (2009), 12, pp. 5479 – 5490.
[18] D. Silva, E. S. Rogers, F. R. Kschishang, R. Koetter, A Rank-Metric Approach to Error Control in Random Network Coding. IEEE Transactions on Information Theory, 54
(2008), 9, pp. 3951 – 3967.
[19] P. Stanley, Enumerative Combinatorics, vol. 1, Cambridge Stud. Adv. Math., vol. 49.
Cambridge University Press (2012).
16
| 7cs.IT
|
arXiv:1711.11487v1 [cs.SY] 30 Nov 2017
FRAPpuccino: Fault-detection through Runtime Analysis of Provenance
Xueyuan Han, Thomas Pasquier, Tanvi Ranjan, Mark Goldstein, Margo Seltzer
Harvard University
Abstract
We present FRAPpuccino (or FRAP), a provenancebased fault detection mechanism for Platform as a Service (PaaS) users, who run many instances of an application on a large cluster of machines. FRAP models,
records, and analyzes the behavior of an application and
its impact on the system as a directed acyclic provenance
graph. It assumes that most instances behave normally
and uses their behavior to construct a model of legitimate
behavior. Given a model of legitimate behavior, FRAP
uses a dynamic sliding window algorithm to compare a
new instance’s execution to that of the model. Any instance that does not conform to the model is identified
as an anomaly. We present the FRAP prototype and experimental results showing that it can accurately detect
application anomalies.
1 Introduction
Platform as a service (PaaS) clouds have become increasingly popular for their efficient use of computational resources, providing users with an abstracted environment
on which to easily deploy customized applications. Various market research companies estimate the growth of
the PaaS market at about 30% annually for the next few
years [28, 36]. However, PaaS cloud applications face
two major challenges: 1) As an increasing number of
businesses, enterprises, and organizations adopt cloud
computing, cloud applications inevitably become a major target of cyber-attacks. For example, a recent DDoS
attack against a top security blogger delivered by hijacked botnet was so aggressive that Akamai had to cancel the account [19]. According to the RightScale 2017
State of the Cloud Report [38], cloud security remains
one of the top 5 challenges among cloud users; 2) PaaS
clouds make it possible to build large-scale applications
that can serve millions of users. A run-time fault introduced in an application can potentially render it useless.
For example, a simple bug can repeatedly crash a server,
making the service appear unavailable ( § 4).
We introduce FRAPpuccino (Fault-detection through
Runtime Analysis of Provenance or FRAP for short),
a fault/intrusion detection framework. FRAP detects
anomalous behavior of cloud applications by using runtime provenance data to model correct program execution and detecting deviations from the model to identify
potentially malicious behavior. Provenance describes
system behavior as a labelled directed acyclic graph
(DAG) representing interactions between system-level
entities (e.g., file, sockets, pipes), activities (i.e., processes) and agents (i.e., users, groups). Provenance data
can be abundant, so FRAP implements a streaming algorithm, using a dynamic sliding window, to avoid storing
this data. From each running instance, the algorithm extracts a feature vector, which is a projection of the graph
as a point into an n-dimensional space. We assume that
most instances exhibit legitimate behavior so that clustering on these features will clearly divide the instances into
good and bad sets. Thus, our application model includes
two parts: the extracted features and the parameters of
the clusters. Once FRAP has constructed such a model,
it monitors program executions, extracts features from
them, and reports any instances whose features deviate
significantly from good behavior.
Unlike most behavioral-based intrusion/fault detection
systems [29] that rely on system-call usage [14, 16, 41,
43, 47] to profile legitimate application behavior, FRAP
uses provenance data that provides a more comprehensive view of program activities, including their effects on
the underlying system. Prior research has shown that understanding the context of a program’s execution, which
the provenance DAG provides, leads to greater accuracy
in detecting program anomalies [15, 40, 47]. Moreover,
the provenance records provide data that can be analyzed to assist in root cause analysis, ideally providing
actionable information. Our use of end-to-end provenance capture to detect intrusions or faults differs from
other end-to-end tracing approaches in two major ways:
1) by using runtime graphical and statistical analysis on
provenance DAGs of normal instances of an application,
FRAP requires no application instrumentation or annotation, while systems such as Pip [37] need developerprovided specifications of expected behavior; 2) FRAP
analyzes interactions between potentially all executing
applications and the system as naturally presented by
provenance DAGs, while systems such as Magpie [6]
and SpectroScope [39] use event logs, which represent a
carefully curated subset of system activity, which may or
may not capture the key actions. Our goal is to show an
alternative approach and evidence of its efficacy in tackling a long-standing problem of intrusion/fault detection.
The contributions of this work are: 1) a novel approach
that combines provenance and graphical and statistical
analysis to model the behavior of cloud applications; 2)
a dynamic sliding window algorithm that allows efficient
processing of large provenance data to achieve online detection; and 3) an implementation of our framework with
demonstrated accuracy.
2 Background
We build FRAP using two existing open-source tools: 1)
CamFlow [33, 34], a state-of-the-art provenance capture
system; and 2) GraphChi [2, 26], a highly-efficient graph
processing framework. However, the concepts are not
tied to either implementation.
CamFlow [34]: Provenance records the chronology of
ownership, change, and movement of an object or a resource. We use provenance data to understand the interactions between the monitored application, other applications, and the underlying operating system. There are
many provenance capture systems available, including
PASS [31], Hi-Fi [35], Linux Provenance Module [7],
and CamFlow [34]. We chose to use CamFlow, because
it tracks multiple applications, their interactions with the
system, and their interactions with each other. Moreover, it both limits the amount of information captured
(i.e., you can specify which applications to trace) and
ensures completeness by propagating capture to any programs that a traced application invokes. The capture is
built upon Linux Security Module (LSM) hooks that provide completeness guarantees [13, 17, 22]. CamFlow
also provides a facility to conveniently stream the provenance data captured through messaging middleware such
as MQTT[3], RabbitMQ[4], or Apache Flume[1].
GraphChi [26]: Provenance data is naturally represented as a DAG in which each node represents an entity,
an activity, or an agent, and each directed edge represents an interaction between two nodes. For example,
a file was generated by a process, or a process used a
packet. Our framework uses type information in a provenance record as labels to construct a labeled DAG. DAGs
can be efficiently processed by graph processing engines.
We chose GraphChi, a vertex-centric graph processing
model, to generate program models and to detect anomalies. GraphChi uses a parallel sliding window method to
achieve efficient computation of vertices.
3 The FRAPpuccino Framework
The FRAP framework consists of three stages: 1) the
learning stage determines the size of the dynamic sliding
window and creates a model of correct program behavior;
2) the detection stage periodically compares instances of
a program’s execution with the model, notifying the user
upon detection of unusual program behavior; and 3) the
revision stage improves the model by incorporating additional information when we encounter a false positive.
FRAP starts with the learning stage, iterates through
the detection stage until it needs to revise the model, and
transitions back to the detection stage after the revision.
3.1 Learning Stage
The learning stage analyzes the provenance DAG of each
instance of a program, creating a model to describe its
legitimate behavior. Provenance data can grow infinitely
large, making it impossible to analyze them as a whole.
However, past research [16, 21] has shown that a program usually has a limited set of interactions with the
system (e.g., writing to a file, sending a packet), repeating them in different orders as it executes. Therefore, we
claim that one can learn most of a program’s behavior
from a subset of its provenance data. We begin with an
overview of our approach, followed by a more detailed
discussion of each step. FRAP uses a simple but effective
algorithm to determine the number of consecutive provenance records it needs to examine to create a program
model, which is the dynamic sliding window size (used in
the detection stage). Using this subset of records, FRAP
transforms the provenance DAG into a multidimensional
numerical feature vector. We construct this feature vector in three steps: First, we run a label propagation algorithm that constructs a label for each node, representing the structure of the graph around the node. Second,
we count the number of instances of each unique label.
Third, we construct a feature vector consisting of all the
label counts. A feature vector is therefore the result of dimensionality reduction, abstracting a program’s behavior
into numerical values. Finally, FRAP clusters the feature vectors with the goal of grouping all well-behaved
instances together and leaving badly behaved instances
in different clusters. Thus, our model consists of the feature vectors of well behaved instances and the parameters
(centroid and radius) of the clusters in which they reside.
Determining the Dynamic Sliding Window Size: The
goal of finding a dynamically sized window is to minimize the number of provenance records needed to characterize an application. We determine the size during
capture by maintaining two counts. The first counts every edge examined until we declare a window size. We
0, 1 c 2 a
0, 3 f
f
0
a
3
3, 0 f
3, NULL
f
c
2, NULL
2, 0 a 1 b 2 b
2
b
b
2, 2 b
2, 1 d
1
1, 2 b 2 d
1, 0 c
3
3, 0 f
3, NULL
d
2
(a) Before relabeling
0, 1 c 2 a → 4
1, 2 b 2 d → 7
2, NULL → 8
2, 2 b → 11
3, 0 f → 13
Relabeling Map
0, 3 f → 5
1, 0 c → 6
2, 0 a 1 b 2 b → 9
2, 1 d → 10
3, NULL → 12
4, 5 → 14
7, 6 → 15
8, 9 → 16
11, 10 → 17
13, 12 → 18
(b) Populating Relabeling map
18, 14
18, NULL
f
14, 15 16
18
14
14, 18 18
a
f
c
b
16, NULL
15, 16 17
18
16
15
16, 14 15 17
15, 14
b
d
18, 14
18, NULL
17, 16
17
17, 15
(c) After relabeling
Figure 1: (a) Before the first iteration, we label each node with its own
label, its neighboring nodes’ labels, and its incident edges’ labels. (b)
Each node inserts its sorted in-edge neighbor label list and its sorted
out-edge neighbor label list into the relabeling map, and gets a new
label based on these two label lists. In-edge node relabeling is shown
in red, out-edge node relabeling is shown in blue, and final relabeling
is shown in black. (c) After relabeling, nodes have both new identities
and new labels while edge labels are unchanged.
calculate the second by the following process: We examine each edge and assign it a triple consisting of the original edge type and the types of each vertex. We count the
number of edges processed until we encounter a triple
we have never seen before, at which point we reset the
count to 0. If the first counter reaches an implementationdefined threshold (§ 4), or if the second counter reaches a
user-defined threshold, we set the dynamic window size
to be the value of the first counter.
Generating a Program Model: To generate a model,
FRAP computes a feature vector for each instance, clusters those vectors, and discards the vectors in isolated
clusters. FRAP generates a feature vector, using the
following iterative algorithm: relabel each vertex in the
DAG using the current vertex label and the labels of its
incoming and outgoing neighbors. During the first iteration, we also incorporate the labels of a vertex’s incident edges ( Fig. 1), but need not include these in later
iterations, because that information is already encoded
in the labels of the neighboring vertices. After each iteration, new labels encode longer sequences of interactions between the program and the system. FRAP generates a vector containing counts of all seen labels, including the ones from previous iterations, and then clusters
the vectors from all instances. We empirically determine
that four iterations produces the best results. To cluster,
FRAP uses symmetric Kullback-Leibler divergence [25]
or Kullback-Leibler Distance (KLD) with back-off probability [30] as the distance metric to measure the similarity between two feature vectors. KLD has been used, for
example, in statistical language modeling [12] and text
categorization [8]. We use two applications of K-means
clustering: first we cluster on distances between feature
vectors, which helps us select K for the second K-means
clustering, which computes the actual model.
We assume that clusters containing many vectors represent legitimate behavior and want to discard feature
vectors in clusters isolated from these good clusters.
Wagstaff et al. [45] have shown that cluster accuracy
improves when additional information is available to the
problem domain. We use our assumption that most instances are well-behaved as this additional information.
Specifically, we hypothesize that there exists an observable difference between inter-cluster and intra-cluster
distances. First we set K equal to the total number of instances we are analyzing and run K-means clustering on
the pairwise KL distance between each pair of instances.
We then set K to the number of populated clusters and
run a second K-means clustering on the feature vectors
themselves. This produces our model consisting of the
set of feature vectors in clusters containing more than
one instance and the parameters of those clusters (e.g.,
centroid and radius).
3.2 Detection Stage
FRAP monitors an instance by taking its provenance data
from a window of execution, generating a feature vector as in § 3.1, and checking whether this vector fits into
any clusters by comparing the distance between the feature vector and the centroids of the clusters. An instance
is considered abnormal if it does not fit into any of the
model’s clusters.
Dynamic Sliding Window: FRAP uses a dynamic
sliding window approach to continuously monitor an instance while it runs uninterrupted. FRAP only stores and
analyzes the provenance DAG within this window. Once
Example Detection Algorithm (Window Size = 4)
0 1 2 3 4 5
provenance data
6 7 8
←−−−−−−−−
streaming in
(a) Learning Stage
0 1 2 3 4 5 6
provenance data
7 8 9
←−−−−−−−−
streaming in
(b) Detection Stage (Right after a tentative model is generated)
0 1 2 3 4 5 6 7
provenance data
8 9 10
←−−−−−−−−
streaming in
(c) Detection Stage (After one run of the detection algorithm)
Figure 2: Black rectangles represent the provenance records in the window. White ones to the left of the arrow are captured but not yet processed, and to the right will be generated in the future.
it determines that the DAG is part of a normal program
execution, FRAP can safely discard the data. The size of
the window is determined in the learning stage (§ 3.1).
As shown in Fig. 2, records 0, 1, 2, 3 in 2a are used to
generate a model while new provenance data (records
4 and 5) are streaming in. The window slides each
time FRAP finishes running the detection algorithm (described in the next paragraph), to include new records for
the next round of analysis ( 2b and 2c). Since the window slides incrementally, each record appears in many
different windows, maintaining a holistic view of program execution. New records may contain vertices related to ones processed in previous windows. This overlapping preserves a vertex’s n-hop neighbors in which n
becomes larger as the overlap increases. This means that
a larger number of iterations becomes more meaningful,
since longer sequences of program-system interactions
can be preserved.
Detection Algorithm: Given a feature vector of a window of execution, FRAP uses the same distance metrics
(i.e., KLD) to see if the instance lies in an existing cluster
(i.e., the distance between it and the centroid is smaller
than the cluster’s radius). If it does not lie in an existing
cluster, FRAP re-runs (the second) K-means clustering,
and only then reports the instance as an anomaly if it
still does not lie within a cluster of legitimate executions.
Since the detection algorithm is entirely local between an
instance and the model, the computation is parallelizable
and scalable.
3.3 Revision Stage
FRAP identifies and collects false-positive instances and
re-clusters, including them in the model. This reclustering is fast, which is important, because instances are
running concurrently. A fast transition from the revision
stage to the detection stage ensures maximal overlap be-
tween the windows before and after revision (§ 3.2).
Our current false-positive identification is somewhat
crude. We assume that most instances are correct and
revise the model only when detecting a large number
of abnormal instances. However, this assumption does
not always hold. For example, a corrupted database may
cause many client requests to fail. FRAP should not learn
such behavior and include it in the model. Another problematic scenario is when a program has normal behavior
(A, B, C), and the model contains only one behavior (A).
When many instances are displaying behavior B and one
instance is displaying C, then only B will be included
in the model while that one instance will be considered
unusual. Developing a more principled approach to this
problem is left for future work.
4 Implementation and Results
We describe our implementation [20], experimental results and future work.
Implementation-specific Window Size:
When we
constantly receive unique provenance records, our
implementation-defined threshold determines the window size (§ 3.1). GraphChi uses a parallel sliding window method to perform computation on the vertices inside the window in memory, write the results to disk, and
move the window to load the next set of vertices to memory. To avoid high-latency I/O operations, the learning
stage takes the provenance data of the maximum size of
the memory for each instance and allows GraphChi to
perform computation on all vertices in-memory.
Program Models and Feature Vectors: GraphChi
performs vertex relabeling to generate a feature vector
for each instance. Our implementation uses a global
map to achieve consistent relabeling. We need to handle GraphChi’s asynchronous model of computation, in
which an update of a vertex label is immediately visible to its neighbors in the same iteration. This means if
a vertex runs its computation after one of its neighbors
updates itself, it will take that neighbor’s updated label,
which should be used in the next iteration. We solve this
problem by alternating the computation with an update
phase and a swap phase. In the update phase (which
consists of one iteration), each vertex computes its new
label without broadcasting it to its neighbors. Therefore,
within the same iteration, all vertices take their neighboring vertices’ labels from the previous update phase. The
next iteration is the swap phase, where all vertices broadcast their new labels to their neighbors so that in the next
update phase, they can all read the latest labels.
Preliminary Results: We conducted a number of experiments to see if FRAP is able to capture instances with
unusual behavior in a pool of well-behaved instances.
We used both Hellinger distance [32] and Euclidian distance, in addition to KLD, to see how different similarity
Metrics
KLD
Hellinger
Euclidean
Captured Bad
Instance
(Learning)
Yes
No
Yes
Captured Bad
Instance
(Detection)
Yes
No
Yes
Table 1: This experiment runs 10 clients sending requests to the server,
one of which causes the server to behave abnormally during learning.
The same bad behavior occurs again during detection.
metrics affect FRAP’s performance. Table 1 shows the
results of one of these experiments. In this experiment,
we set up a Ruby server in a simulated cloud environment. The server handles requests from multiple clients,
and causes an out-of-memory server crash[18] – a known
system level Ruby vulnerability – for certain URLs.
We see that all but Hellinger distance are able to identify the badly behaved Ruby instance. Hellinger distance
does not work well, because it always produces values
between 0 and 1, which make it difficult for K-means
clustering to create meaningful clusters. Hellinger consistently placed all instances in a single cluster. After manual inspection, we also discovered that while
Euclidean distance successfully identified the badly behaved instance, it mistakenly considered two normal instances as abnormal during the learning stage (i.e., false
negatives), resulting in a slightly less accurate model
than KLD. Appendix A provides a reference to the current prototype and other experimental datasets.
Future Work: We plan to develop more sophisticated
ways to identify false-positives as outlined in § 3. We
also want to further optimize our current implementation and identify situations in which more sophisticated
learning algorithms will improve accuracy. One important area of optimization is our global relabeling map.
Since GraphChi processes vertices in parallel, we need to
maintain map consistency and avoid race conditions. Our
current implementation is a single-point of contention
and can easily be a performance bottleneck. We also
have to garbage collect the map to keep its size manageable. We also want to experiment other algorithms to improve our clustering. For example, Principle Component
Analysis [23] might help us further reduce dimensionality of feature vectors and discard misleading features.
Our current prototype does not have all the pieces integrated, so we used manual intervention to achieve the
end to end pipeline. In particular, we do not directly
stream the provenance to the analyzer. Instead, we capture the provenance and then run the analyzer over the
provenance stream. Additionally, we have not yet integrated the revision stage. These will both be available in
the next release of the software.
5 Related Work
A number of systems use sequences of system calls to detect program/system anomalies. pH [43] uses temporally
proximate system call sequences to model the behavior
of a program. More recent systems have proposed more
advanced analyses. For example, MaMaDroid [27] uses
static program analysis to obtain a program’s call graph
and dynamically builds a Markov-chain [24] model of
the graph during runtime. The feature vector of their
program model consists of the probabilities of each state
transition in the Markov chain. CMarkov [47] includes
calling context of system calls when performing static
program analysis, which further refines their Markov
model. FRAP uses provenance to build a program model
without the help of static program analysis. Its relabelling mechanism concisely encodes system call sequences and their contexts. Moreover, FRAP can detect
anomalies in complex applications composed of multiple
processes including distributed ones by capturing, aggregating, and analyzing provenance data from multiple machines. It is not restrained by per-process system call sequences. We leave comparing FRAP performance with
that of other detection systems for future work.
Similar to FRAP, systems such as Magpie [6], Pinpoint [11], Pip [37] and SpectroScope [39] use end-toend tracing to detect anomalies that could indicate bugs.
However, unlike FRAP, they all capture request flows
within distributed systems and analyze event logs either on a per-event basis or on whole paths. More importantly, although these systems are able to infer bugs,
they are mainly designed to diagnose performance problems, not intrusions, in distributed systems. Moreover,
Pinpoint and Pip require manually annotating applications, an error-prone and significant burden on developers, while FRAP can analyze all applications as long as
the underlying system captures provenance.
For systems that do not have detailed end-to-end tracing capabilities, some black-box diagnosis techniques
have been proposed, e.g., using message send/receive
events to deploy black-box performance debugging [5].
Our graph analysis is related to work on graph kernels
[44]. They are widely used in studying relationships between structured graphs [44]. In particular, our relabeling algorithm is based on the subtree Weisfeiler-Lehman
graph kernel [42] and is a variation of the WeisfeilerLehman test of isomorphism [46].
6 Conclusion
We present a novel approach to detecting unusual behavior in programs running on PaaS clouds and demonstrate
its usability via our implementation. We believe current advances in provenance capture systems open a new
landscape for research in cloud computing and computer
systems.
Discussion Topics
We assume that the provenance data we capture are trustworthy and that attackers cannot modify provenance data
to mask the application’s execution trail. What should we
do if this assumption does not hold in practice? Our system requires mitigation techniques to guard against nontrustworthy provenance data or detect provenance data
tampering as a different form of intrusion. This problem
has been explored in the literature [7, 9], and Zhou et al.
[48] presented a solution to secure network provenance.
Can we simply “plug-and-play” those mechanisms in our
system? How do we comprehensively secure various
sources of provenance used in our system?
We propose that user involvement can help build a better model by allowing users to identify false-positives.
What should we do to provide users with meaningful
provenance information to assist their judgement? One
possible solution is to apply differential provenance [10]
to explain the sources of anomalies. However, differential provenance has only been applied to network provenance. How do we apply this technique to other domains?
There are a variety of intrusion detection systems
(IDS) for the cloud environment. Modi et al. [29] categorized them into eight different techniques, identifying both their strengths and weaknesses. FRAP is a
behavioral-based detection system, but unlike other systems in this category, it uses provenance to model the behavior of an application. From what aspects does FRAP
work better than other behavioral-based detection systems and than other cloud IDS’s at large? What kinds of
intrusions are intrinsically hard for FRAP to detect but
easy for other IDS’s?
Acknowledgements
We thank Daniel Margo for inspiring discussions of
graph processing frameworks. This work was supported
by the US National Science Foundation under grant SSI1450277 End-to-End Provenance.
References
[1] Apache flume. https://flume.apache.org.
[2] Graphchi. https://github.com/GraphChi .
[7] BATES , A. M., T IAN , D., B UTLER , K. R., AND M OYER , T.
Trustworthy whole-system provenance for the linux kernel. In
Usenix Security (2015), pp. 319–334.
[8] B IGI , B. Using kullback-leibler distance for text categorization. In European Conference on Information Retrieval (2003),
Springer, pp. 305–319.
[9] B RAUN , U., S HINNAR , A., AND S ELTZER , M. I. Securing
provenance. In HotSec (2008).
[10] C HEN , A., W U , Y., H AEBERLEN , A., Z HOU , W., AND L OO ,
B. T. Differential provenance: Better network diagnostics with
reference events. In Proceedings of the 14th ACM Workshop on
Hot Topics in Networks (2015), ACM, p. 25.
[11] C HEN , Y.-Y. M., A CCARDI , A., K ICIMAN , E., PATTERSON ,
D. A., F OX , A., AND B REWER , E. A. Path-based failure and
evolution management.
[12] D AGAN , I., L EE , L., AND P EREIRA , F. C. Similarity-based
models of word cooccurrence probabilities. Machine learning
34, 1-3 (1999), 43–69.
[13] E DWARDS , A., JAEGER , T., AND Z HANG , X. Runtime verification of authorization hook placement for the linux security modules framework. In Proceedings of the 9th ACM Conference on
Computer and Communications Security (2002), ACM, pp. 225–
234.
[14] F ENG , H. H., G IFFIN , J. T., H UANG , Y., J HA , S., L EE , W.,
AND M ILLER , B. P. Formalizing sensitivity in static analysis for
intrusion detection. In Security and Privacy, 2004. Proceedings.
2004 IEEE Symposium on (2004), IEEE, pp. 194–208.
[15] F ENG , H. H., K OLESNIKOV, O. M., F OGLA , P., L EE , W., AND
G ONG , W. Anomaly detection using call stack information. In
Security and Privacy, 2003. Proceedings. 2003 Symposium on
(2003), IEEE, pp. 62–75.
[16] F ORREST, S., H OFMEYR , S. A., S OMAYAJI , A., AND
L ONGSTAFF , T. A. A sense of self for unix processes. In Security and Privacy, 1996. Proceedings., 1996 IEEE Symposium
on (1996), IEEE, pp. 120–128.
[17] G ANAPATHY, V., JAEGER , T., AND J HA , S. Automatic placement of authorization hooks in the linux security modules framework. In Proceedings of the 12th ACM conference on Computer
and communications security (2005), ACM, pp. 330–339.
[18] G AZIEV,
A.
How
ruby
2.2
can
cause
an
out-of-memory
server
crash,
2015.
https://evilmartians.com/chronicles/ruby-2_2-oom.
[19] G REENE , T.
Largest ddos attack ever delivered by botnet of hijacked iot devices,
2016.
http://www.networkworld.com/article/3123672/security/largest-ddos-
[3] Mqtt. http://mqtt.org.
[4] Rabbitmq. https://www.rabbitmq.com .
[20] H AN ,
X.
michael-hahn/frap:
v1.1.1,
2017.
DOI:10.5281/zenodo.571444,
https://github.com/michael-hahn/frap .
[5] A GUILERA , M. K., M OGUL , J. C., W IENER , J. L.,
R EYNOLDS , P., AND M UTHITACHAROEN , A. Performance debugging for distributed systems of black boxes. ACM SIGOPS
Operating Systems Review 37, 5 (2003), 74–89.
[21] H OFMEYR , S. A., F ORREST, S., AND S OMAYAJI , A. Intrusion
detection using sequences of system calls. Journal of computer
security 6, 3 (1998), 151–180.
[6] BARHAM , P., D ONNELLY, A., I SAACS , R., AND M ORTIER , R.
Using magpie for request extraction and workload modelling. In
OSDI (2004), vol. 4, pp. 18–18.
[22] JAEGER , T., E DWARDS , A., AND Z HANG , X. Consistency analysis of authorization hook placement in the linux security modules framework. ACM Transactions on Information and System
Security (TISSEC) 7, 2 (2004), 175–205.
[23] J OLLIFFE , I. Principal component analysis. Wiley Online Library, 2002.
[24] K EMENY, J. G., S NELL , J. L., ET AL . Finite markov chains,
vol. 356. van Nostrand Princeton, NJ, 1960.
[25] K ULLBACK , S., AND L EIBLER , R. A. On information and sufficiency. The annals of mathematical statistics 22, 1 (1951), 79–86.
[26] K YROLA , A., B LELLOCH , G. E., G UESTRIN , C., ET AL .
Graphchi: Large-scale graph computation on just a pc. In OSDI
(2012), vol. 12, pp. 31–46.
[27] M ARICONTI , E., O NWUZURIKE , L., A NDRIOTIS , P.,
D E C RISTOFARO , E., ROSS , G., AND S TRINGHINI , G.
Mamadroid: Detecting android malware by building markov
chains of behavioral models. arXiv preprint arXiv:1612.04433
(2016).
[40] S EKAR , R., B ENDRE , M., D HURJATI , D., AND B OLLINENI , P.
A fast automaton-based method for detecting anomalous program
behaviors. In Security and Privacy, 2001. S&P 2001. Proceedings. 2001 IEEE Symposium on (2001), IEEE, pp. 144–155.
[41] S HARIF, M., S INGH , K., G IFFIN , J., AND L EE , W. Understanding precision in host based intrusion detection. In International Workshop on Recent Advances in Intrusion Detection
(2007), Springer, pp. 21–41.
[42] S HERVASHIDZE , N., S CHWEITZER , P., L EEUWEN , E. J. V.,
M EHLHORN , K., AND B ORGWARDT, K. M. Weisfeiler-lehman
graph kernels. Journal of Machine Learning Research 12, Sep
(2011), 2539–2561.
[43] S OMAYAJI , A. B. Operating system stability and security
through process homeostasis. PhD thesis, The University of New
Mexico, 2002.
[28] M ARKO ,
K.
Cloud-first
application
plat[44] V ISHWANATHAN , S. V. N., S CHRAUDOLPH , N. N., K ONDOR ,
forms ?
paas tools to watch in 2017, 2016.
R., AND B ORGWARDT, K. M. Graph kernels. Journal of Mahttp://diginomica.com/2016/12/30/cloud-first-application-platforms-paas-tools-watch-2017/
.
chine Learning Research 11, Apr (2010), 1201–1242.
[29] M ODI , C., PATEL , D., B ORISANIYA , B., PATEL , H., PATEL ,
A., AND R AJARAJAN , M. A survey of intrusion detection techniques in cloud. Journal of Network and Computer Applications
36, 1 (2013), 42–57.
[45] WAGSTAFF , K., C ARDIE , C., ROGERS , S., S CHR ÖDL , S.,
ET AL . Constrained k-means clustering with background knowledge. In ICML (2001), vol. 1, pp. 577–584.
[30] M ORI , R. D. Spoken dialogues with computers. Academic Press,
Inc., 1997.
[46] W EISFEILER , B. J., AND L EMAN , A. A reduction of a graph
to a canonical form and an algebra arising during this reduction,
nauchno–technicheskaja informatsia, 9 (1968), 12–16.
[31] M UNISWAMY-R EDDY, K.-K., H OLLAND , D. A., B RAUN , U.,
AND S ELTZER , M. I. Provenance-aware storage systems. In
USENIX Annual Technical Conference, General Track (2006),
pp. 43–56.
[32] N IKULIN , M. S. Hellinger distance. Encyclopedia of Mathematics (2001).
[33] PASQUIER ,
T.
Camflow/camflowdev,
2017.
DOI:10.5281/zenodo.571427,
https://github.com/CamFlow/camflow-dev.
[34] PASQUIER , T. F.-M., S INGH , J., BACON , J., AND E YERS , D.
Information flow audit for paas clouds. In Cloud Engineering
(IC2E), 2016 IEEE International Conference on (2016), IEEE,
pp. 42–51.
[35] P OHLY, D. J., M C L AUGHLIN , S., M C D ANIEL , P., AND B UTLER , K. Hi-fi: collecting high-fidelity whole-system provenance.
In Proceedings of the 28th Annual Computer Security Applications Conference (2012), ACM, pp. 259–268.
[47] X U , K., T IAN , K., YAO , D., AND RYDER , B. G. A sharper
sense of self: Probabilistic reasoning of program behaviors for
anomaly detection with context sensitivity. In Dependable Systems and Networks (DSN), 2016 46th Annual IEEE/IFIP International Conference on (2016), IEEE, pp. 467–478.
[48] Z HOU , W., F EI , Q., N ARAYAN , A., H AEBERLEN , A., L OO ,
B. T., AND S HERR , M. Secure network provenance. In Proceedings of the Twenty-Third ACM Symposium on Operating Systems
Principles (2011), ACM, pp. 295–310.
A
Availability
The work presented in this paper is open-source and
available for download at
https://github.com/michael-hahn/frap
[36] R ESEARCH , Z. M.
Global public cloud platform as a service (paas) market worth usd 9.12
billion by 2021 - zion market research,
2016.
https://globenewswire.com/news-release/2016/09/26/874593/0/en/Global-Public-Cloud-Platform-as-a-Service-PaaS-Market-worth-U
[37] R EYNOLDS , P., K ILLIAN , C. E., W IENER , J. L., M OGUL ,
J. C., S HAH , M. A., AND VAHDAT, A. Pip: Detecting the unexpected in distributed systems. In NSDI (2006), vol. 6, pp. 115–
128.
[38] R IGHT S CALE.
Largest ddos attack ever delivered by botnet of hijacked iot devices,
2017.
http://assets.rightscale.com/uploads/pdfs/RightScale-2017-State-of-the-Cloud-Report.pdf.
[39] S AMBASIVAN , R. R., Z HENG , A. X., D E ROSA , M., K REVAT,
E., W HITMAN , S., S TROUCKEN , M., WANG , W., X U , L., AND
G ANGER , G. R. Diagnosing performance changes by comparing
request flows. In NSDI (2011), pp. 43–56.
| 3cs.SY
|
Thick, syndetic, and piecewise syndetic subsets of
Fraı̈ssé structures
arXiv:1608.04286v2 [math.DS] 20 Mar 2017
Andy Zucker
March 21, 2017
Abstract
We define and undertake a systematic study of thick, syndetic, and piecewise syndetic subsets of a Fraı̈ssé structure. Each of these collections forms a family in the
sense of Akin and Glasner [AG], and we define and study ultrafilters on each of these
families, paying special attention to ultrafilters on the thick sets. In the process, we
generalize many results of Bergelson, Hindman, and McCutcheon [BHM]. We also discuss some abstract questions about families implicit in the work of Akin and Glasner.
1
Introduction
Let G be a discrete group, and let Fin(G) denote the finite subsets of G. Recall the following
definitions (see either [BHM] or [HS]).
Definition 1.1.
• T ⊆ G is thick if for every E ∈ Fin(G), there is g ∈ G with gE ⊆ T ; equivalently, T
is thick if the collection {T g −1 : g ∈ G} has the finite intersection property,
• S⊆G
if G \ S is not thick; equivalently, S is syndetic if there is E ∈ Fin(G)
Sis syndetic
−1
with g∈E Sg = G,
S
• P ⊆ G is piecewise syndetic if there is E ∈ Fin(G) with g∈E P g −1 thick.
Note the left-right switch from the definitions found in [BHM] or [HS]; the reasons for this
switch will be explained later. There are several equivalent formulations of these definitions.
One such formulation uses the space 2G , which we identify with the space of subsets of G.
G acts on 2G by right shift, i.e. for χA ∈ 2G and g, h ∈ G, we have χA · g(h) = χA (gh). We
now have the following equivalent characterizations.
2010 Mathematics Subject Classification. Primary: 37B05; Secondary: 03C15, 03E15, 05D10, 22F50,
54D35, 54D80, 54H20.
Key words and phrases. Fraı̈ssé theory, topological dynamics, ultrafilters, topological semigroups.
The author was partially supported by NSF Grant no. DGE 1252522.
1
Fact 1.2.
• T ⊆ G is thick iff χG ∈ χT · G,
• S ⊆ G is syndetic iff χ∅ 6∈ χS · G,
• P ⊆ G is piecewise syndetic iff there is syndetic S ⊆ G with χS ∈ χP · G.
Another formulation uses βG, the space of ultrafilters on G. We view G as a subspace of
βG by identifying g ∈ G with the principal ultrafilter containing g. To each subset A ⊆ G,
we can form the clopen set A := {p ∈ βG : A ∈ p}, and these clopen sets form a basis for the
topology of βG. We endow βG with the structure of a compact left-topological semigroup.
Compact left-topological semigroups contain minimal right ideals which are always closed.
In βG, each minimal right ideal is also a minimal (right) G-flow. We now have the following.
Fact 1.3.
• T ⊆ G is thick iff T ⊆ βG contains a minimal right ideal,
• S ⊆ G is syndetic iff S ∩ M 6= ∅ for every minimal right ideal M ⊆ βG,
• P ⊆ G is piecewise syndetic iff P ∩ M 6= ∅ for some minimal right ideal M ⊆ βG.
The primary goal of this paper is to generalize these definitions and characterizations to
the setting of countable first-order structures and their groups of automorphisms. We will
be especially interested in ultrahomogeneous structures, structures where every isomorphism
between finite substructures extends to an automorphism of the whole structure. If K is an
ultrahomogeneous structure with automorphism group G, we endow G with the pointwise
convergence topology. If A ⊆ K is a finite substructure, then the pointwise stabilizer GA :=
{g ∈ G : ∀a ∈ A(g · a = a)} is an open subgroup; these open subgroups form a neighborhood
base for G at the identity. Now suppose that f : A → K is an embedding; we can identify
f with the (non-empty!) set of g ∈ G which extend f . Notice that {g ∈ G : g|A = f }
is a left coset of GA , and we can identify Emb(A, K) with the set of left cosets of GA .
Our generalized notions of thick, syndetic, and piecewise syndetic will describe subsets of
Emb(A, K).
There are three difficulties worth previewing now. First, each finite substructure of K
will give rise to different notions of thick, syndetic, and piecewise syndetic; we will need to
understand how these different notions interact with each other. Second, note that when G
is a countable discrete group, G acts on itself on both the left and the right. The right action
is used in Definition 1.1, and the left action, which is needed to define the right shift action
on 2G , is used in Fact 1.2. However, when G = Aut(K) and A ⊆ K is a finite substructure,
there is a natural left G action on Emb(A, K), but no natural notion of right action. Third,
when G = Aut(K), the space of ultrafilters on G is too fine a compactification to use. The
correct compactificaton to use is the Samuel compactification of G, denoted S(G); we will
need an explicit construction of S(G) to allow us to state something similar to Fact 1.3.
A secondary goal of this paper will be to consider the notions of thick, syndetic, and
piecewise syndetic sets as families, collections of subsets of a set X closed upwards under
2
inclusion. Families have been considered by Akin and Glasner [AG] and implicitly by Brian
[B]. We will need to study the family of thick sets in some detail; along the way, we will
define ultrafilters on families and address some general questions about how these ultrafilters
interact with maps.
2
Background
2.1
Fraı̈ssé structures
Recall that S∞ is the group of all permutations of a countable set X. Given a finite B ⊆ X,
the pointwise stabilizer of B is the set NB = {g ∈ S∞ : ∀ b ∈ B (g(b) = b)}. We can endow
S∞ with the pointwise convergence topology, where a basis of open sets at the identity is
given by the collection {NB : B ∈ Pf in (X)}. This turns S∞ into a Polish group. Note that
each NB is a clopen subgroup of S∞ .
Fix now G a closed subgroup of S∞ . A convenient way to describe the G-orbits of finite
tuples from X is given by the notions of a Fraı̈ssé class and structure.
Definition 2.1. Let L be a relational language. A Fraı̈ssé class K is a class of L-structures
with the following four properties.
1. K contains only finite structures, contains structures of arbitrarily large finite cardinality, and is closed under isomorphism.
2. K has the Hereditary Property (HP): if B ∈ K and A ⊆ B, then A ∈ K.
3. K has the Joint Embedding Property (JEP): if A, B ∈ K, then there is C which embeds
both A and B.
4. K has the Amalgamation Property (AP): if A, B, C ∈ K and f : A → B and g : A → C
are embeddings, there is D ∈ K and embeddings r : B → D and s : C → D with
r ◦ f = s ◦ g.
If D is an infinite L-structure, we write Age(D) for the class of finite L-structures which
embed into D. The following is the major fact about Fraı̈ssé classes.
Fact 2.2. If K is a Fraı̈ssé class, there is up to isomorphism a unique countably infinte
L-structure K with Age(K) = K satisfying one of the following two equivalent conditions.
1. K is ultrahomogeneous: if f : A → B is an isomorphism between finite substructures
of K, then there is an automorphims of K extending f .
2. K satisfies the Extension Property: if B ∈ K, A ⊆ B, and f : A → K is an embedding,
there is an embedding h : B → K extending f .
Conversely, if K is a countably infinite L-structure satisfying 1 or 2, then Age(K) is a
Fraı̈ssé class.
3
Given a Fraı̈ssé class K, we write Flim(K), the Fraı̈ssé limit of K, for the unique structure
K as above. We say that K is a Fraı̈ssé structure if K ∼
= Flim(K) for some Fraı̈ssé class.
Our interest in Fraı̈ssé structures stems from the following result.
Fact 2.3. For any Fraı̈ssé structure K, Aut(K) is isomorphic to a closed subgroup of S∞ .
Conversely, every closed subgroup of S∞ is isomorphic to Aut(K) for some Fraı̈ssé structure
K.
Throughout the rest of the paper, fix
S a Fraı̈ssé class K with Fraı̈ssé limit K. Set G =
Aut(K). We also fix an exhaustion K = n An , with each An ∈ K and Am ⊆ An for m ≤ n.
Write Hn = {gGn : g ∈ G}, where Gn = G ∩ NAn is the pointwise stabilizer of An ; we can
identify Hn with Emb(An , K),Sthe set of embeddings of An into K. Note that under this
identification, we have Hn = N ≥n Emb(An , AN ). For g ∈ G, we often write g|n for gGn ,
and we write in for Gn . The group G acts on Hn on the left; if x ∈ Hn and g ∈ G, we have
g · x = g ◦ x. For m ≤ n, we let inm ∈ Emb(Am , An ) be the inclusion embedding.
Each f ∈ Emb(Am , An ) gives rise to a dual map fˆ : Hn → Hm given by fˆ(x) = x ◦ f .
Proposition 2.4.
1. For f ∈ Emb(Am , An ), the dual map fˆ : Hn → Hm is surjective.
2. For every f ∈ Emb(Am , An ), there is N ≥ n and h ∈ Emb(An , AN ) with h ◦ f = iN
m.
Proof. Item 1 is an immediate consequence of the extension property. For item 2, use
ultrahomogeneity to find g ∈ G with g ◦ f = im . Let N ≥ n be large enough so that
ran(g|n ) ⊆ AN , and set h = g|n .
2.2
Topological dynamics and the Samuel compactification
We now want to investigate the dynamical properties of the group G. Though many of the
following definitions and facts hold for more general topological groups, we will only consider
the case G = Aut(K).
A G-flow is a compact Hausdorff space X equipped with a continuous right G-action
a : X × G → X. We suppress the action a by writing a(x, g) = x · g. If X is a G-flow, a
subflow Y ⊆ X is a closed subspace which is G-invariant. The G-flow X is minimal if it
contains no proper subflow; equivalently, X is minimal iff every orbit is dense. By Zorn’s
lemma, every G-flow contains a minimal subflow.
Given two G-flows X and Y , a G-map f is a continuous map f : X → Y satisfying
f (x · g) = f (x) · g for every g ∈ G. A G-flow X is universal if for every minimal G-flow Y ,
there is a G-map f : X → Y . The following is an important fact of topological dynamics.
Fact 2.5. There is up to G-flow isomorphism a unique G-flow M (G) which is both universal
and minimal. We call M (G) the universal minimal flow of M (G). Any G-map from M (G)
to M (G) is an isomorphism.
One way of exhibiting the universal minimal flow is to consider the closely related notion
of an ambit. A G-ambit is a pair (X, x0 ) with X a G-flow and x0 ∈ X a distinguished
4
point with dense orbit. If (X, x0 ) and (Y, y0 ) are G-ambits, a map of ambits f is a G-map
f : X → Y with f (x0 ) = y0 . Note that there is at most one map of ambits from (X, x0 ) to
(Y, y0 ), and such a map is always surjective.
Fact 2.6. There is up to G-ambit isomorphism a unique G-ambit (S(G), 1) which admits
a map of G-ambits onto every other G-ambit. We call (S(G), 1) the greatest G-ambit. The
space S(G) is the Samuel compactification of G.
Now suppose M ⊆ S(G) is a subflow. If Y is any minimal G-flow, pick any y0 ∈ Y and
form the ambit (Y, y0 ). Let f : S(G) → Y be the unique map of ambits. Then f |M is a
G-map. It follows that M ∼
= M (G).
We now proceed with an explicit construction of S(G). Let f ∈ Emb(Am , An ). The dual
map fˆ extends to a continuous map f˜ : βHn → βHm , where for a discrete space X, βX is
the space of ultrafilters on X. If p ∈ βHn and f ∈ Emb(Am , An ), we will sometimes write
p · f for f˜(p). Form the inverse limit lim βHn along the maps ı̃nm . We can identify G with a
←−
dense subspace of lim βHn by associating to each g ∈ G the sequence of ultrafilters principal
←−
on g|n . The space lim βHn turns out to be the Samuel compactification S(G) (see Corollary
←−
3.3 in [P]).
To see that S(G) is the greatest ambit, we need to exhibit a right G-action on S(G).
This might seem unnatural at first; after all, the left G-action on each Hn extends to a left
G-action on βHn , and the maps ı̃nm are all G-equivariant, giving us a left G-action on S(G).
Indeed, let Gd be the group G with the discrete topology; then the map πm : Gd → Hm
given by πm (g) = gGm extends to a continuous map π̃m : βGd → Hm . As πm = ı̃nm ◦ πn for
n ≥ m, we obtain a map π : βGd → S(G), and the left action just described on S(G) makes
π G-equivariant, where βGd is given its standard left G-action. The problem is that the left
action is not continuous when G is given its Polish topology. However, βGd also comes with
a natural right G-action, and there is a unique way to equip S(G) with a right G-action to
make the map π equivariant. This right G-action on S(G) is continuous. For α ∈ lim βHn ,
←−
g ∈ G, m ∈ N, and S ⊆ Hm , we have
S ∈ αg(m) ⇔ {x ∈ Hn : x ◦ g|m ∈ S} ∈ α(n)
where n ≥ m is large enough so that ran(g|m ) ⊆ An . Notice that if g|m = h|m = f ,
then αg(m) = αh(m) := α · f := λαm (f ). By distinguishing the point 1 ∈ lim βHn with
←−
1(m) principal on im , we endow S(G) with the structure of a G-ambit, and (S(G), 1) is the
greatest ambit (see Theorem 6.3 in [Z]).
Using the universal property of the greatest ambit, we can define a left-topological semigroup structure on S(G): Given α and γ in lim βHn , m ∈ N, and S ⊆ Hm , we have
←−
S ∈ αγ(m) ⇔ {f ∈ Hm : S ∈ α · f } ∈ γ(m).
If α ∈ S(G) and S ⊆ Hm , a useful shorthand is to put α−1 (S) = {f ∈ Hm : S ∈ α · f }.
Then the semigroup multiplication can be written as S ∈ αγ(m) iff α−1 (S) ∈ γ(m). Notice
that for fixed α, αγ(m) depends only on γ(m); indeed, if α ∈ lim βHn , p ∈ βHm , and
←−
S ⊆ Hm , we have S ∈ α · p iff α−1 (S) ∈ p. In fact, α · p = λ̃αm (p), where the map λ̃αm is the
continuous extention of λαm to βHm .
5
By continuity of left-multiplication, subflows of the G-flow lim βHn are exactly the closed
←−
right ideals of the semigroup lim βHn .
←−
Recall that if S is a semigroup, then u ∈ S is an idempotent if u · u = u.
Fact 2.7. In compact left-topological semigroups, minimal right ideals exist, are always
compact, and always contain idempotents. If Y, Y 0 ⊆ lim βHn are minimal right ideals, then
←−
any G-map ϕ : Y → Y 0 is of the form ϕ(y) = αy for some α ∈ Y 0 , and each such G-map is
an isomorphism.
Furthermore, write Y = lim Yn , Y 0 = lim Yn0 , with Ym and Ym0 compact subsets of βHm .
←−
←−
Given α ∈ Y 0 as above, we note that since αy(m) depends only on y(m), the isomorphism
ϕ gives rise to homeomorphisms ϕm : Ym → Ym0 satisfying ı̃nm ◦ ϕn = ϕm ◦ ı̃nm for n ≥ m.
Some quick remarks about notation are in order. Given f ∈ Emb(Am , An ) and S ⊆ Hm ,
we will often want to consider the set (fˆ)−1 (S); instead of “dualling twice,” we instead write
f (S). Note that if x ∈ Emb(An , AN ), then we have x(f (S)) = (x ◦ f )(S), so this notation
is justified. We also remark that whenever we write lim Ym for Ym ⊆ Hm closed, we are
←−
implicitly assuming that ı̃nm (Yn ) = Ym .
3
The family of thick sets
If X is an infinite set, a family S on X, or just (X, S), is a non-trivial collection of subsets
of X closed upwards under inclusion. By non-trivial, we demand that X ∈ S and ∅ 6∈ S. An
S-filter F on X is any filter F on X with F ⊆ S. An S-ultrafilter is any maximal S-filter.
By Zorn, any S-filter can be extended to some S-ultrafilter. Write β(S) for the collection of
S-ultrafilters. We borrow the term family from Akin and Glasner [AG], and the definition
of an S-filter is implicit in [B].
If (X, S) is a family and f : X → Y is a map, we can push forward the family S to the
family f (S) := {B ⊆ Y : f −1 (B) ∈ S}. In a similar way, any S-filter F pushes forward
to an f (S)-filter f (F). We call a map f : (X, S) → Y strong if f −1 (f (A)) 6∈ S whenever
A 6∈ S; we call f regular if f (β(S)) ⊆ β(f (S)), i.e. if the push forward of any S-ultrafilter is
an f (S)-ultrafilter. Note that we always have β(f (S)) ⊆ f (β(S)), since given q ∈ β(f (S)),
any p ∈ β(S) extending the S-filter generated by {f −1 (B) : B ∈ q} will satisfy f (p) = q.
Here is a quick proposition to give some intuition about the definitions.
Proposition 3.1. Any strong map is also regular.
Proof. Suppose f : (X, S) → Y is strong, and let p ∈ β(S). To show f (p) ∈ β(f (S)), let
B ⊆ Y with B ∈ f (S) \ f (p). Then f −1 (B) ∈ S \ p. As p ∈ β(S), there must be A ∈ p with
A ∩ f −1 (B) 6∈ S. As f is strong, we have f (A ∩ f −1 (B)) = f (A) ∩ B 6∈ f (S). As f (A) ∈ f (p),
we have shown that f (p) ∈ β(f (S)).
We will be interested in studying the family Tm of thick subsets of Hm , and in particular understanding the members of β(Tm ). Our definitions of thick, syndetic, and piecewise
syndetic will most closely follow the second set of definitions from the introduction. Indeed,
since G acts on the left on Hm , this induces a right shift action on 2Hm
6
Definition 3.2.
• T ⊆ Hm is thick iff χHm ∈ χT · G,
• S ⊆ Hm is syndetic iff χ∅ 6∈ χS · G,
• P ⊆ Hm is piecewise syndetic iff there is syndetic S ⊆ Hm with χS ∈ χP · G.
Each of these notions is upwards closed, so forms a family. Write Sm and Pm for the
families of syndetic and piecewise syndetic subsets of Hm , respectively. Note that T ⊆ Hm
is thick iff Hm \ T is not syndetic, and vice versa.
Let us briefly return to the setting of the introduction, where G is a discrete group.
Definition 1.1 uses the right G-action on itself; given T ⊆ G, we need the right action to
describe the sets T g −1 for g ∈ G. But the definition of “thick” given there also includes
a more combinatorial characterization, where T is thick iff for every E ∈ Fin(G), there is
g ∈ G with gE ⊆ T . When G = Z for instance, this amounts to saying that thick subsets
of the integers are exactly those subsets containing arbitrarily long intervals. Since this
characterization uses the left G-action on itself, we can hope to generalize it. We now return
to the normal setting of the paper, where G = Aut(K).
Proposition 3.3. T ⊆ Hm is thick iff for every n ≥ m, there is x ∈ Hn with
x ◦ Emb(Am , An ) ⊆ T .
Proof. Suppose T ⊆ Hm is thick, and let gn ∈ G be a sequence of group elements with
χT · gn → χHm . By passing to a subsequence, we may assume that for each n ≥ m, we have
χT · gn (f ) = 1 for every f ∈ Emb(Am , An ). It follows that gn |n ◦ Emb(Am , An ) ⊆ T .
Conversely, suppose T ⊆ Hm is a set so that for every n ≥ m, there is xn ∈ Hn with
xn ◦ Emb(Am , An ) ⊆ T . Find gn ∈ G with gn |n = xn . Then χ· gn → χHm , and T is thick.
~ = {xn : n < ω} with each xn ∈ Hn an exhausting sequence. Writing
Call a sequence X
S
~ :=
Emb(Am , X)
n≥m (xn ◦ Emb(Am , An )), we can say that T ⊆ Hm is thick iff there is an
~ ⊆ T.
exhausting sequence with Emb(Am , X)
We now want something akin to the third set of definitions from the introduction. Before
we can attempt a generalization, we need to understand how the thick families on different
levels interact. The following proposition is implicit in section 4 of [Z] (see Lemma 4.2).
Proposition 3.4. For any f ∈ Emb(Am , An ), we have that fˆ(Tn ) = Tm .
Proof. We need to show that T ⊆ Hm is thick iff f (T ) ⊆ Hn is thick. Assume first that T is
~ is an exhausting sequence with Emb(Am , X)
~ ⊆ T . Then Emb(An , X)
~ ⊆ f (T )
thick and X
~
~
(since Emb(An , X) ◦ f ⊆ Emb(Am , X)).
~ is an exhausting sequence with
For the other direction, suppose f (T ) is thick and X
~ ⊆ f (T ). For each N ≥ m, find `N ≥ N large enough so that for each h ∈
Emb(An , X)
Emb(Am , AN ), there is xh ∈ Emb(An , A`N ) with xh ◦ f = h. Let Y~ be the exhausing
sequence given by yN = x`N ◦ i`NN . Then Emb(Am , Y~ ) ⊆ T .
7
A consequence of Proposition 3.4 is that for n ≥ m, f ∈ Emb(Am , An ), and F a Tn -filter
on Hn , we have that fˆ(F ) is a Tm -filter on Hm . This will be especially important when
f = inm .
If T ⊆ Hm , set T = {p ∈ βHm : T ∈ p}, and set T̃ = {α ∈ lim βHn : T ∈ α(m)}. If
←−
Ym ⊆ βHm T
is closed, then the filter of clopen neighborhoods of Ym is the filter Fm on Hm so
that Ym = T ∈Fm T . Conversely, if Fm is a filter on Hm , then the closed set for Fm is the
T
set T ∈Fm T . The following two theorems are the key to generalizing Fact 1.3
Theorem 3.5. Let Y = lim Yn ⊆ lim βHn = S(G) be compact, with Ym ⊆ βHm for m ∈ N.
←−
←−
Let Fm be the filter of clopen neighborhoods of Ym . Then Y is a minimal right ideal iff each
Fm is a member of β(Tm ).
Theorem 3.6. If f ∈ Emb(Am , An ), then fˆ : (Hn , Tn ) → Hm is regular,
As a consequence of both Theorems 3.5 and 3.6, notice that if Fm ∈ β(Tm ), then there is
a minimal right ideal Y = lim Ym ⊆ S(G) with Fm the filter of clopen neighborhoods of Ym .
←−
In particular, since every thick T ⊆ Hm belongs to some F ∈ β(Tm ), we have the following
corollary, a generalization of Theorem 2.9(d) from [BHM].
Corollary 3.7. T ⊆ Hm is thick iff T̃ ⊆ S(G) contains a minimal right ideal.
The first step is to characterize right ideals of lim βHn . The following is a generalization
←−
of Theorem 8.7 from [Z].
Proposition 3.8. Let Y = lim Yn ⊆ lim βHn be compact, with Ym ⊆ βHm for m ∈ N. Let
←−
←−
Fm be the filter of clopen neighborhoods of Ym . Then Y contains a right ideal iff each Fm is
a Tm -filter.
Proof. First assume that each Ym is thick. For W ⊆ G finite, m ∈ N, and S ∈ Fm , let YW,S
consist of those α ∈ Y such that S ∈ αg(m) for each g ∈ W . Notice that YW,S ⊆ Y is closed,
hence compact.
Claim. First, let us show that YW,S is nonempty. Fix n large enough so that g(Am ) ⊆ An
for each g ∈ W∪ {1G }. For
g ∈ G, set Tg = {f ∈ Hn : f ◦ g|m ∈ S}. We will show that the
T
set X := T1G \
g∈W Tg is not thick. If it were, pick N large enough so that g(An ) ⊆ AN
for each g ∈ W ∪ {1G } and find h ∈ HN so that h ◦ Emb(An , AN ) ⊆ X. But now for each
N
g ∈ W , set xg = h ◦ g|n ◦ inm = h ◦ iN
n ◦ g|m . Since g|n and in are both in Emb(An , AN ), we
N
have h ◦ g|
Tn and h ◦ in in X. Since h ◦ g|n ∈ X ⊆ T1G , we have xg ∈ S. But this implies that
N
h ◦ in ∈ g∈W Tg , a contradiction.
T
Since Fn ∈ Tn and since T1G = inm (S) ∈ Yn , this means that ( g∈W ∪{1G } Tg ) ∈ p for some
p ∈ Yn . Now any α ∈ Y with α(n) = p is a member of YW,S . This proves the claim.
Now observe that if W1 , W2 are finite subsets of G, S1 ∈ Fm , and S2 ∈ Fn (m ≤ n), then
letting S3 = inm (S1 ) ∩ S2 ∈ Fn , we have YW1 ,S1 ∩ YW2 ,S2 ⊇ YW1 ∪W2 ,S3 . In particular, since each
YW,S is compact, there is α ∈ Y a member of all of them. Hence α · G ⊆ Y is a subflow of
lim βHn .
←−
For the other direction, suppose there were S ∈ Fm with S 6∈ Tm . Pick α ∈ Y ; we need
to show that for some g ∈ G, S 6∈ αg(m). To the contrary, suppose that S ∈ αg(m) for
8
every g ∈ G, or equivalently, that S ∈ α · f for every f ∈ Hm . But then {x ∈ Hn : x ◦ f ∈
S for every f ∈ Emb(Am , An )} ∈ α(n), so in particular is non-empty. But this implies that
S is thick, a contradiction.
Before proving Theorem 3.5, we need a quick remark about topological dynamics. If X
is a G-flow, then each point x ∈ X gives rise to the ambit (x · G, x). As such, there is a map
of ambits ϕ : (S(G), 1) → (x · G, x), and we write ϕ(α) := x · α. Since ϕ is surjective, we
have x · S(G) = x · G. If α, γ ∈ S(G), we also have x · (αγ) = (x · α) · γ. Now suppose that
X = 2Hm , and T ⊆ Hm . Then if α ∈ S(G), we have χT · α = χα−1 (T ) .
Proof of Theorem 3.5.
(⇐) We prove a slightly stronger statement. Suppose Y = lim Yn is a closed subset of
←−
lim βHn with Fm the filter of clopen neighborhoods of Ym . Say that for infinitely many
←−
m ∈ N, we have Fm ∈ β(Tm ). By Proposition 3.4, we see that for every m ∈ N, Fm is
a Tm -filter, so by Proposition 3.8, Y contains some right ideal. If Z = lim Zn ( Y is any
←−
proper closed subset, then Zm ( Ym for any large enough m ∈ N. So fix a large m ∈ N with
Fm ∈ β(Tm ). The filter Gm of clopen neighborhoods of Zm then properly contains Fm and
thus cannot be in Tm . Hence Z is not a right ideal, so Y is a minimal right ideal.
(⇒) If α ∈ lim βHn , then the left-multiplication map p → α · p on βHm is continuous.
←−
Hence if X ⊆ βHm is a closed subset defined by the filter F , αX ⊆ βHm is also closed. It
is defined by the filter
αF := {S ⊆ Hm : α−1 (S) ∈ F }.
Now suppose that F is a Tm -filter. Then for any α ∈ lim βHn , αF is also a Tm -filter. This
←−
is because α−1 (S) ∈ Tm implies that for some α0 ∈ lim βHn , we have χS · αα0 = χHm . But
←−
this implies that S ∈ Tm .
Let Y = lim Ym be a minimal right ideal, with Fm the filter of clopen neighborhoods of
←−
Ym . Suppose for sake of contradiction that for some m ∈ N, Fm 6∈ β(Tm ). Say Gm ) Fm
with Gm ∈ β(Tm ). Inductively define for each n ≥ m Gn ∈ β(Tn ) so that Gn+1 is any
Tn+1 -ultrafilter with ı̂n+1
n (Gn+1 ) = Gn . If Zm ⊆ βHm is the closed set for Gm , then by the
proof of the right to left implication, lim Zn is a minimal right ideal. Fix α ∈ Z, and consider
←−
αGm . As the left-multiplication map p → α · p is a homeomorphism from Ym to Zm , we
have α(Zm ) ( Zm , so αGm ) Gm . However, αGm is a Tm -filter, and Gm is maximal. Taken
together, this is a contradiction.
We now turn to the proof of Theorem 3.6. First an easy lemma.
Lemma 3.9. If f : (X, S) → Z is a regular map, and f = g ◦ h, with h : X → Y and
g : Y → Z, then g : (Y, h(S)) → Z is regular.
Proof. Let q ∈ β(h(S)), and set r = g(q). Pick p ∈ β(S) with h(p) = q. Then f (p) = r, so
r ∈ β(f (S)) = β(g(h(S))).
Proof of Theorem 3.6.
By Proposition 2.4 and Lemma 3.9, it is enough to show that each ı̂nm is regular. Say
Fn ∈ β(Tn ). Inductively define FN ∈ β(TN ) for each N ≥ n so that ı̂N
N +1 (FN +1 ) = FN ;
9
letting Yn ⊆ βHn be the closed set for Fn , we have seen that lim Yn is a minimal right ideal.
←−
It follows that Fm = ı̂nm (Fn ) is in β(Tm ).
We end this section with a brief “sanity check.” Recall from section 2 the map πm :
Gd → Hm with πm (g) = g|m .
−1
Theorem 3.10. T ⊆ Hm is thick iff πm
(T ) ⊆ Gd is thick.
Proof. Say T ⊆ Hm is thick. Let g1 , ..., gk ∈ Gd . Write fi = πm (gi ), and find n ≥ m so
that for each i ≤ k, we have fi ∈ Emb(Am , An ). By Proposition 3.3, there is x ∈ Hn with
x ◦ Emb(A
n ) ⊆ T . Find g ∈ Gd with g|n = x. Then for each i ≤ k, we have (ggi )|m ∈ T ,
T m, A
−1
−1
so g ∈ i≤k πm
(T )gi−1 6= ∅, and πm
(T ) is thick.
Conversely, say T is not thick, so that S := Hm \ T is syndetic. By Proposition 3.3, this
means that there is n ≥ m so that for every x ∈ Hn , we have S ∩ (x ◦ Emb(Am , An )) 6= ∅.
Let Emb(Am , An ) = {f1 , ..., fk }, and find g1 , ..., gk ∈ Gd with πmS(gi ) = fi . Fix g ∈ Gd .
−1
Then for some i ≤ k, we have (ggi )|m ∈ S. But this implies that i≤k πm
(S)gi−1 = Gd , so
−1
−1
πm (T ) = Gd \ πm (S) is not thick.
4
Piecewise syndetic sets
In the previous section, we successfully proved analogues of Definition 1.1 and Fact 1.3
for thick subsets of Hm using a definition resembling Fact 1.2. It easily follows that these
analogues also hold for syndetic subsets of Hm , in turn providing a generalization of Theorem
2.9(c) from [BHM]. However, we can not yet conclude that these analogues hold for piecewise
syndetic sets.
Let’s briefly consider what this even means. As we have pointed out before, there is no
right action on the set Hm , so it is not immediate what an analogue of Definition 1.1 even
looks like for piecewise syndetic sets. Theorem 3.10 suggests the following statement.
−1
Theorem 4.1. P ⊆ Hm is piecewise syndetic iff πm
(P ) ⊆ Gd is piecewise syndetic.
Proof. The surjection πm : Gd → Hm induces a continuous embedding π̂m : 2Hm → 2Gd ,
−1
where given S ⊆ Hm , we have π̂m (χS ) = χπm
(S) . The group Gd acts on the right on both
spaces, and the map π̂m is Gd -equivariant. The statements from Definition 3.2 do not depend
on the topology of G, so also hold for Gd .
Suppose P ⊆ Hm is piecewise syndetic, and let S ⊆ Hm be syndetic with χS ∈ χP · Gd .
−1
−1
−1
Then χπm
(S) ∈ χπm
(P ) · Gd , and πm (S) ⊆ Gd is syndetic by Theorem 3.10. It follows that
−1
πm
(P ) is piecewise syndetic.
−1
Conversely, suppose P ⊆ Hm is a set with πm
(P ) ⊆ Gd piecewise syndetic. Then there
Hm
−1
is syndetic W ⊆ Gd with W ∈ χπm
is compact, we must have that
(P ) · Gd . But since 2
−1
W = πm (S) for some S ⊆ Hm , and this S is syndetic by Theorem 3.10. It follows that P is
piecewise syndetic.
Let us briefly return to the setting where G is a discrete group. It is a fact that P ⊆ G
is piecewise syndetic iff P = S ∩ T for some syndetic S ⊆ G and thick T ⊆ G. Let us recall
10
one direction of the proof. Say P ⊆ G is piecewise syndetic, and let g1 , ..., gk ∈ G with
T := P g1−1 ∪ · · · ∪ P gk−1 thick; we may assume that k is minimal and that g1 = 1G (if T
is thick, then so is T g1 ). Then setting S = (G \ T ) ∪ P , we have that S is syndetic (since
Sg1−1 ∪ · · · ∪ Sgk−1 = G) and that P = S ∩ T . Notice that T \ P ⊆ G \ S is not thick.
Returning to the setting where G = Aut(K), we can ask whether these statements also
hold for Hm . Given a thick T ⊆ Hm , set Dest(T ) = {P ⊆ Hm : T \ P is not thick};
if P ∈ Dest(T ), we say that P destroys T . The main theorems of this section are the
following, generalizing Theorem 2.4 from [BHM]:
Theorem 4.2. P ⊆ Hm is piecewise syndetic iff P ∈ Dest(T ) for some thick T ⊆ Hm .
Theorem 4.3. P ⊆ Hm is piecewise syndetic iff there are thick T ⊆ Hm and syndetic
S ⊆ Hm with P = S ∩ T .
Let us remark that Theorems 4.2 and 4.3 are not corollaries of Theorem 4.1. The best we
−1
(S)g −1 = πn−1 (g|m (S)) for n ≥ m
can do is to note that given S ⊆ Hm and g ∈ G, we have πm
large enough with g|m (Am ) ⊆ An . Then using Theorem 4.1, we can say that if P ⊆ Hm
is piecewise syndetic, there is some n ≥ m and some thick T ⊆ Hn with inm (P ) ∈ Dest(T ).
However, it is not immediately clear that we can take n = m. In the preprint [Z1], it is
mistakenly asserted (see the discussion at the end of page 23) that Theorem 4.2 is somehow
obvious. So while the result is true, the proof is a bit more difficult.
Let us start by obtaining a more combinatorial characterization of piecewise syndetic
sets. The proof is similar to the proof of Proposition 3.3.
Proposition 4.4. P ⊆ Hm is piecewise syndetic iff there is n ≥ m and an exhausting
~ = {xN : N < ω} so that for every x ∈ Emb(An , X),
~ we have
sequence X
P ∩ (x ◦ Emb(Am , An )) 6= ∅.
Proof. Suppose P ⊆ Hm is piecewise syndetic. Find gN ∈ G a sequence of group elements
so that χP · gN → χS for some syndetic S ⊆ Hm . We may assume that χP · gN (f ) = χS (f )
for every f ∈ Emb(Am , AN ). Since S is syndetic, there is some n ≥ m with S ∩ (x ◦
Emb(Am , An )) 6= ∅ for every x ∈ Hn . Setting xN = gN |N , it follows that the conclusion
holds.
Conversely, suppose P ⊆ Hm is a set with the property described by the proposition.
Find gN ∈ G with gN |N = xN . Pass to a subsequence so that χP · gN |N → χS for some
S ⊆ Hm . This S has the property that S ∩ x ◦ Emb(Am , An ) 6= ∅ for every x ∈ Hn , so S is
syndetic, and P is therefore piecewise syndetic.
Recall that Pm is the collection of piecewise syndetic subsets of Hm . Given a set X and
a family S on X, we say that S is partition regular if given S ∈ S, then whenever we write
S = S1 ∪ S2 , some Si ∈ S.
Theorem 4.5. Pm is partition regular.
Proof. Let P = P1 ∪ P2 be piecewise syndetic; using Proposition 4.4, say this is witnessed by
n ∈ N and (xN )N ≥n , xN ∈ HN . If P1 is not piecewise syndetic, then for each n0 ∈ N, there
are N ∈ N and yn0 ∈ xN ◦ Emb(An0 , An ) with P1 ∩ (yn0 ◦ Emb(Am , An0 )) = ∅. But now we
see that n ∈ N and (yn0 )n0 ≥n witness that P2 is piecewise syndetic.
11
Write Nm := P(Hm ) \ Pm for the collection of not-piecewise-syndetic subsets of Hm . We
have just proven that Nm is an ideal. There is another natural ideal one could consider on
Hm :
Im := {A ⊆ Hm : ∀ T ∈ Tm (T \ A ∈ Tm )}.
In other words, members of Im are exactly those sets which do not destroy any thick set.
We will eventually see that Nm = Im , and we will use this to prove Theorems 4.2 and 4.3.
Proposition 4.6. Nm ⊆ Im .
Proof. Let W ∈ Nm , and fix T ∈ Tm . Find for each N ≥ m an xN ∈ HN so that xN ◦
Emb(Am , AN ) ⊆ T . As W is not piecewise syndetic, we can find for each n ≥ m an
N ≥Sn and a yn ∈ xN ◦ Emb(An , AN ) with W ∩ (y ◦ Emb(Am , An )) = ∅. But now set
U = n yn ◦ Emb(Am , AN ). We see that U is thick and U ⊆ T \ W .
To prove the other inclusion, we need a characterization of piecewise syndetic sets similar
to Fact 1.3. Call S ⊆ Hm minimal if χS · G is a minimal G-flow. Note that any non-empty
minimal set is syndetic.
Proposition 4.7. Let P ⊆ Hm . Then P is piecewise syndetic iff there is some minimal
right ideal Y ⊆ lim βHn and some α ∈ Y with P ∈ α(m). Equivalently, P is piecewise
←−
syndetic iff P̃ ⊆ S(G) meets some minimal right ideal.
Proof. First, suppose there are Y and α ∈ Y as above with P ∈ α(m). Then χP · α · G is
a minimal G-flow, so α−1 (P ) is a minimal set. Since im ∈ α−1 (P ), we see that α−1 (P ) is
syndetic, so P is piecewise syndetic.
For the converse, suppose P is piecewise syndetic, where S is syndetic and χS = χP · α,
α ∈ lim βHn . By replacing α with α · γ for γ ∈ lim βHn in some minimal right ideal, we
←−
←−
can assume that α is in some minimal right ideal Y . Fix g ∈ G with g|m ∈ α−1 (P ). Then
P ∈ αg(m) and α · g ∈ Y .
Lemma 4.8. Let Fm ∈ β(Tm ). Then for any A ∈ Im , we have Hm \ A ∈ Fm .
Proof. Fix A ∈ Im , and let B = Hm \ A. Then for every T ∈ Tm , we have T ∩ B ∈ Tm . In
particular, Fm ∪ {B} generates a thick filter, so by maximality of Fm we have B ∈ Fm .
Theorem 4.9. Nm = Im .
Proof. By Proposition 4.6, it is enough to prove Im ⊆ Nm . Suppose that P ⊆ Hm is
piecewise syndetic. We need to show that P 6∈ Im . By Proposition 4.7, fix a minimal right
ideal Y = lim Yn ⊆ lim βHn and an α ∈ Y with P ∈ α(m). Let Fn be the filter of clopen
←−
←−
neighborhoods of Yn ; by Theorem 3.5, we have Fn ∈ β(Tn ) for each n ∈ N. But suppose P
were in Im ; by Lemma 4.8, we have Hm \ P ∈ Fm , so in particular Hm \ P ∈ α(m). This is
a contradiction.
Note that Theorem 4.9 implies Theorem 4.2. We conclude this section with the (now
short) proof of Theorem 4.3.
12
Proof of Theorem 4.3.
Assume P is piecewise syndetic. As P 6∈ Im , find a thick T ⊆ Hm with T \ P not thick; we
may assume P ⊆ T . So Hm \ (T \ P ) = P ∪ (Hm \ T ) := S is syndetic, and P = T ∩ S.
For the other direction, say T ∈ Tm and S ∈ Sm . Then T \ T ∩ S is not thick, so T ∩ S
is piecewise syndetic.
Remark. The proof of Theorem 4.9 gives us the following strengthening of Corollary 4.3. If
P ⊆ Hm is piecewise syndetic and P is F -positive for some F ∈ β(Tm ), then there is T ∈ F
with T \ P not thick. As we may assume T ⊇ P , we have that Hm \ (T \ P ) := S is syndetic,
and P = T ∩ S.
5
Questions
In this final section, we collect some questions suggested by the previous results. The first
question is an abstract one about families.
Question 5.1. If (X, S) is a family and Y is a set, then what are necessary and sufficient
conditions for ϕ : (X, S) → Y to be regular?
The “map” language we have been using is somewhat misleading; whether or not a map
ϕ : (X, S) → Y is regular or strong depends only on the equivalence relation Eϕ , where
xEϕ x0 iff ϕ(x) = ϕ(x0 ). In particular, we can restrict our attention to the case where ϕ is
surjective.
When thinking about Question 5.1, it might be useful to “turn the problem around.”
Fix a surjection ϕ : X → Y . If T is a family on Y , set
ϕ−1 (T ) = {S : S is a family on X and ϕ(S) = T }.
Two families in ϕ−1 (T ) are worth distinguishing. We set
−1
ϕ−1
min (T ) = {A ⊆ X : ∃B ∈ T (ϕ (B) ⊆ A)},
ϕ−1
max (T ) = {A ⊆ X : ϕ(A) ∈ T }.
−1
It is not hard to check that for any family S ∈ ϕ−1 (T ), we have ϕ−1
min (T ) ⊆ S ⊆ ϕmax (T ).
Also note that ϕ : (X, S) → Y is strong iff S = ϕ−1
max (T ). When T and ϕ are understood
and S ∈ ϕ−1 (T ), call S regular if the map ϕ : (X, S) → Y is regular.
Proposition 5.2. S = ϕ−1
min (T ) is regular.
Proof. Let p ∈ β(S). For A ⊆ X, let à denote the largest Eϕ -invariant subset of A. By our
assumption that S = ϕ−1
min (T ), we have that if A ∈ p, then also à ∈ p.
Suppose towards a contradiction that ϕ(p) ( q for some q ∈ β(T ). Fix B ∈ q \ ϕ(p).
Then ϕ−1 (B) 6∈ p, so find A ∈ p with ϕ−1 (B) ∩ A 6∈ S. We may assume A = Ã. But now
B ∩ ϕ(A) 6∈ T , a contradiction since we have B and ϕ(A) in q.
Given S, S 0 ∈ ϕ−1 (T ), say that S 0 is a conservative over S if S ⊆ S 0 and for every
p ∈ β(S 0 ), we have p ∩ S ∈ β(S). If S is regular and S 0 is conservative over S, then S 0 is
also regular.
13
Proposition 5.3. With T and ϕ as above, S ∈ ϕ−1 (T ) is regular iff S is conservative over
ϕ−1
min (T ).
Proof. One direction is clear. For the other direction, suppose S is regular, and fix p ∈ β(S).
−1
Towards a contradiction, suppose p ∩ ϕ−1
min (T ) ( q with q ∈ β(ϕmin (T )), and let A ∈ q \ p.
We may assume A = Ã, i.e. that A is Eϕ -invariant. But we have ϕ(A) ∈ ϕ(q) = ϕ(p), so
ϕ−1 (ϕ(A)) = A ∈ p, a contradiction.
The next question concerns both abstract families and the families Tm of thick sets on
Hm . Suppose X is a set and F , F 0 are filters on X. In a slight abuse of language, call F
and F 0 disjoint if there are A ∈ F and B ∈ F 0 with A ∩ B = ∅. If S is a family on X, we
say that S has disjointness if whenever p 6= q ∈ β(S), we have p and q disjoint.
Question 5.4. Characterize those families S with disjointness.
Suppose briefly that G is a discrete group. If T is the family of thick subsets of G, then
T has disjointness. This is because the members of β(T ) are exactly the filters of clopen
neighborhoods of minimal right ideals of βG, and minimal right ideals of βG are either
identical or disjoint.
Returning to the case that G = Aut(K), while it is still true that distinct minimal right
ideals of S(G) are disjoint, this is not necessarily true on each level. So for the family Tm ,
Question 5.4 is equivalent to the following:
Question 5.5. Let Y = lim Yn and Y 0 = lim Yn0 be distinct minimal right ideals of S(G) =
←−
←−
lim βHn . If Ym 6= Ym0 , must we have Ym ∩ Ym0 = ∅?
←−
The final question is specific to the families Tm , or at least to reasonably “definable”
families. Theorem 4.2 guarantees that to every piecewise syndetic P ⊆ Hm , there is a thick
T ⊆ Hm so that P destroys T . This question asks whether there is a ”definable” choice of
T.
Question 5.6. Let X = {(P, T ) ∈ 2Hm × 2Hm : T is thick and P ∈ Dest(T )}. Does X admit
a Borel uniformization? In other words, is there a Borel partial function ψ : 2Hm → 2Hm
with domain the piecewise syndetic sets so that ψ(P ) is thick and P destroys ψ(P )?
In the setting where G is a countable discrete group, the answer to the analogous question
is positive. If P is piecewise syndetic, search for a finite E = {g1 , ..., gk } ⊆ G with T :=
P g1 ∪ · · · ∪ P gk thick so that k is minimal. Then set ψ(P ) = T g1−1 .
References
[AG] E. Akin and E. Glasner, Is there a Ramsey-Hindman theorem?, preprint,
http://arxiv.org/pdf/1504.00218v2.pdf, (2015).
[B] W. R. Brian, P -sets and minimal right ideals in N∗ , Fund. Math. 229 (2015), 277–293.
[BHM] V. Bergelson, N. Hindman, and R. McCutcheon, V. Bergelson, Notions of size and
combinatorial properties of quotient sets in semigroups, Topology Proc. 23 (1998), 2360.
14
[HS] N. Hindman and D. Strauss, Algebra in the Stone-Čech Compactification, 2nd Edition,
De Gruyter (2012).
[P] V. Pestov, On free actions, minimal flows, and a problem by Ellis, Trans. Amer. Math.
Soc., 350 (10), (1998), 4149–4165.
[Z] Andy Zucker, Topological dynamics of automorphism groups, ultrafilter combinatorics,
and the Generic Point Problem, Transactions of the American Mathematical Society,
368(9), (2016).
[Z1] Andy Zucker, Topological dynamics of closed subgrops of S∞ , preprint,
http://arxiv.org/pdf/1404.5057v2.pdf, (2014).
Andy Zucker
Carnegie Mellon University
Pittsburgh, PA 15213
[email protected]
15
| 4math.GR
|
Deep Learning for Decoding of Linear Codes A Syndrome-Based Approach
arXiv:1802.04741v1 [cs.IT] 13 Feb 2018
Amir Bennatan* , Yoni Choukroun* and Pavel Kisilev
Huawei Technologies Co.,
Email: {amir.bennatan, yoni.choukroun, pavel.kisilev}@huawei.com
Abstract— We present a novel framework for applying deep
neural networks (DNN) to soft decoding of linear codes at
arbitrary block lengths. Unlike other approaches, our framework
allows unconstrained DNN design, enabling the free application
of powerful designs that were developed in other contexts. Our
method is robust to overfitting that inhibits many competing
methods, which follows from the exponentially large number
of codewords required for their training. We achieve this by
transforming the channel output before feeding it to the network,
extracting only the syndrome of the hard decisions and the
channel output reliabilities. We prove analytically that this
approach does not involve any intrinsic performance penalty, and
guarantees the generalization of performance obtained during
training. Our best results are obtained using a recurrent neural
network (RNN) architecture combined with simple preprocessing
by permutation. We provide simulation results that demonstrate
performance that sometimes approaches that of the ordered
statistics decoding (OSD) algorithm.
I. I NTRODUCTION
Interest in applying neural networks to decoding has existed
since the 1980’s [6], [7], [8]. These early works, however, did
not have a substantial impact on the field due to the limitations
of the networks that were available at the time. More recently,
deep neural networks were studied in [4], [5], [2].
A major challenge facing applications of deep networks to
decoding is the avoidance of overfitting the codewords encountered during training. Specifically, training data is typically
produced by randomly selecting codewords and simulating
the channel transitions. Due to the large number of codewords
(exponential in the block length), it is impossible to account for
even a small fraction of them during training, leading to poor
generalization of the network to new codewords. This issue
was a major obstacle in [4], [5], constraining their networks
to very short block lengths.
Nachmani et al. [2], [3] proposed a deep learning framework which is modeled on the LDPC belief propagation (BP)
decoder, and is robust to overfitting. A drawback of their
design, however, is that to preserve symmetry, the design is
constrained to closely mimic the message-passing structure
of BP. Specifically, the connections between neurons are
controlled to resemble BP’s underlying Tanner graph, as are
the activations at neurons. This severely limits the freedom
available to the neural networks design, and precludes the
application of powerful architectures that have emerged in
recent years [15].
* Both authors contributed equally to this work.
In this paper, we present a method which overcomes this
drawback while maintaining the resilience to overfitting of [2].
Our framework allows unconstrained neural network design,
paving the way to the application of powerful neural network
designs that have emerged in recent years. Central to our
approach is a preprocessing step, which extracts from the
channel output only the reliabilities (absolute values), and
the syndrome of its hard decisions, and feeds them into the
neural network. The network’s output is later combined with
the channel output to produce an estimate of the transmitted
codeword.
Decoding methods that focus on the syndrome are well
known in literature on algebraic decoding (see e.g. [12][Sec.
3.2]). The approach decouples the estimation of the channel
noise from that of the transmitted codeword. In our context
of deep learning, its potential lies in the elimination of the
need to simulate codewords during training, thus overcoming
the overfitting problem. A few early works that used shallow
neural networks have employed syndromes (e.g. [6]). However, these works did not discuss its potential in terms of
overfitting, presumably because the problem was not as acute
in their relatively simple networks. Importantly, their approach
does not apply in cases where the channel includes reliabilities
(mentioned above).
Our approach in this paper extends syndrome decoding
to include channel reliabilities, and applies it to overcome
the overfitting issue mentioned above. We provide a rigorous
analysis which proves that our framework incurs no loss
in optimality, in terms of bit error rate (BER) and mean
square error (MSE). Our analysis utilizes some techniques
developed by Burshtein et al. [14], Wiechman and Sason [13]
and Richardson and Urbanke [11][Sec. 4.11].
Building on our analysis, we propose two deep neural
network architectures for decoding of linear codes. The first
is a vanilla multilayer network, and the second a moreelaborate recurrent neural network (RNN) based architecture.
We also develop a preprocessing technique (beyond the abovementioned computation of the syndrome and reliabilities),
which applies a permutation (an automorphism) to the decoder
input to facilitate the operation of the neural network. This
technique builds on ideas by Fossorier et al. [9], [10] and
Dimnik and Be’ery [17] but does not involve list decoding.
Finally, we provide simulation results for decoding of BCH
codes which for the case of BCH(63,45), demonstrate performance approaches that of the ordered statistics algorithm
(OSD) of [9], [10].
Summarizing, our main contributions are:
1) A novel deep neural network training framework for
decoding, which is robust to overitting. It is based on the
extension of syndrome decoding that is described below.
2) An extension of syndrome decoding which accounts
for reliabilities. As with legacy syndrome decoding, we
define a generic framework, leaving room for a noiseestimation algorithm which is specified separately. We
provide analysis that proves that the framework involves
no loss of optimality, and that regardless of the noiseestimation algorithm, performance is invariant to the
transmitted codeword.
3) Two neural network designs for the noise-estimation algorithm, including an elaborate RNN-based architecture.
4) A simple preprocessing technique that applies permutations to boost the decoder’s performance.
Our work is organized as follows. In Sec. II we introduce
some notations and in Sec III we provide some bakground on
neural networks. In Sec. IV we describe our syndrome-based
framework and provide an analysis of it. In Sec. V we discuss
deep neural network architectures as well as preprocessing by
permutation. In Sec. VI we present simulation results. Sec. VII
concludes the paper.
II. N OTATIONS
We will often use the superscripts b and s (e.g., xb and xs ) to
denote binary values (i.e, over the alphabet {0, 1}) and bipolar
values (i.e., over {±1}), respectively. We define the mapping
from the binary to the bipolar alphabet, denoted bipolar(·), by
0 → 1, 1 → −1 and let bin(·) denote the inverse mapping.
Note that the following identity holds:
s
s
s
s
bin(x · y ) = bin(x ) ⊕ bin(y ),
s
s
∀x , y ∈ {±1}
(1)
where ⊕ denotes XOR. sign(x) and |x| denote the sign
and absolute value of any real-valued x, respectively, In
our analysis below, when applied to a vector, the operations
bipolar(·), bin(·), sign(·) and | · | are assumed to be applied
independently to each of the vector’s components.
III. B RIEF OVERVIEW OF D EEP N EURAL N ETWORKS
We now briefly describe the essentials of neural networks.
Our discussion is in no way comprehensive, and there are
many variations on the simple setup we describe. For an
elaborate discussion, see e.g. [15].
for an associated computational algorithm. The nodes are
called neurons and each performs (i.e., is associated with) a
simple computation on inputs to produce a single output. The
algorithm inputs are fed into the first-layer neurons, whose
outputs are fed as inputs into the next layer and so forth.
Finally, the outputs of the last layer neurons become the
algorithm output. Deep networks are simply neural networks
with many layers.
The inputs and output at each neuron are real-valued numbers. To compute its output, each neuron first computes an
affine function of its input (a weighted sum plus a bias). It then
applies a predefined activation function, which is typically
nonlinear (e.g., sigmoid or hyperbolic tangent), to render the
output.
The power of neural networks lies in their configurability.
Specifically, the weights and biases at each neuron are parameters which can be tailored to produce a diverse range
of computations. Typically, the network is configured by a
training procedure. This procedure relies on a sample dataset,
consisting of inputs, and in the case of supervised training, of
desired outputs (known as labels). There are several training
paradigms available, most of which are variations of gradient
descent.
Overfitting occurs typically when the training dataset is
insufficiently large or diverse to be representative of all valid
network inputs. The resulting network does not generalize well
to inputs that were not encountered in the training set.
IV. T HE P ROPOSED S YNDROME -BASED F RAMEWORK AND
ITS A NALYSIS
A. Framework Definition
We begin by briefly discussing the encoder. We assume
standard transmission that uses a linear code C. We let
m ∈ {0, 1}K denote the input message, which is mapped
to a codeword xb ∈ {0, 1}N (recall that the superscript b
denotes binary vectors). We assume that xb is mapped to a
bipolar vector x using the mapping defined in Sec. II, and x
is transmitted over the channel. We let y denote the channel
output. For convenience, we define m, xb , x and y to be
column vectors.
Decoder
𝐲𝒔
Channel
output
sign
bin
𝐲𝒃
𝐲
abs
Algorithm
inputs
|𝐲|
𝐻
𝐻𝐲 𝒃
Noise
estimation
F
Estimate
x
sign
𝐱
Algorithm
outputs
Fig. 2. Decoder framework. The final sign operation is omitted when the
system is designed to produce soft decisions. In Sec. V, we will implement
F using a neural network.
Layer 1
Fig. 1.
Layer 2
Layer 3
Layer 4
Layer 5
Illustration of a multi-layer neural network.
Fig. 1 depicts a simple multilayered neural network. The
network is a directed graph whose structure is the blueprint
Our decoder framework is depicted in Figure 2. The decoder’s main component is F, whose role is to estimate the
channel noise, and which we will later (Section V) implement
using a deep neural network. The discussion in this section,
however, applies to arbitrary F. The inputs to F are the
absolute value |y| and the syndrome Hyb , where H is a parity
check matrix of the code C, yb is a vector of hard decisions,
yb = bin(sign(y)) where bin(·) is simply the inverse of the
above defined bipolar mapping. The multiplication by H is
modulo-2. The output of F is multiplied (componentwise) by
ys , which is the sign of y. Finally, when interested in hard
decisions, we take the sign of the results and define this to be
the estimate x̂. This final hard decision step (which is depicted
in Fig. 2) is omitted when the system is required to produce
soft decisions.
B. Binary-Input Symmetric-Output (BISO) Channels
Our analysis in the following section applies to a broad class
of channels known as binary-input symmetric-output (BISO)
channels. This class includes binary-input AWGN channels,
binary-symmetric channels (BSCs) and many others. Our
definition below follows Richarson et al. [1][Definition 1].
Definition 1. Consider a memoryless channel with input alphabet {±1}. The channel is BISO if its transition probability
function satisfies:
Pr[Y = y | X = 1] = Pr[Y = −y | X = −1],
(2)
for al y the channel output alphabet, where X and Y denote
the random channel input and output (respectively).
An important feature of BISO channels is that their random
transitions can be modeled by [1][proof of Lemma 1],
Y = X · Z̃,
(3)
where Z̃ is random noise which is independent of the transmitted X. The tilde in Z̃ serves to indicate that this is an
equivalent statistical model, which might differ from the true
physical one. To prove (3), we simply define Z̃ to be a random
variable distributed as Pr[Y | X = 1] and independent of X.
The validity now follows from (2).
Outline. In Part 1 we neglect implementation concerns, and
focus on realizations of F that try to optimally estimate
the multiplicative noise z̃ (see (3)). By (3), given y, such
estimation is equivalent to estimation of x. We argue that
the pair |y| and Hyb is a sufficient statistic for estimation
of z̃. To see this, observe that y is equivalent to the pair |y|
and sign(y). The latter term, in turn, is equivalent to the pair
Hyb and Ayb , where yb is as defined in Sec. II and A is
a pseudo-inverse of the code’s generator matrix G. By (1)
and (3), yb = xb ⊕z̃b and so Ayb is the sum of the transmitted
message m = Axb and Az̃ (the projection of z̃ onto the code
subspace). We argue that Ayb is independent of the noise and
thus irrelevant to its estimation. This follows because m is
independent of z̃, and we assume it to be uniformly distributed
within the message space {0, 1}K .
To prove Part 2 of the theorem, we allow F to be arbitrary
and show that the decoder’s output can be modeled as f (z̃) · x
for some vector-valued function f (·). Thus, its relationship
with x (which determines the BER and MSE) depends on the
noise z̃ alone. To see why this holds, first observe that the
inputs to F (and consequently, its outputs) are dependent on
the noise z̃ only. This follows because the syndrome Hyb
equals Hz̃b . This in turns follows from the relation yb =
xb ⊕ z̃b and the fact that xb is a codeword, and so Hxb = 0.
By (3) and the bipolarity of x, the absolute value |y| equals
|z̃|. It now follows that the output of F is dependent on z̃
alone. Multiplication by sign(y) (see Fig. 2) is equivalent to
multiplying by sign(z̃) · x (by (3)) and the result follows.
V. I MPLEMENTATION USING D EEP N EURAL N ETWORKS
Theorem 1 proves that our framework does not intrinsically
involve a loss of optimality. To realize its potential, we propose
efficient implementations of the function F. In this section,
we discuss deep neural network implementations as well as a
simple preprocessing technique that enables the networks to
achieve improved performance.
A. Deep Neural Network Architectures
C. Analysis
We now show that the decoder framework involves no
penalty in performance in terms of metrics mean-squared-error
(MSE) or bit error rate (BER). That is, the decoder can be
designed to achieve any one of them. Importantly, it addresses
the overfitting problem that was described in Sec. I.
Theorem 1. The follwing holds with respect to the framework
of Sec. IV-A, assuming communication over a BISO channel:
1) The framework incurs no intrinsic loss of optimality, in
the sense that with an appropriately designed F, the
decoder can achieve maximum a-posteriori (MAP) or
minimum MSE (MMSE) decoding.
2) For any choice of F, the decoder’s BER and MSE,
conditioned on transmission of any codeword x, are both
invariant to x.
We provide an outline of the proof here, and defer the details
to Appendix I.
We consider the following two architectures:
1) Vanilla Multi-Layer: With this architecture, the neural
network contains an array of fully-connected layers as
illustrated in Fig. 3. It closely resembles simple designs [15] with the exception that we feed the network
inputs into each of the layers, in addition to the output of
the previous layer (this idea is borrowed from the belief
propagation algorithm).
The network includes 11 fully-connected layers, roughly
equivalent to the 5 LDPC belief-propagation iterations
as in [2]. We use rectified linear unit (ReLU) nonlinear
activation functions [15]. Each of the first 10 layers
consists of the same number of nodes (6N for block
length N = 64 and 15N for block length N = 127).
The final fully-connected layer has N nodes and produces
the network output, using a hyperbolic tangent arctivation
function.
2) Recurrent Neural Network (RNN): With this architecture, we build a deep RNN by stacking multiple recurrent
…
𝐲,
𝐻𝐲 𝑏
𝐳
Network
output
Network
input
Output
layer
Layer 1
Fig. 3.
Layer 2
Layer 10
codeword. With each training batch, we generate a new set of
noise samples. While this procedure produced our best results,
an alternative approach which fixes the training noise and uses
other techniques (e.g., dropout) to overcome overfitting the
noise, is worth exploring.
With the RNN architecture, the network produces multiple
outputs (at each time-step) and we use the following loss
function:
An instance of the Vanilla multi-layer architecture.
L=
hidden states on top of each other [19] as illustrated
in Fig. 4. RNNs realize a design which is equivalent
to a multi-layer architecture by maintaining a network
hidden state (memory) which is updated via feedback
connections. Note that from a practical perspective, this
renders the network more memory-efficient. In many
applications, this structure is useful to enable temporal
processing, and stacking RNNs as in Fig. 4 enables
operation at different timescales. In our setting, this
interpretation does not apply, but we nonetheless continue
to refer to RNN layers as “time steps.” Stacking multiple
RNNs produces an effect that is similar to deepening the
network.
We use Gated Recurrent Unit (GRU) [20] cells which
have shown peformance similar to well-known long shortterm memory (LSTM) cells [21], but have less parameters
due to the lack of a reset gate, making them a faster inference alternative. We use the hyperbolic tangent nonlinear
activation functions, the networks posses 4 RNN stacks
(levels), the hidden state size is set to 5N and the RNN
performs 5 time steps.
T
N
1 X X T −t
γ
HCE (z̃is , ẑis (t)),
N t=1 i=1
where HCE is the cross-entropy function, z̃is is the sign of
component i of the multiplicative noise and where ẑis (t) the
network output corresponding to codebit i at RNN time step
(layer) t. γ < 1 a discount factor (in our simulations, we used
γ = 0.5). The loss for the vanilla architecture is a degenerate
version of the RNN one, with time steps and discount factors
removed.
B. Preprocessing by Permutation
The performance of the implementations described above
can further be improved by applying simple preprocessing and
postprocessing steps at the input and output of the decoder
(respectively). Our approach is depicted in Fig. 5. Preprocessing involves applying a permutation to the components
of the channel output vector, and postprocessing applies the
inverse permutation to the decoder output. The approach draws
on ideas from Fossorier et al. [9], [10] and Dimnik and
Be’ery [17]. Note that the approach deviates from these works
in that it does not involve computing a list of vectors.
Decoder
𝑧(1)
𝑧(2)
𝐺𝑅𝑈𝑛 (1)
𝐺𝑅𝑈𝑛 (2)
…
…
𝑧(𝑇)
…
Select
permutation
𝐺𝑅𝑈𝑛 (𝑇)
…
y
𝐺𝑅𝑈2 (1)
𝐺𝑅𝑈2 (2)
…
𝐺𝑅𝑈2 (𝑇)
𝐺𝑅𝑈1 (1)
𝐺𝑅𝑈1 (2)
…
𝐺𝑅𝑈1 (𝑇)
…
𝒚 , 𝐻𝒚
𝑏
Fig. 4.
Our implementation of stacked RNN layers where GRUi (t)
represents the i-th level cell at RNN time step (layer) t.
To train the networks, we simulate transmission of the allone codeword (assuming the bipolar alphabet, {±1}). We
also simulate the multiplicative noise z̃, which in the case
of an AWGN channel, is distributed as a mean-1 Gaussian
random variable. In our training for Sec. VI, we set Eb /N0 to
4 dB. This was selected arbitrarily, and could potentially be
improved. We use Google’s TensorFlow library and the Adam
optimizer [16]. Testing proceeds in the same lines, except that
we use randomly generated codewords rather than the all-one
Decoder
of
Fig. 2
-1
x^
Fig. 5. Decoder with perprocessing and postprocessing components. Note
that in Appendix II we show that this decoder has an equivalent representation
which is a special case of Fig. 2.
Similar to [9], [10], our decoder selects the preprocessing
permutation so as to maximize the sum of the adjusted reliabilities of the first K components of the permuted channel output
vector. Borrowing an idea from [17], however, we confine our
permutations to subsets of the code’s automorphism group. We
assume that the parity check matrix, by which the syndrome in
Fig. 5 is computed, is arranged so that the last N −K columns
are diagonal and correspond to parity bits of the code.
We define the adjusted reliability of a channel component
yi , denoted R(yi ) by,
∆
R(y) = I(X; Y | |Y | = |y|).
(4)
well as the algorithm of [2]. However, a gap remains to the
OSD algorithm2 , which widens with Eb /N0 .
With respect to the number of codewords simulated, with
our algorithms, we simulated 105 codewords for each Eb /N0
point. With the OSD algorithm, we simulated 104 codewords
for each Eb /N0 point. With the BP algorithm we simulated
103 for each point.
The analysis of [2] also includes an mRRD framework,
into which their algorithm can be plugged to obtain superior
performance. While our algorithms can similarly be plugged
into that framework, our interest in this paper is in methods
whose primary components are neural networks.
1 Their
paper [3] does not include results for this case.
1
10
2
10
3
Belief propagation (BP)
Syndrome-based, vanilla
Syndrome-based, stacked RNN
Nachmani et al. best results
OSD, order 2
1.0
Fig. 6.
BER
VI. S IMULATION R ESULTS
Fig. 6 presents simulation results for communication with
the BCH(63,45) code over an AWGN channel. We simulated
our two architectures, namely syndrome-based vanilla and
stacked-RNN. Note that in this case, we did not simulate
permutations (Sec. V-B). Also plotted are results for the best
method of Nachmani et al. [2],[3], the belief propagation
(BP) algorithm, and for the ordered statistics decoding (OSD)
algorithm [10] of order 2 (for this algorithm we simulated
104 codewords for each Eb /N0 point). As can be seen from
the results, both our architectures substantially outperform the
BP algorithm. Our stacked-RNN architecture, like that of [3],
approaches the OSD algorithm very closely.
Fig. 7 presents results for the BCH(127,64) code. We
simulated our syndrome-based stacked RNN method, with and
without preprocessing. As can be seen, the preprocessing step
renders as a substantial benefit. Also plotted are results for
the BP and the OSD algorithms as well as the best results
of [2]1 . Both our methods outperform the BP algorithm as
10
BER
That is, R(y) equals the mutual information of random
variables X and Y , denoting the channel input and output,
respectively, conditioned on the event that the absolute value
|Y | equals |y|. X is uniformly distributed in {±1} and Y
is related to it via the channel transition probabilities. With
this definition, our permutation selection criterion is equivalent
to concentrating as much as possible of the channel capacity
within the first K channel output components.
Unlike [10], we borrow an idea from [17] and restrict
the set of allowed permutations to the code’s automorphism
group [18]. Permutations in this group have the property that
the permuted version of any codeword is guaranteed to be
a codeword as well. In our context, confinement to such
permutations ensures that the decoder input (the permuted
channel output) continues to obey the communication model,
namely being a noisy valid codeword. The decoder can thus
continue to rely on the code’s structure to decode.
When compared to framework of Fig. 2, the decoder of
Fig. 5 has the added benefit of knowing that the K first channel
outputs are consistently more reliable than the remaining
components. That is, the input exhibits additional structure
that the neural network can rely on.
In Appendix II we discuss permutations for BCH codes like
those we will use in Sec VI below, as well as efficient methods
for computing the optimal permutation. We also discuss formal
aspects related to applying the analysis of Sec. IV to the
framework of Fig. 5. Note that to achieve good results, the
added steps of Fig. 5 need to be included during training of
the neural network.
In Sec. VI we present simulation results for BCH(127,64)
codes with and without the above preprocessing method,
demonstrating the effectiveness of this approach. Note that
with the shorter block length BCH(63,45) codes, preprocessing was not necessary, and we obtained performance that
approaches the ordered statistics algorithm even without it.
2.0
2.5
Eb/No [dB]
3.0
3.5
4.0
BER results for a BCH (63,45) code.
10
1
10
2
10
3
10
4
10
5
Belief propagation (BP)
Nachmani et al. best results
Syndrome-based, stacked RNN, without preprocessing
Syndrome-based, stacked RNN, preprocessing
OSD, order 2
1
Fig. 7.
1.5
2
3
Eb/No [dB]
4
5
6
BER results for a BCH (127,64) code.
2 Note that for E /N of 4 dB or higher, we encountered no OSD errors
0
b
for BCH(127,64), in our simulations.
VII. C ONCLUSION
Our work in this paper presents a promising new framework
for the application of deep learning to error correction coding.
An important benefit of our design is the elimination of the
problem of overfitting to the training codeword set, which was
experienced by [4], [5]. We achieve this by using syndrome
decoding, and by extending it to account for soft channel
reliabilities. Our approach enables the neural network to focus
on the estimating the noise alone, rather than the transmitted
codeword.
It is interesting to compare our framework to that of
Nachmani et al. [2]. While their approach also resolves the
overfitting problem and achieves impressive simulation results,
it is heavily constrained to follow the structure of the LDPC
belief-propagation decoding algorithm. By contrast, our framework allows the unconstrained design of the neural network,
and our architectures are free to draw from the rich experience
that has emerged in recent years on neural network design.
Our simulation results demonstrate that our framework can
be applied to achieve strong performance that approaches
OSD. Further research will examine additional neural network architectures (beyond the RNN-based) and preprocessing
methods, to improve our performance further. It will also
consider the questions of latency and complexity.
We hope that research in these lines will produce powerful
algorithms and have a substantial impact on error correction
coding, rivaling the dramatic effect of deep learning on other
fields.
A PPENDIX I
P ROOF OF T HEOREM 1
In this section we provide the rigorous details of the proof,
whose outline was provided in Sec IV-C.
We begin with the following lemma.
Lemma 1. The following two claims hold with respect to the
framework of Sec. IV-A:
1) There exists a matrix A with dimensions K × N , such
that Axb = m for all m and xb defined as in Sec. IV-A
(recall that m and xb are both column vectors).
2) Let B be a matrix obtained by concatenating the rows of
H and A (i.e, B = [H T , AT ]T ). Then B has full column
rank, and is thus injective (one-to-one)
Note that in this lemma, we allow H to contain redundant,
linear dependent rows, as long as its rank remains N − K.
While we have not used such matrices in our work, the extra
redundant rows could in theory be helpful in the design of
effective neural networks for F.
Proof of Lemma 1. Part 1 of the lemma follows simply from
the properties of the generator matrix of the code C, denoted
here G. This matrix has full column-rank and dimensions N ×
K and satisfies xb = Gm. Thus, we can define A to be its
left-inverse, and Part 1 of the lemma follows.
To prove Part 2 of the lemma, we first observe that we
can assume without loss of generality the matrix H has
full row-rank. This is because by removing redundant (linear
dependent) rows from H we can obtain a full-rank paritycheck matrix, and such removal cannot affect the rank of the
corresponding B. Let D be a right-inverse of the matrix H,
whose existence follows from the fact that H has full rowrank. D has dimensions N × N − K. Consider the matrix
[G, D] (obtained by concatenating the columns of G and D).
H
HG HD
B · [G, D] =
· G D =
A
AG AD
0 IN −K
=
,
IK
AD
where IK and IN −K denote the identity matrices of dimensions K and N − K, respectively. The equality HG = 0
follows from the orthogonality of the generator and parity
matrices, and equalities HD = IN −K , AG = IK follow from
the definitions above of A and D. The resulting matrix has
rank N , and thus B cannot have rank less than N .
We now proceed to prove Part 1 of the theorem. We use
the following notation: Vector are denoted by boldface (e.g.,
x) and scalars by normalface (e.g., x). Random variables are
upper-cased (X) and their instantiations lower-cased (x).
We use the notation of Fig. 2, replacing lowercase with
uppercase wherever we need to denote random variables.
Accordingly, we let X and Y denote random variables corresponding to the channel input and output (respectively). y
is the realized channel output observed at the decoder and
x ∈ {±1} an arbitrary value.
Consider the string of equations ending in (5). In (a), Yis is a
random variable defined as yis (Fig. 2) and equality Yis = yis
follows from the condition Y = y. In (b), we have relied
∆
on (3) to replace Z̃is = Xi · Yis , where Z̃is = sign(Z̃i ). In
∆
(c), |Y| is the absolute value of Y and Ys = sign(Y). In
(d), we have relied on (3) and the bipolarity of X to obtain
|Z̃| = |Y|. We have also defined and Yb and yb as in Fig. 2.
In (e), the matrix B was defined as in Lemma 1 and equality
follows by the fact that B is injective. In (f ), we have used the
definition B = [H T , AT ]T where A is as defined in Lemma 1.
In (g), we have decomposed Yb = Xb ⊕ Z̃b . This follows
from (1) and (3). In (h), we have relied on HXb = 0 which
follows from the fact that Xb is a valid codeword. We have
also replaced AXb = M, where M is the random message
(see Fig. 2), following
i Finally, in (i) we have made
h Lemma 1.
the observation that M ⊕ AZ̃b is independent of Z̃ and can
therefore be omitted from the condition. This follows because
M, being the transmitted message, is uniformly distributed in
{0, 1}K (see Sec. IV-A) and independent of AZ̃b (and Z̃).
The proof now follows from (5). To obtain the MAP
decision for Xi given y we can define the components of
N
F as follows, for s ∈ {0, 1}N −K and a ∈ R+ .
n
o
∆
Fi (a, s) = argmax Pr[Z̃is = z̃ | |Z̃| = a, H Z̃b = s]
z̃∈{±1}
By (5), we now have MAP(Xi ) = yis Fi (|y|, Hyb ). Similarly,
to obtain the MMSE estimate for Xi , we can define:
∆
Fi (a, s) = E[Z̃is | |Z̃| = a, H Z̃b = s]
Pr[Xi = x | Y = y]
(a)
=
Pr[Xi · Yis = xyis | Y = y]
(b)
=
Pr[Z̃is = xyis | Y = y]
(c)
=
Pr[Z̃is = xyis | |Y| = |y|, Ys = ys ]
(d)
=
Pr[Z̃is = xyis | |Z̃| = |y|, Yb = yb ]
(e)
=
Pr[Z̃is = xyis | |Z̃| = |y|, BYb = Byb ]
(f )
=
Pr[Z̃is = xyis | |Z̃| = |y|, HYb = Hyb , AYb = Ayb ]
(g)
=
Pr[Z̃is = xyis | |Z̃| = |y|, H[Xb ⊕ Z̃b ] = Hyb , A[Xb ⊕ Z̃b ] = Ayb ]
(h)
=
Pr[Z̃is = xyis | |Z̃| = |y|, H Z̃b = Hyb , M ⊕ AZ̃b = Ayb ]
(i)
Pr[Z̃is = xyis | |Z̃| = |y|, H Z̃b = Hyb ]
=
Recall from Fig. 2 that in this case the decoder is configured
to produce soft decisions and so the final sign operation is
omitted (see Fig. 2). This concludes the proof of Part 1 of the
theorem.
Turning to Part 2, in Sec IV-C we proved that the decoder’s
output can always be modeled as f (z̃) · x for some vectorvalued function f (·). With respect to the BER metric, the
indices where the vectors f (z̃) · x and x diverge coincide with
the indices where f (z̃) equals -1, and thus are independent of
x. With respect to the MSE metric (recall that in this case,
the sign operation in Fig. 2 is omitted), we have MSE =
kf (z̃)·x−xk2 = kf (z̃)−1k2 and thus the error is independent
of x.
A PPENDIX II
AUTOMORPHISMS OF BCH CODES
In this section, we discuss automorphisms of primitive
narrow-sense binary BCH codes codes [18], which include
the codes used in Sec VI.
With the above codes, the blocklength N equals 2m − 1
for some positive integer m. Codewords are binary vectors
c = [c0 , ..., cN −1 ] (i.e., defined over indices i = 0, ..., N − 1).
Permutations are bijective functions π : {0, ..., N − 1} →
{0, ..., N − 1}. Given a codeword c and a permutation π, we
define the corresponding permuted codeword cπ by cπi = cπ(i) .
The automorphism group of the above-mentioned BCH
codes includes [18] [pp. 233] permutations of the form:
πk,l (i) = 2k i + l
mod N
where k ∈ {0, ...m − 1} and l ∈ {0, ..., N − 1}. The inverse
−1
permutation πk,l
can be shown to equals πs,t shere s = m − k
mod m and t = −2s · l mod N .
We now address the question of efficiently finding the
optimal permutation πk,l in the sense of Sec. V-B, i.e., the
permutation that maximizes the sum of adjusted reliabilities
(see (4)) over the first K components of the permuted codeword. For fixed k = 0, the set of permutations π0,l (i), l =
0, ..., N − 1 coincides with the set of cyclic permutations.
Finding the optimal cyclic permutation can efficiently be
achieved by computing the cumulative sum of the adjusted
reliabilities. The case of arbitrary k is adressed by observing
that πk,l (i) = π0,l (πk,0 (i)). The optimal permutation can
(5)
be achieved by first applying the permutation πk,0 and then
repeating the above procedure for cyclic codes. Finally, the
optimal permutation across all k is computed by combining the
above results for each individual k = 0, ..., m−1. With respect
to computation latency, we note that the cumulative sum can
be computed in logarithmic time by recursively splitting the
range 0, ..., N − 1.
Strictly speaking, the decoder of Fig. 5 violates the framework of Sec. IV, because the preprocessing and postprocessing
steps are not included in that framework. In the case of
BCH codes, however, this formal obstacle is easily overcome
by removing the two steps and redefining F to compensate.
Specifically, the preprocessing permutation can equivalently
be realized within F by permuting the vector |y| and manipulating the syndrome using identities detailed in [18]. The
selection of the optimal permutation (which depends only
on |y|) and the postprocessing step can be redefined to be
included in F. The results of Sec. IV (particularly, resilience
to overfitting) thus carry over to our setting.
R EFERENCES
[1] T. Richardson and R. Urbanke, “The capacity of low-density paritycheck codes under message-passing decoding,” IEEE Trans. Inf. Theory,
vol. 47, pp. 599–618, Feb. 2001.
[2] E. Nachmani, E. Marciano, L. Lugosch, Loren, W.J. Gross, D. Burshtein
and Y. Be’ery, “ Deep learning methods for improved decoding of linear
codes,” arXiv:1706.07043, 2017
[3] E. Nachmani, Y. Bachar, E. Marciano, D. Burshtein and Y. Be’ery
“Near Maximum Likelihood Decoding with Deep Learning,” Int. Zurich
Seminar on Inf. and Comm., 2018
[4] T. J. O’Shea and J. Hoydis, “An introduction to machine learning
communications systems,” arXiv:1702.00832, 2017.
[5] T. Gruber, S. Cammerer, J. Hoydis, and S. t. Brink, “On deep learningbased channel decoding,” 51st Annual Conference on Inf. Sciences and
Systems (CISS), 2017.
[6] L. G. Tallini and P. Cull, “Neural nets for decoding error-correcting
codes,” Proc. IEEE Tech. Applicat. Conf. and Workshops Northcon95,
pp. 89–94, Oct. 1995.
[7] J.-L. Wu, Y.-H. Tseng, and Y.-M. Huang, “Neural network decoders for
linear block codes,” Int. Journ. of Computational Engineering Science,
vol. 3, no. 3, pp. 235–255, 2002.
[8] J. Bruck and M. Blaum, “Neural networks, error-correcting codes, and
polynomials over the binary n-cube.” IEEE Trans. Inf. Theory vol. 35(5),
pp. 976–987, 1989.
[9] M.P. Fossorier, S. Lin and J. Snyders. “Reliability-based syndrome
decoding of linear block codes.” IEEE Trans. Inf. Theory vol. 44(1)
pp. 388–398, Jan. 1998.
[10] M.P. Fossorier and S. Lin, “Soft-decision decoding of linear block codes
based on ordered statistics.” IEEE Trans. Inf. Theory, vol. 41(5), 1379–
1396, Sep. 1995.
[11] T. Richardson and R. Urbanke, “Modern coding theory,” Cambridge
university press. (2008)
[12] S. Lin and D. J. Costello. “Error control coding,” 2nd edition, Prentice
Hall, 2004.
[13] G. Wiechman and I. Sason, I., “Parity-check density versus performance
of binary linear block codes: New bounds and applications,” IEEE Trans.
Inf. Theory vol. 53(2), pp. 550–579, Jan. 2007.
[14] D. Burshtein, M. Krivelevich, S. Litsyn, S. and G. Miller, “Upper bounds
on the rate of LDPC codes,”. IEEE Trans. Inf. Theory vol. 48(9), pp.
2437–2449, Sep. 2002.
[15] I. Goodfellow, Y. Bengio, and A. Courville, “Deep learning.” MIT press,
2016.
[16] D. Kingma and J. Ba, “Adam: A method for stochastic optimization.”
arXiv preprint arXiv:1412.6980. Dec. 2014.
[17] I. Dimnik and Y. Be’ery, “Improved random redundant iterative HDPC
decoding.” IEEE Trans. Commu., 57(7), July 2009.
[18] F. J. MacWilliams and N. J. A. Sloane, The Theory of Error-Correcting
Codes. North-Holland, 1978.
[19] A. Graves, M. Abdel-rahman and G. Hinton, “Speech recognition
with deep recurrent neural networks” IEEE international conference on
acoustics, speech and signal processing (icassp) 2013.
[20] K. Cho, B. Van Merriënboer, C. Gulcehre , D. Bahdanau, F. Bougares
,H. Schwenk and Y. Bengio G. Hinton, “Learning phrase representations
using RNN encoder-decoder for statistical machine translation” arXiv
preprint arXiv:1406.1078 2014.
[21] S. Hochreiter and J. Schmidhuber, “Long short-term memory” Neural
computation 1997.
| 9cs.NE
|
Lazy Local Search Meets Machine Scheduling
Chidambaram Annamalai∗
arXiv:1611.07371v2 [cs.DS] 13 Dec 2016
December 14, 2016
Abstract
We study the restricted case of Scheduling on Unrelated Parallel Machines. In this
problem, we are given a set of jobs J with processing times pj and each job may be scheduled
only on some subset of machines Sj ⊆ M . The goal is to find an assignment of jobs to machines
to minimize the time by which all jobs can be processed. In a seminal paper, Lenstra, Shmoys,
and Tardos [LST87] designed an elegant 2-approximation for the problem in 1987. The question
of whether approximation algorithms with better guarantees exist for this classic scheduling
problem has since remained a source of mystery.
In recent years, with the improvement of our understanding of Configuration LPs, it now
appears an attainable goal to design such an algorithm. Our main contribution is to make
progress towards this goal. When the processing times of jobs are√either 1 or ǫ ∈ (0, 1), we
design an approximation algorithm whose guarantee tends to 1 + 3/2 ≈ 1.8660254, for the
interesting cases when ǫ → 0. This improves on the 2 − ǫ0 guarantee recently obtained by
Chakrabarty, Khanna, and Li [CKL15] for some constant ǫ0 > 0.
Keywords: scheduling, unrelated parallel machines, restricted assignment, configuration linear programs.
∗
Department of Computer Science, ETH Zurich. Email: [email protected].
1
Introduction
We study a special case of the problem of Scheduling on Unrelated Parallel Machines.
An instance I of this problem consists of machines M , jobs J, and a collection of positive processing
times {pij }i∈M,j∈J for every machine–job pair. The goal is to assign all the jobs to the available
machines such that the makespan of the resulting schedule is as small as possible. The makespan
of a schedule σ : J 7→ M is defined as
X
max
pij .
i∈M
j∈σ−1 (i)
It is one of the major open questions [WS11] in the field of approximation algorithms to better
understand the approximability of this problem. Curiously, the best known hardness result and
approximation algorithm for the problem today were simultaneously established in the 1987 paper
of Lenstra, Shmoys and Tardos [LST87].
Given the challenging nature of the problem, the road to a better understanding of its approximability has focused on two special cases that each isolate two difficult aspects associated with
it—on the one hand a given job j ∈ J may be assigned with a finite processing time to an unbounded
number of machines, and on the other hand, its processing time may vary considerably across those
machines. In the graph balancing case, only instances where every job may be assigned to at most
two machines with finite processing time are considered. In the restricted case, often referred to as
the restricted assignment makespan minimization problem, the input processing times obey
the condition pij ∈ {∞, pj } for each i ∈ M and j ∈ J, where pj is a machine independent processing
time for job j. This has the natural interpretation that each job has a fixed processing time but
may only be scheduled on some subset of the machines. The latter special case is the focus of this
work.
The elegant 2-approximation of Lenstra et al. [LST90] works by rounding extreme point solutions
to a linear program called the Assignment LP. As this linear program has a matching integrality
gap, one of the natural directions was to develop a stronger convex relaxation for the restricted
case. An important step forward was made by Bansal and Sviridenko [BS06] who, among other
things, introduced the Configuration LP for the problem, which has exponentially many decision
variables. At the time, however, it was not clear if this new linear program was indeed stronger than
the Assignment LP in the sense of a worst case integrality gap. A breakthrough in this direction
was achieved by Svensson [Sve12] who proved that the integrality of the Configuration LP is no
worse than 33/17 ≈ 1.94, and, therefore, strictly better than the Assignment LP. Tantalizingly,
however, the proof of his result did not lead to an approximation algorithm with the same (or even
similar) guarantee. This is a fairly strange situation for a problem, as we usually expect integrality
gap upper bounds to accompany an approximation algorithm; indeed, it is often established as a
consequence of the latter.
The difficulties in turning the non-constructive aspects of Svensson’s result into an efficient algorithm mirror the situation faced in the restricted max-min fair allocation problem. In
the latter problem, following a long line of work [BS06, Fei08, HSS11, AFS12, PS12, AKS15], a
non-constructive integrality gap upper bound on the Configuration LP by Asadpour, Feige and
Saberi [AFS12] was turned into an approximation algorithm [AKS15]. Although one might reasonably hope for a similar resolution in the restricted case, it has proved to be elusive thus far. Indeed,
there have been works aimed at obtaining better-than-2 approximation algorithms for special cases,
2
and the gap between the 33/17 ≈ 1.94 (recently improved to 11/6 ≈ 1.83 [JR16]) integrality gap
upper bound and the 2 approximation algorithm of Lenstra et al. [LST90] persists. Ebenlendr, Křćal
and Sgall [EKS08] studied the case when jobs may be assigned to at most two machines and gave a
1.75 approximation in this case. Recently, Chakrabarty, Khanna, and Li [CKL15] designed a 2 − ǫ0
approximation algorithm, for some constant ǫ0 > 0, for the so-called (1, ǫ)-case where processing
times of jobs are drawn from a set of size two.
The special significance of (1, ǫ)-case is that it already proves to be hard from the perspective
of the Configuration LP—the best known 1.5 factor integrality gap instances are of this type—and
seems to adequately capture the difficulty of the restricted case in general. It is also interesting in
its own right, and in a sense it is the simplest case of the problem that is not yet fully understood.
Indeed, the case when processing times of all the jobs are equal can be solved optimally in polynomial
time: this amounts to finding a maximum flow in an appropriate flow network with jobs as sources,
machines as sinks, and setting the sink capacities to be uniformly equal to some guess on the
optimum makespan.
Our Results. After normalizing the job sizes, we assume without loss of generality that the jobs
are of size 1 or ǫ for some 0 < ǫ < 1. Our main result is a new purely flow based local search
algorithm for the (1, ǫ)-case.
Theorem 1.1. Let 0 < ǫ < 1. For an arbitrary but fixed ζ > 0, the (1, ǫ)-case of restricted assignment makespan minimization admits a polynomial time 1 + R(ǫ, ζ) approximation algorithm
where
∆ 1 √
3 − 2ǫ + ǫ + ζ.
R(ǫ, ζ) =
2
From the point of view of better-than-2 approximation algorithms, the hard case is when√ǫ → 0.
For this range of values, the approximation ratio guaranteed by Theorem 1.1 tends to 1 + 3/2 ≈
1.87. By balancing against a simple algorithm based on bipartite matching we also derive an
approximation guarantee independent of the size of small jobs.
Theorem 1.2. Let 0 < ǫ < 1. For an arbitrary but fixed ζ > 0, the (1, ǫ)-case of restricted
assignment makespan minimization admits a polynomial time 17/9+ζ approximation algorithm.
Note that 17/9 ≈ 1.89.
Our Techniques. We now give a very high level overview of the ideas behind the proof of Theorem 1.1 assuming that the optimum makespan is 1 for simplicity.
Our local search algorithm continually increases the number of jobs
P scheduled by an assignment
σ : M 7→ J ∪ {TBD}1 that satisfies the 1 + R makespan bound i.e., j∈σ−1 (i) pij ≤ 1 + R for each
i ∈ M . The algorithm takes a job j0 such that σ(j0 ) = TBD and attempts to assign it to one of
the machines M while respecting the makespan bound of 1 + R. In general, it may be required to
modify the assignment σ, which we call a partial schedule, before j0 can be successfully assigned
along with the rest of the jobs in σ −1 (M ). The algorithm identifies a set of machines M0 such
that sufficiently reducing the load on any of the machines i ∈ M0 suffices to assign j0 successfully.
Once again, to reduce the load on one of the machines in M0 a new set of machines M1 that is
disjoint from M0 is identified. In this way, in general there will be a sequence of disjoint machine
sets M0 , M1 , . . . , Mℓ such that reducing the load on some Mi allows the load to be reduced on some
1
TBD for “to be decided”
3
machine in M0 ∪ . . . ∪ Mi−1 . At some point, it may turn out that a lot of machines in Mℓ have
very little load, call them free machines, and therefore, it is possible for many of the jobs currently
scheduled on some machine in M0 ∪ . . . ∪ Mℓ−1 to be relocated to free machines in Mℓ , which
represents progress towards our goal of eventually scheduling j0 .
The first property we require is large progress which ensures that many free machines in Mℓ
implies that proportionally many jobs from σ −1 (M0 ∪ . . . ∪ Mℓ−1 ) can be relocated. Further, such
relocations should be performed in a way as to maintain the same property for the machine sets of
smaller indices.
Large progress by itself would not guarantee that the algorithm terminates quickly if it does
not happen often enough. To ensure that we find many free machines frequently, a second property
ensures |Mi | ≥ µ|M0 ∪ . . . ∪ Mi−1 | so that large progress happens at least once every Oµ (log |M |)
sets that are encountered.
These two properties–and maintaining them as the partial schedule is continually modified by
the algorithm–roughly correspond to the core technical difficulties of our approach. The proof of the
first property (see Section 3.3.4) makes fairly extensive use of properties of maximum flows, while
the second property (see Section 3.3.5) is based on new ideas for constructing dual unboundedness
certificates for the Configuration LP. It is interesting to note that the construction of such certificates
in the analysis, to prove the required property about the local search algorithm, effectively amounts
to a second algorithm that is merely used to determine an assignment of values to the dual variables
in the proof.
Our final algorithm is a highly structured local search that performs job relocations only through
flow computations in two kinds of flow networks. In essence, our approach can be seen as effectively
reducing the scheduling problem we started with to polynomially many maximum flow computations
in a structured fashion. The sets alluded to earlier correspond to certain “reachability graphs”
associated with ǫ jobs whereas the number of such sets at any time is bounded by a function of the
number of 1 jobs in the instance. We also stress that identifying such sets, which allow our analyses
to prove these properties, requires a careful design of the algorithm in the first place, and is not
clear at the outset that this can be achieved.
Organization. The rest of the paper is organized as follows. In Section 2, we briefly go over
some notation and state the Configuration LP for our problem. In Section 3.1 we explain some of
the basic concepts such as flow networks and subroutines for job relocations used in the final local
search algorithm. The algorithm and its running time analysis are presented in Sections 3.2 and 3.3
respectively, followed by the proof of the main theorem in the paper in Section 3.4. Proofs of some
statements missing in the body of the paper appear in Appendix A.
2
2.1
Preliminaries
Notation and Conventions
Let I be the given instance of restricted assignment makespan minimization. By scaling all
∆
the processing times in the instance, we assume without loss of generality that pmax = max{pj | j ∈
J} = 1. We use OPT to denote the optimal makespan for I.
P
P
For a subset of jobs S ⊆ J, we use the notation p(S) and pi (S) to refer to j∈S pj and j∈S pij
respectively. We define Γ : J 7→ 2M to be a function that maps each job j ∈ J to the set of
4
all machines that it can be assigned to with a finite processing time. The input processing times
{pij }i∈M,j∈J and {pj }j∈J satisfy
(
pj , if i ∈ Γ(j),
∀i ∈ M ∀j ∈ J.
pij =
∞, else.
For S
a collection of indexed sets {S0 , . . . , Sℓ } and 0 ≤ i ≤ ℓ we use the notation S≤i to refer to the
set ij=0 Sj .
2.2
The Configuration Linear Program
The Configuration LP is a feasibility linear program parametrized by a guess τ on the value of the
optimal makespan for I, and is simply denoted by CLP (τ ). A configuration for machine i is a set
of jobs with a total processing time at most τ on machine i. The collection of all such configurations
for i is denoted as C(i, τ ). CLP (τ ) ensures that each machine receives at most one configuration
fractionally, while ensuring that every job is assigned, also in the fractional sense. The constraints
of CLP (τ ) are described in (2.1).
X
xiC ≤ 1, ∀i ∈ M,
C∈C(i,τ )
X
X
i∈M C∈C(i,τ ) : j∈C
xiC ≥ 1,
∀j ∈ J,
(2.1)
x ≥ 0.
We can also write the dual of CLP (τ ) as follows.
X
X
max
zj −
yi
j∈J
yi ≥
X
j∈C
i∈M
zj ,
∀i ∈ M and C ∈ C(i, τ ),
(2.2)
y, z ≥ 0.
Let τ ∗ be the smallest value of τ for which CLP (τ ) is feasible. We refer to τ ∗ as the value of
the Configuration LP. Observe that τ ∗ is a lower bound on OPT. As pmax = 1, τ ∗ must be at least 1.
3
The (1, ǫ) Case
Let 0 < ǫ < 1. In the (1, ǫ)-case of restricted assignment makespan minimization, jobs
j ∈ J have one of only two possible sizes: 1 or ǫ. We partition the jobs accordingly into the sets
∆
∆
Jb = {j ∈ J | pj = 1} and Js = J \ Jb , which we will refer to as the sets of big and small jobs
respectively.
For the rest of the section fix some 0 < ǫ < 1 and ζ > 0. As ǫ and ζ are fixed constants,
we refer to R(ǫ, ζ) as simply R. To prove Theorem 1.1 we describe an algorithm that terminates
in polynomial time with a schedule of makespan at most τ ∗ + R for the given instance I. This
algorithm, described in Section 3.2, is a local search algorithm which continually increases the size
of a partial schedule.
5
Definition 3.1 (Partial schedule). A partial schedule is a map σ : J 7→ M ∪ {TBD} such that
(a) ∀ i ∈ M , pi (σ −1 (i)) ≤ τ ∗ + R,
(b) Jb ⊆ σ −1 (M ), and
(c) ∀ i ∈ M , |σ −1 (i) ∩ Jb | ≤ 1.
The size of a partial schedule σ is |σ −1 (M )|.
Remark 3.2. Following the description and analysis of our algorithms, it will become clear that
solving the Configuration LP on the input instance I, in order to determine τ ∗ , is not necessary.
For the moment, however, it may be assumed that it is somehow known. We remark that τ ∗ can
be computed in polynomial time upto any desired accuracy ν > 0 by using the ellipsoid algorithm
with an appropriate separation oracle [BS06].
Of course, a partial schedule σ of size |J| is a schedule of makespan at most τ ∗ + R for our
instance I. The following statement ensures that partial schedules exist in the first place.
Lemma 3.3. Suppose 1 ≤ τ ∗ < 2. Then, there is a map from Jb to M such that i) no two big jobs
are mapped to the same machine, and ii) every big job j ∈ Jb is mapped to a machine ij ∈ M such
that ij ∈ Γ(j). Furthermore, such a map can be computed in polynomial time.
3.1
Flow Networks for Job Relocations
Let σ be some partial schedule. We now define several quantities whose description depends on σ.
Definition 3.4 (Job Assignment Graphs). Gσ = (M ∪ J, E) is a directed bipartite graph with
machines and jobs in I as vertices. The edge set of Gσ is defined as
∆
E = {(i, j) | ∃ i ∈ M, j ∈ J : σ(j) = i} ∪ {(j, i) | ∃i ∈ M, j ∈ J : σ(j) 6= i, i ∈ Γ(j)}.
∆
We define the graph of small job assignments Gsσ = Gσ \ Jb and the graph of big job assignments
∆
Gbσ = Gσ \ Js .
∆
∆
Definition 3.5 (Big and Small Machines). Let Mσb = {i ∈ M | σ −1 (i) ∩ Jb 6= ∅} and Mσs = M \ Mσb ,
which we refer to as big machines and small machines respectively.
We need to define two flow networks which will facilitate the movement of the two kinds of jobs
we have in our instance I. We will speak of the maximum flow in these flow networks, even though
it may not necessarily be unique. In such cases it is implicitly assumed that there is fixed rule to
obtain a particular maximum flow given a flow network. We also assume that flow decompositions
of flows in such networks contain only source to sink paths (no cycles). First, we define the flow
network for big jobs.
Definition 3.6 (Flow Network for Big Jobs). For collections of machines S ⊆ Mσb , and T ⊆ Mσs ,
the flow network Hσb (S, T ) is defined on the directed graph Gbσ as follows. Each machine to job arc
has a capacity of 1 whereas all other arcs have infinite capacity. S and T are the sources and sinks
respectively in this flow network. Sinks have vertex capacities of 1. The value of maximum flow in
this flow network is denoted as |Hσb (S, T )|.
6
The interpretation of the flow network Hσb (S, T ) is that it allows us to compute the maximum
number of vertex disjoint paths in the graph Gbσ between the sets of vertices S and T .
Proposition 3.7. For any S ⊆ Mσb and T ⊆ Mσs , there are |Hσb (S, T )| vertex disjoint paths in Gbσ
with sources in S and sinks in T .
These vertex disjoint paths suggest an update of the partial schedule σ in the natural way.
Algorithm 3.1 formalizes the description of this task and Proposition 3.8 follows easily.
Algorithm 3.1 BigUpdate(σ, X): Update σ using flow paths in X.
Require: σ is a partial schedule and X is a flow in Hσb (Mσb , Mσs ) where
∀f ∈ X,
p(σ −1 (f ∩ Mσs )) ≤ τ ∗ + R − 1.
P ← Vertex disjoint paths corresponding to X as ensured by Proposition 3.7.
for all p = i0 , j0 , . . . , jkp −1 , ikp ∈ P do
for ℓ = 0, . . . , kp − 1 do
σ(jℓ ) ← iℓ+1 .
end for
end for
return σ.
Proposition 3.8. For any partial schedule σ, and flow X in Hσb (Mσb , Mσs ) such that ∀f ∈ X,
p(σ −1 (f ∩ Mσs )) ≤ τ ∗ + R − 1, BigUpdate(σ, X) returns a partial schedule σ ′ such that
(a) σ ′−1 (M ) = σ −1 (M ), and
(b) ∀f = i0 , j0 , . . . , jk−1 , ik ∈ X, σ ′−1 (i0 ) ∩ Jb = ∅.
Now we define the flow network for small jobs.
Definition 3.9 (Flow Network for Small Jobs). For two disjoint collections of machines S ⊆ Mσs
and T ⊆ M , the flow network Hσs (S, T ) is defined on the directed graph Gsσ as follows. The arcs
going from machines to jobs have capacity ǫ while arcs going from jobs to machines have infinite
capacity. S and T are the sources and sinks respectively in this flow network. The sinks have vertex
capacities are set as follows:
(
1 + τ ∗ + R − p(σ −1 (i)) − ǫ, if i ∈ Mσb ,
∀i ∈ T, c(i) =
τ ∗ + R − p(σ −1 (i)),
else.
The value of the maximum flow in this network is denoted as |Hσs (S, T )|.
By construction it is clear that the maximum flow in both flow networks (3.6) and (3.9) is finite.
By the max-flow min-cut theorem, infinite capacity arcs going from the source-side to the sink-side
will therefore not cross any minimum capacity. We will use this fact later in our proof.
Algorithm 3.2 interprets flows in Hσs (S, T ) as a collection of updates for σ. Proposition 3.10 is a
statement about the partial schedule σ ′ output by SmallUpdate(. . .), and the flow X ′ computed
at the end of the while loop in the procedure. For convenience we let f source and f sink denote the
source and sink vertices, respectively, of a flow path f .
7
Algorithm 3.2 SmallUpdate(σ, S, T ): Update σ using Hσs (S, T ).
Require: σ is a partial schedule, S ⊆ Mσs , T ⊆ M \ S.
X0 ← Maximum flow in the network Hσs (S, {i ∈ T | p(σ −1 (i)) ≤ τ ∗ + R − ǫ}).
X ← Augment X0 to a maximum flow in the network Hσs (S, T ).
while ∃f = i0 , j0 , . . . , jkf −1 , ikf ∈ X : p(σ −1 (f sink )) ≤ τ ∗ + R − ǫ do
for ℓ = 0, . . . , kf − 1 do
σ(jℓ ) = iℓ+1 .
end for
X ← X \ {f }.
end while
return σ.
Proposition 3.10. Let S ⊆ Mσs , T ⊆ M \ S, and X be the maximum flow in Hσs (S, T ). Then,
SmallUpdate(σ, S, T ) returns a partial schedule σ ′ and computes a maximum flow X ′ in Hσs′ (S, T )
at the end of the while loop in Algorithm 3.2 such that
(a) σ ′−1 (M ) = σ −1 (M ),
(b) ∀i ∈ S, p(σ ′−1 (i)) − ǫ · |{f ∈ X ′ | f source = i}| = p(σ −1 (i)) − ǫ · |{f ∈ X | f source = i}|, and
(c) ∀f ∈ X, p(σ ′−1 (f ∩ T )) > τ ∗ + R − ǫ.
(d) There is no path in the graph Gsσ′ from S to some machine i ∈ T such that
p(σ ′
−1
(i)) ≤ τ ∗ + R − ǫ.
Proof. The first three properties follow directly from the updates performed in the while loop of
Algorithm 3.2. To see that X ′ is a maximum flow in Hσs′ (S, T ), first observe that for each update of
the partial schedule σ maintained by the algorithm, along some flow path f ∈ X from σ (b) to σ (a)
in a single iteration of the while loop (superscripts for before and after), the graph Gsσ(a) can be
obtained from Gsσ(b) by simply reversing the directions of arcs of the flow path in question. Suppose
that X is a maximum flow in Hσs(b) (S, T ). By the max-flow min-cut theorem, the maximum flow
is associated with a minimum capacity cut of equal value, and the latter observation implies that
both the flow value in X \ {f } and the corresponding cut capacity in the new network Hσs(a) (S, T ) is
less, than the flow value in X and its corresponding cut capacity in Hσs(b) (S, T ), by ǫ. This implies
that X \ {f } is a maximum flow in Hσs(b) (S, T ).
The final property follows from the particular way in which the maximum flow X in Hσs (S, T ) is
constructed. Note that Algorithm 3.2 first computes a flow X0 that maximizes the flow to machines
in {i ∈ T | p(σ −1 (i)) ≤ τ ∗ +R−ǫ}, and then augments X0 to a maximum flow in Hσs (S, T ). Suppose
to the contrary that there is a path P in Gsσ′ from S to a machine i such that p(σ ′−1 (i)) ≤ τ ∗ +R−ǫ.
We know that one can obtain Gsσ′ from Gsσ by reversing the arc directions of all the flow paths in
X \X ′ i.e., the paths that were used to update the partial schedule in the while loop. So each arc in
P is either i) present in some path of X \ X ′ in the opposite direction, or ii) not present in the paths
of X \X ′ in the opposite direction and, therefore, also present in Gsσ . Now consider the residual flow
network of the flow X \ X ′ in Hσs (S, T ). It is now easy to see that P is an augmenting path in this
residual flow network because p(σ −1 (i)) + ǫ · |{f ∈ X \ X ′ | f sink = i}| = p(σ ′ −1 (i)) ≤ τ ∗ + R − ǫ,
and, therefore, c(i) ≥ τ ∗ + R − p(σ −1 (i)) ≥ ǫ · |{f ∈ X \ X ′ | f sink = i}| + ǫ. This, however,
contradicts the maximality of the flow X0 computed in the first step of the algorithm.
8
We now have the tools necessary to state our local search algorithm.
3.2
Flow Based Local Search
In this section we describe the local search algorithm that takes as input τ ∗ , a partial schedule σ, and
a small job j0 ∈ Js \ σ −1 (M ). It outputs a partial schedule σ ′ such that σ ′−1 (M ) = σ −1 (M ) ∪ {j0 }.
The algorithm is parameterized by three constants 0 < µ1 , µ2 , δ ≤ 1. They are defined as:
∆
µ1 = min{1, ζ}/4,
∆
(3.1)
µ2 = min{δ, ζ}/4,
∆ √
3 − 2ǫ − 1 /2.
δ=
The algorithm maintains certain sets of machines in layers L0 , . . . , Lℓ throughout its execution,
where ℓ is some dynamically updated index variable that always points to the last layer. A layer
Li is a tuple (Ai , Bi ) where Ai ⊆ Mσs , and Bi ⊆ M , except L0 which is defined to be ({j0 }, B0 ) for
some B0 ⊆ M . In addition to layers, the algorithm also maintains a collection of machines {Ii }ℓi=0
that will be disjoint from the machines in L≤ℓ .
We will describe the algorithm in a procedural style, and in the course of the execution the
algorithm, sets and other variables will be modified. Function calls are specified in the pseudocode
assuming call-by-value semantics. Concretely, this means that function calls have no side effects
besides the assignment of the returned values at the call site. We abuse notation slightly and use
Li to also refer to Ai ∪ Bi (for i = 0, as A0 is not a set of machines, we use L0 to just mean B0 ),
so that L≤ℓ refers to a set of machines. For a subset of machines N ⊆ M we use Rσs (N ) denote the
set of all machines reachable from vertices N in the graph Gsσ . Note that N ⊆ Rσs (N ) always holds.
The description is now found in Algorithm 3.3. We refer to the while loop in Step 3 of Algorithm 3.3 as the main loop of the algorithm. Observe that, in Step 4, Algorithm 3.4 is used
a subroutine, which constructs and returns a new layer while potentially modifying the partial
schedule σ maintained by Algorithm 3.3.
3.3
Running Time Analysis
∆
The state of the algorithm is defined to be the dynamic tuple S = (σ, ℓ, {Li }ℓi=0 , {Ii }ℓi=0 ) which
contains the variables and sets that are maintained by Algorithm 3.3. In the analysis it will be
useful to compare quantities before and after certain operations performed by the algorithm, and
we will consistently use S and S ′ to refer to the state of the algorithm before and after such an
operation. For example, if S and S ′ denote the states of the algorithm before Step 4 and after
Step 5 in Algorithm 3.3 respectively, then ℓ′ = ℓ + 1, Iℓ′ ′ = ∅, etc.
3.3.1
Basic Invariants of the Algorithm
By observing the description of Algorithms 3.3 and 3.4 we can conclude certain basic properties
which will come in handy when reasoning about its running time.
Proposition 3.11. Consider some state S of the algorithm.
(a) The sets in the collection {Ai }ℓi=1 ∪ {Bi }ℓi=0 ∪ {Ii }ℓi=0 are pairwise disjoint subsets of M .
9
Algorithm 3.3 LocalSearch(τ ∗ , σ, j0 ): Extend the partial schedule σ to include small job j0 .
Require: σ is a partial schedule, j0 ∈ Js \ σ −1 (M ).
s (Γ(j )).
⊲ Construction of layer L0
1: Set A0 ← {j0 }, B0 ← Rσ
0
2: Set ℓ ← 0 and I0 ← ∅.
3: while ∄ i ∈ B0 such that p(σ −1 (i)) ≤ τ ∗ + R − ǫ do
⊲ Main loop
4:
(σ, Aℓ+1 , Bℓ+1 ) ← BuildLayer(σ, {Li }ℓi=0 , {Ii }ℓi=0 ).
⊲ Construction of layer Lℓ+1
5:
Set ℓ ← ℓ + 1 and Iℓ+1 ← ∅.
6:
while ℓ ≥ 1 and |{i ∈ Aℓ | p(σ −1 (i)) ≤ τ ∗ + R − 1}| ≥ µ2 |Aℓ | do
7:
Set I ← {i ∈ Aℓ | p(σ −1 (i)) ≤ τ ∗ + R − 1}.
8:
(I0′ , . . . , Iℓ′ , X) ← CanonicalDecomposition(σ, {Li }ℓi=0 , {Ii }ℓi=0 , I).
9:
Set Ii ← Ii′ for all 1 ≤ i ≤ ℓ.
10:
if ∃ 1 ≤ r ≤ ℓ : |Ir | ≥ µ1 µ2 |Br−1 ∩ Mσb | then
11:
Choose the smallest such r.
12:
σ ← BigUpdate(σ, {f ∈ X | f ∩ Ir 6= ∅}).
13:
σ ← SmallUpdate(σ, Ar−1 , Br−1 ) unless r = 1.
14:
Br−1 ← Rσs (Ar−1 ) \ (Ar−1 ∪ L≤r−2 ∪ I≤r−2 ) unless r = 1.
15:
Discard all layers with indices greater than r − 1.
16:
Set ℓ ← r − 1.
17:
end if
18:
end while
19: end while
20: Update σ using a path from j0 to i in Gsσ where p(σ −1 (i)) ≤ τ ∗ + R − ǫ.
21: return σ.
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
function CanonicalDecomposition(σ, {Li }ℓi=0 , {Ii }ℓi=0 , I)
Let X be the maximum flow in Hσb (B0 ∩ Mσb , I≤ℓ ∪ I).
for 1 ≤ i ≤ ℓ − 1 do
Augment X to a maximum flow in Hσb (B≤i ∩ Mσb , I≤ℓ ∪ I).
end for
for 1 ≤ i ≤ ℓ do
Set Ii′ to be the collection of sinks used by flow paths from X with sources in Bi−1
end for
return (∅, I1′ , . . . , Iℓ′ , X).
end function
10
Algorithm 3.4 BuildLayer(σ, {Li }ℓi=0 , {Ii }ℓi=0 ): Construct and return a new layer.
1: Let S ← ∅.
2: while ∃i ∈ M : IsAddableQ(i, σ, {Li }ℓi=0 , {Ii }ℓi=0 , S) do
3:
σ ← SmallUpdate(σ, S ∪ {i}, T \ {i}) where T ← M \ (L≤ℓ ∪ I≤ℓ ∪ S).
4:
S ← S ∪ {i}.
5: end while
6: Aℓ+1 ← S.
s (A
7: Bℓ+1 ← Rσ
ℓ+1 ) \ (Aℓ+1 ∪ L≤ℓ ∪ I≤ℓ ).
8: return (σ, Aℓ+1 , Bℓ+1 ).
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
function IsAddableQ(i, σ, {Li }ℓi=0 , {Ii }ℓi=0 , S)
⊲ Decide if i can be added to S
if i 6∈ Mσs \ (L≤ℓ ∪ I≤ℓ ∪ S) then
return False.
end if
Set T ← M \ (L≤ℓ ∪ I≤ℓ ∪ S).
if |Hσb (B≤ℓ ∩ Mσb , S ∪ {i})| = |Hσb (B≤ℓ ∩ Mσb , S)| + 1 then
if |Hσs (S ∪ {i}, T \ {i})| ≥ |Hσs (S, T )| + p(σ −1 (i)) − (τ ∗ − 1 + R − δ) then
return True.
end if
end if
return False.
end function
(b) For each i = 1, . . . , ℓ, the sets Ai have not been modified since the last time Li was initialized
in some execution of Step 4 of Algorithm 3.3. Similarly, for i = 0, the sets A0 and B0 have
not been modified since the execution of Step 1.
(c) For each i ∈ B≤ℓ , p(σ −1 (i)) > τ ∗ + R − ǫ.
(d) For every j ∈ σ −1 (L≤ℓ ) ∩ Js , Γ(j) ⊆ L≤ℓ ∪ I≤ℓ .
Proof. (a) For a newly constructed layer Lℓ+1 in Step 4 of Algorithm 3.3, the sets Aℓ+1 and Bℓ+1
satisfy the properties by the construction of the set S and the setting in Step 7 Algorithm 3.4.
Further, Iℓ+1 is initialized to the empty set in Step 5 of Algorithm 3.3. In future updates
of Bℓ+1 (if any) in Step 14 of Algorithm 3.3, let S and S ′ be the states before Step 13 and
after Step 14 of Algorithm 3.3 respectively. From the description of Algorithm 3.2, we see
′
from the
that σ −1 (Ar−1 ) ⊇ σ ′−1 (Ar−1 ) = σ ′−1 (A′r−1 ), which then implies that Br−1 ⊇ Br−1
assignment in Step 14.
(b) This follows directly from the description of the algorithm.
(c) When a new layer Lℓ+1 is constructed during a call to BuildLayer(. . .), at the end of the
while loop in Step 2 of Algorithm 3.4, we show in Claim 3.19, that a maximum flow X in
Hσs (S, T ) is computed where T = M \ (L≤ℓ ∪ I≤ℓ ∪ S). So we can apply Proposition 3.10(d)
and conclude that after the assignment in Step 7, there is no machine i ∈ Bℓ+1 such that
p(σ −1 (i)) ≤ τ ∗ + R − ǫ. We can argue in exactly the same way in Steps 13 and 14 of
Algorithm 3.3.
11
(d) This follows from Step 7 of Algorithm 3.4 and Step 14 of Algorithm 3.3.
Definition 3.12 (Collapsibility of a layer). Layer L0 is collapsible if there is an i ∈ B0 such that
p(σ −1 (i)) ≤ τ ∗ + R − ǫ. For ℓ ≥ 1, Lℓ is collapsible if Aℓ contains at least µ2 |Aℓ | machines i such
that p(σ −1 (i)) ≤ τ ∗ + R − 1.
Note the correspondence between Definition 3.12 and the conditions in Steps 3 and 6 of Algorithm 3.3.
Lemma 3.13. At the beginning of each iteration of the main loop of Algorithm 3.3, none of the
layers L0 , . . . , Lℓ are collapsible. In particular, for all 1 ≤ i ≤ ℓ,
|{i′ ∈ Ai | p(σ −1 (i′ )) ≤ τ ∗ + R − 1}| < µ2 |Ai |.
Proof. In the first iteration of the main loop of the algorithm, ℓ = 0, and the statement follows from
the condition of the main loop of the algorithm. Assume the statement to hold at the beginning
of some iteration of the main loop of the algorithm with the state (σ, ℓ, {Li }ℓi=0 , {Ii }ℓi=0 ). Observe
that a new layer Lℓ+1 is created in Step 4 and it is tested for collapsibility in Step 6. Suppose that
the while loop in the latter step executes at least once (if it happens infinitely many times, we
are done). Let r denote the choice made in Step 11 in the last execution of the step. The layers
L0 , . . . , Lr−2 continue to be non-collapsible by induction and Proposition 3.11(b), and layer Lr−1 is
not collapsible because execution exits the while loop.
Lemma 3.14. At the beginning of each iteration of the main loop of the algorithm, for every
0 ≤ i ≤ ℓ − 1,
|Ii+1 | < µ1 µ2 |Bi ∩ Mσb |.
Proof. The sets I0 , . . . , Iℓ maintained by the algorithm start out initialized to ∅ in Step 5 of Algorithm 3.3 when the corresponding layer is created. They are modified only within the while loop of
Step 6 of Algorithm 3.3 through the computation of the canonical decomposition and assignment
in Step 9. Within this loop, in Step 11, the smallest 1 ≤ r ≤ ℓ such that |Ir | ≥ µ1 µ2 |Br−1 ∩ Mσb | is
chosen; layers with indices greater than r − 1 are discarded in Step 15; and ℓ is set to r − 1 at the
end of the loop in Step 16. Therefore, the claim follows.
3.3.2
Maximum Flows and Canonical Decompositions
We now recall some properties about network flows that will be of use later on in the proof our
main theorem. For basic concepts related to flows, such as residual flow networks and augmenting
paths, we refer the reader to the textbook by Cormen, Leiserson, Rivest and Stein [CLRS09].
Proposition 3.15. Let S ⊆ Mσb and T ⊆ Mσs . Let S ′ ⊆ Mσs and T ′ ⊆ M such that S ′ ∩ T ′ = ∅.
(a) Let X be the maximum flow in Hσb (S, T ) and let CX denote the minimum capacity cut corresponding to X i.e., the set of vertices reachable from S in the residual flow network of X.
Then, |Hσb (S, T ∪ {i})| > |Hσb (S, T )| for all i ∈ CX \ S.
(b) Let Y be the maximum flow in Hσs (S ′ , T ′ ) and let CY denote the minimum capacity cut corresponding to Y . For any i ∈ M \ CY such that i is not used as a sink by a flow path in Y ,
let the corresponding maximum flow in Hσs (S ′ ∪ {i}, T ′ \ {i}) be Y ′ and minimum capacity cut
be CY ′ . Then, CY ⊂ CY ′ and this inclusion is strict.
12
We state some consequences of the description of the procedure CanonicalDecomposition(. . .)
in Algorithm 3.3.
Proposition 3.16. For a given state S and I ⊆ Mσs , let (I0′ , . . . , Iℓ′ , X) be the tuple returned by
CanonicalDecomposition(σ, {Li }ℓi=0 , {Ii }ℓi=0 , I). Then, X is a maximum flow in the network
Hσb (B≤ℓ−1 ∩ Mσb , I≤ℓ ∪ I),
such that
(a) Ii′ is the collection of sinks used by flow paths from X with sources in Bi−1 for all i = 1, . . . , ℓ,
and I0′ = ∅,
′
)| = |Hσb (B≤i ∩ Mσb , I≤ℓ ∪ I)|, for all i = 0, . . . , ℓ − 1, and
(b) |Hσb (B≤i ∩ Mσb , I≤i+1
′
′ )|, for all i = 0, . . . , ℓ − 1.
(c) |Hσb (B≤i ∩ Mσb , I≤i+1
)| = |Hσb (B≤i ∩ Mσb , I≤ℓ
3.3.3
Relating Set Sizes within Layers
Lemma 3.17. Suppose that 1 ≤ τ ∗ < 2. At the beginning of each iteration of the main loop of the
algorithm, |B0 ∩ Mσb | ≥ 1.
Proof. After the execution of Step 1 of Algorithm 3.3, |A0 | = 1 and |B0 ∩ Mσb | ≥ 1. The latter
inequality follows from the feasibility of CLP (τ ∗ ) and the fact that τ ∗ + R − ǫ ≥ τ ∗ . We omit its
proof here since it is similar to Lemma 3.3. In the main loop of the algorithm, consider the first
time (if at all) the schedule of big jobs on machines in B0 is altered in Step 12. Then it must be
that |I1 | ≥ µ1 µ2 |B0 ∩ Mσb | > 0 from Step 10. Using Proposition 3.16(b) in Step 8, X contains
a set of flow paths connecting sources in B0 ∩ Mσb to I1 . Then, Proposition 3.8(b) implies that
|B0′ ∩ Mσb′ | < |B0 ∩ Mσb | after Step 12. Then, the condition of the main loop of the algorithm is no
longer satisfied since p(σ ′−1 (i)) ≤ τ ∗ + R − 1 ≤ τ ∗ + R − ǫ for some i ∈ B0′ . The condition of the
while loop in Step 6 is also not satisfied because ℓ = 0 after Step 16. Therefore, the main loop is
exited in this case.
Lemma 3.18. At the beginning of each iteration of the main loop of the algorithm, for every
1 ≤ i ≤ ℓ,
|Bi ∩ Mσb | > (δ(1 − µ2 ) − 2µ2 ) · |Ai |.
Proof. We first prove a general claim which will then ease the proof the lemma.
Claim 3.19. The while loop in Step 2 of Algorithm 3.4 that iteratively builds S and modifies σ
∆
satisfies the following invariant, where, T = M \ (L≤ℓ ∪ I≤ℓ ∪ S), as defined in Step 3, and X denotes
the maximum flow in the network Hσs (S, T ).
ǫ|X| ≥
X
i′ ∈S
p(σ −1 (i′ )) − (τ ∗ − 1 + R − δ) ,
∀f ∈ X, f sink ∈ Mσb .
13
Proof. Before the first iteration of the while loop, S = ∅ and the statement is vacuously true.
Suppose it is true before some iteration for a set S. Let T and X be as in the claim. If i ∈ M is
chosen in Step 2 then, from the description of the procedure IsAddableQ(. . .) in Algorithm 3.4,
we can conclude that
1. i ∈ Mσs \ (L≤ℓ ∪ I≤ℓ ∪ S),
2. |Hσb (B≤ℓ ∩ Mσb , S ∪ {i})| = |Hσb (B≤ℓ ∩ Mσb , S)| + 1, and
3. |Hσs (S ∪ {i}, T \ {i})| ≥ |Hσs (S, T )| + p(σ −1 (i)) − (τ ∗ − 1 + R − δ) .
By the induction hypothesis and the first property, i cannot be a sink of some flow path in X. So,
X is a valid flow in Hσs (S ∪ {i}, T \ {i}). Using the third property we therefore conclude that X
can be augmented to a maximum flow X ′ in Hσs (S ∪ {i}, T \ {i}) such that
X
p(σ −1 (i′ )) − (τ ∗ − 1 + R − δ) ,
ǫ|X ′ | ≥ ǫ|X| + p(σ −1 (i)) − (τ ∗ − 1 + R − δ) ≥
i′ ∈S∪{i}
where the second inequality uses the induction hypothesis. In Step 3, a call to SmallUpdate(. . .)
is made. In this call, a maximum flow in Hσs (S ∪ {i}, T \ {i}), say X̄, is computed at the beginning
of the while loop in Algorithm 3.2. By using Proposition 3.10(b), we can conclude that a partial
schedule σ ′ and a maximum flow X̄ ′ in Hσs′ (S ∪ {i}, T \ {i}) are computed at the of the while loop
which satisfy the property
X
−1
ǫ|X̄ ′ | ≥
p(σ ′ (i′ )) − (τ ∗ − 1 + R − δ) .
i′ ∈S∪{i}
Furthermore, Proposition 3.10(c) implies that f sink ∈ Mσb for all f ∈ X̄ ′ . This is because the
vertex capacities of small machine sinks i′ is defined to be τ ∗ + R − p(σ ′−1 (i′ )) in Definition 3.9,
p(σ ′−1 (i′ )) > τ ∗ + R − ǫ, and flow paths carry flows of value ǫ.
Let L0 , . . . , Lℓ denote the set of layers at the beginning of the current iteration. Fix some
1 ≤ i ≤ ℓ. By Lemma 3.13,
|{i′ ∈ Ai | p(σ −1 (i′ )) ≤ τ ∗ + R − 1}| < µ2 |Ai |.
Now, consider the iteration (some previous one) in which Li was constructed and let σ (b)
be the partial schedule at the end of Step 6 in Algorithm 3.4 during the corresponding call to
BuildLayer(. . .). Using Claim 3.19, after the assignments in Steps 6 and 7, X is a maximum flow
in Hσs(b) (Ai , Bi ) such that
X
−1
ǫ|X| ≥
p(σ (b) (i)) − (τ ∗ − 1 + R − δ) > δ(1 − µ2 )|Ai | − 2µ2 |Ai |,
i′ ∈Ai
where we use Lemma 3.13 in the final step along with the bound (τ ∗ − 1 + R − δ) ≤ 2. Now consider
−1
a sink f sink ∈ Mσb(b) used by some flow path f ∈ X. By Proposition 3.10(c), p(σ (b) (f ∩ T )) >
τ ∗ + R − ǫ. Definition 3.9 states that the vertex capacity c(f sink ) = 1 + τ ∗ + R − p(σ ′−1 (f ∩ T )) − ǫ
since f sink ∈ Mσb(b) from Claim 3.19. Thus, c(f sink ) < 1. This proves that at the iteration in which
Li was constructed, by flow conservation, |Bi ∩ Mσb(b) | > (δ(1 − µ2 ) − 2µ2 )|Ai |.
14
In the intervening iterations, the variable r in Step 11 of Algorithm 3.3 might have been chosen to
be i+1, and, therefore, |Bi ∩Mσb | may have reduced in Step 12 and |{i′ ∈ Ai | p(σ −1 (i′ )) > τ ∗ +R−1}|
may have reduced in Step 13. In such an event, all the layers following layer Li would have been
discarded in Step 15. But in our current iteration, the set of layers is L0 , . . . , Lℓ . So, it only remains
to prove that |Bℓ ∩ Mσb | > (δ(1 − µ2 ) − 2µ2 ) |Aℓ | in the current iteration after one or more events
where r was chosen to be ℓ + 1 in the intervening iterations. The claim is true in this case too by
arguing as follows. In each intervening iteration, where r was chosen to be ℓ + 1, after Step 12, the
partial schedule changes from σ (b) to σ (a) . Due to the way sink capacities were set in Definition 3.9,
the new flow network Hσs(a) (Aℓ , Bℓ ) can be obtained from the old flow network Hσs(b) (Aℓ , Bℓ ) by
increasing the capacities of the machines Mσb(b) \ Mσb(a) (those machines in Bℓ from which big jobs
were moved in Step 12) by ǫ. Using the same arguments as before after applying Lemma 3.13 proves
the lemma.
3.3.4
Maintaining Multiple Sets of Disjoint Paths
We now prove an invariant of Algorithm 3.3 concerning the updates performed in Step 12 through
the statement σ ← BigUpdate(σ, {f ∈ X | f ∩ Ir 6= ∅}).
Theorem 3.20. At the beginning of each iteration of the main loop of the algorithm, for every
0 ≤ i ≤ ℓ − 1,
|Hσb (B≤i ∩ Mσb , Ai+1 ∪ I≤i+1 )| ≥ |Ai+1 |.
Furthermore, at the beginning of each execution of the while loop in Step 6 of Algorithm 3.3,
for all 0 ≤ i ≤ ℓ − 1, |Hσb (B≤i ∩ Mσb , Ai+1 ∪ I≤i+1 )| ≥ |Ai+1 |.
Proof. Consider the first statement. Before the first iteration of the main loop, there is nothing to
prove. Assume the statement to be true at the beginning of iteration u of the main loop for some
u ≥ 1. We will now show that the statement holds at the end of iteration u as well.
Following Step 4 of iteration u, the newly created layer Lℓ+1 has the property Iℓ+1 = ∅ and
b
|Hσ (B≤ℓ ∩ Mσb , Aℓ+1 )| = |Aℓ+1 |, by the construction of set S in Step 2 of Algorithm 3.4. For
0 ≤ i ≤ ℓ − 1, the statement holds by the induction hypothesis and the fact that σ was not changed
in Step 4 of Algorithm 3.3 in a way that affects the graph Gbσ (indeed, only small jobs are moved).
As in Step 5 of the algorithm, we also update ℓ to ℓ + 1 in this proof, and so we have at the end of
Step 5, for all 0 ≤ i ≤ ℓ − 1,
|Hσb (B≤i ∩ Mσb , Ai+1 ∪ I≤i+1 )| ≥ |Ai+1 |.
(3.2)
Now iteration u of the main loop could potentially involve one or more iterations of the while loop
in Step 6. If there are none, we are done using (3.2). The rest of the proof follows from Claim 3.21
which completes the induction on u, and also proves the second statement in Theorem 3.20.
Claim 3.21. At the end of the execution of some iteration of the while loop in Step 6 of Algorithm 3.3, for all 0 ≤ i ≤ ℓ − 1, |Hσb (B≤i ∩ Mσb , Ai+1 ∪ I≤i+1 )| ≥ |Ai+1 |, assuming it holds at the
beginning of the same iteration of the while loop.
Proof. Assume the statement to be true at the beginning of some iteration of the while loop in
question as stated in the hypothesis. We will now show that the statement holds at the end of that
iteration as well.
15
∆
As in Step 7, let I = {i ∈ Aℓ | p(σ −1 (i)) ≤ τ ∗ + R − 1}. In Step 8, we have the statement
(I0′ , . . . , Iℓ′ , X) ← CanonicalDecomposition(σ, {Li }ℓi=0 , {Ii }ℓi=0 , I),
′ ) which has the properties
which computes a specific maximum flow X in the network Hσb (B≤ℓ−1 , I≤ℓ
′
guaranteed by Proposition 3.16. Recall that the sets Ii are precisely the sinks used by flow paths
from X with sources in Bi−1 . Additionally, for the purposes of this proof, define Xi to be the set
of flow paths from X that end in sinks Ii′ , for each 1 ≤ i ≤ ℓ. Observe that X = X1 ∪ · · · ∪ Xℓ . Let
r be the choice made in Step 11.
In the algorithm, in Step 12, observe that only the flow paths Xr are used to modify the partial
schedule from σ to σ ′ and therefore also change the graph Gbσ to Gbσ′ . Step 13 does not alter the
set Ar−1 , even though it alters the partial schedule σ ′ by moving small jobs. Since it will not be
important to consider these updates for the proof of Claim 3.21, which is a statement about the
flow network for big jobs, we simply denote by σ ′ the partial schedule at the end of this step. After
this update, all layers following layer Lr−1 are discarded in Step 15 and ℓ is updated to r − 1 in
Step 16. So, to prove Claim 3.21, we only need to verify that for all 0 ≤ i ≤ r − 2,
′
)| ≥ |Ai+1 |.
|Hσb ′ (B≤i ∩ Mσb′ , Ai+1 ∪ I≤i+1
(3.3)
Fix i for the rest of the proof. Since 0 ≤ i ≤ r − 2 ≤ ℓ − 2, we have, by hypothesis, |Hσb (B≤i ∩
Mσb , Ai+1 ∪ I≤i+1 )| ≥ |Ai+1 |. So let Y be the maximum flow in Hσb (B≤i ∩ Mσb , Ai+1 ∪ I≤i+1 ), with
at least |Ai+1 | flow paths. We now interpret the flow paths X and Y as vertex disjoint paths in Gbσ
using Proposition 3.7.
For a collection of vertex disjoint paths P in Gbσ let SP ⊆ Mσb and TP ⊆ Mσs denote the set of
source and sink vertices respectively used by paths in P . Now, if it happens that SY ⊆ SX≤i+1 then
we are done. This is because there must be some X ′ ⊆ X≤i+1 such that X ′ has cardinality at least
|SY | = |Y | ≥ |Ai+1 |. Also, i + 1 ≤ r − 1 and therefore the paths Xr used to update the partial
schedule σ are disjoint from the paths X ′ and hence, X ′ continues to be present in the new graph
Gbσ′ following the update.
Our goal is to show this generally in Claim 3.22. Note that it immediately implies that even
after updating the partial schedule to σ ′ we will have a collection of at least |Y | ≥ |Ai+1 | many
′
vertex disjoint paths connecting the sources of Y to sinks in Ai+1 ∪ I≤i+1
in Gbσ′ . This proves (3.3)
and completes the proof of the claim.
Claim 3.22 (Source Alignment Lemma). There is a collection of vertex disjoint paths D in Gbσ such
′
that Xr ⊆ D, SY ⊆ SD\Xr , and each source in SY is connected by D to a sink in Ai+1 ∪ I≤i+1
.
Proof. Recall the two sets of disjoint paths X = X1 ∪ · · · ∪ Xℓ and Y we have defined in the graph
Gbσ . Paths in Xi connect sources in Bi−1 ∩ Mσb to sinks in Ii′ , whereas paths in Y connect sources
from B≤i ∩ Mσb to sinks in Ai+1 ∪ I≤i+1 . We now describe a finite procedure which demonstrates
the existence of paths D as in the claim.
16
D ← X.
while ∃s ∈ SY \ SD do
v ← s.
while ¬(v ∈ SD ∨ v ∈ TY ) do
if there is an incoming arc of some path p ∈ D at v then
Set v to be the previous vertex along p.
else if there is an outgoing arc of some path q ∈ Y at v then
Set v to be the next vertex along q.
end if
end while
Augment D using the set of edges traversed in the above loop.
end while
return D.
Before we analyze this procedure, we wish to point out that the augmenting step is well-defined.
Suppose D and Y are sets of disjoint paths before some iteration of the outer while loop; this is
clearly true before the first iteration. Let s ∈ SY \ SD be chosen as the starting vertex for v. If
at no point during the execution of the inner while loop, the if condition was satisfied, then it is
trivial to see that the augmentation step is well-defined since the set of edges that are traversed
in this case simply corresponds to some path y ∈ Y , whose source is s, and whose edges do not
intersect with the edges of paths from D. On the other hand, consider the first time the if condition
is satisfied. From that point onwards it is easy to see that the inner loop simply traverses the edges
of some particular path x ∈ D in reverse order until it reaches the source of x, and the inner loop
terminates. Here, we made use of the induction hypothesis that D is a collection of vertex disjoint
paths. Therefore, we can conclude that the total set of edges traversed in this case are composed
of two disjoint parts: i) edges from the prefix of the path y ∈ Y , whose source is s, followed by ii)
edges from the prefix of the path x ∈ D in reverse order. Furthermore, the unique vertex, say v ∗ ,
at which the two parts have an incoming arc must be a machine vertex (since job vertices in Gbσ
have at most one incoming arc and the two parts are disjoint). Also, v ∗ ∈ Mσb since v ∗ 6∈ TY , and
the paths y and x must intersect at the unique edge e∗ = (v ∗ , j ∗ ) where j ∗ ∈ Jb : σ(j ∗ ) = v ∗ . Thus,
deleting the set of edges traversed in the second part, and adding the set of edges traversed in the
first part corresponds to a valid augmentation of D.
We now prove that this procedure terminates after finite iterations of the outer while loop. We
claim that, in each iteration, either |SY \SD | decreases, or |SY \SD | stays the same and the quantity
∆
Q=
XX
y∈Y x∈D
N (x, y)
decreases, where N (x, y) is defined as the total number of non-contiguous intersections between a
pair of paths x and y in Gbσ (see Figure 1). Consider a particular iteration of the outer loop. If
the if condition is never satisfied during the execution of the inner loop, then, by the arguments
above, the number of disjoint paths in D increases after the augmentation, and further the vertex
s chosen in the outer loop becomes a new source of D after the augmentation. On the other hand,
suppose that the path chosen for augmentation is composed of two parts arising from two paths
y ∈ Y and x ∈ D as argued before. Further, let sx be the source of x, and suppose that sx ∈ SY , as
otherwise, once again |SY \ SD | decreases after augmenting D. Let y ′ ∈ Y be the path with source
17
x
y
Figure 1: The number N (x, y) of non-contiguous intersections between the pair of paths x ∈ D and
y ∈ Y depicted here is 2. Arcs from Gbσ that are neither present in D nor in Y are shown in black.
P
P
′
′
sx . After augmenting
D
it
is
seen
that
′ N (x, y ) <
x∈D N (x, y ), and for all other paths
x∈D
P
P
y ′′ ∈ Y \ {y ′ }, x∈D′ N (x, y ′′ ) ≤ x∈D N (x, y ′′ ), thereby proving that the procedure eventually
terminates. For an example execution of a single iteration of the outer while loop of the procedure,
see Figure 2.
At the end of the procedure, we have SY ⊆ SD . As SY ⊆ B≤i , by the invariant of the procedure
′
.
that we prove below in Claim 3.23, the paths in D with sources SY end in sinks from Ai+1 ∪ I≤i+1
Also, by Claim 3.23, Xr ⊆ D (because D3 = X≥i+2 ⊇ Xr ) and SY ⊆ SD\Xr (because B≥i+1 ⊇
SD3 ⊇ SXr ), which proves the claim.
Claim 3.23 (The D Invariant). D is a collection of vertex disjoint paths in Gbσ which can be partitioned into D1 ∪ D2 ∪ D3 such that:
′
(a) SD1 ⊆ B≤i , TD1 = I≤i+1
,
(b) SD2 ⊆ B≤i , TD2 ⊆ Ai+1 , and
(c) D3 = X≥i+2 .
Proof. Before the first iteration of the algorithm, D is initialized to X≤ℓ and therefore admits the
decomposition into D1 = X≤i+1 , D2 = ∅, D3 = X≥i+2 which satisfy all of the above invariants.
Notice here that i ≤ r − 2 ≤ ℓ − 2, so that this decomposition is well-defined.
Assume the D invariant to hold for a collection of disjoint paths D at the beginning of some
iteration of outer while loop. Following the augmentation let D ′ be the resulting collection of
18
y 00
v∗
y0
sy 2 S Y n S D
sx
Figure 2: There are three blue paths y, y ′ , y ′′ ∈ Y and one red path x ∈ D shown here. Starting
from the vertex sy ∈ SY \ SD , the outer while loop of the procedure defined in Claim 3.22 traverses
precisely the arcs which are dot-dashed and ends in sx , the source vertex of x, which in this case is
the same as the source vertex of y ′ and therefore sx ∈ SY .
disjoint paths. Let
∆
′
D1′ = {p ∈ D ′ | T{p} ⊆ I≤i+1
},
∆
D2′ = {p ∈ D ′ | T{p} ⊆ Ai+1 },
∆
′
}.
D3′ = {p ∈ D ′ | T{p} ⊆ I≥i+2
First we prove that D1′ , D2′ , D3′ defines a valid partition of D ′ . As it is clear that the sets
′ . Recall that the augmenting path
are disjoint, we only need to prove that TD′ ⊆ Ai+1 ∪ I≤ℓ
either ends in a source vertex of D or a sink vertex of Y . In the first case, no new sinks are
′ . In the second case, the augmenting path ends in a sink
introduced, i.e., TD′ = TD ⊆ Ai+1 ∪ I≤ℓ
′
from SY ⊆ Ai+1 ∪ I≤i+1 . Potentially this could introduce a sink from the set I≤i+1 \ I≤i+1
. But
′
ℓ
in this case the computed canonical decomposition {Ii }i=0 would not be maximal since we now
′
have a collection of |I≤i+1
| + 1 vertex disjoint paths in D ′ with sources from B≤i and sinks in
′
′
I≤i+1 ∪ I≤i+1 ⊆ I≤ℓ ∪ I. That is, this contradicts the fact that |Hσb (B≤i , I≤i+1
)| = |Hσb (B≤i , I≤ℓ ∪ I)|
guaranteed by Proposition 3.16(b).
In the remainder of the proof we show that the defined partition of D ′ satisfies the invariants.
Since we do not lose any sinks in the augmentation of D to D ′ , by a basic property of flow augmen′
′
′
tation, it is true that |D1′ | = |I≤i+1
|, and therefore TD1′ = I≤i+1
. Next, TD2′ ⊆ Ai+1 and TD3′ ⊆ I≥i+2
follow by definition.
Since the path used to augment D started from a vertex in SY \SD and SY ⊆ B≤i it is clear that
SD1′ , SD2′ ⊆ B≤i unless we encountered a path from D3 during the augmentation process. However,
19
that would lead to a contradiction to the property of the canonical decomposition I1′ ∪ · · · ∪ Iℓ′ that
′ )| (note here again that i+1 ≤ r −1 ≤ ℓ−1) by Proposition 3.16(c).
′
)| = |Hσb (B≤i , I≤ℓ
|Hσb (B≤i , I≤i+1
Therefore, we also have that SD1′ , SD2′ ⊆ B≤i . Finally, since we did not encounter any edges of D3
during the augmentation process, we not only have that SD3′ ⊆ B≥i+1 but that D3 = D3′ .
3.3.5
Proportionally Many Options for Big Jobs
Theorem 3.24. Suppose that 1 ≤ τ ∗ < 2. At the beginning of each iteration of the main loop of
the algorithm, for every 0 ≤ i ≤ ℓ − 1,
|Ai+1 | ≥ µ1 |B≤i |.
The statement also holds at the beginning of each iteration of the while loop of Step 6 of Algorithm 3.3.
Proof. Let L0 , . . . , Lℓ be the set of layers maintained by the algorithm at the beginning of iteration
of the main loop. It is sufficient to prove that following the construction of layer Lℓ+1 in Step 4 of
Algorithm 3.3,
|Aℓ+1 | ≥ µ1 |B≤ℓ |.
The rest follows by applying Proposition 3.11(b). Suppose that the set S at the end of the while loop
in Step 2 of Algorithm 3.4 is smaller than µ1 |B≤ℓ |. We now describe an assignment of values to the
variables (y, z) from the dual of CLP (τ ∗ ) (defined in Section 2.2) in four parts.
we will show
P Then,P
that the final setting of dual variables (ȳ, z̄) obtained in this way satisfies j∈J z̄j − i∈M ȳi > 0,
while also respecting the constraints of (2.2). It then follows that the dual of CLP (τ ∗ ) is unbounded
because for any λ > 0, (λȳ, λz̄) is a feasible dual solution as well. Therefore, by weak duality,
CLP (τ ∗ ) must be infeasible, a contradiction. We now proceed to execute this strategy.
Part I: Layers We set positive dual values to all machines that appear in the sets L≤ℓ ∪ I≤ℓ , and
the corresponding jobs, as follows:
−1
(
R − δ, ∃i ∈ L≤ℓ : j ∈ σ (i) ∩ Jb ,
τ ∗ , i ∈ L≤ℓ ∪ I≤ℓ ,
(1)
(1)
zj = ǫ,
yi =
∃i ∈ L≤ℓ : j ∈ σ −1 (i) ∩ Js ,
0,
else.
0,
else.
The objective function of the assignment (y (1) , z (1) ) can be lower bounded as:
X (1) X (1)
zj −
yi ≥ (2R − δ − ǫ − 1)|B≤ℓ | − τ ∗ |I≤ℓ | − (1 − µ2 )(1 − R)|A≤ℓ | − µ2 τ ∗ |A≤ℓ |.
j∈J
i∈M
Let us explain the lower bound. For each machine i ∈ B≤ℓ , p(σ −1 (i)) > τ ∗ + R − ǫ from PropoP
(1)
(1)
sition 3.11(c). This allows us to derive j∈σ−1 (i) zj − yi > τ ∗ + R − ǫ − (1 − (R − δ)) − τ ∗ =
P
(1)
(1)
2R−δ −ǫ−1. For the machines i ∈ I≤ℓ , we have the trivial lower bound j∈σ−1 (i) zj −yi ≥ −τ ∗ .
P
(1)
(1)
Next, for each machine i ∈ A≤ℓ such that p(σ −1 (i)) > τ ∗ − 1 + R, we have j∈σ−1 (i) zj − yi ≥
20
(CX [ Mσs ) n (M (1) [ S)
e
S
i
CX
B≤` \ Mσb
Figure 3: The cut CX arising from the flow X, whose paths are colored red, in the flow network
Hσb (B≤ℓ ∩ Mσb , S) is shown here. All machines except those inside dotted circles are big machines.
Every big machine is matched to a distinct big job except for the machine with the outgoing
machine-job arc e, which crosses CX . The matching edges are dot-dashed and i has unit capacity.
τ ∗ − 1 + R − τ ∗ = −1 + R, whereas for the rest of the machines in A≤ℓ , we use the trivial lower
bound −τ ∗ . Thus, using Lemmas 3.13 and 3.14, we have
X (1) X (1)
zj −
yi ≥ (2R − δ − ǫ − 1 − τ ∗ µ1 µ2 ) |B≤ℓ | − (1 − µ2 )(1 − R)|A≤ℓ | − µ2 τ ∗ |A≤ℓ |. (3.4)
j∈J
i∈M
(1)
At this point we have assigned a positive zj value to big jobs j assigned by σ to machines in
L≤ℓ . However there could potentially be machines i ∈ M \ (L≤ℓ ∪ I≤ℓ ) such that i ∈ Γ(j) as well.
Therefore, the current assignment (y (1) , z (1)
solution
P) does not necessarily constitute a dual feasible
∗
since it might violate the inequality yi ≥ j∈C zj , for a configuration C = {j} ∈ C(i, τ ) consisting
∆
of a single big job. We now fix this in the next part. For convenience, let M (1) = L≤ℓ ∪ I≤ℓ .
Part II: Approximate Matchings Consider the flow network of big jobs Hσb (B≤ℓ ∩ Mσb , S) that
was used to construct the set S. By the construction of the set S in the algorithm, there is a flow
X in this network of value |S|. This flow naturally defines a minimum capacity cut: the cut CX
is defined as the set of reachable jobs and machines from B≤ℓ ∩ Mσb in the residual flow network
∆
corresponding to X in Hσb (B≤ℓ ∩ Mσb , S). Let M (2) = (CX ∩ Mσb ) \ M (1) . We extend the assignment
(y (1) , z (1) ) described in the first part in the following way.
(
(
R − δ, ∃i ∈ M (2) : j ∈ σ −1 (i) ∩ Jb ,
R − δ, i ∈ M (2) ,
(2)
(2)
zj =
yi =
0,
else.
0,
else.
The capacity of the cut CX is |S| by the max-flow min-cut theorem. This in particular implies that
no job-machine arcs can cross this cut as such arcs have infinite capacity. In other words, the only
arcs crossing CX are machine-job arcs and machine-supersink arcs where the machine arises from
S (recall that sink vertices have vertex capacity 1 in Hσb (B≤ℓ ∩ Mσb , S) according to Definition 3.6).
21
ρ ← σ.
U ← S.
V ← M \ (M (1) ∪ U ).
Y ← Maximum flow in Hσs (U, V ).
CY ← Mincut corresponding to Y in Hσs (U, V ).
while ∃ i ∈ (CX ∩ Mρs ) \ (M (1) ∪ CY ) do
Augment Y to a maximum flow in Hρs (U ∪ {i}, V \ {i}).
⊲ This is well-defined
s
CY ← Mincut corresponding to Y in Hρ (U ∪ {i}, V \ {i}).
for f ∈ Y : the sink of f belongs to ((CX ∩ Mρs ) \ (M (1) ∪ CY )) do
Update ρ by using the flow path f .
Y ← Y \ {f }.
end for
⊲ CY is still the mincut corresponding to Y
U ← U ∪ {i}.
V ← V \ {i}.
end while
return CY
Figure 4: Mincut Growing Procedure
For every big machine i that is present in CX , the corresponding big job assigned to i by σ is
also present in CX with the exception of at most |S| big jobs as shown in Figure 3 (corresponding
to the machine-job arcs that cross the cut CX ). Therefore, the total loss incurred in this step is at
most |S|. In other words,
X (2) X (2)
zj −
yi ≥ −(R − δ)|S|.
(3.5)
j∈J
i∈M
Part III: Growing Mincuts In this part and the next, we assign positive dual values to machines
in S, machines in (CX ∩ Mσs ) \ M (1) , and some other machines, to complete the description of our
dual assignment. To make such an assignment, we will use the algorithm described in Figure 4 in
the analysis.
The properties of the above procedure that we require in the proof are encapsulated in the
following claim which we will prove inductively.
Claim 3.25. The while loop of the above procedure maintains the following invariants.
(a) ρ is a partial schedule.
(b) Y is a maximum flow in Hρs (U, V ) and CY is the corresponding mincut.
(c) The value of maximum flow Y can be upper bounded as
X
|Hρs (U, V )| ≤ (τ ∗ + R)|S| +
(p(ρ−1 (i)) − (τ ∗ − 1 + R − δ)).
i∈U \S
(d) There is no flow path f ∈ Y that ends in a sink belonging to (CX ∩ Mρs ) \ (M (1) ∪ CY ).
(e) For each i ∈ (CX ∩ Mρs ) \ (M (1) ∪ CY ),
|Hρs (U ∪ {i}, V \ {i})| < |Hρs (U, V )| + (p(ρ−1 (i)) − (τ ∗ − 1 + R − δ)).
22
CY 0
g Y 00
C Ys
P
i0
i
CX
I = (CX \ Mρs0 ) n (M (1) [ CY 0 )
0 ∆
Figure 5: The induction step in the proof of Claim 3.25. For an i ∈ (CX ∩ Mρss ) \ (M (1) ∪ CYs ), Y ′ is
the maximum flow in Hρss (Us ∪{i}, Vs \{i}) shown in blue. It is partitioned as Y ′′ ∪P . The direction
of the arcs in P is reversed, as shown in red, during the for loop from the procedure defined in Part
III. The mincuts corresponding to the flows Y ′ and Ys obey the inclusion CY ′ ⊃ CYs .
Proof. Before the first iteration of the while loop, Claim 3.25(c) is satisfied because U = S and ρ = σ
is a partial schedule; Claim 3.25(d) is satisfied because Y is a flow that has the properties guaranteed
by Claim 3.19; Claim 3.25(e) follows from the fact that the while loop in Step 2 of Algorithm 3.4 was
exited. The last claim needs some more explanation. Note that every i ∈ (CX ∩ Mσs ) \ (M (1) ∪ CY )
satisfies the following two properties:
• i ∈ Mσs \ (L≤ℓ ∪ I≤ℓ ∪ S), and
• |Hσb (B≤ℓ ∩ Mσb , S ∪ {i})| = |Hσb (B≤ℓ ∩ Mσb , S)| + 1.
The second property follows from Proposition 3.15(a) because i ∈ CX . Therefore, for all such i, it
must be the case that Hσs (S ∪ {i}, T \ {i}) < Hσs (S, T ) + p(σ −1 (i)) − (τ ∗ − 1 + R − δ) .
Suppose that the statement is true until the beginning of some iteration of the while loop. Let
Ys be the maximum flow in Hρss (Us , Vs ) and CYs be the corresponding minimum cut maintained by
this procedure. We now show it holds at the end of that iteration as well. Let the machine chosen in
this iteration be i ∈ (CX ∩Mρss )\(M (1) ∪CYs ). The augmentation step is well defined because Ys is a
flow in Hρss (Us , Vs ) that does not use any flow path with a sink belonging to (CX ∩Mρss )\(M (1) ∪CYs )
as guaranteed by Claim 3.25(d). Therefore, Ys is also a feasible flow in Hρss (Us ∪ {i}, Vs \ {i}), which
can be augmented to a maximum flow in that network, say Y ′ . Let CY ′ be the corresponding mincut
that is computed in the procedure. We remark that CYs ⊂ CY ′ where the inclusion is strict, because
i ∈ CY ′ \ CYs , which follows from Proposition 3.15(b). We use this fact later.
The execution now enters a for loop that modifies the flow Y and the partial schedule ρ maintained by the procedure, which currently assume the values Y ′ and ρs respectively. Let Y ′′ ⊆ Y ′
∆
and ρ′ be the state of Y and ρ at the end of the for loop respectively, so that P = Y ′ \ Y ′′ is
precisely the set of flow paths used to update the partial schedule ρ maintained by the procedure. It
is seen that the Gsρ′ is obtained from the graph Gsρs by reversing the directions of the arcs contained
in the paths of P . For an illustration, see Figure 5.
23
As only flow paths P ⊆ Y ′ ending in small machine sinks were used to update ρ, ρ′ is still
a partial schedule (recall that Y ′ is a flow in the network Hρss (Us ∪ {i}, V \ {i})), which proves
Claim 3.25(a). We also claim that Claim 3.25(b) holds.
Claim 3.26. Y ′′ is a maximum flow in Hρs′ (Us ∪ {i}, Vs \ {i}) and CY ′ is the corresponding minimum
cut.
Proof. The first part is true, because, after every single iteration of the for loop, the value of the
flow Y decreases by ǫ and so does the capacity of the arcs in Hρs (U ∪ {i}, V \ {i}) crossing the cut
CY ; since the value of the flow Y ′ was equal to the capacity of the minimum cut CY ′ before the for
loop, by the max-flow min-cut theorem, the claim follows (we apply the converse of the max-flow
min-cut theorem at the end).
The second part is true as well, because the set of vertices reachable from Us ∪ {i} in the reduced
flow network corresponding to the flow Y ′ in Hρss (Us ∪ {i}, Vs \ {i}) (defined to be CY ′ ) is the same
as the set of vertices reachable from Us ∪ {i} in the reduced flow network corresponding to the flow
Y ′′ in Hρs′ (Us ∪ {i}, Vs \ {i}).
Next, using Claims 3.25(e) and 3.25(c), we see that
(i)) − (τ ∗ − 1 + R − δ))
|Hρss (Us ∪ {i}, Vs \ {i})| < |Hρss (Us , Vs )| + (p(ρ−1
| s
{z
}
(∗)
≤ (τ ∗ + R)|S| +
= (τ ∗ + R)|S| +
X
i′ ∈Us \S
′
∗
(p(ρ−1
s (i )) − (τ − 1 + R − δ)) + (∗)
X
i′ ∈(Us ∪{i})\S
′
∗
(p(ρ−1
s (i )) − (τ − 1 + R − δ)).
In the final equality, note that i 6∈ S because i 6∈ CY ′ and CY ′ ⊇ Us ⊇ S (using the fact mentioned
in our earlier remark). In each iteration of the for loop, we saw that the value of the quantities
|Hρs (U ∪ {i}, V \ {i})| and p(ρ−1 (i)) reduces exactly by ǫ so that at the end of the for loop, we have
|Hρs′ (Us ∪ {i}, Vs \ {i})| ≤ (τ ∗ + R)|S| +
X
i′ ∈(Us ∪{i})\S
(p(ρ′−1 (i′ )) − (τ ∗ − 1 + R − δ)),
which proves Claim 3.25(c).
At the end of the for loop, Claim 3.25(d) is true as well because for every f ∈ Y ′′ the sink of
∆
f does not belong to I ′ = (CX ∩ Mρs′ ) \ (M (1) ∪ CY ′ ) by the postcondition of the loop, and using
Claim 3.26. It only remains to prove Claim 3.25(e). Suppose towards contradiction that there is an
i′ ∈ I ′ such that
|Hρs′ (Us ∪ {i, i′ }, Vs \ {i, i′ })| ≥ |Hρs′ (Us ∪ {i}, Vs \ {i})| + (p(ρ′
−1
(i′ )) − (τ ∗ − 1 + R − δ)).
Let F be some maximum flow in Hρs′ (Us ∪ {i, i′ }, Vs \ {i, i′ }) obtained by augmenting Y ′′ (which is
well-defined because Y ′′ does not have flow paths that use vertices in I ′ as sinks). Construct a new
flow network H ′ from Hρss (Us ∪ {i, i′ }, Vs \ {i, i′ }) by adding a copy of the vertex i′ and call it i′dummy
(with identical neighborhood structure and vertex capacity). Interpret the flow P in H ′ so that
none of the flow paths use i′ as as sink (they may use i′dummy however). In the residual flow network
corresponding to this flow in H ′ , use the flow paths of F to augment the flow. This is well-defined
24
because of the way the graphs Gsρ′ and Gsρs are related (recall that, to obtain the former from the
latter, we just need to reverse the directions of the arcs of paths in P ). It is important to note here
that the resulting flow only contains paths and no cycles. Through this process we obtain a flow of
value |F | + |P | in the network H ′ . By assumption,
|F | + |P | ≥ |Y ′′ | + (p(ρ′
|
′
= |Y | + (∗)
−1
(i′ )) − (τ ∗ − 1 + R − δ)) +|P |
{z
}
(∗)
= |Ys | + fi + (∗).
The first equality follows from the definition of P ; the second equality follows from the fact that Ys
was augmented in the flow network |Hρss (Us ∪ {i}, Vs \ {i})| to Y ′ so that Y ′ has exactly fi value
flow paths with sources at i, and |Ys | value flow paths with sources in Us (here we make use of the
fact that Ys is a maximum flow). Therefore, we have a flow of value at least |Ys | + fi + (∗) in the
network H ′ . Since the latter flow was constructed by augmenting maximum flows, we can deduce
that the it is composed of |Ys | value flow paths originating at Us , fi value flow paths originating at
i and the rest originating at i′ . Deleting all flow paths leading to i′dummy , we have a resulting flow
of value at least |Ys | + fi + (∗) − li′ , where li′ is the value of flow paths in P that end in i′ . Owing
′
to the way in which we updated ρ in the for loop, we can see that li′ = p(ρ′−1 (i′ )) − p(ρ−1
s (i )).
−1
′
∗
Therefore there is a flow of value at least |Ys | + (p(ρs (i )) − (τ − 1 + R − δ)) in the network
Hρss (Us ∪ {i′ }, Vs \ {i′ }), which then implies
|Hρss (Us ∪ {i′ }, Vs \ {i′ })| ≥ |Hρss (Us , Vs )| + (p(ρs −1 (i′ )) − (τ ∗ − 1 + R − δ)),
contradicting Claim 3.25(e).
Returning to our proof, we now run this procedure with one modification: we add i to the set
U maintained by the procedure only if, in addition to the condition in the while loop, the new set
CY′ would have a size at least |CY | + 2. Let ρf and CY,f be the reallocation policy and the cut at
the end of the exeuction of this modified procedure. We have as a postcondition that executing the
body of the while loop once with an i ∈ (CX ∩ Mρsf ) \ (M (1) ∪ CY,f ) would result only in a set of
size |CY,f | + 1 (note that i would be the new element in that case). Extend the dual assignment as
follows.
(
(
∗,
ǫ, j ∈ ρ−1
τ
i
∈
C
,
Y,f
(3)
(3)
f (i) ∩ Js : i ∈ CY,f ,
zj =
yi =
0,
else.
0, else.
We now need to bound the total loss incurred in this part of the proof. Suppose the while
loop in the procedure executes t ≥ 0 times. Let Uf , Vf , and Yf be the state of the (remaining)
variables at the end of the procedure. For convenience assume that Uf = S ∪ {i1 , . . . , it }, where the
numbering follows the order in which the machines are added to the variable U in the while loop.
By Claim 3.25(b), Yf is a maximum flow in Hρsf (Uf , Vf ) and CY,f is the corresponding mincut.
∆
∆
Let P = (CY,f \ Uf ) ∩ Mρsf and Q = CY,f ∩ Mρbf . Since the size of the mincut in the variable CY
increased by at least 2 in each iteration, we have at the end that |P | + |Q| ≥ t.
By the max-flow min-cut theorem, the value of the maximum flow equals the capacity of the
25
minimum cut, and therefore, by Claim 3.25(c),
t
X
∗
(p(ρ−1
(τ + R)|S| +
f (ij )) − (τ − 1 + R − δ))
∗
j=1
X
>
ci +
X
X
(τ ∗ + 1 + R − p(ρ−1
(τ ∗ + R − p(ρ−1
f (i)) − ǫ),
f (i))) +
i∈Q
i∈P
i∈S∪{i1 ,...,it }∪P ∪Q
where ci is the total capacity of machine-job arcs with i as one endpoint crossing the minimum cut
CYf . The terms on the left together upper bound the value of maximum flow in the final network
Hρsf (Uf , Vf ), whereas the terms on the right count the contributions to the minimum cut arising
from machine-job arcs and machine-sink arcs. Splitting the first sum on the right,
t
X
∗
(p(ρ−1
(τ + R)|S| +
f (ij )) − (τ − 1 + R − δ))
∗
j=1
>
X
ci +
i∈S
X
ci +
i∈{i1 ,...,it }∪P
X
ci +
i∈Q
X
X
(τ ∗ + 1 + R − p(ρ−1
(τ ∗ + R − p(ρ−1
f (i)) − ǫ).
f (i))) +
i∈Q
i∈P
After rearranging the terms,
(τ ∗ + R)|S| − t(τ ∗ − 1 + R − δ) >
X
i∈S
ci −
X
i∈{i1 ,...,it }∪P
(p(ρ−1
f (i)) − ci )
X
∗
−
(p(ρ−1
f (i)) − 1 − ci ) + (|P | + |Q|)(τ + R) − |Q|ǫ,
i∈Q
we derive
X
−
ci +
i∈S
X
i∈{i1 ,...,it }∪P
(p(ρ−1
f (i)) − ci ) +
X
(p(ρ−1
f (i)) − 1 − ci )
i∈Q
> −(τ ∗ + R)|S| + t(τ ∗ − 1 + R − δ) + (|P | + |Q|)(τ ∗ + R) − |Q|ǫ
≥ −(τ ∗ + R)|S| + t(τ ∗ − 1 + R − δ) + (|P | + |Q|)(τ ∗ + R − ǫ).
(3.6)
We demonstrate that the assignment (y (3) , z (3) ) amortizes itself locally using (3.6).
26
X (3) X (3)
zj −
yi
j∈J
i∈M
X
=
i∈S∪{i1 ,...,it }∪P
≥−
|
X
ci +
i∈S
(p(ρ−1
f (i)) − ci ) +
X
(p(ρ−1
f (i))
i∈{i1 ,...,it }∪P
X
i∈Q
∗
(p(ρ−1
f (i)) − 1 − ci ) − τ (|S| + t + |P | + |Q|)
− ci ) +
{z
X
i∈Q
∗
(p(ρ−1
f (i)) − 1 − ci ) −τ (|S| + t + |P | + |Q|)
}
(3.6)
> −(τ ∗ + R)|S| + t(τ ∗ − 1 + R − δ) + (|P | + |Q|)(τ ∗ + R − ǫ) − τ ∗ (|S| + t + |P | + |Q|)
= −(τ ∗ + R)|S| + t(τ ∗ − 1 + R − δ) + (|P | + |Q|)(R − ǫ) − τ ∗ (|S| + t)
= −(2τ ∗ + R)|S| + t(R − δ − 1) + (|P | + |Q|)(R − ǫ)
≥ −(2τ ∗ + R)|S| + t(R − δ − 1) + t(R − ǫ)
= −(2τ ∗ + R)|S| + t
(2R − δ − ǫ − 1)
{z
}
|
≥ 0 follows from Claim A.1
∗
≥ −(2τ + R)|S|.
(3.7)
Part IV: The Rest As noted in Part III, we may now have machines i ∈ (CX ∩Mρsf )\(M (1) ∪CY,f )
that increase the size of the set CY,f described in the previous part by one. Let M (4) denote the set
of such machines; note that they must necessarily be a subset of Mσs (which is the same as Mρsf ).
By the postcondition of the modified procedure, we deduce that each machine in M (4) has at least
τ ∗ − 1 + R − δ processing time small jobs assigned to it by ρf such that each of those jobs can be
assigned to only machines in CY,f ∪ M (1) besides itself. Let
∆
(1)
}.
Si = {j ∈ ρ−1
f (i) ∩ Js | Γ(j) ⊆ {i} ∪ CY,f ∪ M
We set the dual values of these machines as follows.
(4)
yi
=
(P
0,
j∈ρ−1
f (i)
(4)
zj ,
i ∈ M (4) ,
(4)
zj
else.
=
(
ǫ,
0,
j ∈ Si : i ∈ M (4) ,
else.
The Dual Assignment Before we describe our final dual assignment (ȳ, z̄), let us note that the
supports of (y (1) , z (1) ), (y (3) , z (3) ) and (y (4) , z (4) ) are disjoint by construction. Further, observe that
the support of (y (2) , z (2) ) may only intersect with the support of (y (3) , z (3) ), and is disjoint from
the other two. However, we can assume without loss of generality that they too are disjoint, as
machines that receive both positive y (2) and y (3) values will only help us in the later arguments.
(3)
(2)
The reasoning is that, for a machine i such that yi = τ ∗ and yi = R − δ, we will only consider
(3)
the contribution of yi to the final assignment ȳ in the feasibility whereas we will count both
contributions towards the objective function i.e., we prove that the dual objective function of the
final assignment is positive even after counting an extra contribution of R − δ for such machines.
Note that there can be no jobs in the intersection of the supports of the dual assignments from the
second and third parts. So we assume that the supports of the dual assignments from the four parts
are disjoint. Set (ȳ, z̄) to be the union of the four assignments in the natural way.
27
P
∗
Feasibility Our assignment (ȳ, z̄) to the dual variables is such that
j∈C z̄j ≤ τ for every
i ∈ M, C ∈ C(i, τ ∗ ) because z̄j ≤ pj for every j ∈ J. Therefore, the constraints of (2.2) involving
machines i for which ȳi = τ ∗ are satisfied.
This leaves us to only consider the machines whose dual values were set in Parts II and IV.
Let i ∈ M (2) and C ∈ C(i, τ ∗ ). By Proposition 3.11(d), the construction of the cut CY,f (note
that infinite capacity job-machine arcs cannot cross this cut), and the dual setting of z (4) (where
(4)
we assigned positive zj values only to jobs in Si for some i ∈ M (4) ), there can be no j ∈ C ∩ Js
such that z̄j > 0. As τ ∗ < 2, there is at most one big job in a configuration. Since it is assigned
(4)
a dual value of R − δ, all constraints involving such machines are satisifed.
P Now let i ∈ M ∗ and
∗
∗
C ∈ C(i, τ ). Recall that ȳi > τ − 1 + R − δ. If C contains a big job then j∈C z̄j ≤ R − δ + τ − 1.
If C does not contain big jobs, then,
X
X
X
z̄j =
z̄j ≤
z̄j = ȳi .
j∈C
j∈C:z̄j >0
j∈σ−1 (i)
The inequality in the middle deserves explanation. This follows from the assertion that any job
j ∈ C ∩ Js such that z̄j > 0 must be part of σ −1 (i) by Proposition 3.11(d), the construction of the
cut CY,f , and the dual setting of z (4) .
P
P
Positivity Now that we have described our dual assignment, we show j∈J z̄j − i∈M ȳi > 0 by
counting the contributions to the objective function from the dual variable settings in each of the
four previous parts.
From (3.4) and (3.5), the total gain in the first and second part is at least
(2R − δ − ǫ − 1 − τ ∗ µ1 µ2 ) |B≤ℓ | − (1 − µ2 )(1 − R)|A≤ℓ | − µ2 τ ∗ |A≤ℓ | − |S|.
In the third part, using (3.7), the total loss is at most (2τ ∗ + R)|S|. In the fourth part there is no
net loss or gain. So, we can lower bound the objective function value of our dual assignment (ȳ, z̄)
as follows, making use of Lemma 3.17 and Lemma 3.18 in the second inequality.
X
X
z̄j −
ȳi
j∈J
i∈M
≥ (2R − δ − ǫ − 1 − τ ∗ µ1 µ2 ) |B≤ℓ | − (1 − µ2 )(1 − R)|A≤ℓ | − µ2 τ ∗ |A≤ℓ | − |S| − (2τ ∗ + R)|S|
≥ (2R − δ − ǫ − 1 − τ ∗ µ1 µ2 − (1 + 2τ ∗ + R)µ1 ) |B≤ℓ | − (1 − µ2 )(1 − R)|A≤ℓ | − µ2 τ ∗ |A≤ℓ |
≥ (2R − δ − ǫ − 1 − τ ∗ µ1 µ2 − (1 + 2τ ∗ + R)µ1 ) (δ(1 − µ2 ) − 2µ2 ) · |A≤ℓ | − ((1 − µ2 )(1 − R) − µ2 τ ∗ ) |A≤ℓ |
≥ (2R − δ − ǫ − 1 − τ ∗ µ1 µ2 − (1 + 2τ ∗ + R)µ1 ) (δ(1 − µ2 ) − 2µ2 ) − (1 − µ2 )(1 − R) − µ2 τ ∗ |A≤ℓ |.
|
{z
}
(∗)
Subtituting the values of R, µ1 , µ2 , δ as defined in the statement of Theorem 1.1 and (3.1), one
can verify (see Claim A.1) that the bracketed expression (∗) is strictly positive for every 1 ≤ τ ∗ < 2,
0 < ǫ < 1 and ζ > 0.
3.3.6
Polynomial Bound on Loop Iterations
Corollary 3.27. Suppose 1 ≤ τ ∗ < 2. In each execution of the while loop in Step 6 of Algorithm 3.3, the if condition in Step 10 is satisfied.
28
Proof. Consider the beginning of some iteration of the while loop in Step 6 of Algorithm 3.3 with
∆
a state S. By the condition in Step 6, ℓ ≥ 1 and |I| ≥ µ2 |Aℓ |, where I = {i ∈ Aℓ | p(σ −1 (i)) ≤
τ ∗ + R − 1}. Applying Theorem 3.20, |Hσb (B≤ℓ−1 ∩ Mσb , Aℓ ∪ I≤ℓ )| ≥ |Aℓ |. Since |I| ≥ µ2 |Aℓ | and
I ⊆ Aℓ this means that
|Hσb (B≤ℓ−1 ∩ Mσb , I ∪ I≤ℓ )| ≥ µ2 |Aℓ | ≥ µ1 µ2 |B≤ℓ−1 ∩ Mσb |,
where the second inequality follows from Theorem 3.24. By Proposition 3.16(a), at least one of
the sets Ii′ computed in Step 8 of Algorithm 3.3 must be of size at least µ1 µ2 |Bi−1 ∩ Mσb | for some
1 ≤ i ≤ ℓ.
Given the state S of the algorithm at some point during its execution, the signature of a layer
Li is defined as
!
i
k
j
1
∆
Bi ∩ Mσb
+ i,
si = log 1
1−µ1 µ2
η
∆
where η = (δ(1 − µ2 ) − 2µ2 ) µ1 > 0 by Claim A.2. The signature vector corresponding to the given
state is then defined as a vector in the following way:
∆
s = (s0 , . . . , sℓ , ∞).
Lemma 3.28. Suppose 1 ≤ τ ∗ < 2. The signature vector satisfies the following properties.
(a) At the beginning of each iteration of the main loop of the algorithm, ℓ = O(log |Jb |).
(b) The coordinates of the signature vector are well-defined and increasing at the beginning of each
iteration of the main loop of the algorithm.
Proof. Consider the beginning of some iteration of the main loop of the local search algorithm. Let
L0 , . . . , Lℓ be the set of layers maintained by the algorithm. Let 0 ≤ i ≤ ℓ. Observe that from the
moment layer Li was constructed until now, Ai remains unmodified (even though the assignment
of jobs by σ to machines in Aℓ may have changed). This is because Ai can be modified only if, in
some intervening iteration of the main loop, the variable r from Step 11 of Algorithm 3.3 is chosen
to be i′ for some i′ ≤ i. But in that case we discard all the layers Li′ , . . . , Lℓ in Step 15 and this
includes layer Li as well. Therefore, for 0 ≤ i ≤ ℓ − 1,
|Bi+1 ∩ Mσb |
Lem 3.18
>
(δ(1 − µ2 ) − 2µ2 ) |Ai+1 |
Thm 3.24
≥
(∗)
(δ(1 − µ2 ) − 2µ2 ) · µ1 |B≤i | ≥ η|B≤i ∩ Mσb |.
The second inequality above uses the fact that the layers L0 , . . . , Lℓ were not modified since construction as argued previously. As the final term in the chain of inequalities above is at least
η|Bi ∩ Mσb |, this proves (b). As the sets B0 , . . . , Bℓ are disjoint by construction,
|B≤i+1 ∩ Mσb | = |Bi+1 ∩ Mσb | + |B≤i ∩ Mσb |
Using (∗)
≥
(1 + η)|B≤i ∩ Mσb |.
As |B0 ∩ Mσb | ≥ 1 by Lemma 3.17, ℓ is O(log1+η |Mσb |) = O(log |Jb |), which proves (a).
Lemma 3.29. Suppose 1 ≤ τ ∗ < 2. Only poly(|Jb |) many signature vectors are encountered during
the execution of Algorithm 3.3.
29
Proof. By Lemma 3.28(a), and the definition of si , each coordinate is at most O(log |Jb |). Lemma 3.28(b)
also implies that the coordinates of the signature vector are increasing at the beginning of the main
loop. So every signature vector encountered at the beginning of the main loop can be unambiguously
described as a subset of a set of size O(log |Jb |).
Lemma 3.30. Suppose 1 ≤ τ ∗ < 2. The signature vector decreases in lexicographic value across
each iteration of the main loop of the local search algorithm.
Proof. Consider the beginning of some iteration of the main loop of the local search algorithm
with the state S. So, L0 , . . . , Lℓ are the set of layers at the beginning of the iteration. During
the iteration, a single new layer Lℓ+1 is created in Step 4, and zero or more layers from the set
{L1 , . . . , Lℓ , Lℓ+1 } are discarded in Step 15. We consider two cases accordingly.
• No layer is discarded. Therefore, at the end of this iteration, we will have layers L0 , . . . , Lℓ+1
and we can apply Lemma 3.28(b) at the beginning of the next iteration to prove this claim.
Note here that we used the converse of Corollary 3.27 to deduce that the while loop in Step 6
of Algorithm 3.3 did not execute since no layer was discarded.
• At least one layer is discarded. During each iteration of the while loop in Step 6, for some
1 ≤ r ≤ ℓ + 1 as chosen in Step 11, the if condition in Step 10 is satisfied by Corollary 3.27.
Therefore, the size of |Br−1 ∩ Mσb | reduces to at most (1 − µ1 µ2 )|Br−1 ∩ Mσb |, and the (r − 1)th coordinate of the signature vector reduces by at least one, whereas the coordinates of the
signature vector of the layers preceding r − 1 are unaffected. In other words, the signature
vector at the beginning of the next iteration of the main loop (if any) would be
s′ = (s0 , . . . , sr−2 , s′r−1 , ∞),
where r ≤ ℓ + 1 and s′r−1 ≤ sr−1 − 1.
An immediate corollary of Lemma 3.29 and Lemma 3.30 is that the local search algorithm
described in Section 3.2 terminates after poly(|Jb |) iterations of the main loop under the assumption
1 ≤ τ ∗ < 2. Notice, however, that all statements proved in Section 3.3.6 also hold merely given the
conclusions of Lemma 3.17 and Theorem 3.24 without necessarily assuming that τ ∗ ∈ [1, 2).
3.4
Proof of the Main Theorem
Proof of Theorem 1.1. Let I be the given instance of the (1, ǫ) case of restricted assignment
makespan minimization and OPT denote the optimum makespan. Assume for the moment that
τ ∗ is known by solving the Configuration LP. If τ ∗ ≥ 2, then the algorithm of Lenstra, Shmoys and
Tardos [LST90] for this problem gives an OPT + pmax approximation guarantee, which is of course
at most 1.5OPT in this case.
Suppose instead that 1 ≤ τ ∗ < 2. Start with a partial schedule σ guaranteed by Lemma 3.3.
Let σ and j0 ∈ Js \ σ −1 (M ) denote the input partial schedule and small job to Algorithm 3.3. From
its description it is clear that the partial schedule maintained by it is modified either in Step 20 or
within the main loop. In the main loop, this occurs in exactly three places: Step 3 of Algorithm 3.4;
Steps 12 and 13 of Algorithm 3.3. From Lemma 3.29 and Lemma 3.30, we deduce that the main
30
loop is exited after poly(|Jb |) iterations. Using Proposition 3.10(a), Proposition 3.8(a), and Step 20
of the local search algorithm, the output partial schedule σ ′ therefore satisfies the property
σ ′−1 (M ) = σ −1 (M ) ∪ {j0 }.
Repeating this algorithm a total of |Js | times yields a schedule of makespan at most τ ∗ + R for I
in polynomial time.
However, it is not necessary to know τ ∗ in advance by solving the Configuration LP. Suppose
that τ ∈ [1, 2) is a guess on the value of τ ∗ . Let A(τ ) denote the following algorithm. Run the
procedure outlined above after substituting τ in place of τ ∗ with two modifications to Algorithm 3.3:
if |B0 ∩ Mσb | = 0 after Step 1, or if |Aℓ+1 | < µ1 |B≤ℓ | in any execution of Step 4, then terminate the
procedure with an error claiming that the guessed value τ < τ ∗ .
Suppose A(τ ) returns a schedule during a binary search over the range τ ∈ [1, 2), then it is
guaranteed to have makespan at most τ + R. Note that the range of possible values for τ ∗ is discrete
(1 + kǫ or kǫ for k ∈ Z). As the running time analysis in Section 3.3.6 of Algorithm 3.3 depends
only on conclusions of Lemma 3.17 and Theorem 3.24, A(τ ) is always guaranteed to terminate in
polynomial time irrespective of whether a schedule is returned or not. If A(τ ) does not return
a meaningful result during the binary search then τ ∗ ≥ 2, and it suffices to return the schedule
computed by the algorithm of Lenstra et al. [LST90].
3.5
Balancing Against Bipartite Matching
The approximation guarantee of the local search algorithm from Section 3.2 deteriorates with increasing ǫ. There is however a simple algorithm that performs better for the case of large ǫ.
Theorem 3.31. Let 0 < ǫ < 1. The (1, ǫ) case of restricted assignment makespan minimization admits a 2 − ǫ approximation algorithm.
Proof. Let OPT denote the makespan of an optimum solution to the input instance. Guess OPT
through binary search. Construct a bipartite graph with ⌊OPT/ǫ − 1⌋ small nodes and 1 big node
for each machine in the input instance. Each small job is connected by an edge to all the nodes of
all the machines it can be assigned to with a finite processing time. Each big job is connected by an
edge to all the big nodes of all the machines it can be assigned to with a finite processing time. It is
easy to see that there is a perfect matching in this bipartite graph which corresponds to a schedule
of makespan at most
j
k
OPT
− 1 ǫ + 1 ≤ OPT − ǫ + 1 ≤ OPT + (1 − ǫ)OPT = (2 − ǫ)OPT.
ǫ
Proof of Theorem 1.2. Run the algorithm in Theorem 1.1 with a parameter ζ ′ on the input instance
to obtain a schedule with makespan at most (1 + R(ǫ, ζ ′ ))OPT. Run the algorithm in Theorem 3.31
to get a (2 − ǫ)OPT makespan schedule. The better of the two schedules has an approximation
guarantee that is no worse than
√
1
′
min 1 +
ǫ + 3 − 2ǫ + ζ , 2 − ǫ .
2
31
Figure 6: The previous profile of the approximation guarantee as a function of 1 ≤ τ ∗ ≤ 3 and
0 < ǫ ≤ 3/4. The two surfaces making up the profile correspond to the guarantees 2 − ǫ from
Theorem 3.31 and 1 + 1/τ ∗ from the algorithm of Lenstra, Shmoys and Tardos [LST90]. The work
of Chakrabarty, Khanna and Li [CKL15] provided a 2 − ǫ0 guarantee for some positive ǫ0 > 0, which
is indicated as a red dot at the apex of the profile.
Suppose that 2 − ǫ ≥ 17/9 + ζ. Then, ǫ ≤ 1/9 − ζ. So,
√
√
1
1
ǫ + 3 − 2ǫ + ζ ′ ≤ 1 +
ǫ + 3 − 2ǫ
1+
2
2
1
+ζ =1+ ·
2
ǫ=1/9
′
1 5
+
9 3
+ ζ′ =
17
+ ζ,
9
for ζ ′ = ζ.
4
Conclusion
In this paper we presented a purely flow based local search algorithm for the (1, ǫ)-case of restricted assignment makespan minimization. The guarantees achieved by our algorithm improve significantly over the previous best one due to Chakrabarty et al. [CKL15]. For an illustration
of the approximation profile for the problem, see Figures 6 and 7.
We remark that the ideas presented in this paper do not crucially depend on the fact that
the instances contain exactly two different job sizes. Nevertheless, we have chosen to present our
results in the (1, ǫ)-case as there are still certain obstructions which prevent us from achieving 2 − ǫ0
guarantees for the restricted case in general. A second reason is that the algorithm for the case
with two different job sizes admits a clean description in terms of maximum flows, as we have seen
earlier in Section 3.2.
32
Figure 7: Now, following Theorem 1.2, the worst case guarantee is greatest (roughly 1.89) for
instances with τ ∗ = 1 and ǫ ≈ 1/9 as shown in the figure. The third surface arises from the
guarantee of 1 + R(ǫ, ζ)/τ ∗ for 1 ≤ τ ∗ < 2 from the proof of Theorem 1.1.
References
[AFS12]
Arash Asadpour, Uriel Feige, and Amin Saberi. Santa claus meets hypergraph matchings.
ACM Transactions on Algorithms (TALG), 8(3):24, 2012. 2
[AKS15]
Chidambaram Annamalai, Christos Kalaitzis, and Ola Svensson. Combinatorial algorithm for restricted max-min fair allocation. In Proceedings of the Twenty-Sixth Annual
ACM-SIAM Symposium on Discrete Algorithms, SODA 2015, San Diego, CA, USA,
January 4-6, 2015, pages 1357–1372, 2015. 2
[BS06]
Nikhil Bansal and Maxim Sviridenko. The santa claus problem. In Proceedings of the
thirty-eighth annual ACM symposium on Theory of computing, pages 31–40. ACM, 2006.
2, 6
[CKL15]
Deeparnab Chakrabarty, Sanjeev Khanna, and Shi Li. On (1, ε)-restricted assignment
makespan minimization. In Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 1087–1101. SIAM, 2015. 1, 3, 32
[CLRS09] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and Clifford Stein. Introduction to Algorithms (3. ed.). MIT Press, 2009. 12
[EKS08]
Tomáš Ebenlendr, Marek Křćal, and Jiří Sgall. Graph balancing: a special case of
scheduling unrelated parallel machines. In Proceedings of the nineteenth annual ACMSIAM symposium on Discrete algorithms, pages 483–490. Society for Industrial and Applied Mathematics, 2008. 3
33
[Fei08]
Uriel Feige. On allocations that maximize fairness. In Proceedings of the nineteenth
annual ACM-SIAM symposium on Discrete algorithms, pages 287–293. Society for Industrial and Applied Mathematics, 2008. 2
[HSS11]
Bernhard Haeupler, Barna Saha, and Aravind Srinivasan. New constructive aspects of
the lovasz local lemma. Journal of the ACM (JACM), 58(6):28, 2011. 2
[JR16]
Klaus Jansen and Lars Rohwedder. On the configuration-lp of the restricted assignment
problem. arXiv preprint arXiv:1611.01934, 2016. 3
[LST87]
Jan Karel Lenstra, David B Shmoys, and Eva Tardos. Approximation algorithms for
scheduling unrelated parallel machines. In Foundations of Computer Science, 1987.,
28th Annual Symposium on, pages 217–224. IEEE, 1987. 1, 2
[LST90]
Jan Karel Lenstra, David B Shmoys, and Éva Tardos. Approximation algorithms for
scheduling unrelated parallel machines. Mathematical programming, 46(1-3):259–271,
1990. 2, 3, 30, 31, 32
[PS12]
Lukas Polacek and Ola Svensson. Quasi-polynomial local search for restricted max-min
fair allocation. In Automata, Languages, and Programming, pages 726–737. Springer,
2012. 2
[Sve12]
Ola Svensson. Santa claus schedules jobs on unrelated machines. SIAM Journal on
Computing, 41(5):1318–1341, 2012. 2
[WS11]
David P Williamson and David B Shmoys. The design of approximation algorithms.
Cambridge University Press, 2011. 2
A
Appendix
Proof of Lemma 3.3. Consider the bipartite graph G = (M ∪ Jb , E) where there is an edge {i, j} ∈
E if and only if i ∈ Γ(j). A perfect matching in G of size |Jb | corresponds to such a map. If there
is no such perfect matching, by Hall’s condition, there is a set S ⊆ Jb such that |NG (S)| < |S|.
Consider the following setting (y ∗ , z ∗ ) of variables in the dual of CLP (τ ∗ ).
(
(
1, if i ∈ S,
1, if i ∈ NG (S),
∗
∗
and zj =
yi =
0, else.
0, else.
It is now easily verified that (y ∗ , z ∗ ) is a feasible solution to the dual of CLP (τ ∗ ) defined in (2.2).
We use here the fact that configurations C ∈ C(τ ∗ , i) for any
P machine
Pi ∈ M∗ contain at most
∗
∗
one big job since τ < 2. As the objective function value j∈J zj − i∈M yi attained by this
feasible solution (y ∗ , z ∗ ) is strictly positive, it follows that the dual of CLP (τ ∗ ) is unbounded–for
any λ > 0, (λy ∗ , λz ∗ ) is a feasible dual solution as well. Therefore, by weak duality, CLP (τ ∗ ) must
be infeasible, a contradiction.
Proof of Proposition 3.7. By flow decomposition, the maximum flow in Hσb (S, T ) has flow paths
p1 , . . . , p|Hσb (S,T )| , each of which sends one unit of flow from some vertex in S to some vertex in T .
The flow paths may not share a vertex in T as sinks have unit vertex capacities in Hσb (S, T ) as
34
defined in Definition 3.6. Each machine i ∈ M has at most one outgoing edge with unit capacity
in Gbσ due to Definition 3.1(c) and Definition 3.4. So the flow paths may also not intersect in some
vertex in M \ T since there is at most one outgoing arc with unit capacity. Similarly, they may not
share a vertex in Jb as there is only one incoming arc of unit capacity to a vertex in Jb in Gbσ using
Definition 3.1 and Definition 3.4 .
Claim A.1. ∀1 ≤ τ ∗ < 2, 0 < ǫ < 1, ζ > 0,
(2R − δ − ǫ − 1 − τ ∗ µ1 µ2 − (1 + 2τ ∗ + R)µ1 ) · (δ(1 − µ2 ) − 2µ2 ) − (1 − µ2 )(1 − R) − µ2 τ ∗ > 0,
where µ1 = min{1, ζ}/4, µ2 = min{δ, ζ}/4, δ =
√
√
3 − 2ǫ − 1 /2, and R = ǫ + 3 − 2ǫ /2 + ζ.
Proof. The statement is true if it is true for τ ∗ = 2. To prove that the bracketed expression is
positive we substitute the values of µ1 , µ2 , δ and R from (3.1) and the statement of Theorem 1.1,
and additionally set τ ∗ = 2 to get the statement
√
1
1 √
1 √
1
min
2ζ + ǫ + 3 − 2ǫ − 2
3 − 2ǫ − 1 , ζ −
3 − 2ǫ − 1 , ζ − 1
2
2
4
2
√
1 1 √
1 √
3 − 2ǫ + 3 min
3 − 2ǫ − 1 , ζ − 3 − 2ǫ + 1 ×
+
16 4
2
√
√
1 √
3 − 2ǫ − 1 , ζ + 2ζ + ǫ + 3 − 2ǫ + 10 − 4 4ζ + 3 − 2ǫ − 1 > 0.
min{1, ζ} min
2
1
− min
2
Using a standard computer algebra system for eliminating quantifiers over reals, we can verify the
truth of the above statement for all 0 < ǫ < 1 and ζ > 0.
Claim A.2. ∀0 < ǫ < 1, ζ > 0,
where µ2 = min{δ, ζ}/4, and δ =
√
δ(1 − µ2 ) − 2µ2 > 0,
3 − 2ǫ − 1 /2.
Proof. Substituting the values of µ2 and δ, the statement reads
√
√
1 √
4 3 − 2ǫ − 1 >
3 − 2ǫ + 3 min
3 − 2ǫ − 1 , ζ .
2
It suffices to verify that statement
√ assuming that the min term always evaluates to the first argument,
which then reduces
to ǫ + 3 3 − 2ǫ > 4. Let f (ǫ) denote the expression on the left. Then,
√
f ′ (ǫ) = 1 − 3/( 3 − 2ǫ) is negative over the range [0, 1], f (0) − 4 > 0 and f (1) − 4 = 0.
35
| 8cs.DS
|
Causalities of the Taiwan Stock Market
Julian Juhi-Lian Ting∗
arXiv:cond-mat/0304132v1 [cond-mat.stat-mech] 6 Apr 2003
No.38, Lane 93, Sec.2, Leou-Chuan E. Rd., Taichung, Taiwan 40312, ROC.
(Dated: February 2, 2008)
Volatility, fitting with first order Landau expansion, stationarity, and causality of the Taiwan
stock market (TAIEX) are investigated based on daily records. Instead of consensuses that consider
stock market index change as a random time series we propose the market change as a dual time
series consists of the index and the corresponding volume. Therefore, causalities between these two
time series are investigated.
PACS numbers: 05.45.Tp,89.90.+n,05.40.-a
I.
INTRODUCTION
Physicists are interested in studying stock markets as complex systems. Almost all questions asked can be summarized as searching for price formation theories. Previous researches shown the distribution of price change has
pronounced tail distribution in contrast to Gaussian distribution expected. Furthermore, the auto-correlation of price
change decays exponentially with a characteristic time scale around 5 minutes. Stock crash or rally have also been
identified by physicists as a kind of herd behaviour[1].
Johansen and Sornette considered fitting most stock markets for the bubbles using Landau expansions of the
index[2]. They showed evidence that market crashes as well as large corrections are preceded by speculative bubbles
with two main characteristics: a power law acceleration of the market price decorated with log-periodic oscillations.
For most markets the log-frequency ω/2π is close to unity. However, most data analysis were done for the index of
more mature market like S & P 500. Not all market indexes are defined in the same way. Could different weighting
methods reach the same conclusion? Will government intervention play a role in the conclusion? Johansen and
Sornette further found emergent markets have larger fluctuations. They extended the expansion up to third order
and successfully predicted Nikkei raise in the year 2000[3]. However, why do these fluctuations exist? The logperiodic oscillation appeared in a wide class of out of equilibrium dynamic systems, like ruptures in heterogeneous
media, historic analysis of earthquakes data, and world population. Canessa tried to establish universality for the
exponents from a renormalization group theory[4], and used a stochastic theory to show that the log-periodicities are
a consequence of transient clusters introduced by an entropy-like term. As entropy in thermodynamics corresponds
to the information in an information theory. The possibility to arrive at the log-periodic oscillation therefore suggests
the log-periodic oscillation is a consequence of information exchange between different species of a large system[5].
The effect of volume is less analyzed in the literature. Volume is a measure of market liquidity while index means
the price. Gopikrishnan et al.[6] analyzed the statistical properties of number of shares traded of a particular stock
at a given time interval from an empirical rule saying that it takes volume to push the index. Bonanno et al.[7] also
analyzed the number of shares traded of selected stocks and find a power spectrum of approximately 1/f .
In the present work we considered the volume effects of Taiwan stock market (TAIEX). In particular, the causeeffect relation between the volume time series and the index time series is analyzed. Taiwan stock market is one of
the largest emerging markets. Johansen and Sornette did not studied it because of availability of trading information.
In the present work we tried to make up this missing piece.
When people talk about the volume involved, sometimes they mean the number of shares traded, sometimes the
amount of money involved. We consider money flow to be more important than share number flow if the whole market
is considered, since the person or institute involved should have fixed amount of money.
II.
BASIC PROPERTIES
Taiwan has two stock markets. TAIEX is the major market[8]. TAIEX definition is taken the yearly average of
1966 as 100. All stocks traded are taken into account.
∗ Electronic
address: [email protected]; URL: http://www.chez.com/jlting
2
We analyzed our daily data of TAIEX from January 3, 1991 to December 30, 2000, which include 2814 trading days
for a ten years period. Intra-day time is treated as continuous. Within our data time, several significant events can
be identified. In August 26, 1997, the market reached a local peak 10116.84. The highest index happened at February
17, 2000 with an index 10202.2. The lowest point is at January 7, 1993 with an index 3135.56.
In order to compare Taiwan stock market with other countries we need to know some basic properties studied in
other markets. The first properties are of course daily index and volume histories. They are plotted in Fig.(1).
10000
9000
8000
7000
TAIEX
6000
5000
4000
3000
920930
940630
960402
971223
991102
971223
991102
date
3500
volume (x 10 8NT$)
3000
2500
2000
1500
1000
500
0
920930
940630
960402
date
FIG. 1: Histories of TAIEX daily variations and the corresponding volumes.
One of the most studied properties is volatility. It is found to be a good measure for long time behaviour and risk[9].
It is also the key input parameter for all option pricing models. Many different quantitative definitions of volatility
are used in the literature. Following Liu et al. we define the price change G(t) as the change in the logarithm of the
index,
Z(t + ∆t) − Z(t)
G(t) ≡ ln Z(t + ∆t) − ln Z(t) ∼
,
=
Z(t)
(1)
where ∆t is the sampling time interval. The absolute value of G(t) describes the amplitude of the fluctuation, as
shown in Fig. (2). The large values of |G(t)| correspond to the crashes and big rallies. None of the |G(t)| ever be
larger than 0.07 because government regulation limited maximum daily changes to 7%.
We define the volatility as the average of |G(t)| over a time window T = n · ∆t, i.e.,
VT (t) ≡
t+n−1
1 X
|G(t′ )| ,
n ′
t =t
(2)
3
0.07
0.06
fluctuation |G(t)|
0.05
0.04
0.03
0.02
0.01
0.00
920930
940630
960402
971223
991102
date
FIG. 2: Daily fluctuations amplitude, |G(t)|.
in which n is the moving window size. Fig. (3) shows the calculated volatility VT (t) for a large averaging window
T = 30 days.
0.035
0.030
volatility
0.025
0.020
0.015
0.010
0.005
0.000
921019
940720
960422
980114
991122
date
FIG. 3: Volatility history over a time window T = 30 days. Dates are labeled at the centers of the average windows.
Fig. (4) shows the probability density function P (VT ) of the volatility for several values of T with ∆t = 1 day. The
10-day data is further fitted with Gaussian distribution in Fig. (5). Apparently the data distribution deviates from
the Gaussian curve in many places. The right hand side heel has much higher probability than the Gaussian fit.
An important property need to be investigated for any time series analysis is its stationarity. Most empirical
analysis assumes the time series to be stationary. A time series is stationary if its mean value and its variance do
not vary systematically over time[10]. For TAIEX we choose a window of 1000 trading days, say, and calculate the
corresponding mean value and variance. The result is plotted in Fig. (6). We find the averages of both index and
volume have clear upward trend, while the variances of index oscillate. The volume average and its variance curves
have a turning point around the year 1995, which marks a fundamental change of Taiwan economics. The index curve
is further fitted with a linear curve. The slope is 1.6, which means if an investor holds a index portfolio for over 1000
days her expectation to profit is 1.6 point/day.
PAnother test for stationarity is based on sample autocorrelation function, ρk = γk /γ0 , in which γk =
(Yt − Ȳ )(Yt+k − Ȳ ) is the covariance at lag k. If ρk decays very fast with increasing k, the time series is stationary. The resulting correlogram is plotted in Fig.(7). Apparently, the index time series is more stochastic than the
4
6 days
10 days
21 days
200
180
160
140
probability
120
100
80
60
40
20
0
-20
0.01175
0.02175
0.03175
0.04175
volatility
FIG. 4: Volatility distribution for various time windows with ∆t = 30 min.
180
160
140
frequency count
120
100
80
60
40
20
0
-20
0.01175
0.02175
0.03175
0.04175
volatility
FIG. 5: Gaussian fit for the volatility distribution for T = 10 days and ∆t = 1 days.
volume time series. The results of Fig.(6) and Fig.(7) are consistent.
III.
CRASHES
In this section, we follow Johansen and Sornette to fit the index changes and obtain future trends prediction.
However, our analysis differs from theirs in labeling the date. They convert all dates into decimal numbers. We
labeled only by trading dates. Our labeling method will be more reliable in dating, since trading days spread
irregularly amount years.
Johansen and Sornette wrote down the first, second, and third order Landau expansion of the index starting or
ending at the historical high tc . The first order expansion reads:
p1 (t) ≈ A + Bτ α + Cτ α cos [ω ln (τ ) + φ] .
(3)
in which τ = |t − tc | is the absolute value of time difference from the critical time of the highest index tc . The
parameter φ correspond to the time unit used, while the parameter A, B,and C are units determined by the index as
5
(a)
index average
volume average
3
8x10
3
7x10
3
TAIEX or Volume
6x10
3
5x10
3
4x10
3
3x10
3
2x10
3
1x10
910901
930113
940528
951010
970221
980706
date
2.4
2.2
(b)
index variance
volume variance
2.0
1.8
variance (x 10 6)
1.6
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0
910901
930113
940528
951010
970221
980706
date
FIG. 6: Stationarity test for TAIEX with a window size T = 1000 days for (a) average of the index and the corresponding
volume (b) variance of the index and the corresponding volume. The window is labeled by its central date.
well as the historical period. Only tc , α, and ω are the key parameters: α stands for the power law acceleration, while
ω quantifies for the log-periodic oscillation. Numerically, A is roughly the index at tc .
Our fitting start from January 7, 1993 to February 17, 2000. The result is shown in Fig.(8). The fitting does not
appears good, since it is only a first order approximation while the fitting period is about 7 years. There exits a fitting
solution with tc far beyond the highest index point. However, we cannot take that solution because it is obviously
incorrect. The fitted tc has to come before the highest point.
A theory will only be useful if it can also predict for the future. We used the same equation to fit from February
17,2000 in Fig.(9). We find the future is promising. However, in both cases the ω were found to be much higher than
the previous authors’.
IV.
CAUSALITY TEST
It is well known to the practitioner, as an empirical rule, that volume changes come before price change. A rule
of thumb is that higher index will come after huge volume, and if the volume is reaching a minimum the bottom of
the index will not be far. However, how much ahead? What kinds of correlation do they posses? Why is there a lag?
This kind of dual series with one influencing the other has been analyzed in the econometric literature for GNP (gross
national product)-money supply relation, for example[10, 11]. It is called causality as in the literature of physics. We
have already known our time series are non-stationary. In order for the causality test to be meaningful, the two time
series have to be cointegrated, i.e. their wave-lengths of variation have to be of the same order. Fig.(6)(b) and Fig.(7)
6
1.05
1.00
0.95
index
Correlogram
0.90
0.85
0.80
0.75
0.70
volume
0.65
0.60
0
10
20
30
40
50
lag (k)
FIG. 7: Correlogram plot for both index and the corresponding volume time series.
20000
TAIEX
10000
9000
8000
7000
6000
5000
4000
3000
2000
500
1000
1500
2000
2500
date
FIG. 8: TAIEX fit by first order Landau expansion of the index starting from January 7,1993 to February 17,2000. Fitted
parameters are: tc = 2147.999, α = 1.276, ω = 11.013, φ = 7.037, A = 8250.314, B = −0.49, and C = −0.236.
show their time scale differs but not far.
In the econometrics literature a linear functional form is used for the regression of these two time series. Furthermore,
because the future cannot cause the past, we can write down a functional form for the future index, It and the future
trading volume, Vt , as
P
P
(4)
It = i ai Vt−i + j bj It−j + ut
P
P
(5)
Vt = i li Vt−i + j kj It−j + vt .
We fitted for 90-days from the history using the above functional form. The result is plotted in Fig.(10). Both
the index fitting and the volume fitting show stronger dependence on the index history than on the volume history.
However, only the previous trading day shows significance. Another plot is made for the index difference and the
corresponding volume for the future index difference, It and the future trading volume, Vt , as
P
P
(6)
δIt = i ai Vt−i + j bj δIt−j + ut
P
P
(7)
Vt = i li Vt−i + j kj δIt−j + vt .
The result is plotted in Fig.(11). The figure suggests the index difference time series indeed received slightly higher
7
11000
10000
9000
TAIEX
8000
7000
6000
5000
4000
50
100
150
200
250
date
FIG. 9: TAIEX fit by first order Landau expansion of the index starting from February 17, 2000. Fitted parameters are:
tc = −0.491, α = 1.106, ω = 12.650, φ = 5.558, A = 9776.675, B = −9.079, and C = 2.926.
influence from the previous index difference and the previous volume. However, these effects are pretty small. Furthermore, these peaks will be changed with the truncation of the time series.
V.
CONCLUSION
We tried to find the correlation between the volume of transactions involved and stock market fluctuations beyond
simply considering temporal price series, using Taiwan stock market as an example in the present work.
We find TAIEX volatility spread relatively uniformly in the past ten years, which can also be found in the stationarity
test. The distribution of price change shows fat-heel behaviour.
A slightly higher influence coefficient is found in the index difference fitting with previous index and previous
volume. Johansen and Sornette’s result suggests us to try regression with log-periodic functional forms in the future.
However, it does not seem meaningful to go beyond 90-days for the time lag, since companies will have their seasonal
reports by that time.
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
Victor M. Eguiluz and Martin G. Zimmermann, Phys. Rev. Lett., 85, 5659 (2000).
A. Johansen and D. Sornette, Int. J. Theor. and Applied. Finance, cond-mat/9810071.
A. Johansen and D. Sornette, Int. J. Mod. Phys. C, 10, 563 (1999).
E. Canessa, Euro. Phys. J. B, 20, 591 (2001) cond-mat/0008300.
E. Canessa, J. Phys. A, 33, 9131 (2000) cond-mat/0012031.
P. Gopikrishnan, V. Plerou, X. Gabaix, and H.E. Stanley, Phys. Rev. E, 62, R4493 (2000).
G. Bonanno, F. Lillo, and R. N. Mantegna, cond-mat/9912006 Physica A.
Official site for TAIEX located at http://www.tse.com.tw , see also http://www.sfi.org.tw/newsfi/intdb/menu/MainFrame.asp?Mitem=E
Y. Liu, P. Gopikrishnan, P. Cizeau, M. Meyer, C.-K. Peng and H.E. Stanley, Phys. Rev. E, 60, 1390 (1999).
D. N. Gujarati, Basic Econometrics, 3ed, (McGraw-Hill 1995).
G. S. Maddala, Introduction to Econometrics, 2ed, (Macmillan Publishing Co. 1992).
8
bj
ai
1.2
1.0
index coefficients
0.8
0.6
0.4
0.2
0.0
-0.2
0
20
40
60
80
lag/day
0.8
kj
li
0.6
volume coefficients
0.4
0.2
0.0
-0.2
-0.4
-0.6
-0.8
0
20
40
60
80
lag/day
FIG. 10: Causality coefficients for index and for volume.
9
bj
ai
1
standardized index coefficients
0.3
0.2
79
0.1
0.0
-0.1
-0.2
-0.3
0
20
40
60
80
lag/day
kj
li
0.6
standardized volume coefficients
0.5
0.4
0.3
0.2
0.1
0.0
-0.1
0
20
40
60
80
lag/day
FIG. 11: Causality coefficients for index difference and for the corresponding volume.
| 5cs.CE
|
Interactive Generative Adversarial Networks for Facial Expression Generation
in Dyadic Interactions
Behnaz Nojavanasghari
University of Central Florida
arXiv:1801.09092v2 [cs.CV] 30 Jan 2018
[email protected]
Yuchi Huang
Educational Testing Service
[email protected]
Saad Khan
Educational Testing Service
[email protected]
Abstract
A social interaction is a social exchange between two or
more individuals, where individuals modify and adjust their
behaviors in response to their interaction partners. Our social interactions are one of most fundamental aspects of
our lives and can profoundly affect our mood, both positively and negatively. With growing interest in virtual reality and avatar-mediated interactions, it is desirable to make
these interactions natural and human like to promote positive effect in the interactions and applications such as intelligent tutoring systems, automated interview systems and
e-learning. In this paper, we propose a method to generate facial behaviors for an agent. These behaviors include
facial expressions and head pose and they are generated
considering the users affective state. Our models learn semantically meaningful representations of the face and generate appropriate and temporally smooth facial behaviors
in dyadic interactions.
1. Introduction
Designing interactive virtual agents and robots have
gained a lot of attention in recent years [10, 21, 43]. The
popularity and growing interest in these agents is partially
because of their wide applications in real world scenarios.
They can be used for variety of applications from education [30, 31], training [36] and therapy [7, 34] to elderly
care [5] and companionship [6]. One of the most important aspects in human communication is social intelligence
[3]. A part of social intelligence depends on understanding other people’s affective states and being able to respond
appropriately. To have a natural human- machine interaction, it is critical to enable machines with social intelligence
as well. The first step towards this capability is observing human communication dynamics and learning from the
occurring behavioral patterns [33, 32]. In this paper, we
focus on modeling dyadic interactions between two partners. The problem we are solving is given the affective
Figure 1. The problem we are solving in this paper is, given affective state of one person in a dyadic interaction, we generate the
facial behaviors of the other person. By facial behaviors, we refer to facial expressions, movements of facial landmarks and head
pose.
state of one partner, we are interested in generating nonverbal facial behaviors of the other partner. Figure 1 shows
the overview of our problem. The affective states that we
consider in generating facial expressions include joy, anger,
surprise, fear, contempt, disgust, sadness and neutral. The
non-verbal cues that we are generating in this work, are facial expressions and head pose. These cues help us generate
head nods, head shakes, tilted head and various facial expressions such as smile which are behaviors that happen in
our daily face to face communications. A successful model
should learn meaningful factors of the face for generating
of most appropriate responses for the agent. We have de-
signed our methodology based on the following factors: 1)
humans react to each others’ affective states from an early
age and adjust their behavior accordingly [42]. 2) Each affective state is characterized by particular facial expressions
and head pose[9]. 3) Each affective state can have different
intensities, such as happy vs very happy. These factors (e.g.
affective state and their intensities) effect how fast or slow
or facial expressions and head pose changes over time[9].
There has been great deal of research on facial expression analysis where the effort is towards describing the facial expressions [40]. Facial expressions are a result of
movements of facial landmarks such as raising eyebrows,
smiling and wrinkling nose. A successful model for our
problem, should learn the movements of these points in response to different affective states.
We are using FACET SDK for extracting 8-dimensional
affective state of one partner in the interaction. We then use
it as a condition in generation of non-verbal facial behaviors of the other person. Our generation process relies on
a conditional generative adversarial networks [29], where
the conditioning vector is affective state of one party in the
interaction. Our proposed methodology is able to extract
meaningful information about the face and head pose and
integrate the encoded information with our network to generate expressive and temporally smooth face images for the
agent.
2. Related work
Humans communicate using not only words but also
non-verbal behaviors such as facial expressions, body posture and head gestures. In our daily interactions, we are
constantly adjusting our non-verbal behavior based on other
peoples behavior. In collaborative activities such as interviews and negotiations, people tend to unconsciously mimic
other people’s behavior [4, 26].
With increasing interest in social robotics and virtual reality and their wide applications in social situations such as
automated interview systems [17], companionship for elder
care [5] and therapy for autism [7], there is a growing need
to enable computer systems to understand, interpret and respond to people’s affective states appropriately.
As a person’s face discloses important information about
their affective state, it is very important to generate appropriate facial expressions in interactive systems as well.
There has been a number of studies focusing on automatic generation of facial expressions. These efforts can
be categorized into three main groups: 1) Rule based approaches where affective states are mapped into a predefined 2D or 3D face model [15]. 2) Statistical approaches
where face shape is modeled as linear combination of prototypical expression basis [27]. 3) Deep belief networks
where the models learn the variation of facial expressions
in presence of various affective states and produce convinc-
ing samples [39].
While generating facial expressions using previous approaches can result in promising results, it is important to
consider other people’s affective states while generating facial expressions in dyadic interactions. There has been previous work in this area, [18] where researchers use conditional generative adversarial networks for generating facial expressions in dyadic interactions. While the proposed
approach results in appropriate facial expression for one
frame, it does not consider the temporal consistency in a sequence, which results in generating non-smooth facial expressions over time. Also, some of the attributes such as
head position and orientation that are important cues for
recognizing tilted head, head nods, head shakes are not captured by the model.
In this paper, we address the problem of generating facial
expressions in dyadic interactions while considering temporal constraints and additional facial behaviors such as head
pose. Our proposed model extracts semantically meaningful encodings of the face regarding movements of facial
landmarks and head pose and integrate it with the network
which results in generation of appropriate and smooth facial
expressions.
3. Methodology
Generative adversarial networks (GANs) [13] have been
widely used in the field of computer vision and machine
learning for various applications such as image to image
translation [20], face generation [12], semantic segmentation [28] and etc. GANs are a type of generative model
which learn to generate based on generative and discriminative networks that are trained and updated at the same time.
The generative network tries to find the true distribution of
the data and generate realistic samples and the discriminator network tries to discriminate between the samples that
are generated by generated and samples from real data. The
general formulation of conditional GANs are as follows:
min max L (D, G) = Ex,y∼pdata (x,y) [log D(x, y)]+
G
D
Ex∼pdata (x),z∼pz (z) [log(1 − D(x, G(x, z)))]. (1)
Conditional GANs are generative models that learn a mapping from random noise vector z to output image y conditioned on auxiliary information x : G : {x, z} ⇒ y. A conditional GAN consists of a generator G(x, z) and a discriminator D(x, y) that compete in a two-player minimax game:
the discriminator tries to distinguish real training data from
generated images, and the generator tries to fail the discriminator. That is, D and G play the following game on V (D,
G). As you can see the goal of these networks is to increase
the probability of samples generated by generative networks
to resemble the real data so that the discriminator network
Figure 2. Overview of our affect-sketch network. In first stage of our two-stage facial expression generation, we generate face sketches
conditioned on the affective state of the interviewee and using a z vector that carries contains semantically meaningful information about
facial behaviors. This vector is sampled from generated distributions of rigid and non-rigid face shape parameters using our proposed
methods.
Figure 3. Overview of our sketch-image network. In the second
stage of our two-stage facial expression generation, we generate
face images conditioned on the generator sketch of the interviewer
which is generated in the first stage of our network.
fail to differentiate between the generated samples and the
real data.
Conditional generative adversarial networks (CGANs)
are a type of GAN that generate samples considering a conditioning factor. As our problem is generation of facial expressions conditioned on affective states, we found CGANs
as perfect candidate for addressing our problem. We chose
to use CGANs to generate facial expressions of one party,
conditioned on the affective state of the other partner in
a dyadic interaction.The GAN formulation uses a continuous input noise vector z, that is an input to the generator. However, this vector is a source of randomness to the
model and does not capture semantically meaningful information. However, using a meaningful distribution which
corresponds to our objective and desired facial behaviors,
can help the model integrate meaningful information with
the network and improve the results of generation. For example, in case of generating facial behaviors, the important factors are the variation in locations of facial landmarks
such as smiling and head pose such as tilted head.
In this paper, we present a modification to the sampling process of the latent vector z which is an input to the
CGAN network. We sample z from the meaningful distributions of facial behaviors that are generated by two different
approaches 1) Affect-shape dictionary and 2) Conditional
LSTM. This step helps the network to learn interpretable
Figure 4. This figure shows 68 landmarks of the face that we intend to generate using our model. In the bottom row you see a
visualization of non-rigid shape parameters which correspond to
movements of facial landmarks such as opening the mouth. Our
model aims to learn the distribution of these parameters and use
them in generating temporally smooth sequences.
and meaningful representation of facial behaviors. In the
next sections we will first explain the two stage architecture
of our network. Then we will explain preprocessing of the
frames by going over the intuition behind our choice of face
representation. Finally, we provide detailed description of
our methods for generating distributions for vector z which
is the input to our CGAN networks.
3.1. Two-Stage Conditional Generative Adversarial
Network
To generate facial behaviors for agent, we use a two stage
architecture, both of which are based on conditional gener-
ative adversarial networks. Figure 2 shows an overview of
our first network. In the first stage we use a conditional generative adversarial network which takes face shape parameter as input for vector z as well as the conditioning vector
which is the 8 affective states of one partner in the interaction, and it generates sketches of faces for the other partner
in the interaction. Note that the z vector will be sampled
using one of our proposed strategies, which are explained
in the following sections. In the second stage, we input
the generated sketches in the first stage as input to GAN
and generate the final facial expressions. Figure 3 shows an
overview of our second network.
3.2. Data Preprocessing and Face Representation
Facial landmarks are the salient points on face located
at the corners, tips or midpoints of facial components such
as eyes, nose, mouth, etc [41]. Movements of these points
form various facial expressions that can convey different affective states. For example, widening the eyes can communicate surprise and a smile can be an indication of happiness. Since the locations of facial landmarks alone are not
particularly meaningful in communicating affect but rather
their movements and change over time which plays a rule in
affect recognition [38], we decided to measure their movements and change over time. To do so, we are projecting a
3D point distribution model on each image. This model has
been trained on in the wild data [25]. The following equation is used to place a single feature point of the 3D PDM
in a given input image:
x̄i = s.R2D .(X̄i + φi q) + t
In this equation, s shows scale, R shows the head rotation
and t is the head translation which are known as rigid- shape
parameters. X̄i = [x̄i , ȳi , z̄i ] is the mean value of i th feature ,
φi is principle component matrix and q is a matrix controlling the non-rigid shape parameters, which correspond to
deviations of landmarks locations from an average (neutral)
face.
As described, shape parameters of the face carry rich information about head pose and facial expressions. Hence,
we use them as our face representations. Figure 4 shows
68 facial landmarks that we have considered in this work.
We have also included a visualization of non-rigid shape
parameters at the bottom row for better understanding of
these parameters.
3.3. Affect-Shape Dictionary
Different facial expressions are indicative of different affective states [23]. Depending on the affective state and its
intensity, the temporal dynamics and changes in landmarks
can happen in various facial expressions such as a smile vs
a frown and they can happen slow or fast depending on how
intense is the affect [8]. Based on this knowledge, we propose the following strategies to learn meaningful distribu-
tions of face shape parameters. Figure 5 shows an overview
of our affect-shape dictionary generation.
3.3.1
Affect Clustering
While in state of fear landmarks can change very sudden, in
state of sadness or neutral the changes in locations of landmarks will be smaller and slower. In order to account for
this point, we first cluster the calculated shape parameters
for all input frames into 8 clusters, using k-means clustering [14] where clusters denote the following affective states:
Joy, Anger, Surprise, Fear, Contempt, Disgust, Sadness and
Neutral. This stage gives us distributions of possible facial
expressions for each affective state.
3.3.2
Inter Affect Clustering
The output of the first clustering step will give us groupings of shape parameters in 8 clusters corresponding to each
category of affective states. Each affective state can have
various arousal levels (a.k.a intensity)[24]. We did a second level of clustering of shape parameters inside each affect cluster. To do so, we used a hierarchical agglomerative
clustering approach [11], where we put the constraint that
each cluster should contain at least 100 videos. The result
of this clustering gives us 3 to 9 sub clusters corresponding
to different arousal levels of each affective state. Having
information about the possible facial expressions that can
happen in a particular affect category and what has been expressed in previous frames by the agent, we can sample z
from an appropriate distribution which results in an expressive and temporally smooth sequence of facial expressions.
To account for the temporal constraints, for generating nth
frame, we pick z from the appropriate distribution based on
the affective state which is the closest to the previously chosen z in terms of euclidean distance.
3.4. Conditional LSTM
Long Short Term Memory networks (LSTM) have been
known for their ability in learning long-term dependencies
[16]. We are interested in learning the dynamics of facial
expression, hence LSTMs are an appropriate candidate for
our problem. As the problem we are solving is a conditional problem, we will use a conditional LSTM (C-LSTM)
[2]. The input to CLSTM will be the concatenated vector of
affective state of the interviewer and the shape parameters
of the interviewee for n previous frames and the generated
output will be the future shape parameters of the interviewee. We fix n to be 100 frames in our experiments. We
use conditional LSTMS in two different settings which are
described in the following subsections. Figure 6 shows an
overview of our C-LSTM approach.
Overlapping C-LSTM: In overlapping conditional
LSTMS, we consider information across 100 frames as in-
Figure 5. Overview of our affect-shape approach for generating clusters of face shape parameters. We first extract shape descriptors for all
faces in the dataset. Then we cluster them into 8 affect classes corresponding to Joy, Anger, Surprise, Fear, Contempt, Disgust, Sadness
and Neutral. Then, we do a further inter class clustering for each affect class. The purpose of the second clustering is to find the inter
affect clusters, such as very joyful, joyful or little joyful.These clusters form the distributions from which the z vector is sampled in our
generative models.
Figure 6. Overview of our conditional LSTM approach for generating facial shape parameters. In this approach we consider the affective
state of the interviewee and past frames of the interviewer’s facial expressions as conditioning vector and previous history. We then learn
to generate the future frames of interviewer’s facial expressions using a conditional LSTM. The output result of this generation is the input
for z vector in our generative models.
put of the model and the output of the model is generation
of shape parameters for one future frame. In each step, the
newly generated frame is added to the history and the information from the first input frame is removed instead ( in
first in, first out manner) , to predict futures frames.
4. Experiments
Non-overlapping C-LSTM: In non-overlapping CLSTMs the input to the models is information across 100
frames and the output is generation of shape parameters for
future 100 frames, with no overlap between the frames.
4.1. Dataset
To assess the effectiveness of our proposed approaches
in generating appropriate and temporally smooth facial expressions, we performed a number of experiments using the
proposed frameworks approaches. Our experiments show
the comparison between our models in terms of result and
their pros and cons.
The dataset used in this paper is 31 pair of dyadic interactions (videos) which are interviews for undergraduate
Figure 7. Examples of generated sketches using both of our approaches. The first sequence corresponds to the result of affect-shape
dictionary approach and the second row is generated using the conditional LSTM approach. Overall, our first approach generates smoother
sequences.
admissions process. The purpose of interviews is to assess
English speaking ability of the prospective college. There
are 16 male and 15 female candidates and each candidate
is interviewed by the same interviewer (Caucasian female)
who followed a predetermined set of academic and nonacademic questions designed to encourage open conversation.
The interviews were conducted using Skype video conferencing so the participants could see and hear each other and
the video data from each dyadic interaction was captured.
The duration of interviews varies from 8 to 37 minutes and
a total of 24 hours of video data. We have used 70,000 short
video clips for training data and 7000 videos for evaluation.
4.2. Affect Recognition Framework
For estimating the affective state of the interviewee, we
have used Emotients Facet SDK [1] to process the frames
and estimate 8- dimensional affect descriptor vectors, representing the likelihood of the following classes: joy, anger,
surprise, fear, contempt, disgust, sadness and neutral.
4.3. Implementation Details
To create our training set we randomly sampled 70,000
video clips of 100 frames each (3.3 seconds) from the interviewee videos and extracted their affect descriptor using
Facet SDK. These affect descriptors are then combined into
a single vector using the approach proposed by Huang et
al. [18]. For each interviewee video clip a single frame
from the corresponding interviewer video is also sampled
for training. Our approach is to train the model to generate
a single frame of the interviewer conditioned on facial expression descriptors from the preceding 100 frames of the
interviewee. All face images are aligned and landmarks are
estimated by the method proposed by kazem et al. [22] to
generate ground truth face sketches. For creating our test
data, we randomly sampled 7000 interviewee video clips of
400 frames each (no overlap with the training set) and used
these as input to our two model to generate 7000 facial expression video clips.
For training the first stage of our network (affect to
sketch), we first estimate the landmarks of the interviewer
using the proposed method by Kazemi et al. [22] which
gives us 68 facial landmarks (See Figure 4). We then connect these points by linear lines of one pixel width to generate the sketch images. The generated sketches are used
in a pair with the corresponding image of the interviewer
for training the second network. We masked all training image pairs by an oval mask after roughly aligned all faces,
to reduce the effect of appearance and lightening variation
of the the face in generating our videos. In the generator,
a sketch image is passed through an encoder-decoder network [20] each containing 8 layers of down sampling and
up sampling, in order to generate the final image. We have
adopted the idea proposed by Ronneberger et al. [37] and
have connected feature maps at layer i and layer n-i where
n is total number of layers in the network. In this network
receptive fields after convolution are concatenated with the
receptive fields in up-convolving process, which allows network to use original features in addition to features after upconvolution and results in overall better performance than a
network that has access to only features after up-sampling.
We adopted the architecture proposed by Radford et. al
[35] for our generation framework and deep convolutional
structure for generator and discriminator. We used modules
of the form convolution-BatchNorm-ReLu [19] to stabilize
optimization. In the training phase, we used mini-batch
SGD and applied the Adam solver. To avoid the fast con-
Figure 8. Examples of generated face images using the second stage of our architecture. The first sequence corresponds to a single step
generation without considering the sketches and generating face images directly from the affect vector and noise. The second row shows
the results related to generation of faces using our proposed two stage architecture that we have used in our experiments. Note that the
quality of images in the second row is much better than the first row.
vergence of discriminators, generators were updated twice
for each discriminator update, which differs from original
setting [35] in that the discriminator and generator update
alternately.
In order to generate temporally smooth sequences, for
each new frame, z is sampled from a meaningful distribution and based on the previous frames. Note that the appropriateness of z for a frame is decided based on the possible variation between two adjacent frames in distribution of
face shape parameters training data which correspond to the
movements of facial landmarks and head pose.
4.4. Results and Discussions
We generated sequences of non-verbal facial behaviors
for an agent, based on affect-shape strategy and conditional
LSTM. Figure 7 shows the output sketches generated by
these two models. The first row corresponds to the affectshape dictionary approach and the second row shows the
results based on Conditional LSTM. We observed that the
results from dictionary based is more smooth. In results of
conditional LSTM there seems to be grouping of smooth sequences followed by sudden jumps between the sequences
(In Figure 7 see the transition between last and one before
last frame in the second row). We think that this can happen
because of number of frames we have considered as history,
as in LSTM frameworks all of the history gets summarized
and in unrolling step the time steps are not captured accurately and the generation is more based on an average of all
frames. We compare the two proposed strategies in Table
1. The advantage of the dictionary based approach is that it
does not require previous history of the face and it can run
in real time. However, since z is randomly sampled from the
possible distribution space, it generates a different sequence
each time and it is hard to evaluate this approach.
In table 2 we show an evaluation of conditional LSTM
approach for generating the facial expressions. We have
generated the facial expressions using C-LSTM by considering 100 frames as history and 1) generating the 101 st
frame (complete overlap). 2) generating the next 100 frames
(No overlap). As this method, tries to generate the closest
sequence to ground truth, we have calculated mean squared
error as an evaluation metric.
As you can see, C-LSTM with 100 frame overlap has a
smaller error value compared to the C-LSTM with no overlap. Intuitively, it is harder to predict future with no overlap
than having overlap.
Properties
History
Diversity
Evaluation
Speed
Affect-Shape Dictionary
No need for history.
Generates diverse sequence.
Hard to evaluate.
Real time
Conditional LSTM
Needs history.
Generates one particular sequence.
Easier to evaluate.
Not real time
Table 1. Comparison between Affect-shape dictionary and conditional LSTM approaches in generating face shape parameters.
Method
C-LSTM w overlap
C- LSTM w/o overlap
MSE
0.101
0.183
Table 2. Comparison between performance of conditional LSTMs
with and without overlap in history.
We have also generated the final face images using our
CGAN network. Figure 8 shows the generated images of
the interviewer which is the final result of the our second
network, sketch-image GAN. The top row shows the results
of such images if we only had one network and directly
went from affect vector and z into face images. The second row shows the results of a two step network where we
first generate face sketches and then face images. Note that
quality of the images in second row is significantly better
than the first row. Also, as our final goal is to transfer these
expressions to an avatar’s face, our mid-level results can be
used to transfer the expressions to various virtual characters.
5. Conclusions and Future Work
Human beings react to each other’s affective state and
adjust their behaviors accordingly. In this paper we have
proposed a method for generating facial behaviors in a
dyadic interaction. Our model learns semantically meaningful facial behaviors such as head pose and movements
of landmarks and generates appropriate and temporally
smooth sequences of facial expressions.
In our future work, we are interested in transferring the
generated facial expressions to an avatar’s face and generating sequences of expressive avatars reacting to the interviewee’s affective state using the proposed approaches in
this work. We then will run a user study using pairs of interviewee and generated avatar’s videos to study the most
appropriate responses for each video and will use the result in building our final models of emotionally intelligent
agents.
References
[1] imotions inc. manual of emotients Facet SDK. 2013.
[2] I. Augenstein, T. Rocktäschel, A. Vlachos, and
K. Bontcheva.
Stance detection with bidirectional conditional encoding.
arXiv preprint
arXiv:1606.05464, 2016.
[3] R. Bar-On, D. Tranel, N. L. Denburg, and A. Bechara.
Emotional and social intelligence. Social neuroscience: key readings, 223, 2004.
[4] S. Bilakhia, S. Petridis, and M. Pantic. Audiovisual detection of behavioural mimicry. In Affective
Computing and Intelligent Interaction (ACII), 2013
Humaine Association Conference on, pages 123–128.
IEEE, 2013.
[5] J. Broekens, M. Heerink, H. Rosendal, et al. Assistive
social robots in elderly care: a review. Gerontechnology, 8(2):94–103, 2009.
[6] K. Dautenhahn, M. Walters, S. Woods, K. L. Koay,
C. L. Nehaniv, A. Sisbot, R. Alami, and T. Siméon.
How may i serve you?: a robot companion approaching a seated person in a helping context. In Proceedings of the 1st ACM SIGCHI/SIGART conference
on Human-robot interaction, pages 172–179. ACM,
2006.
[7] K. Dautenhahn and I. Werry. Towards interactive
robots in autism therapy: Background, motivation
and challenges. Pragmatics & Cognition, 12(1):1–35,
2004.
[8] J. Decety, L. Skelly, K. J. Yoder, and K. A. Kiehl. Neural processing of dynamic emotional facial expressions in psychopaths. Social neuroscience, 9(1):36–
49, 2014.
[9] P. Ekman. An argument for basic emotions. Cognition
& emotion, 6(3-4):169–200, 1992.
[10] T. Fong, I. Nourbakhsh, and K. Dautenhahn. A survey of socially interactive robots. Robotics and autonomous systems, 42(3):143–166, 2003.
[11] P. Franti, O. Virmajoki, and V. Hautamaki. Fast
agglomerative clustering using a k-nearest neighbor
graph. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 28(11):1875–1881, 2006.
[12] J. Gauthier. Conditional generative adversarial nets
for convolutional face generation. Class Project for
Stanford CS231N: Convolutional Neural Networks for
Visual Recognition, Winter semester, 2014(5):2, 2014.
[13] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu,
D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In Advances in neural information processing systems, pages 2672–2680,
2014.
[14] J. A. Hartigan and M. A. Wong. Algorithm as
136: A k-means clustering algorithm. Journal of the
Royal Statistical Society. Series C (Applied Statistics),
28(1):100–108, 1979.
[15] D. Heylen, M. Poel, A. Nijholt, et al. Generation of facial expressions from emotion using a fuzzy rule based
system. In Australian Joint Conference on Artificial
Intelligence, pages 83–94. Springer, 2001.
[16] S. Hochreiter and J. Schmidhuber. Lstm can solve
hard long time lag problems. In Advances in neural information processing systems, pages 473–479, 1997.
[17] M. E. Hoque, M. Courgeon, J.-C. Martin, B. Mutlu,
and R. W. Picard. Mach: My automated conversation
coach. In Proceedings of the 2013 ACM international
joint conference on Pervasive and ubiquitous computing, pages 697–706. ACM, 2013.
[18] Y. Huang and S. M. Khan. Dyadgan: Generating facial
expressions in dyadic interactions. In Computer Vision
and Pattern Recognition Workshops (CVPRW), 2017
IEEE Conference on, pages 2259–2266. IEEE, 2017.
[19] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine
Learning, pages 448–456, 2015.
[20] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros. Imageto-image translation with conditional adversarial networks. arXiv preprint arXiv:1611.07004, 2016.
[21] T. Kanda, T. Hirano, D. Eaton, and H. Ishiguro. Interactive robots as social partners and peer tutors for
children: A field trial. Human-computer interaction,
19(1):61–84, 2004.
[22] V. Kazemi and J. Sullivan. One millisecond face alignment with an ensemble of regression trees. In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 1867–1874, 2014.
[23] D. Keltner, P. Ekman, G. C. Gonzaga, and J. Beer. Facial expression of emotion. 2003.
[24] E. A. Kensinger and S. Corkin. Two routes to
emotional memory: Distinct neural processes for
valence and arousal. Proceedings of the National
Academy of Sciences of the United States of America,
101(9):3310–3315, 2004.
[25] M. Koestinger, P. Wohlhart, P. M. Roth, and
H. Bischof. Annotated facial landmarks in the wild:
A large-scale, real-world database for facial landmark
localization. In Computer Vision Workshops (ICCV
Workshops), 2011 IEEE International Conference on,
pages 2144–2151. IEEE, 2011.
[26] J. L. Lakin, V. E. Jefferis, C. M. Cheng, and T. L.
Chartrand. The chameleon effect as social glue:
Evidence for the evolutionary significance of nonconscious mimicry. Journal of nonverbal behavior,
27(3):145–162, 2003.
[27] C. L. Lisetti and D. J. Schiano. Automatic facial expression interpretation: Where human-computer interaction, artificial intelligence and cognitive science intersect. Pragmatics & cognition, 8(1):185–235, 2000.
[28] P. Luc, C. Couprie, S. Chintala, and J. Verbeek.
Semantic segmentation using adversarial networks.
arXiv preprint arXiv:1611.08408, 2016.
[29] M. Mirza and S. Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014.
[30] B. Nojavanasghari, T. Baltrušaitis, C. E. Hughes, and
L.-P. Morency. Emoreact: a multimodal approach and
dataset for recognizing emotional responses in children. In Proceedings of the 18th ACM International
Conference on Multimodal Interaction, pages 137–
144. ACM, 2016.
[31] B. Nojavanasghari, T. Baltrušaitis, C. E. Hughes, and
L.-P. Morency. The future belongs to the curious: Towards automatic understanding and recognition of curiosity in children. In Workshop on Child Computer
Interaction, pages 16–22, 2016.
[32] B. Nojavanasghari, D. Gopinath, J. Koushik, T. Baltrušaitis, and L.-P. Morency. Deep multimodal fusion
for persuasiveness prediction. In Proceedings of the
18th ACM International Conference on Multimodal
Interaction, pages 284–288. ACM, 2016.
[33] B. Nojavanasghari, C. Hughes, T. Baltrusaitis, L.p. Morency, et al. Hand2face: Automatic synthesis
and recognition of hand over face occlusions. arXiv
preprint arXiv:1708.00370, 2017.
[34] B. Nojavanasghari, C. E. Hughes, and L.-P. Morency.
Exceptionally social: Design of an avatar-mediated interactive system for promoting social skills in children
with autism. In Proceedings of the 2017 CHI Conference Extended Abstracts on Human Factors in Computing Systems, pages 1932–1939. ACM, 2017.
[35] A. Radford, L. Metz, and S. Chintala.
Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint
arXiv:1511.06434, 2015.
[36] J. Rickel. Intelligent virtual agents for education and
training: Opportunities and challenges. In Intelligent
virtual agents, pages 15–22. Springer, 2001.
[37] O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention,
pages 234–241. Springer, 2015.
[38] T. Senechal, V. Rapp, and L. Prevost. Facial feature
tracking for emotional dynamic analysis. Advances
Concepts for Intelligent Vision Systems, pages 495–
506, 2011.
[39] J. M. Susskind, G. E. Hinton, J. R. Movellan, and
A. K. Anderson. Generating facial expressions with
deep belief nets. In Affective Computing. InTech,
2008.
[40] Y. Tian, T. Kanade, and J. F. Cohn. Facial expression
recognition. In Handbook of face recognition, pages
487–519. Springer, 2011.
[41] Y. Tie and L. Guan. Automatic landmark point detection and tracking for human facial expressions.
EURASIP Journal on Image and Video Processing,
2013(1):8, 2013.
[42] E. Tronick. Still face experiment. Retrieved October,
22:2016, 2009.
[43] B. Ulicny and D. Thalmann. Crowd simulation for
interactive virtual environments and vr training systems. Computer Animation and Simulation 2001,
pages 163–170, 2001.
| 1cs.CV
|
GENERAL HYPERPLANE SECTIONS OF THREEFOLDS
IN POSITIVE CHARACTERISTIC
arXiv:1703.00770v1 [math.AG] 2 Mar 2017
KENTA SATO AND SHUNSUKE TAKAGI
Dedicated to Professor Lawrence Ein on the occasion of his sixty birthday.
Abstract. In this paper, we study the singularities of a general hyperplane section
H of a three-dimensional quasi-projective variety X over an algebraically closed field
of characteristic p > 0. We prove that if X has only canonical singularities, then H
has only rational double points. We also prove, under the assumption that p > 5,
that if X has only klt singularities, then so does H.
1. Introduction
Reid proved in his historic paper [25] that if a quasi-projective variety over an algebraically closed field of characteristic zero has only canonical singularities, then its
general hyperplane section has only canonical singularities, too. This can be deduced
from the Bertini theorem for base point free linear systems, and the same argument
works for other classes of singularities in the minimal model program such as terminal,
klt, log canonical singularities. Then what if the variety is defined over an algebraically
closed field of positive characteristic? As a test case, we consider the following question:
Question 1.1. Let X be a three-dimensional quasi-projective variety over an algebraically closed field k of positive characteristic and H be a general hyperplane section
of X. If X has only terminal singularities, is H smooth? If X has only canonical (resp.
klt, log canonical) singularities, does H have only rational double points (resp. klt
singularities, log canonical singularities)?
Since the Bertini theorem for base point free linear systems fails in positive characteristic, Reid’s argument does not work in this setting. Thanks to the recent development
in birational geometry, we are able to find a resolution of singularities of X ([6], [7]) and
to run the minimal model program if X has only klt singularities and the characteristic
of k is larger than 5 ([12], [3], [4]), but we do not know in general how to overcome the
difficulty arising from the lack of the Bertini theorem.
The terminal case in Question 1.1 is proved affirmatively by the fact that threedimensional terminal singularities are isolated singularities. The other cases are much
more subtle. For example, Reid proved, as a corollary of the result mentioned above,
that all closed points of a complex canonical threefold, except finitely many of them,
have an analytic neighborhood which is nonsingular or isomorphic to the product of a
2010 Mathematics Subject Classification. 14B05, 14J17, 13A35.
Key words and phrases. Bertini theorem, canonical singularities, klt singularities, MJ-canonical
singularities, strongly F -regular singularities.
1
2
K. SATO and S. TAKAGI
rational double point and A1C . This does not hold in positive characteristic: HirokadoIto-Saito [18] and Hirokado [17] gave counterexamples in characteristic two and three.
However, the canonical case in Question 1.1 has remained open.
In this paper, we give an affirmative answer to the canonical and klt cases in Question
1.1, using jet schemes and F -singularities, while we need to assume that the characteristic of k is larger than 5 in the klt case. In both cases, we look at every codimension
two point x of X. The normal surface singularity Spec OX,x is canonical or klt but
is not defined over an algebraically closed field. In the canonical case, we observe, as
in the case where the base field is algebraically closed, that Spec OX,x is a hypersurface singularity. In the klt case, we deduce from the following theorem, which is a
generalization of a result of Hara [13], that Spec OX,x is strongly F -regular when the
characteristic of k is larger than 5.
Theorem 1.2 (Proposition 2.8, Theorem 5.7). Let (s ∈ S) be an F -finite normal
surface singularity (not necessarily defined over an algebraically closed field) of characteristic p > 5 and B be an effective Q-divisor on S whose coefficients belong to the
standard set {1 − 1/m | m ∈ Z>1 }. Then (s ∈ S, B) is klt if and only if (s ∈ S, B) is
strongly F -regular.
In the canonical case, we see from the above observation that X is MJ-canonical
except at finitely many closed points. MJ-canonical singularities are defined in terms of
Mather-Jacobian discrepancies instead of the usual discrepancies and can be viewed as
a jet scheme theoretic counterpart of canonical singularities. Similarly, in the klt case,
we see that if the characteristic of k is larger than 5, then X is strongly F -regular except
at finitely many closed points. Strongly F -regular singularities are defined in terms of
the Frobenius morphism and can be viewed as an F -singularity theoretic counterpart
of klt singularities. Applying the Bertini theorem for MJ-canonical singularities due to
Ishii-Reguera [20] to the canonical case and the Bertini theorem for strongly F -regular
singularities due to Schwede-Zhang [28] to the klt case, we obtain the assertion. To
be precise, we can prove a more general result involving a boundary divisor, which is
stated as follows.
Main Theorem (Proposition 4.1, Theorems 4.2 and 5.9). Let X be a three-dimensional
normal quasi-projective variety over an algebraically closed field k of characteristic
p > 0 and ∆ be an effective Q-divisor on X such that KX + ∆ is Q-Cartier. Let H be
a general hyperplane section of X.
(1) If (X, ∆) is terminal, then (H, ∆|H ) is terminal.
(2) If (X, ∆) is canonical, then (H, ∆|H ) is canonical.
(3) Suppose in addition that p > 5 and the coefficients of ∆ belong to the standard
set {1 − 1/m | m ∈ Z>1 }. If (X, ∆) is klt, then (H, ∆|H ) is klt.
Acknowledgments. The authors are indebted to Shihoko Ishii for helpful discussions, especially
on the proof of Theorem 4.2. They are also grateful to Mircea Mustaţă and Hiromu Tanaka for
useful conversations. The first author was partially supported by the Program for Leading
Graduate Schools, MEXT, Japan. The second author was partially supported by JSPS
KAKENHI 26400039.
GENERAL HYPERPLANE SECTIONS OF THREEFOLDS
3
Convention. Throughout this paper, all rings are assumed to be commutative and
with a unit element and all schemes are assumed to be noetherian and separated. A
variety over a field k means an integral scheme of finite type over k.
2. Preliminaries
2.1. F -singularities. In this subsection, we recall the basic notions of F -singularities,
which we will need in §4.
Let X be a scheme of prime characteristic p > 0. We say that X is F -finite if the
Frobenius morphism F : X −→ X is a finite morphism. When X = Spec R is an affine
scheme, R is said to be F -finite if X is F -finite. For example, a field k of positive
characteristic p > 0 is F -finite if and only if [k : k p ] < ∞. It is known by [22] that
every F -finite scheme is locally excellent.
Suppose in addition that X is a normal integral scheme. For an integer e > 1 and a
Weil divisor D on X, we consider the following composite map:
(F e )#
F eι
OX −−−→ F∗e OX −−∗→ F∗e OX (D),
where F e : X −→ X is the e-th iteration of Frobenius and ι : OX −→ OX (D) is the
natural inclusion.
Strong F -regularity is one of the most basic classes of F -singularities.
Definition 2.1. Let (R, m) be an F -finite normal local ring of positive characteristic
p > 0 and ∆ be an effective Q-divisor on X = Spec R. Then the pair (R, ∆) is said to
be strongly F -regular if for every nonzero element c ∈ R, there exists an integer e > 1
such that the map
OX −→ F∗e OX (⌈(pe − 1)∆⌉ + divX (c))
splits as an OX -module homomorphism. When y is a point in an F -finite normal
integral scheme Y and B is an effective Q-divisor on Y , we say that (y ∈ Y, B) is
strongly F -regular if (OY,y , By ) is strongly F -regular. We say that the normal singularity (y ∈ Y ) is strongly F -regular if (OY,y , 0) is strongly F -regular.
Remark 2.2. We have the following hierarchy of properties of normal singularities:
regular =⇒ strongly F -regular =⇒ Cohen-Macaulay.
Global F -regularity is defined as a global version of strong F -regularity.
Definition 2.3. Let X be an F -finite normal integral scheme and ∆ be an effective
Q-divisor on X. Then (X, ∆) is said to be globally F -regular if for every Cartier divisor
D on X, there exists an integer e > 1 such that the map
OX −→ F∗e OX (⌈(pe − 1)∆⌉ + D)
splits as an OX -module homomorphism.
Remark 2.4. If X is an affine scheme, then (X, ∆) is globally F -regular if and only
if (x ∈ X, ∆) is strongly F -regular for all x ∈ X. However, in general, the former
is much stronger than the latter. For example, a smooth projective variety of general
type over an algebraically closed field of positive characteristic is not globally F -regular
but strongly F -regular at all points.
4
K. SATO and S. TAKAGI
2.2. Singularities in the minimal model program. In this subsection, we recall
the definition of singularities in the minimal model program. Note that we do not
assume that the singularities are defined over an algebraically closed field. We start
with the definition of canonical divisors on normal schemes.
Suppose that π : Y −→ X is a separated morphism of finite type between schemes.
If X has a dualizing complex (see [16, Chapter V, §2] for the definition of dualizing
•
complexes), then ωY• := π ! ωX
is a dualizing complex on Y , where π ! is the twisted
inverse image functor associated to π obtained in [16, Chapter VII, Corollary 3.4 (a)].
Since any F -finite affine scheme has a dualizing complex by [10], any scheme of finite
type over an F -finite local ring has a dualizing complex.
•
Let X be an excellent integral scheme with a dualizing complex ωX
. The canonical
•
sheaf ωX associated to ωX is the coherent OX -module defined as the first non-zero co•
homology module of ωX
. It is well-known that ωX satisfies the Serre’s second condition
(S2 ) (see for example [1, (1.10)]). If X is a variety over an algebraically closed field k
with structure morphism f : X −→ Spec k, then the canonical sheaf associated to f ! k
coincides with the classical definition of the canonical sheaf of X. When X is a normal
•
scheme, a canonical divisor KX on X associated to ωX
is any Weil divisor KX on X
∼
such that OX (KX ) = ωX .
•
Suppose that X is a normal scheme with a dualizing complex ωX
and fix a canonical
•
divisor KX on X associated to ωX . Given a proper birational morphism π : Y −→ X
from a normal scheme Y , we always choose a canonical divisor KY on Y which is
•
associated to π ! ωX
and coincides with KX outside the exceptional locus of f .
Definition 2.5. (i) We say that (x ∈ X) is a normal singularity if X = Spec R is an
affine scheme with a dualizing complex where R is an excellent normal local ring and
if x is the unique closed point of X.
(ii) A proper birational morphism f : Y −→ X between schemes is said to be a
resolution of singularities of X, or a resolution of X for short, if Y is regular. Suppose
that X is a normal integral scheme and ∆ is a Q-divisor on X. A resolution f : Y −→ X
is said to be a log resolution if the union of f −1 (Supp(∆)) and the exceptional locus of
f is a simple normal crossing divisor.
We are now ready to state the definition of singularities in the minimal model program.
Definition 2.6. (i) Suppose that (x ∈ X) is a normal singularity with a dualizing
•
complex ωX
and ∆ is an effective Q-divisor on X such that KX + ∆ is Q-Cartier.
Given a proper birational morphism π : Y −→ X from a normal scheme Y , we choose
•
a canonical divisor KY on Y associated to π ! ωX
such that
X
KY + π∗−1 ∆ = π ∗ (KX + ∆) +
ai Ei ,
i
π∗−1 ∆
where
is the strict transform of ∆ by π, the ai are rational numbers and the Ei are
π-exceptional prime divisors on Y . We say that (x ∈ X, ∆) is terminal (resp. canonical,
klt, plt) if all the ai > 0 (resp. all the ai > 0, all the ai > −1 and ⌊∆⌋ = 0, all the
ai > −1) for every proper birational morphism π : Y −→ X from a normal scheme Y .
We say that (x ∈ X) is terminal (resp. canonical, klt, plt) if so is (x ∈ X, 0).
GENERAL HYPERPLANE SECTIONS OF THREEFOLDS
5
(ii) Let X be an excellent integral normal scheme with a dualizing complex and ∆
be an effective Q-divisor on X such that KX + ∆ is Q-Cartier. Then (X, ∆) is terminal
(resp. canonical, klt, plt) if so is the normal singularity (x ∈ X = Spec OX,x , ∆x ) for
every x ∈ X.
Remark 2.7. (i) If X is defined over a field of characteristic zero or dim X 6 3, then
it is enough to check the condition in Definition 2.6 only for one f , namely, for a log
resolution of (X, ∆).
(ii) We have the following hierarchy of properties of Q-Gorenstein normal singularities:
regular =⇒ terminal =⇒ canonical =⇒ klt.
Strong F -regularity implies being klt.
Proposition 2.8 ([15, Theorem 3.3]). Let (x ∈ X) be an F -finite normal singularity
and ∆ be an effective Q-divisor on X such that KX + ∆ is Q-Cartier. If (x ∈ X, ∆)
is strongly F -regular, then it is klt.
3. Brief review on surface singularities
In this section, we briefly review the theory of surface singularities. All results are
well-known if the singularity is defined over an algebraically closed field. Some of the
results may be known to experts (even if it is not defined over an algebraically closed
field), but we include their proofs here, because we have not been able to find any
reference to them.
Definition 3.1. We say that a scheme X is a surface if X is a two-dimensional excellent
•
separated integral scheme with a dualizing complex ωX
and that a normal singularity
(x ∈ X) is a surface singularity if dim OX,x = 2.
Remark 3.2. From now on, we often use the results in [21] and [29]. We remark that
the base scheme is assumed to be regular in these references, but this assumption is
unnecessary in our framework, because our base scheme is a normal surface singularity
and its dualizing complex is unique up to isomorphism.
First we recall the definition and basic properties of intersection numbers. Let (x ∈
X) be a normal surface singularity and f : Y −→ X beSa proper birational morphism
from a normal surface Y withP
exceptional curves E = i Ei . For a Cartier divisor D
on Y and a Weil divisor Z = i ai Ei on Y , We define the intersection number D · Z
as follows:
X
D·Z =
ai degEi /k(x) (OY (D)|Ei ),
i
where degEi /k(x) : Pic(Ei ) −→ Z is the degree morphism defined as in [9, Definition
1.4].
Proposition 3.3 ([9, Proposition 2.5], [21, Theorem 10.1]). With the notation above,
the following hold.
6
K. SATO and S. TAKAGI
(1) The intersection pairing
M
Div(X) ×
ZEi −→ Z,
(D, Z) 7−→ D · Z
i
is a bilinear map.
(2) If Z is Cartier and Supp D ⊆ E, then D · Z = Z · D.
(3) If Y is regular, then the intersection matrix (Ei · Ej )i,j is a negative-definite
symmetric matrix.
Definition 3.4. Let (x ∈ X) be a normal surface singularity
Swith a dualizing complex
•
ωX
. A resolution f : Y −→ X with exceptional curves E = i Ei is said to be minimal
•
if a canonical sheaf KY on Y associated to the dualizing complex f ! ωX
is f -nef, that
is, KY · Ei > 0 for every i. A minimal resolution of X always exists by [21, Theorem
2.25].
We also recall the definition of rational singularities. A normal surface singularity
(x ∈ X) is said to be a rational singularity if R1 f∗ OY = 0 for every resolution f :
Y −→ X.
Proposition 3.5. Let (x ∈ X) be a normal surface singularity.
(1) ([23, Proposition 1.2]) (x ∈ X) is a rational singularity if and only if R1 f∗ OY =
0 for some resolution f : Y −→ X.
(2) ([21, Proposition 2.28]) If there exists an effective Q-divisor ∆ on X such that
(x ∈ X, ∆) is plt, then (x ∈ X) is a rational singularity.
In the rest of this section, we work in the following setting.
Setting 3.6. Suppose that (x ∈ X) is a normal surface singularity with a dualizing
•
complex
curves
S ωX and f : Y −→ X is a (nontrivial) resolution with ! exceptional
•
E = i Ei . Let KY denote a canonical divisor on Y associated to f ωX
. Let Zf denote
the fundamental cycle of f , that is, a unique minimal nonzero effective divisor Z on Y
such that Supp Z ⊆ E and −Z is f -nef (such a divisor always exists by [21, Definition
and Claim 10.3.6]).
Let D be a divisor on Y and Z be a nonzero effective divisor on Y such that Supp Z ⊆
E. The Euler characteristic χ(Z, D) is defined by
χ(Z, OY (D)|Z ) = dimk(x) H 0 (Z, OY (D)|Z ) − dimk(x) H 1 (Z, OY (D)|Z ).
The following is the Riemann-Roch theorem for curves embedded in a regular surface.
Proposition
3.7. For each divisor D on Y and each nonzero effective divisor Z =
P
a
E
on
Y
such that Supp Z ⊆ E, we have
i i i
(2D − KY − Z) · Z
= χ(Z, 0) + D · Z.
2
P
Proof. We will prove the assertion by induction on a := i ai . If a = 1, then Z = Ei for
some i and the assertion immediately follows from [29, Theorem 2.5 (1) and Theorem
χ(Z, D) =
GENERAL HYPERPLANE SECTIONS OF THREEFOLDS
7
2.7] (see also Remark 3.2). Suppose that a > 2. Then there exists i such that ai > 1,
so set Z ′ := Z − Ei and consider the exact sequence
0 −→ OEi (−Z ′ ) −→ OZ −→ OZ ′ −→ 0.
By the additivity of the Euler characteristic and applying the induction hypothesis to
Z ′ and Ei , one has
χ(Z, D) = χ(Z ′ , D) + χ(Ei , D − Z ′ )
(2D − KY − Z ′ ) · Z ′ (2(D − Z ′ ) − KY − Ei ) · Ei
+
2
2
(2D − KY − Z) · Z
=
.
2
=
Proposition 3.8 ([2]). Suppose that (x ∈ X) is a rational singularity. Let mx denotes the maximal ideal on X corresponding to x and e(x ∈ X) be the Hilbert-Samuel
multiplicity of OX,x . Then the following hold for each integer n > 0.
(1) mnx = H 0 (Y, OY (−nZf )).
(2) dimk(x) mn /mn+1 = −nZf2 + 1.
(3) e(x ∈ X) = −Zf2 .
Proof. First note that mnx OY = OY (−nZf ), which follows from essentially the same
argument as for [2, Theorem 4], where X is assumed to be defined over an algebraically
closed field. Since in a rational surface singularity, the product of integrally closed ideals
is again integrally closed (see [23, Theorem 7.2]), mnx is integrally closed. Therefore,
mnx = H 0 (Y, mnx OY ) = H 0 (Y, OY (−nZf )).
For (2), we consider the exact sequence
0 −→ OY (−(ℓ + 1)Zf ) −→ OY (−ℓZf ) −→ OZf (−ℓZf ) −→ 0
for each integer ℓ > 0. Since H 1 (Y, OY (−ℓZf )) = H 1 (Y, OY (−(ℓ + 1)Zf )) = 0 by [21,
Proposition 10.9 (1)],
χ(Zf , −ℓZf ) = dimk(x) H 0 (Zf , OZf (−ℓZf ))
= dimk(x) H 0 (Y, OY (−ℓZf ))/H 0(Y, OY (−(ℓ + 1)Zf ))
= dimk(x) mℓx /mℓ+1
x ,
where the last equality follows from (1). In particular, χ(Zf , 0) = 1. Thus, wee see
from Proposition 3.7 that
dimk(x) mnx /mn+1
= χ(Zf , −nZf ) = 1 − nZf2 .
x
Finally, (3) immediately follows from (2), because
dimk(x) OX,x /mnx =
n−1
X
i=0
dimk(x) mix /mi+1
=
x
−Zf2 2
n + O(n).
2
8
K. SATO and S. TAKAGI
We will use the following result to prove the canonical case of the main theorem in
§3.
Proposition 3.9. A normal surface singularity (x ∈ X) is canonical if and only if
(x ∈ X) is either a regular point or a rational double point. In particular, a canonical
surface singularity is a hypersurface singularity.
Proof. We may assume by Proposition 3.5 that (x ∈ X) is a rational singularity. Since
the assertion is obvious when X is regular, we can also assume that (x ∈ X) is not
regular.
Suppose that f : Y −→ X is a minimal resolution, and let ∆ = f ∗ KX − KY . Then
∆ is an anti-f -nef Q-divisor on Y , so ∆ > 0 by [21, Claim 10.3.5]. On the other hand,
it follows from Proposition 3.7 that
2 = 2χ(Zf , 0) = −(KY + Zf ) · Zf = −KY · Zf − Zf2 = ∆ · Zf − Zf2 ,
which implies that ∆ · Zf = Zf2 + 2.
If (x ∈ X) is a canonical singularity, then we have ∆ = 0 and hence Zf2 = −2. We
see by Proposition 3.8 (3) that e(x ∈ X) = 2, that is, (x ∈ X) is a rational double
point. Using Proposition 3.8 (2), (3), we can verify that every rational double point is
a hypersurface singularity.
Conversely, if (x ∈ X) is a rational double point, then Zf2 = −2 by Proposition 3.8
(3) and therefore ∆ · Zf = 0. Since −∆ is f -nef and Supp Zf = E, the intersection
number ∆ · Ei has to be zero for all i. This means by [21, Claim 10.3.5] that ∆ = 0. It
then follows from the fact that KY + ∆ = f ∗ KX and (Y, ∆) = (Y, 0) is canonical that
(x ∈ X) is a canonical singularity.
4. Terminal and canonical singularities
The terminal case of the main theorem is an immediate consequence of a Bertini
theorem for Hilbert-Samuel multiplicity.
Proposition 4.1. Let X be a three-dimensional normal quasi-projective variety over
an algebraically closed field of characteristic p > 0 and ∆ be an effective Q-divisor on
X such that KX + ∆ is Q-Cartier. Suppose that the pair (X, ∆) is terminal. Then
(H, ∆|H ) is terminal for a general hyperplane section H of X. In particular, H is
smooth.
Proof. Let U be the locus of the points x ∈ X such that X is regular at x and
multx (∆) < 1. Since the regular locus of X is open and Hilbert-Samuel multiplicity
is upper semicontinuous, U is an open subset of X. Then every codimension two
point x ∈ X lies in U by [21, Theorem 2.29 (1)], because (Spec OX,x , ∆x ) is a twodimensional terminal pair. Hence, X \ U consists of only finitely many closed points
and a general hyperplane section H of X is contained in U. It follows from a Bertini
theorem for Hilbert-Samuel multiplicity [8, Proposition 4.5], which holds in arbitrary
characteristic, that H is smooth and multx (∆|H ) = multx (∆) < 1 for all x ∈ Supp ∆|H .
Thus, applying [21, Theorem 2.29 (1)] again, we see that (H, ∆|H ) is terminal.
GENERAL HYPERPLANE SECTIONS OF THREEFOLDS
9
The canonical case of the main theorem can be deduced from a Bertini theorem for
MJ-canonical singularities. The proof was inspired by a discussion with Shihoko Ishii,
whom we thank.
Theorem 4.2. Let X be a three-dimensional normal quasi-projective variety over an
algebraically closed field k of characteristic p > 0 and ∆ be an effective Q-divisor on
X such that KX + ∆ is Q-Cartier. Suppose that the pair (X, ∆) is canonical. Then
(H, ∆|H ) is canonical for a general hyperplane section H of X. In particular, H has
only rational double points.
Proof. First we will show the case where ∆ = 0. Since X has only canonical singularities, Spec OX,x is a two-dimensional scheme with only canonical singularities for every
codimension two point x ∈ X. It then follows from Proposition 3.9 that OX,x is a
hypersurface singularity for every codimension two point x ∈ X. Since the l.c.i. locus
of X is open by [11], X is a l.c.i. except at finitely many closed points. By the fact that
l.c.i. canonical singularities are MJ-canonical (see for example [20, Remark 2.7 (iv)]),
X has only MJ-canonical singularities except at finitely many closed points. Thus,
applying a Bertini theorem for MJ-canonical singularities [20, Corollary 4.11], we see
that H has only rational double points.
Next we consider the case where ∆ 6= 0. Let U1 be the locus of the points x ∈ X
such that X is regular at x and multx (∆) 6 1, and U2 = X \ Supp ∆. Note that U1
and U2 are both open subsets of X (the openness of U1 follows from the openness of
the regular locus of X and the upper-semicontinuity of Hilbert-Samuel multiplicity).
Then every codimension two point x ∈ X lies in U1 ∪ U2 by [21, Theorem 2.29 (2)],
because (Spec OX,x , ∆x ) is a two-dimensional canonical pair. Hence, we may assume
that X = U1 ∪ U2 . We have already seen that the assertion holds when ∆ = 0, so
(H ∩ U2 , ∆|H∩U2 ) = (H ∩ U2 , 0) is canonical for a general hyperplane section H of X.
On the other hand, it follows from a Bertini theorem for Hilbert-Samuel multiplicity
[8, Proposition 4.5] that H ∩ U1 ( U1 is smooth and multx (∆|H∩U1 ) = multx (∆) 6 1
for all x ∈ Supp ∆|H∩U1 . Thus, applying [21, Theorem 2.29 (2)] again, we see that
(H ∩ U1 , ∆|H∩U1 ) is canonical.
Corollary 4.3. Let X be a three-dimensional quasi-projective variety over an algebraically closed field k of characteristic p > 5 with only canonical singularities. Then
there exists a zero-dimensional closed subscheme Z ⊂ X with the following property: for
[
every closed point x ∈ X \ Z, the maximal-adic completion O
X,x of the local ring of X
b k k[[t]]
at x is either regular or isomorphic to the formal tensor product k[[u, v, w]]/(f )⊗
of a rational double point k[[u, v, w]]/(f ) and the one-dimensional formal power series
ring k[[t]].
Proof. An immediate application of Theorem 4.2 to [18, Theorem 3] yields the desired
result.
Remark 4.4. When the characteristic p is less than 5, there exist counterexamples
to Corollary 4.3 due to Hirokado-Ito-Saito [18] and Hirokado [17]. In particular, in
characteristic three, an exhaustive list of examples is given in [18, Theorem 3].
10
K. SATO and S. TAKAGI
5. Klt singularities
In this section, we will prove the klt case of the main theorem. Since we often use
the results of [21] and [29] in this section, the reader is referred to Remark 3.2.
First we prepare some definitions and lemmas for Theorem 1.2.
Definition 5.1. Let I ⊆ [0, 1] be a subset. We define the subset D(I) ⊆ [0, 1] by
)
(
P
m − 1 + nj=1 ij
m ∈ Z>1 , n ∈ Z>0 , ij ∈ I for j = 1, . . . , n ∩ [0, 1].
D(I) :=
m
In particular, D(∅) = {1 − 1/m | m ∈ Z>1 } is the set of standard coefficients.
Lemma 5.2 (cf. [24, Lemma 4.3]). Let X be a normal surface and ∆ be an effective
Q-divisor on X such that (X, ∆) is plt. Let C be a regular curve which is an irreducible
component of ⌊∆⌋ and Diff C (B) be the different of B := ∆−C on C (see [21, Definition
2.34] for the definition of the Q-divisor Diff C (B)). If the coefficients of B belong to a
subset I ⊂ [0, 1] ∩ Q, then the coefficients of Diff C (B) belong to D(I).
Proof. We may assume that (x ∈ X) is a surface singularity, that is, X = Spec R for
a two-dimensional normal local ring.SLet f : Y −→ X be a minimal log resolution of
(X, C) with exceptional curves E = i Ei defined as in [21, Definition 2.25 (b)] and m
be the determinant of the negative of the intersection matrix of the Ei . Then by [21,
(3.36.1)],
m−1
[x] + B|C .
Diff C (B) =
m
P
Let B = j bj Bj be the irreducible decomposition of B. Since mD is Cartier for every
Weil divisor D on X by [21, Proposition 10.9 (3)], we see that ord[x] (mBj )|C is an
integer. Thus,
m − 1 X bj
ord[x] Diff C (B) =
+
ord[x] (mBj )|C ∈ D(I),
m
m
j
which completes the proof of the proposition.
A log Fano pair (X, ∆) is a pair of a normal projective variety over a field k and an
effective Q-divisor on X such that −(KX + ∆) is ample and (X, ∆) is klt.
Proposition 5.3 (cf. [5, Corollary 4.1], [31, Theorem 4.2]). Let I ⊆ [0, 1] ∩ Q be a
finite set. There exists a positive constant p0 (I) depending only on I with thePfollowing
property: if k is an F -finite field of characteristic p > p0 (I) and (P1k , B = m
i=1 ai Pi )
is a log Fano pair such that every coefficient ai belongs to D(I) and every point Pi is
a k-rational point of P1k , then (P1k , B) is globally F -regular. Moreover, when I = ∅, we
may take the constant p0 (I) to be 5.
Proof. The proof is essentially the same as in the case where k is algebraically closed.
When k is algebraically closed, the first assertion is nothing but [5, Theorem 4.1] and
the second one follows from [31, Theorem 4.2].
GENERAL HYPERPLANE SECTIONS OF THREEFOLDS
11
Definition 5.4. A curve X over a field k is a one-dimensional integral scheme of finite
type over k. The arithmetic genus g(X) of a complete curve X over k is defined to be
dimk H 1 (X, OX )/ dimk H 0 (X, OX ). We remark that g(X) is independent of the choice
of the base field k.
We extend Proposition 5.3 to the case of an arbitrary curve of genus zero.
Proposition 5.5. Let I ⊆ [0, 1]∩Q be a finite set. There exists a positive constant p1 (I)
depending only on I with the following property: if k is an F -finite field of characteristic
p > p1 (I) and (C, B) is a log Fano pair such that C is a complete curve over k and
the coefficients of B belong to D(I), then (C, B) is globally F -regular. Moreover, when
I = ∅, we may take p1 (I) to be 5.
Proof. Let p0 (I) be the constant as in Proposition 5.3. Set a(I) = min{a ∈ D(I) | a 6=
0} and p1 (I) = max{p0 , ⌈2/a(I)⌉}. If I = ∅, then a(I) = 1/2, so we can take p1 (∅) = 5
by Proposition 5.3. We say that a pair (C, B) satisfies the condition (⋆) if C is a
complete curve over an F -finite field of characteristic p > p1 (I) and (C, B) is a log
Fano pair such that the coefficients of B belong to D(I). Note that if (C, B) satisfies
(⋆), then C is a complete regular curve with g(C) = 0 by [29, Corollary 2.6].
Let (C, B) be a log Fano pair over a field k satisfying the condition (⋆). First, we will
show that we can reduce to the case where C ∼
= P1k . Replacing k by its finite extension
if necessary, we may assume that k = H 0 (C, OC ). Suppose that C ∼
6= P1k . Then C is
2
isomorphic to a conic in Pk by [21, Lemma 10.6 (3)], and hence there exists a closed
point P ∈ C whose residue field k(P ) is a quadratic extension of k. Let l := k(P )
and consider the pullback f : Cl := C ×Spec k Spec l −→ C of C. Note that Cl is
connected, because [l : k] = 2 and l is not contained in the function field K(C) of C,
so K(C) ⊗k l is a domain. Since p > p1 > 2 = [l : k], the pullback f is a surjective
étale morphism. It follows from [21, Lemma 10.6 (3)] again, together with the fact
that a regular conic with an l-rational point is isomorphic to P1l , that Cl ∼
= P1l . By
the fact that being klt is preserved under surjective étale morphisms ([21, Proposition
2.15]), the pair (Cl , Bl := f ∗ B) satisfies the condition (⋆). On the other hand, a pair is
globally F -regular if and only if its affine cone is strongly F -regular ([26, Proposition
5.3]), and strong F -regularity is preserved under finite étale morphisms of degree not
divisible by p ([14, Theorem 3.3], [27, Corollary 6.31 and Proposition 7.4]). Therefore,
(C, B) is globally F -regular if and only if so is (Cl , Bl ). Thus, replacing k by l and C
by Cl , we may assume that C = P1k .
Next, we will show that we can
P reduce to the case where every point in Supp B
is a k-rational point. Let B = i a
i Pi be the irreducible decomposition of B. Since
P
1
(C = Pk , B) is log Fano, we have i ai [k(Pi ) : k] = degP1k /k (B) < 2. In particular,
[k(Pi ) : k] < 2/a(I) 6 p1 (I) < p for every i. Hence, there exists a Galois extension K/k
such that k(Pi ) ⊆ K for every i and p does not divide [K : k]. Since the condition (⋆)
and global F -regularity are preserved under finite Galois base field extensions of degree
not divisible by p as we have seen above, replacing k by K and (P1k , B) by (P1K , BK )
if necessary, we may assume that every point in Supp B is a k-rational point. The
assertion is now an immediate consequence of Proposition 5.3.
12
K. SATO and S. TAKAGI
Lemma 5.6 (cf. [5, Proposition 2.13]). Let (x ∈ X) be a normal surface singularity
•
with a dualizing complex ωX
and B be an effective Q-divisor on X such that KX + B
is Q-Cartier. If (x ∈ X, B) is klt, then there exists a proper birational morphism
f : Y −→ X from a normal surface Y such that
(1) the exceptional locus C of f is a complete regular curve with g(C) = 0,
(2) (Y, BY + C) is plt, where BY = f∗−1 B is the strict transform of B by f , and
(3) −(KY + BY + C) is f -ample, where KY is a canonical divisor on Y associated
•
to the dualizing complex f ! ωX
.
Proof. We can apply essentially the same argument as in the case where X is defined
over an algebraically closed field [5, Proposition 2.13], using the minimal model program
for excellent surfaces [29, Theorem 1.1] instead of that for surfaces over an algebraically
closed field. Then (2) and (3) are easily verified. Hence, it is enough to check (1). Since
(Y, BY + C) is plt, the complete curve C has to be regular by [21, Theorem 2.31]. The
pair (Y, BY + C) being plt also implies by Proposition 3.5 that Y has only rational
singularities, from which it follows that g(C) = 0 (see [21, Lemma 10.8 (3)]).
Theorem 1.2 is a consequence of Proposition 2.8 and the following theorem, which
gives a generalization of [5, Theorem 1.1] (see also [13]) to the case where the base field
is not necessarily algebraically closed.
Theorem 5.7 (cf. [13], [5, Theorem 1.1]). Let I ⊆ [0, 1]∩Q be a finite set. There exists
a positive constant p1 (I) depending only on I with the following property: if (x ∈ X)
be an F -finite normal surface singularity of characteristic p > p1 (I) and (x ∈ X, B) is
a klt pair such that the coefficients of B belong to D(I), then (x ∈ X, B) is strongly
F -regular. Moreover, when I = ∅, we can take p1 (I) to be 5.
Proof. We employ the same strategy as the proof of [5, Theorem 1.1]. Let p1 (I) be the
constant as in Proposition 5.5. Suppose that (x ∈ X, B) is a klt pair such that (x ∈ X)
is an F -finite normal surface singularity of characteristic p > p1 (I) and the coefficients
of B belong to D(I). Take a proper birational morphism f : Y −→ X with exceptional
prime divisor C as in Lemma 5.6, and let BY = f∗−1 B be the strict transform of B by
f and BC = Diff C (BY ) be the different of BY on C. Note that by Lemma 5.2 and [24,
Lemma 4.4], the coefficients of BC belong to D(D(I)) = D(I) ∪ {1}. Since (C, BC )
is log Fano by adjunction ([21, Theorem 3.36 and Lemma 4.8]), the coefficients of BC
are less than 1, and then (C, BC ) is globally F -regular by Proposition 5.5. Applying
the same argument as in the case where X is defined over an algebraically closed field
([5, Proposition 2.11] and [12, Lemma 2.12]), we can conclude that (X, B) is strongly
F -regular.
Example 5.8 (cf. [21, 2.26]). Let k be a non-algebraically closed field of characteristic
p > 0 with an element a ∈ k that is not a cubic power in k. Let (0 ∈ X) be the origin
of the hypersurface (x2 = y 3 − az 3 + y 4 + z 4 ) ⊂ A3k , which is a klt surface singularity.
Its minimal resolution has two exceptional curves C1 , C2 and their dual graph is
1
7654
0123
2
3
0123
7654
2 ,
GENERAL HYPERPLANE SECTIONS OF THREEFOLDS
13
where the numbers above the circles are the dimk H 0 (Ci , OCi ) and the numbers inside
the circles are the −Ci2 / dimk H 0 (Ci , OCi ). Note that this graph does not appear in the
list of the dual graphs of exceptional curves for the minimal resolutions of klt surface
singularities over an algebraically closed field (see [30]). On the other hand, (0 ∈ X)
is strongly F -regular if and only if p > 3.
We are now ready to prove the klt case of the main theorem.
Theorem 5.9. Let X be a three-dimensional normal quasi-projective variety over an
algebraically closed field of characteristic p > 5 and ∆ be an effective Q-divisor on X
whose coefficients belong to the standard set {1−1/m | m ∈ Z>1 }. Suppose that KX +∆
is Q-Cartier and (X, ∆) is klt. Then (H, ∆|H ) is klt for a general hyperplane section
H of X.
Proof. Let U be the locus of the points x ∈ X such that (X, ∆) is strongly F -regular.
Note that U is an open subset of X (the ∆ = 0 case was proved in [19, Theorem 3.3]
and the general case follows from a very similar argument). Every codimension two
point x ∈ X lies in U, because (Spec OX,x , ∆x ) is strongly F -regular by Theorem 5.7.
Hence, X \ U consists of only finitely many closed points and a general hyperplane
section H of X is contained in U. It then follows from a Bertini theorem for strongly
F -regular pairs [28, Corollary 6.6] that (H, ∆|H ) is strongly F -regular. Since strongly
F -regular pairs are klt by Proposition 2.8, we obtain the assertion.
References
[1] Y. Aoyama, Some basic results on canonical modules, J. Math. Kyoto. Univ. 23 (1983), 85–94.
[2] M. Artin, On isolated rational singularities of surfaces, Amer. J. Math. 88 (1966), 129–136.
[3] C. Birkar, Existence of flips and minimal models for 3-folds in char p, Ann. Sci. Éc. Norm. Supér.
(4) 49 (2016), 169–212.
[4] C. Birkar and J. Waldron, Existence of Mori fibre spaces for 3-folds in char p, arXiv:1410.4511,
preprint (2014).
[5] P. Cascini, Y. Gongyo and K. Schwede, Uniform bounds for strongly F -regular surfaces, Trans.
Amer. Math. Soc. 368 (2016), 5547–5563.
[6] V. Cossart, O. Piltant, Resolution of singularities of threefolds in positive characteristic. I, J.
Algebra 320 (2008), 1051–1082.
[7] V. Cossart, O. Piltant, Resolution of singularities of threefolds in positive characteristic. II, J.
Algebra 321 (2009), 1836–1976.
[8] T. de Fernex, L. Ein, M. Mustaţă, Bounds for log canonical thresholds with applications to
birational geometry, Math. Res. Lett. 10 (2003), 219–236.
[9] W. Fulton, Intersection theory, second edition, Ergebnisse der Mathematik und ihrer Grenzgebiete, 3. Folge, Band 2, Springer, Berlin, 1998.
[10] O. Gabber, Notes on some t-structures, Geometric aspects of Dwork theory. Vol. I, II, pp. 711–
734, Walter de Gruyter GmbH & Co. KG, Berlin, 2004.
[11] S. Greco, M. G. Marinari, Nagata’s criterion and openness of loci for Gorenstein and complete
intersection, Math. Z. 160 (1978), 207–216.
[12] C. Hacon and C. Xu, On the three dimensional minimal model program in positive characteristic,
J. Amer. Math. Soc. 28 (2015), 711–744.
[13] N. Hara, Classification of two-dimensional F -regular and F -pure singularities, Adv. Math. 133
(1998), 33–53.
[14] N. Hara and S. Takagi, On a generalization of test ideals, Nagoya Math. J. 175 (2004), 59–74.
14
K. SATO and S. TAKAGI
[15] N. Hara and K.-i. Watanabe, F -regular and F -pure rings vs. log terminal and log canonical
singularities, J. Alg. Geom. 11 (2002), 363–392.
[16] R. Hartshorne, Residues and duality, Lecture notes of a seminar on the work of A. Grothendieck,
given at Harvard 1963/64, With an appendix by P. Deligne, Lecture Notes in Mathematics, No.20
Springer-Verlag, Berlin-New York 1966.
[17] M. Hirokado, Canonical singularities of dimension three in characteristic 2 which do not follow
Reid’s rules, arXiv:1607.08664, preprint (2016).
[18] M. Hirokado, H. Ito and N. Saito, Three dimensional canonical singularities in codimension two
in positive characteristic, J. Algebra 373 (2013), 207–222.
[19] M. Hochster and C. Huneke, Tight closure and strong F -regularity, Mém. Soc. Math. France 38
(1989), 119–133.
[20] S. Ishii and J. Reguera, Singularities in arbitrary characteristic via jet schemes, arXiv:1510.05210,
preprint (2015).
[21] J. Kollár, Singularities of the minimal model program, Cambridge Tracts in Mathematics, 200,
Cambridge University Press, Cambridge, 2013.
[22] E. Kunz, On Noetherian rings of characteristic p, Amer. J. Math. 98 (1976), 999–1013.
[23] J. Lipman, Rational singularities, with applications to algebraic surfaces and unique factorization,
Inst. H́autes Etudes Sci. Publ. Math., No. 36, 1969, 195–279.
[24] J. McKernan and Y. Prokhorov, Threefold thresholds, Manuscripta Math. 114 (2004), 281–304.
[25] M. Reid, Canonical 3-folds, in: A. Beauville (Ed.), Journées de géométrie algébrique d’Angers
1979, Sijthoff & Noordhoff, Alphen, 1980, pp. 273–310.
[26] K. Schwede and K. E. Smith, Globally F -regular and log Fano varieties, Adv. Math. 224 (2010),
863–894.
[27] K. Schwede and K. Tcuker, On the behavior of test ideals under finite morphisms, J. Algebraic
Geom. 23 (2014), 399–443.
[28] K. Schwede and W. Zhang, Bertini theorems for F -singularities, Proc. London Math. Soc. 107
(2013), 851–874.
[29] H. Tanaka, Minimal model programme for excellent surfaces, arXiv:1608.07676, preprint (2016).
[30] K. Watanabe, Plurigenera of normal isolated singularities. I, Math. Ann. 250 (1980), 65–94.
[31] K.-i. Watanabe, F -regular and F -pure normal graded rings, J. Pure Appl. Algebra 71 (1991),
341–350.
Graduate School of Mathematical Sciences, University of Tokyo, 3-8-1 Komaba,
Meguro-ku, Tokyo 153-8914, Japan
E-mail address: [email protected]
Graduate School of Mathematical Sciences, University of Tokyo, 3-8-1 Komaba,
Meguro-ku, Tokyo 153-8914, Japan
E-mail address: [email protected]
| 0math.AC
|
1
Modulus Zero-Forcing Detection for MIMO
Channels
arXiv:1801.00947v2 [cs.IT] 10 Jan 2018
Sha Hu and Fredrik Rusek
Abstract
We propose a modulus arithmetic based zero-forcing (MZF) detector for multi-input multi-output
(MIMO) channels. Traditionally, a ZF detector completely eliminates interference from other symbol
layers when detecting a particular symbol layer, which results in suboptimal performance due to noiseenhancement. The only constraint for application of our proposed MZF detector is that the transmitter
must employ a finite cardinality M quadrature-amplitude-modulation (QAM) alphabet. With that, the
modus operandi of the MZF is to allow for integer-valued interference and then remove it by modulus
arithmetic operations.
Index Terms
Multi-input multi-output (MIMO), modulus, zero-forcing (ZF), quadrature-amplitude-modulation
(QAM), pulse-amplitude-modulation (PAM), linear minimum-mean-square-error (LMMSE), sphere-decoding
(SD), lattice-reduction (LR), Lenstra-Lenstra-Lovász (LLL).
I. I NTRODUCTION
We consider a standard multi-input multi-output (MIMO) channel model with a received signal
ỹ expressed as
ỹ = H̃ x̃ + ñ,
(1)
where H̃, x̃ and ñ are the complex-valued MIMO channel, transmitted symbols and Gaussian
noise, respectively.
The authors are with the Department of Electrical and Information Technology, Lund University, Lund, Sweden (email:
{firstname.lastname}@eit.lth.se).
2
Given received signal model (1), detecting x̃ is referred to as a MIMO detection problem,
which has a history that can be traced back about half a century and a review on it can be found
in e.g., [1]. In general, maximum likelihood (ML) detection [2] yields optimal performance
but with prohibitive complexity when the MIMO dimension and/or the input alphabet has large
cardinality. Effective implementations of ML detection, such as sphere-decoding (SD) [3], [4] can
significantly reduce the complexity, but not overcome an exponential complexity in the number
of symbol layers CITE JALDEN’S PAPER. On the other hand, linear detectors [2] such as zeroforcing (ZF) and linear minimum-mean-square-error (LMMSE), have low complexities, but also
suboptimal performances. One direction for improving linear detectors is lattice-aided-reduction
(LAR) [5] based approaches, which use lattice-reduction (LR) algorithms, e.g., Lenstra-LenstraLovász (LLL), to find a short and nearly orthogonal basis for the lattice induced by the MIMO
channel [6].
Other than the existing approaches [1], as the transmitted symbols are drawn from finite alphabets such as quadrature-amplitude-modulation (QAM) and pulse-amplitude-modulation (PAM)
symbols, the modulus can also be used in MIMO detection for improving the detection performance. The modulus operation has been used in Tomlinson-Harashima precoding (THP) [7],
[8] as suboptimal approximation for dirty-paper coding (DPC) [9], and recently it has also be
considered in the designs of integer-forcing (IF) receives for MIMO channles [13], [15], [16].
The IF scheme in [13] requires the transmitter to employ a same lattice code [14] for each
transmitted layer, which does not apply in most of current communication systems. Besides,
when higher-oder modulations such as M -PAM are used, designing lattice code over ZM is
challenging [14]. Simpler IF receivers dealing with linear binary codes such as turbo and LDPC
are proposed in [15], [16], and the designs follow the same as in [13]. One disadvantage of
the IF receivers in [13], [16] is that, each transmit antenna needs a separate encoding/decoding
process, which is not the case in practical LTE systems where one codeword is split among
transmit-antennas. Moreover, the IF design in [15] needs a separate encoding/decoding process
per transmit antenna and per bit-layer in higher-order modulations. Another advantage is that,
the receiver has to detect the linear combinations of codewords for all transmit-antennas first,
followed by a matrix inversion (over a finite-field) process to recover the original codeword on
each layer.
To overcome these disadvantages in previous IF receiver designs, we consider a new approach
to improve linear detection with modulus operation, namely, the proposed modulus ZF (MZF)
3
detection. Note that, MZF is conceptually different from previous IF receivers, although they
share quite some similarities. The fundamental difference is that, with MZF, there is no encoding/decoding process needed which are implemented on a finite-filed with IF receivers to recover
the transmit symbols (i.e., the linear combinations of codewords across all transmit-antennas).
Alternatively, we design MZF detector such that the transmit symbols on each transmit-antenna
can be recovered directly by modulating away the interferences from the remaining transmitantennas, and the symbol detection on different transmit-antenna is independent and fully in
parallel. Such a principle simplifies the operation and can be well cooperated into practical
systems such as LTE. To achieve such a nice property, with MZF the modulus matrix is carefully
designed and optimized according to the specific modulation-order.
II. P RELIMINARIES
We start with reviewing the standard ZF detection. Before proceeding, without loss of generality, the matrix H̃ is always assumed to be a square matrix, obtained by a QR factorization or
padding zero rows to the matrix if necessary. With the following definitions,
R{ỹ}
R{x̃}
R{ñ}
R{H̃} −I{H̃}
, x =
, n =
, H =
,
y=
I{ỹ}
I{x̃}
I{ñ}
I{H̃} R{H̃}
(2)
we can rewrite (1) as a real-valued model
y = Hx + n
(3)
where the K ×K channel matrix H is known to the receiver, x = [x1 . . . xK ]T contains PAM
√
symbols from an alphabet A = {±1, ±3, . . . , ±( M − 1)}, and n is random Gaussian noise
with a covariance matrix (N0 /2)I. As the transmit power depends on M , the signal-to-noise
(SNR) is defined as
SNR = 2E[|xk |2 ]/N0 .
(4)
x̂ = QA H + y ,
(5)
The ZF detector is given by
where QA (·) denotes entry-wise quantization to the nearest point in A. This can be slightly
rewritten as
x̂k = QA (rk ),
1≤k≤K
4
with
rk = δ k H + y
(6)
and
δ k = [0| .{z
. . 0} 1 0| .{z
. . 0}].
k−1
K−k
For later use, we note that we may just as well replace the “1” with any arbitrary scalar value
and equivalently work with
rk = τ xk + wk ,
(7)
where wk is zero-mean complex Gaussian noise with variance τ 2 (N0 /2)kδ k H + k2 . Accordingly,
the post-processing SNR, which is independent of τ , becomes
γk =
SNR
,
kδ k H + k2
(8)
where SNR is defined in (4).
III. D ESCRIPTION OF THE P ROPOSED MZF D ETECTION
A main issue with ZF is that kδ k H + k2 in (8) is typically large and results in noise-enhancement.
To combat that, we make use of the underlying idea of THP but apply it to detection, without
any involvement of the transmitter. We propose to replace (6) with
rk = (τ δ k + q k )H + y
(9)
where q k = [qk1 , qk2 , . . . qkK ] and qk` ∈ 2Z, i.e., the even integers. With that,
rk = τ xk +
K
X
qk` x` + wk ,
(10)
`=1
and the post processing SNR changes to
γk =
τ 2 SNR
,
k(τ δ k + q k )H + k2
(11)
rather than (8).
Note that (10) coincides with the received signal per user in a vector perturbation (VP) system
[10], [11]. Therefore, further processing of (10) follows the same steps as those for VP. Based
on (10), hard-output detection of xk from rk can be obtained based on the following property.
5
Property 1. Let
y =z+α
M
X
p m bm ,
m=1
where α ≥ 1, |z| < 2, and pm , (bm −1) ∈ 2Z. Then,
M
P
(y mod 4α) − 2α,
if 12
pm is odd,
z=
m=1
(y + 2α) mod 4α − 2α, otherwise.
(12)
Proof. See Appendix A.
Let us now consider α = 1 (other values are considered in Section IV-A). In view of Property
1, we see that qk` and x` in (10) qualify as pm and bm . Further, from the condition |z| < 2 in
Property 1, we have that τ must be selected so that
√
τ max |a| = τ ( M − 1) < 2.
a∈A
To finalize the detector, we let
K
P
(rk mod 4) − 2,
if 21
qk` is odd,
zk =
`=1
(r + 2) mod 4 − 2, otherwise,
k
(13)
(14)
which can be expressed as
zk = τ xk + w̃k ,
where w̃k has a complicated distribution due to the modulus operation. The detected symbol x̂k
can now be obtained as
x̂k = Qτ A (zk ),
(15)
where the quantization is implemented on τ A, with the definition τ A = {τ x : x ∈ A}.
√
We remark that, the choice τ = 2/( M − 1) is not suitable in (13). This is so since if
√
τ xk +wk = 2+, for some small > 0, then zk = −2+. However, provided that τ 2/( M −1),
such wrap seldomly happens at high SNR and w̃k = wk with high probability. Observe that
for constellation points xk with small magnitude, then w̃k = wk with much higher probability
than for constellation points xk of large magnitude. To ensure equal error probability for all
constellation points, we design τ such that:
The distance from 2 to the largest constellation point in τ A is half the distance between two
points in τ A.
6
Following this rule results in
√
τ = 2(1−log2 (
M ))
,
(16)
√
where with out loss of generality we assume that log2 ( M ) is an integer in the rest of the
paper. With that, we have that w̃k is “nearly” Gaussian at high SNR; see VP-PAPER for further
details.
To optimize the receiver, we should solve
= arg max γk
q opt
k
qk
= arg min k(τ δ k + q k )H + k2
(17)
qk
where elements of q k are even integers. We rewrite (17) as
+
+ 2
q opt
k = arg min kτ δ k H − q k (−H )k
qk
= arg min kbk − q k Bk2 ,
(18)
qk
which is an instance of sphere detection over the integers [4], and
bk = τ δ k H + ,
(19)
B = −H + .
(20)
Without any further extensions, pseudo-code for implementation of the MZF detector is given
in Algorithm 1. We remind the reader that the inputs H and y to the algorithm are assumed to
be real-valued, while M denotes the cardinality of the complex-valued QAM constellation. We
provide a worked out example of MZF detection in Appendix B to illustrate the process.
A. Some Remarks on the MZF Detection
With the principle of MZF detection introduced, we have a few important remarks as follows.
Remark 1. The MZF is an extension of the ZF, where the latter is the special case of the former
when
q opt
k − δk
where
q opt
k
2
= 0,
(21)
denotes the Hadamard product. Therefore, from the perspective of post-processing SNR
γk , MZF is always superior to ZF.
7
Remark 2. Following Remark 1, when (21) holds, the modulus operation degrades performance
and should be removed.
Remark 3. In general the minimum value achieved by q opt
in (18) increases as τ decreases.
k
That is, for alphabets with large cardinality, the gain of MZF decreases.
To resolve the issue in Remark 3 and to further improve performance, some useful extensions
will be introduced in Sec. IV.
B. Discussion on the Diversity-Multiplexing Trade-off (DMT)
In the designs of IF receivers [13], [15], [16], the target is to directly optimize
+ 2
popt
k = arg min kpk H k ,
(22)
pk
given the constraint that P ∈ ZK×K is full-rank1 over R, where P it the matrix comprises all
vectors pk . Comparing to (17), it can be seen that, with MZF we constrain pk in (22) to be
1
pk = δ k + q k ,
τ
(23)
where q k comprises even integers. With (23) it holds that
1
P = I + Q,
τ
(24)
where τ < 1 and the elements of Q are even integers. Therefore, with MZF, the degrees-offreedom (DoFs) in designing P is less than that of an IF receiver2 . Nevertheless, we have the
following Property 2 (which is an extension of [13, Theorem 5]) that shows that, the MZF also
achieves the optimal DMT [18] as IF receiver does. Therefore, from an information-theoretic
perspective, the MZF detection does not scarifies much performance compared to an IF receiver,
while the latter one has much higher encoding/decoding complexity.
1
The rationale behind is that, in order to have P mod p to be full-rank over Zp (to recover the codewords from their linear
combinations after decoding on Zp ), it suffices to check whether P is full-rank over R, given that the magnitudes of the elements
of P are upper-bounded by a constant [17, Th. 11].
2
However, with MZF matrix P is not required to be full-rank.
8
Property 2. For a complex-valued MIMO channel with K transmit-antennas, N ≥ K receiveantennas, and independent and identically distributed (i.i.d.) Rayleigh fading, the achievable
DMT with the MZF detection3 is
r
dMZF = N 1 −
, 0 ≤ r ≤ K.
K
(25)
Proof. See Appendix C.
IV. E XTENSIONS OF THE MZF D ETECTION
In this section we introduce some extensions to the basic MZF detector for further improving
its performance. While Extension 1 and 4 are generalizations of the basic algorithm, Extension
2 is to resolve the issue mentioned for large cardinality alphabets and improve the performance
for weak bit-layers, while Extension 3 is a decision feedback version of Extension 2.
A. Extension 1: A scaled modulus
This first extension arises from a slight relaxation of α = 1 in the MZF detector. From Property
1, we can replace (14) as
K
P
(rk mod 4α) − 2α,
qk` is odd,
if 12
zk =
`=1
(r + 2α) mod 4α − 2α, otherwise.
k
(26)
This requires us to optimize, instead of (17),
opt
(q opt
) = arg min k(τ δ k + αq k )H + k2 .
k ,α
(27)
|α|≥1, q k
Solving (27) is harder than solving (17) since it can be regarded as an instance of non-coherent
sphere detection. Instead, we solve (17) first, and then plug the optimal solution into (27) and
solve for the optimal α. That is, q opt
is obtained with (17), and
k
+ 2
αopt = arg min k(τ δ k + αq opt
k )H k ,
|α|≥1
(28)
opt
where we sligthly abused notation since the pair (q opt
) is in general not jointly optimal
k ,α
in the sense of (27). Although Extension 3 is intuitive, the gain seems marginal according to
numerical results.
3
In this case, the size of H corresponding to the real-valued model in (3) is 2N ×2K.
9
B. Extension2: Bit-wise Modulus Zero-Forcing
An underlying assumption of this extension is that the bit-mapping to the symbols in A is
such that the constellation has an additive structure. By this we mean that a PAM symbol xk
should be of the form
√
log2 ( M )
X
xk =
ukb 2b−1 ,
(29)
b=1
where ukb ∈ {±1} correspond to information bits.
Using Algorithm 1, the bits ukb are determined by the output x̂k = Qτ A (zk ), with setting
√
τ = 2(1−log2 (
M ))
. As M increases, τ decreases and so are the gains of MZF detection. To
resolve this for large values of M , we extend the symbol-based MZF detector in Algorithm 1
to a bit-wise version.
Note that we can rewrite (10) as
√
log2 ( M )
X
rk = τ
b−1
ukb 2
+
K
X
qk` x` + wk .
(30)
`=1
b=1
Assuming that we are interested in the n-th bit ukn , we let
x̃k =
n
X
ukb 2b−1 ,
(31)
b=1
which belongs to a 2n -PAM alphabet. Setting τ (n) = 21−n in (30) yields
√
log2 ( M )
1−n
rk = 2
X
x̃k +
b−n
ukb 2
b=n+1
It can be easily seen that
1
2
Plog2 (√M )
b=n+1
+
K
X
qk` x` + wk .
(32)
`=1
ukb 2b−n is an odd integer so it qualifies as a valid value
of qk` , and ukn can be detected as
ûkn = sign(zk ).
(33)
Therefore, for each bit-layer, a different value of τ is used and only a sign operation is needed
for detection. Extension 2 has a complexity increment over Algorithm 1 since an optimization
to find q k is needed for each bit-layer.
Note that, according to Remark 2, when detecting the last bit-layer and if (21) holds, the ZF
estimate shall be used for detection, while for detecting the other layers, modulus operations are
still needed to cancel the transmitted bits corresponding to higher bit-layers. Psuedo-code for
the MZF with this extension is summarized in Algorithm 2.
10
C. Extension 3: A decision feedback version of Extension 2
An obstacle with Extension 2 is that τ (n) decreases as n grows, and as previously mentioned,
performance deteriorates. Small values of n correspond to weak bit-layers, and large n correspond
to strong bit-layers. Thus, with Extension 2, predominantly the weak bit-layers can gain by the
MZF, while the gain could be minuscule for strong bit-layers. A gain for weak bit-layers is
important since it is typically these bit-layers that limit ultimate performance. However, we can
also harvest a gain for strong bit-layers via a decision feedback mechanism. To prevent error
propagation in decision feedback equalization, strong bits are typically detected first and then
canceled. That option is not available for MZF, rather we detect the weakest bit-layer first and
then move on to stronger ones.
The method works as follows. First set n = 1 and follow Extension 2 verbatim to obtain
û1 = [û11 . . . ûK1 ]T . For notational convenience, define y 1 = y. Now construct
y2 =
1
(y − H û1 ) .
2 1
(34)
Provided that û1 is correct, y 2 is described with the same MIMO channel as y 1 , but with
√
√
M /2-PAM rather than M -PAM inputs. Next, move on to n = 2 and keep τ (2) = 1. Since nor
the value of τ (2) neither the channel H has changed, the optimal vector q k for n = 2 coincides
with that already found for n = 1. We then have that for y 2
√
log2 ( M )
rk = uk2 +
X
b−2
ukb 2
b=3
+
K
X
1
qk` (x` − ûk1 ) + wk ,
2
`=1
(35)
and ûk2 is obtained by taking the sign of zk as in (33). We proceed by
y3 =
1
(y − H û2 ) ,
2 2
(36)
and continue the process until all bit-layers have been detected.
Similarly, according to Remark 2, whenever (21) holds, the ZF estimate shall be used. Pseudocode is provided in Algorithm 3.
Extension 3 is similar to Extension 2 in the sense that, the detection for all bit-layers only
needs to take the signs of zk as in (33), but it has less complexity since only one optimization of
(17) is needed which is shared for all bit-layers. A drawback with Extension 3 is that, as for all
decision-feedback based detectors, the processing of the bit-layers cannot be parallelized, which
is however, possible with Extension 2. Another drawback is potential error-propagation at low
SNRs.
11
D. Extension 4: Replacing ZF by LMMSE
So far we have introduced modulus arithmetic detection using ZF, however, H + can also be
replaced by other linear detectors4 such as LMMSE, which sets
−1
H + = H H HH H + N0 I .
(37)
In vector form, and with the introduction of a matrix T , the received signal after equalization is
T H + y = T x + T H + H − I x + T H + n,
where T = (τ I +Q). The target of optimizing q k in this case, is to minimize the interference
plus noise power that equals
q̃ k = arg min k(τ δ k + q k )Ek2 ,
(38)
E = H + H − I, N0 H + .
(39)
qk
where
Note that, when H + is the pseudo-inverse of H such as with ZF, E degrades to B, which
shows the generalization of the MZF concept. The reason for introducing Extension 4 is that,
the ZF detector is suboptimal to LMMSE at low SNRs, in which case the modulus operation
based on LMMSE can improve performance. Therefore, it is beneficial to use LMMSE instead
of ZF. Since only H + is replaced by LMMSE equalization in Extension 4, all Algorithms 1-3
still apply verbatim.
There are also many other possible variations of the MZF, but which we will not pursue any
further. Next we put an interest on comparing the MZF to a traditional LAR detector. The reason
is that, solving (17) involves significant complexity, and we put forth an approximated solution
based on LR with less computational efforts.
V. A S OLUTION BASED ON ,
AND A
C OMPARISON TO , L ATTICE R EDUCTION
Except for approximately solving (17) with LR, another reason for comparing MZF with LR
detection is that, the obtained MZF allows for a direct comparison to LAR detectors. In LAR
as well as the MZF, the most burdening task is to execute the LLL algorithm (or other similar
algorithms), thus the complexities of LAR and MZF become virtually identical. As we will
demonstrate, the detection-performance of MZF is superior in some cases.
4
This is also known as regularized perturbation in VP [10].
12
A. A quick review of LAR
Given (3), LAR starts by performing the LLL algorithm on H, so that we obtain H̄ = HT
where T is unimodular and H̄ is nearly orthogonal. With z = T −1 x we have
y = H̄z + n.
(40)
Performing ZF based on H̄ and quantizing to the nearest integers gives
ẑ = QZ (H̄
−1
y)
(41)
from which one can obtain
x̂ = QA (T ẑ).
Clearly, once H̄ has been established, the remaining steps are of miniscule complexity.
At this point, a reasonable question is, what the relation between LAR and MZF is, and
whether they are equivalent? The answers to these questions are that, they are closely related,
but not equivalent. Prior to quantization in (41), we can write
r = H̄
−1
y
= T −1 x + w.
(42)
Since T is unimodular, so is T −1 .
On the other hand, written in vector form, (9) equals
r = (τ I + Q) H −1 y
= (τ I + Q) x + w.
(43)
Comparing (42) and (43) with τ = 1, we see that in both cases r equals an integer-valued
matrix multiplied with the data symbols, plus noise. However, the matrix T −1 in (42) has no
particular structure (besides being unimodular) so the modulus operation in (14) is not available.
This makes LAR, i.e., (42) and MZF, i.e., (43) fundamentally different, as the structure of (42)
requires further processing in the form of (41) while (43) allows for further processing via (14).
B. An approximate solution to (17) based on LLL
In (17) we have the following problem to solve
q opt = arg min kb − qBk2 ,
q
(44)
13
where we removed the subscript k, and the vectors are row-vectors. Perform the LLL algorithm
to B T so that we have
B̄ = B T T .
Since B = −H + the LLL algorithm needs, similar to LAR, to be executed only once per
coherence interval. We can now proceed as in the LAR case,
−1
ẑ = QZ (B̄ bT )
followed by
q opt = [Q2Z (T ẑ)]T .
(45)
Note that, the optimization (44) itself is also a MIMO detection problem (but only needs
to run once per coherence-interval), therefore, there are also other low-complexity suboptimal
algorithms to solve (44), such as using ZF or partial marginalization [12]. In the simulations,
we will focus on the optimal SD and the suboptimal LLL solutions for (44), respectively.
VI. N UMERICAL R ESULTS
In this section, we show some numerical results of the proposed MZF detector, as well as
its extensions. In all tests, we test with K ×K real-valued MIMO channels (each element is an
independent and identically distributed Gaussian variable with a zero-mean and unit-variance)
√
with M -PAM modulated symbols that are transfered from K/2×K/2 complex-valued MIMO
channels and M -QAM modulated symbols. We simulate 50,000 channel realizations for each of
the tests.
A. SINR improvements
In Fig. 1 we show the post-processing SNR improvements with the MZF detector using
Algorithm 1, and compare to a traditional ZF detector with different PAM modulations (i.e., τ
values). As can be seen, the SNRs are greatly improved, especially for low-order modulations (or
the weak bit-layers of high-order modulations with Extension 2). When τ decreases, the gains
become smaller. We also test the MZF with Extension 1, where we can observe only marginal
gains (not shown in Fig. 1). Therefore, in the remaining tests we set α = 1.
14
1
0.9
0.8
0.7
CDF
0.6
0.5
0.4
0.3
0.2
=1
= 1/2
= 1/4
0.1
0
0
2
4
6
8
10
12
14
16
SNR gain of MZF over ZF in dB
Fig. 1.
SNR gains under real-valued 12×12 MIMO with 2-PAM, 4-PAM and 8-PAM modulations, respectively.
100
10-1
BER
10-2
10-3
ZF, bit-layer 1
ZF, bit-layer 2
MZF-SD, bit-layer 1
MZF-SD, bit-layer 2
MZF-SD-Ext.2, bit-layer 1
MZF-SD-Ext.3, bit-layer 2
ML, bit-layer 1
ML, bit-layer 2
10-4
10-5
0
5
10
15
20
25
SNR [dB]
Fig. 2.
Uncoded BER under real-valued 6×6 MIMO with 4-PAM modulation.
B. Uncoded bit-error-rate (BER)
Next we show uncoded BER performance. In Fig. 2 we compare MZF with ZF and ML
under 6×6 MIMO with 4-PAM modulation. The MZF uses SD to find the optimal q k . As can be
seen, the MZF without extensions outperforms the ZF with more than 2 dB at 0.1% BER. With
Extension 2, the BER of the first bit-layer (weaker layer) is greatly improved by more than 4 dB
at 0.1% BER and outperforms the second bit-layer, which justifies the application of Extension
15
100
10-1
BER
10-2
10-3
ZF, bit-layer 1
ZF, bit-layer 2
ZF, bit-layer 3
MZF-SD-Ext.2, bit-layer 1
MZF-SD-Ext.2, bit-layer 2
ML, bit-layer 1
ML, bit-layer 2
ML, bit-layer 3
10-4
10-5
5
10
15
20
25
30
SNR [dB]
Fig. 3.
Uncoded BER under real-valued 4×4 MIMO with 8-PAM modulation.
100
ZF
LAR
MZF-LLL
MZF-SD
LMMSE
MZF-SD-Ext.4
10-1
BER
10-2
10-3
10-4
10-5
0
2
4
6
8
10
12
14
16
SNR [dB]
Fig. 4.
Uncoded BER under real-valued 8×8 MIMO with 2-PAM modulation.
3. With Extension 3, where the feedbacks of the first bit-layer are used, the BER of the second
bit-layer is also improved by more than 3 dB at 0.1% BER compared to MZF with Extension
2. The gaps between ZF and ML are significantly reduced by the MZF, and the slopes of BER
with MZF are also much steeper than the ZF, and close to those with the ML. However, as
also can be observed, the MZF has only marginal gains at low SNRs, and the decision-feedback
approach performs even worse due to inaccurate feedbacks. This obstacle can be relieved by
16
using LMMSE based approaches, i.e., Extension 4.
In Fig. 3 we repeat the tests in Fig. 2 under 4×4 MIMO with 8-PAM modulation, that is,
three bit-layers are considered. The MZF with Extension 2 using SD is compared to ZF and
ML. As already shown in Fig. 1, setting τ = 1/4 for detecting the third-layer (strongest layer)
only has small gains, and the BER performance is also close to ZF and therefore are not shown
in Fig. 3. Nevertheless, the BER of the first and second bit-layers are significantly boosted by
MZF. As can be seen, the MZF performs around 3 dB better than the ZF at 0.1% BER for the
second bit-layer, and 7 dB better for the first bit-layer. Since the weakest bit-layer usually has a
stronger impact on the decoding performance, the gains for the first bit-layer is of importance.
C. Comparison with LAR
In Fig. 4 we compare the MZF with the LAR under 8×8 MIMO with 2-PAM modulation.
The MZF uses both SD and the approximate LLL method to find the optimal q k . As can be
seen, MZF outperforms LAR with more than 1.5 dB at 0.1% BER, with a similar complexity
for running LLL algorithm for LR.
Moreover, with Extension 4 (the LMMSE based detection), the BERs at low SNRs are
also improved with the MZF which is inferior to the original ZF based MZF at high SNRs.
Nevertheless, with Extension 4, the MZF is 4 dB better than a normal ZF, and more than 2dB
better than a normal LMMSE detector at 0.1% BER. Another observation is that, the SD based
MZF is more than 2 dB better than the LLL based MZF, which shows that optimal selection of
q k is important.
VII. S UMMARY
We have proposed a novel modulus arithmetic based zero-forcing (MZF) detector for multiinput multi-output (MIMO) channels, with a number of possible extensions of the basic algorithm.
The MZF detector shows significant gains in terms of post-processing signal-to-noise-ratio (SNR)
and bit-error-rate (BER) compared to traditional linear detectors, at medium and high SNR
scenarios and in particular for weak bit-layers. At low SNRs and with large cardinalities of the
input alphabet, we have provided several possible extensions to improve the performance of the
MZF. Finding the optimal modulus matrix itself is a burdening MIMO detection problem, but
it needs to be done only once per a coherence-interval of the MIMO channel using spheredecoding (SD) or other suboptimal algorithms. In particular, with a similar complexity, MZF
17
with lattice-reduction (LR) based approaches outperforms the traditional lattice-aided-reduction
(LAR) detector, which justifies its potential in MIMO detection.
A PPENDIX A: P ROOF OF P ROPERTY 1
Since pm , (bm −1) ∈ 2Z, we let pm = 2p̃m and bm = 2b̃m +1, where p̃m , b̃m ∈ Z. Then,
y = z+α
M
X
p m bm
m=1
= z + 4α
M
X
p̃m b̃m + 2α
m=1
Since |z| < 2 and α > 1, it holds that z +2α > 0. If
M
X
p̃m .
(46)
m=1
1
2
M
P
M
P
pm =
m=1
p̃m is odd, we have
m=1
y mod 4α = z + 2α;
Otherwise, if
M
P
(47)
p̃m is even, it also holds that
m=1
(y + 2α) mod 4α = z + 2α.
(48)
Combing (47) and (48), z can be obtained as in (12).
A PPENDIX B: A 4×4 EXAMPLE FOR APPLYING THE MZF
DETECTION
Below we give a 4×4 real-valued MIMO example with 4-PAM modulation to illustrate the
process of MZF detection, with assuming the channel, transmitted symbol vector and received
signal vector as
−6
0
−1
5
1
3
−3 −2 −1 1
−1
1
H =
,
x
=
,
y
=
,
1 −5 −6 0
−1
15
1 −1 −3 −2
1
11
respectively. Then it can be shown that
−5
−55
30
−40
35 −59 −25 58
1
H+ =
,
185 −30 40 −5 −55
25 −58 35 −59
18
and the ZF estimate of x equals
x̃ZF
−60
309
1
= H +y =
,
185 −730
−107
where only the third symbol is correctly detected.
Next we use the basic MZF detection with Algorithm 1. Setting τ = 1 and run SD for
optimization (17) yields an optimal Q as
−2
0
Q=
0
2
0
0
0
0
.
0 −2 0
0 0 −2
0
2
We first see that, the MZF shall reuse the ZF estimates for the first and third layers based on (21).
Then, we see that with Q, the post-processing SNR (assuming the noise power equals 1) for the
second bit-layer (which is identical to the fourth bit-layer) is increased from 1/kδ 2 H + k = 185/47
to 1/k(δ 2 + q 2 )H + k2 = 185/27. Next we compute estimates with the MZF for the these two
layers.
For the second layer, according to (14) we have
r2 = (δ 2 + q 2 )x̃ZF
−60
309 −1151
1
[0 1 2 0]
,
=
=
−730
185
185
−107
and
z2 = (r2 mod 4) − 2 =
−7
.
38
Similarly, for the fourth layer we have
r4 = (δ 4 + q 4 )x̃ZF
−60
309 −13
1
=
[2 0 0 − 1]
,
=
−730 185
185
−107
19
and
z4 = (r4 mod 4) − 2 =
357
.
185
As can be seen, the MZF corrects both detections for the second and the fourth layers where
the ZF fails.
A PPENDIX C: P ROOF OF P ROPERTY 2
In the case where each transmit antenna encodes an independent data stream, the optimal
DMT [18] is
r
dML (r) = N 1 −
, 0 ≤ r ≤ K,
K
(49)
which can be achieved by joint maximum likelihood (ML) decoding, while the ZF (and also the
LMMSE) receiver attains a DMT
r
dZF (r) = (N − K + 1) 1 −
, 0 ≤ r ≤ K.
K
(50)
Since H̃ has i.i.d. Rayleigh entries, H̃ is full column rank with probability 1, and so is
H ∈ Z2N ×2K which is the real-valued representation H̃. To prove that the MZF detection can
achieve the optimal DMT, we follows a similar proof in [13] for IF receiver, which builds on a
result in [23] that showed that, uncoded signaling coupled with LR can achieve the full diversity
and with a multiplexing gain of zero.
We define a lattices Λ and its dual Λ∗ generated by 2N ×2K matrix H and 2K ×2N matrix
H + as [19]
Λ = Hp, p ∈ Z2K ,
Λ∗ = (H T )+ p, p ∈ Z2K ,
(51)
(52)
respectively. From [19]–[21], the i-th successive minimum i (Λ) of a lattice Λ is defined as the
smallest length r (with respect to the Euclidean norm) such that, there are i vectors in Λ of
length at most r that are linearly independent (with respect to R). By this definition, it holds
that [13]
21 (Λ) = min kpH T k2 ,
p∈Z2K
22K (Λ∗ ) =
min
P ∈Z2K×2K
rank(P )=2K
max kpk H + k2 .
1≤k≤2K
(53)
(54)
20
Then, we state the following Lemma 1, which follows the proof in [13, Appendix C].
Lemma 1. Let Λ be the lattice generated by H lattice generated by H according to (51). Then,
for a sufficiently large s > 0, the successive minimum 2K (Λ) of Λ satisfies
2
K (2K + 3)
∗
2
>s
P(2K (Λ ) > s) ≤ P
21 (Λ)
= P 21 (Λ) < s̃
(55)
(56)
≤ −γs̃2N (ln s̃)1+N ,
(57)
where 21 (Λ) is the first successive minimum for Λ, and
r
2K + 3
s̃ = K
.
s
(58)
The inequality (55) follows from [13, Lemma 4], and the equality (56) holds by switching
21 (Λ) to the other side, and the inequality (57) is a result of [13, Lemma 5] when s̃ > 0 and is
sufficiently small.
With the post-processing SNR γk in MZF detection in (11), we can show that for a given
target rate r ln SNR, the outage probability equals
P (R < r ln SNR) = P
max
Q∈2Z2K×2K
≤P
max
2K
X
1
k=1
2
2K
X
1
Q∈2Z2K×2K
rank(Q)=2K k=1
2
!
ln(1 + γk ) < r ln SNR
ln(1 + γk ) < r ln SNR
≤P
max
Q∈2Z2K×2K
min γk < SNRr/K
1≤k≤2K
rank(Q)=2K
=P
2
max
min
Q∈2Z2K×2K 1≤k≤2K
rank(Q)=2K
τ SNR
< SNRr/K
k(τ δ k + q k )H + k2
=P
min
Q∈2Z2K×2K
max k(τ δ k + q k )H + k2 > τ 2 SNR1−r/K .
1≤k≤2K
(59)
rank(Q)=2K
Using the inequality
k(τ δ k + q k )H + k2 ≤ kq k H + k2 + τ 2 kδ k H + k2 ,
(60)
21
and inserting (60) back into (59) yields
P (R < r ln SNR) ≤ P
min
Q∈2Z2K×2K
max kq k H + k2 > τ 2 SNR1−r/K
1≤k≤2K
− kδ k H + k2
rank(Q)=2K
τ
= P min
max kpk H + k2 >
SNR1−r/K − kδ k H + k2
2K×
2K
1≤k≤2K
4
P ∈Z
rank(P )=2K
τ2
+ 2
1−r/K
2
∗
= P 2K (Λ ) >
SNR
− kδ k H k
,
(61)
4
2
where the successive minimum 22K (Λ∗ ) is defined in (54). Comparing (61) to [13, Eq. (43)], it
can be seen that the outage probability with MZF detection could be inferior to that with the IF
receiver. However, the slope of the outage probability in high SNR regimes remains the same.
Using Lemma 1, it holds that
dMZF (r) = lim
− ln P (R < r ln SNR)
ln SNR
+ 2
1−r/K
τ2
∗
2
− ln P 2K (Λ ) > 4 SNR
− kδ k H k
SNR→∞
= lim
SNR→∞
2N
ln SNR
1+N
− ln −γŝ (ln ŝ)
SNR→∞
ln SNR
≥ lim
,
(62)
where
2K
ŝ =
τ
s
2K + 3
SNR1−r/K − kδ k H + k2
2K
≥
τ
r
2K + 3
.
SNR1−r/K
(63)
Inserting (63) into (62), it holds that
r
dMZF (r) ≥ N 1 −
.
K
(64)
But as the MZF detection is inferior to the ML detection, so is the DMT. Therefor, the equality
in (64) holds.
R EFERENCES
[1] S. Yang and L. Hanzo, “Fifty years of MIMO detection: The road to large-scale MIMOs, ” IEEE Commun. Surveys &
Tutorials, vol. 17, no. 4, pp. 1941-1988, Sep. 2015.
[2] S. M. Kay, “Fundamentals of statistical signal processing, volume I: Estimation theory,” Prentice Hall signal processing
series, 1993.
22
[3] A. Ghasemmehdi and E. Agrell, “Faster recursions in sphere decoding,” IEEE Trans. Inf. Theory, vol. 57, no. 6, pp.
3530-3536, Jun. 2011.
[4] E. Agrell, T. Eriksson, A. Vardy and K. Zeger, “Closest point search in lattices,” IEEE Trans. Inf. Theory, vol. 48, no. 8,
pp. 2201-2214, Aug. 2002.
[5] D. Wubben, R. Bohnke, V. Kuhn and K. Kammeyer, “Near-maximum-likelihood detection of MIMO systems using MMSEbased lattice reduction,” IEEE Int. Conf. Commun. (ICC), Jun. 2004, pp. 798-802.
[6] A. K. Lenstra, H. W. Lenstra and L. Lovász, “ Factoring polynomials with rational coefficients” Mathematische Annalen,
vol. 261 no. 4, pp. 515-534, Jul. 1982.
[7] M. Tomlinson, “New automatic equaliser employing modulo arithmetic,” Electronics Lett., vol. 7, no. 5, pp. 138-139, Mar.
1971.
[8] H. Harashima and H. Miyakawa, “Matched-transmission technique for channels With intersymbol interference,” IEEE
Trans. Commun., vol. 20, no. 4, pp. 774-780, Aug. 1972.
[9] M. H. Costa, “Writing on dirty paper,” IEEE Trans. Inf. Theory, vol. 29, no. 2, pp. 439-441, May 1983.
[10] B. M. Hochwald, C. B. Peel and A. L. Swindlehurst, “A vector-perturbation technique for near-capacity multiantenna
multiuser communication - Part II: Perturbation,” IEEE Trans. Commun., vol. 53, no. 3, pp. 537-544, Mar. 2005.
[11] J. Maurer, J. Jaldén, D. Seethaler and G. Matz, “Vector perturbation precoding revisited,” IEEE Trans. Signal Process.,
vol. 59, no. 1, pp. 315-328, Jan. 2011.
[12] S. Hu and F. Rusek, “A soft-output MIMO detector with achievable information rate based partial marginalization,” IEEE
Trans. Signal Processing, vol. 65, no.6, pp. 1622-1637, Mar. 2017.
[13] J. Zhan, B. Nazer, U. Erez, and M. Gastpar, “Integer-forcing linear receivers,” IEEE Trans. Inform. Theory, vol. 60, no.
12, pp. 7661-7685, Dec. 2014.
[14] S.-N. Hong and G. Caire, “Compute-and-forward strategies for cooperative distributed antenna systems,” IEEE Trans.
Inform. Theory, vol. 59, no. 9, pp. 5227-5243, Sep. 2013.
[15] S. H. Chae, M. Jang, and S.-K. Ah, “Multilevel coding scheme for integer-forcing MIMO receivers with binary codes”,
IEEE Trans. Wireless. Commun., vol. 16, no. 8, pp. 5428-5441, Aug. 2017.
[16] O. Ordentlich and U. Erez, “Achieving the gains promised by Integer-Forcing equalization with binary codes”, IEEE Conv.
Elect. and Electron. Eng. in Israel (IEEEI), Nov. 2010, pp. 000703-000707.
[17] B. Nazer and M. Gastpar, “Compute-and-forward: Harnessing interference through structured codes”, IEEE Trans. Inform.
Theory, vol. 57, no. 10,pp. 6463-6486, Oct. 2011.
[18] L. Zheng and D. N. C. Tse, “Diversity and multiplexing: A fundamental tradeoff in multiple-antenna channels, IEEE Trans.
Inf. Theory,” IEEE Trans. Inform. Theory, vol. 49, no. 5, pp. 1073-1096, May 2003.
[19] J. C. Lagarias, H. W. Lenstra, Jr., and C. P. Schnorr, “KorkinZolotarev bases and successive minima of a lattice and its
reciprocal lattice,” Combinatorica, vol. 10, no. 4, pp. 333-348, Dec. 1990.
[20] S. Lang, Introduction to Diophantine approximations, New York, NY, USA, Springer-Verlag, 1995.
[21] I. Aliev and M. Henk, “Successive minima and best simultaneous Diophantine approximations,” Monatshefte Math., vol.
147, no. 2, pp. 95-101, 2006.
[22] M. Taherzadeh, A. Mobasher, and A. K. Khandani, “Communication over MIMO broadcast channels using lattice-basis
reduction”, IEEE Trans. Inform. Theory, vol. 53, no. 12, pp. 4567-4582, Dec. 2007.
[23] M. Taherzadeh, A. Mobasher, and A. K. Khandani, “LLL reduction achieves the receive diversity in MIMO decoding”,
IEEE Trans. Inform. Theory, vol. 53, no. 12, pp. 4801-4805, Dec. 2007.
23
Algorithm 1 Modulus Zero-Forcing (MZF) Algorithm
H is K × K real-valued
y is K × 1 real-valued
1:
M is cardinality of QAM constellation
function x̂ =M ODULAR ZF(H, y, M )
√
2:
τ = 2(1−log2 (
3:
B = −H +
M ))
Preprocessing for each coherence interval
4:
for k = 1 to K
5:
bk = τ δ k H +
6:
2
Solve : q opt
k = arg min kbk − q k Bk
qk
7:
end for
Executed for every channel observation
8:
9:
10:
for k = 1 to K
+
rk = (τ δ k + q opt
k )H y
if q opt
k −δ k
if
opt
`=1 qk`
is odd then
zk = ((rk + 2) mod 4) − 2
16:
end if
17:
21:
PK
else
15:
20:
1
2
zk = (rk mod 4) − 2
14:
19:
= 0 then
else
13:
18:
2
zk = bk y
11:
12:
q opt
k
end if
x̂k = Qτ A (zk )
end for
end function
24
Algorithm 2 MZF Algorithm with Extension 2
H is K × K real-valued
y is K × 1 real-valued
1:
2:
M is cardinality of QAM constellation
function x̂ =M ODULAR ZFE XT 2(H, y, M )
√
N = log 2( M )
3:
B = −H +
4:
for n = 1 to N
5:
6:
τ (n) = 21−n
end for
Preprocessing for each coherence interval
7:
for k = 1 to K
8:
bk = δ k H +
9:
for n = 1 to N
10:
bkn = τ (n)bk
11:
2
Solve : q opt
k,n = arg min kbkn − q k Bk
qk
12:
13:
end for
end for
Executed for every channel observation
14:
15:
for k = 1 to K
for n = 1 to N
if n = N and q opt
k −δ k
16:
q opt
k
zk = bk y
17:
else
18:
19:
+
rk = (τ (n)δ k + q opt
k )H y
20:
zk = (rk mod 4) − 2
end if
21:
ûkn = sign(zk )
22:
23:
24:
25:
end for
end for
end function
2
= 0, then
25
Algorithm 3 MZF Algorithm with Extension 3
H is K × K real-valued
y is K × 1 real-valued
1:
2:
3:
4:
M is cardinality of QAM constellation
function x̂ =M ODULAR ZFE XT 3(H, y, M )
τ =1
√
N = log2 ( M )
B = −H +
Preprocessing for each coherence interval
5:
for k = 1 to K
6:
bk = τ δ k H +
7:
2
Solve : q opt
k = arg min kbk − q k Bk
qk
8:
end for
Executed for every channel observation
9:
ŷ = y
10:
for n = 1 to N
11:
ûn = 0
12:
for k = 1 to K
if q opt
k −δ k
13:
q opt
k
2
= 0, then
zk = τ δ k H + ŷ
14:
else
15:
16:
+
rk = (τ δ k + q opt
k )H ŷ
17:
zk = (rk mod 4) − 2
18:
end if
19:
ûkn = sign(zk )
20:
end for
21:
ûn = [û1n û2n . . . ûKn ]T
22:
ŷ = (ŷ − H uˆn )/2
23:
24:
end for
end function
| 7cs.IT
|
Convergence Rate of Riemannian Hamiltonian Monte Carlo and
Faster Polytope Volume Computation
arXiv:1710.06261v1 [cs.DS] 17 Oct 2017
Yin Tat Lee∗, Santosh S. Vempala†
October 18, 2017
Abstract
We give the first rigorous proof of the convergence of Riemannian Hamiltonian Monte Carlo,
a general (and practical) method for sampling Gibbs distributions. Our analysis shows that
the rate of convergence is bounded in terms of natural smoothness parameters of an associated
Riemannian manifold. We then apply the method with the manifold defined by the log barrier
function to the problems of (1) uniformly sampling a polytope and (2) computing its volume, the
latter by extending Gaussian cooling to the manifold setting. In both cases, the total number
2
of steps needed is O∗ (mn 3 ), improving the state of the art. A key ingredient of our analysis is
a proof of an analog of the KLS conjecture for Gibbs distributions over manifolds.
Contents
.
.
.
.
.
3
3
5
6
7
7
2 Basics of Hamiltonian Monte Carlo
2.1 Hamiltonian Monte Carlo on Riemannian manifolds . . . . . . . . . . . . . . . . . . .
8
10
3 Convergence of Riemannian Hamiltonian Monte Carlo
3.1 Basics of geometric Markov chains . . . . . . . . . . . . .
3.2 Overlap of one-step distributions . . . . . . . . . . . . . .
3.2.1 Proof Outline . . . . . . . . . . . . . . . . . . . . .
3.2.2 Variation of Hamiltonian curve . . . . . . . . . . .
3.2.3 Local Uniqueness of Hamiltonian Curves . . . . . .
3.2.4 Smoothness of one-step distributions . . . . . . . .
3.3 Convergence bound . . . . . . . . . . . . . . . . . . . . . .
13
13
15
15
16
19
21
23
1 Introduction
1.1 Results . . . .
1.2 Approach and
1.3 Practicality .
1.4 Notation . .
1.5 Organization
. . . . . . . .
contributions
. . . . . . . .
. . . . . . . .
. . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4 Improved analysis of the convergence
24
4.1 Improved one-to-one correspondence for Hamiltonian curve . . . . . . . . . . . . . . 25
4.2 Improved smoothness of one-step distributions . . . . . . . . . . . . . . . . . . . . . . 26
∗
†
University of Washington and Microsoft Research, [email protected]
Georgia Tech, [email protected]
1
5 Gibbs sampling on manifolds
29
5.1 Isoperimetry for Hessian manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.2 Sampling with the log barrier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
6 Polytope volume computation: Gaussian cooling on manifolds
6.1 Algorithm: cooling schedule . . . . . . . . . . . . . . . . . . . . .
6.2 Correctness of the algorithm . . . . . . . . . . . . . . . . . . . . .
6.2.1 Initial and terminal conditions . . . . . . . . . . . . . . .
6.2.2 Variance of Yx . . . . . . . . . . . . . . . . . . . . . . . .
6.2.3 Main lemma . . . . . . . . . . . . . . . . . . . . . . . . . .
6.3 Volume computation with the log barrier . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
32
32
32
32
35
37
38
7 Logarithmic barrier
7.1 Riemannian geometry on ML (G2 ) . . . .
7.2 Hamiltonian walk on ML . . . . . . . . .
7.3 Randomness of the Hamiltonian flow (ℓ0 )
7.4 Parameters R1 , R2 and R3 . . . . . . . . .
7.5 Stability of L2 + L4 + L∞ norm (ℓ1 ) . . .
7.6 Mixing Time . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
39
39
40
42
46
53
55
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
A Matrix ODE
59
B Concentration
61
C Calculus
61
D Basic definitions of Riemannian geometry
62
D.1 Curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
D.2 Hessian manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2
1
Introduction
Hamiltonian dynamics provide an elegant alternative to Newtonian mechanics. The Hamiltonian
H, which captures jointly the potential and kinetic energy of a particle, is a function of its position
and velocity. First-order differential equations describe the change in both.
∂H(x, v)
dx
=
,
dt
∂v
dv
∂H(x, v)
=−
.
dt
∂x
As we review in Section 2, these equations preserve the Hamiltonian H.
Riemannian Hamiltonian Monte Carlo (or RHMC) [26, 25][6, 7] is a Markov Chain Monte Carlo
method for sampling from a desired distribution. The target distribution is encoded in the definition
of the Hamiltonian. Each step of the method consists of the following: At a current point x,
1. Pick a random velocity y according to a local distribution defined by x (in the simplest setting,
this is the standard Gaussian distribution for every x).
2. Move along the Hamiltonian curve defined by Hamiltonian dynamics at (x, y) for time (distance) δ.
For a suitable choice of H, the marginal distribution of the current point x approaches the desired
target distribution. Conceptually, the main advantage of RHMC is that it does not require a
Metropolis filter (as in the Metropolis-Hastings method) and its step sizes are therefore not severely
limited even in high dimension.
Over the past two decades, RHMC has become very popular in Statistics and Machine Learning,
being applied to Bayesian learning, to evaluate expectations and likelihood of large models by sampling from the appropriate Gibbs distribution, etc. It has been reported to significantly outperform
other known methods [2, 25] and much effort has been made to make each step efficient by the use
of numerical methods for solving ODEs.
In spite of all these developments and the remarkable empirical popularity of RHMC, analyzing
its rate of convergence and thus rigorously explaining its success has remained an open question.
1.1
Results
In this paper, we analyze the mixing rate of Hamiltonian Monte Carlo for a general function f as a
Gibbs sampler, i.e., to generate samples from the density proportional to e−f (x) . The corresponding
Hamiltonian is H(x, v) = f (x) + 21 log((2π)n det g(x)) + 12 v T g(x)−1 v for some metric g. We show
that for x in a compact manifold, the conductance of the Markov chain is bounded in terms of
a few parameters of the metric g and the function f . The parameters and resulting bounds are
given in Corollary 28 and Theorem 30. Roughly speaking, the guarantee says that Hamiltonian
Monte Carlo mixes in polynomial time for smooth Hamiltonians. We note that these guarantees
use only the smoothness and Cheeger constant (expansion) of the function, without any convexity
type assumptions. Thus, they might provide insight in nonconvex settings where (R)HMC is often
applied.
We then focus on logconcave densities in Rn , i.e., f (x) is a convex function. This class of
functions appears naturally in many contexts and is known to be sampleable in polynomial-time
given access to a function value oracle. For logconcave densities, the current fastest sampling
algorithms use n4 function calls, even for uniform sampling [20, 23], and n2.5 oracle calls given a
3
warm start after appropriate rounding (linear transformation) [14]. In the prototypical setting of
uniform sampling from a polytope Ax ≥ b, with m inequalities, the general complexity is no better,
with each function evaluation taking O(mn) arithmetic operations, for an overall complexity of
n4 · mn = mn5 in the worst case and n2.5 · mn after rounding from a warm start. The work of
Kannan and Narayanan [11] gives an algorithm of complexity mn2 · mnω−1 from an arbitrary start
and mn · mnω−1 from a warm start (here ω is the matrix multiplication exponent), which is better
than the general case when the number of facets m is not too large. This was recently improved
to mn0.75 · mnω−1 from a warm start [13]; the subquadratic complexity for the number of steps is
significant since all known general oracle methods cannot below a quadratic number of steps. The
leading algorithms and their guarantees are summarized in Table 1.
Year
1997 [10]
2003 [21]
2009 [11]
2016 [13]
2016 [15]
This paper
Algorithm
Ball walk#
Hit-and-run#
Dikin walk
Geodesic walk
Ball walk#
RHMC
Steps
n3
n3
mn
3
mn 4
n2.5
2
mn 3
Cost per step
mn
mn
mnω−1
mnω−1
mn
mnω−1
Table 1: The complexity of uniform polytope sampling from a warm start, where each step of
e
every algorithm uses O(n)
bit of randomness. The entries marked # are for general convex bodies
presented by oracles, while the rest are for polytopes.
In this paper, using RHMC, we improve the complexity of sampling polytopes. In fact we do
this for a general family of Gibbs distributions, of the form e−αφ(x) where φ(x) is a convex function
over a polytope. When φ(x) is the standard logarithmic barrier function and g(x) is its Hessian,
1
1
we get a sampling method that mixes in only n 6 m 2 +
1
2
2
n5 m5
1
1
α 5 +m− 5
+
n3
α+m−1
steps from a warm start!
When α = 1/m, the resulting distribution is very close to uniform over the polytope.
Theorem 1. Let φ be the logarithmic barrier for a polytope M with m constraints and n variables.
Hamiltonian Monte Carlo applied to the function f = exp(−αφ(x)) and the metric given by ∇2 φ
with appropriate step size mixes in
!
1
2
1
3
3n3
1
1
n
m
e
+ m2 n6
O
+
α + m−1 α 31 + m− 13
steps where each step is the solution of a Hamiltonian ODE.
In recent independent work, Mangoubi and Smith [24] analyze Euclidean HMC in the oracle setting, i.e., assuming an oracle for evaluating φ. Their analysis formally gives a dimension-independent
convergence rate based on certain regularity assumptions such as strong convexity and smoothness
of the Hamiltonian H. Unfortunately, these assumptions do not hold for the polytope sampling
problem.
An important application of sampling is integration. The complexity of integration for general
logconcave functions is also n4 oracle calls. For polytopes, the most natural questions is computing
its volume. For this problem, the current best complexity is n4 · mn, where the factor of O(mn) is
the complexity of checking membership in a polytope. Thus, even for explicitly specified polytopes,
the complexity of estimating the volume from previous work is asymptotically the same as that
4
for a general convex body given by a membership oracle. Here we obtain a volume algorithm with
2
complexity mn 3 · mnω−1 , improving substantially on previous algorithms. The volume algorithm
is based using Hamiltonian Monte Carlo for sampling from a sequence of Gibbs distributions over
polytopes. We remark that in the case when m = O(n)1 , the final complexity is o(n4 ) arithmetic
operations, improving by more than a quadratic factor in the dimension over the previous best
complexity of Õ(n6 ) operations for arbitrary polytopes. These results and prior developments are
given in Table 2.
Year
1989 [5]
1989-93 [17, 4, 1, 18, 19]
1997 [10]
2003 [22]
2015 [3]
This paper
Algorithm
DFK
many improvements
DFK, Speedy walk, isotropy
Annealing, hit-and-run
Gaussian Cooling*
RHMC + Gaussian Cooling
Steps
n23
n7
n5
n4
n3
2
mn 3
Cost per step
mn
mn
mn
mn
mn
mnω−1
e
Table 2: The complexity of volume estimation, each step uses O(n)
bit of randomness, all except
*
the last for general convex bodies (the result marked is for well-rounded convex bodies). The
current paper applies to general polytopes, and is the first improvement utilizing their structure.
Theorem 2. For any polytope P = {x : Ax ≥ b} with m constraints and n variables, and any
ε > 0, the Hamiltonian
volume algorithm estimates the volume of P to within 1 ± ε multiplicative
2
−2
e
3
factor using O mn ε
steps where each step consists of solving a first-order ODE and takes time
e mnω−1 LO(1) logO(1) 1 and L is the bit complexity2 of the polytope.
O
ε
A key ingredient in the analysis of RHMC is a new isoperimetric inequality for Gibbs distributions over manifolds. This inequality can be seen as an evidence of a manifold version of the
KLS hyperplane conjecture. For the family of Gibbs distributions induced by convex functions with
convex Hessians, the expansion is within a constant factor of that of a hyperplane cut. This result
might be of independent interest.
1.2
Approach and contributions
Traditional methods to sample from distributions in Rn are based on random walks that take straight
line steps (grid walk, ball walk, hit-and-run). While this leads to polynomial-time convergence for
logconcave distributions, the length of each step has to be small due to boundary effects, and a
Metropolis filter (rejection sampling) has to be applied to ensure the limiting
distribution is the
desired one. These walks cannot afford a step of length greater than δ = O √1n for a distribution
in isotropic position, and take a quadratic number of steps even for the hypercube. The Dikin walk
for polytopes [11], which explicitly takes into account the boundary of polytope at each step, has a
varying step size, but still runs into similar issues and the bound on its convergence rate is O(mn)
for a polytope with m facets.
1
We suspect that the LS barrier [12] might be used to get a faster algorithm even in the regime even if m is
sub-exponential. However, our proof requires a delicate estimate of the fourth derivative of the barrier functions.
Therefore, such a result either requires a new proof or a unpleasantly long version of the current proof.
2
L = log(m + dmax + kbk∞ ) where dmax is the largest absolute value of the determinant of a square sub-matrix of
A.
5
In a recent paper [13], we introduced the geodesic walk. Rather than using straight lines in
Euclidean space, each step of the walk is along a geodesic (locally shortest path) of a Riemannian
metric. More precisely, each step first makes a deterministic move depending on the current point
(drift), then moves along a geodesic in a random initial direction and finally uses a Metropolis filter.
Each step can be computed by solving a first-order ODE. Due to the combination of drift and
geodesic, the local 1-step distributions are smoother than that of the Dikin walk and larger steps
can be taken while keeping a bounded rejection probability for the filter. For sampling polytopes,
3
the manifold/metric defined by the standard log barrier gives a convergence rate of mn 4 , going
below the quadratic (or higher) bound of all previous sampling methods.
One major difficulty with geodesic walk is ensuring the stationary distribution is uniform. For
high dimensional problems, this necessitates taking a sufficiently small step size and then rejecting
some samples according to the desired transition probabilities according to Metropolis filter. Unfortunately, computing these transition probabilities can be very expensive. For the geodesic walk,
it entails solving an n × n size matrix ODE.
Hamiltonian Monte Carlo bears some similarity to the geodesic walk — each step is a random
(non-linear) curve. But the Hamiltonian-preserving nature of the process obviates the most expensive ingredient, Metropolis filter. Due to this, the step size can be made longer, and as a result we
2
obtain a faster sampling algorithm for polytopes that mixes in mn 3 steps (the per-step complexity
remains essentially the same, needing the solution of an ODE).
To get a faster algorithm for volume computation, we extends the analysis to a general family of
Gibbs distributions, including f (x) = e−αφ(x) where φ(x) is the standard log-barrier and α > 0. We
show that the smoothness we need for the sampling corresponding to a variant of self-concordance
defined in Definition 46. Furthermore, we establish an isoperimetric inequality for this class of
functions. This can be viewed as an extension of the KLS hyperplane conjecture from Euclidean to
Riemannian metrics (the analogous case in Euclidean space to what we prove here is the isoperimetry
of the Gaussian density function multiplied by any logconcave function, a case for which the KLS
conjecture holds). The mixing rate for this family of functions is sublinear for α = Ω(1).
Finally, we study the Gaussian Cooling schedule of [3]. We show that in the manifold setting,
2
the Gaussian distribution e−kxk /2 can be replaced by e−αφ(x) . Moreover, the speed of Gaussian
Cooling depends on the “thin-shell” constant of the manifold and classical self-concordance of φ.
Combining all of these ideas, we obtain a faster algorithm for polytope volume computation.
The resulting complexity of polytope volume computation is the same as that of sampling uniformly
2
from a warm start: mn 3 steps. To illustrate the improvement, for polytopes with m = O(n) facets,
5
the new bound is n 3 while the previous best bound was n4 .
1.3
Practicality
From the experiments, the ball walk/hit-and-run seem to mix in n2 steps, the geodesic walk seems
to mix in sublinear number of steps (due to the Metropolis filter bottleneck) and RHMC seems to
mix in only polylogarithmic number of steps. One advantage of RHMC compared to the geodesic
walk is that it does not require the expensive Metropolis filter that involves solving n × n matrix
ODEs. In the future, we plan to do an empirical comparison study of different sampling algorithms.
We are hopeful that using RHMC we might finally be able to sample from polytopes in millions of
dimensions after more than three decades of research on this topic!
6
1.4
Notation
Throughout the paper, we use lowercase letter for vectors and vector fields and uppercase letter for
d
matrices and tensors. We use ek to denote coordinate vectors. We use dt
for the usual derivative,
df (c(t))
∂
is the derivative of some function f along a curve c parametrized by t, we use ∂v
for
e.g.
dt
k
th
the usual partial derivative. We use D f (x)[v1 , v2 , · · · , vk ] for the k directional derivative of f at
x along v1 , v2 , · · · , vk . We use ∇ for the usual gradient and the connection (manifold derivative,
defined in Section D which takes into account the local metric), Dv for the directional derivative
of a vector with respect to the vector (or vector field) v (again, defined in Section D), and Dt if
the curve v(t) is clear from the context. We use g for the local metric. Given a point x ∈ M , g
is a matrix with entries gij . Its inverse has entries g ij . Also, n is the dimension, m the number of
inequalities. We use dT V for the total variation (or L1 ) distance between two distributions.
1.5
Organization
In Section 2, we define the Riemannian Hamiltonian Monte Carlo and study its basic properties
such as time-reversibility. In Section 3, we give the the first convergence rate analysis of RHMC.
However, the convergence rate is weak for the sampling applications (it is polynomial, but not better
than previous methods). In Section 4, we introduce more parameters and use them to get a tighter
analysis of RHMC. In Section 5, we study the isoperimetric constant of f (x) = e−αφ(x) under the
metric ∇2 φ(x). In Section 6, we study the generalized Gaussian Cooling schedule and its relation
to the thin-shell constant. Finally, in Section 7, we compute the parameters we need for the log
barrier function.
7
2
Basics of Hamiltonian Monte Carlo
In this section, we define the Hamiltonian Monte Carlo method for sampling from a general distribution e−H(x,y) . Hamiltonian Monte Carlo uses curves instead of straight lines and this makes
the walk time-reversible even if the target distribution is not uniform, with no need for a rejection
sampling step. In contrast, classical approaches such as the ball walk require an explicit rejection
step to converge to a desired stationary distribution.
Definition 3. Given a continuous, twice-differentiable function H : M×Rn ⊂ Rn ×Rn → R (called
the Hamiltonian, which often corresponds to the total energy of a system) where M is the x domain
of H, we say (x(t), y(t)) follows a Hamiltonian curve if it satisfies the Hamiltonian equations
∂H(x, y)
dx
=
,
dt
∂y
dy
∂H(x, y)
=−
.
dt
∂x
(2.1)
def
We define the map Tδ (x, y) = (x(δ), y(δ)) where the (x(t), y(t)) follows the Hamiltonian curve with
the initial condition (x(0), y(0)) = (x, y).
Hamiltonian Monte Carlo is the result of a sequence of randomly generated Hamiltonian curves.
Algorithm 1: Hamiltonian Monte Carlo
Input: some initial point x(1) ∈ M.
for i = 1, 2, · · · , T do
R
1
(k)
Sample y (k+ 2 ) according to e−H(x ,y) /π(x(k) ) where π(x) = Rn e−H(x,y) dy.
1
With probability 12 , set (x(k+1) , y (k+1) ) = Tδ (x(k) , y (k+ 2 ) ).
1
Otherwise, (x(k+1) , y (k+1) ) = T−δ (x(k) , y (k+ 2 ) ).
end
Output: (x(T +1) , y (T +1) ).
Lemma 4 (Energy Conservation). For any Hamiltonian curve (x(t), y(t)), we have that
d
H(x(t), y(t)) = 0.
dt
Proof. Note that
∂H dx ∂H dy
∂H ∂H
∂H ∂H
d
H(x(t), y(t)) =
+
=
−
= 0.
dt
∂x dt
∂y dt
∂x ∂y
∂y ∂x
Lemma 5 (Measure Preservation). For any t ≥ 0, we have that
det (DTt (x, y)) = 1
where DTt (x, y) is the Jacobian of the map Tt at the point (x, y).
8
Proof. Let (x(t, s), y(t, s)) be a family of Hamiltonian curves given by Tt (x + sdx , y + sdy ). We write
u(t) =
∂
∂
x(t, s)|s=0 , v(t) =
y(t, s)|s=0 .
∂s
∂s
By differentiating the Hamiltonian equations (2.1) w.r.t. s, we have that
du
∂ 2 H(x, y)
∂ 2 H(x, y)
=
u+
v,
dt
∂y∂x
∂y∂y
dv
∂ 2 H(x, y)
∂ 2 H(x, y)
=−
u−
v,
dt
∂x∂x
∂x∂y
(u(0), v(0)) = (dx , dy ).
This can be captured by the following matrix ODE
∂ 2 H(x(t),y(t))
∂y∂x
2
− ∂ H(x(t),y(t))
∂x∂x
dΦ
=
dt
∂ 2 H(x(t),y(t))
∂y∂y
2
− ∂ H(x(t),y(t))
∂x∂y
!
Φ(t)
dx
dy
Φ(0) = I
using the equation
DTt (x, y)
dx
dy
=
u(t)
v(t)
= Φ(t)
.
Therefore, DTt (x, y) = Φ(t). Next, we observe that
d
−1 d
log det Φ(t) = Tr Φ(t)
Φ(t) = Tr
dt
dt
∂ 2 H(x(t),y(t))
∂y∂x
2
− ∂ H(x(t),y(t))
∂x∂x
∂ 2 H(x(t),y(t))
∂y∂y
2
− ∂ H(x(t),y(t))
∂x∂y
!
= 0.
Hence,
det Φ(t) = det Φ(0) = 1.
Using the previous two lemmas, we now show that Hamiltonian Monte Carlo indeed converges
to the desired distribution.
Lemma 6 (Time reversibility). Let px (x′ ) denote the probability density of one step of the Hamiltonian Monte Carlo starting at x. We have that
π(x)px (x′ ) = π(x′ )px′ (x)
R
for almost everywhere in x and x′ where π(x) = Rn e−H(x,y) dy.
Proof. Fix x and x′ . Let Fδx (y) be the x component of Tδ (x, y). Let V+ = {y : Fδx (y) = x′ } and
x (x) = x′ )}. Then,
V− = {y : F−δ
π(x)px (x′ ) =
1
2
Z
y∈V+
1
e−H(x,y)
+
x
2
det DFδ (y)
Z
y∈V−
e−H(x,y)
.
x (y)
det DF−δ
We note that this formula assumed that DFδx is invertible. Sard’s theorem showed that Fδx (N ) is
def
measure 0 where N = {y : DFsx (y) is not invertible}. Therefore, the formula is correct except for
a measure zero subset.
9
By reversing time for the Hamiltonian curve, we have that for the same V± ,
Z
Z
′ ′
′ ′
1
e−H(x ,y )
e−H(x ,y )
1
+
π(x′ )px′ (x) =
x′ (y ′ )
2 y∈V+ det DF−δ
2 y∈V− det DFδx′ (y ′ )
(2.2)
where y ′ denotes the y component of Tδ (x, y) and T−δ (x, y) in the first
and second
sum respectively.
A B
We compare the first terms in both equations. Let DTδ (x, y) =
. Since Tδ ◦ T−δ = I
C D
and Tδ (x, y) = (x′ , y ′ ), the inverse function theorem shows that DT−δ (x′ , y ′ ) is the inverse map of
DTδ (x, y). Hence, we have that
−1
A B
· · · −A−1 B(D − CA−1 B)−1
′ ′
DT−δ (x , y ) =
.
=
···
···
C D
′
x (y ′ ) = −A−1 B(D − CA−1 B)−1 . Hence, we have that
Therefore, we have that Fδx (y) = B and F−δ
−1
x′
=
det DF−δ
(y ′ ) = det A−1 det B det D − CA−1 B
|det B|
.
A B
det
C D
A B
= 1 (Lemma 5), we have that
C D
x′
det DF−δ
(y ′ ) = |det (DFδx (y))| .
Using that det (DTt (x, y)) = det
Hence, we have that
1
2
Z
y∈V+
1
e−H(x,y)
=
x
2
det DFδ (y)
=
1
2
−H(x′ ,y ′ )
Z
Z
y∈V+
e−H(x,y)
x′ (y ′ )
det DF−δ
′
y∈V+
′
e−H(x ,y )
x′ (y ′ )
det DF−δ
where we used that e−H(x,y) = e
(Lemma 4) at the end.
For the second term in (2.2), by the same calculation, we have that
Z
Z
′ ′
1
e−H(x,y)
e−H(x ,y )
1
=
x (y)
2 y∈V− det DF−δ
2 y∈V+ det DFδx′ (y ′ )
Combining both terms we have the result.
The main challenge in analyzing Hamiltonian Monte Carlo is to bound its mixing time.
2.1
Hamiltonian Monte Carlo on Riemannian manifolds
Suppose we want to sample from the distribution e−f (x) . We define the following energy function
H:
1
1
def
(2.3)
H(x, v) = f (x) + log((2π)n det g(x)) + v T g(x)−1 v.
2
2
One can view x as the location and v as the velocity. The following lemma shows that the first
variable x(t) in the Hamiltonian curve satisfies a second-order differential equation. When we view
the domain M as a manifold, this equation is simply Dt dx
dt = µ(x), namely, x acts like a particle
under the force field µ. (For relevant background on manifolds, we refer the reader to Appendix D).
10
Lemma 7. In Euclidean coordinates, The Hamiltonian equation for (2.3) can be rewritten as
Dt
dx
=µ(x),
dt
dx
(0) ∼N (0, g(x)−1 )
dt
where µ(x) = −g(x)−1 ∇f (x) − 12 g(x)−1 Tr g(x)−1 Dg(x) and Dt is the Levi-Civita connection on
the manifold M with metric g.
Proof. From the definition of the Hamiltonian curve, we have that
dx
= g(x)−1 v
dt
1 dx T
dv
1
dx
= −∇f (x) − Tr g(x)−1 Dg(x) +
Dg(x) .
dt
2
2 dt
dt
Putting the two equations together, we have that
dx
dv
d2 x
= − g(x)−1 Dg(x)[ ]g(x)−1 v + g(x)−1
2
dt
dt
dt
1
dx dx
1
dx T
dx
= − g(x)−1 Dg(x)[ ]
− g(x)−1 ∇f (x) − g(x)−1 Tr g(x)−1 Dg(x) + g(x)−1
Dg(x) .
dt dt
2
2
dt
dt
Hence,
T
dx dx 1
1
d2 x
dx
−1
−1 dx
+
g(x)
Dg(x)[
]
−
g(x)
= − g(x)−1 ∇f (x) − g(x)−1 Tr g(x)−1 Dg(x) .
Dg(x)
2
dt
dt dt
2
dt
dt
2
(2.4)
Using the formula of Christoffel symbols
Dt
dx
d2 x X dxi dxj k
= 2 +
Γ ek
dt
dt
dt dt ij
where
Γkij =
1 X kl
g (∂j gli + ∂i glj − ∂l gij ),
2
l
ijk
we have that
Dt
X dxi dxj
d2 x 1
dx
= 2 + g(x)−1
(∂j gli + ∂i glj − ∂l gij )el
dt
dt
2
dt dt
ijl
=
T
d2 x
dx dx 1
dx
−1
−1 dx
+
g(x)
Dg(x)[
]
−
g(x)
Dg(x) .
dt2
dt dt
2
dt
dt
Putting this into (2.4) gives
Dt
dx
1
= − g(x)−1 ∇f − g(x)−1 Tr g(x)−1 Dg(x) .
dt
2
Motivated by this, we define the Hamiltonian map as the first component of the Hamiltonian
dynamics operator T defined earlier. For the reader familiar with Riemannian geometry, this is
similar to the exponential map (for background, see Appendix D).
11
Definition 8. Let Hamx,δ (vx ) = γ(δ) where γ(t) be the solution of the Hamiltonian equation
′
Dt dγ
dt = µ with initial conditions γ(0) = x and γ (0) = vx . We also denote Hamx,1 (vx ) by Hamx (vx ).
We now give two examples of Hamiltonian Monte Carlo.
Example 9. When g(x) = I, the Hamiltonian curve acts like stochastic gradient descent for the
function f with each random perturbation drawn from a standard Gaussian.
Dt
dx
= −∇f (x).
dt
When g(x) = ∇2 f (x), the Hamiltonian curve acts like a stochastic Newton curve for the function
f + ψ:
−1
dx
Dt
= − ∇2 f (x)
∇(f (x) + ψ(x))
dt
where the volumetric function ψ(x) = log det ∇2 f (x).
Next we derive a formula for the transition probability in Euclidean coordinates.
Lemma 10. For any x ∈ M ⊂ Rn and s > 0, the probability density of the 1-step distribution from
x is given by
s
X
1
det (g(y))
2
−1
exp − kvx kx
(2.5)
px (y) =
|det(DHamx,δ (vx ))|
(2π)n
2
vx :Hamx,δ (vx )=y
where DHamx,δ (vx ) is the Jacobian of the Hamiltonian map Hamx,δ .
Proof. We prove the formula by separately considering each vx ∈ Tx M s.t. Hamx,δ (vx ) = y, then
summing up. In the tangent space Tx M, the point vx follows a Gaussian step. Therefore, the
probability density of vx in Tx M is as follows:
1
1
2
exp
−
kv
k
pTxx M (vx ) =
x x .
2
(2π)n/2
Let y = Hamx,δ (vx ) and F : Tx M → Rn be defined by F (v) = idM→Rn ◦ Hamx,δ (v). Here Rn is
the same set as M but endowed with the Euclidean metric. Hence, we have
DF (vx ) = DidM→Rn (y)DHamx,δ (vx ).
The result follows from px (y) = |det(DF (vx ))|−1 pxTx M (vx ) and
det DF (vx ) = det (DidM→Rn (y)) det (DHamx,δ (vx ))
= det(g(y))−1/2 det (DHamx,δ (vx )) .
12
3
Convergence of Riemannian Hamiltonian Monte Carlo
Hamiltonian Monte Carlo is a Markov chain on a manifold whose stationary stationary distribution
has density q(x) proportional to exp(−f (x)). We will bound the conductance of this Markov chain
and thereby its mixing time to converge to the stationary distribution. Bounding conductance
involves showing (a) the induced metric on the state space satisfies a strong isoperimetric inequality
and (b) two points that are close in metric distance are also close in probabilistic distance, i.e., the
one-step distributions from two nearby points have large overlap. In this section and the next, we
present general conductance bounds using parameters determined by the associated manifold. In
Section 7, we bound these parameters for the manifold corresponding to the logarithmic barrier in
a polytope.
3.1
Basics of geometric Markov chains
For completeness, we will discuss some standard techniques in geometric random walks in this subsection. For a Markov chain with state space M, stationary distribution q and next step distribution
pu (·) for any u ∈ M, the conductance of the Markov chain is
R
def
S pu (M \ S)dq(u)
φ = inf
.
S⊂M min {q(S), q(M \ S)}
The conductance of an ergodic Markov chain allows us to bound its mixing time, i.e., the rate of
convergence to its stationary distribution, e.g., via the following theorem of Lovász and Simonovits.
Theorem 11 ([19]). Let qt be the distribution of the current point after t steps of a Markov chain
with stationary distribution q and conductance at least φ, starting from initial distribution q0 . For
any ε > 0,
s
t
dq0 (x)
φ2
1
.
Ex∼q0
dT V (qt , q) ≤ ε +
1−
ε
dq(x)
2
Definition 12. The isoperimetry of a metric space M with target distribution q is
R
d(S,x)≤δ q(x)dx − q(S)
def
ψ = inf min
δ>0 S⊂M δ min {q(S), q(M \ S)}
where d is the shortest path distance in M.
The proof of the following theorem follows the standard outline for geometric random walks (see
e.g., [29]).
Lemma 13. Given a metric space M and a time-reversible Markov chain p on M with stationary
distribution q. Fix any r > 0. Suppose that for any x, y ∈ M with d(x, z) < r, we have that
dT V (px , py ) ≤ 0.9. Then, the conductance of the Markov chain is Ω(rψ).
Proof. Let S be any measurable subset of M. Then our goal is to bound the conductance of the
Markov chain
R
S px (M \ S) dq(x)
= Ω (rψ) .
min {q(S), q(M \ S)}
R
R
Since the Markov chain is time-reversible (For any two subsets A, B, A px (B) dq(x) = B px (A) dq(x)),
we can write the numerator of the left hand side above as
!
Z
Z
1
px (S) dq(x) .
px (M \ S) dq(x) +
2
M\S
S
13
Define
S1 = {x ∈ S : px (M \ S) < 0.05}
S2 = {x ∈ M \ S : px (S) < 0.05}
S3 = M \ S1 \ S2 .
Without
loss of generality, we can assume that q(S1 ) ≥ (1/2)q(S) and q(S2 ) ≥ (1/2)q(M \ S) (if
R
not, S px (M \ S) dq(x) = Ω(1) and hence the conductance is Ω(1).)
Next, we note that for any two points x ∈ S1 and y ∈ S2 , dT V (px , py ) > 0.9. Therefore, by the
assumption, we have that d(x, y) ≥ r. Therefore, by the definition of ψr , we have that
Z
q(x)dx − q(S1 )
q(S3 ) ≥
d(S1 ,x)≤r
≥ rψ min {q(S1 ), q(M \ S1 )}
≥ rψ min {q(S1 ), q(S2 )} .
Going back to the conductance,
1
2
Z
S
px (M \ S) dq(x) +
Z
!
px (S) dq(x)
M\S
1
≥
2
Z
(0.05)dq(x)
S3
= Ω (rψ) min{q(S1 ), q(S2 )}
= Ω (rψ) min{q(S), q(M \ S)}.
Therefore, the conductance of the Markov chain is Ω(rψ).
Combining Theorem 13 and Lemma 13 gives the following result for bounding mixing time of
general geometric random walk.
Lemma 14. Given a metric space M and a time-reversible Markov chain p on M with stationary
distribution q. Suppose that there exist r > 0 and ψ > 0 such that
1. For any x, y ∈ M with d(x, z) < r, we have that dT V (px , py ) ≤ 0.9.
2. For any S ⊂ M, we have that
Z
0<d(S,x)≤r
q(x)dx ≥ rψ min {q(S), q(M \ S)} .
Let qt be the distribution of the current point after t steps of a Markov chain with stationary distribution q starting from initial distribution q0 . For any ε > 0,
s
t
1
dq0 (x)
Ex∼q0
1 − Ω(r 2 ψ 2 ) .
dT V (qt , q) ≤ ε +
ε
dq(x)
14
3.2
Overlap of one-step distributions
The mixing of the walk depends on smoothness parameters of the manifold and the functions f, g
used to define the Hamiltonian. Since each step of our walk involves a Gaussian vector, many
smoothness parameters depend on choices of the random vector. Formally, let γ be the Hamiltonian curve used in a step of Hamiltonian Monte Carlo. In the analysis, we need a large fraction
of Hamiltonian curves from any point on the manifold to be well-behaved. A Hamiltonian curve
can be problematic when its velocity or length is too large and this happens with non-zero probability. Rather than using supremum bounds for our smoothness parameters, it suffices to use large
probability bounds, where the probability is over the random choice of Hamiltonian curve at any
point x ∈ Ω. To capture the notion that “most Hamiltonian curves are well-behaved”, we use an
auxiliary function ℓ(γ) ≥ 0 which assigns a real number to each Hamiltonian curve γ and measures
how “good” the curve is. The smoothness parameters assume that this function ℓ is bounded and
Lipshitz. One possible choice of such ℓ is ℓ(γ) = kγ ′ (0)kγ(0) which measures the initial velocity, but
this will give us a weaker bound. Instead, we use the following which jointly bounds the change in
position (first term) and change in velocity (second term).
Definition 15. An auxiliary function ℓ is a non-negative real-valued function on the set of Hamiltonian curves, i.e., maps γ : [0, δ] → M , with bounded parameters ℓ0 , ℓ1 such that
1. For any variation γs of a Hamiltonian curve (see Definition 18) with ℓ(γs ) ≤ ℓ0 , we have
!
d
d
′
ℓ(γs ) ≤ ℓ1
γs (0)
+ δ Ds γs (0) γs (0) .
ds
ds
γs (0)
1
min 1, ℓℓ10δ where γ ∼ x indicates a ran2. For any x ∈ M , Pγ∼x (ℓ(γ) ≤ 21 ℓ0 ) ≥ 1 − 100
dom Hamiltonian curve starting at x, chosen by picking a random Gaussian initial velocity
according to the local metric at x.
3.2.1
Proof Outline
To bound the conductance of HMC, we need to show that one-step distributions from nearby points
have large overlap for reasonably large step size δ. To this end, recall that the probability density
of going from x to y is given by the following formula
s
X
1
det (g(y))
2
−1
exp − kvx kx .
px (y) =
|det (DHamx,δ (vx ))|
(2π)n
2
vx :Hamx,δ (vx )=y
In Section 3.2.2, we introduce the concept of variations of Hamiltonian curves and use it to bound
|det (DHamx,δ (vx ))|−1 . We can show that px (y) is in fact close to
s
X
1
det (g(y))
1
2
·
exp − kvx kx .
(3.1)
pex (y) =
δn
(2π)n
2
vx :Hamx,δ (vx )=y
To compare px (y) with pz (y), we need to relate vx and vz that map x and z to y respectively. In
Section 3.2.3, we shows that if x and z are close enough, for every vx , there is a unique vz such that
vx is close to vz and that Hamz,δ (vz ) = Hamx,δ (vx ). Combining these facts, we obtain our main
theorem for this section, stated in Subsection 3.2.4.
15
In the analysis, we use three important operators from the tangent space to itself. The motivation for defining these operators comes directly from Lemma 19, which studies the variation in
Hamiltonian curves as the solution of a Jacobi equation. In words, the operator R(.) below allows
us to write the change in the Hamiltonian curve as an ODE.
Definition 16. Given a Hamiltonian curve γ, let R(γ, t), M (γ, t) and Φ(γ, t) be the operators from
T M to T M defined by
R(t)u = R(u, γ ′ (t))γ ′ (t),
M (t)u = Du µ(γ(t)),
Φ(t)u = M (t)u − R(t)u.
When γ is explicit from the context, we simply write them as R(t), M (t) and Φ(t).
The key parameter R1 we use in this section is a bound on the Frobenius norm of Φ formally
defined as follows.
Definition 17. Given a manifold M with metric g and an auxiliary function ℓ with parameters
ℓ0 , ℓ1 , we define the smoothness parameter R1 depending only on M and the step size δ such that
kΦ(γ, t)kF,γ(t) ≤ R1
for any γ such that ℓ(γ) ≤ ℓ0 and any 0 ≤ t ≤ δ where the Frobenius norm kAkF,γ(t) is defined by
kAk2F,γ(t) = Eα,β∼N (0,g(x)−1 ) (αT Aβ)2 .
The above definitions are related to but different from our previous paper analyzing the geodesic
walk [13].
3.2.2
Variation of Hamiltonian curve
To bound the determinant of the Jacobian of Hamx , we study variations of Hamiltonian curves.
Definition 18. We call γs (t) a Hamiltonian variation if γs (·) satisfies the Hamiltonian equation for
s
every s. We call ∂γ
∂s a Jacobi field.
The following lemma shows that a Jacobi field satisfies the following Jacobi equation.
Lemma 19. Given a path c(s), let γs (t) = Hamc(s) (t(v + sw)) be a Hamiltonian variation. The
def
Jacobi field ψ(t) =
∂
∂s γs (t)|s=0
satisfies the following Jacobi equation
Dt2 ψ(t) = Φ(t)ψ(t)
(3.2)
Let Γt parallel transport from Tγ(t,0) M to Tγ(0,0) M and ψ(t) = Γt ψ(t). Then, ψ(t) satisfies the
following ODE on the tangent space Tγ(0,0) M:
′′
ψ (t) = Γt Φ(t)Γ−1
t ψ(t)
′
ψ (0) = w,
ψ(0) = Ds c(0).
16
∀t ≥ 0,
(3.3)
Proof. Taking derivative Ds on both sides of Dt ∂γ
∂t = µ(γ), and using Fact 67, we get
∂γ
∂t
∂γ ∂γ ∂γ
∂γ
+ R( ,
)
= Dt Ds
∂t
∂s ∂t ∂t
∂γ
∂γ ∂γ ∂γ
= Dt2
+ R( ,
) .
∂s
∂s ∂t ∂t
Ds µ(γ) = Ds Dt
In short, we have Dt2 ψ(t) = Φ(t)ψ(t). This shows (3.2).
Equation (3.3) follows from the fact that
Dt v(t) = Γt
d
Γ−1
t v(t)
dt
′
for any vector field on γ0 (t) (see Definition 11 in the appendix) applied to v(t) = ψ (t).
We now proceed to estimate the determinant of the Jacobian of Hamx . For this we will use the
following elementary lemmas describing the solution of the following second-order matrix ODE:
d2
Ψ(t) = Φ(t)Ψ(t),
dt2
d
Ψ(0) = B,
dt
Ψ(0) = A.
(3.4)
Lemma 20. Consider the matrix ODE (3.4). Let λ = max0≤t≤ℓ kΦ(t)k2 . For any t ≥ 0, we have
that
√
√
kBk
kΨ(t)k2 ≤ kAk2 cosh( λt) + √ 2 sinh( λt).
λ
Lemma 21. Consider the matrix ODE (3.4). Let λ = max0≤t≤ℓ kΦ(t)kF . For any 0 ≤ t ≤
have that
t3
2
kΨ(t) − A − BtkF ≤ λ t kAk2 + kBk2 .
5
√1 ,
λ
we
In particular, this shows that
Ψ(t) = A + Bt +
with kE(s)kF ≤ λ s2 kAk2 +
s3
5
kBk2 .
Z
t
0
(t − s)Φ(s)(A + Bs + E(s))ds
The proofs of these lemmas are in Appendix A. We continue with the main proof here.
Lemma 22. Let γ(t) = Hamx (tvx ) be a Hamiltonian curve and step size δ satisfy 0 < δ2 ≤ R11
where R1 = max0≤t≤h kΦ(t)kF,γ(t) . Then DHamx,δ is invertible with kDHamx,δ − δIkF,γ(δ) ≤ 5δ .
Also, we have,
log det
2
Z δ
δ 2 R1
t(δ − t)
1
DHamx,δ (vx ) −
TrΦ(t)dt ≤
.
δ
δ
10
0
17
(3.5)
Proof. We want to compute DHamx,δ (vx )[w] for some w ∈ Tx M . By definition, we have that
DHamx,δ (vx )[w] =
∂
γ(t, s)|t=δ,s=0
∂s
(3.6)
where γ(t, s) = Hamx (t(vx + sw)). Define ψ(t) as in Lemma 19 with c(s) = x. So, Ds c(0) = 0, i.e.,
ψ(0) = 0. Then, by the lemma,
′′
ψ (t) = Γt Φ(t)Γ−1
t ψ(t),
′
ψ (0) = w,
ψ(0) = 0.
Now, we define Ψ be the solution of the matrix ODE
Ψ′′ (t) = Γt Φ(t)Γ−1
t Ψ(t) ∀t ≥ 0,
Ψ′ (0) = I,
Ψ(0) = 0.
By the definition of Ψ, we see that ψ(t) = Ψ(t)w. Therefore, we have that
∂
−1
γ(t, s)|s=0 = Γ−1
t ψ(t) = Γt Ψ(t)w.
∂s
Combining it with (3.6), we have that DHamx,δ (vx ) = Γ−1
t Ψ(δ). Since Γt is an orthonormal matrix,
we have that
log det (DHamx,δ (vx )) = log det Ψ(δ).
(3.7)
Note that Γt Φ(t)Γ−1
t
that
F,γ(t)
= kΦ(t)kF,γ(t) ≤ R1 for all 0 ≤ t ≤ δ. Using this, Lemma 21 shows
1
Ψ(δ) − I
δ
F,x
≤ R1
δ2
kIk2
5
≤
1
5
(3.8)
Hence, Ψ(δ) is invertible, and so is DHamx .
By Lemma 64, we have that
1
log det( Ψ(δ)) − Tr
δ
1
Ψ(δ) − I
δ
≤
1 2
δ R1
5
2
.
(3.9)
Now we need to estimate Tr(Ψ(δ) − δI). Lemma 21 shows that
Z δ
Z δ
(δ − t)Φ(t)E(t)dt
t(δ − t)Φ(t)dt +
Ψ(δ) = δI +
0
0
3
with kE(t)kF,γ(t) ≤ t 5R1 for all 0 ≤ t ≤ δ. Since
Z δ
Z δ
δ5
(δ − t) kΦ(t)kF,x kE(t)kF,x dt ≤ R12 ,
(δ − t)Φ(t)E(t)dt ≤
Tr
20
0
0
we have that
Tr
1
Ψ(δ) − I −
δ
Z
δ
0
t(δ − t)
Φ(t)dt
δ
≤
δ4 2
R .
20 1
Combining (3.9) and (3.10), we have
2
2
Z δ
δ 2 R1
t(δ − t)
1 2
δ4 2
1
TrΦ(t)dt ≤
δ R1 + R1 ≤
.
log det( Ψ(δ)) −
δ
δ
5
20
10
0
Applying (3.7), we have the result.
18
(3.10)
3.2.3
Local Uniqueness of Hamiltonian Curves
Next, we study the local uniqueness of Hamiltonian curves. We know that for every pair x, y, there
can be multiple Hamiltonian curves connecting x and y. Due to this, the probability density px
at y in M is the sum over all possible Hamiltonian curves connecting x and y. The next lemma
establishes a 1-1 map between Hamiltonian curves connecting x to y as we vary x.
Lemma 23. Let γ(t) = Hamx (tvx ) be a Hamiltonian curve and let the step size δ satisfy 0 < δ2 ≤
1
R1 , where R1 = max0≤t≤δ kΦ(t)kF,γ(t) . Let the end points be x = γ(0) and y = γ(δ). Then there is
an unique smooth invertible function v : U ⊂ M → V ⊂ T M such that
y = Hamz,δ (v(z))
for any z ∈ U where U is a neighborhood of x and V is a neighborhood of vx = v(x). Furthermore,
5
we have that k∇η v(x)kx ≤ 2δ
kηkx and
1
η + ∇η v(x)
δ
x
≤
3
R1 δ kηkx .
2
Let γs (t) = Hamc(s) (t · v(c(s))) where c(s) is any path with c(0) = x and c′ (0) = η. Then, for
all 0 ≤ t ≤ δ, we have that
∂
∂s
γs (t)
s=0
γ(t)
≤ 5 kηkx
and
Ds γs′ (t)
s=0 γ(t)
≤
10
kηkx .
δ
Proof. Consider the smooth function f (z, w) = Hamz,δ (w). From Lemma 22, the Jacobian of w
at (x, vx ) in the w variables, i.e., DHamx,δ (vx ), is invertible. Hence, the implicit function theorem
shows that there is a open neighborhood U of x and a unique function v on U such that f (z, v(z)) =
f (x, vx ), i.e. Hamz,δ (v(z)) = Hamx,δ (vx ) = y.
To bound ∇η v(x), we let γs (t) = Hamc(s) (t · v(c(s))) and c(s) be any path with c(0) = x and
′
c (0) = η. Let Γt be the parallel transport from Tγ(t) M to Tγ(0) M. Define
ψ(t) = Γt
∂
∂s
γs (t).
s=0
Lemma 19 shows that ψ(t) satisfies the following ODE
′′
ψ (t) = Γt Φ(t)Γ−1
t ψ(t) ∀t ≥ 0,
′
ψ (0) = ∇η v(x),
ψ(0) = η.
Moreover, we know that ψ(δ) = 0 because γs (δ) = Hamc(s) (δ · v(c(s))) = Hamc(s),δ (v(c(s))) = y for
small enough s.
To bound k∇η v(x)kx , we note that Lemma 21 shows that
t
2
(3.11)
ψ(t) − η − t · ∇η v(x) x ≤ R1 t kηkx + k∇η v(x)kx .
5
Since ψ(δ) = 0 and δ2 ≤
1
R1 ,
we have that
kη + δ · ∇η v(x)kx ≤ kηkx +
19
δ
k∇η v(x)kx
5
which implies that
δ k∇η v(x)kx − kηkx ≤ kηkx +
Therefore, k∇η v(x)kx ≤
5
2δ
δ
k∇η v(x)kx .
5
kηkx . More precisely, from (3.11), we have that
1
η + ∇η v(x)
δ
x
≤
3
R1 δ kηkx .
2
Putting this into (3.11), for t ≤ δ we get
ψ(t)
x
5
1
2
≤ kηkx + kηkx + R1 t kηkx + kηkx ≤ 5 kηkx .
2
2
Now, apply the conclusion of Lemma 21 after taking a derivative, we have that
Z t
′
Φ(s)(η + s · ∇η v(x) + E(s))ds
ψ (t) = ∇η v(x) +
0
where kE(s)kx ≤ R1 s2 kηkx +
that t ≤ δ, we have
′
ψ (t)
x
s3
5
k∇η v(x)kx
≤
3
2
kηkx . Hence, bounding each term and noting
5
10
5 3
≤
kηkx + δR1 1 + +
kηkx .
kηkx ≤
2δ
2 2
δ
When we vary x, the Hamiltonian curve γ from x to y varies and we need to bound ℓ(γ) over
the variation.
Lemma 24. Given a Hamiltonian curve γ(t) = Hamx (t · vx ) with step size δ satisfying δ2 ≤ R11 , let
c(s) be any geodesic starting at γ(0). Let x = c(0) = γ(0) and y = γ(δ). Suppose that the auxiliary
ℓ0
dc
function ℓ satisfies ds
≤ 7ℓ
and ℓ(γ) ≤ 21 ℓ0 . Then, there is a unique vector field v on c such
c(0)
1
that
y = Hamc(s),δ (v(s)).
Moreover, this vector field is uniquely determined by the geodesic c(s) and any v(s) on this vector
field. Also, we have that ℓ(Hamc(s),δ (v(s))) ≤ ℓ0 for all s ≤ 1.
Proof. Let smax be the supremum of s such that v(s) can be defined continuously such that y =
Hamc(s),δ (v(s)) and ℓ(γs ) ≤ ℓ0 where γs (t) = Hamc(s) (t · v(s)). Lemma 23 shows that there is a
neighborhood N at x and a vector field u on N such that for any z ∈ N , we have that
y = Hamz,δ (u(z)).
Also, this lemma shows that u(s) is smooth and hence the parameter ℓ1 shows that ℓ(γs ) is Lipschitz
in s. Therefore, ℓ(γs ) ≤ ℓ0 in a small neighborhood of 0. Hence smax > 0.
Now, we show smax > 1 by contradiction. By the definition of smax , we have that ℓ(γs ) ≤ ℓ0 for
dc
5
5
any 0 ≤ s < smax . Hence, we can apply Lemma 23 to show that kDs v(s)kγ(s) ≤ 2δ
ds γ(s) = 2δ L
where L is the length of c up to s = 1 (since the speed is constant on any geodesic and the curve
is defined over [0, 1]). Therefore, the function v is Lipschitz and hence v(smax ) is well-defined and
ℓ(γsmax ) ≤ ℓ0 by continuity. Hence, we can apply Lemma 23 at ℓ(smax ) and extend the domain of
ℓ(s) beyond smax .
20
To bound ℓ(γs ) beyond smax , we note that kDs γs′ kγ(s) = kDs v(s)kγ(s) ≤
d
ds c γ(0)
= L. Hence,
d
ds ℓ(γs )
5
2δ L
and
≤ (L + 25 L)ℓ1 by the definition of ℓ1 . Therefore, if L ≤
that ℓ(γs ) ≤ ℓ(γ) + 21 ℓ0 ≤ ℓ0 for all s ≤ 1.01 wherever v(s) is
the assumption that smax is the supremum. Hence, smax > 1.
d
ds c γ(s)
ℓ0
7ℓ1 ,
=
we have
defined. Therefore, this contradicts
The uniqueness follows from Lemma 23.
3.2.4
Smoothness of one-step distributions
Lemma 25. For δ2 ≤
px , pz from x, z satisfy
1
√
100 nR1
and δ3 ≤
√ℓ0
,
100 nR1 ℓ1
the one-step Hamiltonian walk distributions
1
1
dTV (px , pz ) = O
d(x, z) + .
δ
25
ℓ0
. Let c(s) be a unit speed geodesic connecting x and
Proof. We first consider the case d(x, y) < 7ℓ
1
ℓ0
z of length L < 7ℓ1 .
def
Let ℓe = min 1, ℓℓ10δ . By the definition of ℓ0 , with probability at least 1 − ℓe in paths γ start at
x, we have that ℓ(γ) ≤ 21 ℓ0 . Let Vx be the set of vx such that ℓ(Hamx (t · vx )) ≤ 21 ℓ0 . Since the
ℓ0
and ℓ(γ) ≤ 12 ℓ0 , for those vx , Lemma 24 shows there is a family
distance from x to z is less than 7ℓ
1
of Hamiltonian curves γs that connect c(s) to y, and ℓ(γs ) ≤ ℓ0 for each of them.
For any vx ∈ V , we have that ℓ(γs ) ≤ ℓ0 . When ℓ(γs ) ≤ ℓ0 , by the definition of R1 , we indeed
have that kΦ(t)k ≤ R1 and hence Lemma 22 shows that
2
Z δ
δ 2 R1
t(δ − t)
1
DHamx,δ (vx ) ≤
TrΦ(t)dt +
log det
δ
δ
10
0
2
2
2
δ R1
1 e
δ √
nR1 +
≤
ℓ
≤
6
10
600
where we used our assumption on δ. We use p(vx ) to denote the probability density of choosing vx
and
s
det (g(Hamx (δ · vx )))
1
def
2
exp − kvx kx .
pe(vx ) =
(2πδ2 )n
2
Hence, we have that
C −1 · p(vx ) ≤ pe(vx ) ≤ C · p(vx )
(3.12)
1 e
ℓ. As we noted, for every vx , there is a corresponding vz such that Hamz (δ·vz ) = y
where C = 1+ 600
and ℓ(Hamz (t · vz )) ≤ ℓ0 . Therefore, we have that
def
(1 − C 2 )p(vx ) + C(e
p(vx ) − pe(vz )) ≤ p(vx ) − p(vz ) ≤(1 − C −2 )p(vx ) + C −1 (e
p(vx ) − pe(vz )).
R
1
Since Vx p(vx )dvx ≥ 1 − 100
min 1, ℓℓ10δ , we have that
Z
ℓe
|p(vx ) − p(vz )| dvx
+
100
Vx
Z
Z
ℓe
≤
|e
p(vx ) − pe(vz )| dvx
p(vx )dvx ) + 2
(1 +
100
Vx
Vx
Z Z
d
ℓe
+2
pe(vc(s) ) dsdvx
≤
50
Vx s ds
dT V (px , pz ) ≤
21
(3.13)
Note that
d
pe(v ) =
ds c(s)
−
1 d
kv(s)k2c(s) pe(vc(s) ).
2 ds
Using (3.12), we have that pe(vc(s) ) ≤ 2 · p(vc(s) ) and hence
Z
Vx
d
pe(v ) dy ≤
ds c(s)
Z
d
kv(s)k2c(s) p(vc(s) )dy
Vx ds
d
kv(s)k2c(s)
≤Eℓ(γs )≤ℓ0
ds
(3.14)
Using that k∂s c(s)k = 1, Lemma 23 shows that
1
∂s c(s) + Ds v(s)
δ
x
3
≤ R1 δ.
2
Therefore, we have that
d
kv(s)k2c(s) = 2 hv(s), Ds v(s)ic(s)
ds
2
hv(s), ∂s c(s)ic(s) + 3R1 δ kv(s)kc(s) .
≤
δ
Since v(s) is a random Gaussian vector from the local metric, we have that hv(s), ∂s c(s)ic(s) = O(1)
√
and kv(s)kc(s) = O( n) with high probability. Putting it into (3.14), we have that
Eℓ(γs )≤ℓ0
d
kv(s)k2c(s) = O
ds
√
1
+ δR1 n .
δ
Putting this into 3.14, we have that
Z
√
d
1
pe(vc(s) ) dy = O
+ δR1 n .
δ
Vx ds
Putting this into (3.13), we get
dTV (px , pz ) = O
√
ℓe
1
+ δR1 n L +
δ
50
ℓ0
for any L < 7ℓ
. By taking a minimal length geodesic, and summing over segment of length
1
any x and z, we have
√
1
1
+ δR1 n d(x, z) +
dTV (px , pz ) = O
δ
25
1
1
=O
d(x, z) + .
δ
25
22
ℓ0
8ℓ1 ,
for
3.3
Convergence bound
Combining Lemma 25 and Lemma 14, we have the following result.
Theorem 26. Given a manifold M. Let ℓ0 , ℓ1 , R1 be the parameters of the Hamiltonian Monte
Carlo defined in Definition 15 and 17. Let qt be the distribution of the current point after t steps
Hamiltonian Monte Carlo with step size δ satisfying
δ2 ≤
1
√
100 nR1
δ3 ≤
and
ℓ
√0
,
100 nR1 ℓ1
starting from initial distribution q0 . Let q be the distribution proportional to e−f . For any ε > 0,
we have that
s
t
dq0 (x)
(δψ)2
1
Ex∼q0
1−
dT V (qt , q) ≤ ε +
ε
dq(x)
2
where ψ is the conductance of the manifold defined in Definition 12.
For ℓ(γ) = kγ ′ (0)kγ(0) , we can bound ℓ0 and ℓ1 as follows:
√
Lemma 27. For the auxiliary function ℓ(γ) = kγ ′ (0)kγ(0) , we have that ℓ0 = 10 n and ℓ1 = O( 1δ ).
Furthermore, we have that
γ ′ (t) γ(t) ≤ γ ′ (0) γ(0) + R0 t
where R0 = supx∈M kµ(x)kx .
√
Proof. For ℓ0 , we note that γ ′ (0) ∼ N (0, g(x)−1 ) and hence kγ ′ (0)kγ(0) ≤ 5 n with probability
√
e−O( n) .
For ℓ1 , we note that
d
d
γ ′ (0)
ℓ(γs ) =
ds
ds s
=
γs (0)
≤
1
2
hDs γs′ (0), γs′ (0)iγs (0)
d
ds
kγs′ (0)k2γs (0)
kγs′ (0)kγs (0)
kγs′ (0)kγs (0)
≤ Ds γs′ (0)
γs (0)
.
Hence, we have that ℓ1 = 1δ .
Next, we note that
d
γ ′ (t)
dt
2
γ(t)
= 2 Dt γ ′ (t), γ ′ (t)
γ(t)
≤ 2 kµ(γ(t))kγ(t) γ ′ (t)
γ(t)
.
Therefore, we get the last result.
Combining Lemma 27 and Theorem 26, we have the following result. This result might be
more convenient to establish an upper bound on the rate of convergence, as it depends on only two
worst-case smoothness parameters. In the next section, we will see a more refined bound that uses
the randomness of Hamiltonian curves via additional parameters.
23
Corollary 28. Given a manifold M. Let qt be the distribution of the current point after t steps
Hamiltonian Monte Carlo with step size δ and q be the distribution proportional to e−f . Let R0 and
R1 be parameters such that
1. kµ(x)kx ≤ R0 for any x ∈ M where µ is defined in Lemma 7.
2. Eα,β∼N (0,g(x)−1 ) hDα µ(x), βi2x ≤ R12 for any x ∈ M.
3. Eα,β∼N (0,g(x)−1 ) hR(α, v)v, βi 2x ≤ R12 for any x ∈ M and any kvkx ≤
Riemann curvature tensor of M.
Suppose that δ ≤
√
n
R0
and δ2 ≤
1
√
,
100 nR1
√
n where R is the
then for any ε > 0, we have that
dT V (qt , q) ≤ ε +
s
1
dq0 (x)
Ex∼q0
ε
dq(x)
1−
(δψ)2
2
t
where ψ is the conductance of the manifold defined in Definition 12. In short, the mixing time of
Hamiltonian Monte Carlo is
2
√
R
−2
0
e ψ
.
nR1 +
O
n
Proof. The statement is basically restating the√ definition of R1 and R0 used in Lemma 27 and
Theorem 26. The only difference is that if δ ≤ Rn0 , then we know that
γ ′ (t)
γ(t)
≤ γ ′ (0)
γ(0)
√
+ R0 t = O( n)
for all 0 ≤ t ≤ δ. Therefore, we can relax the constraints ℓ(γ) ≤ ℓ0 in the definition of R1 to simply
√
kvkx ≤ n. It allows us to use R1 without mentioning the auxiliary function ℓ.
4
Improved analysis of the convergence
Corollary 28 gives a polynomial mixing time for the log barrier function. There are two bottlenecks
to improving the bound. First, the auxiliary function ℓ it used does not capture the fact each curve
γ in Hamiltonian Monte Carlo follows a random initial direction. Second, Lemma 23 also does not
take full advantage of the random initial direction. In this section, we focus on improving Lemma
23.
Our main theorem for convergence can be stated as follows in terms of ψ and additional parameters ℓ0 , ℓ1 , R1 , R2 , R3 (see Definitions 12, 15, 16, 17, 33 and 31). It uses the following key
lemma.
Theorem 29. For δ2 ≤
x, z satisfy
1
R1
and δ5 ≤
ℓ0
,
R21 ℓ1
the one-step Hamiltonian walk distributions px , pz from
1
1
dTV (px , pz ) = O δ R2 + + δR3 d(x, z) + .
δ
25
2
Remark. The constant term at the end is an artifact that comes from bounding the probability of
some bad events of the Hamiltonian walk, and can be made to arbitrary small.
24
We now prove this theorem. In a later section, we specialize to sampling distributions over
polytopes using the logarithmic barrier, by defining a suitable auxiliary function and bounding all
the parameters. The key ingredient is Theorem 29 about the overlap of one-step distributions, which
we prove in the next section.
Here is the consequence of Theorem 29 and Theorem 13.
Theorem 30. Given a manifold M. Let ℓ0 , ℓ1 , R1 , R2 , R3 be the parameters of the Hamiltonian
Monte Carlo defined in Definition 15, 17, 33 and 31. Let qt be the distribution of the current point
after t steps Hamiltonian Monte Carlo starting from initial distribution q0 . Let q be the distribution
proportional to e−f . Suppose that the step size δ satisfies
δ2 ≤
1
ℓ0
, δ5 ≤ 2 and δ3 R2 + δ2 R3 ≤ 1.
R1
R1 ℓ 1
For any ε > 0, we have that
s
dT V (qt , q) ≤ ε +
dq0 (x)
1
Ex∼q0
ε
dq(x)
(δψ)2
1−
2
t
where ψ is the conductance of the manifold.
4.1
Improved one-to-one correspondence for Hamiltonian curve
In the previous section, we only used R1 to analyze how much a Hamiltonian curves change as one
end point varies. Here we derive a more refined analysis of Lemma 23 using an additional parameter
R3 .
Definition 31. For a manifold M and auxiliary function ℓ, R3 is a constant such that for any
Hamiltonian curve γ(t) of step size δ with ℓ(γ0 ) ≤ ℓ0 , if ζ(t) is the parallel transport of the vector
γ ′ (0) along γ(t), then we have
sup kΦ(t)ζ(t)kγ(t) ≤ R3 .
0≤t≤δ
Lemma 32. Under the same assumptions as Lemma 23, we have that
δ
∇η kv(x)k2x ≤ |hvx , ηix | + 3δ2 R3 kηkx .
2
Proof. Let χ = ∇η v(x). Define ψ(t) as in the proof of Lemma 23. Using Lemma 21 we get that
0 = ψ(δ) = η + δχ +
Z
0
i.e.,
−δχ = η +
with
Z
δ
0
kE(s)kx ≤ R1
δ
(δ − s)Γs Φ(s)Γ−1
s (η + sχ + e(s))ds
(δ − s)Γs Φ(s)Γ−1
s (η + sχ + E(s))ds
δ3
kχkx
δ kηkx +
5
2
for all 0 ≤ s ≤ δ where we used k∇η v(x)kx ≤
5
2δ
≤ 2R1 δ2 kηkx
kηkx at the end (by Lemma 23).
25
Therefore,
δ |hv(x), χix | ≤ |hv(x), ηix | +
δ2
sup hv(x), Γs Φ(s)Γ−1
s (η + sχ + E(s))ix .
2 0≤s≤δ
Noting that
kη + sχ + E(s)kx ≤ kηkx + δ kχkx + kE(s)kx ≤ (1 +
5
+ 2R1 δ2 ) kηkx ≤ 6 kηkx ,
2
δ |hv(x), χix | ≤ |hv(x), ηix | + 3δ2 sup
x
we have
0≤s≤δ
= |hv(x), ηix | + 3δ2 sup
0≤s≤δ
γ ′ (0)T Γs Φ(s)Γ−1
s
Φ(s)Γs γ ′ (0)
γ(s)
kηkx
kηkx
≤ |hv(x), ηix | + 3δ2 R3 kηkx .
Finally, we note that
4.2
δ
∇η kv(x)k2x = δ hv(x), ∇η v(x)ix .
2
Improved smoothness of one-step distributions
The proof of Theorem 29 is pretty similar to Lemma 25. First, we show that px (y) is in fact close
to
s
Z δ
X
t(δ − t)
1
det (g(y))
2
exp −
TrΦ(t)dt − kvx kx .
(4.1)
pex (y) =
(2πδ2 )n
δ
2
0
vx :Hamx,δ (vx )=y
Note that this is a more refined estimate than (3.1). We use Lemma 32 to bound the change of
kvx k2x . For the change of TrΦ(t), we defer the calculation until the end of this section.
Theorem 29. For δ2 ≤
x, z satisfy
1
R1
and δ5 ≤
ℓ0
,
R21 ℓ1
the one-step Hamiltonian walk distributions px , pz from
1
1
dTV (px , pz ) = O δ R2 + + δR3 d(x, z) + .
δ
25
Proof. We first consider the case d(x, y) <
δ2 ≤ R11 and δ5 ≤ Rℓ20ℓ , we have that
2
ℓ0
7ℓ1 .
By a similar argument as Lemma 25, using that
1 1
Z Z
d
ℓe
+2
pe(vc(s) ) dsdvx
dT V (px , pz ) ≤
50
Vx s ds
def
where ℓe = min 1, ℓℓ10δ . By direct calculation, we have
d
pe(v ) =
ds c(s)
Z
−
δ
0
t(δ − t) d
1 d
2
′
TrΦ(γs (t))dt −
kv(s)kc(s) pe(vc(s) ).
δ
ds
2 ds
By similar argument as (3.12), we have that pe(vc(s) ) ≤ 2 · p(vc(s) ) and hence
Z δ
t(δ − t) d
1 d
d
2
′
pe(v ) ≤ 2
TrΦ(γs (t))dt +
kv(s)kc(s) p(vc(s) ).
ds c(s)
δ
ds
2 ds
0
26
(4.2)
Since ℓ(γs ) ≤ ℓ0 , we can use Lemma 34 to get
Z
δ
0
Hence,
t(δ − t) d
TrΦ(γs′ (t))dt ≤ O δ2 R2 .
δ
ds
Z
Z
Z
d
2
pe(vc(s) ) dvx ≤O(δ R2 )
p(vc(s) )dvx +
Vx ds
Vx
Vx
R
For the first term, we note that Vx p(vc(s) )dvx ≤ 1.
For the second term, we have that
Z
d
kv(s)k2c(s) p(vc(s) )dvx ≤Eℓ(γs )≤ℓ0
Vx ds
d
kv(s)k2c(s) p(vc(s) )dvx .
ds
(4.3)
d
kv(s)k2c(s) .
ds
(4.4)
By Lemma 32, we have that
δ d
kv(s)k2c(s) ≤ |hv(s), ∂s cix | + 3δ2 R3 k∂s ckx .
2 ds
Since k∂s ckx = 1 and |hv(s), ∂s cix | = O(1) with high probability since v(s) is a Gaussian vector
from the local metric. We have
δ d
kv(s)k2c(s) ≤ O(1) + 3δ2 R3 .
2 ds
Putting it into (4.4) and (4.3), we have that
Z
d
1
2
pe(vc(s) ) dvx = O δ R2 + + δR3 .
δ
Vx ds
Putting this into 4.2, we get
1
ℓe
dTV (px , pz ) = O δ R2 + + δR3 L +
δ
50
2
ℓ0
. By taking a minimal length geodesic, and summing over segment of length
for any L < 7ℓ
1
any x and z, we have
1
1
2
dTV (px , pz ) = O δ R2 + + δR3 d(x, z) + .
δ
25
ℓ0
8ℓ1 ,
for
Definition 33. Given a Hamiltonian curve γ(t) with ℓ(γ0 ) ≤ ℓ0 . Let R2 be a constant depending
on the manifold M and the step size δ such that for any 0 ≤ t ≤ δ, any curve c(s) starting from
γ(t) and any vector field v(s) on c(s) with v(0) = γ ′ (t), we have that
!
dc
d
+ δ k Ds v|s=0 kγ(t) R2 .
TrΦ(v(s))|s=0 ≤
ds
ds s=0 γ(t)
27
Lemma 34. For δ2 ≤
1
R1
and ℓ(γs ) ≤ ℓ0 , we have
Z
δ
0
t(δ − t) d
TrΦ(γs′ (t))dt ≤ O δ2 R2
δ
ds
where γs is a family of Hamiltonian curve that connect c(s) to y defined in Lemma 24.
Proof. By Definition 33, we have that
d
TrΦ(γs′ (t)) ≤
ds
∂
∂s
γs (t)
s=0
+δ
γ(t)
Ds γs′ (t) s=0 γ(t)
!
· R2 .
(4.5)
d
d
d
γs (0) = ds
Hamc(s) (0) = ds
c(s) is a unit vector and hence
By definition of γs , we have that ds
d
ds γs (0) γ(0) = 1. Since Hamc(s),δ (v(s)) = y, Lemma 23 shows that
∂
∂s
γs (t)
s=0
γ(t)
≤ 5 and
Ds γs′ (t)
Therefore, we have that
d
TrΦ(γs′ (t)) ≤ 15R2 .
ds
28
s=0 γ(t)
≤
10
.
δ
5
5.1
Gibbs sampling on manifolds
Isoperimetry for Hessian manifolds
Here we derive a general isoperimetry bound, assuming that the manifold is defined by the Hessian
of a convex function, and that the directional fourth directive is non-negative, a property satisfied,
e.g., by the standard logarithmic barrier.
Lemma 35. Let φ : [a, b] → R be a convex function such that φ′′ is also convex. For any x ∈ [a, b],
we have that
Z b
Z x
p
−φ(t)
−φ(t)
−φ(x)
′′
e
dt .
e
dt,
e
≥ 0.372 φ (x) min
a
x
Let f (x) be the logconcave density proportional to
VarX∼f (x) X ≤
e−φ(x) .
Then, we have that
O(1)
.
minx∈[a,b] φ′′ (x)
p
Proof. Case 1)p|φ′ (x)| ≥ a φ′′ (x) (we will pick a at the end). Without loss of generality, we have
that φ′ (x) ≥ a φ′′ (x). Since φ is convex, we have that
φ(t) ≥ φ(x) + φ′ (x)(t − x).
Therefore, we have that
Z
Case 2) |φ′ (x)| ≤ a
we have that
p
b
x
e−φ(t) dt ≤
Z
∞
′
e−φ(x)−φ (x)(t−x) dt
x
e−φ(x)
e−φ(x)
= ′
≤ p
.
φ (x)
a φ′′ (x)
φ′′ (x). Without loss of generality, we have that φ′′′ (x) ≥ 0. Since φ′′ is convex,
φ′′ (t) ≥ φ′′ (x) + φ′′′ (x)(t − x) ≥ φ′′ (x)
for t ≥ x. Hence, we have that
′
φ(t) = φ(x) + φ (x)(t − x) +
Z
t
x
(t − s)φ′′ (t)ds
1
≥ φ(x) + φ (x)(t − x) + φ′′ (x)(t − x)2
2
′
for all t ≥ x. Therefore, we have that
Z
Z b
−φ(t)
e
dt ≤
x
∞
′
1
′′
2
e−φ(x)−φ (x)(t−x)− 2 φ (x)(t−x) ds
−∞
Z ∞
1 ′′
2
′
= e−φ(x)
e−φ (x)s− 2 φ (x)s ds
−∞
s
2
2π −φ(x)+ 21 φφ′′′(x)
(x)
=
e
φ′′ (x)
s
2π −φ(x)+ a2
2 .
e
≤
′′
φ (x)
29
Hence, we have
Z
−φ(x)
√
a2 e
.
2πe 2 p
φ′′ (x)
x
2
− a2
e√
Combining both cases, the isoperimetric ratio is min a, 2π . Setting a to be the solution of
p
a2
ae 2 = √12π , this minimum is achieved at W (1/2π) > 0.372 where W is the inverse Lambert
b
e−φ(t) dt ≤
1
. This proves the first result.
function, i.e., W (x)eW (x) = 2π
The variance of f follows from the fact that f is logconcave.
This generalizes to higher dimension with no dependence on the dimension using localization,
which we review next. Define an exponential needle E = (a, b, γ) as a segment [a, b] ⊆ Rn and γ ∈ R
corresponding to the weight function eγt applied the segment [a, b]. The integral of an n-dimensional
function h : Rn → R over this one dimensional needle is
Z |b−a|
Z
b−a
.
h(a + tu)eγt dt
where u =
h=
|b − a|
0
E
Theorem 36 (Theorem 2.7 in [9]). Let f1 , f2 , f3 , f3 be four nonnegative continuous functions defined
on Rn , and c1 , c2 > 0. Then, the following are equivalent:
1. For every logconcave function F defined on Rn with compact support,
c 1 Z
c 2 Z
c 1 Z
Z
F (x)f3 (x) dt
≤
F (x)f2 (x) dx
F (x)f1 (x) dx
Rn
Rn
Rn
Rn
F (x)f4 (x) dx
c2
2. For every exponential needle E,
c 2
c 1 Z
c 2 Z
c 1 Z
Z
f4
f3
≤
f2
f1
E
E
E
E
Lemma 37. Let φ : K ⊂ Rn → R be a convex function defined over a convex body K such
that D 4 φ(x)[h, h, h, h] ≥ 0 for all x ∈ K and h ∈ Rn . Given any partition S1 , S2 , S3 of K with
d = minx∈S1 ,y∈S2 d(x, y), i.e., the minimum distance between S1 and S2 in the Riemannian metric
induced by φ. For any α > 0, we have that
R
−αφ(x) dx
√
S3 e
o = Ω( α · d)
nR
R
min S1 e−αφ(x) dx, S2 e−αφ(x) dx
Proof. By rescaling φ, we can assume α = 1. We write the desired inequality as follows, for a
constant C, with χS being the indicator of set S:
Z
Z
Z
Z
−φ(x)
−φ(x)
−φ(x)
e−φ(x) χS3 (x) dx.
e
dx ·
e
χS2 (x) dx ≤
e
χS1 (x) dx ·
Cd
Rn
Rn
Rn
Rn
Using the localization lemma for exponential needles (Theorem 36), with fi (x) being Cde−φ(x) χS1 (x),
e−φ(x) χS2 (x), e−φ(x) and e−φ(x) χS3 (x) with respectively, it suffices to prove the following onedimensional inequality for functions φ defined on an interval and shifted by a linear term:
Z 1
Z 1
e−φ((1−t)a+tb) e−ct χS2 ((1 − t)a + tb) dt
e−φ((1−t)a+tb) e−ct χS1 ((1 − t)a + tb) dt ·
Cd
0
0
Z 1
Z 1
e−φ((1−t)a+tb) e−ct χS3 ((1 − t)a + tb) dt.
e−φ((1−t)a+tb) e−ct dt ·
≤
0
0
30
Each Ti = {t : (1 − t)a + tb ∈ Si } is a union of intervals. By a standard argument (see [19]),
it suffices to consider the case when each Si is a single interval and add up over all intervals in
S3 . Thus it suffices to prove the statement in one dimension for all convex φ with convex φ′′ . In
one-dimension, we have
Z yp
d(x, y) =
φ′′ (t) dt.
x
Taking T3 =
is
[a′ , b′ ]
⊂ [a, b], the inequality we need to prove is that for any convex φ with convex φ′′
R a′ −φ(t) R b −φ(t)
−φ(t) dt
e
dt b′ e
a′ e
≥ Ω(1) a
Rb
R b′ p
−φ(t) dt
φ′′ (t) dt
a e
a′
R b′ −φ(t)
e
dt
e−φ(x)
≥ minx∈[a′ ,b′ ] √
and
that R ba′′ √ ′′
φ′′ (x)
φ
(t)
dt
a′
R b′
which is implied by noting
5.2
dt
applying Lemma 35.
Sampling with the log barrier
For any polytope M = {Ax > b}, the logarithmic barrier function φ(x) is defined as
φ(x) = −
m
X
i=1
log(aTi x − bi ).
Theorem 1. Let φ be the logarithmic barrier for a polytope M with m constraints and n variables.
Hamiltonian Monte Carlo applied to the function f = exp(−αφ(x)) and the metric given by ∇2 φ
with appropriate step size mixes in
!
1
2
1
3
3n3
1
1
n
m
e
+ m2 n6
O
+
α + m−1 α 31 + m− 13
steps where each step is the solution of a Hamiltonian ODE.
√
Proof. Lemma 37 shows that the isoperimetric coefficient ψ is Ω( α). Also, we know that isoperi1
metric coefficient ψ is at worst Ω(m− 2 ) [16]. Lemma 61 shows that the condition of Theorem 30 is
satisfied, thus implying the bound
√
e max( α, m− 12 )−2 δ−2
O
√
e max( α, m− 12 )−2 max n 32 , α 23 m 13 n 31 , αm 12 n 61
=O
!
2
1
1
1
1
2
3 + α 3 m 3 n 3 + αm 2 n 6
n
e
=O
α + m−1
!
1
2
1
3
3n3
1
1
n
m
e
+ m2 n6
=O
+
α + m−1 α 13 + m− 13
To implement the walk, we solve this ODE using the collocation method as described in [13].
The key lemma used in that paper is that the slack of the geodesic does not change by more than
1
). Similarly in Lemma 53, we proved
a constant multiplicative factor up to the step size δ = O( n1/4
the the Hamiltonian flow does not change by more than a constant multiplicative factor up to
−1/3
n−1/4
√ ). Since the step size we use is Θ( n √ ), we can apply the
the step size δ = O( 11/4 ) = O( 1+
α
1+ α
M1
31
collocation method as described in [13] and obtain an algorithm to compute the Hamiltonian flow
ω−1 logO(1) 1 ) time with additive error η distance in the local norm. Due to the exponential
e
in O(mn
η
convergence, it suffices for the sampling purpose with only a polylogarithmic overhead in the total
running time.
6
Polytope volume computation: Gaussian cooling on manifolds
The volume algorithm is essentially the Gaussian cooling algorithm introduced in [3]. Here we apply
it to a sequence of Gibbs distributions rather than a sequence of Gaussians. More precisely, for a
convex body K and a convex barrier function φ : K → R, we define
(
exp −σ −2 φ(x)
if x ∈ K
2
f (σ , x) =
0
otherwise
and
2
F (σ ) =
x∗
Z
f (σ 2 , x) dx
Rn
where
is the minimizer of φ (the center of K). Let µi be the probability distribution proportional
to f (σi2 , x) where σi is the tempature of the Gibbs distribution to be fixed. The algorithm estimates
each ratio in the following telescoping product:
e−σ
−2 φ(x∗ )
vol(K) ≈ F (σk ) = F (σ0 )
k
2 )
Y
F (σi+1
i=1
F (σi2 )
for some large enough σk .
2 , x)/f (σ 2 , x). Then,
Let x be a random sample point from µi and let Yx = f (σi+1
i
Ex∼µi (Yx ) =
6.1
Algorithm: cooling schedule
6.2
Correctness of the algorithm
2 )
F (σi+1
.
F (σi2 )
In this subsection, we prove the correctness of the algorithm. We separate the proof into two parts.
In the first part, we estimate how small σ0 we start with should be and how large σk we end with
should be. In the second part, we estimate the variance of the estimator Yx .
6.2.1
Initial and terminal conditions
First, we need the following lemmas about self-concordance functions and logconcave functions:
Lemma 38 ([27, Thm 2.1.1]). Let φ be a self-concordant function and x∗ be its minimizer. For any
x such that φ(x) ≤ φ(x∗ ) + r 2 with r ≤ 12 , we have that
φ(x) = φ(x∗ ) +
1 ± Θ(r)
(x − x∗ )T ∇2 φ(x∗ )(x − x∗ ).
2
32
Algorithm 2: Volume(M , ε)
Let σ02 = Θ(ε2 n−3 log−3 (n/ε)),
2
σi+1
σ 2 1 + √1
if ϑ ≤ nσi2
i
n
=
.
σ 2 1 + min( √σi , 1 )
otherwise
i
2
ϑ
and
ki =
( √
Θ( ε2n log nε )
Θ((
√
ϑ
σi
if ϑ ≤ nσi2
+ 1)ǫ−2 log nε ) otherwise
.
Set i = 0. Compute x∗ = arg min φ(x).
Assume that φ(x∗ ) = 0 by shifting the barrier function.
Sample k0 points {X1 , . . . , Xk0 } from the Gaussian distribution centered at the minimizer x∗
−1
.
of φ with covariance σ02 ∇2 φ(x∗ )
nϑ
ϑ
2
while σi ≤ Θ(1) ε log ε do
Sample ki points {X1 , . . . , Xk } using Hamiltonian Monte Carlo with target density
f (σi2 , X) and the previous {X1 , . . . , Xk } as warm start.
Compute the ratio
ki
2 ,X )
f (σi+1
1 X
j
Wi+1 =
·
2, X ) .
ki
f
(σ
j
i
j=1
Increment i.
end
n
1
Output: (2πσ02 ) 2 det(∇2 φ(x∗ ))− 2 W1 . . . Wi as the volume estimate.
33
Lemma 39 ([27, Prop 2.3.2]). For any ϑ-self-concordance barrier function φ on convex K, for any
interior point x and y in K, we have that
φ(x) ≤ φ(y) + ϑ ln
1
1 − πy (x)
where πy (x) = inf{t ≥ 0|y + t−1 (x − y) ∈ K}.
Lemma 40 ([23, Lem 5.16]). For any logconcave distribution f on Rn and any β ≥ 2, we have
Px∼f (f (x) ≤ e−βn max f (y)) ≤ e−O(βn) .
y
3
3
Lemma 41 (Large and small σ 2 ). Let φ be a ϑ-self concordant barrier function for K. If σn 2 log 2
1
1
1 and σn 2 log 2 σ1 ≤ 1, then we have that
1
σ
≤
3
1
3 1
n
−2
∗
ln F (σ) − ln e−σ φ(x ) (2πσ 2 ) 2 det(∇2 φ(x∗ ))− 2 ≤ O(σn 2 log 2 ).
σ
If σ 2 ≥ ϑ, then
−2
∗
ln F (σ) − ln e−σ φ(x ) vol(K) ≤ O(σ −2 ϑ ln(σ 2 n/ϑ)).
Proof. We begin with the first inequality. Let S be the set of x such that φ(x) ≤ φ(x∗ ) + βσ 2 n for
some β ≥ 2 to be determined. Therefore,
f (σ 2 , x) ≥ e−βn f (σ 2 , x∗ )
for all x ∈ S. Since f (σ 2 , x) is a logconcave function, Lemma 40 shows that Px∼f (S) ≤ e−O(βn) .
Therefore,
Z
Z
Z
f (σ 2 , x) dx ≤ f (σ 2 , x) dx ≤ (1 + e−O(βn) ) f (σ 2 , x) dx.
S
S
In short, we have
−O(βn)
F (σ) = (1 ± e
)
Z
f (σ 2 , x) dx.
S
By the construction of S, Lemma 38 and the fact that φ is self-concordant, if βσ 2 n ≤ 14 , we have
that
√
1 ± Θ(σ βn)
φ(x) = φ(x∗ ) +
(x − x∗ )T ∇2 φ(x∗ )(x − x∗ )
2
for all x ∈ S. Hence,
Z
√
σ −2
−2
∗
∗ T 2
∗
∗
F (σ) = (1 ± e−O(βn) )e−σ φ(x ) e−(1±O(σ βn)) 2 (x−x ) ∇ φ(x )(x−x ) dx.
S
Now we note that
Z
Z
σ −2
σ −2
∗ T 2
∗
∗
∗ T 2
∗
∗
e−Θ( 2 (x−x ) ∇ φ(x )(x−x )) dx = e−O(βn) e− 2 (x−x ) ∇ φ(x )(x−x ) dx
Sc
because
S
σ −2
(x − x∗ )T ∇2 φ(x∗ )(x − x∗ ) = Ω(βn)
2
34
outside S. Therefore,
F (σ) = (1 ± e−O(βn) )e−σ
−O(βn)
= (1 ± e
−2 φ(x∗ )
Z
3
2
e−(1±O(σ
√
Rn
−σ−2 φ(x∗ )
± O(σ(βn) ))e
3
= (1 ± e−O(βn) ± O(σ(βn) 2 ))e−σ
−2 φ(x∗ )
−2
βn)) σ 2 (x−x∗ )T ∇2 φ(x∗ )(x−x∗ )
Z
e−
dx
σ −2
(x−x∗ )T ∇2 φ(x∗ )(x−x∗ )
2
dx
Rn
n
1
(2πσ 2 ) 2 det(∇2 φ(x∗ ))− 2
1
3
where we used that σ(βn) 2 = O(1) in first sentence and σ(βn) 2 = O(1) in the second sentence.
Setting β = Θ(log σ1 ), we get the first result.
For the second inequality, for any 0 ≤ t < 1 and any x ∈ x∗ + t(K − x∗ ), we have that πx∗ (x) ≤ t
(πx∗ is defined in Lemma 39). Therefore, Lemma 39 shows that
φ(x) ≤ φ(x∗ ) + ϑ ln
1
.
1−t
Note that Px∼µ (x ∈ x∗ + t(K − x∗ )) = tn where µ is the uniform distribution in K. Therefore, for
any 0 < β < 1, we have that
Px∼µ (φ(x) ≤ φ(x∗ ) + ϑ ln
1
}) ≥ (1 − β)n .
β
Hence,
vol(K) · e−σ
−2 φ(x∗ )
≥ F (σ 2 )
n
≥ vol(K) · (1 − β) exp −σ
−2
1
(φ(x ) + ϑ ln ) .
β
∗
Setting β = σ −2 n−1 ϑ, we get the second result.
6.2.2
Variance of Yx
Our goal is to estimate Ex∼µi (Yx ) within a target relative error. The algorithm estimates the
quantity Ex∼µi (Yx ) by taking random sample points x1 , . . . , xk and computing the empirical estimate
for Ex∼µi (Yx ) from the corresponding Yx1 , . . . , Yxk . The variance of Yxi divided by its expectation
squared will give a bound on how many independent samples xi are needed to estimate Ex∼µi (Yx )
within the target accuracy. We have
R
σ2 σi2
2φ(x)
φ(x)
−
dx
)
exp
F ( 2σ2i+1
2
2
2
K
σ
σ
i
i+1
i −σi+1
Ex∼µi (Yx2 ) =
=
R
φ(x)
F (σi2 )
dx
K exp − σ2
i
and
σ2
σ2
i
)
F (σi2 )F ( 2σ2i+1
2
Ex∼µi (Yx2 )
i −σi+1
=
2 )2
Ex∼µi (Yx )2
F (σi+1
2
If we let σ 2 = σi+1
and σi2 = σ 2 /(1 + r), then we can further simplify as
2 2
σ
σ
2
F
1+r F 1−r
Ex∼µi (Yx )
=
.
Ex∼µi (Yx )2
F (σ 2 )2
35
(6.1)
Lemma 42. For any 1 > r ≥ 0, we have that
2 2
σ
σ
Z r Z 1+t
F 1−r
F 1+r
1
= 4
Varx∼µs φ(x)dsdt
ln
σ 0 1−t
F (σ 2 )2
2
where µs be the probability distribution proportional to f ( σs , x).
2
Proof. Fix σ 2 . Let g(t) = ln F ( σt ). Then, we have that
2 2
2 2
σ
σ
σ
σ
Z r
F 1−r
F 1+r
d F 1+t F 1−t
=
dt
ln
ln
F (σ 2 )2
F (σ 2 )2
0 dt
Z r
d
d
=
g(1 + t) − g(1 − t)dt
dt
0 dt
Z r Z 1+t 2
d
=
g(s)dsdt.
2
ds
0
1−t
For
d2
g(s),
ds2
(6.2)
we have that
Z
s
exp − 2 φ(x) dx
σ
K
R
1 d K φ(x) · exp − σs2 φ(x) dx
R
=− 2 ·
s
σ ds
K exp − σ2 φ(x) dx
2 !
R
2 R
s
2 (x) · exp − s φ(x) dx
φ(x)
dx
φ(x)
·
exp
−
φ
1
2
2
K R
σ
σ
− KR
=
2
s
2
s
σ
K exp − σ2 φ(x) dx
K exp − σ2 φ(x) dx
1
1
= 4 Ex∼µs φ2 (x) − (Ex∼µs φ(x))2 = 4 Varx∼µs φ(x).
σ
σ
d2
d2
g(s)
=
ln
ds2
ds2
Putting it into (6.2), we have the result.
Now, we bound Varx∼µs φ(x). This can be viewed as a manifold version of the thin shell or
variance hypothesis estimate.
Lemma 43 (Thin shell estimates). Let φ be a ϑ-self concordant barrier function for K. Then, we
have that
2
σ
Varx∼µs φ(x) = O
ϑ .
s
def
Proof. Let Kt = {x ∈ K such that φ(x) ≤ t} and m be the number such that µs (Km ) = 21 . Let
Km,r = {x such that d(x, y) ≤ r and y ∈ Km }. By repeatedly applying Lemma 37, we have that
µs (Km,r ) = 1 − e−Ω(
√
s
r)
σ
.
√
By our assumption on φ, for any x and y, we have that |φ(x) − φ(y)| ≤ ϑd(x, y). Therefore,√for
√
s
any x ∈ Km,r , we have that φ(x) ≤ m + ϑr. Therefore, with probability at least 1 − e−Ω( √σ r)
√
√
s
φ(x) ≥ m − ϑr. Hence, with 1 − e−Ω( σ r)
in µs , it follows that φ(x) ≤ m + ϑr. Similarly,
√
probability in µs , we have that |φ(x) − m| ≤ ϑr. The bound on the variance follows.
36
Now we are ready to prove the key lemma.
Lemma 44. Let φ be a ϑ-self concordant barrier function for K. For any
Ex∼µi (Yx2 )
= O r 2 min
2
Ex∼µi (Yx )
1
2
ϑ
,n .
σi2
> r ≥ 0, we have that
Proof. Using Lemma 43 and Lemma 42, we have that
2 2
σ
σ
2
F 1−r
F 1+r
r ϑ
=O
.
ln
2
σ2
F (σ 2 )
(6.3)
This bounds is useful when σ 2 is large.
case σ 2 is small, we recall that for any logconcave function f , the function a →
RFor the
n
a
a
f (x) dx is logconcave (Lemma 3.2 in [8]). In particular, this shows that an F a1 is logconcave
1 1−r
in a. Therefore, with a = 1+r
σ2 , σ2 , σ2 ,we have
1
F (σ 2 )2 ≥
σ 4n
1+r
σ2
n
F
Rearranging the term, we have that
2 2
σ
σ
F 1−r
F 1+r
F (σ 2 )2
Therefore, we have that
ln
F
σ2
1+r
F
F
σ2
1+r
≤
σ2
1−r
(σ 2 )2
1−r
σ2
F
n
.
n
1
(1 + r)(1 − r)
σ2
1−r
.
Combining (6.1), (6.3) and (6.4), we have the result.
6.2.3
= O nr 2 .
(6.4)
Main lemma
Lemma 45. Given any ϑ-self-concordance barrier φ on a convex set K and 0 < ε < 12 , the algorithm
Volume(M , ε) outputs the volume of K to within a 1 ± ε multiplicative factor.
−2
∗
n
1
Proof. By our choice of ε, Lemma 41 shows that e−σ0 φ(x ) (2πσ02 ) 2 det(∇2 φ(x∗ ))− 2 is an 1± 4ε mul−2
∗
tiplicative approximation of F (σ0 ) and that e−σk φ(x ) vol(K) is a 1± 4ε multiplicative approximation
of F (σk ). Note that we shifted the function φ such that φ(x∗ ) = 0. Therefore,
k
2
n
1 Y F (σi+1 )
ε
vol(K) = (1 ± )(2πσ02 ) 2 det(∇2 φ(x∗ ))− 2
2) .
2
F
(σ
i
i=1
2 , X)/f (σ 2 , X) is upper
In Lemma 44, we showed that the variance of the estimator Y = f (σi+1
i
q
√
2
bounded by O(1)(EY ) . Note that the algorithm takes O( n) iterations to double σi if ϑn ≤ σi
√
and O( ϑσi−1 ) iterations otherwise. By a simple analysis of variance, to have relative error ε, it
e i ) samples in each phase.
suffices to have O(k
37
6.3
Volume computation with the log barrier
In this section, we prove the Theorem 2, restated below for convenience.
Theorem 2. For any polytope P = {x : Ax ≥ b} with m constraints and n variables, and any
ε > 0, the Hamiltonian
volume algorithm estimates the volume of P to within 1 ± ε multiplicative
2
−2
e mn 3 ε
factor using O
steps where each step consists of solving a first-order ODE and takes time
e mnω−1 LO(1) logO(1) 1 and L is the bit complexity3 of the polytope.
O
ε
1
m
e
n− 3 ). Since the number
n , the mixing time of HMC is O(m ·√
√
e 2n ), the total number of
σ 2 is O( n) and since we samples O(
ǫ
Proof. In the first part, when σ 2 ≤
of sampling phases to double such
steps of HMC is
√
√
n
− 31
e
e
e
O mn
=O
×O n ×O
2
ǫ
In the second part, when σ 2 ≥
m
n,
2
m · n3
ǫ2
!
.
√
1
2
e −2n 3 −1 +
the mixing time of HMC is O(
σ +m
1
m3 n3
σ
−2
3
1
1
3
+m−√
1
+ m 2 n 6 ).
m
−2
e
Since the number of sampling phases to double σ 2 is O(1+ σm ) and since we sample O((
σ +1)ε )
in each phase, the total number of steps of HMC is
!
!
√
√
1
2
2
1
3
3 n3
3
1
1
m
m
·
n
n
m
m
−2
e (
e
e
+ m2 n6 × O 1 +
+
+ 1)ε
.
×O
=O
O
σ −2 + m−1 σ − 32 + m− 13
σ
σ
ǫ2
Combining both parts, the total number of steps of HMC is
!
2
3
m
·
n
e
.
O
ǫ2
3
A.
L = log(m + dmax + kbk∞ ) where dmax is the largest absolute value of the determinant of a square sub-matrix of
38
7
Logarithmic barrier
For any polytope M = {Ax > b}, the logarithmic barrier function φ(x) is defined as
φ(x) = −
m
X
i=1
log(aTi x − bi ).
We denote the manifold induced by the logarithmic barrier on M by ML . The goal of this section
is to analyze Hamiltonian Monte Carlo on ML . In Section 7.1, we give explicit formulas for various
Riemannian geometry concepts on ML . In Section 7.2, we describe the HMC specialized to ML .
In Sections 7.3 to 7.5, we bound the parameters required by Theorem 30, resulting in Theorem 1.
The following parameters that are associated with barrier functions will be convenient.
Definition 46. For a convex function f , let M1 , M2 and M3 be the smallest numbers such that
−1
1. M1 ≥ maxx∈M (∇f (x))T ATx Ax
∇f (x) and M1 ≥ n.
2. ∇2 f M2 · ATx Ax .
3. Tr((ATx Ax )−1 ∇3 f (x)[v]) ≤ M3 kvkx for all v.
For the case f = φ are the standard logarithmic barrier, these parameters are n, 1,
tively.
7.1
√
n respec-
Riemannian geometry on ML (G2 )
We use the following definitions throughout this section.
Definition 47. For any matrix A ∈ Rm×n and vectors b ∈ Rm and x ∈ Rn , define
1. sx = Ax − b, Sx = Diag(sx ), Ax = Sx−1 A.
2. sx,v = Ax v, Sx,v = Diag(Ax v).
(2)
= (Px )2ij .
3. Px = Ax (ATx Ax )−1 ATx , σx = diag(Px ), Σx = Diag(P ), Px
ij
4. Gradient of φ: φi = −
P
ℓ
eTℓ Ax ei .
−1
P T
ej .
eℓ Ax ei eTℓ Ax ej , gij = eTi ATx Ax
5. Hessian of φ and its inverse: gij = φij = ATx Ax ij =
P
6. Third derivatives of φ: φijk = −2 ℓ eTℓ Ax ei eTℓ Ax ej eTℓ Ax ek .
7. For brevity (overloading notation), we define sγ ′ = sγ,γ ′ , sγ ′′ = sγ,γ ′′ , Sγ ′ = Sγ,γ ′ and
Sγ ′′ = Sγ,γ ′′ for a curve γ(t).
In this section, we will frequently use the following identities derived from elementary calculus
(using only the chain/product rules and the formula for derivative of the inverse of a matrix). For
reference, we include proofs in Appendix C.
Fact 48. For any matrix A and any curve γ(t), we have
dAγ
= −Sγ ′ Aγ ,
dt
dPγ
= −Sγ ′ Pγ − Pγ Sγ ′ + 2Pγ Sγ ′ Pγ ,
dt
dSγ ′
= Diag(−Sγ ′ Aγ γ ′ + Aγ γ ′′ ) = −Sγ2′ + Sγ ′′ ,
dt
39
We also use these matrix inequalities: Tr(AB) = Tr(BA), Tr(P AP ) ≤ Tr(A) for any psd
matrix A; Tr(ABAT ) ≤ Tr(AZAT ) for any B Z; the Cauchy-Schwartz, namely, Tr(AB) ≤
1
1
Tr(AAT ) 2 Tr(BB T ) 2 . We note Px2 = Px because Px is a projection matrix.
Since the manifold ML is naturally embedded in Rn , we can identify Tx ML with Euclidean
coordinates. We have that
hu, vix = uT ∇2 φ(x)v = uT ATx Ax v.
We will use the following two lemmas proved in [13].
Lemma 49. Let w(t) be a vector field defined on a curve z(t) in ML . Then,
∇z ′ w =
−1 T
−1 T
dw
dw
Az Sz ′ sz,w =
Az Sz,w sz ′ .
− ATz Az
− ATz Az
dt
dt
In particular, the equation for parallel transport on a curve γ(t) is given by
−1 T
d
Aγ Sγ ′ Aγ v.
v(t) = ATγ Aγ
dt
(7.1)
Lemma 50. Given u, v, w, x ∈ Tx ML , the Riemann Curvature Tensor at x is given by
R(u, v)w =
ATx Ax
def
−1
ATx (Sx,v Px Sx,w − Diag(Px sx,v sx,w )) Ax u
and the Ricci curvature Ric(u) = TrR(u, u) is given by
Ric(u) = sTx,u Px(2) sx,u − σxT Px s2x,u
where R(u, u) is the operator defined above.
7.2
Hamiltonian walk on ML
We often work in Euclidean coordinates. In this case, the Hamiltonian walk is given by the formula
in the next lemma. To implement the walk, we solve this ODE using the collocation method as
described in [13], after first reducing it to a first-order ODE. The resulting complexity is Õ(mnω−1 )
per step.
Lemma 51. The Hamiltonian curve at a point x in Euclidean coordinates is given by the following
equations
γ ′′ (t) =
ATγ Aγ
′
γ (0) = w,
−1
ATγ s2γ ′ + µ(γ(t))
∀t ≥ 0
γ(0) = x.
where µ(x) = ATx Ax
−1
ATx σx − (ATx Ax )−1 ∇f (x) and w ∼ N (0, (ATγ Aγ )−1 ).
Proof. Recall from Lemma 7 that the Hamiltonian walk is given by
Dt
dγ
=µ(γ(t)),
dt
dγ
(0) ∼N (0, g(x)−1 )
dt
40
where µ(x) = −g(x)−1 ∇f (x)− 12 g(x)−1 Tr g(x)−1 Dg(x) . By Lemma 49, applied with w(t) = γ ′ (t),
z(t) = γ(t), we have
−1 T 2
dγ
Dt
Aγ sγ ′ .
= γ ′′ (t) − ATγ Aγ
dt
For the formula of µ, we note that
∂
1
1X
(ATx Ax )−1 ij
Tr g(x)−1 Dg(x) k =
ATx Ax ji
2
2
∂xk
ij
X
X
(ATx Ax )−1 ij
eTℓ Ax ei eTℓ Ax ej eTℓ Ax ek
by Defn.47(6) = −
ij
=−
=
X
ℓ
ATx (ATx Ax )−1 Ax
ℓ
−ATx σx .
ℓℓ
eTℓ Ax ek
Therefore,
µ(x) = −(ATx Ax )−1 ∇f (x) + ATx Ax
−1
ATx σx .
Many parameters for Hamiltonian walk depends on the operator Φ(t). Here, we give a formula
of Φ(t) in Euclidean coordinates.
Lemma 52. Given a curve γ(t), in Euclidean coordinates, we have that
Φ(t) = M (t) − R(t)
where
−1
ATγ Sγ ′ Pγ Sγ ′ Aγ − ATγ Diag(Pγ s2γ ′ )Aγ
−1 T
Ax Sx,µ − 3Σx + 2Px(2) Ax − ∇2 f (x) .
M (t) = ATγ Aγ
R(t) = ATγ Aγ
Proof. Lemma 50 with v = w = γ ′ , gives the formula for R(t).
For M (t), Lemma 49 with w = µ(x), z ′ = u shows that
Du µ(x) = ∇u µ(x) − (ATx Ax )−1 ATx Sx,µ Ax u.
For the first term ∇u µ(x), we note that
−1
−1 T
ATx Sx,u Ax ATx Ax
Ax σx
−1
ATx Sx,u σx
− 3 ATx Ax
−1 T
+ 2 ATx Ax
Ax diag(Px Sx,u Px )
−1
−1
∇f (x)
ATx Sx,u Ax ATx Ax
− 2 ATx Ax
∇u µ(x) =2 ATx Ax
− (ATx Ax )−1 ∇2 f (x)u.
41
Therefore, we have that
Du µ(x)
−1 T (2)
−1 T
Ax Px Ax u
Ax Σx Ax u + 2 ATx Ax
ATx Diag(Px σx )Ax u − 3 ATx Ax
−1
−1 T
∇f (x) Ax u − (ATx Ax )−1 ∇2 f (x)u − (ATx Ax )−1 ATx Sx,µ Ax u
Ax Diag Ax ATx Ax
− 2 ATx Ax
−1
−1 T
∇f (x) − Sx,µ − 3Σx + 2Px(2) Ax − ∇2 f (x) u
Ax 2Diag(Px σx ) − 2Diag Aγ ATx Ax
= ATx Ax
−1 T
= ATx Ax
Ax 2Sx,µ − Sx,µ − 3Σx + 2Px(2) Ax − ∇2 f (x) u.
−1 T
= ATx Ax
Ax Sx,µ − 3Σx + 2Px(2) Ax − ∇2 f (x) u.
=2 ATx Ax
−1
where we used the facts that
µ = ATx Ax
−1
ATx σx − (ATx Ax )−1 ∇f (x)
Sx,µ = Ax µ = Diag(Px σx − Ax (ATx Ax )−1 ∇f (x)).
Remark. Note that R(t) and M (t) is symmetric in h·, ·iγ , but not in h·, ·i2 . That is why the formula
does not look symmetric.
7.3
Randomness of the Hamiltonian flow (ℓ0 )
Many parameters of a Hessian manifold relate to how fast a Hamiltonian curve approaches the
boundary of the polytope. Since the initial velocity of the
curve is drawn from a
Hamiltonian
1
Gaussian distribution, one can imagine that sγ ′ (0) ∞ = O √m sγ ′ (0) 2 (each coordinate of sγ ′
measures the relative rate at which the curve is approaching the corresponding facet). So the walk
initial approaches/leaves every facet of the polytope at roughly the same slow pace. If this holds
for the entire walk, it would allow us to get very tight bounds on various parameters. Although we
are not able to prove that sγ ′ (t) ∞ is stable throughout 0 ≤ t ≤ δ, we will show that sγ ′ (t) 4 is
stable and thereby obtain a good bound on sγ ′ (t) ∞ .
Throughout this section, we only use the randomness of the walk to prove that both sγ ′ (t) 4 and
sγ ′ (t) ∞ are small with high probability. Looking ahead, we will show that sγ ′ (t)
√
√
and sγ ′ (t) ∞ = O( log n + M1 δ) (Lemma 55), we define
sγ ′ (t)
def
ℓ(γ) = max
0≤t≤δ
n1/2
+
2
1/4
M1
+
sγ ′ (t)
1/4
M1
4
sγ ′ (t)
√∞
+√
+
log n + M1 δ
sγ ′ (0)
n1/2
2
+
sγ ′ (0)
n1/4
4
1/4
4
= O(M1 )
sγ ′ (0) ∞
+ √
log n
!
to capture this randomness involves in generating the geodesic walk. This allows us to perturb the
geodesic (Lemma 24) without worrying about the dependence on randomness.
We first prove the the walk is stable in the L4 norm and hence ℓ(γ) can be simply approximated
by sγ ′ (0) 4 and sγ ′ (0) ∞ .
Lemma 53. Let γ be a Hamiltonian flow in ML starting at x. Let v4 =
1
0≤t≤
1/4 , we have that
12(v4 +M1
)
42
sγ ′ (0)
4
. Then, for
1.
sγ ′ (t)
1/4
4
≤ 2v4 + M1 .
2. kγ ′′ (t)k2γ(t) ≤ 128v44 + 30M1 .
Proof. Let u(t) = sγ ′ (t)
4
. Then, we have (using Holder’s inequality in the first step),
d
du
≤
Aγ γ ′
dt
dt
≤ Aγ γ
′′
4
4
2
= Aγ γ ′′ − Aγ γ ′
+ u (t).
2
4
(7.2)
Under the Euclidean coordinates, by Lemma 51 the Hamiltonian flow is given by
γ ′′ (t) = ATγ Aγ
with µ(x) = ATx Ax
−1
γ ′′
−1
ATγ s2γ ′ + µ(γ(t))
ATx σx − (ATx Ax )−1 ∇f (x). Hence, we have that
2
γ
−1 T T −1 T 2
Aγ sγ ′
Aγ Aγ Aγ Aγ
Aγ ATγ Aγ
−1 T T −1 T
+ 3σγT Aγ ATγ Aγ
Aγ Aγ Aγ Aγ
Aγ σγ
−1
−1
+ 3(∇f (x))T ATγ Aγ
ATγ Aγ ATγ Aγ
∇f (x)
X
X
−1
≤3
(s4γ ′ )i + 3
(σγ2 )i + 3(∇f (x))T ATγ Aγ
∇f (x)
≤3 s2γ ′
T
i
4
i
≤3u (t) + 3(n + M1 ) ≤ 3u4 (t) + 6M1
Therefore, we have
Aγ γ ′′
4
≤ Aγ γ ′′
2
Plugging it into (7.2), we have that
1/4
(7.3)
p
≤ 2u2 (t) + 3 M1 .
p
du
≤ 3u2 (t) + 3 M1 .
dt
Note that when u ≤ 2v4 + M1 , we have that
p
du
1/4
≤ 12v42 + 9 M1 ≤ 12(v4 + M1 )2 .
dt
Since u(0) = v4 , for 0 ≤ t ≤
1
1/4 ,
12(v4 +M1 )
1/4
we have that u(t) ≤ 2v4 + M1
and this gives the first
inequality.
Using (7.3), we get the second inequality.
We can now prove that ℓ(γ) is small with high probability.
Lemma 54. Assume that δ ≤
1
1/4 ,
36M1
ℓ1 = Ω(n1/4 δ) and n is large enough, we have that
ℓ0
1
min 1,
Pγ∼x (ℓ(γ) ≥ 128) ≤
.
100
ℓ1 δ
Therefore, we have ℓ0 ≤ 256.
43
Proof. From the definition of the Hamiltonian curve (Lemma 7), we have that
Aγ γ ′ (0) = Bz
−1/2
and z ∼ N (0, I).
where B = Aγ ATγ Aγ
First, we estimate kAγ γ ′ (t)k4 . Lemma 65 shows that
Pz∼N (0,I) kBzk44 ≤ 3
Note that
P
eTi B
i
4
2
=
P
2
i (σγ )i
Pγ ′ (0)
def
X
eTi B
i
4
2
!1/4
4
2
s
+ kBk2→4 s ≤ 1 − exp(− ).
2
≤ n and kBk2→4 ≤ kBk2→2 = 1. Putting s =
′
Aγ γ (0)
4
4
n1/4
2 ,
we have that
√
n
).
≤ 11n ≤ 1 − exp(−
8
Therefore, we have that v4 = kAγ γ ′ (0)k4 ≤ 2n1/4 with probability at least 1 − exp(−
apply Lemma 53 to get that
1/4
1/4
sγ ′ (t) 4 ≤ 2v4 + M1 ≤ 5M1
√
n
8 ).
Now, we
1
1/4 .
12(v4 +M1 )
for all 0 ≤ t ≤
Next, we estimate kAγ γ ′ (t)k∞ . Since eTi Aγ γ ′ (0) = eTi Bx ∼ N (0, σi ), we have
2
√
t
T
′
.
Pγ ′ (0) ei Aγ γ (0) ≥ σi t ≤ 2 exp −
2
Hence, we have that
Pγ ′ (0)
′
Aγ γ (0)
∞
X
p
2 log n
≥ 2 log n ≤ 2
exp −
σi
i
P
P
n
log n
Since i exp − 2 log
is
concave
in
σ,
the
maximum
of
exp
−
on the feasible set {0 ≤
i
σi
σi
P
σ ≤ 1, σi = n} occurs on its vertices. Hence, we have that
p
2
Pγ ′ (0) Aγ γ ′ (0) ∞ ≥ 2 log n ≤ 2n exp (−2 log n) = .
n
√
√
Lemma 53 shows that kAγ γ ′′ k∞ ≤ kγ ′′ kγ(t) ≤ 46 n + 6 M1 . Hence, for any 0 ≤ t ≤ δ, we have
that
sγ ′ (t)
∞
Z
t
Aγ(t) γ ′′ (r) ∞ dr
≤ Aγ(t) γ (0) ∞ +
0
Z δ
sγ(t),i
′
≤ max
Aγ(r) γ ′′ (r)
Aγ(0) γ (0) ∞ +
i,0≤s≤t sγ(s),i
0
p
sγ(t),i p
√
≤ max
2 log n + (46 n + 6 M1 )δ
i,0≤s≤t sγ(s),i
p
sγ(t),i p
≤ max
2 log n + 52 M1 δ .
i,0≤s≤t sγ(s),i
′
44
∞
dr
(7.4)
sγ(t),i
sγ(s),i
Let z(t) = maxi,0≤s≤t
sγ(t),i = sγ(s),i exp
Z
. Note that
sγ ′ (α),i dα
t
r
since
(Aγ(t)−b)i = (Aγ(r)−b)i exp
Z
t
r
ai · γ ′ (α)
dα
(Aγ(α) − b)i
Hence, we have that
z ′ (t) ≤ z(t) sγ ′ (t)
∞
Solving this, since z(0) = 1, we get
z(t) ≤
Since t ≤ δ ≤
1
1/4 ,
36M1
1
.
√
√
1 − 2 log n + 52 M1 δ t
we have that z(t) ≤ 1.05. Putting this into (7.4), we have that
sγ ′ (t)
Finally, we estimate sγ ′ (t)
2
X
Pz∼N (0,I) kBzk22 ≤
P
i
eTi B
2
2
=
P
∞
p
p
≤ 3 log n + 55 M1 δ.
= kAγ γ ′ (t)k2 . Lemma 65 shows that
Note that
p
p
≤ z 2 (t) 2 log n + 52 M1 δ .
eTi B
i
2
2
!1/2
2
2
r
+ kBk2→2 r ≤ 1 − exp(− ).
2
1/2
≤ n and kBk2→2 ≤ 1. Putting s = n 3 , we have that
n
2
Pγ ′ (0) Aγ γ ′ (0) 2 ≤ 2n ≤ 1 − exp(− ).
18
i (σγ )i
Therefore, kAγ γ ′ (0)k22 ≤ 2n with high probability. Next, we note that
d
Aγ γ ′
dt
2
2
= Aγ γ ′ , Aγ γ ′′ − s2γ ′
E
D
−1 T 2
Aγ sγ ′ + Aγ µ(γ(t)) − Aγ s2γ ′
= Aγ γ ′ , Aγ ATγ Aγ
E
D
−1 T 2
Aγ sγ ′ − s2γ ′ + Aγ γ ′ , Aγ µ(γ(t))
= Aγ γ ′ , Aγ ATγ Aγ
= Aγ γ ′ , Aγ µ(γ(t)) .
Using that µ(x) = ATx Ax
d
Aγ γ ′
dt
2
2
=
−1
ATx σx − (ATx Ax )−1 ∇f (x), we have that
X
X
(γ ′ )i (∇f )i
(sγ ′ )i (σγ )i −
i
i
sX
sX
qX
q
2
2
≤
(sγ ′ )i
(σγ )i +
(sγ ′ )2i (∇f )T (ATγ Aγ )−1 (∇f )
i
≤ 2 Aγ γ ′
2
p
i
M1 .
√
≤ M1 . Since δ ≤
d
1
′
Therefore, we have that dt
kAγ γ ′ k2
1/4 , we have that kAγ γ (t)k2 ≤
36M1
1/4
1/4
√
√
M1
M1
1/4
kAγ γ ′ (0)k2 + 36
≤ 2n + 36
. Therefore, we have that kAγ γ ′ (t)k2 ≤ 2 n + M1 with
n
probability at least 1 − exp(− 18 ).
45
Combining our estimates on sγ ′ (t)
we have that
sγ ′ (t)
2
1/4
1/2
n + M1
max
P
0≤t≤δ
+
sγ ′ (t)
4
1/4
M1
2
, sγ ′ (t)
+√
√
n
2
n
≤ exp(− ) + exp(−
)+ .
18
8
n
sγ ′ (t)
log n +
4
and sγ ′ (t)
√∞
M1 δ
+
∞
sγ ′ (0)
n1/2
and using the assumption on δ,
2
+
sγ ′ (0)
n1/4
4
sγ ′ (0) ∞
+ √
log n
!
!
≥ 128
In Lemma 59, we indeed have that ℓ1 = Ω(n1/4 δ) and hence ℓℓ10δ = Ω(n−1/4 δ−2 ) = Ω(1). Therefore,
1
min 1, ℓℓ10δ when n is large enough.
the probability is less than 100
Here, we collect some simple consequences of small ℓ(γ) that we will use later.
Lemma 55. Given a Hamiltonian flow γ on ML with ℓ(γ) ≤ ℓ0 ≤ 256. For any 0 ≤ t ≤ δ,
1/4
1. kAγ γ ′ (t)k2 ≤ 256(n1/2 + M1 ), kAγ γ ′ (0)k2 ≤ 256n1/2 .
1/4
2. kAγ γ ′ (t)k4 ≤ 256M1 , kAγ γ ′ (0)k4 ≤ 256n1/4 .
√
√
√
3. kAγ γ ′ (t)k∞ ≤ 256 log n + M1 δ , kAγ γ ′ (0)k∞ ≤ 256 log n.
4. kγ ′′ (t)k2γ ≤ 1013 M1 .
Proof. The first three inequalities simply follow from the definition of ℓ(γ). Since kAγ γ ′ (0)k4 ≤
1/4
256M1 , Lemma 53 shows the last inequality.
7.4
Parameters R1 , R2 and R3
Lemma 56. For a Hamiltonian curve γ on ML with ℓ(γ) ≤ ℓ0 , we have that
sup kΦ(t)kF,γ(t) ≤ R1
0≤t≤ℓ
√
√
with R1 = O( M1 + M2 n).
Proof. Note that Φ(t) = M (t) − R(t) where
ATγ Sγ ′ Pγ Sγ ′ Aγ − ATγ Diag(Pγ s2γ ′ )Aγ ,
−1 T
M (t) = ATγ Aγ
Aγ (Sγ,µ − 3Σγ + 2Pγ(2) )Aγ − ∇2 f .
R(t) = ATγ Aγ
−1
We bound the Frobenius norm of Φ(t) separately.
For kR(t)kF,γ , we note that
kR(t)k2F,γ
≤2 (ATγ Aγ )−1/2 ATγ Sγ ′ Pγ Sγ ′ Aγ (ATγ Aγ )−1/2
=2TrPγ S Pγ S Pγ S Pγ S +
γ′
≤4 sγ ′
γ′
γ′
γ′
2
F
+ 2 (ATγ Aγ )−1/2 ATγ Diag(Pγ s2γ ′ )Aγ (ATγ Aγ )−1/2
2TrPγ Diag(Pγ s2γ ′ )Pγ Diag(Pγ s2γ ′ )
4
.
4
46
2
F
For kM (t)kF,γ , we note that
kM (t)k2F,γ
≤2 (ATγ Aγ )−1/2 ATγ (Sγ,µ − 3Σγ + 2Pγ(2) )Aγ (ATγ Aγ )−1/2
2
F
+ 2 (ATγ Aγ )−1/2 ∇2 f (ATγ Aγ )−1/2
2
F
≤6TrPγ Sγ,µ Pγ Sγ,µ + 54TrPγ Σγ Pγ Σγ + 24TrPγ Pγ(2) Pγ Pγ(2) + 2M22 n
2
≤6TrSγ,µ
+ 54TrΣ2γ + 24Tr(Pγ(2) )2 + 2M22 n.
(2)
where we used that diag(Pγ )
2
≤ kσγ k2 ≤ n and
TrPγ M Pγ M = TrPγ M Pγ M Pγ = TrPγ M M Pγ
= TrM Pγ Pγ M ≤ TrM 2 .
Note that
2
2
ksγ,µ k22 = kµ(γ)k2γ = Pγ σγ − Aγ (ATγ Aγ )−1 ∇f (γ)
and
Aγ ATγ Aγ
Therefore, we have that
−1
∇f (γ(t))
2
2
= ∇f (γ(t))T ATγ Aγ
≤ 2n + 2M1 ≤ 4M1
−1
∇f (γ(t)) = M1 .
kM (t)k2F,γ ≤ 102M1 + 2M22 n.
The claim follows from Lemma 55.
Lemma 57. Let γ be a Hamiltonian curve on ML with ℓ(γ) ≤ ℓ0 . Assume that δ2 ≤ √1n . For any
0 ≤ t ≤ δ, any curve c(r) starting from γ(t) and any vector field v(r) on c(r) with v(0) = γ ′ (t), we
have that
!
dc
d
+ δ k Dr v|r=0 kγ(t) .
TrΦ(v(r))|r=0 ≤ R2
dr
dr r=0 γ(t)
where
1/4
p
√
M
R2 = O( nM1 + nM1 δ2 + 1 +
δ
√
n log n √
+ nM2 + M3 ).
δ
Proof. We first bound TrR(t). By Lemma 50, we know that
(2)
T
Ric(v(r)) = sTc(r),v(r) Pc(r) sc(r),v(r) − σc(r)
Pc(r) s2c(r),v(r)
= Tr(Sc(r),v(r) Pc(r) Sc(r),v(r) Pc(r) ) − Tr(Diag(Pc(r) s2c(r),v(r) )Pc(r) ).
For simplicity, we suppress the parameter r and hence, we have
Ric(v) = Tr(Sc,v Pc Sc,v Pc ) − Tr(Diag(Pc s2c,v )Pc ).
47
d
d
c = c′ and dr
v = v ′ (in Euclidean coordinates). Since
We write dr
d
and dr
Sc,v = −Sc′ Sc,v + Sc,v′ , we have that
d
dr Pc
= −Sc′ Pc − Pc Sc′ + 2Pc Sc′ Pc
d
Ric(v)
dr
= −2Tr(Sc,v Sc′ Pc Sc,v Pc ) − 2Tr(Sc,v Pc Sc′ Sc,v Pc ) + 4Tr(Sc,v Pc Sc′ Pc Sc,v Pc )
−2Tr(Sc′ Sc,v Pc Sc,v Pc ) + 2Tr(Sc,v′ Pc Sc,v Pc )
+Tr(Diag(Pc s2c,v )Sc′ Pc ) + Tr(Diag(Pc s2c,v )Pc Sc′ ) − 2Tr(Diag(Pc s2c,v )Pc Sc′ Pc )
+Tr(Diag(Pc Sc′ s2c,v )Pc ) + Tr(Diag(Sc′ Pc s2c,v )Pc ) − 2Tr(Diag(Pc Sc′ Pc s2c,v )Pc )
+2Tr(Diag(Pc Sc,v Sc′ sc,v )Pc ) − 2Tr(Diag(Pc Sc,v sc,v′ )Pc )
= −6Tr(Sc,v Sc′ Pc Sc,v Pc ) + 4Tr(Sc,v Pc Sc′ Pc Sc,v Pc ) + 2Tr(Sc,v′ Pc Sc,v Pc )
+3Tr(Diag(Pc s2c,v )Sc′ Pc ) − 2Tr(Diag(Pc s2c,v )Pc Sc′ Pc )
+3Tr(Diag(Pc Sc′ s2c,v )Pc ) − 2Tr(Diag(Pc Sc′ Pc s2c,v )Pc )
−2Tr(Diag(Pc Sc,v sc,v′ )Pc ).
d
Let dr
Ric(v) = (1) + (2) where (1) is the sum of all terms not involving v ′ and (2) is the sum of
other terms.
For the first term (1), we have that
|(1)| ≤ 6 |Tr(Sc,v Sc′ Pc Sc,v Pc )| + 4 |Tr(Sc,v Pc Sc′ Pc Sc,v Pc )|
+3 Tr(Diag(Pc s2c,v )Sc′ Pc ) + 2 Tr(Diag(Pc s2c,v )Pc Sc′ Pc )
+3 Tr(Diag(Pc Sc′ s2c,v )Pc ) + 2 Tr(Diag(Pc Sc′ Pc s2c,v )Pc )
sX
sX
2
(sc,v )i
(sc,v )2i + 4 ksc′ k∞ |Tr(Pc Sc,v Pc Sc,v Pc )|
≤ 6 ksc′ k∞
i
i
sX
sX
sX
4
4
(sc,v )i kSc′ k2 + 2
(sc,v )i
(Pc Sc′ Pc )2ii
+3
i
i
i
sX
sX
sX
sX
(sc,v )4i
(Pc )2ii + 2 ksc′ k∞
(sc,v )4i
(Pc )2ii
+3 ksc′ k∞
i
≤
≤
i
i
10 ksc′ k∞ ksc,v k22 + 3 ksc,v k24 ksc′ k2
√
20 ksc′ k2 ksc,v k24 n.
+
√
7 ksc′ k∞ ksc,v k24 n
i
1/2
Since sc,v = sγ ′ at r = 0, we have that ksc,v k24 = O(M1 ) and hence
p
|(1)| = O
nM1 ksc′ k2 .
For the second term (2), we have that
|(2)| ≤ 2 Tr(Sc,v′ Pc Sc,v Pc ) + 2 Tr(Diag(Pc Sc,v sc,v′ )Pc )
sX
√
(sc,v′ sc,v )2i
≤ 2 sc,v′ 2 ksc,v k2 + 2 n
i
p
p
1/4
n log n + nM1 δ sc,v′
≤ O n1/2 + M1
sc,v′ 2 + O
p
p
1/4
= O M1 + n log n + nM1 δ sc,v′ 2
48
2
where we used ksc,v k∞ = sγ ′
1/4
M1 )
∞
=O
√
log n +
√
M1 δ and at r = 0, we have ksc,v k2 = Aγ(0) γ ′ (0)
sγ ′ 2 = O(n1/2 +
in the second-to-last line.
Note that at r = 0, by Lemma 49, we have
Dr v =
Therefore,
−1 T
dv
− ATc Ac
Ac Sc′ sc,v .
dr
sc,v′ = Ac v ′ = Ac (Dr v) − Ac ATc Ac
and hence
sc,v′
2
≤ kDr vk + Ac ATc Ac
≤ kDr vk + sγ ′
−1
−1
ATc Sc′ sc,v
ATc Sc′ sc,v
2
ksc′ k2
∞
Therefore,
p
p
p
p
1/4
|(2)| = O M1 + n log n + nM1 δ kDr vk +
log n + M1 δ ksc′ k2 .
Therefore, we have
d
Ric(v(r))|s=0
dr
p
p
p
1/4
=O
nM1 ksc′ k2 + O M1 + n log n + nM1 δ kDr vk
p
p
p
p
1/4
log n + M1 δ ksc′ k2
+ O M1 + n log n + nM1 δ
p
p
p
√
√
1/4
1/4
=O
nM1 + M1
log n + M1
M1 δ + n log n + nM1 δ2 ksc′ k2
!
√
1/4
M1
n log n p
+
+ nM1 δ kDr vk
+O
δ
δ
!
√
1/4
p
√
n log n
M1
2
=O
(ksc′ k2 + δ kDr vk) .
+
nM1 + nM1 δ +
δ
δ
√
√
3/4
where we used M1 δ = O( nM1 δ2 + nM1 ) at the last line.
Next, we bound TrM (t). Lemma 52 shows that
TrM (r) =Tr((ATc Ac )−1 ATc (Sc,µ − 3Σc + 2Pc(2) )Ac − (ATc Ac )−1 ∇2 f )
=Tr(Pc (Sc,µ − 3Σc + 2Pc(2) )) − Tr((ATc Ac )−1 ∇2 f )
=σcT Pc σc − σcT Ac (ATc Ac )−1 ∇f − 3Tr(Σ2c ) + 2Tr(Pc(3) ) − Tr((ATc Ac )−1 ∇2 f ).
where in the last step we used
Sc,µ = Ac µ = Diag(Pc σc − Ac (ATc Ac )−1 ∇f ).
49
2
=
Since
d
dr Pc
= −Sc′ Pc − Pc Sc′ + 2Pc Sc′ Pc and
d
dr Ac
= −Sc′ Ac , we have that
d
TrM (r)
dr
= − σcT Sc′ Pc σc − σcT Pc Sc′ σc + 2σcT Pc Sc′ Pc σc
− 4σcT Pc Sc′ σc + 4σcT Pc diag(Pc Sc′ Pc )
+ 3σcT Sc′ Ac (ATc Ac )−1 ∇f − 2diag(Pc Sc′ Pc )T Ac (ATc Ac )−1 ∇f
− 2σcT Ac (ATc Ac )−1 ATc Sc′ Ac (ATc Ac )−1 ∇f − σcT Ac (ATc Ac )−1 ∇2 f · c′
+ 12Tr(Sc′ Σ2c ) − 12Tr(Σc diag(Pc Sc′ Pc ))
− 6Tr(Pc(2) Sc′ Pc + Pc(2) Pc Sc′ ) + 12Tr(Pc(2) Pc Sc′ Pc )
− 2Tr((ATc Ac )−1 ATc Sc′ Ac (ATc Ac )−1 ∇2 f ) − Tr((ATc Ac )−1 ∇3 f [c′ ]).
Simplifying it, we have
d
TrM (r)
dr
= − 6σcT Sc′ Pc σc + 2σcT Pc Sc′ Pc σc + 4σcT Pc Pc(2) sc′
+ 3σcT Sc′ Ac (ATc Ac )−1 ∇f − 2sTc′ Pc(2) Ac (ATc Ac )−1 ∇f
− 2σcT Ac (ATc Ac )−1 ATc Sc′ Ac (ATc Ac )−1 ∇f − σcT Ac (ATc Ac )−1 ∇2 f · c′
+ 12Tr(Sc′ Σ2c ) − 12σcT Pc(2) sc′
− 6Tr(Pc(2) Sc′ Pc + Pc(2) Pc Sc′ ) + 12Tr(Pc(2) Pc Sc′ Pc )
− 2Tr((ATc Ac )−1 ATc Sc′ Ac (ATc Ac )−1 ∇2 f ) − Tr((ATc Ac )−1 ∇3 f [c′ ]).
d
TrM (r) = (3) + (4) where (3) is the sum of all terms not involving f and (4) is the sum of
Let dr
other terms with f .
For the first term (3), we have that
|(3)| ≤6 σcT Sc′ Pc σc + 2 σcT Pc Sc′ Pc σc + 4 σcT Pc Pc(2) sc′ + 12 Tr(Sc′ Σ2c ) + 12 σcT Pc(2) sc′
+ 6 Tr(Pc(2) Sc′ Pc + Pc(2) Pc Sc′ ) + 12 Tr(Pc(2) Pc Sc′ Pc )
q
√
√
√
√
≤6 σcT Sc2′ σc n + 2 ksc′ k∞ n + 4 n ksc′ k2 + 12 ksc′ k2 n + 12 kSc′ k2 n
+ 6 diag(Pc Pc(2) )
2
ksc′ k2 + 6 diag(Pc(2) Pc )
≤36n ksc′ k2 .
50
2
ksc′ k2 + 12 diag(Pc Pc(2) Pc )
2
ksc′ k2
For the second term (4), we have that
(4) ≤3 σcT Sc′ Ac (ATc Ac )−1 ∇f + 2 sTc′ Pc(2) Ac (ATc Ac )−1 ∇f
+ 2 σcT Ac (ATc Ac )−1 ATc Sc′ Ac (ATc Ac )−1 ∇f + σcT Ac (ATc Ac )−1 ∇2 f · c′
+ 2 Tr((ATc Ac )−1 ATc Sc′ Ac (ATc Ac )−1 ∇2 f ) + Tr((ATc Ac )−1 ∇3 f [c′ ])
q
q
q
q
(2)
(2)
T
T
T
T
−1
≤3 sc′ Σc sc′ ∇f (Ac Ac ) ∇f + sc′ Pc Pc Pc sc′ ∇f T (ATc Ac )−1 ∇f
q
q
q
q
T
T
T
−1
T
+ 2 σc Pc Sc′ Pc Sc′ Pc σc ∇f (Ac Ac ) ∇f + σc Pc σc c′ ∇2 f (ATc Ac )−1 ∇2 f · c′
+ 2 diag(Ac (ATc Ac )−1 ∇2 f (ATc Ac )−1 ATc ) 2 ksc′ k2 + Tr((ATc Ac )−1 ∇3 f [c′ ])
p
p
√
≤4 ksc′ k2 M1 + 2 ksc′ k∞ nM1 + 3 nM2 ksc′ k2 + M3 ksc′ k2
p
√
≤ 6 M1 + 3 nM2 + M3 ksc′ k2 .
Therefore,
p
√
d
TrM (s) ≤ O n + M1 + nM2 + M3 ksc′ k2 .
dr
Lemma 58. Let γ be a Hamiltonian curve on ML with ℓ(γ) ≤ ℓ0 . Assume that δ ≤
ζ(t) be the parallel transport of the vector γ ′ (0) on γ(t). Then,
1
1/4 .
36M1
sup kΦ(t)ζ(t)kγ(t) ≤ R3
0≤t≤δ
where
1
2
R3 = O M1
p
3
4
1
4
log n + M1 n δ + M2 n
1
2
.
Proof. By Lemma 52, we have that
−1 T
Aγ (Sγ,µ − 3Σγ + 2Pγ(2) − Sγ ′ Pγ Sγ ′ + Diag(Pγ s2γ ′ ))Aγ − ∇2 f
Φ(t) = ATγ Aγ
= (1) + (2)
where (2) is the last term − ATγ Aγ
For the first term, we have that
−1
∇2 f .
k(1)ζkγ = Pγ (Sγ,µ − 3Σγ + 2Pγ(2) − Sγ ′ Pγ Sγ ′ + Diag(Pγ s2γ ′ ))sγ,ζ
≤ (Sγ,µ − 3Σγ +
2Pγ(2)
− Sγ ′ Pγ Sγ ′ +
Diag(Pγ s2γ ′ ))sγ,ζ
Pγ(2) sγ,ζ
2
2
+ Sγ ′ Pγ Sγ ′ sγ,ζ 2 + Diag(Pγ s2γ ′ )sγ,ζ
≤ kSγ,µ sγ,ζ k2 + 3 kΣγ sγ,ζ k2 + 2
2
≤ ksγ,ζ k∞ ksγ,µ k2 + kσγ k2 + Pγ s2γ ′ 2 + 2 Pγ(2) sγ,ζ + sγ ′ ∞ Sγ ′ sγ,ζ 2
2
p
√
2
(2)
≤ ksγ,ζ k∞ 2 M1 + n + sγ ′ 4 + 2 Pγ sγ,ζ + sγ ′ ∞ sγ ′ 4 ksγ,ζ k4
2
51
2
Let
where we used that ksγ,µ k22 = kµ(x)k2x ≤ 4M1 , kσγ k2 ≤ n. Now, we note that
Pγ(2) sγ,ζ
2
2
=
X
i
2
X
(Pγ )2ij (sγ,ζ )j
j
≤ ksγ,ζ k2∞
= ksγ,ζ k2∞
X
i
X
X
(Pγ )2ij
j
σγ2
i
i
= n ksγ,ζ k2∞ .
√
√
1/4
′
log
n
+
M1 δ), we have that
=
O(M
)
and
s
=
O(
γ
1
4
∞
p
2
k(1)ζkγ ≤ ksγ,ζ k∞ 5 M1 + sγ ′ 4 + sγ ′ ∞ sγ ′ 4 ksγ,ζ k4
p
1/2
1/4
3/4
≤ O M1
ksγ,ζ k∞ + O( log nM1 + M1 δ) ksγ,ζ k4 .
Using also that sγ ′
For the second term, we have that
k(1)ζkγ = Pγ ∇2 f ζ
2
2
≤ ζ T ∇2 f · ζ ≤ M2 ksγ,ζ k2 .
Combining both terms, we have that
p
1/4
3/4
1/2
kΦ(t)ζkγ = O M1
ksγ,ζ k∞ + O( log nM1 + M1 δ) ksγ,ζ k4 + M2 ksγ,ζ k2 .
(7.5)
Now, we bound ksγ,ζ k2 , ksγ,ζ k4 and ksγ,ζ k∞ . (7.1) shows that
−1 T
d
ζ(t) = ATγ Aγ
Aγ Sγ ′ Aγ ζ.
dt
Let wp (t) = kAγ ζ(t)kp . Then, we have that
d
d
wp (t) ≤
Aγ ζ(t)
dt
dt
p
d
ζ(t)
dt
≤ Sγ ′ Aγ ζ(t)
p
+ Aγ
≤ Sγ ′ Aγ ζ(t)
p
+ Pγ Sγ ′ Aγ ζ
≤ sγ ′
d
w2 (t) ≤ 2 sγ ′
For p = 2, we have that dt
1
that t ≤ δ ≤
1/4 , we have
∞
w (t) + Sγ ′ Aγ ζ
∞ p
w2 (t) ≤ e512(
= O(n
where we used that ζ(0) = γ ′ (0) and t ≤
O(n1/2 ).
√
√
log n+ M1 δ)t
1/2
p
.
2
w2 (t). Using that sγ ′
36M1
p
∞
≤ 256
√
log n +
√
M1 δ and
w2 (0)
)
1
1/4
12(v4 +M1 )
52
at the end. Therefore, we have that ksγ,ζ k2 =
For p = 4, we note that
d
w4 (t) ≤ sγ ′ ∞ w4 (t) + sγ ′ 4 ksγ,ζ k4
dt
1/4
≤ 2 sγ ′ 4 w4 (t) = O(M1 w4 (t)).
1
1/4 , we have again
36M1
O(n1/4 ), we have that
that w4 (t) = O(w4 (0)). Since w4 (0) = kAγ ζ(0)k4 =
Since t ≤ δ ≤
kAγ γ ′ (0)k4 =
w4 (t) = O(n1/4 ).
For p = ∞, we note that
d
w∞ (t) ≤ sγ ′ ∞ w∞ (t) + sγ ′ 4 ksγ,ζ k4
dt
1/4
1/4
≤ O(M1 w∞ (t)) + O(M1 n1/4 ).
Again using t ≤ δ ≤
1
1/4 ,
36M1
√
1/4
we have that w∞ (t) ≤ O( log n + M1 n1/4 δ).
Combining our bounds on w2 , w4 , w∞ to (7.5), we get
1
3
3
1
p
p
1
1
1
1
2
4
kΦ(t)ζkγ = O M1 log n + M1 n 4 δ + O( log nM14 n 4 + M14 n 4 δ) + O M2 n 2
1
3
p
1
1
2
4
4
2
= O M1 log n + M1 n δ + M2 n
.
7.5
Stability of L2 + L4 + L∞ norm (ℓ1 )
Lemma 59. Given a family of Hamiltonian curves γr (t) on ML with ℓ(γ0 ) ≤ ℓ0 , for δ2 ≤
1 √
√
, we have that
M +M n
1
2
d
1
1/4
ℓ(γr ) ≤ O M1 δ + √
dr
δ log n
1/4
Hence, ℓ1 = O M1 δ +
Proof. For brevity, all
d
dr
√1
δ log n
d
γr (0)
dr
.
are evaluated at r = 0. Since
d
sγ ′
dr r
2
≤ sγr′
∞
Aγr
d
γr
dr
d
′
dr sγr
+ δ Dr γr′ (0)
γr (0)
γr (0)
!
(7.6)
.
d ′
= −Sγr , d γr sγr′ + Aγr dr
γr , we have
dr
+ Aγr
2
d ′
γ
dr r
.
2
For the last term, we note that
Dr γr′ =
Hence, we have
Aγr
d ′
γ
dr r
2
≤ Dr γr′
γr
−1 T
d ′
γr − ATγr Aγr
Aγr Sγr , d γr sγr′ .
dr
dr
+ Sγr , d γr sγ ′
dr
53
2
≤ Dr γr′
γr
+ sγr′
∞
sγr , d γr
dr
2
.
Therefore, we have that
d
sγ ′
dr r
2
d
+ Dr γ ′
γr
dr
γr
p
d
p
=O
log n + M1 δ
γr
dr
≤ 2 sγr′
∞
γ
+ Dr γr′ .
(7.7)
γr
Since γr is a family of Hamiltonian curves, Lemma 19 shows that
′′
ψ (t) = Γt Φ(t)Γ−1
t ψ(t)
where ψ(t) is the parallel transport of
d
ψ(t) = γr (0) + Dr γr′ (0)t +
dr
d
dr γr (t)
Z
0
t
from γr (t) to γr (0). By Lemma 49, we have that
(t − r)Γr Φ(r)Γ−1
r (
d
γr (0) + Dr γr′ (0)r + E(r))dr
dr
def
d
with kE(r)kF ≤ O(1)∆ and ∆ = dr
γr (0) γr (0) + δ kDr γr′ (0)kγr (0) where we used that kΦ(t)kF,γ =
√
√
1 √
O( M1 + M2 n) (Lemma 56) and that s2 ≤ δ2 ≤ √M +M
.
1
2 n
Therefore, we have that
Z t
d
γr (t)
= ψ(t) γr (0) ≤ ∆ + O(∆) (t − s) Γr Φ(s)Γ−1
dr
r
γr (0)
dr
0
γr (t)
≤ O(∆)
√
√
where we used again kΦ(s)kF,γ = O( M1 + M2 n) and s2 ≤ δ2 ≤
′
Similarly, we have that kDr γr′ (t)kγr (t) = ψ (t)
√1 ,
n
Putting these into (7.7) and using h ≤
d
sγ ′
dr r
We write
ℓ(γr ) = max
0≤t≤δ
sγr′ (t)
n1/2 +
2
1/4
M1
+
sγr′ (t)
1 √
.
M1 +M2 n
≤ O( ∆
δ ).
we have
∆
M1 δ ∆ +
δ
p
1
M1 δ +
=O
∆.
δ
=O
2
γr (0)
√
4
1/4
M1
p
log n +
sγr′ (t)
+√
log n +
p
√∞
M1 δ
+
sγr′ (0)
(7.8)
2
n1/2
+
According to same calculation as (7.7), we can improve the estimate on
d
sγ ′ (0)
dr r
2
≤
p
log n
d
γr (0)
dr
∆
≤ .
δ
sγr′ (0)
n1/4
d
′
dr sγr 2
4
sγ ′ (0) ∞
+ √r
log n
!
.
for t = 0 and get
+ Dr γr′ (0)
γr
(7.9)
54
Using (7.8) and (7.9), we have that
d
d
′
′
dr sγr (t) 2
dr sγr (0) 2
√
max
+
1/4
0≤t≤δ
log n
M1
!
√
M1 δ + 1δ
1
∆
+ √
1/4
δ log n
M1
d
ℓ(γr ) = O
dr
=O
=O
7.6
1/4
M1 δ
1
+ √
δ log n
!
∆.
Mixing Time
Lemma 60. If f (x) = α · φ(x) (logarithmic barrier), then, we have that M1 = n + α2 m, M2 = α
√
and M3 = 2α · n.
−1
Proof. For M1 , we note that (∇f (x))T ATx Ax
∇f (x) = α2 1T Ax (ATx Ax )−1 ATx 1 ≤ α2 m. Hence,
M1 = n + α2 m.
For M2 , it directly follows from the definition.
For M3 , we note that
Tr((ATx Ax )−1 ∇3 f (x)[v]) = −2αTr((ATx Ax )−1 ATx Sx,v Ax )
X
σx,i (sx,v )i .
= −2α
i
Hence, we have
Tr((ATx Ax )−1 ∇3 f (x)[v])
sX
sX
√
2
σx,i
(sx,v )2i ≤ 2α n kvkx .
≤ 2α
i
i
Using Theorem 29, we have the following
Lemma 61. There is a universal constant c > 0 such that if the step size
1
1
1
1
1
1
1
δ ≤ c · min n− 3 , α− 3 m− 6 n− 6 , α− 2 m− 4 n− 12 ,
then, all the δ conditions for Theorem 30 are satisfied.
Proof. In the previous section, we proved that if δ ≤
1
1/4
36M1
and n is large enough,
1. ℓ0 = 256 (Lemma 54)
1/4
1
(Lemma 59)
2. ℓ1 = O M1 δ + δ√log
n
√
√
3. R1 = O( M1 + M2 n) (Lemma 56)
√
√
4. R2 = O( nM1 + nM1 δ2 +
1/4
M1
δ
+
√
n log n
δ
+
55
√
nM2 + M3 ) (Lemma 57)
3
1√
1
1
5. R3 = O(M12 log n + M14 n 4 δ + M2 n 2 ) (Lemma 58)
1
1
1
1
1
Substituting the value of M1 , M2 and M3 and using that δ . n− 3 , δ . α− 2 n− 3 and δ . α− 2 m− 4 ,
we have that
1 1
1
1. ℓ1 = O α 2 m 4 δ + δ√log
n
√
√
2. R1 = O( n + α m)
1 √
√
√
4
3. R2 = O(n + α nm + αm +δ n log n )
√
√
3
3
1
4. R3 = O( n log n + α m log n + nδ + α 2 m 4 n 4 δ)
Now, we verify all the δ conditions for Theorem 30.
1
1
1
Using δ . n− 3 and δ . α− 2 m− 4 , we have that δ2 .
Using δ . n
− 13
− 12
and δ . α
m
− 41
1
R1 .
, we have that
1
1
1
δ5 R12 ℓ1 . δ5 (n + α2 m) · (α 2 m 4 δ + √
) ≤ ℓ0 .
δ log n
For the last condition, we note that
p
√
√
1
δ3 R2 + δ2 R3 .nδ3 + α nmδ3 + αm 4 δ2 + n log nδ2
p
p
3
1
3
+ n log nδ2 + α m log nδ2 + nδ3 + α 2 m 4 n 4 δ3
p
p
√
√
3
1
1
3
. αm 4 δ2 + n log nδ2 + α m log nδ2 + nδ3 + α nmδ3 + α 2 m 4 n 4 δ3
p
p
√
3
3
1
. n log nδ2 + α m log nδ2 + nδ3 + α nmδ3 + α 2 m 4 n 4 δ3
where we used that
Therefore, if
√
√
1
αm 4 . (1 + α m) at the end
1
1
1
1
1
1
1
δ ≤ c · min n− 3 , α− 3 m− 6 n− 6 , α− 2 m− 4 n− 12
for small enough constant, then all the δ conditions for Theorem 30 are satisfied.
Acknowledgement. We thank Ben Cousins for helpful discussions. This work was supported in part
by NSF awards CCF-1563838, CCF-1717349 and CCF-1740551.
References
[1] D. Applegate and R. Kannan. Sampling and integration of near log-concave functions. In
STOC, pages 156–163, 1991.
[2] M. Betancourt. A Conceptual Introduction to Hamiltonian Monte Carlo. ArXiv e-prints,
January 2017.
[3] B. Cousins and S. Vempala. Bypassing KLS: Gaussian cooling and an O ∗ (n3 ) volume algorithm.
In STOC, pages 539–548, 2015.
[4] M. E. Dyer and A. M. Frieze. Computing the volume of a convex body: a case where randomness provably helps. In Proc. of AMS Symposium on Probabilistic Combinatorics and Its
Applications, pages 123–170, 1991.
56
[5] M. E. Dyer, A. M. Frieze, and R. Kannan. A random polynomial time algorithm for approximating the volume of convex bodies. In STOC, pages 375–381, 1989.
[6] M. Girolami, B. Calderhead, and S. A. Chin. Riemannian Manifold Hamiltonian Monte Carlo.
ArXiv e-prints, July 2009.
[7] Mark Girolami and Ben Calderhead. Riemann manifold langevin and hamiltonian monte carlo
methods. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 73(2):123–
214, 2011.
[8] A. T. Kalai and S. Vempala. Simulated annealing for convex optimization. Math. Oper. Res.,
31(2):253–266, 2006.
[9] R. Kannan, L. Lovász, and M. Simonovits. Isoperimetric problems for convex bodies and a
localization lemma. Discrete & Computational Geometry, 13:541–559, 1995.
[10] R. Kannan, L. Lovász, and M. Simonovits. Random walks and an O ∗ (n5 ) volume algorithm
for convex bodies. Random Structures and Algorithms, 11:1–50, 1997.
[11] R. Kannan and H. Narayanan. Random walks on polytopes and an affine interior point method
for linear programming. In STOC, pages 561–570, 2009.
[12] Yin Tat Lee and Aaron Sidford. Path finding methods for linear programming: Solving linear
programs in õ (vrank) iterations and faster algorithms for maximum flow. In Foundations
of Computer Science (FOCS), 2014 IEEE 55th Annual Symposium on, pages 424–433. IEEE,
2014.
[13] Yin Tat Lee and Santosh S. Vempala. Geodesic walks in polytopes. CoRR, abs/1606.04696,
2016.
[14] Yin Tat Lee and Santosh Srinivas Vempala. Eldan’s stochastic localization and the KLS hyperplane conjecture: An improved lower bound for expansion. CoRR, abs/1612.01507, 2016.
[15] Yin Tat Lee and Santosh Srinivas Vempala. Eldan’s stochastic localization and the KLS hyperplane conjecture: An improved lower bound for expansion. In Proc. of IEEE FOCS, 2017.
[16] Yin Tat Lee and Santosh Srinivas Vempala. Geodesic walks in polytopes. In Proceedings of the
49th Annual ACM SIGACT Symposium on Theory of Computing, STOC 2017, Montreal, QC,
Canada, June 19-23, 2017, pages 927–940, 2017.
[17] L. Lovász and M. Simonovits. Mixing rate of Markov chains, an isoperimetric inequality, and
computing the volume. In ROCS, pages 482–491, 1990.
[18] L. Lovász and M. Simonovits. On the randomized complexity of volume and diameter. In Proc.
33rd IEEE Annual Symp. on Found. of Comp. Sci., pages 482–491, 1992.
[19] L. Lovász and M. Simonovits. Random walks in a convex body and an improved volume
algorithm. In Random Structures and Alg., volume 4, pages 359–412, 1993.
[20] L. Lovász and S. Vempala. Fast algorithms for logconcave functions: sampling, rounding,
integration and optimization. In FOCS, pages 57–68, 2006.
[21] L. Lovász and S. Vempala. Hit-and-run from a corner. SIAM J. Computing, 35:985–1005, 2006.
57
[22] L. Lovász and S. Vempala. Simulated annealing in convex bodies and an O ∗ (n4 ) volume
algorithm. J. Comput. Syst. Sci., 72(2):392–417, 2006.
[23] L. Lovász and S. Vempala. The geometry of logconcave functions and sampling algorithms.
Random Struct. Algorithms, 30(3):307–358, 2007.
[24] Oren Mangoubi and Aaron Smith. Rapid mixing of hamiltonian monte carlo on strongly logconcave distributions. arXiv preprint arXiv:1708.07114, 2017.
[25] Radford M. Neal. MCMC using Hamiltonian dynamics. Handbook of Markov Chain Monte
Carlo, 54:113–162, 2010.
[26] R.M. Neal. Bayesian Learning for Neural Networks. Lecture Notes in Statistics. Springer New
York, 1996.
[27] Yurii Nesterov, Arkadii Nemirovskii, and Yinyu Ye. Interior-point polynomial algorithms in
convex programming, volume 13. SIAM, 1994.
[28] Burt Totaro. The curvature of a hessian metric.
15(04):369–391, 2004.
International Journal of Mathematics,
[29] S. Vempala. Geometric random walks: A survey. MSRI Combinatorial and Computational
Geometry, 52:573–612, 2005.
58
A
Matrix ODE
In this section, we prove Lemmas (20) and (21) for the solution of the ODE (3.4), restated below
for convenience.
d2
Ψ(t) = Φ(t)Ψ(t),
dt2
d
Ψ(0) = B,
dt
Ψ(0) = A.
Lemma 62. Consider the matrix ODE (3.4). Let λ = max0≤t≤ℓ kΦ(t)k2 . For any t ≥ 0, we have
that
√
√
kBk
kΨ(t)k2 ≤ kAk2 cosh( λt) + √ 2 sinh( λt).
λ
Proof. Note that
Z
′
t
Ψ(t) = Ψ(0) + tΨ (0) +
(t − s)Ψ′′ (s)ds
0
Z t
= A + tB +
(t − s)Φ(s)Ψ(s)ds.
(A.1)
0
Let a(t) = kΨ(t)k2 , then we have that
a(t) ≤ kAk2 + t kBk2 + λ
Z
t
(t − s)a(s)ds.
0
Let a(t) be the solution of the integral equation
a(t) = kAk2 + t kBk2 + λ
Z
t
0
(t − s)a(s)ds.
By induction, we have that a(t) ≤ a(t) for all t ≥ 0. By taking derivatives on both sides, we have
that
a′′ (t) = λa(t), a(0) = kAk2 , a′ (0) = kBk2 .
Solving these equations, we have
√
√
kBk
kΨ(t)k2 = a(t) ≤ a(t) = kAk2 cosh( λt) + √ 2 sinh( λt)
λ
for all t ≥ 0.
Lemma 63. Consider the matrix ODE (3.4). Let λ = max0≤t≤ℓ kΦ(t)kF . For any 0 ≤ t ≤
have that
t3
2
kΨ(t) − A − BtkF ≤ λ t kAk2 + kBk2 .
5
In particular, this shows that
Ψ(t) = A + Bt +
with kE(s)kF ≤ λ s2 kAk2 +
s3
5
kBk2 .
Z
t
0
(t − s)Φ(s)(A + Bs + E(s))ds
59
√1 ,
λ
we
Proof. Recall from (A.1) that
Ψ(t) = A + tB +
Z
t
0
(A.2)
(t − s)Φ(s)Ψ(s)ds.
Let E(t) = Ψ(t) − (A + tB). Using Lemma 20, we have that
kE(t)kF =
Z
≤λ
t
0
Z
Z
(t − s)Φ(s)Ψ(s)ds
F
t
(t − s) kΨ(s)k2 ds
0
√
√
kBk
(t − s) kAk2 cosh( λs) + √ 2 sinh( λs)ds
λ
0
√
√
√
kBk2
= λ kAk2 (cosh( λt) − 1) + √ (sinh( λt) − λt) .
λ
≤λ
t
√
√
√
Since 0 ≤ t ≤ √1λ , we have that cosh( λt) − 1 ≤ λt2 and sinh( λt) − λt ≤
the result.
The last equality follows again from (A.2)
λ3/2 t3
5 .
This gives
Next, we have an elementary lemma about the determinant.
Lemma 64. Suppose that E is a matrix (not necessarily symmetric) with kEk2 ≤ 41 , we have
|log det(I + E) − TrE| ≤ kEk2F .
Proof. Let f (t) = log det(I + tE). Then, by Jacobi’s formula, we have
f ′ (t) = Tr (I + tE)−1 E ,
f ′′ (t) = −Tr((I + tE)−1 E(I + tE)−1 E).
Since kEk2 ≤ 14 , we have that (I + tE)−1
f ′′ (t)
=
≤
2
≤
4
3
and hence
Tr((I + tE)−1 E(I + tE)−1 E)
T
Tr(E T (I + tE)−1 (I + tE)−1 E)
≤ 2 Tr(E T E) = 2 kEk2F .
The result follows from
Z
1
(1 − s)f ′′ (s)ds
f (1) = f (0) + f (0) +
0
Z 1
(1 − s)f ′′ (s)ds.
= Tr(E) +
′
0
60
B
Concentration
Lemma 65 ([13, Ver 3, Lemma 90]). For p ≥ 1, we have
p
!1/p
2
p+1 X
p/2
2 Γ( 2 )
t
p
p
√
+ kAk2→p t
Px∼N (0,I) kAxkp ≤
kai k2
.
≤ 1 − exp −
π
2
i
In particular, we have
and
Px∼N (0,I) kAxk44 ≤ 3
Px∼N (0,I) kAxk22 ≤
C
X
i
X
i
kai k42
kai k22
!1/4
!1/2
4
2
t
+ kAk2→4 t
≤ 1 − exp −
2
2
2
t
.
+ kAk2→2 t ≤ 1 − exp −
2
Calculus
Proof of Fact 48. Recall Definition 47 and write
dSγ−1
dAγ
=
A
dt
dt
dSγ −1
Sγ A
= −Sγ−1
dt
d(Aγ − b)
−1
= −Sγ Diag
Aγ
dt
= −Diag Sγ−1 Aγ ′ Aγ
= −Diag(Aγ γ ′ )Aγ = −Sγ ′ Aγ .
For the second, using the first,
dAγ (ATγ Aγ )−1 ATγ
dPγ
=
dt
dt
d(ATγ Aγ )−1 T
dAγ T
dAγ
=
(Aγ Aγ )−1 ATγ + Aγ (ATγ Aγ )−1
+ Aγ
Aγ
dt
dt
dt
d(ATγ Aγ ) T
(Aγ Aγ )−1 ATγ
= −Sγ ′ Pγ − Pγ Sγ ′ − Aγ (ATγ Aγ )−1
dt
= −Sγ ′ Pγ − Pγ Sγ ′ + 2Aγ (ATγ Aγ )−1 ATγ Sγ ′ Aγ (ATγ Aγ )−1 ATγ
= −Sγ ′ Pγ − Pγ Sγ ′ + 2Pγ Sγ ′ Pγ .
And for the last,
dSγ ′
= Diag
dt
dAγ ′
γ + Aγ γ ′′
dt
= Diag(−Sγ ′ Aγ γ ′ + Aγ γ ′′ ) = −Sγ2′ + Sγ ′′ .
61
D
Basic definitions of Riemannian geometry
Here we recall basic notions of Riemannian geometry. One can think of a manifold M as a ndimensional “surface” in Rk for some k ≥ n.
1. Tangent space Tp M : For any point p, the tangent space Tp M of M at point p is a linear
subspace of Rk of dimension n. Intuitively, Tp M is the vector space of possible directions that
are tangential to the manifold at x. Equivalently, it can be thought as the first-order linear
d
c(t) is tangent
approximation of the manifold M at p. For any curve c on M , the direction dt
′
to M and hence lies in Tc(t) M . When it is clear from context, we define c (t) = dc
dt (t). For any
n
n
open subset M of R , we can identify Tp M with R because all directions can be realized by
derivatives of some curves in Rn .
2. Riemannian metric: For any v, u ∈ Tp M , the inner product (Riemannianq
metric) at p is given
by hv, uip and this allows us to define the norm of a vector kvkp = hv, vip . We call a
manifold a Riemannian manifold if it is equipped with a Riemannian metric. When it is clear
from context, we define hv, ui = hv, uip . In Rn , hv, uip is the usual ℓ2 inner product.
3. Differential (Pushforward) d: Given a function f from a manifold M to a manifold N , we
define df (x) as the linear map from Tx M to Tf (x) N such that
df (x)(c′ (0)) = (f ◦ c)′ (0)
for any curve c on M starting at x = c(0). When M and N are Euclidean spaces, df (x) is the
Jacobian of f at x. We can think of pushforward as a manifold Jacobian, i.e., the first-order
approximation of a map from a manifold to a manifold.
4. Hessian manifold: We call M a Hessian manifold (induced by φ) if M is an open subset of Rn
with the Riemannian metric at any point p ∈ M defined by
hv, uip = v T ∇2 φ(p)u
where v, u ∈ Tp M and φ is a smooth convex function on M .
5. Length: For any curve c : [0, 1] → M , we define its length by
L(c) =
Z
1
0
d
c(t)
dt
dt.
c(t)
6. Distance: For any x, y ∈ M , we define d(x, y) be the infimum of the lengths of all paths
connecting x and y. In Rn , d(x, y) = kx − yk2 .
7. Geodesic: We call a curve γ(t) : [a, b] → M a geodesic if it satisfies both of the following
conditions:
(a) The curve γ(t) is parameterized with constant speed. Namely,
for t ∈ [a, b].
d
dt γ(t) γ(t)
is constant
(b) The curve is the locally shortest length curve between γ(a) and γ(b). Namely, for any
family of curve c(t, s) with c(t, 0) = γ(t) and c(a, s) = γ(a) and c(b, s) = γ(b), we have
Rb d
d
c(t, s) c(t,s) dt = 0.
that ds
s=0 a dt
62
Note that, if γ(t) is a geodesic, then γ(αt) is a geodesic for any α. Intuitively, geodesics are
local shortest paths. In Rn , geodesics are straight lines.
8. Exponential map: The map expp : Tp M → M is defined as
expp (v) = γv (1)
where γv is the unique geodesic starting at p with initial velocity γv′ (0) equal to v. The
exponential map takes a straight line tv ∈ Tp M to a geodesic γtv (1) = γv (t) ∈ M . Note that
expp maps v and tv to points on the same geodesic. Intuitively, the exponential map can be
thought as point-vector addition in a manifold. In Rn , we have expp (v) = p + v.
9. Parallel transport: Given any geodesic c(t) and a vector v such that hv, c′ (0)ic(0) = 0, we
define the parallel transport Γ of v along c(t) by the following process: Take h to be infinitesimally small and v0 = v. For i = 1, 2, · · · , 1/h, we let vih be the vector orthogonal to c′ (ih)
that minimizes the distance on the manifold between expc(ih) (hvih ) and expc((i−1)h) (hv(i−1)h ).
Intuitively, the parallel transport finds the vectors on the curve such that their end points are
closest to the end points of v. For general vector v ∈ Tc′ (0) , we write v = αc′ (0) + w and we
define the parallel transport of v along c(t) is the sum of αc′ (t) and the parallel transport of
w along c(t). For a non-geodesic curve, see the definition in Fact 66.
10. Orthonormal frame: Given vector fields v1 , v2 , · · · , vn on a subset of M , we call {vi }ni=1 is an
orthonormal frame if hvi , vj ix = δij for all x. Given a curve c(t) and an orthonormal frame at
c(0), we can extend it on the whole curve by parallel transport and it remains orthonormal
on the whole curve.
11. Directional derivatives and the Levi-Civita connection: For a vector v ∈ Tp M and a vector
field u in a neighborhood of p, let γv be the unique geodesic starting at p with initial velocity
γv′ (0) = v. Define
u(h) − u(0)
∇v u = lim
h→0
h
where u(h) ∈ Tp M is the parallel transport of u(γ(h)) from γ(h) to γ(0). Intuitively, LeviCivita connection is the directional derivative of u along direction v, taking the metric into
d
u(x + tv). When u is defined on a curve
account. In particular, for Rn , we have ∇v u(x) = dt
d
d
n
c, we define Dt u = ∇c′ (t) u. In R , we have Dt u(γ(t)) = dt
u(γ(t)). We reserve dt
for the usual
derivative with Euclidean coordinates.
We list some basic facts about the definitions introduced above that are useful for computation and
intuition.
Fact 66. Given a manifold M , a curve c(t) ∈ M , a vector v and vector fields u, w on M , we have
the following:
1. (alternative definition of parallel transport) v(t) is the parallel transport of v along c(t) if and
only if ∇c′ (t) v(t) = 0.
2. (alternative definition of geodesic) c is a geodesic if and only if ∇c′ (t) c′ (t) = 0.
3. (linearity) ∇v (u + w) = ∇v u + ∇v w.
4. (product rule) For any scalar-valued function f, ∇v (f · u) =
63
∂f
∂v u
+ f · ∇v u.
5. (metric preserving)
d
dt
hu, wic(t) = hDt u, wic(t) + hu, Dt wic(t) .
∂c
6. (torsion free-ness) For any map c(t, s) from a subset of R2 to M , we have that Ds ∂c
∂t = Dt ∂s
where Ds = ∇ ∂c and Dt = ∇ ∂c .
∂s
∂t
7. (alternative definition of Levi-Civita connection) ∇v u is the unique linear mapping from the
product of vector and vector field to vector field that satisfies (3), (4), (5) and (6).
D.1
Curvature
Roughly speaking, curvature measures the amount by which a manifold deviates from Euclidean
space. Given vector u, v ∈ Tp M , in this section, we define uv be the point obtained from moving
from p along direction u with distance kukp (using geodesic), then moving along direction “v” with
distance kvkp where “v” is the parallel transport of v along the path u. In Rn , uv is exactly p + u + v
and hence uv = vu, namely, parallelograms close up. For a manifold, parallelograms almost close
up, namely, d(uv, vu) = o(kuk kvk). This property is called being torsion-free.
1. Riemann curvature tensor: Three-dimensional parallelepipeds might not close up, and the
curvature tensor measures how far they are from closing up. Given vector u, v, w ∈ Tp M , we
define uvw as the point obtained by moving from uv along direction “w” for distance kwkp
where “w” is the parallel transport of w along the path uv. In a manifold, parallelepipeds do
not close up and the Riemann curvature tensor how much uvw deviates from vuw. Formally,
for vector fields v, w, we define τv w be the parallel transport of w along the vector field v for
one unit of time. Given vector field v, w, u, we define the Riemann curvature tensor by
R(u, v)w =
d d −1 −1
τ τ τsu τtv w
ds dt su tv
.
(D.1)
t,s=0
Riemann curvature tensor is a tensor, namely, R(u, v)w at point p depends only on u(p), v(p)
and w(p).
2. Ricci curvature: Given a vector v ∈ Tp M , the Ricci curvature Ric(v) measures if the geodesics
starting around p in direction v converge together. Positive Ricci curvature indicates the
geodesics converge while negative curvature indicates they diverge. Let S(0) be a small shape
around p and S(t) be the set of point obtained by moving S(0) along geodesics in the direction
v for t units of time. Then,
volS(t) = volS(0)(1 −
Formally, we define
Ric(v) =
t2
Ric(v) + smaller terms).
2
X
ui
(D.2)
hR(v, ui )ui , vi
where ui is an orthonormal basis of Tp M . Equivalently, we have Ric(v) = Eu∼N (0,I) hR(v, u)u, vi.
kvk2 .
For Rn , Ric(v) = 0. For a sphere in n + 1 dimension with radius r, Ric(v) = n−1
r2
Fact 67 (Alternative definition of Riemann curvature tensor). Given any M -valued function c(t, s),
∂c
we have vector fields ∂c
∂t and ∂s on M . Then, for any vector field z,
R(
∂c ∂c
, )z = ∇ ∂c ∇ ∂c z − ∇ ∂c ∇ ∂c z.
∂t
∂s
∂s
∂t
∂t ∂s
Equivalently, we write R(∂t c, ∂s c)z = Dt Ds z − Ds Dt z.
64
Fact 68. Given vector fields v, u, w, z on M ,
hR(v, u)w, zi = hR(w, z)v, ui = − hR(u, v)w, zi = − hR(v, u)z, wi .
D.2
Hessian manifolds
2
Recall that a manifold is called Hessian if it is a subset of Rn and its metric is given by gij = ∂x∂i ∂xj φ
for some smooth convex function φ. We let gij be entries of the inverse matrix of gij . For example,
P
2
∂3
we have j gij gjk = δik . We use φij to denote ∂x∂i ∂xj φ and φijk to denote ∂xi ∂x
j ∂xk φ.
Since a Hessian manifold is a subset of Euclidean space, we identify tangent spaces Tp M by
Euclidean coordinates. The following lemma gives formulas for the Levi-Civita connection and
curvature under Euclidean coordinates.
Lemma 69 ([28]). Given a Hessian manifold M , vector fields v, u, w, z on M , we have the following:
1. (Levi-Civita connection) ∇v u =
and the Christoffel symbol
P
∂uk
ik vi ∂xi ek
Γkij =
+
P
k
ijk vi uj Γij ek
where ek are coordinate vectors
1 X kl
g φijl .
2
l
2. (Riemann curvature tensor) hR(u, v)w, zi =
Rklij =
3. (Ricci curvature) Ric(v) =
1
4
P
P
ijlk
Rklij ui vj wl zk where
1 X pq
g (φjkp φilq − φikp φjlq ) .
4 pq
ijlkpq g
pq g jl
(φjkpφilq − φikp φjlq ) vi vk .
65
| 8cs.DS
|
arXiv:1607.03010v1 [math.GR] 11 Jul 2016
INTERSECTING FREE SUBGROUPS IN FREE PRODUCTS OF
LEFT ORDERED GROUPS
S. V. IVANOV
Abstract. A conjecture of Dicks and the author on rank of the intersection
of factor-free subgroups in free products of groups is proved for the case of left
ordered groups.
1. Introduction
Recall that the Hanna Neumann conjecture [16] claims that if F is a free group
of rank r(F ), r̄(F ) := max(r(F ) − 1, 0) is the reduced rank of F , and H1 , H2 are
finitely generated subgroups of F , then r̄(H1 ∩ H2 ) ≤ r̄(H1 )r̄(H2 ). It was shown
by Hanna Neumann [16] that r̄(H1 ∩ H2 ) ≤ 2r̄(H1 )r̄(H2 ). For more discussion and
results on this problem, theQreader is referred to [2], [3], [6], [15], [17], [18].
∗
More generally, let F = α∈I Gα be the free product of some groups Gα , α ∈ I.
According to the classic
theorem [13], every subgroup H of F is
Q∗Kurosh subgroup
−1
sα,β Hα,β sα,β
, where Hα,β is a subgroup of Gα , sα,β ∈ F,
a free product F (H) ∗
and F (H) is a free subgroup of F such that, for every s ∈ F and γ ∈ I, it is
true that F (H) ∩ sGγ s−1 = {1}. We say that H is a factor-free subgroup of F
if H = F (H) in the above form of H, i.e., for every s ∈ F and γ ∈ I, we have
H ∩ sGγ s−1 = {1}. Since a factor-free subgroup H of F is free, the reduced rank
r̄(H) := max(r(H) − 1, 0) of H is well defined. Let q ∗ = q ∗ (Gα , α ∈ I) denote the
minimum of orders > 2 of finite subgroups of groups Gα , α ∈ I, and q ∗ := ∞ if
∗
there are no such subgroups. If q ∗ = ∞, define q∗q−2 := 1. It was shown by Dicks
and the author [4] that if H1 , H2 are finitely generated factor-free subgroups of F,
then
∗
r̄(H1 ∩ H2 ) ≤ 2 q∗q−2 r̄(H1 )r̄(H2 ).
It was conjectured by Dicks and the author [4] that if groups Gα , α ∈ I, contain
no involutions then, similarly to the Hanna Neumann conjecture, the coefficient 2
could be left out and
∗
r̄(H1 ∩ H2 ) ≤ q∗q−2 r̄(H1 )r̄(H2 ).
(1)
A special case of this generalization of the Hanna Neumann conjecture is established by Dicks and the author [5] by proving that if H1 , H2 are finitely generated
factor-free subgroups of the free product F all of whose factors are groups of order
∗
3, then, indeed, r̄(H1 ∩ H2 ) ≤ q∗q−2 r̄(H1 )r̄(H2 ) = 3r̄(H1 )r̄(H2 ). We remark that it
follows from results of [4] that the last inequality (as well as (1)) is sharp and may
not be improved. Here is another special case when the conjectured inequality (1)
holds true.
2010 Mathematics Subject Classification. Primary 20E06, 20E07, 20F60, 20F65.
Supported in part by the NSF under grant DMS 09-01782.
1
2
S. V. IVANOV
Theorem
1. Suppose that Gα , α ∈ I, are left (or right) ordered groups, F =
Q∗
G
is
their free product, and H1 , H2 are finitely generated factor-free subα
α∈I
groups of F. Then
r̄(H1 ∩ H2 ) ≤
q∗
q∗ −2 r̄(H1 )r̄(H2 )
= r̄(H1 )r̄(H2 ).
Moreover, let S(H1 , H2 ) denote a set of representatives of those double cosets H1 tH2
of F, t ∈ F, that have the property H1 ∩ tH2 t−1 6= {1}. Then
X
r̄(H1 ∩ sH2 s−1 ) ≤ r̄(H1 )r̄(H2 ).
r̄(H1 , H2 ) :=
s∈S(H1 ,H2 )
We remark that Antolı́n, Martino, and Schwabrow [1] proved a more general
result on Kurosh rank of the intersection of subgroups of free products of right
ordered groups by utilizing the Bass–Serre theory of groups acting on trees and
some ideas of Dicks [3]. Our proof of Theorem 1 seems to be of independent
interest as it uses explicit geometric construction of graphs, representing subgroups
of free products, that are often more suitable for counting arguments, see [9], [11].
It is fairly easy to see that Theorem 1 implies both the Hanna Neumann conjecture and the strengthened Hanna Neumann conjecture, put forward by W. Neumann [17], see Sect. 5. The strengthened Hanna Neumann conjecture claims that
if H1 , H2 are finitely generated subgroups of a free group F , then
X
r̄(H1 ∩ sH2 s−1 ) ≤ r̄(H1 )r̄(H2 ),
s∈S(H1 ,H2 )
where the set S(H1 , H2 ) is defined as in Theorem 1 with F in place in F. Recall
that Friedman [6] proved the strengthened Hanna Neumann conjecture by making
use of sheaves on graphs and Mineyev [15] gave a proof to the strengthened Hanna
Neumann conjecture by using Hilbert modules and group actions, see also Dicks’s
proof [3]. Similarly to Dicks [3] and Mineyev [15], we also use the idea of group
ordering and special sets of edges, however, our arguments deal directly with core
graphs of subgroups of free products that are analogous to Stallings graphs [18]
representing subgroups of free groups.
In Sect. 2, we introduce necessary definitions and basic terminology. In Sect. 3,
we define and study strongly positive words in free products of left ordered groups.
Sect. 4 contains the proof of Theorem 1. In Sect. 5, we briefly look at the case of
free groups.
2. Preliminaries
Q
Let Gα , α ∈ I, be nontrivial groups, F = ∗α∈I Gα be their free product, and
H a finitely generated factor-free subgroup of F, H 6= {1}. Consider an alphabet
A = ∪α∈I Gα , where Gα ∩ Gα′ = {1} if α 6= α′ .
Similarly to the graph-theoretic approach of article [9], in a simplified version
suitable for finitely generated factor-free subgroups of F, see [7], [8], [10], we first
define a labeled A-graph Ψ(H) which geometrically represents H in a fashion analogous to the way the Stallings graph represents a subgroup of a free group, see
[18].
If Γ is a graph, V Γ denotes the vertex set of Γ and EΓ denotes the set of oriented
edges of Γ. If e ∈ EΓ, e− , e+ denote the initial, terminal, resp., vertices of e and e−1
is the edge with opposite orientation, where e−1 6= e for every e ∈ EΓ, (e−1 )− = e+ ,
(e−1 )+ = e− . A finite path p = e1 . . . ek in Γ is a sequence of edges ei such that
INTERSECTING FREE SUBGROUPS IN FREE PRODUCTS
3
(ei )+ = (ei+1 )− , i = 1, . . . , k − 1. Denote p− := (e1 )− , p+ := (ek )+ , and |p| := k,
where |p| is the length of p. We allow the possibility |p| = 0 and p = {p− } = {p+ }.
A finite path p is called closed if p− = p+ . An infinite path p = e1 . . . ek . . . is
an infinite sequence of edges ei such that (ei )+ = (ei+1 )− for all i = 1, 2, . . . . If
p = e1 . . . ek . . . and q = f1 . . . fℓ . . . are infinite paths such that (e1 )− = (f1 )− ,
then q −1 p := . . . fℓ−1 . . . f1−1 e1 . . . ek . . . is a biinfinite path. A path p is reduced if
p contains no subpath of the form ee−1 , e ∈ EΓ. A closed path p = e1 . . . ek is
cyclically reduced if |p| > 0 and both p and the cyclic permutation e2 . . . ek e1 of p
are reduced paths. The core of a graph Γ, denoted core(Γ), is the minimal subgraph
of Γ that contains every edge e which can be included into a cyclically reduced path
in Γ.
Let Ψ be a graph whose vertex set V Ψ consists of two disjoint parts VP Ψ, VS Ψ,
so V Ψ = VP Ψ ∪ VS Ψ. Vertices in VP Ψ are called primary and vertices in VS Ψ are
termed secondary. Every edge e ∈ EΨ connects primary and secondary vertices,
hence Ψ is a bipartite graph. Ψ is called a labeled A-graph, or simply A-graph if Ψ
is equipped with a map ϕ : EΨ → A, called labeling, such that, for every e ∈ EΨ,
ϕ(e) ∈ A = ∪α∈I Gα , ϕ(e−1 ) = ϕ(e)−1 and if e+ = f+ ∈ VS Ψ, then ϕ(e), ϕ(f ) ∈ Gα
for the same α = θ(e+ ) ∈ I, called the type of the vertex e+ ∈ VS Ψ and denoted
α = θ(e+ ). If e+ ∈ VS Ψ, define θ(e) := θ(e+ ) and θ(e−1 ) := θ(e+ ). Thus, for every
e ∈ Ψ, we have defined an element ϕ(e) ∈ A, called the label of e, and θ(e) ∈ I,
the type of e.
The reader familiar with van Kampen diagrams over a free product of groups,
see [14], will recognize that our labeling function ϕ : EΨ → A is defined in the
way analogous to labeling functions on van Kampen diagrams over free products of
groups. Recall that van Kampen diagrams are planar 2-complexes whereas graphs
are 1-complexes, however, apart from this, the ideas of cancelations and edge foldings work equally well for both diagrams and graphs.
An A-graph Ψ is called irreducible if properties (P1)–(P3) hold true:
(P1) If e, f ∈ EΨ, e− = f− ∈ VP Ψ, and e+ 6= f+ , then θ(e) 6= θ(f ).
(P2) If e, f ∈ EΨ, e 6= f , and e+ = f+ ∈ VS Ψ, then ϕ(e) 6= ϕ(f ) in Gθ(e) .
(P3) Ψ has no multiple edges, degΨ v > 0 for every v ∈ V Ψ, and there is at most
one vertex of degree 1 in Ψ which, if exists, is primary.
Suppose Ψ is a connected irreducible A-graph and a primary vertex o ∈ VP Ψ is
distinguished so that degΨ o = 1 if Ψ happens to have a vertex of degree 1. Then o
is called the base vertex of Ψ = Ψo .
Q
As usual, elements of the free product F = ∗α∈I Gα are regarded as words over
the alphabet A = ∪α∈I Gα . A syllable of a word W over A is a maximal subword
of W all of whose letters belong to the same factor Gα . The syllable length kW k
of W is the number of syllables of W , whereas the length |W | of W is the number
of all letters in W . A nonempty word W over A is called reduced if every syllable
of W consists of a single letter. Clearly, |W | = kW k if W is reduced. An arbitrary
nontrivial element of the free product F can be uniquely written as a reduced word.
A word W is called cyclically reduced if W 2 is reduced. We write U = W if words
U , W are equal as elements of F. The literal (or letter-by-letter) equality of words
U , W is denoted U ≡ W .
The significance of irreducible A-graphs for geometric interpretation of factorfree subgroups H of F is given in the following.
4
S. V. IVANOV
Lemma
Q∗ 1. Suppose H is a finitely generated factor-free subgroup of the free product
F = α∈I Gα , H 6= {1}. Then there exists a finite connected irreducible A-graph
Ψ = Ψo (H), with the base vertex o, such that a reduced word W over the alphabet A
belongs to H if and only if there is a reduced path p in Ψo (H) such that p− = p+ = o,
ϕ(p) = W in F, and |p| = 2|W |.
Proof. The proof is based on Stallings’s folding techniques and is somewhat analogous to the proof of van Kampen lemma for diagrams over free products of groups,
see [14] (in fact, it is simpler because foldings need not preserve the property of
being planar for the diagram). A more general approach, suitable for an arbitrary
subgroup of F, is discussed in Lemmas 1, 4 [9].
Let H = hV1 , . . . , Vk i be generated by reduced words V1 , . . . , Vk ∈ F. Consider
e which consists of k closed paths p1 , . . . , pk such that they have a single
a graph Ψ
common vertex o = (pi )− , and |pi | = 2|Vi |, i = 1, . . . , k. Furthermore, we dise and call o primary, the vertices adjacent to o
tinguish o as the base vertex of Ψ
are called secondary vertices and so on. The labeling function ϕ on pi is defined
so that ϕ(pi ) = Vi , i = 1, . . . , k, where ϕ(p) := ϕ(e1 ) . . . ϕ(eℓ ) if p = e1 . . . eℓ and
e
e1 , . . . , eℓ ∈ E Ψ.
e
e o is a finite connected A-graph with the base vertex o that has
Clearly, Ψ = Ψ
the following property
e o such that
(Q) A word W ∈ F belongs to H if and only if there is a path p in Ψ
p− = p+ = o and ϕ(p) = W .
e o need not be irreducible and we will do foldings of edges in Ψ
e o which
However, Ψ
preserve property (Q) and which aim to achieve properties (P1)–(P2).
e o so that
Assume that property (P1) fails for edges e, f with e− = f− ∈ VP Ψ
′
e+ 6= f+ and θ(e) = θ(f ). Let us redefine the labels of all edges e with e′+ = e+ so
that ϕ(e′ )ϕ(e)−1 does not change and ϕ(e) = ϕ(f ) in Gθ(e) . Now we identify the
edges e, f and vertices e+ , f+ . Observe that this folding preserves property (Q)
e o |.
((P2) might fail) and decreases the edge number |E Ψ
If property (P2) fails for edges e, f and ϕ(e) = ϕ(f ) in Gθ(e) , then we identify the
e o|
edges e, f . Note property (Q) still holds ((P1) might fail) and the number |E Ψ
decreases.
e o such that
Suppose property (P3) fails and there are two distinct edges e, f in Ψ
e o . By property (Q), it follows from H being factor-free
e− = f− , e+ = f+ ∈ VS Ψ
that ϕ(e) = ϕ(f ) in Gθ(e) . Therefore, we can identify the edges e, f , thus preserving
e o |. If property (P3) fails so that there
property (Q) and decreasing the number |E Ψ
is a vertex v of degree 1, different from o, then we delete v along with the incident
e o | decreases.
edge. Clearly, property (Q) still holds and the number |E Ψ
e o |, in polynomial time of size of input, which is the
Thus, by induction on |E Ψ
Pk
total length i=1 |Vi |, we can effectively construct an irreducible A-graph Ψo with
property (Q). Other stated properties of Ψo are straightforward.
The following lemma further elaborates on the correspondence between finitely
generated factor-free subgroups of the free product F and irreducible A-graphs.
Lemma 2. Let Ψo be a finite connected irreducible A-graph with the base vertex
o, and H = H(Ψo ) be a subgroup of F that consists of all words ϕ(p), where p
INTERSECTING FREE SUBGROUPS IN FREE PRODUCTS
5
is a path in Ψo with p− = p+ = o. Then H is a factor-free subgroup of F and
r̄(H) = −χ(Ψo ), where χ(Ψo ) = |V Ψo | − 21 |EΨo | is the Euler characteristic of Ψo .
Proof. This follows from the facts that the fundamental group π1 (Ψo , o) of Ψo at
o is free of rank −χ(Ψo ) + 1 and that the homomorphism π1 (Ψo , o) → F, given by
p → ϕ(p), where p is a path with p− = p+ = o, has the trivial kernel following from
properties (P1)–(P2).
Suppose
H is a nontrivial finitely generated factor-free subgroup of a free product
Q∗
F = α∈I Gα , and Ψo = Ψo (H) is an irreducible A-graph for H as in Lemma 1.
Let Ψ(H) denote the core of Ψo (H). Clearly, Ψ(H) has no vertices of degree ≤ 1
and Ψ(H) is also an irreducible A-graph. It is easy to see that the graph Ψo (H)
of H can be obtained back from the core graph Ψ(H) of H by attaching a suitable
path p to Ψ(H) so that p starts at a primary vertex o, ends in p− ∈ VP Ψ(H), and
then by doing foldings of edges as in the proof of Lemma 1.
Now suppose H1 , H2 are nontrivial finitely generated factor-free subgroups of
F. Consider a set S(H1 , H2 ) of representatives of those double cosets H1 tH2 of
F, t ∈ F, that have the property H1 ∩ tH2 t−1 6= {1}. For every s ∈ S(H1 , H2 ),
define the subgroup Ks := H1 ∩ sH2 s−1 . Analogously to the case of free groups, see
[17], [18], we now construct a finite irreducible A-graph Ψ(H1 , H2 ) whose connected
components are core graphs Ψ(Ks ), s ∈ S(H1 , H2 ).
First we define an A-graph Ψ′o (H1 , H2 ). The set of primary vertices of Ψ′o (H1 , H2 )
is VP Ψ′o (H1 , H2 ) := VP Ψo1 (H1 ) × VP Ψo2 (H2 ). Let
τi : VP Ψ′o (H1 , H2 ) → VP Ψoi (Hi )
denote the projection map, τi ((v1 , v2 )) = vi , i = 1, 2.
The set of secondary vertices VS Ψ′o (H1 , H2 ) of Ψ′o (H1 , H2 ) consists of equivalence
classes [u]α , where u ∈ VP Ψ′o (H1 , H2 ), α ∈ I, with respect to the minimal equivaα
α
lence relation generated by the following relation ∼ on VP Ψ′o (H1 , H2 ): Define v ∼ w
if there are edges ei , fi ∈ EΨoi (Hi ) such that (ei )− = τi (v), (fi )− = τi (w), (ei )+ =
(fi )+ , i = 1, 2, the edges ei , fi have type α, and ϕ(e1 )ϕ(f1 )−1 = ϕ(e2 )ϕ(f2 )−1 in
α
Gα . It is easy to see that ∼ is symmetric and transitive on distinct pairs and triples
(but it could lack reflexive property).
The edges in Ψ′o (H1 , H2 ) are defined so that u ∈ VP Ψ′o (H1 , H2 ) and [v]α ∈
VS Ψ′o (H1 , H2 ) are connected by an edge if and only if u ∈ [v]α .
The type of a vertex [v]α ∈ VS Ψ′o (H1 , H2 ) is α and if e ∈ EΨ′o (H1 , H2 ), e− = u,
e+ = [v]α , then ϕ(e) := ϕ(e1 ), where e1 ∈ EΨo1 (H1 ) is an edge of type α with
(e1 )− = τ1 (u), when such an e1 exists, and ϕ(e1 ) := bα 6= 1, bα ∈ Gα , otherwise.
It follows from the definitions and properties (P1)–(P2) of Ψoi (Hi ), i = 1, 2,
that Ψ′o (H1 , H2 ) is an A-graph with properties (P1)–(P2). Hence, taking the core
of Ψ′o (H1 , H2 ), we obtain an irreducible A-graph which we denote Ψ(H1 , H2 ).
In addition, it is not difficult to see that, when taking the connected component
Ψ′o (H1 , H2 , o) of Ψ′o (H1 , H2 ) that contains the vertex o = (o1 , o2 ) and inductively
removing from Ψ′o (H1 ∩ H2 , o) vertices of degree 1 different from o, we obtain an
irreducible A-graph Ψo (H1 ∩ H2 ) with the base vertex o that corresponds to the
intersection H1 ∩ H2 as in Lemma 1.
Observe that it follows from the definitions and property (P1) for Ψ(Hi ), i = 1, 2,
that, for every edge e ∈ EΨ(H1 , H2 ) with e− ∈ VP Ψ(H1 , H2 ), there are unique
edges ei ∈ EΨ(Hi ) such that τi (e− ) = (ei )− , i = 1, 2. Hence, by setting τi (e) = ei ,
6
S. V. IVANOV
τi (e+ ) = (ei )+ , i = 1, 2, we extend τi to the graph map
τi : Ψ(H1 , H2 ) → Ψ(Hi ),
i = 1, 2.
It follows from definitions that τi is locally injective and τi preserves syllables of
ϕ(p) for every path p with primary vertices p− , p+ .
Lemma 3. Suppose H1 , H2 are finitely generated factor-free subgroups of the free
product F and the set S(H1 , H2 ) is not empty. Then the connected components of
the graph Ψ(H1 , H2 ) are core graphs Ψ(H1 ∩ sH2 s−1 ) of subgroups H1 ∩ sH2 s−1 ,
s ∈ S(H1 , H2 ). In particular,
X
r̄(H1 ∩ sH2 s−1 ) = −χ(Ψ(H1 , H2 )).
r̄(H1 , H2 ) =
s∈S(H1 ,H2 )
Proof. As in Lemma 1, let Ψoi (Hi ) be an irreducible A-graph, corresponding to
the subgroup Hi of F, i = 1, 2, and Ψ(Hi ) denote the core of Ψoi (Hi ). Let
vi ∈ VP Ψ(Hi ), i = 1, 2, and q(vi ) denote a path in Ψoi (Hi ) with q(vi )− = oi
and q(vi )+ = vi . Suppose Xi ∈ F, i = 1, 2, and H1X1 ∩ H2X2 6= {1}, where
HiXi := Xi Hi Xi−1 . Consider an irreducible A-graph Ψui (HiXi ), i = 1, 2. Note
that the core graph Ψ(H1X1 ∩ H2X2 ) can be identified with a connected component,
denoted Ψ(X1 ,X2 ) (H1 , H2 ), of the irreducible A-graph Ψ(H1 , H2 ). In addition, if
w ∈ VP Ψ(X1 ,X2 ) (H1 , H2 ), then there are paths pi (w) in Ψui (HiXi ), i = 1, 2, such
that (pi (w))− = ui , (pi (w))+ = τi (w), and ϕ(p1 (w)) = ϕ(p2 (w)). Furthermore, it
follows from the definitions that Xi ϕ(q(τi (w)))ϕ(pi (w))−1 ∈ HiXi , i = 1, 2. Therefore, there are words Vi ∈ Hi , i = 1, 2, such that Xi Vi ϕ(q(τi (w))) = ϕ(pi (w)).
Since ϕ(p1 (w)) = ϕ(p2 (w)), we further obtain
X1−1 X2 = V1 ϕ(q(τ1 (w)))ϕ(q(τ2 (w)))−1 V2−1 .
(2)
Now we can draw the following conclusion. For every pair (X1 , X2 ) ∈ F × F
such that H1X1 ∩ H2X2 6= {1} and a vertex w ∈ VP Ψ(X1 ,X2 ) (H1 , H2 ), there are
words Vi ∈ Hi , i = 1, 2, such that the equality (2) holds. Since the paths q(τi (w)),
i = 1, 2, in (2) depend only on a connected component of Ψ(H1 , H2 ), it follows from
(2) that if (1, X), (1, Y ) are some pairs such that Ψ(1,X) (H1 , H2 ) = Ψ(1,Y ) (H1 , H2 ),
then X ∈ H1 Y H2 ⊆ F.
Conversely, if X ∈ H1 Y H2 , then the equality Ψ(1,X) (H1 , H2 ) = Ψ(1,Y ) (H1 , H2 )
is obviously true. Thus, the set S(H1 , H2 ) is in bijective correspondence with
connected components of Ψ(H1 , H2 ) and, by Lemma 2, we have r̄(H1 ∩ sH2 s−1 ) =
−χ(Ψ(1,s) (H1 , H2 )) for every s ∈ S(H1 , H2 ). Adding up over all s ∈ S(H1 , H2 ), we
arrive to the required equality r̄(H1 , H2 ) = −χ(Ψ(H1 , H2 )).
3. Strongly positive words in free products of left ordered groups
Recall that G is called a left ordered group if G is equipped with a total order
≤ which is left invariant, i.e., for every triple a, b, c ∈ G, the relation a ≤ b implies
ca ≤ cb. If G is left ordered, then G can also be right ordered (and vice versa).
Indeed, if ≤ is a left order on G then, setting a b if and only if a−1 ≤ b−1 , we
obtain a right order on G.
Q
Let Gα , α ∈ I, be nontrivial left (or right) ordered groups and let F = ∗α∈I Gα
be their free product. Since it will be more convenient to work with left order, we
assume that Gα , α ∈ I, are left ordered. It is well known, see [12], and fairly easy
INTERSECTING FREE SUBGROUPS IN FREE PRODUCTS
7
to show that there exists a total order on F which extends the left orders on
groups Gα , α ∈ I, and which turns F into a left ordered group.
A reduced word W ∈ F is called positive if W ≻ 1. A reduced word W is called
strongly positive, denoted W ≻≻ 1, if every nonempty suffix of W is positive, i.e.,
if W ≡ W1 W2 with |W2 | > 0, then W2 ≻ 1. Clearly, a strongly positive word is
positive. Note if U, W are strongly positive and U W is reduced, then U W ≻≻ 1. A
word U is (resp. strongly) negative if U −1 is (resp. strongly) positive.
Lemma 4. Suppose S, T are strongly positive words and the word S −1 T is reduced.
Then S −1 T is either strongly positive or strongly negative.
Proof. Let S ≡ S1 S2 , T ≡ T1 T2 , where |S2 |, |T2 | > 0. Then S2 ≻ 1, T2 ≻ 1
by S, T ≻≻ 1. Since S −1 T is reduced, we have S −1 T 6= 1, hence S −1 T ≻ 1 or
S −1 T ≺ 1. Assume S −1 T ≻ 1. Then T ≻ S = S1 S2 or S1−1 T ≻ S2 ≻ 1, hence, in
view of T2 ≻ 1, all nonempty suffixes of S −1 T are positive. This implies S −1 T ≻≻ 1.
If S −1 T ≺ 1, then, switching S and T , we can show as above that T −1 S ≻≻ 1, hence
S −1 T is strongly negative.
Lemma 5. Suppose W is a reduced word. Then there exists a factorization W ≡
U1 U2−1 such that |U1 |, |U2 | ≥ 0 and each of U1 , U2 is either empty or strongly
positive.
Proof. Consider a factorization W ≡ U1ε1 . . . Ukεk , where, for every j, Uj ≻≻ 1 and
εj = ±1, that would be minimal with respect to k. Since for every letter a of W
either a ≻≻ 1 or a−1 ≻≻ 1, it follows that such a factorization exists and k ≤ |W |.
Note if εj = εj+1 = 1, then Uj Uj+1 is reduced and so Uj Uj+1 ≻≻ 1. Similarly, if
−1
εj = εj+1 = −1, then Uj−1 Uj+1
is reduced and so Uj+1 Uj ≻≻ 1. Hence, it follows
from the minimality of k that εj 6= εj+1 for all j = 1, . . . , k − 1. If now εj = −1
and εj+1 = 1 for some j, then we can use Lemma 4 and conclude that either
−1
Uj−1 Uj+1 ≻≻ 1 or Uj+1
Uj ≻≻ 1, contrary to minimality of k. Thus, it is proven that
either k = 1 or k = 2 and ε1 = 1, ε2 = −1, as required.
Lemma 6. Suppose W is a cyclically reduced word. Then there exists a factorization W ≡ W1 W2 such that the cyclic permutation W̄ ≡ W2 W1 of W is either
strongly positive or strongly negative.
Proof. By Lemma 5, W ≡ U1 U2−1 , where |U1 |, |U2 | ≥ 0 and Uj ≻≻ 1 if |Uj | > 0, j =
1, 2. Since W 2 is reduced, U2−1 U1 is reduced and, by Lemma 4, either U2−1 U1 ≻≻ 1
or U2−1 U1 ≺≺ 1. Hence, W̄ ≡ U2−1 U1 is a desired cyclic permutation of W .
4. Proving Theorem 1
Let Gα , α ∈ I, be nontrivial left (or right) ordered groups and let F be their free
product equipped with a left order . Also, fix a total order ≤ on the index set I.
Let Ψ be a finite irreducible A-labeled graph, where A = ∪α∈I Gα . An edge
e ∈ EΨ is called maximal if there are reduced infinite paths p = p(e) = e1 e2 . . . ,
q = q(e) = f1 f2 . . . in Ψ, where ej , fj ∈ EΨ, such that e = e1 , (e1 )− = (f1 )−
is primary, θ(e1 ) > θ(f1 ), and, for every j ≥ 1, both ϕ(e1 . . . e2j ) ≺ 1 and
ϕ(f1 . . . f2j ) ≺ 1. Note that the vertices (e1 . . . e2j )+ , (f1 . . . f2j )+ are primary
and q −1 p = . . . f2−1 f1−1 e1 e2 . . . is a reduced biinfinite path.
Lemma 7. Suppose Ψ is a finite irreducible A-labeled graph whose Euler characteristic is negative, χ(Ψ) < 0. Then Ψ contains a maximal edge.
8
S. V. IVANOV
Proof. Since χ(Ψ) < 0, Ψ has a connected component Ψ1 with χ(Ψ1 ) < 0. Without
loss of generality, we may assume that core (Ψ1 ) = Ψ1 . It is not difficult to see from
χ(Ψ1 ) < 0 and from core (Ψ1 ) = Ψ1 that, for every pair h, h′ ∈ EΨ1 , there is a
reduced path p = h . . . h′ whose first, last edges are h, h′ , resp.. Pick a primary
vertex o in Ψ1 and two distinct edges t1 , u1 with (t1 )− = (u1 )− = o. Let t, u be
some reduced paths such that first edges of t, u are t1 , u1 , resp., and t+ , u+ have
degree > 2. Then it follows from the above remark that there are closed paths
r0 , s0 starting at t+ , u+ , resp., such that the path tr02 t−1 us20 u−1 is reduced. Since
Ψ1 is irreducible and r0 , s0 are reduced, it follows ϕ(r0 ) 6= 1, ϕ(s0 ) 6= 1 in F.
Let r, s be some cyclic permutations of the closed paths r0 , s0 , resp., that start
at some primary vertices and R = ϕ(r), S = ϕ(s) be reduced words. Clearly, R, S
are cyclically reduced and |R| = |r|/2 > 1, |S| = |s|/2 > 1. By Lemma 6, there
are cyclic permutations R̄, S̄ of R, S, resp., such that R̄εr , S̄ εs are strongly positive,
where εr , εs ∈ {±1}. Switching from r0 , s0 to r0−1 , s−1
0 , resp., if necessary, we may
assume that εr = εs = −1, i.e., R̄−1 , S̄ −1 ≻≻ 1. Let r̄, s̄ denote cyclic permutations
of r, s, resp., such that ϕ(r̄) = R̄, ϕ(s̄) = S̄. Also, let r̄ = r̄1 r̄2 , s̄ = s̄1 s̄2 be
factorizations of r̄, s̄, resp., defined by vertices t+ , u+ , resp..
Consider two infinite paths starting at o = t− = u− and defined as follows. Let
T = tr0+∞ whose prefixes are tr0k , k ≥ 0, and U = us+∞
whose prefixes are usℓ0 ,
0
ℓ ≥ 0. It follows from the definitions that T starts at t− = o, goes along t to t+ and
then cycles around r0 infinitely many times, in particular, T is reduced. Similarly,
U starts at u− = o, goes along u to u+ and then cycles around s0 .
Denote T = t1 t2 . . . , where tj ∈ EΨ1 , and U = u1 u2 . . . , where uk ∈ EΨ1 .
Let T (j1 , j2 ) := tj1 . . . tj2 , where j1 ≤ j2 , denote the subpath of T that starts at
(tj1 )− and ends in (tj2 )+ . It is convenient to set T (j, j − 1) := {(tj )− } for all j ≥ 1.
Similarly, U (j1 , j2 ) := uj1 . . . uj2 , where j1 ≤ j2 , and U (j, j − 1) := {(uj )− } if j ≥ 1.
Suppose 2j > |t| + |r̄2 |. Then 2j − |t| − |r̄2 | > 0. Let m be the remainder of
2j − |t| − |r̄2 | when divided by |r|. Set mr := m if m > 0 and mr := |r| if m = 0.
Note T (1, 2j) = T (1, 2j − mr )T (2j − mr + 1, 2j) and ϕ(T (2j − mr + 1, 2j)) ≡ R̄3 ,
where R̄3 is a prefix of ϕ(r̄) ≡ R̄3 R̄4 of even length mr > 0. Recall ϕ(r̄) = R̄ and
R̄−1 ≻≻ 1, hence R̄3−1 ≻ 1 and R̄3 ≺ 1. Note R̄3 = ϕ(T (2j − mr + 1, 2j)) ≺ 1
implies, by left invariance of the order , that
ϕ(T (1, 2j)) ≺ ϕ(T (1, 2j − mr )).
(3)
′
Now suppose 2j > |u| + |s̄2 |. Let m be the remainder of 2j − |t| − |s̄2 | > 0 when
divided by |s|. Set ms := m′ if m′ > 0 and ms := |s| if m′ = 0. Then we can derive
from S̄ −1 ≻≻ 1, similar to (3), that
ϕ(U (1, 2j)) ≺ ϕ(U (1, 2j − ms )).
(4)
The comparisons (3)–(4) prove that a maximal element of the infinite set
{ϕ(T (1, 2j)), ϕ(U (1, 2k)) | j ≥ 0, k ≥ 1} ⊆ F
(5)
exists and it is the maximal element of the finite set
{ ϕ(T (1, 2j)), ϕ(U (1, 2k)) | 0 ≤ 2j ≤ |t| + |r̄2 |, 0 < 2k ≤ |u| + |s̄2 | } ⊂ F.
Let ϕ(Q(1, 2jM )), where jM ≥ 0, Q ∈ {T, U }, denote the maximal element of
the set (5). Note vM = Q(1, 2jM )+ is primary. Observe that elements ϕ(T (1, 2j)),
ϕ(U (1, 2k)), j ≥ 0, k ≥ 1, in (5) are distinct and represent ϕ-labels of subpaths of
the biinfinite path U −1 T that connect the primary vertex o = t− to all primary
vertices of U −1 T along U −1 T (or its inverse). If we take another primary vertex v on
INTERSECTING FREE SUBGROUPS IN FREE PRODUCTS
9
U −1 T and consider the set of labels of subpaths that connect v to primary vertices
of U −1 T as above, then the resulting set can be obtained from (5) by multiplication
on the left by ϕ(h(o, v))−1 , where h(o, v) = T (1, 2jv ) if v = T (1, 2jv )+ , 2jv ≥ 0,
and h(o, v) = U (1, 2kv ) if v = U (1, 2kv )+ , 2kv > 0. Since left multiplication
preserves the order, these remarks imply that the vertex vM = Q(1, 2jM )+ defines
a factorization of the biinfinite path U −1 T = q −1 p into infinite paths q, p, where
p = e1 e2 . . . , q = f1 f2 . . . , ej , fj are edges, j ≥ 1, so that, for every j ≥ 1, we have
ϕ(e1 e2 . . . e2j ) ≺ 1 and ϕ(f1 f2 . . . f2j ) ≺ 1. Therefore, if θ(e1 ) > θ(f1 ), then e1 is a
maximal edge of Ψ and if θ(f1 ) > θ(e1 ), then f1 is maximal in Ψ.
Suppose Γ is a finite graph. A set D ⊆ EΓ of edges is called good (for cutting)
if the graph Γ \ (D ∪ D−1 ) consists of connected components whose Euler characteristics are 0. Clearly, Γ contains a good edge set if and only if no connected
component of Γ is a tree.
Lemma 8. Suppose Ψ is a finite connected irreducible A-graph with χ(Ψ) < 0.
Then the set of all maximal edges of Ψ is good (for cutting).
Proof. Arguing on the contrary, assume that D ⊆ EΨ is the set of all maximal
edges of Ψ and D is not good. Then the graph Ψ \ (D ∪ D−1 ) contains a connected
component Ψ′1 with either χ(Ψ′1 ) < 0 or χ(Ψ′1 ) > 0. If χ(Ψ′1 ) < 0, then the core
Ψ1 = core(Ψ′1 ) of Ψ′1 is a finite irreducible A-graph with χ(Ψ1 ) < 0. By Lemma 7,
Ψ1 contains a maximal edge e. However, it follows from the definition that e is also
maximal for Ψ, hence, e ∈ D. This contradiction shows that χ(Ψ′1 ) > 0, hence Ψ′1
is a tree which we denote T .
Let C denote the set that consists of all c ∈ EΨ such that c+ ∈ V T and c 6∈ ET .
It follows from the definitions that if c ∈ C then c ∈ D or c−1 ∈ D. Since every
d ∈ D is maximal, there are infinite reduced paths p(d) = e1 (d)e2 (d) . . . and q(d) =
f1 (d)f2 (d) . . . such that e1 (d)− = f1 (d)− ∈ VP Ψ, e1 (d) = d, θ(e1 (d)) > θ(f1 (d))
and, for every j ≥ 1, ϕ(e1 (d) . . . e2j (d)) ≺ 1 and ϕ(f1 (d) . . . f2j (d)) ≺ 1.
Pick an arbitrary c ∈ C. Suppose c− is primary. Since d− ∈ VP Ψ if d ∈ D
and c or c−1 is in D, it follows that c ∈ D. Consider a shortest path of the form
h(c) := e1 (c) . . . e2ℓ (c), ℓ ≥ 1, such that either e2ℓ (c)−1 ∈ C or e2ℓ (c)−1 ∈ T and
e2ℓ+1 (c)−1 ∈ C. Define σ(c) := e2ℓ (c)−1 if e2ℓ (c)−1 ∈ C and σ(c) := e2ℓ+1 (c)−1
if e2ℓ+1 (c)−1 ∈ C. Since T is a finite tree, such a path h(c) exists, |h(c)| > 0,
σ(c) 6= c, and ϕ(h(c)) ≺ 1. Note h(c)− , h(c)+ are primary vertices of c, σ(c),
resp., and h(c) = chT (c)σ(c)−εσ(c) , where hT (c) is a subpath of h(c) in T with
hT (c)− = c+ , hT (c)+ = σ(c)+ , and εσ(c) = 1 if σ(c)+ is secondary and εσ(c) = 0 if
σ(c)+ is primary.
Now assume that c+ is primary. Then c−1 = dc ∈ D. Consider a shortest path
of the form h(c) := f1 (dc ) . . . f2ℓ−2 (dc ), where ℓ ≥ 1 and if ℓ = 1 then h(c) := {c+ },
such that either f2ℓ−2 (dc )−1 ∈ C or f2ℓ−2 (dc )−1 ∈ T (or f2ℓ−2 (dc ) is undefined if
ℓ = 1) and f2ℓ−1 (dc )−1 ∈ C. Define σ(c) := f2ℓ−2 (dc )−1 if f2ℓ−2 (dc )−1 ∈ C and
σ(c) := f2ℓ−1 (dc )−1 if f2ℓ−1 (dc )−1 ∈ C. Since T is a finite tree, such a path h(c)
exists, |h(c)| ≥ 0, σ(c) 6= c, and ϕ(h(c)) 1. In addition, the equality ϕ(h(c)) = 1
implies that h(c) = {c+ }, σ(c) = f1 (c)−1 and θ(σ(c)) = θ(f1 (dc )) < θ(e1 (dc )) =
θ(c). As above, we remark that h(c)− , h(c)+ are primary vertices of c, σ(c), resp.,
and h(c) = hT (c)σ(c)−εσ(c) , where hT (c) is in T with hT (c)− = c+ , hT (c)+ = σ(c)+ ,
and εσ(c) = 1 if σ(c)+ is secondary and εσ(c) = 0 if σ(c)+ is primary.
10
S. V. IVANOV
Let us summarize. For every c ∈ C, we have defined an edge σ(c) ∈ C, σ(c) 6= c,
hence, σ : C → C is a function. Furthermore, there is a reduced path h(c) such that
h(c) = cεc hT (c)σ(c)−εσ(c) , where hT (c) is in T with hT (c)− = c+ , hT (c)+ = σ(c)+ ,
εc = 1 if c+ ∈ VS Ψ and εc = 0 if c+ ∈ VP Ψ. In addition, εσ(c) = 1 if σ(c)+ ∈ VS Ψ
and εσ(c) = 0 if σ(c)+ ∈ VP Ψ. Also, ϕ(h(c)) 1 and ϕ(h(c)) = 1 implies that
h(c) = {c+ } = {σ(c)+ } and θ(c) > θ(σ(c)). Finally, h(c)− , h(c)+ are primary
vertices of c, σ(c), resp., whence h(c)+ = h(σ(c))− for every c ∈ C.
Since C is finite, there is a cycle c, σ(c), . . . , σ k (c) = c, k ≥ 2, for some c ∈ C.
Consider the closed path pc = h(c)h(σ(c)) . . . h(σ k−1 (c)). Since ϕ(h(σ j (c))) 1 for
every j, we obtain that ϕ(pc ) 1 and the equality ϕ(pc ) = 1 implies ϕ(h(σ j (c))) =
1 and h(σ j (c)) = {σ j (c)+ } for every j. On the other hand,
pc = cεc hT (c)σ(c)−εσ(c) σ(c)εσ(c) hT (σ(c))σ 2 (c)−εσ2 (c) . . .
σ k−1 (c)εσk−1 (c) hT (σ k−1 (c))σ k (c)−εσk (c) = cεc hT (c)hT (σ(c)) . . . hT (σ k−1 (c))c−εc
following from σ k (c) = c. Since hT (c)hT (σ(c)) . . . hT (σ k−1 (c)) is a closed path in
the tree T , we have ϕ(pc ) = 1 in F. Therefore, h(σ j (c)) = {σ j (c)+ } = {σ(c)+ } and
θ(σ j (c)) > θ(σ j+1 (c)) for every j = 0, 1, . . . , k − 1, implying θ(c) > θ(σ k (c)) = θ(c).
This contradiction completes the proof.
Proof of Theorem 1. As in Sect. 2, consider a finite irreducible A-graph Ψ(H1 , H2 )
whose connected components correspond to core graphs of subgroups H1 ∩ sH2 s−1 ,
s ∈ S(H1 , H2 ). Without loss of generality, we may assume that −χ(Ψ(H1 , H2 )) > 0.
Let D be the set of all maximal edges in Ψ(H1 , H2 ). It is easy to see from the definitions that if d ∈ D, then τi (d) is maximal in Ψoi (Hi ), i = 1, 2. Hence, τi (D) ⊆ Di ,
where Di is the set of maximal edges of Ψoi (Hi ), i = 1, 2. By Lemma 8, Di is good
for Ψoi (Hi ) and it follows from Lemma 3 and definitions that
r̄(H1 , H2 ) = |D| ≤ |τ1 (D)| · |τ2 (D)| ≤ |D1 | · |D2 | = r̄(H1 ) · r̄(H2 ),
as desired.
5. The Free Group Case
Suppose H1 , H2 are finitely generated subgroups of a free group F = F (A),
where A = {a1 , . . . , am } is a set of free generators of F . Let F (a, b) be a free
group of rank 2 with free generators a, b. Note that the map µ : ai → ai bi a−i b−i ,
i = 1, . . . , m, extends to a monomorphism µ : F (A) → F (a, b) such that µ(H1 ),
µ(H2 ) are factor-free subgroups of the free product F (a, b) = A ∗ B, where A =
hai, B = hbi are infinite cyclic groups generated by a, b. We may assume that
µ(S(H1 , H2 )) ⊆ S(µ(H1 ), µ(H2 )). Since a cyclic group is left ordered, it follows
from Theorem 1 that
X
X
r̄(µ(H1 ) ∩ tµ(H2 )t−1 )
r̄(H1 ∩ sH2 s−1 ) ≤
r̄(H1 , H2 ) =
s∈S(H1 ,H2 )
t∈S(µ(H1 ),µ(H2 ))
≤ r̄(µ(H1 ))r̄(µ(H2 )) = r̄(H1 )r̄(H2 ).
(6)
We remark that there is a more direct way to prove the inequality (6) by repeating
verbatim the arguments of Sect. 4 with a few changes in basic definitions. To do this,
±1 ±1
consider a graph U with V U = {oP , oS } and EU = {a±1
1 , a2 , a3 }, where (aj )− =
oP and (aj )+ = oS , j = 1, 2, 3. The fundamental group F2 = π1 (U, oP ) of U at
−1
−1
−1
oP is free of rank 2, and F2 = ha1 a−1
2 , a1 a3 i is freely generated by a1 a2 , a1 a3 .
Let H1 , H2 be finitely generated subgroups of F2 , Xi be the Stallings graph of
INTERSECTING FREE SUBGROUPS IN FREE PRODUCTS
11
Hi , i = 1, 2, and W be the core of the pullback X1 × X2 of X1 , X2 over U , see
U
[18]. If Q ∈ {X1 , X2 , W, U }, there is a canonical graph map ϕQ : Q → U which is
locally injective and which we call labeling. If v ∈ V Q and ϕQ (v) = oP , v is called
primary. If ϕQ (v) = oS , v is secondary. The image ϕQ (e) = a±1
is the label of an
j
edge e ∈ EQ and θ(e) := j ∈ {1, 2, 3} = I is the type of e. With this terminology,
the definitions and arguments of Sect. 4 for graphs Q, W , X1 , X2 and the group
F2 in place of Ψ, Ψ(H1 , H2 ), Ψo1 (H1 ), Ψo2 (H2 ) and F, resp., are retained.
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
Y. Antolı́n, A. Martino, and I. Schwabrow, Kurosh rank of intersections of subgroups of free
products of right-orderable groups, preprint, http://arxiv.org/abs/1109.0233v3
W. Dicks, Equivalence of the strengthened Hanna Neumann conjecture and the amalgamated
graph conjecture, Invent. Math. 117(1994), 373–389.
W. Dicks, Simplified Mineyev, preprint, http://mat.uab.cat/∼dicks/SimplifiedMineyev.pdf
W. Dicks and S. V. Ivanov, On the intersection of free subgroups in free products of groups,
Math. Proc. Cambridge Phil. Soc. 144(2008), 511–534.
W. Dicks and S. V. Ivanov, On the intersection of free subgroups in free products of groups
with no 2-torsion, Illinois J. Math. 54(2010), 223–248.
J. Friedman, Sheaves on graphs, their homological invariants, and a proof of the Hanna Neumann conjecture: with an appendix by Warren Dicks, Mem. Amer. Math. Soc. 233(2014),
no. 1100. xii+106 pp.
S. V. Ivanov, On the intersection of finitely generated subgroups in free products of groups,
Internat. J. Algebra and Comp. 9(1999), 521–528.
S. V. Ivanov, Intersecting free subgroups in free products of groups, Internat. J. Algebra
and Comp. 11(2001), 281–290.
S. V. Ivanov, On the Kurosh rank of the intersection of subgroups in free products of groups,
Adv. Math. 218(2008), 465–484.
S. V. Ivanov, A property of groups and the Cauchy-Davenport theorem, J. Group Theory
13(2010), 21–39.
S. V. Ivanov, Linear programming and the intersection of free subgroups in free products of
groups, submitted.
V. M. Kopytov and N. Y. Medvedev, Right ordered groups, Plenum Publ., New York, 1996.
A. G. Kurosh, The theory of groups, Chelsea, 1956.
R. C. Lyndon and P. E. Schupp, Combinatorial group theory, Springer-Verlag, 1977.
I. Mineyev, Submultiplicativity and the Hanna Neumann conjecture, Ann. Math. 175(2012),
393–414.
H. Neumann, On the intersection of finitely generated free groups, Publ. Math. 4(1956),
186–189; Addendum, Publ. Math. 5(1957), 128.
W. D. Neumann, On the intersection of finitely generated subgroups of free groups, Lecture
Notes in Math. (Groups-Canberra 1989) 1456(1990), 161–170.
J. R. Stallings, Topology of finite graphs, Invent. Math. 71(1983), 551–565.
Department of Mathematics, University of Illinois, Urbana, IL 61801, U.S.A.
E-mail address: [email protected]
| 4math.GR
|
ON THE STRUCTURE OF INVOLUTIONS AND SYMMETRIC
SPACES OF QUASI DIHEDRAL GROUP
arXiv:1401.0112v3 [math.GR] 4 Jul 2017
ZAHID RAZA, IMRAN AND BIJAN DAVVAZ
Abstract. Let G = QD8k be the quasi-dihedral group of order 8n and θ be an
automorphism of QD8k of finite order. The fixed-point set H of θ is defined as
Hθ = Gθ = {x ∈ G | θ(x) = x} and generalized symmetric space Q of θ is given
by Qθ = {g ∈ G | g = xθ(x)−1 for some x ∈ G}.
The characteristics of the sets H and Q have been calculated. It is shown that
for any H and Q, H.Q 6= QD8k . The H-orbits on Q are obtained under different
conditions. Moreover, the formula to find the order of v-th root of unity in Z2k
for QD8k has been calculated. The criteria to find the number of equivalence
classes denoted by C4k of the involution automorphism has also been constructed.
Finally, the set of twisted involutions R = Rθ = { x ∈ G | θ(x) = x−1 } has been
explored.
1. Introduction
Serious work on groups generated by reflections began in the 19th century. The
finite subgroups of O(3) generated by isometric reflections on the 2-sphere (or equivalently, by orthogonal linear reflections on R3 ) had been determined by Möbius in
1852. In the second half of the 19th century, work also began on finite reflection
groups on S n for n > 2 (or equivalently, finite linear reflection groups on Rn+1 ).
By the late 19′ s other groups generated by isometric reflections across the edges of
polygons (with 3 or more edges) in the hyperbolic plane had been studied by Klein
and Poincare. The group of symmetries of such a root system was a finite group
generated by reflections had been shown by Weyl in 1925. This intimate connection
with the classification of semi-simple Lie groups cemented reflection groups into a
central place in mathematics. The two lines of research were united in (Coxeter,
1930). He classified discrete groups generated by reflections on the n-dimensional
sphere or Euclidean space. The notion of an abstract reflection group introduced
by Jacques Tits which he called a coxeter group. He considered pairs (W, S) with
W a group and S a set of involutions which generate W so that the group has a
presentation of the following form: the set of generators is S and the set of relations
2010 Mathematics Subject Classification. 20F28 .
Key words and phrases. Automorphisms, involutions, fixed-point set, symmetric spaces, quasidihedral group.
1
2
ZAHID RAZA, IMRAN AND BIJAN DAVVAZ
is {(st)m(s,t) }, where m(s, t) denotes the order of st and the relations range over all
unordered pairs s, t ∈ S with m(s, t) 6= ∞. The pair (W, S) is a coxeter system
and W is a coxeter group. This was demonstrated in (Bourbaki, 1968; Bjorner and
Francesco, 2005).
The early work by Helminck on symmetric spaces concerned some algebraic results
about their structure(Helminck, 1988). Classification of the real symmetric spaces
and their fine structure had been included in (Helminck, 1994). By considering an
open subgroup H of the fixed point group Gθ of an involution θ of G, the result was
proved that ‘‘ For k a field of characteristic not two, let G be a connected semi-simple
algebraic k-group and θ an involution of G. Then θ is defined over k if and only if
H θ is defined over k” (Helminck and Wang, 1993). They also proved that ‘‘ if G is a
connected semi-simple algebraic group, θi involution of G and Hi = Gθi , (i = 1, 2).
Let H1θ1 and H2θ2 be the identity components of H1 and H2 respectively. If H1θ1 = H2θ2
then θ1 = θ2 .”
A mini-workshop on generalizations of symmetric spaces was organized by Ralf
Kohl at.el in 2012. Mathematicians in this area brought together to present their current research projects or trigger new collaboration via comparison, analogy, transfer,
generalization and unification of method in this workshop.
The structure of automorphism group of dihedral group has been found by Cunningham at.el (2012). They also characterized the automorphism group with θ2 = id
and calculated H and Q for each automorphism for the family of finite dihedral
groups. They also proved that the set of involutions is partitioned in such a way
that there are at most two distinct equivalence classes of involutions in each of the
partition. In involution case, Q is always a subgroup of G. Moreover, the infinite
dihedral group is also discussed.
Let G = Dcn be the dicyclic group of order 4n. Let φ be an automorphism of G
of order k. Bishop at.el (2013) describe φ and the generalized symmetric space Q of
G associated with φ. When φ is an involution, they describe its fixed point group H
along with the H-orbits and G-orbits of Q corresponding to the action of φ-twisted
conjugation.
In this paper, we extent the study to quasi-dihedral group QD8k .
2. Preliminaries
The quasi-dihedral group is denoted by QD8k , where QD8k = ha, b : a4k = b2 =
1, ba = a2k−1 bi. In reset of the whole paper, we denote it by QD2m , where m = 4k
and n = m2 = 2k. That is QD2m = ha, b : am = b2 = 1, ba = an−1 bi. It is
clear from the presentation that ai bj ∈ QD2m is in normal form for some integer
0 ≤ i < m and j ∈ {0, 1}. In the whole study, Zn denotes the additive group
of integers modulo n and Um denotes the multiplicative group of units of Zn and
ON THE STRUCTURE OF INVOLUTIONS AND SYMMETRIC SPACES OF QUASI DIHEDRAL GROUP
3
r ∈ Um if and only if gcd(r, m) = 1. The v-th root of unity in Zn which can be
v
described by Rm
= {r ∈ Um | r v = 1 (mod n)}.
Let G be a group, then an element a is a conjugate with b in G if and only
if there exists an element x such that a = xbx−1 . For a ∈ G, the function
φa (x) = axa−1 ∀ x ∈ G is called an inner automorphism of G. Let θ1 , θ2 be
two automorphisms, then θ1 ∼ θ2 if and only if they are conjugate to each other. If
there exists another automorphism σ such that σθ1 σ −1 = θ2 .
Definition 3. For any c ∈ Zn , we define ZDiv (c) = {y ∈ Zn | cy ≡ 0 (mod n)} and
easy to see that | ZDiv (c) |= gcd(c, n)
Notation 4. Let us fixed the following notations, which will be used in the whole
paper.
v
• For fixed r ∈ Rm
. The number of equivalence classes with fixed leading
co-efficient r is denoted by Nr =| {θ̄ | θ = rx + s ∈ Autv (D2m )} |
• Let us consider the subset of automorphism denoted by Autv (QD2m ) and is
defined as Autv (QD2m ) = {θ ∈ Aut(QD2m ) | θv = id }.
The following lemma is easy to prove.
Lemma 5. If QD2m be the quasi-dihedral group, then
(1) bai = ai(n−1) b for all i = 0, 1, . . . , (m − 1),
(2) | a2i b |= 2 and | a2i+1 b |= 4 for all i = 0, 1, . . . , (n − 1).
Lemma 6. The automorphism group of QD2m is isomorphic to the group of affine
linear transformations of Zn . In particular, we have
Aut(QD2m ) ∼
= Af f (Zn ) = {rx + s : Zn → Zn | r ∈ Um , s ∈ Zn }
and the action of rx + s on elements of QD2m in normal form is given by
(6.1)
(rx + s)(aj bt ) = arj+[(t−1)n−(t−2)]s bt
Proof. Let θ be an automorphism, then θ(a) = ar and θ(b) = (as b), where gcd(r, m) =
1, s = 0, 2, 4, . . . , (m − 2). Consider arbitrary element aj bt ∈ QD2m , and θ being
homomorphism,
(6.2)
θ(aj bt ) = θ(aj )θ(bt ),
then θ(aj ) = arj and θ(bt ) = (as b)(as b)(as b) . . . (as b) = as (bas )(bas )b . . . (as b).
By Lemma 5, we have
θ(bt ) = as [as(n−1) b](bas )(bas )b . . . (as b) = as+s(n−1) (b2 as )(bas )b . . . (as b)
θ(bt ) = asn (b2 as )(bas )b . . . (as b) = asn (as(n−1) b2 )(bas )b . . . (as b)
θ(bt ) = asn+sn−sb3 (bas )b . . . (as b) = a2sn−s b3 (bas )b . . . (as b) after t-steps, we have
θ(bt ) = a(t−1)ns−(t−2)s bt . Put values in equation 6.2, we get
θ(aj bt ) = arj .a(t−1)ns−(t−2)s bt = arj+(t−1)ns−(t−2)s bt = arj+[(t−1)n−(t−2)]s bt . Since b2 =
4
ZAHID RAZA, IMRAN AND BIJAN DAVVAZ
1, so t ∈ {0, 1}. If t = 0, then θ(aj ) = arj+(2−n)s =⇒ rx + (2 − n)s ∈ Af f (Zn ). If
t = 1, then θ(aj b) = arj+s b =⇒ rx + s ∈ Af f (Zn ).
Conversely, For f ∈ Af f (Zn ), that is f = rx + s where r ∈ Um , s ∈ Zn and it is
given as a group action f (aj bt ) = arj+[(t−1)n−(t−2)]s bt for all aj bt ∈ QD2m .
Now, we will show that f is an automorphism. For this let us define θ : QD2m −→
QD2m by θ(aj bt ) = f (aj bt ) = arj+[(t−1)n−(t−2)]s bt . We will show that θ is a bijective
homomorphism. For this we will show that θ is one to one and onto.
one to one: Suppose that θ(aj1 bt1 ) = θ(aj2 bt2 ), then arj1 +[(t1 −1)n−(t1 −2)]s bt1 =
arj2 +[(t2 −1)n−(t2 −2)]s bt2 ⇒ arj1 +[(t1 −1)n−(t1 −2)]s−rj2−[(t2 −1)n−(t2 −2)]s bt1 −t2 = e.
Case 1 :
rj1 + [(t1 − 1)n − (t1 − 2)]s − rj2 − [(t2 − 1)n − (t2 − 2)]s = 0 and t1 − t2 = 2, since,
t ∈ {0, 1} therefore t1 − t2 6= 2 not possible.
Case 2 :
rj1 + [(t1 − 1)n − (t1 − 2)]s − rj2 − [(t2 − 1)n − (t2 − 2)]s = m and t1 − t2 = 0
⇒ r(j1 −j2 )+ns(t1 −t2 )−s(t1 −t2 ) = m ⇒ r(j1 −j2 ) = m ⇒ r | m but gcd(r, m) = 1,
not possible. Now r(j1 − j2 ) = m will be only possible when j1 − j2 = 0, r 6= 0
j1 = j2 ⇒ aj1 = aj2 ⇒ aj1 bt1 = aj2 bt2 ⇒ θ is one to one.
onto: It is cleared from the definition of the group action that for every arj+[(t−1)n−(t−2)]s bt
there exists aj bt ∈ QD2m such that θ(aj bt ) = arj+[(t−1)n−(t−2)]s bt ⇒ θ is onto.
Hence Aut(QD2m ) ∼
= Af f (Zn ).
7. Structure of Automorphisms of QD2m
Proposition 7.1. For any integer v ≥ 1 and θ = rx + s be an automorphism of
v
QD2m then θv =id if and only if r ∈ Rm
and s ∈ ZDiv(r v−1 + r v−2 + · · · + r + 1).
Proof. We will use the mathematical induction on v to show first the following
statement: θv (x) = r v x + (r v−1 + r v−2 + · · · + r + 1)s.
For v = 2, 3 θ2 (x) = θ[θ(x)] = θ(rx + s) = r(rx + s) + s = r 2 x + rs + s =
r 2 x + (r + 1)s. Similarly, θ3 (x) = r 3 x + r 2 s + rs + s = r 3 x + (r 2 + r + 1)s.
Suppose that the statement is true for v = i that is
i
θ (x) = r i x + (r i−1 + r i−2 + · · · + r + 1)s. We will show that the statement is
true for v = i + 1. Take θi+1 (x) = θ[θi (x)] = θ[r i x + (r i−1 + r i−2 + · · · + r + 1)s]
θi+1 (x) = r[r i x + (r i−1 + r i−2 + · · · + r + 1)s] + s = r i+1 x + (r i + r i−1 + · · · + r + 1)s].
Hence , the statement is true for all positive values of v. We have identified the
identity automorphism with x ∈ Af f (Zn ), thus r v ≡ 1 (mod n) and (r v−1 + r v−2 +
v
· · · + r + 1)s ≡ 0 (mod n) ⇔ r ∈ Rm
and s ∈ ZDiv(r v−1 + r v−2 + · · · + r + 1) as
required.
3
Example 7.2. Let G = QD56 and v = 3. In this case R28
= {1, 9, 11, 15, 23, 25},
take r = 9, then ZDiv(92 + 9 + 1) = ZDiv(91) = {0, 2, 4, 6, 8, 10, 12}. Now, consider
ON THE STRUCTURE OF INVOLUTIONS AND SYMMETRIC SPACES OF QUASI DIHEDRAL GROUP
5
θ1 = 9x+6 ∈ Aut(QD56 ), then after simple calculation we get θ13 = x =id. Similarly,
for θ2 = 9x + 12 ∈ Aut(QD56 ), we have θ23 = x =id.
Proposition 7.3. For any m and v ≥ 1, we have
X
gcd(r v−1 + r v−2 + · · · + r + 1, n)
|Autv (QD2m )| =
r∈Rvm
v
Proof. This follows from above Proposition: for any r ∈ Rm
there are
|ZDiv(r v−1 + r v−2 + · · · + r + 1)| = gcd(r v−1 + r v−2 + · · · + r + 1, n)
element s such that (r v−1 +r v−2 +· · ·+r +1)s ≡ 0 (mod n) and every automorphism
θ ∈ Autv (QD2m ) must be of this form.
Proposition 7.4. Let θ1 = r1 x+s1 and θ2 = r2 x+s2 be two different automorphism
of QD2m . Then θ1 ∼ θ2 if and only if r1 = r2 and f s1 − s2 ∈ hr1 − 1i ≤ Zn for some
f ∈ Um .
Proof. Given that θ1 = r1 x + s1 and θ2 = r2 x + s2 ∈ Aut(QD2m ). Suppose that
σ = f x + g be any arbitrary automorphism of QD2m and define σ −1 = f −1 x − f −1 g.
We will check that σ ·σ −1 = x. For this take σ ·σ −1 = σ(σ −1 (x)) = σ(f −1 x−f −1 g) =
f (f −1 x − f −1 g) + g = x − g + g = x. Now, we calculate
(7.1)
σθ1 σ −1 = σ · (θ1 · σ −1 )
Take θ1 · σ −1 = θ1 [σ −1 (x)] = θ1 [f −1 x − f −1 g] = r1 (f −1 x − f −1 g) + s1 = r1 f −1 x −
r1 f −1 g + s1 put in 7.1, we get σθ1 σ −1 = σ(r1 f −1 x − r1 f −1 g + s1 ) = f (r1 f −1 x −
r1 f −1 g+s1 )+g σθ1 σ −1 = r1 x−r1 g+f s1 +g = r1 x+f s1 −g(r1 −1). Now, by definition
σθ1 σ −1 = θ2 if and only if r1 = r2 and f s1 − g(r1 − 1) = s2 ⇒ f s1 − s2 = g(r1 − 1)
⇒ f s1 − s2 ∈ hr1 − 1i.
Example 7.5. Let G = QD64 and consider θ1 = 7x + 3, θ2 = 7x + 14, θ3 = 7x + 11
three automorphisms of QD64 . Then it is easy to check that θ1 ≁ θ2 because there
does not exist such f ∈ U32 for which 3f − 14 ∈ h6i. And θ1 ∼ θ3 because there
exist some f ∈ U32 for which 3f − 14 ∈ h6i.
Proposition 7.6. Suppose m is fixed and let v ≥ 1.
v
(1) For any r ∈ Rm
; hr − 1i ≤ ZDiv(r v−1 + r v−2 + · · · + r + 1).
v−1 +r v−2 +···+r+1)
v
(2) For any r ∈ Rm
; Um acts on the cosets ZDiv(r hr−1i
.
(3) The set Autv (QD2m ) is partitioned into equivalence classes indexed by pair
v−1 +r v−2 +···+r+1)
v
(r, A), where r ∈ Rm
and A is an orbit of Um on ZDiv(r hr−1i
.
Proof.
(1) We know that (r − 1).(r v−1 + r v−2 + · · · + r + 1) = r v − 1 ≡ 0 (mod n)
⇒ (r − 1).(r v−1 + r v−2 + · · · + r + 1) ≡ 0 (mod n) ⇒ (r − 1) ∈ ZDiv(r v−1 +
r v−2 + · · · + r + 1) ⇒ hr − 1i ≤ ZDiv(r v−1 + r v−2 + · · · + r + 1).
6
ZAHID RAZA, IMRAN AND BIJAN DAVVAZ
(2) We simply recognize that Um = Aut(Zn ) acts on Zn by multiplication and
since every subgroup is cyclic, Um must stabilize the subgroups of Zn .
v−1 +r v−2 +···+r+1)
.
(3) It is simply Proposition 7.4 in terms of the Um -action on ZDiv(r hr−1i
Example 7.7. Let G = QD40 ,and v = 2. Then U20 = {1, 3, 7, 9, 11, 13, 17, 19} and
Z10 = {0, 1, 2, . . . , 9}.
2
(1) Since v = 2, so R20
= {1, 9, 11, 19}. Take r = 9 then h9 − 1i = h8i =
{0, 2, 4, 6, 8} and ZDiv(10) = {0, 1, 2, . . . , 9}. So, {0, 2, 4, 6, 8} ⊆ {0, 1, 2, . . . , 9}
(2) For U20 action on the cosets ZDiv(r+1)
. Take r = 9 then ZDiv(r+1)
= ZDiv(10)
=
hr−1i
hr−1i
h8i
E={0,1,2,...,9}
F ={0,2,4,6,8}
= x + F ∀ x ∈ E provide two sets F and F1 = {1, 3, 5, 7, 9}.
Now, the action of U20 on the cosets
(3)
ZDiv(10)
h8i
has two sets.
ZDiv(10)
h8i
= y.F =
F ∀ y ∈ U20 and ZDiv(10)
= y.F1 = F1 ∀ y ∈ U20
h8i
ZDiv(10)
The action of U20 on h8i = F, F1 . Then the number of equivalence classes
2
of Aut(QD40 ) indexed by the pair (r, A) where r ∈ R20
and A is an orbit of
ZDiv(10)
U20 on h8i is 8.
v
Theorem 7.8. If m is a fixed positive integer and r ∈ Rm
, then the number
v−1
v−2
ZD iv(r
+r
+···+r+1)
of orbits of Um on
is equal to the number of divisors of
hr−1i
gcd(r−1,n). gcd(r v−1 +r v−2 +···+r+1,n)
.
n
Proof. The Um -orbit on Zn are indexed by the subgroups of Zn . Thus the Um -orbit
v−1 +r v−2 +···+r+1)
on ZDiv(r hr−1i
are indexed by subgroups L ≤ Zn such that
hr − 1i ≤ L ≤ ZDiv(r v−1 + r v−2 + · · · + r + 1).
It is well known that the subgroup lattice of Zn is isomorphic to the divisor lattice
of n. The subgroup hr − 1i corresponds to the divisor gcd(r − 1, n) and the subgroup
n
,
ZDiv(r v−1 + r v−2 + · · · + r + 1) correspond to the divisor gcd(rv−1 +rv−2
+···+r+1,n)
and the subgroups between these groups correspond to the divisor of n between
n
gcd(r − 1, n) and gcd(rv−1 +rv−2
in the divisor lattice. Finally, it is known
+···+r+1,n)
that this sub-lattice of the divisor of n is isomorphic to the divisor lattice of
gcd(r − 1, n)
gcd(r − 1, n). gcd(r v−1 + r v−2 + · · · + r + 1, n)
=
.
n
n
gcd(r v−1 +r v−2 +···+r+1,n)
8. Fixed Ssts and Symmetric Spaces of Automorphisms
In this section if θ is an automorphism the fixed-point set Hθ and the symmetric
space Qθ will be discussed, where Hθ = {x ∈ G | θ(x) = x} and Qθ = {g ∈ G | g =
ON THE STRUCTURE OF INVOLUTIONS AND SYMMETRIC SPACES OF QUASI DIHEDRAL GROUP
7
xθ(x)−1 for some x ∈ G}. When θ is understood to be fixed the subscript from our
notation has been dropped.
The following theorem characterizes the sets H and Q in the case of quasi-dihedral
groups.
Theorem 8.1. If θ = rx + s is an automorphism of G = QD2m of finite order, then
′
′
H = {aj | j(r + 2s − 1) ≡ 0 (mod n)} ∪ {aj b | j (r + 2s − 1) ≡ −s (mod n)} and
′
Q = {aj | j ∈ hr + 2s − 1i ∪ [l(1 − r) − s] (mod n)} where, j and l are even numbers.
Proof. Since, θ is an automorphism and θ(aj bt ) = arj+[(t−1)n−(t−2)]s bt . First we will
find set the H. Let us consider aj ∈ G, then by definition of H, we have θ(aj ) = aj ,
θ(aj ) = a[r+(2−n)s]j = aj , ⇒ n | j(r + 2s − 1).
′
′
′
′
′
Now consider aj b ∈ G, then θ(aj b) = aj b so, aj b = a[r+(2−n)s]j as b, thus n |
′
′
′
(r+2s−1)j +s. Hence, H = {aj | j(r+2s−1) ≡ 0 (mod n)}∪{aj b | j (r+2s−1) ≡
′
−s (mod n)}, where j is even number.
Now we will find the set Q. L aj ∈ G be an arbitrary element and let ai be a fixed
element of G, where j 6= i, then by definition of Q, aj = ai θ(ai )−1 = ai θ(a−i ) =
ai .a−[r+(2−n)s]i = ai[1−r−(2−n)s] ⇒ j ≡ i[1 − (r + 2s)] (mod n) ⇒ j ∈ hr + 2s − 1i.
Now consider al b ∈ G, where l is even number, then aj = (al b)θ(al b)−1 =
(al b)θ(al b) = (al b)(arl+s b) = al (barl+s )b = al a(rl+s)(n−1) bb = al+(rl+s)(n−1) ⇒ j =
l + (rl + s)(n − 1) ⇒ j = l[1 + r(n − 1)] + s(n − 1) ⇒ j ≡ l(1 − r) − s (mod n).
So, Q = {aj | j ∈ hr + 2s − 1i ∪ [l(1 − r) − s] (mod n)}, where l is an even
numbers.
Example 8.2. Let us consider G = QD64 and θ = 3x + 12 ∈ Aut(D64 ), then
′
′
H = {aj | 10j ≡ 0 (mod 16)} ∪ {aj b | 10j ≡ −4 (mod 16)} = {1, a8 , a2 b, a10 b} and
′
Q = {aj | j ∈ h10i ∪ (14j + 4) (mod 16)} = {1, a2 , a4 , . . . , a14 }
Using the descriptions of H and Q , we obtain further results as well.
Corollary 8.3. If θ = rx + s is an automorphism of QD2m of finite order, where
m = 4k, n = 2k and k is even, then H will be cyclic or non-cyclic according to the
following conditions
(1) Let gcd(r + 2s − 1, n) = d1 . If s = 2w + 1 is odd and d1 ∤ w then H will be
cyclic that is H ∼
= ZDiv(r + 2s − 1) and | H |= d1
(2) Let gcd(r + 2s − 1, k) = d2 . If s = 2w is even and d2 | w then we have two
cases
a): If w is odd, then H will be cyclic | H |= d1 and H ∼
= ZDiv(r +2s−1).
b): If w is even
i): If d2 ∤ w then H will be cyclic | H |= d1 and H ∼
= ZDiv(r +
2s − 1).
ii): If d2 | w then H will be non-cyclic. We have two cases
8
ZAHID RAZA, IMRAN AND BIJAN DAVVAZ
k
I-): if 4 ∤ k then | H |= 23 d1 H = {ha 2 i, a2l b, a2q b | l 6=
q for some l, q ∈ {1, 2, ...k}}. It is not a subgroup of G.
II-): if 4 | k then | H |= 2d1 and H ∼
= ZDiv(r + 2s − 1) × Z2 .
Example 8.4. Let G = QD48 , then gcd(r, m) = 1 ⇒ r ∈ U24 = {1, 5, 7, 11, 13, 17, 19, 23}
and s ∈ Z12 = {0, 1, 2, . . . , 11}.
(1) Consider θ1 = 7x + 11 ∈ Aut(QD48 ) then r = 7, s = 2(5) + 1 = 11
⇒ r + 2s − 1 = 4 and d1 = gcd(28, 12) = 4 ⇒ 4 ∤ 5 then H1 = {1, a3, a6 , a9 }
is cyclic. and ZDiv4 = {0, 3, 6, 9} ⇒ H1 ∼
= ZDiv4 and O(H1) = d1 = 4
(2) a): Take θ2 = 11x + 10 ∈ Aut(QD48 ) then r = 11, s = 2(5) = 10
⇒ r + 2s − 1 = 6 and d1 = gcd(6, 12) = 6, d2 = gcd(6, 6) = 6 then
H2 = {1, a2 , a4 , . . . , a10 } is cyclic and ZDiv6 = {0, 2, 4, 6, 8, 10} ⇒ H2 ∼
=
ZDiv6 and O(H2) = d1 = 6
b):
i): For θ3 = 11x + 4 ∈ Aut(QD48 ) then r = 11, s = 2(2) = 4
⇒ r + 2s − 1 = 6 and d1 = gcd(6, 12) = 6, d2 = gcd(6, 6) = 6
⇒ 6 ∤ −2 then H3 = {1, a2 , a4 , . . . , a10 } is cyclic and ZDiv6 =
{0, 2, 4, 6, 8, 10} ⇒ H3 ∼
= ZDiv6 and O(H3 ) = d1 = 6
ii): I-: For θ4 = 5x + 8 ∈ Aut(QD48 ) then r = 5, s = 2(4) = 8
⇒ r + 2s − 1 = 8 and d1 = gcd(8, 12) = 4, d2 = gcd(8, 6) = 2
⇒ 2 | −4 then H4 = {1, a3 , a6 , a9 , a2 b, a8 b} is non-cyclic and
O(H4 ) = 32 .d1 = 6
II-: Take θ5 = 7x + 4 ∈ Aut(QD48 ) then r = 7, s = 2(2) = 4
⇒ r + 2s − 1 = 2 and d1 = gcd(2, 12) = 2, d2 = gcd(2, 6) =
2 ⇒ 2 | −2 then H5 = {1, a6 , a4 b, a10 b} is non-cyclic and
O(H5 ) = 2.d1 = 4
Corollary 8.5. If m = 4k, n = 2k and k is odd, then H will be cyclic or non-cyclic
according to the following conditions
(1) Let gcd(r + 2s − 1, n) = d1 . If s = 2w + 1 is odd then H will be cyclic
| H |= d1 and H ∼
= ZDiv(r + 2s − 1).
(2) Let gcd(r + 2s − 1, n) = d1 . If s = 2w is even then
a): If d1 = n then H will be cyclic | H |= d1 = n and H ∼
= ZDiv(r +
2s − 1).
b): If d1 = 2 then H will be non-cyclic | H |= 23 d1 = 23 (2) = 3 and
H = {1, ak , a2l b} for some l ∈ {1, 2, ..., k} is not a subgroup of G.
Example 8.6. Let G = QD40 where m = 20, n = 10, k = 5. Since gcd(r, m) = 1
⇒ r ∈ U20 = {1, 3, 7, 9, 11, 13, 17, 19} and s ∈ Z10 = {0, 1, 2, . . . , 9}.
(1) Consider θ1 = 3x + 5 ∈ Aut(QD40 ) then r = 3, s = 5 ⇒ r + 2s − 1 = 2
and gcd(2, 10) = 2 then H1 = {1, a5 } is cyclic and ZDiv2 = {0, 5} ⇒ H1 ∼
=
ZDiv2 and | H1 |= 2
ON THE STRUCTURE OF INVOLUTIONS AND SYMMETRIC SPACES OF QUASI DIHEDRAL GROUP
9
(2)
a): For θ2 = 7x + 2 ∈ Aut(QD40 ) then r = 7, s = 2 ⇒ r + 2s − 1 = 0
and gcd(0, 10) = 10 then H2 = {1, a2 , a3 , . . . , a9 } is cyclic and ZDiv0 =
{0, 1, 2, . . . , 9} ⇒ H2 ∼
= ZDiv0 and | H2 |= 10
b): Take θ3 = 7x+4 ∈ Aut(QD40 ) then r = 7, s = 4 ⇒ r+2s−1 = 4 and
gcd(4, 10) = 2 then H3 = {1, a5 , a4 b} is non-cyclic and | H3 |= 32 .2 = 3
Corollary 8.7. If θ = rx + s ∈ Autv (QD2m ) be of finite order, where m = 4k, n =
2k and k is even number, then Q will be cyclic or non-cyclic according to the
following conditions
(1) If s = 2w + 1 is odd, then Q will be non-cyclic.
(2) Let gcd(r + 2s − 1, n) = d1 . If s = 2w is even, then
′
a): If d1 | (1 − r)j − s and
i): if r + 2s − 1 6≡ k (mod n) then Q will be cyclic.
ii): If r + 2s − 1 ≡ k (mod n) we have two cases
• if 1 − r ≡ k (mod n) then Q will be cyclic.
• If 1 − r 6≡ k (mod n) then Q will be non-cyclic.
′
b): If d1 ∤ (1 − r)j − s and
i): if −s 6≡ k (mod n) then Q will be non-cyclic.
ii): If −s ≡ k (mod n) then Q will be cyclic.
Example 8.8. Let G = QD48 where m = 24, n = 12, k = 6. Since gcd(r, m) = 1
⇒ r ∈ U24 = {1, 5, 7, 11, 13, 17, 19, 23} and s ∈ Z12 = {0, 1, 2, . . . , 11}.
(1) Consider θ1 = 7x + 11 ∈ Aut(QD48 ) then r = 7, s = 11 ⇒ r + 2s − 1 = 4
then Q1 = {1, a, a4 , a8 } is non-cyclic.
(2) a):
i): For θ2 = 5x + 8 ∈ Aut(QD48 ) then r = 5, s = 8 ⇒ r + 2s − 1 =
′
′
8 6≡ k and d1 = gcd(8, 12) = 4, (1 − r)j − s = −4j − 8 at
′
j = 2 yeilds 4 | −16 then Q2 = {1, a4 , a8 } is cyclic.
ii): Take θ3 = 7x + 6 ∈ Aut(QD48 ) then r = 7, s = 6 ⇒ r + 2s − 1 =
′
′
6 ≡ k and d1 = gcd(6, 12) = 6, (1 − r)j − s = −6j − 6 at
′
j = 2 yeilds 6 | −18 and 1 − r = −6 ≡ 6 ≡ k then Q3 = {1, a6 }
is cyclic. And for θ4 = 11x + 4 ∈ Aut(QD48 ) then r = 11, s = 4
′
⇒ r + 2s − 1 = 6 ≡ k and d1 = gcd(6, 12) = 6, (1 − r)j − s =
′
′
−10j − 4 at j = 2 yeilds 6 | −24 and 1 − r = −10 ≡ 2 6≡ k then
Q4 = {1, a4 , a6 , a8 } is non-cyclic.
b):
i): Consider θ5 = x + 10 ∈ Aut(QD48 ) then r = 1, s = 10 ⇒
′
′
r + 2s − 1 = 8 and d1 = gcd(8, 12) = 4, (1 − r)j − s = 0j − 10 ≡
2 ⇒ 4 ∤ 2 and −s = −10 ≡ 2 6≡ k then Q5 = {1, a2 , a4 , a8 } is
non-cyclic.
ii): For θ6 = 5x+ 6 ∈ Aut(QD48 ) then r = 5, s = 6 ⇒ r + 2s −1 = 4
′
′
′
and d1 = gcd(4, 12) = 4, (1 − r)j − s = −4j − 6 for any j ,
10
ZAHID RAZA, IMRAN AND BIJAN DAVVAZ
′
d1 ∤ −4j −6 and −s = −6 ≡ 6 ≡ k then Q6 = {1, a2 , a4 , . . . , a10 }
is cyclic.
Corollary 8.9. If m = 4k, n = 2k and k is odd number, then Q will be cyclic or
non-cyclic according to the following conditions
(1) If s = 2w is even then Q will be cyclic.
(2) If s = 2w + 1 is odd
a): Q will be non-cyclic if either r + 2s − 1 ≡ 0 (mod n) or 1 − r ≡
0 (mod n).
b): Q will be cyclic
i): if 1 − r ≡ 0 and s ≡ k (mod n).
ii): if r + 2s − 1 6≡ 0 and 1 − r 6≡ 0 (mod n).
Example 8.10. Let G = QD40 where m = 20, n = 10, k = 5. Since gcd(r, m) = 1
⇒ r ∈ U20 = {1, 3, 7, 9, 11, 13, 17, 19} and s ∈ Z10 = {0, 1, 2, . . . , 9}.
(1) For θ1 = 7x + 6 ∈ Aut(QD40 ) then r = 7, s = 6 ⇒ r + 2s − 1 = 8 then
Q1 = {1, a2 , a4 , a6 , a8 } is cyclic.
(2) a): Take θ2 = x + 7 ∈ Aut(QD40 ) then r = 1, s = 7 ⇒ r + 2s − 1 =
4 and 1 − r = 0 then Q2 = {1, a2 , a3 , a4 , a6 , a8 } is non-cyclic.
b):
i): Consider θ3 = 11x + 5 ∈ Aut(QD40 ) then r = 11, s = 5
⇒ r + 2s − 1 = 0, 1 − r = 0, s = 5 ≡ k then Q3 = {1, a5 } is
cyclic.
ii): For θ4 = 7x + 5 ∈ Aut(QD40 ) then r = 7, s = 5 ⇒ r + 2s − 1 =
6, 1 − r = −6 6≡ 0 then Q4 = {1, a, a2 , . . . , a9 } is cyclic.
Corollary 8.11. Let θ = rx + s ∈ Autv (QD2m ) be of finite order, where m =
4k, n = 2k and k is even number. Then Q is isomorphic to a cyclic group as
follows:
(1) If | Q |= 2 then Q = hak : an = 1i and Q ∼
= hki.
k
(2) If | Q |= 2 then
a): Let gcd(r + 2s − 1, n) = d1 . If n ∤ r + 2s − 1 then Q = had1 : an = 1i
and Q ∼
= hd1 i.
b): If n | r + 2s − 1 and gcd(r + 2s − 1, n) = n then Q = ha4 : an = 1i
and Q ∼
= h4i.
(3) If | Q |= k and gcd(r + 2s − 1, n) = d1 then
′
a): if (r + 2s − 1) ∩ (1 − r)j − s = ∅ then Q = had1 /2 : an = 1i
and Q ∼
= h d21 i.
′
b): if (r+2s−1)∩(1−r)j −s 6= ∅ then Q = had1 : an = 1i and Q ∼
= hd1 i.
Corollary 8.12. Let θ = rx + s ∈ Autv (QD2m ) be of finite order, if k is odd
number. Then the conditions for Q to be cyclic and isomorphism relationship are:
(1) If s is even, then
ON THE STRUCTURE OF INVOLUTIONS AND SYMMETRIC SPACES OF QUASI DIHEDRAL GROUP
11
a): if r = 1, s = 0 | Q |= 1 then Q = {1} and Q ∼
= h0i.
2
n
∼
b): if | Q |= k then Q = ha : a = 1i = h2i.
(2) If s is odd, then
a): Let n | 1 − r, n | s − k. If | Q |= 2 then Q = hak : an = 1i ∼
= hki.
b): if r + 2s − 1 6≡ 0 and 1 − r 6≡ 0 (mod n) and | Q |= n then Q =
ha : an = 1i ∼
= h1i.
Corollary 8.13. Let θ = rx + s ∈ Aut(QD2m ) be a fixed automorphism. Then for
any H and Q, HQ =
6 QD2m .
Example 8.13.1. Consider the involution θ1 = 17x + 8 of QD48 . Then H1 =
{1, a3 , a6 , a9 , a2 b, a8 b} and Q1 = {1, a4 , a8 } ⇒ H1 Q1 6= QD48 . Now, consider
the automorphism θ2 = 5x + 3 of order 4. Then H2 = {1, a6 b} and Q2 =
{1, a2 , a4 , . . . , a10 , a, a5 , a9 } ⇒ H2 Q2 6= QD48 .
Proposition 8.14. Let θ = rx + s ∈ Autv (QD2m ) and for any al ∈ Q.
′
i): If l is odd, then H\Q = {{aj } | j ∈ hr + 2s − 1i ∪ [(1 − r)j − s] (mod n)}
′
ii): If l is even then H\Q = {{aj , a−j } | j ∈ hr+2s−1i∪[(1−r)j −s] (mod n)}
′
where, j is any even number. In either case G\Q = {Q}, that is there is only a
single G-orbit on Q.
Proof. Since, H is fixed by θ thus the action of H on Q is simply by conjugation.
Notice that Q ⊆ hai ≤ QD2m so, the action of H on general element al of Q. Let
ai ∈ H then, ai .al .a−i = ai+l−i = al . So, hai fixes Q point wise.
′
′
′
′
′
′
′
Suppose ai b ∈ H, then (ai b).al .(ai b)−1 = ai (bal )b−1 a−i = ai (al(n−1) b)b−1 a−i =
′
′
′
l
l
ai +l(n−1)−i = al(n−1) = aln .a−l = (a2n ) 2 .a−l = (am ) 2 .a−l = a−l so, aj b takes
′
al to a−l ⇒ {aj , a−j } is an orbit ⇔ aj b ∈ H which is true ⇔ l is even.
Now, for every element of Q is in G-orbit of 1 ∈ Q. Since, every element of Q are
of the form al ∈ Q. and G has two type of elements ai and ai b. Now let us supposed
that q1 = al = ai .θ(ai )−1 for some ai ∈ G, thus q1 = ai .θ(a−i ) = ai .a−ri+[2−n]y =
ai−ri+[2−n]y = a−i(r−1)+(2−n)y = av(r−1)+(2−n)y , where v = −i.
Also, al = (ai b).θ(ai b)−1 for some ai b ∈ G ⇒ al = (ai b).θ(ai b) = (ai b).(ari+y b) =
i
a (bari+y )b = ai .a(ri+y)(n−1) b.b = ai+(ri+y)(n−1) b2 Let us denote this element q2 =
al = a−v+(y−vr)(n−1) , where v = −i.
For a−v ∈ G, we have a−v .1.θ(a−v )−1 = a−v .θ(av ) = a−v .arv+(2−n)y = av(r−1)+(2−n)y =
q1 . Now for a−v b ∈ G, we have (a−v b).1.θ(a−v b)−1 = (a−v b).θ(a−v b) = (a−v b).a−rv+y b =
a−v (ba−rv+y )b = a−v .a(−rv+y)(n−1) b.b = a−v+(y−rv)(n−1) = q2 .
Therefore, for any q ∈ Q there exist g ∈ G such that g.1.θ(g)−1 = q. Hence
G\Q = {Q}, there is a single G-orbit on Q.
Example 8.14.1. Revisiting θ1 , θ2 of Example 8.13.1 applying Proposition 8.14 to
obtain that for θ1 . Since all l are even, therefore H1 \Q1 = {{1}, {a4 , a−4 }, {a8, a−8 }},
12
ZAHID RAZA, IMRAN AND BIJAN DAVVAZ
for θ2 there exists even and odd powers of elements of Q. For all even powers
H2 \Q2 = {{1}, {a2 , a−2 }, . . . , {a10 , a−10 }, {a, a−1 }, {a5 , a−5 }, {a9 , a−9 }} and odd
powers H2 \Q2 = {{1}, {a2 }, . . . , {a10 }, {a}, {a5 }, {a9 }}.
9. Involutions in Aut(QD2m )
If θ = rx + s is an automorphism such that θ2 =id, then Proposition 7.1 gives
2
r ∈ Rm
and s ∈ ZDiv(r + 1).
Proposition 9.1. Suppose that m = 4pt11 pt22 , where the pi are distinct odd primes.
2
Then | Rm
|= 8
Proof. Since, m = 4pt11 pt22 ⇒ n = 2pt11 pt22 . Suppose that r ∈ Zn with r 2 ≡ 1 (mod n)
∵ gcd(r, m) = 1 ⇒ gcd(r, 4pt11 pt22 ) = 1. But m is even therefore, r is odd. Assume
that r = 2q + 1 for some 0 ≤ q < (n − 1). Now, r 2 = (2q + 1)2 = 4q 2 + 4q + 1
⇒ 4q 2 +4q ≡ 0 (mod n) ⇒ 4q(q+1) = l.(2pt11 pt22 ) ∵ gcd(4, 2pt11 pt22 ) = 2 ⇒ 2q(q+1) =
l.(pt11 pt22 ) (mod pt11 pt22 ). Either q or q + 1 is odd.
l
.pt11 pt22 ⇒ q = h.pt11 pt22 where 1 ≤ h ≤ 4. Now,
(1) If q is even, then q = 2(q+1)
r = 2q + 1 ⇒ r = 2h.pt11 pt22 + 1 ⇒ r ≡ 1(mod n)
(2) If q + 1 is even, then q + 1 = 2ql .pt11 .pt22 ⇒ q + 1 = h.pt11 pt22 ⇒ q = h.pt11 pt22 − 1
where 1 ≤ h ≤ 4. Now, r = 2q + 1 ⇒ r = 2h.pt11 pt22 − 1 ⇒ r ≡ −1(mod n)
⇒ r ≡ ±1(mod n),
t
Proposition 9.2. Suppose that m = 2α pt11 pt22 . . . pqq where the pi are distinct odd
primes. Then
2q+1 α = 2
2
2q+2 α = 3
| Rm
|=
q+3
2
α ≥ 4.
Proof. Similar to the Propositions 10.1.
Example 9.3. Let G = QD192 , then U96 = {1, 5, 7, 11, 13, 17, 19, 23, 25, 29, 31, 35, 37,
2
41, 43, 47, 49, 53, 55, 59, 61, 65, 67, 71, 73, 77, 79, 83, 85, 89, 91, 95} and R96
= {1, 7, 17,
2
4
23, 25, 31, 41, 47, 49, 55, 65, 71, 73, 79, 89, 95} thus, | R96 |= 2 = 16
Corollary 9.4.
|Aut2 (QD2m )| =
X
gcd(r + 1, n).
r∈R2m
2
Theorem 9.5. Let r ∈ Rm
, then the followings hold:
(1) hr − 1i ≤ ZDiv(r + 1)
iv(r+1) |≤ 2
(2) Nr =| ZDhr−1i
ON THE STRUCTURE OF INVOLUTIONS AND SYMMETRIC SPACES OF QUASI DIHEDRAL GROUP
13
(3) Nr = 2 if and only if m = 2α l ; α ≥ 2, and l is odd, r ≡ ±1 (mod 2α−1 ).
2
(1) Given that r ∈ Rm
, then (r − 1).(r + 1) = r 2 − 1 ≡ 0 ⇒ (r − 1) ≤
ZDiv(r + 1) ⇒ hr − 1i ≤ ZDiv(r + 1).
(2) Suppose
Proof.
|
(9.1)
ZDiv(r + 1)
|= j
hr − 1i
n
It is well known that | hr − 1i |= gcd(r−1,n)
and | ZDiv(r + 1) |= gcd(r + 1, n)
put values in 9.1
gcd(r + 1, n)
=j
n
gcd(r−1,n)
(9.2)
⇒ gcd(r + 1, n). gcd(r − 1, n) = n.j
t
Now, suppose that m = 2α pt11 pt22 . . . pqq where α ≥ 2 ⇒ α − 1 ≥ 1
t
⇒ n = 2α−1 pt11 pt22 . . . pqq where pi are distinct odd primes and ti ≥ 1. Then
gcd(r + 1, n) = 2α1 pa11 pa22 . . . paq q
with α1 ≤ (α − 1) and 0 ≤ ai ≤ ti ∀ i and
gcd(r − 1, n) = 2α2 pb11 pb22 . . . pbqq
with α2 ≤ (α − 1) and 0 ≤ bi ≤ ti . Now, for all i ∈ {1, 2, . . . q} either ai = 0
or bi = 0. Indeed, if ai > 0 and bi > 0 then pi divides both (r + 1) and
(r − 1) which is impossible. ∵ (pi > 2)
Similarly, either α1 ≤ 1 or α2 ≤ 1. Otherwise 2min{α1 ,α2 } divides (r + 1)
and (r − 1) which is impossible.
Since, from 9.2
gcd(r + 1, n). gcd(r − 1, n) = n.j
(2α1 pa11 pa22 . . . paq q ).(2α2 pb11 pb22 . . . pbqq ) = j.2α−1 pt11 pt22 . . . ptqq
(2α1 +α2 pa11 +b1 pa22 +b2 . . . paq q +bq ) = j.2α−1 pt11 pt22 . . . ptqq
Now, since ∀ i either ai = 0 or bi = 0 ⇒ ai + bi = ti
t
2
(9.3)
α1 +α2
= j.2
α−1
pt11 pt22 . . . pqq
a +bq
pa11 +b1 pa22 +b2 . . . pq q
2α1 +α2 = j.2α−1 .
As α1 ≤ (α − 1) and α2 ≤ (α − 1) and since either α1 ≤ 1 or α2 ≤ 1,
⇒ (α − 1) ≤ (α1 + α2 ) ≤ α
equation 9.3 ⇒ j = 2α1 +α2 −α+1 ⇒ 1 ≤ j ≤ 2.
14
ZAHID RAZA, IMRAN AND BIJAN DAVVAZ
(3) By Theorem 7.8, Nr is the number of divisors of
gcd(r − 1, n). gcd(r + 1, n)
n
and Nr is the number of divisor of j computed above. Therefore,
Nr = 1 if j = 1 and Nr = 2 if j = 2.
In either case,
ZDiv(r + 1)
|
|≤ 2.
hr − 1i
Finally according to the proof of (2), j = 2 if and only if
α1 = α − 1, α2 = 1 or α1 = 1, α2 = α − 1.
If α1 = α − 1, α2 = 1 then 2α−1 divides (r + 1).
If α1 = 1, α2 = α − 1 then 2α−1 divides (r − 1).
⇒ r ≡ ±1 (mod 2α−1 ).
Example 9.6. Let G = QD48 where m = 23 .3, α = 3, then
2
U24 = {1, 5, 7, 11, 13, 17, 19, 23} and R24
= {1, 5, 7, 11, 13, 17, 19, 23}.
2
(1) Let r = 5 ∈ R24 then hr − 1i = h4i = {0, 4, 8} = B1 and
ZDiv(r + 1) = ZDiv6 = {0, 2, 4, 6, 8, 10} = B2 clearly, h4i ≤ ZDiv6.
iv(6) = {0,2,4,6,8,10} = B2 = x + B ∀ x ∈ B
(2) For r = 5, ZDh4i
1
2
{0,4,8}
B1
0 + B1 = {0, 4, 8} = B1 , 2 + B1 = {2, 6, 10} = B3
4 + B1 = {4, 8, 0} = B1 , 6 + B1 = {6, 10, 2} = B3
8 + B1 = {8, 0, 4} = B1 , 10 + B1 = {10, 2, 6} = B3
= {B1 , B2 } ⇒ Nr = N5 = 2.
U24 action on ZDiv6
h4i
3
(3) Since, m = 2 .3 ⇒ l = 3 > 1 is odd and r = 5 ≡ 1 (mod 22 ).
Thus N5 = 2 as calculated in (2).
t
Corollary 9.7. Suppose that m = 2α pt11 pt22 . . . pqq , where the pi are distinct odd
primes and α ≥ 2. Then the number of equivalence classes C4k of Aut2 (QD8k ) is
given by
q+1
2
α=2
C4k =
q+2
2
α ≥ 3.
2
Example 9.8. Let G = QD384 . Thus R192
= {1, 17, 31, 47, 49, 65, 79, 95, 97, 113, 127,
2
|= 24 . It is easy to check that the values r =
143, 145, 161, 175, 191} ⇒| R192
1, 65, 97, 161 satisfy r ≡ 1 (mod 25 ) and the values r = 31, 95, 127, 191 satisfy
r ≡ −1 (mod 25 ). So, Cm =| {1, 31, 65, 95, 97, 127, 161, 191} |= 23 .
Now for θ an involution, the set of twisted involutions is given in the followings
porosities:
ON THE STRUCTURE OF INVOLUTIONS AND SYMMETRIC SPACES OF QUASI DIHEDRAL GROUP
15
Proposition 9.9. If θ2 =id, then the set of twisted involutions is given as
R = {ai | (r + 1)i ≡ −2s (mod n)} ∪ {ai b | (r − 1)i ≡ −s (mod n)}
2
Proof. Since, θ is an involution therefore, r ∈ Rm
and s ∈ ZDiv(r + 1). To find the
i
i
i −1
ri+(2−n)s
set R, , let a ∈ G, then θ(a ) = (a ) ⇒ a
= a−i ⇒ ri + (2 − n)s = −i
⇒ (r + 1)i + (2 − n)s = 0 ⇒ (r + 1)i ≡ −2s (mod n).
Now for ai b ∈ G, we have, θ(ai b) = (ai b)−1 ⇒ ari+s b = ai b ⇒ ri + s = i
⇒ (r − 1)i ≡ −s (mod n), hence R = {ai | (r + 1)i ≡ −2s (mod n)} ∪ {ai b |
(r − 1)i ≡ −s (mod n)}.
Corollary 9.10. If θ2 =id, m = 4k, n = 2k and k is even number, then Q is a
subgroup of hai and the natural isomorphism ψ : hai −→ Zn gives using Theorem
8.1 and Corollary 8.7 and 8.11
(1) If | Q |= 2, then ψ(Q) = hki.
(2) If | Q |= k2 then
hd1 i If n ∤ r + 2s − 1 and d1 = gcd(r + 2s − 1, n)
ψ(Q) =
h4i If n | r + 2s − 1 and n = gcd(r + 2s − 1, n).
(3) If | Q |= k and gcd(r + 2s − 1, n) = d1 then
ψ(Q) =
h d21 i If {(r + 2s − 1) ∩ [l(1 − r) − s]} = ∅
hd1 i If {(r + 2s − 1) ∩ [l(1 − r) − s]} =
6 ∅,
where l is an even number.
Corollary 9.11. If θ2 =id, m = 4k, n = 2k and k is even number, then Q is a
subgroup of hai and the natural isomorphism ψ : hai −→ Zn gives using Theorem
8.1 and Corollary 8.9 and 8.12
(1) If s is even, then
h0i If r = 1, s = 0, | Q |= 1
ψ(Q) =
h2i If | Q |= k.
(2) If s is odd, then
hki | Q |= 2 If n | 1 − r and n | s − k
ψ(Q) =
h1i | Q |= n If n ∤ r + 2s − 1 and n ∤ 1 − r.
Corollary 9.12. In any Case R 6= Q.
References
[1] A. Bishop, C. Cyr, J. Hutchens, C. May, N. Schwartz, and B. Turner (2013). ”On involutions
and generalized symmetric spaces of dicyclic groups”. arXiv:1310.0121v1
[2] Anders Bjorner and Francesco Brenti (2005). ”Combinatorics of coxter groups” . Gradate Text
in Mathematics, vol.231, Springer, New York.
16
ZAHID RAZA, IMRAN AND BIJAN DAVVAZ
[3] Nicolas Bourbaki (2002). ”Lie groups and Lie algebras”. Chapter 4-6, Elements of Mathematics
(Berlin), Springer-Verlag , Berlin , Translated from the (1968) French original by Andrew
Pressley.
[4] K. K. A. Cunningham, T. J. Edgar, A. G. Helminck, B. F. Jones, OH, R. Schwell , J.F.
Vasquez. (2012). ”On the Structure of Involution and Symmetric Spaces of Dihedral Groups”.
arXiv:1205.3207v1.
[5] A. G. Helminck (1994). Symmetric k- varities. Proc. Sympos. Pure Math. 56, no. 1, 233-279.
[6] A. G. Helminck (1988). ”Algebraic groups with a computing pair of involutions and semisimple symmetric spaces”. Adv. in Math 71 , 21-91.
[7] A. G. Helminck and S.P. Wang (1993). ”On rationality properties of involutions of reductive
groups”. Adv. in Math. 99, 26-96.
[8] Gary L. Walls (1986). ”Automorphism groups”. Amer. Math. Monthly 93, no. 6, 459-462.
Department of Mathematics, College of Sciences,University of Sharjah, UAE.
E-mail address: [email protected]
Department of Mathematics, National University of Computer and Emerging
Sciences, Lahore, Pakistan.
E-mail address: imran [email protected]
Department of Mathematics, Yazd University, Yazd, Iran
E-mail address: [email protected]
| 4math.GR
|
Tree-based Arithmetic and Compressed
Representations of Giant Numbers
Paul Tarau
arXiv:1301.0114v1 [cs.PL] 1 Jan 2013
Department of Computer Science and Engineering
University of North Texas
e-mail: [email protected]
Abstract. Can we do arithmetic in a completely different way, with a
radically different data structure? Could this approach provide practical benefits, like operations on giant numbers while having an average
performance similar to traditional bitstring representations?
While answering these questions positively, our tree based representation
described in this paper comes with a few extra benefits: it compresses
giant numbers such that, for instance, the largest known prime number as
well as its related perfect number are represented as trees of small sizes.
The same also applies to Fermat numbers and important computations
like exponentiation of two become constant time operations.
At the same time, succinct representations of sparse sets, multisets and
sequences become possible through bijections to our tree-represented natural numbers.
1
Introduction
If extraterrestrials were to do arithmetic computations, would one assume that
their numbering system should look the same? At a first thought, one might
be inclined to think so, maybe with the exception of the actual symbols used
to denote digits or the base of the system likely to match numbers of fingers
or some other anthropocentric criteria. After some thinking about the endless
diversity of the universe (or unconventional models of Peano’s axioms), one might
also consider the possibility that the departure from our well-known number
representations could be more significant.
With more terrestrial uses in mind, one would simply ask: is it possible to do
arithmetic differently, possibly including giant numbers and a radically different
underlying data structure, while maintaining the same fundamentals – addition,
multiplication, exponentiation, primality, Peano’s axioms behaving in the same
way? Moreover, can such exotic arithmetic be as efficient as binary arithmetic,
while possibly supporting massively parallel execution?
We have shown in [1] that type classes and polymorphism can be used to
share fundamental operations between natural numbers, finite sequences, sets
and multisets. As a consequence of this line of research, we have discovered
that it is possible to transport the recursion equations describing binary number
arithmetics on natural numbers to various tree types.
However, the representation proposed in this paper is different. It is arguably
simpler, more flexible and likely to support parallel execution of operations. It
also allows to easily compute average and worse case complexity bounds. We
will describe it in full detail in the next sections, but for the curious reader, it is
essentially a recursively self-similar run-length encoding of bijective base-2 digits.
The paper is organized as follows. Section 2 describes our type class used to
share generic properties of natural numbers. Section 3 describes binary arithmetic operations that are specialized in section 4 to our compressed tree representation. Section 5 describes efficient tree-representations of some important
number-theoretical entities like Mersenne, Fermat and perfect numbers. Section
6 shows that sparse sets, multisets and lists have succinct tree representation.
Section 7 describes generic isomorphisms between data types, centered around
transformations of instances of our type class to their corresponding sets, multisets and lists. Section 8 shows interesting complexity reductions in other computations and section 9 compares the performance of our tree-representation with
conventional ones. Section 10 discusses related work. Section 11 concludes the
paper and discusses future work.
To provide a concise view of our compressed tree data type and its operations, we will use the strongly typed functional language Haskell as a precise means to provide an executable specification. We have adopted a literate
programming style, i.e. the code contained in the paper forms a self-contained
Haskell module (tested with ghc 7.4.1), also available as a separate file at http:
//logic.cse.unt.edu/tarau/research/2013/giant.hs . Also, a Scala package implementing the same tree-based computation is available from http:
//code.google.com/p/giant-numbers/. We hope that this will encourage the
reader to experiment interactively and validate the technical correctness of our
claims.
We mention, for the benefit of the reader unfamiliar with Haskell, that a
notation like f x y stands for f (x, y), [t] represents sequences of type t and a
type declaration like f :: s -> t -> u stands for a function f : s×t → u (modulo Haskell’s “currying” operation, given the isomorphism between the function
spaces s × t → u and s → t → u). Our Haskell functions are always represented
as sets of recursive equations guided by pattern matching, conditional to constraints (simple arithmetic relations following | and before the = symbol). Locally
scoped helper functions are defined in Haskell after the where keyword, using
the same equational style. The composition of functions f and g is denoted f .
g. It is also customary in Haskell, when defining functions in an equational style
(using =) to write f = g instead of f x = g x (“point-free” notation). The use
of Haskell’s “call-by-need” evaluation allows us to work with infinite sequences,
like the [0..] infinite list notation, corresponding to the set N itself.
Our literate Haskell program is organized as the module Giant depending on
a few packages, as follows:
{-# LANGUAGE NoMonomorphismRestriction #-}
-- needed to define toplevel isomorphisms
module Giant where
-- cabal install data-ordlist is required before importing this
import Data.List.Ordered
import Data.List hiding (unionBy)
import System.CPUTime
2
Sharing axiomatizations with type classes
Haskell’s type classes [2] are a good approximation of axiom systems as they
describe properties and operations generically i.e. in terms of their action on
objects of a parametric type. Haskell’s type instances approximate interpretations [3] of such axiomatizations by providing implementations of the primitive
operations, with the added benefit of refining and possibly overriding derived
operations with more efficient equivalents.
We start by defining a type class that abstracts away properties of binary
representations of natural numbers.
The class N assumes only a theory of structural equality (as implemented by
the class Eq in Haskell). It implements a representation-independent abstraction
of natural numbers, allowing us to compare our tree representation with “ordinary” natural numbers represented as non-negative arbitrary large Integers in
Haskell, as well as with a binary representation using bijective base-2 [4].
class Eq n ⇒ N n where
An instance of this class is required to implement the following 6 primitive
operations:
e :: n
o,o’,i,i’ :: n→n
o_ :: n→Bool
The constant function e can be seen as representing the empty sequence of binary
digits. With the usual representation of natural numbers, e will be interpreted
as 0. The constructors o and i can be seen as applying a function that adds a
0 or 1 digit to a binary string on {0,1}∗ . The deconstructors o’ and i’ undo
these operations by removing the corresponding digit. The recognizer o detects
that the constructor o is the last one applied, i.e. that the“string ends with the
0 symbol. It will be interpreted on N as a recognizer of odd numbers.
This type class also endows its instances with generic implementations of the
following derived operations:
e_,i_ :: n→Bool
e_ x = x == e
i_ x = not (e_ x | | o_ x)
with structural equality used implicitly in the definition of the recognizer predicate for empty sequences e and with the assumption that the domain is exhausted by the three recognizers in the definition of the recognizer i of sequences
ending with 1, representing even positive numbers in bijective base 2.
Successor s and predecessor s’ functions are implemented in terms of these
operations as follows:
s,s’ :: n→n
s x | e_ x = o x
s x | o_ x = i (o’ x)
s x | i_ x = o (s (i’ x))
s’ x | x == o e = e
s’ x | i_ x = o (i’ x)
s’ x | o_ x = i (s’ (o’ x))
By looking at the code, one might notice that our generic definitions of operations
mimic recognizers, constructors and destructors for bijective base-2 numbers, i.e.
sequences in the language {0, 1}∗ , similar to binary numbers, except that 0 is
represented as the empty sequence and left-delimiting by 1 is omitted.
Proposition 1 Assuming average constant time for recognizers, constructors
and destructors e ,o o,i ,i,o’,i’, successor and predecessor s and s’ are constant time, on the average.
Proof. Clearly the first two rules are constant time for both s and s’ as they do
not make recursive calls. To show that the third rule applies recursion a constant
number of times on the average, we observe that the recursion steps are exactly
given by the number of 0s or 1s that a (bijective base-2 number) ends with. As
only half of them end with a 0 and another half of those end with another 0 etc.
one can see that the average number of 0s is bounded by 21 + 14 + . . . = 1. The
same reasoning applies to the average number of 1s a number can end with.
The infinite stream of generic natural numbers is generated by iterating over the
successor operation s:
allFrom :: n→[n]
allFrom x = iterate s x
3
Efficient arithmetic operations, generically
We will first show that all fundamental arithmetic operations can be described
in this abstract, representation-independent framework. This will make possible
creating instances that, on top of symbolic tree representations, provide implementations of these operations with asymptotic efficiency comparable to the
usual bitstring operations.
We start with addition (add) and subtraction (sub):
add
add
add
add
add
add
:: n→n→n
x y | e_ x
x y | e_ y
x y | o_ x
x y | o_ x
x y | i_ x
=y
=x
&& o_ y = i (add (o’ x) (o’ y))
&& i_ y = o (s (add (o’ x) (i’ y)))
&& o_ y = o (s (add (i’ x) (o’ y)))
add x y | i_ x && i_ y = i (s (add (i’ x) (i’ y)))
sub
sub
sub
sub
sub
sub
:: n→n→n
x y | e_ y
y x | o_ y
y x | o_ y
y x | i_ y
y x | i_ y
=x
&& o_
&& i_
&& o_
&& i_
x
x
x
x
=
=
=
=
s’ (o (sub (o’ y) (o’ x)))
s’ (s’ (o (sub (o’ y) (i’ x))))
o (sub (i’ y) (o’ x))
s’ (o (sub (i’ y) (i’ x)))
It is easy to see that addition and subtraction are implemented generically, with
asymptotic complexity proportional to the size of the operands. Comparison
provides the expected total order of N on our type class:
cmp :: n→n→Ordering
cmp x y | e_ x && e_ y
cmp x _ | e_ x = LT
cmp _ y | e_ y = GT
cmp x y | o_ x && o_ y
cmp x y | i_ x && i_ y
cmp x y | o_ x && i_ y
down EQ = LT
down r = r
cmp x y | i_ x && o_ y
up EQ = GT
up r = r
= EQ
= cmp (o’ x) (o’ y)
= cmp (i’ x) (i’ y)
= down (cmp (o’ x) (i’ y)) where
= up (cmp (i’ x) (o’ y)) where
And based on it one can define the minimum min2 and maximum max2 functions
as follows:
min2,max2 :: n→n→n
min2 x y = if LT==cmp x y then x else y
max2 x y = if LT==cmp x y then y else x
Next, we define multiplication:
mul
mul
mul
mul
m
m
m
:: n→n→n
x _ | e_ x
_ y | e_ y
x y = s (m
x y | e_ x
x y | o_ x
x y | i_ x
=e
=e
(s’ x) (s’ y)) where
=y
= o (m (o’ x) y)
= s (add y (o (m (i’ x) y)))
as well as double of a number db and half of an even number hf, having both
simple expressions:
db,hf :: n→n
db = s’ . o
hf = s . i’
Power is defined as follows:
pow :: n→n→n
pow _ y | e_ y = o e
pow x y | o_ y = mul x (pow (mul x x) (o’ y))
pow x y | i_ y = mul (mul x x) (pow (mul x x) (i’ y))
together with more efficient special instances, exponent of 2 (exp2) and multiplication by a power of 2 (leftshift):
exp2 :: n→n
exp2 x | e_ x = o e
exp2 x = db (exp2 (s’ x))
leftshift :: n→n→n
leftshift x y = mul (exp2 x) y
Finally, division and reminder on N is a bit trickier but can be expressed generically as well:
div_and_rem :: n→n→(n,n)
div_and_rem x y | LT == cmp x y = (e,x)
div_and_rem x y | not (e_ y) = (q,r) where
(qt,rm) = divstep x y
(z,r) = div_and_rem rm y
q = add (exp2 qt) z
divstep :: N n ⇒ n→n→(n,n)
divstep n m = (q, sub n p) where
q = try_to_double n m e
p = mul (exp2 q) m
try_to_double x y k =
if (LT==cmp x y)
then s’ k
else try_to_double x (db y) (s k)
divide,reminder :: n→n→n
divide n m = fst (div_and_rem n m)
reminder n m = snd (div_and_rem n m)
And for the reader curious by now about how this maps to “arithmetic as
usual”, here is an instance built around the (arbitrary length) Integer type,
also usable as a witness on the time/space complexity of our operations.
instance N Integer where
e=0
o_ x = odd x
o x = 2∗x+1
o’ x | odd x && x > 0 = (x-1) ‘div‘ 2
i
x = 2∗x+2
i’ x | even x && x > 0 = (x-2) ‘div‘ 2
An instance mapping our abstract operations to actual constructors, follows in
the form of the datatype B
data B = B | O B | I B deriving (Show, Read, Eq)
instance N B where
e=B
o=O
i=I
o’ (O x) = x
i’ (I x) = x
o_ (O _) = True
o_ _ = False
One can try out various operations on these instances:
*Giant> mul 10 5
50
*Giant> exp2 5
32
*Giant> add (O B) (I (O B))
O (I B)
4
Computing with our compressed tree representations
We will now show how our shared axiomatization framework can be implemented
as a new, somewhat unusual instance, that brings the ability to do arithmetic
computations with trees.
First, we define the data type for our tree represented natural numbers:
data T = T | V T [T] | W T [T] deriving (Eq,Show,Read)
The intuition behind this “union type” is the following:
– The type T corresponds to an empty sequence
– the type V x xs counts the number x of o applications followed by an alternation of similar counts of i and o applications
– the type W x xs counts the number x of i applications followed by an alternation of similar counts of o and i applications
– the same principle is applied recursively for the counters, until the empty
sequence is reached
One can see this process as run-length compressed bijective base-2 numbers, represented as trees with either empty leaves or at least one branch, after applying
the encoding recursively. First we define the 6 primitive operations:
instance N T where
e=T
o T = V T []
o (V x xs) = V (s x) xs
o (W x xs) = V T (x:xs)
i T = W T []
i (V x xs) = W T (x:xs)
i (W x xs) = W (s x) xs
o’ (V T []) = T
o’ (V T (x:xs)) = W x xs
o’ (V x xs) = V (s’ x) xs
i’ (W T []) = T
i’ (W T (x:xs)) = V x xs
i’ (W x xs) = W (s’ x) xs
o_ (V _ _ ) = True
o_ _ = False
Next, we override two operations involving exponents of 2 as follows
exp2 = exp2’ where
exp2’ T = V T []
exp2’ x = s (V (s’ x) [])
leftshift = leftshift’ where
leftshift’ _ T = T
leftshift’ n y | o_ y = s (vmul n (s’ y))
leftshift’ n y | i_ y = s (vmul (s n) (s’ y))
The leftshift’ operation uses an efficient implementation, specialized for the
type T, of the repeated application (n times) of constructor o, over the second
argument of the function vmul:
vmul
vmul
vmul
vmul
T
n
n
n
y=y
T = V (s’ n) []
(V y ys) = V (add (s’ n) y) ys
(W y ys) = V (s’ n) (y:ys)
Note that such overridings take advantage of the specific encoding, as a result of simple number theoretic observations. For instance, the operation exp2’
works in time proportional to s and s’, that can be shown to be constant on the
average. The more complex leftshift’ operation observes that repeated application of the o operation, when adjusted based on being even or odd, provides
an efficient implementation of multiplication with an exponent of 2.
It is convenient at this point, as we target a diversity of interpretations materialized as Haskell instances, to provide a polymorphic converter between two
different instances of the type class N as well as their associated lists, implemented by structural recursion over the representation to convert. The function
view allows importing a wrapped object of a different instance of N, generically.
view
view
view
view
:: (N a,N b)⇒a→b
x | e_ x = e
x | o_ x = o (view (o’ x))
x | i_ x = i (view (i’ x))
We can specialize view to provide conversions to our three data types, each
denoted with the corresponding lower case letter, tt t, b and n for the usual
natural numbers.
t :: (N n) ⇒ n → T
t = view
b :: (N n) ⇒ n → B
b = view
n :: (N n) ⇒ n → Integer
n = view
One can try them out as follows:
*Giant> t 42
W (V T []) [T,T,T]
*Giant> b it
I (I (O (I (O B))))
*Giant> n it
42
While space constraints forbid us from providing the correctness proofs of
operations like exp2’ and leftshift’, we are able to illustrate their expected
usage as follows:
*Giant> t 5
V T [T]
*Giant> exp2 it
W T [V (V T []) []]
*Giant> n it
32
*Giant> t 10
W (V T []) [T]
*Giant> leftshift it (t 1)
W T [W T [V T []]]
*Giant> n it
1024
5
Efficient representation of some important
number-theoretical entities
Let’s first observe that Fermat, Mersenne and perfect numbers have all compact
expressions with our tree representation of type T.
fermat n = s (exp2 (exp2 n))
mersenne p = s’ (exp2 p)
perfect p = s (V q [q]) where q = s’ (s’ p)
And one can also observe that this contrasts with both the Integer representation and the bijective base-2 numbers B:
*Giant> mersenne (b 127)
O (O (O (O (O (O (O (O (O (O (O (O (O (O (O
... a few lines of Os and Is
)))))))))))))))))))))))))))))))))))))))))
*Giant> mersenne (n 127)
170141183460469231731687303715884105727
*Giant> mersenne (t 127)
V (W (V T [T]) []) []
The largest known prime number, found by the GIMP distributed computing
project [5] is the 45-th Mersenne prime = 243112609 − 1. It is defined in Haskell
as follows:
-- its exponent
prime45 = 43112609 :: Integer
-- the actual Mersenne prime
mersenne45 = s’ (exp2 (t p)) where
p = prime45::Integer
While it has a bitsize of 43112609, we have observed that its compressed tree
representation using our type T is rather small:
*Giant> mersenne45
V (W T [V (V T []) [],T,T,T,W T [],V T [],T,W T [],W T [],T,V T [],T,T]) []
One the other hand, displaying it with a decimal or binary representation would
take millions of digits.
And by folding replicated subtrees to obtain an equivalent DAG representation, one can save even more memory. Figure 1 shows this representation,
involving only 6 nodes.
It is interesting to note that similar compact representations can also be
derived for perfect numbers. For instance, the largest known perfect number,
derived from the largest known Mersenne prime as 243112609−1 (243112609 − 1), is:
perfect45 = perfect (t prime45)
V
0
W
1
6
0
2
3
4
7
10
5
11
8
9
V
12 13
0
V
W
0
0
0
0 0 0
T
Fig. 1. Largest known prime number: the 45-th Mersenne prime, represented as
a DAG
W
1
2
W
0
2
0
5
8
10
11
0
V
1
4
9
3
W
1
0
6
7
0
5
10 4
V
1 0
0
7 8
W
0 0 0
0
0
0 0 0 0
T
Fig. 2. Largest known perfect number
1
2 3
6
9
11 12
Fig. 2 shows its DAG representation involving only 7 nodes. Similarly, the
11
largest Fermat number that has been factored so far, F11=22 + 1 is compactly
represented as
*Giant> fermat (t 11)
V T [T,V T [W T [V T []]]]
By contrast, its (bijective base-2) binary representation consists of 2048 digits.
6
Representing sparse sets, multisets and lists
We will now describe bijective mappings between collection types as well as
a Gödel numbering scheme putting them in bijection with natural numbers.
Interestingly, natural number encodings for sparse instances of these collections
will have space-efficient representations as natural numbers of type T, in contrast
with bitstring or conventional Integer-based representations.
The type class Collections will convert between natural numbers and lists,
by using the bijection f (x, y) = 2x (2y + 1), implemented by the function c and
its first and second projections c’ and c’’, inverting it.
class (N n) ⇒ Collections n where
c :: n→n→n
c’,c’’ :: n→n
c x y = mul (exp2 x) (o y)
c’ x | not (e_ x) = if o_ x then e else s (c’ (hf x))
c’’ x | not (e_ x) = if o_ x then o’ x else c’’ (hf x)
The bijection between natural numbers and lists of natural numbers, to list
and its inverse from list apply repeatedly c and respectively c’ and c’’.
to_list :: n→[n]
to_list x | e_ x = []
to_list x = (c’ x) : (to_list (c’’ x))
from_list:: [n]→n
from_list [] = e
from_list (x:xs) = c x (from_list xs)
Incremental sums are used to transform arbitrary lists to multisets and sets,
inverted by pairwise differences.
list2mset, mset2list, list2set, set2list :: [n]→[n]
list2mset ns = tail (scanl add e ns)
mset2list ms = zipWith sub ms (e:ms)
list2set = (map s’) . list2mset . (map s)
set2list = (map s’) . mset2list . (map s)
By composing with natural number-to-list bijections, we obtain bijections to
multisets and sets.
to_mset, to_set :: n→[n]
from_mset, from_set :: [n]→n
to_mset = list2mset . to_list
from_mset = from_list . mset2list
to_set = list2set . to_list
from_set = from_list . set2list
We will add the usual instances to the type class Collections. A simple numbertheoretic observation connecting 2n and n applications of the constructor i,
implemented by the function vmul, allows a shortcut that speeds up the bijection
from lists to natural numbers, by overriding the functions c, c’, c’’ in instance
T.
instance Collections B
instance Collections Integer
instance Collections T where
c = cons where
cons n y = s (vmul n (s’ (o y)))
c’ = hd where
hd z | o_ z = T
hd z = s x where
V x _ = s’ z
c’’ = tl where
tl z | o_ z = o’ z
tl z = f xs where
V _ xs = s’ z
f [] = T
f (y:ys) = s (i’ (W y ys))
As the following example shows, trees of type T offer a significantly more compact
representation of sparse sets.
*Giant> from_set (map t [1,100,123,234])
W (V T []) [V T [T,W T [],T],T,V T [V T [],T],T,V T [W T [],T,T]]
*Giant> from_set (map n [1,100,123,234])
27606985387162255149739023449108112443629669818608757680508075841159170
*Giant> from_set (map b [1,100,123,234])
I (I (O (O (O (O (O (O (O (O (O (O (O (O (O (O (O (O (O ...
... a few lines ...
..)))))))))))
Note that a similar compression occurs for sets of natural numbers with only a
few elements missing, as they have the same representation size with type T as
the dual of their sparse counterpart.
Giant> from_set ([1,3,5]++[6..220])
3369993333393829974333376885877453834204643052817571560137951281130
*Giant> t it
W (V T []) [T,T,T,W (W T []) [T,T,T,T]]
*Giant> dual it
V (V T []) [T,T,T,W (W T []) [T,T,T,T]]
*Giant> to_set it
[T,V T [],W T [T],W T [T,W T [],T,T]]
*Giant> map n it
[0,1,4,220]
As an application, we can define bitwise operations on our natural numbers
by borrowing the corresponding ordered set operations, provided in Haskell by
the package Data.List.Ordered.
First we define the type class BitwiseOperations and the higher order function l op transporting a binary operation from ordered sets to natural numbers.
class Collections n ⇒ BitwiseOperations n where
l_op :: ([n]→[n]→[n])→n→n→n
l_op op x y = from_set (op (to_set x) (to_set y))
Next we define the bitwise and, or and xor operations:
l_and,l_or,l_xor,l_dif:: n→n→n
l_and = l_op (isectBy cmp)
l_or = l_op (unionBy cmp)
l_xor = l_op (xunionBy cmp)
l_dif = l_op (minusBy cmp)
More complex operations like the ternary if-the-else can be defined as a
combination of binary operations:
l_ite
l_ite
ite
c
e
:: n→n→n→n
x y z = from_set (ite (to_set x) (to_set y) (to_set z)) where
d a b = xunionBy cmp e b where
= xunionBy cmp a b
= isectBy cmp c d
Finally, bitwise negation (requiring additional parameter l defining the bitlength
of the operand) can be defined using set complement with respect to 2l − 1,
corresponding to the set of all elements up to l.
l_not :: Integer→n→n
l_not l x | xl ≤ l = from_set (minusBy cmp ms xs) where
xs = to_set x
xl = genericLength xs
ms = genericTake l (allFrom e)
We will next generalize the iso-functor mechanism implemented generically by
l op that transports operations back an forth between data types as a special
data type, consisting of two higher order functions inverse to each other.
7
Isomorphisms between data types, generically
Along the lines of [6] we can define isomorphisms between data types as follows:
data Iso a b = Iso (a→b) (b→a)
from (Iso f _) = f
to (Iso _ f’) = f’
Morphing between data types as well as “lending” operations from one to another
is provided by the combinators as, land1 and land2:
as that this x = to that (from this x)
lend1 op1 (Iso f f’) x = f’ (op1 (f x))
lend2 op2 (Iso f f’) x y = f’ (op2 (f x) (f y))
Assuming that the Haskell option NoMonomorphismRestriction is set on, we
can now define generically “virtual types” centered around type class N:
nat = Iso id id
list = Iso from_list to_list
mset = Iso from_mset to_mset
set = Iso from_set to_set
This results in a small “embedded language” that morphs between various instances of type class N and their corresponding list, multiset and set types, as
follows:
*Giant> as set nat 1234
[1,4,6,7,10]
*Giant> map t it
[V T [],W T [T],W (V T []) [],V (W T []) [],W (V T []) [T]]
*Giant> as nat set it
W (V T []) [V T [],T,T,V T [],V T []]
*Giant> n it
1234
*Giant> lend1 s set [0,2,3]
[1,2,3]
*Giant> lend2 add set [0,2,3] [4,5]
[0,2,3,4,5]
8
Complexity reduction in other computations
A number of other, somewhat more common computations also benefit from
our data representations. The type class SpecialComputations groups them
together and provides their bitstring inspired generic implementations.
The function dual flips o and i operations for a natural number seen as
written in bijective base 2.
class Collections
dual :: n→n
dual x | e_ x =
dual x | o_ x =
dual x | i_ x =
n ⇒ SpecialComputations n where
e
i (dual (o’ x))
o (dual (i’ x))
The function bitsize computes the number of applications of the o and i operations:
bitsize
bitsize
bitsize
bitsize
:: n→n
x | e_ x = e
x | o_ x = s (bitsize (o’ x))
x | i_ x = s (bitsize (i’ x))
The function repsize computes the representation size, which defaults to the
bit-size in bijective base 2:
-- representation size - defaults to bitsize
repsize :: n→n
repsize = bitsize
The functions decons and cons provide bijections between N−{0} and N×N and
can be used as an alternative mechanism for building bijections between lists,
multisets and sets of natural numbers and natural numbers. They are based on
separating o and i applications that build up a natural number represented in
bijective base 2.
decons ::n→(n,n)
cons :: (n,n)→n
decons z | o_ z = (x,y) where
x0 = s’ (ocount z)
y = otrim z
x = if e_ y then (s’.o) x0 else x0
decons z | i_ z = (x,y) where
x0 = s’ (icount z)
y = itrim z
x = if e_ y then (s’.i) x0 else x0
cons
cons
cons
cons
cons
(x,y)
(x,y)
(x,y)
(x,y)
(x,y)
|
|
|
|
|
e_
o_
i_
o_
i_
x
x
x
y
y
&& e_ y = s e
&& e_ y = itimes (s (i’ (s x))) e
&& e_ y = otimes (s (o’ (s x))) e
= itimes (s x) y
= otimes (s x) y
Implementing decons and cons requires counting the number of applications of
o and i provided by ocount and icount, as well trimming the applications of o
and i, performed by otrim and itrim.
ocount,icount,otrim,itrim :: n→n
ocount x | o_ x = s (ocount (o’ x))
ocount _ = e
icount x | i_ x = s (icount (i’ x))
icount _ = e
otrim x | o_ x = otrim (o’ x)
otrim x = x
itrim x | i_ x = itrim (i’ x)
itrim x = x
otimes,itimes :: n→n→n
otimes x y | e_ x = y
otimes x y = otimes (s’ x) (o y)
itimes x y | e_ x = y
itimes x y = itimes (s’ x) (i y)
An alternative bijection between natural numbers and lists of natural numbers, to list’ and its inverse from list’ is obtained by applying repeatedly
cons and respectively decons.
to_list’ :: n→[n]
to_list’ x | e_ x = []
to_list’ x = hd : (to_list’ tl) where (hd,tl)=decons x
from_list’ :: [n]→n
from_list’ [] = e
from_list’ (x:xs) = cons (x,from_list’ xs)
One can observe the significant reduction of asymptotic complexity with respect to the default operations provided by the type class SpecialComputations
when overriding with tbitsize and tdual in instance T.
instance SpecialComputations Integer
instance SpecialComputations B
instance SpecialComputations T where
bitsize = tbitsize where
tbitsize T = T
tbitsize (V x xs) = s (foldr add1 x xs)
tbitsize (W x xs) = s (foldr add1 x xs)
add1 x y = s (add x y)
dual = tdual where
tdual T = T
tdual (V x xs) = W x xs
tdual (W x xs) = V x xs
repsize = tsize where
tsize T = T
tsize (V x xs) = s (foldr add T (map tsize (x:xs)))
tsize (W x xs) = s (foldr add T (map tsize (x:xs)))
The replacement with special purpose code for the cons / decons functions is
even more significant:
decons
decons
decons
decons
cons
cons
cons
cons
cons
(V
(V
(W
(W
x
x
x
x
[]) = ((s’.o) x,T)
(y:ys)) = (x,W y ys)
[]) = ((s’.i) x,T)
(y:ys)) = (x,V y ys)
(T,T) = V T []
(x,T) | o_ x =
(x,T) | i_ x =
(x,V y ys) = W
(x,W y ys) = V
W (i’ (s x)) []
V (o’ (s x)) []
x (y:ys)
x (y:ys)
One can also see that their complexity is now proportional to s and s’ given
that the V and W operations perform in constant time the work of otimes and
itimes. The following example illustrates their work:
*Giant> map to_list’ [0..20]
[[],[0],[1],[2],[0,0],[0,1],[3],[4],[0,2],[0,0,0],[1,0],[1,1],
[0,0,1],[0,3],[5],[6],[0,4],[0,0,2],[1,2],[1,0,0],[0,0,0,0]]
*Giant> map from_list’ it
[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20]
Note the shorter lists created close to powers of 2, coming from the longer sequences of consecutive o and im operations in that region.
9
A performance comparison
Our performance measurements (run on a Mac Air with 8GB of memory and an
Intel i7 processor) serve two objectives:
1. to show that, on the average, our tree based representations perform on a
blend of arithmetic computations within a small constant factor compared
with conventional bitstring-based computations
2. to show that on interesting special computations they outperform the conventional ones due to the much lower asymptotic complexity of such operations on data type T.
Objective 1 is served by the Ackerman function that exercises the successor
and predecessor functions quite heavily, the prime generation and the LucasLehmer primality test for Mersenne numbers that exercise a blend of arithmetic
operations.
Objective 2 is served by the other benchmarks that take advantage of the
overriding by instance T of operations like exp2 and bitsize, as well as the
compressed representation of large numbers like the 45-th Mersenne prime and
perfect numbers. In some cases the conventional representations are unable to
Benchmark
Integer binary type B tree type T
Ackermann 3 7
9418
7392
12313
exp2 (exp2 14)
23
315
0
sparse set encoding
7560
2979
45
bitsize of Mersenne 45
?
?
0
bitsize of Perfect 45
?
?
2
generating primes
2722
2567
3591
Mersenne prime tests
6925
6431
15037
Fig. 3. Time (in ms.) on a few benchmarks
run these benchmarks within existing computer memory and CPU-power limitations (marked with ? in the comparison table of Fig. 3). In other cases, like
in the sparse set encoding benchmark, data type T performs significantly faster
than binary representations.
Together they indicate that our tree-based representations are likely to be
competitive with existing bitstring-based packages on typical computations and
significantly outperform them on some number-theoretically interesting computations. While the code of the benchmarks is omitted due to space constraints,
it is part of the companion Haskell file at http://logic.cse.unt.edu/tarau/
Research/2013/giant.hs.
10
Related work
We will briefly describe here some related work that has inspired and facilitated
this line of research and will help to put our past contributions and planned
developments in context.
Several notations for very large numbers have been invented in the past.
Examples include Knuth’s arrow-up notation [7] covering operations like the
tetration (a notation for towers of exponents). In contrast to our tree-based natural numbers, such notations are not closed under addition and multiplication,
and consequently they cannot be used as a replacement for ordinary binary or
decimal numbers.
The first instance of a hereditary number system, at our best knowledge,
occurs in the proof of Goodstein’s theorem [8], where replacement of finite numbers on a tree’s branches by the ordinal ω allows him to prove that a “hailstone
sequence” visiting arbitrarily large numbers eventually turns around and terminates.
Numeration systems on regular languages have been studied recently, e.g. in
[9] and specific instances of them are also known as bijective base-k numbers [4].
Arithmetic packages similar to our bijective base-2 view of arithmetic operations
are part of libraries of proof assistants like Coq [10] and the corresponding regular
languages have been used as a basis of decidable arithmetic systems like (W)S1S
[11] and (W)S2S [12].
Arithmetic computations based on recursive data types like the free magma
of binary trees (isomorphic to the context-free language of balanced parentheses)
are described in [13] and [14], where they are seen as Gödel’s System T types,
as well as combinator application trees. In [1] a type class mechanism is used to
express computations on hereditarily finite sets and hereditarily finite functions.
An emulation of Peano and conventional binary arithmetic operations in
Prolog, is described in [15]. Their approach is similar as far as a symbolic representation is used. The key difference with our work is that our operations
work on tree structures, and as such, they are not based on previously known
algorithms.
Arithmetic computations with types expressed as C++ templates are described
in [16] and in online articles by Oleg Kiselyov using Haskell’s type inference
mechanism. However, the algorithm described there is basically the same as
[15], focusing on Peano and binary arithmetics.
Efficient representation of sparse sets are usually based on a dedicated data
structure [17] and they cannot be at the same time used for arithmetic computations as it is the case with our tree-based encoding.
In [18] integer decision diagrams are introduced providing a compressed representation for sparse integers, sets and various other data types. However likewise [13] and [1], and by contrast to those proposed in this paper, they do not
compress dense sets or numbers.
Ranking functions (bijections between data types and natural numbers) can
be traced back to Gödel numberings [19] associated to formulae. Together with
their inverse unranking functions they are also used in combinatorial generation
algorithms [20,21].
As a fresh look at the topic, we mention recent work in the context of functional programming on connecting heterogeneous data types through bijective
mappings and natural number encodings [22,23,24].
11
Conclusion and future work
We have seen that the average performance of arithmetic computations with
trees of type T is comparable, up to small constant factors, to computations
performed with the binary data type B and it outperforms them by an arbitrarily
large margin on the interesting special cases favoring the tree representations.
Still, does that mean that such binary trees can be used as a basis for a
practical arbitrary integers package?
Native arbitrary length integer libraries like GMP or BigInteger take advantage of fast arithmetic on 64 bit words.
To match their performance, we plan to switch between bitstring representations for numbers fitting in a machine word and a a tree representation for
numbers not fitting in a machine word.
We have shown that some interesting number-theoretical entities like Mersenne,
Fermat and perfect numbers have significantly more compact representations
with our tree-based numbers. One may observe their common feature: they are
all represented in terms of exponents of 2, successor/predecessor and specialized
multiplication operations.
The fundamental theoretical challenge raised at this point is the following:
can other number-theoretically interesting operations, with possible applications
to cryptography be also expressed succinctly in terms of our tree-based data type?
Is it possible to reduce the complexity of some other important operations, besides
those found so far?
The methodology to be used relies on two key components, that have been
proven successful so far, in discovering succinct representations for Mersenne,
Fermat and perfect numbers, as well as low complexity algorithms for operations
like bitsize and exp2:
– partial evaluation of functional programs with respect to known data types
and operations on them, as well as the use of other program transformations
– salient number-theoretical observations, provable by induction, that relate
operations on our tree data types to known identities and number-theoretical
algorithms
References
1. Tarau, P.: Declarative modeling of finite mathematics. In: PPDP ’10: Proceedings
of the 12th international ACM SIGPLAN symposium on Principles and practice
of declarative programming, New York, NY, USA, ACM (2010) 131–142
2. Wadler, P., Blott, S.: How to make ad-hoc polymorphism less ad-hoc. In: POPL.
(1989) 60–76
3. Kaye, R., Wong, T.L.: On Interpretations of Arithmetic and Set Theory. Notre
Dame J. Formal Logic Volume 48(4) (2007) 497–510
4. Wikipedia: Bijective numeration — wikipedia, the free encyclopedia (2012) [Online; accessed 2-June-2012].
5. Wikipedia: Great internet mersenne prime search — wikipedia, the free encyclopedia (2012) [Online; accessed 9-December-2012].
6. Tarau, P.: “Everything Is Everything” Revisited: Shapeshifting Data Types with
Isomorphisms and Hylomorphisms. Complex Systems (18) (2010) 475–493
7. Knuth, D.E.: Mathematics and Computer Science: Coping with Finiteness. Science
194(4271) (1976) 1235 –1242
8. Goodstein, R.: On the restricted ordinal theorem. Journal of Symbolic Logic (9)
(1944) 33–41
9. Rigo, M.: Numeration systems on a regular language: arithmetic operations, recognizability and formal power series. Theoretical Computer Science 269(12) (2001)
469 – 498
10. The Coq development team: The Coq proof assistant reference manual. LogiCal
Project. (2004) Version 8.0.
11. Büchi, J.R.: On a decision method in restricted second order arithmetic. International Congress on Logic, Method, and Philosophy of Science 141 (1962) 1–12
12. Rabin, M.O.: Decidability of second-order theories and automata on infinite trees.
Transactions of the American Mathematical Society 141 (1969) 1–35
13. Tarau, P., Haraburda, D.: On Computing with Types. In: Proceedings of SAC’12,
ACM Symposium on Applied Computing, PL track, Riva del Garda (Trento), Italy
(March 2012) 1889–1896
14. Tarau, P.: Declarative Specification of Tree-based Symbolic Arithmetic Computations. In Russo, C., Zhou, N.F., eds.: Proceedings of PADL’2012, Practical Aspects
of Declarative Languages, Philadelphia, PA, Springer, LNCS7149 (January 2012)
273–289
15. Kiselyov, O., Byrd, W.E., Friedman, D.P., Shan, C.c.: Pure, declarative, and
constructive arithmetic relations (declarative pearl). In: FLOPS. (2008) 64–80
16. Kiselyov, O.: Type arithmetics: Computation based on the theory of types. CoRR
cs.CL/0104010 (2001)
17. Briggs, P., Torczon, L.: An efficient representation for sparse sets. ACM Letters
on Programming Languages and Systems 2 (1993) 59–69
18. Vuillemin, J.: Efficient Data Structure and Algorithms for Sparse Integers, Sets and
Predicates. In: Computer Arithmetic, 2009. ARITH 2009. 19th IEEE Symposium
on. (june 2009) 7 –14
19. Gödel, K.: Über formal unentscheidbare Sätze der Principia Mathematica und
verwandter Systeme I. Monatshefte für Mathematik und Physik 38 (1931) 173–
198
20. Martinez, C., Molinero, X.: Generic algorithms for the generation of combinatorial
objects. In Rovan, B., Vojtas, P., eds.: MFCS. Volume 2747 of Lecture Notes in
Computer Science., Berlin Heidelberg, Springer (2003) 572–581
21. Knuth, D.:
The Art of Computer Programming, Volume 4, draft (2006)
http://www-cs-faculty.stanford.edu/∼knuth/taocp.html.
22. Vytiniotis, D., Kennedy, A.: Functional Pearl: Every Bit Counts. ICFP 2010 :
The 15th ACM SIGPLAN International Conference on Functional Programming
(September 2010) ACM Press.
23. Kobayashi, N., Matsuda, K., Shinohara, A.: Functional Programs as Compressed
Data. ACM SIGPLAN 2012 Workshop on Partial Evaluation and Program Manipulation (January 2012) ACM Press.
24. Tarau, P.: A Unified Formal Description of Arithmetic and Set Theoretical Data
Types. In Auxtier, S., ed.: Intelligent Computer Mathematics, 17th Symposium,
Calculemus 2010, 9th International Conference AISC/Calculemus/MKM 2010,
Paris, Springer, LNAI 6167 (July 2010) 247–261
Appendix
This appendix contains some additional code, used for testing and benchmarking
our functions, grouped in the type class Benchmarks. First we define a prime
number generator working with all our instances.
class SpecialComputations n ⇒ Benchmarks n where
primes :: [n]
primes = s (s e) : filter is_prime (odds_from (s (s (s e)))) where
odds_from x = x : odds_from (s (s x))
is_prime p = [p]==to_primes p
to_primes n = to_factors n p ps where
(p:ps) = primes
to_factors
to_factors
(q,r) =
to_factors
n p ps | cmp (mul p p) n == GT = [n]
n p ps | e_ r = p : to_factors q p ps where
div_and_rem n p
n p (hd:tl) = to_factors n hd tl
Next we define the Lucas-Lehmer fast primality test for Mersenne numbers:
lucas_lehmer :: n → Bool
lucas_lehmer p = e_ y where
p_2 = s’ (s’ p)
four = i (o e)
m = exp2 p
m’ = s’ m
y = f p_2 four
f k n | e_ k = n
f k n = r where
x = f (s’ k) n
y = s’ (s’ (mul x x))
--r = reminder y m’
r = fastmod y m
-- fast computation of k mod 2^p-1
fastmod k m | k == s’ m = e
fastmod k m | LT == cmp k m = k
fastmod k m = fastmod (add q r) m where
(q,r) = div_and_rem k m
-- exponents leading to Mersenne primes
mersenne_prime_exps :: [n]
mersenne_prime_exps = filter lucas_lehmer primes
-- actual Mersenne primes
mersenne_primes :: [n]
mersenne_primes = map f mersenne_prime_exps where
f p = s’ (exp2 p)
The Ackerman function is a good benchmark for successor and predecessor operations:
ack
ack
ack
ack
:: n→n→n
x n | e_ x = s n
m1 x | e_ x = ack (s’ m1) (s e)
m1 n1 = ack (s’ m1) (ack m1 (s’ n1))
Next we define a variant of the 3x+1 problem / Collatz conjecture / Syracuse
function (see http://en.wikipedia.org/wiki/Collatz_conjecture) that, some-
what surprisingly, can be expressed as a mix of arithmetic operations and reflected list operations, to test the relative performance of some of our instances.
It is easy to show that the Collatz conjecture is true iff the function nsyr, implementing the n-th iterate of the Syracuse function, always terminates:
syracuse :: n→n
-- n→c’’ (3n+2)
syracuse n = c’’ (add n (i n))
nsyr :: n→[n]
nsyr n | e_ n = [e]
nsyr n = n : nsyr (syracuse n)
Finally we close our type class with the usual instance declarations:
instance Benchmarks Integer
instance Benchmarks B
instance Benchmarks T
The following example illustrates the first 8 sequences of the Syracuse function:
*Giant> map nsyr [0..7]
[[0],[1,2,0],[2,0],[3,5,8,6,2,0],[4,3,5,8,6,2,0],
[5,8,6,2,0],[6,2,0],[7,11,17,26,2,0]]
Our generic benchmark function measures the CPU time for running a no
argument toplevel function f received as a parameter.
benchmark mes f = do
x←getCPUTime
print f
y←getCPUTime
let time=(y-x) ‘div‘ 1000000000
return (mes++" :time="++(show time))
The following benchmarks provide the code used in the section 9.
bm1t = benchmark "ack 3 7 on t" (ack (t (toInteger 3)) (t (toInteger 7)))
bm1b = benchmark "ack 3 7 on b" (ack (b (toInteger 3)) (b (toInteger 7)))
bm1n = benchmark "ack 3 7 on n" (ack (n (toInteger 3)) (n (toInteger 7)))
bm2t = benchmark "exp2 t" (exp2 (exp2 (t (toInteger 14))))
bm2b = benchmark "exp2 b" (exp2 (exp2 (b (toInteger 14))))
bm2n = benchmark "exp2 n" (exp2 (exp2 (n (toInteger 14))))
bm3 tvar = benchmark "sparse_set on a type" (n (bitsize (from_set ps)))
where ps = map tvar [101,2002..100000]
bm4t =benchmark "bitsize of Mersenne 45" (n (bitsize mersenne45))
bm5t = benchmark "bitsize of Perfect 45" (n (bitsize perfect45))
bm6t = benchmark "large leftshift" (leftshift n n) where
n = t prime45
bm3’ tvar m = benchmark "to/from list on a type"
(n (bitsize (from_list (to_list (from_list ps)))))
where ps = map tvar [101,2002..3000+m]
bm3’’ tvar m = benchmark "to/from list on a type"
(n (bitsize (from_list (to_list (from_list ps)))))
where ps = map (dual.tvar) [101,2002..3000+m]
bm7t = benchmark "primes on t"
(last (take 100 ps)) where ps = primes :: [T]
bm7b = benchmark "primes on b"
(last (take 100 ps)) where ps = primes :: [B]
bm7n = benchmark "primes on n"
(last (take 100 ps)) where ps = primes :: [Integer]
bm8t = benchmark "mersenne
(last (take 7 ps)) where
bm8b = benchmark "mersenne
(last (take 7 ps)) where
bm8n = benchmark "mersenne
(last (take 7 ps)) where
on
ps
on
ps
on
ps
t"
= mersenne_primes :: [T]
b"
= mersenne_primes :: [B]
n"
= mersenne_primes :: [Integer]
The following tests the syracuse / Collatz conjecture up to m
test_syr tvar m = maximum (map length (map (nsyr . tvar) [0..m]))
compress_syr tvar m = r where
nss = map (nsyr . tvar) [0..m]
r = maximum (map (n.bitsize) (map from_list nss))
-- overflows for m>2 except for tvar=t
compress_syr_twice tvar m = r where
nss = map (nsyr . tvar) [0..m]
r = (n.bitsize) (from_list (map from_list nss))
bm9 tvar = benchmark "test syracuse" (test_syr tvar 2000)
bm10 tvar = benchmark "compress syracuse" (compress_syr tvar 100)
bm11 tvar = benchmark "compress syracuse_twice" r where
r = compress_syr_twice tvar 20
The following function computes the size of a tree-represented natural number:
tsize T = 1
tsize (V x xs) = 1+ sum (map tsize (x:xs))
tsize (W x xs) = 1+ sum (map tsize (x:xs))
The function kth computes the k-th iteration of a function application.
kth _ k x | e_ k = x
kth f k x = f (kth f (s’ k) x)
The following assertions are used for testing some of our operations:
-- relation between iterations of o,i and power of 2
a1 k = pow (i e) k == s (kth o k e)
a2 k = pow (i e) k == s (s (kth i (s’ k) e))
-a3
m
u
v
relations between power operations and multiplication
n b = (u==v,u,v) where
= pow (i e) n
= kth o n b
= s’ (mul m (s b))
a4 x y = (a==b,a,b) where
a = mul (pow (i e) x) y
b = s (kth o x (s’ y))
| 6cs.PL
|
arXiv:cs/0405097v1 [cs.LO] 26 May 2004
A Coalgebraic Approach to Kleene Algebra with Tests∗
Riccardo Pucella
Cornell University
Ithaca, NY 14853
[email protected]
Hubie Chen
Cornell University
Ithaca, NY 14853
[email protected]
Abstract
Kleene algebra with tests is an extension of Kleene algebra, the algebra of regular expressions, which can be used to reason about programs. We develop a coalgebraic theory of Kleene
algebra with Tests, along the lines of the coalgebraic theory of regular expressions based on deterministic automata. Since the known automata-theoretic presentation of Kleene algebra with
tests does not lend itself to a coalgebraic theory, we define a new interpretation of Kleene algebra with tests expressions and a corresponding automata-theoretic presentation. One outcome
of the theory is a coinductive proof principle, that can be used to establish equivalence of our
Kleene algebra with tests expressions.
1 Introduction
Kleene algebra (KA) is the algebra of regular expressions [Conway 1971; Kleene 1956]. As is well
known, the theory of regular expressions enjoys a strong connection with the theory of finite-state
automata. This connection was used by Rutten [1998] to give a coalgebraic treatment of regular
expressions. One of the fruits of this coalgebraic treatment is coinduction, a proof technique for
demonstrating the equivalence of regular expressions [Rutten 2000]. Other methods for proving
the equality of regular expressions have previously been established—for instance, reasoning by
using a sound and complete axiomatization [Kozen 1994; Salomaa 1966], or by minimization of
automata representing the expressions [Hopcroft and Ullman 1979]. However, the coinduction
proof technique can give relatively short proofs, and is fairly simple to apply.
Recently, Kozen [1997] introduced Kleene algebra with tests (KAT), an extension of KA designed for the particular purpose of reasoning about programs and their properties. The regular
expressions of KAT allow one to intersperse boolean tests along with program actions, permitting
the convenient modelling of programming constructs such as conditionals and while loops. The utility of KAT is evidenced by the fact that it subsumes propositional Hoare logic, providing a complete
deductive system for Hoare-style inference rules for partial correctness assertions [Kozen 1999].
The goal of this paper is to develop a coalgebraic theory of KAT, paralleling the coalgebraic
treatment of KA. Our coalgebraic theory yields a coinductive proof principle for demonstrating the
equality of KAT expressions, in analogy to the coinductive proof principle for regular expressions.
∗
This paper is essentially the same as one that will appear in Theoretical Computer Science. A preliminary version appeared in the Proceedings of the Sixth International Workshop on Coalgebraic Methods in Computer Science, Electronic
Notes in Theoretical Computer Science, Volume 82.1, 2003.
1
The development of our coalgebraic theory proceeds as follows. We first introduce a form of deterministic automaton and define the language accepted by such an automaton. Next, we develop
the theory of such automata, showing that coinduction can be applied to the class of languages representable by our automata. We then give a class of expressions, which play the same role as the
regular expressions in classical automata theory, and fairly simple rules for computing derivatives
of these expressions.
The difficulty of our endeavor is that the known automata-theoretic presentation of KAT [Kozen
2003] does not lend itself to a coalgebraic theory. Moreover, the notion of derivative, essential to
the coinduction proof principle in this context, is not readily definable for KAT expressions as they
are defined by Kozen [1997]. Roughly, these difficulties arise from tests being commutative and
idempotent, and suggest that tests need to be handled in a special way. In order for the coalgebraic
theory to interact smoothly with tests, we introduce a type system along with new notions of strings,
languages, automata, and expressions, which we call mixed strings, mixed languages, mixed automata, and mixed expressions, respectively. (We note that none of these new notions coincide with
those already developed in the theory of KAT.) All well-formed instances of these notions can be
assigned types by our type system. Our type system is inspired by the type system devised by Kozen
[1998, 2002] for KA and KAT, but is designed to address different issues.
This paper is structured as follows. In the next section, we introduce mixed strings and mixed
languages, which will be used to interpret our mixed expressions. In Section 3, we define a notion of
mixed automaton that is used to accept mixed languages. We then impose a coalgebraic structure on
such automata. In Section 4, we introduce a sufficient condition for proving equivalence that is more
convenient than the condition that we derive in Section 3. In Section 5, we introduce our type system
for KAT, and connect typed KAT expressions with the mixed language they accept. In Section 6, we
give an example of how to use the coalgebraic theory, via the coinductive proof principle, to establish
equivalence of typed KAT expressions. In Section 7, we show that our technique is complete, that
is, it can establish the equivalence of any two typed KAT expressions that are in fact equivalent. We
conclude in Section 8 with considerations of future work.
2 Mixed Languages
In this section, we define the notions of mixed strings and mixed languages that we will use throughout the paper. Mixed strings are a variant of the guarded strings introduced by Kaplan [1969] as an
abstract interpretation for program schemes; sets of guarded strings were used by Kozen [2003] as
canonical models for Kleene algebra with tests. Roughly speaking, a guarded string can be understood as a computation where atomic actions are executed amidst the checking of conditions, in the
form of boolean tests. Mixed strings will be used as an interpretation for the mixed expressions we
introduce in Section 5.
Mixed strings are defined over two alphabets: a set of primitive programs (denoted P) and a
set of primitive tests (denoted B). We allow P to be infinite, but require that B be finite. (We will
see in Section 3 where this finiteness assumption comes in. Intuitively, this is because our automata
will process each primitive test individually.) Primitive tests can be put together to form more
complicated tests. A literal l is a primitive test b ∈ B or its negation b; the underlying primitive
test b is said to be the base of the literal, and is denoted by base(l). When A is a subset of B,
lit(A) denotes the set of all literals over A. A test is a nonempty set of literals with distinct bases.
2
Intuitively, a test can be understood as the conjunction of the literals it comprises. The base of a test
t, denoted by base(t), is defined to be the set {base(l) : l ∈ t}, in other words, the primitive tests
the test t is made up from. We extend the notion of base to primitive programs, by defining the base
of a primitive program p ∈ P as ∅.
Example 2.1: Let P = {p, q}, and B = {b, c, d}. The literals lit(B) of B are {b, b, c, c, d, d}. Tests
include {b, c, d} and {b, d}, but {b, b, c} is not a test, as b and b have the same base b. The base of
{b, c, d} is {b, c, d}.
Primitive programs and tests are used to create mixed strings. A mixed string is either the empty
string, denoted by ǫ, or a sequence σ = a1 . . . an (where n ≥ 1) with the following properties:
(1) each ai is either a test or primitive program,
(2) for i = 1, . . . , n − 1, if ai is a test, then ai+1 is a primitive program,
(3) for i = 1, . . . , n − 1, if ai is a primitive program, then ai+1 is a test, and
(4) for i = 2, . . . , n − 1, if ai is a test, then base(ai ) = B.
Hence, a mixed string is an alternating sequence of primitive programs and tests, where each test
in the sequence is a “complete” test, except possibly if it occurs as the first or the last element of
the sequence. This allows us to manipulate mixed strings on a finer level of granularity; we can
remove literals from the beginning of a mixed strings and still obtain a mixed string. The length of
the empty mixed string ǫ is 0, while the length of a mixed string a1 . . . an is n.
Example 2.2: Let P = {p, q}, and B = {b, c, d}. Mixed strings include ǫ (of length 0), {b} and
p (both of length 1), and {b}p{b, c, d}q{d} (of length 5). The sequence {b}p{b, d}q{d} is not a
mixed string, since base({b, d}) 6= B.
We define the concatenation of two mixed strings σ and σ ′ , denoted by σ · σ ′ , as follows. If one
of σ, σ ′ is the empty string, then their concatenation is the other string. If both σ = a1 . . . an and
σ ′ = b1 . . . bm have non-zero length, their concatenation is defined as:
(1) τ = a1 . . . an b1 . . . bm if exactly one of an , b1 is a primitive program and τ is a mixed string;
(2) τ = a1 . . . an−1 (an ∪ b1 )b2 . . . bm if an and b1 are tests such that base(an ) ∩ base(b1 ) = ∅
and τ is a mixed string; and is
(3) undefined otherwise.
Intuitively, concatenation of the two strings is obtained by concatenating the sequence of string
elements, possibly by combining the last test of the first string with the first test of the second string,
provided that the result is a valid mixed string. We note that concatenation of strings is an associative
operation.
Example 2.3: Let P = {p, q}, and B = {b, c, d}. The concatenation of the mixed strings p and
{b, c, d}q is p{b, c, d}q. Similarly, the concatenation of the mixed strings {b}p{b, c} and {d}q{d}
is the mixed string {b}p{b, c, d}q{d}. However, the concatenation of {b}p{b, c} and {b, d}q is not
defined, as {b, c} ∩ {b, d} =
6 ∅. The concatenation of {b}p{b, c} and q is also not defined, as
base({b, c}) 6= B, and thus {b}p{b, c}q is not a mixed string.
3
We assign one or more types to mixed strings in the following way. A type is of the form
A → B, where A and B are subsets of B. Intuitively, a mixed string has type A → B if the first
element of the string has base A, and it can be concatenated with an element with base B. It will be
the case that a mixed string of type A → B can be concatenated with a mixed string of type B → C
to obtain a mixed string of type A → C.
The mixed string ǫ has many types, namely it has type A → A, for all A ∈ ℘(B). A mixed
string of length 1 consisting of a single test t has type base(t) ∪ A → A, for any A ∈ ℘(B) such
that A ∩ base(t) = ∅. A mixed string of length 1 consisting of a single program p has type ∅ → B.
A mixed string a1 . . . an of length n > 1 has type base(a1 ) → B \ base(an ).
Example 2.4: Let P = {p, q}, and B = {b, c, d}. The mixed string p{b, c, d} has type ∅ → ∅.
The mixed string {d}p has type {d} → B. The mixed string {b}p{b, c, d}q{b, c} has type {b} →
{d}. The concatenation of {b}p{b, c, d}q{b, c} and {d}p, namely {b}p{b, c, d}q{b, c, d}p, has type
{b} → B.
A mixed language is a set of mixed strings, and is typeable, with type A → B, if all of the
mixed strings it contains have type A → B. In this paper, we will only be concerned with typeable
mixed languages.
We will be interested in different operations on mixed languages in the following sections. When
L1 ,L2 , and L are mixed languages, we use the notation L1 · L2 to denote the set {σ1 · σ2 : σ1 ∈
L1 , σ2 ∈ L2 }, L0 to denote the set {ǫ}, and for n ≥ 1, Ln to denote the set L · Ln−1 . The following
two operations will be useful in Section 5. The operator T , defined by
T (L) = {σ : σ ∈ L, |σ| = 1, σ is a test}
extracts from a language all the mixed strings made up of a single test. The operator ǫ, defined by
ǫ(L) = L ∩ {ǫ}
essentially checks if the empty mixed string ǫ is in L, since ǫ(L) is nonempty if and only if the
empty mixed string is in L.
3 Mixed Automata
Having introduced a notion of mixed strings, we now define a class of deterministic automata that
can accept mixed strings. Mixed strings enforce a strict alternation between programs and tests,
and this alternation is reflected in our automata. The transitions of the automata are labelled with
primitive programs and literals. Given a mixed string, mixed automaton can process the tests in the
string in many different orders; this reflects the fact that the tests that appear in mixed strings are
sets of literals.
A mixed automaton over the set of primitive programs P and set of primitive tests B is a 3-tuple
M = (hSA iA∈℘(B) , o, hδA iA∈℘(B) ), consisting of a set SA of states for each possible base A 6= ∅
of a test as well as a set S∅ of program states, an output function o : SS
∅ → {0, 1}, and transition
functions δ∅ : S∅ × P → SB and (for A 6= ∅) δA : SA × lit(A) → A∈℘(B) SA , subject to the
following two conditions:
A1. δA (s, l) ∈ SA\{base(l)} , and
4
A2. for every state s in SA , for every test t with base A, and for any two orderings hx1 , . . . , xm i,
y1
ym
x1
xm
hy1 , . . . , ym i of the literals in t, if s −→
. . . −→
s1 and s −→ . . . −→ s2 then s1 = s2 .
l
(For convenience, we write s −→ s′ if δA (s, l) = s′ for A the base of s.)
We give an example of a mixed automaton in Example 3.2. Intuitively, a state in SA can process
a mixed string of type A → B, for some B. Condition A1 enforces the invariant that, as a string is
being processed, the current state is in SA , for A the base of the first element of the string. Condition
A2 is a form of “path independence”: regardless of the order in which we process the literals of a
test, we end up in the same program state. Condition A2, and basing transitions on literals rather
than tests, allow the manipulation of mixed expressions at a finer level of granularity. This is related
to a similar choice we made when allowing mixed strings to start with a test that is not “complete”.
This flexibility will be useful when we analyze mixed expressions in Section 5.
The accepting states are defined via the output function o(s), viewed as a characteristic function.
Accepting states are in S∅.
As in the coalgebraic treatment of automata [Rutten 1998], and contrary to standard definitions,
we allow both the state spaces SA and the set P of primitive programs to be infinite. We also do not
force mixed automata to have initial states, for reasons that will become clear.
We now define the mixed language accepted by a state of a mixed automaton. Call a sequence
µ = e1 . . . em of primitive programs and literals a linearization of a mixed string σ = a1 . . . an if
µ can be obtained from σ by replacing each test ai in σ with a sequence of length |ai | containing
exactly the literals in ai .
Example 3.1: Let P = {p, q}, and B = {b, c}. The mixed string {b}p{b, c}q{b, c} (of type
{b} → ∅) has four linearizations: bpbcqbc, bpcbqbc, bpbcqcb, and bpcbqcb.
Intuitively, a mixed string σ is accepted by an automaton if a linearization of σ is accepted by
the automaton according to the usual definition. Formally, a mixed string σ is accepted by a state s
of an automaton M if either
(1) σ is ǫ and s is a program state with o(s) = 1 (i.e., s is an accepting program state), or
e
e
1
m
(2) there exists a linearization e1 . . . em of σ such that s −→
. . . −→
s′ , s′ is a program state,
′
and o(s ) = 1.
If σ is accepted (by a state s) in virtue of satisfying the second criterion, then every linearization is
a witness to this fact— in other words, the existential quantification in the second criterion could
be replaced with a universal quantification (over all linearizations of σ) without any change in the
actual definition. This is because of condition A2 in the definition of a mixed automaton.
We define the mixed language accepted by state s of automaton M , written LM (s), as the set
of mixed strings accepted by state s of M . It is easy to verify that all the strings accepted by a state
have the same type, namely, if s is in SA , then every string in LM (s) has type A → ∅, and hence
LM (s) has type A → ∅.
Example 3.2: Let P = {p, q}, and B = {b, c}. Consider the mixed automaton over P and B
5
b
s1,{b}
p
s1,ø
s2,{b,c}
c
b
s2,{b}
s2,{c}
b
c
s2,ø
Figure 1: A mixed automaton
pictured in Figure 1, given by M = (hSA iA∈℘(B) , o, hδA iA∈℘(B) ), where:
S{b,c} = {s2,{b,c} , ssink,{b,c} }
S{b} = {s1,{b} , s2,{b} , ssink,{b} }
S{c} = {s2,{c} , ssink,{c} }
S∅ = {s1,∅, s2,∅, ssink,∅}
and
o(s1,∅) = 1
o(s2,∅) = 1
o(ssink,∅) = 0.
The transition function δA can be read off from Figure 1; note that the sink states ssink,A as well as the
transitions to the sink states are not pictured. Intuitively, any transition not pictured in the automaton
can be understood as going to the appropriate sink state. For instance, we have δ{b,c} (s2,{b,c} , c) =
ssink,{b} . We can check that the two conditions A1 and A2 hold in M . The language accepted by
state s1,{b} is LM (s1,{b} ) = {{b}, {b}p{b, c}}. The language accepted by state s1,∅ is LM (s1,∅) =
{ǫ, p{b, c}}.
We define a homomorphism between mixed automata M and M ′ to be a family f = hfA iA∈℘(B)
′ such that:
of functions fA : SA → SA
6
′ (f (s), p),
(1) for all s ∈ S∅, o(s) = o′ (f∅(s)), and for all p ∈ P, fB (δ∅(s, p)) = δ∅
∅
′ (f (s), l).
(2) for all s ∈ SA (where A 6= ∅) and all l ∈ lit(A), fA\{base(l)} (δA (s, l)) = δA
A
A homomorphism preserves accepting states and transitions. We write f : M → M ′ when f is a
homomorphism between automata M and M ′ . For convenience, we often write f (s) for fA (s) when
the type A of s is understood. It is straightforward to verify that mixed automata form a category
(denoted MA), where the morphisms of the category are mixed automata homomorphisms.
We are interested in identifying states that have the same behaviour, that is, that accept the same
mixed language. A bisimulation between two mixed automata M = (hSA iA∈℘(B) , o, hδA iA∈℘(B) )
′ i
′
′
′
and M ′ = (hSA
A∈℘(B) , o , hδA iA∈℘(B) ) is a family of relations hRA iA∈℘(B) where RA ⊆ SA × SA
such that the following two conditions hold:
′ , if sR s′ , then o(s) = o′ (s′ ) and for all p ∈ P, δ (s, p)R δ ′ (s′ , p),
(1) for all s ∈ S∅ and s′ ∈ S∅
∅
∅
B ∅
and
′ (where A 6= ∅), if sR s′ , then for all l ∈ lit(A), δ (s, l)R
′
′
(2) for all s ∈ SA and s′ ∈ SA
A
A
A\{base (l)} δA (s , l).
A bisimulation between M and itself is called a bisimulation on M . Two states s and s′ of M
having the same type B are said to be bisimilar, denoted by s ∼M s′ , if there exists a bisimulation
hRA iA∈℘(B) such that sRB s′ . (We simply write s ∼ s′ when M is clear from the context.) For each
M , the relation ∼M is the union of all bisimulations on M , and in fact is the greatest bisimulation
on M .
Proposition 3.3: If s is a state of M and s′ is a state of M ′ with s ∼ s′ , then LM (s) = LM ′ (s′ ).
Proof: We show, by induction on the length of mixed strings that for all mixed strings σ, and for all
states s, s′ such that s ∼ s′ , then σ ∈ LM (s) if and only if σ ∈ LM ′ (s′ ). For the empty mixed string
ǫ, we have ǫ ∈ LM (s) if and only if o(s) = 1 if and only if o′ (s′ ) = 1 (by definition of bisimilarity)
if and only if ǫ ∈ LM ′ (s′ ). Assume inductively that the results holds for mixed strings of length n.
Let σ be a mixed string of length n + 1, of the form aσ ′ . Assume σ ∈ LM (s). By definition, there
e1
em
is a linearization e1 . . . em of a and a state s1 such that s −→
. . . −→
s1 and σ ′ ∈ LM (s1 ). By the
e1
em
′
′
′
definition of bisimilar states, we have s −→ . . . −→ s1 and s1 ∼ s1 . By the induction hypothesis,
σ ′ ∈ LM ′ (s′1 ). By the choice of s′1 , we have that σ ∈ LM ′ (s′ ), as desired.
Conditions (1) and (2) of the definition of a bisimulation are analogous to the conditions in the
definition of a homomorphism. Indeed, a homomorphism can be viewed as a bisimulation.
Proposition 3.4: If f : M → M ′ is a mixed automataon homomorphism, then hRA iA∈℘(B) , defined
by RA = {(s, fA (s)) : s ∈ SA } is a bisimulation.
Proof: First, for all s ∈ S∅, sR∅s′ implies s′ = f∅(s), and o(s) = o′ (f∅(s)) = o′ (s′ ). Moreover,
′ (s′ , p) = δ ′ (f (s), p) = f (δ (s, l)), so that δ (s, l)R δ ′ (s′ , l), as
for all p ∈ P, we have δ∅
B ∅
∅
B ∅
∅ ∅
required. Similarly, let s ∈ SA (where A 6= ∅); sRA s′ implies s′ = fA (s), and thus for all
′ (s′ , l) = δ ′ (f (s), l) = f
′
′
l ∈ lit(A), δA
A\{base(l)} (δA (s, l)), so that δA (s, l)RA\{base (l)} δA (s , l),
A A
as required, proving that hRA iA∈℘(B) is a bisimulation.
An immediate consequence of this relationship is that homomorphisms preserve accepted languages.
7
Proposition 3.5: If f : M → M ′ is a mixed automaton homomorphism, then LM (s) = LM ′ (f (s))
for all states s of M .
Proof: Immediate from Propositions 3.4 and 3.3.
It turns out that we can impose a mixed automaton structure on the set of all mixed languages
with type A → ∅. We take as states mixed languages of type A → ∅. A state is accepting if the
empty string ǫ is in the language. It remains to define the transitions between states; we adapt the
idea of Brzozowski derivatives [Brzozowski 1964]. Our definition of derivative depends on whether
we are taking the derivative with respect to a program element or a literal.
If the mixed language L has type ∅ → B and p ∈ P is a primitive program, define
Dp (L) = {σ : p · σ ∈ L}.
If the mixed language L has type A → B (for A 6= ∅) and l ∈ lit(A) is a literal, then
Dl (L) = {σ : {l} · σ ∈ L}.
Define LA to be the set of mixed languages of type A → ∅. Define L to be (hLA iA∈℘(B) , oL , hδA iA∈℘(B) ),
where oL (L) = 1 if ǫ ∈ L, and 0 otherwise; δ∅(L, p) = Dp (L); and δA (L, l) = Dl (L), for A 6= ∅
and l ∈ lit(A). It is easy to verify that L is indeed a mixed automaton. The following properties of
L are significant.
Proposition 3.6: For a mixed automaton M with states hSA iA∈℘(B) , the maps fA : SA → L
mapping a state s in SA to the language LM (s) form a mixed automaton homomorphism.
Proof: We check the two conditions for the family hfA iA∈℘(B) to be a homomorphism. First, given
s ∈ S∅, o(s) = 1 if and only if ǫ ∈ LM (s), which is equivalent to oL (f∅(s)) = 1. Moreover, given
p ∈ B, fB (δ∅(s, p)) = LM (δ∅(s, p)) = {σ : p · σ ∈ LM (s)} = Dp (LM (s)) = Dp (f∅(s)),
as required. Similarly, given s ∈ SA (where A 6= ∅), and l ∈ lit(A), fA\{base(l)} (δA (s, l)) =
LM (δA (s, l)) = {σ : {l} · σ ∈ LM (s)} = Dl (LM (s)) = Dl (fA (s)), as required.
Proposition 3.7: For any mixed language L in L, the mixed language accepted by state L in L is L
itself, that is, LL (L) = L.
Proof: We prove by induction on the length of linearizations of σ that for all mixed strings σ,
σ ∈ L if and only if σ ∈ LL (L). For the empty mixed string ǫ, we have ǫ ∈ L ⇔ oL (L) = 1 ⇔ ǫ ∈
LL (L). For σ of the form pσ ′ , we have σ = p·σ ′ , and thus we have p·σ ′ ∈ L ⇔ σ ′ ∈ Dp (L), which
by the induction hypothesis holds if and only if σ ′ ∈ LL (Dp (L)) ⇔ σ ′ ∈ Dp (LL (L)) (because
LL is a mixed automaton homomorphism from L to L), which is just equivalent to p · σ ′ ∈ LL (L).
For σ with a linearization le1 . . . em , letting σ ′ denote a string with linearization e1 . . . em , we have
σ = {l} · σ ′ , and we can derive in an exactly similar manner that {l} · σ ′ ∈ L ⇔ σ ′ ∈ Dl (L) ⇔
σ ′ ∈ LL (Dl (L)) ⇔ σ ′ ∈ Dl (LL (L)) ⇔ {l} · σ ′ ∈ LL (L) ⇔ σ ∈ LL (L).
These facts combine into the following fundamental property of L, namely, that L is a final
automaton.
8
Theorem 3.8: L is final in the category MA, that is, for every mixed automaton M , there is a
unique homomorphism from M to L.
Proof: Let M be a mixed automaton. By Proposition 3.6, there exists a homomorphism f from M
to the final automaton L, mapping a state s to the language LM (s) accepted by that state. Let f ′ be
another homomorphism from M to L. To establish uniqueness, we need to show that for any state
s of M , we have f (s) = f ′ (s):
f (s) = LM (s)
(by definition of f )
′
= LL (f (s))(by Proposition 3.5)
= f ′ (s)
(by Proposition 3.7).
Hence, f is the required unique homomorphism.
The finality of L gives rise to the following coinduction proof principle for language equality,
in a way which is by now standard [Rutten 2000].
Corollary 3.9: For two mixed languages K and L of type A → ∅, if K ∼ L then K = L.
In other words, to establish the equality of two mixed languages, it is sufficient to exhibit a
bisimulation between the two languages when viewed as states of the final automaton L. In the
following sections, we will use this principle to analyze equality of languages described by a typed
form of KAT expressions.
4 Pseudo-Bisimulations
The “path independence” condition (A2) in the definition of a mixed automaton gives mixed automata a certain form of redundancy. It turns out that due to this redundancy, we can define a
simpler notion than bisimulation that still lets us establish the bisimilarity of states.
A pseudo-bisimulation (relative to the ordering b1 , . . . , b|B| of the primitive tests in B) between
′ i
′
′
two mixed automata M = (hSA iA∈℘(B) , o, hδA iA∈℘(B) ) and M ′ = (hSA
A∈℘(B) , o , hδA iA∈℘(B) )
′
is a family of relations hRi ii=0,... ,|B| where Ri ⊆ SAi × SAi (with Ai denoting {bj : j ≤ i, j ∈
{1, . . . , |B|}}) such that the following two conditions hold:
′ , if sR s′ , then o(s) = o′ (s′ ) and for all p ∈ P, δ (s, p)R δ ′ (s′ , p),
(1) for all s ∈ S∅ and s′ ∈ S∅
0
∅
|B| ∅
and
′ , if sR s′ , then for all l ∈ lit(b ),
(2) for all i = 1, . . . , |B|, for all s ∈ SAi and s′ ∈ SA
i
i
i
′ (s′ , l).
δAi (s, l)Ri−1 δA
i
The sense in which pseudo-bisimulation is weaker than a bisimulation is that there need not
be a relation for each element of ℘(B). As the following theorem shows, however, we can always
complete a pseudo-bisimulation to a bisimulation.
Theorem 4.1: If hRi ii=0,... ,|B| is a pseudo-bisimulation (relative to the ordering b1 , . . . , b|B| of the
′ i such that R′ = R for all i = 0, . . . , |B|
primitive tests in B), then there exists a bisimulation hRA
i
Ai
(with Ai denoting {bj : j ≤ i, j ∈ {1, . . . , |B|}}).
9
Proof: Let hRi ii=0,... ,|B| be a pseudo-bisimulation (relative to the ordering on primitive tests
′ ⊆ S × S ′ for each A ∈ ℘(B), and show
b1 , . . . , b|B| ). We define a family of relations RA
A
A
that it forms a bisimulation with the required property. The proof relies on the path independence
condition A2 of mixed automata in a fundamental way. Given A ∈ ℘(B), let i(A) be the largest i ∈
{1, . . . , |B|} such that {b1 , . . . , bi } ⊆ A, and let c(A) be the relative complement of {b1 , . . . , bi(A) }
defined by A \ {b1 , . . . , bi(A) }. We say that a sequence of literals l1 , . . . , lk is exhaustive over a set
′ as follows: sR′ s′ holds if and
of bases A if A = {base(l1 ), . . . , base(lk )} and |A| = k. Define RA
A
l
l
l
1
1
k
only if for all literal sequences l1 , . . . , lk exhaustive over c(A), we have s −→
. . . −→
s1 , s′ −→
l
k
′ =R
. . . −→
s′1 , and s1 Ri(A) s′1 . Clearly, if A = {b1 , . . . , bi(A) }, then RA
i(A) , as required. We now
′ s′ , then sR s′ , and hence
′
′
check that hRA iA∈℘(B) is a bisimulation. Clearly, since R∅ = R0 , if sR∅
0
o(s) = o′ (s′ ), and for all p ∈ P, it holds that δ∅(s, p)R|B| δ∅(s′ , p), implying δ∅(s, p)RB′ δ∅(s′ , p).
′ , l ∈ lit(A), and assume sR′ s′ . Consider the following cases:
Now, let A 6= ∅, s ∈ SA , s′ ∈ SA
A
′ s′ , then sR
′
Case A = {b1 , . . . , bi(A) }, base(l) = bi(A) : Since sRA
i(A) s , and by the properties
′
of pseudo-bisimulations, we have δA (s, l)Ri(A)−1 δA (s′ , l), which is exactly δA (s, l)RA\{base(l)}
δA (s′ , l).
′
′
′
Case A = {b1 , . . . , bi(A) }, base(l) = bj , j < i(A): Since sRA s , then sRi(A) s . let l1 , . . . , lk
′
′
be an arbitrary exhaustive sequence of literals over {bi(A) , . . . , bj+1 }. Let li(A)
, . . . , lj+1
be the
′
′
′
arrangement of l1 , . . . , lk such that base(lm ) = bm . Consider the states s1 , s2 , s1 , s2 such that
′
li(A)
′
lj+1
′
li(A)
l
′
lj+1
l
s −→ . . . −→ s1 −→ s2 , and s′ −→ . . . −→ s′1 −→ s′2 . By the definition of pseudo-bisimulation,
′
li(A)
l
we have that s2 Rj−1 s′2 . Now, by condition A2, we have states s3 , s′3 such that s −→ s3 −→
′
lj+1
′
li(A)
l
′
lj+1
l
l
1
k
. . . −→ s2 and s′ −→ s′3 −→ . . . −→ s′2 . By condition A2 again, we have that s3 −→
. . . −→
s2
l
l
k
1
and s′3 −→
s′2 . Since l1 , . . . , lk was arbitrary, s2 Rj−1 s′2 and i(A \ {base(l)}) = j − 1, we
. . . −→
′
′
′
have s3 RA\{base(l)} s′3 , that is, δA (s, l)RA\{base
(l)} δA (s , l).
Case A ⊃ {b1 , . . . , bi(A) }, base(l) ∈ c(A)): Pick an arbitrary sequence l1 , . . . , lk of literals
l
l
l
k
1
that is exhaustive over c(A \ {base(l)}), and states s1 , s2 , s′1 , s′2 such that s −→ s2 −→
s1 ,
. . . −→
l
l
l
1
k
′ , we have s R
′
. . . −→
s′1 . By definition of RA
and s′ −→ s′2 −→
1 i(A) s1 . Since the sequence of
′
literals l1 , . . . , lk was arbitrary, and since i(A) = i(A \ {base(l)}), we have that s2 RA\{base(l)}
s′2 ,
′
that is, δA (s, l)RA\{base(l)}
δA (s′ , l).
Case A ⊃ {b1 , . . . , bi(A) }, base(l) = bi(A) : Pick an arbitrary sequence l1 , . . . , lk of literals
l
l
l
l
1
1
k
k
. . . −→
s1 and s′ −→
. . . −→
s′1 .
that is exhaustive over c(A), and states s1 , s′1 such that s −→
l
′ , we have s R
′
By definition of RA
1 i(A) s1 . By definition of pseudo-bisimulation, if s1 −→ s2 and
l
l
s′1 −→ s′2 , then we have s2 Ri(A)−1 s′2 . By condition A2, we have that for states s3 , s′3 , s −→
l
l
l
l
l
1
k
1
k
. . . −→
s′2 . Thus, since l1 , . . . , lk was arbitrary, and
s3 −→
. . . −→
s2 and s′ −→ s′3 −→
′
′
δA (s′ , l).
i(A \ {base(l)}) = i(A) − 1, we have s3 RA\{base(l)}
s′3 , that is, δA (s, l)RA\{base(l)}
Case A ⊃ {b1 , . . . , bi(A) }, base(l) = bj , j < i(A): Pick an arbitrary sequence l1 , . . . , lk of
literals that is exhaustive over c(A)∪{bi(A) , . . . , bj+1 }. Let l1′ , . . . , lk′ ′ be the elements of l1 , . . . , lk
with bases in c(A). Let l1′′ , . . . , lk′′′′ be the elements of l1 , . . . , lk with bases in {bi(A) , . . . , bj+1 }. Let
′′′ , . . . , l′′′ be the arrangement of l′′ , . . . , l′′ such that base(l′′′ ) = b . Consider states s , s′
li(A)
m
1 1
m
1
j+1
k ′′
l′
l′ ′
l′
l′ ′
1
k
1
k
′ , we have s R
′
. . . −→
. . . −→
such that s −→
s1 and s′ −→
s′1 . By definition of RA
1 i(A) s1 . Now,
10
′′′
li(A)
′′′
lj+1
l
′′′
li(A)
′′′
lj+1
l
l
l
l
l
l
consider states s2 , s3 , s′2 , s′3 such that s1 −→ . . . −→ s2 −→ s3 and s′1 −→ . . . −→ s′2 −→ s′3 . By
the definition of pseudo-bisimulation, since s1 Ri(A) s′1 , we have that s3 Rj−1 s′3 . Now, by condition
l
1
k
1
k
. . . −→
s′3 .
. . . −→
s3 and s′ −→ s′4 −→
A2, we have states s4 , s′4 such that s −→ s4 −→
′
Since l1 , . . . , lk was arbitrary, and i(A \ {base(l)}) = j − 1, we have s4 RA\{base(l)}
s′4 , that is,
′
′
δA (s, l)RA\{base (l)} δA (s , l).
Let us say that two states s, s′ are pseudo-bisimilar if they are related by some Ri in a pseudobisimulation hRi i; it follows directly from Theorem 4.1 that pseudo-bisimilar states are bisimilar.
5 Mixed Expressions and Derivatives
A mixed expression (over the set of primitive programs P and the set of primitive tests B) is any
expression built via the following grammar:
e ::= 0 | 1 | p | l | e1 + e2 | e1 · e2 | e∗
(with p ∈ P and l ∈ lit(B)). For simplicity, we often write e1 e2 for e1 · e2 . We also freely
use parentheses when appropriate. Intuitively, the constants 0 and 1 stand for failure and success,
respectively. The expression p represents a primitive program, while l represents a primitive test.
The operation + is used for choice, · for sequencing, and ∗ for iteration. These are a subclass
of the KAT expressions as defined by Kozen [1997]. (In addition to allowing negated primitive
tests, Kozen also allows negated tests.) We call them mixed expressions to emphasize the different
interpretation we have in mind.
In a way similar to regular expressions denoting regular languages, we define a mapping M
from mixed expressions to mixed languages inductively as follows:
M (0) = ∅
M (1) = {ǫ}
M (p) = {p}
M (l) = {{l}}
M (e1 + e2 ) = M (e1 ) ∪ M (e2 )
M (e1 · e2 ) = M (e1 ) · M (e2 )
[
M (e∗ ) =
M (e)n .
n≥0
The mapping M is a rather canonical homomorphism from mixed expressions to mixed languages. (It is worth noting that we have not defined any axioms for deriving the “equivalence” of
mixed expressions, and it is quite possible for distinct mixed expressions to give rise to the same
mixed language.)
Inspired by a type system devised by Kozen [1998, 2002] for KA and KAT expressions, we
impose a type system on mixed expressions. The types have the form A → B, where A, B ∈ ℘(B),
the same types we assigned to mixed strings in Section 2. We shall soon see that this is no accident.
We assign a type to a mixed expression via a type judgment written ⊢ e : A → B. The following
11
inference rules are used to derive the type of a mixed expression:
⊢0:A→B
⊢1:A→A
⊢p:∅→B
⊢ l : A ∪ {base(l)} → A \ {base(l)}
⊢ e1 : A → B ⊢ e2 : A → B
⊢ e1 + e2 : A → B
⊢ e1 : A → B ⊢ e2 : B → C
⊢ e1 · e2 : A → C
e:A→A
.
e∗ : A → A
It is clear from these rules that any subexpression of a mixed expression having a type judgment
also has a type judgment.
The typeable mixed expressions (which intuitively are the “well-formed” expressions) induce
typeable mixed languages via the mapping M , as formalized by the following proposition.
Proposition 5.1: If ⊢ e : A → B, then M (e) is a mixed language of type A → B.
Proof: A straightforward induction on the structure of mixed expressions.
Our goal is to manipulate mixed languages by manipulating the mixed expressions that represent
them via the mapping M . (Of course, not every mixed language is in the image of M .) In particular,
we are interested in the operations T (L) and ǫ(L), as defined in Section 2, as well as the language
derivatives Dp and Dl introduced in the last section.
We now define operators on mixed expressions that capture those operators on the languages
denoted by those mixed expressions. We define T̂ inductively on the structure of mixed expressions,
as follows:
T̂ (0) = 0
T̂ (1) = 1
T̂ (p) = 0
T̂ (l) = l
T̂ (e1 + e2 ) = T̂ (e1 ) + T̂ (e2 )
T̂ (e1 · e2 ) = T̂ (e1 ) · T̂ (e2 )
T̂ (e∗ ) = T̂ (e)∗
(where p ∈ P and l ∈ lit(B)). The operator T̂ “models” the operator T (L), as is made precise in
the following way.
Proposition 5.2: If ⊢ e : A → B, then T̂ (e) is a typeable mixed expression such that T (M (e)) =
M (T̂ (e)).
12
Proof: A straightforward induction on the structure of mixed expressions.
We define ǫ̂ inductively on the structure of mixed expressions, as follows:
ǫ̂(0) = 0
ǫ̂(1) = 1
ǫ̂(p) = 0
ǫ̂(l) = 0
0
ǫ̂(e1 + e2 ) =
1
1
ǫ̂(e1 · e2 ) =
0
if ǫ̂(e1 ) = ǫ̂(e2 ) = 0
otherwise
if ǫ̂(e1 ) = ǫ̂(e2 ) = 1
otherwise
ǫ̂(e∗ ) = 1
(where p ∈ P and l ∈ lit(B)). Note that ǫ̂(e) is always the mixed expression 0 or 1. In analogy to
Proposition 5.2, we have the following fact connecting the ǫ and ǫ̂ operators.
Proposition 5.3: If ⊢ e : A → B, then ǫ̂(e) is a typeable mixed expression such that ǫ(M (e)) =
M (ǫ̂(e)).
Proof: A straightforward induction on the structure of mixed expressions.
Finally, we define, by induction on the structure of mixed expressions, the derivative operator
D̂ for typeable mixed expressions. There are two forms of the derivative, corresponding to the two
forms of derivative for mixed languages: the derivative D̂l with respect to a literal l ∈ lit(B), and
the derivative D̂p with respect to a primitive program p ∈ P. The two forms of derivative are
defined similarly, except on the product of two expressions. (Strictly speaking, since the definition
of the derivative depends on the type of the expressions being differentiated, D̂ should take type
derivations as arguments rather than simply expressions. To lighten the notation, we write D̂ as
though it took mixed expressions as arguments, with the understanding that the appropriate types
are available.)
The derivative D̂p with respect to a primitive program p ∈ P is defined as follows:
D̂p (0) = 0
D̂p (1) = 0
1 if p = q
D̂p (q) =
0 otherwise
D̂p (l) = 0
D̂p (e1 + e2 ) = D̂p (e1 ) + D̂p (e2 )
D̂p (e1 ) · e2
if B 6= ∅
D̂p (e1 · e2 ) =
D̂p (e1 ) · e2 + ǫ̂(e1 ) · D̂p (e2 ) otherwise
where ⊢ e1 : A → B and ⊢ e2 : B → C
D̂p (e∗ ) = D̂p (e) · e∗ .
13
The derivative D̂l with respect to a literal l ∈ lit(B) is defined as follows:
D̂l (0) = 0
D̂l (1) = 0
D̂l (p) = 0
1 if l = l′
′
D̂l (l ) =
0 otherwise
D̂l (e1 + e2 ) = D̂l (e1 ) + D̂l (e2 )
D̂l (e1 ) · e2
if base(l) ∈
/B
D̂l (e1 · e2 ) =
D̂l (e1 ) · e2 + T̂ (e1 ) · D̂l (e2 ) otherwise
where ⊢ e1 : A → B and ⊢ e2 : B → C
∗
D̂l (e ) = D̂l (e) · e∗ .
We have the following proposition, similar to the previous two, connecting the derivative D̂ to
the previously defined derivative D on mixed languages.
Proposition 5.4: Suppose that ⊢ e : A → B.
If A = ∅, then for all p ∈ P, Dp (M (e)) = M (D̂p (e)).
If A 6= ∅, then for all l ∈ lit(A), Dl (M (e)) = M (D̂l (e)).
Proof: The proof is by induction on the structure of the mixed expression e. To illustrate the proof
technique, we give one case of the proof.
Suppose that ⊢ e1 : A → B and ⊢ e2 : B → C, and e = e1 · e2 . Suppose further that
l ∈ lit(B) is a literal such that base(l) ∈ A and base(l) ∈ B. We will show that the proposition
holds for the expression e, assuming (by the induction hypothesis) that the proposition holds for all
subexpressions of e.
We first establish three claims that will be needed.
Claim 1: If t is a test which (as a mixed string) can be judged to have type A → B, then {t} · {σ :
{l} · σ ∈ M (e2 )} = {σ ′ : {l} · σ ′ ∈ {t} · M (e2 )}.
First suppose that σ is a mixed string such that {l} · σ ∈ M (e2 ). Then σ can be judged to
have type B \ {base(l)} → C, and so {l} · {t} · σ = {t} · {l} · σ ∈ {t} · M (e2 ). It follows that
t · σ ∈ {σ ′ : {l} · σ ′ ∈ {t} · M (e2 )}. For the other direction, suppose that σ ′ is a mixed string such
that {l} · σ ′ ∈ {t} · M (e2 ). Then there exists a mixed string τ ∈ M (e2 ) such that {l} · σ ′ = {t} · τ .
Since t can be judged to have type A → B and base(l) ∈ A ∩ B, base(l) ∈
/ t and there exists a
mixed string σ such that {l} · σ ′ = {t} · τ = {l} · {t} · σ. Thus σ ′ = {t} · σ where {l} · σ ∈ M (e2 ).
Claim 2: If σ is a mixed string such that l · σ ∈ M (e1 ), then l · σ ∈ M (e1 ) \ T (M (e1 )).
This claim holds because {l} · σ ∈ M (e1 ) implies that σ has type A \ {base(l)} → B; since
B 6⊆ A′ , by the definition of the type of a mixed string, |σ| > 1 and so |{l} · σ| > 1.
Claim 3: {σ : {l}·σ ∈ M (e1 )\T (M (e1 ))}·M (e2 ) = {σ : {l}·σ ∈ (M (e1 )\T (M (e1 )))·M (e2 )}
The ⊆ direction is straightforward. For the ⊇ direction, let σ be a mixed string in the second
set; then, there exist strings τ1 ∈ M (e1 ) \ T (M (e1 )) and τ2 ∈ M (e2 ) such that {l} · σ = τ1 · τ2 .
All strings in M (e1 ) have type A → B; since base(l) ∈ B, there are no strings in M (e1 ) of length
one consisting of a primitive program, and so |τ1 | > 3. Hence σ = σ ′ · τ2 for some mixed string σ ′
such that {l} · σ ′ ∈ M (e1 ) \ T (M (e1 )).
14
Using these three claims, we show that Dl (M (e)) = M (D̂l (e)):
M (D̂l (e1 · e2 ))
= M (D̂l (e1 ) · e2 + T̂ (e1 ) · D̂l (e2 ))
(by definition of D̂l )
= M (D̂l (e1 )) · M (e2 ) ∪ M (T̂ (e1 )) · M (D̂l (e2 )) (by definition of M )
= Dl (M (e1 )) · M (e2 ) ∪ M (T̂ (e1 )) · D(M (e2 ))
(by induction hypothesis)
= Dl (M (e1 )) · M (e2 ) ∪ T (M (e1 )) · D(M (e2 ))
(by Proposition 5.2)
= {σ : {l} · σ ∈ M (e1 )} · M (e2 ) ∪
T (M (e1 )) · {σ : {l} · σ ∈ M (e2 ))}
(by definition of Dl )
= {σ : {l} · σ ∈ M (e1 )} · M (e2 ) ∪
{σ : {l} · σ ∈ T (M (e1 )) · M (e2 ))}
(by Claim 1)
= {σ : {l} · σ ∈ M (e1 ) \ T (M (e1 ))} · M (e2 ) ∪
{σ : {l} · σ ∈ T (M (e1 )) · M (e2 ))}
(by Claim 2)
= {σ : {l} · σ ∈ (M (e1 ) \ T (M (e1 ))) · M (e2 )} ∪
{σ : {l} · σ ∈ T (M (e1 )) · M (e2 ))}
(by Claim 3)
= {σ : {l} · σ ∈ M (e1 ) · M (e2 )}
= Dl (M (e1 ) · M (e2 ))
(by definition of Dl )
= Dl (M (e1 · e2 ))
(by definition of M ).
The other cases are similar.
6 Example
In this section, we use the notions of pseudo-bisimulation and the coinduction proof principle
(Corollary 3.9), along with the derivative operator D̂, to prove the equivalence of two mixed languages specified as mixed expressions.
Fix P to be the set of primitive programs {p, q}, and B to be the set of primitive tests {b, c}. Let
[b] be a shorthand for (b + b). Define α to be the mixed expression
(bp([b]cq)∗ c)∗ b
and β to be the mixed expression
bp([b]cq + bcp)∗ cb + b.
Our goal is to prove that α and β are equivalent, in the sense that they induce the same language
via the mapping M . In other words, we want to establish that M (α) = M (β). This example
demonstrates the equivalence of the program
while b do {
p;
while c do q
}
15
and the program
if b then {
p;
while b + c do
if c then q else p
}
This equivalence is a component of the proof of the classical result that every while program can be
simulated by a while program with at most one while loop, as presented by Kozen [1997]. We refer
the reader there for more details.
There are a few ways to establish this equivalence. One is to rely on a sound and complete axiomatization of the equational theory of KAT, and derive the equivalence of α and β algebraically
[Kozen and Smith 1996]. Another approach is to first construct for each expression an automaton
that accepts the language it denotes, and then minimize both automata [Kozen 2003]. Two expressions are then equal if the two resulting automata are isomorphic.
In this paper, we describe a third approach, using the coinductive proof principle for mixed
languages embodied by Corollary 3.9. Since the theory we developed in Section 3 applies only to
mixed languages of type A → ∅, we verify that indeed we have ⊢ α : {b} → ∅ and ⊢ β : {b} → ∅,
so that, by Proposition 5.1, M (α) and M (β) are languages of type {b} → ∅.
We prove the equivalence of α and β by showing that the mixed languages M (α) and M (β)
are pseudo-bisimilar, that is, they are related by some pseudo-bisimulation. More specifically, we
exhibit a pseudo-bisimulation, relative to the ordering b1 = b, b2 = c, on the final automaton
L, such that M (α) and M (β) are pseudo-bisimilar. This is sufficient for proving equivalence,
since by Theorem 4.1, the languages M (α) and M (β) are then bisimilar, and by Corollary 3.9,
M (α) = M (β).
Define α′ to be the mixed expression
([b]cq)∗ cα
and define β ′ to be the mixed expression
([b]cq + bcp)∗ cb.
Notice that β = bpβ ′ + b.
We note that (using the notation of the definition of pseudo-bisimulation), A0 = ∅, A1 = {b},
and A2 = {b, c}. We claim that the following three relations form a pseudo-bisimulation:
R2 = {(M (α′ ), M (β ′ )),
(M (0), M (0))}
R1 = {(M ([b]qα′ ), M ([b]qβ ′ )),
(M (α), M (β))}
R0 = {(M (pα′ ), M (pβ ′ )),
(M (qα′ ), M (qβ ′ )),
(M (1), M (1)),
(M (0), M (0))}.
16
It is straightforward to verify that hR0 , R1 , R2 i is a pseudo-bisimulation on L, using the operators defined in the previous section. For instance, consider Db (M (α)), which is equal to M (D̂b (α))
by Proposition 5.4. We compute D̂b (α) here:
D̂b (α) = D̂b ((bp([b]cq)∗ c)∗ )b + T̂ ((bp([b]cq)∗ c)∗ )D̂b (b)
= D̂b (bp([b]cq)∗ c)(bp([b]cq)∗ c)∗ b + T̂ ((bp([b]cq)∗ c)∗ )0
= p([b]cq)∗ c(bp([b]cq)∗ c)∗ b
= pα′ .
Hence, Db (M (α)) = M (D̂b (α)) = M (pα′ ). The other cases are similar.
As we shall see shortly, there is a way to mechanically construct such a bisimulation to establish
the equivalence of two mixed expressions.
We remark that an alternative approach to establish equivalence of while programs based on
coalgebras is described by Rutten [1999]. This approach uses the operational semantics of the
programs instead of an algebraic framework.
7 Completeness
Thus far, we have established a coinductive proof technique for establishing the equality of mixed
languages (Section 3), and illustrated its use by showing the equality of two particular mixed languages specified by mixed expressions (Section 6), making use of the derivative calculus developed
in Section 5. A natural question about this proof technique is whether or not it can establish the
equivalence of any two mixed expressions that are equivalent (in that they specify the same mixed
language). In this section, we answer this question in the affirmative by formalizing and proving
a completeness theorem for our proof technique. In particular, we show that given two equivalent
mixed expressions, a finite bisimulation relating them can be effectively constructed, by performing only simple syntactic manipulations. In fact, we exhibit a deterministic procedure for deciding
whether or not two mixed expressions are equivalent.
In order to state our completeness theorem, we need a few definitions. We say that two mixed
ACI
expressions e1 and e2 are equal up to ACI properties, written e1 = e2 , if e1 and e2 are syntactically
equal, up to the associativity, commutativity, and idempotence of +. That is, e1 and e2 are equal
up to ACI properties if the following three rewriting rules can be applied to subexpressions of e1 to
obtain e2 :
e + (f + g) = (e + f ) + g
e+f =f +e
e + e = e.
Given a relation R̂ between mixed expressions, we define an induced relation R̂ACI as follows:
ACI
ACI
e1 R̂ACI e2 if and only if there exists e′1 , e′2 such that e1 = e′1 , e2 = e′2 , and e′1 R̂e′2 .
We define a syntactic bisimulation between two mixed expressions e1 and e2 having the same
type B → ∅ (for some B ⊆ B) to be a family R̂ = hR̂A iA∈℘(B) of relations such that
(1) for all mixed expressions e, e′ , if eR̂A e′ , then ⊢ e : A → ∅ and ⊢ e′ : A → ∅,
17
(2) eR̂B e′ ,
(3) for all mixed expressions e, e′ , if eR̂∅e′ , then ǫ̂(e) = ǫ̂(e′ ), and for all p ∈ P, D̂p (e)R̂BACI D̂p (e′ ),
and
ACI
′
(4) for all mixed expressions e, e′ , if eR̂A e′ (for A 6= ∅), then for all l ∈ lit(A), D̂l (e)R̂A\{base
(l)} D̂l (e ).
A syntactic bisimulation resembles a bisimulation, but is defined over mixed expressions, rather
than over mixed languages. The next theorem shows that any two equivalent mixed expressions are
related by a finite syntactic bisimulation, that is, a syntactic bisimulation R̂ where the number of
pairs in each relation R̂A is finite.
Theorem 7.1: For all mixed expressions e1 , e2 , of type A → ∅, M (e1 ) = M (e2 ) if and only if
there exists a finite syntactic bisimulation between e1 and e2 .
Proof: (⇐) It is easy to check that a syntactic bisimulation R̂ induces a bisimulation R such that
e1 R̂A e2 if and only if M (e1 )RA M (e2 ). The result then follows by Corollary 3.9.
(⇒) We first show how to construct, for every mixed expression e with ⊢ e : Ae → Be , a finitestate automaton M = (hSA iA∈℘(B)S, hδA iA∈℘(B) ) with transition functions δ∅ : S∅ × P → SB and
(for A 6= ∅) δA : SA × lit(A) → A∈℘(B) SA , satisfying the conditions
(1) δA (s, l) ∈ SA\{base(l)} ,
(2) the states of SA are mixed expressions having type A → Be ,
(3) e is a state of SAe ,
ACI
(4) if δ∅(s1 , p) = s2 , then D̂p (s1 ) = s2 , and
ACI
(5) if δA (s1 , l) = s2 , then D̂l (s1 ) = s2 .
We define the automaton by induction on the structure of e. The cases for 0, 1, p, l are straightforward. We focus on the remaining cases:
Case e = e1 + e2 : Assume by induction that we have automata M1 , M2 for e1 and e2 . Define:
SA = {f1 + f2 : f1 ∈ S1,A , f2 ∈ S2,A }
δ∅(f1 + f2 , p) = δ∅(f1 , p) + δ∅(f2 , p)
δA (f1 + f2 , l) = δ1,A (f1 , l) + δ2,A (f2 , l), for A 6= ∅, l ∈ lit(A).
Case e = e1 · e2 : Let ⊢ e1 : A1 → B1 . Assume by induction that we have automata M1 ,M2 for
18
e1 and e2 . Define:
P
SA = {f · e2 +
P
t·g+
δ∅(f
g :
g∈G
(t,g)∈E
f ∈ S1,A , E ⊆ Tests(A → B1 ) × S2,B1 , G ⊆ S2,A }
P
g, p) =
· e2 +
g∈G
P
δ2,∅(g, p) if B = ∅, ǫ̂(f ) = 1
δ1,∅(f, p) · e2 + δ2,∅(e2 , p) +
g∈G
P
δ2,∅(g, p)
otherwise
δ1,∅(f, p) · e2 +
g∈G
P
g, l) =
t·g+
δA (f · e2 +
g∈G
(t,g)∈E
P
P
δ2,A (g, l)
if base(l) ∈ A \ B1
δ1,A (f, l) · e2 +
Dl (t) · g +
g∈G
(t,g)∈E
P
P
δ1,A (f, l) · e2 +
δ2,A (g, l) if base(l) 6∈ A ∪ B1
t · δ2,B1 (g, l) +
P
g∈G
(t,g)∈E
if base(l) ∈ B1
δ1,A (f,
P l) · e2 + T̂ (f ) · δ2,B
P1 (e2 , l)+
δ2,A (g, l)
t
·
δ
(g,
l)
+
2,B
1
g∈G
(t,g)∈E
for A 6= ∅, l ∈ lit(A).
Case e = e∗1 : Let ⊢ e1 : A1 → A1 . Assume by induction that we have an automaton M1 for e1 .
Define:
P
∗
f · e∗1 : γ ∈ {0, 1}, F ⊆ S1,A1 } if A = A1
{γ · e1 +
f
∈F
P
SA =
otherwise
f · e∗1 : F ⊆ S1,A }
{
f ∈F
P
f · e∗1 , p) =
δ∅(γ · e∗1 +
f ∈F
P
P
ǫ̂(f ) · δ1,∅(e, p) · e∗1 ,
δ1,∅(f, p) · e∗1 +
γ · δ1,∅(e1 , p) · e∗1 +
f ∈F
f ∈F
for A = A1
X
X
δ∅(
f · e∗1 , p) =
δ1,∅(f, p) · e∗1 , for A 6= A1 ,
f ∈F
f ∈F
f · e∗1 , l) =
f ∈F
P
P
ǫ̂(f ) · δ1,A (e, l) · e∗1 ,
δ1,A (f, l) · e∗1 +
γ · δ1,A (e1 , l) · e∗1 +
δA (γ ·
e∗1
+
P
f ∈F
f ∈F
for A 6= ∅, A = A1 , l ∈ lit(A)
X
X
δA (
f · e∗1 , l) =
δ1,A (f, l) · e∗1 , for A 6= ∅, A 6= A1 , l ∈ lit(A).
f ∈F
f ∈F
It is straightforward (if tedious) to verify that the resulting automaton satisfies properties (1)-(5)
given above.
This completes the construction of the finite state mixed automaton corresponding to e.
Given equivalent mixed expressions e1 and e2 of type A → ∅, a finite syntactic bisimulation
R̂ can be constructed as follows. First, construct the automata M1 and M2 corresponding to e1 and
19
e2 . Then, initialize R̂ to contain the pair (e1 , e2 ), and iterate the following process: for every (e, e′ )
in R̂, add the pairs (δ1,B (e, x), δ2,B (e′ , x)) (where e, e′ have type B → ∅), for all x. Perform this
iteration until no new pairs are added to R̂. This must terminate, because there are finitely many
pairs of states (e, e′ ) with e in M1 and e′ in M2 . It is straightforward to check that R̂ is a syntactic
bisimulation, under the assumption that M (e1 ) = M (e2 ).
The procedure described in the proof of Theorem 7.1 can in fact be easily turned into a procedure
for deciding if two mixed expressions are equivalent. To perform this decision, construct R̂, and
verify that at all pairs of states (e, e′ ) in R̂, ǫ̂(e) = ǫ̂(e′ ). If this verification fails, then the two mixed
expressions are not equivalent; otherwise, they are equivalent.
The bisimulation in Section 6 is indeed a bisimulation induced by a syntactic bisimulation on
the mixed expressions α and β.
8 Conclusions and Future Work
We believe that proofs of equivalence between mixed expressions such as α and β via bisimulation
are in general more easily derived than ones obtained through a sound and complete axiomatization
of KAT. Given two equivalent mixed expressions, we can exhibit a bisimulation using the purely
mechanical procedure underlying Theorem 7.1: use the derivative operators to construct a finite
bisimulation in which the two expressions are paired. In contrast, equational reasoning typically
requires creativity.
The “path independence” of a mixed automaton (condition A2) gives any mixed automaton
a certain form of redundancy. This redundancy persists in the definition of bisimulation, and is
the reason why a pseudo-bisimulation, a seemingly weaker notion of bisimulation, gives rise to
a bisimulation. An open question is to cleanly eliminate this redundancy; a particular motivation
for doing this would be to make proofs of expression equivalence as simple as possible. Along
these lines, it would be of interest to develop other weaker notions of bisimulation that give rise to
bisimulations; pseudo-bisimulations require a sort of “fixed variable ordering” that does not seem
absolutely necessary.
Another issue for future work would be to give a class of expressions wider than our mixed expressions for which there are readily understandable and applicable rules for computing derivatives.
In particular, a methodology for computing derivatives of the KAT expressions defined by Kozen
[1997] would be nice to see. Intuitively, there seems to be a tradeoff between the expressiveness
of the regular expression language and the simplicity of computing derivatives (in the context of
KAT). Formal work towards understanding this tradeoff could potentially be quite useful.
Acknowledgments
The authors wish to thank Dexter Kozen for helpful comments on a draft of this paper, as well
as the anonymous referees who helped improve the presentation of the results. The second author was supported by NSF under grant CTC-0208535, by ONR under grants N00014-00-1-03-41
and N00014-01-10-511, and by the DoD Multidisciplinary University Research Initiative (MURI)
program administered by the ONR under grant N00014-01-1-0795.
20
References
Brzozowski, J. A. (1964). Derivatives of regular expressions. Journal of the ACM 11(4), 481–
494.
Conway, J. H. (1971). Regular Algebra and Finite Machines. London, UK: Chapman and Hall.
Hopcroft, J. E. and J. D. Ullman (1979). Introduction to Automata Theory, Languages, and Computation. Addison Wesley.
Kaplan, D. (1969). Regular expressions and the equivalence of programs. Journal of Computer
and Systems Sciences 3, 361–386.
Kleene, S. C. (1956). Representation of events in nerve nets and finite automata. In C. E. Shannon
and J. McCarthy (Eds.), Automata Studies, pp. 3–41. Princeton, NJ: Princeton University
Press.
Kozen, D. (1994). A completeness theorem for Kleene algebras and the algebra of regular events.
Information and Computation 110, 366–390.
Kozen, D. (1997). Kleene algebra with tests. Transactions on Programming Languages and Systems 19(3), 427–443.
Kozen, D. (1998). Typed Kleene algebra. Technical Report 98-1669, Computer Science Department, Cornell University.
Kozen, D. (1999). On Hoare logic and Kleene algebra with tests. In Proceedings of the Conference on Logic in Computer Science (LICS’99), pp. 167–172. IEEE Computer Society Press.
Kozen, D. (2002). On Hoare logic, Kleene algebra, and types. In Scope of Logic, Methodology,
and Philosophy of Science: Volume 1 of the 11th Int. Congress Logic, Methodology and
Philosophy of Science, Cracow, August 1999, Volume 315 of Studies in Epistemology, Logic,
Methodology, and Philosophy of Science, pp. 119–133. Kluwer.
Kozen, D. (2003). Automata on guarded strings and applications. Matématica Contemporânea 24, 117–139.
Kozen, D. and F. Smith (1996). Kleene algebra with tests: Completeness and decidability. In
Proceedings of the 10th Workshop on Computer Science Logic (CSL’96), Volume 1258 of
Lecture Notes in Computer Science, pp. 244–259. Springer-Verlag.
Rutten, J. J. M. M. (1998). Automata and coinduction (an exercise in coalgebra). In Proceedings
of CONCUR’98, Volume 1466 of Lecture Notes in Computer Science, pp. 193–217.
Rutten, J. J. M. M. (1999). A note on coinduction and weak bisimilarity for while programs.
Theoretical Informatics and Applications (RAIRO) 33, 393–400.
Rutten, J. J. M. M. (2000). Universal coalgebra: a theory of systems. Theoretical Computer
Science 249(1), 3–80.
Salomaa, A. (1966). Two complete axiom systems for the algebra of regular events. Journal of
the ACM 13(1), 158–169.
21
| 2cs.AI
|
Block Sparse Memory Improved
Proportionate Affine Projection Sign
Algorithm
gl n
h n h0 n h1 n
sgn
hˆ n [hˆ 0 n , hˆ 1 n ,
y n y n y n 1
(2)
gk n
in which 1P
hˆ k n
2
(12)
1
2L
1
2 N i 0
N 1
, g N 1 n 1P (13)
hˆ k n
2
hˆ i n
,
(14)
2
is a P-length column vector of all ones, and
P
hˆ
2
j 1 kN j
n
, k 0,1,
, N 1 . The weight
update equation for BS-MIP-ASPA is
M most
P n g n
x n , P1 n 1 ,
x gs n P n sgn e n ,
(4)
(5)
hˆ n hˆ n 1
y n M 1 , (6)
T
x gs n
xTgs n x gs n
where P1 n 1 also contains the first
where M is called the projection order. In [5], MIP-APSA proposed the
following weight update:
, g L 1 n ,
, hˆ N 1 n ],
g n g0 n 1P , g1 n 1P ,
recent input vectors together:
e n y n X n hˆ n 1 ,
is a small
then the control matrix g n in (7)-(8) is be replaced by
hˆ n be the L1 adaptive filter coefficient vector which estimates the
T
M 1 columns of P n 1 ,
and L N P ,
time index, and v n is the background noise plus near-end signals. Let
x n M 1 ,
denotes the Hadamard
Algorithm design: In network echo cancellation, the network echo path
is typically characterized by a bulk delay dependent on network loading,
encoding, and jitter buffer delays and an “active” dispersive region in the
range of 8-12 ms duration [1]. Meanwhile, it is well-known that NLMS
is preferred over PNLMS for dispersive system. Therefore, considering
the block-sparse characteristic of the network impulse response, the BSPNLMS algorithm was proposed to improve the PNLMS algorithm by
exploiting this special block-sparse characteristic, in which BS-PNLMS
used the same step-size within each block and the step-sizes for each
block were proportionate to their relative magnitude [7]. We propose to
take in account the block-sparse characteristic and partition the MIPAPSA adaptive filter coefficients into N groups with group-length P ,
T denotes transposition, L is the filter length, n is the
X n x n x n 1
(11)
, L 1 , is a small positive constant
takes the sign of each element of a vector, and
(3)
true echo path vector h n at iteration n , and group the
,
positive constant. Compared with RIP-PAPSA, MIP-PAPSA takes into
account the ‘proportionate history’ from the last M moments of time.
More details can be found in [5]-[6].
(1)
T
(10)
xTgs n x gs n
product, P1 n 1 contains the first
T
hL 1 n ,
(9)
x gs n
that avoids division by zero, the operation
filtered through the echo path h n to get the desired signal y n .
x n L 1 ,
(8)
x n , P1 n 1 ,
where 1 1 , l 0,1,
Review of MIP-APSA: For echo cancellation, the far-end signal x n is
x n x n x n 1
.
L 1
2 i 0 hˆi n
hˆ n 1 hˆ n
Introduction: Adaptive filters have been widely used in various
applications of system identification in which the normalized least mean
square (NLMS) algorithm is well-known due to its simplicity, but suffers
from slow convergence for colored input [1]. The affine projection
algorithm (APA) provides better convergence for colored input compared
with NLMS [2]. Meanwhile, the family of affine projection sign
algorithm (APSA) has been proposed to improve the performance of
APA under impulsive noise together with lower complexity [3]. In order
to exploit the sparsity of some echo paths, the real-coefficient improved
proportionate APSA (RIP-APSA) was proposed [4], and a memory
improved proportionate APSA (MIP-APSA) was further proposed to
achieve improved steady-state misalignment with similar computational
complexity compared with RIP-APSA [5]. Recently, the block-sparse
improved proportionate NLMS (BS-IPNLMS) algorithm was proposed
to improve the performance of IPNLMS for identifying block-sparse
systems [7]. In this Letter, motived by both BS-PNLMS and MIP-APSA,
we will propose a block sparse memory improved proportionate APSA
(BS-MIP-APSA) algorithm, which not only inherits the performance
improvement for block-sparse system identification, but also achieves
robustness to impulsive noise and the efficiency of MIP-APSA.
y(n) xT (n)h(n) v(n)
1 hˆl n
x gs n P n sgn e n ,
A block sparse memory improved proportionate affine projection sign
algorithm (BS-MIP-APSA) is proposed for block sparse system
identification under impulsive noise. The new BS-MIP-APSA not only
inherits the performance improvement for block-sparse system
identification, but also achieves robustness to impulsive noise and the
efficiency of the memory improved proportionate affine projection sign
algorithm (MIP-APSA). Simulations indicate that it can provide both
faster convergence rate and better tracking ability under impulsive
interference for block sparse system identification as compared to APSA
and MIP-APSA.
g n g0 n , g1 n ,
2L
P n g n
J. Liu, and S. L. Grant
super-script
1
(15)
(16)
(17)
M 1 columns of P n 1 .
It should be noted that the proposed BS-MIP-APSA includes both APSA
and MIP-APSA. The MIP-APSA algorithm is a special case of proposed
BS-MIP-APSA with group length P 1 . Meanwhile, when P is
chosen as L , the BS-MIP-APSA algorithm degenerates to APSA.
(7)
1
Complexity: Compared with traditional RIP-APSA and MIP-APSA, the
extra computational complexity of the BS-MIP-APSA arises from the
computation of the l2 norm in (14), which requires
10
L multiplications
0
Normalized Misalignment
and N square roots. The complexity of the square root could be reduced
through a look up table or Taylor series [7]. Meanwhile, the increase in
complexity can be offset by the performance improvement as shown in
the simulation results.
Simulation results: In our simulation, the echo path is a L 512 finite
impulse response (FIR) filter, and the adaptive filter is the same length.
We generated colored input signals by filtering white Gaussian noise
through a first order system with a pole at 0.8. Independent white
Gaussian noise is added to the system background with a signal-to-noise
ratio (SNR) of 40 dB. The impulsive noise with signal-to-interference
ratio (SIR) of 0 dB is generated as a Bernoulli-Gaussian (BG) distribution.
BG is a product of a Bernoulli process and a Gaussian process, and the
probability for Bernoulli process is 0.1. The performance was evaluated
through the normalized misalignment: 10log10 ( h hˆ
2
2
-10
-20
-30
-50
Normalized Misalignment
0
-0.5
100
150
200
250
(a)
300
350
400
450
2
3
4
Iterations
5
6
7
8
x 10
4
0
0.5
50
1
5
1
0
0
Fig. 2 Normalized misalignment of APSA, MIP-APSA, and BS-MIPAPSA for colored input signal.
2
h 2 ) . In
order to evaluate the tracking ability, we switch the echo path from the
one-cluster block-sparse system of Fig. 1(a) to the two-cluster blocksparse system of Fig. 1(b).
-1
APSA
MIP-APSA
BS-MIP-APSA
-40
-5
-10
-15
-20
-25
APSA
MIP-APSA
BS-MIP-APSA
500
-30
0.4
-35
0.2
0
1
2
3
4
Iterations
5
6
7
8
x 10
4
Fig. 3 Normalized misalignment of APSA, MIP-APSA, and BS-MIPAPSA for speech input signal.
-0.2
-0.4
0
0
50
100
150
200
250
(b)
300
350
400
450
References
500
Fig. 1 Two block-sparse systems used in the simulations: (a) one-cluster
block-sparse system, (b) two-cluster block-sparse system.
1
The APSA and MIP-APSA algorithms are compared with BS-MIPAPSA. The parameters are 0.001 , 0.01 , 0.01 , 0 ,
2
M 2 , and P 4 . In the first case, we show the normalized
misalignment for colored input in Fig. 2. We could see that the proposed
BS-MIP-APSA achieves both faster convergence rate and better tracking
ability. In Fig. 3, the performance of BS-MIP-APSA is compared with
APSA and MIP-APSA for speech input signal, and we found that our
proposed algorithm demonstrates better performance too.
3
4
Conclusion: We have proposed a block-sparse memory improved affine
projection sign algorithm to improve the performance of block-sparse
system identification. Simulations demonstrate the proposed algorithm
has both faster convergence speed and tracking ability for block-sparse
system identification compared with APSA and MIP-APSA algorithms.
5
6
Acknowledgments: This work is supported by Wilkens Missouri
Endowment.
7
J. Liu, and S. L. Grant (Department of Electrical and Computer
Engineering, Missouri University of Science and Technology, Rolla, MO,
USA)
E-mail: [email protected]
2
Benesty, J., Gänsler, T., Morgan, D.R., Sondhi, M.M, and Gay,
S.L., Advances in network and acoustic echo cancellation.
Springer, 2001.
Ozeki, K., and Umeda, T.: ‘’An adaptive filtering algorithm
using an orthogonal projection to an affine subspace and its
properties’, Electron. Commun. Jpn., 1984, 67-A, (5), pp. 1927
Shao, T., Zheng, Y. R., and Benesty, J.: ‘An affine projection
sign algorithm robust against impulsive interferences’, IEEE
Signal Process. Lett., 2010, 17, (4), pp. 327–330.
Yang, Z., Zheng, Y.R., and Grant, S.L.: ‘Proportionate Affine
Projection Sign Algorithms for Network Echo Cancellation’,
IEEE Trans. Audio, Speech, Lang. Process., 19, (8), pp. 22732284.
Albu, F., and Kwan, H.K.: ‘Memory improved proportionate
affine projection sign algorithm’, Electron. Lett., 2012, 48,
(20), pp. 1279-1281.
Paleologu, C., Ciochina, S., and Benesty, J.: ‘An efficient
proportionate affine projection algorithm for echo
cancellation’, IEEE Signal Process. Lett., 2010, 17, (2), pp.
165–168
Liu, J., and Grant, S.L.: ‘Proportionate adaptive filtering for
block-sparse system identification’, arXiv preprint
arXiv:1508.04172, 2015
| 3cs.SY
|
arXiv:1311.5107v1 [math.AC] 20 Nov 2013
FACTORIZATION IN THE SELF-IDEALIZATION OF A PID
GYU WHAN CHANG AND DANIEL SMERTNIG
a b
Abstract. Let D be a principal ideal domain and R(D) = {
|
0 a
a, b ∈ D} be its self-idealization. It is known that R(D) is a commutative
noetherian ring with identity, and hence R(D) is atomic (i.e., every nonzero
nonunit can be written as a finite product of irreducible elements). In this
paper, we completely characterize the irreducible elements of R(D). We then
use this result to show how to factorize each nonzero nonunit of R(D) into
irreducible elements. We show that every irreducible element of R(D) is a
primary element, and we determine the system of sets of lengths of R(D).
1. Introduction
Let R be a commutative noetherian ring. Then R is atomic, which means that
every nonzero nonunit element of R can be written as a finite product of atoms
(irreducible elements) of R. The study of non-unique factorizations has found
a lot of attention. Indeed this area has developed into a flourishing branch of
Commutative Algebra (see some surveys and books [3, 6, 8, 5]). However, the focus
so far was almost entirely on commutative integral domains, and only first steps
were done to study factorization properties in rings with zero-divisors (see [2, 7]).
In the present note we study factorizations in a subring of a matrix ring over a
principal ideal domain, which will turn out to be a commutative noetherian ring
with zero-divisors.
To begin with, we fix our notation and terminology. Let R be a commutative
ring with identity and U (R) be the set of units of R. Two elements a, b ∈ R are said
to be associates if aR = bR. Clearly, if a = ub for some u ∈ U (R), then a and b are
associates. An a ∈ R is said to be irreducible if a = bc implies that either b or c is
associated with a. We say that R is atomic if every nonzero nonunit of R is a finite
product of irreducible elements. It is clear that noetherian rings are atomic (cf. [2,
Theorem 3.2]) and that 0 ∈ R is irreducible if and only if R is an integral domain.
A ring R is a half-factorial ring (HFR) (resp., bounded factorization ring (BFR))
if R is atomic and two factorizations of a nonzero nonunit into irreducible elements
have the same length (resp., for each nonzero nonunit x ∈ R, there is an integer
N (x) ≥ 1 so that for any factorization x = x1 · · · xn , where each xi is a nonunit,
we have n ≤ N (x)). R is a FFR (finite factorization ring) if R is atomic and each
nonzero nonunit has only finitely many factorizations into irreducibles, up to order
and associates. A nonzero nonunit x ∈ R is said to be prime (resp., primary) if xR
Date: 24.5.2013.
2010 Mathematics Subject Classification. 13A05, 13F15, 20M13, 20M14.
Key words and phrases. non-unique factorizations, matrix ring, noetherian ring, sets of lengths.
1
2
G.W. CHANG AND D. SMERTNIG
is a prime (resp., primary) ideal. Hence a prime element is primary but not vice
versa (for example, if Z is the ring of integers, then 4 ∈ Z is primary but not prime).
We say that R is a unique factorization ring (UFR) if every nonzero principal ideal
of R can be written as a product of principal prime ideals (cf. [2, Theorem 4.9]).
Clearly, a prime element is irreducible, and so a UFR is atomic.
For x ∈ R a nonzero nonunit, its set of lengths is defined as
L(x) = { k ∈ N | there exist irreducibles u1 , . . . , uk ∈ R with x = u1 · . . . · uk }.
Clearly, x is irreducible if and only if L(x) = { 1 }. If x ∈ U (R), we set L(x) = { 0 }.
The system of sets of lengths is defined as L(R) = L(x) | x ∈ R \ { 0 } . Sets
of lengths and invariants derived from them are some of the classical invariants
considered in non-unique factorization theory (see [8, Ch. 1.4]). The reader is
referred to [8] for undefined definitions and notations.
Let M be an R-module. The idealization R(+)M of M is a ring, which is defined
as an abelian group R ⊕ M , whose ring multiplication is given by (a, b) · (x, y) =
(ax, ay + bx) for all a, x ∈ R and b, y ∈ M . It is known that R(+)M is a noetherian
ring if and only if R is noetherian and M is finitely generated [4, Theorem 4.8].
Let D be an
integral
domain, Mat2×2 (D) be the ring of 2 × 2 matrices over D, and
a b
R(D) = {
| a, b ∈ D}. It is easy to show that R(D) is a commutative
0 a
ring with identity under the usual matrix addition and multiplication;
in particular,
a 0
embeds D into
R(D) is a subring of Mat2×2 (D). Clearly, the map a 7→
0 a
a b
R(D), and the map ϕ : D(+)D → R(D), given by ϕ(a, b) =
, is a ring
0 a
isomorphism. In view of this isomorphism, R(D) is called the self-idealization of D
(cf.[13]). There is also an isomorphism D[X]/hX 2 i → R(D) mapping a+bX +hX 2 i
a b
to
. Some factorization properties of R(+)M have been studied in [2,
0 a
Theorem 5.2]. For more on basic properties of R(+)M (and hence of R(D)), see
[4] or [11, Section 25].
Let D be a principal ideal domain (PID). Then R(D) is noetherian, and thus
R(D) is atomic. In Section 2, we first characterize the irreducible elements of R(D),
and we then use this result to show how to factorize each nonzero nonunit
of R(D)
0 1
into irreducible elements via the factorization of D. We show that
is
0 0
the unique prime element (up to associates) of R(D). We prove that every nonzero
nonunit of R(D) can be written as a product of primary elements. Finally, in
Section 3, we completely describe the system of sets of lengths L(R(D)).
We write N = { 1, 2, 3, . . . } for the set of positive integers, and N0 = N ∪ { 0 } for
the set of non-negative integers.
2. Factorization in R(D) when D is a PID
Let D be an integral domain, and
a b
R(D) = {
| a, b ∈ D}
0 a
FACTORIZATION IN THE SELF-IDEALIZATION OF A PID
3
1 0
be the self-idealization of D. Clearly,
is the identity of R(D).
0 1
a b
If α =
∈ R(D), then nr(α) = a is the norm, and this is a ring homo0 a
morphism R(D) → D. Observe that α is a zero-divisor if and only if a = 0. We
write R(D)• for the monoid of non-zero-divisors of R(D).
We begin this paper by characterizing the units of R(D), which is very useful in
the proof of Theorem 5.
Lemma 1. (cf. [11, Theorem 25.1(6)]) If D is an integral domain, then α =
a b
∈ R(D) is a unit of R(D) if and only if a is a unit of D.
0 a
x y
1 0
x y
Proof. If α is a unit, then α ·
=
for some
∈ R(D).
0 x
0 1
0 x
Thus ax = 1, and
so a∈ U (D). Conversely,
assume
that
a is a unit,
and let
a b
u −bu2
1 0
u −bu2
−1
u = a . Then
=
and
∈ R(D).
0 a
0
u
0 1
0
u
Thus α is a unit.
For an arbitrary commutative ring R, there can be two elements a, b ∈ R such
that a and b are associates but a 6= ub for all u ∈ U (R) (see, for example, [2,
Example 2.3]). This cannot happen in the self-idealization of an integral domain.
Lemma 2. Let
D
a
such that α =
0
be an integral
domain
and α, β ∈ R(D) and let a, b, x, y ∈ D
b
x y
and β =
. The following statements are equivalent.
a
0 x
(1) α and β are associates.
(2) There exists θ ∈ U (R(D)) such that β = θα.
(3) There exists u ∈ U (D) such that x = au and y ≡ bu mod a.
Proof. (1) ⇒ (2): If α and β are associates,
are some
so
then there
γ, δ ∈ R(D)
a1 b 1
x1 y1
that α = βγ and β = αδ. Hence if γ =
and δ =
, then
0 a1
0 x1
a = xa1 and x = ax1 , and
a1 , x1 ∈ U (D). Thus γ, δ ∈ U (R(D)) by Lemma 1.
so
u v
(2) ⇒ (3): Let θ =
. By Lemma 1, u ∈ U (D). From β = θα it follows
0 u
that x = au and y = av + bu ≡ bu mod a.
u v
(3) ⇒ (2) and (1): Let v ∈ D be such that y = bu + av. Define θ =
.
0 u
Then θ ∈ U (R(D)) by Lemma 1 and β = θα. Thus, α and β are associates.
We write α ≃ β if α, β ∈ R(D) are associates.
a b
Lemma 3. Let D be a PID and let α =
∈ R(D)• . If a = cd with coprime
0 a
c, d ∈ D, then there exist γ, δ ∈ R(D) with α = γδ and nr(γ) = c, nr(δ) = d.
This representation is unique in the sense that, if γ ′ , δ ′ ∈ R(D) with α = γ ′ δ ′ and
nr(γ ′ ) ≃ c, nr(δ ′ ) ≃ d, then γ ≃ γ ′ and δ ≃ δ ′ .
4
G.W. CHANG AND D. SMERTNIG
Proof. Existence: Since 1 ∈
GCD(c,
d) and Dis a PID,
there exist e, f ∈ D such
c e
d f
that b = cf + ed. Then γ =
and δ =
are as claimed.
0 c
0 d
′
d′ f ′
c e′
with c′ , e′ , d′ , f ′ ∈ D and
and δ ′ =
Uniqueness: Let γ ′ =
′
0 d′
0 c
suppose that α = γ ′ δ ′ . Let u, v ∈ U (D) such that c′ = cu and d′ = dv. Since
c′ d′ = cd, necessarily v = u−1 . Since cf + ed = c′ f ′ + e′ d′ = c(f ′ u) + d(e′ v), we
have c(f ′ u) ≡ cf mod d and f ′ u ≡ f mod d, i.e., f ′ ≡ f v mod d. Therefore
δ ′ ≃ δ and similarly γ ′ ≃ γ.
Corollary 4. Let D be a PID and let α ∈ R(D)• \ U (R(D)). Then there exist
β1 , . . . , βn ∈ R(D)• of pairwise distinct prime power norm, such that α = β1 ·. . .·βn .
This representation is unique up to order and associates.
Proof. Let nr(α) = pe11 · . . . · penn with n ≥ 0, p1 , . . . , pn ∈ D pairwise distinct prime
elements and e1 , . . . , en ≥ 1. By induction on n and the previous lemma, there
exist β1 , . . . , βn ∈ R(D)• such that α = β1 · . . . · βn and nr(βi ) = pei i for all i ∈ [1, n].
′
Suppose α = β1′ · . . . βm
is another such factorization. Since D is a UFD, then
′
m = n and there exists a permutation π ∈ Sn such that nr(βπ(i)
) ≃ nr(βi ) for all
i ∈ [1, n]. The uniquenes statement of the previous lemma implies βi′ ≃ βi for all
i ∈ [1, n].
As a consequence, to study factorizations of α ∈ R(D)• , it is sufficient to study
factorizations of α ∈ R(D)• with prime power norm.
We next give the first main result of this paper, which completely characterizes
the irreducible elements of R(D) when D is a PID.
a b
Theorem 5. Let D be a PID and α =
∈ R(D). Then α is irreducible
0 a
if and only if either (i) a = 0 and b ∈ U (D), (ii) a = p or (iii) a = upn and
1 ∈ GCD(a, b) for some prime p ∈ D, u ∈ U (D), and integer n ≥ 2.
b 0
0 1
Proof. Necessity. Assume that a = 0, and let β =
and γ =
.
0 b
0 0
Then α = β · γ and αR(D)
6= βR(D)
because b 6= 0. Hence αR(D) = γR(D), and
x y
so γ = α · δ for some δ =
∈ R(D). Thus bx = 1.
0 x
Next, assume that a 6= 0. If a is not of the form upn , then Lemma 3 implies
that α = β · γ with nr(β) and nr(γ) nonzero nonunits. Hence α is not irreducible, a
contradiction. Thus a = upn for some prime p ∈ D, u ∈ U (D), and integer n ≥ 1.
If n = 1, then up is also a prime element of D and we have case (ii).
Finally, assume that n ≥ 2 and pk ∈ GCD(a, b) for someinteger
k ≥ 1. Let
p
0
and ξ =
b = b1 pk , where b1 ∈ D. Then α = θ · ξ, where θ =
0 p
upn−1 b1 pk−1
, but θ, ξ 6∈ αR(D), a contradiction. This completes the proof.
0
upn−1
x2 y2
x1 y1
. We
and γ =
Sufficiency. Let α = β · γ, where β =
0 x2
0 x1
will show that β or γ is a unit, and thus α is irreducible.
FACTORIZATION IN THE SELF-IDEALIZATION OF A PID
5
Case 1. a = 0 and b ∈ U (D). Note that x1 = 0 or x2 = 0; so for convenience,
let x2 = 0. Then x1 y2 = b, and since b ∈ U (D), we have x1 ∈ U (D). Thus β is a
unit of R(D) by Lemma 1.
Case 2. a = p for a prime p ∈ D. Then α = β · γ implies that either x1 or x2 is
a unit in D. Hence β or γ is a unit in R(D) by Lemma 1.
Case 3. a = upn for a prime p ∈ D,u ∈ U (D), n≥ 2 and 1
∈ GCD(a, b). Since p
wpn−k
y
vpk
x
and γ =
is a prime and α = β · γ, we have β =
k
0
wpn−k
0 vp
for some 0 ≤ k, n − k ≤ n, x, y ∈ D, and v, w ∈ U (D) with vw = u. Hence
b = pk vy + pn−k wx, and thus k = 0 or n − k = 0 because a and b are coprime.
Therefore β or γ is a unit in R(D) by Lemma 1.
a b
Corollary 6. Let D be a PID and α =
∈ R(D) be a nonzero nonunit
0 a
such that c ∈ GCD(a, b), a = ca1 , and b = cb1 for some c, a1 , b1 ∈ D. Let c =
km
upe11 · · · penn and a1 = q1k1 · · · qm
(when a 6= 0) be prime factorizations of c and a1 ,
respectively, where u ∈ U (D). The following is a factorization of α into irreducible
elements.
ei
0 u Qn
pi 0
(1) If a = 0, then α =
.
i=1
0 0
0 pi
!
ei
k
Qm
qj j cj
u 0 Qn
pi 0
) for some
(2) If a 6= 0, then α =
( i=1
)( j=1
k
0 u
0 pi
0 qj j
cj ∈ D with 1 ∈ GCD(cj , qj ).
Proof. (1) Clear.
(2) We first note that α =
c 0
0 c
=
c
0
u 0
0 u
0
c
a1
0
p1
0
0
p1
b1
a1
e1
···
and
pn
0
0
pn
en
.
Next, assume that a1 = b2 d2 for some b2 , d2 ∈ D with 1 ∈ GCD(b2 , d2 ). Then
there are some
x, y ∈ D such
that b
2 (xb
1 ) + d2 (yb1 ) = b1 because D is a PID,
a1 b 1
b2 yb1
d2 xb1
and hence
=
. Note that 1 ∈ GCD(a1 , b1 );
0 a1
0 b2
0 d2
hence 1 ∈ GCD(b2 , yb1 ) and 1 ∈ GCD(d
! 2 , xb1 ). So by repeating this process, we
kj
Qm
cj
q
a1 b 1
for some cj ∈ D with 1 ∈ GCD(cj , qj ).
= j=1 j
have
k
0 a1
0 qj j
Corollary 7. If D is a PID, then
associates) of R(D).
0 1
0 0
is the unique prime element (up to
Proof. Clearly, prime elements are
and hence by Theorem 5, we have
irreducible,
a b
three cases to consider. Let α =
∈ R(D) be irreducible.
0 a
6
G.W. CHANG AND D. SMERTNIG
0 1
Case 1. a = 0 and b ∈ U (D). Note that if we set I =
, then α =
0 0
b 0
b 0
I·
and
∈ U (R(D)) by Lemma 1; so α and I are associates.
0 b
0 b
x y
c d
Let β =
,γ =
∈ R(D). Then βγ ∈ IR(D) if and only if
0 x
0 c
xc = 0; so if x = 0 (for convenience), then β ∈ IR(D). Thus I is a prime.
Cases 2 and 3. a 6= 0. Note that
2
2
a b−1
a 2a(b − 1)
=
0
a
0
a2
a b
a b−2
=
∈ αR(D),
0 a
0
a
a b−1
but
6∈ αR(D) because a 6∈ U (D). Thus α is not a prime.
0
a
For zero-divisors and elements with prime power norm, the following lemma
further refines Corollary 6, by giving all possible factorizations, up to order and
associates. The general case can be obtained in combination with Corollary 4.
a b
Lemma 8. Let D be a PID, and let α =
∈ R(D) \ { 0 } with a, b ∈ D.
0 a
(1) Suppose a = 0 and b = q1 · . . . · qn , with (possibly associated) prime powers
q1 , . . . , qn ∈ D. Then, for every choice of a1 , . . . , an ∈ D,
n
0 1 Y qi ai
α=
,
0 qi
0 0
i=1
and this is a factorization into irreducibles if and only if for all i ∈ [1, n]
either qi is prime or 1 ∈ GCD(qi , ai ).
(2) Suppose a = pn with p ∈ D a prime element and n ∈ N. For all l ∈ [1, n]
let ml ∈ N0 and for all j ∈ [1, ml ] let al,j ∈ D. Then
ml l
n Y
Y
p al,j
α=
0 pl
l=1 j=1
Pn
Pn n−l Pml
if and only if n =
( j=1 al,j ). This is a
l=1 ml l and b =
l=1 p
product of irreducibles if and only if 1 ∈ GCD(al,j , p) for all l ∈ [2, n] and
j ∈ [1, ml ].
Up to order and associativity of the factors, all the factorizations of α are of this
form.
Proof. This is checked by a straightforward calculation. The statement about the
irreducibles follows from the characterization of the irreducible elements in Theorem
5. That every representation of α as a product of irreducibles is, up to order and
associates, one of the stated ones also follows from this characterization.
Corollary 9. Let D be a PID.
(1) R(D) is a BFR.
FACTORIZATION IN THE SELF-IDEALIZATION OF A PID
7
(2) R(D) is a FFR if and only if D/pD is finite for all prime elements p ∈ D.
(3) If D is a field, then every nonzero nonunit of R(D) is a prime, and hence
R(D) is a UFR with a unique nonzero (prime) ideal.
a b
Proof. (1) By Corollary 6, R(D) is atomic, and if α =
∈ R(D), then
0 a
the lengths of factorizations of α into irreducible elements are less than or equal to
that of the prime factorization of a or b in D, plus at most one. Thus R(D) is a
BFR.
(2) Suppose first that D/pD is finite for all prime elements p ∈ D. Then also
D/pn D is finite for all n ≥ 1 and all prime elements p ∈ D. Hence, by the Chinese
Remainder Theorem, D/cD is finite for all nonzero c ∈ D.
Let c ∈ D• . By Lemma 2(3) there exist, up to associativity, only finitely many
elements γ ∈ R(D) with nr(γ) ≃ c. If α ∈ R(D)• and γ|α, then nr(γ)| nr(α),
and therefore there are, up to associativity, only finitely many irreducibles that can
possibly divide α. Together with (1), this implies that every α ∈ R(D)• has only
finitely many
factorizations.
0 b
If α =
∈ R(D) is a zero-divisor, then every factorization has exactly one
0 0
0 1
factor associated to
and if γ is any other factor in the factorization then
0 0
nr(γ) | b (cf. Lemma 8(1)). By the same argument as before, α has only finitely
many factorizations.
For the converse, suppose that p ∈ D is a prime element and |D/pD| = ∞. Since
2
p
0
p a
p −a
=
,
0 p2
0 p
0 p
2
p
0
for all a ∈ D,
has infinitely many (non-associated) factorizations in
0 p2
R(D).
a b
(3) Let α =
∈ R(D) be a nonzero nonunit. Since D is a field, by
0 a
0 1
Lemma 1, a = 0 and b ∈ U (D). Hence α is associated with I :=
, and
0 0
so α is a prime by the proof of Corollary 7. Thus R(D) is a UFR and IR(D) is a
unique nonzero (prime) ideal of R(D).
If D is a PID but not a field, we will see in Corollary 15 that R(D) is not a UFR,
even when D is the ring of integers.
We next prove that every nonunit of R(D) can be written as a (finite) product
of primary elements.
√
Lemma 10. Let R be a commutative ring. If a ∈ R is such that aR is a maximal
ideal, then aR is primary.
√
√
Proof. Let x, y ∈ R be such
√
√ that xy ∈ aR but x√ 6∈ aR. Note that aR (
aR + xR; so aR + xR = aR + xR = R because aR is a maximal ideal. Hence
1 = as + xt for some s, t ∈ R. Thus y = y(as + xt) = a(ys) + (xy)t ∈ aR.
8
G.W. CHANG AND D. SMERTNIG
Corollary 11. If D is a PID, then every irreducible element of R(D) is primary.
In particular, each nonzero nonunit of R(D) can be written as a finite product of
primary elements.
a b
Proof. Let α =
∈ R(D) be irreducible. By Theorem 5, there are three
0 a
cases that we have to consider.
Case 1. a = 0 and b ∈ U (D). By Corollary 7, α is a prime, and hence a primary
element.
Cases 2 and 3. a = upn for some prime
pelement p ∈ D, u ∈ U (D), and n ∈ N.
By Lemma 10, it suffices to show that αR(D) is a maximal ideal. Let β =
p
0 d
x y
∈ R(D), then δ 2 = 0,
∈ R(D) \ αR(D). Note that if δ =
0 0
0 x
p
p
x 0
upn
0
6∈ αR(D) and
∈
and hence δ ∈ αR(D). Hence
0 x
0
upn
p
p
p 0
αR(D). But then
∈
αR(D). Note also that if x ∈ pD, then
0 p
p
x 0
p 0
x1 0
x = px1 for some x1 ∈ D, and so
=
∈ αR(D),
0 x
0 p
0 x1
a contradiction.
So
x
∈
6
pD,
and
hence
xz
+
pz
=
1
for
some
z1 , z
1
2
2 ∈ D. Thus
z1 0
1 0
z2 0
p 0
0 −yz1
=β·
+
+
∈ βR(D) +
0 z1
0 1
0 z2
0 p
0
0
p
p
αR(D). Therefore αR(D) is maximal.
Remark 12. In view of Corollary 11, Corollary 4 in fact corresponds to the (unique)
primary decomposition of αR(D), as every prime ideal of R(D), except for 0(+)D,
is maximal (cf. [4, Theorem 3.2]).
Associativity is a congruence relation on (R(D)• , ·), and we denote by R(D)•red
the corresponding quotient monoid. Corollary 4 may also be viewed as a monoid
`
isomorphism R(D)•red ∼
= p R(D(p) )•red , where the coproduct is taken over all associativity classes of prime elements p of D, and D(p) is the localization at pD.
3. The sets of lengths in R(D) when D is a PID
a b
Let D be an integral domain and R(D) = {
| a, b ∈ D}. In this
0 a
section, we characterize the sets of lengths in R(D) when D is a PID.
Lemma 13. Let D be a PID and α, β ∈ R(D).
(1) If αβ 6= 0, then L(α) + L(β) ⊂ L(αβ).
(2) If nr(α) and nr(β) are coprime, then L(α) + L(β) = L(αβ).
Proof. (1) Clear.
(2) Let n ∈ L(αβ). Then there exist irreducible γ1 , . . . , γn ∈ R(D)• such
that αβ = γ1 · . . . · γn . Then also nr(α) nr(β) = nr(γ1 ) · . . . · nr(γn ). Since
1 ∈ GCD(nr(α), nr(β)), we may without loss of generality assume nr(α) ≃ nr(γ1 ) ·
. . . · nr(γk ) and nr(β) ≃ nr(γk+1 ) · . . . · nr(γn ) for some k ∈ [0, n]. By Lemma 3,
therefore α ≃ γ1 ·. . .·γk and β ≃ γk+1 ·. . .·γn , and n = k+(n−k) ∈ L(α)+L(β).
FACTORIZATION IN THE SELF-IDEALIZATION OF A PID
9
For a prime element p ∈ D we denote by vp : D → N0 ∪ { ∞ } the corresponding
valuation, i.e., vp (0) = ∞ and vp (apk ) = k if k ∈ N0 and a ∈ D• is coprime to p.
a b
Theorem 14. Let D be a PID, α ∈ R(D) and suppose α =
with a, b ∈ D.
0 a
(1) If a = 0, and b = pe11 · . . . · penn with pairwise non-associated prime elements
p1 , . . . , pn ∈ D and e1 , . . . , en ∈ N, then L(α) = [1 + n, 1 + e1 + . . . + en ].
(2) Let p ∈ D be a prime element, n ∈ N and suppose a = pn and vp (b) = k ∈
N0 ∪ { ∞ }. Then L(α) = { 1 } if and only if k = 0 or n = 1. If k ≥ n − 1,
then
[3, n − 2] ∪ { n } ⊂ L(α) ⊂ [2, n − 2] ∪ { n },
and if k ∈ [1, n − 2], then
[3, k + 1] ⊂ L(α) ⊂ [2, k + 1].
Moreover, if k ≥ 1, then 2 ∈ L(α) if and only if n is even or k <
n
2.
Proof. (1) This is clear from Lemma 8(1), as every factorization of b into prime
powers gives a factorization of α (choose ai = 1), and conversely.
(2) The cases k = 0 and n = 1 are clear from Theorem 5, so from now on we
assume k ≥ 1 and n > 1. Let b = upk with u ∈ D and 1 ∈ GCD(u, p). We
repeatedly make use of Lemma 8(2), and the notation used there to describe a
factorization, without explicitly mentioning this fact every time.
Claim A: L(α) ⊂ [2, min{ k + 1, n }].
Proof. Because α is not an atom, 1 6∈ L(α). Any factorization of α is associated to
one in Lemma 8(2); we fix a factorization of α with notation as in the lemma. The
P
P
length of the factorization is then t = nl=1 ml . Since nl=1 ml l = n, clearly t ≤ n.
Pn
P ml
Moreover, necessarily ml = 0 for all l > n−(t−1). Since b = l=1 pn−l ( j=1
al,j ),
therefore k = vp (b) ≥ vp (pn−(n−t+1) ) = t − 1, i.e., t ≤ k + 1.
Claim B: 2 ∈ L(α) if and only if n is even or k <
n
2.
Proof. Suppose 2 ∈ L(α) and n is odd. Then n = l + (n − l) and b = pn−l al,1 +
pl an−l,1 with 1 ∈ GCD(al,1 , p) and 1 ∈ GCD(an−l,1 , p). Since n is odd, then
n − l 6= l and therefore k = vp (b) = min{ n − l, l } < n2 .
For the converse suppose first 1 ≤ k < n2 . Then n = k + (n − k), n − k > k and
b = pn−k · 1 + pk (u − pn−2k ) with 1 ∈ GCD(u − pn−2k , p). If n is even and k ≥ n2 ,
n
n
n
then n = n2 + n2 and b = p 2 (1 + (upk− 2 − 1)) with 1 ∈ GCD(upk− 2 − 1, p).
Claim C: If n − 1 ∈ L(α), then k = n − 2.
Proof. For a corresponding factorization we must have m1 = n − 2, m2 = 1,
and ml = 0 for all l > 2. Then b = pn−1 (a1,1 + . . . + a1,n−2 ) + pn−2 a2,1 with
1 ∈ GCD(a2,1 , p), whence k = vp (b) = n − 2.
Claim D: Let n ≥ 3 and k ≥ 2. If either k = 2 or n 6= 4, then 3 ∈ L(α).
Proof. Suppose first that n
p
(1)
α=
0
is odd and set b′ = b/p. Then
n−1
0
p
b′
α′ with α′ =
,
p
0
pn−1
10
G.W. CHANG AND D. SMERTNIG
and, by Claim B, 2 ∈ L(α′ ). Therefore 3 ∈ L(α).
If n is even, n ≥ 6, and k ≥ 3, then
2
p
α=
0
u
p2
n−2
p
0
u(pk−2 − pn−4 )
,
pn−2
where the first factor is irreducible and the second has a factorization of length 2
by Claim B.
If k = 2, then
2 n−2
p 0
p
u
α=
0 p
0
pn−2
is a factorization of length 3.
Claim E: If k ≥ n − 1, then n ∈ L(α).
Proof. We use Lemma 8(2). Set m1 = n, a1,1 = upk−(n−1) and a1,2 = . . . = al,n =
0. Then pn−1 (upk−(n−1) + 0 + . . . + 0) = b.
Claim F: If k ∈ [1, n − 2], then [3, k + 1] ⊂ L(α).
Proof. If n ≤ 3 or k = 1, then the claim is trivially true, so we may assume k ≥ 2.
We proceed by induction on n. Suppose n ≥ 4, and that the claim is true for n − 1.
Let b′ = b/p and let α′ be as in (1). We have vp (b′ ) = k − 1 ≥ 1.
′
If k = 2, then 1 = k − 1 < n−1
2 , and hence 2 ∈ L(α ) (by Claim B). Therefore
′
{ 3 } = [3, k + 1] ⊂ { 1 } + L(α ) ⊂ L(α).
If k ≥ 3, then by induction hypothesis, [3, k] ⊂ L(α′ ), and thus [4, k + 1] =
{ 1 } + L(α′ ) ⊂ L(α), and by Claim D, also 3 ∈ L(α).
Claim G: If k ≥ n − 1, then [3, n − 2] ⊂ L(α).
Proof. If n ≤ 4, then the claim is trivially true. We again proceed by induction on
n. Suppose n ≥ 5 (then k ≥ 4), and that the claim is true for n − 1.
Let b′ = b/p and let α′ be as in (1). Again, vp (b′ ) = k − 1 ≥ 3 and by induction
hypothesis [3, n − 3] ⊂ L(α′ ). Therefore [4, n − 2] ⊂ L(α) and by Claim D also
3 ∈ L(α).
If k ≥ n − 1, then the claim of the theorem follows from claims A, B, C, E and
G. If k ∈ [2, n − 2], then the claim of the theorem follows from claims A, B and
F.
If α ∈ R(D) is a nonzero nonunit, and L(α) = { l1 , l2 , . . . , lk }, then theset
of distances of α is defined as ∆(α) = { li − li−1 | i ∈ [2, k] }, and ∆ R(D) =
S
∆(α). For k ∈ N≥2 , set Uk R(D) = S
α∈R(D),k∈L(α) L(α).
α∈R(D)\ { 0 }∪U(R(D))
Corollary 15. If D is a PID, but not a field, then U2 R(D) = N≥2 and ∆ R(D) =
{ 1, 2 }.
Proof. This follows directly from Theorem 14.
FACTORIZATION IN THE SELF-IDEALIZATION OF A PID
11
Corollary 16. Suppose D is a PID that has infinitely many pairwise non-associated
prime elements. Then
n
o
n
o
L(R(D)) = { 0 }, { 1 } ∪ [m, n] | m ∈ [2, n]
n
o
∪ [m, n] ∪ { n + 2 } | m ∈ [2, n] and n even
n
o
∪ [m, n] ∪ { n + 2 } | m ∈ [3, n] and n odd
n
∪ m + 2[0, n] | with m ∈ N≥2n and n ∈ N }.
Proof. The sets { 0 } and { 1 } correspond to units and irreducibles. For zerodivisors, the sets of lengths are discrete intervals and completely described in Theorem 14(1). By Corollary 4 and Lemma 13(2), the sets of lengths of nonunit
non-zero-divisors are arbitrary sumsets of sets as in Theorem 14(2), i.e., of sets of
the form { 1 }, [2, n] (for n ≥ 2), [3, n] (for n ≥ 3), [2, n] ∪ { n + 2 } for even n ≥ 2,
and [3, n] ∪ { n + 2 } for odd n ≥ 3.
Finally, we remark that other important invariants of factorization theory (their
definitions readily generalize to the zero-divisor case) are easily determined for
R(D) using the characterization of sets of lengths and Corollary 4.
Corollary 17. Let D be
ring with
a PID but not a field. R(D)
is a locally tame
catenary degree c R(D) = 4. In particular, ∆ R(D) = [1, c R(D) − 2].
Proof. We first observe that the catenary degree (see [8, Chapter 1.6] for the definition in the non-zero-divisor case) of R(D) is 4: Let first α ∈ R(D) with nr(α) 6= 0.
Using Corollary 4, we can reduce to the case where nr(α) is a prime power. Since
then min L(α) ≤ 3, we can argue as in bifurcus semigroups (cf. [1, Theorem 1.1]),
to find c(α) ≤ 4. In view of Lemma 8(1), and with a similar argument, the catenary
degree of a zero-divisor is at most 2. Together this gives c(R(D)) ≤ 4. Since there
exists an element with set of lengths { 2, 4 }, also c(R(D)) ≥ 4.
We still have to show that R(D) is locally tame (see [8, Chapter 1.6] or [10] for
definitions). For this we have to show t(R(D), γ) < ∞ for all irreducible γ ∈ R(D).
Let α ∈ R(D) and γ ∈ R(D) be irreducible. If γ is prime, then t(R(D), γ) = 0,
hence we may suppose that γ is associated to one of the non-prime irreducibles
from Theorem 5, and hence there exist a prime element p ∈ D and n ∈ N such
that nr(γ) = pn . If α ∈ R(D) is a zero-divisor, then t(α, γ) = n follows easily from
Lemma 8(1).
A standard technique allows us to show t(R(D)• , γ) < ∞: By [10, Proposition
3.8], it suffices to show that two auxiliary invariants, ω(R(D)• , γ) and τ (R(D)• , γ)
are finite.
Suppose I ⊂ (R(D)• , ·) is a divisorial ideal. If we denote by R(D) hIi the ideal
of R(D) generated by I, one checks that R(D) hIi ∩ R(D)• = I. Since R(D) is
noetherian, R(D)• is therefore v-noetherian. By [10, Theorem 4.2], ω(R(D)• , γ) is
finite.
Recalling the definition of τ (α, γ) (from [10, Definition 3.1]), it is immediate from
Theorem 14 together with Corollary 4, that τ (R(D)• , γ) ≤ 3. Altogether, therefore
t(R(D), γ) < ∞.
12
G.W. CHANG AND D. SMERTNIG
Remark 18. Suppose D is a PID but not a field.
(1) Trivially, Theorem 14(2) holds true for R(D)• .
(2) Let K be the quotient field of D, and H = R(D)• . We have
n a b
o
| b ∈ D, a ∈ D• ,
H=
0 a
and the complete integral closure of H is equal to
o
n a b
b=
H
| b ∈ K, a ∈ D•
0 a
n n
a b
a
nan−1 b
because
for all a, b ∈ K and n ∈ N. This shows
=
0 a
0
an
b and even more we have f = (H : H)
b = ∅ (note that (D : K) = ∅).
H 6= H,
Thus the monoid under discussion is neither a Krull nor a C-monoid, which
have been extensively studied in recent literature (see [8, Chapters 2.9, 3.3,
and 4.6], [9], [12]).
Acknowledgements. The first author’s work was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded
by the Ministry of Education, Science and Technology (2010-0007069). The second
author was supported by the Austrian Science Fund (FWF): W1230. We thank
the anonymous referee for his/her careful reading and all the helpful comments, in
particular on the concept of self-idealization.
References
[1] D. Adams, R. Ardila, D. Hannasch, A. Kosh, H. McCarthy, V. Ponomarenko, and R. Rosenbaum, Bifurcus semigroups and rings, Involve 2 (2009) 351–356.
[2] D.D. Anderson and S. Valdes-Leon, Factorization in commutative rings with zero divisors,
Rocky Mountain J. Math. 26 (1996) 439–480.
[3] D.D. Anderson (ed.), Factorization in Integral Domains, Lect. Notes Pure Appl. Math., vol.
189, Marcel Dekker, 1997.
[4] D.D. Anderson and M. Winders, Idealization of a Module, J. Commutative Algebra 1 (2009)
3–56.
[5] N.R. Baeth and R.A. Wiegand, Factorization theory and decomposition of modules, Am.
Math. Mon. 120 (2013), 3–34.
[6] S.T. Chapman (ed.), Arithmetical Properties of Commutative Rings and Monoids, Lect.
Notes Pure Appl. Math., vol. 241, Chapman & Hall/CRC, 2005.
[7] C. Frei and S. Frisch, Non-unique factorization of polynomials over residue class rings of the
integers, Comm. Algebra 39 (2011) 1482–1490.
[8] A. Geroldinger and F. Halter-Koch, Non-unique factorizations, vol. 278 of Pure and Applied Mathematics (Boca Raton), Chapman & Hall/CRC, Boca Raton, FL, 2006. Algebraic,
combinatorial and analytic theory.
[9] A. Geroldinger and W. Hassler, Arithmetic of Mori domains and monoids, J. Algebra 329
(2008) 3419–3463.
[10] A. Geroldinger and W. Hassler, Local tameness of v-noetherian monoids, J. Pure Appl.
Algebra 212 (2008) 1509–1524.
[11] J. Huckaba, Commutative rings with zero divisors, Dekker, New York, 1988.
[12] A. Reinhart, On integral domains that are C-monoids, Houston J. Math., to appear.
[13] L. Salce, Transfinite self-idealization and commutative rings of triangular matrices, in Commutative algebra and its applications, 333–347, Walter de Gruyter, Berlin, 2009.
FACTORIZATION IN THE SELF-IDEALIZATION OF A PID
13
(Chang) Department of Mathematics, University of Incheon, Incheon 406-772, Korea.
E-mail address: [email protected]
(Smertnig) Institut für Mathematik und Wissenschaftliches Rechnen, Karl-FranzensUniversität Graz, Heinrichstraße 36, 8010 Graz, Austria
E-mail address: [email protected]
| 0math.AC
|
arXiv:1705.02809v2 [math.GR] 1 Feb 2018
APPLICATIONS OF L SYSTEMS TO GROUP THEORY
LAURA CIOBANU, MURRAY ELDER, AND MICHAL FEROV
Abstract. L systems generalise context-free grammars by incorporating parallel rewriting, and generate languages such as EDT0L and ET0L
that are strictly contained in the class of indexed languages. In this paper we show that many of the languages naturally appearing in group
theory, and that were known to be indexed or context-sensitive, are in
fact ET0L and in many cases EDT0L. For instance, the language of
primitives and bases in the free group on two generators, the BridsonGilman normal forms for the fundamental groups of 3-manifolds or orbifolds, and the co-word problem of Grigorchuk’s group can be generated
by L systems. To complement the result on primitives in rank 2 free
groups, we show that the language of primitives, and primitive sets, in
free groups of rank higher than two is context-sensitive. We also show
the existence of EDT0L and ET0L languages of intermediate growth.
1. Introduction
In this paper we show that many of the context-sensitive or indexed languages arising in problems in group theory and combinatorics are in fact
ET0L or even EDT0L. The merit of giving this new formal language characterisation is that ET0L and EDT0L languages are a strict subclass of
the indexed ones, and also, the descriptions of these languages are simpler
and more algebraic than those based on indexed grammars or nested-stack
automata that were given in the literature for the sets considered here.
Both EDT0L and ET0L belong to the languages generated by L systems, which were introduced by Lindenmayer in the late 1960s in order to
model the growth of various organisms. The acronym ET0L (respectively
EDT0L) refers to Extended, Table, 0 interaction, and Lindenmayer (respectively Deterministic). ET0L and EDT0L languages have only recently
featured in group theory; their first prominent appearance was in the work
of the first two authors with Volker Diekert, in the context of equations in
groups. The language of solutions of equations in free groups as tuples of
reduced words is EDT0L [5], as are solutions in appropriate normal forms in
virtually free groups [9] and partially commutative groups [10]. The result
in [5] implies that the pattern languages studied by Jain, Miasnikov and
Stephan are EDT0L as well [19].
Date: February 5, 2018.
2010 Mathematics Subject Classification. 20F10; 20F65; 68Q42.
Key words and phrases. Free group; primitive; normal form; co-word problem; Grigorchuk
group; indexed language; ET0L language; EDT0L language.
1
In Section 2 we give the necessary background on formal languages. In
Section 3 we examine the gap between ET0L and indexed languages, and
in Section 4 we revisit an example of Grigorchuk and Machı̀ of a language
of intermediate growth and show that it is EDT0L. In the remaining sections we show why Lindenmeyer languages are relevant for group theory by
presenting several instances where L systems appear naturally. In Section
5 we describe the set of primitives in the free group on two generators as
EDT0L, improving on the context-sensitive characterisation by Silva and
Weil [27]. Furthermore, we show that the set of primitives, and primitive
sets, in free groups of rank higher than 2 is context-sensitive. In Section
6 we construct an explicit grammar to prove that the co-word problem in
Grigorchuks group is ET0L, and thus take a different approach to that of
Holt and Röver [18], who used nested-stack automata to describe the same
set. In Section 7 we show that the Bridson-Gilman normal forms for the
fundamental groups of 3-manifolds or orbifolds, proved in [2] to be indexed,
are in fact ET0L. We conclude the paper with a list of open problems.
2. ET0L and EDT0L languages
L systems were introduced by Lindenmayer in order to model the growth
of various organisms and capture the fact that growth happens in parallel
everywhere in the organism. Therefore the rewriting system had to incorporate parallelism, as opposed to the sequential behavior of context-free
grammars. The difference between sequential and parallel grammars is well
illustrated by the following example (page 2 in [26]).
Example 1. Suppose we have an alphabet A = {a} and a single rewriting
rule a −→ a2 , which is to be applied to a3 . If we apply this rule to one
a inside a3 at a time, we get the set {ai | i ≥ 3}. If we apply the rule
simultaneously to each a in a3 we obtain a6 after one rewrite, and the set
i
of words obtained via parallel rewriting is {a3·2 | i ≥ 0}.
There is a vast literature on Lindenmayer systems, see for example [26, 25],
with various acronyms such as D0L, DT0L, ET0L, HDT0L and so forth. The
following inclusions hold: EDT0L ( ET0L ( indexed, and context-free (
ET0L. Furthermore, the classes of EDT0L and context-free languages are
incomparable.
Let V be a finite alphabet. A table for V is a finite subset of V × V ∗ ,
that can be represented as in Figure 1.
V
V∗
a −→ a, b, aba
b −→ bbabb
# −→
#, b
Figure 1. A table for V = {a, b, #}.
2
If (c, v) is in some table t, we say that (c, v) is a rule for c and use the
convention that if for some c ∈ V no rule for c is specified in t, then t
contains the rule (c, c). We express the rewriting corresponding to the rule
(c, v) in t as c −→t v.
Definition 1 (ET0L). An ET0L-system is a tuple H = (V, A, T, I), where
(1) V is a finite alphabet,
(2) A ⊆ V is the subset of terminal symbols,
(3) T is a finite set of tables for V , that is, each t ∈ T is a finite subset
of V × V ∗ , and
(4) I ⊆ V ∗ is a finite set of words called axioms.
Let t ∈ T . We will write u −→t v to denote that a word v ∈ V ∗ can
be produced from u ∈ V ∗ using the rules in t; that is, if u = c1 · · · cm and
v = v1 · · · vm , where ci ∈ V and vi ∈ V ∗ , we write u −→t v to mean v was
obtained via rules cj −→t vj from table t, applied to each cj appearing in
u. More generally, u −→ v signifies u −→t v for some t ∈ T . If there exist
u0 , . . . , uk ∈ V ∗ with ui −→ ui+1 for 0 ≤ i ≤ k−1, then we write u0 −→∗ uk .
The language generated by H is defined as
L(H) = {v ∈ A∗ | w −→∗ v for some w ∈ I}.
A language is ET0L if it is equal to L(H) for some ET0L system H.
Definition 2 (EDT0L). An EDT0L-system is an ET0L system where in
each table there is exactly one rule for each letter in V . A language is
EDT0L if it is equal to L(H) for some EDT0L system H.
ET0L languages form a full AFL (abstract family of languages), that is,
they are closed under homomorphisms, inverse homomorphisms, intersection with regular languages, union, concatenation and Kleene closure, while
EDT0L are closed under all of the above except inverse homomorphism so
do not form a full AFL [7, 1].
If for some words u1 , u2 , u3 and tables t1 , t2 we have u1 −→t1 u2 and
u2 −→t2 u3 we will write u1 −→t1 t2 u3 to denote the composition of the
rewriting. This can be naturally extended to any finite sequence of rewrites.
Given a regular expression R over a finite set T = {t1 , . . . tn }, where t1 , . . . , tn
are tables, we will sometimes abuse the notation and use u1 −→R u2 to denote that there is a word r in the language generated by the regular expression R such that u1 −→r u2 . Furthermore, if the system is deterministic,
then every table is in fact a homomorphism on the free monoid V ∗ , and using this more algebraic notation we can give Asveld’s equivalent definition
for E(D)T0L languages as follows [1].
Definition 3. Let A be an alphabet and L ⊆ A∗ . We say that L is an
ET0L language if there is an alphabet C with A ⊆ C, a set H of tables (i.e.
finite subsets of C × C ∗ ), a regular language R ⊆ H ∗ and a letter c ∈ C
such that
L = {w ∈ A∗ | c −→r w for some r ∈ R}.
3
In the case when every table h ∈ H is deterministic, i.e. each h ∈ H is in
fact a homomorphism, we write r(c) = w and say that L is EDT0L.
The set R is called the regular (or rational) control, the symbol c the start
symbol and C the extended alphabet.
Convention. In any description of rational control in this paper, the maps
are always applied left to right, but in algebraic settings where f and g are
morphisms f g(a) := f (g(a)).
3. Non-ET0L languages
For completeness we present in this section a short survey of examples
of languages which are not ET0L. The first examples turn out not to be
indexed either.
Let ϕ : N+ → N+ be such that limn→∞ ϕ(n) = ∞ and let U be an arbitrary infinite subset of N. Set K(ϕ, U ) = {(baϕ(k) )k | k ∈ U }. This
construction gives us an infinite family of languages. It was proved in [11,
Theorem 2] that K(ϕ, U ) is not ET0L regardless of the choice of ϕ and U .
We show in Lemma 5 that these languages are not indexed.
Lemma 4. [15, Theorem A] (Shrinking lemma) Let L be an indexed language over a finite alphabet Σ and let m > 0 be a given integer. There is a
constant k > 0 such that each word w ∈ L with |w| ≥ k can be factorised as
a product w = w1 . . . wr such that the following conditions hold:
(i) m < r ≤ k,
(ii) wi 6= ǫ for every i ∈ {1, . . . , r},
(iii) each choice of m factors is included in a proper subproduct which
lies in L.
Lemma 5. The language K(ϕ, U ) is not indexed for any choice of ϕ and
U.
Proof. The proof follows that of [15, Corollary 4]. Assume that K(ϕ, U ) is
indexed and fix m = 1. Let k ∈ N be given by Lemma 4.
Pick k′ ∈ U such that k′ > k and k′ = min(ϕ−1 (ϕ(k′ ))). Note that such
′
′
′
k exists as limn→∞ ϕ(n) = +∞. Then w = (baϕ(k ) )k can be factorised as
w = w1 . . . wr , where 1 < r ≤ k < k ′ . As r < |w|b = k′ , by the pigeonhole
principle we see that there is some 1 ≤ i ≤ r such that |wi |b ≥ 2, i.e.
′
wi contains baϕ(k ) b as a subword. By Lemma 4, wi can be included in a
′
proper subproduct v which lies in K(ϕ, U ). However, v contains baϕ(k ) b, so
′′
′′
v = (baϕ(k ) )k for some k′′ ∈ N such that ϕ(k′ ) = ϕ(k′′ ). As v is a proper
subproduct of w we see that k′′ < k ′ . However, this is a contradiction with
k′ = min(ϕ−1 (ϕ(k′ ))), thus K(ϕ, U ) is not indexed.
Note that when ϕ is the identity function and U = N this was already
established in [17, Theorem 5.3] and later in [15, Corollary 4].
An explicit example of an indexed language which is not ET0L was given
in [14] as the language of tree-cuts. Informally speaking, a tree-cut is a
4
sequence of binary strings encoding the list of leaves of a proper rooted
binary tree, i.e. a tree in which every vertex has exactly zero or two children.
Formally, tree cuts can be defined in a recursive manner:
(i) the sequence containing only the empty string, i.e. (ǫ), is a tree-cut;
(ii) suppose that u1 , . . . , uk , v1 , . . . , vl ∈ {0, 1}∗ such that the sequences
(u1 , . . . , uk ) and (v1 , . . . , vl ) are tree-cuts, then the sequence
(0u1 , . . . , 0uk , 1v1 , . . . , 1vl )
is a tree-cut.
A sequence (v1 , . . . , vk ), where v1 , . . . , vk ∈ {0, 1} is a tree-cut if and only if
it can be obtained by repeatedly applying to the above mentioned rules. For
example, the sequence (000, 001, 01, 10, 11) is a tree-cut, but the sequence
(000, 111) is not. For more detail on tree-cuts see for example [14, Section
3].
Let a, b denote symbols distinct from 0 and 1. The language of cuts is
then defined as
L0 = {av1 0bv1 1 . . . avk 0bvk 1 | (v1 , . . . , vk ) is a cut} ⊆ {0, 1, a, b}∗ .
Carefully checking the proof of [14, Lemma 3.3] one can verify that the language L0 is accepted by a nested stack automaton and hence L0 is indexed.
It is then proved in [14, Lemma 3.4] that L0 is not ET0L.
Another example of an indexed language that is not ET0L is given in [13,
Corollary 2]. In fact, the paper suggests an infinite family of such languages.
However, the statement of [13, Theorem 3] contains a misprint, so we give
the correct statement here. For a word w ∈ Σ∗ let w denote its mirror
image, i.e. if w = x1 . . . xn , where x1 , . . . , xn ∈ Σ, then w = xn . . . x1 .
Theorem 6. (see [13, Theorem 3]) Let Σ be a finite alphabet and let Σ′ be
a copy of Σ distinct from Σ. Let h : Σ∗ → Σ′∗ be a homomorphism defined
by h(x) = x′ for every x ∈ Σ.
Let K be a context-free language over Σ such that K is not EDT0L. Then
the language
MK = {kh(k) | k ∈ K} ⊆ (Σ ∪ Σ′ )∗
is indexed but not ET0L.
It is a well known fact (see [12, Theorem 9]) that if K is the Dyck language
on at least 8 letters, then K is context-free but not EDT0L, hence MK is
indexed but not ET0L. In fact we observe below in Proposition 26 that the
word problem for free groups of rank at least 2 is not EDT0L.
4. Growth of languages
The growth of a language L ⊆ Σ∗ is the function f : N → N such that
f (n) is the number of words in L of length n. If Σ is finite then the growth
function is at most exponential. A language has intermediate growth if for
any α, β > 1 there is an integer N ∈ N such that for all n > N we have
5
nα < f (n) < β n . Bridson and Gilman showed that there are no contextfree languages of intermediate growth [3], whereas Grigorchuk and Machı̀
[16, Theorem 1] give the following example of a language of intermediate
growth which is recognisable by a one-way deterministic non-erasing stack
automaton (1DNESE), so is indexed. They define the set
A := {abi1 abi2 . . . abik | 0 ≤ i1 ≤ · · · ≤ ik , k ∈ N}
over the alphabet {a, b}.
Proposition 7. The language A is EDT0L.
Proof. We show that A is generated by an EDT0L system as in Definition
3, with the following data: the extended alphabet is {a, b, q, q ′ }, the start
symbol q, and the maps are ha , hb , h$ with ha (q) = qaq ′ , hb (q) = qb,
hb (q ′ ) = q ′ b, h$ (q) = h$ (q ′ ) = ǫ.
The rational control is given by R = {ha , hb }∗ ha h$ .
Let g ∈ {ha , hb }∗ . We can easily prove by induction on the length of g as a
word over {ha , hb } that g(q) = qbi1 aq ′ bi2 a . . . aq ′ bik where 0 ≤ i1 ≤ · · · ≤ ik ,
i.e the word starts with q, and then only q ′ appears. Then applying ha h$ to
g(q) produces a word which starts with a and contains no more q, q ′ .
Corollary 8. There exist EDT0L (and ET0L) languages of intermediate
growth.
5. Primitives and bases in free groups
Given an alphabet Σ, we will always assume that a tuple (w1 , . . . , wt ),
where w1 , . . . , wn ∈ Σ∗ , is encoded as a string w1 # . . . #w2 , where the symbol # is distinct from the ones contained in Σ.
A free basis of a free group is a tuple of elements that freely generate the
group, and a primitive (element) is an element that belongs to some free
basis of the group. In this section we show that the set of bases and the
set of primitives, written as reduced words over the standard free basis, is
EDT0L for free groups of rank 2, and context-sensitive for higher rank. It is
an open question whether the higher rank context-sensitive characterisation
can be lowered to indexed or E(D)T0L.
5.1. Bases and primitives in the free group F2 are EDT0L. In [27]
Silva and Weil showed that the set of primitives in the free group F2 on
generators {a, b}, written as reduced words over {a, b, a−1 , b−1 }, is contextsensitive and not context-free. In this section we show that both the set of
bases and the set of primitives are EDT0L.
Proposition 9. The set of free bases and the set of primitives in F2 , as
reduced words, are EDT0L.
Proof. It is a classical result due to Nielsen that two elements g, h ∈ F2
form a basis of F2 if and only if the commutator [g, h] is conjugate either to
6
[a, b] or [b, a]. Hence the set of bases in F2 can be seen as the values X, Y
satisfying the equations
Z[X, Y ]Z −1 = [a, b]
or
Z[X, Y ]Z −1 = [b, a]
where Z can take any value.
By Corollary 2.2 in [5], for any equation in a free group the set of solutions,
or any projections thereof, written as tuples in reduced words, is EDT0L.
In this case the values of (X, Y ) will provide the set of bases, and the values
of X the set of primitives.
5.2. Primitives in free groups of higher rank. While in F2 the set of
free bases can be expressed in terms of solutions to equations, this does
not hold in higher rank. It was shown in [21] that the set of primitives in
free groups of higher rank is not definable in the first order theory of the
group, and thus we cannot use our previous approach to give an EDT0L
characterisation in this case. Another approach is to produce the set of
primitives by applying all the maps in Aut(Fk ) to a free basis element.
Since Aut(Fk ) is finitely generated, we would apply all the maps in Aut(Fk ),
written as words over the generators of Aut(Fk ) - which can be seen as
rational control, to a basis element which would play the role of the axiom.
The set thus obtained would be EDT0L and contain all the primitives, but
neither as reduced nor unique words.
Since we are interested in establishing a formal language characterisation for primitives as reduced words, we use a different approach, based on
Stallings’ folding of labeled graphs, and an algorithm for testing primitivity
in a free group given in [6], to show here that they are context-sensitive. For
the sake of keeping this note succinct, we do not define here the Stallings
graph or what is meant by folding, but refer the reader to references such as
[28] and [20]. If Γ is the Stallings graph of a finitely generated subgroup of
a free group F , a pinch will signify the identification of two distinct vertices
of Γ.
A primitive set (as opposed to the set of primitives) in a free group F is
a set of elements that can be extended to a basis of F .
Lemma 10. [6, Theorem 4.4] Let n ≤ k and let W = {w1 , . . . , wn } be a
set of reduced words in Fk = F (X), where X = {x1 , . . . , xk }, such that for
every x ∈ X either x or x−1 appears in some w ∈ W . Let Γ be a graph
representing the subgroup of Fk generated by W . Then W is a primitive set
if and only if there is a sequence of pinches and folds, containing exactly
k − n pinches, that transforms Γ into the elementary wedge on X.
In the case that not every generator or inverse appears in the list of wi
then the Lemma 10 can still decide primitivity by taking a smaller value for
k.
7
A standard way to represent a Stallings graph Γ = (V Γ, EΓ) over an
alphabet X would be as a collection of labeled oriented edges, where an
edge γ ∈ EΓ is given as a triple (i(γ), t(γ), l(γ)), with i(γ) ∈ V Γ the initial
vertex of γ, t(γ) ∈ V Γ the terminal vertex and l(γ) ∈ X ∪ X −1 the label
of γ. Obviously, a word w ∈ F (X) can be represented by a graph Γw such
that |V Γw |, |EΓw | ≤ n, where n = |w|. However, as there might be up to
n vertices, and one needs log(n) space to encode the names of the vertices,
the naı̈ve implementation of Γ would require n log(n) space.
We propose a different representation of the Stallings graph, along with
two additional types of moves, and show that Lemma 10 can be realised by
a linear space algorithm. Instead of labelling an edge by a single symbol, we
will label edges by subwords, that is, an edge γ ∈ EΓ will be represented by
a triple (i(γ), t(γ), s(γ)), where i(γ) and t(γ) are as before and s(γ) ∈ F (X)
is a reduced word. In this case we say that Γ is a segment graph. Adapting
the terminology of Stallings graphs, we say that a segment graph Γ is folded
if it does not contain a pair of incident edges whose labels share a common
prefix or suffix. If a vertex v is incident to two edges with a common prefix
or suffix, then we say that v is unfolded, and otherwise call it folded.
Folding at a vertex v applies to segment graphs in three possible ways,
depending on whether two segments incident to v have exactly the same
label, or only common proper prefixes. More precisely, let γ1 , γ2 ∈ EΓ satisfy
i(γ1 ) = v = i(γ2 ), s(γ1 ) = uw1 and s(γ2 ) = uw2 for some u, w1 , w2 ∈ F (X).
I. If w1 = w2 = ǫ, then this is a usual folding of graphs.
II. If w1 = ǫ and w2 6= ǫ, then we split γ2 into γ2′ and γ2′′ by introducing
a new vertex v ′ on γ2 such that i(γ2′ ) = v, t(γ2′ ) = v ′ and s(γ2′ ) = u,
and fold γ1 with γ2′ as in case I.
III. If w1 , w2 6= ǫ, then we introduce a new vertex v ′ and a new edge
γ, such that i(γ) = v, t(γ) = v ′ , s(γ) = u, and replace edges γ1 , γ2
by new edges γ1′ , γ2′ such that i(γ1′ ) = v ′ = i(γ2′ ), t(γ1′ ) = t(γ1 ),
t(γ2′ ) = t(γ2 ), s(γ1′ ) = w1 and s(γ2′ ) = w2 .
Pinching a segment graph is defined as follows. Let γ, γ ′ ∈ EΓ satisfy
s(γ) = w1 w2 and s(γ ′ ) = w1′ w2′ for some w1 , w2 , w1′ , w2′ ∈ F (X). We remove
the edges γ, γ ′ , introduce new vertex v and up to four new edges γ1 , γ2 , γ1′ , γ2′
given by
γ1 :
γ2 :
γ1′ :
γ2′ :
i(γ1 ) = i(γ), t(γ1 ) = v, s(γ1 ) = w1 ;
i(γ2 ) = v, t(γ2 ) = t(γ), s(γ2 ) = w2 ;
i(γ1′ ) = i(γ ′ ), t(γ1′ ) = v, s(γ1′ ) = w1′ ;
i(γ2′ ) = v, t(γ2′ ) = t(γ ′ ), s(γ2 ) = w2′ .
If one of w1 , w2 , w1′ , w2′ is empty, i.e. when we are pinching a segment with
a vertex, we don’t need to introduce the vertex v.
Both foldings of type III and pinches can cause the numbers of edges and
vertices to increase; the resulting graph could potentially not fit on a linear
size tape, but the discussion below shows that this is not the case.
8
Recall that if Γ is a Stallings graph corresponding to some finite set W
then the core of Γ (or the core graph for hW i), is the smallest subgraph of Γ
in which we can read all the elements of hW i as reduced words along loops
based at the base vertex of Γ. In particular, if Γ is finite then the core of Γ
does not contain any vertices of degree one, with the exception of the base
vertex of Γ. The core of a segment graph is defined in an analogous manner.
Remark 1. (1) A set W = {w1 , . . . , wn } ⊆ Fk is primitive if and only if
the set W g = {gw1 g−1 , . . . , gwn g−1 } is primitive for some (and thus for all)
g ∈ Fk . As choosing a different base vertex in the core graph of hW i is
the same as taking a conjugate of hW i, we will assume that the base vertex
in any segment graph has degree at least 2. In fact, unless the graph is
homeomorphic to a loop, we may assume that the base vertex is of degree
at least three.
(2) Furthermore, as the core graph will never contain any leaves other
than the base vertex and as was already mentioned, we can always assume
that the base vertex is of degree at least two, hence without loss of generality
we may assume that the segment graph does not contain any leaves; this
can be achieved by pruning, i.e. repeatedly removing vertices of degree 1
(along with corresponding edges). Note that for finite graphs this procedure
always terminates.
We say that a segment graph Γ is topological if all its vertices have degree
≥ 3, unless it consists of a single vertex, or a loop attached at a degree
two vertex. Following Remark 1, we see that every segment graph can be
transformed into a topological one (representing a subgroup conjugate to
the original) by picking a different base vertex, pruning and merging edges:
suppose γ1 , γ2 ∈ EΓ such that t(γ1 ) = v = i(γ2 ) and deg(v) = 2; then we can
replace the edges γ1 , γ2 by a new edge γ such that i(γ) = i(γ1 ), t(γ) = t(γ2 )
and s(γ) = s(γ1 )s(γ2 ), and remove vertex v.
The following lemma implies that pruning will need to be done at most
once in the entire process corresponding to Lemma 10.
Lemma 11. Let Γ be a folded graph that does not contain any vertex of
degree 1 and let Γ′ be a graph obtained from Γ by performing a pinch and a
folding. Then Γ′ does not contain vertices of degree 1.
Proof. Suppose that there is v ′ ∈ V Γ′ such that deg(v ′ ) = 1. Without loss of
generality we may assume that v ′ is the initial vertex of its unique adjacent
edge γ ∈ EΓ′ . Note that the pinch and the subsequent folding uniquely
define a surjective morphism of graphs f : V Γ0 → V Γ′ , where Γ can be
obtained from Γ0 by merging. Note that Γ0 is folded and it does not contain
vertices of degree 1. Let v ∈ f −1 (v ′ ) be arbitrary. Again, without loss of
generality we may assume that v is the initial vertex of all of its adjacent
edges. It follows that all the edges adjacent to v must share a common
prefix with γ, meaning that v is not folded folded unless deg(v) = 1. As Γ0
is folded, deg(v) = 1 for every v ∈ f −1 (v ′ ), which contradicts the fact that
9
deg(u) > 1 for every u ∈ V Γ0 . Thus Γ′ does not contain vertices of degree
1.
The topological rank of a graph is the rank of its fundamental group.
Lemma 12. Let Γ be a finite topological graph of topological rank r with
r > 1. Then |EΓ| ≤ 3r − 3 and |V Γ| ≤ 2r − 2.
Proof. As r > 1 we see that deg(v) ≥ 3 for every v ∈ V Γ. From the Euler
characteristic
formula r = |EΓ| − |V Γ| + 1, and since in any graph 2|EΓ| =
P
deg(v)
by assumption 2|EΓ| ≥ 3|V Γ|, which gives the inequalities
v∈V Γ
above.
Remark 2. Folding does not cause the topological rank of a graph to increase, but pinching can increase it by 1.
Corollary 13. Let Γ be a finite topological segment graph of topological
rank r with a single unfolded vertex, and let Γ′ be the folded topological
graph obtained from Γ by folding, merging and pruning.
If r = 1 then |V Γ′ | = |EΓ′ | = 1, otherwise |V Γ′ | ≤ 2r − 2 and |EΓ′ | ≤
3r − 3.
Proof. Let r ′ denote the topological rank of the graph Γ′ ; note that r ′ ≤ r
by Remark 2.
If r ′ = 1 then, following the definition of topological graphs, we see that
Γ′ is a single vertex with an attached loop and |V Γ′ | = |EΓ′ | = 1.
If r ′ > 1 then the result follows by Lemma 12.
Proposition 14. Let Fk be a free group of rank k ≥ 2. Then the primitive
sets in Fk can be recognised in linear space.
Proof. We represent a graph with t edges on the tape of a Turing machine
by the string
#vi1 |vi2 |w1 ## . . . ##vi2t−1 |vi2t |wt #,
where vij are binary numbers and the factor #v|v ′ |w# represents the segment γ with i(γ) = v, t(γ) = v ′ and s(γ) = w.
On input a list of words w1 , w2 , . . . , wn , if n ≥ k we return NO, else we
write on the tape
#v0 |v0 |w1 ##v0 |v0 |w2 ## . . . ##v0 |v0 |wn #
which describes the bouquet of loops labeled wi at a vertex v0 . Let k′ be
the rank of the free group generated by the letters appearing in the wi , so
k′ ≤ k.
We perform folds of types I, II, III by modifying the tape as follows:
I. Scan the tape to find factors #vi |vj |u# and #vi |vk |u#; erase the
second factor and replace vk by vj everywhere on the tape.
II. Scan the tape to find #vi |vj |u# and #vi |vk |up#; overwrite the second factor with #vj |vk |p#.
10
III. Scan the tape to find #vi |vj |uap# and #vi |vk |ubq# with a 6= b, a, b ∈
X ±1 ; erase both factors and write #vi |v|u##v|vj |ap##v|vk |bq#
where the binary number v is some value not already in use.
For pruning we scan the tape to find a factor #v|v ′ |u# such that the v
(or v ′ ) does not appear in any other factor. If such a factor is found, we
delete it.
Similarly, for merging we scan the tape for a pair of factors #v|v ′ |u1 # and
#v ′ |v ′′ |u2 # (or #v ′′ |v ′ |u2 #) such that v ′′ does not occur in any other factor.
If such a pair is found, we replace them by #v|v ′′ |u1 u2 # (or #v|v ′′ |u1 u−1
2 #,
respectively).
We perform a pinch by choosing either:
(1) two distinct vertices vi , vj and replacing vj by vi everywhere on the
tape;
(2) a vertex v and a segment (vi , vj , w1 w2 ) with |wi | > 0 and replacing
#vi |vj |w1 w2 # by #vi |v|w1 ##v|vj |w2 #;
(3) two segments (vi , vj , w1 w2 ), (vp , vq , w3 w4 ) with |wi | > 0, and replacing their encodings by #vi |v|w1 ##v|vj |w2 ##vp |v|w3 ##v|vq |w4 #
where v is some value not already in use.
Note that folding and pruning moves decrease the number of letters appearing as labels of segments, and pinching and merging moves preserve this
number.
The procedure starts by performing folding moves I, II, III in any order exhaustively, and each folding is followed by all possible pruning and
merging. Then a pinch is applied, and the previous two steps repeat k′ − n
times.
Following Lemma 11 we see that we will only need to perform pruning
moves before the first pinch. Note that during this process, the number of
letters present in the labels might decrease. In this case we need to decrease
k′ accordingly.
Return YES if the tape contains
#v|v|ai1 ## . . . ##v|v|aik′ −n #
with all ij distinct and aij ∈ X ±1 , else return NO.
Termination of the algorithm is guaranteed since each fold strictly decreases the number of letters from X ±1 appearing on the tape as labels
of segments. The algorithm accepts precisely the primitive sets in Fk by
Remark 1 and Lemma 10.
Since the rank of all topological segment graphs considered is smaller
than the rank of the ambient free group, by Corollary 13 the number of
vertices in use at any time is at most 2k − 2, so in binary notation each
vertex requires at most log(2k − 2) space, the total number of segments is
at most 3k P
− 3, and the number of letters from X ±1 appearing on the tape
is at most ni=1 |wi | = N . Thus the amount of space required at any time
11
in the process is at most
(3k − 3)(2 log(2k − 2) + 4) +
n
X
|wi |,
i=1
which is linear in the input size.
From the proposition it follows that for every k there is a linearly bounded
Turing machine Tk which recognises primitive sets in Fk , hence we can state
the following corollary.
Corollary 15. Let Fk = F (X) be a free group over X, where |X| = k and
n ≤ k. Then
Pk,n = {w1 # . . . #wn | {w1 , . . . wn } is a primitive set in Fk }
∗
⊆ X ∪ X −1 ∪ {#}
is a context-sensitive language.
In particular, the set Pk (=Pk,1 ) of primitive elements is context-sensitive.
6. Co-word problem for Grigorchuk group is ET0L
In this section we show that the co-word problem for the Grigorchuk group
is ET0L, improving on Holt and Röver’s result in [18], where they showed
it is indexed. It is still an open question whether the co-word problem for
the Grigorchuk’s group is context-free.
6.1. Generators and the word problem. We refer the reader to [8] for
more details, here we give the essentials for our proof. Let T denote the
infinite rooted binary tree and let A = Aut(T). Note that A ≃ A≀C2 ≃ (A×
A) ⋊ C2 , where C2 = hα | α2 = 1i, so every g ∈ A can be uniquely expressed
as g = (gL , gR )αg for some gL , gR ∈ A and αg ∈ C2 . The Grigorchuk group
is G = ha, b, c, di ≤ A, where the generators a, b, c, d ∈ A are given by
a = (1, 1)α, b = (a, c)1, c = (a, d)1, d = (1, b)1.
One can easily verify that the following identities hold in G:
(1)
a2 = b2 = c2 = d2 = 1,
bc = cb = d, cd = dc = b, db = bd = c.
Starting with an arbitrary word w ∈ {a, b, c, d}∗ , one can rewrite w via the
identities in (1) to a word w′ which represents the same element in G and
does not contain any of
aa, bb, cc, dd, bc, cb, bd, db, cd, dc
as a subword. Thus w′ has the form w′ = x0 ax1 . . . xn−1 axn , where xi ∈
{1, b, c, d} for 0 ≤ i ≤ n and xi 6= 1 for 1 < i < n − 1, so we say that w′ is
alternating or reduced. Every non-trivial element g ∈ G can be represented
by a reduced word. The following remark has an analogous proof to that of
the uniqueness of reduced words in free groups (see [23, Section I.1]).
12
Remark 3. Every word w ∈ {a, b, c, d}∗ can be rewritten via the identities
in (1) to a unique reduced word.
However, not every reduced word represents a nontrivial element: the
word dadadada is reduced, yet it represents the identity in G.
Let G1 be the subgroup of G consisting of all elements that fix the first
level of T. Then |G : G1 | = 2, G1 = hb, c, d, aba, aca, adai and
aba = (c, a)1, aca = (d, a)1, ada = (d, 1)1.
Every element g ∈ G1 can be expressed as an alternating sequence g =
x0 xa1 . . . xn−1 xan , where xi ∈ {1, b, c, d} and xai ∈ {1, aba, aca, ada} for 0 ≤
i ≤ n such that xi 6= 1 if i > 0 and xai 6= 1 if i < n.
Let φL , φR : G1 → G be the group homomorphisms defined as
b
→ c,
b
→ a,
→ d,
→ a,
c
c
d
d
→ b,
→ 1,
φR :
φL :
aba
→ a,
aba
→
c,
aca → a,
aca → d,
ada → 1,
ada → b,
and define φ : G1 → G × G as φ(g) = (φL (g), φR (g)). Then φ is injective,
and |φi (w)| < 12 |w| + 1 < |w| for every reduced w ∈ {a, b, c, d}∗ representing
some element in G1 such that |w| > 1. Let LS be the set of words with an
odd number of a’s, that is,
LS = {w ∈ {a, b, c, d}∗ | |w|a ≡ 1
mod 2}.
Notice that the parity of |w|a is invariant under the rewrite rules (1). Indeed,
if w ∈ LS then w 6=G 1.
The following is the outline of the word problem algorithm in Grigorchuk’s
group (see [8, Section VIII.E] for more details and Figure 2 for an example).
We use ǫ to denote the empty string.
(1) reduce w
(2) if w = ǫ answer YES (meaning w =G 1),
(3) if w ∈ LS answer NO (meaning w 6=G 1),
(4) otherwise answer (φL (w) =G 1 and φR (w) =G 1).
Obviously, w 6=G 1 if and only if there are sequences w0 , . . . , wn ∈ {a, b, c, d}
and φ1 , . . . , φn ∈ {φL , φR } such that wn =G w, wi−1 =G φi (wi ) for i =
1, . . . , n and w0 ∈ LS . The main idea behind our grammar is to invert this
process, i.e. start with a word in w0 ∈ LS and generate a sequence of words
w1 , . . . , wn such that wi−1 =G φi (wi ).
6.2. ET0L grammar. In this subsection we introduce the ET0L grammar
used to generate the co-word problem for Grigorchuk’s group and give an
informal explanation of the roles of the corresponding tables.
13
Our grammar works over the extended alphabet
Σ = {S0 , S1 , a, b, c, d, δ, #},
where S0 is the start symbol, along with tables s, p, hL , hR , u, t and rational
control
R = s∗ {p∗ {hL , hR }u∗ t}∗ .
The process of generating words can be split in three phases. It is important to note that the second and third phase can occur multiple times, as
can be seen in Example 2.
Phase 1: Generate LS . First we generate the language LS of words with
an odd number of occurrences of a. This is done by table s, which can be
seen as the grammar version of a two state finite automaton producing the
regular language LS . We call the words in LS seeds.
(s)
Generate seeds
S0 → aS1 , bS0 , cS0 , dS0
S1 → aS0 , bS1 , cS1 , dS1 , ǫ
∗
Lemma 16. Let w ∈ {a, b, c, d}∗ . Then w ∈ LS if and only if S0 −→s w.
∗
In particular, if S0 −→s w then w 6=G 1.
Phase 2: Invert φL and φR . Once we have produced a seed w0 , we generate
a sequence of words w1 , . . . , wn such that wi−1 =G φi (wi ), and we create
wi from wi−1 by ‘inverting’ the maps φL and φR ; this is achieved via tables
hL , hR :
(hL , hR )
Invert
a →
b →
c →
d →
δ →
Invert
a →
b →
c →
d →
δ →
φL
b, c
ada
aba
aca
d
φR
aba, aca
d
b
c
ada
Table p introduces a new symbol δ, which serves as a placeholder for the
empty word that resulted from applying φL and φR (recall that φL (d) = 1,
φR (ada) = 1).
(p)
Insert
a →
b →
c →
d →
δ
a, δa, aδ
b, δb, bδ
c, δc, cδ
d, δd, dδ
Lemma 17. Let w, w′ ∈ {a, b, c, d}∗ and let i ∈ {L, R} be given. Then
∗
∗
w =G φi (w′ ) if and only if w −→p hi w′ . In particular, if w −→p hi w′ and
w 6=G 1 then w′ 6=G 1.
The proof follows immediately from an analysis of the tables.
14
Phase 3: Use group relations and insert trivial subwords. The following table
inverts the reduction process induced by the identities in (1). We introduce
the symbol # as a placeholder for the reduction process. This means that
# signifies the position of a subword that reduces to the trivial word.
Unreduce
a → a, #a, a#
b → b, #b, b#, c#d, d#c
c → c, #c, c#, b#d, d#b
d → d, #d, d#, b#c, c#b
# → #, a#a, b#b, c#c, d#d
(u)
At this stage we remove all the occurrences of the placeholder #. This is
achieved by table t:
Tidy #
# → ǫ
(t)
Lemma 18. Let w, w′ ∈ {a, b, c, d}∗ be arbitrary. Then w can be obtained
from w′ by reducing rules (induced by the identities in (1)) if and only if
∗
∗
w −→u t w′ . In particular, if w −→u t w′ and w 6=G 1 then w′ 6=G 1.
Example 2. Figure 2 demonstrates the word problem algorithm on input
bcddacbabcaa, showing it to be nontrivial.
bcddacbabcaa
φL dadad φR
1b1
b1b
1
φL
b
φR
c
a ∈ LS
Figure 2. Word problem algorithm applied to bcddacbabcaa.
This word can be obtained from our grammar as follows.
15
Phase 1:
S0 −→s aS1 −→s a
Phase 2:
a −→hL b
Phase 3:
b −→t b
Phase 2:
b −→p δb −→p δbδ −→hR dadad
Phase 3:
dadad −→u b#c#ac#bab#c −→u b#cd#dac#bab#c#
−→u b#cd#dac#bab#ca#a −→t bcddacbabcaa.
6.3. Proof that the ET0L grammar generates exactly the co-word
problem. Let L′ be the language generated by the ET0L grammar with
alphabet Σ, tables s, p, hL , hR , u, c given above and rational control
R = s∗ {p∗ {hL , hR }u∗ t}∗ .
Note that the rational control R is equivalent to s∗ {p∗ hL u∗ t, p∗ hR u∗ t}∗ .
Following Definition 3, L′ is ET0L. Now consider L = L′ ∩{a, b, c, d}∗ . As the
class of ET0L languages is closed under intersection with regular languages,
L is an ET0L language. Intersecting L′ with {a, b, c, d}∗ effectively discards
all words containing symbols S0 , S1 . With this in mind, we will assume that
we are only working with words that do not contain symbols S0 and S1 .
Combining Lemmas 16, 17 and 18 we can immediately show that our
grammar produces words that represent non-trivial elements in the group.
Lemma 19. Let w ∈ {a, b, c, d}∗ be arbitrary. If w ∈ L then w 6=G 1.
Proof. If w ∈ L then, by definition, there is a sequence of words
w0 , w1 , w1′ , . . . , wn , wn′ ∈ {a, b, c, d}∗
and a sequence of tables h1 , . . . , hn ∈ {hL , hR } such that
∗
(1) s −→s w0 ,
∗
(2) wi−1 −→p hi wi′ for i = 1, . . . , n
∗
(3) wi′ −→u c wi ,
(4) wn = w.
We use induction on n. If n = 0 then by Lemma 16 we see that w0 ∈ LS ,
hence w0 6=G 1.
Now suppose that the result has been established for all words that can be
obtained via sequences of length ≤ n − 1. Then wn−1 6=G 1 by the induction
hypothesis. Using Lemma 17 we see that wn−1 = φi (wn′ ), where φi = φL
if hi = hL and φi = φR if hi = hR , thus wn′ 6=G 1. Using Lemma 18 we
see that wn′ =G wn and hence wn 6=G 1, so w indeed represents a nontrivial
element of Grigorchuk’s group.
The next lemma establishes the completeness of our grammar.
16
Lemma 20. Let {a, b, c, d}∗ be arbitrary. If w 6=G 1 then w ∈ L.
Proof. Suppose that w 6=G 1. Following the algorithm described in Subsection 6.1 there is a sequence of words w0 , w0′ , . . . , wn , wn′ and a sequence of
maps φ1 , . . . , φn ∈ {φL , φR } such that
(1) wi is the reduced word obtained by reducing wi′ for i = 0, . . . , n,
(2) |wi |a is even for i ≥ 1 and odd for i = 0,
′
is obtained from wi by applying φi ,
(3) wi−1
′
(4) wn = w.
Again, we proceed by induction on n.
Suppose that n = 0, i.e. w0′ = w. As the parity of the number of
occurrences of the symbol a is invariant with respect to the reduction rules,
∗
we see that |w|a is odd. It follows by Lemma 16 that S0 −→s w and thus
w ∈ L.
Now suppose that the statement holds for all w′ ∈ {a, b, c, d}∗ for which
the word problem algorithm uses up to n−1 levels of recursion. In particular,
S0 −→s
∗ {p∗ h
Lu
∗ t,p∗ h
Ru
∗ t}∗
′
wn−1
.
∗
′
−→p hn wn , where hn = hL if φn = φL
Using Lemma 17 we see that wn−1
∗
and hn = hR if φn = φR . Similarly, using Lemma 18 we see that wn −→u t
∗
∗
′
−→p hn u t w and therefore w ∈ L.
wn′ . Altogether we see that wn−1
Combining Lemma 19 and Lemma 20 we see that a word w ∈ {a, b, c, d}∗
represents a nontrivial element of Grigorchuk’s group if and only if w ∈ L,
which implies the main result of this section.
Theorem 21. The co-word problem in Grigorchuk’s group is an ETOL
language.
7. ET0L languages and 3-manifold groups
The goal of this section is to prove Theorem 22, which is a strengthening
of Theorem B in [2], proved there for indexed instead of ET0L languages.
In 1996 Bridson and Gilman stated the theorem for all manifolds satisfying
the geometrisation conjecture, but since then Perelman [24] proved that all
compact 3-manifolds do, so we can state the result in full generality.
Theorem 22. Let M be a compact 3-manifold or orbifold, and let µ : Σ⋆ →
π1 M be a choice of generators. Then there exists a set of normal forms
L ⊆ Σ⋆ which satisfies the asynchronous fellow-traveler property and is an
ET0L language.
We follow in the footsteps of Bridson and Gilman, whose proof relies on (1)
closure properties of AFL languages, and (2) showing that an appropriate
set of normal forms for Z2 is ET0L. Note that standard regular normal
forms for Z2 will not produce an appropriate language of normal forms for
the extension Z2 ⋊ Z, which needs to satisfy the properties detailed in [2,
page 541].
17
For the sake of completeness we recall the relevant results on AFL languages. As in the paper of Bridson and Gilman (and much of the literature),
we will call a set of normal forms satisfying the asynchronous fellow-traveler
property a combing.
Proposition 23 ([2]). Let A be a full AFL class of languages (such as
regular, context-free, indexed, or ET0L).
(1) (Prop. 2.9) If G1 and G2 both have an asynchronous A-combing,
then so does the free product G1 ∗ G2 .
(2) (Theorem 2.16) Let G be a finitely generated group, and H a subgroup of finite index. Then G admits an asynchronous A-combing if
and only if H admits an asynchronous A-combing.
We first recall the crossing sequence κ(m, n) of Bridson and Gilman, which
gives the EDT0L normal form for Z2 . Let (m, n) ∈ Z2 have m > 0, n ≥ 0,
and consider the line l(m, n+ ) in the plane from (0, 0) to (m, n+ ), where
n+ is chosen slightly larger than n, but small enough to ensure that (1)
l(m, n+ ) does not contain any lattice points except (0, 0), and (2) there are
no lattice points in the interior of the triangle with vertices (0, 0), (m, n)
and (m, n+ ). For the line l(m, n+ ), the sequence formed by recording an
h each time a horizontal line in the plane is crossed and a v each time a
vertical line is crossed is called the crossing sequence κ(m, n). For example,
κ(2, 3) = hvhhv.
Theorem 24. The set L = {κ(m, n) | m > 0, n ≥ 0} is an EDT0L language.
That is, the indexed combing for Z2 in [2] is in fact EDT0L.
Proof. The proof only focuses on the first quadrant in Z2 , and it can be
easily extended to all of Z2 . As is described in [2], L can be generated by
starting with an arbitrary v k , k > 0, and alternately replacing all v’s by hi v
and all h’s by v j h. The EDT0L grammar thus first has to produce v k , and
then apply maps which mimic the morphisms described. Let {q, v, h} be the
extended alphabet, with q the start symbol. Let φq , φv , φh and φs be the
maps defined by φq (q) = qv, φv (v) = hv, φh (h) = vh and φs (q) = v.
Then φs φqk−1 (q) = v k generates the starting point of the crossing sequence, and then we apply any map in {φv , φh }⋆ to v k and obtain the set
L. Thus by Definition 3 the set L is an EDT0L language.
We remark that the EDT0L characterisation for the Z2 combing cannot
be lifted to Z2 ⋊Z and other more general groups, because EDT0L languages
do not form a full AFL, which is essential in several proofs in [2].
Proposition 25. Every semidirect product of the form Z2 ⋊ Z admits an
asynchronous ET0L combing.
Proof. The proof is exactly as that of Corollary 3.5 in [2]. More precisely,
let t be a generator of Z and L be the language of normal forms for Z2 from
Theorem 24. By [2, Theorem 3.1], the language L0 = {t∗ ∪ (t−1 )∗ }L is an
18
asynchronous combing of Z2 ⋊Z, and since ET0L languages are closed under
concatenation with a regular language and finite unions, we get that L0 is
ET0L.
Proof. (of Theorem 22) The work of Thurston, Epstein and Perelman implies
that any π1 M as in the hypothesis is commensurable to the free product of
an automatic group and (possibly) finite extensions of groups of the form
Z2 ⋊Z. Thus the proof follows immediately from Propositions 23 and 25.
8. Open problems
Among the formal languages naturally appearing in group theory none is
more prominent than the word problem, that is, the set of words representing
the trivial element in a finitely generated group. Since EDT0L languages are
not closed under inverse homomorphism, a priori a group may have EDT0L
word problem for one finite generating set but not for another. However,
we do not know of any infinite group which has EDT0L word or co-word
problem for some finite generating set. Since EDT0L languages are relatively
close in complexity to context-free languages, one might wonder whether the
groups with context-free word problem have EDT0L word problem. A first
negative answer is given below.
Proposition 26. Let F be the free group of rank at least two. Then the
word problem in F is not EDT0L.
Proof. It was proved in [22] that if a language L is a context-free generator,
i.e. for every context-free language K there is a regular language RK and
a homomorphism hK such that K = hK (L ∩ RK ), then L is not EDT0L.
It follows by the Chomsky-Schützenberger representation theorem [4] that
every Dyck language on at least two letters is a context-free generator. It can
be easily seen that the word problem in Fn , the free group on n generators,
is isomorphic to Dn⋆ , the symmetric Dyck Language on n letters. It follows
that if n > 1 then the word problem in Fn is not EDT0L.
Question 8.1. Is the word problem for Z (for some or every finite generating
set) EDT0L?
A related open problem is to determine the class of groups having ET0L
word problem; a well known problem is to find a non-virtually free group
with indexed word problem, so a reasonable conjecture here is that the only
groups with ET0L word problem are virtually free.
Conjecture 8.2. A group has EDT0L word problem if and only if it is
finite. A group has ET0L word problem if and only if it is virtually free.
Acknowledgments
We would like to thank Sylvain Salvati for the EDT0L grammar used in
the proof of Proposition 7 and Michel Latteux for the outline for Proposition 26. We also thank Claas Röver for interesting discussions. Further, we
19
would like to thank the anonymous referee for suggesting several simplifications of the submitted manuscript.
The first two authors were partially supported by the Swiss National
Science Foundation grant Professorship FN PP00P2-144681/1, and by a
Follow-On grant of the International Centre of Mathematical Sciences in
Edinburgh. All authors were supported by the Australian Research Council
Discovery Project grant DP160100486.
References
[1] Peter R. J. Asveld. Controlled iteration grammars and full hyper-AFL’s. Information
and Control, 34(3):248–269, 1977.
[2] Martin R. Bridson and Robert H. Gilman. Formal language theory and the geometry
of 3-manifolds. Comment. Math. Helv., 71(4):525–555, 1996.
[3] Martin R. Bridson and Robert H. Gilman. Context-free languages of sub-exponential
growth. J. Comput. System Sci., 64(2):308–310, 2002.
[4] Noam Chomsky and Marcel P Schützenberger. The algebraic theory of context-free
languages. Studies in Logic and the Foundations of Mathematics, 35:118–161, 1963.
[5] Laura Ciobanu, Volker Diekert, and Murray Elder. Solution sets for equations over
free groups are EDT0L languages. Internat. J. Algebra Comput., 26(5):843–886, 2016.
[6] A. Clifford and R. Z. Goldstein. Sets of primitive elements in a free group. J. Algebra,
357:271–278, 2012.
[7] Karel Culik, II. On some families of languages related to developmental systems.
Internat. J. Comput. Math., 4:31–42, 1974.
[8] Pierre de la Harpe. Topics in geometric group theory. Chicago Lectures in Mathematics. University of Chicago Press, Chicago, IL, 2000.
[9] V. Diekert and M. Elder. Solutions of twisted word equations, EDT0L languages, and
context-free groups. ArXiv e-prints, January 2017. ICALP 2017.
[10] V. Diekert, A. Jeż, and M. Kufleitner. Solutions of Word Equations over Partially
Commutative Structures. ArXiv e-prints, March 2016.
[11] A. Ehrenfeucht and G. Rozenberg. On proving that certain languages are not ET0L.
Acta Informat., 6(4):407–415, 1976.
[12] Andrzej Ehrenfeucht and Grzegorz Rozenberg. On some context free languages which
are not ET0L languages. 1974.
[13] Andrzej Ehrenfeucht, Grzegorz Rozenberg, and Sven Skyum. A relationship between
ET0L and EDT0L languages. Theoretical Computer Science, 1(4):325–330, 1976.
[14] Joost Engelfriet, Erik Meineche Schmidt, and Jan van Leeuwen. Stack machines and
classes of nonnested macro languages. J. Assoc. Comput. Mach., 27(1):96–117, 1980.
[15] Robert H. Gilman. A shrinking lemma for indexed languages. Theoret. Comput. Sci.,
163(1-2):277–281, 1996.
[16] R. I. Grigorchuk and A. Machı̀. An example of an indexed language of intermediate
growth. Theoret. Comput. Sci., 215(1-2):325–327, 1999.
[17] Takeshi Hayashi. On derivation trees of indexed grammars: an extension of the
uvwxy-theorem. Publ. Res. Inst. Math. Sci., 9:61–92, 1973/74.
[18] Derek F. Holt and Claas E. Röver. Groups with indexed co-word problem. Internat.
J. Algebra Comput., 16(5):985–1014, 2006.
[19] Sanjay Jain, Alexei Miasnikov, and Frank Stephan. The complexity of verbal languages over groups. In Proceedings of the 2012 27th Annual ACM/IEEE Symposium
on Logic in Computer Science, pages 405–414. IEEE Computer Soc., Los Alamitos,
CA, 2012.
[20] Ilya Kapovich and Alexei Myasnikov. Stallings foldings and subgroups of free groups.
J. Algebra, 248(2):608–668, 2002.
20
[21] Olga Kharlampovich and Alexei Myasnikov. Definable sets in a hyperbolic group.
Internat. J. Algebra Comput., 23(1):91–110, 2013.
[22] Michel Latteux. Sur les générateurs algébriques et linéaires. Acta Inform., 13(4):347–
363, 1980.
[23] Roger C. Lyndon and Paul E. Schupp. Combinatorial group theory. Springer-Verlag,
Berlin, 1977. Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 89.
[24] Grigoriy Perelman. Ricci flow with surgery on three-manifolds. 2013.
https://arxiv.org/abs/math/0303109.
[25] G. Rozenberg and A. Salomaa, editors. Handbook of formal languages. Vol. 1.
Springer-Verlag, Berlin, 1997. Word, language, grammar.
[26] Grzegorz Rozenberg and Arto Salomaa. The Book of L. Springer, 1986.
[27] Pedro V. Silva and Pascal Weil. Automorphic orbits in free groups: words versus
subgroups. Internat. J. Algebra Comput., 20(4):561–590, 2010.
[28] John R. Stallings. Topology of finite graphs. Invent. Math., 71(3):551–565, 1983.
School of Mathematical and Computer Sciences, Heriot-Watt University,
Edinburgh EH14 4AS, Scotland
E-mail address: [email protected]
University of Technology Sydney, Ultimo NSW 2007, Australia
E-mail address: [email protected]
University of Technology Sydney, Ultimo NSW 2007, Australia
E-mail address: [email protected]
21
| 4math.GR
|
arXiv:1606.08995v4 [math.GR] 31 Jan 2017
MULTIFRACTION REDUCTION II: CONJECTURES FOR ARTIN-TITS
GROUPS
PATRICK DEHORNOY
Abstract. Multifraction reduction is a new approach to the word problem for Artin-Tits
groups and, more generally, for the enveloping group of a monoid in which any two elements
admit a greatest common divisor. This approach is based on a rewrite system (“reduction”)
that extends free group reduction. In this paper, we show that assuming that reduction
satisfies a weak form of convergence called semi-convergence is sufficient for solving the word
problem for the enveloping group, and we connect semi-convergence with other conditions
involving reduction. We conjecture that these properties are valid for all Artin-Tits monoids,
and provide partial results and numerical evidence supporting such conjectures.
1. Introduction
Artin-Tits groups are those groups that admit a positive presentation with at most one
relation s... = t... for each pair of generators s, t and, in this case, the relation has the form
stst... = tsts..., both sides of the same length [4, 23]. It is still unknown whether the word
problem is decidable for all Artin–Tits groups as, at the moment, decidability was established
for particular families only: braid groups (E. Artin [3] in 1947), spherical type (P. Deligne [20]
and E. Brieskorn–K. Saito [5] in 1972), large type (K.I. Appel–P.E. Schupp [2] in 1983), trianglefree (S. Pride [29] in 1986), FC type (J. Altobelli [1] and A. Chermak [7] in 1998). Later on,
some of these groups were proved to be biautomatic or automatic [22, 6].
This paper, which follows [13] but is self-contained, continues the investigation of multifraction reduction, a new approach to the word problem for Artin-Tits groups and, more generally,
for the enveloping group U(M ) of a cancellative monoid M in which any two elements admit a
left and a right greatest common divisor (“gcd-monoid”). This approach is based on a certain
algebraic rewrite system, called reduction, which unifies and (properly) extends all previous
rewrite systems based on exploiting the Garside structure of Artin–Tits monoids [9, 30, 27].
It is proved in [13] that, if the monoid M satisfies various properties involving the divisibility
relations, all true in every Artin–Tits monoid, together with an additional assumption called
the 3-Ore condition, then reduction is convergent and every element of the enveloping group
of M admits a unique representation by an irreducible multifraction, directly extending the
classical result by Ø. Ore about representation by irreducible fractions.
In the current paper, we address the case of a general gcd-monoid, when the 3-Ore condition
is not assumed. In this case, reduction is not convergent, and there is no unique representation
of the elements of the group U(M ) by irreducible multifractions. However, we introduce a
new, weaker condition called semi-convergence, and prove that most of the applications of the
convergence of reduction can be derived from its semi-convergence, in particular a solution of
the word problem for U(M ) whenever convenient finiteness conditions are satisfied. This makes
the following conjecture crucial:
Conjecture A. Reduction is semi-convergent for every Artin-Tits monoid.
1991 Mathematics Subject Classification. 20F36, 20F10, 20M05, 68Q42, 18B40.
Key words and phrases. Artin-Tits monoid; Artin-Tits group; gcd-monoid; enveloping group; word problem; multifraction; reduction; semi-convergence; cross-confluence; tame reduction; van Kampen diagram;
embeddability.
1
2
PATRICK DEHORNOY
A proof of Conjecture A would imply the decidability of the word problem for every ArtinTits group. The reasons for believing in Conjecture A are multiple. One abstract reason is that
reduction is really specific and uses both the whole Garside structure of Artin–Tits monoids and,
for the finiteness of the set of basic elements, some highly nontrivial properties of the associated
Coxeter groups [15, 21]: this may be seen as more promising than a generic approach based
on, say, a “blind” Knuth–Bendix completion. Next, we state several related conjectures (“B”,
“C”, “Cunif ”), of which some partial cases are proven and which suggest the existence of a
rich rigid structure. Another reason is the existence of massive computer tests supporting all
the conjectures and, at the same time, efficiently discarding wrong variations and dead-ends.
Finally, the existence of a proof in the special case of FC type is a positive point. In the same
direction, a weak version of Conjecture A (sufficient for solving the word problem) was recently
established for all Artin–Tits groups of sufficiently large type [18]: although saying nothing
about A itself, this shows that reduction is relevant for a new family of Artin–Tits groups.
We present below a state-of-the-art description of the known results involving multifraction
reduction, and report about computer experiments supporting Conjecture A and its variants.
The paper is organized as follows. We gather in Section 2 the needed prerequisites about
multifractions, gcd-monoids, and reduction. Semi-convergence and Conjecture A are introduced
in Section 3, and their consequences are established. In Section 4, we analyze specific cases of
reduction, namely divisions and their extensions, tame reductions. This leads to a new property,
stated as Conjecture B, which is stronger than Conjecture A but easier to experimentally check
and maybe to establish. Then, we introduce in Section 5 cross-confluence, a new property
of reduction that involves both reduction and a symmetric counterpart of it. This leads to
Conjecture C and its uniform version Cunif , again stronger than Conjecture A but possibly
more accessible. In Section 6, we analyze the case of small depth multifractions. We prove in
particular that semi-convergence for multifractions of depth 2 is equivalent to M embedding
into its enveloping group, and semi-convergence for multifractions of depth 4 is equivalent to a
unique decomposition property for fractions in U(M ). Finally, we gather in Section 7 reports
about computer experiments and a few comments about further possible developments.
Acknowledgments. The author thanks Friedrich Wehrung for many discussions about the
content of this paper. In particular, the notion of a lcm-expansion mentioned in Sec. 7.1
appeared during our joint work of interval monoids [19]. The author also thanks both the
editor and the referee, whose many suggestions certainly improved the exposition significantly.
2. Multifraction reduction
In this introductory section, we recall the notions of a multifraction and of a gcd-monoid, as
well as the definition of multifraction reduction [13].
2.1. Multifractions. If M is a monoid, we denote by U(M ) and ι the enveloping group of M
and the canonical morphism from M to U(M ), characterized by the universal property that
every morphism from M to a group factors through ι. By definition, every element g of U(M )
can be expressed as
(2.1)
ι(a1 )ι(a2 )−1 ι(a3 )ι(a4 )−1 ···
or
ι(a1 )−1 ι(a2 )ι(a3 )−1 ι(a4 )···,
with a1 , ..., an in M . We shall investigate U(M ) using such expressions. In [13], without loss of
generality, we only considered, expressions (2.1) where the first term ι(a1 ) is positive (possibly
trivial, that is, equal to 1). Here, in particular in view of Section 5, we skip that condition, and
allow for both signs in the first entry.
Definition 2.2. Let M be a monoid. Let M be a disjoint copy of M ; call the elements of M
(resp., M ) positive (resp., negative). For n > 1, a positive (resp., negative) n-multifraction
on M is a finite sequence (a1 , ..., an ) with entries in M ∪ M , alternating signs, and a1 in M
MULTIFRACTION REDUCTION II
3
(resp., M ). The set of all multifractions (resp., all positive multifractions) completed with the
empty sequence ∅ is denoted by FM± (resp., FM ). A multiplication is defined by
for an and b1 in M ,
(a1 , ..., an−1 , an b1 , b2 , ..., bp )
(a1 , ..., an ) · (b1 , ..., bp ) = (a1 , ..., an−1 , b1 an , b2 , ..., bp )
for an and b1 in M ,
(a1 , ..., an , b1 , ..., bp )
otherwise,
extended with a · ∅ = ∅ · a = a for every a.
We use a, b, ... as generic symbols for multifractions, and ai for the ith entry of a counted
from 1. For a in FM± , the length of a (number of entries) is called its depth, written kak. We
identify an element a of M with the depth one positive multifraction (a). Multfractions will
play the role of iterated fractions, and the following convention is then convenient:
Notation 2.3. For a1 , ..., an in M , we put
(2.4)
a1 /···/an := (a1 , a2 , a3 , a4 , ...)
and /a1 /···/an := (a1 , a2 , a3 , a4 , ...).
We say that i is positive (resp., negative) in a if ai (resp., ai ) occurs in the expansion of a.
With this convention, we recover the notation of [13], where only positive multifractions
are considered and M remains hidden. Multifractions are adequately illustrated by associating
with every element a of M an arrow labeled a, concatenating arrows to represent the product
in M , and associating with every multifraction the path made of (the arrows of) the successive
entries with alternating orientations. The rules for the multiplication of a and b can be read in
the following diagrams:
an
b1
an b 1
- n positive in a, 1 positive in b:
·
=
an
b1
an b 1
- n positive in a, 1 negative in b:
·
=
an
b1
an b 1
- n negative in a, 1 positive in b:
·
=
a
b
b 1 an
n
1
- n negative in a, 1 negative in b:
·
=
Proposition 2.5. (i) The set FM± equipped with · and ∅ is a monoid generated by M ∪ M ,
and FM is the submonoid of FM± generated by M ∪ {1/a | a ∈ M }. The family of depth one
multifractions is a submonoid isomorphic to M .
(ii) Let ≃± be the congruence on FM± generated by (1, ∅) and the pairs (a/a, ∅) and (/a/a, ∅)
with a in M . For a in FM± , let ι(a) be the ≃± -class of a. Then the group U(M ) is (isomorphic
to) FM± /≃± and, for all a1 , ..., an in M , we have
(2.6)
ι(a1 /···/an ) = ι(a1 )ι(a2 )−1 ι(a3 )···,
ι(/a1 /···/an ) = ι(a1 )−1 ι(a2 )ι(a3 )−1 ···.
(iii) The restriction of ≃± to FM is the congruence ≃ generated by (1, ∅) and the pairs
(a/a, ∅) and (1/a/a, ∅) with a in M . The group U(M ) is also isomorphic to FM /≃. The
translation a 7→ 1 · a maps FM± onto FM and, for all a, b in FM± , the relation a ≃± b is equivalent
to 1 · a ≃ 1 · b.
Proof. The argument is similar to the proof of [13, Proposition 2.4], and we only point the
differences due to using signed multifractions. For (i), associativity is checked directly, and the
generating subsets for FM± and FM follow from the equalities
a1 /···/an = a1 · a2 · a3 · a4 · ··· = a1 · 1/a2 · a3 · 1/a4 · ··· :
(2.7)
±
both hold in FM , and the second only involves positive multifractions.
For (ii) and (iii), for every a in FM± , the definition of ≃± implies ι(a) = ι(a)−1 and, writing
+
ι (a) for the ≃-class of a, that of ≃ implies ι+ (1/a) = ι+ (a)−1 . Hence both FM± /≃± and FM /≃
are groups generated by M . One easily checks that the latter groups satisfy the universal
property defining U(M ), and are therefore isomorphic to U(M ). Then (2.6) directly follows
from (2.7).
Next, for every a in FM± , the product 1 · a belongs to FM . Then, for a, b in FM± , write a ≈ b
for 1 · a ≃ 1 · b. By considering all sign combinations and using relations like 1/ab ≃ 1/b · 1/a,
4
PATRICK DEHORNOY
one checks that ≈ is a congruence on FM± , and it contains the pairs (1, ∅), (a/a, ∅), and
(/a/a, ∅) that generate ≃± , as one finds for instance 1 · /a/a = 1/a/a ≃ ∅ = 1 · ∅. Hence ≈
includes ≃± . In the other direction, ≃± contains pairs that generate ≃ and, being compatible
with multiplication in FM± , it is in particular compatible with multiplication in FM . So ≃ is
included in ≃± , and a ≃± b implies 1 · a ≃ 1 · b since 1 is invertible mod ≃± . Hence, ≃± is
included in ≈ and, finally, ≃± and ≈ coincide, which means that, for all a, b in FM± , we have
(2.8)
a ≃± b
⇐⇒
1 · a ≃ 1 · b.
For a positive, we have 1 · a = a, so (2.8) implies in particular that ≃ is the restriction of ≃±
to FM , and the rest follows easily.
Hereafter, we identify U(M ) with FM± /≃± and FM /≃. This representation is redundant in
that, for every a in M , the inverse ι(a)−1 of ι(a) is represented both by the depth 1 negative
multifraction /a and the depth 2 positive multifraction 1/a.
We conclude this introduction with some terminology that will be used frequently:
Definition 2.9. A multifraction a is called unital if a ≃± 1 holds, i.e., if a represents 1 in the
group U(M ). It is called trivial if all entries are equal to 1 or 1. For n > 0, we write 1n for
1/···/1, n terms; for n < 0, we write 1n for /1/1/···/1, |n| terms; in practice, we shall often
omit the index and write 1 for a trivial multifraction.
2.2. Gcd-monoids. The reduction process we investigate requires that the ground monoid is
a gcd-monoid. We recall the basic definitions, referring to [13] (and [16]) for more details.
Let M be a monoid. For a, b in M , we say that a left divides b or, equivalently, that b is
a right multiple of a, written a 6 b, if ax = b holds for some x in M . If M is a cancellative
monoid and 1 is the only invertible element in M , the left divisibility relation is a partial order
on M . In this case, when they exist, the greatest common 6-lower bound of two elements a, b
is called their left gcd, denoted by a ∧ b, and their least common 6-upper bound is called their
right lcm, denoted by a ∨ b.
Symmetrically, we say that a right divides b or, equivalently, that b is a left multiple of a,
e b, if xa = b holds for some x. Under the same hypotheses, 6
e is a partial order
written a 6
e and ∨
e.
on M , with the derived right gcd and left lcm written ∧
Definition 2.10. We say that M is a gcd-monoid if M is a cancellative monoid, 1 is the only
invertible element in M , and any two elements of M admit a left gcd and a right gcd.
Typical examples of gcd-monoids are Artin-Tits monoids. Many more examples are known.
In particular, every Garside or preGarside monoid [11, 16, 25] is a gcd-monoid.
Standard arguments [13, Lemma 2.15] show that a gcd-monoid admits conditional right and
left lcms: it need not be true that any two elements admit a right lcm, but any two elements
that admit a common right multiple admit a right lcm, and similarly on the left.
The gcd and lcm operations of a gcd-monoid are connected in several ways with the product.
We refer to [16, section II.2] for the (easy) proof of the rule for an iterated lcm:
Lemma 2.11. If M is a gcd-monoid and a, b, c belong to M , then a ∨ bc exists if and only if
a ∨ b and a′ ∨ c do, where a′ is defined by a ∨ b = ba′ , and then we have
(2.12)
a ∨ bc = a · b′ c′ = bc · a′′ .
with a ∨ b = ba′ = ab′ and a′ ∨ c = a′ c′ = ca′′ .
This implies in particular that a 6 bc holds if and only if a ∨ b exists and a′ 6 c holds, with a′
defined by a ∨ b = ba′ .
Lemma 2.13. If M is a gcd-monoid and a, b, a′ , b′ , c belong to M and satisfy ab′ = ba′ , then
a ∧ b = a′ ∧ c = 1 implies a ∧ bc = 1.
Proof. Assume x 6 a and x 6 bc. By Lemma 2.11, b ∨ x must exist and, writing b ∨ x = bx′ , we
must have x′ 6 c. On the other hand, x 6 a implies x 6 ab′ = ba′ . So ba′ is a common right
MULTIFRACTION REDUCTION II
5
multiple of b and x, hence it is a right multiple of their right lcm bx′ . As M is left cancellative,
bx′ 6 ba′ implies x′ 6 a′ . Hence x′ left divides both a′ and c, and a′ ∧ c = 1 implies x′ = 1,
whence x 6 b. Then x left divides both a and b, and a∧b = 1 implies x = 1, hence a∧bc = 1.
We shall also need the notion of a noetherian monoid. If M is a gcd-monoid, we use < for
the proper version of left divisibility: a < b holds if we have b = ax for some non-invertible x,
e vs. 6.
e
i.e., for x 6= 1, and similarly for <
e are well-founded,
Definition 2.14. A monoid M is called noetherian if the relations < and <
e
i.e., every nonempty subset of M has a <-minimal element and a <-minimal
element.
Note that a monoid hS | Ri+ is noetherian whenever each relation in R is homogeneous, i.e.,
it has the form u = v with u, v of the same length: indeed, a < b implies that any word in S
representing a is shorter than any word representing b, and an infinite <-descending sequence
cannot exist. Artin-Tits monoids are typical examples.
2.3. Reduction of multifractions. Introduced in [13], our tool for investigating the congruence ≃± on FM± is reduction, a family of partial depth-preserving transformations that, when
defined, map a multifraction to a ≃± -equivalent multifraction. These transformations are written as an action on the right: when defined, a • R is the result of applying R to a.
Definition 2.15. If M is a gcd-monoid and a, b lie in FM± , then, for i > 1 and x in M , we
declare that b = a • Ri,x holds if we have kbk = kak, bk = ak for k 6= i − 1, i, i + 1, and there
exists x′ (necessarily unique) satisfying
for
for
for
for
i=1
i=1
i>2
i>2
positive in a:
negative in a:
positive in a:
negative in a:
bi−1
bi−1
b i x = ai ,
xbi = ai ,
e ai ,
= x′ ai−1 , bi x = x′ ai = x ∨
= ai−1 x′ , xbi = ai x′ = x ∨ ai ,
bi+1 x = ai+1 ,
xbi+1 = ai+1 ,
bi+1 x = ai+1 ,
xbi+1 = ai+1 .
We write a ⇒ b if a • Ri,x holds for some i and some x 6= 1, and use ⇒∗ for the reflexive–
±
transitive closure of ⇒. The rewrite system RM
so obtained on FM± is called reduction, and
its restriction to FM (positive multifractions) is denoted by RM . A multifraction a is called
R-reducible if a ⇒ b holds for at least one b, and R-irreducible otherwise.
The system RM is the one investigated in [13], where only positive multifractions are consid±
ered: the only difference between [13, Def. 3.4] and Def. 2.15 is the adjunction in RM
of a rule for
the reduction at level 1 of a negative multifraction. As Ri,x preserves the sign of multifractions,
no specific notation is needed for the restriction of Ri,x to positive multifractions.
±
The reduction systems RM and RM
extend free reduction (deletion of factors x−1 x or xx−1 ):
applying Ri,x to a consists in removing x from ai+1 and pushing it through ai using an lcm
operation, see Figure 1. A multifraction a is eligible for R1,x if and only if x divides both a1
and a2 , on the side coherent with their signs, and, eligible for Ri,x with i > 2 if and only if x
divides ai+1 and admits a common multiple with ai , on the due side again.
ai
ai−1
...
x′
bi−1
⇐
bi
...
x
bi+1
ai
ai−1
ai+1
x′
bi−1
⇐
bi
ai+1
...
x
bi+1
Figure 1. The reduction rule Ri,x : starting from a (grey), we extract x from ai+1 ,
push it through ai by taking the lcm of x and ai (indicated by the small round arc),
and incorporate the remainder x′ in ai−1 to obtain b = a • Ri,x (colored). The left
hand side diagram corresponds to the case when i is negative in a, i.e., ai is crossed
negatively, the right hand one to the case when i is positive in a, i.e., ai is crossed
positively, with opposite orientations of the arrows.
6
PATRICK DEHORNOY
e 2 , here written
Example 2.16. Let M be the Artin-Tits monoid of type A
ha, b, c | aba = bab, bcb = cbc, cac = acai+ ,
and let a := 1/c/aba. Then a is eligible for R2,a and R2,b , since a and b left divide aba and admit
a common right multiple with c. We find a • R2,a = ac/ca/ba and a • R2,b = bc/cb/ab. The
latter are eligible for no reduction R1,x , since ac and ca (resp., bc and cb) admit no nontrivial
common right divisor, and for no reduction R2,x , since the only nontrivial left divisors of ba
(resp., ab) are b and ba (resp., a and ab), which admit no common right multiple with ca (resp.,
ac). Hence these multifractions are R-irreducible.
We now state the basic properties of reduction needed below. They directly extend those
established for positive multifractions in [13]. Verifying them in the general case is easy.
Lemma 2.17. Assume that M be a gcd-monoid.
(i) The relation ⇒∗ is included in ≃± , i.e., a ⇒∗ b implies a ≃± b.
(ii) The relation ⇒∗ is compatible with the multiplication of FM± .
(iii) For all a, b and p, q, the relation a ⇒∗ b is equivalent to 1p · a · 1q ⇒∗ 1p · b · 1q .
Proof. (i) It directly follows from the definition (and from Fig. 1) that b = a • Ri,x implies
ι(ai−1 )ι(ai )−1 ι(ai+1 ) = ι(bi−1 )ι(bi )−1 ι(bi+1 )
(resp., ι(ai−1 )−1 ι(ai )ι(ai+1 )−1 = ι(bi−1 )−1 ι(bi )ι(bi+1 )−1 ) for i negative (resp., positive) in a.
(ii) Assume b = a • Ri,x , and let c be an r-multifraction. If the signs of r in c and 1 in a are
different, then c · a is the concatenation of c and a, similarly c · b is the concatenation of c and b,
and c · b = (c · a) • Ri+r,x trivially holds.
If r is positive in c and 1 is positive in a, we have c · a = c1 /···/cr−1 /cr a1 /a2 /··· and
c · b = c1 /···/cr−1 /cr b1 /b2 /···, and we obtain c · b = (c · a) • Ri+r−1,x : the point is that, if x
both right divides a1 and a2 , it a fortiori right divides cr a1 and a2 .
Finally, assume that r negative in c and 1 is negative in a. Then we find c·a = c1 /···/a1 cr /a2 /···
and c · b = c1 /···/b1 cr /b2 /···: the argument is the same as above, mutatis mutandis: the assumption that a • R1,x is defined means that x both left divides a1 and a2 , which implies that it a
fortiori left divides a1 cr and a2 . Hence (c·a)•Ri+r−1,x is defined and we find c·b = (c·a)•Ri+r−1,x
again. This completes compatibility with left multiplication.
The compatibility on the right is similar: adding extra entries cannot destroy the eligibility
for reduction. Let n = kak. Everything is trivial for i < n − 1, so we assume i = n − 1. If the
signs of n in a and of 1 in c are different, the multiplication is a concatenation, and we obtain
b · c = (a · c) • Rn−1,x trivially. If n is positive in a and 1 is positive in c, the argument is the
same as for Rn−1,x . Finally, assume that n is negative in a and 1 is negative in c. Then we
find a · c = a1 /···/an−1 /c1 an /c2 /··· and b · c = b1 /···/bn−1 /c1 bn /c2 /···. The assumption that
e an−1 exists and x right divides an , which implies that it a
a • Rn−1,x is defined means that x ∨
fortiori right divides c1 an . Hence (a · c) • Rn−1,x is defined, yielding b · c = (a · c) • Rn−1,x again.
Thus reduction is compatible with multiplication on the right.
(iii) That a ⇒ b implies 1p · a · 1q ⇒ 1p · b · 1q follows from (ii) directly. Conversely, assume
1p · b · 1q = (1p · a • 1q ) • Ri,x with x 6= 1. Let n = kak = kbk, and assume that the entries of a
occur in 1p · b · 1q from r + 1 to r + n (with r = p or r = p − 1 according to the sign of p in 1p
and that of 1 in a). Then we necessarily have r + 1 6 i < r + n. Indeed, i < r and i > r + n
are impossible, since the (i + 1)st entry of 1p · a · 1q is trivial. Moreover, in the case i = r + 1,
the element x necessarily divides a1 , since, otherwise, the rth entry of 1p · b · 1q could not be
trivial. Hence, a • Ri,x is defined, and it must be equal to b.
Finally, we have a simple sufficient condition for termination.
±
Lemma 2.18. [13, Proposition 3.13] If M is a noetherian gcd-monoid, then RM
is terminating:
every sequence of reductions leads in finitely many steps to an R-irreducible multifraction.
MULTIFRACTION REDUCTION II
7
We skip the proof, which is exactly the same in the signed case as in the positive case, and
consists in observing that a ⇒ b forces b to be strictly smaller than a for some antilexicographical
ordering on FM± (comparing multifractions starting from the highest entry).
±
2.4. The convergent case. The rewrite system RM
(as any rewrite system) is called convergent if every element, here every multifraction a, admits a unique R-irreducible reduct, usually
denoted by red(a). The main technical result of [13] is
Proposition 2.19. If M is a noetherian gcd-monoid satisfying the 3-Ore condition:
If three elements of M pairwise admit a common right (resp. left) multiple,
(2.20)
then they admit a common right (resp. left) multiple,
±
then RM
is convergent.
When a monoid M is eligible for Proposition 2.19, one easily deduces that two multifractions a, b represent the same element of U(M ) if and only if red(a) · 1p = red(b) · 1q holds
for some p, q and, from there, that the monoid M embeds in its enveloping group U(M ) and
b ± is
b
every element of U(M ) is represented by a unique R-irreducible
multifraction, where R
M
±
obtained from RM by adding a rule that removes trivial final entries. It is also proved in [13]
that, under mild additional finiteness assumptions on M (see Section 3 below), the relation ⇒∗
on M is decidable and, from there, so is the word problem for U(M ) when the 3-Ore condition
is satisfied.
The above results are relevant for a number of gcd-monoids. We recall that an Artin-Tits
monoid M = hS | Ri+ is said to be of spherical type if the Coxeter group obtained by adding
to (S, R) the relation s2 = 1 for each s in S is finite [5]. And M is said to be of type FC if,
for every subfamily S ′ of S such that, for all s, t in S ′ , there is a relation s... = t... in R, the
submonoid of M generated by S ′ is spherical [1, 24].
Proposition 2.21. [13, Proposition 6.5] An Artin-Tits monoid satisfies the 3-Ore condition if
and only if it is of type FC.
However, a number of Artin-Tits monoids fail to be of type FC and therefore are not eligible
e 2 considered in Example 2.16. So we are
for Proposition 2.19, typically the monoid of type A
left with the question of either weakening the assumptions for Proposition 2.19, or using a
conclusion weaker than convergence.
3. Semi-convergence
After showing in Subsection 3.1 that the 3-Ore assumption cannot be weakened when proving
±
±
the convergence of RM
, we introduce in Subsection 3.2 a new property of RM
called semiconvergence, which, as the name suggests, is weaker than convergence. We conjecture that,
±
for every Artin-Tits monoid, the system RM
is semi-convergent (“Conjecture A”). We prove
±
in Subsection 3.3 that most of the consequences known to follow from the convergence of RM
follow from its semi-convergence, in particular in terms of controlling the group U(M ) from
inside the monoid M and solving its word problem. Finally, we describe in Subsection 3.4
several variants of semi-convergence.
3.1. The strength of the 3-Ore condition. A first attempt for improving Proposition 2.19
±
could be to establish the convergence of RM
from an assumption weaker than the 3-Ore condition. This approach fails, as the latter turns out to be not only sufficient, but also necessary.
Hereafter we say that a monoid M satifies the right (resp., left) 3-Ore condition when (2.20) is
valid for right (resp., left) multiples. First, we recall
Lemma 3.1. [13, Lemma 2.12] If M is a gcd-monoid, and a, b, c, d are elements of M satisfying
e d = 1.
ad = bc, then ad is the right lcm of a and b if and only if c and d satisfy c ∧
Proposition 3.2. Let M be a gcd-monoid.
(i) If RM is convergent, then M satisfies the right 3-Ore condition.
±
(ii) If RM
is convergent, then M satisfies the 3-Ore condition.
8
PATRICK DEHORNOY
Proof. (i) Assume that x, y, z belong to M and pairwise admit common right multiples, hence
right lcms. Write
x ∨ y = xy ′ = yx′ ,
y ∨ z = yz ′ = zy ′′ ,
x ∨ z = xz ′′ = zx′′ .
Let a := 1/x/y ∨ z. As in Example 2.16, we find
b = a • R2,y = y ′ /x′ /z ′
and c = a • R2,z = z ′′ /x′′ /y ′′ .
The assumption that RM is convergent implies that b and c both reduce to d := red(a). By
construction, d is of depth 3, and d3 must be a common right divisor of b3 and c3 , which are z ′
e y ′′ = 1, whence d3 = 1. Therefore, there exist d1 and d2
and y ′′ . Now, Lemma 3.1 implies z ′ ∧
∗
in M satisfying 1/x/y ∨ z ⇒ d1 /d2 /1. By Lemma 2.17, we deduce 1/x/y ∨ z ≃± d1 /d2 /1,
hence, by (2.6), x−1 (y ∨ z) = d1 d−1
in U(M ). This implies xd1 = (y ∨ z)d2 in U(M ), hence
2
in M , since the assumption that RM is convergent implies that M embeds in U(M ). It follows
that x and y ∨ z admit a common right multiple, hence that x, y, and z admit a common right
multiple. Hence M satisfies the right 3-Ore condition.
(ii) The argument is symmetric, with negative multifractions. Assume that x, y, z belong
to M and pairwise admit common left multiples, hence left lcms. Write
e y = y ′ x = x′ y,
x∨
e z = z ′ y = y ′′ z,
y∨
Let a := /1/x/y ∨ z, in FM± \ FM . Then we have
b = a • R2,y = /y ′ /x′ /z ′
e z = z ′′ x = x′′ z.
x∨
and c = a • R2,z = /z ′′ /x′′ /y ′′ .
±
The assumption that RM
is convergent implies that b and c admit a common R-reduct, say d.
By construction, d is of depth 3, and d3 must be a common left divisor of b3 and c3 , which are
z ′ and y ′′ . By the symmetric counterpart of Lemma 3.1, we have z ′ ∨ y ′′ = 1, whence d3 = 1.
e ∗ /d1 /d2 /1. By Lemma 2.17, we deduce
Therefore, there exist d1 and d2 in M satisfying a ⇒
±
−1
e z ≃ /d1 /d2 /1, hence, by (2.6), x(y ∨
e z) = d−1
e
/1/x/y ∨
1 d2 in U(M ), whence d1 x = d2 (y ∨ z)
e z admit a common left multiple, hence that
in U(M ), hence in M . This shows that x and y ∨
x, y, and z admit a common left multiple. Hence M satisfies the left 3-Ore condition.
Note that the argument for (i) cannot be used for (ii), because one should start with a :=
e z and then we know nothing about the first entry(ies) of red(a).
1/1/x/y ∨
In principle, the right 3-Ore condition is slightly weaker than the full 3-Ore condition, and
±
the convergence of RM might be weaker than that of RM
. However, when M is an Artin-Tits
monoid, all the above conditions are equivalent to M being of type FC and, therefore, none is
weaker. So it seems hopeless to improve Proposition 2.19.
3.2. Semi-convergence. We are thus led to explore the other way, namely obtaining useful
±
information about U(M ) from a property weaker than the convergence of RM or RM
. This is
the approach we develop in the rest of this paper.
When the system RM is not convergent, a ≃-class may contain several R-irreducible multifractions, and there is no distinguished one: in Example 2.16, the automorphism that exchanges a and b exchanges the two R-irreducible reducts of a, making them indiscernible.
However, a direct consequence of convergence is
±
Lemma 3.3. If M is a gcd-monoid and RM
is convergent, then, for every a in FM± ,
(3.4)
If a is unital, then a ⇒∗ 1 holds.
±
Indeed, if RM
is convergent, a ≃± 1 implies (and, in fact, is equivalent to) red(a) = red(1),
hence red(a) = 1, since 1 is R-irreducible. Note that, by Lemma 2.17(i), the converse implication of (3.4) is always true: a ⇒∗ 1 implies that a and 1 represent the same element of U(M ),
hence that a represents 1.
±
When RM
is not convergent, (3.4) still makes sense, and it is a priori a (much) weaker
condition than convergence. This is the condition we shall investigate below:
MULTIFRACTION REDUCTION II
9
±
Definition 3.5. If M is a gcd-monoid, we say that RM
(resp., RM ) is semi-convergent if (3.4)
±
holds for every a in FM (resp., in FM ).
±
Thus Lemma 3.3 states that RM
is semi-convergent whenever it is convergent. By Propo±
sition 2.21, RM and RM are semi-convergent for every Artin-Tits monoid M of type FC. But,
as can be expected, semi-convergence is strictly weaker than convergence. We refer to [19] for
±
the construction of monoids for which RM
is semi-convergent but not convergent.
The main conjecture we propose is:
Conjecture A. For every Artin-Tits monoid M , the system RM is semi-convergent.
We shall report in Section 7 about experimental data supporting Conjecture A. For the
moment, we just mention one example illustrating its predictions.
e 2 . We saw in Example 2.16 that
Example 3.6. Let M be the Artin-Tits monoid of type A
±
RM
is not convergent: a = 1/c/aba admits the two distinct irreducible reducts ac/ca/ba and
bc/cb/ab. When we multiply the former by the inverse of the latter (see Subsection 3.3 below),
we obtain the 6-multifraction b = ac/ca/ba/ab/cb/bc which, by construction, is unital. Then
Conjecture A predicts that b must reduce to 1. This is indeed the case: we find
b • R3,ab R4,cb R5,bc R1,ac R2,cbc R3,bc R1,bc = 1
(as well as b • R5,bc R3,ac R1,ac R3,b R4,c R2,c = 1: the reduction sequence is not unique).
More generally, we can observe that, for every gcd-monoid M , if a′ and a′′ are two reducts of
an n-multifraction a, then the 2n-multifraction b obtained by concatenating a′ and the inverse
of a′′ is always reducible whenever a′′ is nontrivial: by construction, we have bn = a′n and
bn+1 = a′′n . An obvious induction shows that a ⇒∗ a′ implies that a′n divides an (on the left or
on the right, according to the sign of n in a) and, similarly, a′′n divides an . Hence bn and bn+1
admit a common multiple and, therefore, b is eligible for some reduction Rn,x with x 6= 1
whenever bn+1 , i.e., a′′n , is not 1. Finally, if a′′n = 1 holds, then b is eligible for R2n−m,a′′m , where
m is the largest index such that a′′m is nontrivial.
3.3. Applications of semi-convergence. Most of the consequences of convergence already
follow from semi-convergence—whence the interest of Conjecture A. We successively consider
the possibility of controlling the congruence ≃± , the decidability of the word problem, and what
is called Property H.
Controlling ≃± . In order to investigate ≃± without convergence, we introduce a new operation
on multifractions to represent inverses.
e = ∅, and, for every n-multifraction a,
Notation 3.7. We put ∅
(
/an /···/a1 if n is positive in a,
(3.8)
e
a :=
an /···/a1
if n is negative in a.
Lemma 3.9. (i) For all multifractions a, b, we have ag
· b = eb · e
a.
a) = ι(1 · e
a) = ι(a)−1 .
(ii) For every multifraction a, we have ι(e
Proof. (i) If a written as a sequence in M ∪ M is (x1 , ..., xn ), then e
a is (xn , ..., x1 ) (where we
· b = eb · e
a directly follows from the definition of the product.
put a = a for a in M ), and then ag
Point (ii) comes from (2.6).
Lemma 3.9(ii) says that, if a represents an element g of U(M ), then both e
a and 1 · e
a represent g −1 . By definition, ke
ak = kak always holds, and the operation e on FM± is involutive.
a is posHowever, this operation does not restrict to FM : if a is a positive multifraction, then e
itive if and only if kak is even. In order to represent inverses inside FM , one can compose e
with a left translation by 1, thus representing the inverse of a by 1 · e
a. The inconvenience is
that involutivity is lost: for kak odd and b = 1 · e
a, we find 1 · eb = 1 · a 6= a.
10
PATRICK DEHORNOY
Lemma 3.10. If M is a gcd-monoid and RM is semi-convergent, then a ≃± b is equivalent
to 1 · a · eb ⇒∗ 1 for all a, b in FM± .
Proof. As we have 1 ≃± ∅ and b · eb ≃± ∅ by Lemma 3.9, a ≃± b is equivalent to 1 · a · eb ≃± ∅,
hence, by (2.6), to 1 · a · eb being unital, i.e., to ι(1 · a · eb) = 1. By construction, 1 · a · eb lies in FM .
So, if RM is semi-convergent, ι(1 · a · eb) = 1 is equivalent to 1 · a · eb ⇒∗ 1.
As a direct application, we obtain
Proposition 3.11. If M is a gcd-monoid and RM is semi-convergent, then M embeds in its
enveloping group U(M ).
Proof. Assume a, b ∈ M and ι(a) = ι(b), i.e., a ≃± b. By Lemma 3.10, we must have a · eb ⇒∗ 1,
which is a/b ⇒∗ 1/1. By definition of reduction, this means that there exists x in M satisfying
a/b • R1,x = 1/1. This implies a = b ( = x ).
The word problem for U(M ). If S is any set, we denote by S ∗ the free monoid of all words in S,
using ε for the empty word. For representing group elements, we consider words in S ∪ S, where
S is a disjoint copy of S consisting of one letter s for each letter s of S, due to represent s−1 . If
w is a word in S ∪ S, we denote by w the signed word obtained from w by exchanging s and s
and reversing the order of letters. If M is a monoid, S is included in M , and w is a word in S,
we denote by [w]+ the evaluation of w in M . We extend this notation to words in S ∪ S by
defining [w]+ to be the multifraction [w1 ]+ /···/[wn ]+ , where (w1 , ..., wn ) is the unique sequence
of words in S such that w can be decomposed as w1 w2 w3 w4 ··· with wi 6= ε for 1 < i 6 n.
Lemma 3.12. [13, Lemma 2.5] For every monoid M , every generating family S of M , and
every word w in S ∪ S, the following are equivalent:
(i) The word w represents 1 in U(M );
(ii) The multifraction [w]+ satisfies [w]+ ≃± 1 in FM± .
Thus solving the word problem for the group U(M ) with respect to the generating set S
amounts to deciding the relation [w]+ ≃± 1, which takes place in FM± , hence essentially inside M ,
as opposed to U(M ).
A few more definitions are needed. First, a gcd-monoid M is called strongly noetherian if
there exists a map λ : M → N satisfying, for all a, b in M ,
(3.13)
λ(ab) > λ(a) + λ(b),
and λ(a) > 0 for a 6= 1.
This condition is stronger than noetherianity, but it still follows from the existence of a presentation by homogeneous relations (same length on both sides): in this case, the word length
induces a map λ as in (3.13). So every Artin-Tits monoid is strongly noetherian.
Next, we need the notion of a basic element. Noetherianity implies the existence of atoms,
namely elements that cannot be expressed as the product of two non-invertible elements. One
shows [16, Corollary II.2.59] that, if M is a noetherian gcd-monoid, then a subfamily S of M
generates M if and only if it contains all atoms of M .
Definition 3.14. [11] If M is a noetherian gcd-monoid, an element a of M is called right basic
if it belongs to the smallest family X that contains the atoms of M and is such that, if a, b
belong to X and a ∨ b exists, then the element a′ defined by a ∨ b = ba′ still belongs to X.
Left-basic elements are defined symmetrically. We say that a is basic if it is right or left basic.
Note that, in the above definition, nothing is required when a ∨ b does not exist. The key
technical result is as follows:
Lemma 3.15. [13, Prop 3.27] If M is a strongly noetherian gcd-monoid with finitely many
basic elements and atom set S, then the relation [w]+ ⇒∗ 1 on (S ∪ S)∗ is decidable.
This result is not trivial, because deciding whether a multifraction is eligible for some reduction requires to decide whether two elements of the ground monoid admit a common multiple,
MULTIFRACTION REDUCTION II
11
and this is the point, where the finiteness of the number of basic elements occurs crucially, as
it provides an a priori upper bound on the size of this possible common multiple. Then, we
immediately deduce:
Proposition 3.16. If M is a strongly noetherian gcd-monoid with finitely many basic elements
and RM is semi-convergent, then the word problem for U(M ) is decidable.
Proof. Let S be the atom set of M . By Lemma 3.15, the relation [w]+ ⇒∗ 1 on words in S ∪ S
is decidable. By (3.4) (and by Lemma 2.17(i)), [w]+ ⇒∗ 1 is equivalent to [w]+ ≃± 1. Finally,
by Lemma 3.12, [w]+ ≃± 1 is equivalent to w representing 1 in U(M ). Hence the latter relation
is decidable.
Note that, because the multifraction [w]+ is always defined to be positive, we only need
semi-convergence for RM in the above argument.
In the particular case of Artin-Tits monoids, we deduce
Corollary 3.17. If Conjecture A is true, then the word problem for every Artin-Tits group is
decidable.
Proof. Let M be an Artin-Tits monoid. We noted that M is a gcd-monoid [5], and that it is
strongly noetherian. Next, M has finitely many basic elements: this follows from (and, actually,
is equivalent to) the result that every Artin-Tits monoid has a finite Garside family [15, 21].
Hence M is eligible for Proposition 3.16.
Let us conclude with algorithmic complexity. Lemma 3.15 says nothing about the complexity
of reduction. We show now the existence of an upper bound for the number of reductions.
Lemma 3.18. If M is a strongly noetherian gcd-monoid with finitely many basic elements,
then the number of reduction steps from an n-multifraction a is at most Fn (λ(a1 ), ..., λ(an )),
where λ satisfies (3.13), C is the maximum of λ(a) + 1 for a basic in M , and Fn is inductively
defined by F1 (x) = x + 2 and Fn (x1 , ..., xn ) = (x1 + 1)C Fn−1 (x2 ,... ,xn) .
Proof. An easy induction shows that the function Fn is increasing with respect to each variable
and, for every 1 6 i < n, it satisfies the inequality
(3.19)
Fn (Cx1 , ..., Cxi , xi+1 − 1, xi+2 , ..., xn ) < Fn (x1 , ..., xn ).
For a an n-multifraction, write T (a) for the maximal number of reduction steps from a and λ(a)
for (λ(a1 ), ..., λ(an )). We prove using induction on ⇒ the inequality T (a) 6 Fn (λ(a)) for every
n-multifraction a. Assume a • Ri,x = b with x an atom of M (what can assumed without loss
of generality). We compare the sequences λ(a) and λ(b). By definition, bi+1 is a proper divisor
of ai+1 , which implies λ(bi+1 ) < λ(ai+1 ). Next, ai is the product of at most λ(ai ) basic elements
of M , hence so is bi , implying λ(bi ) 6 Cλ(ai ). Finally, ai−1 is the product of at most λ(ai−1 )
basic elements of M , hence bi−1 is the product of at most λ(ai−1 ) + 1 basic elements, implying
λ(bi−1 ) 6 Cλ(ai−1 ). Then the induction hypothesis implies T (b) 6 Fn (b), so, plugging the
upper bounds for bi , and using that Fn is increasing and (3.19), we find
T (b) 6 Fn (Cλ(a1 ), ..., Cλ(ai ), λ(ai+1 )−1, λ(ai+2 ), ..., λ(an )) < Fn (λ(a1 ), ..., λ(an )) = Fn (λ(a)),
and T (a) 6 T (b) + 1 implies T (b) 6 Fn (b).
The upper bound of Lemma 3.18 is not polynomial (very far from statistical data, which
suggest a quadratic bound), but it is not very high either in the hierarchy of fast growing
functions (it is “primitive recursive”). From there, one can easily deduce a similar upper bound
(tower of exponentials) for the word problem for U(M ) when RM is semi-convergent.
12
PATRICK DEHORNOY
Property H. One says [12, 17, 26] that Property H is true for a presentation (S, R) of a monoid M
if a word w in S ∪ S represents 1 in U(M ) if and only if the empty word can be obtained
from w using special transformations, namely positive and negative equivalence and left and
right reversing. Positive equivalence means replacing a positive factor of w (no letter s) with an
R-equivalent word, negative equivalence means replacing the inverse of a positive factor with
the inverse of an R-equivalent word, whereas right reversing consists in deleting some length
two factor ss or replacing some length two factor st with vu such that sv = tu is a relation
of R, and left reversing consists in deleting some length two factor ss or replacing some length
two factor st with uv such that vs = ut is a relation of R. Roughly speaking, Property H
says that a word representing 1 can be transformed into the empty word without introducing
new trivial factors ss or ss, a situation directly reminiscent of Dehn’s algorithm for hyperbolic
groups, see [17, Section 1.2].
Say that a presentation (S, R) of a monoid M is a right lcm presentation if R consists of
one relation su = tv for each pair of generators s, t that admit a common right multiple, with
su and tv representing s ∨ t. The standard presentation of an Artin-Tits monoid is a right lcm
presentation, and, symmetrically, a left lcm presentation.
Proposition 3.20. If M is a gcd-monoid and RM is semi-convergent, Property H is true for
every presentation of M that is an lcm presentation on both sides.
The point is that applying a rule Ri,x to a multifraction [w]+ can be decomposed into a
sequence of special transformations as defined above. The argument is the same as in the case
±
when RM
is convergent [13, Proposition 5.19], and we do not repeat it.
Thus Conjecture A would imply the statement conjectured in [12]:
Corollary 3.21. If Conjecture A is true, Property H is true for every Artin-Tits presentation.
3.4. Alternative forms. Here we mention several variants of semi-convergence.
±
Proposition 3.22. If M is a noetherian gcd-monoid, then RM
(resp., RM ) is semi-convergent
±
if and only if for every a in FM (resp., FM ),
(3.23)
If a is unital, then a is either trivial or reducible.
±
Proof. Assume that RM
is semi-convergent, and let a be a nontrivial unital multifraction in FM± .
∗
By definition, a ⇒ 1 holds. As a is nontrivial, the reduction requires at least one step, so a
cannot be R-irreducible. Hence, (3.23) is satisfied.
±
Conversely, assume (3.23). As M is noetherian, the rewrite system RM
is terminating, i.e.,
the relation ⇒ admits no infinite descending sequence. Hence we can use induction on ⇒ to
establish (3.4). So let a be a unital multifraction in FM± . If a is ⇒-minimal, i.e., if a is Rirreducible, then, by (3.23), a must be trivial, i.e., we have a = 1, whence a ⇒∗ 1. Otherwise,
a is R-reducible, so there exist i, x such that b = a • Ri,x is defined. By construction, b is ≃± equivalent to a, hence it is unital. By the induction hypothesis, we have b ⇒∗ 1. By transitivity
±
of ⇒∗ , we deduce a ⇒∗ 1. Hence RM
is semi-convergent.
The proof is similar for RM .
Condition (3.23) can be restricted to more special unital multifractions.
Definition 3.24. Call a multifraction a prime if, for every i that is positive (resp., negative)
in a, the entries ai and ai+1 admit no nontrivial common right (resp., left) divisor.
Since dividing adjacent entries by a common factor is a particular case of reduction, an
R-irreducible multifraction must be prime. The converse need not be true: for instance, the
6-multifraction b of Example 3.6 is prime, and it is R-reducible.
±
Proposition 3.25. If M is a noetherian gcd-monoid, then RM
(resp., RM ) is semi-convergent
±
if and only if for every a in FM (resp., FM ),
(3.26)
If a is unital and prime, then a is either trivial or reducible.
MULTIFRACTION REDUCTION II
13
Proof. By Proposition 3.22, the condition is necessary, since (3.26) is subsumed by (3.23).
For the converse implication, assume (3.26). As for Proposition 3.22, we establish (3.4) using
induction on ⇒. Let a be a unital multifraction in FM± . If a is R-irreducible, then it must be
prime, for, otherwise, it is eligible for at least one division, which is a special case of reduction.
Hence, a must be 1 by (3.26). Otherwise, a is R-reducible, there exist i, x such that b = a • Ri,x is
±
is semi-convergent.
defined, the induction hypothesis implies b ⇒∗ 1, hence a ⇒∗ 1. Hence RM
The proof for RM is similar.
Corollary 3.27. Conjecture A is true if and only if (3.26) holds for every Artin-Tits monoid M
and every a in FM .
We turn to another approach. Whenever the ground monoid M is noetherian, the rewrite
±
systems RM
and RM are terminating, hence they are convergent if and only if they are confluent,
meaning that
(3.28)
If we have a ⇒∗ b and a ⇒∗ c, there exists d satisfying b ⇒∗ d and c ⇒∗ d
(“diamond property”). We now observe that semi-convergence is equivalent to a weak form of
confluence involving the unit multifractions 1.
±
Proposition 3.29. If M is a gcd-monoid, then RM
(resp., RM ) is semi-convergent if and only
±
if for every a in FM (resp., FM ),
(3.30)
The conjunction of a ⇒∗ b and a ⇒∗ 1 implies b ⇒∗ 1.
Relation (3.30) can be called 1-confluence for a, since it corresponds to the special case c = 1
of (3.28): indeed, (3.28) with c = 1 claims the existence of d satisfying b ⇒∗ d and 1 ⇒∗ d, and,
as 1 is R-irreducible, we must have d = 1, whence b ⇒∗ 1, as asserted in (3.30). In order to
establish Proposition 3.29, we need an auxiliary result, which connects ≃± with the symmetric
closure of ⇒∗ and is a sort of converse for Lemma 2.17.
Lemma 3.31. If M is a gcd-monoid and a, b belong to FM± , then a ≃± b holds if and only if
there exist a finite sequence c0 , ..., c2r in FM± and p, q in Z satisfying
(3.32)
a · 1p = c0 ⇒∗ c1 ∗⇐ c2 ⇒∗ ···
∗
⇐ c2r = b · 1q .
b ∗ b if a· ⇒∗ b · 1p holds for some p. By Lemma 2.17, a ⇒∗ b
Proof. For a, b in FM± , write a ⇒
∗
b ∗ is transitive. It is also compatible with
implies a · 1r ⇒ b · 1r and, therefore, the relation ⇒
multiplication: on the left, this follows from Lemma 2.17 directly. On the right, a ⇒∗ b · 1p
implies a · c ⇒∗ b · 1p · c for every c, and we observe that 1p · c ⇒∗ c · 1p always holds. Hence, the
b ∗ is a congruence on FM± . As we have 1 ⇒∗ ∅ · 1, a/a ⇒∗ ∅ · 12 and
symmetric closure ≈ of ⇒
∗
/a/a ⇒ ∅ · 1−2 for every a in M , the congruence ≈ contains pairs that generate ≃± . Hence
b ∗ and its inverse connecting a to b. Taking the
a ≃± b implies the existence of a zigzag in ⇒
maximum of |r| for 1r occurring in the zigzag, one obtains (3.32).
±
Proof of Proposition 3.29. Assume that RM
is semi-convergent, and we have a ⇒∗ b and a ⇒∗ 1.
±
is semi-convergent, this
By Lemma 2.17, we have a ≃± b and a ≃± 1, hence b ≃± 1. As RM
±
implies b ⇒∗ 1. So (3.30) is satisfied, and RM is 1-confluent.
±
Conversely, assume that RM
is 1-confluent. We first show using induction on k that, when
0
∗ 1 ∗
2
we have a zigzag c ⇒ c ⇐ c ⇒∗ c3 ∗⇐ c4 ⇒∗ ···, then c0 = 1 implies ck ⇒∗ 1 for every k.
For k = 0, this is the assumption. For k even non-zero, we obtain ck ⇒∗ ck−1 ⇒∗ 1 using the
induction hypothesis, whence ck ⇒∗ 1 by transitivity of ⇒∗ . For k odd, we have ck−1 ⇒∗ 1 by
the induction hypothesis and ck−1 ⇒∗ ck , whence ck ⇒∗ 1 by 1-confluence.
Now assume that a is unital. Lemma 3.31 provides p, q, r and c0 , ..., c2r satisfying
1p = c0 ⇒∗ c1 ∗⇐ c2 ⇒∗ ···
∗
⇐ c2r = a · 1q .
±
As shown above, we deduce a · 1q ⇒∗ 1, whence a ⇒∗ 1 by Lemma 2.17. Hence RM
is semiconvergent.
Once again, the proof for FM is the same.
14
PATRICK DEHORNOY
Corollary 3.33. Conjecture A is true if and only if (3.30) holds for every Artin-Tits monoid M
and every a in FM .
Proposition 3.29 is important for testing Conjecture A, because it shows that, if a ⇒∗ 1
holds, then every sequence of reductions from a inevitably leads to 1. In other words, any
reduction strategy may be applied without loss of generality.
4. Divisions and tame reductions
When reduction is not convergent, it is not confluent either, and a multifraction may admit
several reducts with no subsequent common reduct. However, by restricting to particular
reductions, we can retrieve a (weak) form of confluence and let distinguished reducts appear.
This is the approach we explore in this section. We start in Subsection 4.1 with divisions,
which are particular reductions with good, but too weak properties. Then, in Subsection 4.2,
we extend divisions into what we call tame reductions, which are those reductions that, in
a sense, exclude no subsequent opportunities. Extending the example of divisions to tame
reductions leads us in Subsection 4.3 to the natural notion of a maximal tame reduction and
to Conjecture B about tame reductions from unital multifractions, which is stronger but more
precise than Conjecture A.
We feel that the many technical details, examples, and counter-examples appearing in this
section and the next one are important, because they illustrate how subtle the mechanism of
reduction is. Skipping such details would induce a superficial view and misleadingly suggest
that things are more simple than they really are, possibly leading to naive attempts with no
chance of success.
4.1. Divisions. Divisions are the most direct counterparts of free reductions in free monoids.
They are the special cases of reduction when no remainder appears. No confluence result can be
expected for divisions in a non-free monoid, but we shall see in Proposition 4.4, the main result
of this subsection, that, for every multifraction a, there exists a unique, well-defined maximal
reduct ∂a accessible from a by divisions.
Following the model of reductions, we first fix notation for divisions.
Definition 4.1. If M is a gcd-monoid and a, b belong to FM± , we declare that b = a • Di,x holds
if we have b = a • Ri,x and, in addition, x right (resp., left) divides ai if i is positive (resp.,
±
negative) in a. We use DM
for the family of all Di,x with x 6= 1, write a ⇒div b if some rule
±
∗
for the reflexive–transitive closure of ⇒div .
of DM maps a to b, and ⇒div
So a • Di,x is defined if and only if x divides ai and ai+1 on the due side, and applying Di,x
means dividing ai and ai+1 by x. By definition, a is D-irreducible if and only if it is prime
(Definition 3.24), i.e., the gcds of adjacent entries (on the relevant side) are trivial.
±
Except in degenerated cases, e.g., free monoids, the system DM
is not convergent: typically,
+
for M = ha, b | aba = babi (Artin’s 3-strand braid monoid) and a = a/aba/b, we find
a • D2,b = a/ab/1 and a • D1,a = 1/ab/b, with no further division, and confluence can be
restored only at the expense of applying some reduction Ri,x , here a • D2,b = a • D1,a R2,b .
However, we shall see now that, for every multifraction a, there exists a unique, well-defined
maximal R-reduct of a that can be obtained using divisions.
The first step is to observe that, for each level i, there is always a maximal division at level i,
namely dividing by the gcd of the ith and (i + 1)st entries (on the due side). Indeed, assuming
for instance i positive in a, the multifraction a • Di,x is defined if and only if x right divides
e ai+1 .
both ai and ai+1 , hence if and only if x right divides the right gcd ai ∧
Notation 4.2. If M is a gcd-monoid and a is a multifraction on M , then, for i < kak, we write
a • Dimax for a • Di,x with x the gcd of ai and ai+1 on the due side.
Next, we observe that, contrary to general irreducibility, (local) primeness is robust, in that,
once obtained, it cannot be destroyed by subsequent divisions:
MULTIFRACTION REDUCTION II
15
Lemma 4.3. Say that a multifraction a is j-prime if a • Dj,y is defined for no y 6= 1. If a is
j-prime, then so is a • Di,x for all i, x.
Proof. Assume for instance j positive in a, and let b = a • Di,x . We have either bj = aj (for
i < j − 1 and i > j) or bj 6 aj (for i = j − 1); similarly, we have either bj+1 = aj+1 (for i < j
e ai+1 = 1
and i > j + 1) or bj+1 6 aj+1 (for i = j + 1). Hence, in all cases, the assumption ai ∧
e
implies bi ∧ bi+1 = 1.
Hence, if we start with a multifraction a and apply, in any order, maximal divisions Dimax in
such a way that every level between 1 and kak − 1 is visited at least one, we always finish with
a prime multifraction. The latter may depend on the order of the divisions, but we shall now
see that there exists a preferred choice.
Proposition 4.4. Let M be a gcd-monoid. For every n-multifraction a on M , put
max
max
∂a := a • Dn−1
Dn−2
···D1max .
(4.5)
Then, ∂a is prime, and, for every multifraction b on M ,
∗
a ⇒div
b
(4.6)
implies
∗
b ⇒div
∂b ⇒∗ ∂a.
Thus ∂a is a reduct of every multifraction obtained from a using division. The proof of
Proposition 4.4 is nontrivial and requires to precisely control the way divisions and reductions
can be commuted. We begin with a confluence result. By [13, Lemma 4.6], there always exists
a confluence solution for any two reductions at level i and i + 1. This applies of course when
one of the reductions is a division, but, in that case, we can say more.
Lemma 4.7. Assume that both a • Ri+1,x and a • Di,y are defined. Then we have a • Ri+1,x Di,z =
a • Di,y Ri+1,x , where z is determined by the equalities ai = ay, x ∨ a = av, and vz = v ∨ y
e a = va, and zv = v ∨
e y) for i positive (resp., negative) in a. Moreover, if y
(resp., ai = ya, x ∨
is maximal for a (i.e., y is the gcd of ai and ai+1 ), then z is maximal for a • Ri+1,x .
Proof. (Figure 2) Assume that i is positive in a, so ai+1 is negative in a. Put b := a • Ri+1,x
and c := a • Di,y . By definition, there exists x′ satisfying
bi−1 = ai−1 , bi = ai x′ , xbi+1 = ai+1 x′ = x ∨ ai+1 , xbi+2 = ai+2 ,
ci−1 = ai−1 , ci y = ai , ci y = ai , ci+2 = ai+2 .
As ai+1 is ci+1 y, Lemma 2.11 implies the existence of u, v, and z satisfying
(4.8)
bi+1 = uz
with
ci+1 v = xu = x ∨ ci+1
and yx′ = vz = y ∨ v.
By construction, we have bi = ai x′ = ci yx′ = ci vz, which shows that z right divides both bi
and bi+1 . Hence d := b • Di,z is defined, and we have
(4.9)
di−1 = ai−1 ,
di = cv ,
di+1 = u,
di+2 = bi+2 .
On the other hand, by assumption, x left divides ci+2 , which is ai+2 , and x and ci+1 admit a
common right multiple, namely their right lcm xu. Hence, c • Ri,x is defined, and comparing
with (4.9) directly yields the expected equality c • Ri,x = b • Di,z .
e ai+1 .
It remains to prove that, if y is maximal for a, then z is maximal for b. So assume y = ai ∧
e
We deduce ci ∧ ci+1 = 1. On the other hand, by Lemma 3.1, the assumption ci+1 v = ci+1 ∨ x
e v = 1. Then (the symmetric counterpart of) Lemma 2.13 implies ci v ∧
e u = 1, whence
implies u ∧
e
bi ∧ bi+1 = z.
A symmetric argument applies when i is negative in a.
Next, we observe that reduction and division commute when performed at distant levels: the
result is easy for very distant levels, slightly more delicate when the levels are closer.
Lemma 4.10. Assume b = a • Ri,x , and j 6= i, i + 1, i − 2 (resp., j = i − 2). Put a′ := a • Djmax
max
and b′ := b • Djmax . Then we have b′ = a′ • Ri,x (resp., b′ = a′ • Ri,x Di−2
).
16
PATRICK DEHORNOY
ai+2
bi+2
x
ci
ci+1 a
v
y
ai−1
ai
u
bi+1
z
x′
bi
Figure 2. Proof of Lemma 4.7: if a is eligible both for Ri+1,x and Di,y , we can
start with either and converge to the colored path.
Proof. For j 6 i−3 or j > i+2, the reduction Ri,x does not change the jth and (j +1)st entries,
so the greatest division at level j remains the same, and commutation is straightforward. For
j = i − 1, Lemma 4.7 gives b′ = a′ • Ri,x .
Assume j = i − 2. Then the reduction Ri,x does not change the jth entry, but it possibly
increases the (j + 1)st entry. So, if Dj,y is the maximal j-division for a, then Dj,y applies
to b, but it need not be the maximal j-division for b. Expanding the definitions, we obtain the
commutation relation a • Ri,x Dj,y = a • Dj,y Ri,x , meaning b • Dj,y = a′ • Ri,x , together with
max
, as expected.
b′ = (b • Dj,y )Djmax , whence b′ = a′ • Ri,x Di−2
The last preliminary result, needed for the end of the proof of Proposition 4.4, connects ∂a
and ∂b in the (very special) case when b is an elementary reduct of a and a is prime at every
sufficiently large level.
Lemma 4.11. If a and b are j-prime for j > i, then b = a • Ri+1,x implies ∂a ⇒∗ ∂b.
max
max
Proof. By assumption, we have ∂a = a • Di−1
···D1max and ∂b = b • Di−1
···D1max . Put ai := a,
i
j
j+1
j
j+1
max
max
• D
,b =b
• D
for j decreasing from i − 1 to 1,
b := b and, inductively, a = a
j
j
1
1
yielding ∂a = a and ∂b = b . We prove using induction on j decreasing from i to 1 that, for
every y, there exist x0 , x1 , ..., xk with i−2k > j satisfying bj = aj • Ri+1,x0 Ri−1,x1 ···Rx−2k+1,xk .
By assumption, the property is true for j = i, with k = 0 and x0 = x. Assume i > j > 1. By
induction hypothesis, we have bj+1 = aj+1 • Ri+1,x0 Ri−1,x1 ···Ri−2k+1,xk for some x0 , ..., xk with
i−2k > j+1. By repeated applications of Lemma 4.10, we deduce that each reduction Ri−2ℓ+1,xℓ
commutes with Djmax , except the last one in the case i − 2k = j + 2, in which case Lemma 4.10
prescribes to add one more reduction (a division) Ri−2k−1,xk+1 . In this way, we obtain either
bj = aj • Ri+1,x0 Ri−1,x1 ···Ri−2k+1,xk , or bj = aj • Ri+1,x0 Ri−1,x1 ···Ri−2k+1,xk Ri−2k−1,xk+1 , and
the induction continues.
We can now complete the argument for Proposition 4.4. The proof that ∂a is prime is easy,
but that of the relation (4.4) is more delicate.
max
Proof of Proposition 4.4. It follows from the definition that a • Dn−1
is (n − 1)-prime, then that
max
max
a • Dn−1 Dn−2 is (n − 1)- and (n − 2)-prime, etc., hence that ∂a is i-prime for 1 6 i < n, hence
it is prime.
∗
b implies ∂b ⇒∗ ∂a. For an induction, it is
We now establish (4.6), i.e., prove that a ⇒div
∗
sufficient to prove that a ⇒div b implies ∂b ⇒ ∂a. So, we assume b = a • Di,z , and aim at
∗
∂a. Put kak = n. By definition, ∂a and ∂b are obtained by performing n − 1
proving ∂b ⇒div
successive divisions, and we shall establish a step-by-step connection summarized in Figure 3.
Put an := a, bn := b, and let ai (resp., bi ) be obtained from ai+1 (resp., bi+1 ) by applying Dimax ,
so that we finally have ∂a = a1 and ∂b = b1 . We assume that i is positive in a.
Consider first j > i + 2. Applying Lemma 4.10, we inductively obtain bj = aj • Di,z , implying
the commutativity of the n − i − 2 left hand squares in the diagram of Figure 3.
MULTIFRACTION REDUCTION II
17
Now let x = ai+1 ∧ ai+2 . By definition, we have ai+1 = ai+2 • Di+1,x . By Lemma 4.7, there
exists c and z ′ satisfying c = ai+2 • Di+1,x Di,z′ = ai+2 • Di,z Ri+1,x , which reads c = ai+1 Di,z′ =
bi+2 • Ri+1,x .
Next, c is obtained from ai+1 by some i-division, whereas ai is obtained from ai+1 by the
maximal i-division, hence ai must be obtained from c by some further i-division, namely the
maximal i-division for c. So, ai = c • Dimax holds.
On the other hand, bi+1 is obtained from bi+2 by the maximal (i + 1)-division, namely Di+1,y
with y = bi+1 ∧ bi+2 . As we have x = ai+2 ∧ ai+2 and bi+2 = ai+2 , the relation bi+1 6 ai+1
implies y 6 x, say x = yx′ . If follows that reducing x at level i + 1 in bi+2 amounts to first
dividing by y and then reducing x′ , i.e., we have b • Ri+1,x = b • Di+1,y Ri+1,x′ = bi+1 • Ri+1,x′ .
Now, two reductions apply to bi+1 , namely Ri+1,x′ , which leads to c, and Dimax , which is,
say, Di+1,y and leads to bi . Applying Lemma 4.7 again, we obtain the existence of y ′ satisfying
bi+1 • Ri+1,x′ Di,y′ = bi+1 • Di,y Ri+1,x′ , which boils down to c • Di,y′ = bi • Ri+1,x′ . Moreover, as
Di,y is the maximal i-division applying to bi+1 , Lemma 4.7 implies that Di,y′ is the maximal
i-division applying to c. We obtained above ai = c • Dimax , so we deduce ai = bi • Ri+1,x′ .
From there, we are in position for applying Lemma 4.11 (with a and b interchanged): we have
ai = bi • Ri+1,x′ , and, by construction, ai and bi are j-prime for every j > i. Then Lemma 4.11
ensures b1 ⇒∗ a1 , which is ∂b ⇒∗ ∂a.
an
max
Dn−1
an−1
Di,z
b
n
max
Dn−1
ai+2
Di,z
b
n−1
max
Di+1
ai+1
Di,z
b
Dimax
ai
Di,z′
i+2
c
Ri+1,x
Dimax
a
i
Ri+1,x′
b
i+2
i+1
max
Di+1
b
Dimax
max
Di−1
ai−1
Ri+1,x′
b
i
max
Di−1
a2
D1max
Lemma 4.11
b
i−1
b
2
D1max
∂a
∗
∂b
Figure 3. Comparing the computations of ∂a and ∂(a • Di,x ).
Let us denote by Irr(a) the family of all R-irreducible reducts of a. The failure of confluence
means that Irr(a) may content more than one element, and controlling Irr(a) is one of the main
challenges in the current approach. As, for every multifraction a, we have now a distinguished
∗
reduct ∂a, a natural task is to compare Irr(∂a) with Irr(a). By definition, a ⇒div
∂a implies
∗
Irr(∂a) ⊆ Irr(a), and, by Proposition 4.4, a ⇒div b implies b ⇒ ∂a, whence Irr(∂a) ⊆ Irr(∂b).
The next examples show these easy inclusions are the best we can expect in general.
e 2 , let a = ab/aba/aca. Then one finds
Example 4.12. In the Artin-Tits monoid of type A
Irr(a) = {a1 , a2 }, with a1 = a • D2,a = ab/ba/ca and a2 = a • D1,ab R2,c = cb/bc/ac. Now,
we obtain ∂a = a1 , whence Irr(∂a) = {a1 }: so, by performing divisions, we lost one of the
irreducible reducts of a. On the other hand, for b = a • D1,ab = 1/b/aca, we find ∂b = b and
Irr(∂b) = Irr(b) = {a1 , a2 } = Irr(a): so a ⇒div b does not imply Irr(∂b) ⊆ Irr(∂a).
Remark 4.13. The order of divisions is important in the definition of ∂a and, even at the
expense of using reductions instead of divisions, we cannot start from low levels in general.
Indeed, for kak = 3, Lemma 4.7 implies ∂a = a • R1,x1 R2,x2 , with xi the (relevant) gcd of ai
Q
and ai+1 , but this expression of ∂a as a • i=kak−1
Ri,xi with xi gcd of ai and ai+1 does not work
i=1
for kak > 4: for instance, for a := a/a/a/a, one finds a • R1,x1 R2,x2 R3,x3 = a/a/1/1 6= ∂a = 1.
4.2. Tame reductions. What motivates studying divisions specifically is that the latter satisfy
a form of confluence: by Lemma 4.7 and the results of [13], if a multifraction a is eligible for a
division Di,x and for another reduction Rj,y , a common reduct for a • Di,x and a • Rj,y always
exists. Moreover, we saw in Proposition 4.4 that there always exists a maximal div-reduct
with good compatibility properties of the associated operator ∂. However, because many prime
18
PATRICK DEHORNOY
multifractions are not irreducible, and, in particular, many prime unital multifractions are not
trivial, it is hopeless to analyze reduction in terms of divisions exclusively, making it natural to
try to extend the family of divisions so as to preserve its main property, namely “guaranteed
confluence”. This leads to tame reductions, and the main result here is that, exactly as in
the case of divisions, there exists for each multifraction a and each level i a maximal tame
i-reduction applying to a.
Definition 4.14. If a is a multifraction, we say that x is an i-reducer for a if a • Ri,x is defined;
we then say that an i-reducer x is tame for a if, for all j, y such that a • Rj,y is defined, a • Ri,x
and a • Rj,y admit a common reduct; otherwise, x is called wild for a.
Thus x is a tame i-reducer for a if reducing x in a leaves all possibilities for further reductions
open, whereas reducing a wild i-reducer excludes at least one subsequent confluence.
±
Example 4.15. If M is a gcd-monoid satisfying the 3-Ore condition, the system RM
is confluent
and, therefore, every reducer is tame for every multifraction it applies to. By contrast, in the
e 2 , for a = 1/c/aba, both a and b are 2-reducers for a, but a • R2,a and
Artin-Tits of type A
a • R2,b admit no common reduct, hence a and b are wild 2-reducers for a.
To prove the existence of the maximal tame i-reducer in Proposition 4.18 below, we shall
use convenient characterizations of tame reducers established in Lemmas 4.16 and 4.17.
Lemma 4.16. If M is a gcd-monoid, a is a multifraction on M , and a • Ri,x is defined, then
x is a tame i-reducer for a if and only if for i positive (resp., negative) in a, the elements x, y,
and ai admit a common right (resp., left) multiple whenever a • Ri,y is defined.
Proof. Assume that x is a tame i-reducer for a, and let y be an i-reducer for a. By definition,
a • Ri,x and a • Ri,y admit a common reduct, which is necessarily of the form a • Ri,z for some z.
Then there exist u, v satisfying
a • Ri,z = (a • Ri,x ) • Ri,u = (a • Ri,y ) • Ri,v .
Assuming i negative in a, we deduce z = xu = yv. Hence z is a right multiple of x ∨ y and,
therefore, a • Ri,x∨y is defined as well, implying that ai , x, and y admit a common right multiple.
The argument is symmetric when i is positive in a.
Conversely, assume that x is an i-reducer for a and, for every i-reducer y, the elements ai ,
x, and y admit a common multiple, say a common right multiple, assuming that i is negative
in a. Then x ∨ y left divides ai+1 since x and y do, and ai and x ∨ y admit a common right
multiple. Hence a • Ri,x∨y is defined. Then, writing x ∨ y = xy ′ = yx′ , we have
a • Ri,x∨y = (a • Ri,x ) • Ri,y′ = (a • Rj,y ) • Ri,x′ ,
which shows that a • Ri,x and a • Ri,y admit a common reduct. On the other hand, for j 6= i,
Lemmas 4.18 and 4.19 from [13] imply that a • Ri,x and a • Rj,y always admit a common reduct.
Therefore, x is a tame i-reducer for a.
If M is a noetherian gcd-monoid, every nonempty family X of left divisors of an element a
necessarily admits <-maximal elements, i.e., elements z such that there is no x with z < x in
e
the family: take z so that z −1 a is <-minimal
in {x−1 a | x ∈ X}. Hence, in particular, for every
multifraction a and every level i, there exist maximal i-reducers for a.
Lemma 4.17. If M is a noetherian gcd-monoid and a is a multifraction on M , an i-reducer x
for a is tame if and only if x divides every maximal i-reducer for a.
Proof. Assume that x is a tame i-reducer for a, and y is a maximal i-reducer for a. By
Lemma 4.16, ai , x, and y admit a common multiple, hence an lcm, and, therefore, the lcm of x
and y is again an i-reducer for a. The assumption that y is maximal implies that this lcm is y,
i.e., that x divides y (on the due side).
Conversely, assume that x is an i-reducer for a that divides every maximal i-reducer. Let y
be an arbitrary i-reducer for a. As M is noetherian, y divides at least one maximal i-reducer,
MULTIFRACTION REDUCTION II
19
say z. By assumption, x divides z, hence so does the lcm of x and y. Since z is an i-reducer
for a, so is its divisor the lcm of x and y and, therefore, a • Ri,x and a • Ri,y admit a common
reduct. Hence x is tame for a.
Proposition 4.18. If M is a noetherian gcd-monoid and a is a multifraction on M , then, for
every i < kak, there exists a unique greatest tame i-reducer for a, namely the gcd of all maximal
i-reducers for a. The latter is a multiple of the gcd of ai and ai+1 .
Proof. Let z be the gcd (on the relevant side) of all maximal i-reducers for a. Then z, and
every divisor x of z, divides every maximal i-reducer for a, hence, by Lemma 4.17, it is a tame
i-reducer for a.
Conversely, if x is a tame i-reducer for a, then, by Lemma 4.17, x divides every maximal
i-reducer for a, hence it divides their gcd z. Hence z is the greatest tame i-reducer for a.
Finally, assume that x divides ai and ai+1 . Then a • Di,x , hence a fortiori a • Ri,x is defined,
so x is an i-reducer for a. Let y be any i-reducer for a. Then the lcm of ai and y is a common
multiple of ai , x, and y. Hence, by Lemma 4.16, x is a tame i-reducer for a. This applies in
particular when x is the gcd of ai and ai+1 .
On the shape of what we did with divisions, we introduce
Definition 4.19. If M is a noetherian gcd-monoid and a is a multifraction on M , then, for
i < kak, the unique element x whose existence is stated in Proposition 4.18 is called the greatest
tame i-reducer for a; we then write a • Rimax for a • Ri,x .
Example 4.20. When M satisfies the 3-Ore condition, every reducer is tame, and a • Rimax
coincides with the maximal i-reduct of a, as used in [13, Section 6]. Otherwise, wild reducers
e 2,
may exist and a • Rimax need not be a maximal i-reduct of a: in the Artin-Tits monoid of type A
for a = 1/c/aba, we have a • R2max = a, since there is no nontrivial tame 2-reducer.
By definition, the greatest tame i-reducer for a only depends on the entries ai and ai+1 ,
and on the sign of i in a. By Lemma 4.17, it can be computed easily as a gcd of maximal
reducers. Note that the greatest tame i-reducer for a may be strictly larger than the gcd of ai
e 2 , there exist two maximal 2-reducers
and ai+1 : for instance, in the Artin-Tits monoid of type A
for a := 1/a/cabab, namely caa and cab, both wild, and the greatest tame 2-reducer is their
left gcd ca, a proper multiple of the left gcd of a and cabab, which is 1.
If M is a noetherian gcd-monoid, starting from an arbitrary multifraction a and repeatedly
performing (maximal) tame reductions leads in finitely many steps to a ≃± -equivalent multifraction that is tame-irreducible, meaning eligible for no tame reduction. By Proposition 4.18,
a division is always tame, so a tame-irreducible multifraction is prime. Adapting the proof of
Proposition 3.25 yields:
±
Proposition 4.21. If M is a noetherian gcd-monoid, then RM
(resp., RM ) is semi-convergent
±
if and only if, for every a in FM (resp., FM ),
(4.22)
If a is unital and tame-irreducible, then a is either trivial or reducible.
Corollary 4.23. Conjecture A is true if and only if (4.22) holds for every Artin-Tits monoid M
and every a in FM .
The above results suggest to investigate tame-irreducible multifractions more closely. By
definition, only wild reductions may apply to a tame-irreducible multifraction. A possible
approach for establishing (4.22) could be to study the irreducible reducts of tame-irreducible
multifractions. It happens frequently that, if a is tame-irreducible and admits several (wild)
reducts a1 , ..., am , then the reducts of the various ak s are pairwise disjoint, as if every such
reduct kept a trace of ak . If true, such a property might lead to a proof of Conjecture A using
Corollary 4.23 and arguments similar to those alluded to in the proof of Proposition 7.3. But
e 2,
the assumption is not readily correct. Indeed, always in the Artin-Tits monoid of type A
the 6-multifraction a := 1/c/aba/bc/a/bcb is tame-irreducible, it admits four wild reducers,
20
PATRICK DEHORNOY
namely a and b at level 2, and b and c at level 5, and the associated 2-reducts of a admit a
common reduct
(a • R2,a ) • R5,c R3,b R1,ac R2,b R3,c = bc/accb/ca/ab/ca/cb = (a • R2,b ) • R5,c R3,cb .
However, in this example, there is no confluence for the 5-reducts, and, more generally, we have
no example where the highest level wild reducts of a tame-irreducible multifraction admit a
common reduct.
4.3. The operator redt and Conjecture B. Our main claim in this section is that, for every
multifraction a, there exists a distinguished tame reduct of a, denoted redt (a), that can be
computed easily, and that should be 1 whenever a is unital: this is what we call Conjecture B.
Just mimicking the approach of Section 4.1 and trying to identify a unique maximal tame
reduct on the shape of ∂a cannot work, because the tame reducts of a multifraction need
e 2 again, a and b are tame 4-reducers
not admit a common reduct: in the context of type A
for a := 1/c/1/1/aba, but a • R4,a and a • R4,b admit no common reduct (by the way, in this
case, the two irreducible reducts of a can be reached using tame reductions only, respectively
R4,a R2,a R4,ba and R4,b R2,b R4,ab .
However, it is shown in [13, Section 6] that, if M is a noetherian gcd-monoid satisfying
the 3-Ore condition, hence in a case when all reductions are tame, there exists, for each n, a
universal sequence of levels U (n) such that, starting with any n-multifraction a and applying
the maximal (tame) reduction at the successive levels prescribed by U (n) inevitably leads to
the unique R-irreducible reduct red(a) of a. It is then natural to copy the recipe in the general
case, and to introduce:
Definition 4.24. If M is a noetherian gcd-monoid, then, for every depth n multifraction a
max
on M , we put redt (a) := a • RU(n)
, where U (n) is empty for n = 0, 1 and is (1, 2, ..., n − 1)
···Rimax
.
followed by U (n − 2) for n > 2, and, for i = (i1 , ..., iℓ ), we write a • Rimax for a • Rimax
1
ℓ
Thus, by [13, Proposition 6.7], if M is a noetherian gcd-monoid satisfying the 3-Ore condition,
red(a) = redt (a) holds for every a in FM . In this case, b ⇒∗ redt (a) holds for every reduct b of a,
and redt (a) is always R-irreducible. It is easy to see that, in the general case, these properties
do not extend to arbitrary (namely, not necessarily unital) multifractions.
e 2 , let a := 1/c/aba. Then U (3) = (1, 2)
Example 4.25. In the Artin-Tits monoid of type A
leads to redt (a) = a • R1max R2max = a, since 1 and c have no nontrivial common divisor, and
there is no tame 2-reducer for a. On the other hand, both a and b are 2-reducers for a, and
neither a • R2,a ⇒∗ redt (a) nor a • R2,b ⇒∗ redt (a) holds.
Next, let b := ac/aca/aba. We find redt (b) = b • D1,ac = 1/c/aba, to be compared with
∗ ′
b does not imply b′ ⇒∗ redt (b).
∂b = b • D2,a = ac/ca/ba. Then ∂b ⇒∗ redt (b) fails, so b ⇒div
Finally, let c := 1/c/aba/cb. We find redt (c) = c • D3,b = 1/c/ba/c, which is not irreducible,
nor even tame-irreducible: we have red2t (c) = redt (c) • R2,b R3,c = bc/cb/a/c.
However, these negative facts say nothing about tame reductions starting from a unital
multifraction, and, in spite of many tries, no counter-example was ever found so far to:
Conjecture B. If M is an Artin-Tits monoid, then redt (a) = 1 holds for every unital multifraction a in FM .
By definition, a ⇒∗ redt (a) holds, so redt (a) = 1 is a strengthening of a ⇒∗ 1 in which we
assert not only that a reduces to 1 but also that it goes to 1 in some prescribed way. Thus:
Fact 4.26. Conjecture B implies Conjecture A.
Although Conjecture B is more demanding than Conjecture A, it might be easier to establish
(or to contradict), as it predicts a definite equality rather than an existential statement. As
recalled above, [13, Proposition 6.7] implies that Conjecture B is true for every Artin-Tits
monoid of FC type. On the other hand, an example of a gcd-monoid (but not an Artin–Tits
one) for which (the counterpart of) Conjecture A is true but (that of) Conjecture B is false is
given in [19].
MULTIFRACTION REDUCTION II
21
5. Cross-confluence
Besides tame reductions and Conjecture B of Section 4, we now develop another approach to
e of R. The
Conjecture A, involving both the reduction system R and a symmetric counterpart R
properties of reduction and its counterpart are just symmetric, but interesting features appear
when both are used simultaneously, in particular what we call cross-confluence, a completely
novel property to the best of our knowledge. We are then led to a new statement, Conjecture C,
which would imply Conjecture A and, from there, the decidability of the word problem.
The section comprises four subsections. First, right reduction, the symmetric counterpart of
(left) reduction, is introduced in Subsection 5.1, and its basic properties are established. Next,
cross-conflence, which combines reduction and its counterpart, is introduced in Subsection 5.2,
and partial results are established. Then Conjecture C and its uniform version Cunif are stated
and discussed in Subsection 5.3. Finally, we briefly study in an Appendix the termination of
the joint system obtained by merging left and right reduction, a natural topic with nontrivial
results, but not directly connected so far to our main conjectures.
5.1. Right reduction. By definition, the reduction rule Ri,x of Definition 2.15 consists in
pushing a factor x to the left (small index entries) in the multifraction it is applied to: for
this reason, we shall call it a left reduction. From now on, we shall also consider symmetric
counterparts, naturally called right reductions, where elements are pushed to the right.
Definition 5.1. If M is a gcd-monoid and a, b lie in FM± , then, for i > 1 and x in M , we declare
ei,x holds if we have kbk = kak, bk = ak for k 6= i − 1, i, i + 1, and there exists x′
that b = a • R
(necessarily unique) satisfying
for
for
for
for
i < kak
i < kak
i = kak
i = kak
positive in a:
negative in a:
positive in a:
negative in a:
xbi−1 = ai−1 , xbi = ai x′ = x ∨ ai , bi+1 = ai+1 x′ ,
e ai , bi+1 = x′ ai+1 ,
bi−1 x = ai−1 , bi x = x′ ai = x ∨
xbi−1 = ai−1 ,
xbi = ai ,
bi−1 x = ai−1 ,
b i x = ai .
ei,x holds for some i and some x 6= 1, and use ⇒
e b if a • R
e ∗ for the reflexive–
We write a ⇒
±
±
eM so obtained on FM is called right reduction,
transitive closure of ⇒.
e The rewrite system R
eM .
and its restriction to FM (positive multifractions) is denoted by R
ei,x is symmetric of that of Ri,x : one extracts x from ai−1 , lets it cross ai
The action of R
using an lcm, and incorporates the resulting remainder in ai+1 , thus carrying x from level i − 1
en,x for n = kak is adapted to
to level i + 1, see Figure 4. As in the case of R1,x , the action of R
avoid creating a (n + 1)st entry.
Ri,x :
...
ei,x :
R
ai
ai−1
x′
⇐
...
x
ai
ai−1
ai+1
x′
⇐
ai+1
...
x
bi−1
bi
bi+1
bi−1
bi
bi+1
ai−1
ai
ai+1
ai−1
ai
ai+1
...
x
bi−1
⇒
bi
...
x′
bi+1
x
bi−1
⇒
bi
...
x′
bi+1
ei,x : in the
Figure 4. Comparing the left reduction Ri,x and the right reduction R
first case (top), one pushes the factor x from ai+1 to ai−1 through ai , in the second
case (bottom), one pushes x from ai−1 to ai+1 through ai .
22
PATRICK DEHORNOY
Remark 5.2. Right reduction is not an inverse of left reduction: when the reduced factor x
crosses (in one direction or the other) the entry ai , using the lcm operation cancels common
ei,x amount to dividing by x and,
factors. Typically, if x divides ai , then both Ri,x and R
therefore, their actions coincide. See Remark 5.26 for more on this.
The rest of this subsection is devoted to the basic properties of right reduction and their
connection with those of left reduction, in particular with respect to convergence and semiconvergence. As can be expected, the convenient tool is an operation exchanging left and right
reduction, in this case the duality map e of Notation 3.7.
Lemma 5.3. If M is a gcd-monoid, then, for all a, b in FM± ,
(5.4)
a ⇒ b is equivalent to e
a ⇒
e eb.
Proof. Comparing the definitions shows that, if a and b have depth n, then b = a • Ri,x is
en+1−i,x , whence (5.4).
a•R
equivalent to eb = e
The following properties of right reduction follow almost directly from their counterpart
involving left reduction. The only point requiring some care is the lack of involutivity of e,
itself resulting from the lack of surjectivity of the map a 7→ 1 · a.
Lemma 5.5. Assume that M is a gcd-monoid.
e ∗ b implies a ≃± b.
(i) The relation ⇒
e ∗ is included in ≃± , i.e., a ⇒
∗
(ii) The relation ⇒
e is compatible with the multiplication of FM± .
e ∗ b is equivalent to 1p · a · 1q ⇒
e ∗ 1p · b · 1q .
(iii) For all a, b and p, q, the relation a ⇒
(iv) If a, b belong to FM , then a ⇒ b is equivalent to 1 · e
a ⇒
e 1 · eb, and a ⇒
e b is equivalent
e
to 1 · e
a ⇒ 1 · b.
Proof. (i) By (5.4), a ⇒
e ∗ b implies e
a ⇒∗ eb, whence ι(e
a) = ι(eb) by Lemma 2.17(i), hence
±
ι(a) = ι(b) by Lemma 3.9, i.e., a ≃ b.
(ii) For all c, d, the relation a ⇒
e ∗ b implies e
a ⇒∗ eb, whence e
d·e
a ·ec ⇒∗ e
d·eb·ec by Lemma 2.17(ii),
∗
which is (c · a · d)e ⇒ (c · b · d)e by Lemma 3.9, and finally c · a · d ⇒∗ c · b · d by (5.4) again.
(iii) Using duality as above, the result follows now from Lemma 2.17(iii).
(iv) By (5.4), a ⇒ b is equivalent to e
a ⇒
e eb, hence, by (iii), it is also equivalent to 1 · a ⇒ 1 · b.
Similarly, by (5.4) again, a ⇒
e b is equivalent to e
a ⇒ eb, hence, by Lemma 2.17, it is also
equivalent to 1 · e
a ⇒ 1 · eb.
By (5.4), an infinite sequence of right reductions would provide an infinite sequence of left
±
±
eM
reductions, and vice versa, so R
is terminating if and only if RM
is. Comparing irreducible
elements is straightforward:
Lemma 5.6. If M is a gcd-monoid, a multifraction a is R-irreducible if and only if e
a is
e
e
R-irreducible.
For a positive with kak even, a is R-irreducible
if and only if e
a is R-irreducible.
e
e b. By (5.4), we deduce a = e
a ⇒ eb, hence a is not R-irreducible. So a
Proof. Assume e
a ⇒
e
being R-irreducible implies that e
a is R-irreducible.
Conversely, assume e
a ⇒ b. By (5.4), we
e
e
e
e
deduce a = e
a ⇒
e b, hence a is not R-irreducible. So a being R-irreducible implies that e
a is
R-irreducible. For the second equivalence, use the involutivity of e on positive multifractions
of even depth (but we claim nothing for positive multifractions of odd length).
Using the above technical results, we can compare convergence and semi-convergence for left
and right reduction. Below, observe the difference between (i) and (ii), which involve only one
direction but multifractions of both signs, and (iii), which only involves positive multifractions
and requires using both left and right reductions. This distinction is one of the reasons for
considering both positive and negative multifractions in this paper (contrary to [13]).
MULTIFRACTION REDUCTION II
23
Proposition 5.7. For every gcd-monoid M , the following are equivalent:
±
(i) The system RM
is convergent (resp., semi-convergent);
±
eM
(ii) The system R
is convergent (resp., semi-convergent);
eM are convergent (resp., semi-convergent).
(iii) The systems RM and R
±
Proof. We begin with convergence. Assume that RM
is convergent. Let a belong to FM± . Let
a). By definition, we have e
a ⇒∗ b. Then, by Lemmas 5.5 and 5.6, we have a ⇒
e ∗ eb
b = red(e
e
e
and eb is R-irreducible.
Assume that c is R-irreducible
and a ⇒
e ∗ c holds. By Lemmas 5.5
±
∗
and 5.6 again, we deduce e
a ⇒ ec and ec is R-irreducible. The assumption that RM
is convergent
±
e
e
eM
implies ec = b, whence c = eb. Hence eb is the only R-irreducible
R-reduct
of a. Therefore, R
is convergent, and (i) implies (ii). The converse implication is proved in the same way, so (i) is
equivalent to (ii).
e
Since all R-reducts and R-reducts
of a positive multifraction are positive, it is clear that (i)
eM is convergent. Conversely, assume that
implies that RM is convergent, and (ii) implies that R
e
both RM and RM are convergent. Let a be an arbitrary multifraction on M . Assume first that
a is positive. Let b be the unique R-irreducible reduct of a. Assume now that a is negative.
e
e
Then e
a is positive. Let b be the unique R-irreducible
R-reduct
of e
a. By Lemmas 5.5 and 5.6, eb
∗
is R-irreducible, and a ⇒ b holds. Now assume that c is R-irreducible and a ⇒∗ c holds. Then
eM is convergent, we deduce ec = b,
ec is R-irreducible and e
a ⇒
e ∗ ec holds. As e
a is positive and R
±
whence c = eb. Hence RM is convergent, and (iii) implies (i). This completes the argument for
convergence.
±
Assume now that RM
is semi-convergent. Let a be a unital multifraction. By Lemma 2.17, e
a
±
∗ e
∗
e
is unital as well, hence we must have e
a ⇒ 1, which, by (5.4), implies a ⇒
e 1 = 1. Hence RM is
semi-convergent. The converse implication is similar, so (i) and (ii) are equivalent in this case as
eM
well. On the other hand, (i) and (ii) clearly imply (iii). Finally, assume that both RM and R
are semi-convergent. Let a be a unital multifraction. If a is positive, the assumption that RM is
eM
semi-convergent implies a ⇒∗ 1. If a is negative, then e
a is positive, and the assumption that R
∗
±
∗ e
is semi-convergent implies e
a ⇒
e 1, whence a ⇒ 1 = 1 by (5.4). Hence RM is semi-convergent.
So (iii) implies (i), which completes the argument for semi-convergence.
f
= (red(e
a))e.
In the convergent case, the above proof implies, with obvious notation, red(a)
Remark 5.8. It is shown in [13, Sec. 3] that, when left reduction is considered, trimming
final trivial entries essentially does not change reduction. This is not true for right reduction:
deleting trivial final entries can change the reducts, as deleting trivial initial entries does in the
case of left reduction.
5.2. The cross-confluence property. We now introduce our main new notion, a variant of
confluence that combines left and right reduction.
Definition 5.9. If M is a gcd-monoid, we say that RM is cross-confluent if, for all a, b, c in FM ,
(5.10)
e ∗ b and a ⇒
e ∗ c, there exists d satisfying b ⇒∗ d and c ⇒∗ d.
If we have a ⇒
So cross-confluence for RM means that left reduction provides a solution for the confluence
±
pairs of right reduction. We shall naturally say that RM
is cross-confluent if (5.10) holds for
±
eM is cross-confluent if right reduction provides
all a, b, c in FM . On the other hand, we say that R
a solution for the confluence pairs of left reduction, that is, if
(5.11)
e ∗ d and c ⇒
e ∗ d.
If we have a ⇒∗ b and a ⇒∗ c, there exists d satisfying b ⇒
±
eM
is cross-confluent if (5.11) holds for all a, b, c in FM± .
holds for all a, b, c in FM . Finally, R
Remark 5.12. The definition of cross-confluence involves both left and right reduction. But,
owing to the equivalence (5.4), it can alternatively be stated as a property involving left reduc±
tion exclusively. Indeed, in the case of RM
, cross-confluence is equivalent to
(5.13)
a ⇒∗ ec, there exists d satisfying b ⇒∗ d and c ⇒∗ d,
If we have e
a ⇒∗ eb and e
24
PATRICK DEHORNOY
±
eM
and similarly for R
. In the case of RM , when we restrict to positive multifractions, we also
have an alternative statement involving ⇒∗ exclusively, but it takes the less symmetric form
(5.14)
a ⇒∗ 1 · ec,
If we have 1 · e
a ⇒∗ 1 · eb and 1 · e
there exists d satisfying b ⇒∗ d and c ⇒∗ d,
because the e operation is not involutive in this case.
e ± and that of their positive
Using duality, we obtain for the cross-confluences of R± and R
versions the same connection as in the case of convergence and semi-convergence:
Proposition 5.15. For every gcd-monoid M , the following are equivalent:
±
(i) The system RM
is cross-confluent;
±
eM
(ii) The system R
is cross-confluent;
eM are cross-confluent.
(iii) The systems RM and R
±
Proof. Assume that RM
is cross-confluent. Applying (5.10) to e
a, eb, and ec and using (5.4), we
±
e
deduce (5.11), so RM is also cross-confluent. Doing the same from (5.11) returns to (5.10). So
(i) and (ii) are equivalent.
±
e
Next, all R- and R-reducts
of a positive multifraction are positive, so, if RM
is cross±
e
confluent, its retriction to FM is also cross-confluent. Similarly, if RM is cross-confluent, its
eM
retriction to FM is cross-confluent. Hence (i) implies (iii). Finally, assume that RM and R
∗
∗
±
are cross-confluent, and a, b, c belong to FM and satisfy a ⇒
e b and a ⇒
e c. If a, hence b
and c as well, are positive, the assumption that RM is cross-confluent implies the existence of d
satisfying (5.10). Assume now that a, hence b and c, are negative. Then e
a, eb, ec are positive, and
∗ e
∗
eM is cross-confluent implies the existence
we have e
a ⇒ b and e
a ⇒ ec. The assumption that R
of d satisfying eb ⇒
e ∗ d and ec ⇒
e ∗ d. By (5.4), we deduce b ⇒∗ e
d and c ⇒∗ e
d. So (5.10) holds,
±
RM is cross-confluent, and (iii) implies (i).
In view of our main purpose, namely establishing the semi-convergence of (left) reduction,
the main result is the following connection, which locates cross-confluence as an intermediate
between convergence and semi-convergence:
Proposition 5.16. Assume that M is a noetherian gcd-monoid.
±
±
(i) If RM
is convergent, then RM
is cross-confluent.
±
±
(ii) If RM is cross-confluent, then RM
is semi-convergent.
(iii) Mutatis mutandis, the same implications hold for RM .
We begin with an auxiliary result:
Lemma 5.17. Assume that M is a gcd-monoid.
±
e ∗ 1 are
is cross-confluent, then, for every a in FM± , the relations a ⇒∗ 1 and a ⇒
(i) If RM
equivalent.
(ii) If RM is cross-confluent, then, for every a in FM , the relation a ⇒
e ∗ 1 implies a ⇒∗ 1.
±
implies
Proof. (i) Assume a ∈ FM± and a ⇒∗ 1. By Proposition 5.15, the cross-confluence of RM
±
∗
eM . By definition, we also have a ⇒ a. Then (5.11) implies the existence of d satisfying
that of R
e
e ∗ d and a ⇒
e ∗ d. By definition, 1 is R-irreducible,
so 1 ⇒
e ∗ d implies d = 1, whence a ⇒
e ∗ 1.
1 ⇒
∗
∗
Conversely, assume a ⇒
e 1. We have a ⇒
e a, and (5.10) implies the existence of d satisfying
a ⇒∗ d and 1 ⇒∗ d, whence d = 1, and a ⇒∗ 1.
(ii) When a lies in FM , the latter argument remains valid, and it shows again that a ⇒
e∗ 1
∗
implies a ⇒ 1. (By contrast, the former argument need not extend, as there is a priori no
eM .)
reason why the cross-confluence of RM should imply that of R
Proof of Proposition 5.16. (i) Let a belong to FM± , and assume a ⇒
e ∗ b and a ⇒
e ∗ c. By
±
±
∗
∗
Lemma 5.5, we have a ≃ b ≃ c, whence b ⇒ d and c ⇒ d, where d is the (unique)
R-irreducible reduct of a. Hence (5.10) is satisfied.
MULTIFRACTION REDUCTION II
25
±
(ii) Assume that RM
is cross-confluent, and we have a ⇒∗ 1 and a ⇒∗ b for some a in FM± .
±
By (5.11), which is valid since, by Proposition 5.15, the cross-confluence of RM
implies that
±
∗
∗
eM , there exists d satisfying b ⇒
e
of R
e d and 1 ⇒
e d. Since 1 is R-irreducible,
we must have
±
e ∗ 1. By Lemma 5.17, we deduce b ⇒∗ 1. Hence RM
is 1-confluent and,
d = 1, whence b ⇒
therefore, by Proposition 3.29, it is semi-convergent.
(iii) For (i), the argument is the same in the case a ∈ FM . For (ii), assume that RM is
cross-confluent, and we have a ⇒∗ 1 and a ⇒∗ b for some a in FM . By Lemma 5.5(iv), we have
1·e
a ⇒
e ∗ 1 · 1 = 1 and 1 · a ⇒
e ∗ 1 · b, and 1 · e
a is positive. As RM is cross-confluent, we deduce the
∗
e
existence of d satisfying 1 ⇒ d and 1 · b ⇒∗ d. As 1 is R-irreducible, we have d = 1, whence
1 · eb ⇒∗ 1, and, by Lemma 5.5(iv) again, b ⇒
e ∗ 1. By Lemma 5.17(ii), we deduce b ⇒∗ 1. Hence
RM is 1-confluent and, by Proposition 3.29, it is semi-convergent.
5.3. Conjectures C and Cunif . We thus arrive at what we think is the main conjecture in
this paper:
Conjecture C. For every Artin-Tits monoid M , the system RM is cross-confluent.
By Proposition 5.16, Conjecture C implies Conjecture A, whence the decidability of the
word problem of the group, and it is true for every Artin–Tits monoid of type FC. Note that
Conjecture C is different from Conjectures A and B in that it predicts something for all
multifractions, not only for unital ones. So, in a sense, it is a more structural property, which
we think is interesting independently of any application.
No proof of Conjecture C is in view so far in the general case, but we now observe that local
cross-confluence, namely cross-confluence with single reduction steps on the left, is always true.
Proposition 5.18. If M is a gcd-monoid, then, for all a, b, c in FM± ,
(5.19)
(5.20)
If we have a ⇒
e b and a ⇒
e c, there exists d satisfying b ⇒∗ d and c ⇒∗ d.
If we have a ⇒ b and a ⇒ c, there exists d satisfying b ⇒
e ∗ d and c ⇒
e ∗ d.
The proof relies on the following preparatory result:
Lemma 5.21. Assume that M is a gcd-monoid and a is a multifraction on M such that a • Ri,x
b be defined by ai x′ = ai ∨ x and
is defined. If i is negative (resp., positive) in a, let x′ and x
′
e
e
x
b = ai ∧ x (resp., x ai = ai ∨ x and x
b = ai ∧ x). Then, we have
ei,x′ = a • Di,bx .
(5.22)
a • Ri,x R
Proof. (Figure 5) Let a′ = a • Ri,x . Assuming i negative in a, we have
(5.23)
a′i−1 = ai−1 x′ ,
xa′i = ai x′ = ai ∨ x,
xa′i+1 = ai+1 .
By construction, x′ right divides a′i−1 , and x′ and a′i admit a common left multiple, namely ai x′ .
ei,x′ is defined, and it is determined by
Hence a′′ = a′ • R
(5.24)
a′′i−1 x′ = a′i−1 ,
e x′ ,
a′′i x′ = x′′ a′i = a′i ∨
a′′i+1 = x′′ a′i+1 .
By definition, the left lcm x′′ a′i of x′ and a′i left divides their common left multiple xa′i , which
implies the existence of x
b satisfying x = x
bx′′ , and, from there, ai x′ = xa′i = x
bx′′ a′i = x
ba′′i x′ ,
′′
′′
′′
ba′′i+1 =
bai = ai , and x
whence ai = x
bai . Merging (5.23) and (5.24), we deduce ai−1 = ai−1 , x
′′
ai+1 , which shows that a is obtained from a by left dividing the ith and (i + 1)st entries by x
b,
i.e., by applying the division Di,bx .
The argument is symmetric when i is positive in a.
ei,x is defined, a symmetric argument gives
When a • R
(5.25)
ei,x Ri,x′ = a • Di−1,bx ,
a•R
e x and x
e x (resp., ai x′ = ai ∨ x and
where x′ and x
b are now defined by x′ ai = ai ∨
b = ai ∧
x
b = ai ∧ x) if i is negative (resp., positive) in a. The index of the division is shifted (i − 1
ei+1,x .
instead of i) relatively to (5.22), because, as a set of pairs, Di,x is included in R
We can now complete the argument for Proposition 5.18.
26
PATRICK DEHORNOY
ai
ai−1
x′
a′i−1
a′′i
x
b
x′′
a′i
ai+1
x
a′i+1
ei,x′
Figure 5. Composing the left reduction Ri,x and the inverse right reduction R
amounts to performing the division Di,bx where x
b is the gcd of ai and x.
ei,x and c = a • R
ej,y . By Lemma 5.21, we have
Proof of Proposition 5.18. Assume b = a • R
b • Ri,x′ = a • Di,bx
and c′ • Rj,y′ = a • Dj,by
for some x′ , x
b and y ′ , yb. By Proposition 4.4, ∂a is a common reduct of all multifractions obtained
from a by a division, hence in particular of a • Di,x′′ and a • Dj,y′′ . Thus b ⇒∗ ∂a and c ⇒∗ ∂b
hold, hence (5.19) is satisfied with d = ∂a.
The proof of (5.20) is symmetric, using (5.25) instead of (5.22), and d = ∂a again.
In the case of a single rewrite system, local confluence implies confluence whenever the system
is terminating (by Berman’s well known diamond lemma). There is no hope of a similar result
e is not terminating in general—see (5.30) below—and
here, both because the union of R and R
because, in the definition of cross-confluence, the arrows ⇒ and ⇒
e are not in a position for a
natural induction.
Remark 5.26. Right reduction is close to being an inverse of left reduction. Indeed, provided
the ground monoid is noetherian, every reduction is a product of atomic reductions, namely
ei,x with x an atom. Now, if x is an atom, the gcd of x and ai
reductions of the form Ri,x or R
is either ai , meaning that x divides ai , or 1. In the former case, a • Ri,x is a • Di,x , whereas,
ei,x′ , i.e., left reducing x in a is the inverse
in the latter, Lemma 5.21 implies a = (a • Ri,x ) • R
′
of right reducing x . Thus, writing D for the family of divisions and Rat for that of atomic left
−1
e is generated by D ∪ Rat
reductions, R is generated by D ∪ Rat , whereas R
. By Lemma 4.7 and
the results of [13], confluence between D and Rat is always true, whereas confluence between Rat
−1
and Rat
is trivial. Therefore, one might hope that cross-confluence diagrams can always be
constructed by assembling the various types of elementary confluence diamonds. This is not
true: using a tedious case-by-case argument, one can indeed establish cross-confluence in the
case when, in (5.10), b and c are obtained from a by two atomic reduction steps, but there is no
hope to go very far in this direction, both because of the counter-examples of Example 5.27, and
because, in any case, cross-confluence cannot be true for an arbitrary noetherian gcd-monoid,
since there exist such monoids for which the counterparts of Conjectures A and C fail [19]:
if true, cross-confluence has to be a specific property of Artin–Tits monoids, or at least of a
restricted family of gcd-monoids.
The following examples are given to show that naive attempts to extend the local crossconfluence result of Proposition 5.18 are due to fail.
Example 5.27. Proposition 5.18 shows that, if b is obtained by applying one right reductions
to a, then applying one well chosen left reduction to b provides a multifraction c obtained by
one division from a. The result fails when 1 is replaced by k > 2. Indeed, in the Artin-Tits
e5,b = 1/1/ca/cb/1, (which is
e3,a R
e 2 , consider a = 1/a/ca/cb/b and b = a • R
monoid of type A
e
R-irreducible).
The only way to left reduce b is to start with R1,ca , leading to c = ca/1/cb/1,
not reachable from a by two, or any number, of divisions.
In the above case, we have c = a • R4,b R2,cac , and therefore there is no contradiction with the
weaker conclusion that c is obtained both from a by applying k left reductions. The following
example (with k = 3) shows that this is not true either. Indeed, consider a = ca/cb/bc/ba and
e4,a = 1/1/ac/ab. As predicted by Conjecture C, a and b admit common left
e2,ca R
e3,bc R
b = a•R
reducts, but the latter are c and c • D2,c , with c = ac/cab/c/1 = a • R2,b R3,a D2,a R3,b D1,aa D2,b ,
not reachable from a using less than six left reductions. What is surprising here is that, if we
MULTIFRACTION REDUCTION II
27
e2,ca = 1/1/aca/aba, then b′ left reduces to a, whereas b = b′ • D3,a only left
e3,bc R
put b′ := a • R
reduces to c, very far from a: one single division may change left reducts completely.
We conclude with one more conjecture. The conjunction of Propositions 4.4 and 5.18 shows
not only that any two right reducts b, c of a multifraction a admit a common left reduct d, but
even that there exists d only depending on a, namely ∂a, that witnesses for all elementary right
reducts of a simultaneously. This suggests to consider a strong version of cross-confluence:
Definition 5.28. If M is a gcd-monoid, we say that RM is uniformly cross-confluent if there
exists a map ∇ from FM to itself such that, for every a in FM , the relation b ⇒∗ ∇a holds for
every right reduct b of a.
In the case when reduction is convergent, defining ∇a = red(a) provides a convenient witness, and therefore the conclusion of Proposition 5.16 can be strengthened to uniform crossconfluence. We propose:
Conjecture Cunif . For every Artin-Tits monoid M , the system RM is uniformly cross-confluent.
By the above observation, Conjecture Cunif is true for every Artin–Tits monoid of type FC,
and no counter-example could be found so far in any other type. It implies Conjecture C and,
therefore, Conjecture A, but it is more demanding. However, if an explicit definition of ∇a
could be found, one can reasonably hope that the proof of Conjecture Cunif would then reduce
to a series of verifications. But, here again, naive attempts fail: two natural candidates for ∇a
could be either redt (a) (which works in the convergent case), or (if it always exists) a maximal
common ancestor of all irreducible reducts in the tree of all left reducts of a, but the example
of Figure 6 shows that neither of these choices works in every case.
Appendix: Mixed termination. Although it is not directly connected with cross-confluence,
we mention here one further result involving both left and right reduction. First, we know
that, in a noetherian context, (left) reduction is terminating, meaning that there is no infinite
sequence of reductions. It turns out that a stronger finiteness result holds:
Proposition 5.29. If M is a finitely generated noetherian gcd-monoid, then every multifraction a on M admits only finitely many left reducts, and finitely many right reducts.
Proof. As M is noetherian and contains no nontrivial invertible element, a subfamily of M
is generating if and only if it contains all atoms. Hence, the assumption that M is finitely
generated implies that the atom set A of M is finite. Let a belong to FM± . We construct a
tree Ta , whose nodes are pairs (b, s), where b is a left reduct of a and s is a finite sequence
in A × N: the root of Ta is (a, ε), and, using ⌢ for concatenation, the sons of (b, s) are all
pairs (b • Ri,x , s⌢ (i, x)) such that b • Ri,x is defined. As A is finite, for every multifraction b,
the number of pairs (i, x) with x in A and b • Ri,x defined is finite. Hence each node in Ta has
finitely many immediate successors. On the other hand, the assumption that M is noetherian
±
implies that RM
is terminating and, therefore, the tree Ta has no infinite branch. Hence, by
König’s lemma, Ta is finite. As every left reduct of a appears (maybe more than once) in Ta ,
the number of such reducts is finite.
The argument for right reducts is symmetric.
Thus, it makes sense to wonder whether, starting from a multifraction a, the family of
all multifractions that can be obtained from a using left and right reduction is finite. The
argument for Proposition 5.29 does not extend, because the well-orders witnessing for the
termination of left and right reductions are not the same, and it is easy to see that the result
e 2 , we find
itself fails in general: starting from a := 1/a/bc/1 in the Artin-Tits of type A
e
a • R2,b R3,a = ba/b/ca/ac, whence, repeating three times,
(5.30)
e3,c = bacbac/a/bc/acbacb = bacbac · a · acbacb.
e3,b R2,a R
e3,a R2,c R
a • R2,b R
e3,c )p make for p > 0 an infinite none3,b R2,a R
e3,a R2,c R
Hence the multifractions a • (R2,b R
e
terminating (and non-periodic) sequence with respect to R ∪ R.
28
PATRICK DEHORNOY
a9 : /cbac/bcbb/aca
R2,bc
a12 : /ac/caba/acaab
D3,b
D3,c
e2,a
R
R2,ca
e2,a
R
D2,c
a13 : /ac/cba/acaa
a11 : a/c/ba/acaab
a1 : a/bac/bb/aca
D3,a
a10 : /bac/bcb/aca D
2,b
e3,a
R
a14 : /ac/cb/aca
R3,ab
D2,b
R3,c
e2,a
R
R2,c
R3,a
a2 : a/ac/b/aca
D3,c
e3,b
R
R3,c
R3,c
a3 : a/bcbac/ccb/ca
D1,a
a4 : a/babac/aab/ac
D2,c
a5 : /cbac/ccb/ca
a6 : a/bcac/cb/ca
D2,c
a7 : /bac/cb/ca
D1,a
R3,a
D2,a
a8 : a/baac/ab/ac
Figure 6. Left and right reducts of a1 := a/bac/bb/aca in the Artin-Tits monoid of
e 2 : there are 8 left reducts (in grey), among which a7 and a8 are R-irreducible,
type A
e
and 10 right reducts (dashed lines), among which a14 is R-irreducible.
Plain (resp.,
dashed) arrows correspond to left (resp., right) reductions that are not divisions,
double arrows correspond to divisions (which are both left and right reductions); all
left reductions decrease the distance to the bottom. As Conjecture Cunif predicts,
there exists a common left reduct for all right reducts: in this case, there is only one,
namely a7 , and it is neither redt (a1 ) nor the maximal common ancestor of a7 and a8 ,
both equal to a2 .
By contrast, let us mention without detailed proof a finiteness result valid whenever the
ground monoid M is a Garside monoid [11, 16], i.e., a strongly noetherian gcd-monoid possessing
in addition an element ∆ (“Garside element”) whose left and right divisors coincide, generate M ,
and are finite in number.
Proposition 5.31. If M is a Garside monoid, then, for every multifraction a on M , the family
e
of all (R ∪ R)-reducts
of a is finite.
Proof (sketch). Let ∆ be a Garside element in M and let a be a multifraction on M . Then
there exists a positive integer d such that the path associated with a can be drawn in the finite
fragment of the Cayley graph of M made of the divisors of ∆d : this is the notion of a path “drawn
in Div(∆d )” as considered in [10]. Then the family of all paths drawn in Div(∆d ) is closed under
the special transformations alluded to in Subsection 3.3, and, therefore, all multifractions that
e are drawn in the same finite fragment Div(∆d ) of the
can be derived from a using ⇒ and ⇒
Cayley graph. As we consider multifractions with a fixed depth, only finitely many of them can
be drawn in a finite fragment of a Cayley graph.
The argument extends to every Artin-Tits monoid M of type FC, replacing the finite family
of divisors of the Garside element ∆ with the union of the finitely many finite families of divisors
of the Garside elements ∆I , where I is a family of atoms of M that generates a spherical type
submonoid of M .
MULTIFRACTION REDUCTION II
29
6. Finite approximations
±
The semi-convergence of RM and Conjectures A, B, and C, involve multifractions of arbitrary depth. Further results appear in the particular case of small depth multifractions. The
cases of depth 2 and, more interestingly, of depth 4 are addressed in Subsections 6.1 and 6.2,
where connections with the embeddability in the group and the uniqueness of fractional decompositions, respectively, are established. An application of the latter to partial orderings of
the group is established in Subsection 6.3. Finally, we describe in Subsection 6.4 a connection
between Conjecture B and van Kampen diagrams for unital n-multifractions.
±
6.1. The n-semi-convergence property. The rewrite system RM
has been called semiconvergent if (3.4) holds for every multifraction on M , i.e., if a being unital implies a ⇒∗ 1.
±
Definition 6.1. If M is a gcd-monoid, we say that RM
(resp., RM ) is n-semi-convergent if (3.4)
±
holds for every n-multifraction a in FM (resp., FM ).
Accordingly, we shall use Conjecture An for the restriction of Conjecture A to depth n
±
multifractions, and similarly for B and C. Some easy connections exist. Of course, if RM
is
n-semi-convergent, then so is its subsystem RM .
±
Lemma 6.2. If M is a gcd-monoid and RM
(resp., RM ) is n-semi-convergent, it is p-semiconvergent for p < n.
Proof. Let a be a nontrivial unital p-multifraction, with p < n. There exists r (equal to n − p
±
or n − p + 1) such that a · 1r has width n, and it is also nontrivial and unital. As RM
is
±
∗
∗
is
n-semi-convergent, we have a · 1r ⇒ 1, which implies a ⇒ 1 by Lemma 2.17(iii). So RM
p-semi-convergent.
On the other hand, by repeating the proof of Proposition 3.22, we obtain
±
Lemma 6.3. If M is a noetherian gcd-monoid, then RM
(resp., RM ) is n-semi-convergent if
and only if (3.23) holds for every n-multifraction a in FM± (resp., FM ), i.e., if a is unital, then
it is either trivial or reducible.
We now address the cases of small depth. The case of depth one is essentially trivial, in
that it follows from a sufficiently strong form of noetherinity and does not really involve the
algebraic properties of the monoid:
Proposition 6.4. Assume that M is a gcd-monoid that admits a length function, namely a
map λ : M → N satisfying, for all a, b in M ,
(6.5)
λ(ab) = λ(a) + λ(b),
and
λ(a) > 0 for a 6= 1.
±
Then the system RM is 1-semi-convergent.
Condition (6.5) is strong noetherianity (3.13) with > replaced by =. It holds in every ArtinTits monoid and, more generally, in every monoid with a homogeneous presentation.
P
P
Proof. Extend the map λ to FM± by λ(a) := i positive in a λ(ai ) − i negative in a λ(ai ). Then λ
is a homomorphism from the monoid FM± to (N, +), and, for every a in M , we have λ(a/a) =
λ(/a/a) = λ(1) = λ(∅) = 0. By Proposition 2.5, the latter pairs generate ≃± as a congruence,
hence λ is invariant under ≃± . Then a 6= 1 implies λ(a) 6= λ(1), whence a 6≃± 1. Hence the only
±
unital 1-multifraction is 1, and RM
is 1-semi-convergent.
The cases of depths 2 and 3 turn out to be directly connected with the embeddability of the
considered monoid in its enveloping group.
Proposition 6.6. If M is a gcd-monoid, the following are equivalent:
(i) The system RM is 2-semi-convergent.
±
(ii) The system RM
is 2-semi-convergent.
±
is 3-semi-convergent.
(iii) The system RM
(iv) The monoid M embeds in U(M ).
30
PATRICK DEHORNOY
Proof. Assume that RM is 2-semi-convergent. Let a, b two elements of M satisfying ι(a) = ι(b).
By (2.6), we have a/b ≃± 1. The assumption that RM is 2-semi-convergent implies a/b ⇒∗ 1.
By definition, this means that there exists x in M satisfying a = x = b, whence a = b. So M
embeds in U(M ). Hence (i) implies (iv).
Clearly (ii) implies (i), and (iii) implies (ii) by Lemma 6.2.
Finally, assume that M embeds in U(M ), and let a be a positive nontrivial unital 3multifraction. By assumption, we have a ≃± 1, whence ι(a1 )ι(a2 )−1 ι(a3 ) = 1 in U(M ) by (2.6),
and, therefore, ι(a2 ) = ι(a3 )ι(a1 ) = ι(a3 a1 ) in U(M ). As M embeds in U(M ), this implies
a2 = a3 a1 in M . Then a has the form a1 /a3 a1 /a3 , implying a • D1,a1 D2,a3 = 1/1/1. Thus
RM is 3-semi-convergent. The argument is the same for a negative 3-multifraction a, finding
±
is
now ι(a1 )−1 ι(a2 )ι(a3 )−1 = 1, whence a2 = a1 a3 , and a • D1,a1 D2,a3 = /1/1/1. Hence RM
3-semi-convergent. So (iv) implies (iii).
±
Corollary 6.7. For every Artin-Tits monoid M , the system RM
is n-semi-convergent for
n 6 3.
±
Proof. By Lemma 6.2, RM
is 1-semi-convergent. Next, it is known [28] that M embeds
±
into U(M ). Hence, by Proposition 6.6, RM
is 2- and 3-semi-convergent.
In other words, Conjecture An is true for n 6 3.
6.2. Multifractions of depth 4. We now address 4-semi-convergence, which turns out to give
rise to interesting phenomena. We begin with preliminary results about unital multifractions
that are in some sense the simplest ones.
Definition 6.8. If M is a monoid and a is an n-multifraction on M ,
with n even, we say that (x1 , ..., xn ) is a central cross for a if we have
(
xi xi+1 for i positive in a,
ai =
xi+1 xi for i negative in a,
a2
a3
x2
x1
x3
x4
a1
a4
with the convention xn+1 = x1 .
The diagram of Definition 6.8 shows that a multifraction that admits a central cross is unital:
if (x1 , ..., xn ) is a central cross for a positive multifraction a, we find
ι(a) = ι(x1 x2 ) ι(x3 x2 )−1 ι(x3 x4 ) ··· (x1 xn )−1 = 1,
and similarly when a is negative. It follows from the definition that a sequence (x1 , ..., xn )
is a central cross for a positive multifraction a if and only if (x2 , ..., xn , x1 ) is a central cross
for /a2 /···/an /a1 . So, we immediately obtain
Lemma 6.9. For every monoid M and every even n, a positive n-multifraction a admits a
central cross if and only if the negative multifraction /a2 /···/an /a1 does.
Multifractions with a central cross always behave nicely in terms of reduction:
Lemma 6.10. If M is a gcd-monoid and a is a multifraction on M that admits a central cross,
then redt (a) = 1 holds.
Proof. We prove the result using induction on n > 2 even, and assuming a positive. Assume
that (x1 , ..., xn ) is a central cross for a. For n = 2, the assumption boils down to a1 = a2 = x1 x2 ,
e x3 , with
directly implying a • R1max = 1/1. Assume n > 4 and, say, a positive. Let x := x1 ∧
e a2 = xx2 , whence a • R1max = x′1 /x′3 /a3 /···/an . As
x1 = x′1 x and x3 = x′3 x. Then we have a1 ∧
a3 = x3 x4 expands into a3 = x′3 xx4 , this can be rewritten as a • R1max = x′1 /x′3 /x′3 xx4 /···/an .
We deduce
(6.11)
a • R1max R2max = x′1 xx4 /1/1/a4/···/an = x1 x4 /1/1/a4/···/an
with a4 = x1 x4 for n = 4, and a4 = x5 x4 for n > 6. In every case, the subsequent action
max
of R3max ···Rn−1
is to push a4 , then a5 , etc. until an , through 1/1, leading to
(6.12)
max
a • R1max ···Rn−1
= x1 x4 /a4 /···/an /1/1.
MULTIFRACTION REDUCTION II
31
For n = 4, (6.12) reads a • R1max R2max R3max = x1 x4 /x1 x4 /1/1, and a further application of R1max
yields 1. For n > 6, the assumption that (x1 , ..., xn ) is a central cross for a implies that
(x1 , x4 , ..., xn ) is a central cross for the (n − 2)-multifraction x1 x4 /a4 /···/an . The induction
hypothesis for the latter gives redt (x1 x4 /a4 /···/an ) = 1, which expands into
max
= 1n−2 .
x1 x4 /a4 /···/an • RU(n−2)
(6.13)
max
By Lemma 2.17(iii), (6.13) implies x1 x4 /a4 /···/an /1/1 • RU(n−2)
= 1n . Merging with (6.12),
max
max
max
we deduce a • R1 ··· n−1 RU(n−2) = 1n , which is a • RU(n) = 1n , i.e., redt (a) = 1.
The argument is similar when a is negative.
We now concentrate on 4-multifractions. A sort of transitivity of central crosses holds.
Lemma 6.14. Assume that M is a gcd-monoid, a, b are 4-multifractions admitting a central
cross, and c1 a4 = c4 b1 and c2 a3 = c3 b2 holds. Then c1 a1 /c2 a2 /c3 b3 /c4 b4 admits a central cross.
In particular, a1 /a2 /b3 /b4 admits a central cross for a4 = b1 and a3 = b2 .
Proof. (Figure 7) Let (x1 , ..., x4 ) and (y1 , ..., y4 ) be central crosses for a and b, respectively.
By assumption, we have c2 x3 x4 = c2 a3 = c3 b2 = c3 y3 y2 , so x4 and y2 admit a common left
e y2 = xx4 = yy2 , and there exists z3 satisfying c2 a3 = z3 xx4 = z3 yy2 = c3 b2 .
multiple, say x4 ∨
Then a3 = x3 x4 and b2 = y3 y2 respectively imply
(6.15)
c2 x3 = z3 x
and
c3 y3 = z3 y.
Arguing similarly from c1 x1 x4 = c1 a4 = c4 b1 = c4 y1 y2 , we deduce the existence of z1 satisfying
c1 a4 = z1 xx4 = z1 yy2 = c4 b1 , leading to
(6.16)
c1 x1 = z1 x
and
c4 y1 = z1 y.
Then (z1 , xx2 , z3 , yy4 ) is a central cross for c1 a1 /c2 a2 /c3 b3 /c4 b4 .
c2
a2
a3
x3
x
x2
x4
c3
z3
b2
y2
x
a1
x1
y
y3
b3
y4
y
a4
z1 b1
y1
b4
c1
c4
Figure 7. Transitivity of the existence of a central cross.
We deduce that reduction preserves the existence of a central cross in both directions.
Lemma 6.17. Assume that M is a gcd-monoid and a, b are 4-multifractions on M satisfying
a ⇒∗ b. Then a admits a central cross if and only if b does.
Proof. It is enough to prove the result for a ⇒ b, say b = a • Ri,x . Assume that a is positive
and (x1 , ..., x4 ) is a central cross for a. Our aim is to construct a central cross for b from that
for a. Consider the case i = 2, see Figure 8. Let a2 x′ = x ∨ a2 . By definition, we have
(6.18)
b1 = a1 x′ ,
xb2 = a2 x′ ,
xb3 = a3 ,
b 4 = a4 .
By assumption, x and x3 admit a common right multiple, namely a3 , hence they admit a right
lcm, say x ∨ x3 = xy3 = x3 y. We have xb2 = a2 x′ = x3 (x2 x′ ), hence the right lcm of x and x3
left divides xb2 , i.e., we have xy3 6 xb2 , whence y3 6 b2 , say b2 = y3 y2 .
Next, we have x3 x2 x′ = a2 x′ = xb2 = xy3 y2 = x3 yy2 , whence x2 x′ = yy2 by left cancelling x3 . Put y1 = x1 y. We find b1 = a1 x′ = x1 x2 x′ = x1 yy2 = y1 y2 .
32
PATRICK DEHORNOY
On the other hand, by assumption, as both x and x3 left divide a3 , their right lcm x3 y left
divide a3 , which is x3 x4 . So we have x3 y 6 x3 x4 , whence y 6 y4 , say x4 = yy4 . Then we find
xb3 = a3 = x3 x4 = x3 yy4 = xy3 y4 , whence b3 = y3 y4 .
Finally, we have b4 = a4 = x1 x4 = x1 yy4 = y1 y4 . So (y1 , ..., y4 ) is a central cross for b.
The argument for i = 3 is similar, mutatis mutandis, and so is the one for i = 1: in this case,
the counterpart of x′ is trivial, which changes nothing. Finally, the case when a is negative is
treated symmetrically. So b admits a central cross whenever a does.
For the other direction, assume again that a and b are positive, and that (y1 , ..., y4 ) is a
central cross for b. Assume again i = 2, and (6.18). The equality xb2 = a2 x implies that
(a1 , 1, a2 , x′ ) is a central cross for a1 /a2 /xb2 /b1 . On the other hand, (y1 , y2 , xy3 , y4 ) is a central
cross for b1 /xb2 /a3 /a4 . So both a1 /a2 /xb2 /b1 and b1 /xb2 /a3 /a4 admit a central cross. By
Lemma 6.14, this implies that a1 /a2 /a3 /a4 admits a central cross. So a admits a central cross
whenever b does.
The above proof does not use the assumption that xb2 is the lcm of x and a2 , but only
the fact that the equalities of (6.18) hold for some x, x′ , which is the case, in particular, for
ei,x′ . So it also shows that every right reduct of a 4-multifraction with a central cross
b = a•R
admits a central cross, and conversely.
a2
a3
x
x3
b2
b3
y3 x4
x2
x′
y
a1
y2
b1 x
1
y4
y1
a4 = b 4
Figure 8. Construction of a central cross for a • Ri,x starting from one for a.
We deduce a complete description of the 4-multifractions that reduce to 1:
Proposition 6.19. If M is a gcd-monoid, then, for every 4-multifraction a on M , the following
are equivalent:
(i) The relation a ⇒∗ 1 holds.
(ii) The relation redt (a) = 1 holds.
(iii) The multifraction a admits a central cross.
Proof. The 4-multifractions 1/1/1/1 and /1/1/1/1 both admit the central cross (1, 1, 1, 1).
Hence, by Lemma 6.17, every 4-multifraction satisfying a ⇒∗ 1/1/1/1 or a ⇒∗ /1/1/1/1 admits
a central cross as well. Hence (i) implies (iii).
Next, Lemma 6.10 says that redt (a) = 1 holds for every multifraction a that admits a central
cross, so (iii) implies (ii).
Finally, (ii) implies (i) by definition.
±
We return to the study of n-semi-convergence for RM
, here for n = 4, 5.
Proposition 6.20. If M is a gcd-monoid, the following are equivalent:
(i) The system RM is 4-semi-convergent.
±
(ii) The system RM
is 4-semi-convergent.
±
(iii) The system RM
is 5-semi-convergent.
(iv) Every unital 4-multifraction in FM admits a central cross.
MULTIFRACTION REDUCTION II
33
Proof. Assume that RM is 4-semi-convergent. Let a be a unital positive 4-multifraction. As
RM is 4-semi-convergent, we have a ⇒∗ 1. By Proposition 6.19, we deduce that a admits a
central cross. Hence (i) implies (iv).
By definition, (ii) implies (i), and, by Lemma 6.2, (iii) implies (ii).
Finally, assume (iv), and let a be a unital 5-multifraction on M . Assume first that a is
positive. Then the positive 4-multifraction a5 a1 /a2 /a3 /a4 is unital as well, hence, by assumption, it admits a central cross (x1 , ..., x4 ), which expands into a = a1 /x3 x2 /x3 x4 /x1 x4 /a5 with
a5 a1 = x1 x2 . Then we obtain
a ⇒∗ a1 /x2 /1/x1 /a5
via D2,x3 D3,x4
∗
via R3,x1 R4,a5
∗
via D1,a1 D2,a5 .
⇒ a1 /x1 x2 /a5 /1/1 = a1 /a5 a1 /a5 /1/1
⇒ 1/1/1/1/1
If a is negative, then a2 /a3 /a4 /a1 a5 is unital and positive, hence admits a central cross by
assumption, leading to a = /a1 /x1 x2 /x3 x2 /x3 x4 /a5 with a1 a5 = x1 x4 , and to a ⇒∗ 1, this
time via D2,x2 D3,x3 R3,x4 R4,a5 D1,a1 D2,a5 . Thus, every unital 5-multifraction on M reduces
±
is 5-semi-convergent. So (iv) implies (iii).
to 1, and RM
We now establish alternative forms for the point (iv) in Proposition 6.20, connected with the
uniqueness of the expression by irreducible fractions.
Proposition 6.21. For every gcd-monoid M , the following are equivalent:
(i) Every unital 4-multifraction in FM admits a central cross.
(ii) For all a, b, c, d in M satisfying ι(a/b) = ι(c/d), there exist x, y in M satisfying
(6.22)
e b),
a = x(a ∧
e b),
b = y(a ∧
e d),
c = x(c ∧
e d).
d = y(c ∧
e b = 1 satisfy a 6 c and b 6 d.
(iii) All a, b, c, d in M satisfying ι(a/b) = ι(c/d) and a ∧
e b = c∧
e d = 1 satisfy a = c and b = d.
(iv) All a, b, c, d in M satisfying ι(a/b) = ι(c/d) and a ∧
e b = b ∧ c = 1 and ι(a/b/c) ∈ ι(M ) satisfy b = 1.
(v) All a, b, c in M satisfying a ∧
Before establishing Proposition 6.21, we begin with two characterizations of 4-multifractions
with a central cross:
Lemma 6.23. If M is a gcd-monoid, a positive 4-multifraction a on M admits a central cross
if and only if there exist x, y in M satisfying
(6.24)
e a2 ),
a1 = x(a1 ∧
′
′
e a2 ),
a2 = y(a1 ∧
if and only if there exist x , y in M satisfying
(6.25)
e a4 )x′ ,
a1 = (a1 ∧
e a3 )x′ ,
a2 = (a2 ∧
e a4 ),
a3 = y(a3 ∧
e a3 )y ′ ,
a3 = (a2 ∧
e a4 ),
a4 = x(a3 ∧
e a4 )y ′ .
a4 = (a1 ∧
e x3 )
Proof. Assume that a is positive and (x1 , ..., x4 ) is a central cross for a. Write x1 = x(x1 ∧
e x3 ). Then we have x ∧
e y = 1, and a1 = x(x1 ∧
e x3 )x2 , a2 = y(x1 ∧
e x3 )x2 ,
and x3 = y(x1 ∧
e a2 = (x1 ∧
e x3 )x2 , and a1 = x(a1 ∧
e a2 ), a2 = y(a1 ∧
e a2 ). On the other hand, we
whence a1 ∧
e x3 )x4 , a4 = x1 x4 = x(x1 ∧
e x3 )x4 , whence a3 ∧
e a4 = (x1 ∧
e x3 )x4 ,
also find a3 = x3 x4 = y(x1 ∧
e a4 ), d = x(a3 ∧
e a4 ). So we found x, y satisfying (6.24).
and a3 = y(a3 ∧
The proof for (6.25) is similar, writing x2 = (x2 ∧ x4 )x′ and x4 = (x2 ∧ x4 )y ′ and deducing
a1 ∧ a4 = x1 (x2 ∧ x4 ) and a2 ∧ a3 = x3 (x2 ∧ x4 ).
e a2 , y, a3 ∧
e a4 ) is a central cross for a,
In the other direction, if (6.24) is satisfied, (x, a1 ∧
whereas, if (6.25) is satisfied, so is (a1 ∧ a4 , x′ , a2 ∧ a3 , y).
Proof of Proposition 6.21. If ι(a/b) = ι(c/d) holds, the 4-multifraction a/b/d/c is unital, hence
it admits a central cross if (i) is true. In this case, Lemma 6.23 gives (6.22). So (i) implies (ii).
e b = 1 gives a = x and b = y, whence a 6 c and b 6 d. If,
Next, applying (ii) in the case a ∧
e d = 1, we similarly obtain c 6 a and d 6 b, whence a = c and b = d. So
in addition, we have c ∧
(ii) implies (iii) and (iv). On the other hand, (iv) implies (iii). Indeed, assume ι(a/b) = ι(c/d)
e b = 1. Let e = c ∧
e d, with c = c′ e and c = d′ e. Then we have c′ ∧
e d′ = 1 and
with a ∧
′
′
′
′
ι(a/b) = ι(c /d ). Then (iv) implies a = c 6 c and b = d 6 d.
34
PATRICK DEHORNOY
e b = b ∧ c = 1 and ι(a/b/c) = ι(d) for some d in M . Then
Now, let a, b, c in M satisfy a ∧
ι(a/b) = ι(d/c) holds so, if (ii) holds, we have a 6 d and b 6 c. Then the assumption b ∧ c = 1
implies b = 1. So (iii) implies (v).
Finally, assume that a is a unital 4-multifraction in FM . Write a2 = xa′2 and a3 = xa′3 with
e a′2 ,
x = a2 ∧ a3 , so that a′2 ∧ a′3 = 1 holds. Next, write a1 = a′1 y and a′2 = a′′2 with y = a1 ∧
′ e ′′
′
′′
′
′
′
so that a1 ∧ a2 = 1 holds. Then we have ι(a1 /a2 /a3 ) = ι(a1 /a2 /a3 ) = ι(a1 /a2 /a3 ), and the
assumption ι(a) = 1 implies ι(a1 /a2 /a3 ) = ι(a4 ) ∈ ι(M ). Hence ι(a′1 /a′′2 /a′3 ) lies in ι(M ) with
e a′′2 = 1 and a′2 ∧ a′3 = 1, whence a fortiori a′′2 ∧ a′3 = 1. If (v) is true, we deduce a′′2 = 1.
a′1 ∧
This means that (a′1 , y, x, a′3 ) is a central cross for a. So (v) implies (i).
Putting things together, we obtain:
Corollary 6.26. Conjectures A4 and B4 are equivalent, and they are equivalent to the property
that, for every Artin-Tits monoid M , every element of U(M ) of the form ab−1 with a, b in M
e b = 1.
admits only one such expression with a ∧
Proof. Let M be an Artin-Tits monoid and a be a unital 4-multifraction on M . By Proposition 6.19, a ⇒∗ 1 implies redt (a) = 1, so Conjecture A4 implies Conjecture B4 . On the other
hand, by Proposition 6.19 again, the property “a unital implies a ⇒∗ 1” is equivalent to “ a
unital implies a admits a central cross” and, by Proposition 6.21, the latter is equivalent to the
uniqueness of fractional decompositions.
6.3. An application to partial orderings on U(M ). If M is a Garside monoid, it is
known [16, Sec. II.3.2] that the left and right divisibility relations of M can be extended into
e h) for
well-defined partial orders on the enveloping group U(M ) by declaring g 6 h (resp., g 6
g −1 h ∈ M (resp., gh−1 ∈ M ). The construction extends to every gcd-monoid, and we show
±
that lattice properties are preserved whenever RM
is 4-semi-convergent.
Proposition 6.27. (i) If M is a gcd-monoid that embeds into U(M ), then declaring g 6 h for
g −1 h ∈ ι(M ) provides a partial order on U(M ) that extends left divisibility on M .
±
(ii) If RM
is 4-semi-convergent, any two elements of U(M ) that admit a common 6-lower
bound (resp., a common 6-upper bound) admit a greatest one (resp., a lowest one).
Proof. (i) We identify M with its image under ι. As M is a semigroup in U(M ), the relation 6
is transitive on U(M ), and it is antisymmetric, as 1 is the only invertible element of M . Hence
6 is a partial order on U(M ). By definition, it extends left divisibility on M .
(ii) Assume that g and h admit a common 6-lower bound f (see Figure 9 left). This means
that we have g = f x and h = f y for some x, y in M . Let f0 := f (x ∧ y). Write x = (x ∧ y)x0 ,
y = (x ∧ y)y0 . Then we have g = f0 x0 and h = f0 y0 , whence f0 6 g and f0 6 h. Now
assume that f1 is any common 6-lower bound of g and h, say g = f1 x1 and h = f1 y1 . In the
−1
group U(M ), we have x0 x−1
1 = y0 y1 . Hence, the 4-multifraction x0 /x1 /y1 /y0 is unital, so, by
assumption, it admits a central cross. Then Lemma 6.23 provides x′ , y ′ satisfying
x0 = (x0 ∧ y0 )x′ ,
x1 = (x1 ∧ y1 )x′ ,
y1 = (x1 ∧ y1 )y ′ ,
y0 = (x0 ∧ y0 )y ′ .
By definition of f0 , we have x0 ∧ y0 = 1, whence x′ = x0 and y ′ = y0 , and, from there,
f1 (x1 ∧ y1 ) = f0 , hence f1 6 f0 , in U(M ). So f0 is a greatest 6-lower bound for g and h.
The argument for lowest 6-upper bound is symmetric. Assume that g and h admit a common
e y) and y = y0 (x ∧
e y).
6-upper bound f (see Figure 9 right). Write f = gx = hy, then x = x0 (x ∧
e
e
Put f0 := gx0 . We have f0 (x ∧ y) = f = hy = hy0 (x ∧ y), whence hy0 = f0 . So f0 is a common
6-upper bound of g and h. Now assume that f1 is any common 6-upper bound of g and h,
say f1 = g1 x1 = hy1 . In U(M ), we have x0 y0−1 = x1 y1−1 , i.e., the 4-multifraction x0 /y0 /y1 /x1
is unital, hence it admits a central cross. Then Lemma 6.23 provides x, y satisfying
e y0 ),
x0 = x(x0 ∧
e y0 ),
y0 = y(x0 ∧
e y1 ),
x1 = x(x1 ∧
e y1 ).
y1 = y(x1 ∧
e y0 = 1, whence x = x0 and y = y0 , leading in U(M ) to
The definition of f0 gives x0 ∧
e y1 ) = f1 , hence f0 6 f1 . So f0 is a lowest 6-upper bound for g and h.
f0 (x1 ∧
MULTIFRACTION REDUCTION II
f
y
x∧
f1
y f
0
x0
x
y0
∧y
0
x0
′
x
g
h
x1
g
′
x1
y
∧y
y1
y1
x1 e
∧y
1
h
x
x0
1
x1
35
y
x0 e
∧y
0
x
f1
y0
f0
x∧
e
y
y
f
Figure 9. Conditional lattice property for the poset (U(M ), 6): greatest lower
bound on the left, lowest upper bound on the right.
Of course, we have a symmetric extension for the right divisibility relation.
Question 6.28. Is the assumption that M embeds in U(M ) sufficient to ensure that conditional
greatest lower bounds and lowest upper bounds for 6 exist in U(M )?
6.4. Depth 6 and beyond. The results established in the case of 4-multifractions do not
extend to depth 6 and beyond. In particular, unital 6-multifractions need not admit a central
e 2 , let a := aba/aca/cac/cbc/bcb/bab and b :=
cross: in the Artin-Tits monoid of type A
ab/ac/ca/cb/bc/ba. We have a ⇒ b ⇒∗ 1, and a admits a central cross, but b does not, as it
is prime. So the counterparts of Lemma 6.17 and Proposition 6.19 fail. However, we shall see
now that, for every n, Conjecture Bn implies a geometrical property of van Kampen diagrams
that directly extends Prop 6.19.
What the latter says is that, if we define (Γ4 , ∗) to be the pointed
graph on the right, then, for every 4-multifraction a that reduces to 1,
there exists an M -labeling of the edges of Γ4 such that the outer labels
from ∗ are a1 , ..., a4 and the labels in each triangle induce equalities in M .
∗
If (Γ, ∗) is a finite, simply connected pointed graph, let us say that a multifraction a on a
monoid M admits a van Kampen diagram of shape Γ if there is an M -labeling of Γ such that
the outer labels from ∗ are a1 , ..., an and the labels in each triangle induce equalities in M .
This notion is a mild extension of the usual one: if S is any generating set for M , then
replacing the elements of M with words in S and equalities with word equivalence provides a
van Kampen diagram in the usual sense for the word in S ∪ S then associated with a. Then,
Proposition 6.19 says that every 4-multifraction reducing to 1 admits a van Kampen diagram
of shape Γ4 . Conjecture B predicts similar results for every depth.
Definition 6.29. For n > 6 even, let (Γn , ∗) be the graph obtained by appending n−2 adjacent
copies of Γ4 around (Γn−2 , ∗) starting from ∗, with alternating orientations, and connecting the
last copy of (Γ4 , ∗) with the first one, see Figure 10.
Proposition 6.30. Let M be an Artin-Tits monoid. If Conjecture B is true, then every unital
n-multifraction on M (with n even) admits a van Kampen diagram of shape Γn , see Figure 10.
max
= 1.
Proof. Let a be a unital n-multifraction on M . Conjecture B predicts the equality a • RU(n)
We shall see that the latter (and, more generally, any equality of the form a • RU(n),x = 1, with
x maximal or not) implies that a admits a van Kampen diagram of shape Γn . In view on
an induction, assume a • R1,x1 ···Rn−1,xn−1 = b · 12 , for some (necessarily unital) (n − 2)multifraction b. Let a0 := a and, inductively, ai := ai−1 • Ri,xi . We start from a loop of
edges with alternating orientations labeled a0 . Then we inductively complete the graph using
n − 1 steps of the type
aii
aii
aii−1
aii−1
ai
ai
i+1
or
xi
ai+1
i−1
ai+1
i
i+1
ai+1
i+1
xi
ai+1
i−1
ai+1
i
ai+1
i+1
36
PATRICK DEHORNOY
ac
aca
ca
ca
1
ab
ba
1
bc bc
c
ba
b
cb
ba ba
1
c
a
a
ac
ab
cb
ab
1 cb
1
1
bc
bc
1
ab
ab
cb
cb
∗
∗
Figure 10. The graph Γ6 : on the left, the naked graph, with four juxtaposed
copies of Γ4 around one (colored) copy of Γ4 ; it has fourteen vertices, namely four
springs (one inner), five wells (two inner), and five 4-prongs; on the right, for M
e 2 , the M -labeling of Γ6 that results from the first
the Artin-Tits monoid of type A
reduction in Example 3.6; it provides a van Kampen diagram for the 6-multifraction
ab/ba/ca/ac/bc/ab. Conjecture B predicts that every 6-multifraction representing 1
in an Artin-Tits monoid admits a van Kampen diagram of shape Γ6 .
according to the sign of i in a. Because the final two entries of an−1 are trivial, the last two
n−1
steps take a simpler form: ann−1 = 1 is equivalent to an = xn−1 , whereas an−1
= 1 is equivalent
−1
e an .
to xn−2 an−1 6
a4
a5
We thus obtain an M -labeling of an annux
4
x3
lar graph made of n − 2 copies of Γ4 , whose
outer boundary is labeled a, and whose ina3
a6
x5
ner boundary is labeled b, see the diagram
x2
b4
b3
on the right, here in the case going from 8
b2
b5
to 6. Then we repeat the process with b,
x6
a
a7
2
etc., until a 4-multifraction is reached, and
b1
b6
x1
we conclude using Proposition 6.19. This
construction exactly corresponds to the ina8 = x7
a1
ductive definition of Γn .
It is well-known that, if a multifraction represents 1 in a group, then it admits a van Kampen
diagram in the sense defined above. What is remarkable here is the existence of one single
universal shape, with prescribed springs, wells, and 4-prongs, that works for every unital nmultifraction at the same time.
Finally, one may wonder whether some counterparts of Lemma 6.10 and 6.17 might hold
with Γn replacing Γ4 : maybe they do, but the natural argument for proving them requires that
the ground monoid satisfies the 3-Ore condition, in which case it is known that every unital
n-multifraction admits a van Kampen diagram of shape Γn , thus making the results trivial.
7. Miscellanea
The main three properties addressed in this paper are Conjectures A, B, and C (together
with the uniform version Cunif of the latter), which involve arbitrary Artin-Tits monoids, and
are known to be true for those of FC type. Testing these statements with a computer is easy,
and we report about experiments that, alltogether, support the conjectures and provide some
experimental evidence. The involved program is available at [14], and the experiments are easy
to repeat and confirm.
We begin with a few remarks about implementation options (Subsection 7.1), then report
about the obtained data (Subsection 7.2). Finally, we conclude with a few hints about further
properties of reduction, including several counter-examples (Subsection 7.3).
MULTIFRACTION REDUCTION II
37
7.1. Implementation options.
±
Choice of the monoid. We are interested in Artin-Tits monoids M such that RM
is not convergent, hence not of FC type. It is natural to look for monoids with a maximal convergence
defect, meaning that the proportion of multifractions with more than one irreducible reduct is
e 2 , and,
maximal. As can be expected, the ratio is maximal for the Artin-Tits monoid of type A
more generally, those with all relations of length 3 exactly: as divergent reducts may arise only
with counter-examples to the 3-Ore property, it is natural that this happens more frequently
e 2 and, more generwhen all atoms give rise to such counter-examples. Another advantage of A
ally, Kn,3 , the Artin-Tits monoid whose Dynkin diagram is the complete graph with n vertices
and all edges labeled 3, is the existence of an explicit description of basic elements (namely 1,
atoms, and products of two distinct atoms) providing a better efficiency (and 100% correctness
with no termination problem) for the implementation of the monoid operations (equality, lcms,
e 2 , considered as the critical type (but any
gcds, etc). Therefore, we mostly concentrated on A
other choice is possible with [14]).
Generation of random multifractions. Exhaustively enumerating multifractions up to a given
length (sum of the lengths of the entries) is difficult, as, even in the case of 3 atoms, there
are more than 2.6 × 109 multifractions of length up to 12. Therefore it is more realistic to use
samples of random multifractions. Generating random elements of the monoid and, from there,
random multifractions, is easy via random words in the atom alphabet (with a bias due to the
relations).
Generating random unital multifractions is more delicate. As the density of unital multifractions is negligible, generating random multifractions and selecting those that are unital is not a
e2
good option (in addition, it requires a prior solution to the word problem, which exists for A
but not in general). Two methods have been used. The first one (“brownian motion”) is to
follow the definition of ≃, thus starting with an empty word and randomly adding or deleting
pairs s/s and applying the Artin-Tits relations. Inserting right and left reversing steps (the
special transformations of Property H, see Subsection 3.3) improves the efficiency.
The second method (“lcm-expansions”) consists in starting from a multifraction that admits
a random central cross, hence is unital of a very special type, and deriving new, more generic,
unital multifractions as follows:
Definition 7.1. (Figure 11) If M is a gcd-monoid and a is a unital n-multifraction on M , with
n even, we say that b is an lcm-expansion of a if kbk = kak holds and, for each i, there exist
decompositions ai = a′i a′′i , bi = b′i b′′i satisfying a′i b′′i−1 = a′i+1 b′′i = a′i ∨ a′i+1 for i negative in a,
e a′′i+1 for i positive in a, with indices modulo n, i.e., n + 1 means 1.
and b′i−1 a′′i = b′i a′′i+1 = a′′i ∨
...
...
b′i−2
bi−2
b′′i−2
a′′i−1 a′i−1
ai−1
b′′i−1
bi−1
a′i
ai
bi
b′i−1
b′i
a′′i
a′′i+1 a′i+1
ai−1
b′′i
b′′i+1
bi+1
a′i+2
b′i+1
a′′i+2
ai
...
...
Figure 11. Lcm-expansion, here for i positive in a; the shift of the indices for b
ensures that b has the same sign as a. Note the symmetry: starting with right divisors
leads to the same notion.
For b an lcm-expansion of a, one reads on Figure 11 the equality ι(b) = ι(a′1 b′′n )−1 ι(a)ι(a′1 b′′n ),
hence lcm-expansion preserves unitality. Constructing lcm-expansions of a is easy: with the
notation of Definition 7.1, once a left divisor a′i of ai is chosen for each i, all the remaining
elements a′′i , b′i , b′′i are determined, and then so is b (but the choice leads to an lcm-expansion
only if the lcms exist). The advantage is that the depth is controlled (which is more difficult with
brownian motion), the inconvenience is that there is no guarantee that generic multifractions
are obtained (but the expansion procedure can be iterated).
38
PATRICK DEHORNOY
Maximal vs. atomic reduction steps. Implementing reduction is straightforward, once the lattice
operations of the monoid are available. As a composition of i-reductions is again an i-reduction,
one might think of restricting to maximal reduction steps, i.e., considering reducts a • Ri,x where
x is a maximal i-reducer for a. This is not a good option, as some reducts may be missed:
e 2 , let a = ab/ba/ca/bcbc. The only maximal
Example 7.2. In the Artin-Tits monoid of type A
reductions from a are R3,cc followed by D1,ab , leading to the unique irreducible 1/ab/ca/cb.
However, one also finds a • R3,c D1,ab R2,c R3,b = cb/abbc/ba/bc, a second irreducible reduct
of a not reachable by maximal steps.
This however does not contradict the following (surprising?) result:
Proposition 7.3. If M is a strongly noetherian gcd-monoid, and b, c are irreducible reducts
of some multifraction, then there exists a finite sequence of maximal reductions and inverses of
maximal reductions connecting b to c.
Proof (sketch). Write b ⊲⊳ c when there exists a sequence as in the statement. Using induction
on the ordinal λ(a), where λ satisfies (3.13), one proves the following general criterion: If ⊲⊳ is an
equivalence relation on FM± such that, for every multifraction a that is i-prime and j-irreducible
for j 6= i and every 1-reduct a′ of a, there exists a irreducible reduct b of a′ satisfying a ⊲⊳ b,
then, for every a in FM± , any two irreducible reducts of a are ⊲⊳-equivalent.
7.2. Experimental data.
Conjectures A and B. Testing the two of them is essentially the same thing: one generates a
random unital multifraction a, and one checks a ⇒∗ 1 in the former case, redt (a) = 1 is the
latter. Because semi-convergence implies 1-confluence, one can fix any reduction strategy for
checking a ⇒∗ 1, for instance looking at each step for the smallest level i and the first atom x
such that Ri,x applies. Although computing redt is slightly slower, as, at each step, all i-reducers
have to be determined in order to take their gcds, both computations are fast. Precise numbers
e 2 or K4,3 , the typical order of magnitude is 5 × 104
are not really significant here; in type A
4
(resp., 1.5 × 10 ) random unital multifractions of length 20 and depth 4 (resp., length 40 and
e 3 , or C
e 2 ), efficiency is diminished by
depth 6) per hour of computation. In other types (e.g., A
a factor 10 approximately.
No counter-example to Conjecture A or B was ever found. As the density of visited (unital)
multifractions becomes negligible when the length grows, the significance of such data is questionable. However, it may be noticed that, for the many properties considered in this paper
and discarded by counter-examples, the length of the latter (all found by random search) is
never more than 12 or so: this does not say anything about a possible counter-example to
Conjecture A or B but, at the least, this shows that the considered lengths are not ridiculous.
Conjecture A4 . For the special case of Conjecture A4 (and of Conjecture B4 , which, by Corollary 6.26, is equivalent), an exhaustive search makes sense, by systematically considering all
possibilities for the central cross (up to a certain length).
e 2 , Conjecture A4 is true for all lcm-expansions
Fact 7.4. For the Artin-Tits monoid of type A
of all multifractions that admits a central cross with entries of length at most 2.
By contrast, the procedure applied to the monoid MC,4 of [19, Proposition 6.9] duly finds a
counter-example, namely an irreducible lcm-expansion of a 4-multifraction with a central cross
(with rays of length 1). What seems to discard a similar counter-example in an Artin-Tits
monoid is the fact that Artin-Tits relations preserve the atoms occurring in an element (the
“support”), but, even for this weak form of Conjecture A4 , we have no proof so far.
Conjectures C and Cunif . Testing these conjectures is easier in that it involves arbitrary, not
necessarily unital multifractions, but it is more difficult in that it requires to determine all
right reducts of a multifraction a and then, for each pair of them, to determine all their left
reducts. The complexity of constructing the tree Ta (as used in the proof of Proposition 5.29)
MULTIFRACTION REDUCTION II
39
and its right counterpart Tea increases quickly with the depth. Beyond depth 4, the total number
of reducts often becomes large (usually a few ones, but possibly several thousands), resulting
in a huge computation time. Typically, in the current version of the program, one can test
Conjecture Cunif for about 5 × 103 (resp., 4 × 103 ) random multifractions of length 20 (resp.,
length 30) and depth 3 per hour of computation. Going to depth 4 diminishes the speed by a
factor 50.
To overcome these bounds, we also tested (without size limitation) the following instance
of Conjecture C: starting with a multifraction a, the determine the (not necessarily distinct)
right reducts b1 , ..., b4 and left reducts c1 , ..., c4 of a obtained using the four natural strategies
(levels from bottom or top, atoms in lexicographical or antilexicographical order), and check
the property ∃k ∀j (bj ⇒∗ ck ), thus checking Conjecture C for {b1 , ..., b4 }. The cost is then
comparable as the one for Conjecture B, with about 5 × 104 (resp., 5 × 103 ) random tries for
length 20 (resp., 30) per hour of computation. The stronger conclusion ∀k ∀j (bj ⇒∗ ck ) is
almost always valid but, as in the case of Figure 6, exceptions occur.
7.3. Further questions. We point to a few natural questions involving reduction. Most of
them remain open, or gave rise to counter-examples.
Normal forms for reductions. Distinguished expressions for sequences of reductions could be
obtained by identifying skew commutation relations, typically of the form Ri,x Rj,y ⇛ Rj,y′ Ri,x′ ,
meaning that, if a • Ri,x Rj,y is defined, then so is a • Rj,y′ Ri,x′ and the results are equal.
Typically, one could try to push divisions to one side, so that all remaining steps are invertible.
This approach does not work well, as exceptions always appear. In the same vein, in view of
a possible induction and building on the universal scheme U (n) that works in the 3-Ore case,
one could conjecture that every reduction sequence is equivalent to one where the highest level
occurs only once, or that, if an n-multifraction a is i-irreducible for i < n − 1, then reducing a
can be done by a sequence of the form Rn−1,x1 Rn−3,x3 Rn−5,x5 ···. This need not be the case.
e 2 , let a = 1/ba/cb/ca/ab. Then a is i-irreducible
Example 7.5. In the Artin-Tits of type A
for i 6 3, but the only sequence from a to an irreducible reduct is R4,a R2,bc R3,a R4,b , discarding
the above two conjectures. Of course, Conjecture B is not contradicted, because a is not unital.
In the same direction, one can study local confluence between left reductions Ri,x and right
ej,y , with the hope of obtaining normal forms useful for cross-conflence. In almost
reductions R
all cases, there exists indeed local confluence solutions. However, the case j = i + 2 remains
problematic in general. Moreover, using local confluence for an induction is unclear, because
there is no common well-founded relation underlying both left and right reduction and, except
in type FC, a multifraction may admit infinitely many left-right reducts.
Homomorphisms. As reduction is constructed using multiplication and lcm operations, it is
preserved by morphisms preserving these operations, namely lcm-morphisms [8]: if φ is an lcmmorphism from a gcd-monoid M to a gcd-monoid M ′ , then a ⇒∗ b implies φ(a) ⇒∗ φ(b) for all
a, b in FM± . This however is not easy to use for, say, Conjecture A, because the implications
±
go in the wrong direction. If if we study the semi-convergence of RM
, mapping M to a gcd′
monoid M that satisfies the 3-Ore condition does not help: if a is unital in FM± , then φ(a) is
unital in FM±′ , so φ(a) ⇒∗ 1 holds, but deducing a ⇒∗ 1 is problematic. In the other direction,
if M ′ satisfies the 3-Ore condition and φ′ is an lcm-morphism from M ′ to M , then Imφ′ is
±
included in some part of M where the 3-Ore condition is satisfied, and knowing that RM
′ is
′
(semi)-convergent will not help for multifractions on M outside Imφ .
Morphisms (lcm-morphisms or not) might be useful for establishing particular properties,
like Conjecture A4 , i.e., the fact that every unital 4-multifraction admits a central cross. If M is
e 2 , one may think of using the classical embedding from M to the
the Artin-Tits monoid of type A
Artin-Tits group of type B3 , namely ha, b, c | abab = baba, bcb = cbc, ac = cai, that maps a
to a−1 b−1 cba, and b and c to themselves, but the image is not included in the monoid. A
probably better choice is to map M to the Artin-Tits monoid M ′ of type D4 , namely ha, b, c, d |
40
PATRICK DEHORNOY
ada = dad, bdb = dbd, cdc = dcdi+ , by φ(s) = ds for s = a, b, c. If a is a unital 4-multifraction
on M , then φ(a) admits a central cross in M ′ : to deduce that a admits a central cross in M ,
it suffices to show that at least one of the central crosses for φ(a) lies in Imφ. The problem is
that Imφ is not closed under right divisor in M ′ : for instance, φ(a) 6 φ(bac) holds in M ′ , but
a 6 bac fails in M .
By the way, the following natural question seems to be open:
Question 7.6 (F. Wehrung). Does the above morphism φ induce an embedding from the
e 2 into the Artin-Tits group of type D4 ?
Artin-Tits group of type A
Reduction graphs. Almost nothing is known about the structural properties of the graph formed
by the reducts of a multifraction, for instance their possible lattice properties: if we have a ⇒∗ b
and a ⇒∗ c and if b and c admit a common reduct, does there exist a common reduct d of b
and c such that every common reduct of b and c is a reduct of d? This is frequently true, but
not always:
e 2 , consider a = 1/a/bcb/bcb/a (which is
Example 7.7. In the Artin-Tits monoid of type A
unital). One finds b = a • R2,b = ba/ab/cb/bcb/a and c = a • R2,c = ca/ac/bc/bcb/a. Then b
and c admit the two symmetric maximal common reducts, namely
d′ = b • D3,b = c • R3,b D1,ca R2,b = ba/ab/c/bc/a,
d′′ = b • R3,c D1,ba R2,c = c • D3,c = ca/ac/b/cb/a,
and there is no common reduct d of b and c of which d′ and d′′ are reducts.
A possible conclusion in view of the long list of counter-examples described in this paper
could be that there is no hope for many further general properties of reduction, implying that
a possible proof of semi-convergence has to involve the specific properties of the ground monoid
in a deep way. In particular, a proof of Conjecture A, B, or C should require developing new
specific tools for Artin-Tits groups. The results of [19] may suggest approaches.
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
J. Altobelli. The word problem for Artin groups of type FC. J. Pure Appl. Algebra, 129:1–22, 1998.
K.I. Appel and P.E. Schupp. Artin groups and infinite Coxeter groups. Invent. Math., 72:201–220 1983.
E. Artin. Theory of Braids. Ann. of Math., 48:101–126, 1947.
E. Brieskorn. Sur les groupe de tresses. In Séminaire Bourbaki 1971/72, volume 317 of Lectures Notes in
Math., pages 21–44. Springer, 1973.
E. Brieskorn and K. Saito. Artin-Gruppen und Coxeter-Gruppen. Invent. Math., 17:245–271, 1972.
R. Charney. Problems related to Artin Groups. Amer. Inst. of Math., Problems in Geometric Group Theory
wiki, http://people.brandeis.edu/∼charney/papers/Artin probs3.pdf, http://aimath.org/pggt.
A. Chermak. Locally non-spherical Artin groups. J. Algebra, 200:56–98, 1998.
J. Crisp. Injective maps between Artin groups. In Geom. Group Theory Down Under, pages 119–137. de
Gruyter, Berlin, 1999. (Canberra 1996).
P. Dehornoy. Preuve de la conjecture d’irréflexivité pour les structures distributives libres. C. R. Acad. Sci.
Paris, 314:333–336, 1992.
P. Dehornoy. A fast method for comparing braids. Adv. in Math., 125:200–235, 1997.
P. Dehornoy. Groupes de Garside. Ann. scient. Éc. Norm. Sup. 4e série, 35:267–306, 2002.
P. Dehornoy. The word reversing method. Intern. J. Alg. and Comput., 21:71–118, 2011.
P. Dehornoy. Multifraction reduction I: The 3-Ore condition and Artin–Tits groups of type FC. J. Comb.
Algebra, to appear, arXiv:1606.08991.
P. Dehornoy. MoKa. Executable MacOS/Docker binaries; www.math.unicaen.fr/∼dehornoy/programs.
P. Dehornoy, M. Dyer, and C. Hohlweg. Garside families in Artin–Tits monoids and low elements in Coxeter
groups. Comptes Rendus Math., 353:403–408, 2015.
P. Dehornoy, with F. Digne, E. Godelle, D. Krammer, and J. Michel. Foundations of Garside Theory.
Number 22 in EMS Tracts in Mathematics. Europ. Math. Soc., 2015.
P. Dehornoy and E. Godelle. A conjecture about Artin–Tits groups. J. Pure Appl. Algebra, 217:741–756,
2013.
P. Dehornoy, D. Holt, and S. Rees. Multifraction reduction IV: Padding and Artin–Tits groups of sufficiently
large type. arXiv:1701.06413.
P. Dehornoy and F. Wehrung. Multifraction reduction III: The case of interval monoids. J. Comb. Algebra,
to appear, arXiv:1606.09018.
MULTIFRACTION REDUCTION II
41
[20] P. Deligne. Les immeubles des groupes de tresses généralisés. Invent. Math., 17:273–302, 1972.
[21] M. Dyer and C. Hohlweg. Small roots, low elements, and the weak order in Coxeter groups. Advances in
Math., 301:769–784, 2016.
[22] D. Epstein, with J. Cannon, D. Holt, S. Levy, M. Paterson & W. Thurston. Word Processing in Groups
Jones & Bartlett Publ (1992).
[23] E. Godelle and L. Paris. Basic questions on Artin-Tits groups. Geometry, Combinatorics, and Topology,
pages 399–311, 2012.
[24] E. Godelle and L. Paris. K(π, 1) and word problems for infinite type Artin-Tits groups, and applications
to virtual braid groups. Math. Z., 272:1339–1364, 2012.
[25] E. Godelle and L. Paris. PreGarside monoids and groups, parabolicity, amalgamation, and FC property.
Int. J. Alg. Comput., 23:1431–1467, 2013.
[26] E. Godelle and S. Rees. Rewriting systems in sufficiently large Artin-Tits groups. arXiv:1511.02789.
[27] S. Hermiller and J. Meier. Artin groups, rewriting systems and three-manifolds. J. Pure Appl. Algebra,
36(2):141–156, 1999.
[28] L. Paris. Artin monoids embed in their groups. Comment. Math. Helv., 77(3):609–637, 2002.
[29] S.J. Pride. On Tits’ conjecture and other questions concerning Artin and generalized Artin groups. Invent.
Math., 86(2):347–356, 1986.
[30] K. Tatsuoka. An isoperimetric inequality for Artin groups of finite type. Trans. Amer. Math. Soc.,
339(2):537–551, 1993.
Laboratoire de Mathématiques Nicolas Oresme, CNRS UMR 6139, Université de Caen, 14032 Caen
cedex, France, and Institut Universitaire de France
E-mail address: [email protected]
URL: www.math.unicaen.fr/∼dehornoy
| 4math.GR
|
Nilpotent Evolution Algebras over arbitrary fields
arXiv:1508.06860v1 [math.AC] 27 Aug 2015
A. S. Hegazi, Hani Abdelwahab
Mathematics department, Faculty of Science, Mansoura University, Egypt
[email protected]
[email protected]
Abstract
The paper is devoted to the study of annihilator extensions of evolution algebras and suggests
an approach to classify finite-dimensional nilpotent evolution algebras. Subsequently nilpotent
evolution algebras of dimension up to four are classified.
Keywords and phrases: Nilpotent evolution algebra, Annihilator extension, Automorphism
group.
2010 Mathematics Subject Classification: 17D92, 17D99.
1
Introduction
Evolution algebras were introduced in 2006 by J.P. Tian and P. Vojtechovsky in their paper “Mathematical concepts of evolution algebras in non-Mendelian genetics ” (see [10]). Later on, Tian laid the
foundations of evolution algebras in his monograph [11]. These algebras present many connections with
other mathematical fields including graph theory, group theory, Markov chains, dynamical systems,
knot theory, 3-manifolds and the study of the Riemann-Zeta function (see [11]).
Evolution algebras are in general non-associative and do not belong to any of the well-known classes
of non-associative algebras such as Lie algebras, alternative algebras, or Jordan algebras. Therefore,
the research on these algebras follows different paths (see [2], [3], [4], [8], [12]).
One of the classical problems in study of any class of algebras is to know how many different (up
to isomorphism) algebras exist for each dimension. In this way in [1], [6], [7] and [9], the classifications
of nilpotent Jordan algebras, nilpotent Lie superalgebras, nilpotent associative algebras and nilpotent
Lie algebras of low dimensions were given.
In this paper we study the class of nilpotent evolution algebras. Our aim is to describe a method
for classifying nilpotent evolution algebras. In [3], the equivalence between nilpotent evolution algebras
and evolution algebras which are defined by upper triangular matrices is proved. In [11], J. Tian defined
an evolution algebra associated to any directed graph. In [5], A. Elduque and A. Labra considered the
reverse direction, a directed graph is attached to any evolution algebra and they proved that nilpotency
of an evolution algebra is equivalent to the nonexistence of oriented cycles in the attached directed
graph.
The paper is organized as follows. In Section 2, we give some basic concepts about evolution algebras. Section 3 is devoted to the description of construction of evolution algebras with non-trivial
annihilator as annihilator extensions of evolution algebras of lower dimensions. In Section 4, we describe
a method for classifying nilpotent evolution algebras. In Section 5, we classify nilpotent evolution algebras of dimension up to three over arbitrary fields. In Section 6, four-dimensional nilpotent evolution
algebras are classified over an algebraic closed field of any characteristic and over R.
2
Preliminaries
Definition 1 [11] An evolution algebra is an algebra E containing a basis (as a vector space) B =
{e1 , . . . , en } such that ei ej = 0 for any 1 ≤ i < j ≤ n. A basis with this property is called a natural
basis.
Given a natural basis B = {e1 , . . . , en } of an evolution algebra E,
e2i =
n
X
j=1
1
αij ej
for some scalars αij ∈ F, 1 ≤ i, j ≤ n. The matrix A = αij is the matrix of structural constants of
the evolution algebra E, relative to the natural basis B.
Definition 2 An ideal I of an evolution algebra E is an evolution algebra satisfying EI ⊆ I.
Given an evolution algebra E, consider its annihilator
ann (E) := {x ∈ E : xE = 0} .
Lemma 3 [5, Lemma 2.7] Let B = {e1 , . . . , en } be a natural basis of an evolution algebra E. Then
ann (E) = span ei : e2i = 0 .
Hence ann (E) is an ideal of the evolution algebra E. Let B ′ = {er+1 , . . . , en } be a natural basis of ann (E). Then the quotient E/ann (E) is an evolution algebra with a natural basis
B ′′ = {ēi = ei + ann (E) : 1 ≤ i ≤ r}.
Note that we can reorder the natural basis B of an evolution algebra E so that B ′ = {er+1 , . . . , en }
be a natural basis of ann (E) .
Given an evolution algebra, we introduce the following sequence of subspaces:
E h1i = E, E hk+1i = E hki E.
Definition 4 An algebra E is called nilpotent if there exists n ∈ N such that E hni = 0, and the minimal
such number is called the index of nilpotency.
One can easily verify that if E is nilpotent then ann (E) 6= 0. Moreover, E is nilpotent if and only
if E/ann (E) is nilpotent.
3
Constructing Nilpotent evolution Algebra
Let E be an evolution algebra with a natural basis BE = {e1 , . . . , em } and V be a vector space with
a basis BV = {em+1 , . . . , en }. Let θ : E × E −→V be a bilinear map on E, set Eθ = E⊕V define a
multiplication on Eθ by:
ei ⋆ ej = ei ej |E +θ (ei , ej ) if 1 ≤ i, j ≤ m, and otherwise ei ⋆ ej = 0.
Then Eθ is an evolution algebra with a natural basis BEθ = {e1 , . . . , en } if and only if θ (ei , ej ) = 0 for
all i 6= j. Denote the space of all bilinear θ such that θ (ei , ej ) = 0 for all i 6= j by Z (E × E, V ). For
θ ∈ Z (E × E, V ) with dim V = n − m, the evolution algebra Eθ = E⊕V is called a (n − m)-dimensional
annihilator extension of E by V.
Now, for a linear map f ∈ Hom (E, V ), if we define θ f : E × E −→V by θ f (ei , ej ) = f (ei ej ),
then θf ∈ Z (E × E, V ). One can easily verify that the space defined by B (E × E, V ) = {θf : f ∈
Hom (E, V )} is a subspace of Z (E × E, V ).
Lemma 5 Let E be an evolution algebra with a natural basis BE = {e1 , . . . , em } and V be a vector
space with a basis BV = {em+1 , . . . , en }. Suppose that θ ∈ Z (E × E, V ) and θf ∈ B (E × E, V ). Then
Eθ ∼
= Eθ+θf .
Proof. Let x =
n
P
αi ei ∈ Eθ , define a linear map σ : Eθ −→ Eθ+θf by
i=1
σ
n
X
i=1
αi ei
!
n
X
α i ei + f
=
i=1
m
X
α i ei
i=1
Then σ is an invertible linear transformation. Moreover, let x =
n
P
!
α i ei , y =
i=1
σ (x ⋆Eθ y) =
σ
m
X
i=1
=
m
X
i=1
αi β i e2i
.
n
P
β i ei ∈ Eθ then
i=1
!
m
X
αi β i e2i + θ (ei , ei ) + f e2i
=
+ θ (ei , ei )
i=1
αi β i e2i + (θ + θf ) (ei , ei ) = σ (x) ⋆Eθ+θf σ (y) .
2
Hence Eθ ∼
= Eθ+θf .
Hence the isomorphism type of Eθ only depends on the element θ + B (E × E, V ). Therefore we
consider the quotient space
H (E × E, V ) = Z (E × E, V ) /B (E × E, V ) .
The annihilator extension corresponding to θ = 0, and consequently corresponding to any element in
B (E × E, V ) is called a trivial extension.
Suppose that E is an evolution algebra with a natural basis BE = {e1 , . . . , em } and V be a vector
space with a basis BV = {em+1 , . . . , en }. Then a symmetric bilinear map θ ∈ Z (E × E, V ) can be
n−m
P
θj (ei , ei ) em+j , where θj ∈ Z (E × E, F). Further, θ ∈ B (E × E, V ) if and
written as θ (ei , ei ) =
j=1
only if all θj ∈ B (E × E, F).
Lemma 6 dim H (E × E, F) = dim E − dim E h2i .
Proof. We claim that dim B (E × E, F) = dim E h2i . Let x1 , x2 , . . . , xk be a basis of E h2i , then
x∗1 , x∗2 , . . . , x∗k are linearly independent in E ∗ where x∗i is defined by x∗i (xi ) = 1 and x∗i (xj ) = 0 if
i 6= j. Let θ f ∈ B (E × E, F), then θf = α1 θ x∗1 + α2 θ x∗1 + · · · + αk θx∗k and θx∗1 , θx∗1 , . . . , θx∗k are linearly independent in B (E × E, F). Hence dim B (E × E, F) = dim E h2i and dim H (E × E, F) = dim E − dim E h2i .
Suppose that E is an evolution algebra with a natural basis BE = {e1 , . . . , em }. Let us denote
by Sym (E) the space of all symmetric bilinear forms on E. Then Sym (E) = δ ei ,ej : 1 ≤ i ≤ j ≤ m
while Z (E × E, F) = hδ ei ,ei : 1 = 1, 2, . . . , mi, where a bilinear form δ ei ,ej : E × E −→F is defined by
δ ei ,ej (el , em ) = 1 if {i, j} = {l, m}, and otherwise it takes the value zero.
Example 7 Let E : e21 = e2 be a 2-dimensional nilpotent evolution algebra. Here Z (E × E, F) =
hδ e1 ,e1 , δ e2 ,e2 i, by Lemma 6, B (E × E, F) is spanned by θe∗2 . Let θe∗2 = αδ e1 ,e1 + βδ e2 ,e2 ∈ Z (E × E, F).
It follows that α = θe∗2 (e1 , e1 ) = e∗2 e21 = 1, and β = θe∗2 (e2 , e2 ) = e∗2 e22 = 0 then θe∗2 = δ e1 ,e1 .
Hence B (E × E, F) = hδ e1 ,e1 i and H (E × E, F) = hδ e2 ,e2 i.
Example 8 Let E : e21 = e3 , e22 = e3 be a 3-dimensional nilpotent evolution algebra. Here
3
P
αi δ ei ,ei .
Z (E × E, F) = hδ e1 ,e1 , δ e2 ,e2 , δ e3 ,e3 i, by Lemma 6, B (E × E, F) is spanned by θ e∗3 . Let θ e∗3 =
i=1
Then α1 = α2 = 1, and α3 = 0.
hδ e1 ,e1 , δ e3 ,e3 i.
Hence B (E × E, F) = hδ e1 ,e1 + δ e2 ,e2 i and H (E × E, F) =
Suppose that Eθ = E⊕V is an evolution algebra then V V = 0 and so ann(Eθ ) ⊇ V 6= 0. It means
that by this way we construct evolution algebras with ann(Eθ ) 6= 0. Further, any evolution algebras
with ann (E) 6= 0 can be obtained by this way.
Lemma 9 Let E be an evolution algebra and ann (E) 6= 0. Then E is an annihilator extension of a
lower dimensional evolution algebra.
Proof. Suppose that E is an evolution algebra with a natural basis B = {e1 , . . . , en }. Let B ′ =
{er+1 , . . . , en } be a natural basis of ann (E) then the quotient E/ann (E) is an evolution algebra with
r
P
αi ēi ∈ E/ann (E) then we define
a natural basis B ′′ = {ēi = ei + ann (E) : 1 ≤ i ≤ r}. If x̄ =
an injective linear map σ : E/ann (E) −→ E by σ (x̄) =
r
P
i=1
αi ei . For any x̄, ȳ ∈ E/ann (E) we
i=1
have σ(x̄)σ(ȳ) − σ(x̄ȳ) ∈ ann (E) (let π : E −→ E/ann (E) be a projection map then π (σ (x̄)) = x̄
and hence σ(x̄)σ(ȳ) − σ(x̄ȳ) ∈ ker π). Then θ : E/ann (E) × E/ann (E) −→ ann (E) defined by
θ(x̄, ȳ) = σ(x̄)σ(ȳ)−σ(x̄ȳ) ∈ Z (E × E, V ). It remains to show that (E/ann (E))θ = E/ann (E)⊕ann (E)
is isomorphic to E. If x ∈ E, then x can be uniquely written as x = σ(ȳ) + z, where ȳ ∈ E/ann (E) and
z ∈ ann (E). Defining φ : E −→ (E/ann (E))θ by φ(x) = φ (σ(ȳ) + z) = ȳ + z, then φ is bijective and
φ(x1 x2 ) =
=
=
φ ((σ(ȳ1 ) + z1 )(σ(ȳ2 ) + z2 )) = φ(σ(ȳ1 )σ(ȳ2 ))
φ(σ(ȳ1 ȳ2 ) + θ(ȳ1 , ȳ2 )) = ȳ1 ȳ2 + θ(ȳ1 , ȳ2 )
(ȳ1 + z1 ) ⋆ (ȳ2 + z2 ) = φ(x1 ) ⋆ φ(x2 ).
Then φ is an isomorphism.
So in particular all nilpotent evolution algebras can be obtained by this way.
3
4
The Classification Method
In Section 3, Lemma 9, we show that a nilpotent evolution algebra E can be obtained as an annihilator
extension of M = E/ann (E) by V = ann (E). Now, suppose that Eθ is an annihilator extension of E by
V , if ann(Eθ ) 6= V then Eθ can be obtained as an annihilator extension of E/ann (E) by ann (E). So we
get Eθ as annihilator extensions of different algebras E and E/ann (E). Our aim is to avoid construct
the same algebra by different algebras.
Definition 10 Let θ ∈ Z (E × E, V ) then θ⊥ = {ei ∈ E : θ(ei , ei ) = 0} is called the radical of θ. If
n−m
P
θj (ei , ei ) em+j where θj ∈ Z (E × E, F), then
θ(ei , ei ) =
j=1
⊥
⊥
θ⊥ = θ⊥
1 ∩ θ 2 ∩ · · · ∩ θ n−m .
Lemma 11 Let θ ∈ Z (E × E, V ). Then ann(Eθ ) = θ ⊥ ∩ ann (E) ⊕ V.
2
2
Proof.
If ei ∈ ann(E
θ ) then ei ⋆ ei = ei + θ (ei , ei ) = 0, hence ei = θ (ei , ei ) = 0. It follows that
⊥
ei ∈ θ ∩ ann (E) ⊕ V . On the other hand, note that V ⊂ ann(Eθ ). If ei ∈ θ⊥ ∩ ann (E) then
ei ⋆ ei = e2i + θ (ei , ei ) = 0 and hence ei ∈ ann(Eθ ).
Corollary 12 Let θ ∈ Z (E × E, V ). Then ann(Eθ ) = V if and only if
θ⊥ ∩ ann (E) = 0.
Corollary 13 Let θ ∈ Z (E × E, V ) such that θ⊥ ∩ ann (E) = 0. Then Eθ can not be obtained as an
annihilator extension of a different algebra.
Proof. Consider ϑ ∈ Z(M × M, W ) such that ϑ⊥ ∩ ann(M) = 0. If Mϑ ∼
= Eθ , then ann (Mϑ ) ∼
=
ann (Eθ ). Hence M ∼
= E.
= Eθ /ann (Eθ ) ∼
= Mϑ /ann (Mϑ ) ∼
Thus to avoid constructing the same algebra as an annihilator extension of different algebras, we
restrict θ to be such that θ⊥ ∩ ann (E) = 0.
Suppose that E is an evolution algebra. Denote Aut (E) for the set of all automorphisms of E. Let
φ ∈ Aut (E). For θ ∈ Z (E × E, V ) defining φθ (ei , ei ) = θ (φ (ei ) , φ (ei )) then φθ ∈ Z (E × E, V ) if and
only if θ (φ (ei ) , φ (ej )) = 0 for all i 6= j.
m
P
cii δ ei ,ei ∈ Z (E × E, F). Then θ can be
Consider the matrix representation and assume that θ =
i=1
represented by an m × m diagonal matrix cii , and then φθ = φt cii φ. Hence φθ ∈ Z (E × E, V ) if
and only if φθ be a diagonal matrix.
Now, suppose that E be an evolution algebra with a natural basis BE = {e1 , . . . , em } and V
n−m
P
θ i (el , el )em+i , ϑ(el , el ) =
be a vector space with a basis BV = {em+1 , . . . , en }. Let θ(el , el ) =
i=1
n−m
P
⊥
⊥
ϑi (el , el )em+i ∈ H (E × E, V ) and θ ∩ ann (E) = ϑ ∩ ann (E) = 0, i.e., ann(Eθ ) = ann(Eϑ ) = V .
i=1
Assume that Eθ and Eϑ are isomorphic, let φ : Eθ −→ Eϑ be an isomorphism then φ (ann(Eθ )) =
ann(Eϑ ), i.e., ψ = φ |V ∈ Gl (V ). If φ (ei ) = yi + vi where yi ∈ E, vi ∈ V , then φ induces an
isomorphism φ0 : E −→ E defined by φ0 (ei ) = yi and a linear map f : E −→ V defined by f (ei ) = vi .
So, we can realize φ as a matrix of the form:
φ0 0
φ=
; φ0 ∈ Aut (E) , f ∈ Hom (E, V ) and ψ = φ |V ∈ Gl (V ) .
f ψ
Now, let x =
n
P
ai ei ∈ Eθ = E ⊕ V then
i=1
φ
n
X
i=1
ai e i
!
= φ0
m
X
i=1
ai e i
!
+f
m
X
i=1
4
ai e i
!
+ψ
n
X
i=m+1
ai e i
!
.
According to the definition of φ we have
φ
n
X
n
X
ai e i ⋆ Eθ
ai e i
i=1
i=1
L.H.S. =
!
m
X
φ
m
X
i=1
m
X
=
i=1
R.H.S. =
=
"
φ0
m
X
i=1
2
!
n
X
⋆ Eϑ φ
ai e i
i=1
e2i + θ(ei , ei )
!
.
!
a2i φ0 e2i + f e2i + ψ (θ(ei , ei ))
a2i φ0 e2i + (θf + ψ ◦ θ) (ei , ei ) .
ai e i
i=1
m
X
ai e i
i=1
ai
i=1
=
=φ
n
X
!#2
+ ϑ φ0
m
X
ai e i
i=1
!
, φ0
m
X
ai e i
i=1
!!
a2i φ0 e2i + φ0 ϑ(ei , ei ) .
This yields to:
φ0 ϑ(ei , ei ) = (θ f + ψ ◦ θ) (ei , ei ) for i = 1, 2, ..., m.
(1)
i.e., φ0 υ = θf + ψ ◦ θ ∈ Z (E × E, V ). Thus Eθ and Eϑ are isomorphic if and only if there exist
φ0 ∈ Aut (E) and ψ ∈ Gl (V ) such that φ0 υ = ψ ◦ θ modulo B (E × E, V ).
Note that for all i 6= j we have
0 = φ (ei ⋆Eθ ej ) = φ (ei ) ⋆Eϑ φ (ej ) = ϑ (φ0 (ei ) , φ0 (ej )) .
i.e., φ0 υ ∈ H (E × E, V ) .
n−m
P
Now, for 1 ≤ i ≤ n − m, if ψ (em+i ) =
aij em+j , then
j=1
n−m
X
φ0 ϑi (el , el )em+i =
θ i (el , el )ψ (em+i ) =
n−m
X n−m
X
aij θ i (el , el )em+j .
j=1 i=1
i=1
i=1
Hence φ0 ϑj =
n−m
X
n−m
P
aij θi modulo B (E × E, V ). It follows that the φ0 ϑi span the same subspace of
i=1
H (E × E, V ) as the θi . We have proved
Lemma 14 Let θ(el , el ) =
n−m
P
θi (el , el )em+i and ϑ(el , el ) =
n−m
P
ϑi (el , el )em+i be two elements of
i=1
i=1
⊥
H (E × E, V ). Suppose that θ ∩ ann (E) = ϑ⊥ ∩ ann (E) = 0. Then Eθ ∼
= Eϑ if and only if there is a
φ ∈ Aut (E) such that the φϑi span the same subspace of H (E × E, F) as the θi .
In case of θ = ϑ, we obtain from Equation (1) the following description of Aut(Eθ ).
Lemma 15 Let E be a nilpotent evolution algebra and θ ∈ H (E × E, V ) such that θ ⊥ ∩ ann (E) = 0.
Then Aut (Eθ ) consists of all linear operators of the matrix form
φ0 0
; φ0 ∈ Aut (E) , f ∈ Hom(E, V ) and ψ = φ |V ∈ Gl (V ) .
φ=
f ψ
such that φ0 θ(ei , ei ) = f e2i + ψ (θ(ei , ei )) for i = 1, 2, ..., m.
5
Let E be an evolution algebra with a natural basis BE = {e1 , . . . , em }.
If E =
he1 , .., er−1 , er+1 , .., em i ⊕ Fer is the direct sum of two nilpotent ideals, the 1-dimensional ideal Fer
is called an annihilator component of E. Any m-dimensional nilpotent evolution algebra with annihilator component is a 1-dimensional trivial extension of (m − 1)-dimensional nilpotent evolution algebra.
Our aim is to avoid construct nilpotent algebras with annihilator components, and hence when we
construct (n + 1)-dimensional nilpotent algebras, we have to add algebras E = I ⊕ Fen+1 where I be
a nilpotent algebra with a natural basis BI = {e1 , . . . , en }.
Lemma 16 Let θ(el , el ) =
n−m
P
θi (el , el )em+i and θ⊥ ∩ ann (E) = 0. Then Eθ has an annihilator
i=1
component if and only if θ1 , θ2 , . . . , θs are linearly dependent in H (E × E, F).
Proof. Suppose that Eθ has an annihilator component. Then we can write Eθ = I ⊕ Fem+r where
em+r ∈ V . Then
θ : E × E −→ hem+1 , . . . , em+r−1 , em+r+1 , . . . , en−m i ,
and hence θr = 0. Then θ 1 , θ2 , . . . , θn−m are linearly dependent. On the other hand, suppose that
s
P
ϑi (el , el )em+i where hϑi : i = 1, 2, . . . , si is a
θ1 , θ2 , . . . , θn−m are linearly dependent, let ϑ(el , el ) =
i=1
proper subspace of hθi : i = 1, 2, . . . , n − mi and span the same space as θi . Then by Lemma 14, we
have Eθ ∼
= Eϑ . It follows that Eθ has an annihilator component.
Suppose that E is an evolution algebra. For θ ∈ H (E × E, F), define a subgroup Sθ (E) of Aut (E)
by
Sθ (E) = {φ ∈ Aut (E) : θ (φ (ei ) , φ (ej )) = 0 ∀i 6= j} .
Now, we have a procedure takes a nilpotent evolution algebra E of dimension m with a natural
basis BE = {e1 , . . . , em } gives us all nilpotent evolution
algebras Ẽ of dimension n with a natural basis
B = {e1 , . . . , em , em+1 , . . . , en } such that Ẽ/ann E˜ ∼
= E and Ẽ has no annihilator components. The
Ẽ
procedure runs as follows:
1. Determine Z (E × E, F) , B (E × E, F) and H (E × E, F).
2. Let V be a vector space with a basis BV = {em+1 , . . . , en }. Consider θ ∈ H (E × E, V ) with
θ(el , el ) =
n−m
X
θ i (el , el )em+i ,
i=1
where the θ i ∈ H (E × E, F) are linearly independent, and θ⊥ ∩ ann (E) = 0. Use the action of
Sθ (E) on θ by hand calculations to get a list of orbit representatives as small as possible.
3. For each orbit found, construct the corresponding algebra.
5
Nilpotent evolution algebras of dimension ≤ 3
In this section we present a complete classification of nilpotent evolution algebras of dimension ≤ 3.
We denote the j-th algebra of dimension i by Ei,j and among the natural basis of Ei,j the multiplication
is specified by giving only the nonzero products.
5.1
Nilpotent evolution algebras of dimension 1
Let E be an evolution algebra of dimension 1 with natural basis {e1 }. If E is nilpotent then ann (E) =
he1 i. Therefore, there is only one nilpotent evolution algebra of dimension 1, that is E1,1 with a natural
basis {e1 } and
E
E1,1
Multiplication Table
—————–
H (E × E, F)
hδ e1 ,e1 i
ann (E)
E1,1
Aut (E)
GL (E1,1 )
Table 1: One-dimensional nilpotent evolution algebras
6
5.2
Nilpotent evolution algebras of dimension 2
Here we consider the classification of 2-dimensional nilpotent algebras.
5.2.1
Nilpotent algebras with annihilator component
We consider the 1-dimensional trivial extensions ( corresponding to θ = 0) of algebras of dimension 1.
We get E2,1 = E1,1 ⊕ E1,1 .
5.2.2
Nilpotent algebras with no annihilator component
Here we consider 1-dimensional extensions of E1,1 . We get only one algebra E2,2 : e21 = e2 corresponding
to θ = δ e1 ,e1 .
Theorem 17 Up to isomorphism there exactly two nilpotent evolution algebras of dimension 2 with a
natural basis B = {e1 , e2 }over a field F which are isomorphic to the following pairwise non-isomorphic
algebras:
E
E2,1
Multiplication Table
—————–
B (E × E, F)
0
H (E × E, F)
hδ e1 ,e1 , δe2 ,e2 i
ann (E)
E2,1
E2,2
e21 = e2 .
hδ e1 ,e1 i
hδ e2 ,e2 i
he2 i
Aut (E)
GL
(E2,1 )
a
0
φ = 11
a21 a211
Table 2: Two-dimensional nilpotent evolution algebras
5.3
Nilpotent evolution algebras of dimension 3
Here we classify three-dimensional nilpotent evolution algebras over arbitrary fields based on nilpotent
evolution algebras of dimension up to two.
5.3.1
Nilpotent algebras with annihilator component
Here we consider the 1-dimensional trivial extensions of algebras of dimension 2. We get E3,1 = E2,1 ⊕
E1,1 and E3,2 = E2,2 ⊕ E1,1 .
5.3.2
Nilpotent algebras with no annihilator component
1-dimensional extension of E2,1 :
Let θ = αδ e1 ,e1 + βδ e2 ,e2 and αβ 6= 0, otherwise θ⊥ ∩ ann(E2,1 ) 6= 0. The α−1 θ span the same
subspace in H (E × E, F) as θ, so we may assume α = 1. For any α, β ∈ F∗ , let θα = δ e1 ,e1 +αδ e2 ,e2 , θβ =
δ e1 ,e1 + βδ e2 ,e2 and assume that (E2,1 )θα is isomorphic to (E2,1 )θβ then by Lemma 14, there exist a
1 0
1 0
, it follows that
φ = λ
φ ∈ Aut (E2,1 ) and λ ∈ F∗ such that φθ α = λθβ . Thus φt
0 β
0 α
2
1 0
2
det φ
, then
α. On the other hand, assume β = δ α. Then φθ α = θβ where φ =
β = δ= λ
0 δ
by Lemma 14 (E2,1 )θα is isomorphic to (E2,1 )θβ . Hence (E2,1 )θα is isomorphic to (E2,1 )θβ if and only if
there is a δ ∈ F∗ such that β = δ 2 α. So we get the algebras:
α
E3,3
: e21 = e3 , e22 = αe3 ; α ∈ F∗ /F∗2 .
1-dimensional extension of E2,2 :
We get only one algebra E3,4 : e21 = e2 , e22 = e3 corresponding to θ = δ e2 ,e2 .
2-dimensional extension of E1,1 :
Here we have H(E1,1 × E1,1 , F) = hδ e1 ,e1 i, so there is no 2-dimensional extension of E1,1 .
7
Theorem 18 Up to isomorphism any 3-dimensional nilpotent evolution algebras with a natural basis
B = {e1 , e2 , e3 } over a field F is isomorphic to one of the following pairwise non-isomorphic nilpotent
evolution algebras:
E
E3,1
Multiplication Table
—————–
E3,2
e21 = e2 .
α∈F
E3,3
∗
F∗2
e21 = e3 , e22 = αe3 .
e21 = e2 , e22 = e3 .
E3,4
B (E × E, F)
0
H (E × E, F)
Sym (E3,1 )
ann (E)
E3,1
hδ e1 ,e1 i
hδ e2 ,e2 , δ e3 ,e3 i
he2 , e3 i
hδ e1 ,e1 + δ e2 ,e2 i
hδ e1 ,e1 , δ e3 ,e3 i
he3 i
hδ e1 ,e1 , δ e2 ,e2 i
hδ e3 ,e3 i
he3 i
Aut (E)
GL (E3,1 )
#
"
a11 0
0
φ = a21 a211 a23
a
0 a33
" 31
#
a11 a12 0
φ = a21 a22 0 ,
a31 a32 a33
a211 + αa221 = a33 ,
a212 + αa222 = αa33 ,
a11 a12
+ αa21 a22 = 0.
a11 0
0
φ = 0 a211 0
a31 0 a411
Table 3: Three-dimensional nilpotent evolution algebras
From Theorem 18, up to isomorphism, the number of 3-dimensional nilpotent evolution algebras:
Over an algebraic closed field F is
Over R is
Over Q is
4
5
∞
Over any finite field is
5 if characteristic F is odd
4 if characteristic F is even
Table 4: Number of 3-dimensional nilpotent Evolution algebras
6
Nilpotent evolution algebras of dimension 4
In this section we present a complete classification of 4-dimensional nilpotent evolution algebras over
an algebraic closed field and R. We denote real nilpotent evolution algebras of dimension 4 by E4,j .
6.1
Nilpotent algebras with annihilator component
Here we consider the 1-dimensional trivial extension corresponding to θ = 0 of algebras of dimension
α=−1
α=1
3. We get E4,1 = E4,1 = E3,1 ⊕ E1,1 , E4,2 = E4,2 = E3,2 ⊕ E1,1 , E4,3 = E4,3 = E3,3
⊕ E1,1 , E4,4 = E3,3
⊕
E1,1 and E4,4 = E4,5 = E3,4 ⊕ E1,1 .
6.2
Nilpotent algebras with no annihilator component
1-dimensional extension of E3,1 :
Let θ = αδ e1 ,e1 + βδ e2 ,e2 + γδ e3 ,e3 and αβγ 6= 0. After dividing we may assume that θ = δ e1 ,e1 +
βδ e2 ,e2 + γδ e3 ,e3 and βγ 6= 0. Consider φ1 , φ2 , φ3 , φ4 ∈ Sθ (E3,1 ) such that
φ1
φ4
1
0
1
= 0 β − 2
0
0
0
0
=
1
(−γ)− 2
1
0
0 , φ2 = 0
1
γ− 2
0
0
1
− 21
0 .
(−β)
0
0
0
β
− 21
0
0
1
0 , φ3 = 0
−1
0
(−γ) 2
0
0
1
γ− 2
0
1
(−β)− 2 ,
0
• Over an algebraic closed field F:
Here β, γ are squares then φ1 θ = δ e1 ,e1 + δ e2 ,e2 + δ e3 ,e3 . So we get the algebra E4,5 : e21 = e22 =
e23 = e4 .
• Over the real field R:
8
–
–
–
–
If
If
If
If
β
β
β
β
> 0, γ
> 0, γ
< 0, γ
< 0, γ
>0
<0
>0
<0
then
then
then
then
φ1 θ
φ2 θ
φ3 θ
φ4 θ
= θ1 = δ e1 ,e1 + δ e2 ,e2 + δ e3 ,e3 .
= θ2 = δ e1 ,e1 + δ e2 ,e2 − δ e3 ,e3 .
= θ2 = δ e1 ,e1 + δ e2 ,e2 − δ e3 ,e3 .
= −θ2 = − (δ e1 ,e1 + δ e2 ,e2 − δ e3 ,e3 ).
Therefore we have two representatives θ1 , θ2 . It remains to decide isomorphism between
(E3,1 )θ1 , (E3,1 )θ2 . By Lemma 14, (E3,1 )θ1 is isomorphic to (E3,1 )θ2 if and only if there exist a
φ ∈ Aut (E3,1 ) and λ ∈ R∗ such that φθ1 = λθ 2 . This amounts to (among others) the equations:
a211 + a221 + a231 = λ, a212 + a222 + a232 = λ, a213 + a223 + a233 = −λ,
which never have a solution in R. Thus (E3,1 )θ1 , (E3,1 )θ2 are non-isomorphic. So we get only two
algebras:
E4,6 : e21 = e22 = e23 = e4 .
E4,7 : e21 = e22 = e4 , e23 = −e4 .
1-dimensional extension of E3,2 :
Let θ = αδ e2 ,e2 + βδ e3 ,e3 and αβ 6= 0. We can divide to get α = 1. So we may assume that
θ = δ e2 ,e2 + βδ e3 ,e3 . Consider φ1 , φ2 ∈ Sθ (E3,2 ) such that
1 0
0
1 0
0
0 .
0 , φ2 = 0 1
φ1 = 0 1
1
−1
0 0 β− 2
0 0 (−β) 2
• Over an algebraic closed field F:
Here β is a square then φ1 θ = δ e2 ,e2 + δ e3 ,e3 . So we get the algebra E4,6 : e21 = e2 , e22 = e23 = e4 .
• Over the real field R:
– If β > 0 then φ1 θ = θ 1 = δ e2 ,e2 + δ e3 ,e3 .
– If β < 0 then φ2 θ = θ 2 = δ e2 ,e2 − δ e3 ,e3 .
Suppose that φ ∈ Aut (E3,2 ) and λ ∈ R∗ such that φθ1 = λθ 2 . Then we get (among others) the
equations
a222 + a232 = λ, a233 = −λ,
which never have a solutions in R. Thus (E3,2 )θ1 , (E3,2 )θ2 are not isomorphic. So we get only
two algebras:
E4,8 : e21 = e2 , e22 = e23 = e4 .
E4,9 : e21 = e2 , e22 = e4 , e23 = −e4 .
α
1-dimensional extension of E3,3
:
Let θ = γδ e1 ,e1 + βδ e3 ,e3 and β 6= 0. After dividing we may assume that θ = γ (δ e1 ,e1 + αδ e2 ,e2 ) +
α
such that
δ e3 ,e3 . Consider φ1 , φ2 , φ3 ∈ Sθ E3,3
1
1
1
(−γ) 2
0
0
0
0 0
0
(−γ) 2
γ2
1
1
1
φ1 = 0
(−γ) 2
0 , φ2 = 0 γ 2 0 , φ3 = (−γ) 2
0
0 .
0
0 γ
0
0
−γ
0
0
−γ
• Over an algebraic closed field F. In this case α = 1.
If γ = 0 then we get θ1 = δ e3 ,e3 . If γ 6= 0 then φ2 θ = γ 2 (δ e1 ,e1 + δ e3 ,e3 ). Then after dividing
α=1
and λ ∈ F∗ such that
we may assume that θ2 = δ e1 ,e1 + δ e3 ,e3 . Suppose that φ ∈ Aut E3,3
φθ1 = λθ2 .
a231 − a232 = λ, a233 = λ, a31 a32 = a31 a33 = a32 a33 = 0,
α=1
α=1
.Therefore we get the
is not isomorphic to E3,3
which never have a solution. So, E3,3
θ2
θ1
following non-isomorphic algebras:
E4,7 : e21 = e3 , e22 = e3 , e23 = e4 .
E4,8 : e21 = e3 + e4 , e22 = e3 , e23 = e4 .
9
• Over the real field R. In this case α = ±1.
Case1. α = 1:
– If γ = 0, then we get θ1 = δ e3 ,e3 .
α=1
×E α=1
– If γ < 0 then φ3 θ = γ 2 (−δ e2 ,e2 + δ e3 ,e3 ). Since B E3,3
3,3 , R = hδ e1 ,e1 + δ e2 ,e2 i,
α=1
2
×E α=1
δ e1 ,e1 = −δ e2 ,e2 in H E3,3
3,3 , R . So, φ3 θ = γ (δ e1 ,e1 + δ e3 ,e3 ). So after dividing we
may assume that θ2 = δ e1 ,e1 + δ e3 ,e3 .
– If γ > 0 then φ2 θ = γ 2 (δ e1 ,e1 + δ e3 ,e3 ). Therefore we a gain get θ2 = δ e1 ,e1 + δ e3 ,e3 .
α=1
and λ ∈ F∗ such that φθ1 = λθ 2 . This then amounts to the
Suppose that φ ∈ Aut E3,3
following equations:
a231 − a232 = λ, a233 = λ, a31 a32 = a31 a33 = a32 a33 = 0,
α=1
α=1
which never have a solution in R. So, E3,3
is not isomorphic to E3,3
.Therefore we get
θ1
θ2
the following non-isomorphic algebras:
E4,10 : e21 = e3 , e22 = e3 , e23 = e4 .
E4,11 : e21 = e3 + e4 , e22 = e3 , e23 = e4 .
Case 2. α = −1:
– If γ = 0 then we get θ1 = δ e3 ,e3 .
– If γ < 0 then φ3 θ = −γ 2 (δ e1 ,e1 − δ e3 ,e3 ). So after dividing we may assume that θ2 =
δ e1 ,e1 − δ e3 ,e3 .
– If γ > 0 then φ2 θ = γ 2 (δ e1 ,e1 + δ e3 ,e3 ). So after dividing we may assume that θ3 =
δ e1 ,e1 + δ e3 ,e3 .
α=−1
and λ ∈ R∗ such that φθ1 = λθ2 . Then we get the
Firstly, suppose that φ ∈ Aut E3,3
following equations
a231 + a232 = λ, a233 = −λ, a31 a32 = a31 a33 = a32 a33 = 0,
α=−1
α=−1
.
is not isomorphic to E3,3
which never have a solution in R. So E3,3
θ2
θ1
α=−1
and λ ∈ R∗ such that φθ 1 = λθ3 . Then we get the equations
Next, suppose that φ ∈ Aut E3,3
a231 + a232 = λ, a233 = λ, a31 a32 = a31 a33 = a32 a33 = 0,
α=−1
α=−1
which never have a solution in R. So E3,3
is not isomorphic to E3,3
.
θ1
θ3
α=−1
and λ ∈ R∗ such that φθ3 = λθ2 . Then we get (among
Finally, suppose that φ ∈ Aut E3,3
others) the equations
a211 + a231 + a212 + a232 = λ, a233 = −λ,
α=−1
α=−1
. Hence we get
is not isomorphic to E3,3
which never have a solution in R. So E3,3
θ3
θ2
the following pairwise non-isomorphic algebras:
E4,12 : e21 = e3 , e22 = −e3 , e23 = e4 .
E4,13 : e21 = e3 + e4 , e22 = −e3 , e23 = −e4 .
E4,14 : e21 = e3 + e4 , e22 = −e3 , e23 = e4 .
1-dimensional extension of E3,4 :
• Over an algebraic closed field F:
Since H (E3,4 ×E 3,4 , F) = hδ e3 ,e3 i, we get only one algebra E4,9 : e21 = e2 , e22 = e3 , e23 = e4 .
• Over the real field R:
Also, H (E3,4 ×E 3,4 , R) = hδ e3 ,e3 i. So, we get only one algebra E4,15 : e21 = e2 , e22 = e3 , e23 = e4 .
2-dimensional extension of E2,1 :
10
• Over an algebraic closed field F:
As H (E2,1 ×E 2,1 , F) = hδ e1 ,e1 , δ e2 ,e2 i, we get only one algebra E4,10 : e21 = e3 , e22 = e4 .
• Over the real field R:
Also, H (E2,1 ×E 2,1 , R) = hδ e1 ,e1 , δ e2 ,e2 i and so we get only one algebra E4,16 : e21 = e3 , e22 = e4 .
2-dimensional extension of E2,2 :
Here H (E2,2 × E2,2 , F) = hδ e2 ,e2 i and hence E2,2 has no 2-dimensional extension.
3-dimensional extension of E1,1 :
Here H (E1,1 × E1,1 , F) = hδ e1 ,e1 i and hence E1,1 has no 3-dimensional extension.
Theorem 19 Up to isomorphism there exist 10 nilpotent evolution algebras of dimension four with a
natural basis B = {e1 , e2 , e3 , e4 } over an algebraic closed field F which are isomorphic to one of the
following pairwise non-isomorphic nilpotent evolution algebras:
• E4,1 : All products are zeros.
• E4,2 : e21 = e2 .
• E4,3 : e21 = e3 , e22 = e3 .
• E4,4 : e21 = e2 , e22 = e3 .
• E4,5 : e21 = e4 , e22 = e4 , e23 = e4 .
• E4,6 : e21 = e2 , e22 = e4 , e23 = e4 .
• E4,7 : e21 = e3 , e22 = e3 , e23 = e4 .
• E4,8 : e21 = e3 + e4 , e22 = e3 , e23 = e4 .
• E4,9 : e21 = e2 , e22 = e3 , e23 = e4 .
• E4,10 : e21 = e3 , e22 = e4 .
Theorem 20 Up to isomorphism there exist 16 nilpotent evolution algebras of dimension four with a
natural basis B = {e1 , e2 , e3 , e4 } over R which are isomorphic to one of the following pairwise nonisomorphic nilpotent evolution algebras:
• E4,1 : All products are zeros.
• E4,2 : e21 = e2 .
• E4,3 : e21 = e3 , e22 = e3 .
• E4,4 : e21 = e3 , e22 = −e3 .
• E4,5 : e21 = e2 , e22 = e3 .
• E4,6 : e21 = e4 , e22 = e4 , e23 = e4 .
• E4,7 : e21 = e4 , e22 = e4 , e23 = −e4 .
• E4,8 : e21 = e2 , e22 = e4 , e23 = e4 .
• E4,9 : e21 = e2 , e22 = e4 , e23 = −e4 .
• E4,10 : e21 = e3 , e22 = e3 , e23 = e4 .
• E4,11 : e21 = e3 + e4 , e22 = e3 , e23 = e4 .
• E4,12 : e21 = e3 , e22 = −e3 , e23 = e4 .
• E4,13 : e21 = e3 + e4 , e22 = −e3 , e23 = −e4 .
• E4,14 : e21 = e3 + e4 , e22 = −e3 , e23 = e4 .
• E4,15 : e21 = e2 , e22 = e3 , e23 = e4
• E4,16 : e21 = e3 , e22 = e4 .
11
References
[1] J. Ancochea Bermudez, J. Fresan, J. Bentabol, Contractions of Low-Dimensional Nilpotent Jordan
Algebras, Communications In Algebra. 39 (2011) 1139–1151.
[2] L. Camacho, J. Gomez, B. Omirov, R. Turdibaev, Some properties of evolution algebras, Bulletin
Of The Korean Mathematical Society. 50 (2013) 1481–1494.
[3] J. Casas, M. Ladra, B. Omirov, U. Rozikov, On Evolution Algebras, Algebra Colloq. 21 (2014)
331–342.
[4] J. Casas, M. Ladra, B. Omirov, U. Rozikov, On nilpotent index and dibaricity of evolution
algebras, Linear Algebra And Its Applications. 439 (2013) 90–105.
[5] A. Elduque, A. Labra, Evolution algebras and graphs, arxiv.org. (2013).
[6] A. Hegazi, Classification of Nilpotent Lie Superalgebras of Dimension Five. I, International Journal
Of Theoretical Physics. 38 (1999) 1735–1739.
[7] G. Mazzola, Generic finite schemes and Hochschild cocycles, Commentarii Mathematici Helvetici.
55 (1980) 267–293.
[8] U. Rozikov, J. Tian, Evolution algebras generated by Gibbs measures, Lobachevskii Journal Of
Mathematics. 32 (2011) 270–277.
[9] C. Seeley, 7-Dimensional Nilpotent Lie Algebras, Transactions Of The American Mathematical
Society. 335 (1993) 479.
[10] J. Tian, P. Vojtechovsky, Mathematical concepts of evolution algebras in non-Mendelian genetics,
in: Quasigroups and Related Systems, 2006: pp. 111–122.
[11] J. Tian, Evolution algebras and their applications, Springer, Berlin, 2008.
[12] J. Tian, Y. Zou, Finitely generated nil but not nilpotent evolution algebras, J. Algebra Appl. 13
(2014) 1350070.
12
| 0math.AC
|
Regularization and the small-ball method I: sparse recovery
Guillaume Lecué1,3
Shahar Mendelson2,4,5
arXiv:1601.05584v2 [math.ST] 3 Jan 2017
January 4, 2017
Abstract
We obtain bounds on estimation error rates for regularization procedures of the form
!
N
X
1
2
(Yi − f (Xi )) + λΨ(f )
fˆ ∈ argmin
N i=1
f ∈F
when Ψ is a norm and F is convex.
Our approach gives a common framework that may be used in the analysis of learning
problems and regularization problems alike. In particular, it sheds some light on the
role various notions of sparsity have in regularization and on their connection with the
size of subdifferentials of Ψ in a neighbourhood of the true minimizer.
As ‘proof of concept’ we extend the known estimates for the LASSO, SLOPE and
trace norm regularization.
1
Introduction
The focus of this article is on regularization, which is one of the most significant methods
in modern statistics. To give some intuition on the method and on the reasons behind its
introduction, consider the following standard problem.
Let (Ω, µ) be a probability space and set X to be distributed according to µ. F is a
class of real-valued functions defined on Ω and Y is the unknown random variable that one
would like to approximate using functions in F . Specifically, one would like to identify the
best approximation to Y in F , say in the L2 sense, and find the function f ∗ that minimizes
in F the squared loss functional f → E(f (X) − Y )2 ; that is,
f ∗ = argminf ∈F E(f (X) − Y )2 ,
with the underlying assumption that f ∗ exists and is unique.
Unlike problems in approximation theory, neither the target Y nor the underlying measure µ are known. Therefore, computing the L2 distance between functions in F and Y is
1
CNRS, CREST, ENSAE, 3, avenue Pierre Larousse, 92245 MALAKOFF. France.
Department of Mathematics, Technion, I.I.T., Haifa, Israel and Mathematical Sciences Institute, The
Australian National University, Canberra, Australia
3
Email: [email protected]
4
Email: [email protected]
5
Supported by the Israel Science Foundation.
2
1
impossible. Instead, one is given partial information: a random sample (Xi , Yi )N
i=1 , selected
independently according to the joint distribution of X and Y .
Because of the random nature of the sample and the limited information it provides,
there is no real hope of identifying f ∗ , but rather, only of approximating it. In an estimation
problem one uses the sample to produce a random function fˆ ∈ F , and the success of the
choice is measured by the distance between fˆ and f ∗ in the L2 (µ) sense. Thus, one would
like to ensure that with high probability with respect to the samples (Xi , Yi )N
i=1 , the error
rate
2
2
= E fˆ(X) − f ∗ (X) |(Xi , Yi )N
fˆ − f ∗
i=1
L2 (µ)
is small. More accurately, the question is to identify the way in which the error rate depends
on the structure of the class F and scales with the sample size N and the required degree
of confidence (probability estimate).
It is not surprising (and rather straightforward to verify) that the problem becomes
harder the larger F is. In contrast, if F is small, chances are that f ∗ (X) is very far from
Y , and identifying it, let alone approximating it, is pointless.
In situations we shall refer to as learning problems, the underlying assumption is that
F is indeed small, and the issue of the approximation error – the distance between Y and
f ∗ is ignored.
While the analysis of learning problems is an important and well-studied topic, the
assumption that F is reasonably small seems somewhat restrictive; it certainly does not
eliminate the need for methods that allow one to deal with very large classes.
Regularization was introduced as an alternative to the assumption on the ‘size’ of F .
One may consider large classes, but combine it with the belief that f ∗ belongs to a relatively
small substructure in F . The idea is to penalize a choice of a function that is far from that
substructure, which forces the learner to choose a function in the ‘right part’ of F .
Formally, let E be a vector space, assume that F ⊂ E is a closed and convex set and
let Ψ : E → R+ be the penalty. Here, we will only consider the case in which Ψ is a norm
on E.
Let λ > 0 and for a sample (Xi , Yi )N
i=1 , set
fˆ ∈ argmin
f ∈F
!
N
1 X
(Yi − f (Xi ))2 + λΨ(f ) ;
N
i=1
fˆ is called a regularization procedure, Ψ is the regularization function and λ is the regularization parameter.
In the classical approach to regularization, the substructure of f ∗ is quantified directly
by Ψ. The underlying belief is that Ψ(f ∗ ) is not ‘too big’ and one expects the procedure
to produce fˆ for which Ψ(fˆ) is of the order of Ψ(f ∗ ). Moreover, the anticipated error rate
kfˆ − f ∗ kL2 (µ) depends on Ψ(f ∗ ). In fact, an optimistic viewpoint is that regularization
could perform as well as the best learning procedure in the class {f : Ψ(f ) ≤ Ψ(f ∗ )}, but
without knowing Ψ(f ∗ ) beforehand.
2
Among the regularization schemes that are based on the classical approach are reproducing kernel Hilbert spaces (RKHS), in which the RKHS norm serves as the penalty. Since
RKHS norms capture various notions of smoothness, in RKHS regularization one is driven
towards a choice of a smooth fˆ – as smooth as f ∗ is.
In more modern regularization problems the situation is very different. Even when
penalizing with a norm Ψ, one no longer cares whether or not Ψ(f ∗ ) is small; rather, one
knows (or at least believes) that f ∗ is sparse in some sense, and the hope is that this sparsity
will be reflected in the error rate.
In other words, although one uses certain norms as regularization functions – norms
that seemingly have nothing to do with ‘sparsity’ – the hope is that the sparse nature of f ∗
will be exposed by the regularization procedure, while Ψ(f ∗ ) will be of little importance.
The most significant example in the context of sparsity-driven regularization is the
celebrated LASSO estimator [33]. Let F = { t, · : t ∈ Rd } and set t∗ to be a minimizer in
Rd of the functional t → E( t, X − Y )2 . The LASSO is defined by
t̂ ∈ argmin
t∈Rd
N
1 X
N
i=1
t, Xi − Yi
2
+ λΨ(t)
Pd
for the choice Ψ(t) = ktk1 = i=1 |ti |.
The remarkable property of the LASSO (see [8] and [3]) is that for a well-chosen regularization parameter λ, if t∗ is supported on at most s coordinates (and under various
assumptions on X and Y to which we will return later), then with high probability,
kt̂ − t∗ k22 .
s log(ed)
.
N
Thus, the error rate of the LASSO does not depend on Ψ(t∗ ) = kt∗ k1 , but rather on the
degree of sparsity of t∗ , measured here by the cardinality of its support kt∗ k0 = |{i : t∗i 6= 0}|.
This fact seems almost magical, because to the naked eye, the regularization function
ktk1 has nothing to do with sparsity; yet ℓ1 regularization leads to a sparsity-driven error
rate.
A standard (yet somewhat unconvincing) explanation of this phenomenon is that the
penalty ktk1 is a convexified version of ktk0 = |{i : ti 6= 0}|, though this loose connection
hardly explains why kt∗ k0 has any effect on the error rate of the LASSO.
A similar phenomenon occurs for other choices of Ψ, such as the SLOPE and tracenorm regularization, which will be explored in detail in what follows. In all these cases
and others like them, the regularization function is a norm that does not appear to be
connected to sparsity, nor to other natural notions of low-dimensional structures for that
matter. Yet, and quite mysteriously, the respective regularization procedure emphasizes
those very properties of t∗ .
The aim of this note is to offer a framework that can be used to tackle standard learning
problems (small F ) and regularized problems alike. Moreover, using the framework, one
may explain how certain norms lead to the emergence of sparsity-based bounds.
3
In what follows we will show that two parameters determine the error rate of regularization problems. The first one captures the ‘complexity’ of each set in the natural hierarchy
in F
Fρ = {f ∈ F : Ψ(f − f ∗ ) ≤ ρ}.
Applying results from [20, 22, 19], the ‘complexity’ of each Fρ turns out to be the optimal
(in the minimax sense) error rate of the learning problem in that set. To be more precise,
the main ingredient in obtaining a sharp error rate of a learning problem in a class H is an
accurate analysis of the empirical excess squared loss functional
f → PN Lf =
N
N
1 X
1 X ∗
(f (Xi ) − Yi )2 −
(f (Xi ) − Yi )2 .
N
N
i=1
(1.1)
i=1
Since the minimizer fˆ of the functional (1.1) satisfies PN Lfˆ ≤ 0, one may obtain an estimate
on the error rate by showing that with high probability, if kf −f ∗ kL2 (µ) ≥ r then PN Lf > 0.
This excludes functions in the set {f ∈ H : kf − f ∗ kL2 (µ) ≥ r} as potential empirical
minimizers. That ‘critical level’ turns out to be the correct (minimax) error rate of a learning
problem in H. That very same parameter is of central importance in regularization problems
— specifically, the ‘critical level’ r(ρ) for each one of the sets {f ∈ F : Ψ(f − f ∗ ) ≤ ρ} (see
Section 2.1 for an accurate definition of r(ρ) and its role in the analysis of learning problems
and regularization problems).
The second parameter, which is the main ingredient in our analysis of regularization
problems, measures the ‘size’ of the subdifferential of Ψ in points that are close to f ∗ : recall
that the subdifferential of Ψ in f is
(∂Ψ)f = {z ∗ ∈ E ∗ : Ψ(f + h) ≥ Ψ(f ) + z ∗ (h) for every h ∈ E}
where E ∗ is the dual space of the normed space (E, Ψ), and that if f 6= 0, the subdifferential
consists of all the norm one linear functionals z ∗ for which z ∗ (f ) = Ψ(f ).
Fix ρ > 0 and let Γf ∗ (ρ) be the collection of functionals that belong to the subdifferential
(∂Ψ)f for some f ∈ F that satisfies Ψ(f − f ∗ ) ≤ ρ/20. Set
Hρ = {f ∈ F : Ψ(f − f ∗ ) = ρ and kf − f ∗ kL2 (µ) ≤ r(ρ)}
and let
∆(ρ) = inf
sup
h∈Hρ z ∗ ∈Γ
f ∗ (ρ)
z ∗ (h − f ∗ ).
/ {f ∈ F : Ψ(f − f ∗ ) ≤ ρ/20}
Hence, Γf ∗ (ρ) is a subset of the unit sphere of E ∗ when 0 ∈
∗
and it is the entire unit ball of E otherwise. And, since Hρ consists of functions whose Ψ
norm is ρ, it is evident that ∆(ρ) ≤ ρ. Therefore, if ∆(ρ) ≥ αρ for a fixed 0 < α ≤ 1 then
Γf ∗ (ρ) is rather large: for every h ∈ Hρ there is some z ∗ ∈ Γf ∗ (ρ) for which z ∗ (h) is ‘almost
extremal’—that is, at least αρ.
Our main result (Theorem 3.2 below) is that if Γf ∗ (ρ) is large enough to
ensure that ∆(ρ) ≥ 4ρ/5, and the regularization parameter λ is set to be of the
4
r 2 (ρ)
ρ ,
then with high probability, the regularized minimizer in F , fˆ,
satisfies that kfˆ − f ∗ kL2 (µ) ≤ r(ρ) and Ψ(fˆ − f ∗ ) ≤ ρ.
order of
Theorem 3.2 implies that one may analyze regularization problems by selecting ρ wisely,
keeping in mind that points in a Ψ-ball of radius ∼ ρ around f ∗ must generate a sufficiently
large subdifferential. And the fact that functionals in Γf ∗ (ρ) need to be ‘almost extremal’
only for points in Hρ rather than for the entire sphere is crucial; otherwise, it would have
forced Γf ∗ (ρ) to be unreasonably large – close to the entire dual sphere.
As will be clarified in what follow, sparsity, combined with the right choice of Ψ, contributes in two places: firstly, if f ∗ is sparse in some sense and Ψ is not smooth on sparse
elements, then Γf ∗ (ρ), which contains the subdifferential (∂Ψ)f ∗ , is large; secondly, for
the right choice of ρ the ‘localization’ Hρ consists of elements that are well placed: if
Ψ(f − f ∗ ) = ρ and kf − f ∗ kL2 (µ) ≤ r(ρ), there is some z ∗ ∈ Γf ∗ (ρ) for which z ∗ (f − f ∗ ) is
large enough. The fact that Hρ is well placed is an outcome of some compatibility between
Ψ and the L2 (µ) norm.
Of course, to find the right choice of ρ one must first identify r(ρ), which is, in itself, a
well-studied yet nontrivial problem.
Before we dive into technical details, let us formulate some outcomes of our main result.
We will show how it can be used to obtain sparsity-driven error rates in three regularization
procedures: the LASSO, SLOPE and trace norm regularization. In all three cases our
results actually extend the known estimates in various directions.
The LASSO.
The LASSO is defined for the class of linear functional F = { t, · : t ∈ Rd }. For a fixed
t0 ∈ Rd , the goal
P is to identify t0 using linear measurements, the regularization function is
Ψ(t) = ktk1 = di=1 |ti |, and the resulting regularization procedure produces
N
1 X
t̂ ∈ argmin
N
t∈Rd
i=1
t, Xi − Yi
2
+ λktk1 .
The LASSO has been studied extensively in the last two decades. Even though some
recent advances [39, 35, 27] have shown the LASSO to have its limitation, historically, it
has been the benchmark estimator of high-dimensional statistics — mainly because a high
dimensional parameter space does not significantly affect its performance as long as t0 is
sparse. This was shown for example, in [3, 15, 36, 37, 17, 26, 34] in the context of estimation
and sparse oracle inequalities, in [16, 40, 2] for support recovery results; and in various other
instances as well; we refer the reader to the books [5, 8] for more results and references on
the LASSO.
SLOPE.
In some sense, SLOPE, introduced in [4, 31], is actually an extension of the LASSO,
even though it has been introduced as an extension of multiple-test procedures. Again,
the underlying class is F = { t, · : t ∈ Rd }, and to define the regularization function let
5
β1 ≥ β2 ≥ ... ≥ βd > 0 and set
Ψ(t) =
d
X
βi t♯i ,
i=1
(t♯i )di=1
where
denotes the non-increasing re-arrangement of (|ti |)di=1 . Thus, the SLOPE
norm is a sorted, weighted ℓ1 -norm, and for (β1 , ..., βd ) = (1, ..., 1), SLOPE regularization
coincides with the LASSO.
Trace-norm regularization.
Consider the trace inner-product on Rm×T . Let F = { A, · : A ∈ Rm×T } and given a
target Y put A∗ to be the matrix that minimizes A → E( A, X − Y )2 . The regularization
function is the trace norm.
Definition 1.1 Let A be a matrix and setP(σi (A)) to be its singular values, arranged in a
non-increasing order. For p ≥ 1, kAkp = ( σip (A))1/p is the p-Schatten norm.
Note that the trace-norm is simply the 1-Schatten norm, the Hilbert-Schmidt norm is the
2-Schatten norm and the operator norm is the ∞-Schatten norm.
The trace norm regularization procedure is
N
1 X
 ∈ argmin
(Yi − Xi , A )2 + λkAk1
A∈Rm×T N
i=1
and it was introduced for the reconstruction of low-rank, high-dimensional matrices [28, 9,
29, 6, 7, 25].
As will be explained in what follows, our main result holds in rather general situations
and may be implemented in examples once the ‘critical levels’ r(ρ) are identified. Since the
examples we present serve mainly as “proof of concept”, we will focus only on one scenario
in which r(ρ) may be completely characterized for an arbitrary class of functions.
Definition 1.2 Let ℓM
2 be an M -dimensional inner product space and let µ be a measure
on ℓM
.
The
measure
µ
is isotropic if for every x ∈ ℓM
2
2 ,
Z
2
x, t dµ(t) = kxk2ℓM ;
2
it is L-subgaussian if for every p ≥ 2 and every x ∈ ℓM
2 ,
√
k x, · kLp (µ) ≤ L pk x, · kL2 (µ) .
Hence, the covariance structure of an isotropic measure coincides with the inner product in
ℓM
2 , and if µ is an L-subgaussian measure then the Lp (µ) norm of a linear form does not
grow faster than the Lp norm of the corresponding Gaussian variable.
6
Assumption 1.1 Assume that the underlying measure µ is isotropic and L-subgaussian,
and that for f ∗ = t∗ , · (or f ∗ = A∗ , · in the matrix case), the noise1 ξ = f ∗ (X) − Y
belongs to Lq for some q > 2.
When dealing with the LASSO and SLOPE, the natural Euclidean structure is the
standard one in Rd , and for trace norm regularization, the natural Euclidean structure is
endowed by the trace inner product in Rm×T .
Remark 1.3 In the supplementary material we study a general X without assuming it is
isotropic, which means dealing with less natural Euclidean structures in the examples we
present. It is also possible to go beyond the subgaussian case, we refer the reader to [13]
where other moment assumptions on X are considered.
The second part of Assumption 1.1, that ξ ∈ Lq for some q > 2, is rather minimal.
Indeed, for the functional f → E(f (X) − Y )2 to be well defined, one must assume that
f (X) − Y ∈ L2 ; the assumption here is only slightly stronger.
Applying our main result we will show the following:
Theorem 1.4 Consider the LASSO under Assumption 1.1. Let 0 < δ < 1. Assume that
there is some v ∈ Rd supported on at most s coordinates for which
r
log(ed)
∗
.
kt − vk1 ≤ c1 (δ)kξkLq s
N
p
If λ = c2 (L, δ)kξkLq log(ed)/N and N ≥ s log(ed/s), then with probability at least 1 − δ
the LASSO estimator with regularization parameter λ satisfies that for every 1 ≤ p ≤ 2
r
log(ed)
∗
1/p
t̂ − t p ≤ c3 (L, δ)kξkLq s
.
N
The error rate in Theorem 1.4 coincides with the standard estimate on the LASSO (cf. [3]),
but in a broader context: t∗ need not be sparse but only approximated by a sparse vector;
the target Y is arbitrary and the noise ξ may be heavy tailed and need not be independent
of X.
Turning to SLOPE, let us recall the estimates from [31], where the setup is somewhat
restricted: Let X be a Gaussian vector on Rd , set W to be a Gaussian random variable
with variance σ 2 that is independent of X and put Y = t∗ , X + W . Consider some
q ∈ (0, 1), let Φ−1 (α) be the α-th quantile of the standard normal distribution and put
βi = Φ−1 (1 − iq/(2d)).
1
In what follows we will refer to ξ as ‘the noise’ even though it depends in general on Y and X. The
reason for using that term comes from the situation in which Y = f ∗ (X) − W for a symmetric random
variable W that is independent of X (independent additive noise); thus ξ = W . We have opted to call ξ ‘the
noise’ because its role in the general case and its impact on the error rate is rather similar to what happens
for independent noise.
7
Theorem 1.5 [31] Let 1 ≤ s ≤ d satisfy that s/d
√ = o(1) and (s log d)/N = o(1) when
N → ∞. If 0 < ε < 1, N → ∞ and λ = 2σ/ N , the SLOPE estimator with weights
(βi )di=1 and regularization parameter λ satisfies
N t̂ − t∗ 2
2
>
1
+
3ε
→ 0.
sup P r
2σ 2 s log(d/s)
kt∗ k0 ≤s
Note that Theorem 1.5 is asymptotic in nature and not ‘high-dimensional’. Moreover,
it only holds for a Gaussian X, independent Gaussian noise W , a specific choice of weights
(βi )di=1 and t∗ that is s-sparse.
We consider a more general situation. Let βi ≤ C
p
log(ed/i) and set Ψ(t) =
♯
i=1 βi ti .
Pd
Theorem 1.6 There exists constants c1 , c2 and c3 that depend only on L, δ and C for which
the following holds. Under Assumption 1.1, if there is v ∈ Rd that satisfies |supp(v)| ≤ s
and
ed
s
∗
√
,
Ψ(t − v) ≤ c1 kξkLq
log
s
N
√
then for N ≥ c2 s log(ed/s) and with the choice of λ = c2 kξkLq / N , one has
ed
ed
s
s
2
and
t̂ − t∗ 2 ≤ c3 kξk2Lq log
Ψ(t̂ − t∗ ) ≤ c3 kξkLq √ log
s
N
s
N
with probability at least 1 − δ.
Finally, let us consider trace norm regularization.
Theorem 1.7 Under Assumption 1.1 and if there is V ∈ Rm×T that satisfies that rank(V ) ≤
s and
r
max{m, T }
kA∗ − V k1 ≤ c1 kξkLq s
,
N
q
}
. Then with
one has the following. Let N ≥ c2 s max{m, T } and λ = c3 kξkLq max{m,T
N
probability at least 1 − δ, for any 1 ≤ p ≤ 2
r
max{m, T }
.
 − A∗ ≤ c4 kξkLq s1/p
N
p
The constants c1 , c2 , c3 and c4 depends only on L and δ.
A result of a similar flavour to Theorem 1.7 is Theorem 9.2 from [8].
Theorem 1.8 Let X be an isotropic and L-subgaussian vector, and W that is mean-zero,
independent of X and belongs to the Orlicz space Lψα for some α ≥ 1. If Y = A∗ , X + W
and
)
(
r
kξk √m(t + log N )(t + log m)
m(t + log m)
ψα
1/α
,
, kξkψα log
λ ≥ c1 (L) max kξk2
N
kξkL2
N
then with probability at least 1 − 3 exp(−t) − exp(−c2 (L)N )
2
 − A∗ ≤ c3 min λ kA∗ k1 , λ2 rank(A∗ ) .
2
8
(1.2)
Clearly, the assumptions of Theorem 1.8 are more restrictive than those of Theorem
1.7, as the latter holds for a heavy tailed ξ that need not be independent of X, and for A∗
that can be approximated by a low-rank matrix. Moreover, if kA∗ k1 is relatively large and
the error rate in Theorem 1.8 is the sparsity-dominated λ2 rank(A∗ ), then the error rate in
Theorem 1.7 is better by a logarithmic factor.
The proofs of the error rates in all the three examples will be presented in Section 5.
1.1
Notation
We end the introduction with some standard notation.
Throughout, absolute constants are denoted by c, c1 ..., etc. Their value may change from
line to line. When a constant depends on a parameter α it will be denoted by c(α). A . B
means that A ≤ cB for an absolute constant c, and the analogous two-sided inequality is
denoted by A ∼ B. In a similar fashion, A .α B implies that A ≤ c(α)B, etc.
Let E ⊂ L2 (µ) be a vector space and set Ψ to be a norm on E. For a set A ⊂ E, t ∈ E
and r > 0, let rA + t = {ra + t : a ∈ A}.
Denote by BΨ = {w ∈ E : Ψ(w) ≤ 1} the unit ball of (E, Ψ) and set SΨ = {f ∈ E :
Ψ(f ) = 1} to be the corresponding unit sphere. BΨ (ρ, f ) is the ball of radius ρ centred in
f and SΨ (ρ, f ) is the corresponding sphere. Also, set D to be the unit ball in L2 (µ), S is
the unit sphere there, and D(ρ, f ) and S(ρ, f ) are the ball and sphere centred in f and of
radius ρ, respectively.
A class of spaces we will be interested in consists of ℓdp , that is, Rd endowed with the ℓp
norm; Bpd denotes the unit ball in ℓdp and S(ℓdp ) is the unit sphere.
For every x = (xi )di=1 , (x♯i )di=1 denotes the non-increasing rearrangement of (|xi |)di=1 .
1 PN
Finally, if (Xi , Yi )N
i=1 is a sample, PN h = N
i=1 h(Xi , Yi ) is the empirical mean of h.
2
Preliminaries: The regularized functional
Let F ⊂ E be a closed and convex class of functions. Recall that for target Y , f ∗ is the
minimizer in F of the functional f → E(f (X) − Y )2 . Since F is closed and convex, the
minimum exists and is unique.
Let Lf (X, Y ) = (f (X) − Y )2 − (f ∗ (X) − Y )2 be the excess squared loss functional and
for λ > 0 let
Lλf (X, Y ) = Lf + λ(Ψ(f ) − Ψ(f ∗ ))
be its regularized counterpart. Thus, for a random sample (Xi , Yi )N
i=1 , the empirical (regularized) excess loss functional is
PN Lλf
N
1 X
=
Lf (Xi , Yi ) + λ(Ψ(f ) − Ψ(f ∗ )),
N
i=1
Note that if ℓf (x, y) = (y − f (x))2 and fˆ minimizes PN ℓf + λΨ(f ) then fˆ also minimizes
PN Lλf . Moreover, since Lλf∗ = 0, it is evident that PN Lλˆ ≤ 0.
f
9
This simple observation shows that the random set {f ∈ F : PN Lλf > 0} may be
excluded from our considerations, as it does not contain potential minimizers. Therefore, if
one can show that with high probability,
{f ∈ F : PN Lλf ≤ 0} ⊂ {f ∈ F : kf − f ∗ kL2 (µ) ≤ r},
then on that event, kfˆ − f ∗ kL2 (µ) ≤ r.
We will identify when PN Lλf > 0 by considering the two parts of the empirical functional:
the empirical excess loss PN Lf and the regularized part λ(Ψ(f ) − Ψ(f ∗ )).
Because of its crucial role in obtaining error estimates in learning problems, the functional f → PN Lf has been studied extensively using the small-ball method, (see, e.g.,
[20, 22, 19]). Thus, the first component in the machinery we require for explaining both
learning problems and regularization problems is well understood and ready-to-use; its details are outlined below.
2.1
The natural decomposition of PN Lf
Set ξ = ξ(X, Y ) = f ∗ (X) − Y and observe that
Lf (X, Y ) =(f − f ∗ )2 (X) + 2(f − f ∗ )(X) · (f ∗ (X) − Y )
=(f − f ∗ )2 (X) + 2ξ(f − f ∗ )(X).
Since F is convex, the characterization of the nearest point map in a Hilbert space shows
that
E(f − f ∗ )(X) · (f ∗ (X) − Y ) ≥ 0
for every f ∈ F . Hence, setting ξi = f ∗ (Xi ) − Yi , one has
PN Lλf
N
N
1 X
1 X
∗ 2
≥
(f − f ) (Xi ) + 2
ξi (f − f ∗ )(Xi ) − Eξ(f − f ∗ )(X)
N
N
i=1
i=1
∗
+λ(Ψ(f ) − Ψ(f )).
To simplify notation, for w ∈ L2 (µ) set Qw = w2 and Mw = ξw − Eξw. Thus, for every
f ∈ F,
(2.1)
PN Lλf ≥ PN Qf −f ∗ + 2PN Mf −f ∗ + λ(Ψ(f ) − Ψ(f ∗ )).
The decomposition of the empirical excess loss to the quadratic component (Qf −f ∗ ) and
the multiplier one (Mf −f ∗ ) is the first step in applying the small-ball method to learning
problems. One may show that on a large event, if kf − f ∗ kL2 (µ) is larger than some critical
level then PN Qf −f ∗ ≥ θkf − f ∗ k2L2 and dominates PN Mf −f ∗ ; hence PN Lf > 0.
To identify this critical level, let us define the following parameters:
Definition 2.1 Let H ⊂ F be a convex class that contains f ∗ . Let (εi )N
i=1 be independent,
symmetric, {−1, 1}-valued random variables that are independent of (Xi , Yi )N
i=1 .
10
For γQ , γM > 0 set
rQ (H, γQ ) = inf
(
r>0:E
sup
h∈H∩D(r,f ∗ )
let
φN (H, s) =
and put
sup
h∈H∩D(s,f ∗ )
N
1 X
εi (h − f ∗ )(Xi ) ≤ γQ r
N
i=1
)
,
N
1 X
√
εi ξi (h − f ∗ )(Xi ) ,
N i=1
o
n
√
rM (H, γM , δ) = inf s > 0 : P r φN (H, s) ≤ γM s2 N ≥ 1 − δ .
The main outcome of the small-ball method is that for the right choices of γM and γQ ,
r = max{rM , rQ } is the above-mentioned ‘critical level’ in H, once H satisfies a weak
small-ball condition.
Assumption 2.1 (The small ball condition) Assume that there are constants κ > 0
and 0 < ε ≤ 1, for which, for every f, h ∈ F ∪ {0},
P r |f − h| ≥ κkf − hkL2 (µ) ≥ ε.
There are numerous examples in which the small-ball condition may be verified for κ and ε
that are absolute constants. We refer the reader to [12, 19, 10, 21, 22, 30] for some of them.
Theorem 2.2 ([22]) Let H be a closed, convex class of functions that contains f ∗ and
satisfies Assumption 2.1 with constants κ and ε. If θ = κ2 ε/16 then for every 0 < δ < 1,
with probability at least 1 − δ − 2 exp(−N ε2 /2) one has:
• for every f ∈ H,
|PN Mf −f ∗ | ≤
n
o
θ
2
max kf − f ∗ k2L2 (µ) , rM
(H, θ/10, δ/4) ,
8
• If f ∈ H and kf − f ∗ kL2 (µ) ≥ rQ (H, κε/32) then
PN Qf −f ∗ ≥ θkf − f ∗ k2L2 (µ) .
In particular, with probability at least 1 − δ − 2 exp(−N ε2 /2),
θ
PN Lf ≥ kf − f ∗ k2L2 (µ)
2
for every f ∈ H that satisfies
kf − f ∗ kL2 (µ) ≥ max {rM (H, θ/10, δ/4) , rQ (H, κε/32)} .
From now on, we will assume that F satisfies the small-ball condition with constants κ
and ε, and that θ = κ2 ε/16.
11
Definition 2.3 Let ρ > 0 and set
rM (ρ) = rM F ∩ BΨ (ρ, f ∗ ),
θ δ
,
10 4
and
rQ (ρ) = rQ F ∩ BΨ (ρ, f ∗ ),
κε
.
32
In what follows we will abuse notation and omit the dependence of rM and rQ on f ∗ , κ, ε
and δ.
Let r(·) be a function that satisfies r(ρ) ≥ supf ∗ ∈F max{rQ (ρ), rM (ρ)}. Finally, put
O(ρ) =
sup
f ∈F ∩BΨ (ρ,f ∗ )∩D(r(ρ),f ∗ )
PN Mf −f ∗ .
Theorem 2.2 implies the following:
Corollary 2.4 ([22]) Using the notation introduced above, on an event of probability at
least 1 − δ − 2 exp(−N ε2 /2), if f ∈ F ∩ BΨ (ρ, f ∗ ) and kf − f ∗ kL2 (µ) ≥ r(ρ) then
θ
PN Lf ≥ kf − f ∗ k2L2 (µ) .
2
Moreover, on the same event,
O(ρ) ≤
θ 2
r (ρ).
8
Remark 2.5 Let us stress once again that r(ρ) plays a central role in the analysis of empirical risk minimization in the set F ∩ BΨ (ρ, f ∗ ). Theorem 2.2 implies that with high
probability, the empirical risk minimizer h̃ in F ∩ BΨ (ρ, f ∗ ) satisfies
kh̃ − h∗ kL2 (µ) ≤ r(ρ).
Moreover, it follows from [11] and [23] that under mild structural assumptions on F , r(ρ)
is the best possible error rate of any learning procedure in F ∩ BΨ (ρ, f ∗ ) – i.e., the minimax
rate in that class.
Let A be the event from Corollary 2.4 and set
γO (ρ) = sup O(ρ).
w∈A
γO will be of little importance in what follows, because it may be upper bounded by
(θ/8)r 2 (ρ). However, it will be of the utmost importance in [13], where complexity-based
regularization is studied (see Section 6 for more details).
3
The main result
Let us turn to the second part of the regularized functional – namely, λ(Ψ(f ) − Ψ(f ∗)). Let
E ∗ be the dual space to (E, Ψ) and set Ψ∗ to be the dual norm. BΨ∗ and SΨ∗ denote the
dual unit ball and unit sphere, respectively; i.e., BΨ∗ consists of all the linear functionals
z ∗ on E for which supΨ(x)=1 |z ∗ (x)| ≤ 1.
12
Definition 3.1 The functional z ∗ ∈ SΨ∗ is a norming functional for z ∈ E if z ∗ (z) = Ψ(z).
In the language of Convex Analysis, a functional is norming for x if and only if it belongs
to (∂Ψ)x , the subdifferential of Ψ in x.
Let Γf ∗ (ρ) be the collection of functionals that are norming for some f ∈ BΨ (ρ/20, f ∗ ).
In particular, Γf ∗ (ρ) contains all the norming functionals of f ∗ .
Set
∆(ρ) = inf
sup z ∗ (h − f ∗ ),
h∈H z ∗ ∈Γ ∗ (ρ)
f
where the infimum is taken in the set
H = F ∩ SΨ (ρ, f ∗ ) ∩ D(r(ρ), f ∗ ) = {h ∈ F : Ψ(h − f ∗ ) = ρ and kh − f ∗ kL2 (µ) ≤ r(ρ)}.
Note that if z ∗ ∈ Γf ∗ (ρ) and h ∈ SΨ (ρ, f ∗ ) then |z ∗ (h − f ∗ )| ≤ Ψ(h − f ∗ ) = ρ. Thus,
a lower bound of the form ∆(ρ) ≥ (1 − δ)ρ implies that Γf ∗ (ρ) is a relatively large subset
of the dual unit sphere: each point in F ∩ SΨ (ρ, f ∗ ) ∩ D(r(ρ), f ∗ ) has an ‘almost norming’
functional in Γf ∗ (ρ).
Our main result is that if Γf ∗ (ρ) is indeed large enough to ensure that ∆(ρ) ≥ 4/5ρ
then with high probability kfˆ − f ∗ kL2 (µ) ≤ r(ρ) and Ψ(fˆ − f ∗ ) ≤ ρ.
Theorem 3.2 Assume that F is closed and convex. Let ρ > 0 and set A to be an event on
which Corollary 2.4 holds. If ∆(ρ) ≥ 4ρ/5 and
3
γO (ρ)
θ r 2 (ρ)
≤λ< ·
,
ρ
2
ρ
then on the event A, a regularized empirical minimizer fˆ ∈ argminf ∈F PN Lλf satisfies
Ψ(fˆ − f ∗ ) ≤ ρ and kfˆ − f ∗ kL2 (µ) ≤ r(ρ).
Moreover, since rO (ρ) ≤ (θ/8)r 2 (ρ), the same assertion holds if
θ r 2 (ρ)
3θ r 2 (ρ)
·
≤λ< ·
.
8
ρ
2
ρ
The proof of the theorem follows in three steps: first, one has to show that PN Lλf is
positive on the set F ∩ SΨ (ρ, f ∗ ). Second, thanks to certain homogeneity properties of the
functional, it is positive in F \BΨ (ρ, f ∗ ), because it is positive on the ‘sphere’ F ∩ SΨ (ρ, f ∗ ).
Finally, one has to study the functional in F ∩ BΨ (ρ, f ∗ ) and verify that it is positive in
that set, provided that kf − f ∗ kL2 (µ) ≥ r(ρ).
Proof. Fix h ∈ F ∩ SΨ (ρ, f ∗ ) and we shall treat two different cases: when kh − f ∗ kL2 (µ) ≥
r(ρ) and when kh − f ∗ kL2 (µ) ≤ r(ρ).
If kh − f ∗ kL2 ≥ r(ρ), then by the triangle inequality for Ψ,
Ψ(h) − Ψ(f ∗ ) = Ψ(h − f ∗ + f ∗ ) − Ψ(f ∗ ) ≥ −Ψ(h − f ∗ ).
13
D(r(ρ), f ∗ )
f∗
SΨ (ρ, f ∗ )
h
f
Hence, for (Xi , Yi )N
i=1 ∈ A and by the upper estimate in the choice of λ,
θ
θ
PN Lλh ≥ kh − f ∗ k2L2 (µ) − λΨ(h − f ∗ ) ≥ r 2 (ρ) − λρ > 0.
2
2
(3.1)
Next, if kh − f ∗ kL2 (µ) ≤ r(ρ) then
PN Lλh ≥ −2O(ρ) + λ(Ψ(h) − Ψ(f ∗ )).
Consider u, v ∈ E that satisfy f ∗ = u + v and Ψ(u) ≤ ρ/20. Let z ∗ be any norming
functional of v; thus, z ∗ ∈ SΨ∗ and z ∗ (v) = Ψ(v). Since Ψ(h) = supx∗ ∈BΨ∗ x∗ (h) it follows
that
Ψ(h) − Ψ(f ∗ ) ≥Ψ(h) − Ψ(v) − Ψ(u) ≥ z ∗ (h − v) − Ψ(u) ≥ z ∗ (h − f ∗ ) − 2Ψ(u).
This holds for any v ∈ BΨ (ρ/20, f ∗ ), and by the definition of ∆(ρ) and for an optimal choice
of z ∗ ,
PN Lλh ≥ −2O(ρ) + λ(z ∗ (h − f ∗ ) − 2Ψ(u)) ≥ −2O(ρ) + λ(∆(ρ) − ρ/10) > 0,
(3.2)
where the last inequality holds because ∆(ρ) ≥ 4ρ/5 and λ ≥ 3γO (ρ)/ρ. Also, since
γO (ρ) ≤ (θ/8)r 2 (ρ), it suffices that λ ≥ (3θ/8)r 2 (ρ)/ρ to ensure that PN Lλh > 0 in (3.2).
This completes the proof of the first step – that PN Lλh > 0 on F ∩ SΨ (ρ, f ∗ ).
Turning to the second step, one has to establish a similar inequality for functions outside
BΨ (ρ, f ∗ ). To that end, let f ∈ F \BΨ (ρ, f ∗ ). Since F is convex and Ψ is homogeneous,
f = f ∗ + α(h − f ∗ ) for some h ∈ F ∩ SΨ (ρ, f ∗ ) and α > 1. Therefore,
PN Qf −f ∗ = α2 PN Qh−f ∗ and PN Mf −f ∗ = αPN Mh−f ∗ ;
moreover, Ψ(f − f ∗ ) = αΨ(h − f ∗ ) and for every functional z ∗ , z ∗ (f − f ∗ ) = αz ∗ (h − f ∗ ).
Thus, by (3.1), when kh − f ∗ kL2 (µ) ≥ r(ρ), PN Lλf > 0, and when kh − f ∗ kL2 (µ) ≤ r(ρ),
PN Lλf ≥ α2 PN Qh−f ∗ + 2αPN Mh−f ∗ + λ(αz ∗ (h − f ∗ ) − 2Ψ(u))
≥ α PN Qh−f ∗ + 2PN Mh−f ∗ + λ(z ∗ (h − f ∗ ) − 2Ψ(u)) > 0.
Finally, when h ∈ F ∩ BΨ (ρ, f ∗ ) and kh − f ∗ kL2 (µ) ≥ r(ρ), (3.1) shows that PN Lλf > 0.
14
Remark 3.3 Note that if ρ ≥ Ψ(f ∗ ) there is no upper limitation on the choice of λ. Indeed,
if kh − f ∗ kL2 (µ) ≥ r(ρ) and Ψ(h) = ρ ≥ Ψ(f ∗ ) then λ(Ψ(h) − Ψ(f ∗ )) ≥ 0, and PN Lλh > 0
just as in (3.1). The rest of the proof remains unchanged.
It follows from the proof that the quadratic component PN Qf −f ∗ and the regularization
one λ(Ψ(f ) − Ψ(f ∗ )) dominate the multiplier component 2PN Mf −f ∗ in different parts of
F . The behaviour of PN Qf −f ∗ allows one to exclude the set (F ∩ Bψ (ρ, f ∗ ))\D(r(ρ), f ∗ ), as
well as any point in F for which the interval [f, f ∗ ] intersects (F ∩ Sψ (ρ, f ∗ ))\D(r(ρ), f ∗ ).
This exclusion is rather free-of-charge, as it holds with no assumptions on the norm Ψ.
The situation is more subtle when trying to exclude points for which the interval [f, f ∗ ]
intersects F ∩ Sψ (ρ, f ∗ ) ∩ D(r(ρ), f ∗ ). That is precisely the region in which the specific
choice of Ψ is important and the regularization component is the reason why PN Lλf > 0.
Figure 1 shows this idea: PN Lλf > 0 for two different reasons: either Q > M – the
quadratic component dominates the multiplier component, or R > M – the regularization
component dominates the multiplier component.
Note that an output of the sparsity equation is that the descent cone TΨ (f ∗ ) = ∪τ >0 {h :
Ψ(f ∗ +τ h) ≤ Ψ(f ∗ )} does not intersect SΨ (ρ, f ∗ )∩D(r(ρ), f ∗ ) when the “sparsity condition”
∆(ρ) ≥ 4ρ/5 is satisfied (cf. Figure 2).
Q>M
R>M
TΨ (f ∗ )
SΨ (ρ, f ∗ )
R>M
D(r(ρ), f ∗ )
Q>M
Q>M
f∗
R>M
f∗
R>M
Q>M
Figure 1: The “Q > M and R > M ” decomposition.
4
Figure 2: TΨ (f ∗ ) ∩ SΨ (ρ, f ∗ ) ∩ D(r(ρ), f ∗ ) = ∅.
The role of ∆(ρ)
It is clear that ∆(ρ) plays a crucial role in the proof of Theorem 3.2, and that the larger
Γf ∗ (ρ) is, the better the lower bound on ∆(ρ).
Having many norming functionals of points in BΨ (ρ/20, f ∗ ) can be achieved somewhat
artificially, by taking ρ ∼ Ψ(f ∗ ). If ρ is large enough, BΨ (ρ/20, f ∗ ) contains a Ψ-ball centred
15
in 0. Therefore, Γf ∗ (ρ) is the entire dual sphere and ∆(ρ) = ρ. This is the situation when
one attempts to derive complexity-based bounds (see Section 6 and [13]), i.e., when one
wishes to find fˆ that inherits some of f ∗ ’s ‘good qualities’ that are captured by Ψ(f ∗ ).
Here, we are interested in cases in which ρ may be significantly smaller than Ψ(f ∗ ) and
enough norming functionals have to be generated by other means.
If Ψ is smooth, each f 6= 0 has a unique norming functional, and for a small ρ, the
norming functionals of points in BΨ (ρ/20, f ∗ ) are close to the (unique) norming functional
of f ∗ ; hence there is little hope that Γf ∗ (ρ) will be large enough to ensure that ∆(ρ) ∼ ρ.
It is therefore reasonable to choose Ψ that is not smooth in f ∗ or in a neighbourhood of f ∗ .
Another important fact is that Γf ∗ (ρ) need not be as large as the entire dual sphere to
ensure that ∆(ρ) ∼ ρ. Indeed, it suffices if Γf ∗ (ρ) contains ‘almost norming’ functionals
only to points that satisfy kwkL2 (µ) ≤ r(ρ)/ρ and Ψ(w) = 1, rather than to every point in
the sphere SΨ .
4.1
∆(ρ) and sparsity
It turns out that the combination of the right notion of sparsity with a wise choice of a
norm Ψ ensures that Γf ∗ (ρ) contains enough ‘almost norming’ functionals precisely for the
subset of the sphere one is interested in.
To give an indication of how this happens, let us show the following:
Lemma 4.1 Let Z ⊂ SΨ∗ , W ⊂ SΨ and 0 < η1 , η2 < 1. If every w ∈ W can be written as
w = w1 + w2 , where Ψ(w1 ) ≤ η1 Ψ(w) and supz ∗ ∈Z z ∗ (w2 ) ≥ (1 − η2 )Ψ(w2 ), then
inf sup z ∗ (w) ≥ (1 − η1 )(1 − η2 ) − η1
w∈W z ∗ ∈Z
In particular, if η1 , η2 ≤ 1/20 then inf w∈W supz ∗ ∈Z z ∗ (w) ≥ 4/5.
Proof. Let w = w1 + w2 and observe that Ψ(w2 ) ≥ Ψ(w) − Ψ(w1 ) ≥ (1 − η1 )Ψ(w). Thus,
for the optimal choice of z ∗ ∈ Z,
z ∗ (w1 + w2 ) ≥(1 − η2 )Ψ(w2 ) + z ∗ (w1 ) ≥ (1 − η2 )Ψ(w2 ) − η1 Ψ(w).
≥ (1 − η1 )(1 − η2 ) − η1 Ψ(w),
and the claim follows because w ∈ SΨ .
Let E = Rd viewed as a class of linear functionals on Rd . Set µ to be an isotropic
2
measure on Rd ; thus {t ∈ Rd : E t, X ≤ 1} = B2d .
Assume that for t ∈ Rd that is supported on I ⊂ {1, ..., d}, the set of its norming
functionals consists of functionals of the form z0∗ + (1 − η2 )u∗ for some fixed z0∗ that is
supported on I and any u ∈ BΨ∗ that is supported on I c (such is the case, for example,
when E = ℓd1 ).
For every such t, consider w ∈ ρSΨ and set w1 = PI w and w2 = PI c w, the coordinate
projections of w onto span(ei )i∈I and span(ei )i∈I c , respectively. Hence, there is a functional
z ∗ = z0∗ + (1 − η2 )u∗ that is norming for t and also satisfies
z ∗ (w2 ) = (1 − η2 )u∗ (w2 ) = (1 − η2 )Ψ(w2 ).
16
Therefore, Lemma 4.1 may be applied once Ψ(PI w) ≤ η1 Ψ(w).
Naturally, such a shrinking phenomenon need not be true for every w ∈ SΨ ; fortunately,
it is only required for w ∈ SΨ ∩ (r(ρ)/ρ)D – and we will show that it is indeed the case
in the three examples we present. In all three, the combination of sparsity and the right
choice of the norm helps in establishing a lower bound on ∆(ρ) in two ways: firstly, the set
Γt∗ (ρ) consists of functionals that are ‘almost norming’ for any x whose support is disjoint
from the support of t∗ ; and secondly, a coordinate projection ‘shrinks’ the Ψ norm of points
in ρSΨ ∩ r(ρ)D.
4.2
∆(ρ) in the three examples
Let us show that in the three examples, the LASSO, SLOPE and trace norm regularization,
∆(ρ) ≥ (4/5)ρ for the right choice of ρ, and that choice depends on the degree of sparsity
in each case.
In all three examples, we will assume that the underlying measure is isotropic; thus the
L2 (µ) norm coincides with the natural Euclidean structure: the ℓd2 norm for the LASSO
and SLOPE, and the Hilbert-Schmidt norm for trace-norm regularization.
The LASSO.
Observe that if f ∗ = t∗ , · is the true minimizer of the functional t, · → E( t, X −Y )2
in F = { t, · : t ∈ Rd }, then any function ht = t, · for which kht − f ∗ kL2 ≤ r(ρ) and
Ψ(ht − f ∗ ) = ρ is of the form ht = t, · = w + t∗ , · , where w ∈ ρS(ℓd1 ) ∩ r(ρ)B2d . Recall
that the dual norm to k · k1 is k · k∞ , and thus
∆(ρ) =
sup
inf
w∈ρS(ℓd1 )∩r(ρ)B2d z∈Γt∗ (ρ)
z, w ,
where Γt∗ (ρ) is the set of all vectors z ∗ ∈ Rd that satisfy
kz ∗ k∞ = 1 and z ∗ (v) = kvk1 for some v for which kv − t∗ k1 ≤ ρ/20.
Lemma 4.2 If t∗ = v+u for u ∈ (ρ/20)B1d and 100|supp(v)| ≤ (ρ/r(ρ))2 then ∆(ρ) ≥ 4ρ/5.
In other words, if t∗ is well approximated with respect to the ℓd1 norm by some v ∈ Rd that
is s-sparse, and s is small enough relative to the ratio (ρ/r(ρ))2 , then ∆(ρ) ≥ (4/5)ρ.
Just as noted earlier, we shall use two key properties of the ℓ1 norm and sparse vectors:
firstly, that if x and y have disjoint supports, there is a functional that is simultaneously
d for which
norming for x and y, i.e., z ∗ ∈ B∞
z ∗ (x) = kxk1 and z ∗ (y) = kyk1 ;
(4.1)
secondly, that if kxk1 = ρ and kxk2 is significantly smaller than ρ, a coordinate projection
‘shrinks’ the ℓd1 norm: kPI xk1 is much smaller than kxk1 .
Proof. Let w ∈ ρS(ℓd1 ) ∩ r(ρ)B2d . Since kt∗ − vk1 ≤ ρ/20 there exists z ∗ ∈ Γt∗ (ρ) that is
norming for v. Moreover, if I = supp(v), then according to (4.1) one can choose z ∗ that is
also norming for PI c w. Thus, kPI c wk1 = z ∗ (PI c w) and
z ∗ (w) = z ∗ (PI w) + z ∗ (PI c w) ≥ kPI c wk1 − kPI wk1 ≥ kwk1 − 2 kPI wk1 .
17
√
√
s kPI wk2 ≤ sr(ρ). Therefore,
√
z, w ≥ ρ − 2 sr(ρ) ≥ 4ρ/5
Since kwk2 ≤ r(ρ), one has kPI wk1 ≤
when 100s ≤ (ρ/r(ρ))2 .
SLOPE.
P
Let β1 ≥ β2 ≥ ... ≥ βd > 0 and recall that Ψ(t) = di=1 βi t∗i .
Note that Ψ(t) = supz∈Z z, t , for
Z=
( d
X
εi βπi ei :
(εi )di=1
i=1
)
d
∈ {−1, 1} , π is a permulation of {1, ..., d} .
P
Therefore, the extreme points of the dual unit ball are of the form di=1 εi βπi ei .
Following the argument outlined above, let us show that if x is supported on a reasonably small I ⊂ {1, ..., d}, the set of norming functionals of x consists of ‘almost norming’
functionals for any y that is supported on I c . Moreover, and just like the ℓd1 norm, if
Ψ(x) = ρ and kxk2 is significantly smaller than ρ, a coordinate projection of x ‘shrinks’ its
Ψ norm.
√
P
Lemma 4.3 Let 1 ≤ s ≤ d and set Bs = i≤s βi / i. If t∗ is ρ/20 approximated (relative
to Ψ) by an s-sparse vector and if 40Bs ≤ ρ/r(ρ) then ∆(ρ) ≥ 4ρ/5.
Proof. Let t∗ = u + v, for v that is supported on at most s coordinates and u ∈ (ρ/20)BΨ .
Set I ⊂ {1, ..., d} to be the support of v and let z = (zi )di=1 be a norming functional for v
to be specified later; thus, z ∈ Γt∗ (ρ).
Given t for which Ψ(t − t∗ ) = ρ and kt − t∗ k2 ≤ r(ρ), one has
z(t − t∗ ) = z(t − v) − z(u) = z(PI c (t − v)) + z(PI (t − v)) − z(u)
X
X
≥
zi (t − v)i +
zi (t − v)i − Ψ(u)
≥
=
i∈I c
i∈I
X
X
i∈I c
X
i∈I c
zi (t − v)i −
zi (t − v)i −
i≤s
X
i≤s
βi (t − v − u)♯i − 2Ψ(u)
βi (t − t∗ )♯i − 2Ψ(u) = (∗).
Since v is supported in I, one may optimize the choice of z by selecting the right permutation
of the coordinates in I c , and
X
X
X
zi (t − v)i ≥
βi (t − v)♯i ≥
βi (t − v − u)♯i − Ψ(u)
i∈I c
=
d
X
i=1
i>s
βi (t − t∗ )♯i −
X
i≤s
i>s
βi (t − t∗ )♯i − Ψ(u).
18
Therefore,
(∗) ≥
d
X
i=1
βi (t − t∗ )♯i − 2
X
i≤s
βi (t − t∗ )♯i − 3Ψ(u) ≥
X
17
ρ−2
βi (t − t∗ )♯i .
20
i≤s
√
Since kt − t∗ k2 ≤ r(ρ), it is evident that (t − t∗ )♯i ≤ r(ρ)/ i, and
s
X
i=1
βi (t
− t∗ )♯i
s
X
β
√i = r(ρ)Bs .
≤ r(ρ)
i
i=1
Hence, if ρ ≥ 40r(ρ)Bs then ∆(ρ) ≥ 4ρ/5.
Trace-norm regularization.
The trace norm has similar properties to the ℓ1 norm. Firstly, one may show that the
dual norm to k · k1 is k · k∞ , which is simply the standard operator norm. Moreover, one
may find a functional that is simultaneously norming for any two elements with ‘disjoint
support’ (and of course, the meaning of ‘disjoint support’ has to be interpreted correctly
here). Finally, it satisfies a ‘shrinking’ phenomenon for matrices whose Hilbert-Schmidt
norm is significantly smaller than their trace norm.
Lemma 4.4 If A∗ = V + U , where kU k1 ≤ ρ/20 and 400rank(V ) ≤ (ρ/r(ρ))2 , then
∆(ρ) ≥ 4ρ/5.
The fact that a low-rank matrix has many norming functionals is well known and follows,
for example, from [38].
Lemma 4.5 Let V ∈ Rm×T and assume that V = PI V PJ for appropriate orthogonal
projections onto subspaces I ⊂ Rm and J ⊂ RT . Then, for every W ∈ Rm×T there is a
matrix Z that satisfies kZk∞ = 1, and
Z, V = kV k1 ,
Z, PI ⊥ W PJ ⊥ = kPI ⊥ W PJ ⊥ k1 ,
Z, PI W PJ ⊥ = 0 and
Z, PI ⊥ W PJ = 0.
Lemma 4.5 describes a similar phenomenon to the situation in ℓd1 , but with a different notion
of ‘disjoint support’: if V is low-rank and the projections PI and PJ are non-trivial, one
may find a functional that is norming both for V and for the part of W that is ‘disjoint’ of
V . Moreover, the functional vanishes on the ‘mixed’ parts PI W PJ ⊥ and PI ⊥ W PJ .
Proof of Lemma 4.4. Recall that S1 is the unit sphere of the trace norm and that B2 is
the unit ball of the Hilbert-Schmidt norm. Hence,
∆(ρ) =
inf
sup
W ∈ρS1 ∩r(ρ)B2 Z∈ΓA∗ (ρ)
Z, W
where ΓA∗ (ρ) is the set of all matrices Z ∈ Rm×T that satisfy kZk∞ = 1 and Z, V
for some V for which kA∗ − V k1 ≤ ρ/20.
19
=1
Fix a rank-s matrix V = PI V PJ , for orthogonal projections PI and PJ that are onto
subspaces of dimension s. Consider W ∈ Rm×T for which kW k1 = ρ and kW k2 ≤ r(ρ) and
put Z to be a norming functional of V as in Lemma 4.5. Thus, Z ∈ ΓA∗ (ρ) and
Z, W = Z, PI ⊥ W PJ ⊥ + Z, PI W PJ = kPI ⊥ W PJ ⊥ k1 − kPI W PJ k1
≥kW k1 − kPI W PJ ⊥ k1 − kPI ⊥ W PJ k1 − 2kPI W PJ k1 .
All that remains is to estimate the trace norms of the three components that are believed
to be ‘low-dimension’ - in the sense that their rank is at most s.
Recall that (σi (A)) are the singular values of A arranged in a non-increasing order.
It is straightforward to verify (e.g., using the characterization of the singular values via
low-dimensional approximation), that
σi (PI W PJ ⊥ ), σi (PI ⊥ W PJ ), σi (PI W PJ ) ≤ σi (W ).
Moreover, kW k2 ≤ r(ρ), therefore, being rank-s operators, one has
kPI W PJ ⊥ k1 , kPI ⊥ W PJ k1 , kPI W PJ k1 ≤
implying that
s
X
i=1
s
1/2 √
√ X
σi (W ) ≤ s
σi2 (W )
≤ sr(ρ),
i=1
√
Z, W ≥ ρ − 4r(ρ) s.
Therefore, if 400s ≤ (ρ/r(ρ))2 , then ∆(ρ) ≥ 4ρ/5.
5
The three examples revisited
The estimates on ∆(ρ) presented above show that in all three examples, when f ∗ is well
approximated by a function whose ‘degree of sparsity’ is . (ρ/r(ρ))2 , then ∆(ρ) ≥ 4ρ/5
and Theorem 3.2 may be used. Clearly, the resulting error rates depend on the right choice
of ρ, and thus on r(ρ).
Because r(ρ) happens to be the minimax rate of the learning problem in the class
F ∩ BΨ (ρ, f ∗ ), its properties have been studied extensively. Obtaining an estimate on r(ρ)
involves some assumptions on X and ξ, and the one setup in which it can be characterized
for an arbitrary class F is when the class is L-subgaussian and ξ ∈ Lq for some q > 2 (though
ξ need not be independent of X). It is straightforward to verify that an L-subgaussian class
satisfies the small-ball condition of Assumption 2.1 for κ = 1/2 and ε = c/L4 where c
is an absolute constant. Moreover, if the class is L-subgaussian, the natural complexity
parameter associated with it is the expectation of the supremum of the canonical Gaussian
process indexed by the class.
Definition 5.1 Let F ⊂ L2 (µ) and set {Gf : f ∈ F } to be the canonical Gaussian process
indexed by F ; that is, each Gf is a centred Gaussian variable and the covariance structure
of the process is endowed by the inner product in L2 (µ). The expectation of the supremum
of the process is defined by
ℓ∗ (F ) = sup{E sup Gf : F ′ ⊂ F is finite}.
f ∈F ′
20
It follows from a standard chaining argument that if F is L-subgaussian then
N
1 X
ℓ∗ (F )
E sup
εi (h − f ∗ )(Xi ) . L √ .
N
f ∈F N
i=1
Therefore, if
Fρ,r = F ∩ BΨ (ρ, f ∗ ) ∩ D(r, f ∗ )
then for every ρ > 0 and f ∗ ∈ F
n
√ o
rQ (ρ) ≤ inf r > 0 : ℓ∗ (Fρ,r ) ≤ C(L)r N .
Turning to rM , we shall require the following fact from [18].
Theorem 5.2 (Corollary 1.10 in [18]) Let q > 2 and L ≥ 1. For every 0 < δ < 1 there
is a constant c = c(δ, L, q) for which the following holds. If H is an L-subgaussian class
and ξ ∈ Lq , then with probability at least 1 − δ,
N
1 X
sup √
εi ξi h(Xi ) ≤ ckξkLq ℓ∗ (H).
N i=1
h∈H
The complete version of Theorem 5.2 includes a sharp estimate on the constant c. However,
obtaining accurate probability estimates is not the main feature of this note and deriving
such estimates leads to a cumbersome presentation. To keep our message to the point, we
have chosen not to present the best possible probability estimates in what follows.
A straightforward application of Theorem 5.2 shows that
n
√ o
rM (ρ) ≤ inf r > 0 : kξkLq ℓ∗ (Fρ,r ) ≤ cr 2 N
for a constant c that depends on L, q and δ.
Recall that we have assumed that X is isotropic, which means that the L2 (µ) norm
coincides with the natural Euclidean structure on the space: the standard ℓd2 norm for the
LASSO and SLOPE and the Hilbert-Schmidt norm for trace norm regularization. Since the
covariance structure of the indexing Gaussian process is endowed by the inner product, it
follows that
ℓ∗ (ρBΨ ∩ rD) = E
sup
G, w
w∈ρBΨ ∩rB2
for the standard Gaussian vector G = (g1 , ..., gd ) in the case of the LASSO and SLOPE and
the Gaussian matrix G = (gij ) in the case of trace norm minimization. Hence, one may
obtain a bound on r(ρ) by estimating this expectation in each case.
The LASSO and SLOPE. Let (βi )di=1 be a non-increasing positive sequence and set
P
Ψ(t) = di=1 t♯i βi .
Since the LASSO corresponds to the choice of (βi )di=1 = (1, ..., 1), it suffices to identify
ℓ∗ (ρBΨ ∩ rB2d ) for the SLOPE norm and a general choice of weights.
21
Lemma 5.3 There exists an absolute constant C for which the following holds. If β and
Ψ are as above, then
)
( r
p
ed
log(ed/i)
+ ρ max
E
sup
G, w ≤ C min r (k − 1) log
i≥k
k
k−1
βi
w∈ρBΨ ∩rB d
2
(and if k = 1, the first term is set to be 0).
Proof. Fix 1 ≤ k ≤ d. Let J be the set of indices of the k largest coordinates of (|gi |)di=1 ,
and for every w let Iw be the sets of indices of the k largest coordinates of (|wi |)di=1 . Put
Jw = J ∪ Iw and note that |Jw | ≤ 2k. Hence,
sup
d
X
w∈ρBΨ ∩rB2d i=1
wi gi ≤ sup
.r
X
wi gi + sup
w∈ρBΨ i∈J c
w∈rB2d i∈Jw
X
i<k
(gi♯ )2
!1/2
X
wi gi
w
+ sup
X
w∈ρBΨ i≥k
g♯
wi♯ βi i
βi
.r
X
(gi♯ )2
i<k
!1/2
+ ρ max
i≥k
gi♯
.
βi
As a starting point, note that a standard binomial estimate shows that
d
p
p
♯
P r i |g| ≥ t log(ed/i)
P r gi ≥ t log(ed/i) ≤
i
≤2 exp(i log(ed/i) − i log(ed/i) · t2 /2).
Applying the union bound one has that for t ≥ 4, with probability at least 1−2 exp(−(t2 /2)k log(ed/k)),
p
(5.1)
gi♯ ≤ c3 t log(ed/i) for every i ≥ k.
The same argument shows that E(gi♯ )2 . log(ed/i).
Let Uk be the set of vectors on the Euclidean sphere that are supported on at most k
coordinates. Set
X
1/2
kxk[k] =
(x♯i )2
= sup x, u
u∈Uk
i≤k
and recall that by the Gaussian concentration of measure theorem (see, e.g., Theorem 7.1
in [14]),
1/q
√
√
≤ EkGk[k] + c q sup k G, u kL2 ≤ EkGk[k] + c1 q.
EkGkq[k]
u∈Uk
Moreover, since
E(gi♯ )2
. log(ed/i), one has
1/2 p
X
EkGk[k] ≤ E
(gi♯ )2
. k log(ed/k).
i≤k
Therefore, by Chebyshev’s inequality for q ∼ k log(ed/k), for t ≥ 1, with probability at
least 1 − 2t−c1 k log(ed/k) ,
1/2
X
p
(gi♯ )2
≤ c2 t k log(ed/k).
i≤k
22
Turning to the ‘small coordinates’, by (5.1),
gi♯
max
. t max
i≥k βi
i≥k
p
log(ed/i)
.
βi
It follows that for every choice of 1 ≤ k ≤ d,
gi♯
E
sup
G, w . rE
+ ρE max
i≥k βi
w∈ρBΨ ∩rB2d
i<k
p
p
log(ed/i)
,
.r (k − 1) log(ed/(k − 1)) + ρ max
i≥k
βi
X
(gi♯ )2
1/2
and, if k = 1, the first term is set to be 0.
d
√ If β = (1, ..., 1) (which corresponds
√ to the LASSO), then BΨ = B1 , and one may select
k ∼ ρ/r, provided that r ≤ ρ ≤ r d. In that case,
p
E
sup
G, w . ρ log(edr 2 /ρ2 ).
w∈ρB1d ∩rB2d
√
The estimates when r ≥ ρ or r d ≤ ρ are straightforward. Indeed, if r ≥ ρ then ρB1d ⊂ rB2d
and
p
ℓ∗ (ρB1d ∩ rB2d ) = ℓ∗ (ρB1d ) ∼ ρ log(ed),
√
while if r d ≤ ρ then rB2d ⊂ ρB1d , and
√
ℓ∗ (ρB1d ∩ rB2d ) = ℓ∗ (rB2d ) ∼ r d.
The LASSO.
A straightforward computation shows that
kξk2Lq d
N
2
rM (ρ) .L,q,δ
r
ekξk d
L
1
ρ kξkL
√ q
log
N
q
ρ N
and
2
rQ
(ρ)
.L
(
ρ2
N
log
0
c(L)d
N
if ρ2 N &L,q,δ kξk2Lq d2
otherwise,
if N &L d
otherwise.
Proof of Theorem 1.4. We will actually prove a slightly stronger result, which gives an
improved estimation error if one has prior information on the degree of sparsity.
Using the estimates on rM and rQ , it is straightforward to verify that the sparsity
condition of Lemma 4.2 holds when N &L,q,δ s log(ed/s) and for any
r
ed
1
ρ &L,q,δ kξkLq s
.
log
N
s
23
It follows from Lemma 4.2 that if there is an s-sparse vector that belongs to t∗ + (ρ/20)B1d ,
then ∆(ρ) ≥ 4ρ/5. Finally, Theorem 3.2 yields the stated bounds on kt̂ − t∗ k1 and kt̂ − t∗ k2
once we set
r
ed
1
r 2 (ρ)
λ∼
.
∼L,q,δ kξkLq
log
ρ
N
s
The estimates on kt̂ − t∗ kp for 1 ≤ p ≤ 2 can be easily verified because
−1+2/p
kxkp ≤ kxk1
2−2/p
kxk2
.
In case one has no prior information on s, one may take
r
1
log(ed)
ρ ∼L,q,δ kξkLq s
N
and
λ ∼L,q,δ kξkLq
r
log(ed)
.
N
The rest of the argument remains unchanged.
SLOPE
p
Assume that βi ≤ C log(ed/i), which is the standard assumption for SLOPE [4, 31].
By considering the cases k = 1 and k = d,
√
(5.2)
G, w . min{Cρ, dr}.
E
sup
w∈ρBΨ ∩rB2d
Thus, one may show that
2
rQ
(ρ) .L
0
ρ2
N
if N &L d
2
and rM
(ρ) .L,q,δ
otherwise,
2 d
kξkLq N
if ρ2 N &L,q,δ kξk2Lq d2
kξkL √ρ
q
N
otherwise.
p
√
P
Proof of Theorem 1.6. Recall that Bs = i≤s βi / i, and when βi ≤ C log(ed/i), one
may verify that
p
Bs . C s log(ed/s).
Hence, the condition Bs . ρ/r(ρ) holds when N &L,q,δ s log(ed/s) and
ed
s
.
ρ &L,q,δ kξkLq √ log
s
N
It follows from Lemma 4.3 that ∆(ρ) ≥ 4ρ/5 when there is an s-sparse vector in t∗ +
(ρ/20)BΨ ; therefore, one may apply Theorem 3.2 for the choice of
λ∼
kξkL
r 2 (ρ)
∼L,q,δ √ q .
ρ
N
24
The trace-norm.
Recall that B1 is the unit ball of the trace norm, that B2 is the unit ball of the HilbertSchmidt norm, and that the canonical Gaussian vector here is the Gaussian matrix G =
(gij ). Since the operator norm is the dual to the trace norm,
p
ℓ∗ (B1 ) = Eσ1 (G) . max{m, T },
and clearly,
ℓ∗ (B2 ) = E kGk2 .
√
mT .
Thus,
p
√
ℓ∗ (ρBΨ ∩ rB2 ) = ℓ∗ (ρB1 ∩ rB2 ) ≤ min ρℓ∗ (B1 ), rℓ∗ (B2 ) . min{ρ max{m, T }, r mT }.
Therefore,
2
rQ
(ρ) .L
0
}
ρ2 max{m,T
N
if N &L mT
otherwise,
and
2
rM
(ρ) .L,q,δ
kξk2Lq
ρkξkLq
q
mT
N
if ρ2 N &L,q,δ kξk2Lq mT (min {m, T })2
max{m,T }
N
otherwise.
Proof of Theorem 1.7. It is straightforward to verify that if N &L,q,δ s max{m, T } then
s . (ρ/r(ρ))2 when
r
max{m, T }
ρ &L,q,δ kξkLq s
N
as required in Lemma 4.4. Moreover, if there is some V ∈ Rm×T for which kV − A∗ k1 . ρ
and rank(V ) ≤ s, it follows that ∆(ρ) ≥ 4ρ/5. Setting
r
max{m, T }
r 2 (ρ∗ )
,
∼L,q,δ kξkLq
λ∼
∗
ρ
N
Theorem 3.2 yields the bounds on k − A∗ k1 and k − A∗ k2 . The bounds on the Schatten
−1+2/p
2−2/p
norms k − A∗ kp for 1 ≤ p ≤ 2 hold because kAkp ≤ kAk1
kAk2
.
6
Concluding Remarks
As noted earlier, the method we present may be implemented in classical regularization
problems as well, leading to an error rate that depends on Ψ(f ∗ ) – by applying the trivial
bound on ∆(ρ) when ρ ∼ Ψ(f ∗ ).
The key issue in classical regularization schemes is the price that one has to pay for not
knowing Ψ(f ∗ ) in advance. Indeed, given information on Ψ(f ∗ ), one may use a learning
procedure taking values in {f ∈ F : Ψ(f ) ≤ Ψ(f ∗ )} such as Empirical Risk Minimization.
This approach would result in an error rate of r(cΨ(f ∗ )), and the hope is that the error rate
25
of the regularized procedure is close to that – without having prior knowledge on Ψ(f ∗ ).
Surprisingly, as we show in [13], that is indeed the case.
The problem with applying Theorem 3.2 to the classical setup is the choice of λ. One has
no information on Ψ(f ∗ ), and thus setting λ ∼ r 2 (ρ)/ρ for ρ ∼ Ψ(f ∗ ) is clearly impossible.
A first attempt of bypassing this obstacle is Remark 3.3: if ρ & Ψ(f ∗ ), there is no upper
2
constraint on the choice of λ. Thus, one may consider λ ∼ supρ>0 r ρ(ρ) , which suits any
ρ > 0. Unfortunately, that choice will not do, because in many important examples the
supremum happens to be infinite. Instead, one may opt for the lower constraint on λ and
select
γO (ρ)
,
(6.1)
λ ∼ sup
ρ
ρ>0
which is also a legitimate choice for any ρ, and is always finite.
We will show in [13] that the choice in (6.1) leads to optimal bounds in many interesting
examples – thanks to the first part of Theorem 3.2.
An essential component in the analysis of regularization problems is bounding r(ρ),
and we only considered the subgaussian case and completely ignored the question of the
probability estimate. In that sense, the method we presented falls short of being completely
satisfactory.
Addressing both these issues requires sharp upper estimates on empirical and multiplier
processes, preferably in terms of some natural geometric feature of the underlying class.
Unfortunately, this is a notoriously difficult problem. Indeed, the final component in the
chaining-based analysis used to study empirical and multiplier processes is to translate
a metric complexity parameter (e.g., Talagrand’s γ-functionals) to a geometric one (for
example, the mean-width of the set). Such estimates are known almost exclusively in the
Gaussian case – which is, in a nutshell, Talagrand’s Majorizing Measures theory [32].
The chaining process in [18] is based on a more sensitive metric parameter than the
standard Gaussian one. This leads to satisfactory results for other choices of random vectors that are not necessarily subgaussian, for example, unconditional log-concave random
vectors. Still, it is far from a complete theory – as a general version of the Majorizing
Measures Theorem is not known.
Another relevant fact is from [24]. It turns out that if V is a class of linear functionals
on Rd that satisfies a relatively minor symmetry property, and X is an isotropic random
vector for which
√
(6.2)
sup k X, t kLp ≤ L p for 2 ≤ p . log d,
t∈S d−1
then the empirical and multiplier processes indexed by V behave as if X were a subgaussian
vector. In other words, for such “symmetric” problems it suffices to have a subgaussian
moment growth up to p ∼ log d to ensure a subgaussian behaviour.
This fact is useful because all the indexing sets considered here (and in many other
sparsity-based regularization procedures as well) satisfy the required symmetry property.
Finally, a word about the probability estimate in Theorem 5.2. The actual result from
[18] leads to a probability estimate governed by two factors: the Lq space to which ξ belongs
26
and the ‘effective dimension’ of the class. For a class of linear functionals on Rd and an
isotropic vector X, this effective dimension is
∗
ℓ (V ) 2
D(V ) =
,
d2 (V )
where ℓ∗ (V ) = E supv∈V | G, v | and d2 (V ) = supv∈V kvkℓd .
2
One may show that with probability at least
1 − c1 w−q N −((q/2)−1) logq N − 2 exp(−c2 u2 D(V )),
N
1 X
sup √
ξi V, Xi − Eξ X, v
. LwukξkLq ℓ∗ (V ).
N i=1
v∈V
(6.3)
If ξ has better tail behaviour, the probability estimate improves; for example, if ξ is subgaussian then (6.3) holds with probability at least 1 − 2 exp(−cw2 N ) − 2 exp(−cu2 D(V )).
The obvious complication is that one has to obtain a lower bound on the effective dimension D(V ). And while it is clear that D(v) & 1, in many cases (including our three
examples) a much better bound is true.
Let us mention that the effective dimension is perhaps the most important parameter in
Asymptotic Geometric Analysis. Milman’s version of Dvoretzky’s Theorem (see, e.g., [1])
shows that D(V ) captures the largest dimension of a Euclidean structure hiding in V . In
fact, this geometric observation exhibits why that part of the probability estimate in (6.3)
cannot be improved.
References
[1] Shiri Artstein-Avidan, Apostolos Giannopoulos, and Vitali D. Milman. Asymptotic geometric analysis. Part I, volume 202 of Mathematical Surveys and Monographs. American Mathematical Society,
Providence, RI, 2015.
[2] Francis R. Bach. Structured sparsity-inducing norms through submodular functions. In Advances in
Neural Information Processing Systems 23: 24th Annual Conference on Neural Information Processing
Systems 2010. Proceedings of a meeting held 6-9 December 2010, Vancouver, British Columbia, Canada.,
pages 118–126, 2010.
[3] Peter J. Bickel, Ya’acov Ritov, and Alexandre B. Tsybakov. Simultaneous analysis of lasso and Dantzig
selector. Ann. Statist., 37(4):1705–1732, 2009.
[4] Malgorzata Bogdan, Ewout van den Berg, Chiara Sabatti, Weijie Su, and Emmanuel J. Candès.
SLOPE—Adaptive variable selection via convex optimization. Ann. Appl. Stat., 9(3):1103–1140, 2015.
[5] Peter Bühlmann and Sara van de Geer. Statistics for high-dimensional data. Springer Series in Statistics.
Springer, Heidelberg, 2011. Methods, theory and applications.
[6] Emmanuel J. Candès and Yaniv Plan. Tight oracle inequalities for low-rank matrix recovery from a
minimal number of noisy random measurements. IEEE Trans. Inform. Theory, 57(4):2342–2359, 2011.
[7] David Gross. Recovering low-rank matrices from few coefficients in any basis. IEEE Trans. Inform.
Theory, 57(3):1548–1566, 2011.
[8] Vladimir Koltchinskii. Oracle inequalities in empirical risk minimization and sparse recovery problems,
volume 2033 of Lecture Notes in Mathematics. Springer, Heidelberg, 2011. Lectures from the 38th
Probability Summer School held in Saint-Flour, 2008, École d’Été de Probabilités de Saint-Flour.
[Saint-Flour Probability Summer School].
27
[9] Vladimir Koltchinskii, Karim Lounici, and Alexandre B. Tsybakov. Nuclear-norm penalization and
optimal rates for noisy low-rank matrix completion. Ann. Statist., 39(5):2302–2329, 2011.
[10] Vladimir Koltchinskii and Shahar Mendelson. Bounding the Smallest Singular Value of a Random
Matrix Without Concentration. Int. Math. Res. Not. IMRN, (23):12991–13008, 2015.
[11] Guillaume Lecué and Shahar Mendelson. Learning subgaussian classes: Upper and minimax bounds.
Technical report, CNRS, Ecole polytechnique and Technion, 2013.
[12] Guillaume Lecué and Shahar Mendelson. Sparse recovery under weak moment assumptions. Technical report, CNRS, Ecole Polytechnique and Technion, 2014. To appear in Journal of the European
Mathematical Society.
[13] Guillaume Lecué and Shahar Mendelson. Regularization and the small-ball method ii: complexity-based
bounds. Technical report, CNRS, ENSAE and Technion, I.I.T., 2015.
[14] Michel Ledoux. The concentration of measure phenomenon, volume 89 of Mathematical Surveys and
Monographs. American Mathematical Society, Providence, RI, 2001.
[15] Karim Lounici. Sup-norm convergence rate and sign concentration property of Lasso and Dantzig
estimators. Electron. J. Stat., 2:90–102, 2008.
[16] Nicolai Meinshausen and Peter Bühlmann. High-dimensional graphs and variable selection with the
lasso. Ann. Statist., 34(3):1436–1462, 2006.
[17] Nicolai Meinshausen and Bin Yu. Lasso-type recovery of sparse representations for high-dimensional
data. Ann. Statist., 37(1):246–270, 2009.
[18] Shahar Mendelson. Upper bounds on product and multiplier empirical processes. Technical report. To
appear in Stochastic Processes and their Applications.
[19] Shahar Mendelson. Learning without concentration for general loss function. Technical report, Technion, I.I.T., 2013. arXiv:1410.3192.
[20] Shahar Mendelson. Learning without concentration. In Proceedings of the 27th annual conference on
Learning Theory COLT14, pages pp 25–39. 2014.
[21] Shahar Mendelson. A remark on the diameter of random sections of convex bodies. In Geometric
aspects of functional analysis, volume 2116 of Lecture Notes in Math., pages 395–404. Springer, Cham,
2014.
[22] Shahar Mendelson. Learning without concentration. J. ACM, 62(3):Art. 21, 25, 2015.
[23] Shahar Mendelson. ’local vs. global parameters’, breaking the gaussian compexity barrier. Technical
report, Technion, I.I.T., 2015.
[24] Shahar Mendelson. On multiplier processes under weak moment assumptions. Technical report, Technion, I.I.T., 2015.
[25] Sahand Negahban and Martin J. Wainwright. Restricted strong convexity and weighted matrix completion: optimal bounds with noise. J. Mach. Learn. Res., 13:1665–1697, 2012.
[26] Sahand N. Negahban, Pradeep Ravikumar, Martin J. Wainwright, and Bin Yu. A unified framework for
high-dimensional analysis of M -estimators with decomposable regularizers. Statist. Sci., 27(4):538–557,
2012.
[27] Richard Nickl and Sara van de Geer. Confidence sets in sparse regression. Ann. Statist., 41(6):2852–
2876, 2013.
[28] Benjamin Recht, Maryam Fazel, and Pablo A. Parrilo. Guaranteed minimum-rank solutions of linear
matrix equations via nuclear norm minimization. SIAM Rev., 52(3):471–501, 2010.
[29] Angelika Rohde and Alexandre B. Tsybakov. Estimation of high-dimensional low-rank matrices. Ann.
Statist., 39(2):887–930, 2011.
[30] Mark Rudelson and Roman Vershynin. Small ball probabilities for linear images of high dimensional
distributions. Technical report, University of Michigan, 2014. International Mathematics Research
Notices, to appear. [arXiv:1402.4492].
28
[31] Weijie Su and Emmanuel Candès. SLOPE is adaptive to unknown sparsity and asymptotically minimax.
Ann. Statist., 44(3):1038–1068, 2016.
[32] Michel Talagrand. Upper and lower bounds for stochastic processes, volume 60 of Ergebnisse der
Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics [Results
in Mathematics and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics]. Springer,
Heidelberg, 2014. Modern methods and classical problems.
[33] Robert Tibshirani. Regression shrinkage and selection via the lasso. J. Roy. Statist. Soc. Ser. B,
58(1):267–288, 1996.
[34] Sara van de Geer. Weakly decomposable regularization penalties and structured sparsity. Scand. J.
Stat., 41(1):72–86, 2014.
[35] Sara van de Geer, Peter Bühlmann, Ya’acov Ritov, and Ruben Dezeure. On asymptotically optimal
confidence regions and tests for high-dimensional models. Ann. Statist., 42(3):1166–1202, 2014.
[36] Sara A. van de Geer.
The deterministic lasso.
http://www.stat.math.ethz.ch/ geer/lasso.pdf.
Technical report, ETH Zürich, 2007.
[37] Sara A. van de Geer. High-dimensional generalized linear models and the lasso. Ann. Statist., 36(2):614–
645, 2008.
[38] G. Alistair Watson. Characterization of the subdifferential of some matrix norms. Linear Algebra Appl.,
170:33–45, 1992.
[39] Cun-Hui Zhang and Stephanie S. Zhang. Confidence intervals for low dimensional parameters in high
dimensional linear models. J. R. Stat. Soc. Ser. B. Stat. Methodol., 76(1):217–242, 2014.
[40] Peng Zhao and Bin Yu. On model selection consistency of Lasso. J. Mach. Learn. Res., 7:2541–2563,
2006.
7
Supplementary material: non-isotropic design
An inspection of Theorem 3.2 reveals no mention of an isotropicity assumption. There is
no choice of a Euclidean structure, and in fact, the statement itself is not even finite dimensional. All that isotropicity has been used for was to bound the “complexity function” r(·)
and the “sparsity function” ∆(·) in the three applications — the LASSO (in Theorem 1.4),
SLOPE (in Theorem 1.6) and the trace norm regularization (in Theorem 1.7). We may
apply Theorem 3.2 to situations that do not involve an isotropic vector and here we give
an example of how this may be done.
To simplify our presentation we will only consider ℓ1 and SLOPE regularization, which
may both be written as
d
X
βj t♯j ,
Ψ(t) =
j=1
where β1 ≥ · · · ≥ βd > 0 and t♯1 ≥ · · · ≥ t♯d ≥ 0 is the nondecreasing rearrangement of
(|tj |). As mentioned previously, the LASSO
case is recovered for β1 = · · · = βd = 1 and the
p
SLOPE norm is obtained for βj = C log(ed/j) for some constant C. We also denote by
BΨ (resp. SΨ ) the unit ball (resp. sphere) associated with the Ψ-norm.
Let Σ ∈ Rd×d be the covariance matrix of X and set D = {x ∈ Rd : Σ1/2 x 2 ≤ 1} to
be the corresponding ellipsoid. Naturally, if X is not isotropic than Σ is not the identity
matrix.
29
In order to apply Theorem 3.2, we need to bound from above the expectation of the
supremum of the Gaussian process indexed by ρBΨ ∩ rD:
ℓ∗ (ρBΨ ∩ rD) = E
Σ1/2 G, w
sup
(7.1)
w∈ρBΨ ∩rD
where G is a standard Gaussian vector in Rd .
We also need to solve the “sparsity equation”—that is, find ρ∗ > 0 for which ∆(ρ∗ ) ≥
∗
4ρ /5 where, for every ρ > 0,
∆(ρ) =
inf
sup
h∈ρSΨ ∩rD g∈Γ ∗ (ρ)
t
h, g
and Γt∗ (ρ) is the collection of all subgradients of Ψ of vectors in t∗ + (ρ/20)BΨ .
We will show that the same results that have been obtained for the LASSO and SLOPE
in Theorem 1.4 and Theorem 1.6 actually hold under the following assumption.
1/2
Assumption 7.1 Let j ∈ {1, ..., d} and denote by Σj• the j-th row of Σ1/2 . Let s ∈
P
√
{1, . . . , d} and set Bs = sj=1 βj / j.
1/2
1. There exists σ > 0 such that for all j ∈ {1, . . . , d}, Σj•
2. For all x ∈ (20Bs SΨ ) ∩ D, 2 Σ1/2 x
2
2
≤ σ.
≥ sup|J|≤s kxJ k2 .
SLOPE.
We first control the Gaussian mean width in (7.1) when Ψ(·) is the SLOPE norm.
p
Lemma 7.1 Set βj = C log(ed/j) and let Σ be a d × d symmetric nonnegative matrix for
1/2
which maxj Σj•
E
2
≤ σ. If D = {x ∈ Rd : Σ1/2 x
sup
w∈ρBΨ ∩rD
Σ
1/2
G, w ≤ min
(
ρ
C
2
≤ 1} then
)
√
r !
√
3 6σ
π
+r
,r d
8
2
Proof. Note that
E
sup
w∈ρBΨ ∩rD
Σ
1/2
G, w ≤ rE sup G, Σ
w∈D
1/2
w = rE kGk2 ≤ r
E kGk22
1/2
√
≤ r d.
Next, let H : Rd → R be defined by H(u) = sup Σ1/2 u, w : w ∈ ρBΨ ∩ rD and recall
that G is the standard Gaussian vector in Rd . It is straightforward to verify that
H(G) ≤ sup
w∈ρBΨ
Σ
1/2
ξj♯
ρ
max p
G, w ≤
C 1≤j≤d log(ed/j)
30
where we set (ξj )dj=1 = Σ1/2 G and (ξj♯ )dj=1 is the non-increasing rearrangement of (|ξj |)dj=1 .
Observe that for u, v ∈ Rd ,
|H(u) − H(v)| ≤
sup
w∈ρBΨ ∩rD
| Σ1/2 (u − v), w | ≤ r sup | u − v, w | = r ku − vk2 ,
w∈B2d
implying that H is a Lipschitz function with constant r; thus, it follows from p. 21 in
Chapter 1 of [14] that
r
π
EH(G) ≤ Med(H(G)) + r
,
(7.2)
2
where Med(H(G)) is the median of H(G).
Hence, to obtain the claimed bound on E supw∈ρBΨ Σ1/2 G, w it suffices to establish
a suitable upper estimate on the median of max1≤j≤d √
ξj♯
log(ed/j)
. With that in mind, let
ξ1 , . . . , ξN be mean-zero Gaussian variables and assume that for every j = 1, . . . , d,
E exp(ξj2 /L2 ) ≤ e
(7.3)
for some L > 0. Note that in our case, ξ1 , . . . , ξN satisfying (7.3) for L = 3σ/8.
By Jensen’s inequality,
!
!
2
j
j
d
♯
(ξk♯ )2
ξk
1X
1X
ed
1 X (ξk )2
E exp
E exp
≤
≤
≤ ;
E exp
2
2
2
j
L
j
L
j
L
j
k=1
k=1
k=1
hence,
j
Pr
♯
1 X (ξk )2
≥ 2 log (ed/j)
j
L2
k=1
!
≤ exp (− log (ed/j)) =
j
.
ed
(7.4)
Let q ≥ 0 be the integer that satisfies 2q ≤ d < 2q+1 . It follows from (7.4) that with
probability at least
q−1 ℓ
X
2
2q − 1
1
1−
=1−
> ,
ed
ed
2
ℓ=0
for every ℓ = 0, · · · , q − 1,
ℓ
(ξ2♯ ℓ )2
2
1 X ♯ 2
≤ ℓ
(ξj ) ≤ 2L2 log(ed/2ℓ ).
2
j=1
Moreover, for 2ℓ ≤ j < 2ℓ+1 , we have ξj♯ ≤ ξ2♯ ℓ and log(ed/2ℓ ) ≤ 2 log(ed/j); also for
2q ≤ j ≤ d, we have ξj♯ ≤ ξ2♯ q−1 and log(ed/2ℓ ) ≤ 3 log(ed/j). Therefore,
Pr
max p
1≤j≤d
ξj♯
log(ed/j)
proving the requested bound on Med(H(G)).
31
≤
√
6L
!
1
> ,
2
(7.5)
Observe that up to constant σ, we actually recover the same result as in (5.2); therefore,
one may choose the same “complexity function” r(·) as in the proof of Theorem 1.6.
Let us turn to a lower bound on the “sparsity function”.
Lemma 7.2 There exists an absolute
constant 0 < c < 80 for which the following holds.
P
√
Let s ∈ {1, . . . , d} and set Bs = j≤s βj / j. Assume that for every x ∈ (80Bs Sψ ) ∩ D one
has Σ1/2 x 2 ≥ (1/2) sup|J|≤s kxJ k2 . Let ρ > 0 and assume further that there is a s-sparse
vector in t∗ + (ρ/20)BΨ . If 80Bs ≤ ρ/r(ρ) then
∆(ρ) =
inf
sup
h∈ρSΨ ∩r(ρ)D g∈Γt∗ (ρ)
h, g ≥
4ρ
.
5
Proof. Let h ∈ ρSΨ ∩ r(ρ)D and denote by (h♯j ) the non-increasing rearrangement of (|hj |).
It follows from the proof of Lemma 4.3 that
X
17ρ
sup h, g ≥
−2
βj h♯j .
20
g∈Γt∗ (ρ)
j≤s
Let h♯,s be the s-sparse vector with coordinates given by h♯j for 1 ≤ j ≤ s and 0 otherwise.
We have
r(ρ)
1
h
∈ SΨ ∩
D
D ⊂ SΨ ∩
ρ
ρ
80Bs
√
implying that 2 Σ1/2 h 2 ≥ h♯,s 2 . Furthermore, since h♯j ≤ h♯ 2 / j for every 1 ≤ j ≤ s,
we have
X
βj h♯j ≤ h♯,s Bs ≤ 2Bs Σ1/2 h ≤ 2Bs r(ρ).
2
2
j≤s
Hence, if ρ ≥ 80r(ρ)Bs then ∆(ρ) ≥ 4ρ/5.
We thus recover the same condition as in Lemma 4.3, implying that Theorem 1.6 actually holds under the weaker Assumption 7.1: let X be an L-subgaussian random vector
whose covariance matrix satisfies
Assumption 7.1. The SLOPE estimator with regulariza√
tion parameter λ ∼ kξkLq / N satisfies, with probability at least 1 − δ − exp(−c0 N L8 ),
ed
s
and
Ψ(t̂ − t∗ ) ≤ c3 kξkLq √ log
s
N
Σ1/2 (t̂ − t∗ )
2
2
≤ c3 kξk2Lq
ed
s
log
N
s
when N ≥ c4 s log(ed/s) and when there is a s-sparse vector close enough to t∗ .
The LASSO.
P
√
√
Here, for every 1 ≤ j ≤ d, βj = 1; BΨ = B1d ; and Bs = sj=1 1/ j ≤ 2 s.
1/2
Lemma 7.3 Let Σ be a d × d symmetric nonnegative matrix for which maxj Σj•
If D = {x ∈ Rd : Σ1/2 x
E
≤ 1} then every ρ > 0 and r > 0,
o
n √
p
Σ1/2 G, w ≤ min r d, ρσ log(ed) .
sup
2
w∈ρB1d ∩rD
32
2
≤ σ.
Proof. Note that
E
sup
w∈ρB1d ∩rD
1/2
√
Σ1/2 G, w ≤ rE sup G, Σ1/2 w = rE kGk2 ≤ r E kGk22
≤ r d.
w∈D
Next, set (ξj )dj=1 = Σ1/2 G and let (ξj♯ )dj=1 be the non-increasing rearrangement of (|ξj |).
Therefore, ξ1 , . . . , ξd are mean-zero Gaussian variables and satisfy E exp(ξj2 /L2 ) ≤ e for
L = 3σ/8. It is evident that
!
!!
!
d
X
ξj2
(ξ1♯ )2
(ξ1♯ )2
≤ log (ed) ,
E exp
≤ log E exp
≤ log
E
L2
L2
L2
j=1
and therefore,
E
sup
w∈ρB1d ∩rD
1/2 3ρσL p
Σ1/2 G, w ≤ E sup Σ1/2 G, w ≤ ρE (ξ1♯ )2
≤
log (ed).
8
w∈ρB d
1
Lemma 7.3 leads to a slightly different result than in the isotropic case (Lemma 5.3), and
as a consequence, r(·) has to be slightly modified. A straightforward computation shows
that
!
r
kξkLq d
log(ed)
2
rM (ρ) .L,q,δ min
, ρσ kξkLq
N
N
and
2
rQ
(ρ)
.L
(
ρ2 σ2
N
0
log c(L)d
N
if N &L d
otherwise.
and still r(ρ) = max{rM (ρ), rQ (ρ)}.
Finally, let us prove the sparsity condition.
Lemma 7.4 There exists an absolute
constant 0 < c < 80 for which the following holds.
P
√
Let s ∈ {1, . . . , d} and set Bs = j≤s 1/ j. Assume that for every x ∈ 20Bs S(ℓd1 ) ∩ D
one has Σ1/2 x 2 ≥ (1/2) sup|J|≤s kxJ k2 . Let ρ > 0 and assume further that there is a
s-sparse vector in t∗ + (ρ/20)B1d . If 20Bs ≤ ρ/r(ρ) then
∆(ρ) =
inf
sup
h∈ρS(ℓd1 )∩r(ρ)D g∈Γt∗ (ρ)
h, g ≥
4ρ
.
5
Proof. Let h ∈ ρS(ℓd1 ) ∩ r(ρ)D and denote by (h♯j ) the non-increasing rearrangement of
(|hj |). It follows from the proof of Lemma 4.2 that
X ♯
sup h, g ≥ ρ − 2
hj .
g∈Γt∗ (ρ)
j≤s
33
Let h♯,s be the s-sparse vector with coordinates given by h♯j for 1 ≤ j ≤ s and 0 otherwise.
Observe that
r(ρ)
1
h
d
d
∈ S(ℓ1 ) ∩
D,
D ⊂ S(ℓ1 ) ∩
ρ
ρ
40Bs
and therefore 2 Σ1/2 h 2 ≥ h♯,s
It follows that
X ♯ √
hj ≤ s h♯
j≤s
2
2
.
≤ Bs h♯
2
≤ 2Bs Σ1/2 h
2
≤ 2Bs r(ρ),
and in particular, if ρ ≥ 20r(ρ)Bs then ∆(ρ) ≥ 4ρ/5.
Using the estimate on r(·) and Lemma 7.4, it is evident that when N &L,q,δ sσ 2 log(ed),
one has ∆(ρ∗ ) ≥ 4ρ∗ /5 for
r
log(ed)
∗
ρ ∼L,q,δ kξkLq s
N
and if there is a s-sparse vector in t∗ + (ρ∗ /20)B1d .
Finally, one may choose the regularization parameter by setting
r
log(ed)
r 2 (ρ∗ )
∼L,q,δ kξkLq s
.
λ∼
∗
ρ
N
It follows that if X is an L-subgaussian random vector that satisfies Assumption 7.1 then
with probability larger than 1 − δ − 2 exp(−c0 N L8 ),
r
r
log(ed)
s log(ed)
1/2
∗
∗
∗
∗
and Σ (t̂ − t ) ≤ r(ρ ) = c2 (L, δ) kξkLq
.
t̂ − t 1 ≤ ρ = c1 (δ) kξkLq s
N
N
2
34
| 10math.ST
|
Geometric operations implemented by conformal geometric algebra neural nodes
Eckhard HITZER
University of Fukui
Abstract: Geometric algebra is an optimal frame work for calculating with vectors. The geometric algebra
of a space includes elements that represent all the its subspaces (lines, planes, volumes, ...). Conformal geometric algebra expands this approach to elementary representations of arbitrary points, point pairs, lines,
circles, planes and spheres. Apart from including curved objects, conformal geometric algebra has an elegant unified quaternion like representation for all proper and improper Euclidean transformations, including
reflections at spheres, general screw transformations and scaling. Expanding the concepts of real and complex neurons we arrive at the new powerful concept of conformal geometric algebra neurons. These neurons
can easily take the above mentioned geometric objects or sets of these objects as inputs and apply a wide
range of geometric transformations via the geometric algebra valued weights.
arXiv:1306.1358v1 [cs.CV] 6 Jun 2013
1.
Introduction
The co-creator of calculus W. Leibniz dreamed of a new
type of mathematics in which every number, every operation
and every relation would have a clear geometric counterpart.
Subsequently the inventor of the concept of vector space and
our modern notion of algebra H. Grassmann was officially
credited to fulfill Leibniz’s vision. Contemporary to Grassmann was W. Hamilton, who took great pride in establishing the algebra of rotation generators in 3D, which he himself called quaternion algebra. About 30 years later W. Clifford successfully fused Grassmann’s and Hamilton’s work
together in what he called geometric algebra. Geometric algebra can be understood as an algebra of a vector space and
all its subspaces equipped with an associative and invertible
geometric product of vectors.
It is classic knowledge that adding an extra dimension for
the origin point to a vector space leads to the projective geometry of rays, where Euclidean points correspond to rays.
Adding one more dimension for the point of infinity allows
to treat lines as circles through infinity, and planes as spheres
through infinity and unifies their treatment. This socalled
conformal geometric algebra represents geometric points,
spheres and planes by 5D vectors. The inner product of two
conformal points yields their Euclidean distance. Orthogonal
transformations preserve inner products and in the 5D model
Euclidean distances, provided that they also keep the point at
infinity invariant. According to Cartan and Dieudonné all orthogonal transformations are products of reflections. Proper
and improper Euclidean transformations can therefore be expressed nowadays in conformal geometric algebra as elegant
as complex numbers express rotations in 2D. And the language is not limited to 3+2D, n+2D follows the very same
principles.
Transformation groups generated by products of reflections in geometric algebra are known as Clifford (or Lipschitz or versor) groups [7, 10]. Versors (Clifford group or
Lipschitz elements) are simply the geometric products of the
normal vectors to the (hyper) planes of reflection. These
versors assume the role of geometric weights in concept of
conformal geometric algebra neural nodes. Precursors for
these nodes are complex [1], quaternion [3], and Clifford
spinor [2, 6] neurons. They have also been named versor
Clifford neurons [13], but regarding their fundamental geometric nature even the term geometric neurons seems fully
justified.
2.
Geometric algebra
Definition 1 (Clifford geometric algebra). A Clifford geometric algebra Gp,q is defined by the associative geometric
product of elements of a quadratic vector space Rp,q , their
linear combination and closure. Gp,q includes the field of
real numbers R and the vector space Rp,q as subspaces. The
geometric product of two vectors is defined as
ab = a · b + a ∧ b,
(1)
where a · b indicates the standard inner product and the
bivector a ∧ b indicates Grassmann’s antisymmetric outer
product. a ∧ b can be geometrically interpreted as the oriented parallelogram area spanned by the vectors a and b.
Geometric algebras are graded, with grades (subspace dimensions) ranging from zero (scalars) to n = p + q (pseudoscalars, n-volumes).
For example geometric algebra G3 = G3,0 of threedimensional Euclidean space R3 = R3,0 has an eightdimensional basis of scalars (grade 0), vectors (grade 1),
bivectors (grade 2) and trivectors (grade 3). Trivectors in
G3 are also referred to as oriented volumes or pseudoscalars.
Using an orthonormal basis {e1 , e2 , e3 } for R3 we can write
the basis of G3 as
{1, e1 , e2 , e3 , e2 e3 , e3 e1 , e1 e2 , i = e1 e2 e3 }.
(2)
In (2) i is the unit trivector, i.e. the oriented volume of a unit
cube. The even subalgebra G3+ of G3 is isomorphic to the
quaternions H of Hamilton. We therefore call elements of
G3+ rotors, because they rotate all elements of G3 . The role
of complex (and quaternion) conjugation is naturally taken
by reversion (a1 , a2 , . . . , as ∈ Rp,q , s ∈ N)
(a1 a2 . . . as )∼ = as . . . a2 a1 .
(3)
The inverse of a non-null vector a ∈ Rp,q is
a−1 =
a
.
a2
(4)
A reflection at a hyperplane normal a is
x0 = −a−1 xa.
(5)
A rotation by the angle θ in the plane of a unit bivector i
can thus be given as the product R = ab of two vectors a,
b from the i-plane (i.e. geometrically as a sequence of two
reflections) with angle θ/2,
Rp,q 3 x → R−1 xR ∈ Rp,q ,
e0 · e∞ = −1,
(7)
e0 E = −e0 , Ee0 = e0 , e∞ E = e∞ ,
Ee∞ = −e∞ , E 2 = 1.
Extracting a certain grade part from the geometric product of
two blades Ak and Bl has a deep geometric meaning.
One example is the grade l − k part (contraction [8]) of
the geometric product Ak Bl , that represents the orthogonal
complement of a k-blade Ak in an l-blade Bl , provided that
Ak is contained in Bl
(9)
Another important grade part of the geometric product of Ak
and Bl is the maximum grade l + k part, also called the outer
product part
Ak ∧ Bl = hAk Bl il+k .
(10)
If Ak ∧ Bl is non-zero it represents the union of the disjoint
(except for the zero vector) subspaces represented by Ak and
Bl .
The dual of a multivector A is defined by geometric division with the pseudoscalar I = A∗ = A I − 1, which maps
k-blades into (n − k)-blades, where n = p + q. Duality transforms inner products (contractions) to outer products
and vice versa. The outer product null space representation
(OPNS) of (8) is therefore transformed by duality into the
socalled inner product null space (IPNS) representation
x ∧ A = 0 ⇐⇒ x · A∗ = 0.
(11)
2.1 Conformal geometric algebra
Conformal geometric algebra embeds the geometric algebra
G3 of R3 in the geometric algebra of R4,1 = R3+1,0+1 Given
an orthonormal basis for R4,1
{e1 , e2 , e3 , e+ , e− }
b=
A
n
X
(−1)k hAik .
A Clifford (or Lipschitz) group is a subgroup in Gp,q generated by non-null vectors x ∈ Rp,q
Γp,q = {m ∈ Gp,q | ∀x ∈ Rp,q , m
b −1 x m ∈ Rp,q } (19)
For every m ∈ Γp,q we have mm
e ∈ R. Examples are m = a
and m = R = ab for reflections and rotations, respectively.
Clifford groups include Pin(p, q), Spin(p, q), and
Spin+ (p, q) groups as covering groups of orthogonal
O(p, q), special orthogonal SO(p, q) and SO+ (p, q) groups,
respectively. Conformal transformation groups C(p, q)
preserve inner products (angles) of vectors in Rp,q up to a
change of scale. C(p, q) is isomorphic to O(p + 1, q + 1).
The metric affine group (orthogonal transformations and
translations) of Rp,q is a subgroup of O(p + 1, q + 1), and
can be implemented as a Clifford group in Gp+1,q+1 .
2.2 Geometric objects
The conformal geometric algebra G4,1 provides us with a superb model [7–9] of Euclidean geometry. The basic geometric objects in conformal geometric algebra are homogeneous
conformal points
1
P = p + p2 e∞ + e0 ,
(20)
2
p
where p ∈ R3 , p =
p2 . The +e0 term shows that we
include projective geometry. The second term + 12 p2 e∞ ensures, that conformal points are isotropic vectors
P 2 = P P = P · P = 0.
e21 = e22 = e23 = e2+ = −e2− = 1,
(13)
we introduce a change of basis for the two additional dimensions {e+ , e− } by
1
(e− − e+ ),
2
e∞ = e− + e+ .
(14)
The vectors e0 and e∞ are isotropic vectors, i.e.
e20 = e2∞ = 0,
(18)
k=0
(12)
with
e0 =
(17)
As we will now see, conformal geometric algebra is advantageous for the unified representation of many types of
geometric transformations. In the next section we will further
consider the unified representation of eight different types of
Euclidean geometric objects possible in conformal geometric
algebra.
Combining reflections (5) leads to an overall sign (parity)
for odd and even numbers of (reflection plane) vectors a, b,
etc. Therefore we define the grade involution
x ∈ V = span[a1 , . . . ak ] ⇔ x∧a1 ∧a2 ∧. . .∧ak = 0. (8)
Bl = hAk Bl il−k .
(16)
We further have the following useful relationships
where |a|, and |b| are the lengths of a, b. This description
corresponds exactly to using quaternions.
Blades of grade k, 0 ≤ k ≤ n = p + q are the outer
products of k vectors al (1 ≤ l ≤ k) and directly represent
the k-dimensional vector subspaces V spanned by the set of
vectors al (1 ≤ l ≤ k). This is also called the outer product
null space (OPNS) representation.
Ak
E = e∞ ∧ e0 = e+ ∧ e− .
(6)
where the vectors a, b are in the plane of the unit bivector
i ∈ Gp,q if and only if a ∧ i = b ∧ i = 0. The rotor R can
also be expanded as
R = ab = |a||b|exp(θi/2),
and have inner and outer products of
(15)
(21)
In general the inner product of two conformal points P1 and
P2 gives their Euclidean distance
1
P1 · P2 = − (p1 − p2 )2 .
2
(22)
Orthogonal transformations preserve this distance. The outer
product of two conformal points spans a conformal point pair
(in OPNS) as in Fig. 1
P p = P1 ∧ P2 .
(23)
Fig. 1 Conformal point pair P1 ∧ P2 with center C.
This and the following illustrations were produced with the
OpenSource visual software CLUCalc [14], which is also
based on conformal geometric algebra.
The outer product three conformal points gives a circle (cf.
Fig. 2)
Circle = P1 ∧ P2 ∧ P3 .
(24)
Fig. 4 Conformal line P1 ∧P2 ∧e∞ with 3D direction vector
d.
also called bivector moment of the line. Conformal planes
(spheres through infinity, cf. Fig. 5) are represented by
P lane = P1 ∧ P2 ∧ P3 ∧ e∞ = Circle ∧ e∞ .
(27)
Further flattened objects are
Fig. 2 Conformal circle P1 ∧ P2 ∧ P3 with center C.
The conformal outer product of four conformal points
gives a sphere (cf. Fig. 3)
Sphere
= P1 ∧ P2 ∧ P3 ∧ P4 .
(25)
Fig. 5 Conformal plane P1 ∧ P2 ∧ P3 ∧ e∞ .
P ∧ e∞ ,
(28)
a (flat) finite–infinite point pair, and
−is E = Sphere ∧ e∞ ∼ I5 = iE,
(29)
the in 5D embedded (flat) 3D Euclidean space R3 .
2.3 General reflection and motion operators (motors)
As in G3 , 3D rotations around the origin in the conformal
model G4,1 are still represented by rotors (6). The standard
IPNS representation of a conformal plane, i.e. the dual of
the direct OPNS representation (27) results in a vector, that
both describes the unit normal direction n = c⊥ /|c⊥ | and
the position (signed distance d = c⊥ · n from the origin) of
a plane by
µ = P lane∗ = n + de∞ .
Fig. 3 Conformal sphere P1 ∧ P2 ∧ P3 ∧ P4 with center C.
If one of the points is at infinity, we get conformal lines
(circles through infinity, cf. Fig. 4)
Line = P1 ∧ P2 ∧ e∞ = P p ∧ e∞
= 2r(d ∧ ce∞ − dE) ∝ m e∞ − dE,
(26)
where d is the direction vector of the line and c the 3D midpoint between P1 and P2 , 2r = |p2 − p2 |. c can be replaced by any point p on the line. m = d ∧ c = d ∧ p is
(30)
We can reflect the conformal point X at the general plane
(30) similar to (5) with
X 0 = −µ−1 Xµ.
(31)
A reflection of a general conformal object O (point, point
pair, line, ..., sphere) at the general plane (30) similar to (31)
with
b
O0 = µ−1 Oµ,
(32)
where the grade involution takes care of the sign changes.
Just as we obtained rotations (6) by double reflections we
now obtain rotations around arbitrary axis (lines of intersection of two planes) by double reflections at planes µ1 and µ2
X 0 = R−1 XR,
R = µ1 µ2 .
(33)
And if the two planes µ1 and µ2 happen to be parallel, i.e.
n1 = ±n2 , we instead get a translation by twice the Euclidean distance t/2 between the planes
X 0 = T −1 XT,
1
1
T (t) = µ1 µ2 = 1 + te∞ = exp( te∞ ).
2
2
Table 2 Summary of motions and scaling O → W −1 OW in
conformal GA. O is a general conformal object (point, point
pair, circle, sphere, flat point, line, plane, 3D space). The
conformal versor W represents the motion operator.
Motion operator Versor W
Rotor
exp(θi/2)
Translator
exp( 12 te∞ )
Motor
TR
Scale operator
T −1 (c) exp (E 21 log s) T (c)
(34)
A general motion operator (motor) results from combining
rotations (33) and translations (34) to
M = T R.
(35)
The standard IPNS representation of a sphere results in the
vector
1
σ = S ∗ = C − r2 e∞ ,
2
(36)
which is exactly the dual S ∗ = S i−1
3 E of (25). The expression (36) shows that in conformal GA a point can be regarded
as a sphere with zero radius. We can reflect (invert) the conformal point X at the sphere (36) [or at a conformal point]
similar to (5) and (31) simply by
X 0 = −σ −1 Xσ.
(37)
The double reflection at two concentric spheres σ1 and σ2
(centered at c) results in a rescaling operation [8] with factor
s and center c
X 0 = Z −1 XZ,
Z = σ1 σ2 = T −1 (c) exp (E
Fig. 6 Examples of combinations of geometric transformations: rotor R, translator T , scaling operator Z, motor and
scaling op. Ms .
1
log s) T (c).
2
(38)
The following illustrations have been produced with the
socalled Space Group Visualizer [15], a symmetry visualization program for crystallographic space group, which is also
based on conformal geometric algebra software.
Figure 7 shows a glide reflection, i.e. a reflection combined with a translation parallel to the plane of reflection.
Figure 8 shows a point inversion. Figure 9 shows a screw
Table 1 summarizes the various reflections possible in conformal GA. Table 2 summarizes the motions and scaling posb in conforTable 1 Summary of reflections O → −W −1 OW
mal GA. O is a general conformal object (point, point pair,
circle, sphere, flat point, line, plane, 3D space). The conformal versor W represents the mirror.
Mirror Versor W
plane
n + de∞
point
p + 12 p2 e∞ + e0
sphere C − 12 r2 e∞
line
m e∞ − dE
Fig. 7 Examples of glide reflection, which is a reflection
combined with a translation parallel to the plane of reflection.
sible in conformal GA.
2.4 Combining elementary transformations
All operations in Tables 1 and 2 can be combined to give further geometric transformations, like rotoinversions, glide reflections, etc.! Algebraically the combination is represented
by simply computing the product of the multivector versors
of the elementary transformations explained above. Figure
6 shows some examples of combinations of geometric transformations applied to a point pair P1 ∧ P2 . The overbars
abbreviate the inverse of an operator.
transformation, which is a rotation followed by a translation
along the axis of the rotation. Figure 10 shows a rotoinversion, which is a combination of point inversion and rotation.
A rotoinversion is also equivalent to a rotation followed by a
reflection at a plane perpendicular to the axis of rotation.
3.
Geometric neurons
Conformal versors V describe in conformal Clifford group
[10] representations the above mentioned transformations
Fig. 8 Examples of point inversion.
Fig. 10 Examples of a 90 degree rotoinversion, which is a
combination of point inversion and a 90 degree rotation.
4.
Fig. 9 Examples of a screw transformation.
of arbitrary conformal geometric object multivectors X ∈
G(R3+1,1 ) of section 2.2
X 0 = (−1)v V −1 XV,
(39)
where the versor is a geometric product of v invertible vectors ∈ R3+1,1 .
The geometric neuron (GN) is characterized by a twosided multiplication of a single multivector weight versor
W ∈ G(R3+1,1 ), input multivectors X ∈ G(R3+1,1 ), and
multivector thresholds Θ ∈ G(R3+1,1 )
Y = (−1)w W −1 XW + Θ,
Conclusions
We have introduced the concept of geometric algebra as the
algebra of a vector space and all its subspaces. We have
shown how conformal geometric algebra embeds and models
Euclidean geometry. Outer products of points (including the
point at infinity) model the elementary geometric objects of
points, point pairs, flat points, circles, lines, spheres, planes
and the embedded 3D space itself.
The unified representation of affine Euclidean transformations (including translations) by orthogonal transformations
in the conformal model allows the construction of a geometric neuron, whose versor weights can learn these transformations precisely. The transformations were illustrated in
detail.
(40)
where w represents the number of vector factors (parity) in
W . The theory, optimization and example simulation of such
geometric neurons has been studied in [13]. It was shown
that e.g. the inversion at a sphere can be learned exactly by a
geometric neuron, outperforming linear networks and multilayer perceptrons with the same or higher number of degrees
of freedom.
Acknowledgments
The author would like to thank K. Tachibana (COE FCS,
Nagoya) and S. Buchholz (Kiel). He further wishes to thank
God, the creator: Soli Deo Gloria, and his very supportive
family.
References
[1] A. Hirose, Complex-Val. NNs, Springer, Berlin, 2006.
[2] S. Buchholz, G. Sommer, Quat. Spin. MLP, Proc.
ESANN 2000, d-side pub., pp. 377–382, 2000
[3] N. Matsui, et al, Quat. NN with geome. ops., J. of Intel.
& Fuzzy Sys. 15 pp. 149-1164 (2004).
[4] E. Hitzer, Multivector Diff. Calc., Adv. Appl. Cliff. Alg.
12(2) pp. 135-182 (2002).
[5] D. Hestenes, G. Sobczyk, Cliff. Alg. to Geom. Calc.,
Kluwer, Dordrecht, 1999.
[6] S. Buchholz, A Theory of Neur. Comp. with Cliff. Alg.,
TR No. 0504, Univ. of Kiel, May 2005.
[7] D. Hestenes, H. Li, A. Rockwood, New Alg. Tools for
Class. Geom., in G. Sommer (ed.), Geom. Comp. with
Cliff. Alg., Springer, Berlin, 2001.
[8] L. Dorst, D. Fontijne, S. Mann, Geom. Alg. for Comp.
Sc., Morgan Kaufmann Ser. in Comp. Graph., San Francisco, 2007.
[9] P. Angles, Conf. Groups in Geom. and Spin Struct.,
PMP, Birkhauser, Boston, 2007.
[10] P. Lounesto, Cliff. Alg. and Spinors, 2nd ed., CUP,
Cambridge, 2006.
[11] S. Buchholz, K. Tachibana, E. Hitzer, Optimal Learning Rates for Cliff. Neurons, in LNCS 4668, Springer,
Berlin, 2007, pp. 864–873.
[12] E. Hitzer, et al. Carrier method f. gen. eval. & control of
pose, molec. conform., tracking, and the like acc. Adv.
in Appl. Cliff. Algs., 26 pp. (2008).
[13] S. Buchholz, E. Hitzer, K. Tachibana, Coordinate independent update formulas for versor Clifford neurons
Proceedings of Joint Conference SCIS & ISIS 2008,
Nagoya, Japan, pp. 814–819, (2008).
[14] C.
Perwass,
Visual
Calculator
CLUCalc,
http://www.clucalc.info
[15] C. Perwass, E. Hitzer, Space Group Visualizer,
http://www.spacegroup.info
| 9cs.NE
|
IEEE TRANSACTIONS ON BIG DATA, VOL. XX, NO. X, JUNE 2016
1
Towards Real-Time Detection and Tracking of
Spatio-Temporal Features: Blob-Filaments in
Fusion Plasma
arXiv:1505.03532v3 [cs.DC] 2 Jul 2016
Lingfei Wu, Kesheng Wu, Alex Sim, Michael Churchill,
Jong Y. Choi, Andreas Stathopoulos, CS Chang, and Scott Klasky
Abstract—A novel algorithm and implementation of real-time identification and tracking of blob-filaments in fusion reactor data is
presented. Similar spatio-temporal features are important in many other applications, for example, ignition kernels in combustion and
tumor cells in a medical image. This work presents an approach for extracting these features by dividing the overall task into three
steps: local identification of feature cells, grouping feature cells into extended feature, and tracking movement of feature through
overlapping in space. Through our extensive work in parallelization, we demonstrate that this approach can effectively make use of a
large number of compute nodes to detect and track blob-filaments in real time in fusion plasma. On a set of 30GB fusion simulation
data, we observed linear speedup on 1024 processes and completed blob detection in less than three milliseconds using Edison, a
Cray XC30 system at NERSC.
Index Terms—Big data analytics, spatio-temporal feature, real-time detection and tracking, blob-filaments, fusion plasma.
F
1
I NTRODUCTION
A
W ide variety of “big data” such as simulations of
diesel combustion and images of tissue from biopsy,
are spatiotemporal in nature [1], [2]. When analyzing these
data sets, a common task is to find coherent structure in
space and time, for example, ignition kernels in combustion,
and cancerous cells in medical images. There are many
possible approaches to identify such a feature based on the
application requirements [3], [4], [5], [6]. However, when
faced with tight time constraints many of these techniques
are too slow to produce an satisfactory answer.
Our work was originally motivated by the need to detect
spatio-temporal feature associated with instabilities in fusion plasma. Magnetic confinement fusion has the potential
to be an inexhaustible source of clean energy; and billions of
dollars have been invested in developing fusion reactors,
like the ITER project [7]. However, steady-state plasma
confinement is often interrupted by blob filaments driven by
the edge turbulence. A blob filament (or blob) is a magneticfield-aligned plasma structure that appears near the edge
of the confined plasma, and has significantly higher density
and temperature than the surrounding plasma [8]. Blobs can
•
•
•
•
L. Wu and A. Stathopoulos are with the Department of Computer Science,
College of William and Mary, Williamsburg, VA, 23185.
E-mail: {lfwu,andreas}@cs.wm.edu
K. Wu and A. Sim are with Lawrence Berkeley National laboratory,
Berkeley, CA, 94720.
E-mail: {KWu,ASim}@lbl.gov
M. Churchill and C. Chang are with Princeton Plasma Physics Laboratory, Princeton, NJ, 08536.
E-mail: [email protected], [email protected]
J. Choi and S. Klasky are with Oak Ridge National Laboratory, Oak Ridge,
TN, 37831.
E-mail: {choij,klasky}@ornl.gov
Manuscript received in May 4th, 2015. Revised manuscript received in April
18th, 2016.
also be considered as outliers because they are rare events
that convect filaments of plasma outwards towards the
containment wall, causing substantial heat loss, degradation
of the plasma confinement, and erosion of the containment
wall. By identifying and tracking these blob filaments from
fusion plasma data streams, physicists can improve their
understanding of the dynamics and interactions of such
coherent structures (blobs) with edge turbulence.
Fusion experiments and simulations could easily produce many terabytes per second; and features such as blobs
have to be detected in milliseconds in order for the control
system to have a chance to take mitigating actions. Though
there are many well known feature extraction methods for
detecting outliers, they often have some shortcomings. Classical multi-dimensional outlier detection techniques are designed to detect global outliers. However, these techniques
do not distinguish between non-spatial attributes and spatial attributes and do not consider apriori information about
the statistical distribution of the data [9]. Since spatiotemporal data types have unique characteristics and their
relations are more complicated than ordinary data, dedicated outlier detection techniques are typically required to
examine anomalies in data across space and time [5]. In this
work, we propose an approach for detecting and tracking
spatio-temporal features such as blobs by breaking down
the process into three steps: (1) find cells that satisfying
application specific requirements, (2) group cells into spatial
features, and (3) track features by the amount of overlap
in space. By varying the first step, this procedure could be
applying to different applications. Earlier, this approach was
applied to data from regular meshes [2], [10]. In this work,
we will demonstrated that it can also be effectively applied
to irregular mesh data.
This work addresses several challenges exemplified by
IEEE TRANSACTIONS ON BIG DATA, VOL. XX, NO. X, JUNE 2016
Blob Detection
Data Stream
2
could satisfy the millisecond time requirement. Additional
contributions of this work include:
•
Data Generation
(Experiments or
Simulations)
Data Hub
Blob Detection
•
Blob Detection
Fig. 1: A real-time data analysis frame for finding blobfilaments in fusion plasma data streams
the detection of blobs in fusion plasma. First off, fusion
experiments and numerical simulations can easily generate
massive amounts of data per run. During a magnetic fusion
device experiment (or ”shot”), terabytes of data are generated over short time periods (on the order of hundreds
of seconds). In the XGC1 fusion simulation [11], [12], a
few tens of terabytes can be generated per second. Timely
access to this amount of data can already be a challenge [13],
[14], but analyzing all this data in real time is impractical.
Currently, there are three types of analyses in most of
fusion experiments: in-shot-analysis, between-shot-analysis,
and post-run-analysis. All existing blob detection methods
address post-run-analysis, but in this work, we focus on
the more challenging first two cases to provide a real-time
analysis so that scientists can monitor the progress of fusion
experiments. Figure 1 presents a real-time analysis frame
for finding blob-filaments in fusion plasma data streams. To
perform this data analysis in real time, we utilize effectively
modern supercomputers to address the high volume and
velocity challenges arising from fusion plasma big data.
This work has been integrated into the International
Collaboration Framework for Extreme Scale Experiments
(ICEE), a wide-area in-transit data analysis framework for
near real-time scientific applications [15]. ICEE takes advantage of an efficient IO solution ADIOS [16], and a cuttingedge indexing solution FastBit [10], to design and construct
a real-time remote data processing framework over widearea networks for international collaborations such as the
ITER project. In this system, a blob detection algorithm
is used to monitor the health of fusion experiments at
the Korea Superconducting Tokamak Advanced Research
(KSTAR). However, existing data analysis approaches are
often single-threaded, only for post-run analysis, and take
a long time to produce results. Also, compared to the
simulation data, the resolution of the raw camera data may
be coarse, but interesting features can still be identified after
normalization. In order to meet real-time feedback requirement, we develop a real-time blob detection method, which
can leverage in-situ raw data in the ICEE server and find
blob-filaments efficiently during fusion experiments. Our
blob detection algorithm is not limited to KSTAR only, and
can be applied to other fusion experiments and simulations.
In this research, we apply the three-step approach to
detect and track blob structures in fusion data, with the
goal of achieving millisecond response time on terabytes
of data. With this response space, it is possible for the
control system of the magnetic confinement fusion reactor to
implement mitigating strategies in real time. To the best of
our knowledge, this is the first time a blob detection method
•
•
•
We illustrate how to adopt the three-step approach
to detect and track blob-filaments as an example of
spatial-temporal feature on an irregular mesh.
We propose a two-phase region outlier detection
method for finding blob-filaments. In the first phase,
we apply a distribution-based outlier detection
scheme to identify blob candidate points. In the
second phase, we adopt a fast two-pass connected
component labeling (CCL) algorithm from [17] to find
different region outliers on an irregular mesh.
We develop a high-performance blob detection approach to meet real-time feedback requirements by
exploiting many-core architectures in a large cluster.
We propose a scheme to efficiently track the movement of region outliers by linking the centers of the
region outlier over consecutive frames.
We have implemented our blob detection algorithm
with hybrid MPI/OpenMP, and demonstrated the
effectiveness and efficiency of our implementation
with a set of data from the XGC1 fusion simulations. Our tests show that we can achieve linear time
speedup and complete blob detection in two or three
milliseconds using a cluster at NERSC. In addition,
we demonstrate that our method is more robust
than recently developed state-of-the-art blob detection methods in [18], [19].
The rest of paper is organized as follows. In Section
II, we give the problem formulation of the blob detection
and discuss related work. In Section III we describe in
detail our three-step approach consisting of a two-phase
region outlier detection algorithm and a tracking scheme
for identifying and tracking blobs. We then present a realtime blob detection approach by leveraging MPI/OpenMP
parallelization in a large cluster in Section IV. The blob
detection and tracking results and its real time evaluation
are shown in Section V. We conclude the paper, and give
our future plans in Section VI.
2
P ROBLEM D EFINITION AND R ELATED W ORK
In this section, we introduce our problem definition and
discuss previous work related to our study. For related
work, we first discuss existing research work on outlier
detection, and then review previous work on blob detection
in fusion plasma domain.
2.1
Problem Definition
Extracting spatial-temporal features play an important role
in analyzing scientific and engineering applications, including behavior recognition [20], bioinformatics [21], video
analysis c [22], and health informatics [23]. Depending on
the applications, mining spatial features in one time frame
and relationships among spatial objects in and across time
frames are extremely challenging tasks due to three reasons.
First, the extent and shape of a feature could be an important
indicator in determining its influence. However, due to
various data type (regular and irregular), it is not easy
to apply a generic approach for all applications. Second,
IEEE TRANSACTIONS ON BIG DATA, VOL. XX, NO. X, JUNE 2016
3
2.2
Fig. 2: A contour plot of the local normalized density in the
region of interests in one time frame in fusion experiments
or numerical simulations. A cross-section of the torus is
called a poloidal plane. R and Z are cylindrical coordinates
and the major radius of the torus is denoted by R.
effectively incorporating the temporal information in the
overall analysis is a necessity to uncover interesting upcoming events. Finally, how to process very large data sets
in real time demands appropriately responding to extreme
scale computing and big data challenges. In this work, we
attack this problem by presenting a three-step approach
for detecting and tracking spatio-temporal features in the
context of blob-filament detection in fusion plasma.
The definition of a blob is varied in the literature depending on fusion experiments or simulations as well as
available diagnostic information for measurements [8]. This
makes blob detection a challenging task. Figure 2 plots local
normalized density distribution in the regions of interest in
one time frame. We can observe that there are two reddish
spots located at the left portion of the figure, which are
associated with blob-filaments and are significantly different
from their surrounding neighbors. It is clear that a reddish
spot is not a single point but a group of connected points or a
region. Therefore, we formulate the blob detection problem
as a region outlier detection problem. Similar to the spatial
outlier [9], a region outlier is a group of spatial connected
objects whose non-spatial attribute values are significantly
different from those of other spatial surrounding objects in
its spatial neighborhood. Figure 2 shows blobs are region
outliers. The number of region outliers detected is determined by pre-defined criteria provided by domain experts.
The problem is to design an efficient and effective
approach to detect and track different shapes of region
outliers simultaneously in fusion plasma data streams. By
identifying and monitoring these blob-filaments (region
outliers), scientists can gain a better understanding about
this phenomena. In addition, a data stream is an ordered
sequence of data that arrives continuously and has to be
processed online. Due to the high arrival rate of data, the
blob detection must finish processing before the next data
chunk arrives [24]. Therefore, another critical problem is
to develop a high-performance blob detection approach in
order to meet the real-time requirements.
Outlier Detection
The problem of outlier detection has been extensively
studied and can be generally classified into four categories: distance-based, density-based, clustering-based, and
distribution-based approaches [3], [25].
Distance-based methods [26] use a distance metric to
measure the distances among data points. If the number
of data points within a certain distance from the given
point is less than pre-defined threshold, then this point is
determined as an outlier. This approach could be very useful
with accurate pre-defined threshold. However, it may not
be proper to use a simple threshold if different densities in
various regions of the data exhibit across space or time.
Density-based methods [27] assign a local outlier factor
(LOF) to each sample based on their local density. The LOF
determines the degree of outlierness, where samples with
high LOF value are identified as outliers. This approach
does not require any prior knowledge of underlying distribution of the data. However, it has a high computational
complexity since pair-wise distances have to be computed
to obtain each local density value.
Clustering-based methods [28] conduct clustering-based
techniques on the sample points of the data to characterize
the local data behavior. Since this method does not focus on
outlier detection, the outliers are produced as by-products
and it is not optimized for outlier detection.
Distribution-based methods [9] applies machine learning
techniques to estimate a probability distribution over the
data and develop a statistical test to detect outliers. These
methods use all dimensions to define a neighborhood for
comparison and typically do not distinguish non-spatial
attributes from spatial attributes.
In the context of data streams, a line of research has been
devoted to develop efficient outlier detection techniques
[24], [29], [30], [31], [32], [33]. But their main focus is to solve
the problem of event detection in sensor network [29], query
processing [30], [32], clustering [31], and graph outliers [33].
Therefore, these methods cannot be easily generalized to
region outlier detection problems. In addition, the problem
of blob detection presents a special challenge, because the
spatiotemporal attributes of the blob-filaments has to be
considered together to study their various characteristics
including speed, direction, movement, and size. More importantly, these methods are mostly single-threaded which
cannot cope with real-time requirements in fusion plasma.
A number of distributed outlier detection methods have
also been studied in [29], [34], [35], [36], [37]. Most of
these methods are seeking an efficient way to parallelize
classical outlier detection methods such as distance-based
outliers [35], [37], distribution-based outliers [29], densitybased outliers [35], density-based outliers [36], and PCAbased techniques [34]. However, there methods are not
generally applicable to region outlier detection and tracking.
In particular, in order to tackle high volume and velocity
challenges arising from fusion plasma big data, specialized
outlier detection scheme and suitable high performance
computing technique are demanded to complete blob detection in the order of milliseconds.
In the first two steps of our proposed approach, we
apply distribution-based outlier detection to detect outlier
points by considering only non-spatial attributes and then
IEEE TRANSACTIONS ON BIG DATA, VOL. XX, NO. X, JUNE 2016
leverage fast CCL to construct the region outliers by taking
into account spatial-attributes. We choose distribution-based
outlier detection since it can solve the problem of finding
outliers efficiently if an accurate approximation of a data
distribution can be properly found [9], [29]. Normally the
distribution of the stream data may change over time [5].
However, this assumption may not hold in fusion experiments since a fusion experiment lasts very short time period
from a few seconds to hundreds of seconds. Therefore,
we consider the simpler problem of fixed distribution parameters, noting that several fusion devices have shown
similar distribution functions of blob events. Then we can
perform exploratory data analysis to compute best fitted
distribution parameters offline and then build an accurate
online distribution model. We leave the more complicated
problem of real-time distribution estimation for future work.
2.3
Blob Detection in Fusion Plasma
Independently, fusion blob detection problems have been
researched by the physics community in the context of coherent structures in fusion plasma [8]. Various post-run blob
detection methods have been proposed to identify and track
these structures, to study the impact of the size, movement
and dynamics of blobs. A plasma blob is most commonly
determined by some threshold, computed statistically in the
local plasma density signal [38], [39], [40]. However, the
exact criteria have varied from one experiment to another,
which reflects the intrinsic variability and complexity of the
blob structures. In [38], a conditional averaging approach is
applied to analyze spatio-temporal fluctuation data. When
the vorticity is larger than one standard deviation at some
time frame, a blob is considered to be detected by the probe.
In [39], the conditional averaging technique is also used to
study the evolution of the blob-filaments using Langmuir
probes and a fast camera.
Without using a conditional averaging technique, [40]
searches for blob structures can be done using local measurements of the 2D density data obtained from a 2D probe
array. Identification of a blob is based on the choices of
several constraints such as the threshold intensity level, the
minimum distance of blob movement, and the maximum
allowed blob movement between successive frames. The
trajectories of the different blobs can be computed with
the blob centers based on identification results in each time
frame. The seminal work by Zweben, et. al. [40] was the first
attempt to take only individual time frame data into account
to detect blobs and track their movements, although the
process of identification of a blob was somewhat arbitrary
and oversimplified.
Due to the emergence of fast cameras and beam emission
spectroscopy in the last decade, the situations of insufficient
diagnostic access and limited spatial and temporal resolution have been greatly improved. In the context of computer
version, a number of methods have been developed to tackle
blob detection problem, which is aimed to detect points or
regions in the image that either brighter or darker than the
surrounding [41]. Among them, scale-space methods based
on the Laplacian of Gaussian [42], [43], [44] and Watershed
detection methods based on local extrema in the intensity
landscape [45] are two main classes of blob detectors. In
4
[46], Love and Kumath made the first attempt to apply an
image analysis using Watershed techniques for identifying
blobs in fusion plasma. The images are first processed to
remove the noise spikes, followed by further smoothing
using a Gaussian filter, and then identified by various image
segmentation techniques. However, due to noise and lack of
a ground truth image, this approach can be sensitive to the
setting of parameters, and it is hard to use generic method
for all images. In addition, the output from visualization is
not convenient to feed into other analysis [2]. The regions of
interest computed from this work can be more conveniently
fed into other analyses. For instance, one can compute blobs
in the regions of interest very quickly and transmit these
compact meta information over internet to remote domain
scientists for real-time analysis.
Recently, several researchers [18], [19], [47] have developed a blob-tracking algorithm that uses raw fast camera
data directly with GPI technique. In [18], [19], they leverage a contouring method, database techniques and image
analysis software to track the blob motion and changes
in the structure of blobs. After normalizing each frame
by an average frame created from roughly one thousand
frames around the target time frame, the resulting images
are contoured and the closed contours satisfying certain size
constraints are determined as blobs. Then, an ellipse is fitted
to the contour midway between the smallest level contours
and the peak. All information about blobs are added into
a SQL database for more data analysis. This method is
close to our approach but it can not be used for real-time
blob detection since they compute time-averaged intensity
to normalize the local intensity. Additionally, only closed
contours are treated as blobs, which may miss blobs at
the edges of the regions of interest. Finally, these methods
are still post-run-analysis, which cannot provide real-time
feedback in fusion experiments.
3
O UR PROPOSED APPROACH
Given a fusion data stream, which consists of a time ordered sequence of sample frames that arrive continuously
from fusion experiments or numerical simulations through
remote direct memory access protocols. Our data sets are
simulated electron density from the fusion simulation code
XGC1 [11], [12]. In the present data sets, simulation data is
captured every 2.5 microseconds for a total time window of
2.5 milliseconds. Each point si ∈ S in a time frame t has a
spatial attribute (r, z, t) which defines its location in a triangulated measurement grid, and some non-spatial attributes
including all important plasma quantities such as electron
density ne (r, z, t) as well as connectivity information in a
poloidal plane. The spatial neighborhoods are defined for
each point from the connectivity database in a triangulated
grid. Formally, an region outlier responding to a blob is
defined as a spatial area in the regions of interest where
a subset Bi ⊆ S is a group of connected points si .
Our overall goal is to develop an approach to detect
and track spatial region outliers (blobs) over time using
a stream of fusion data. To achieve this, we break down
the process into three steps: (1) find outlier points in the
region of interests, (2) group these outlier points as different
region outliers, and (3) track these regions outliers by the
IEEE TRANSACTIONS ON BIG DATA, VOL. XX, NO. X, JUNE 2016
5
Density distribution fitting using 50 bins
4
7
Fusion data stream
x 10
7
5
4
3
2
1
Outliers: (si, ne(ri, zi, t))
0
0
Fig. 3: Two-phase region outlier detection for finding blobs
Magnetic Fields in Poloidal Plane
Poloidal Plane
Region of Interests
Reinfed
Original
0.25
1
0.2
0.15
0.1
Z (m)
Z (m)
0.5
0
0.05
0
−0.05
-0.5
−0.1
−0.15
-1
−0.2
-1.5
1
1.2
1.4
1.6
1.8
2
2.2
R (m)
(a) Regions of interest
2.4
−0.25
2.25
2.26
2.27
2.28
2.29
R (m)
2.3
2.31
2.32
(b) Refined mesh
Fig. 4: An example of the regions of interest and the comparison between refined and original triangular mesh vertices
in the R (radial) direction and the Z (poloidal) direction.
overlapping in space. We address the first two steps by
presenting a two-phase approach, as shown in Figure 3.
In the first phase, we apply a distribution-based outlier
detection algorithm to the fusion data stream in order to
detect outlier points which have significantly higher nonspatial attributes than other points. The outputs of this step
are tuples (si , ne (ri , zi , t)), the 2D spatial attributes, and
non-spatial attributes such as electron density. These tuples,
as well as connectivity information, are used as input for the
second phase, where region outlier are detected by applying
a fast CCL [17] to efficiently find different connected components on the triangular mesh. The outputs of the CCL-based
region outlier detection algorithm are a set of connected
components with outlier points inside, which are associated
with blobs if some criteria are satisfied. We address the last
step by proposing an efficient blob tracking algorithm by
leveraging cues from changes of blobs area and distance
of center of blobs. Note that, by varying the first step, this
procedure could be applying to different applications.
In the following section, we describe the proposed twophase region outlier detection in detail.
3.1
4
3
2
4
6
8
Normalized electron density (n_e/n_e0)
10
12
0
0
2
4
6
8
Normalized electron density (n_e/n_e0)
10
12
(b) Log Normal
Fig. 5: An example of exploratory data analysis to analyze
the underlying distribution of the local normalized density
over all poloidal planes and time frames.
Region outliers:
Blobs
1.5
5
1
2
(a) Extreme Value
CCL-based region outlier
detection
x 10
6
Number of points in each bin
Number of points in each bin
6
Distribution-based outlier
detection
Density distribution fitting using 50 bins
4
Distribution-Based Outlier detection
The main task of this phase is to perform efficient outlier
detection to determine outlier points which form the region
outliers associated with blobs. To facilitate this goal, we propose a novel distribution-based outlier detection algorithm
based on the electron density with various criteria for fusion
plasma data streams. We separate spatial attributes from
non-spatial attributes and consider the statistical distribution of the non-spatial attributes to develop a test based on
distribution properties, since it is more suitable for detecting
spatial outliers [9]. As claimed in [29], it is very efficient to
find outliers by using a data distribution approximation if
we estimate the underlying distribution of data accurately.
Values for various criteria are determined by domain experts or subjectively by examining the resulting plotting and
adjusting them until satisfied.
In the proposed outlier detection we firstly preprocess
the sample frame to compute needed quantities in the region
of interests, as shown in Figure 4a. Then it is analyzed
by normalizing the total electron density ne (r, z, t) (which
includes fluctuations) with respect to the initial background
electron density, ne (r, z, 1) (if using real diagnostic data
from, e.g. GPI, actual emission intensity I(r, z, t) would
be used instead of electron density). Note that using the
initial time frame as the benchmark is an important factor
to achieve real-time blob detection. The normalized electron density in the subsequent time frames can be easily
computed, especially compared to the time-average electron
density with a long time interval [19].
Algorithm 1 Triangular mesh refinement algorithm
1:
2:
3:
4:
5:
Input/output:
triGrid: connectivity array of the triangular mesh
(r, z): spatial coordinate of each point
ne : normalized electron density of each point
Compute unique edges E and indices vector IE by
sorting and removing duplicates based on triGrid
Compute spatial coordinate of each new vertices in the
middle of E based on (r, z)
Compute electron density of each new vertices on E by
performing linear interpolation based on ne
Compute indices for each new vertices by adding vector
index IE with the number of original points
Compute a new triangular mesh by assigning appropriate indices from each new and old vertices
To obtain meaningful region outliers using the CCL
method, it is necessary to have fine grained connectivity
information. This particular simulation mesh has coarse
vertical resolution, so resolution enhancement techniques
are applied to generate a higher resolution triangular mesh
based on the original triangulated mesh. As shown in Algorithm 1, the resulting triangular mesh is refined to achieve
IEEE TRANSACTIONS ON BIG DATA, VOL. XX, NO. X, JUNE 2016
6
four times better granularity. We split each original triangle
into four smaller ones by linking three middle points of
the original mesh edges in each triangle. The corresponding
density of generated vertices can be obtained using linear
interpolation of the original triangular mesh. This step can
be applied recursively until the satisfactory resolution of the
triangular mesh is computed. Figure 4b shows the resulting
triangular mesh vertices after applying the triangular mesh
refinement algorithm once.
In order to apply an appropriate predefined quantile in
two-phase distribution-based outlier detection, it is advised
to perform exploratory data analysis to exploit main characteristics of the data sets. Figure 5 reveals that extreme value
distribution and log normal distribution are fitted best with
one of our sample data sets (after comparing over sixteen
different common distributions). After analyzing the underlying distribution, a novel outlier detection is performed to
determine outlier points in the regions of interest. The basic
idea of the proposed two-step outlier detection is motivated
from the observations that there are relatively high density
areas (a half banded ellipse area with cyan color) in the edge
and several significantly high density small regions (a few
small areas with reddish yellow color) in these relatively
high density areas, as shown in Figure 2. The proposed
outlier detection method extends the previous approach
that applies statistical detection with conditional averaging
intensity value [38], [39], and applies more intelligent outlier
detection with only considering individual time frame data.
Compared to traditional single threshold segmentation approach, our approach is more generic, flexible and easier to
tune a satisfactory result.
In the first step, the standard deviation σ and the expected value µ are computed over all sixteen poloidal planes
in one time frame. Using the best fitted distribution, we
apply first step outlier detection to identify the relative high
density areas with a specified predefined quantile:
N (ri , zi , t) − µ > α ∗ σ, ∀(ri , zi ) ∈ Γ
(1)
where N is the normalized electron density, α is the multiple
of σ associated to the specified predefined quantile and Γ
is the domain in the region of interests. Once the relative
high density regions are determined, we compute another
standard deviation σ2 and the expected value µ2 in these
areas. Then we employ second step outlier detection to
identify the outlier points in the relative high density areas
with an appropriately chosen predefined quantile:
N (ri , zi , t) − µ2 > β ∗ σ2 , ∀(ri , zi ) ∈ Γ2
(2)
where β is the multiple of σ2 associated to the judiciously
chosen confidence level and Γ2 is the domain of blob candidates. In practice, α and β can be chosen to be same or
different, depending on the characteristics of blob-filaments.
In our experience, the α value is generally greater than β
since the standard deviation σ over the region of interests
is much smaller than the standard deviation σ2 from the
relative high density areas.
However, two-step outlier detection alone cannot be
used to distinguish the blob candidates since identified blob
candidates may actually have small density, which does not
satisfy traditional definition of blobs. Therefore, the density
of the mesh points in the outlier points smaller than a
certain minimum absolute value criterion need to be filtered
out. On the other hand, it is also possible that the middle
areas between surrounding plasmas and outlier points have
density higher than the given minimum absolute value
criterion. Thus, we also apply a minimum relative value
criterion to remove these unwanted points. To combine
these two rules together, we have a more robust and flexible
criterion:
N (ri , zi , t) > max(dma , (dmr ∗ µ2 )), ∀(ri , zi ) ∈ Γ3
(3)
where dma and dmr are minimum absolute value and minimum relative value respectively, and Γ3 is the domain of
good blob candidates.
3.2
CCL-Based Region Outlier Detection
The main task of the second phase is to apply an efficient
connected component labeling algorithm adopted from [17]
on a refined triangular mesh to find different blob candidate components. A connected component labeling algorithm generally considers the problem of labeling binary 2D
images with either 4-connectedness or 8-connectedness. It
performs an efficient scanning technique, and fills the label
array labels so that the neighboring object pixels have the
same label. Wu [17] presents an efficient two-pass labeling algorithm that is much faster than other state-of-theart methods and theoretically optimal. However, since we
process a refined triangular mesh rather than the traditional
2D images, we have modified their CCL algorithm to take
the special features of a triangular mesh into account. As
shown in Algorithm 2, each triangle is scanned first rather
than a point. Since we know the three vertices in a triangle
are connected, we can reduce unnecessary memory accesses
once any vertex in a triangle is found to be connected with
another vertex in a different triangle. Then we compute the
current minimum parent label in this triangle, and assign
each vertex a parent label if its label has already filled or a
label if its label has not initialized yet. If all three vertices
in a triangle are scanned for the first time, then a new
label number is issued and assigned to their labels and the
associated parent label. After the label array is filled full,
we need flatten the union and find tree. Finally, a second
pass is performed to correct labels in the label array, and all
blob candidates components are found. Note that to perform
efficient union-find operations, the union-find data structure
is implemented with a single array as suggested in [17].
After all blob candidates are determined, a blob is
claimed to be found if the median of a blob candidate
component satisfies a certain minimum absolute median
value criterion. The reason we are setting this constraint to
select the blobs is that the minimum value criterion has to
be a reasonably small value in order to produce more blob
candidate components. It is possible that if the minimum
absolute median value criterion is too large, it may also
remove the blobs. On the other hand, it is also possible if this
value is too small, it does not have effect on filtering out unwanted components. Therefore, with the same philosophy
of measurement, a minimum relative median value criterion
is also applied to determine the blobs. However, in order to
protect the blobs from being removed due to the extremely
large mean value µ2 , we also set the maximum absolute
IEEE TRANSACTIONS ON BIG DATA, VOL. XX, NO. X, JUNE 2016
7
Algorithm 2 Connected component labeling algorithm on
triangular mesh to find various blob candidates components
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
Input:
triGrid: connectivity array of the triangular mesh
Output:
Bc : Region structure of each blob candidate
Initialize label, parentLabel, and labnum
for Scanning each triangle until the end of triGrid do
if label of three vertices are all zero then
Assign a new labnum to all three vertices
Update label and parentLabel with labnum
else
Find the minimum parentLabel of all three vertices
Update their label and parentLabel with this value
end if
end for
for Scanning until the end of parentLabel do
Update parentLabel by flattening union-find tree
end for
for Scanning until the end of Label do
Update Label with latest parentLabel
end for
Find each Bc of points with same parentLabel
median value criterion to limit the power of minimum
relative median value criterion. We unify these three rules
to be one:
N (ri , zi , t) > max(dˆma , min((dˆmr ∗ µ2 ), dˆxa )),
∀(ri , zi ) ∈ Γ4
(4)
where dˆma , dˆmr and dˆxa are minimum absolute and relative median values and maximum absolute median value
respectively and Γ4 is the domain of blobs.
3.3
Tracking Region Outliers
The objective of the third step is to track the direction and
speed of the detected blobs over time. The blob tracking
algorithm has to cope with the problem of tracking multiple
region outliers simultaneously even when the blobs merge
together or split into separated ones. On the other hand, the
blob tracking method should be simple and efficient to meet
real-time requirements. To achieve this goal, we propose an
efficient blob tracking algorithm by leveraging cues from
changes of blobs area and distance of center of blobs. We
compute the correspondence between previously tracked
blobs and currently detected blobs, and then recover the
trajectories of the tracked blobs.
To identify the location center of detected blob, we compute the density-weighted average of the spatial coordinates
of each point inside a blob.
(rc , zc ) =
n
1 X
(r, z)ne
M i=1
(5)
where M is summation of ne of all points in a blob. The
density-weighted average is used to better capture the center of density of a blob. We track the movement of these
detected blobs by linking the centers in consecutive time
frames. In order to obtain the boundary of region outliers
(blobs), we compute the convex hull [48] of a set of points
in a blob. The area of a blob is computed by counting the
number of points in a blob.
Algorithm 3 Efficient blob tracking algorithm
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
Input:
B : Current detected blobs
T : Previous blob tracks
Output:
T : Updated blob tracks with B appended
Initialize hull, cen, and area
hull = getBoundary(B )
cen = getCenter(B )
area = getArea(B )
for Scanning until the end of B do
cenDis = getCenterDis(B ,T )
areaDif = getAreaDif(B ,T )
if cenDis ≤ maxJump ∧ areaDif ≤ maxDif then
Find a blob track T with smallest cenDis
Append current blob into this blob track T
end if
end for
Update T with hull, cen, area, and computed speed
As shown in Algorithm 3, the input parameters are
current detected blobs and the previous blob tracks. The
data structure of a blob track is composed of the track ID,
the length of track, the area of previous blob, the timestamps, the center points, the boundary points, and the
velocity. There are two heuristics to verify whether a blob
is associated with an existing blob track. The first heuristic
is based on the fact that the area of a blob between consecutive time frames cannot decrease or increase significantly.
The second heuristic takes into account the distance of the
centers of a blob does not change dramatically over very
short time period (microseconds). The proper thresholds
for these two heuristics are provided by domain experts.
Since blobs can appear, disappear, merge together or split,
a greedy scheme is applied to find the best matching pair
of blob and track based on closest distance of the centers
of current detected blob and the latest blob in a blob track.
Based on computed correspondence between a blob track
and the currently detected blobs, existing blob tracks are
automatically processed through corresponding operations
such as adding a blob into a track, creating a new track,
and a track ending. If the length of a track is smaller than
3 consecutive time frames, the track will be treated an
anomaly and deleted due to errors in data or inappropriate
blob detection thresholds. The speed and direction of the
blobs can thus be computed from two consecutive center
points. Finally, we can recover the trajectories of the tracked
blobs by monitoring the movement of blob centers.
4
A REAL - TIME BLOB DETECTION APPROACH
Existing blob detection approaches cannot tackle the two
challenges of the large amount of data produced in a shot
and the real-time requirement. In addition, existing data
analysis approaches are often operated in a single thread,
only for post-run analysis and often take a few hours to generate the results [49]. In order to meet the real-time feedback
IEEE TRANSACTIONS ON BIG DATA, VOL. XX, NO. X, JUNE 2016
8
Total number 24 of MPI × OpenMP
Simulation or experiment data
8
MPI
Blob Det
Blob Det + Comm
7
Pi
OpenMP
T1
T2
…
Tm
Pn
OpenMP
T1
T2
…
Tm
Normalized Runtime
P0
OpenMP
T1
T2
…
Tm
Fig. 6: Hybrid MPI/OpenMP parallelization
requirement, we address these challenges by developing
a high performance blob detection approach, which can
leverage in situ raw data and find blob-filaments efficiently
in fusion experiments or numerical simulations.
4.1
A hybrid MPI/OpenMP parallelization
The key idea is to exploit many cores in a large cluster
system by running MPI to allocate n processes to process
the data in one or several time frames at the high level,
and by leveraging OpenMP to accelerate the computations
using m threads at the low level. Figure 6 shows our hybrid
MPI/OpenMP parallelization for blob detection. Using this
approach, we can complete our blob detection in a few
milliseconds with in situ evaluation.
In order to achieve blob detection in real time, the goal
is to minimize data movements in the memory and speed
up computation. Ideally, the performance is optimal without
any communication if we can perform the job correctly. The
proposed blob detection algorithm in the previous section
supports embarrassingly parallel since we only need the
initial time frame and the target time frame to do the computation. This is an important difference between our blob
detection method and recently developed methods [18], [19]
in terms of real-time requirement. Furthermore, we explore
many-core processor architectures to speed up the computation of each MPI task by taking full advantage of multithreading in the shared memory. Therefore, our real-time
blob detection approach based on hybrid MPI/OpenMP
parallelization is a natural choice and is expected to provide
the optimal performance for fusion plasma data streams.
A practical interesting issue is how to tune the number
of MPI processes and OpenMP threads for the best performance by taking both analysis speed and memory size into
account. As shown in Figure 7, we vary the number of MPI
processes and OpenMP threads but fix the total number to
be 24 for investigating the performance when processing the
same amount of time frames data. A faster analysis speed
is achieved when increasing the number of MPI processes
since more data frames can be processed simultaneously. On
the other hand, the analysis speed remains constant with
a few OpenMP threads and degrades with more OpenMP
threads due to lack of enough computation in one time
frame. However, more OpenMP threads could significantly
reduce the memory demands. Therefore, in this study, we
choose the number of OpenMP threads to be four for each
MPI task, to achieve a good trade off between analysis speed
and memory savings.
6
5
4
3
2
1
0
0
5
10
15
Number of MPI
20
25
Fig. 7: Investigate the performance of hybrid MPI/OpenMP
parallelization when varying number of MPI processes and
OpenMP threads. The blue triangle denotes only normalized blob detection time. The red star denotes the normalized total time including both blob detection time and initial
communication time for broadcasting the first time frame to
all analysis nodes for normalization.
4.2
Outline of the implementation
We implement our blob detection algorithm in C with a
hybrid MPI/OpenMP parallelization. Algorithm 4 summarizes the proposed blob detection algorithm without considering OpenMP. Users can specify the regions of interest
by (Rmin, Rmax, Zmin, Zmax), the range of target time
frames by (t start, t end), and the location of the data sets.
However, with in situ evaluation, there is no need to specify
the file location since all data are already in memory. We
use static scheduling to evenly divide the number of time
frames for each MPI task for efficiency. The n MPI processes
are allocated to process one or several time frames and m
OpenMP threads are launched to accelerate the computation
in one time frame. Note that the MPI process is also the
master thread in the runtime environment. At the beginning, the initial time frame data is broadcasted to all MPI
processes so that normalization can be performed with new
coming time frames. Then each MPI process embarrassingly
process the data in each time frame with multithreading
in the shared memory. Only detected blobs information are
maintained and added into local database. Since these local
blobs information are very compact, they can be efficiently
transmitted over internet to remote servers for real-time
analysis by domain scientists.
5
E XPERIMENTS AND R ESULTS
In this section we present experimental evaluations of our
blob detection and tracking algorithms, and report the performance of the real-time blob detection under both strong
and weak scaling. Before showing experimental results in
the next section, we briefly introduce our experimental
environment, data sets, and parameter setting in our blob
detection and tracking algorithms. We have tested our implementation on the NERSC’s newest supercomputer Edison, where each compute node has two Intel “Ivy Bridge”
processors (2.4GHz with 12 cores) and 64 GB of memory.
IEEE TRANSACTIONS ON BIG DATA, VOL. XX, NO. X, JUNE 2016
Algorithm 4 A real-time blob detection approach
1:
2:
3:
4:
5:
6:
7:
8:
9:
Input:
Rmin, Rmax, Zmin, Zmax: specify region of interest
t start, t end : start and end time frames
FileDir: location where data sets locate
Output:
B : Detected region outliers (blobs)
Apply static scheduling to assign equal amount of n
time frames data to each MPI process
Broadcast the initial time frame to all MPI processes
for t = 1 : n do
Process i loads raw data in one frame and computes
normalized density ne (r, z, t) in region of interest
Refine the triangular mesh. See Algorithm 1
Apply two-phase distribution-based outlier detection
to identify outliers with various criteria
Apply CCL-based region outlier detection on a triangular mesh to find blob components. See Algorithm 2
A blob is added into B if certain criteria is satisfied
end for
Our base data sets are simulation data sets with 1024 time
frames based on the XGC1 simulation [11] [12] from the
Princeton Plasma Physics Laboratory, which last around
2.5 milliseconds. One of our main goals is that we can
control analysis speed by varying the number of processes
to complete the blob detection on the entire data set in a
time close to 2.5 milliseconds. It would indicate that our
algorithm could monitor fusion experiments in real time
(neglecting data transfer latency). If we consider internet
transfer latency in real experiments or numerical simulation,
the system needs at least 1 to 25 milliseconds to transfer one
time frame data depending on size of data, which may give
us more time for data analysis.
Another goal is to validate the effectiveness of the proposed algorithms. In Algorithm 4, we apply various criteria
to identify the blobs. The parameters for blob detection
and tracking in our experiments are given in Table 1. One
criterion we have not mentioned in the previous section is
parameter “minArea”. This parameter is used to decide how
many points a blob should have, which is used to remove
impossibly small blobs. In our experiment, this parameter is
set to three since there are at least three vertices connected
as a 2D component in a triangular mesh. Another criteria
are parameters “maxFrames” and “minFrames”, which are
used to control the length of a blob track and remove noisy
tracks. It is important to note that these parameters need
to be tuned in order to achieve optimal performance in
different fusion experiments or numerical simulations. The
reasons for this uncertainty in the context of blob detection
are from the intrinsic variability and complexity of the blob
structures observed in different experiments [8].
5.1
Performance comparison
We first conduct experiments to compare our method with
recently developed state-of-the-art blob detection methods
in [18], [19]. Since their methods are based on the contouring methods and thresholding, we call their methods
the contouring-based methods. We have to point out that
9
TABLE 1: Parameters setting for the proposed blob detection
and tracking algorithms on XGC1 simulation data sets
detection criteria
minArea
3
minRden (dma )
1.2
minAbsden (dmr ) 2.05
maxAbsMden (dˆxa ) 2.75
minMden (dˆma )
1.3
minAbsMden (dˆmr ) 2.15
tracking criteria
maxAreaChange
25
maxJump 0.04
maxFrames
100
minFrames
3
strictly quantitative comparisons are not possible since the
blob itself is not well-defined [8]. Due to this reason, there
are rarely direct comparisons between any new proposed
method and existing ones in the literature in the domain
of fusion plasma [18], [19], [38], [39], [40], [46], [47], [49].
However, in order to demonstrate that our methods is more
robust than the contouring-based methods, we compare
these two methods in two typical cases to shed light on their
performance in terms of the detection accuracy.
Figure 8 shows the comparison of the blob detection
results between our region outlier detection method and the
contouring-based methods in two different time frames. As
shown in Figures 8a and 8b, we can see that our region
outlier detection method does not miss detecting the blob
at the edge of the regions of interest while the contouringbased methods fail the detection. The reason is that the
contouring-based methods require the computed contours
are closed, which do not exist at the edge of the regions of interest. In Figures 8c and 8d, we notice that our region outlier
detection method can accurately detect all blobs. However,
the contouring-based methods either yield the blobs with
incorrect areas (much larger or smaller), or misdetect the
wrong area as a blob. This is because that it is hard to
use one single threshold to identify the blobs for various
time frames even in the same experimental data. Our region
outlier detection method does not have such problem since
we use more flexible distribution-based outlier detection.
5.2
More blob detection results
We perform more experiments to comprehensively examine
the blob detection results in five continuous time frames and
four different poloidal planes as shown in Figure 9. As we
can see from the figure, our region outlier detection method
can provide consistently good results in different situations.
In addition, our method does not miss any blobs at the edge
of the regions of interest, as shown in subfigures 9b, 9g, 9c
and 9h. It is interesting to see that large-scale blob structures
are often generated, which could cause substantial plasma
transport [40]. As pointed out in [50], these large-scale
structures are mainly contributed by the low-frequency and
long-wavelength fluctuating components, which may be
responsible for the observations of long-range correlations.
We also noticed that different poloidal planes may display
significant diversity in edge turbulence, even in the same
time frame. We have shown that we are able to effectively
detect the blobs and reveal some interesting results to help
physicists improve their understanding of the characteristic
of blobs and their correlation with other plasma properties.
IEEE TRANSACTIONS ON BIG DATA, VOL. XX, NO. X, JUNE 2016
10
Blob Detection: time frame 45 and Poloidal plane 1
4
0.25
3.5
0.2
3
0.15
Z value
0.1
2.5
0.05
2
0
1.5
−0.05
−0.1
1
−0.15
0.5
−0.2
−0.25
2.25
2.26
2.27
2.28
R value
2.29
2.3
0
2.31
(a) Contouring-based methods
(b) Region outlier detection method
Blob Detection: time frame 87 and Poloidal plane 1
4
0.25
3.5
0.2
3
0.15
Z value
0.1
2.5
0.05
2
0
1.5
−0.05
−0.1
1
−0.15
0.5
−0.2
−0.25
2.25
2.26
2.27
2.28
R value
2.29
2.3
2.31
0
(c) Contouring-based methods
(d) Region outlier detection method
Fig. 8: Two examples of comparing our region outlier detection method with the Contouring-based methods in the R
(radial) direction and the Z (poloidal) direction. The separatrix position is shown by a white line and the different pink and
blue circles denote blobs.
5.3
Blob tracking results
We investigate the blob tracking results in two different
situations. Figure 10a exhibits a 2D trajectory of a blob.
Again, the trajectory is generated by plotting the location of
the density peak of the detected blobs over five consecutive
time frames. We can see that our blob tracking algorithm
can track two separate blobs simultaneously. The blob size
can grow when they move towards confined plasma in the
right region of separatrix. Figure 10b shows a 3D trajectory
for a detected blob over fifteen consecutive time frames.
In this case, the blob seems to maintain its size for a few
time frames, then gradually decreases, and eventually disappears. Through these interesting results, physicists may
be able to understand the characteristics of blobs better.
5.4
Real-time blob detection under strong scaling
We have illustrated the effectiveness and robustness of
the proposed blob detection and tracking methods. Next,
we perform a set of experiments to demonstrate the performance of our real-time blob detection approach under
strong scaling and weak scaling. We define the speedup
of our parallel implementation on heterogeneous multi-core
architecture as follows:
runtime of Blob detection using single core
speedup =
runtime using P cores
Our most encouraging results are that we can complete
blob detection on the simulation data set described above in
around 2 milliseconds with MPI/OpenMP using 4096 cores
and in 3 milliseconds with MPI using 1024 cores. In Figure
11, we can achieve linear time speedup in blob detection
time under strong scaling. The MPI and the MPI/OpenMP
implementations accomplish 800 and 1200 times speedup
respectively, when the number of processes is scaled to 1024.
Also, we can see that the hybrid MPI/OpenMP implementation is about two times faster than the MPI implementation
when varying the number of processes from 1 to 512. With
1024 processes, both of them achieve similar performance,
but the MPI/OpenMP one is slightly better. This demonstrates that we are able to control analysis speed by varying
the number of processes to achieve real-time analysis.
5.5
Real-time blob detection under weak scaling
In this experiment, we evaluate the performance of our
real-time blob detection under weak scaling. We replicate
existing data sets (30GB) in order to obtain adequate experimental data sets (4.3TB). The basic unit data contains
IEEE TRANSACTIONS ON BIG DATA, VOL. XX, NO. X, JUNE 2016
11
(a) Frame 82 and plane 1 (b) Frame 83 and plane 1 (c) Frame 84 and plane 1 (d) Frame 85 and plane 1 (e) Frame 86 and plane 1
(f) Frame 82 and plane 2 (g) Frame 83 and plane 2 (h) Frame 84 and plane 2 (i) Frame 85 and plane 2 (j) Frame 86 and plane 2
(m) Frame 84 and plane
(k) Frame 82 and plane 3 (l) Frame 83 and plane 3 3
(n) Frame 85 and plane 3 (o) Frame 86 and plane 3
(p) Frame 82 and plane 4 (q) Frame 83 and plane 4 (r) Frame 84 and plane 4 (s) Frame 85 and plane 4 (t) Frame 86 and plane 4
Fig. 9: An example of the blob detection in five continuous time frames and four different poloidal planes in the R (radial)
direction and the Z (poloidal) direction. The separatrix position is shown by a white line and the different blue circles
denote blobs.
Trajectory of region outlier (blobs)
158
156
Time frame
154
152
150
148
146
144
0.2
0.1
2.3
0
2.28
−0.1
Z value
(a) 2D trajectory for detected blobs
−0.2
2.26
R value
(b) 3D trajectory for detected blobs
Fig. 10: 2D and 3D center trajectories for detected blobs over consecutive time frames. The red solid polygon indicates the
starting times of the blobs tracked while the blue broken polygons indicate subsequent times of the same blobs tracked. The
centers of the moving blobs are linked to show their trajectories of the blob motion. The pink line represents the separatrix.
128 time frames and the size of data increases linearly with
the number of processes. In Figure 12, the blob detection
time remains almost constant under weak scaling, which
indicates that our implementations scale very well to solve
much larger problems. Also, both MPI and MPI/OpenMP
implementations achieve high parallel efficiency as the
IEEE TRANSACTIONS ON BIG DATA, VOL. XX, NO. X, JUNE 2016
12
Blob Detection Runtime: Strong Scaling
1
10
4
MPI Speedup
MPI/OpenMP Speedup
MPI Runtime
MPI/OpenMP Runtime
Speedup over sequatial
0
10
Runtime (Second)
Blob Detection Speedup: Strong Scaling
10
−1
10
−2
10
2
10
1
10
0
−3
10
3
10
0
10
1
10
2
3
10
10
Number of processes
10 0
10
4
10
1
10
(a) Time
2
3
10
10
Number of processes
4
10
(b) Speedup
Fig. 11: Blob detection time and speedup with MPI and MPI/OpenMP varying number of processes under strong scaling
Blob Detection Runtime: Weak Scaling
0.5
Blob Detection Efficiency: Weak Scaling
1
MPI Runtime
MPI/OpenMP Runtime
0.8
Parallel Efficiency
Runtime (Second)
0.4
0.3
0.2
0.1
0.6
0.4
0.2
MPI Efficiency
MPI/OpenMP Efficiency
0 0
10
1
10
2
3
10
10
Number of processes
4
10
(a) Time
0 0
10
1
10
2
3
10
10
Number of processes
4
10
(b) Speedup
Fig. 12: Blob detection time and speedup with MPI and MPI/OpenMP varying number of processes under weak scaling
number of processes increases from 1 to 1024.
6
C ONCLUSION AND FUTURE WORK
Near real-time extraction of spatio-temporal features in very
large-scale irregular data presents both opportunities and
challenges responding to extreme scale computing and big
data in many applications. In this paper, we propose, for the
first time, a real-time blob detection and tracking approach
for finding blob-filaments in fusion experiments or numerical simulations. The key idea of the proposed approach is
to break down the overall process into three steps. The first
two steps are based on a distribution-based outlier detection
scheme with various criteria and a fast CCL method to find
blob components. In the third step, an efficient blob tracking
scheme is presented to recover the trajectories of the motions of blobs. Our hybrid MPI/OpenMP implementations
demonstrate the effectiveness and efficiency of the proposed
approach with a set of fusion plasma simulation data. Our
tests show that we can achieve linear time speedup and
complete blob detection in two or three milliseconds using
a cluster at NERSC.
We are currently working on integrating our blob detection algorithm into the ICEE system for consuming fusion
plasma data streams where the blob detection function is
used in a central data analysis component and the resulting
detection results are monitored and controlled from portable
devices, such as an iPad. We plan to test the proposed
method in both simulations and real fusion experiments.
ACKNOWLEDGMENTS
The authors would like to thank the referees for their
valuable comments. This work was supported by the Office of Advanced Scientific Computing Research, Office of
Science, of the U.S. Department of Energy under Contract
No. DE-AC02-05CH11231 and partially supported by NSF
under grants No. CCF 1218349 and ACI SI2-SSE 1440700,
and by DOE under a grant No. DE-FC02-12ER41890. The
blobby plasma turbulence simulation was performed using
resources of the Oak Ridge Leadership Computing Facility,
which is a DOE Office of Science User Facility supported
under Contract DE-AC05-00OR22.
IEEE TRANSACTIONS ON BIG DATA, VOL. XX, NO. X, JUNE 2016
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
N. Cressie and C. K. Wikle, Statistics for Spatio-Temporal Data.
Wiley, 2011.
K. Wu, R. R. Sinha, C. Jones, S. Ethier, S. Klasky, K.-L. Ma,
A. Shoshani, and M. Winslett, “Finding regions of interest on
toroidal meshes,” Computational Science & Discovery, vol. 4, no. 1,
p. 015003, 2011.
V. Chandola, A. Banerjee, and V. Kumar, “Anomaly detection: A
survey,” ACM computing surveys (CSUR), vol. 41, no. 3, p. 15, 2009.
J. Han, M. Kamber, and J. Pei, Data Mining: Concepts and Techniques,
3rd ed. San Francisco, CA, USA: Morgan Kaufmann Publishers
Inc., 2011.
M. Gupta, J. Gao, C. C. Aggarwal, and J. Han, “Outlier detection
for temporal data: A survey,” IEEE Transactions on Knowledge and
Data Engineering, vol. 26, no. 9, pp. 1–1, 2014.
H. Yang, S. Parthasarathy, and S. Mehta, “A generalized framework for mining spatio-temporal patterns in scientific data,” in
Proceedings of the eleventh ACM SIGKDD international conference on
Knowledge discovery in data mining. ACM, 2005, pp. 716–721.
R. Aymar, P. Barabaschi, and Y. Shimomura, “The iter design,”
Plasma Physics and Controlled Fusion, vol. 44, no. 5, p. 519, 2002.
D. DIppolito, J. Myra, and S. Zweben, “Convective transport by intermittent blob-filaments: Comparison of theory and experiment,”
Physics of Plasmas (1994-present), vol. 18, no. 6, p. 060501, 2011.
S. Shekhar, C.-T. Lu, and P. Zhang, “A unified approach to detecting spatial outliers,” GeoInformatica, vol. 7, no. 2, pp. 139–166,
2003.
K. Wu, S. Ahern, E. W. Bethel, J. Chen, H. Childs, E. CormierMichel, C. Geddes, J. Gu, H. Hagen, B. Hamann et al., “Fastbit:
interactively searching massive data,” in Journal of Physics: Conference Series, vol. 180, no. 1. IOP Publishing, 2009, p. 012053.
C. Chang, S. Ku, P. Diamond, Z. Lin, S. Parker, T. Hahm, and
N. Samatova, “Compressed ion temperature gradient turbulence
in diverted tokamak edgea),” Physics of Plasmas (1994-present),
vol. 16, no. 5, p. 056108, 2009.
S. Ku, C. Chang, and P. Diamond, “Full-f gyrokinetic particle
simulation of centrally heated global itg turbulence from magnetic
axis to edge pedestal top in a realistic tokamak geometry,” Nuclear
Fusion, vol. 49, no. 11, p. 115021, 2009.
B. Dong, S. Byna, and K. Wu, “Expediting scientific data analysis
with reorganization of data,” in Cluster Computing (CLUSTER),
2013 IEEE International Conference on. IEEE, Sept 2013, pp. 1–8.
——, “Sds: A framework for scientific data services,” in Proceedings
of the 8th Parallel Data Storage Workshop, ser. PDSW ’13. ACM,
2013, pp. 27–32.
J. Y. Choi, K. Wu, J. C. Wu, A. Sim, Q. G. Liu, M. Wolf, C. Chang,
and S. Klasky, “Icee: Wide-area in transit data processing framework for near real-time scientific applications,” in 4th SC Workshop
on Petascale (Big) Data Analytics: Challenges and Opportunities in
conjunction with SC13, 2013.
J. F. Lofstead, S. Klasky, K. Schwan, N. Podhorszki, and C. Jin,
“Flexible io and integration for scientific codes through the adaptable io system (adios),” in Proceedings of the 6th international workshop on Challenges of large applications in distributed environments.
ACM, 2008, pp. 15–24.
K. Wu, E. Otoo, and K. Suzuki, “Optimizing two-pass connectedcomponent labeling algorithms,” Pattern Analysis and Applications,
vol. 12, no. 2, pp. 117–135, 2009.
W. Davis, M. Ko, R. Maqueda, A. Roquemore, F. Scotti, and
S. Zweben, “Fast 2-d camera control, data acquisition, and
database techniques for edge studies on nstx,” Fusion Engineering
and Design, vol. 89, no. 5, pp. 717–720, 2014.
J. Myra, W. Davis, D. D’Ippolito, B. LaBombard, D. Russell, J. Terry,
and S. Zweben, “Edge sheared flows and the dynamics of blobfilaments,” Nuclear Fusion, vol. 53, no. 7, p. 073013, 2013.
P. Dollár, V. Rabaud, G. Cottrell, and S. Belongie, “Behavior recognition via sparse spatio-temporal features,” in 2005 IEEE International Workshop on Visual Surveillance and Performance Evaluation of
Tracking and Surveillance. IEEE, 2005, pp. 65–72.
J. Winkler, A. Seybert, L. König, S. Pruggnaller, U. Haselmann,
V. Sourjik, M. Weiss, A. S. Frangakis, A. Mogk, and B. Bukau,
“Quantitative and spatio-temporal features of protein aggregation
in escherichia coli and consequences on protein quality control
and cellular ageing,” The EMBO journal, vol. 29, no. 5, pp. 910–
923, 2010.
13
[22] Q. V. Le, W. Y. Zou, S. Y. Yeung, and A. Y. Ng, “Learning hierarchical invariant spatio-temporal features for action recognition with
independent subspace analysis,” in Computer Vision and Pattern
Recognition (CVPR), 2011 IEEE Conference on.
IEEE, 2011, pp.
3361–3368.
[23] K. Aminian, B. Najafi, C. Büla, P.-F. Leyvraz, and P. Robert,
“Spatio-temporal parameters of gait measured by an ambulatory system using miniature gyroscopes,” Journal of biomechanics,
vol. 35, no. 5, pp. 689–699, 2002.
[24] S. Sadik and L. Gruenwald, “Research issues in outlier detection
for data streams,” ACM SIGKDD Explorations Newsletter, vol. 15,
no. 1, pp. 33–40, 2014.
[25] V. J. Hodge and J. Austin, “A survey of outlier detection methodologies,” Artificial Intelligence Review, vol. 22, no. 2, pp. 85–126,
2004.
[26] E. M. Knox and R. T. Ng, “Algorithms for mining distancebased
outliers in large datasets,” in Proceedings of the International Conference on Very Large Data Bases. Citeseer, 1998, pp. 392–403.
[27] M. M. Breunig, H.-P. Kriegel, R. T. Ng, and J. Sander, “Lof:
identifying density-based local outliers,” in ACM sigmod record,
vol. 29, no. 2. ACM, 2000, pp. 93–104.
[28] Z. He, X. Xu, and S. Deng, “Discovering cluster-based local outliers,” Pattern Recognition Letters, vol. 24, no. 9, pp. 1641–1650, 2003.
[29] S. Subramaniam, T. Palpanas, D. Papadopoulos, V. Kalogeraki,
and D. Gunopulos, “Online outlier detection in sensor data using
non-parametric models,” in Proceedings of the 32nd international
conference on Very large data bases. VLDB Endowment, 2006, pp.
187–198.
[30] D. Pokrajac, A. Lazarevic, and L. J. Latecki, “Incremental local
outlier detection for data streams,” in Computational Intelligence
and Data Mining, 2007. CIDM 2007. IEEE Symposium on. IEEE,
2007, pp. 504–515.
[31] M. Elahi, K. Li, W. Nisar, X. Lv, and H. Wang, “Efficient clusteringbased outlier detection algorithm for dynamic data stream,” in
Fuzzy Systems and Knowledge Discovery, 2008. FSKD’08. Fifth International Conference on, vol. 5. IEEE, 2008, pp. 298–304.
[32] F. Angiulli and F. Fassetti, “Detecting distance-based outliers in
streams of data,” in Proceedings of the sixteenth ACM conference on
Conference on information and knowledge management. ACM, 2007,
pp. 811–820.
[33] C. C. Aggarwal, Y. Zhao, and P. S. Yu, “Outlier detection in graph
streams,” in Data Engineering (ICDE), 2011 IEEE 27th International
Conference on. IEEE, 2011, pp. 399–409.
[34] H. Dutta, C. Giannella, K. D. Borne, and H. Kargupta, “Distributed
top-k outlier detection from astronomy catalogs using the demac
system.” in SDM. SIAM, 2007, pp. 473–478.
[35] E. Lozano and E. Acufia, “Parallel algorithms for distance-based
and density-based outliers,” in Data Mining, Fifth IEEE International Conference on. IEEE, 2005, pp. 4–pp.
[36] M. E. Otey, A. Ghoting, and S. Parthasarathy, “Fast distributed
outlier detection in mixed-attribute data sets,” Data Mining and
Knowledge Discovery, vol. 12, no. 2-3, pp. 203–228, 2006.
[37] E. Hung and D. W. Cheung, “Parallel mining of outliers in large
database,” Distributed and Parallel Databases, vol. 12, no. 1, pp. 5–26,
2002.
[38] M. Xu, P. Diamond, G. Tynan, C. Holland, P. Manz, N. Fedorczak,
S. C. Thakur, J. Yu, K. Zhao, J. Dong et al., “Turbulent eddymediated particle, momentum, and vorticity transport in the edge
of hl-2a tokamak plasma,” in 24th IAEA Fusion Energy Conference,
San Diego, 2012.
[39] G. Fuchert, G. Birkenmeier, B. Nold, M. Ramisch, and U. Stroth,
“The influence of plasma edge dynamics on blob properties in the
stellarator tj-k,” Plasma Physics and Controlled Fusion, vol. 55, no. 12,
p. 125002, 2013.
[40] S. Zweben, “Search for coherent structure within tokamak plasma
turbulence,” Physics of Fluids, vol. 28, no. 3, p. 974, 1985.
[41] H. Kong, H. C. Akakin, and S. E. Sarma, “A generalized laplacian
of gaussian filter for blob detection and its applications,” IEEE
transactions on cybernetics, vol. 43, no. 6, pp. 1719–1733, 2013.
[42] I. Laptev and T. Lindeberg, “Space-time interest points,” International Conference on Computer VisionICCV’03, pp. 432–439, 2003.
[43] T. Lindeberg, “Feature detection with automatic scale selection,”
International Journal of Computer Vision, vol. 30, no. 2, pp. 79–116,
1998.
[44] R. T. Collins, “Mean-shift blob tracking through scale space,” in
Computer Vision and Pattern Recognition, 2003. Proceedings. 2003
IEEE TRANSACTIONS ON BIG DATA, VOL. XX, NO. X, JUNE 2016
[45]
[46]
[47]
[48]
[49]
[50]
IEEE Computer Society Conference on, vol. 2. IEEE, 2003, pp. II–
234.
L. Vincent and P. Soille, “Watersheds in digital spaces: an efficient
algorithm based on immersion simulations,” IEEE transactions on
pattern analysis and machine intelligence, vol. 13, no. 6, pp. 583–598,
1991.
N. S. Love and C. Kamath, “Image analysis for the identification
of coherent structures in plasma,” Optical Engineering and Applications, pp. 66 960D–66 960D, 2007.
R. Kube, O. E. Garcia, B. LaBombard, J. Terry, and S. Zweben,
“Blob sizes and velocities in the alcator c-mod scrape-off layer,”
Journal of Nuclear Materials, vol. 438, pp. S505–S508, 2013.
T. M. Chan, “Optimal output-sensitive convex hull algorithms in
two and three dimensions,” Discrete & Computational Geometry,
vol. 16, no. 4, pp. 361–368, 1996.
S. Müller, A. Diallo, A. Fasoli, I. Furno, B. Labit, G. Plyushchev,
M. Podestà, and F. Poli, “Probabilistic analysis of turbulent structures from two-dimensional plasma imaging,” Physics of Plasmas
(1994-present), vol. 13, no. 10, p. 100701, 2006.
G. Xu, B. Wan, W. Zhang, Q. Yang, L. Wang, and Y. Wen, “Multiscale coherent structures in tokamak plasma turbulence,” Physics
of Plasmas (1994-present), vol. 13, no. 10, p. 102509, 2006.
Lingfei Wu is a 6th year Ph.D. candidate in
the computer science department at College
of William and Mary, advised by Dr. Andreas
Stathopoulos. His research interests are in the
areas of high-performance scientific computing,
large-scale machine learning, and big data analytics. In summer 2014 and 2015, Lingfei was
a summer research intern at Lawrence Berkeley National Laboratory and IBM T.J.Watson
Research Center, respectively. Before joining
William and Mary, Lingfei received his M.S. from
University of Science and Technology of China (Hefei, 2010), following
his B.E. from Auhui University (Hefei, 2007)
Kesheng (John) Wu is a Senior Computer
Scientist at Lawrence Berkeley National Laboratory. works actively on a number of topics
in data analysis, data management, and highperformance computing. His recent algorithmic
research work includes bitmap indexing techniques for searching large datasets, statistical
methods for extract features from a variety of
data, and restarting strategies for computing extreme eigenvalues. He authored and coauthored
more than 100 technical publications, nine of
which have more than 100 citations each.
Alex Sim is a Senior Computing Engineer at
Lawrence Berkeley National Laboratory. His current R& D activities focus on data mining and
modeling, data analysis methods, distributed
resource management, and high performance
data systems. He authored and coauthored
more than 100 technical publications, and released a few software packages under open
source license.
14
Michael Churchill received his PhD in Nuclear Science and Engineering from the Massachusetts Institute of Technology, and his Bachelor’s degree in Electrical and Computer Engineering from Brigham Young University. He currently works as a postdoc under Dr. C.S. Chang
at the Princeton Plasma Physics Laboratory, focusing on data management and analysis with
the large-scale XGC plasma simulation codes.
Jong Y. Choi is a researcher working in Scientific Data Group, Computer Science and Mathematics Division, Oak Ridge National Laboratory (ORNL), Oak Ridge, Tennessee, USA. He
earned his Ph.D. degree in Computer Science
at Indiana University Bloomington in 2012 and
his MS degree in Computer Science from New
York University in 2004. His areas of research
interest span data mining and machine learning algorithms, high-performance data-intensive
computing, parallel and distributed systems for
Cloud and Grid computing.
Andreas Stathopoulos received a bachelors degree in Mathematics from University of
Athens, Greece in 1989, and an M.S. and Ph.D.
degrees in Computer Science from Vanderbilt
University, USA in 1991 and 1995 respectively.
In 1995, he obtained an NSF CISE postdoctoral
fellowship to work with Prof. Yousef Saad at the
University of Minnesota, USA. In 1997, he joined
the faculty of Computer Science at the College
of William and Mary, USA, where he currently
holds a full professor position. His research interests span the general area of parallel and high performance scientific
computing. He is an expert on solving eigenvalue problems and has
developed the state-of-the-art eigenvalue software PRIMME.
Choong-Seock Chang is the head of the multiinstitutional multi-disciplinary US SciDAC Center
for Edge Physics Simulation (EPSI), headquartered at Princeton Plasma Physics Laboratory,
Princeton University, awarded by US Department of Energy, Office of Fusion Energy Science and Office of Advanced Scientific Computing Research, jointly. He has over 150 publications in peer reviewed internationally recognized
journals, and has given countless invited talks,
keynote speeches and tutorial lectures at major
scientific conferences.
Scott Klasky is currently a senior research scientist at Oak Ridge National Laboratory, and
head of the end to end task in the scientific
computing group at the National Center Computational Sciences. He has 20 years of experience
in research and development of middleware for
use in high performance computing, and is a
author/co-author of over 40 papers in the field
of physics and high performance computing. He
received his PhD in physics from the University
of Texas at Austin, and then went on to Syracuse
University, and then the Princeton Plasma Physics Laboratory.
| 5cs.CE
|
Maximizing Non-monotone/Non-submodular Functions by
Multi-objective Evolutionary Algorithms
Chao Qian1 , Yang Yu2 , Ke Tang3 , Xin Yao3 , Zhi-Hua Zhou2∗
arXiv:1711.07214v1 [cs.NE] 20 Nov 2017
1 Anhui
Province Key Lab of Big Data Analysis and Application, University of Science and Technology of China, Hefei, China
2 National Key Lab for Novel Software Technology,
3 Shenzhen
Nanjing University, Nanjing, China
Key Lab of Computational Intelligence, Southern University of Science and Technology, Shenzhen, China
Abstract
Evolutionary algorithms (EAs) are a kind of nature-inspired general-purpose optimization algorithm, and have shown empirically good performance in solving various real-word optimization
problems. However, due to the highly randomized and complex behavior, the theoretical analysis of EAs is difficult and is an ongoing challenge, which has attracted a lot of research attentions.
During the last two decades, promising results on the running time analysis (one essential theoretical aspect) of EAs have been obtained, while most of them focused on isolated combinatorial
optimization problems, which do not reflect the general-purpose nature of EAs. To provide a general theoretical explanation of the behavior of EAs, it is desirable to study the performance of EAs
on a general class of combinatorial optimization problems. To the best of our knowledge, this direction has been rarely touched and the only known result is the provably good approximation guarantees of EAs for the problem class of maximizing monotone submodular set functions with matroid constraints, which includes many NP-hard combinatorial optimization problems. The aim of
this work is to contribute to this line of research. As many combinatorial optimization problems
also involve non-monotone or non-submodular objective functions, we consider these two general
problem classes, maximizing non-monotone submodular functions without constraints and maximizing monotone non-submodular functions with a size constraint. We prove that a simple multiobjective EA called GSEMO can generally achieve good approximation guarantees in polynomial
expected running time.
Key words: Evolutionary algorithms, submodular optimization, multi-objective evolutionary
algorithms, running time analysis, computational complexity
∗ Corresponding
author
Email addresses: [email protected] (Chao Qian1 ), [email protected] (Yang Yu2 ), [email protected] (Ke Tang3 ),
[email protected] (Xin Yao3 ), [email protected] (Zhi-Hua Zhou2 )
Preprint submitted for review
November 21, 2017
1. Introduction
Evolutionary algorithms (EAs) [3] are a kind of randomized metaheuristic optimization algorithm,
inspired by the evolution process of natural species, i.e., natural selection and survival of the fittest.
Starting from a random population of solutions, EAs iteratively apply reproduction operators to generate a set of offspring solutions from the current population, and then apply a selection operator
to weed out bad solutions. EAs have been applied to diverse areas (e.g., antenna design [18], bioinformatics [23] and data mining [24]) and can produce human-competitive results [20]. Compared
with the application, the theoretical analysis of EAs is, however, far behind. Many researchers thus
have been devoted to understanding the behavior of EAs from a theoretical point of view, which is
still an ongoing challenge.
In the last two decades, a lot of progress has been made on the running time analysis of EAs, which
is one essential theoretical aspect. The running time measures how many objective (i.e., fitness)
function evaluations an EA needs until finding an optimal solution or an approximate solution.
The running time analysis of EAs started with artificial example problems. In [7, 8], a simple singleobjective EA called (1+1)-EA was shown able to solve two well-structured pseudo-Boolean problems
OneMax and LeadingOnes in Θ(n log n) and Θ(n2 ) (where n is the problem size) expected running
time, respectively. These two problems are to maximize the number of 1-bits of a solution and the
number of consecutive 1-bits counting from the left of a solution, respectively. Both of them have
a short path with increasing fitness to the optimum. For some problems (e.g., SPC) where there is
a short path with constant fitness to the optimum, the (1+1)-EA can also find an optimal solution
in polynomial time [19]. But when the problem (e.g., Trap) has a deceptive path (i.e., a path with
increasing fitness away from the optimum), the (1+1)-EA will need exponential running time [15].
More results can refer to [2].
The analysis on simple artificial problems disclosed theoretical properties of EAs (e.g., which problem structures are easy or hard for EAs), and also helped to develop approaches for analyzing more
complex problems. The running time analysis of EAs was then extended to combinatorial optimization problems. For some P-solvable problems, EAs were shown able to find an optimal solution in polynomial time. For example, the minimum spanning tree problem can be solved by
the (1+1)-EA and a simple multi-objective EA called GSEMO in O(m2 (log n + log wmax )) [29] and
O(mn(n + log wmax )) [28] expected time, respectively. Note that m, n and wmax are the number of
edges, the number of nodes and the maximum edge weight of a graph, respectively. For some NPhard problems, EAs were shown able to achieve good approximation ratios in polynomial time. For
example, for the partition problem, the (1+1)-EA can achieve a (4/3)-approximation ratio in O(n2 )
expected time [36]; for the minimum set cover problem, the expected running time of the GSEMO
2
until obtaining a (log m + 1)-approximation ratio is O(m2 n + mn(log n + log cmax )) [11], where m,
n and cmax denote the size of the ground set, the number of subsets and the maximum cost of a
subset, respectively. More running time results of EAs on combinatorial optimization problems can
refer to [30].
For the analysis of the GSEMO (which is a multi-objective EA) on single-objective optimization
problems (e.g., minimum spanning tree and minimum set cover), the original single-objective problem is transformed into a multi-objective problem, which is then solved by the GSEMO. Note that
multi-objective optimization here is just an intermediate process (which might be beneficial [11, 28,
31, 33]), and we still focus on the quality of the best solution w.r.t. the original single-objective problem, in the population found by the GSEMO. Running time analysis of EAs on real multi-objective
optimization problems has also been investigated, where the running time is measured by the number of fitness evaluations until finding the Pareto front (which represents different optimal tradeoffs
between the multiple objectives) or an approximation of the Pareto front. For example, Giel [12]
proved that the GSEMO can solve the bi-objective pseudo-Boolean problem LOTZ in O(n3 ) expected time; for the NP-hard bi-objective minimum spanning tree problem, it was shown that the
GSEMO can obtain a 2-approximation ratio in pseudo-polynomial time [27, 32].
The analysis on combinatorial optimization problems helped to reveal the ability of EAs. However,
most of the previous promising results were obtained for isolated problems, while EAs are known to
be general-purpose optimization algorithms, which can be applied to various problems. Thus, it is
more desirable to provide a general theoretical explanation of the behavior of EAs, that is, to theoretically study the performance of EAs on a general class of combinatorial optimization problems.
To the best of our knowledge, only two pieces of work in this direction have been reported. Reichel
and Skutella [35] first studied the problem class of maximizing linear functions with k matroid constraints, which includes some well-known combinatorial optimization problems such as maximum
matching, Hamiltonian path, etc. They proved that the (1+1)-EA can obtain a (1/k)-approximation
ratio in O(nk+2 (log r + log wmax )) expected running time, where n, r and wmax denote the size of
the ground set, the minimum rank of the ground set w.r.t. one matroid and the maximum weight
of an element, respectively. Later, a generalization of this problem class, i.e., the objective function
is relaxed to satisfy the monotone and submodular property, was considered in [10]. The (1+1)-EA
1
)-approximation ratio in O( 1ǫ n2p(k+1)+1 k log n) expected time,
was shown able to achieve a ( k+1/p+ǫ
where p ≥ 1 and ǫ > 0. Friedrich and Neumann [10] also studied a specific non-monotone case for
the objective function, i.e., the objective function is symmetric. They proved that the expected run1
ning time until the GSEMO obtains a ( (k+2)(1+ǫ)
)-approximation ratio for maximizing symmetric
submodular functions with k matroid constraints is O( 1ǫ nk+6 log n).
3
The aim of this paper is to contribute to this line of research. Considering that the objective function of many combinatorial optimization problems can be non-monotone (not necessarily symmetric) or non-submodular, we study the performance of EAs on these two general problem classes,
maximizing non-monotone submodular functions without constraints and maximizing monotone
non-submodular functions with a size constraint. Note that the objective function is a set function
f : 2V → R which maps a subset of the ground set V to a real value, and a size constraint means that
the size of a subset is not larger than a budget k. We prove that for both problem classes, the GSEMO
can obtain a good approximation guarantee in polynomial expected running time. Our main results
can be summarized as follows.
• For the problem class of maximizing non-monotone submodular functions without constraints,
with special instances including maximum cut [13], maximum facility location [1] and variants
of the maximum satisfiability problem [14], we prove that the GSEMO can achieve a constant
approximation ratio of
1
3
−
ǫ
n
in O( 1ǫ n4 log n) expected running time (i.e., Theorem 1), where n
is the size of the ground set V and ǫ > 0.
• For the problem class of maximizing monotone non-submodular functions with a size constraint, with special instances including sparse regression [6], dictionary selection [21] and
robust influence maximization [5, 16], we prove the approximation guarantee of the GSEMO
w.r.t. each notion of “approximate submodularity”, which measures how close a general set
function f is to submodularity.
(1) In [22], a set function f is ǫ-approximately submodular if the diminishing return property
holds with some deviation ǫ ≥ 0, i.e., for any X ⊆ Y ⊆ V and v ∈
/ Y , f (X ∪ {v}) − f (X) ≥
f (Y ∪{v})−f (Y )−ǫ. f is submodular iff ǫ = 0. We prove that the GSEMO within O(n2 (log n+k))
expected time can find a subset X with f (X) ≥ (1 − e−1 ) · (OP T − kǫ) (i.e., Theorem 2), where
e is the base of the natural logarithm and OP T denotes the optimal function value.
(2) In [6], the approximate submodularity notion of a set function f is characterized by a quantity γ called submodularity ratio. f is submodular iff γ = 1. We prove that the GSEMO within
O(n2 (log n + k)) expected time can find a subset X with f (X) ≥ (1 − e−γ ) · OP T (i.e., Theorem 3).
(3) In [17], a set function f is ǫ-approximately submodular if there exists a submodular set
function g such that ∀X ⊆ V , (1 − ǫ)g(X) ≤ f (X) ≤ (1 + ǫ)g(X). f is submodular iff ǫ = 0.
We prove that the GSEMO within O(n2 (log n + k)) expected time can find a subset X with
f (X) ≥
1
2kǫ
1+ 1−ǫ
1−ǫ k
(1 − e−1 ( 1+ǫ
) ) · OP T (i.e., Theorem 4).
Note that since EAs are general-purpose algorithms which utilize a small amount of problem knowledge, we cannot expect them to beat the best problem-specific algorithm. For the first problem
class, the approximation ratio nearly 1/3 obtained by the GSEMO is worse than the best known one
4
1/2, which was previously obtained by the double greedy algorithm [4]. For the second problem
class, the approximation ratio obtained by the GSEMO w.r.t. each notion of approximate submodularity reaches the best known one, which was previously obtained by the standard greedy algorithm [22, 6, 17]. Particularly, when the objective function is submodular, the obtained approximation ratio by the GSEMO becomes 1 − 1/e, which is optimal in general [25], and also consistent with
the previous result in [10]. Our analytical results on general problem classes together with the previous ones [35, 10] provide a theoretical explanation for the empirically good behaviors of EAs in
diverse applications.
The rest of this paper is organized as follows. Section 2 introduces some preliminaries. The running
time analyses for non-monotone and non-submodular cases are then presented in Sections 3 and 4,
respectively. Section 5 concludes the paper.
2. Preliminaries
In this section, we first introduce the two problem classes studied in this paper, and then introduce
the multi-objective evolutionary algorithm GSEMO and also how to optimize the studied problem
classes by the GSEMO.
2.1. Non-monotone/Non-submodular Function Maximization
Let R and R+ denote the set of reals and non-negative reals, respectively. Given a finite nonempty
set V = {v1 , v2 , . . . , vn }, we study the functions f : 2V → R defined on subsets of V . A set function
f : 2V → R is monotone if for any X ⊆ Y , f (X) ≤ f (Y ). Without loss of generality, we assume that
the monotone function is normalized, i.e., f (∅) = 0. A set function f is submodular [26] if for any
X ⊆ Y ⊆ V and v ∈
/ Y,
f (X ∪ {v}) − f (X) ≥ f (Y ∪ {v}) − f (Y );
(1)
or equivalently for any X ⊆ Y ⊆ V ,
f (Y ) − f (X) ≤
X
v∈Y \X
f (X ∪ {v}) − f (X) .
(2)
We assume that a set function f is given by a value oracle, i.e., for a given subset X, an algorithm
can query an oracle to obtain the value f (X).
5
Our first studied problem class as presented in Definition 1 is to maximize a non-monotone submodular set function. Without loss of generality, we assume that the objective function f is nonnegative. This problem generalizes many NP-hard combinatorial optimization problems including maximum cut [13], maximum facility location [1], variants of the maximum satisfiability problem [14], etc. The best known approximation guarantee is 1/2, which was achieved by the double
greedy algorithm [4].
Definition 1 (Non-monotone Submodular Function Maximization without Constraints)
Given a non-monotone submodular function f : 2V → R+ , it is to find a subset X ⊆ V such that
arg maxX⊆V
f (X).
The second studied problem class is presented in Definition 2. The goal is to find a subset with
at most k elements such that a given monotone non-submodular set function is maximized. This
problem generalizes many NP-hard problems including sparse regression [6], dictionary selection [21]
and robust influence maximization [5, 16], etc. When the objective function is submodular, it was
well known that the standard greedy algorithm achieves the optimal approximation guarantee 1 −
1/e [25, 26]. For non-submodular cases, several notions of “approximate submodularity” [22, 6, 17]
were introduced to measure to what extent a general set function f has the submodular property.
For each approximately submodular notion, the best known approximation guarantee was achieved
by the standard greedy algorithm [22, 6, 17], which iteratively adds one element with the largest f
improvement until k elements are selected.
Definition 2 (Monotone Non-submodular Function Maximization with a Size Constraint)
Given a monotone non-submodular function f : 2V → R+ and a budget k , it is to find a subset
X ⊆ V such that
arg maxX⊆V
f (X) s.t. |X| ≤ k.
In [22], the approximate submodularity as presented in Definition 3 was defined based on the diminishing return property (i.e., Eq. (1)). That is, the approximately submodular degree depends
on how large a deviation of ǫ the diminishing return property can hold with. It is easy to see that
f is submodular iff ǫ = 0. The standard greedy algorithm was proved to find a subset X with
f (X) ≥ (1 − e−1 ) · (OP T − kǫ) [22], where OP T denotes the optimal function value.
Definition 3 (ǫ-Approximately Submodular [22])
Let ǫ ≥ 0. A set function f : 2V → R is ǫ-approximately submodular if for any X ⊆ Y ⊆ V and v ∈
/ Y,
f (X ∪ {v}) − f (X) ≥ f (Y ∪ {v}) − f (Y ) − ǫ.
In [6], the submodularity ratio as presented in Definition 4 was introduced to measure the closeness
of a set function f to submodularity. It is easy to see from Eq. (2) that f is submodular iff γU,k (f ) = 1
6
for any U and k. When f is clear in the paper, we will use γU,k shortly. The standard greedy algorithm
was proved to find a subset X with f (X) ≥ (1 − e−γX,k ) · OP T [6].
Definition 4 (Submodularity Ratio [6])
Let f : 2V → R be a set function. The submodularity ratio of f with respect to a set U ⊆ V and a
parameter k ≥ 1 is
γU,k (f ) =
min
L⊆U,S:|S|≤k,S∩U=∅
P
∪ {v}) − f (L))
.
f (L ∪ S) − f (L)
v∈S (f (L
The above two notions of approximate submodularity are based on the equivalent statements Eqs. (1)
and (2) of submodularity, while in [17], the approximate submodularity of a set function f as presented in Definition 5 was defined based on the closeness to other submodular functions. It is easy
to see that f is submodular iff ǫ = 0. The standard greedy algorithm was proved to find a subset X
with f (X) ≥
1
4kǫ
1+ (1−ǫ)
2
2k
(1 − e−1 ( 1−ǫ
1+ǫ ) ) · OP T [17].
Definition 5 (ǫ-Approximately Submodular [17])
Let ǫ ≥ 0. A set function f : 2V → R is ǫ-approximately submodular if there exists a submodular set
function g such that ∀X ⊆ V ,
(1 − ǫ) · g(X) ≤ f (X) ≤ (1 + ǫ) · g(X).
2.2. Multi-objective Evolutionary Algorithms
To investigate the performance of EAs optimizing the two problem classes in Definitions 1 and 2,
we consider a simple multi-objective EA called GSEMO, which has been widely used in previous
theoretical analyses [11, 31, 32, 37]. The GSEMO as presented in Algorithm 1 is used for maximizing
multi-objective pseudo-Boolean problems with m objective functions fi : {0, 1}n → R (1 ≤ i ≤ m).
Note that a pseudo-Boolean function f : {0, 1}n → R naturally characterizes a set function f : 2V →
R, since a subset X of V can be naturally represented by a Boolean vector x ∈ {0, 1}n, where the
i-th bit xi = 1 means that vi ∈ X, and xi = 0 means that vi ∈
/ X. Throughout the paper, we will not
distinguish x ∈ {0, 1}n and its corresponding subset for notational convenience.
Before introducing the GSEMO, we first introduce some basic concepts in multi-objective maximization. Since the objectives to be maximized are usually conflicted, there is no canonical complete order on the solution space. The comparison between two solutions relies on the domination
relationship. For two solutions x and x′ , x weakly dominates x′ (i.e., x is better than x′ , denoted by
x x′ ) if ∀1 ≤ i ≤ m, fi (x) ≥ fi (x′ ); x dominates x′ (i.e., x is strictly better than x′ , denoted by
x ≻ x′ ) if x x′ and fi (x) > fi (x′ ) for some i. The domination relationship between two solutions
x and x′ can be formally stated as follows:
7
Algorithm 1 GSEMO Algorithm
Given m pseudo-Boolean objective functions f1 , f2 , . . . , fm , where fi : {0, 1}n → R, the GSEMO
consists of the following steps:
1:
Choose x ∈ {0, 1}n uniformly at random
2:
P ← {x}
3:
repeat
4:
Choose x from P uniformly at random
5:
Create x′ by flipping each bit of x with probability 1/n
6:
Create x′′ ← V \ x′
7:
for y ∈ {x′, x′′ }
8:
9:
10:
11:
12:
if ∄z ∈ P such that z ≻ y
P ← (P \ {z ∈ P | y z}) ∪ {y}
end if
end for
until some criterion is met
• x x′ if ∀1 ≤ i ≤ m, fi (x) ≥ fi (x′ );
• x ≻ x′ if x x′ and fi (x) > fi (x′ ) for some i.
But if neither x is better than x′ nor x′ is better than x, we say that they are incomparable. A solution
is Pareto optimal if there is no other solution that dominates it. The set of objective vectors of all the
Pareto optimal solutions constitutes the Pareto front. The goal of multi-objective optimization is to
find the Pareto front, that is, to find at least one corresponding solution for each objective vector in
the Pareto front.
The procedure of the GSEMO algorithm is presented in Algorithm 1. Starting from a random solution (lines 1-2), it iteratively tries to improve the quality of the solutions in the population P (lines 312). In each iteration, a new solution x′ is generated by randomly flipping bits of an archived solution x selected from the current population P (lines 4-5); the complementary set x′′ = V \ x′ of x′ is
also generated (line 6); these two newly generated solutions are then used to update the population
P (lines 7-11). In the updating procedure, if y ∈ {x′ , x′′ } is not dominated by (i.e., not strictly worse
than) any previously archived solution (line 8), it will be added into P , and meanwhile those previously archived solutions weakly dominated by (i.e., worse than) y will be removed from P (line 9).
It is easy to see that the population P will always contain a set of incomparable solutions due to
the domination-based comparison. Note that the GSEMO here is a little different from its original
version [11, 31, 32]. In the original GSEMO algorithm, only x′ is generated and used to update the
population in each iteration, while here we also generate the complementary set of x′ and use both
8
of them to update the population.
For optimizing the two problems in Definitions 1 and 2 by the GSEMO, each problem is transformed
into a bi-objective maximization problem
arg maxx∈{0,1}n
(f1 (x), f2 (x)),
where f1 (x) = f (x) and f2 (x) = −|x|. That is, the GSEMO is to maximize the objective function
Pn
f and minimize the subset size |x| simultaneously. Note that |x| = i=1 xi denotes the number of
1-bits of a solution x. When the GSEMO terminates after running a number of iterations, the best
solution w.r.t. the original single-objective problem in the resulting population P will be returned.
For the first problem in Definition 1, the solution with the largest f value in P (i.e., arg maxx∈P f (x))
will be returned. For the second problem in Definition 2, the solution with the largest f value satisfying the size constraint in P (i.e., arg maxx∈P,|x|≤k f (x)) will be returned. The running time of the
GSEMO is measured by the number of fitness evaluations until the best solution w.r.t. the original single-objective problem in the population reaches some approximation guarantee for the first
time. Since only the new solutions x′ and x′′ need to be evaluated in each iteration of the GSEMO,
the number of fitness evaluations is just the double of the number of iterations of the GSEMO.
Note that multi-objective optimization here is just an intermediate process, which has been shown
helpful for solving some single-objective combinatorial optimization problems [11, 28, 31, 33]. We
still focus on the quality of the best solution w.r.t. the original single-objective problem, in the population found by the GSEMO, rather than the quality of the population w.r.t. the transformed multiobjective optimization problem.
3. Non-monotone Submodular Function Maximization without Constraints
In this section, we theoretically analyze the performance of the GSEMO for maximizing non-monotone
submodular functions without constraints. We prove in Theorem 1 that the GSEMO can achieve a
constant approximation ratio of nearly 1/3 in polynomial time. Note that OP T denotes the optimal
function value. Inspired from the proof of Theorem 4 in [10], our proof idea is to follow the behavior
of the local search algorithm [9], which iteratively tries to improve a solution by inserting or deleting one element. The proof relies on Lemma 1, which shows that it is always possible to improve a
solution until a good approximation has been achieved. This lemma is extracted from Lemma 3.4
in [9].
Lemma 1
[9] Let x ∈ {0, 1}n be a solution such that no solution x′ with the objective value f (x′ ) > (1+ nǫ2 )·f (x)
9
can be achieved by inserting one element into x or deleting one element from x, where ǫ > 0. Then
max{f (x), f (V \ x)} ≥ ( 13 − nǫ ) · OP T .
Theorem 1
For maximizing a non-monotone submodular function without constraints, the expected running
time of the GSEMO until finding a solution x with f (x) ≥ ( 31 − nǫ ) · OP T is O( 1ǫ n4 log n), where ǫ > 0.
Proof. We divide the optimization process into three phases: (1) starts from an initial random solution and finishes until finding the special solution {0}n (i.e., ∅); (2) starts after phase (1) and finishes
until finding a solution with the objective value at least OP T /n; (3) starts after phase (2) and finishes until finding a solution with the desired approximation guarantee. We analyze the expected
running time of each phase, respectively, and then sum up them to get an upper bound on the total
expected running time of the GSEMO.
For phase (1), we consider the minimum number of 1-bits of the solutions in the population P ,
denoted by Jmin . That is, Jmin = min{|x| | x ∈ P }. Assume that currently Jmin = i > 0, and let
x be the corresponding solution, i.e., |x| = i. It is easy to see that Jmin cannot increase because
x cannot be weakly dominated by a solution with more 1-bits. In each iteration of the GSEMO, to
decrease Jmin , it is sufficient to select x in line 4 of Algorithm 1 and flipping only one 1-bit of x in
line 5. This is because the newly generated solution x′ now has the smallest number of 1-bits (i.e.,
|x′ | = i−1) and no solution in P can dominate it; thus it will be included into P . Let Pmax denote the
largest size of P during the run of the GSEMO. The probability of selecting x in line 4 of Algorithm 1
is
1
|P |
i
n (1
≥
−
1
Pmax
1 n−1
n)
due to uniform selection, and the probability of flipping only one 1-bit of x in line 5 is
≥
i
en ,
since x has i 1-bits. Thus, the probability of deceasing Jmin by at least 1 in each
iteration of the GSEMO is at least
i
enPmax .
Note that Jmin ≤ n. We can then get that the expected
number of iterations of phase (1) (i.e., Jmin reaches 0) is at most
n
X
enPmax
i=1
i
= O(nPmax log n).
Note that the solution {0}n will be always kept in P once generated, since it has the smallest subset
size 0 and no other solution can weakly dominate it.
For phase (2), it is sufficient that in one iteration of the GSEMO, the solution {0}n is selected in line 4,
and only a specific 0-bit corresponding to the best single element v ∗ (i.e., v ∗ ∈ arg maxv∈V f ({v})) is
flipped in line 5. That is, the solution {v ∗ } is generated. Since the objective function f is submodular
and non-negative, we easily have f ({v ∗ }) ≥ OP T /n. After generating the solution {v ∗ }, it will then
be used to update the population P , which makes P always contain a solution z weakly dominating
{v ∗ }, i.e., f (z) ≥ f ({v ∗ }) ≥ OP T /n and |z| ≤ |{v ∗ }| = 1. Thus, we only need to analyze the expected number of iterations of the GSEMO until generating the solution {v ∗ }. Since the probability
10
of selecting {0}n in line 4 of the GSEMO is at least
0-bit in line 5 is
1
n (1
−
1 n−1
n)
≥
1
en ,
1
Pmax
and the probability of flipping only a specific
the expected number of iterations of phase (2) is O(nPmax ).
As in [9], we call a solution x a (1+α)-approximate local optimum if f (x\{v}) ≤ (1+α)·f (x) for any
v ∈ x and f (x ∪ {v}) ≤ (1 + α) · f (x) for any v ∈
/ x. According to Lemma 1, we know that a (1 +
ǫ
n2 )-
approximate local optimum x satisfies that max{f (x), f (V \ x)} ≥ ( 31 − nǫ ) · OP T . For phase (3), we
thus only need to analyze the expected number of iterations until generating a (1+ nǫ2 )-approximate
local optimum x′ in line 5 of Algorithm 1. This is because both x′ and V \ x′ will be used to update
the population P , and then for either one of x′ and V \ x′ , P will always contain one solution weakly
dominating it, which implies that max{f (x) | x ∈ P } ≥ max{f (x′ ), f (V \ x′ )} ≥ ( 13 −
ǫ
n)
· OP T .
We then consider the largest f value of the solutions in the population P , denoted by Jmax . That is,
Jmax = max{f (x) | x ∈ P }. After phase (2), Jmax ≥ OP T /n, and let x be the corresponding solution,
i.e., f (x) = Jmax . It is obvious that Jmax cannot decrease, because x cannot be weakly dominated by
a solution with a smaller f value. As long as x is not a (1 + nǫ2 )-approximate local optimum, we know
that a new solution x′ with f (x′ ) > (1 +
ǫ
n2 )f (x)
= (1 + nǫ2 )Jmax can be generated through selecting
x in line 4 of Algorithm 1 and flipping only one specific 1-bit (i.e., deleting one specific element from
x) or one specific 0-bit (i.e., adding one specific element into x) in line 5, the probability of which is
at least
1
Pmax
· n1 (1 − n1 )n−1 ≥
1
enPmax .
Since x′ now has the largest f value and no other solution in
P can dominate it, it will be included into P . Thus, Jmax can increase by at least a factor of 1 +
with probability at least
1
enPmax
ǫ
n2
in each iteration. Such an increase on Jmax is called a successful step.
Thus, a successful step needs at most enPmax expected number of iterations. It is also easy to see
that until generating a (1 +
most log1+
ǫ
n2
OP T
OP T /n
ǫ
n2 )-approximate
local optimum, the number of successful steps is at
= O( 1ǫ n2 log n). Thus, the expected number of iterations of phase (3) is at most
enPmax · O
1 2
n log n
ǫ
=O
1 3
n Pmax log n .
ǫ
From the procedure of the GSEMO, we know that the solutions maintained in P must be incomparable. Thus, each value of one objective can correspond to at most one solution in P . Because the
second objective f2 (x) = −|x| can only belong to {0, −1, . . . , −n}, we have Pmax ≤ n + 1. Hence,
the expected running time of the GSEMO for finding a solution with the objective function value at
least ( 13 − nǫ ) · OP T is
O(nPmax log n) + O(nPmax ) + O
1 3
n Pmax log n
ǫ
11
=O
1 4
n log n .
ǫ
4. Monotone Non-submodular Function Maximization with a Size Constraint
In this section, we analyze the performance of the GSEMO for maximizing monotone non-submodular
functions with a size constraint. We prove the polynomial-time approximation guarantee of the
GSEMO w.r.t. each notion of approximate submodularity (i.e., Definitions 3-5), respectively. Theorem 2 gives the approximation guarantee of the GSEMO w.r.t. the ǫ-approximately submodular
notion in Definition 3, where OP T denotes the optimal function value. Note that this approximation guarantee obtained by the GSEMO reaches the best known one, which was previously obtained
by the standard greedy algorithm [22]. Inspired from the proof of Theorem 2 in [10], our proof idea is
to follow the behavior of the standard greedy algorithm, which iteratively adds one element with the
currently largest improvement on f . The proof relies on the property of f in Lemma 2, that for any
x ∈ {0, 1}n , there always exists another element, the inclusion of which can bring an improvement
on f roughly proportional to the current distance to the optimum.
Lemma 2
Assume that a set function f is monotone and ǫ-approximately submodular as in Definition 3. Then,
/ x such that
for any x ∈ {0, 1}n, there exists one element v ∈
f (x ∪ {v}) − f (x) ≥
1
(OP T − f (x)) − ǫ.
k
Proof. Let x∗ be an optimal solution, i.e., f (x∗ ) = OP T . We denote the elements in x∗ \ x by
v1∗ , v2∗ , . . . , vl∗ , where |x∗ \ x| = l ≤ k. Then, we have
f (x∗ ) − f (x) ≤ f (x ∪ x∗ ) − f (x)
=
l
X
j=1
≤
l
X
j=1
∗
f (x ∪ {v1∗ , . . . , vj∗ }) − f (x ∪ {v1∗ , . . . , vj−1
})
f (x ∪ {vj∗ }) − f (x) + ǫ ,
where the first inequality is by the monotonicity of f , and the last inequality is derived by Definition 3 since f is ǫ-approximately submodular. Let v ∗ = arg maxv∈x∗ \x f (x ∪ {v}). Then, we have
f (x ∪ {v ∗ }) − f (x) ≥
1
1
(f (x∗ ) − f (x)) − ǫ ≥ (OP T − f (x)) − ǫ.
l
k
Theorem 2
For maximizing a monotone non-submodular function f with a size constraint, where f is ǫ-approximately
submodular as in Definition 3, the expected running time of the GSEMO until finding a solution x
with |x| ≤ k and f (x) ≥ (1 − e−1 ) · (OP T − kǫ) is O(n2 (log n + k)).
12
Proof. We divide the optimization process into two phases: (1) starts from an initial random solution and finishes until finding the special solution {0}n ; (2) starts after phase (1) and finishes until
finding a solution with the desired approximation guarantee. As the analysis of phase (1) in the proof
of Theorem 1, we know that the population P will contain the solution {0}n after O(nPmax log n) iterations in expectation.
For phase (2), we consider a quantity Jmax , which is defined as
(
Jmax = max j ∈ [0, k] | ∃x ∈ P, |x| ≤ j ∧ f (x) ≥
1
1− 1−
k
j !
)
· (OP T − kǫ) .
That is, Jmax denotes the maximum value of j ∈ [0, k] such that in the population P , there exists a
solution x with |x| ≤ j and f (x) ≥ (1 − (1 − k1 )j ) · (OP T − kǫ). We analyze the expected number of
iterations until Jmax = k, which implies that there exists one solution x in P satisfying that |x| ≤ k
and f (x) ≥ (1 − (1 − k1 )k ) · (OP T − kǫ) ≥ (1 − e−1 ) · (OP T − kǫ). That is, the desired approximation
guarantee is reached.
The current value of Jmax is at least 0, since the population P contains the solution {0}n , which will
always be kept in P once generated. Assume that currently Jmax = i < k. Let x be a corresponding
solution with the value i, i.e., |x| ≤ i and f (x) ≥ (1 − (1 − k1 )i ) · (OP T − kǫ). It is easy to see that
Jmax cannot decrease because cleaning x from P (line 9 of Algorithm 1) implies that x is weakly
dominated by a newly generated solution y, which must satisfy that |y| ≤ |x| and f (y) ≥ f (x). By
Lemma 2, we know that flipping one specific 0 bit of x (i.e., adding a specific element) can generate
a new solution x′ , which satisfies that f (x′ ) − f (x) ≥ k1 (OP T − f (x)) − ǫ. Then, we have
i+1 !
1
1
1
′
f (x ) ≥ 1 −
f (x) + · OP T − ǫ ≥ 1 − 1 −
· (OP T − kǫ),
k
k
k
where the last inequality is derived by f (x) ≥ (1−(1− k1 )i )·(OP T −kǫ). Since |x′ | = |x|+1 ≤ i+1, x′
will be included into P ; otherwise, x′ must be dominated by one solution in P (line 8 of Algorithm 1),
and this implies that Jmax has already been larger than i, which contradicts with the assumption
Jmax = i. After including x′ , Jmax ≥ i + 1. Thus, Jmax can increase by at least 1 in one iteration with
probability at least
1
Pmax
· n1 (1 − n1 )n−1 ≥
selecting x in line 4 of Algorithm 1 and
1
enPmax ,
1
n (1
where
1
Pmax
is a lower bound on the probability of
− n1 )n−1 is the probability of flipping a specific bit of x
while keeping other bits unchanged in line 5. Then, it needs at most enPmax expected number of
iterations to increase Jmax . Thus, after at most k · enPmax iterations in expectation, Jmax must have
reached k.
As the proof of Theorem 1, we know that Pmax ≤ n + 1. Thus, by summing up the expected running
time of two phases, we get that the expected running time of the GSEMO for finding a solution x
with |x| ≤ k and f (x) ≥ (1 − e−1 ) · (OP T − kǫ) is O(nPmax log n + knPmax ) = O(n2 (log n + k)).
13
Theorem 3 gives the approximation guarantee of the GSEMO w.r.t. the submodularity ratio in Definition 4. Note that it was proved that the standard greedy algorithm can find a subset x with |x| = k
and f (x) ≥ (1 − e−γx,k ) · OP T [6]. Thus, Theorem 3 shows that the GSEMO can achieve nearly this
best known approximation guarantee. The proof of Theorem 3 is similar to that of Theorem 2. The
main difference is that a different inductive inequality on f is used in the definition of the quantity
Jmax , due to the change of the adopted notion of approximate submodularity. For concise illustration, we will mainly show the difference in the proof of Theorem 3.
Lemma 3
[34] Assume that a set function f is monotone and non-submodular. Then, for any x ∈ {0, 1}n, there
/ x such that
exists one element v ∈
f (x ∪ {v}) − f (x) ≥
γx,k
(OP T − f (x)),
k
where γx,k is the submodularity ratio of f w.r.t. x and k as in Definition 4.
Theorem 3
For maximizing a monotone non-submodular function f with a size constraint, the expected running time of the GSEMO until finding a solution x with |x| ≤ k and f (x) ≥ (1 − e−γmin ) · OP T is
O(n2 (log n + k)), where γmin = minx:|x|=k−1 γx,k and γx,k is the submodularity ratio of f w.r.t. x and
k as in Definition 4.
Proof. The proof is similar to that of Theorem 2. We use a different Jmax , which is defined as
γmin j
· OP T .
Jmax = max j ∈ [0, k] | ∃x ∈ P, |x| ≤ j ∧ f (x) ≥ 1 − 1 −
k
It is easy to verify that Jmax = k implies that the desired approximation guarantee is reached, since
there must exist one solution x in P satisfying that |x| ≤ k and f (x) ≥ (1 − (1 −
(1 − e
−γmin
γmin k
k ) )
· OP T ≥
) · OP T . Assume that currently Jmax = i < k and x is a corresponding solution, i.e.,
|x| ≤ i and f (x) ≥ (1 − (1 −
γmin i
k ) )
· OP T . We then only need to show that flipping one specific
0 bit of x can generate a new solution x′ with f (x′ ) ≥ (1 − (1 −
γmin i+1
)
k )
· OP T . By Lemma 3,
we know that flipping one specific 0 bit of x can generate a new solution x′ , which satisfies that
γx,k
k (OP T
− f (x)). Then, we have
γx,k
γx,k
γmin i
γx,k
f (x′ ) ≥ 1 −
f (x) +
1−
· OP T
· OP T ≥ 1 − 1 −
k
k
k
k
γmin i+1
· OP T,
≥ 1− 1−
k
f (x′ ) − f (x) ≥
where the second inequality is by f (x) ≥ (1 − (1 −
γmin i
k ) )
· OP T , and the last inequality is by γx,k ≥
γmin , which can be easily derived from |x| < k and γx,k decreasing with x. Thus, the theorem holds.
14
Theorem 4 gives the approximation guarantee of the GSEMO w.r.t. the ǫ-approximately submodular
ratio in Definition 5. Note that the standard greedy algorithm obtains the best known approximation
guarantee f (x) ≥
1
4kǫ
1+ (1−ǫ)
2
2k
(1 − (1 − k1 )k ( 1−ǫ
1+ǫ ) ) · OP T [17]. Comparing with this, the approximation
guarantee of the GSEMO is slightly better, since
k
k !
i
k−1
X
1
1
1 1−ǫ
1−ǫ
1−ǫ
1
−
1
−
1
−
=
·
2kǫ
k
1+ǫ
k(1 + ǫ) i=0
k 1+ǫ
1 + 1−ǫ
!
i
2
k
2k !
k−1
1
1
1−ǫ
1
1−ǫ
(1 − ǫ)2 X
1−
1− 1−
.
=
·
≥
4kǫ
k(1 + ǫ)2 i=0
k
1+ǫ
k
1+ǫ
1 + (1−ǫ)
2
The proof of Theorem 4 is also similar to that of Theorem 2, except that a different inductive inequality on f is used in the definition of the quantity Jmax , due to the change of the adopted notion
of approximate submodularity.
Lemma 4
Assume that a set function f is monotone and ǫ-approximately submodular as in Definition 5. Then,
/ x such that
for any x ∈ {0, 1}n, there exists one element v ∈
1−ǫ
1−ǫ
f (x) ≥
(OP T − f (x)).
1+ǫ
k(1 + ǫ)
f (x ∪ {v}) −
Proof. Let x∗ be an optimal solution, i.e., f (x∗ ) = OP T . Let v ∗ = arg maxv∈x∗ \x f (x∪{v}). Since f is
ǫ-approximately submodular as in Definition 5, we use g to denote one corresponding submodular
function satisfying that for all x ∈ {0, 1}n, (1 − ǫ)g(x) ≤ f (x) ≤ (1 + ǫ)g(x). Then, we have
g(x∗ ∪ x) − g(x) ≤
X
v∈x∗ \x
g(x ∪ {v}) − g(x)
X 1
≤
f (x ∪ {v}) − g(x)
1−ǫ
v∈x∗ \x
1
f (x ∪ {v ∗ }) − g(x) ,
≤k
1−ǫ
where the first inequality is by the submodularity of g (i.e., Eq. (2)), the second inequality is by (1 −
ǫ)g(x) ≤ f (x) for any x, and the last inequality is by the definition of v ∗ and |x∗ | ≤ k. By reordering
the terms, we get
1
1−ǫ
∗
(1 − ǫ)g(x).
g(x ∪ x) + 1 −
f (x ∪ {v }) ≥
k
k
∗
Since g(x) ≥
1
∗
1+ǫ f (x) and g(x ∪x)
≥
1
∗
1+ǫ f (x ∪x)
≥
1
∗
1+ǫ f (x )
=
1
1+ǫ OP T
(where the last inequality
is by the monotonicity of f ), we have
f (x ∪ {v ∗ }) ≥
1−ǫ
1 1−ǫ
OP T + 1 −
f (x).
k(1 + ǫ)
k 1+ǫ
By reordering the terms, the lemma holds.
15
Theorem 4
For maximizing a monotone non-submodular function f with a size constraint, where f is ǫ-approximately
submodular as in Definition 5, the expected running time of the GSEMO until finding a solution
x with |x| ≤ k and f (x) ≥
1
2kǫ
1+ 1−ǫ
(1 − (1 −
1 k 1−ǫ k
k ) ( 1+ǫ ) )
· OP T ≥
1
2kǫ
1+ 1−ǫ
k
(1 − e−1 ( 1−ǫ
1+ǫ ) ) · OP T is
O(n2 (log n + k)).
Proof. The proof is similar to that of Theorem 2. We use a different Jmax , which is defined as
(
)
j
j !
1
1−ǫ
1
Jmax = max j ∈ [0, k] | ∃x ∈ P, |x| ≤ j ∧ f (x) ≥
· OP T .
1− 1−
2kǫ
k
1+ǫ
1 + 1−ǫ
It is easy to verify that Jmax = k implies that the desired approximation guarantee is reached. Assume that currently Jmax = i < k and x is a corresponding solution, i.e., |x| ≤ i and f (x) ≥
1
2kǫ
1+ 1−ǫ
1−ǫ i
) ) · OP T . We then only need to show that flipping one specific 0 bit of x
(1 − (1 − k1 )i ( 1+ǫ
can generate a new solution x′ with f (x′ ) ≥
1
2kǫ
1+ 1−ǫ
i+1
) · OP T . By Lemma 4,
(1 − (1 − k1 )i+1 ( 1−ǫ
1+ǫ )
we know that flipping one specific 0 bit of x can generate a new solution x′ , which satisfies that
f (x′ ) −
1−ǫ
1+ǫ f (x)
f (x′ ) ≥
≥
≥
1−
1−ǫ
k(1+ǫ) (OP T
1
k
1
2kǫ
1 + 1−ǫ
− f (x)). Then, we have
1−ǫ
1−ǫ
f (x) +
· OP T
1+ǫ
k(1 + ǫ)
i+1
i+1 !
1
1−ǫ
1− 1−
· OP T,
k
1+ǫ
where the last inequality is by f (x) ≥
1
2kǫ
1+ 1−ǫ
1−ǫ i
(1 − (1 − k1 )i ( 1+ǫ
) ) · OP T . Thus, the theorem holds.
Thus, we have shown that for each notion of approximate submodularity, the GSEMO can always
obtain the best known approximation guarantee. Particularly, when the objective function is submodular, the quantity characterizing the approximately submodular degree in Definitions 3-5 becomes ǫ = 0, γU,k = 1 and ǫ = 0, respectively; thus the approximation guarantees obtained by the
GSEMO in Theorems 2-4 all become 1 − 1/e, which is optimal in general [25], and also consistent
with the previous result in [10].
5. Conclusion
This paper theoretically studies the approximation performance of EAs for two general classes of
combinatorial optimization problems, maximizing non-monotone submodular functions without
constraints and maximizing monotone non-submodular functions with a size constraint. We prove
that in polynomial expected running time, a simple multi-objective EA called GSEMO can achieve
a constant approximation guarantee of nearly 1/3 for the first problem class, and can achieve the
16
best known approximation guarantee for the second problem class. These results together with the
previous ones for the problem class of maximizing monotone submodular functions with matroid
constraints [10, 35] provide a theoretical explanation for the empirically good performance of EAs
in various applications. A question that will be investigated in the future is to investigate whether
simple single-objective EAs such as the (1+1)-EA can achieve good approximation guarantees on the
studied two problem classes. Most of the existing studies considered the problems where the objective function is a set function. We will also try to investigate the performance of EAs for optimizing
functions over the integer lattice or continuous domains.
References
[1] A. A. Ageev and M. I. Sviridenko. An 0.828-approximation algorithm for the uncapacitated
facility location problem. Discrete Applied Mathematics, 93(2):149–156, 1999.
[2] A. Auger and B. Doerr. Theory of Randomized Search Heuristics: Foundations and Recent Developments. World Scientific, Singapore, 2011.
[3] T. Bäck. Evolutionary Algorithms in Theory and Practice: Evolution Strategies, Evolutionary
Programming, Genetic Algorithms. Oxford University Press, Oxford, UK, 1996.
[4] N. Buchbinder, M. Feldman, J. Seffi, and R. Schwartz. A tight linear time (1/2)-approximation
for unconstrained submodular maximization. SIAM Journal on Computing, 44(5):1384–1402,
2015.
[5] W. Chen, T. Lin, Z. Tan, M. Zhao, and X. Zhou. Robust influence maximization. In Proceedings
of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining
(KDD’16), pages 795–804, San Francisco, CA, 2016.
[6] A. Das and D. Kempe. Submodular meets spectral: Greedy algorithms for subset selection,
sparse approximation and dictionary selection. In Proceedings of the 28th International Conference on Machine Learning (ICML’11), pages 1057–1064, Bellevue, WA, 2011.
[7] S. Droste, T. Jansen, and I. Wegener. A rigorous complexity analysis of the (1+1) evolutionary
algorithm for separable functions with Boolean inputs. Evolutionary Computation, 6(2):185–
196, 1998.
[8] S. Droste, T. Jansen, and I. Wegener. On the analysis of the (1+1) evolutionary algorithm. Theoretical Computer Science, 276(1-2):51–81, 2002.
[9] U. Feige, V. S. Mirrokni, and J. Vondrak. Maximizing non-monotone submodular functions.
SIAM Journal on Computing, 40(4):1133–1153, 2011.
17
[10] T. Friedrich and F. Neumann. Maximizing submodular functions under matroid constraints by
evolutionary algorithms. Evolutionary Computation, 23(4):543–558, 2015.
[11] T. Friedrich, J. He, N. Hebbinghaus, F. Neumann, and C. Witt. Approximating covering problems by randomized search heuristics using multi-objective models. Evolutionary Computation, 18(4):617–633, 2010.
[12] O. Giel. Expected runtimes of a simple multi-objective evolutionary algorithm. In Proceedings
of the 2003 IEEE Congress on Evolutionary Computation (CEC’03), pages 1918–1925, Canberra,
Australia, 2003.
[13] M. X. Goemans and D. P. Williamson. Improved approximation algorithms for maximum cut
and satisfiability problems using semidefinite programming. Journal of the ACM, 42(6):1115–
1145, 1995.
[14] J. Håstad. Some optimal inapproximability results. Journal of the ACM, 48(4):798–859, 2001.
[15] J. He and X. Yao. Drift analysis and average time complexity of evolutionary algorithms. Artificial Intelligence, 127(1):57–85, 2001.
[16] X. He and D. Kempe. Robust influence maximization. In Proceedings of the 22nd ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining (KDD’16), pages 885–894,
San Francisco, CA, 2016.
[17] T. Horel and Y. Singer. Maximization of approximately submodular functions. In Advances In
Neural Information Processing Systems 29 (NIPS’16), pages 3045–3053, Barcelona, Spain, 2016.
[18] G. S. Hornby, A. Globus, D. S. Linden, and J. D. Lohn. Automated antenna design with evolutionary algorithms. In Proceedings of the 2006 American Institute of Aeronautics and Astronautics Conference on Space, pages 19–21, San Jose, CA, 2006.
[19] T. Jansen and I. Wegener. Evolutionary algorithms-how to cope with plateaus of constant fitness and when to reject strings of the same fitness. IEEE Transactions on Evolutionary Computation, 5(6):589–599, 2001.
[20] J. R. Koza, M. A. Keane, and M. J. Streeter. What’s AI done for me lately? Genetic programming’s
human-competitive results. IEEE Intelligent Systems, 18(3):25–31, 2003.
[21] A. Krause and V. Cevher. Submodular dictionary selection for sparse representation. In Proceedings of the 27th International Conference on Machine Learning (ICML’10), pages 567–574,
Haifa, Israel, 2010.
18
[22] A. Krause, A. Singh, and C. Guestrin. Near-optimal sensor placements in Gaussian processes:
Theory, efficient algorithms and empirical studies. Journal of Machine Learning Research, 9:
235–284, 2008.
[23] P. C. Ma, K. C. Chan, X. Yao, and D. K. Chiu. An evolutionary clustering algorithm for gene
expression microarray data analysis. IEEE Transactions on Evolutionary Computation, 10(3):
296–314, 2006.
[24] A. Mukhopadhyay, U. Maulik, S. Bandyopadhyay, and C. A. C. Coello. A survey of multiobjective
evolutionary algorithms for data mining: Part I. IEEE Transactions on Evolutionary Computation, 18(1):4–19, 2014.
[25] G. L. Nemhauser and L. A. Wolsey. Best algorithms for approximating the maximum of a submodular set function. Mathematics of Operations Research, 3(3):177–188, 1978.
[26] G. L. Nemhauser, L. A. Wolsey, and M. L. Fisher. An analysis of approximations for maximizing
submodular set functions – I. Mathematical Programming, 14(1):265–294, 1978.
[27] F. Neumann. Expected runtimes of a simple evolutionary algorithm for the multi-objective
minimum spanning tree problem. European Journal of Operational Research, 181(3):1620–
1629, 2007.
[28] F. Neumann and I. Wegener. Minimum spanning trees made easier via multi-objective optimization. Natural Computing, 5(3):305–319, 2006.
[29] F. Neumann and I. Wegener. Randomized local search, evolutionary algorithms, and the minimum spanning tree problem. Theoretical Computer Science, 378(1):32–40, 2007.
[30] F. Neumann and C. Witt. Bioinspired Computation in Combinatorial Optimization: Algorithms
and Their Computational Complexity. Springer-Verlag, Berlin, Germany, 2010.
[31] F. Neumann, J. Reichel, and M. Skutella. Computing minimum cuts by randomized search
heuristics. Algorithmica, 59(3):323–342, 2011.
[32] C. Qian, Y. Yu, and Z.-H. Zhou. An analysis on recombination in multi-objective evolutionary
optimization. Artificial Intelligence, 204:99–119, 2013.
[33] C. Qian, Y. Yu, and Z.-H. Zhou. On constrained Boolean Pareto optimization. In Proceedings
of the 24th International Joint Conference on Artificial Intelligence (IJCAI’15), pages 389–395,
Buenos Aires, Argentina, 2015.
[34] C. Qian, J.-C. Shi, Y. Yu, K. Tang, and Z.-H. Zhou. Parallel Pareto optimization for subset selection. In Proceedings of the 25th International Joint Conference on Artificial Intelligence (IJCAI’16), pages 1939–1945, New York, NY, 2016.
19
[35] J. Reichel and M. Skutella. Evolutionary algorithms and matroid optimization problems. Algorithmica, 57(1):187–206, 2010.
[36] C. Witt. Worst-case and average-case approximations by simple randomized search heuristics. In Proceedings of the 22nd Annual Symposium on Theoretical Aspects of Computer Science
(STACS’05), pages 44–56, Stuttgart, Germany, 2005.
[37] Y. Yu, X. Yao, and Z.-H. Zhou. On the approximation ability of evolutionary optimization with
application to minimum set cover. Artificial Intelligence, 180:20–33, 2012.
20
| 9cs.NE
|
Ray: A Distributed Framework for Emerging AI Applications
arXiv:1712.05889v1 [cs.DC] 16 Dec 2017
Philipp Moritz∗, Robert Nishihara,∗ Stephanie Wang, Alexey Tumanov, Richard Liaw,
Eric Liang, William Paul, Michael I. Jordan, Ion Stoica
UC Berkeley
Abstract
with learning to operate continuously within an uncertain environment [44]. RL-based applications have already led to remarkable results, such as Google’s AlphaGo beating a human world champion [43], and are
finding their way into self-driving cars, UAVs [33], and
robotic manipulation [23, 47].
There are three characteristics that distinguish RL applications from traditional supervised learning applications. First, they often rely heavily on simulations to
explore states and discover the consequences of actions.
A simulator could encode the rules of a computer game
[11], the Newtonian dynamics of a physical system such
as a robot [45], or the hybrid dynamics of virtual environments. This generally requires massive amounts of
computation; for example, a realistic application might
perform hundreds of millions of simulations. Second,
the computation graph of an RL application is heterogeneous and evolves dynamically. A simulation can take
from a few milliseconds to a few minutes, and the result of a simulation can determine the parameters of future simulations. Third, many RL applications, such as
robotic control or autonomous driving, require actions
to be taken quickly in response to a constantly changing
environment. Furthermore, to choose the best action, the
application may need to perform more simulations in real
time. In summary, we need a computation framework
that supports heterogeneous and dynamic computation
graphs, while handling millions of tasks per second with
millisecond-level latencies.
Existing cluster computing frameworks fall short
of adequately satisfying these requirements. MapReduce [18], Apache Spark [50], Dryad [25], Dask [38],
and CIEL [32] support neither the throughput nor the latencies required by general RL applications, while TensorFlow [5], Naiad [31], MPI [21], and Canary [37] generally assume static computation graphs.1
The next generation of AI applications will continuously
interact with the environment and learn from these interactions. These applications impose new and demanding systems requirements, both in terms of performance
and flexibility. In this paper, we consider these requirements and present Ray—a distributed system to address
them. Ray implements a dynamic task graph computation model that supports both the task-parallel and the
actor programming models. To meet the performance
requirements of AI applications, we propose an architecture that logically centralizes the system’s control state
using a sharded storage system and a novel bottom-up
distributed scheduler. In our experiments, we demonstrate sub-millisecond remote task latencies and linear
throughput scaling beyond 1.8 million tasks per second.
We empirically validate that Ray speeds up challenging
benchmarks and serves as both a natural and performant
fit for an emerging class of reinforcement learning applications and algorithms.
1
Introduction
Artificial intelligence is currently emerging as the
workhorse technology for a range of real-world applications [27]. To date, however, these applications have
largely been based on a fairly restricted supervised learning paradigm in which a model is trained offline and
deployed to serve predictions online. As the field matures, it will be necessary to consider a broader setting
than standard supervised learning. Instead of making
and serving a single prediction, machine learning applications must increasingly operate in dynamic environments, react to changes in the environment, and take sequences of actions to accomplish a goal [6, 34]. These
broader requirements are naturally framed within the
paradigm of reinforcement learning (RL), which deals
∗ equal
1 While deep learning frameworks like TensorFlow or MXNet are
critical components of RL systems, they are not sufficient on their own.
Existing distributed RL applications [3, 20, 39] that use these systems
contribution
1
In this paper, we propose Ray, a cluster computing
framework that satisfies these requirements. To support
the heterogeneous and dynamic workloads imposed by
these applications, Ray implements a dynamic task graph
computation model, similar to CIEL [32]. However, Ray
also provides an actor programming abstraction on top
of this execution model, in addition to the task-parallel
abstraction provided by CIEL. The actor abstraction enables Ray to support stateful components, such as thirdparty simulators.
To achieve the stringent performance targets while
supporting dynamic computation graphs, Ray employs
a new distributed architecture that is horizontally scalable. The architecture is based on two key ideas. First,
we store all of the control state of the system in a global
control store, which enables all of the other components
in the system to be stateless. As a result, each component
can easily be scaled horizontally and restarted in the case
of failures. In turn, the global control store can be scaled
via sharding and made fault tolerant via replication.
Second, we introduce a new bottom-up distributed
scheduler, where tasks are submitted by workers and
drivers to local schedulers (there is one local scheduler
per node). Local schedulers may choose to schedule
tasks locally or to forward tasks to a replicated global
scheduler. This decreases task latency by allowing local
decisions, and increases the system throughput by reducing the burden on the global scheduler. We make the
following contributions:
simple RL application to illustrate the key requirements
for Ray.
An RL system consists of an agent that interacts repeatedly with the environment (see Figure 1(a)). The
goal of the agent is to learn a policy that maximizes some
reward. A policy is a mapping from the state of the environment to an action to take. The definitions of environment, agent, state, action, and reward are applicationspecific (Table 1).
Figure 2 shows an example of the pseudocode used
by an agent to learn a policy. A typical procedure consists of two steps: (1) evaluate the current policy, and
(2) improve the policy. To evaluate the policy, the pseudocode invokes rollout(environment, policy) to generate a set of rollouts, where a rollout is a trajectory of
states and rewards collected by interacting with the environment using environment.step(action). An action is
computed given the current policy and environment state
via policy.compute(state). As trajectories are generated,
train policy() uses completed trajectories to improve the
current policy via policy.update(trajectories). This process is repeated until the policy converges.
While simple, this application illustrates the key requirements of emerging AI applications. We group these
requirements into three categories.
Flexibility. The flexibility of a system is typically
measured in terms of the diversity of the workloads it can
support. We consider two aspects of flexibility: the heterogeneity of concurrently executing tasks and the generality and dynamicity of the execution graph.
Concurrent, heterogeneous tasks. Concurrent tasks
can be heterogeneous along three dimensions:
• We specify the systems requirements for emerging
AI applications: support for (a) heterogeneous, concurrent computations, (b) dynamic task graphs, (c)
high-throughput and low-latency scheduling, and
(d) transparent fault tolerance.
• Functionality. In the case of a robot, assessing the
environment’s state (e.g., environment.step(action))
involves processing the inputs of multiple sensors,
such as video, microphone, and radar. This requires
running multiple tasks in parallel, each implementing a different computation (see Figure 1(b)).
• We provide the actor abstraction–in addition to the
task-parallel programming abstraction–on top of a
dynamic task graph computation model.
• Duration. The time it takes to compute a trajectory
can vary significantly (see rollout(policy, environment)). For example, in the case of a game, it could
take just a few actions (moves) to lose the game, or
it could take hundreds of moves to win it.
• We propose a horizontally scalable architecture to
meet the above requirements, and build Ray, a cluster computing system that realizes this architecture.
2
• Resource types. Computing an action by evaluating
the policy (e.g., policy.compute(state)) is in many
cases implemented by a deep neural network, which
typically requires the use of GPUs. On the other
hand, most of the other computations use CPUs.
Motivation and Requirements
While Ray can support a variety of workloads—as it provides both the task-parallel and the actor abstractions—
we focus on RL workloads in this paper, as they are representative of emerging AI applications, and were the
primary driver behind Ray’s design. Here, we consider a
Note that these requirements are not naturally satisfied by
the Bulk Synchronous Parallel (BSP) model [46], which
is implemented by many of today’s popular cluster computing frameworks [18, 50]. With BSP, all tasks within
rely on ad-hoc solutions or other distributed frameworks to implement
task-parallelism and fault tolerance and to integrate stateful simulators.
2
(a)
(b)
(c)
Figure 1: (a) An RL system. (b) Task graph for processing sensor inputs. (c) Task graph for learning policy.
Application
Go (game)
Atari Pong
Robot moving an object
Environment
Board and opponent
Program
Physical world
Agent
Player
Player
Control program
State
Board position
Sequence of video frames
Object and robot pose
Action
Place a piece
Control joystick
Actuate joints
Reward
Game outcome
Pong score
Object moved
Table 1: Example RL applications.
// generate a trajectory under a given policy
rollout(policy, environment):
trajectory ← []
state ← environment.initial state()
while (not environment.has terminated()):
action ← policy.compute(state)
state, reward ← environment.step(action)
trajectory.append(state, reward)
return trajectory
the environment and compute a new action in a matter
of milliseconds. Similarly, simulations might take on
the order of milliseconds as well. As such, we need to
be able to schedule tasks in less than one millisecond.
Otherwise the scheduling overhead could be significant.
Given that clusters with tens of thousands of cores are
common, we need to be able to schedule hundreds of
thousands or even millions of tasks per second. Consider a cluster consisting of 100 servers, each with 32
cores, and suppose each task takes 5ms to execute. To
fully utilize this cluster, we would need to schedule 640K
tasks/sec.
// learn a policy in a given environment
train policy(environment):
policy ← initial policy()
while (policy has not converged):
trajectories ← []
// generate k rollouts and use them to update policy
for i from 1 to k:
trajectories.append(rollout(policy, environment))
policy = policy.update(trajectories)
return policy
Ease of development. Since writing parallel applications is non-trivial, and since ML developers prefer to
focus on their applications rather than on systems programming, simplifying development is paramount for the
success of such a system.
Deterministic replay and fault tolerance. The ability
to deterministically replay a job dramatically simplifies
debugging. Transparent fault tolerance obviates the need
for users to handle faults explicitly. It also enables users
to use cheap preemptible resources (e.g., spot instances
on AWS), leading to substantial cost savings when running in a public cloud.
Easy parallelization of existing algorithms. This involves providing a simple API and supporting existing
languages, tools, and libraries. First, we need to provide support for Python, as Python is the language of
choice for AI developers. Second, we need to provide
tight integration with the wide range of available thirdparty libraries. These libraries include simulators such as
OpenAI gym [13], DeepMind Lab [10], and the Mujoco
physics simulator [45] as well as deep learning frameworks like TensorFlow [5], Theano [12], PyTorch [4],
and Caffe [26]. As we will see, this requires augmenting
the task-parallel model with an actor-like abstraction to
wrap these third-party services.
Figure 2: Pseudocode for a typical RL training application.
the same stage2 typically perform the same computations
(albeit on different data partitions) and take roughly the
same amounts of time.3
Dynamic task graphs. Consider the train policy()
function. Though not shown in Figure 2, it would be natural to update the policy as soon as a subset of rollouts
finish (instead of waiting for all of them) and to launch
new rollouts to maintain a pool of executing rollouts as in
Figure 1(c). This makes the execution graph dynamic, as
we cannot predict the order in which rollouts will complete or which rollouts will be used for a particular policy
update.
Performance. In the case of a robot interacting with
the physical environment, we need to infer the state of
2A
stage is the unit of parallelism in BSP.
straggler mitigation techniques depend on this assumption [8].
3 Many
3
3
Programming and Computation Model
Ray provides basic support for stateful components in
the form of actors. In Ray, an actor is a stateful process
that exposes a set of methods that can be invoked as remote functions and that executes these methods serially.
Ray implements a dynamic task graph computation
model. On top, Ray provides both an actor and a taskparallel programming abstraction. This dual abstraction
differentiates Ray from related systems, such as CIEL
which only provides a task-parallel abstraction, and Orleans, which primarily provides an actor abstraction [14].
3.1
3.2
Computation Model
Ray employs a dynamic task graph computation
model [19], in which the execution of both remote functions and actor methods is automatically triggered by
the system when their inputs become available. In
this section, we describe how the computation graph
(Figure 3(b)) is constructed from a user program (Figure 3(a)). This program uses the API in Table 2 to implement the pseudocode from Figure 2.
Ignoring actors first, there are two types of nodes in
a computation graph: data objects and remote function
invocations, or tasks. There are also two types of edges:
data edges and control edges. Data edges capture the
dependencies between data objects and tasks. More precisely, if data object D is an output of task T , we add a
data edge from T to D. Similarly, if D is an input to T ,
we add a data edge from D to T . Control edges capture
the computation dependencies that result from nested remote functions (Section 3.1): if task T1 invokes task T2 ,
then we add a control edge from T1 to T2 .
Actor method invocations are also represented as
nodes in the computation graph. They are identical to
tasks with one key difference. To capture the state dependency across subsequent method invocations on the
same actor, we add a third type of edge: a stateful edge.
If method M j is called right after method Mi on the same
actor, then we add a stateful edge from Mi to M j . Thus,
all methods invoked on the same actor object form a
chain connected by stateful edges (Figure 3(b)). This
chain captures the order in which these methods were
invoked.
Stateful edges help us embed actors in an otherwise
stateless task graph, as they capture the implicit data dependency between successive method invocations sharing the internal state of an actor. Stateful edges also enable us to maintain lineage. As in other dataflow systems [50], we track data lineage to enable reconstruction. By explicitly including stateful edges in the lineage graph, we can easily reconstruct lost data, whether
produced by remote functions or actor methods (Section 4.2.3).
Programming Model and API
At its core, Ray provides a task-parallel programming
model. Table 2 shows Ray’s API. When a remote function is invoked, a future that represents the result of the
task is returned immediately. Futures can be retrieved using ray.get() and composed, i.e., a future can be passed
as an argument into another remote function. This allows the user to express parallelism while capturing data
dependencies.
Remote functions operate on immutable objects, and
are expected to be stateless and side-effect free: their
outputs are determined solely by their inputs. This
implies idempotence, which simplifies fault tolerance
through function re-execution on failure.
To satisfy the requirements for heterogeneity, flexibility, and ease of development given in Section 2, we augment the task-parallel programming model in four ways.
First, to handle concurrent tasks with heterogeneous
durations, we introduce ray.wait(). This call takes in
a list of futures and returns the subset whose results
are available, either after a timeout or when at least k
are available. In contrast, ray.get() blocks until all requested futures are available. This is highly beneficial
for RL applications, as simulations may have widely different durations, but complicates fault tolerance due to
introduced nondeterminism.
Second, to handle resource-heterogeneous tasks, we
enable developers to specify resource requirements so
that the Ray scheduler can efficiently manage resources.
The resources specified for a remote function are only
allocated during the function’s execution.
Third, to improve flexibility, we enable nested remote functions, meaning that remote functions can invoke other remote functions. This is also critical for
achieving high scalability (see Section 4), as it enables
multiple processes to invoke remote functions in parallel (otherwise the driver becomes a bottleneck for task
invocations).
Finally, and most importantly, for ease of development
and efficiency, we enhance our programming model with
an actor abstraction. One limitation we encountered
early in our development with stateless tasks was the
inability to wrap third-party simulators, which do not
expose their internal state. To address this limitation,
4
Architecture
4.1
Application Layer
The application layer consists of three components:
• Driver: A process executing the user program.
4
Name
futures list = f.remote(args)
obj list = ray.get(futures list)
futures list done =
ray.wait(futures list, k, timeout)
actor = Class.remote(args)
futures list = actor.method.remote(args)
Description
Execute function f () remotely. f () takes either object values or futures as arguments,
and returns a list of futures. This is a non-blocking call.
Return the values associated with a list of futures. This is a blocking call.
Given a list of futures, return the futures whose corresponding tasks have completed
as soon as either k of the tasks have completed or the timeout expires.
Instantiate class Class as a remote actor, and return a reference to it. Call a method
on the remote actor and return a list of futures. This is a non-blocking call.
Table 2: Ray API
T0
train_policy
@ray.remote
def create_policy():
# Initialize the policy randomly.
return policy
@ray.remote(num_gpus=2)
def update_policy(policy, *rollouts):
# Update the policy.
return policy
@ray.remote(num_gpus=1)
@ray.remote
class Simulator(object):
def train_policy():
def __init__(self):
# Create a policy.
# Initialize the environment.
policy_id = create_policy.remote()
self.env = Environment()
# Create 10 actors.
def rollout(self, policy, num_steps):
simulators = [Simulator.remote() for _ in range(10)]
observations = []
# Do 100 steps of training.
observation = self.env.current_state()
for _ in range(100):
for _ in range(num_steps):
# Perform one rollout on each actor.
action = compute(policy, observation)
rollout_ids = [s.rollout.remote(policy) for s in simulators]
observation = self.env.step(action)
# Update the policy with the rollouts.
observations.append(observation)
policy_id = update_policy.remote(policy_id,
return observations
*rollout_ids)
return ray.get(policy_id)
A10
Simulator
T1
create_policy
A20
Simulator
A11
rollout
policy1
A21
rollout
rollout21
rollout11
T2
update_policy
policy2
A12
rollout
rollout22
rollout12
…
object
data edges
A22
rollout
T3
update_policy
…
control edges
(a)
…
task/method
stateful edges
(b)
Figure 3: (a) Python code implementing the example in Figure 2 in Ray. Note that @ray.remote indicates remote functions and
actors. Invocations of remote functions and actor methods return futures, which can be passed to subsequent remote functions or
actor methods to encode task dependencies. Each actor has an environment object self.env shared between all of its methods. (b)
The task graph corresponding to an invocation of train policy.remote(). Remote function calls and the actor method calls correspond
to tasks in the task graph. The figure shows two actors. The method invocations for each actor (the tasks labeled A1i and A2i ) have
stateful edges between them indicating that they share the mutable actor state. There are control edges from train policy to the tasks
that it invokes. To train multiple policies in parallel, we could call train policy.remote() multiple times.
• Worker: A stateless process that executes tasks
(remote functions) invoked by a driver or another
worker. Workers are started automatically and assigned tasks by the system layer. When a remote
function is declared, the function is automatically
published to all workers. A worker executes tasks
serially.
contrast, an actor is a stateful process in that the results of
a method invocation can depend on the previous methods
executed by that actor.
4.2
System Layer
The system layer enables us to meet the performance and
fault tolerance goals, as discussed in Section 2, by employing an architecture in which each component is horizontally scalable and fault-tolerant. This layer consists
of three major components: a global control store, a distributed scheduler, and a distributed object store.
• Actor: A stateful process that executes, when invoked, the methods it exposes. Unlike a worker,
an actor is explicitly instantiated by a worker or a
driver. Like workers, actors execute methods serially.
Note that workers are stateless in that they do not
maintain local state across tasks. Assuming deterministic execution, invoking the same remote function with
the same arguments will return the same results no matter whether it is executed on the same worker or not. In
4.2.1
Global Control Store (GCS)
At its core, our system leverages a global control store
(GCS), which stores all up-to-date metadata and control
state information in the system. This includes (1) a speci5
Figure 5: Bottom-up distributed scheduler. Tasks are submitted bottom-up, from drivers and workers to a local scheduler and forwarded to the global scheduler only if needed (Section 4.2.2). The thickness of each arrow is proportional to its
request rate.
Figure 4: Ray’s architecture consists of two parts: an application layer and a system layer. The application layer implements
the API and the computation model described in Section 3, the
system layer implements task scheduling and data management
to satisfy the performance and fault-tolerance requirements.
schedulers (e.g., Canary [37]); (3) parallel scheduling,
where multiple global schedulers schedule tasks concurrently on all worker nodes (e.g., Sparrow [36]). Unfortunately, none of these approaches addresses Ray’s demands. Batch scheduling still requires the global scheduler to handle every task, which limits its scalability, hierarchical scheduling assumes the task graph is known
in advance (i.e., the graph is static), and parallel scheduling assumes that every global scheduler schedules independent jobs. In contrast, we require a highly scalable
scheduler to handle dynamic task graphs, possibly generated by a single job.
Like existing hierarchical scheduling solutions, we
employ a global scheduler and per-node local schedulers.
However, unlike prior solutions, tasks created on a node
are submitted to the node’s local scheduler first, not to a
global scheduler (Figure 5). A local scheduler schedules
tasks locally, unless the node is overloaded, or it cannot satisfy a task’s requirements (e.g., lacks GPU), or
a task’s inputs are remote. If a local scheduler doesn’t
schedule a task, it sends the task to the global scheduler. To determine the load, the local scheduler checks
the current length of its task queue. If the length exceeds some configurable threshold, it concludes that the
local node is overloaded. The value of this threshold enables the scheduling policy to span the continuum from
centralized, when all tasks are handed over to the global
scheduler, to decentralized, when all tasks are handled
locally.
Each local scheduler sends periodic heartbeats (e.g.,
every 100ms) to the GCS containing its load information. The GCS records this information and forwards
it to the global scheduler. Upon receiving a task, the
global scheduler uses the latest load information from every node along with the locations and sizes of the task’s
inputs (from the GCS’s object metadata) to decide which
node to assign the task to. If the global scheduler be-
fication of every task, (2) the code for every remote function, (3) the computation graph, (4) the current locations
of all objects, and (5) every scheduling event. The GCS
also provides publish-subscribe infrastructure to facilitate communication between components.
By storing and managing the entire control state in a
centralized fashion, the GCS enables every other component to be stateless. This not only simplifies the support
for fault tolerance (i.e., on failure, components restart
and read the latest state from the GCS), but also makes it
easy to horizontally scale every other component, as all
the state shared by the component’s replicas or shards is
accessible via the GCS.
To scale the GCS, we use sharding. Since we can associate pseudo-random IDs to virtually every data entry
in the GCS (e.g., objects, tasks, functions), it is relatively
easy to balance the load across multiple shards. To provide fault-tolerance, we use a hot replica for each shard.
Centralizing the system control information allows us
to easily build debugging, profiling, and visualization
tools on top of the GCS. The minimalist tools we’ve built
so far have already proven useful in our development.
4.2.2
Bottom-Up Distributed Scheduler
Many existing cluster computing frameworks, such
as Apache Spark [50], CIEL [32], Dryad [25] and
Hadoop [49], implement a centralized scheduler. While
this simplifies the design, it hurts scalability.
There are several approaches to improve scheduling
scalability: (1) batch scheduling, where the scheduler
submits tasks to worker nodes in batches to amortize
fixed overheads related to task submission (e.g., Drizzle [48]); (2) hierarchical scheduling, where a global
scheduler partitions the task graph across per-node local
6
N1
comes a bottleneck, we can instantiate more replicas and
have each local scheduler randomly pick a replica to send
its tasks to. This makes our scheduler architecture highly
scalable.
4.2.3
@ray.remote
def add(a, b):
return a + b
0
idC = add.remote(a, b)
c =ray.get(id)
ida
idb
1
@ray.remote
def add(a, b):
return a + b
Object Table
Object store
9
N1
N2
Object store
6
7
ida a
To minimize task latency, we implement an in-memory
distributed storage system to store the inputs and outputs
of every task. This allows workers and actors to share
data efficiently. On each node, we implement the object store via shared memory. This allows zero-copy data
sharing between tasks running on the same node. Additionally, we use Apache Arrow [1], an efficient memory
layout that is becoming the de facto standard in data analytics.
If a task’s inputs are not local, the inputs are replicated to the local object store on the same node before
execution. Tasks also write all outputs to the local object
store. Replication eliminates the potential bottleneck due
to hot data objects and minimizes task execution time as
a task only reads and writes data in local memory. This
increases throughput for computation-bound workloads,
a profile shared by many AI applications.
As with existing cluster computing frameworks, such
as Apache Spark [50], and Dryad [25], the object store
is limited to immutable data, which significantly simplifies the system design by obviating the need for complex
consistency protocols (as it avoids concurrent updates),
and by simplifying support for fault tolerance.
For simplicity, our object store does not build in support for distributed objects, that is, each object fits on a
single node. Distributed objects like large matrices or
trees can be implemented at a higher level (e.g., the application level) as collections of futures.
Worker
Function Table
@ray.remote
def add(a, b):
return a + b
In-Memory Distributed Object Store
N2
Global Control State (GCS)
Driver
3
Local Scheduler
2
Global Scheduler
8
ida a idb b
5
Local Scheduler
4
(a) Executing a task remotely
N1
N2
Global Control State (GCS)
Driver
Worker
Function Table
@ray.remote
def add(a, b):
return a + b
@ray.remote
def add(a, b):
return a + b
idC = add.remote(a, b)
c =ray.get(id)
7
Object Table
2
1
ida a idc c
@ray.remote
def add(a, b):
return a + b
5
ida
idb
idc
N1
N2
N2, N1
3
4
idc c ida a idb b
6
Local Scheduler
Global Scheduler
Local Scheduler
(b) Returning the result of a remote task
Figure 6: An end-to-end example that adds a and b and returns
c. Solid lines are data plane operations and dotted lines are
control plane operations. (a) The function add() is registered
with the GCS by node 1 (N1), invoked on N1, and executed
on N2. (b) N1 gets add()’s result using ray.get(). The Object
Table entry for c is created in step 4 and updated in step 6 after
c is copied to N1.
been lost. By walking backwards along data and stateful
edges we reach A10 which has no input.4 Thus, to reconstruct rollout12 , we need to re-instantiate the actor by
executing A10 , and then executing the methods A11 and
A12 in order.
Note that for any object whose lineage includes stateful edges, reconstruction will require reinstantiating the
actor (e.g., A10 ) and replaying a possibly long chain of
stateful edges (e.g., A11 , A12 , etc). Since actors are commonly used to wrap third-party simulators, which have
a finite lifetime, we expect these chains to be bounded.
However, we’ve also found actors to be useful for managing more general forms of state. To improve recovery
time in such cases, we checkpoint the actor’s state periodically and allow the actor to recover from checkpoints.
For low latency, we keep objects entirely in memory
and evict them as needed to disk using a least-recentlyused eviction policy.
In Section A, we briefly explain how our design satisfies the requirements outlined in Section 2.
Object reconstruction. Component failure may result in object loss, which Ray recovers from through
lineage re-execution. Ray tracks lineage by recording
task dependencies in the GCS during execution. This is
similar to the solution employed by other cluster computing systems like Apache Spark [50] and CIEL [32].
Also, like these systems, Ray assumes that objects are
immutable and operators (i.e., remote functions and actor methods) are deterministic. However, unlike these
systems, Ray adds support for stateful operator (i.e., actor) reconstruction. By integrating stateful edges directly
into the computation graph, we can leverage the same reconstruction mechanism for both remote functions and
actors.
To reconstruct a lost object, we walk backward along
data and stateful edges until we find tasks whose inputs
are all present in the object store. We then replay the
computation subgraph rooted at these inputs. Consider
the example in Figure 3(b), and assume that rollout12 has
4 Since it has no input, all of its inputs are trivially in the object store.
7
4.3
Putting Everything Together
The bottom-up distributed scheduler (Section 4.2.2) is
3.2KLoC and will undergo significant development as
we refine Ray’s scheduling policies. In this section,
we focus on the implementation details for achieving
the performance targets dictated by real-time AI applications: (a) scheduler performance, (b) object store performance, and (c) end-to-end system performance.
Figure 6 illustrates how Ray works end-to-end with a
simple example that adds two objects a and b, which
could be scalars or matrices, and returns result c. The remote function add() is automatically registered with the
GCS upon initialization and distributed to every worker
in the system (step 0 in Figure 6(a)).
Figure 6(a) shows the step-by-step operations triggered by a driver invoking add.remote(a, b), where a
and b are stored on nodes N1 and N2, respectively. The
driver submits add(a, b) to the local scheduler (step 1),
which forwards it to a global scheduler (step 2).5 Next,
the global scheduler looks up the locations of add(a, b)’s
arguments in the GCS (step 3) and decides to schedule
the task on node N2, which stores argument b (step 4).
The local scheduler at node N2 checks whether the local object store contains add(a, b)’s arguments (step 5).
Since the local store doesn’t have object a, it looks up
a’s location in the GCS (step 6). Learning that a is stored
at N1, N2’s object store replicates it locally (step 7). As
all arguments of add() are now stored locally, the local
scheduler invokes add() at a local worker (step 8), which
accesses the arguments via shared memory (step 9).
Figure 6(b) shows the step-by-step operations triggered by the execution of ray.get() at N1, and of add()
at N2, respectively. Upon ray.get(idc )’s invocation, the
driver checks the local object store for the value c, using
the future idc returned by add() (step 1). Since the local
object store doesn’t store c, it looks up its location in the
GCS. At this time, there is no entry for c, as c has not
been created yet. As a result, N1’s object store registers
a callback with the Object Table to be triggered when c’s
entry has been created (step 2). Meanwhile, at N2, add()
completes its execution, stores the result c in the local
object store (step 3), which in turn adds c’s entry to the
GCS (step 4). As a result, the GCS triggers a callback to
N1’s object store with c’s entry (step 5). Next, N1 replicates c from N2 (step 6), and returns c to ray.get() (step
7), which finally completes the task.
While this example involves a large number of RPCs,
in many cases this number is much smaller, as most tasks
are scheduled locally, and the GCS replies are cached by
the global and local schedulers.
5
Bottom-up distributed scheduler. We implement
both the local and global scheduler as event-driven,
single-threaded processes. Internally, local schedulers
maintain cached state for local object metadata, tasks
waiting for inputs, and tasks ready for dispatch to a
worker. As object dependencies become available, tasks
become ready for dispatch. Worker availability triggers
dispatch of as many tasks as possible under the node’s
capacity constraints.
The local scheduler sends periodic heartbeats (every
100ms) to the global schedulers, routed through the GCS
via a publish-subscriber mechanism, containing the dispatch queue length and resource availability. This enables the global scheduler to balance the load across
nodes.
Object store. Ray’s object store is also implemented
as a single-threaded event loop. It uses shared memory
so workers on the same node can read data without copying it. Objects are immutable. An object is only made
visible after a worker finishes creation. To minimize object creation overhead, the store pre-allocates a pool of
large memory-mapped files. We use a SIMD-like memory copy to maximize the throughput of copying data
from a worker to the object store’s shared memory. We
also parallelize computation of an object’s content hash,
which is used to detect non-deterministic computations.
Ray uses Apache Arrow [1] to achieve high performance
when serializing/deserializing Python objects.
Global control store. We implement Ray’s Global
Control Store (GCS) using one Redis [40] key-value
store per shard (Redis can be easily swapped with other
key-value stores). We shard the GCS tables by object
and task IDs to scale, and we replicate every shard for
fault tolerance. As we scale up the experiment, we distribute the shards across multiple nodes. Though our
GCS implementation uses multiple Redis servers, our
performance and fault tolerance requirements could also
be met by existing systems like RAMCloud [35].
Implementation
Ray is implemented in ≈ 40K lines of code (LoC), 72%
in C++ for the system layer, 28% in Python for the
application layer. The object store and our zero-copy
serialization libraries have been factored out as standalone projects that can be used independently of Ray.
Lastly, Ray’s monitor tracks system component liveness and reflects component failures in the GCS. Tasks
and objects on failed cluster nodes are marked as lost,
and objects are later reconstructed with lineage information, as necessary.
5 Note
that the N1 local scheduler could also decide to schedule the
task locally.
8
tasks (thousands)
10 20 30 40 50 60
number of nodes
5
10
nodes
15
20
Figure 8: Ray maintains balanced load. A driver on the first
node submits 100K tasks, which are rebalanced by the global
scheduler across the 21 available nodes.
a linear fashion, leveraging the GCS and bottom-up distributed
scheduler. Ray reaches 1 million tasks per second throughput
with 60 m4.16xlarge nodes and processes 100 million tasks in
under a minute. We omit x ∈ {70, 80, 90} due to cost.
20000
15
IOPS
15000
Evaluation
10
10000
5
5000
In this section, we demonstrate three key points. First,
we examine the scalability of the system as a whole as
well as the performance of its individual components
(Section 6.1). Second, we demonstrate robustness and
fault tolerance (Section 6.2). Third, we demonstrate that
Ray is a natural fit for reinforcement learning applications, both in terms of performance and ease of development (Section 6.3).
All experiments were run on
Amazon Web Services. The specific instance types are
reported below.
6.1
executed tasks
submitted tasks
100
Figure 7: End-to-end scalability of the system is achieved in
6
100
80
60
40
20
00
0
throughput (GB/s)
tasks per second (millions)
1.8
1.6
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0
1KB 10KB 100KB 1MB 10MB 100MB 1GB
object size
0
Figure 9: Object store write throughput and IOPS. From a
single client, throughput exceeds 15GB/s (red) for large objects and 18K IOPS (cyan) for small objects on a 16 core instance (m4.4xlarge). It uses 8 threads to copy objects larger
than 0.5MB and 1 thread for small objects. Bar plots report
throughput with 1, 2, 4, 8, 16 threads. Results are averaged
over 5 runs.
Scalability and Performance
Object store performance. We track two metrics for
object store performance: IOPS (for small objects) and
write throughput (for large objects) in Figure 9. As object size increases, the write throughput from a single
client reaches 15GB/s. For larger objects, copying the
object from the client dominates the time spent on object
creation. For smaller objects, completion time is dominated by serialization overhead and IPC between the
client and object store. The object store peaks at 18K
IOPS, which corresponds to 56µs per operation.
End-to-end scalability. One of the key benefits of the
Global Control Store (GCS) is the ability to horizontally
scale the system. We evaluate this ability in this section. In Fig. 7, we benchmark an embarrassingly parallel
workload, increasing the cluster size on the x-axis. We
observe near-perfect linearity in progressively increasing
task throughput. Ray exceeds 1 million tasks per second
throughput at 60 nodes and continues to scale linearly
beyond 1.8 million tasks per second at 100 nodes. The
rightmost datapoint shows that Ray can process 100 million tasks in less than a minute (54s). Variability (shown
with black error bars) is minimal. As expected, increasing task duration reduces throughput proportionally to
mean task duration, but the overall scalability remains
linear.
The global scheduler’s primary responsibility is to
maintain balanced load throughout the system. In Figure 8, 100K tasks submitted on a single node are rebalanced across the available resources. Note that the
node where the load originates processes more tasks as it
maximizes utilization of the local node before forwarding tasks to the global scheduler.
6.2
Fault Tolerance
Recovering from object failures. In Figure 10, we
demonstrate Ray’s ability to transparently recover from
worker node failures and elastically scale. The driver
submits rounds of tasks where each task is dependent on
a task in the previous round. As worker nodes are killed
(at 25s, 50s, 100s), the surviving local schedulers automatically trigger reconstruction of the lost objects. During periods of reconstruction, the tasks originally submitted by the driver stall, since their dependencies cannot be
satisfied. However, the overall task throughput remains
9
Original tasks
Reexecuted tasks
0
50
100
150
Time since start (s)
200
method calls. The initial throughput is comparable to
that without checkpointing. After node failure, the majority of reconstruction is done by executing checkpoint
tasks to reconstruct the actor’s state (t = 210-270s). As
a result, only 500 tasks need to be re-executed, and new
method calls stall for 60s, versus 10K re-executions and
120s without checkpointing, respectively. In the future,
we hope to further reduce actor reconstruction time, e. g.
by allowing user annotations for read-only methods.
Overhead from GCS replication. To make the GCS
fault tolerant, we replicate each of the database shards.
When a client writes to one of the shards of the GCS,
it duplicates the writes to all replicas. For workloads in
which we artificially make the GCS the bottleneck by reducing the number of GCS shards, the overhead of twoway replication is less than 10%. In most real workloads
the slowdown is undetectable.
60
50
40
30
20
10
0
Number of nodes
Throughput (tasks/s)
2000
1500
1000
500
0
Figure 10: Fully transparent fault tolerance for distributed
tasks. The dashed line represents the number of nodes in the
cluster. Curves show the throughput of new tasks (cyan) and
re-executed tasks (red). The driver continually submits and retrieves rounds of 10000 tasks. Each task takes 100ms and depends on a task in the previous round. Each task has input and
output of size 100KB.
600
500
400
300
200
100
0
Original tasks
Reexecuted tasks
Throughput (tasks/s)
Throughput (tasks/s)
700
100 200 300 400
Time since start (s)
(a) without checkpointing
700
600
500
Original tasks
Reexecuted tasks
Checkpoint tasks
6.3
Given the diverse and demanding requirements of reinforcement learning applications described in Section 2, reinforcement learning algorithms today are implemented on top of special-purpose ad-hoc systems that
typically require substantial engineering effort to develop
and which do not generalize to other algorithms.
In this section, we implement two types of reinforcement learning algorithms in Ray and show that we are
able to match or outperform the performance of specialized systems built specifically for these algorithms. Furthermore, using Ray to distribute these algorithms over
clusters requires changing only a few lines of code in serial implementations of the algorithms.
In addition, we test Ray in a latency-sensitive setting
in which Ray is used to control a simulated robot under
varying real-time requirements.
400
300
200
100
0
RL Applications
100 200 300 400
Time since start (s)
(b) with checkpointing
Figure 11: Fully transparent fault tolerance for actor methods.
The driver continually submits tasks to the actors in the cluster.
At t = 200s, we kill 2 of the 10 nodes, causing 400 of the 2000
actors in the cluster to be recovered on the remaining nodes.
stable, fully utilizing the available resources until the lost
dependencies are reconstructed. Furthermore, as more
nodes are added back to the system at 210s, Ray is able
to fully recover to its initial throughput.
Recovering from actor failures. Next, we demonstrate Ray’s ability to transparently recover lost actors.
By encoding each actor’s method calls into the dependency graph, we can reuse the same object reconstruction mechanism as in Figure 10. The workload in Figure 11a demonstrates the extreme case where no intermediate actor state is saved. All previous method calls
for each lost actor must be re-executed serially (t = 210330s). Lost actors are automatically redistributed across
the available nodes, and throughput fully recovers after
reconstruction.
To improve reconstruction time for long-lived actors,
we provide transparent checkpointing of intermediate actor state. Figure 11b shows the same workload, but with
an automatic checkpoint task on each actor every 10
6.3.1
Evolution Strategies
To evaluate Ray on large-scale RL workloads, we implement the evolution strategies (ES) algorithm and compare to the reference implementation [39], which is a
special-purpose system built for this algorithm. It uses
a hierarchy of Redis servers as message buses and relies
on low-level multiprocessing libraries for sharing data.
As shown in Figure 12, a straightforward implementation on top of Ray is more scalable, scaling to 8192
physical cores, whereas the special-purpose system stops
running after 1024 cores. The Ray implementation runs
in a median time of 3.7 minutes, which is more than
twice as fast as the best published result (10 minutes).
The Ray implementation was also substantially simpler
to develop. Parallelizing a serial implementation using
Ray required modifying only 7 lines of code. In contrast,
10
Mean time to solve (minutes)
80
Reference ES
Ray ES
60
ing p2.16xlarge (GPU) and m4.16xlarge (high CPU) instances, each of which has 32 physical cores.
Ray’s API made it easy to take advantage of heterogeneous resources decreasing costs by a factor of 4.5
[2]. Ray tasks and actors can specify distinct resource
requirements, allowing CPU-only tasks to be scheduled
on cheaper high-CPU instances. In contrast, MPI applications often exhibit symmetric architectures, in which
all processes run the same code and require identical resources, in this case preventing the use of CPU-only machines for scale-out.
As shown in Figure 13, the Ray implementation outperforms the optimized MPI implementation in all experiments (hyperparameters listed in Section D) with a fraction of the GPUs. As with ES, we were able to parallelize
PPO using Ray with minimal changes to the structure of
the serial program.
40
20
0
256
512
x
1024
2048
Number of CPUs
x4096 x8192
Figure 12: Time to reach a score of 6000 in the Humanoid-
Mean time to solve (minutes)
v1 task [13]. The Ray ES implementation scales well to 8192
cores. The special-purpose system failed to run beyond 1024
cores. With 8192 cores, we achieve a median time of 3.7 minutes, which is over twice as fast as the best published result.
ES is faster than PPO (Section 6.3.2) on this benchmark but
exhibits greater runtime variance.
500
MPI PPO
Ray PPO
400
6.3.3
300
We show that Ray can meet soft real-time requirements
by controlling a simulated robot in real time. A Ray
driver runs the simulated robot and takes actions at fixed
time steps varying from 1 millisecond to 30 milliseconds
to simulate different real-time requirements. The driver
submits tasks that compute actions to be taken using a
policy that was trained offline. However, the actions are
only taken if they are received by the driver within the
relevant time step (otherwise the prior action is repeated).
Latency budgets for real robots are on the order of 10
milliseconds, and we find that even if we run the simulation faster than real time (using a 3 millisecond time
step), Ray is able to produce a stable walk. Table 3 shows
the fraction of tasks that did not arrive fast enough to be
used by the robot.
200
100
0
8x1
16x2
Controlling a Simulated Robot
32x4 64x8 128x16 256x32 512x64
CPUs x GPUs
Figure 13: Time to reach a score of 6000 in the Humanoid-v1
task [13]. The Ray PPO implementation outperforms a specialized MPI implementation [3] with fewer GPUs and at a fraction
of the cost. The MPI implementation required 1 GPU for every
8 CPUs, whereas the Ray version required at most 8 GPUs (and
never more than 1 GPU per 8 CPUs).
the reference implementation required several hundred
lines of code to develop a custom protocol for communicating tasks and data between workers and could not be
easily adapted to different algorithms or communication
patterns. We include pseudocode illustrating this point in
Section B.1.
Time budget (ms)
% actions dropped
Stable walk?
30
0
Yes
20
0
Yes
10
0
Yes
5
0
Yes
3
0.4
Yes
2
40
No
1
65
No
Table 3: Low-latency robot simulation results
6.3.2
Proximal Policy Optimization
To evaluate Ray on single node and small cluster RL
workloads, we implement Proximal Policy Optimization
(PPO) [41] in Ray and compare to a highly-optimized
reference implementation [3] that uses OpenMPI communication primitives.6 All experiments were run us-
7
Related Work
Dynamic task graphs. Ray is closely related to
CIEL [32]. They both support dynamic task graphs with
nested tasks, implement the futures abstraction, and provide lineage-based fault tolerance. However, they differ in two important aspects. First, Ray extends the task
model with an actor abstraction. Second, Ray employs
a fully distributed control plane and scheduler, instead
of relying on a single master. In addition, Ray adds the
ray.wait() method, employs an in-memory (instead of a
6 Both implementations use TensorFlow to define neural networks
but rely on the distributed frameworks (OpenMPI and Ray) for communication between processes. Though this algorithm could be implemented using TensorFlow alone, single-machine performance would be
slow due to the Python global interpreter lock and distributed versions
would look similar to MPI implementations (sending and receiving data
via queues).
11
file-based) object store, and extends an existing programming language (Python), while CIEL provides its own
scripting language (Skywriting). Ray is also closely related to Dask [38], which supports dynamic task graphs,
including a wait-like primitive, and employs the futures
abstraction in a Python environment. However, Dask
uses a centralized scheduler, doesn’t offer an actor-like
abstraction, and doesn’t provide fault tolerance.
Data flow systems. Popular dataflow systems, such
as MapReduce [18], Spark [51], and Dryad [25] have
widespread adoption for analytics and ML workloads,
but their computation model is more restrictive. Spark
and MapReduce implement the BSP execution model,
which assumes that tasks within the same stage perform
the same computation and take roughly the same amount
of time. Dryad relaxes this restriction but lacks support
for dynamic task graphs. Furthermore, none of these
systems provide an actor abstraction, nor implement a
distributed scalable control plane and scheduler. Finally,
Naiad [31] is a dataflow system that provides improved
scalability for some workloads, but only supports static
task graphs.
Actor systems. Orleans [14] provides a virtual actorbased abstraction. Actors are perpetual and their state
persists across invocations. For scaling out, Orleans also
allows multiple instances of an actor to run in parallel
when the actor operates on immutable state or has no
state. These stateless actors can act as tasks in Ray. However, unlike Ray, the Orleans developer must explicitly
checkpoint actor state and intermediate responses. In
addition, Orleans provides at-least-once semantics. In
contrast, Ray provides transparent fault tolerance and
exactly-once semantics, as each method call is logged in
the GCS and both arguments and results are immutable.
We find that in practice these limitations do not affect the
performance of our applications.
Erlang [9] and C++ Actor Framework (CAF) [16], two
other actor-based systems, also require the application to
explicitly handle fault tolerance. Also, Erlang’s global
state store is not suitable for sharing large objects such as
ML models, while CAF does not support data sharing.
Global control state and scheduling. The concept
of logically centralizing the control plane has been previously proposed in software defined networks (SDNs)
[15], distributed file systems (e.g., GFS [22]), resource
management (e.g., Omega [42]), and distributed frameworks (e.g., MapReduce [18], BOOM [7]), to name a
few. Ray draws inspiration from these pioneering efforts,
but provides significant improvements. In contrast with
SDNs, BOOM, and GFS which couple the control plane
data and computation, Ray decouples the storage of the
control plane information (e.g., GCS) from the logic implementation (e.g., schedulers). This allows both storage
and computation layers to scale independently, which is
key to achieving our scalability targets. Omega uses a
distributed architecture in which schedulers coordinate
via globally shared state. To this architecture, Ray adds
global schedulers to balance load across local schedulers,
and targets ms-level, not second-level, task scheduling.
Ray implements a unique distributed bottom-up
scheduler that is horizontally scalable, and can handle
dynamically constructed task graphs. Unlike Ray, most
existing cluster computing systems [18, 50, 32] use a
centralized scheduler architecture. While Sparrow [36]
is decentralized, its schedulers make independent decisions, limiting the possible scheduling policies, and all
tasks of a job are handled by the same global scheduler. Mesos [24] implements a two-level hierarchical
scheduler, but its top-level scheduler can be a bottleneck.
Canary [37] achieves impressive performance by having each scheduler instance handle a portion of the task
graph, but does not handle dynamic computation graphs.
Machine learning frameworks. TensorFlow [5] and
MXNet [17] target deep learning workloads and efficiently leverage both CPUs and GPUs. While they
achieve great performance for workloads consisting of
static DAGs of linear algebra operations, they have
limited support for more general workloads. TensorFlow Fold [29] provides some support for dynamic task
graphs, as well as MXNet through its internal C++ APIs,
but neither fully supports the ability to modify the DAG
during execution in response to task progress, task completion times, or faults. TensorFlow and MXNet in principle achieve generality by allowing the programmer to
simulate low-level message-passing and synchronization
primitives, but the pitfalls and user experience in this
case are similar to those of MPI. OpenMPI [21] can
achieve high performance, but it is relatively hard to program as it requires explicit coordination to handle heterogeneous and dynamic task graphs. Furthermore, it forces
the programmer to explicitly handle fault tolerance.
8
Discussion and Experiences
Since we released Ray several months ago, over one hundred people have downloaded and used it. Here we discuss our experience developing and using Ray, as well as
some of the feedback we’ve received from early users.
API. In designing the API, we have emphasized minimalism. Initially we started with a basic task abstraction. Later, we added the wait() primitive to accommodate rollouts with heterogeneous durations and the actor
abstraction to accommodate third-party simulators and
amortize the overhead of expensive initializations. While
the resulting API is relatively low-level, it has proven
both powerful and simple to use. Indeed, some teams
report instructing developers to first write serial implementations and then to parallelize them using Ray.
12
To illustrate this point, next we briefly describe our experience with two other algorithms: Asynchronous Advantage Actor Critic (A3C) and hyperparameter search.
A3C [30] is a state-of-the-art RL algorithm which
leverages asynchronous policy updates to significantly
improve training times over previous algorithms. To
scale out this algorithm, we use a simple hierarchical
scheme where multiple instances of A3C are trained
in parallel and periodically aggregated to form an improved model. Implementing hierarchical A3C in Ray
was straightforward, requiring 20 lines of Python code
to extend the non-hierarchical version. Furthermore,
this simple extension improved performance on the same
hardware by 30%.
We were able to implement a state-of-the-art hyperparameter search algorithm [28] in roughly 30 lines of
Python code using Ray. Ray’s support for nested tasks
was critical because multiple experiments had to be run
in parallel, and each experiment typically used parallelism internally. The wait() primitive allowed us to process the results of experiments in the order that they completed and to adaptively launch new ones. The actor abstraction allowed us to pause and resume stateful experiments based on the progress of other experiments (see
Section B.3). In contrast, most existing implementations
have to wait for all experiments in the round to complete,
leading to inefficient resource utilization.
Ray’s API is still a work in progress. Based on early
user feedback, we are considering enhancing the API to
include higher level primitives, such as simple aggregation and map. These could also inform scheduling decisions in the Ray system layer (Section 4.2).
Limitations. Given the workload generality, specialized optimizations are hard. For example, we must make
scheduling decisions without full knowledge of the computation graph. Scheduling optimizations in Ray might
require more complex runtime profiling. In addition,
storing lineage for each task requires the implementation
of garbage collection policies to bound storage costs in
the GCS, a feature we are actively developing.
Fault tolerance. We are often asked if fault tolerance
is really needed for AI applications. After all, due to
the statistical nature of many AI algorithms, one could
simply ignore failed rollouts. Based on our experience,
our answer is an unqualified “yes”. First, the ability to
ignore failures makes applications much easier to write
and reason about. Second, our particular implementation
of fault tolerance via deterministic replay dramatically
simplifies debugging as it allows us to easily reproduce
most errors. This is particularly important since, due to
their stochasticity, AI algorithms are notoriously hard to
debug. Third, fault tolerance helps save money since it
allows us to run on cheap resources like spot instances on
AWS. Furthermore, as workloads scale, we expect fault
tolerance to become even more important. Of course,
this comes at the price of some overhead. However, we
found this overhead to be minimal for our target workloads.
GCS and Horizontal Scalability. The GCS dramatically simplified Ray development and debugging. Basic failure handling and horizontal scaling for all other
components took less than a week to implement. The
GCS enabled us to query the entire system state while
debugging (instead of having to manually expose internal component state). This helped us find numerous bugs
and generally understand system behavior.
The GCS is instrumental to Ray’s horizontal scalability. In the experiments reported in Section 6.1, we were
able to scale the results by adding more shards whenever
the GCS became a bottleneck. The GCS also enables the
global scheduler to scale by simply adding more replicas.
While currently we are manually configuring the number
of GCS shards and global schedulers, we are planning to
develop adaptive algorithms in the future. Due to these
advantages, we believe that centralizing control state will
be a key design component of future distributed systems.
9
Conclusion
Emerging AI applications present challenging computational demands. To meet these demands, Ray introduces a global control store and a bottom-up distributed
scheduler. Together, this architecture implements dynamic task graph execution, which in turn supports both
a task-parallel and an actor programming model. This
programming flexibility is particularly important for RL
workloads, which produce tasks diverse in resource requirements, duration, and functionality. Our evaluation
demonstrates linear scalability past 1M tasks per second,
transparent fault tolerance, and substantial performance
improvements on several contemporary RL workloads.
Thus, Ray provides a powerful combination of flexibility, performance, and ease of use for the development of
future AI applications.
References
[1] Apache Arrow. https://arrow.apache.org/.
[2] EC2 Instance Pricing.
pricing/on-demand/.
https://aws.amazon.com/ec2/
[3] OpenAI Baselines: high-quality implementations of reinforcement learning algorithms. https://github.com/openai/
baselines.
[4] PyTorch: Tensors and dynamic neural networks in python with
strong gpu acceleration. http://pytorch.org/.
[5] A BADI , M., BARHAM , P., C HEN , J., C HEN , Z., DAVIS , A.,
D EAN , J., D EVIN , M., G HEMAWAT, S., I RVING , G., I SARD ,
13
M., ET AL . TensorFlow: A system for large-scale machine learning. In Proceedings of the 12th USENIX Symposium on Operating
Systems Design and Implementation (OSDI). Savannah, Georgia,
USA (2016).
[19] D ENNIS , J. B., AND M ISUNAS , D. P. A preliminary architecture for a basic data-flow processor. In Proceedings of the 2Nd
Annual Symposium on Computer Architecture (New York, NY,
USA, 1975), ISCA ’75, ACM, pp. 126–132.
[6] AGARWAL , A., B IRD , S., C OZOWICZ , M., H OANG , L., L ANG FORD , J., L EE , S., L I , J., M ELAMED , D., O SHRI , G., R IBAS ,
O., S EN , S., AND S LIVKINS , A. A multiworld testing decision
service. arXiv preprint arXiv:1606.03966 (2016).
[20] D UAN , Y., C HEN , X., H OUTHOOFT, R., S CHULMAN , J., AND
A BBEEL , P. Benchmarking deep reinforcement learning for continuous control. In Proceedings of the 33rd International Conference on International Conference on Machine Learning - Volume
48 (2016), ICML’16, JMLR.org, pp. 1329–1338.
[7] A LVARO , P., C ONDIE , T., C ONWAY, N., E LMELEEGY, K.,
H ELLERSTEIN , J. M., AND S EARS , R. BOOM Analytics: exploring data-centric, declarative programming for the cloud. In
Proceedings of the 5th European conference on Computer systems (2010), ACM, pp. 223–236.
[21] G ABRIEL , E., FAGG , G. E., B OSILCA , G., A NGSKUN , T.,
D ONGARRA , J. J., S QUYRES , J. M., S AHAY, V., K AMBADUR ,
P., BARRETT, B., L UMSDAINE , A., C ASTAIN , R. H., DANIEL ,
D. J., G RAHAM , R. L., AND W OODALL , T. S. Open MPI:
Goals, concept, and design of a next generation MPI implementation. In Proceedings, 11th European PVM/MPI Users’ Group
Meeting (Budapest, Hungary, September 2004), pp. 97–104.
[8] A NANTHANARAYANAN , G., K ANDULA , S., G REENBERG , A.,
S TOICA , I., L U , Y., S AHA , B., AND H ARRIS , E. Reining in
the outliers in map-reduce clusters using mantri. In Proceedings
of the 9th USENIX Conference on Operating Systems Design and
Implementation (Berkeley, CA, USA, 2010), OSDI’10, USENIX
Association, pp. 265–278.
[22] G HEMAWAT, S., G OBIOFF , H., AND L EUNG , S.-T. The Google
file system. In ACM SIGOPS operating systems review (2003),
vol. 37, ACM, pp. 29–43.
[9] A RMSTRONG , J., V IRDING , R., W IKSTR ÖM , C., AND
W ILLIAMS , M. Concurrent programming in ERLANG.
[23] G U *, S., H OLLY *, E., L ILLICRAP, T., AND L EVINE , S.
Deep reinforcement learning for robotic manipulation with asynchronous off-policy updates. In IEEE International Conference
on Robotics and Automation (ICRA 2017) (2017).
[10] B EATTIE , C., L EIBO , J. Z., T EPLYASHIN , D., WARD , T.,
WAINWRIGHT, M., K ÜTTLER , H., L EFRANCQ , A., G REEN , S.,
VALD ÉS , V., S ADIK , A., ET AL . Deepmind lab. arXiv preprint
arXiv:1612.03801 (2016).
[24] H INDMAN , B., KONWINSKI , A., Z AHARIA , M., G HODSI ,
A., J OSEPH , A. D., K ATZ , R., S HENKER , S., AND S TOICA ,
I. Mesos: A platform for fine-grained resource sharing in the
data center. In Proceedings of the 8th USENIX Conference on
Networked Systems Design and Implementation (Berkeley, CA,
USA, 2011), NSDI’11, USENIX Association, pp. 295–308.
[11] B ELLEMARE , M. G., NADDAF, Y., V ENESS , J., AND B OWL ING , M. The arcade learning environment: An evaluation platform for general agents. J. Artif. Intell. Res.(JAIR) 47 (2013),
253–279.
[25] I SARD , M., B UDIU , M., Y U , Y., B IRRELL , A., AND F ETTERLY, D. Dryad: Distributed data-parallel programs from
sequential building blocks. In Proceedings of the 2nd ACM
SIGOPS/EuroSys European Conference on Computer Systems
2007 (New York, NY, USA, 2007), EuroSys ’07, ACM, pp. 59–
72.
[12] B ERGSTRA , J., B REULEUX , O., BASTIEN , F., L AMBLIN ,
P., PASCANU , R., D ESJARDINS , G., T URIAN , J., WARDE FARLEY, D., AND B ENGIO , Y. Theano: A cpu and gpu math
compiler in python. In Proc. 9th Python in Science Conf (2010),
pp. 1–7.
[13] B ROCKMAN , G., C HEUNG , V., P ETTERSSON , L., S CHNEIDER ,
J., S CHULMAN , J., TANG , J., AND Z AREMBA , W. OpenAI
gym. arXiv preprint arXiv:1606.01540 (2016).
[26] J IA , Y., S HELHAMER , E., D ONAHUE , J., K ARAYEV, S., L ONG ,
J., G IRSHICK , R., G UADARRAMA , S., AND DARRELL , T.
Caffe: Convolutional architecture for fast feature embedding.
arXiv preprint arXiv:1408.5093 (2014).
[14] B YKOV, S., G ELLER , A., K LIOT, G., L ARUS , J. R., PANDYA ,
R., AND T HELIN , J. Orleans: Cloud computing for everyone.
In Proceedings of the 2nd ACM Symposium on Cloud Computing
(2011), ACM, p. 16.
[27] J ORDAN , M. I., AND M ITCHELL , T. M. Machine learning:
Trends, perspectives, and prospects. Science 349, 6245 (2015),
255–260.
[15] C ASADO , M., F REEDMAN , M. J., P ETTIT, J., L UO , J., M C KEOWN , N., AND S HENKER , S. Ethane: Taking control of the enterprise. SIGCOMM Comput. Commun. Rev. 37, 4 (Aug. 2007),
1–12.
[28] L I , L., JAMIESON , K., D E S ALVO , G., ROSTAMIZADEH ,
A., AND TALWALKAR , A. Hyperband: A novel banditbased approach to hyperparameter optimization. arXiv preprint
arXiv:1603.06560 (2016).
[16] C HAROUSSET, D., S CHMIDT, T. C., H IESGEN , R., AND
W ÄHLISCH , M. Native actors: A scalable software platform for
distributed, heterogeneous environments. In Proceedings of the
2013 workshop on Programming based on actors, agents, and
decentralized control (2013), ACM, pp. 87–96.
[29] L OOKS , M., H ERRESHOFF , M., H UTCHINS , D., AND N ORVIG ,
P. Deep learning with dynamic computation graphs. arXiv
preprint arXiv:1702.02181 (2017).
[30] M NIH , V., BADIA , A. P., M IRZA , M., G RAVES , A., L ILLI CRAP, T. P., H ARLEY, T., S ILVER , D., AND K AVUKCUOGLU ,
K. Asynchronous methods for deep reinforcement learning. In
International Conference on Machine Learning (2016).
[17] C HEN , T., L I , M., L I , Y., L IN , M., WANG , N., WANG , M.,
X IAO , T., X U , B., Z HANG , C., AND Z HANG , Z. MXNet: A
flexible and efficient machine learning library for heterogeneous
distributed systems. In NIPS Workshop on Machine Learning
Systems (LearningSys’16) (2016).
[31] M URRAY, D. G., M C S HERRY, F., I SAACS , R., I SARD , M.,
BARHAM , P., AND A BADI , M. Naiad: A timely dataflow system.
In Proceedings of the Twenty-Fourth ACM Symposium on Operating Systems Principles (New York, NY, USA, 2013), SOSP ’13,
ACM, pp. 439–455.
[18] D EAN , J., AND G HEMAWAT, S. MapReduce: Simplified data
processing on large clusters. Commun. ACM 51, 1 (Jan. 2008),
107–113.
14
[32] M URRAY, D. G., S CHWARZKOPF, M., S MOWTON , C., S MITH ,
S., M ADHAVAPEDDY, A., AND H AND , S. CIEL: A universal execution engine for distributed data-flow computing. In Proceedings of the 8th USENIX Conference on Networked Systems Design and Implementation (Berkeley, CA, USA, 2011), NSDI’11,
USENIX Association, pp. 113–126.
[47] VAN D EN B ERG , J., M ILLER , S., D UCKWORTH , D., H U , H.,
WAN , A., F U , X.-Y., G OLDBERG , K., AND A BBEEL , P. Superhuman performance of surgical tasks by robots using iterative learning from human-guided demonstrations. In Robotics
and Automation (ICRA), 2010 IEEE International Conference on
(2010), IEEE, pp. 2074–2081.
[33] N G , A., C OATES , A., D IEL , M., G ANAPATHI , V., S CHULTE ,
J., T SE , B., B ERGER , E., AND L IANG , E. Autonomous inverted helicopter flight via reinforcement learning. Experimental
Robotics IX (2006), 363–372.
[48] V ENKATARAMAN , S., PANDA , A., O USTERHOUT, K., G H ODSI , A., A RMBRUST, M., R ECHT, B., F RANKLIN , M., AND
S TOICA , I. Drizzle: Fast and adaptable stream processing at
scale. In Proceedings of the Twenty-Sixth ACM Symposium on
Operating Systems Principles (2017), SOSP ’17, ACM.
[34] N ISHIHARA , R., M ORITZ , P., WANG , S., T UMANOV, A.,
PAUL , W., S CHLEIER -S MITH , J., L IAW, R., N IKNAMI , M.,
J ORDAN , M. I., AND S TOICA , I. Real-time machine learning:
The missing pieces. In Workshop on Hot Topics in Operating
Systems (2017).
[49] W HITE , T. Hadoop: The Definitive Guide. O’Reilly Media, Inc.,
2012.
[50] Z AHARIA , M., C HOWDHURY, M., DAS , T., DAVE , A., M A ,
J., M C C AULEY, M., F RANKLIN , M. J., S HENKER , S., AND
S TOICA , I. Resilient distributed datasets: A fault-tolerant abstraction for in-memory cluster computing. In Proceedings of the
9th USENIX conference on Networked Systems Design and Implementation (2012), USENIX Association, pp. 2–2.
[35] O USTERHOUT, J., G OPALAN , A., G UPTA , A., K EJRIWAL , A.,
L EE , C., M ONTAZERI , B., O NGARO , D., PARK , S. J., Q IN , H.,
ROSENBLUM , M., ET AL . The RAMCloud storage system. ACM
Transactions on Computer Systems (TOCS) 33, 3 (2015), 7.
[36] O USTERHOUT, K., W ENDELL , P., Z AHARIA , M., AND S TO ICA , I. Sparrow: Distributed, low latency scheduling. In Proceedings of the Twenty-Fourth ACM Symposium on Operating
Systems Principles (New York, NY, USA, 2013), SOSP ’13,
ACM, pp. 69–84.
[51] Z AHARIA , M., X IN , R. S., W ENDELL , P., DAS , T., A RM BRUST, M., DAVE , A., M ENG , X., ROSEN , J., V ENKATARA MAN , S., F RANKLIN , M. J., G HODSI , A., G ONZALEZ , J.,
S HENKER , S., AND S TOICA , I. Apache Spark: A unified engine for big data processing. Commun. ACM 59, 11 (Oct. 2016),
56–65.
[37] Q U , H., M ASHAYEKHI , O., T EREI , D., AND L EVIS , P. Canary:
A scheduling architecture for high performance cloud computing.
arXiv preprint arXiv:1602.01412 (2016).
[38] ROCKLIN , M. Dask: Parallel computation with blocked algorithms and task scheduling. In Proceedings of the 14th Python in
Science Conference (2015), K. Huff and J. Bergstra, Eds., pp. 130
– 136.
[39] S ALIMANS , T., H O , J., C HEN , X., AND S UTSKEVER , I. Evolution strategies as a scalable alternative to reinforcement learning.
arXiv preprint arXiv:1703.03864 (2017).
[40] S ANFILIPPO , S. Redis: An open source, in-memory data structure store. https://redis.io/, 2009.
[41] S CHULMAN , J., W OLSKI , F., D HARIWAL , P., R ADFORD , A.,
AND K LIMOV, O. Proximal policy optimization algorithms.
arXiv preprint arXiv:1707.06347 (2017).
[42] S CHWARZKOPF, M., KONWINSKI , A., A BD -E L -M ALEK , M.,
AND W ILKES , J. Omega: Flexible, scalable schedulers for large
compute clusters. In Proceedings of the 8th ACM European Conference on Computer Systems (New York, NY, USA, 2013), EuroSys ’13, ACM, pp. 351–364.
[43] S ILVER , D., H UANG , A., M ADDISON , C. J., G UEZ , A.,
S IFRE , L., VAN D EN D RIESSCHE , G., S CHRITTWIESER , J.,
A NTONOGLOU , I., PANNEERSHELVAM , V., L ANCTOT, M.,
ET AL . Mastering the game of Go with deep neural networks
and tree search. Nature 529, 7587 (2016), 484–489.
[44] S UTTON , R. S., AND BARTO , A. G. Reinforcement Learning:
An Introduction, vol. 1. MIT press Cambridge, 1998.
[45] T ODOROV, E., E REZ , T., AND TASSA , Y. Mujoco: A physics
engine for model-based control. In Intelligent Robots and
Systems (IROS), 2012 IEEE/RSJ International Conference on
(2012), IEEE, pp. 5026–5033.
[46] VALIANT, L. G. A bridging model for parallel computation.
Communications of the ACM 33, 8 (1990), 103–111.
15
A
Satisfying the Requirements
In this section, we briefly explain how our design satisfies the requirements outlined in Section 2.
Flexibility: Ray extends the already general dynamic task model [32], by adding the ray.wait() primitive to efficiently handle tasks whose completion times are not known in advance, and the actor abstraction to handle third part
simulators and amortize expensive setups. Finally, Ray supports heterogeneous resources, such as GPUs.
Performance: Table 4 summarizes techniques for scaling each component and the associated overhead. The last
column shows the average number of requests/sec that each component should handle as a function of system and
workload parameters, where N is the number of nodes, s the number of GCS shards, g the number of global schedulers,
w the average number of tasks/sec generated by a node, and ε, the percentage of tasks submitted to the global scheduler.
For instance, on average, the local scheduler handles a number of requests proportional to w tasks/sec, where α1 ,
includes assigning the task to a worker or sending/receiving the task to/from the global scheduler. Similarly, the global
scheduler handles a number of request proportional to the average number of tasks/sec it receives (εNw/g) times α, a
factor that includes querying the GCS for task’s inputs’ locations; h represents the number of heartbeats/sec.
Note that given the cluster size, N, and the average load generated by a node, w, we could pick the number of GCS
shards, s and global schedulers, s to bound the load on each shard or global scheduler. This makes both GCS and the
global scheduler horizontally scalable.
Ease of development: Ray handles a variety of component failures, relieving the developer from writing complex software to handle failures. Ray achieves fault tolerance by using a variety of techniques, including stateless components,
replication and, replay, summarized in Table 5. This also enables a user to run the system on cheap infrastructures
where nodes may be preempted (e.g., AWS spot instances). The API shown in Table 2, while low-level, has proven to
be a good fit for the RL applications in our experience. The fact that the API is provided in Python, the most popular
language in the AI community, has been a big plus.
Component
Local Sched.
Object Store
Scaling Techniques
By default, use 1 per node.
By default, use 1 per node.
GCS
Use more shards (s).
Global Sched.
Use more replicas (g).
Data Object
Replicate objects to local
node before task execution.
Use nested tasks; spawn
tasks from workers.
Driver
Requests/sec
α1 w
α2 w
β Ns w + h Ng
s
α ε Ng s + h
N/A
N/A
Table 4: Achieving scalability in Ray. By increasing the number of GCS shards s and the number of global scheduler replicas
g, we can reduce the load on each GCS shard and global scheduler replica to a constant. N is the total number of nodes, w avg.
number of tasks/node/sec, h avg. number of node heartbeats/sec, and α, α1 , α2 , β , and ε are constants.
Component
Global Control Store (GCS)
Local and Global
Schedulers, Workers
Object Store, Actor
Driver
Description
Replicate each shard.
Stateless processes; restart and
read state from the GCS.
Use lineage stored in the GCS to
reconstruct objects (Section 4.2.3).
Restart it; can reuse objects already
present in the Object Store.
Table 5: Techniques for achieving fault tolerance in Ray
16
B
Ease of Development
B.1
Evolution Strategies
In this section, we give code examples to illustrate why the evolution strategies algorithm from Section 6.3.1 is so easy
to implement and customize on top of Ray and so difficult in the special-purpose reference system that we compare
to. Suppose we wish to parallelize the following serial Python code.
1
2
3
4
5
6
7
def f ():
return result
def g ( N ):
return [ f () for _ in range ( N )]
results = [ g ( N ) for _ in range ( M )]
Attempting to implement this code in MPI, distributed TensorFlow, or many specialized systems would involve
writing a single program that is run by all workers. As shown in the code below, programs would need to branch based
on the role of that worker and would likely only work for predetermined values of M and N.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
if WORKER_INDEX == 0:
results = []
for i in range ( M ):
child_index = 1 + i
results . append ( RECEIVE ( child_index ) , i )
elif 1 <= WORKER_INDEX < M :
i n t e rmediate_results = []
for i in range ( N ):
child_index = M + 1 + WORKER_INDEX + M * i
intermediate_results . append ( RECEIVE ( child_index ))
SEND ( intermediate_results , 0)
else :
result = f ()
parent_index = ( WORKER_INDEX - M - 1) % M
SEND ( result , parent_index )
This is complex for a number of reasons. Each worker has to decide what tasks to execute and has to explicitly
dictate logic for optimizations such as batching, which normally a task scheduler could take care of. Secondly, allowing
variable values of M and N would substantially complicate the program, but hard coding the values M and N will make
it even more difficult to extend the code to deeper hierarchies in the future.
Finally, this violates the lockstep assumptions that these frameworks make and would sacrifice performance by
restricting the algorithm to a predefined scheduling layout.
In contrast, the Ray version is clean, simple, and extensible.
1
2
3
4
5
6
7
8
9
@ray.remote
def f ():
return result
@ray.remote
def g ( N ):
return ray.get([ f.remote() for _ in range ( N )])
results = ray.get([ g.remote( N ) for _ in range ( M )])
We believe that Ray’s API will allow developers to easily develop more complex distributed schemes.
17
B.2
Tree Reduce
Consider the scenario where one wants to perform an aggregation operation on multiple data points in a distributed
fashion. With Ray, the programmer would be able to easily aggregate data in an efficient manner.
1
2
3
4
5
6
7
8
dataset = [ data1 , data2 , data3 , data4 , data5 , data6 , data7 , data8 ]
dataset_1 = [ aggregate . remote ( dataset [ i ] , dataset [ i +1])
for i in range (0 , 8 , step =2)]
dataset_2 = [ aggregate . remote ( dataset_1 [ i ] , dataset_1 [ i +1])
for i in range (0 , 4 , step =2)]
result = ray.get( aggregate . remote ( dataset_2 [0] , dataset_2 [1]))
Note that this implementation can be written in a more concise fashion.
1
2
3
4
while len ( data ) > 1:
data = data [2:] + [ aggregate.remote( data [0] , data [1])]
result = ray.get( data [0])
In contrast, systems such as MPI and Spark provide specialize tree-reduction operators (i.e. MPI Allreduce and
rdd.treeAggregate), since hierarchical computations are not easily expressed in their APIs.
18
B.3
Hyperparameter Search
Ray enables developers to build hyperparameter search programs in a readable and simple manner.
In this example, assume we have an experiment class with the following interface. We add a single decorator to the
class to convert it into an actor.
1
2
3
4
5
6
7
8
@ray.remote
class Experiment ():
def __init__ ( self , hyperparameters ):
# ...
def training_step ( self ):
# ...
return accuracy
Assume also that we have these helper functions predefined.
is promising(results) returns True if the model is doing well and False otherwise. In practice, this function will
may more information than the current result.
generate hyperparameters defines a queue for many hyperparameter configurations, enumerating the space that
the programmer wants to search over.
1
2
3
4
5
def is_promising ( results ):
return True
def g e n e r a t e_ h y p er p a r am e t e rs ():
return hyperparameter_queue
Here is a simple and straightforward implementation of a hyperparameter search program. Notice that more complicated management schemes can be implemented fairly easily, but we chose to leave those out in favor of readability.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
hp_queue = g e n e ra t e _ hy p e r pa r a m et e r s ()
experiment_list = [ Experiment.remote( next ( hp_queue )) for i in range (10)]
paused = []
current_exps = { exp . training_step.remote(): exp for exp in experiment_list }
while True :
ex pe riment_futures = current_exps . keys ()
[ ready_id ] , remaining = @ray.wait( experiment_futures , num_returns =1)
current_exp = current_exps [ ready_id ]
current_accuracy = ray.get( ready_id )
if is_promising ( current_accuracy ):
current_exps [ current_exp . training_step.remote()] = current_exp
else :
paused . append ( current_exp )
new_exp = Experiment.remote( next ( hp_queue ))
current_exps [ new_exp . training_step.remote()] = new_exp
19
C
Hyperparameters for Evolution Strategies
Hyperparameter
Noise standard deviation
Adam stepsize
L2 coefficient
Episodes per batch
D
Value
0.02
0.01
0.005
5000
Hyperparameters for Proximal Policy Optimization
Hyperparameter
Timesteps per batch
SGD minibatch size
SGD epochs per iteration
Adam stepsize
PPO clip param
GAE parameter (λ )
Discount (γ)
20
Value
320000
32768
20
1e-4
0.2
0.95
0.995
| 2cs.AI
|
Searching critical values for floating-point
programs⋆
Hélène Collavizza, Claude Michel, and Michel Rueher
arXiv:1511.01080v2 [cs.PL] 5 Aug 2016
University of Nice–Sophia Antipolis, I3S/CNRS
BP 121, 06903 Sophia Antipolis Cedex, France
[email protected]
Abstract. Programs with floating-point computations are often derived
from mathematical models or designed with the semantics of the real
numbers in mind. However, for a given input, the computed path with
floating-point numbers may significantly differ from the path corresponding to the same computation with real numbers. As a consequence, developers do not know whether the program can actually produce very unexpected outputs. We introduce here a new constraint-based approach that
searches for test cases in the part of the over-approximation where errors
due to floating-point arithmetic could lead to unexpected decisions.
1
Introduction
In numerous applications, programs with floating-point computations are derived
from mathematical models over the real numbers. However, computations on
floating-point numbers are different from calculations in an idealised semantics1
of real numbers [8]. For some values of the input variables, the result of a sequence
of operations over the floating-point numbers can be significantly different from
the result of the corresponding mathematical operations over the real numbers.
As a consequence, the computed path with floating-point numbers may differ
from the path corresponding to the same computation with real numbers. This
can entail wrong outputs and dangerous decisions of critical systems. That’s
why identifying these values is a crucial issue for programs controlling critical
systems.
Abstract interpretation based error analysis [3] of finite precision implementations computes an over-approximation of the errors due to floating-point operations. The point is that state-of-the-art tools [6] may generate numerous false
alarms. In [16], we introduced a hybrid approach combining abstract interpretation and constraint programming techniques that reduces the number of false
alarms. However, the remaining false alarms are very embarrassing since we
cannot know whether the predicted unstable behaviors will occur with actual
data.
⋆
1
This work was partially supported by ANR COVERIF (ANR-15-CE25-0002)
That’s to say, computations as close as possible to the mathematical semantics of
the real numbers; for instance, computations with arbitrary precision or computer
algebra systems.
More formally, consider a program P , a set of intervals I defining the expected input values of P , and an output variable x of P on which depend critical decisions, e.g., activating an anti-lock braking system. Let [xR , xR ] be a sharp
approximation over the set of real numbers R of the domain of variable x for any
input of P . [xF , xF ] stands for the domain of variable x in the over-approximation
computed over the set of floating-point F for input values of I. The range [xR , xR ]
can be determined by calculation or from physical limits. It includes a small tolerance to take into account approximation errors, e.g. measurement, statistical,
or even floating-point arithmetic errors. This tolerance – specified by the user–
defines an acceptable loss of accuracy between the value computed over the
floating-point numbers and the value calculated over the real numbers. Values
outside the interval [xR , xR ] can lead a program to misbehave, e.g. take a wrong
branch in the control flow.
The problem we address in this paper consists of verifying whether there exist critical values in I for which the program can actually produce a result value
of x inside the suspicious intervals [xF , xR ) and (xR , xF ]. To handle this problem, we introduce a new constraint-based approach that searches for test cases
that hit the suspicious intervals in programs with floating-point computations.
In other words, our framework reduces this test case generation problem to a
constraint-solving problem over the floating-point numbers where the domain
of a critical decision variable has been shrunk to a suspicious interval. A constraint solver –based on filtering techniques designed to handle constraints over
floating-point numbers– is used to search values for the input data. Preliminary
results of experiments on small programs with classical floating-point errors are
encouraging.
The CPBPV FP, the system we developed, outperforms generate and test
methods for programs with more than one input variable. Moreover, these search
strategies can prove in many cases that no critical value exists.
2
Motivating example
Before going into the details, we illustrate our approach on a small example.
Assume we want to compute the area of a triangle from the lengths of its sides
a, b, and c with Heron’s formula:
p
s ∗ (s − a) ∗ (s − b) ∗ (s − c)
where s = (a + b + c)/2. The C program in Fig. 1 implements this formula, when
a is the longest side of the triangle.
The test of line 5 ensures that the given lengths form a valid triangle.
Now, suppose that the input domains are a ∈ [5, 10] and b, c ∈ [0, 5]. Over the
real numbers, s is greater than any of the sides of the triangle and squared_area
cannot be negative. Moreover, squared_area cannot be greater than 156.25 over
the real numbers since the triangle area is maximized for a right triangle with
/∗ Pre−condition : a ≥ b and a ≥ c ∗/
float heron(float a, float b, float c) {
3
float s, squared_area;
4
squared_area = 0.0f;
5
if (a <= b + c) {
6
s = (a + b + c) / 2.0f;
7
squared_area = s*(s-a)*(s-b)*(s-c);
8
}
9
return sqrt(squared_area);
10 }
1
2
Fig. 1. Heron
√
b = c = 5 and a = 5 2. However, these properties may not hold over the floatingpoint numbers because absorption and cancellation phenomena can occur2 .
Tools performing value analysis over the floating-point numbers [6,15] approximate the domain of squared_area to the interval [−1262.21, 979.01]. Since
this domain is an over-approximation, we do not know whether input values
leading to squared_area < 0 or squared_area > 156.25 actually exist. Note that
input domains –here a ∈ [5,10] and b, c ∈ [0,5]– are usually provided by the user.
Assume the value of the tolerance3 ε is 10−5 , the suspicious intervals for
squared_area are [−1262.21, −10−5) and (156.25001, 979.01]. CPBPV FP, the
system we developed, generated test cases for both intervals:
– a = 5.517474, b = 4.7105823, c = 0.8068917, and squared_area equals
−1.0000001 · 10−5 ;
– a = 7.072597, b = c = 5, and squared_area equals 156.25003.
CPBPV FP could also prove the absence of test cases for a tolerance ε =
10−3 with squared_area > 156.25 + ε.
In order to limit the loss of accuracy due to cancellation [8], line 7 of Heron’s
program can be rewritten as follows:
squared_area = ((a+(b+c))*(c-(a-b))*(c+(a-b))*(a+(b-c)))/16.0f;
However, there are still some problems with this optimized program. Indeed,
CPBPV FP found the test case a = 7.0755463, b = 4.350216, c = 2.72533, and
squared_area equals −1.0000001·10−5 for interval [−1262.21, −10−5) of squared_area .
There are no more problems in the interval (156.25001, 979.01] and CPBPV FP
did prove it.
2
3
Let’s remind that absorption in an addition occurs when adding two numbers of
very different order of magnitude, and the result is the value of the biggest number,
i.e., when x + y with y 6= 0 yields x. Cancellation occurs in s − a when s is so close
to a that the subtraction cancels most of the significant digits of s and a.
Note that even this small tolerance may lead to an exception in statement 9.
3
Framework for generating test cases
This section details the framework we designed to generate test cases reaching
suspicious intervals for a variable x in a program P with floating-point computations.
The kernel of our framework is FPCS [14,13,1,12], a solver for constraints
over the floating-point numbers; that’s to say a symbolic execution approach for
floating-point problems which combines interval propagation with explicit search
for satisfiable floating-point assignments. FPCS is used inside the CPBPV
bounded model checking framework [5]. CPBPV FP is the adaptation of CPBPV
for generating test cases that hit the suspicious intervals in programs with
floating-point computations.
The inputs of CPBPV FP are: P , an annotated program; a critical test ct
for variable x; [xF , xR ) or (xR , xF ], a suspicious interval for x. Annotations of P
specify the range of the input variables of P as well as the suspicious interval
for x. The latter assertion is just posted before the critical test ct.
To compute the suspicious interval for x, we approximate the domain of
x over the real numbers by [xR , xR ], and over the floating-point numbers by
[xF , xF ]. These approximations are computed with rAiCp [16], a hybrid system
that combines abstract interpretation and constraint programming techniques
in a single static and automatic analysis. The current implementation of rAiCp
is based upon the abstract interpreter Fluctuat [6], the constraint solver over
the reals RealPaver [10] and FPCS. The suspicious intervals for variable x are
denoted [xF , xR ) and (xR , xF ].
CPBPV FP performs first some pre-processing: P is transformed into DSAlike form4 . If the program contains loops, CPBPV FP unfolds loops k times
where k is a user specified constant. Loops are handled in CPBPV and rAiCp
with standard unfolding and abstraction techniques5 . So, there are no more loops
in the program when we start the constraint generation process. Standard slicing
operations are also performed to reduce the size of the control flow graph.
In a second step, CPBPV FP searches for executable paths reaching ct. For
each of these paths, the collected constraints are sent to FPCS, which solves
the corresponding constraint systems over the floating point numbers. FPCS
returns either a satisfiable instantiation of the input variables of P , or ∅.
As said before, FPCS [14,13,1,12] is a constraint solver designed to solve a set of
constraints over floating-point numbers without losing any solution. It uses 2Bconsistency along with projection functions adapted to floating-point arithmetic
[13,1] to filter constraints over the floating-point numbers. FPCS also provides
stronger consistencies like kB-consistencies, which allow better filtering results.
The search of solutions in constraint systems over floating numbers is trickier
than the standard bisection-based search in constraint systems over intervals of
4
5
DSA stands for Dynamic Single Assignment. In DSA-like form, all variables are
assigned exactly once in each execution path
In bounded model checking, k is usually increased until a counter-example is found
or until the number of time units is large enough for the application.
float slope(float x0, float h) {
float x1 = x0 + h; float x2 = x0 - h;
float fx1 = x1*x1; float fx2 = x2*x2;
float res = (fx1 - fx2) / (2.0*h);
return res;
}
Fig. 2. Approximation of the derivative of x2 by a slope
float polynomial(float a, float b, float c) {
float poly = (a*a + b + 1e-5f) * c;
return poly;
}
Fig. 3. Computation of polynomial (a2 + b + 10−5 ) ∗ c
real numbers. Thus, we have also implemented different strategies combining
selection of specific points and pruning. Details on theses strategies are given in
the experiments section.
CPBPV FP ends up with one of the following results:
– a test case proving that P can produce a suspicious value for x;
– a proof that no test case reaching the suspicious interval can be generated:
this is the case if the loops in P cannot be unfolded beyond the bound k
(See [5] for details on bounded unfolding) ;
– an inconclusive answer: no test case could be generated but the loops in
P could be unfolded beyond the bound k. In other words, the process is
incomplete and we cannot conclude whether P may produce a suspicious
value.
4
Preliminary experiments
We experimented with CPBPV FP on six small programs with cancellation and
absorption phenomena, two very common pitfalls of floating-point arithmetic.
The benchmarks are listed in the first two columns of table 1.
First two benchmarks concern the heron program and the optimized heron
program with the suspicious intervals described in the section 1.
Program slope (see Fig. 2) approximates the derivative of the square function
f (x) = x2 at a given point x0 . More precisely, it computes the slope of a nearby
(x0 −h)
secant line with a finite difference quotient: f ′ (x0 ) ≈ f (x0 +h)−f
. Over the
2h
real numbers, the smaller h is, the more accurate the formula is. For this function,
the derivative is given by f ′ (x) = 2x which yields exactly 26 for x = 13. Over
the floats, Fluctuat [6] approximates the return value of the slope program to
the interval [0, 25943] when h ∈ [10−6 , 10−3 ] and x0 = 13.
Program polynomial in Fig. 3 illustrates an absorption phenomenon. It computes the polynomial (a2 +b+10−5)∗c. For input domains a ∈ [103 , 104 ], b ∈ [0, 1]
and c ∈ [103 , 104 ], the minimum value of the polynomial over the real numbers
is equal to 1000000000.01.
simple interpolator and simple square are two benches extracted from
[9]. The first bench computes an interpolator, affine by sub-intervals while the
second is a rewrite of a square root function used in an industrial context.
All experiments were done on an Intel Core 2 Duo at 2.8 GHz with 4 GB
of memory running 64-bit Linux. We assume C programs handling IEEE 754
compliant floating-point arithmetic, intended to be compiled with GCC without
any optimization option and run on a x86 64 architecture managed by a 64bit Linux operating system. Rounding mode was to the nearest, i.e., where ties
round to the nearest even digit in the required position.
4.1
Strategies and solvers
We run CPBPV FP with the following search strategies for the FPCS solver:
– std: standard prune & bisection-based search used in constraint-systems over
intervals : splits the selected variable domain in two domains of equal size;
– fpc: splits the domain of the selected variable in five intervals:
• Three degenerated intervals containing only a single floating point number: the smallest float l, the largest float r, and the mid-point m;
• Two open intervals (l, m) and (m, r);
– fp3s: selects 3 degenerated intervals containing only a single floating point
number: the smallest float l, the largest float r, and the mid-point m. Hence,
fp3s is an incomplete method that might miss some solutions.
For all these strategies, we select first the variables with the largest domain
and we perform a 3B−consistency filtering step before starting the splitting
process.
We compared CPBPV FP with CBMC [4] and CDFL [7], two state-of-theart software bounded model checkers based on SAT solvers that are able to deal
with floating-point computations. We also run a simple generate & test strategy:
the program is run with randomly generated input values and we test whether
the result is inside the suspicious interval. The process is stopped as soon as a
test case hitting the suspicious interval is found.
4.2
Results
Table 1 reports the results for the other strategies and solvers. Since strategy
fpc3s is incomplete, we indicate whether a test case was found or not. Column
s? specifies whether a test case actually exists. Note that the computation times
of CBMC and CDFL include the pre-processing time for generating the constraint systems; the pre-processing time required by CPBPV is around 0.6s but
CPBPV is a non-optimised system written in java.
Name
Condition
CDFL CBMC
std
fpc
fpc3s s?
area < 10−5
3.874s 0.280s > 180 0.705 0.022 (n) y
area > 156.25 + 10−5 > 180s 34.512s 22.323 7.804 0.083 (n) y
optimized_heron
area < 10−5
7.618s 0.932s > 180 0.148 0.022 (n) y
area > 156.25 + 10−5 > 180s > 180s 8.988 30.477 0.101 (n) n
slope with
dh < 26.0 − 1.0
2.014s 1.548s 0.021 0.012 0.012 (y) y
h ∈ [10−6 , 10−3 ]
dh > 26.0 + 1.0
1.599s 0.653s 0.055 0.011 0.011 (y) y
dh < 26.0 − 10.0
0.715s 1.108s 0.006 0.006 0.007 (n) n
dh > 26.0 + 10.0
1.025s 1.080s 0.006 0.006 0.006 (n) n
polynomial
r < 109 +
0.170s 0.295s 0.022 0.006 0.006 (y) y
0.0099999904 − 10−3
simple_interpolator res < −10−5
0.296s 0.264s 0.018 0.012 0.012 (y) y
simple_square
S > 1.453125
−− 1.079s 0.012 0.012 0.012 (n) n
heron
Table 1. Results of the different solvers and strategies on the benchmarks
5
5.1
Discussion
Results analysis
The generate & test strategy behaves quite well on programs with only one input
variable when a test case exists but it is unable to find any test case for programs
with more than one input variable. More precisely, it found a test case in less
than 0.008s for the 6 suspicious intervals of program slope and for program
simple_interpolator . The generate & test strategy failed to find a test within
180s in all other cases.Of course, this strategy cannot show that there is no test
case reaching the suspicious interval; so, it is of little interest here.
Strategy fpc is definitely the most efficient and most robust one on all these
benchmarks. Note that CBMC and CDFL could neither handle the initial, nor
the optimized version of program heron in a timeout of 20 minutes whereas
FPCS found solutions in a reasonable time.
These preliminary results are very encouraging: they show that CPBPV FP
is effective for generating test cases for suspicious values outside the range of
acceptable values on small programs with classical floating-point errors. More
importantly, a strong point of CPBPV FP is definitely its refutation capabilities.
Of course, experiments on more significant benchmarks and on real applications are still necessary to evaluate the full capabilities and limits of CPBPV FP.
5.2
Related and further work
The goals of software bounded model checkers based on SAT solvers are close
to our approach. The point is that SAT solvers tend to be inefficient on these
problems due to the size of the domains of floating-point variables and the cost
of bit-vector operations [7]. CDFL [7] tries to address this issue by embedding
an abstract domain in the conflict driven clause learning algorithm of a SAT
solver. SAT solvers often use bitwise representations of numerical operations,
which may be very expensive (e.g., thousands of variables for one equation in
CDFL). Brain and al [11,2] have recently introduced a bit-precise decision procedure for the theory of floating-point arithmetic. The core of their approach is
a generalization of the conflict-driven clause-learning algorithm used in modern
SAT solvers. Their technique is significantly faster than a bit-vector encoding approach. Note that the constraint programming techniques used in our approach
are better suited to generate several test cases than these SAT-based approaches.
The advantage of CP is that it provides a uniform framework for representing
and handling integers, real numbers and floats. A new abstract-interpretation
based robustness analysis of finite precision implementations has recently been
proposed [9] for sound rounding error propagation in a given path in presence
of unstable tests.
A close connection between our floating-point solvers and the two abovementioned approaches is certainly worth exploring.
A second direction for further work concerns the integration of our constraintbased approach with new abstract-interpretation based robustness analysis of
finite precision implementations for sound rounding error propagation in a given
path in presence of unstable tests.
References
1. Bernard Botella, Arnaud Gotlieb, and Claude Michel. Symbolic execution
of floating-point computations. Software Testing, Verification and Reliability,
16(2):97–121, 2006.
2. Martin Brain, Vijay D’Silva, Alberto Griggio, Leopold Haller, and Daniel Kroening.
Interpolation-based verification of floating-point programs with abstract cdcl. In
Proc. of 20th Int. Symposium SAS, volume 7935 of LNCS. Springer, 2013.
3. Liqian Chen, Antoine Miné, and Patrick Cousot. A sound floating-point polyhedra
abstract domain. In Proc. of APLAS ’08, pages 3–18. Springer-Verlag, 2008.
4. Edmund Clarke, Daniel Kroening, and Flavio Lerda. A tool for checking ANSI-C
programs. In TACAS, volume 2988 of LNCS, pages 168–176, 2004.
5. Hélène Collavizza, Michel Rueher, and Pascal Van Hentenryck. A constraintprogramming framework for bounded program verification. Constraints Journal,
15(2):238–264, 2010.
6. David Delmas, Eric Goubault, Sylvie Putot, Jean Souyris, Karim Tekkal, and
Franck Védrine. Towards an industrial use of FLUCTUAT on safety-critical avionics software. In FMICS, volume 5825 of LNCS, pages 53–69. Springer, 2009.
7. Vijay D’Silva, Leopold Haller, Daniel Kroening, and Michael Tautschnig. Numeric
bounds analysis with conflict-driven learning. In Proc. TACAS, volume 7214 of
Lecture Notes in Computer Science, pages 48–63. Springer, 2012.
8. David Goldberg. What every computer scientist should know about floating point
arithmetic. ACM Computing Surveys, 23(1):5–48, 1991.
9. Eric Goubault and Sylvie Putot. Robustness analysis of finite precision implementations. In Programming Languages and Systems - 11th Asian Symposium,
APLAS, volume 8301 of LNCS, pages 50–57. Springer, 2013.
10. Laurent Granvilliers and Frédéric Benhamou. Algorithm 852: RealPaver: an interval solver using constraint satisfaction techniques. ACM Transactions on Mathematical Software, 32(1):138–156, 2006.
11. Leopold Haller, Alberto Griggio, Martin Brain, and Daniel Kroening. Deciding
floating-point logic with systematic abstraction. In Formal Methods in ComputerAided Design, FMCAD, pages 131–140. IEEE, 2012.
12. Bruno Marre and Claude Michel. Improving the floating point addition and subtraction constraints. In CP, volume 6308 of LNCS, pages 360–367. Springer, 2010.
13. Claude Michel. Exact projection functions for floating-point number constraints.
In 7th International Symposium on Artificial Intelligence and Mathematics, 2002.
14. Claude Michel, Michel Rueher, and Yahia Lebbah. Solving constraints over
floating-point numbers. In Proc.of CP 2001, LNCS 2239: 524–538, 2001.
15. Olivier Ponsini, Claude Michel, and Michel Rueher. Refining abstract interpretation based value analysis with constraint programming techniques. In Michela
Milano, editor, 18th International Conference on Principles and Practice of Constraint Programming (CP 2012), volume 7514 of Lecture Notes in Computer Science, pages 593–607. Springer, 2012.
16. Olivier Ponsini, Claude Michel, and Michel Rueher. Verifying floating-point programs with constraint programming and abstract interpretation techniques. Autom. Softw. Eng., 23(2):191–217, 2016.
| 8cs.DS
|
Tensor Network alternating linear scheme for MIMO Volterra system
identification
Kim Batselier a , Zhongming Chen a , Ngai Wong a ,
arXiv:1607.00127v2 [cs.NA] 18 Oct 2016
a
The Department of Electrical and Electronic Engineering, The University of Hong Kong
Abstract
This article introduces two Tensor Network-based iterative algorithms for the identification of high-order discrete-time nonlinear multipleinput multiple-output (MIMO) Volterra systems. The system identification problem is rewritten in terms of a Volterra tensor, which is
never explicitly constructed, thus avoiding the curse of dimensionality. It is shown how each iteration of the two identification algorithms
involves solving a linear system of low computational complexity. The proposed algorithms are guaranteed to monotonically converge and
numerical stability is ensured through the use of orthogonal matrix factorizations. The performance and accuracy of the two identification
algorithms are illustrated by numerical experiments, where accurate degree-10 MIMO Volterra models are identified in about 1 second in
Matlab on a standard desktop pc.
Key words: Volterra series; tensors; MIMO; identification methods; system identification
1
Introduction
Volterra series [26] have been extensively studied and applied in applications like speech modeling [17], loudspeaker
linearization [13], nonlinear control [8], active noise control [22], modeling of biological and physiological systems [16], nonlinear communication channel identification
and equalization [4], distortion analysis [24] and many others. Their applicability has been limited however to “weakly
nonlinear systems”, where the nonlinear effects play a nonnegligible role but are dominated by the linear terms. Such
limitation is not inherent to the Volterra series themselves,
as they can represent a wide range of nonlinear dynamical
systems, but is due to the exponentially growing number of
Volterra kernel coefficients as the degree increases. Indeed,
assuming a finite memory M , the dth-order response of
a discrete-time single-input single-output (SISO) Volterra
system is given by
yd (t) =
M
−1
X
hd (k1 , . . . , kd )
k1 ,...,kd =0
d
Y
u(t − ki ),
i=1
where y(t), u(t) are the scalar output and input at time t respectively and the dth-order Volterra kernel hd (k1 , . . . , kd )
Email addresses: [email protected] (Kim
Batselier), [email protected] (Zhongming Chen),
[email protected] (Ngai Wong).
Preprint submitted to Automatica
is described by M d numbers. For a multiple-input multipleoutput (MIMO) Volterra system with p inputs the situation
gets even worse, where the dth-order Volterra kernel for one
particular output is characterized by (pM )d numbers. This
problem is also commonly known as the curse of dimensionality.
The paradigm used in this article to break this curse is to
trade storage for computation. This means that all the
Volterra coefficients are replaced by only a few numbers,
from which all Volterra coefficients can be computed. This
idea is not new, e.g. Volterra kernels have been expanded on
orthonormal basis functions in order to reduce their complexity [2,6]. Tensors (namely, multi-dimensional arrays that
are generalizations of matrices to higher orders) are also
suitable candidates for this purpose. In [7] both the canonical polyadic [3, 10] and Tucker tensor decompositions [23]
were used. The canonical polyadic decomposition can suffer from instability however, and the determination of its
rank is known to be a NP-hard problem [12], which can be
ill-posed [5]. The main disadvantage of the Tucker decomposition of the Volterra kernels, which is in fact an expansion onto a set of orthonormal basis functions, is that it still
suffers from an exponential complexity.
This motivates us to develop and introduce a new description
of discrete-time MIMO Volterra systems in terms of particular Tensor Networks (TN) [18]. For the particular case
of multiple-input-single-output (MISO) systems, these Tensor Networks will turn out to be Trains (TTs) [19]. A TN
19 October 2016
representation does not suffer from any instability or exponential complexity and can represent all Volterra kernels
combined by O(d(pM + 1)r2 ) elements, where r is a to-bedetermined number called the TN-rank. TNs were originally
developed in the physics community. Of particular importance is the Density Matrix Renormalization Group (DMRG)
algorithm [25], which is an iterative algorithm originally developed for the determination of the ground state of a entangled multi-body quantum system. Its applicability however
is not limited to problems in quantum physics, as demonstrated by recent interest in the scientific computing community [11, 20, 21].
i3
i2
i1
21
22
23
24
A d-way or dth-order tensor is denoted A ∈ Rn1 ×n2 ×···×nd
and hence each of its entries ai1 i2 ···id is determined by d
indices. The numbers n1 , n2 , . . . , nd are called the dimensions of the tensor. A tensor is cubical if all its dimensions
are equal. A cubical tensor is symmetric when all its entries
satisfy ai1 i2 ···id = aπ(i1 ,i2 ,...,id ) , where π(i1 , i2 , . . . , id ) is
any permutation of the indices. An example 3-way tensor
with dimensions 4, 3, 2 is shown in Fig. 1. For practical purposes, only real tensors are considered. We use boldface
capital calligraphic letters A, B, . . . to denote tensors, boldface capital letters A, B, . . . to denote matrices, boldface
letters a, b, . . . to denote vectors, and Roman letters a, b, . . .
to denote scalars. The transpose of a matrix A or vector a
are denoted by AT and aT , respectively. The unit matrix of
order n is denoted In .
(1) we derive a new description of discrete-time MIMO
Volterra systems using the TN format,
(2) we derive two iterative MIMO Volterra identification
algorithms that estimate all Volterra kernels in the TN
format from given input-output data.
In each step of the iterative identification a small linear system needs to be solved. The main computational tools are the
singular value decomposition (SVD) and the QR decomposition [9]. These orthogonal matrix factorizations ensure the
numerical stability of the methods [11]. The first identification method, which is called the Alternating Linear Scheme
(ALS) method, has the lowest computational complexity
but assumes that the TN-ranks rk ’s are fixed. The second
identification method, called the Modified Alternating Linear Scheme (MALS), removes this limitation and allows for
the adaptive updating of the TN-ranks during the iterations.
The TN-ranks are determined numerically by means of a
SVD. This is reminiscent of the determination of the order
of linear systems in subspace identification algorithms [14].
Monotonic convergence of both the ALS and MALS methods under certain conditions is discussed in [21].
We now give a brief description of some required tensor
operations and properties. The generalization of the matrixmatrix multiplication to tensors involves a multiplication of
a matrix with a d-way tensor along one of its d possible
modes (equiv. indices/axes).
Definition 2.1 (Tensor k-mode product)( [15, p. 460])
The k-mode product B = A ×k U of a tensor A ∈
Rn1 ×···×nk ×···×nd with a matrix U ∈ Rpk ×nk is defined by
The outline of this article is as follows. In Section 2 we
give a brief overview of important tensor concepts, operations and properties. The MIMO Volterra TN framework is
introduced in Section 3. The two iterative identification algorithms are derived in Section 4 and applied on two examples in Section 5. To our knowledge, this is the only time
where MIMO Volterra systems of degree 10 were identified
in about 1 second on a standard computer. Matlab/Octave
implementations of our algorithms are freely available from
https://github.com/kbatseli/MVMALS.
2.1
5 9
6 10
7 11
8 12
17
18
19
20
Fig. 1. A 4 × 3 × 2 tensor where each entry is determined by three
indices i1 , i2 , i3 .
In this article, we adopt the DMRG method in the TN format for the identification of MIMO discrete-time Volterra
systems. The contributions of this article are twofold:
2
1
2
3
4
13
14
15
16
bi1 ···ik−1 jik+1 ···id =
nk
X
ujik ai1 ···ik−1 ik ik+1 ···id ,
(1)
ik =1
and B ∈ Rn1 ×···×nk−1 ×pk ×nk+1 ×···×nd .
The following illustrative example rewrites the familiar matrix multiplication as a 1-mode and 2-mode product.
Example 1 (Matrix multiplication as mode products) For
matrices A, B, C with matching dimensions we have that
Preliminaries
A ×1 B ×2 C := B A C T .
Tensor basics
An interesting observation is that the definition of the kmode product also includes the multiplication of a tensor
Tensors in this article are multi-dimensional arrays that generalize the notions of vectors and matrices to higher orders.
2
A with d vectors. The following example highlights a very
important case.
n3
n1
Example 2 Consider a symmetric d-way tensor A with dimensions n and a vector x ∈ Rn . The multidimensional
n1
r2
r2
A
r1
(2)
A (3)
n3
n2
is the scalar
A xd := A ×1 xT ×2 xT ×3 · · · ×d xT ,
Fig. 2. The TT-cores of a 3-way tensor A are two matrices
A(1) , A(3) and a 3-way tensor A(2) .
(2)
Definition 2.4 (Vectorization)( [15, p. 460]) The vectorization of a tensor A, denoted vec(A), rearranges all its entries into one column vector.
which is obtained as a homogeneous polynomial of degree
d in the variables x1 , . . . , xn .
Example 4 For the tensor in Fig. 1, we have
The Kronecker product plays a crucial role in our description
of MIMO Volterra systems.
vec(A) = reshape(A, [24, 1]) =
Definition 2.2 (Kronecker product)( [15, p. 461]) If B ∈
Rm1 ×m2 and C ∈ Rn1 ×n2 , then their Kronecker product
B ⊗ C is an m1 × m2 block matrix whose (i3 , i4 )th block
is the n1 × n2 matrix bi3 i4 C
b11 · · · b1n1
b11 C · · · b1n1 C
. .
.
.
..
..
.
..
.. ⊗ C = ..
.
.
.
bm1 1 · · · bm1 n1
A
=
A (1)
n2
T
contraction of A with x = x1 x2 · · · xn
r1
1 2 · · · 24
T
.
The importance of the vectorization lies in the following
equation
vec(A ×1 U1 ×2 · · · ×d Ud ) = (Ud ⊗ · · · ⊗ U1 )vec(A).
(5)
.
bm1 1 C · · · bm1 n1 C
Observe how the order in the Kronecker product is reversed
with respect to the ordering of the mode products. Equation
(5) allows us to rewrite (2) as
(3)
We use the notation x d := x ⊗ x ⊗ · · · ⊗ x for the d-times
repeated Kronecker product. The mixed-product property of
the Kronecker product states that if A, B, C and D are
matrices of such sizes that one can form the matrix products
AC and BD, then
(A ⊗ B) (C ⊗ D) = AC ⊗ BD.
A xd = vec(A)T x d ,
(6)
which tells us how the k-mode products of a tensor A with
a vector x can be computed. When A in (5) is also a matrix,
we then obtain the following useful property
(4)
vec(U1 A U2T ) = (U2 ⊗ U1 ) vec(A).
Definition 2.3 (Reshaping)( [15, p. 460]) Reshaping is another often used tensor operation. The most common reshaping is the matricization, which reorders the entries of
A into a matrix. We adopt the Matlab/Octave reshape operator “reshape(A, [n1 , n2 , n3 · · · ])”, which reshapes the
tensor A into a tensor with dimensions n1 , n2 , n3 , . . .. The
total number of elements of A must be the same as n1 ×
n2 × n3 · · · .
2.2
(7)
Tensor Train decomposition
The Tensor Network representation used in this article lies
conceptually very close to the Tensor Train decomposition [19]. We therefore first discuss the Tensor Train representation. A TT-decomposition represents a d-way tensor A
in terms of d 3-way tensors A(1) , . . . , A(d) , which are called
the TT-cores. The kth TT-core has dimensions rk−1 , nk , rk ,
where rk−1 , rk are called the TT-ranks. Specifically, each
entry of A ∈ Rn1 ×···×nd is determined by
Example 3 We illustrate the reshaping operator on the 4 ×
3 × 2 tensor of Fig. 1
1 5 9 13 17 21
2 6 10 14 18 22
reshape(A, [4, 6]) =
.
3 7 11 15 19 23
4 8 12 16 20 24
(1)
(2)
(d)
ai1 i2 ···id = Ai1 Ai2 · · · Aid ,
(k)
(8)
where Aik is the rk−1 ×rk matrix obtained from specifying
ik . Since ai1 i2 ···id is a scalar, it immediately follows that
r0 = rd = 1. The TT-decomposition is illustrated for a 3way tensor in Fig. 2. Note that when all TT-ranks are equal to
r, then the storage of a cubical d-way tensor with dimensions
n in the TT format needs O(dnr2 ) elements. Small TT-ranks
Probably the most important reshaping of a tensor is the
vectorization.
3
therefore result in a significant reduction of required storage
cost. In order to describe MIMO Volterra systems we will
need to remove the constraint that r0 = 1 and therefore
obtain a slightly more general TN. Observe that the removal
of this constraint implies that the left-hand side of (8) will
then be a r0 × 1 vector. The following multidimensional
contraction in the TN format turns out to be very important
for Volterra systems.
Definition 3.1 (SISO Volterra tensor) Given a discrete-time
SISO Volterra system of degree d and memory M as described in (10), then the d-way cubical Volterra tensor V of
dimension M + 1 is defined by
Lemma 2.1 ( [19, p. 2309]) Given a cubical tensor A ∈
Rn×···×n in the TT format and a vector x ∈ Rn . The multidimensional contraction A xd can then be computed as
T
ut := 1 u(t) u(t − 1) · · · u(t − M + 1)
∈ RM +1 .
y(t) = V udt ,
where
Axd = (A(1) ×2 xT ) (A(2) ×2 xT ) · · · (A(d) ×2 xT ).
(9)
The extension of this definition to the MIMO case is
straightforward. Indeed, the ith output yi (t) is then a multivariate polynomial of degree d in u1 (t), . . . , u1 (t − M +
1), u2 (t), . . . , u2 (t − M + 1), . . . , up (t − M + 1). This implies that the vector ut in Definition 3.1 needs to be extended
with these additional inputs, which results in an increase
in the dimension of the corresponding Volterra tensor. In
addition, the order of the Volterra tensor is incremented to
accommodate for the multiple number of outputs.
Again, in the MIMO case this contraction will result in a
r0 × 1 vector where now each entry is the evaluation of
a homogeneous polynomial. The numerical stability of the
two identification algorithms described in this article relies
on the TN-cores being either left or right orthogonal.
Definition 3.2 (MIMO Volterra tensor) Given a discrete
time p-input l-output Volterra system of degree d and memory M , then the d + 1-way Volterra tensor V of dimensions
l × pM + 1 × · · · × pM + 1 is defined by
Definition 2.5 (Left orthogonal and right orthogonal TNcores)( [11, p. A689]) A TN-core A(i) is left orthogonal if
it can be reshaped into an ri−1 ni × ri matrix A for which
AT A = Iri
y1 (t)
y2 (t)
y(t) := . = V ×2 uTt ×3 uTt · · · ×d+1 uTt , (11)
..
yl (t)
applies. Similarly, a TN-core A(i) is right orthogonal if it
can be reshaped into an ri−1 × ni ri matrix A for which
A AT = Iri−1
where ut is the (pM + 1) × 1 vector with entries
1, u1 (t), u2 (t), . . . , up (t), . . . , u1 (t − M + 1), u2 (t − M +
1), . . . , up (t − M + 1).
applies.
3
Tensor description of MIMO Volterra systems
The MIMO Volterra tensor consists of l (pM + 1)d entries,
which can quickly become practically infeasible to store as
M and d grow. We therefore propose to store the Volterra
tensor V in the TN format V (1) , . . . , V (d) , where the first
TN-core V (1) has dimensions l × pM + 1 × r1 . The other
TN-cores have sizes ri−1 × pM + 1 × ri , with rd = 1 for the
last core. The TN format reduces to a TT for the MISO case
(l = 1). This change in representation reduces the storage
requirement to O((d−1)(pM +1)r2 +(pM +1)lr). Lemma
2.1 then immediately tells us how to simulate the output
samples at time t as
To keep the notation simple, we first consider the following
discrete-time SISO Volterra system of degree d, namely,
y(t) = y0 (t) + y1 (t) + y2 (t) + · · · + yd (t),
= h0 +
d
X
M
−1
X
i=1 k1 ,...,ki =0
hi (k1 , . . . , ki )
i
Y
u(t − kj ),
j=1
(10)
where y(t), u(t) are the scalar output and input at time t respectively, M is the memory length and hi (k1 , . . . , ki ) denotes the ith Volterra kernel. Previous work that uses tensors in Volterra series describes each Volterra kernel as a
separate symmetric tensor [7]. Realizing from (10) that y(t)
is a multivariate polynomial in u(t), . . . , u(t − M + 1), we
define the SISO Volterra tensor as follows.
y(t) = (V (1) ×2 uTt ) (V (2) ×2 uTt ) · · · (V (d) ×2 uTt ),
(12)
with a computational complexity of O(d(pM +1)r+dr3 ). In
[1] a much faster simulation complexity of O(dRN log N )
4
is called a set of persistent exciting inputs of order d.
for computing N samples is obtained by using a symmetric polyadic representation for each of the Volterra kernels
separately. However, the method only works for the SISO
case and the identification of such a representation can be
problematic, since computation of the canonical rank R is
an NP-hard problem [12], which can be ill-posed [5].
4
We will from here on always assume that the inputs are
persistent exciting and that N ≥ rank(U ). The truncated
Volterra series for one particular output is a polynomial
in
+d
pM variables of degree d and therefore has pM
coeffipM
cients. A solution V (1) of the matrix equation (14) that con
+d
sists of only l pM
distinct entries therefore must correpM
spond with l symmetric tensors, where each column of V (1)
corresponds with one particular d-way symmetric tensor. As
a result, persistent exciting inputs and sufficient number of
samples N together with Lemma 4.1 has the following consequences:
MIMO Volterra system identification
In what follows we always consider the MIMO case. Observe now that we can rewrite (11) as
y(t)T = (ut d )T (V (1) )T ,
(13)
(1) the identification problem (14) has an infinite number
of solutions,
(2) the l unique minimal norm solutions of (14) correspond
with l symmetric tensors,
(3) the unicity of the symmetric solutions implies that no
other symmetric solutions exist,
(4) any solution V (1) of (14) can be turned into l minimal
norm solutions by symmetrizing each of the tensors
given by the columns of V (1) .
where V (1) is the Volterra tensor reshaped into a l × (pM +
1)d matrix. Writing out (13) for t = 0, 1, . . . , N − 1 leads
to the following matrix equation
Y = U (V (1) )T ,
(14)
where
T
Y := y(0) y(1) y(2) · · · y(N − 1) ,
T
U := u0 d u1 d u2 d · · · uN −1 d
.
Ideally, one would solve (14) for the minimal norm solutions, which acts as a regularization of the problem such that
the solution is uniquely defined. The two iterative methods
which we describe in the next subsections do not guarantee convergence to the minimal norm solutions. However, in
practice we observe that the norms of the obtained solutions
are quite close to the minimal ones.
Essentially, MIMO Volterra system identification is in its
most basic form solving the matrix equation (14). This immediately reveals the difficulty, as the size of the U matrix
is N × (pM + 1)d , which quickly becomes prohibitive even
for SISO systems and moderate degree d. This motivates us
to solve the following problem.
4.1
Problem 1 For a given set of measured time series
y1 , . . . , yl , u1 , . . . , up , memory M and degree d, solve the
matrix equation (14) for V in the TN format.
Alternating Linear Scheme method
The first method that we derive is the Alternating Linear
Scheme (ALS) method. The key idea of this method is to fix
the TN-ranks r1 , r2 , . . . , rd−1 and choose a particular initial
guess for V (1) , . . . , V (d) . Each of the TN-cores is then updated separately in an iterative fashion until convergence has
been reached. Once the core V (d) is updated, the algorithm
“sweeps” back towards V (1) and so on. It turns out that updating one of the cores is equivalent with solving a much
smaller linear system, which can be done very efficiently.
Before describing the form of the reduced linear system, we
first introduce the following notation
Before presenting the two numerical algorithms that solve
Problem 1, we first give an upper bound on the rank of U ,
together with some important implications.
Lemma
4.1 The rank of the matrix U is upper bounded by
pM +d
pM .
Proof 1 Each row of U consists of (pM +1)d entries formed
d
by the repeated
Kronecker product uk . There are however
pM +d
only pM distinct entries.
vk−1 := (V (1) ×2 uTt ) · · · (V (k−1) ×2 uTt ) ∈ Rl×rk−1 ,
vk+1 := (V (k+1) ×2 uTt ) · · · (V (d) ×2 uTt ) ∈ Rrk ×1 .
Lemma 4.1 motivates us to define a particular set of inputs
such that this upper bound is achieved.
Theorem 4.1 For outputs y1 , . . . , yl described by a MIMO
Volterra system (12) we have that
Definition 4.1 A set of input signals u1 , u2 , . . . , up such
that
pM + d
rank(U ) =
pM
T
y(t) = (vk+1
⊗ uTt ⊗ vk−1 ) vec(V (k) ).
5
(15)
Proof 2 We first rewrite (12) as
downloaded from https://github.com/kbatseli/
MVMALS.
y(t) = vk−1 (V (k) ×2 uTt ) vk+1 .
Algorithm 1 MIMO Volterra ALS Identification
Input: N samples of inputs u1 , . . . , up , outputs y1 , . . . , yl ,
degree d, memory M , TN ranks r1 , . . . , rd−1
Output: TN-cores V (1) , . . . , V (d) that solve Problem 1
Initialize right orthogonal TN-cores V (1) , . . . , V (d) of
prescribed ranks
while termination criterion not satisfied do
for i = 1, . . . , d − 1 do
vec(V (i) ) ← Compute and solve (16)
Vi ← reshape(V (i) , [ri−1 , (pM + 1)ri ])
Compute thin QR decomposition of Vi
V (i) ← reshape(Q, [ri−1 , (pM + 1), ri ])
Vi+1 ← reshape(V (i+1) , ri , (pM + 1)ri+1 ])
V (i+1) ← reshape(RVi+1 , [ri , (pM + 1), ri+1 ])
end for
Repeat the above loop in the reverse order
end while
This equation holds since it is the product of the l × rk−1
matrix vk−1 with the rk−1 × rk matrix (V (k) ×2 uTt ) with
the rk × 1 vector vk+1 , resulting in the l × 1 vector y(t).
We then have that
y(t) = vk−1 (V (k) ×2 uTt ) vk+1 ,
T
= (vk+1
⊗ vk−1 ) vec(V (k) ×2 uTt ),
T
= (vk+1
⊗ vk−1 ) vec(V (k) ×1 Irk−1 ×2 uTt ×3 Irk ),
T
= (vk+1
⊗ vk−1 ) (Irk ⊗ uTt ⊗ Irk−1 ) vec(V (k) ),
T
= (vk+1
⊗ 1 ⊗ vk−1 ) (Irk ⊗ uTt ⊗ Irk−1 ) vec(V (k) ),
T
= (vk+1
⊗ uTt ⊗ vk−1 ) vec(V (k) ),
where the second equation is obtained from (7), the fourth
equation is obtained from using (5) and the final equation
follows from (4). This concludes the proof.
A few remarks on the ALS method are in order.
The importance of Theorem 4.1 lies in the fact that it describes how the reduced linear system to update V (k) can
be constructed. Indeed, suppose we want to update V (k)
and keep all other TN-cores fixed. By using (15) for t =
0, . . . , N − 1 the following linear system is obtained
vec(Y T ) = Uk vec(V (k) ),
• The iterations can be terminated when the solution does
not exhibit any further improvement or when a certain
fixed maximal number of sweeps has been executed. In
our implementation we set a tolerance on the relative
residual ||Y − Ŷ ||2 /||Y ||2 , where Ŷ is the simulated
output from the obtained solution.
• It is proved in [11, 21] that under certain conditions the
ALS method enjoys strictly monotonous convergence.
Convergence to the unique minimal norm solutions cannot be guaranteed however.
• Also in [11, p. A701], it is proved that the QR decomposition step ensures that the condition number of each Uk
matrix in (16) is upper bounded by the condition number
of the large U matrix. This ensures the numerical stability
of the ALS method.
(16)
where Uk is a lN × rk−1 (pM + 1)rk matrix. Next to the
obvious savings in computational complexity compared to
solving (14), this effectively requires much fewer samples
N to perform the identification. Indeed, one only has to
make sure that lN ≥ rk−1 (pM + 1)rk . Computing the minimal norm solution of (16) requires the computation of the
pseudo-inverse of Uk , which requires O(4N l(rk−1 (pM +
1)rk )2 + 8(rk−1 (pM + 1)rk )3 ) computations.
Numerical stability of the ALS algorithm is guaranteed by
the use of an orthogonalization step [11, p. A690]. The key
idea is that all TN-cores are initialized to be right orthogonal
and are kept orthogonal during each step. After updating
V (1) by solving (16), this tensor is reshaped into a r0 (pM +
1) × r1 matrix from which a thin QR decomposition [9,
p. 230] is computed. This takes O(2l(pM + 1)r12 + 2r13 /2)
flops. The orthogonal matrix Q is then chosen as a new left
orthogonal V (1) TN-core. The corresponding R matrix is
then absorbed into the V (2) core by reshaping the core into a
r1 ×(pM +1)r2 matrix V2 and computing R V2 . Next, V (2)
is updated and orthogonalized, after which V (3) is updated
and so forth. When the iterative “sweep” has reached the end
of the TN, it reverses direction and in a similar way all cores
are made right orthogonal by QR decomposition. The whole
algorithm is presented as pseudocode in Algorithm 1. A
Matlab/Octave implementation of Algorithm 1 can be freely
While the ALS method has a small computational complexity, it suffers from the problem that all TN-ranks need to
be specified a priori. Finding a good choice can be quite
difficult when d is large. This is the main motivation for
the development of the Modified Alternating Linear Scheme
(MALS) algorithm, which not only updates the TN-cores
but also adapts the TN-ranks during each step. An additional
benefit is that the TN-cores are guaranteed to be either left
or right orthogonal so that no stabilizing QR step is required
anymore. These benefits, however, come at the cost of a
higher computational complexity.
4.2
Modified Alternating Linear Scheme method
The main idea of the MALS is in fact a simple one, namely,
to update two TN-cores V (k) , V (k+1) at a time by considering them as one “super-core” and keeping all other cores
6
where U , V are orthogonal matrices and S is a diagonal matrix with positive entries s1 ≥ . . . ≥ sq , with q =
min(rk−1 (pM +1), (pM +1)rk+1 ). Its computation requires
2
O(4(rk−1
(pM + 1)3 rk+1 ) + 8(rk−1 (pM + 1)3 rk+1 )2 +
3
9((pM + 1)3 rk+1
) flops. The numerical rank rk can be determined from a given tolerance τ such that s1 ≥ · · · ≥
srk ≥ τ ≥ · · · ≥ sq . If the sweep is going from left to right
then we update V (k) , V (k+1) as
fixed. The updated super-core is then decomposed into one
orthogonal and one non-orthogonal part, which are used as
updates for both V (k) and V (k+1) . This decomposition step
is achieved by computing a singular value decomposition
(SVD) and it is here where the TN-rank rk is updated. This
updating procedure is repeated in the same sweeping fashion
as with the ALS method until the solution has converged.
We now derive the reduced linear system for computing a
new super-core. The entries of the super-core V (k,k+1) are
defined as the contraction of V (k) with V (k+1) by summing
over all possible rk values
(k,k+1)
V i1 [i2 i3 ]i4 =
rk
X
(k)
V (k) := reshape(U , [rk−1 , (pM + 1), rk ]),
V (k+1) := reshape(SV T , [rk , (pM + 1), rk+1 ]),
which makes V (k) left orthogonal. If the sweep is going
from right to left then the updates are
(k+1)
V i1 i2 j V ji3 i4 .
j=1
V (k) := reshape(U S, [rk−1 , (pM + 1), rk ]),
The square brackets around i2 i3 indicate that these indices
need to be interpreted as one single multi-index such that
V (k,k+1) is an rk−1 × (pM + 1)2 × rk+1 3-way tensor.
Observe that the formation of the super-core has removed
all information on rk . It is now straightforward to verify that
V (k+1) := reshape(V T , [rk , (pM + 1), rk+1 ]),
where now V (k+1) is right orthogonal. The pseudocode
for the MALS algorithm is given in Algorithm 2. A Matlab/Octave implementation of Algorithm 2 can also be freely
downloaded from https://github.com/kbatseli/
MVMALS.
(V (k) ×2 uTt )(V (k+1) ×2 uTt ) = V (k,k+1) ×2 (uTt ) 2
holds. Using the same notation as in the ALS method we
now derive the reduced linear system for the MALS.
Algorithm 2 MIMO Volterra MALS Identification
Input: N samples of inputs u1 , . . . , up , outputs y1 , . . . , yl ,
degree d, memory M
Output: TN-cores V (1) , . . . , V (d) that solve Problem 1
Initialize right orthogonal TN-cores V (1) , . . . , V (d) of
ranks 1
while termination criterion not satisfied do
for i = 1, . . . , d − 1 do
vec(V (i,i+1) ) ← Compute and solve (18)
Vi,i+1 ← reshape(V (i,i+1) , [ri−1 (pM +1), (pM +
1)ri+1 ])
Compute the SVD of Vi,i+1
Determine numerical rank ri
V (i) ← reshape(U , [ri−1 , (pM + 1), ri ])
V (i+1) ← reshape(SV T , [ri , (pM + 1), ri+1 ])
end for
Repeat the above loop in the reverse order
end while
Theorem 4.2 For outputs y1 , . . . , yl described by a MIMO
Volterra system (12) we have that
T
y(t) = (vk+2
⊗ (uTt ) 2 ⊗ vk−1 ) vec(V (k,k+1) ).
(17)
Proof 3 By rewriting (12) this time as
y(t) = vk−1 (V (k,k+1) ×2 (uTt ) 2 ) vk+2 ,
it can be clearly seen that the rest of the proof is now identical to the ALS case.
Application of Theorem 4.2 for t = 0, . . . , N −1 now results
in the reduced linear system
vec(Y T ) = Uk,k+1 vec(V (k,k+1) ),
The same remarks as in the ALS case apply. The tolerance τ
for the determination of the numerical rank rk can be chosen
such that the error is below a certain threshold. In our implementation we opted for the default choice in Matlab/Octave,
which is τ = s1 max(ri−1 (pM +1), (pM +1)ri+1 ), where
is the machine precision. If the tolerance on the obtained
solution is set too low then the MALS algorithm will have
the tendency to increase the TN-ranks to very high values,
resulting in higher computational complexity. However, it
does not make sense in system identification to require that
the solution interpolates the measured output to a high accuracy (say up to the machine precision), which is essentially
overfitting.
(18)
where Uk,k+1 is an lN × rk−1 (pM + 1)2 rk+1 matrix.
The minimal norm solution of (18) can by computed in
O(4N l(rk−1 (pM + 1)2 rk+1 )2 + 8(rk−1 (pM + 1)2 rk+1 )3 )
flops.
Once the new super-core has been computed, it can be reshaped into an rk−1 (pM +1)×(pM +1)rk+1 matrix Vk,k+1 .
The SVD of Vk,k+1 is
Vk,k+1 = U S V T ,
7
5
Table 1
Run times, maximal TT-rank and number of estimated Volterra
tensor elements for an increasing degree d.
d
Run time [seconds]
max TT-rank (pM + 1)d
Numerical Experiments
In this section we demonstrate the two proposed identification algorithms. All computations were done on an Intel i5
quad-core processor running at 3.3 GHz with 16 GB RAM.
We are not aware of any other publicly available algorithms
that are able to identify high-degree MIMO Volterra systems, say, at d = 10.
5.1
Decaying multi-dimensional exponentials
First, we demonstrate the validity of Algorithms 1 and 2 by
means of an artificial SISO example. Symmetric Volterra
kernels were generated up to degree d = 10 for a fixed
memory M = 7 and containing exponentially decaying coefficients. The ith symmetric Volterra kernel hi contains the
entries
U †y
ALS
MALS
2
0.017
0.240
0.077
6
64
3
0.253
0.223
0.251
8
512
4
39.36
1.771
0.415
8
4096
5
NA
4.288
0.587
8
32768
6
NA
3.065
0.791
8
262144
7
NA
6.783
0.961
8
2097152
8
NA
13.11
1.199
8
16777216
9
NA
14.42
1.384
8
134217728
10
NA
18.37
1.576
8
1.0737e9
hi (k1 , . . . , ki ) = exp (−k12 − k22 − · · · − ki2 ).
Each of the ki indices attain the values 0, 0.1, 0.2, . . . , 0.6.
For each degree d a random input signal of 5000 samples, uniformly distributed over the interval [0, 1], was generated. The Volterra kernel was estimated by solving (14)
directly using the pseudoinverse of U where possible. The
MALS algorithm was used to determine a solution for which
||y − ŷ||2 /||y||2 < 1e−4 was satisfied. The TT-ranks determined by the MALS algorithm were then used to determine a solution with the same relative accuracy using
the ALS method. Both the MALS and ALS algorithms always ran with the first 700 samples of the input and output.
Table 1 lists the run times in seconds for the three methods, the maximal TT-rank and the total number of identified
Volterra kernel elements for each degree. Starting from degree d = 5 it was not possible anymore to obtain the Volterra
tensor directly since this would require the inversion of a
32768 × 32768 matrix. From d = 3, the maximal TT-rank
stabilizes to 8 for all TT-cores. Even though the ALS method
exhibits low computational complexity, its convergence is
much slower than MALS, resulting in larger run times. The
obtained solutions were validated by simulating the output
using the 4300 remaining input points. Fig. 3 shows both
the real and simulated output from the MALS solution for
samples 1640 up to 1700. The two output signals are almost indistinguishable. No difference between the ALS and
MALS solutions could be observed.
5.2
Fig. 3. Real and simulated output from the MALS solution with
d = 10, M = 7.
are added to the measured RF output, generating signals with
signal-to-noise (SNR) ratios ranging from 11dB up to 25dB.
The first 700 samples of the inputs and the noisy output are
then used to identify an M = 2, d = 11 two-input oneoutput Volterra system using the ALS method. The Volterra
kernel consists in this case of 9765625 entries. The TT-ranks
are all fixed to 2M + 1 = 5. The identified models were
then used to simulate the remaining 4300 samples of the
output. The SNR of the simulated output was computed by
comparing the simulated output with the original noiseless
output. Table 2 lists the SNR of the signals used in the
identification (ID SNR), the relative residual of the simulated
output, the run time of the identification in seconds and the
SNR of the simulated signal (SIM SNR). As expected, a
gradual improvement of the relative residual can be seen
as the SNR of the signals used for identification increases.
Although the residual remains high throughout the different
SNR levels, the SNR of the simulated output is much better,
with a consistent increase of 11dB. The run time varies
between 2 and 6 seconds. Fig. 4 shows the simulated output
Double balanced mixer
In this example we consider a double balanced mixer used
for upconversion. The output radio-frequency (RF) signal is
determined by a 100Hz sine low-frequency (LO) signal and
a 300Hz square-wave intermediate-frequency (IF) signal. A
phase difference of π/8 is present between the LO and IF
signals. All time series were sampled at 5 kHz for 1 second.
We investigate the effect of additive output noise on the
identified models. We define 5 different noise levels which
8
6
This article presented two new and remarkably efficient
identification algorithms for high-order MIMO Volterra systems. The identification problem was rephrased in terms of
the Volterra tensor, which is never explicitly constructed but
instead always stored in the highly economic TT format.
Both proposed identification algorithms are iterative, starting from an initial orthogonal guess for the TT-cores and
updating them until a desired accuracy is acquired. The algorithms are guaranteed to monotonically converge and numerical stability is ensured by retaining orthogonality in the
TT-cores. The efficiency of both identification algorithms
was demonstrated by numerical examples, where reliable
MIMO Volterra systems of degrees 10 were estimated in
only a few seconds. Even though its computational complexity is lower, the ALS method was found to converge slower
than the MALS algorithm when producing solutions with
the same accuracy. The MALS method showed a tendency
to increase the TT-ranks under the presence of high noise
levels. Extending the robustness of these methods to noisy
data will be the subject of further research.
Fig. 4. Real and simulated output from the ALS identified models
under different SNRs.
on the validation data for three Volterra models identified
under three different SNR levels (11dB, 16dB, 25 dB).
Next, the MALS algorithm was run on the same data to
also identify M = 2, d = 11 Volterra models. Based on
the ALS results we set the tolerance on the relative residual
to 0.5. This resulted in all TT-ranks being 5 for all cases.
Table 3 lists the SNR of the signals used in the identification
(ID SNR), the relative residual of the simulated output, the
run time of the identification in seconds and the SNR of
the simulated signal (SIM SNR). The relative residuals and
SNRs of the simulated output are very close to the results
obtained by the ALS identification. Just as in Example 5.1,
MALS is able to finish the identification faster than the ALS
method. Furthermore, the run time does not vary as much
as in the ALS case. Lowering the tolerance for the data with
low SNR resulted in the MALS method increasing the TT
ranks significantly, up to the point that the 5000 samples
were not sufficient anymore. This indicates a tendency of
the MALS method to overfit.
Table 2
ALS identification for 5 different SNR levels.
11dB 13dB 16dB 19dB
ID SNR
References
[1] K. Batselier, Z. Chen, H. Lui, and N. Wong. A tensor-based Volterra
series black-box nonlinear system identification and simulation
framework. In 2016 IEEE/ACM International Conference on
Computer-Aided Design (ICCAD), 2016.
[2] R. J. G. B. Campello, G. Favier, and W. C. do Amaral. Optimal
expansions of discrete-time Volterra models using Laguerre functions.
Automatica, 40(5):815 – 822, 2004.
[3] J. D. Carroll and J. J. Chang. Analysis of individual differences
in multidimensional scaling via an n-way generalization of “EckartYoung” decomposition. Psychometrika, 35(3):283–319, 1970.
[4] C.H. Cheng and E.J. Powers. Optimal Volterra kernel estimation
algorithms for a nonlinear communication system for PSK and QAM
inputs. IEEE Transactions on Signal Processing, 49(1):147–163,
2001.
[5] V. de Silva and L.H. Lim. Tensor rank and the ill-posedness of
the best low-rank approximation problem. SIAM J. on Matrix Anal.
Appl., 30(3):1084–1127, 2008.
25dB
||y−ŷ||2
||y||2
.255
.208
.151
.105
.052
Run time
2.3s
5.3s
6.4s
3.2s
2.2s
SIM SNR
22dB
24dB
27dB
30dB
37dB
[6] C. Diouf, M. Telescu, P. Cloastre, and N. Tanguy. On the Use
of Equality Constraints in the Identification of Volterra-Laguerre
Models. IEEE Signal Processing Letters, 19(12):857–860, Dec 2012.
[7] G. Favier, A. Y. Kibangou, and T. Bouilloc. Nonlinear system
modeling and identification using Volterra-PARAFAC models. Int.
J. Adapt. Control Signal Process, 26(1):30–53, January 2012.
[8] Doyle FJ III, Ronald K Pearson, and Babatunde A Ogunnaike.
Identification and control using Volterra models. Springer Science
& Business Media, 2012.
Table 3
MALS identification for 5 different SNR levels with tolerance of
0.5.
11dB 13dB 16dB 19dB 25dB
ID SNR
||y−ŷ||2
||y||2
.251
.214
.152
.105
.055
Run time
1.3s
1.1s
1.1s
1.1s
1.1s
SIM SNR
24dB
25dB
28dB
31dB
37dB
Conclusions
[9] G.H. Golub and C.F. Van Loan. Matrix Computations. The Johns
Hopkins University Press, 3rd edition, October 1996.
[10] R. A. Harshman. Foundations of the PARAFAC procedure: Models
and conditions for an “explanatory” multi-modal factor analysis.
UCLA Working Papers in Phonetics, 16(1):84, 1970.
[11] S. Holtz, T. Rohwedder, and R. Schneider. The alternating linear
scheme for tensor optimization in the tensor train format. SIAM
Journal on Scientific Computing, 34(2):A683–A713, 2012.
9
[12] J. Håstad. Tensor rank is NP-complete. Journal of Algorithms,
11(4):644–654, 1990.
[13] Y. Kajikawa. Subband parallel cascade Volterra filter for linearization
of loudspeaker systems. In 2008 16th European Signal Processing
Conference, pages 1–5, Aug 2008.
[14] T. Katayama.
Subspace Methods for System Identification.
Communications and Control Engineering. Springer London, 2005.
[15] T. Kolda and B. Bader. Tensor decompositions and applications.
SIAM Review, 51(3):455–500, 2009.
[16] M.J. Korenberg and I.W. Hunter. The identification of nonlinear
biological systems: Volterra kernel approaches. Annals of biomedical
engineering, 24(2):250–268, 1996.
[17] E. Mumolo and D. Francescato. Adaptive predictive coding of
speech by means of Volterra predictors. In IEEE Winter Workshop
on Nonlinear Digital Signal Processing, 1993, pages 2.1.4.1–2.1.4.4,
1993.
[18] R. Orus. A Practical Introduction to Tensor Networks: Matrix Product
States and Projected Entangled Pair States. Annals Phys., 349:117–
158, 2014.
[19] I. Oseledets. Tensor-train decomposition. SIAM Journal on Scientific
Computing, 33(5):2295–2317, 2011.
[20] I. V. Oseledets. DMRG approach to fast linear algebra in the TT–
format. Comput. Meth. Appl. Math., 11(3):382–393, 2011.
[21] T. Rohwedder and A. Uschmajew. On local convergence of
alternating schemes for optimization of convex problems in the tensor
train format. SIAM Journal on Numerical Analysis, 51(2):1134–1162,
2013.
[22] L. Tan and J. Jiang. Adaptive Volterra filters for active control of
nonlinear noise processes. IEEE Transactions on Signal Processing,
49(8):1667–1676, Aug 2001.
[23] L. R. Tucker. Some mathematical notes on three-mode factor analysis.
Psychometrika, 31(3):279–311, 1966.
[24] P. Wambacq and W.M. Sansen. Distortion Analysis of Analog
Integrated Circuits. Kluwer Academic Publishers, Norwell, MA,
USA, 1998.
[25] S R. White. Density matrix formulation for quantum renormalization
groups. Phys. Rev. Lett., 69:2863–2866, Nov 1992.
[26] N. Wiener, J.A. Stratton, and M.I.O. Technology. Nonlinear Problems
in Random Theory. Literary Licensing, LLC, 2013.
10
| 5cs.CE
|
Entropy on abelian groups ∗
arXiv:1007.0533v2 [math.GR] 3 May 2016
Dikran Dikranjan
†
Anna Giordano Bruno
‡§
[email protected]
[email protected]
Dipartimento di Matematica e Informatica,
Dipartimento di Matematica e Informatica,
Università di Udine,
Università di Udine,
Via delle Scienze, 206 - 33100 Udine, Italy
Via delle Scienze, 206 - 33100 Udine, Italy
Dedicated to Professor Justin Peters
Abstract
We introduce the algebraic entropy for endomorphisms of arbitrary abelian groups, appropriately
modifying existing notions of entropy. The basic properties of the algebraic entropy are given, as well
as various examples. The main result of this paper is the Addition Theorem showing that the algebraic
entropy is additive in appropriate sense with respect to invariant subgroups. We give several applications
of the Addition Theorem, among them the Uniqueness Theorem for the algebraic entropy in the category
of all abelian groups and their endomorphisms. Furthermore, we point out the delicate connection of
the algebraic entropy with the Mahler measure and Lehmer Problem in Number Theory.
1
Introduction
Inspired by the notion of entropy invented by Clausius in Thermodynamics in the fifties of the nineteenth
century, Shannon introduced entropy in Information Theory by the end of the forties of the last century.
A couple of years later, Kolmogorov in [31] and Sinai in [42] introduced the measure entropy in Ergodic
Theory. By an appropriate modification of their definition, Adler, Konheim and McAndrew in [2]
obtained the topological entropy htop for continuous self-maps of compact topological spaces. Moreover,
they briefly sketched an idea on how to define an algebraic entropy for endomorphisms of discrete abelian
groups. Weiss in [48] developed this idea and studied the algebraic entropy for endomorphisms of torsion
abelian groups. Later on, Peters in [36] modified this notion of algebraic entropy for automorphisms
of arbitrary abelian groups. The interest in algebraic entropy was renewed after the recent article [16],
where the case of endomorphisms of torsion abelian groups is thoroughly studied.
We introduce some notation in order to discuss in detail the above mentioned notions of algebraic
entropy, with the aim to extend Peters’ definition to arbitrary endomorphisms of abelian groups.
Let G be an abelian group and let φ ∈ End(G). For a non-empty subset F of G and n ∈ N+ , the
n-th φ-trajectory of F is
Tn (φ, F ) = F + φ(F ) + . . . + φn−1 (F ),
∗
MSC2010: 20K30; 37A35; 37B40; 11R06. Keywords: discrete dynamical system, algebraic entropy, group endomorphism,
abelian group, addition theorem, Mahler measure, Lehmer problem.
†
The first named author gratefully acknowledges the FY2013 Long-term visitor grant L13710 by the Japan Society for the
Promotion of Science (JSPS).
‡
The second named author is supported by Programma SIR 2014 by MIUR (project GADYGR, number RBSI14V2LI, cup
G22I15000160008).
§
Work partially supported by the Fondazione Cassa di Risparmio di Padova e Rovigo (Progetto di Eccellenza “Algebraic structures and their applications”) and by the “National Group for Algebraic and Geometric Structures, and Their
Applications” (GNSAGA - INdAM).
1
and the φ-trajectory of F is
X
T (φ, F ) =
φn (F ).
n∈N
For a finite subgroup F of G, the limit
H(φ, F ) = lim
n→∞
log |Tn (φ, F )|
n
exists (see Fact 2.4) and it is the algebraic entropy of φ with respect to F . According to [2, 48], the
algebraic entropy ent of φ is defined as
ent(φ) = sup{H(φ, F ) : F finite subgroup of G}.
Clearly, this notion is more suitable for endomorphisms of torsion abelian groups, as ent(φ) = ent(φ ↾t(G)
), where t(G) is the torsion part of G. We refer to [16] as a convenient reference for the algebraic entropy
ent.
Peters introduced a different notion of algebraic entropy for automorphisms φ of abelian groups,
using instead of Tn (φ, F ) = F + φ(F ) + . . . + φn−1 (F ), the “negative part of the trajectory”
Tn− (φ, F ) = F + φ−1 (F ) + . . . + φ−n+1 (F ) = Tn (φ−1 , F ),
and instead of finite subgroups F of G just non-empty finite subsets F of G. Therefore, Peters’ algebraic
entropy may have non-zero values also for torsion-free abelian groups.
Obviously, this definition can be given using Tn (φ, F ) as well, since Tn− (φ, F ) = Tn (φ−1 , F ). The
approach with the “positive” partial trajectories Tn (φ, F ) has the advantage to be applicable to any
endomorphism φ (whereas Tn− (φ, F ) may be infinite when φ is not injective). Hence, we define the
algebraic entropy h for endomorphisms φ of abelian groups G as follows.
For a non-empty finite subset F of G, the limit
log |Tn (φ, F )|
n→∞
n
H(φ, F ) = lim
(1.1)
still exists (see Lemma 2.2(b)), and we call it the algebraic entropy of φ with respect to F . The algebraic
entropy of φ is
h(φ) = sup{H(φ, F ) : F ∈ [G]<ω },
where [G]<ω denotes the family of all non-empty finite subsets of G.
For endomorphisms of torsion abelian groups, h coincides with ent. More precisely, if G is an abelian
group and φ ∈ End(G), then ent(φ) = ent(φ ↾t(G) ) = h(φ ↾t(G) ).
The definition of h can be extended to the non-abelian case. However, here we prefer to keep the
exposition in the abelian setting, due to the differences and the technical difficulties that naturally appear
considering non-abelian groups. The reader interested in the non-abelian case and in the connection
to the growth rate of finitely generated groups may see [11, 13]. For other related algebraic entropy
functions see [5, 6, 15, 21].
The main result of this paper is the following key additivity property of the algebraic entropy, called
Addition Theorem. The torsion case of this theorem is covered by the Addition Theorem for ent proved
in [16].
Theorem 1.1 (Addition Theorem). Let G be an abelian group, φ ∈ End(G), H a φ-invariant subgroup
of G and φ ∈ End(G/H) the endomorphism induced by φ. Then
h(φ) = h(φ ↾H ) + h(φ).
In terms of the following diagram describing the above situation
H
φ↾H
H
/G
φ
/G
2
/ G/H
φ
/ G/H
(1.2)
the equality (1.2) of the Addition Theorem reads: the algebraic entropy of the mid endomorphism is
the sum of the algebraic entropies of the two side ones. We adopt the notation
ATh (G, φ, H)
for an abelian group G, φ ∈ End(G) and a φ-invariant subgroup H of G, to indicate briefly that “the
equality (1.2) holds for the triple (G, φ, H)”.
A relevant part of the proof of the Addition Theorem is based on the so called Algebraic Yuzvinski
Formula (see Theorem 1.2 below), recently proved in [25]. To state the Algebraic Yuzvinski Formula,
we need to recall the notion of Mahler measure, playing an important role in Number Theory and
Arithmetic Geometry (see [20, Chapter 1] and [29]).
For a primitive polynomial f (t) = a0 + a1 t + . . . + ak tk ∈ Z[t], let α1 , . . . , αk ∈ C be the roots of f (t)
taken with their multiplicity. The Mahler measure of f (t) is
X
log |αi |.
m(f (t)) = log |ak | +
|αi |>1
The Mahler measure of a linear transformation φ of a finite dimensional rational vector space Qn ,
n ∈ N+ , is defined as follows. Let g(t) ∈ Q[t] be the characteristic polynomial of φ. Then there
exists a smallest s ∈ N+ such that sg(t) ∈ Z[t] (so sg(t) is primitive). The Mahler measure of φ is
m(φ) = m(sg(t)).
Theorem 1.2 (Algebraic Yuzvinski Formula). For n ∈ N+ and φ ∈ End(Qn ),
h(φ) = m(φ).
The counterpart of this formula for the topological entropy was proved by Yuzvinski in [50] (see
also [33]). A direct proof (independent from the topological case) of the Algebraic Yuzvinski Formula
was recently given in [25], extending results from [46] and using the properties of the Haar measure on
locally compact abelian groups. A weaker form for the case of zero algebraic entropy is proved in [17]
using basic linear algebra.
The Algebraic Yuzvinski Formula gives one of the two “normalization axioms” that entail uniqueness
of h; in fact, Theorem 1.3(e) concerns the values h(φ) = m(φ), where φ ∈ End(Qn ) and n ∈ N+ . The
other one is furnished by the following family of group endomorphisms. For any abelian group K
the (right) Bernoulli shift βK : K (N) → K (N) is defined by (x0 , x1 , x2 , . . .) 7→ (0, x0 , x1 , . . .). Then
h(βK ) = log |K|, with the usual convention that log |K| = ∞ if K is infinite (see Example 2.10). The
value h(βK ), when K runs over all finite abelian groups, is the condition in Theorem 1.3(d).
The Uniqueness Theorem for the algebraic entropy is our second main result, its proof uses the
Addition Theorem and the Algebraic Yuzvinski Formula (see Theorem 1.3 below, where the algebraic
entropy h is considered as a collection h = {hG : G abelian group} of functions hG : End(G) →
R≥0 ∪ {∞} defined by hG (φ) = h(φ) for every φ ∈ End(G)). It extends the Uniqueness Theorem given
in [16] for ent in the class of torsion abelian groups, that was inspired, in turn, by the Uniqueness
Theorem for the topological entropy of Stojanov (see [44]).
If G and H are abelian groups, φ ∈ End(G) and η ∈ End(H), we say that φ and η are conjugated if
there exists an isomorphism ξ : G → H such that η = ξφξ −1 . Sometimes we specify also the isomorphism
ξ by saying that φ and η are conjugated by ξ.
Theorem 1.3 (Uniqueness Theorem). The algebraic entropy h of the endomorphisms of abelian groups
is the unique collection h = {hG : G abelian group} of functions hG : End(G) → R≥0 ∪ {∞} such that:
(a) if φ ∈ End(G) and η ∈ End(H) are conjugated, then hG (φ) = hH (η) (i.e., h is invariant under
conjugation);
(b) if G is an abelian group, φ ∈ End(G) and G is a direct limit of φ-invariant subgroups {Gi : i ∈ I},
then hG (φ) = supi∈I hGi (φ ↾Gi );
(c) AThG (G, φ, H) for every abelian group G, every φ ∈ End(G) and every φ-invariant subgroup H of
G;
3
(d) hK (N) (βK ) = log |K| for any finite abelian group K;
(e) hQn (φ) = m(φ) for every n ∈ N+ and every φ ∈ End(Qn ).
Another application of the Addition Theorem in this paper concerns Lehmer Problem from Number
Theory (see Problem 1.4), posed by Lehmer in 1933, and still open (see [20, 29, 34, 43]). Let
L := inf{m(f (t)) > 0 : f (t) ∈ Z[t] primitive}
(obviously, L = inf{m(f (t)) > 0 : f (t) ∈ Z[t] monic}).
Problem 1.4 (Lehmer Problem). Is L > 0?
Let us see the connection of Lehmer Problem to algebraic entropy. Note first that the finite values
of the algebraic entropy ent belong to the set log N+ := {log n : n ∈ N+ } (see Fact 2.4). Then
inf{ent(φ) > 0 : G abelian group, φ ∈ End(G)} = log 2.
(1.3)
Consider now the infimum of the positive values of the algebraic entropy h:
ε := inf{h(φ) > 0 : G abelian group, φ ∈ End(G)}.
The next theorem shows that the problem of deciding whether ε > 0 is equivalent to Lehmer Problem.
The counterpart of this theorem for the topological entropy is apparently well-known.
Theorem 1.5. The equalities ε = inf{h(φ) > 0 : n ∈ N+ , φ ∈ Aut(Qn )} = L hold.
Finally, we see that a positive solution to Lehmer Problem is equivalent to the realization of the
finite values of the algebraic entropy, in the following sense.
Definition 1.6. Let G be an abelian group and φ ∈ End(G) with h(φ) < ∞. We say that h(φ) realizes
if there exists F ∈ [G]<ω such that h(φ) = H(φ, F ).
Theorem 1.7 (Realization Theorem). The following conditions are equivalent:
(a) L > 0;
(b) h(φ) realizes for every abelian group G and every φ ∈ End(G) with h(φ) < ∞.
An application of the Addition Theorem and the Uniqueness Theorem is the so called Bridge Theorem, given in [10]. In more detail, both Weiss in [48] and Peters in [36] connected the algebraic entropy
to the topological entropy by using Pontryagin duality. For an abelian group G and φ ∈ End(G), we
b the Pontryagin dual of G and by φb : G
b→G
b the dual endomorphism of φ (see [28, 38]). It
denote by G
b
is known that G is compact, and it is also metrizable when G is countable.
Theorem 1.8 (Bridge Theorem). Let G be an abelian group and φ ∈ End(G).
b
(a) [48] If G is torsion, then ent(φ) = htop (φ).
b
(b) [36] If G is countable and φ is an automorphism, then h(φ) = htop (φ).
b remains true for arbitrary abelian groups
It was proved in [10] that the conclusion h(φ) = htop (φ)
and their endomorphisms φ. Two proofs of this result are given in [10], one is based on the Addition
Theorem, the Algebraic Yuzvinski Formula and their counterparts for the topological entropy, the second
uses the Uniqueness Theorem.
For another extension of Theorem 1.8(a) to the case of (totally disconnected) locally compact abelian
groups, see [14]. We do not discuss here this extension as well as similar results from [37, 47], since
this paper is limited to the discrete case; various definitions of algebraic entropy in the general case of
locally compact abelian groups can be found in [37, 46, 47].
Let us mention, finally, another application of the Addition Theorem given in [8], connecting the
algebraic entropy with the growth rate of abelian groups with respect to endomorphisms. For further
details on this and other applications of the Addition Theorem and the Algebraic Yuzvinski Formula
refer to [11, 13, 19, 26] (see also the survey [19]).
4
The paper is organized as follows.
Section 2 contains first of all the proof of the existence of the limit in (1.1) defining the algebraic
entropy h. Then, the basic properties of h are listed and verified (see Lemma 2.7 and Proposition 2.8);
they are counterpart of the properties of ent proved in [16, 48], as well as of the known properties of
the topological entropy. Moreover, basic examples are given, among them the Bernoulli shift.
In Section 3 we consider various properties of the algebraic entropy of endomorphisms φ of torsionfree abelian groups G. We underline the preservation of the algebraic entropy under the extension of
an endomorphism φ of G to the divisible hull of G (see Proposition 3.7). As examples, the algebraic
entropy of the endomorphisms of Z and Q is computed (see Examples 3.1 and 3.11).
The aim of Section 4 is to reduce the computation of the algebraic entropy to the case of appropriately
small groups. We see in a first reduction that this means endomorphisms of countable abelian groups
(see Lemma 4.6). Moreover, we consider the structure of Z[t]-module induced on an abelian group G
by the action of its endomorphism φ, and we denote by Gφ the group G considered as a Z[t]-module
under the action of φ. One can compute the algebraic entropy of φ as the supremum of the algebraic
entropies of the restrictions of φ to the finitely generated Z[t]-submodules of Gφ (see Lemma 4.4). In
particular, Proposition 4.7 reduces the proof of the Addition Theorem to the case when Gφ is a finitely
generated Z[t]-module and G has finite rank.
Section 5 is dedicated to the Addition Theorem. Its proof is exposed in several steps, which are
partial reductions gradually restricting the class of groups. We start by showing ATh (G, φ, t(G)) (see
Proposition 5.4). This allows us to focus on the case of torsion-free abelian groups. Using the properties
established in the previous sections, we reduce to endomorphisms of finite-rank divisible torsion-free
abelian groups, i.e., to endomorphisms φ : Qn → Qn for some n ∈ N+ . The next step (see Proposition 5.8) is to see that φ can be supposed injective (hence, also surjective). The ultimate case of
automorphisms of Qn can be managed through the Algebraic Yuzvinski Formula (see Theorem 5.9).
In Section 6 we prove the Uniqueness Theorem. We give a direct proof of this result, but we show
also how it can be deduced from a theorem by Vámos on length functions (see Definition 6.1 below, and
[35, 45]). The approach based on length functions is powerful and clarifies the role of the five axioms
in the statement of the Uniqueness Theorem. Indeed, the first three axioms witness the fact that the
algebraic entropy is a length function of the category ModZ[t] of all Z[t]-modules, while the fourth and
the fifth are used as normalizations. More precisely, a length function of ModZ[t] is determined by
its values on the cyclic modules of the form Z[t]/p, where p is a prime ideal of Z[t]. The “Bernoulli
normalization axiom” gives the value of the algebraic entropy in the first case, while the “Yuzvinski
axiom” in the second case. A different relation between algebraic entropy and length functions can be
found in [40].
In the final Section 7 we consider the relation of the algebraic entropy to the Mahler measure and
Lehmer Problem. We give the proof of Theorem 1.5, making use of the Addition Theorem. Moreover,
we end the section and the paper by proving the Realization Theorem.
The deduction of the Addition Theorem from the Algebraic Yuzvinski Formula was obtained already
in [7]. But at that time no direct proof of the Algebraic Yuzvinski Formula was available. On the other
hand, its validity for automorphisms of finite dimensional rational vector spaces could be deduced from
its counterpart for the topological entropy and Theorem 1.8(b). As it was underlined in [7], “the value
of the Addition Theorem would be much higher if a purely algebraic proof of the Algebraic Yuzvinski
Formula would be available”. Indeed, we apply the Algebraic Yuzvinski Formula in the proof of the
Addition Theorem and from these two results we deduce the Uniqueness Theorem and the Realization
Theorem. As mentioned above, another important application is given in [10], where Theorem 1.8(a) is
extended to a general Bridge Theorem for all the endomorphisms of arbitrary abelian groups, obtaining
in this way also a completely new proof of Theorem 1.8(b) (the original proof makes a heavy use of
convolutions). Finally, using the Bridge Theorem and the Uniqueness Theorem for the algebraic entropy,
one can obtain also a Uniqueness Theorem for the topological entropy of the endomorphisms of compact
abelian groups (see [10, Corollary 3.3]).
5
Notation and terminology
We denote by Z, N, N+ , Q and R respectively the set of integers, the set of natural numbers, the set of
positive integers, the set of rationals and the set of reals. Let R≥0 = {r ∈ R : r ≥ 0}. For m ∈ N+ , we
use Z(m) for the finite cyclic group of order m.
Let G be an abelian group. With a slight divergence with the standard use, we denote by [G]<ω the
set of all non-empty finite subsets of G. If H is a subgroup of G, we indicate this by H ≤ G. For a
subset X of G we denote by hXi the subgroup of G generated by X. The free rank of G is denoted by
r(G). The subgroup of torsion elements of G L
is t(G), while D(G) denotes the divisible hull of G. For a
cardinal α we denote by G(α) the direct sum α G of α many copies of G.
Moreover, End(G) is the ring of all endomorphisms of G. We denote by 0G and idG respectively the
endomorphism of G which is identically 0 and the identity endomorphism of G. If G is torsion-free and
φ ∈ End(G), we denote by φe : D(G) → D(G) the unique extension of φ to D(G).
If G is an abelian group, φ ∈ End(G) and H is a φ-invariant subgroup of G, we denote by φH ∈
End(G/H) the endomorphism induced by φ on the quotient G/H. When there is no possibility of
confusion, we write simply φ.
If R is a ring, M is an R-module and r ∈ R, we denote by µr : M → M the group endomorphism of
M defined by µr (x) = rx for every x ∈ M . To avoid confusion, we write also µM
r , when more precision
is necessary.
2
Basic properties of entropy
Let G be an abelian group and φ ∈ End(G). For F ∈ [G]<ω , the φ-trajectory T (φ, F ) need not be a
subgroup of G, so let
V (φ, F ) = hφn (F ) : n ∈ Ni = hT (φ, F )i .
This is the smallest φ-invariant subgroup of G containing F . For g ∈ G, we write simply V (φ, g) in
place of V (φ, {g}). So, for F ∈ [G]<ω , one has
X
V (φ, F ) =
V (φ, g).
g∈F
Definition 2.1. Let G be an abelian group and φ ∈ End(G). We say that G is finitely φ-generated by
F ∈ [G]<ω if G = V (φ, F ).
As a start, we prove that the limit in (1.1) defining the algebraic entropy exists.
Lemma 2.2. Let G be an abelian group, φ ∈ End(G) and F ∈ [G]<ω . For every n ∈ N+ let cn =
log |Tn (φ, F )|.
(a) The sequence {cn }n∈N+ is subadditive, i.e., cn+m ≤ cn + cm for every n, m ∈ N+ .
(b) The limit H(φ, F ) = limn→∞
cn
n
exists and H(φ, F ) = inf n∈N+
cn
n .
Proof. (a) By definition
Tn+m (φ, F ) = F + φ(F ) + . . . + φn−1 (F ) + φn (F ) + . . . + φn+m−1 (F )
= Tn (φ, F ) + φn (Tm (φ, F )).
Consequently,
cn+m = log |Tn+m (φ, F )| ≤ log(|Tn (φ, F )| · |Tm (φ, F )|) = cn + cm .
(b) By item (a) the sequence {cn }n∈N+ is subadditive. Then the sequence { cnn : n ∈ N+ } has limit
and limn→∞ cnn = inf n∈N+ cnn by a folklore fact from Calculus, due to Fekete (see [22]).
Remark 2.3. Let G be an abelian group and φ ∈ End(G).
(a) The function H(φ, −) is non-decreasing, that is, H(φ, F ) ≤ H(φ, F ′ ) for every F, F ′ ∈ [G]<ω with
F ⊆ F ′.
6
(b) By item (a), if F is a cofinal subfamily of [G]<ω with respect to inclusion, then h(φ) = sup{H(φ, F ) :
F ∈ F }. In particular, it is possible to calculate h(φ) supposing without loss of generality that
0 ∈ F , that is,
h(φ) = sup{H(φ, F ) : F ∈ [G]<ω , 0 ∈ F }.
The definition of H(φ, F ) in (1.1) shows that the algebraic entropy H(φ, F ) of φ with respect to
F ∈ [G]<ω measures the growth rate of the partial trajectories Tn (φ, F ) when they approximate the
trajectory T (φ, F ). Our next comment concerns the case when T (φ, F ) = V (φ, F ) is a subgroup.
This occurs for example when F is a finite subgroup of G. In such a case, it was proved in [16] that
H(φ, F ) = log m, where m > 0 is a divisor of |F |. The number m can be found, without any computation
of limits, as follows.
Fact 2.4. Let G be an abelian group, φ ∈ End(G) and F is finite subgroup of G. First, we note that
N = T (φ, F ) is a torsion subgroup of G. Moreover, with φ1 = φ ↾N , one has H(φ, F ) = H(φ1 , F ), i.e.,
the computation of H(φ, F ) can be carried out in the torsion φ-invariant subgroup N of G. The equality
N = F + φ(N ) shows that N/φ(N ) is a quotient of the finite subgroup F , so |N/φ(N )| is a divisor of
|F |. It was proved in [9] that H(φ, F ) = log |N/φ(N )|, provided φ is injective. This formula was stated
without a proof and without asking φ to be injective by Yuzvinski in [49]. A first proof was provided
in [19], where it was pointed out that the formula is wrong if φ is not injective (e.g., when φ = 0G and
|F | > 1). Finally, in [9, Lemma 4.1] ker φ ∩ N was proved to be finite and the following more precise
equality was obtained in the general case:
H(φ, F ) = log
T (φ, F )
− log | ker φ ∩ T (φ, F )|.
φ(T (φ, F ))
We consider now the two most natural examples.
Example 2.5. If G is an abelian group, then h(0G ) = h(idG ) = 0.
The equality h(0G ) = 0 is obvious. To show that also h(idG ) = 0, let F = {f1 , . . . , ft } be a
finite subset of G. As Tn (idG , F ) = F + . . . + F for all n ∈ N+ , every x ∈ Tn (idG , F ) can be written
{z
}
|
n
Pt
Pt
t
as x =
i=1 mi fi , for some mi ∈ N with
i=1 mi = n. Clearly, (m1 , . . . , mt ) ∈ {0, 1, . . . , n} , so
t
|Tn (idG , F )| ≤ (n + 1) . Hence, H(idG , F ) = 0, and consequently h(idG ) = 0 by the arbitrariness of
F ∈ [G]<ω .
For G an abelian group and φ ∈ End(G), the hyperkernel of φ is
[
ker φn .
ker∞ φ =
n∈N+
The subgroup ker∞ φ is φ-invariant and also invariant for inverse images. Hence, the induced endomorphism φ : G/ ker∞ φ → G/ ker∞ φ is injective.
Proposition 2.6. Let G be an abelian group and φ ∈ End(G). Then h(φ ↾ker∞ φ ) = 0.
Proof. Let F ∈ [G]<ω with F ⊆ ker∞ φ and assume without loss of generality that 0 ∈ F (see Remark
2.3(b)). There exists m ∈ N+ such that φm (F ) = 0. Consequently, Tn (φ, F ) = Tm (φ, F ) for every n ∈ N
with n ≥ m. Hence, H(φ, F ) = 0, and we conclude that h(φ ↾ker∞ φ ) = 0 by the arbitrariness of F .
In the next lemma we show that h is monotone under taking restrictions to invariant subgroups and
under taking induced endomorphisms on quotients over invariant subgroups.
Lemma 2.7. Let G be an abelian group, φ ∈ End(G) and H a φ-invariant subgroup of G. Then
h(φ) ≥ max{h(φ ↾H ), h(φ)}.
Proof. For every F ∈ [H]<ω , obviously H(φ ↾H , F ) = H(φ, F ), so H(φ ↾H , F ) ≤ h(φ). Hence, h(φ ↾H
) ≤ h(φ).
Now assume that F ∈ [G/H]<ω and F = π(F0 ) for some F0 ∈ [G]<ω , where π : G → G/H is the
canonical projection. Then π(Tn (φ, F0 )) = Tn (φ, F ) for every n ∈ N+ . Therefore, h(φ) ≥ H(φ, F0 ) ≥
H(φ, F ) and by the arbitrariness of F this proves h(φ) ≥ h(φ).
7
The next proposition collects the basic properties of h, which are also typical properties of the known
entropy functions. Indeed, they are inspired by similar properties of the algebraic entropy ent (see [48]
and [16]) and the topological entropy (see [2] and [44]). In the case of h they were proved in [36]
for automorphisms, here we extend them for endomorphisms. We shall refer to them as, respectively,
Invariance under conjugation, Logarithmic Law, Continuity for direct limits and weak Addition Theorem.
Proposition 2.8. Let G be an abelian group and φ ∈ End(G).
(a) If H is another abelian group, η ∈ End(H) and φ and η are conjugated, then H(φ, F ) = H(η, ξ(F ))
for every F ∈ [G]<ω ; in particular, h(φ) = h(η).
(b) For every k ∈ N+ , h(φk ) = kh(φ). If φ is an automorphism, then h(φk ) = |k|h(φ) for every k ∈ Z.
(c) If G is a direct limit of φ-invariant subgroups {Gi : i ∈ I}, then h(φ) = supi∈I h(φ ↾Gi ).
(d) If G = G1 × G2 , φ = φ1 × φ2 with φi ∈ End(Gi ) and Fi ∈ [Gi ]<ω , i = 1, 2, then H(φ, F1 × F2 ) =
H(φ1 , F1 ) + H(φ2 , F2 ); consequently, h(φ1 × φ2 ) = h(φ1 ) + h(φ2 )
Proof. (a) For F ∈ [G]<ω and n ∈ N+ , Tn (η, ξ(F )) = ξ(F )+ξ(φ(F ))+. . .+ξ(φn−1 (F )), as ξφk = φk ξ for
k = 1, 2, . . . , n − 1. Since ξ is an isomorphism, |Tn (φ, F )| = |Tn (η, ξ(F ))|, and so H(φ, F ) = H(η, ξ(F )).
This proves that h(φ) = h(η).
(b) Fix k ∈ N+ . First, we prove the inequality h(φk ) ≤ kh(φ). Let F ∈ [G]<ω , assuming without
loss of generality that 0 ∈ F (see Remark 2.3(b)). Let n ∈ N+ , then Tn (φk , F ) ⊆ Tkn−k+1 (φ, F ) and so
log |Tn (φk , F )|
log |Tkn−k+1 (φ, F )|
≤ lim
n→∞
n→∞
n
n
kn − k + 1
log |Tkn−k+1 (φ, F )|
· lim
= lim
n→∞
n→∞
kn − k + 1
n
log |Tkn−n+1 (φ, F )|
= k lim
= kH(φ, F ) ≤ kh(φ).
n→∞
kn − k + 1
H(φk , F ) = lim
Therefore, h(φk ) ≤ kh(φ).
To check the inequality h(φk ) ≥ kh(φ), let F ∈ [G]<ω and n ∈ N+ . With F1 = Tk (φ, F ) one has
Tn (φk , F1 ) = Tkn (φ, F ). Then
1
log |Tn (φk , F1 )|
log |Tkn (φ, F )|
1
h(φk ) ≥ H(φk , F1 ) = lim
= lim
= H(φ, F ).
n→∞
n→∞
k
k
kn
kn
Therefore, h(φk ) ≥ kh(φ).
Now assume that φ is an automorphism. It suffices to prove that h(φ−1 ) = h(φ). Let F ∈ [G]<ω
and n ∈ N+ . Then Tn (φ−1 , F ) = φ−n+1 (Tn (φ, F )); in particular, |Tn (φ−1 , F )| = |Tn (φ, F )|, as φ is an
automorphism. This yields H(φ−1 , F ) = H(φ, F ), hence h(φ−1 ) = h(φ).
(c) By Lemma 2.7, h(φ) ≥ h(φ ↾Gi ) for every i ∈ I and so h(φ) ≥ supi∈I h(φ ↾Gi ). To check the
converse inequality, let F ∈ [G]<ω . Since G = lim{Gi : i ∈ I} and {Gi : i ∈ I} is a directed family,
−→
there exists j ∈ I such that F ⊆ Gj . Then H(φ, F ) = H(φ ↾Gj , F ) ≤ h(φ ↾Gj ). This proves that
h(φ) ≤ supi∈I h(φ ↾Gi ).
(d) For every n ∈ N+ ,
Tn (φ, F1 × F2 ) = Tn (φ1 , F1 ) × Tn (φ2 , F2 )
Hence,
H(φ, F1 × F2 ) = H(φ1 , F1 ) + H(φ2 , F2 ).
(2.1)
<ω
Consequently, h(φ) ≥ h(φ1 )+h(φ2 ). Since every F ∈ [G] is contained in some F1 ×F2 , for Fi ∈ [Gi ]<ω ,
i = 1, 2, and so H(φ, F ) ≤ H(φ, F1 × F2 ). Therefore, (2.1) proves also that h(φ) ≤ h(φ1 ) + h(φ2 ).
The following easy remark can be deduced directly either from item (a) or from item (b) of Proposition 2.8: if G is an abelian group and φ ∈ End(G), then
h(−φ) = h(φ).
The next is a direct consequence of Proposition 2.8(b).
8
(2.2)
Corollary 2.9. Let G be an abelian group and φ ∈ End(G). Then:
(a) h(φ) = 0 if and only if h(φk ) = 0 for some k ∈ N+ ;
(b) h(φ) = ∞ if and only if h(φk ) = ∞ for some k ∈ N+ .
The next somewhat technical consequence of Proposition 2.8(a) is frequently applied in the following
sections. Let G be an abelian group, φ ∈ End(G) and H, K two φ-invariant subgroups of G. Consider
φH ∈ End(G/H) and let π : G → G/H be the canonical projection. The subgroup π(K) of G/H
is φH -invariant, and π(K) = K + H/H ∼
= K/H ∩ K. Let φ ↾K ∈ End(K/H ∩ K) be the induced
endomorphism, then φH ↾π(K) is conjugated to φ ↾K , so Proposition 2.8(a) gives the equality
h(φH ↾π(K) ) = h(φ ↾K ).
(2.3)
In the following example we compute the algebraic entropy of the Bernoulli shift.
Example 2.10. For any abelian group K,
h(βK ) = log |K|,
with the usual convention that log |K| = ∞, if |K| is infinite.
It is proved in [16] that h(βK ) = ent(βK ) = log |K| for every finite abelian group K. We verify now
that h(βZ ) = ∞. Indeed, let G = Z(N) ; for every prime p, the subgroup pG of G is βZ -invariant, so βZ
induces an endomorphism βZ : G/pG → G/pG. Since G/pG ∼
= Z(p)(N) , and βZ is conjugated to βZ(p)
through this isomorphism, h(βZ ) = h(βZ(p) ) = log p by Proposition 2.8(a). Therefore, h(βZ ) ≥ log p for
every prime p by Lemma 2.7, and so h(βZ ) = ∞.
Assume that K is an infinite abelian group. If K is non-torsion, then K contains a subgroup
C∼
= Z, so K (N) contains the βK -invariant subgroup C (N) isomorphic to Z(N) . Moreover, βK ↾C (N) = βC
is conjugated to βZ . Hence, by Lemma 2.7, Proposition 2.8(a) and the previous part of this example,
h(βK ) ≥ h(βC ) = h(βZ ) = ∞. If K is torsion, then K contains arbitrarily large finite subgroups H.
Consequently, K (N) contains the βK -invariant subgroup H (N) . By Lemma 2.7 and the first part of this
example, h(βK ) ≥ h(βH ) = log |H| for every H. So h(βK ) = ∞.
In [3], extending the notion of Bernoulli shift, for every self-map λ : X → X of a non-empty set
X, the generalized shift σλ of an arbitrary Cartesian power K X of an abelian group K was introduced
by letting σλ ((xi )i∈X ) = (xλ(i) )i∈X for every (xi )i∈X ∈ K X . In case λ is finitely many-to-one, the
subgroup K (X) is σλ -invariant, so one can consider also the generalized shift σλ : K (X) → K (X) . In [3]
and [24], h(σλ ) was computed on direct sums and on direct products, respectively.
Remark 2.11. There is a remarkable connection between entropy and recurrence in the spirit of
Poincaré recurrence theorem. We do not pursue this topic in the present paper, here we only recall
some recent results from [8].
Let G be an abelian group and φ ∈ End(G). An element x ∈ G is quasi-periodic if there exist n > m
in N such that φn (x) = φm (x); φ is locally quasi-periodic if every x ∈ G is a quasi-periodic point of φ.
Moreover, φ is quasi-periodic if there exist n > m in N such that φn (x) = φm (x) for every x ∈ G. If G
is finitely φ-generated and φ is locally quasi-periodic, then φ is quasi-periodic. The set Q(G, φ) of all
quasi-periodic points of φ in G is a φ-invariant subgroup of G containing ker∞ φ.
It follows from Proposition 2.8(b) that either h(φ) = 0 or h(φ) = ∞ whenever φ is quasi-periodic,
but one can prove more. According to [8], there exists a greatest φ-invariant subgroup P = P (G, φ) of
G such that h(φ ↾P ) = 0. Moreover, P (G/P, φ̄) = 0 and P contains Q(G, φ) (so h(φ) = 0 in case φ is
locally quasi-periodic). On the other hand, φ ∈ End(G) is called algebraically ergodic if Q(G, φ) = 0.
Then φ is algebraically ergodic if and only if P is trivial. Therefore, φ ∈ End(G/P ) is algebraically
ergodic and h(φ) = h(φ) (see [8]).
3
Entropy on torsion-free abelian groups
Here we dedicate more attention to the torsion-free abelian groups. We start by computing the algebraic
entropy of the endomorphisms of Z, for the counterpart for Q see Example 3.11.
9
Example 3.1. (a) For every non-zero k ∈ Z, µk ∈ End(Z) has h(µk ) = log |k|. As h(µk ) = h(µ−k )
by (2.2), we may assume without loss of generality that k ∈ N+ .
For k = 1 the conclusion follows from Example 2.5. Assume that k > 1 and let F0 = {0, 1, . . . , k −
1} ∈ [Z]<ω and n ∈ N+ . Every m ∈ N with m < k n can be uniquely written in the form
m = f0 + f1 k + . . . + fn−1 k n−1 with all fi ∈ F0 . Then Tn (µk , F0 ) = {m ∈ N : m < k n }, and
so |Tn (µk , F0 )| = k n . Consequently H(µk , F0 ) = log k and this yields h(µk ) ≥ log k. To prove
the converse inequality h(µk ) ≤ log k, fix m ∈ N+ and let Fm = {0, ±1, ±2, . . . , ±m} ∈ [Z]<ω
and n ∈ N+ . Then |x| ≤ mk n for every x ∈ Tn (µk , Fm ). So |Tn (µk , Fm )| ≤ 3mk n , hence
H(µk , Fm ) ≤ log k. Since each F ∈ [Z]<ω is contained in Fm for some m ∈ N+ , we obtain
h(µk ) ≤ log k.
(b) For n, k ∈ N+ and µk : Zn → Zn , h(µk ) = n log k. This follows from Proposition 2.8(d) and item
(a).
Item (a) of the above example shows a difference with the torsion case. In fact, for an abelian group
G and φ ∈ End(G), φ is integral if there exists f (t) ∈ Z[t] \ {0} such that f (φ) = 0. According to
[16, Lemma 2.2], if G is torsion and φ is integral, then ent(φ) = 0. On the other hand, for k > 1 the
endomorphism µk : Z → Z in (a) of the above example is integral over Z (as µk (x) − kx = 0 for all
x ∈ Z), nevertheless, h(µk ) = log k > 0.
Example 3.2. Fix k ∈ Z and consider the automorphism φ : Z2 → Z2 defined by φ(x, y) = (x + ky, y)
for all (x, y) ∈ Z2 . Then h(φ) = 0.
Let m ∈ N+ and Fm = {0, ±1, ±2, . . . , ±m} × {0, ±1, ±2, . . . , ±m} ∈ [Z2 ]<ω . Every F ∈ [Z2 ]<ω is
contained in Fm for some m ∈ N. Therefore, it suffices to show that H(φ, Fm ) = 0. One can prove
by induction that, for every n ∈ N+ , Tn (φ, Fm ) is contained in a parallelogram with sides 2nm and
nm(2 + nk − k), so |Tn (φ, Fm )| ≤ 2n2 m2 (2 + nk − k). Thus H(φ, Fm ) = 0 for every m ∈ N and so
h(φ) = 0.
Next we give some properties of the algebraic entropy for endomorphisms φ of (finitely φ-generated)
abelian groups.
Lemma 3.3. Let G be an abelian group and φ ∈ End(G). If V (φ, g) has infinite rank for some g ∈ G,
then V (φ, g) is torsion-free and h(φ ↾V (φ,g) ) = h(φ) = h(φ) = ∞, where φ ∈ End(G/t(G)) is the induced
endomorphism.
Proof. Let H := V (φ, g). To prove that H is torsion-free consider the canonical projection π : H →
H/t(H),
where
H/t(H) = π(V (φ, g)) = V (φ, π(g)) is torsion-free and D
has infinite
D m
E
E rank. Hence,
Lm−1 k
φ (π(g)) ∩Tm (φ, hπ(g)i) = 0 for every m ∈ N+ , so Tm (φ, hπ(g)i) = k=0 φ (π(g)) and π(V (φ, g)) =
D k
E
L
∼
φ
(π(g))
= Z(N) . Suppose that x ∈ H is torsion, then π(x) is torsion as well, so π(x) = 0
k∈N
Pm
Pm
i
i
as H/t(H) is torsion-free. If x =
i=0 ki φ (g), then π(x) =
i=0 ki φ (π(g)) = 0. This entails
k0 = . . . = km = 0, so x = 0. Hence, t(H) = 0, and so H is torsion-free.
Since φ ↾H is conjugated to βZ , by Proposition 2.8(a) and Example 2.10 we conclude that h(φ ↾H
) = h(βZ ) = ∞. Therefore, h(φ) = ∞ and h(φ) = ∞ by Lemma 2.7.
The following are consequences of Lemma 3.3.
Corollary 3.4. Let G be a torsion-free abelian group and φ ∈ End(G).
(a) If G is finitely φ-generated, then G has finite rank if and only if h(φ) < ∞.
(b) If h(φ ↾V (φ,g) ) = ∞ for some g ∈ G, then h(φ ↾V (φ,z) ) = ∞ for every z ∈ V (φ, g) \ {0}.
e < ∞, and so h(φ) ≤ h(φ)
e <∞
Proof. (a) If r(G) = n is finite, then D(G) ∼
= Qn . Theorem 1.2 gives h(φ)
by Lemma 2.7.
Now assume that h(φ) < ∞. Since G is finitely φ-generated, then G is a sum of φ-invariant subgroups
of the form V (φ, gi ) for finitely many gi ∈ G. Lemma 3.3 implies that all r(V (φ, gi )) are finite, hence
r(G) is finite as well.
(b) By item (a), h(φ ↾V (φ,g) ) = ∞ implies r(V (φ, g)) infinite. For z ∈ V (φ, g) \ {0}, it is easy to see
that r(V (φ, z)) is infinite as well, and so h(φ ↾V (φ,z) ) = ∞ by Lemma 3.3.
10
By Corollary 3.4(a), if G is a torsion-free abelian group, φ ∈ End(G) and g ∈ G, then
h(φ ↾V (φ,g) ) = ∞ if and only if r(V (φ, g)) is infinite.
Moreover, in view of Corollary 3.4(b), if h(φ ↾V (φ,g) ) = ∞, then for every φ-invariant subgroup H of
V (φ, g) one has the following striking dichotomy:
either H = 0 or h(φ ↾H ) = ∞.
The next is a consequence of Lemma 3.3 and Corollary 3.4 that completes Lemma 3.3 from the point
of view of the Addition Theorem.
Lemma 3.5. Let G be an abelian group and φ ∈ End(G) such that V (φ, g) has infinite rank for some
g ∈ G. Then h(φ) = ∞ and ATh (G, φ, H) for every subgroup H of G.
Proof. As V (φ, g) has infinite rank, h(φ) = ∞ and V (φ, g) is torsion-free by Lemma 3.3. Let H be a
subgroup of G and suppose that V (φ, g) ∩ H 6= 0. As V (φ, g) is torsion-free, we can apply Corollary
3.4(b) to φ ↾V (φ,g) to deduce that h(φ ↾V (φ,g)∩H ) = ∞. By Lemma 2.7, h(φ ↾H ) = ∞ as well. Then
h(φ) = h(φ ↾H ) = ∞ and in particular ATh (G, φ, H).
Assume now that V (φ, g)∩H = 0. Let π : G → G/H be the canonical projection. Therefore, V (φ, g)
projects injectively in the quotient G/H and so π(V (φ, g)) = V (φ, π(g)) has infinite rank. So h(φ) = ∞
by Lemma 3.3. Hence, h(φ) = h(φ) = ∞ by Lemma 2.7 and in particular ATh (G, φ, H).
A subgroup H of an abelian group G is essential if and only if for every x ∈ G \ {0} there exists
k ∈ Z such that kx ∈ H \ {0}. In case G is torsion-free, H is an essential subgroup of G if and only if
G/H is torsion.
Corollary 3.6. Let G be a torsion-free abelian group, φ ∈ End(G) and let H be an essential φ-invariant
subgroup of G. If h(φ) < ∞, then h(φ) = 0.
Proof. Since G/H is torsion, it suffices to see that h(φ ↾(G/H)[p] ) = 0 for every prime p, by [16, Proposition 1.18]. To this end we have to show that every x ∈ (G/H)[p] has finite trajectory under φ. By
Lemma 3.3, V (φ, x) has finite rank, say n ∈ N. Then there exist ki ∈ Z, i = 0, . . . , n, such that
P
n
i
i=0 ki φ (x) = 0. Since G is torsion-free, we can assume without loss of generality that at least one of
P
i
these coefficient is not divisible by p. Now, projecting in G/H, we conclude that ni=0 ki φ (x) = 0 is a
non-trivial linear combination in (G/H)[p]. Hence, x ∈ (G/H)[p] has finite trajectory under φ.
The next result reduces the computation of the algebraic entropy for endomorphisms of torsion-free
abelian groups to the case of endomorphisms of divisible torsion-free abelian groups. Item (b) was
announced without proof by Yuzvinski in [50].
Proposition 3.7. Let G be a torsion-free abelian group and φ ∈ End(G). Then:
e mF ) = H(φ,
e F );
(a) if F ∈ [D(G)]<ω and m ∈ N+ satisfy mF ⊆ G, then H(φ, mF ) = H(φ,
e
(b) h(φ) = h(φ).
e Hence, Proposition 2.8(a) yields H(φ, mF ) =
Proof. (a) The automorphism µm of D(G) commutes with φ.
e
e
H(φ, mF ) = H(φ, F ).
e by Lemma 2.7. By the arbitrariness of F in item (a), we have also
(b) Obviously, h(φ) ≤ h(φ)
e
h(φ) ≥ h(φ).
Proposition 3.7 may fail if the abelian group G is not torsion-free. Indeed, for G = Z(2)(N) and
βZ(2) ∈ End(G) one has h(βZ(2) ) = ent(βZ(2) ) = log 2. On the other hand, for D(G) = Z(2∞ )(N) the
endomorphism βZ(2∞ ) extends βZ(2) and has h(βZ(2∞ ) ) = ent(βZ(2∞ ) ) = ∞ (see Example 2.10).
The properties from the next lemma, frequently used in the sequel, are well known and easy to prove.
Recall that a subgroup H of an abelian group G is pure in G if H ∩ mG = mH for every m ∈ Z. The
purification H∗ of H in G is the smallest pure subgroup of G containing H.
Lemma 3.8. Let G be an abelian group and H a subgroup of G.
11
(a) If H is divisible, then H is pure.
(b) If G is divisible, then H is divisible if and only if H is pure.
(c) If G is torsion-free, then H is pure if and only if G/H is torsion-free.
Now we see that in the torsion-free case it suffices to compute the algebraic entropy in an essential
invariant subgroup. As a by-product, this proves ATh (G, φ, H) for a torsion-free abelian group G,
φ ∈ End(G) and an essential φ-invariant subgroup H of G.
Corollary 3.9. Let G be a torsion-free abelian group, φ ∈ End(G) and H a φ-invariant subgroup of G.
(a) The purification H∗ of H in G is φ-invariant and h(φ ↾H ) = h(φ ↾H∗ ).
(b) If H is essential in G, then h(φ) = h(φ ↾H ). Consequently, ATh (G, φ, H).
Proof. Even if (b) follows from (a) with H∗ = G, we shall prove first (b). Consider the common divisible
hull D of H and G. Then φe : D → D is the (unique) common extension of φ and φ ↾H . Proposition 3.7
e = h(φ ↾H ). For the second assertion note that
applies to the pairs D, G and D, H, giving h(φ) = h(φ)
this equality implies ATh (G, φ, H) in case h(φ) = ∞. If h(φ) < ∞, then h(φ) = 0 by Corollary 3.6 and
so ATh (G, φ, H) follows again.
Now we can deduce (a) from (b). The verification of the first assertion is immediate (see for example
[41, Lemma 3.3(a)]). The second assertion follows from (b) applied to H∗ , φ ↾H∗ ∈ End(H∗ ) and the
φ-invariant subgroup H of H∗ .
The final part of item (b) of this corollary can be stated equivalently as ATh (G, φ, H) whenever G
is torsion-free and G/H is torsion.
The next corollary is another consequence of Proposition 3.7 and Corollary 3.9. It shows that the
verification of the Addition Theorem for torsion-free abelian groups and their pure subgroups can be
reduced to the case of divisible ones.
Corollary 3.10. Let G be a torsion-free abelian group, φ ∈ End(G) and H a pure φ-invariant subgroup
of G. Considering D(H) as a subgroup of D(G), one has
e h(φ ↾H ) = h(φe ↾D(H) ) and h(φ) = h(φ),
e
h(φ) = h(φ),
where φ ∈ End(G/H) and φe ∈ End(D(G)/D(H)). In particular,
e D(H)).
ATh (G, φ, H) ⇔ ATh (D(G), φ,
(3.1)
e and h(φ ↾H ) = h(φe ↾D(H) ).
h(φ) = h(φ)
(3.2)
Proof. Since the purification H∗ of H in D(G) is divisible in view of Lemma 3.8(b), H∗ = D(H) and
H = D(H) ∩ G, by the purity of H in G. By Proposition 3.7,
Let π : D(G) → D(G)/D(H) be the canonical projection. Then π(G) is essential in D(G)/D(H) and
e = h(φe ↾π(G) ) by Corollary 3.9(b). Since G/H ∼
h(φ)
= π(G), and φ is conjugated to φe ↾π(G) by this
e The conclusion follows
isomorphism, h(φ) = h(φe ↾π(G) ) by Proposition 2.8(a). Hence, h(φ) = h(φ).
from this equality and (3.2).
The validity of the right-hand-side of (3.1), i.e., ATh (D, ψ, K), for a finite-rank divisible torsion-free
abelian group D, ψ ∈ End(D) and a pure ψ-invariant subgroup K of D, is established in Theorem 5.9
below.
We conclude the section computing the algebraic entropy of the endomorphisms of Q.
Example 3.11. (a) Let φ ∈ End(Q). Then there exists r ∈ Q such that φ = µr . If r = 0, ±1, then
h(φ) = 0 by Example 2.5 and (2.2). By (2.2) we can assume that r > 0, and applying Proposition
2.8(b) we may assume also that r > 1. Let r = ab , where (a, b) = 1. We prove that h(φ) = log a.
12
To prove that h(φ) ≥ log a, take F0 = {0, 1, . . . , a − 1} and let n ∈ N+ . It is easy to check
n−1
that all sums f0 + f1 ab + . . . + fn−1 abn−1 , with fi ∈ F0 , are pairwise distinct. This shows that
|Tn (φ, F0 )| = an , and so H(φ, F0 ) = log a, witnessing h(φ) ≥ log a.
To prove the inequality h(φ) ≤ log a, note that the subgroup H of Q formed by all fractions
having as denominators powers of b (i.e., the subring of Q generated by 1b ) is essential and φinvariant.
Hence, h(φ) = h(φ ↾H ) by Corollary 3.9(b). Now for any m ∈ N+ consider Fm =
r
± bm : 0 ≤ r ≤ mbm . So Fm = b1mh ∩ [−m, m],i where the interval [−m, m] is taken in H. Let
us observe that φk (Fm ) ⊆
1
bm+k
k
k
∩ −m abk , m abk , consequently
Tn (φ, Fm ) ⊆ M
. . + M}
| + .{z
n
where M =
1
bm+n−1
i
h
n−1
n−1
n−1
∩ −m abn−1 , m abn−1 . Therefore, |Tn (φ, Fm )| ≤ 2nbm+n−1 m abn−1 . Hence,
log |Tn (φ, Fm )| ≤ log 2n + m log b + (n − 1) log a.
Thus H(φ, Fm ) ≤ log a. Since each F ∈ [H]<ω is contained in Fm for some m ∈ N+ , this proves
that h(φ) = h(φ ↾H ) ≤ log a.
(b) For n ∈ N+ and r = ab ∈ Q with a > b > 0, consider µr : Qn → Qn . Applying Proposition 2.8(d)
and item (a) we conclude that h(µr ) = n log a.
(c) Let G be a torsion-free abelian group and consider µk : G → G for some k ∈ N with k > 1. Then
(
r(G) log k if r(G) is finite,
h(µk ) =
∞
if r(G) is infinite.
According to Proposition 3.7, we can assume without loss of generality that G is divisible. Let
Q(α)
α = r(G). Then G ∼
: Q(α) → Q(α) . If α ∈ N, then
= Q(α) , with µk conjugated to µk
(α)
h(µk ) = h(µQ
) = α log k, by item (b) and Proposition 2.8(a). If α is infinite, by Lemma 2.7 and
k
in view of the finite case, h(µk ) > n log k for every n ∈ N. Hence, h(µk ) = ∞.
In item (a) of the above example we have computed explicitly the algebraic entropy of µr : Q → Q,
with r = ab > 1 and (a, b) = 1. One can also apply the Algebraic Yuzvinski Formula; indeed, the unique
eigenvalue of µr is ab > 1, and so the Algebraic Yuzvinski Formula 1.2 gives h(µr ) = log a. This formula
b
was proved by Abramov in [1] for the topological entropy of the automorphisms of Q.
4
Entropy of finitely generated flows
Consider the category AbGrp of all abelian groups and their homomorphisms. One can introduce the
category FlowAbGrp of flows of AbGrp with objects the pairs (G, φ) with G ∈ AbGrp and φ ∈ End(G)
(named algebraic flows in [12]). A morphism u : (G, φ) → (H, ψ) in FlowAbGrp between two algebraic
flows (G, φ) and (H, ψ) is given by a homomorphism u : G → H such that the diagram
φ
G −−−−→
uy
G
u
y
(4.1)
H −−−−→ H
ψ
in AbGrp commutes. Two algebraic flows (G, φ) and (H, ψ) are isomorphic in FlowAbGrp precisely
when the homomorphism u : G → H in (4.1) is an isomorphism in AbGrp.
If G is an abelian group and φ ∈ End(G), then G admits a structure of Z[t]-module with multiplication determined by tx = φ(x) for every x ∈ G; we denote by Gφ the abelian group G seen as a
Z[t]-module under the action of φ. One has an isomorphism of categories
FlowAbGrp ∼
= ModZ[t] ,
13
(4.2)
given by the functor F : FlowAbGrp → ModZ[t] , associating to an algebraic flow (G, φ) the Z[t]module Gφ (see [12, Theorem 3.2]). In particular, for a morphism u : (G, φ) → (H, ψ) in FlowAbGrp ,
F (u) = u : Gφ → Hψ is a homomorphism of Z[t]-modules. In the opposite direction consider the functor
F ′ : ModZ[t] → FlowAbGrp , which associates to M ∈ ModZ[t] the algebraic flow (M, µt ), where we
denote still by M the underling abelian group of the module M . If u : M → N is a homomorphism in
ModZ[t] , then F ′ (u) = u : (M, µt ) → (N, µt ) is a morphism in FlowAbGrp .
Remark 4.1. By the isomorphism (4.2), every function f : FlowAbGrp → R≥0 ∪ {∞} defined on
endomorphisms of abelian groups can be viewed as a function f : ModZ[t] → R≥0 ∪ {∞} by letting
f (M ) = f (M, µt ) = f (µt ).
In particular, this holds for the algebraic entropy, so we can consider
h : ModZ[t] → R≥0 ∪ {∞},
letting h(Gφ ) = h(φ) for every Gφ ∈ ModZ[t] .
Definition 4.2 (see Definition 2.1). Let (G, φ) be an algebraic flow. We say that (G, φ) is a finitely
generated flow if Gφ is a finitely generated Z[t]-module (i.e., G is finitely φ-generated).
Clearly, the supporting group of a finitely generated flow is countable.
Our aim is to reduce the computation of algebraic entropy to the case of finitely generated flows.
We start from basic properties.
Lemma 4.3. Let G be an abelian group, φ ∈ End(G) and assume that G is finitely φ-generated.
(a) If H is a φ-invariant subgroup of G, then H is finitely φ ↾H -generated and G/H is finitely φgenerated.
(b) There exists m ∈ N+ such that m · t(G) = 0; so there exists a torsion-free subgroup K of G such
that G ∼
= K × t(G).
Proof. (a) The first assertion follows from the fact that the ring Z[t] is Noetherian, the second one is
obvious.
(b) By (a) t(G) is finitely generated as a Z[t]-module, that is, t(G) = V (φ, F ′ ) for some F ′ ∈ [t(G)]<ω .
Since F ′ is finite, there exists m ∈ N+ such that mF ′ = 0. Then m · t(G) = 0. The second assertion
follows from a theorem of Kulikov (see [23, Section 27.5]).
For an abelian group G and φ ∈ End(G), let
F(G, φ) := {V (φ, F ) : F ∈ [G]<ω }
be the family of all finitely φ-generated subgroups of G. All these subgroups are clearly φ-invariant.
Lemma 4.4. Let G be an abelian group and φ ∈ End(G). Then G = lim F(G, φ), and so
−→
h(φ) = sup{h(φ ↾N ) : N ∈ F(G, φ)}.
(4.3)
Proof. The family F(G, φ) is a direct system, as N1 + N2 ∈ F(G, φ) for N1 , N2 ∈ F(G, φ). That
h(φ) = sup{h(φ ↾N ) : N ∈ F(G, φ)} follows from Proposition 2.8(c).
Since h is defined “locally”, Lemma 4.4 permits to reduce the computation of the algebraic entropy
to finitely generated flows.
We see now that for every algebraic flow (G, φ) there exists a countable φ-invariant subgroup S of
G such that h(φ) = h(φ ↾S ).
Definition 4.5. Let G be an abelian group and φ ∈ End(G). An entropy support of (G, φ) is a countable
φ-invariant subgroup S of G such that h(φ ↾S ) = h(φ).
Clearly, every countable φ-invariant subgroup of G containing an entropy support of (G, φ) has the
same property. This means that such a subgroup is not uniquely determined. The family S(G, φ) of all
entropy supports of (G, φ) is always non-empty:
14
Lemma 4.6. Let G be an abelian group and φ ∈ End(G). Then there exists an entropy support of
(G, φ).
<ω
Proof.
such that h(φ) = supn∈N H(φ, Fn ). Then
P By (4.3), there exists a subfamily {Fn }n∈N ⊆ [G]
S = n∈N V (φ, Fn ) is a countable φ-invariant subgroup of G such that
h(φ) = sup H(φ, Fn ) = sup H(φ ↾S , Fn ) = h(φ ↾S ).
n∈N
n∈N
Hence, S ∈ S(G, φ).
The next proposition reduces the proof of the Addition Theorem to the case of finitely φ-generated
abelian groups of finite rank.
Proposition 4.7. Let G be an abelian group, φ ∈ End(G) and H a φ-invariant subgroup of G. Denote
by π : G → G/H the canonical projection. Then:
(a) F(H, φ ↾H ) = {L ∩ H : L ∈ F(G, φ)} and F(G/H, φ) = {π(L) : L ∈ F(G, φ)};
(b) there exists a chain {Ln }n∈N in F(G, φ) such that
h(φ) = lim h(φ ↾Ln ), h(φ ↾H ) = lim h(φ ↾Ln ∩H ) and h(φ) = lim h(φ ↾π(Ln ) );
n→∞
n→∞
n→∞
(4.4)
(c) if ATh (N, φ, N ∩ H) for all N ∈ F(G, φ), then ATh (G, φ, H);
(d) if ATh (N, φ, N ∩ H) for all N ∈ F(G, φ) of finite rank, then ATh (G, φ, H).
Proof. (a) The first equality follows from Lemma 4.3(a) and the second is clear.
(b) By Lemma 4.6, there exist S ∈ S(G, φ), S1 ∈ S(H, φ ↾H ) and S2 ∈ S(G/H, φ). We can
assume without loss of generality that S ⊇ S1 and π(S) ⊇ S2 . Then π(S) ∈ S(G/H, φ). Moreover,
K := ker π ↾S = S ∩ H ⊇ S1 , K ∈ S(H, φ ↾H ), and K is a φ-invariant subgroup of S such that
S/K ∼
= π(S). Let φ ↾S ∈ End(S/K), which is conjugated to φ ↾π(S) . By hypothesis and by Proposition
2.8(a),
(4.5)
h(φ) = h(φ ↾S ), h(φ ↾H ) = h(φ ↾K ) and h(φ) = h(φ ↾π(S) ) = h(φ ↾S ).
Let S = {gn : n ∈ N}, and for every n ∈ N let Fn = {g0 , . . . , gn }. Then S is increasing union of the
subsets Fn and consequently of the subgroups Ln = V (φ, Fn ), n ∈ N. For every F ∈ [S]<ω there exists
n ∈ N such that F ⊆ Fn , hence V (φ, F ) ⊆ Ln and this shows that the countable chain {Ln }n∈N is cofinal
in F(S, φ). Since Ln ∩ K = Ln ∩ H for every n ∈ N, it follows from (a) that {Ln ∩ H}n∈N is a countable
cofinal chain in F(K, φ ↾K ); moreover, {π(Ln )}n∈N is a countable cofinal chain in F(π(S), φ ↾π(S) ).
By the cofinality of {Ln }n∈N in F(S, φ), and by Lemma 4.4, one has h(φ ↾S ) = supn∈N h(φ ↾Ln ).
Since {h(φ ↾Ln ) : n ∈ N} is a non-decreasing sequence, this supremum becomes a limit h(φ ↾S ) =
limn→∞ h(φ ↾Ln ). Analogously, h(φ ↾K ) = limn→∞ h(φ ↾Ln ∩K ) and h(φ ↾π(S) ) = limn→∞ h(φ ↾π(Ln) ).
Now the required equalities follow from (4.5).
(c) According to the hypothesis, for every N ∈ F(G, φ),
h(φ ↾N ) = h(φ ↾N ∩H ) + h(φ ↾N ),
where φ ↾N ∈ End(N/(N ∩ H)) is the induced endomorphism. In view of (2.3),
h(φ ↾N ) = h(φ ↾N/(N ∩H) ).
(4.6)
By item (b), there exists a chain {Ln }n∈N in F(G, φ) with (4.4). Moreover, (4.6) (applied to N = Ln )
entails
h(φ) = lim h(φ ↾π(Ln) ) = lim h(φ ↾Ln ).
n→∞
n→∞
As Ln ∈ F(G, φ), one has
h(φ) = lim h(φ ↾Ln ) = lim h(φ ↾Ln∩H ) + lim h(φ ↾Ln ) = h(φ ↾H ) + h(φ),
n→∞
n→∞
n→∞
that is, ATh (G, φ, H).
(d) If r(V (φ, g)) is infinite for some g ∈ G, then ATh (G, φ, H)) by Lemma 3.5. Hence, we can assume
that r(V (φ, g)) is finite for all g ∈ G. Then r(N ) is finite for all N ∈ F(G, φ). Hence, our hypothesis
entails ATh (N, φ, N ∩ H)) for all N ∈ F(G, φ). Now (c) yields ATh (G, φ, H).
15
5
The Addition Theorem
The aim of this section is to give a complete proof of the Addition Theorem.
5.1
Skew products
Let K and H be abelian groups, and let φ1 ∈ End(K) and φ2 ∈ End(H). The direct product π =
φ1 × φ2 : K × H → K × H is defined by π(x, y) = (φ1 (x), φ2 (y)) for every pair (x, y) ∈ K × H. For a
homomorphism s : K → H, the skew product of φ1 and φ2 via s is φ ∈ End(K × H) defined by
φ(x, y) = (φ1 (x), φ2 (y) + s(x)) for every (x, y) ∈ K × H.
We say that the homomorphism s is associated to the skew product φ.
When s = 0 one obtains the usual direct product endomorphism φ = π = φ1 × φ2 .
Identifying H with the φ-invariant subgroup 0 × H of K × H, the endomorphism induced by φ on
K∼
= (K ×H)/H is precisely φ1 . Let us see that the skew products arise precisely in such a circumstance.
Remark 5.1. If G is an abelian group and φ ∈ End(G), suppose to have a φ-invariant subgroup H of
G that splits as a direct summand, that is G = K × H for some subgroup K of G. Let us see that φ is
a skew product. Indeed, let ι : G/H → K be the natural isomorphism and let φ ∈ End(G/H). Denote
by φ1 : K → K the endomorphism φ1 = ιφι−1 of K, and let φ2 = φ ↾H . It follows from the definition
of φ1 that for every x ∈ K there exists an element sφ (x) ∈ H such that φ(x, 0) − (φ1 (x), 0) = (0, sφ (x)).
The map sφ : K → H is a homomorphism with φ(x, 0) = (φ1 (x), sφ (x)) for every x ∈ K. Hence,
φ(x, y) = (φ1 (x), φ2 (y) + sφ (x)) for every (x, y) ∈ G = K × H. Therefore, φ is the skew product of φ1
and φ2 via sφ .
As the next example shows, a natural instance to this effect are the fully invariant subgroups.
Example 5.2. If G is a abelian group and φ ∈ End(G), then t(G) is fully invariant, so necessarily
φ-invariant. According to Remark 5.1, φ is a skew product when t(G) splits in G, that is, G = K × t(G)
where K is a torsion-free subgroup of G.
In the sequel, for a skew product φ : G = K × H → K × H, we denote by φ1 : K → K the
endomorphism of K conjugated to φ, and we let φ2 = φ ↾H . We refer to πφ = φ1 × φ2 as the direct
product associated to the skew product φ. We can extend to G = K × H the homomorphism sφ : K → H
associated to the skew product, defining it to be 0 on H. This allows us to consider sφ ∈ End(G × H)
and to speak of the composition s2φ = 0, as well as φ = πφ + sφ in the ring End(G). In other words,
the difference sφ = φ − πφ measures how much the skew product φ fails to coincide with its associated
direct product πφ .
Proposition 5.3. Let G be an abelian group and φ ∈ End(G). Assume that G = K × T for some
subgroups K and T of G, with T torsion and φ-invariant, and suppose that φ is a skew product such
that sφ (K) is finite. Then:
(a) for every n ∈ N+ we have Tn (πφ , F1 × F2 ) = Tn (φ, F1 × F2 ) for every F1 ∈ [K]<ω and F2 ∈ [T ]<ω
with F2 a subgroup of T containing sφ (K);
(b) h(φ) = h(πφ ), consequently ATh (G, φ, T ).
Proof. (a) We have πφn (F1 × F2 ) = φn1 (F1 ) × φn2 (F2 ) and so Tn (πφ , F1 × F2 ) = Tn (φ1 , F1 ) × Tn (φ2 , F2 )
for every n ∈ N+ .
One can prove by induction that, for every x ∈ K and every n ∈ N+ ,
φn (x, 0) = (φn1 (x), φ2n−1 (sφ (x)) + φ2n−2 (sφ (φ1 (x))) + . . . + φ2 (sφ (φ1n−2 (x))) + sφ (φ1n−1 (x))).
Since sφ (K) ⊆ F2 , we conclude that
φn (x, 0) ∈ (φn1 (x), 0) + (0 × (φ2n−1 (F2 ) + φ2n−2 (F2 ) + . . . + φ2 (F2 ) + F2 )) = (φn1 (x), 0) + (0 × Tn (φ2 , F2 ));
as 0 × Tn (φ2 , F2 ) = Tn (φ, 0 × F2 ) is a subgroup of G, we deduce that
φn (x, 0) ∈ (φn1 (x), 0) + (0 × Tn (φ2 , F2 )) and (φn1 (x), 0) ∈ φn (x, 0) + Tn (φ, 0 × F2 ).
16
(5.1)
Fix m ∈ N and an m-tuple a0 , a1 , . . . , am−1 ∈ F1 . Applying the first and the second part of (5.1) to an
for n = 0, 1, . . . , m − 1, we get respectively
m−1
X
φn (an , 0) ∈
n=0
and
m−1
X
(φn1 (an ), 0) ∈
n=0
m−1
X
(φn1 (an ), 0) + (0 × Tm (φ2 , F2 )) ⊆ Tm (φ1 , F1 ) × Tm (φ2 , F2 )
(5.2)
φn (an , 0) + Tm (φ, 0 × F2 ) ⊆ Tm (φ, F1 × 0) + Tm (φ, 0 × F2 ).
(5.3)
n=0
m−1
X
n=0
In other words, (5.2) and (5.3) yield
Tm (φ, F1 × 0) ⊆ Tm (φ1 , F1 ) × Tm (φ2 , F2 ) and Tm (φ1 , F1 ) × 0 ⊆ Tm (φ, F1 × 0) + Tm (φ, 0 × F2 ). (5.4)
As 0 × Tm (φ2 , F2 ) = Tm (φ, 0 × F2 ) is a subgroup of G, the two parts of (5.4) give respectively
Tm (φ, F1 × F2 ) = Tm (φ, F1 × 0) + Tm (φ, 0 × F2 ) ⊆ Tm (φ1 , F1 ) × Tm (φ2 , F2 ) = Tm (πφ , F1 × F2 )
and
Tm (πφ , F1 × F2 ) = Tm (φ1 , F1 ) × Tm (φ2 , F2 ) ⊆ Tm (φ, F1 × 0) + Tm (φ, 0 × F2 ) = Tm (φ, F1 × F2 ).
These two inclusions prove the required equality Tn (πφ , F1 × F2 ) = Tn (φ, F1 × F2 ).
(b) Let F ∈ [G]<ω . Then F ⊆ F1 × F2 , for some F1 ∈ [K]<ω and F2 ∈ [T ]<ω . By Remark 2.3(b),
we can assume without loss of generality that (0, 0) ∈ F1 × F2 and that F2 is a subgroup of T with
F2 ⊇ sφ (K). The conclusion follows from item (a) and Remark 2.3(b).
To prove the second assertion of (b), consider φ ∈ End(G/T ). By Proposition 2.8(d), h(πφ ) =
h(φ1 ) + h(φ2 ). Moreover, by definition, φ2 = φ ↾T and φ1 is conjugated to φ, so h(φ) = h(φ1 ) by
Proposition 2.8(a).
Now we show that the Addition Theorem holds with respect to the torsion subgroup.
Proposition 5.4. Let G be an abelian group and φ ∈ End(G). Then ATh (G, φ, t(G)).
Proof. By Proposition 4.7(d), we can assume that r(G) is finite and G is finitely φ-generated. By
Lemma 4.3(b) there exists m ∈ N+ such that m · t(G) = 0 and there exists a torsion-free subgroup
K of G such that G ∼
= K × t(G). Since t(G) is a φ-invariant subgroup of G that splits, by Remark
5.1 this gives rise to a skew product, that is, there exists a homomorphism sφ : K → t(G) such
that φ(x, y) = (φ1 (x), φ2 (y) + sφ (x)) for every (x, y) ∈ G; moreover, φ1 ∈ End(K) is conjugated to
φ ∈ End(G/t(G)) by the isomorphism K ∼
= G/t(G), and φ2 = φ ↾t(G) .
To show that sφ (K) is finite, we notice first that K/mK is finite by [4, Theorem 0.1], since K is
torsion-free of finite rank. As sφ (mK) = msφ (K) ⊆ mt(G) = 0, sφ factorizes through the canonical
projection π : K → K/mK, i.e., there exists a homomorphism ψ : K/mK → t(G) such that sφ = ψ ◦ π.
Therefore, sφ (K) = ψ(K/mK) is finite.
Now Proposition 5.3 applies and gives ATh (G, φ, t(G)).
5.2
Addition Theorem with respect to two invariant subgroups
In the sequel ∧ stays for conjunction.
Proposition 5.5. Let G be an abelian group, φ ∈ End(G) and H, K two φ-invariant subgroups of G.
Then
ATh (G, φ, K) ∧ ATh (H, φ ↾H , H ∩ K) ∧ ATh (K, φ ↾K , H ∩ K)∧
∧ ATh (G/H, φH , (H + K)/H) ∧ ATh (G/K, φK , (H + K)/K) =⇒ ATh (G, φ, H).
17
Proof. The situation is described by the following diagrams involving the triples (G, φ, H) and (G/K, φK , (H+
K)/K).
H
φ↾H
/ H
(H + K)/K
φK ↾(H+K)/K
/ (H + K)/K
G
φ
/G
G/K
φK
/ G/K
G/H
φH
/ G/H
(G/K)/((H + K)/K)
φK
/ (G/K)/((H + K)/K)
Our hypotheses imply that:
(i) h(φ) = h(φ ↾K ) + h(φK );
(ii) h(φ ↾H ) = h(φ ↾H∩K ) + h(φ ↾H );
(iii) h(φ ↾K ) = h(φ ↾H∩K ) + h(φ ↾K );
(iv) h(φH ) = h(φH ↾(H+K)/H ) + h(φH );
(v) h(φK ) = h(φK ↾(H+K)/K ) + h(φK ).
The composition of the isomorphisms (G/K)/((H + K)/K) ∼
= G/(H + K) ∼
= (G/H)/((H + K)/H)
commutes with
φK : (G/K)/((H+K)/K) → (G/K)/((H+K)/K) and φH : (G/H)/((H+K)/H) → (G/H)/((H+K)/H),
i.e., φK and φH are conjugated. So Proposition 2.8(a) yields
h(φH ) = h(φK ),
(5.5)
h(φ ↾H ) = h(φK ↾H+K/K ) and h(φ ↾K ) = h(φH ↾H+K/H ).
(5.6)
Moreover, by (2.3), we have
From (i), (iii), (v), (5.5), (5.6), (ii) and (iv), we deduce
h(φ) = h(φ ↾K ) + h(φK )
= (h(φ ↾H∩K ) + h(φ ↾K )) + (h(φK ↾(H+K)/K ) + h(φK ))
= h(φ ↾H∩K ) + h(φH ↾H+K/H ) + h(φ ↾H ) + h(φH )
= (h(φ ↾H∩K ) + h(φ ↾H )) + (h(φH ↾H+K/H ) + h(φH ))
= h(φ ↾H ) + h(φH ).
Therefore, h(φ) = h(φ ↾H ) + h(φH ).
The following corollary can be deduced easily from Proposition 5.5.
Corollary 5.6. Let G be an abelian group, φ ∈ End(G) and H, K two φ-invariant subgroups of G with
H ⊆ K. Then:
(a) ATh (G, φ, H) ∧ ATh (K, φ ↾K , H) ∧ ATh (G/H, φ, K/H) =⇒ ATh (G, φ, K); and
(b) ATh (G, φ, K) ∧ ATh (K, φ ↾K , H) ∧ ATh (G/H, φ, K/H) =⇒ ATh (G, φ, H).
Proposition 5.5 and Corollary 5.6 have the following useful consequences, that are applied below in
the proof of the Addition Theorem.
18
Lemma 5.7. Let G be an abelian group, φ ∈ End(G) and H a φ-invariant subgroup of G. Then
ATh (G, φ, H) if one of the following conditions holds:
(a) G/H is torsion;
(b) H is torsion.
Proof. (a) Our aim is to apply Proposition 5.5 with K = t(G), so we check that its hypotheses are
satisfied:
(i) ATh (G, φ, t(G)) and ATh (H, φ ↾H , t(H)), according to Proposition 5.4;
(ii) ATh (t(G), φ ↾t(G) , t(H)) and ATh (G/H, φH , (H + t(G))/H), because t(G) and G/H are torsion,
so the Addition Theorem for torsion abelian groups from [16] applies;
(iii) ATh (G/t(G), φt(G) , (H+t(G))/t(G)) by Corollary 3.9, as G/t(G) is torsion-free and (H+t(G))/t(G)
is essential in G/t(G), being G/(H + t(G)) torsion as a quotient of the torsion abelian group G/H.
As t(H) = H ∩ t(G), Proposition 5.5 applies to conclude the proof.
(b) The subgroup t(G)/H of G/H is precisely t(G/H), so both ATh (G/H, φ, t(G)/H) and ATh (G, φ, t(G))
by Proposition 5.4. On the other hand, ATh (t(G), φ ↾t(G) , H) as t(G) is torsion (apply the Addition Theorem for the torsion abelian groups from [16]). Now Corollary 5.6(b) applies to the triple
H ⊆ t(G) ⊆ G.
5.3
The torsion-free case
If the abelian group G is torsion-free, then for any φ ∈ End(G) the subgroup ker∞ φ is pure in G. The
next result reduces the computation of the algebraic entropy of endomorphisms of finite-rank torsion-free
abelian groups to the case of injective ones.
Proposition 5.8. Let G be a torsion-free abelian group of finite rank and φ ∈ End(G). Then h(φ) =
h(φker∞ φ ). Consequently, ATh (G, φ, ker∞ φ).
Proof. We prove that h(φ) = h(φ), then this implies immediately ATh (G, φ, ker∞ φ), since h(φ ↾ker∞ φ
) = 0 by Proposition 2.6.
Suppose first that G is divisible. Since G has finite rank, ker∞ φ has finite rank as well. Then
there exists n ∈ N, such that ker∞ φ = ker φn . Let γ = φn and γ ∈ End(G/ ker γ) the endomorphism
n
induced by γ. Then h(γ) = nh(φ) by Proposition 2.8(b). Since γ = φ , it follows that h(γ) = nh(φ)
again by Proposition 2.8(b). So, if we prove that h(γ) = h(γ), it follows that h(φ) = h(φ). Note that
γ 2 (G) = γ(G).
This shows that we can suppose without loss of generality that ker∞ φ = ker φ and φ2 (G) = φ(G).
Let φ ∈ End(G/ ker φ). From [27, Section 58, Theorem 1] it follows that
G∼
= ker φ × φ(G).
Proposition 2.6 gives h(φ ↾ker φ ) = 0 and so h(φ) = h(φ ↾φ(G) ) by Proposition 2.8(d). Since φ(G) ∼
=
G/ ker φ, and φ ↾φ(G) and φ are conjugated by this isomorphism, h(φ ↾φ(G) ) = h(φ) by Proposition
2.8(a). Hence, h(φ) = h(φ).
We consider now the general case. Since ker∞ φe is pure in D(G), it is divisible by Lemma 3.8(b).
e i.e., there exists n ∈ N+ such that
e Indeed, let x ∈ ker∞ φ,
Moreover, ker∞ φ is essential in ker∞ φ.
n
e
φ (x) = 0. Since G is essential in D(G) there exists k ∈ Z such that kx ∈ G \ {0}. Furthermore,
φn (kx) = φen (kx) = k φen (x) = 0 and so kx ∈ ker∞ φ \ {0}. It follows that ker∞ φe = D(ker∞ φ). By the
e = h(φ),
e is the induced endomorphism. By
e where φe ∈ End(D(G)/ ker∞ φ)
first part of the proof, h(φ)
e
e
Proposition 3.7 h(φ) = h(φ) and by Corollary 3.10 h(φ) = h(φ). Hence, h(φ) = h(φ).
Our next aim is to prove the Addition Theorem for torsion-free abelian groups (see Proposition 5.10).
We start from the divisible case in Theorem 5.9, where we apply the Algebraic Yuzvinski Formula.
Theorem 5.9. Let n ∈ N+ , φ ∈ End(Qn ) and let H be a pure φ-invariant subgroup of Qn . Then
h(φ) < ∞ and ATh (Qn , φ, H).
19
Proof. Let D = Qn . Theorem 1.2 implies h(φ) < ∞.
Assume first that φ is an automorphism. Let r(H) = k ∈ N, that is, H ∼
= Qk . Let B =
{v1 , . . . , vk , vk+1 , . . . , vn } be a basis of D over Q such that BH = {v1 , . . . , vk } is a basis of H over
Q. Then the matrix of φ with respect to B has the following block-wise form:
A1 B
,
A=
0 A2
where A1 is the matrix of φ ↾H with respect to BH . Let π : D → D/H be the canonical projection
and φ ∈ End(D/H) the endomorphism induced by φ. Then B = {π(vk+1 ), . . . , π(vn )} is a basis of
D/H ∼
= Qn−k and A2 is the matrix of φ with respect to B. Let α1 , . . . , αk be the eigenvalues of A1 and
let αk+1 , . . . , αn be the eigenvalues of A2 . Then α1 , . . . , αn are the eigenvalues of A.
Let χ and χ1 , χ2 ∈ Q[t] be the characteristic polynomials of A and A1 , A2 respectively. Then
χ = χ1 χ2 . Let s1 and s2 be the least common multiples of the denominators of the coefficients of χ1
and χ2 respectively. This means that p1 = s1 χ1 and p2 = s2 χ2 ∈ Z[t] are primitive. By Gauss Lemma
p = p1 p2 is primitive and so for s = s1 s2 the polynomial p = sχ ∈ Z[t] is primitive. Now the Algebraic
Yuzvinski Formula 1.2 applied to φ, φ ↾H , φ gives
X
log |αi |
h(φ) = log s +
1≤i≤n,|αi |>1
X
= log(s1 s2 ) +
log |αi | +
1≤i≤k,|αi |>1
= log s1 +
X
1≤i≤k,|αi |>1
= h(φ ↾H ) + h(φ).
X
log |αi |
k+1≤i≤n,|αi |>1
log |αi | + log s2 +
X
k+1≤i≤n,|αi |>1
log |αi |
This proves ATh (Qn , φ, H).
Consider now the general case of φ ∈ End(G) and let π : D → D/ ker∞ φ be the canonical projection.
Then π(H) = (H + ker∞ φ)/ ker∞ φ is a φ-invariant pure (i.e., divisible) subgroup of D/ ker∞ φ and we
have the following two diagrams.
H
φ↾H
/ H
π(H)
D
φ
/D
D/ ker∞ φ
D/H
φH
/ D/H
(D/ ker∞ φ)/π(H)
φ↾π(H)
/ π(H)
φ
/ D/ ker∞ φ
φπ(H)
/ (D/ ker∞ φ)/π(H)
To establish ATh (D, φ, H) we intend to apply Proposition 5.5. To this end we check in the sequel
the validity of its hypotheses:
(i) ATh (D, φ, ker∞ φ), ATh (H, φ ↾H , H ∩ker∞ φ) and ATh (D/H, φH , (H +ker∞ φ)/H) by Proposition
5.8, as H ∩ ker∞ φ = ker∞ φ ↾H and (H + ker∞ φ)/H = ker∞ φ;
(ii) ATh (ker∞ φ, φ ↾ker∞ φ , H ∩ ker∞ φ), as h(φ ↾ker∞ φ ) = 0 by Proposition 2.6;
(iii) ATh (D/ ker∞ φ, φ, (H + ker∞ φ)/ ker∞ φ) by the the first part of the proof (as φ is bijective, being
injective).
Now Proposition 5.5 yields that ATh (D, φ, H).
The next result settles the torsion-free case of the Addition Theorem.
20
Proposition 5.10. Let G be a torsion-free abelian group, φ ∈ End(G) and let H be a φ-invariant
subgroup of G. Then ATh (G, φ, H).
Proof. By Proposition 4.7(d), we can assume that G has finite rank. The purification H∗ of H is
φ-invariant too. We are going to deduce ATh (G, φ, H) from Corollary 5.6(b) applied to the chain
H ⊆ H∗ ⊆ G. The validity of ATh (G, φ, H∗ ) is granted by Theorem 5.9 and Corollary 3.10. Since H is
an essential subgroup of H∗ , Corollary 3.9(a) yields ATh (H∗ , φ ↾H∗ , H). Finally, ATh (G/H, φ, H∗ /H)
follows from Proposition 5.4, as H∗ /H = t(G/H).
Now we can prove the Addition Theorem:
Proof of Theorem 1.1. One has ATh (G, φ, t(G)) and ATh (H + t(G), φ ↾H+t(G) , t(G)) by Proposition
5.4, and ATh (G/t(G), φ, H + t(G)/t(G)) by Proposition 5.10. Then
ATh (G, φ, H + t(G))
by Corollary 5.6(a) applied to the triple t(G) ⊆ H + t(G) ⊆ G. Since H + t(G)/H is torsion, Lemma
5.7(a) yields
ATh (H + t(G), φ ↾H+t(G) , H).
Finally, as the subgroup (H + t(G))/H of G/H is torsion, by Lemma 5.7(b)
ATh (G/H, φ, (H + t(G))/H).
Now Corollary 5.6(b) applies to the triple H ⊆ H + t(G) ⊆ G to conclude the proof.
6
The Uniqueness Theorem
We start this section by proving the Uniqueness Theorem for the algebraic entropy h in the category of all
abelian groups. In other words, we have to show that, whenever a collection h∗ = {h∗G : G abelian group}
of functions h∗G : End(G) → R≥0 ∪ {∞} satisfies (a)–(e) of Theorem 1.3, then h∗ (φ) = h(φ) for every
abelian group G and every φ ∈ End(G).
We shall provide two proofs of this fact, but first we need to point out three easy consequences of
the hypotheses (a)–(e) of Theorem 1.3 to be used in both proofs.
(A) Item (c) implies that h∗ is monotone with respect to taking restrictions to invariant subgroups
and to taking induced endomorphisms of the quotient groups with respect to invariant subgroups.
(B) Item (d) says that h∗ and h coincide on all Bernoulli shifts βK , where K is a finite abelian
group. This can be extended also to βZ , by showing that h∗ (βZ ) = ∞ as follows. Let G = Z(N) .
For every prime p, the subgroup pG of G is βZ -invariant, so βZ induces an endomorphism βZ :
G/pG → G/pG. Since G/pG ∼
= Z(p)(N) and βZ is conjugated to βZ(p) through this isomorphism,
∗
∗
h (βZ ) = h (βZ(p) ) = log p by item (a). Therefore, h∗ (βZ ) ≥ log p for every prime p by (A), and
so h∗ (βZ ) = ∞.
(C) Item (e) means that h∗ and h coincide on all endomorphisms of Qn , when n ∈ N+ varies.
Roughly speaking, (B) and (C) (that are (d) and (e) of Theorem 1.3, respectively) ensure the
coincidence of h and h∗ on all endomorphisms of Qn for every n ∈ N+ and all Bernoulli shifts. We
have to show that along with the properties (a), (b) and (c), this forces h∗ to coincide with h on all
endomorphisms of all abelian groups.
Proof of Theorem 1.3. In the sequel G is an arbitrary abelian group, φ ∈ End(G) and H a φinvariant subgroup of G. To prove that h∗ (φ) = h(φ) we consider various cases depending on the group
G.
(i) If G is torsion, then h∗ (φ) = ent(φ) = h(φ) by the Uniqueness Theorem for ent [16, Theorem
6.1]. Obviously, item (e) becomes vacuous in this case.
(ii) It suffices to consider the torsion-free case. Indeed, assume that h and h∗ coincide on endomorphisms of torsion-free abelian groups. One has h∗ (φ) = h∗ (φ ↾t(G) ) + h∗ (φ) by (c) and h(φ) = h(φ ↾t(G)
21
)+h(φ) by the Addition Theorem 1.1. Since G/t(G) is torsion-free, our hypothesis implies h∗ (φ) = h(φ).
Since h∗ (φ ↾t(G) ) = h(φ ↾t(G) ) by (i), we conclude that h∗ (φ) = h(φ).
(iii) If G is torsion-free and r(G) = n is finite, then h∗ (φ) = h(φ). Indeed, D(G) ∼
= Qn and
D(G)/G is torsion. Let φ ∈ End(D(G)/G) be the endomorphism induced by φe : D(G) → D(G). By
e = h(φ),
e and this value is finite. Since D(G)/G
the Algebraic Yuzvinski Formula 1.2 and by (e) h∗ (φ)
∗ e
∗
∗
is torsion, h (φ) = h(φ) by (i). By (c) h (φ) = h (φ) + h∗ (φ) and by the Addition Theorem 1.1
e = h(φ) + h(φ). Then h∗ (φ) = h∗ (φ)
e − h∗ (φ) = h(φ)
e − h(φ) = h(φ).
h(φ)
(iv) If G is torsion-free, G = V (φ, g) for some g ∈ G and r(G) is infinite, thenLh∗ (φ) = ∞ = h(φ).
The second equality follows from Lemma 3.5. To check the first one, note that G ∼
= Z(N)
= n∈N hφn (g)i ∼
and φ is conjugated to βZ through this isomorphism. Hence, h∗ (φ) = ∞, because h∗ (φ) = h∗ (βZ ) by
(a) and h∗ (βZ ) = ∞ by (B).
(v) If G is torsion-free and G = V (φ, F ) for some F ∈ [G]<ω , then h∗ (φ) = h(φ). To prove this, let
F = {f1 , . . . , fk }. Then G = V (φ, f1 ) + . . . + V (φ, fk ). If r(V (φ, fi )) is finite for every i ∈ {1, . . . , k},
then r(G) is finite as well, and h∗ (φ) = h(φ) by (iii). If r(V (φ, fi )) is infinite for some i ∈ {1, . . . , k},
then h∗ (φ ↾V (φ,fi ) ) = ∞ = h(φ ↾V (φ,fi ) ) by (iv). By (A) and by Lemma 2.7, h∗ (φ) = ∞ = h(φ).
Consider now the general case. By (ii) we can suppose without loss of generality that G is torsionfree. By Lemma 4.4, G = lim{V (φ, F ) : F ∈ [G]<ω }, and by (v), h∗ (φ ↾V (φ,F ) ) = h(φ ↾V (φ,F ) ) for
−→
every F ∈ [G]<ω . Therefore, (b) and Proposition 2.8(c) give h∗ (φ) = supF ∈[G]<ω h∗ (φ ↾V (φ,F ) ) =
supF ∈[G]<ω h(φ ↾V (φ,F ) ) = h(φ).
Note that the Uniqueness Theorem [16, Theorem 6.1] for ent in the torsion case uses one more axiom
that is not present in our list (a)–(e), namely the Logarithmic Law. As shown in [39], in the torsion
case it follows from (a)–(d). We have seen in the Uniqueness Theorem for h that also in the general
case the Logarithmic Law is not among the properties necessary to give uniqueness of h in the category
of all abelian groups, that is, the Logarithmic Law follows automatically from (a)–(e).
It is possible to prove the Uniqueness Theorem also in a less direct way, that is, using a known
theorem by Vámos [45] on length functions. We expose this alternative proof in the remaining part of
the section, considering the algebraic entropy as a function h : ModZ[t] → R≥0 ∪ {∞} (see Remark 4.1).
Let R be a unitary commutative ring, and consider the category ModR of all R-modules and
their homomorphisms. An invariant i : ModR → R≥0 ∪ {∞} is a function such that i(0) = 0 and
i(M ) = i(M ′ ) if M ∼
= M ′ in ModR . For M ∈ ModR , denote by F (M ) the family of all finitely
generated submodules of M .
Definition 6.1. [35, 45] Let R be a unitary commutative ring. A length function L of ModR is an
invariant L : ModR → R≥0 ∪ {∞} such that:
(a) L(M ) = L(M ′ ) + L(M ′′ ) for every exact sequence 0 → M ′ → M → M ′′ → 0 in ModR ;
(b) L(M ) = sup{L(F ) : F ∈ F (M )}.
An invariant satisfying (a) is said additive and an invariant with the property in (b) is called upper
continuous. So a length function is an additive upper continuous invariant of ModR .
Proposition 6.2. If h∗ = {h∗G : G abelian group} is a collection of functions h∗G : End(G) → R≥0 ∪{∞}
satisfying (a)–(e) of Theorem 1.3, then h∗ is a length function of ModZ[t] . In particular, h is a length
function of ModZ[t] .
Proof. In terms of Definition 6.1, h∗ and h are upper continuous and additive invariants of ModZ[t] ,
respectively in view of the properties (a), (b) and (c) of Theorem 1.3, and by Example 2.5, the Addition
Theorem 1.1 and Proposition 2.8(c).
As noted in [45], the values of a length function L of ModR are determined by its values on the
finitely generated R-modules. In case R is a Noetherian commutative ring, L is determined by its values
L(R/p) for prime ideals p of R (see [45, Corollary of Lemma 2]).
22
Second proof of Theorem 1.3. Let h∗ = {h∗G : G abelian group} be a collection of functions h∗G :
End(G) → R≥0 satisfying (a)–(e) of Theorem 1.3. Let R = Z[t]. By Proposition 6.2, h∗ is a length
function of ModR . In view of the above mentioned results from [45], it suffices only to check that
h∗ (R/p) = h(R/p) for all prime ideals p of R.
Let us recall that R has Krull dimension 2, so the non-zero prime ideals p of R are either minimal of
maximal. In particular, if p is a minimal prime ideal of R, then p = (f (t)), where f (t) ∈ R is irreducible
(either f (t) = p is a prime in Z, or f (t) is irreducible with deg f (t) > 0). On the other hand, a maximal
ideal m of R is of the form m = (p, f (t)), where p is a prime in Z and f (t) ∈ R is irreducible modulo p
and deg f (t) > 0.
(i) For p = 0, we prove that h∗ (R) = ∞ = h(R). To this end we have to show that µt : R → R has
h (µt ) = ∞ = h(µt ). The flow (R, µt ) is isomorphic to the flow (Z(N) , βZ ). By (a) and by Proposition
2.8(a) respectively, this yields h∗ (µt ) = h∗ (βZ ) and h(µt ) = h(βZ ). As h∗ (βZ ) = ∞ = h(βZ ) by (B) and
Example 2.10, we are done.
∗
(ii) To see that h∗ (R/m) = h(R/m) for a maximal ideal p = m of R, it suffices to show that R/m is
finite and apply the Uniqueness Theorem for the torsion case [16, Theorem 6.1]. Indeed, m = (p, f (t)),
where p ∈ Z is a prime and f (t) ∈ R is irreducible modulo p. So, R/m ∼
= (Z/pZ)[t]/(fp (t)), where fp (t)
is the reduction of f (t) modulo p. Hence, R/m is finite.
(iii) It remains to see that h∗ (R/p) = h(R/p) when p is a minimal prime ideal of R.
Assume first that p = (p) for some prime p ∈ Z. We show that h∗ (R/p) = log p = h(R/p).
Indeed, R/(p) ∼
= (Z/pZ)[t] and (Z/pZ)[t] ∼
= Z(p)(N) as abelian groups. Moreover, µt : (Z/pZ)[t] →
(Z/pZ)[t] is conjugated to βZ(p) through this isomorphism. By (a), (B) and Proposition 2.8(a), we get
the required equality.
Suppose now that p = (f (t)), where f (t) = a0 + a1 t + . . . + an tn ∈ Z[t] is irreducible (so, primitive)
with deg f (t) = n > 0. Let M = R/p. We verify that h∗ (M ) = h(M ).
Let J be the principal ideal generated by f (t) in Q[t] and D = Q[t]/J. Let π : Q[t] → D be the
canonical projection. Since J ∩ R = p, π induces an injective homomorphism M → D and we can think
without loss of generality that M is a subgroup of D (identifying M with π(R)). Since D ∼
= Qn as
abelian groups and r(M ) ≥ n, M is essential in D. Consider µt : D → D and µt ↾M : M → M . Since
D/M is torsion, h∗ (µt ) = h(µt ) = 0 by the Uniqueness Theorem for the torsion case [16, Theorem
6.1]. By (C), h∗ (µt ) = m(f (t)) = h(µt ) < ∞. Moreover, h∗ (µt ↾M ) = h∗ (µt ) − h∗ (µt ), by (c); while
h(µt ↾M ) = h(µt ) − h(µt ) by the Addition Theorem 1.1. This yields h∗ (µt ↾M ) = h(µt ↾M ), i.e.,
h∗ (M ) = h(M ).
7
7.1
Entropy vs Mahler measure
Computation of entropy via Mahler measure
For an algebraic number α ∈ C, the Mahler measure m(α) of α is the Mahler measure of sf (t), where f (t)
is the minimal polynomial of α over Q and s is the least positive common multiple of the denominators
of f . In order to characterize the algebraic numbers α with m(α) = 0 one need the following:
Theorem 7.1 (Kronecker Theorem). [30] Let f (t) ∈ Z[t] be a monic polynomial with roots α1 , . . . , αk .
If α1 is not a root of unity, then |αi | > 1 for at least one i ∈ {1, . . . , k}.
Corollary 7.2. Let f (t) ∈ Z[t] \ {0} be a primitive polynomial. Then m(f (t)) = 0 if and only if f (t)
is cyclotomic (i.e., all the roots of f (t) are roots of unity). Consequently, if α is an algebraic integer,
then m(α) = 0 if and only if α is a root of unity.
Proof. Let f (t) = a0 + a1 t + . . . + ak tk and deg f (t) = k. Assume that all roots α1 , . . . , αk of f (t) are
roots of unity. Then there exists m ∈ N+ such that every αi is a root of tm − 1. By Gauss Lemma f (t)
divides tm − 1 in Z[t]. Therefore, ak = ±1, and hence m(f (t)) = 0. Suppose now that m(f (t)) = 0.
This entails that f (t) is monic. Moreover, |αi | ≤ 1 for every i ∈ {1, . . . , k}. By Kronecker Theorem 7.1
each αi is a root of unity.
23
This determines completely the case of zero Mahler measure. For the greatest lower bound of the
positive values of the Mahler measure one has Lehmer Problem 1.4, that can be reformulated also as
follows:
Problem 7.3. Is inf{m(α) > 0 : α algebraic integer} > 0?
Obviously, the problem can be formulated equivalently for algebraic numbers, as m(α) ≥ log 2 for
every algebraic number α that is not an algebraic integer.
We prove now that L = ε, that is, that the infimum of the positive values of the Mahler measure
coincides with the infimum of the positive values of the algebraic entropy.
Proof of Theorem 1.5. We start by proving the first equality in Theorem 1.5, namely
ε = inf{h(φ) > 0 : n ∈ N+ , φ ∈ Aut(Qn )}.
(7.1)
We split the proof in three steps of reductions.
(i) Let G be an abelian group, φ ∈ End(G) and φ ∈ End(G/t(G)). Since h(φ ↾t(G) ) = ent(φ ↾t(G) ),
we can suppose that h(φ ↾t(G) ) = 0, otherwise h(φ) ≥ h(φ ↾t(G) ) ≥ log 2 by Lemma 2.7 and (1.3). By
the Addition Theorem 1.1 h(φ) = h(φ). In other words, we can consider only torsion-free abelian groups
G.
(ii) By (i), we can assume that G is a torsion-free abelian group and φ ∈ End(G). If there exists
g ∈ G such that r(V (φ, g)) is infinite, then h(φ) ≥ h(φ ↾V (φ,g) ) = ∞ by Lemma 2.7 and Corollary
3.4(b). Hence, we can assume that r(V (φ, g)) is finite for every g ∈ G. Then r(V (φ, F )) is finite for
every F ∈ [G]<ω . By Lemma 4.4, h(φ) = supF ∈[G]<ω h(φ ↾V (φ,F ) ). So, we can consider only torsion-free
e by Theorem 3.7, and hence we can reduce to
abelian groups of finite rank. Furthermore, h(φ) = h(φ)
divisible torsion-free abelian groups of finite rank.
(iii) By (ii) and by Proposition 2.8(a), we can assume that G = Qn for some n ∈ N+ . Let φ ∈
End(Qn ). By Proposition 5.8 h(φ) = h(φ), where the induced endomorphism φ ∈ End(Qn / ker∞ φ) is
injective (hence, surjective). Moreover, Qn / ker∞ φ ∼
= Qm for some m ∈ N, m ≤ n, as ker∞ φ is pure in
n
n
Q (so Q / ker∞ φ is divisible and torsion-free by Lemma 3.8(c)). Therefore, φ ∈ Aut(Qm ), and this
proves the equality in (7.1).
The equality L = ε follows from (7.1) and the Algebraic Yuzvinski Formula 1.2.
Following [11], for an abelian group G, let Ealg (G) = {h(φ) : φ ∈ End(G)} and let Ealg be the union
of all Ealg (G) when G varies among the class of all abelian groups. Then Ealg is a submonoid of the
monoid (R≥0 ∪ {∞}, +, 0), and L = inf(Ealg \ {0}) by Theorem 1.5. Moreover, let
A = {log |α| : α algebraic number, |α| ≥ 1},
which is a dense countable submonoid of (R≥0 , +, 0). The next theorem collects some equivalent forms
of L = 0 (a proof can be found in [11]).
Theorem 7.4. The following conditions are equivalent:
(a) Ealg = R≥0 ∪ {∞};
(b) L = 0;
S
(c) inf( n∈N+ Ealg (Zn ) \ {0}) = 0.
If L > 0, then Ealg ⊆ A ∪ {∞}, so in particular it is countable.
The inclusion Ealg ⊆ A ∪ {∞} is proper in case L > 0 by the density of the set A in R≥0 .
Remark 7.5. By the Algebraic Yuzvinski Formula 1.2, the algebraic entropy h(φ) of an endomorphism
φ : Qn → Qn , n ∈ N+ , is equal to the Mahler measure of a polynomial f (t) ∈ Z[t], namely f (t) =
sg(t) ∈ Z[t], where g(t) ∈ Q[t] is the characteristic polynomial of φ and s is the least positive common
multiple of the denominators of g(t). One can read this in the opposite direction, that is, the Mahler
24
measure of a polynomial f (t) ∈ Z[t] of degree n is equal to the algebraic entropy of some endomorphism
φ of Qn . Indeed, let f (t) = a0 + a1 t + . . . + an tn ∈ Z[t] with deg f (t) = n and let
0
1
C(f ) =
0
.
..
0
0
0
1
..
.
0
...
...
..
.
..
.
...
0
0
− aan0
a1
− an
a2
− an
..
.
0
..
.
1 − an−1
an
(7.2)
be the companion matrix associated to f (t). The characteristic polynomial of C(f ) is f (t), and C(f )
is associated to an endomorphism φ of Qn (see also Remark 7.6 about this endomorphism). Then
m(f (t)) = h(φ) by the Algebraic Yuzvinski Formula 1.2.
Finally, we present the above situation from a different point of view, namely the algebraic entropy
associated to the multiplication by an element in a commutative ring, defined in [51].
Remark 7.6. Let α ∈ C be an algebraic number of degree n ∈ N+ over Q and let f (t) ∈ Q[t] be its
minimal polynomial, with deg f (t) = n. Let K = Q(α) be the simple field extension of Q associated
to α; then K ∼
= Qn as abelian groups. Inspired by [51], call algebraic entropy h(α) of α the algebraic
entropy h(µα ), where µα is the multiplication by α in K. Let an be the smallest positive integer such
that an f (t) = a0 + a1 t + . . . + an tn ∈ Z[t]. With respect to the basis {1, α, . . . , αn−1 } of K, the matrix
associated to µα is the companion matrix C(f ) in (7.2). The characteristic polynomial of C(f ) is f (t).
By the Algebraic Yuzvinski Formula 1.2, h(µα ) = m(f (t)), i.e., h(α) = m(α). This means that the
algebraic entropy of an algebraic number coincides with its Mahler measure.
7.2
Lehmer Problem and realization of entropy
This section is dedicated to the proof of the Realization Theorem, showing that Lehmer Problem is
equivalent to the “realization problem” of the finite values of the algebraic entropy.
In the next lemma we collect known results showing the validity of the Realization Theorem for
finitely generated flows satisfying some additional restraint.
Lemma 7.7. Let G be an abelian group, φ ∈ End(G) and assume that G is finitely φ-generated by
F ∈ [G]<ω .
(a) [16] If G is torsion, then h(φ) = H(φ, F ).
(b) [8] If H(φ, F ) = 0, then h(φ) = 0 (in particular, h(φ) = H(φ, F )).
The conclusion h(φ) = H(φ, F ) in the above lemma does not hold true in general. Indeed, consider
µ3 : Z → Z and F = {0, 1}; now H(µ3 , F ) ≤ log |F | = log 2, while h(µ3 ) = log 3. This can be generalized
to any non-torsion infinite finitely generated abelian group G with F a finite set of generators of G; in
this case, for every k ∈ N+ with k > |F |, one has h(µk ) ≥ log k > log |F |.
The following is a direct consequence of Proposition 3.7.
e
Corollary 7.8. Let G be a torsion-free abelian group and φ ∈ End(G) with h(φ) < ∞. Then h(φ)
realizes if and only if h(φ) realizes.
e by Proposition 3.7(b), then also h(φ)
e realizes. Assume now
Proof. If h(φ) realizes, since h(φ) = h(φ)
<ω
e
e
e
that h(φ) realizes. Then there exists F ∈ [D(G)]
such that h(φ) = H(φ, F ). By Proposition 3.7 there
e = H(φ,
e F ) = H(φ, mF ), therefore, h(φ) realizes.
exists m ∈ N+ such that mF ⊆ G and h(φ) = h(φ)
The following technical lemma is needed in the proof of Proposition 7.10.
Lemma 7.9. Let G be an abelian group and φ ∈ End(G) with h(φ) < ∞. Let {Vn }n∈N be an increasing
chain of φ-invariant subgroups of G, such that V0 = 0 and h(φ) = supn∈N h(φ ↾Vn ). Let φn = φ ↾Vn and
φn ∈ End(Vn /Vn−1 ) be the induced endomorphism for n ∈ N+ . Assume that η > 0 and that h(φn ) is
either 0 or ≥ η for every n ∈ N+ . Then h(φ) = h(φn ) for some n ∈ N.
25
Proof. For the sake of completeness let φ0 = φ ↾V0 and note that φ1 = φ1 . The sequence {h(φn )}n∈N is
an increasing sequence of non-negative real numbers. For every n ∈ N+ , let
dn = h(φn ) − h(φn−1 ).
By the Addition Theorem 1.1, h(φn ) = h(φn−1 ) + h(φn ) for every n ∈ N+ . Hence, dn = h(φn ), and so
dn is either 0 or ≥ η by hypothesis. Therefore, for every n ∈ N,
h(φn ) =
n
X
di ≥ mn η, where mn = |{di 6= 0 : 1 ≤ i ≤ n}|.
i=1
Since h(φn ) ≤ h(φ) for every n ∈ N, we have that
mn ≤
h(φ)
.
η
Consequently, there are finitely many dn 6= 0, and this means that the sequence {h(φn )}n∈N stabilizes.
By hypothesis, h(φ) = supn∈N h(φn ), hence h(φ) = h(φn ) for every sufficiently large n ∈ N.
From Lemma 7.9 and Corollary 7.8 we deduce respectively item (a) and item (b) of the following
result.
Proposition 7.10. Let G be an abelian group and φ ∈ End(G) with h(φ) < ∞. Then h(φ) realizes
when:
(a) G is torsion;
(b) G is torsion-free of finite rank.
Proof. (a) By Proposition 4.7 there exists a chain {Vn }n∈N in F(G, φ) such that h(φ) = limn→∞ h(φ ↾Vn ).
We can apply Lemma 7.9 with η = log 2 to find n ∈ N such that h(φ) = h(φ ↾Vn ). We have Vn = V (φ, F )
for some F ∈ [G]<ω . Since G is torsion, F ′ = hF i is finite and Vn = V (φ, F ′ ) = T (φ, F ′ ). By Lemma
7.7(a), h(φ ↾Vn ) = H(φ, F ′ ), and hence h(φ) = h(φ ↾Vn ) = H(φ, F ′ ).
(b) If G is a torsion-free abelian group of finite rank, by Corollary 7.8 we can assume that G is
also divisible and then G ∼
= Qn for some n ∈ N+ . By [25, Theorem 4.18] there exist infinitely many
<ω
F ∈ [G]
such that h(φ) = H(φ, F ).
We are now in position to prove the Realization Theorem for the algebraic entropy:
Proof of Theorem 1.7. (b)⇒(a) Assume that L = 0. By Theorem 7.4 this is equivalent to Ealg =
R≥0 ∪ {∞}. Hence, for every
P φn ∈ End(Gn ) such that
Lan abelian group Gn and
L n ∈ N there exist
h(φn ) = 21n . Now let G = n∈N Gn and φ = n∈N φn . Then h(φ) = n∈N 21n = 1 by Proposition
2.8(c,d). To conclude it suffices to note that for every FP∈ [G]<ω there exists n ∈ N such that
n
F ⊆ G0 ⊕ . . . ⊕ Gn , and so H(φ, F ) ≤ h(φ0 ⊕ . . . ⊕ φn ) = i=0 21i < 1 by Proposition 2.8(d). Then
h(φ) = 1, but h(φ) does not realize.
(a)⇒(b) By Proposition 4.7 there exists a chain {Vn }n∈N in F(G, φ) such that h(φ) = limn→∞ h(φ ↾Vn
). As L > 0, by Lemma 7.9 (with η = L) there exists n ∈ N such that h(φ) = h(φ ↾Vn ). Hence, we can
assume without loss of generality that G is finitely φ-generated by some F ∈ [G]<ω .
Since h(φ) < ∞, Lemma 3.3 implies that V (φ, g) has finite rank for every g ∈ G. Then G has finite
rank as well, and by Lemma 4.3(b) there exists a finite-rank torsion-free subgroup K of G such that
G∼
= K × t(G). By Example 5.2, φ is the skew product of φ1 and φ2 , where φ1 : K → K is conjugated
to φ : G/t(G) → G/t(G) by the isomorphism K ∼
= G/t(G), and φ2 = φ ↾t(G) ; moreover, sφ (K) is finite.
By Proposition 7.10(a) there exists F2 ∈ [t(G)]<ω such that h(φ2 ) = H(φ2 , F2 ). We can assume without
loss of generality that sφ (K) ⊆ F2 . Since K ∼
= G/t(G) is torsion-free of finite rank, by Proposition
7.10(b) there exists F1 ∈ [K]<ω such that h(φ1 ) = H(φ1 , F1 ). By Proposition 2.8(d) and Proposition
5.3, for πφ = φ1 × φ2 one has
h(φ) = h(πφ ) = h(φ1 ) + h(φ2 ) = H(φ1 , F1 ) + H(φ2 , F2 ) = H(φ1 × φ2 , F1 × F2 ) = H(φ, F1 × F2 ),
as Tn (πφ , F1 ×F2 ) = Tn (φ, F1 ×F2 ) for every n ∈ N+ . This concludes the proof for the case G = V (φ, F ),
and so the proof of the theorem.
26
References
[1] L. M. Abramov, The entropy of an automorphism of a solenoidal group, Teor. Veroyatnost. i
Primenen 4 (3) (1959) 249–254.
[2] R. L. Adler, A. G. Konheim, M. H. McAndrew, Topological entropy, Trans. Amer. Math. Soc. 114
(1965) 309–319.
[3] M. Akhavin, F. Ayatollah Zadeh Shirazi, D. Dikranjan, A. Giordano Bruno, A. Hosseini, Algebraic
entropy of shift endomorphisms on abelian groups, Quaest. Math. 32 (4) (2009) 529–550.
[4] D. Arnold, Finite Rank Torsion Free Abelian Groups and Rings, L.N.M. n. 931, Springer 1982.
[5] R. Bowen, Entropy and the fundamental group, in: The Structure of Attractors in Dynamical
Systems (Proc. Conf. North Dakota State Univ., Frago, 1977), Lecture Notes in Math. 668 (1978)
21–29.
[6] D. Dikranjan, A. Giordano Bruno, L. Salce, Adjoint algebraic entropy, Journal of Algebra 324
(2010) 442–463.
[7] D. Dikranjan, A. Giordano Bruno, Entropy on abelian groups, arxiv.org/pdf/1007.0533v1.pdf
(2010).
[8] D. Dikranjan, A. Giordano Bruno, The Pinsker subgroup of an algebraic flow, Jour. Pure Appl.
Algebra 216 (2012) 364–376.
[9] D. Dikranjan, A. Giordano Bruno, Limit free computation of entropy, Rend. Istit. Mat. Univ.
Trieste 44 (2012) 297–312.
[10] D. Dikranjan, A. Giordano Bruno, The connection between topological and algebraic entropy, Topology Appl. 159 (13) (2012) 2980–2989.
[11] D. Dikranjan, A. Giordano Bruno, Topological and algebraic entropy for group endomorphisms,
Proceedings ICTA2011 Islamabad, Pakistan July 4-10, 2011 Cambridge Scientific Publishers (2012)
133–214.
[12] D. Dikranjan, A. Giordano Bruno, Entropy in a category, Appl. Categ. Structures 21 (1) (2013)
67–101.
[13] D. Dikranjan, A. Giordano Bruno, Discrete dynamical systems in group theory, Note Mat. 33 (1)
(2013) 1-48.
[14] D. Dikranjan, A. Giordano Bruno, The Bridge Theorem for totally disconnected LCA groups, Topology Appl. 169 (1) (2014) 21–32.
[15] D. Dikranjan, A. Giordano Bruno, L. Salce, S. Virili, Intrinsic algebraic entropy, J. Pure Appl.
Algebra 219 (2015) 2933–2961.
[16] D. Dikranjan, B. Goldsmith, L. Salce, P. Zanardo, Algebraic entropy of endomorphisms of abelian
groups, Trans. Amer. Math. Soc. 361 (2009) 3401–3434.
[17] D. Dikranjan, K. Gong, P. Zanardo, Endomorphisms of abelian groups with small algebraic entropy,
Linear Algebra Appl. 439 (7) (2013) 1894–1904.
[18] D. Dikranjan, I. Prodanov, L. Stoyanov, Topological Groups: Characters, Dualities and Minimal
Group Topologies, Pure and Applied Mathematics, Vol. 130, Marcel Dekker Inc., New York-Basel,
1989.
[19] D. Dikranjan, M. Sanchis, S. Virili, New and old facts about entropy on uniform spaces and topological groups, Topology Appl. 159 (7) (2012) 1916–1942.
[20] G. Everest, T. Ward, Heights of polynomials and entropy in algebraic dynamics, Universitext.
Springer-Verlag London Ltd., London 1999.
[21] K. Falconer, B. Fine, D. Kahrobaei, Growth rate of an endomorphism of a group, Groups Complex.
Cryptol. 3 (2011) no. 2 285–300.
[22] M. Fekete, Über die Verteilung der Wurzeln bei gewisser algebraichen Gleichungen mit ganzzahlingen Koeffizienten, Math. Zeitschr. 17 (1923) 228–249.
[23] L. Fuchs, Infinite Abelian Groups, Vol. I and II, Academic Press, 1970 and 1973.
27
[24] A. Giordano Bruno, Algebraic entropy of shift endomorphisms on products, Comm. Algebra 38 (11)
(2010) 4155–4174.
[25] A. Giordano Bruno, S. Virili, The Algebraic Yuzvinski Formula, J. Algebra 423 (2015) 114–147.
[26] A. Giordano Bruno, S. Virili, About the Algebraic Yuzvinski Formula, Topol. Algebra and its Appl.
3 (1) (2015) 86–103.
[27] P. Halmos, Finite-Dimensional Vector Spaces, Undergraduate Texts in Mathematics, SpringerVerlag, New York - Heidelberg - Berlin, 1987.
[28] E. Hewitt, K. A. Ross, Abstract harmonic analysis I, Springer-Verlag, Berlin-Heidelberg-New York,
1963.
[29] E. Hironaka, What is. . . Lehmer’s number?, Not. Amer. Math. Soc. 56 (3) (2009) 374–375.
[30] L. Kronecker, Zwei Sätze über Gleichungen mit ganzzahligen Coefficienten, Jour. Reine Angew.
Math. 53 (1857) 173–175.
[31] A. N. Kolmogorov, New metric invariants of transitive dynamical systems and automorphisms of
Lebesgue spaces, Doklady Akad. Nauk. SSSR 119 (1958) 861–864 (in Russian).
[32] D. H. Lehmer, Factorization of certain cyclotomic functions, Ann. of Math. 34 (1933) 461–469.
[33] D. Lind, T. Ward, Automorphisms of solenoids and p-adic entropy, Ergodic Theory Dynam. Systems 8 (3) (1988) 411–419.
[34] M. J. Mossinghoff, G. Rhin, Q. Wu, Minimal Mahler measures, Experiment. Math. 17 (4) (2008)
451–458.
[35] D. G. Northcott, M. Reufel, A generalization of the concept of length, Quart. J. of Math. (Oxford)
(2) 16 (1965) 297–321.
[36] J. Peters, Entropy on discrete Abelian groups, Adv. Math. 33 (1979) 1–13.
[37] J. Peters, Entropy of automorphisms on L.C.A. groups, Pacific J. Math. 96 (2) (1981) 475–488.
[38] L. S. Pontryagin, Topological Groups, Gordon and Breach, New York 1966.
[39] L. Salce, Some results on the algebraic entropy, Groups and Model Theory, Contemporary Math.
576 (2012) 297–304.
[40] L. Salce, P. Vámos, S. Virili, Length functions, multiplicities and algebraic entropy, Forum Math.
25 (2) (2013) 255–282.
[41] L. Salce, P. Zanardo, A general notion of algebraic entropy and the rank entropy, Forum Math. 21
(4) (2009) 579–599.
[42] Y. G. Sinai, On the concept of entropy of a dynamical system, Doklady Akad. Nauk. SSSR 124
(1959) 786–781 (in Russian).
[43] C. Smyth, On the product of conjugates outside the unit circle of an algebraic integer, Bull. London
Math. Soc. 3 (1971) 169–175.
[44] L. N. Stoyanov, Uniqueness of topological entropy for endomorphisms on compact groups, Boll. Un.
Mat. Ital. B (7) 1 (3) (1987) 829–847.
[45] P. Vámos, Additive Functions and Duality over Noetherian Rings, Quart. J. of Math. (Oxford) (2)
19 (1968) 43–55.
[46] S. Virili, Entropy for endomorphisms of LCA groups, Topology Appl. 159 (9) (2012) 2546–2556.
[47] S. Virili, Algebraic and topological entropy of group actions, preprint.
[48] M. D. Weiss, Algebraic and other entropies of group endomorphisms, Math. Systems Theory 8 (3)
(1974/75) 243–248.
[49] S. Yuzvinski, Metric properties of endomorphisms of compact groups, Izv. Acad. Nauk SSSR, Ser.
Mat. 29 (1965) 1295–1328 (in Russian). English Translation: Amer. Math. Soc. Transl. (2) 66
(1968) 63–98.
[50] S. A. Yuzvinski, Calculation of the entropy of a group-endomorphism, Sibirsk. Mat. Z̆. 8 (1967)
230–239.
[51] P. Zanardo, Algebraic entropy of endomorphisms over local one-dimensional domains, J. Algebra
Appl. 8 (6) (2009) 759–777.
28
| 4math.GR
|
Social Event Scheduling
?
Nikos Bikakis † , Vana Kalogeraki † , Dimitrios Gunopulos #
†
Athens University of Economics & Business, Greece
arXiv:1801.09973v2 [cs.DS] 6 Mar 2018
[email protected]
#
[email protected]
Abstract—A major challenge for social event organizers (e.g.,
event planning and marketing companies, venues) is attracting
the maximum number of participants, since it has great impact
on the success of the event, and, consequently, the expected gains
(e.g., revenue, artist/brand publicity). In this paper, we introduce
the Social Event Scheduling (SES) problem, which schedules a
set of social events considering user preferences and behavior,
events’ spatiotemporal conflicts, and competing events, in order
to maximize the overall number of attendees. We show that
SES is strongly NP-hard, even in highly restricted instances.
To cope with the hardness of the SES problem we design a
greedy approximation algorithm. Finally, we evaluate our method
experimentally using a dataset from the Meetup event-based
social network.
Keywords-Social Event Planning, Attendance Maximization,
Social Event Arrangement, Event Organizers, Event Participants
I. I NTRODUCTION
The wide adoption of social media and networks has recently given rise to a new type of social networks that focus on
online event management, called Event-based Social Networks
(EBSN) [7]. In the most predominant EBSN platforms, such
as Meetup, Eventbrite or Whova, users organize, manage and
share social events and activities. In conjunction with the
events’ organizers in EBSNs, several entities such as event
planning and marketing companies (e.g., Jack Morton, GPJ),
organizations (e.g., IEEE), as well as venues (e.g., theaters,
night clubs), organize and manage a variety of social events
(music concerts, conferences, promotion parties). A major
challenge for event organizers is attracting the maximum
number of participants, since it has great impact on the
success of the event, and consequently, on the expected gains
from it, for all involved (e.g., revenue, artist/brand publicity).
Consider the following real-world scenario. A company
is going to organize the Summerfest festival. Summerfest is
an 11-day music festival, featuring 11 stages and attracting
more than 800K people each year. Throughout the festival, in
addition to the music concerts, numerous multi-themed events
take place (e.g., theatrical performances). Assume that Alice
enjoys listening to Pop music, and is a fashion lover. On
Monday from 7:00 to 10:00pm, a concert of a famous Pop
band is scheduled to take place at the festival. At the same
day, on a different stage, a fashion show is taking place from
7:00 to 9:00pm. Furthermore, from 6:00 to 8:00pm on that
day, a music concert of a Pop singer has been organized by
a nearby (competing) venue. Despite the fact that Alice is
interested in all three events, she is only able to attend one of
them. In another scenario, assume that a Pop concert is hosted
? This paper appears in 34th IEEE Intl. Conference on Data Engineering (ICDE 2018)
University of Athens, Greece
[email protected]
by the festival on Tuesday evening, but Alice is not capable
of attending this event, because on Tuesdays she works until
late at night.
The above example illustrates the major aspects that should
be considered in events scheduling scenarios. In order to attract
as many attendances as possible, organizers have to carefully
select the events that are going to take place during the festival,
possibly picking among from numerous candidate events, as
well as the date/time on which each event is going to take
place. During the event scheduling process, at least the following aspects have to be considered: user preferences, user habits
(e.g., availability), spatiotemporal conflicts between scheduled
events, and possible third parties events (e.g., organized by a
third party company) which might attract potential attendees
(i.e., competing events).
In this work, we introduce the Social Event Scheduling
(SES) problem, which considers the aforementioned aspects
and the goal is to maximize the overall number of participants
in the scheduled events. In short, given a set of events, a set of
time periods and a set of users, our objective is to determine
how to assign events on the time periods, so that the maximum
participant enrollment is achieved.
Recently, a number of works have been proposed in the
context of event-participant planing [2], [5], [6], [11]–[15].
These works examine a problem from a different perspective:
given a set of pre-scheduled events, they focus on finding the
most appropriate assignments for the users (i.e., participants)
attending the events. The determined user-event assignments
aim at maximizing the satisfaction of the users. However, these
works fail to consider a crucial issue in event management,
which is the “satisfaction” (e.g., revenue, publicity) of the
entities involved in the event organization (e.g., organizer,
artist, sponsors, services’ providers). Here, in contrast to
existing works, our objective is to maximize the satisfaction
of the event-side entities. To this end, instead of assigning
users to events, we assign events to time intervals, so that the
number of events attendees is maximized. Briefly, we study an
“event-centric” problem, while the existing approaches focus
on “user-centric” problems.
Therefore, our objective is substantially different compared
to the existing works. The same holds for the solution; in
our problem, the solution is a set of event-time assignments,
while in existing works is a set of user-event assignments.
Additionally, in order to solve our problem we have to find a
subset from a set of candidate events (i.e., some events may
not be included in the solution), while in other works the
solution contains all the users (i.e., each user is assigned to
events). Finally, beyond the user and event entities which are
also considered in existing works, in our problem more core
entities are involved (e.g., event organizer, competing events).
Thus, overall, the objective, the solution and the setting of our
problem substantially differ from existing works.
II. S OCIAL E VENT S CHEDULING P ROBLEM
In this section we first introduce the Social Event Scheduling
(SES) problem; and then we study its complexity. Before we
formally introduce our problem, we present some necessary
definitions.
Organizer & Time intervals. We assume that the event
organizer (e.g., company, venue) is associated with a number
of (available) resources θ ∈ R+ . For example, as resources we
can consider the agents (i.e., staff) which are responsible to
setup and coordinate the events. Let T be a set of candidate
time intervals, representing time periods that are available for
organizing events. Note that the intervals contained in T are
disjoint.
Candidate Events. Assume a set E of available events to
be scheduled, referred as candidate events. Each e ∈ E is
associated with a location `e representing the place (e.g., a
stage) that is going to host the event. Further, each event
e requires a specific amount of resources ξe ∈ R+
0 for its
organization, referred as required resources.
denoted as µu,h , that models the interest of user u over h. The
interest value (i.e., affinity) can be estimated by considering a
large number of factors (e.g., preferences, social connections)1 .
Moreover, for each user u and time interval t a social activity probability σut is considered, representing the probability
of user u participating in a social activity at t. Formally we
have σ : U × T → [0, 1]. This probability can be estimated by
examining the user’s past behavior (e.g., number of checkins)1 . Note that, user data can either be gathered by analyzing
organizer data (e.g., registered users profiles) or be provided
by a market research company.
Attendance. Assume a user u and a candidate event e ∈ E
that is scheduled by S to take place at time interval t; ρtu,e
denotes the probability of u attending e at t. Considering the
Luce’s choice theory, the probability ρtu,e is influenced by the
social activity probability σ of u at t, and the interest µ of u
over e, Ct and Et (S). We define the probability of u attending
e at t as1 :
∀c∈Ct
Feasibility. A schedule S is said to be feasible if the following
constraints are satisfied: (1) ∀t ∈ T holds that @ei , ej ∈ Et (S)
with
P `ei = `ej (location constraint); and (2) ∀t ∈ T holds that
ξe ≤ θ (resources constraint). In analogy, an assignment
∀e∈Et (S)
αet is said
to be feasible if the aforementioned constraints hold
for t. Further, we call valid assignment, an assignment αet
when the assignment is feasible and e ∈
/ E(S).
(1)
∀p∈Et (S)
Furthermore, considering all users U, we define the expected
attendance for an event e scheduled to take place at t as:
ωet =
αet
denotes that the
Schedule & Assignment. An assignment
candidate event e ∈ E is scheduled to take place at t ∈ T .
An event schedule S is a set of assignments, where there
exist no two assignments referring to the same event. Given
a schedule S, we denote as E(S) the set of all candidate
events that are scheduled by S; and Et (S) the candidate
events that are scheduled by S to take place at t (i.e.,
assigned to t). Formally, E(S) = {ei ∈ E | ateii ∈ S} and
Et (S) = {ei ∈ E | ateii ∈ S with ti = t}. Further, for a
candidate event e ∈ E(S), we denote as te (S) the time interval
on which S assigns e.
µu,e
P
µu,c +
µu,p
ρtu,e = σut P
X
ρtu,e
(2)
∀u∈U
Total Utility. The total utility for a schedule S, denoted as
Ω(S), is computed by considering the expected attendance
over all scheduled events. Thus, we have:
Ω(S) =
X
ωete (S)
(3)
∀e∈E(S)
We formally define the Social Event Scheduling (SES) problem
as follows:
Social Event Scheduling Problem (SES). Given an integer
k, a set of candidate time intervals T ; a set of competing
events C; a set of candidate events E; and a set of users U;
our goal is to find a feasible schedule Sk that determines
how to assign k candidate events such that the total utility
Ω is maximized; i.e., Sk = arg max Ω(S) and |S| = k.
Next, we show that even in highly restricted instances the
SES problem is strongly NP-hard.
Competing Events. Let C be a set of competing events.
As competing events we define events that have already
been scheduled by third parties (e.g., organized by a third
party marketing company), and will possibly attract potential
attendees of the candidate events. Based on its scheduled time,
each competing event c ∈ C is associated with a time interval
tc ∈ T . Further, as Ct we denote the competing events that are
associated with the time interval t; i.e., Ct = {c ∈ C | tc = t}.
P ROOF S KETCH . Our reduction is from the Multiple Knapsack
Problem with Identical bin capacities (MKPI), which is known
to be strongly NP-hard [8]. In MKPI there are multiple items
and bins. Each item has a weight and a profit and all bins
have the same capacity. In the reduction, we use the following
associations between the MKPI and the SES: (1) bins to
Users. Consider a set of users U, for each user u ∈ U and
event h ∈ E ∪ C, there is a function µ : U × (E ∪ C) → [0, 1],
1 Event-based mining methods can be used to compute this value, e.g., [1],
[3], [4], [10], [16]–[18]. However, this is beyond the scope of this work.
Theorem 1. The SES problem is strongly NP-hard.
time intervals; (2) capacity to number of available resources;
(3) items to events; (4) weight to number of required resources; (5) item profit to likeness; and (6) total profit to
expected attendance. Further, in the proof, we consider the
following restricted instance of SES: (1) the users are as
many as the candidate events; (2) there is only one competing
event in each time interval; (3) all users have the same interest
value K over the competing events; (4) each user likes only
one event and each event is liked only by one user; (5) the
K
, where p is item profit; (6) the
interest function is µ = p 1−p
social activity probability is the same for each user and time
interval; and (7) there are no location constraints.
Algorithm 1. GRD (k, T , E, C, U)
Input: k: number of scheduled events; T : time intervals;
E: candidate events; C: competing events; U : users;
Output: S: feasible schedule containing k assignments
Variables: L: assignment list
6
S ← ∅; L ← ∅;
foreach (e, t) ∈ E × T do
compute αte .S
insert αte into L
while |S| < k do
t
αepp ← popTopAssgn(L)
7
if αepp is valid then
1
2
3
4
5
t
10
11
Assignment Score. Assume a schedule S and an assignment
αrt , where r is not previously assigned by S (i.e., r ∈
/ E(S)).
As assignment score (also referred as score) of an assignment
αrt , denoted as αrt .S, we define the gain in the expected
attendance by including αrt in S. The assignment score (based
on Eq. 2) is defined as:
αrt .S =
X
ωe0 jt −
∀ej ∈
Et (S)∪{r}
X
ωet i
(4)
∀ei ∈
Et (S)
Note that, the expected attendance ω 0 of each event after
assigning r, differs from the expected attendance ω before the
assignment. Also, based on Eq. 2 & 4, it is apparent that the
score of an assignment referring to interval t is determined
based on all the events assigned to t. Finally, given a set of
assignments, the term top assignment refers to the assignment
with the largest score.
Algorithm Outline. Here we describe a simple greedy algorithm, referred as Greedy algorithm (GRD). The basic idea of
GRD is that the assignments between all pairs of event and
interval are initially generated. Then, in each step/iteration, the
assignment with the largest score is selected. After selecting
an assignment, a part of the potential assignment’s scores have
to updated. Recall that the assignment’s score is defined w.r.t.
the events assigned in the assignment’s interval (Eq. 4). Thus,
when an assignment αet is selected, we have to recompute
(update) the scores of the assignments referring to t interval.
Algorithm Description. Algorithm 1 presents the pseudocode
of GRD. At the beginning the algorithm calculates the score,
for all possible assignments (line 3). The generated assignments are inserted into list L (line 4). Then, in each step the
t
assignment αepp with the largest score is found and popped
(line 6). If the popped assignment is feasible and the event
t
ep is not previous assigned (i.e., assignment is valid), αepp is
inserted into schedule S (line 8). Until a valid assignment
is selected, the assignment with the largest score is popped
t
and checked. After selecting αepp , the algorithm traverses
L, updating the appropriate assignments (loop in line 11).
14
//find & remove the assignment with largest score
//insert the top and valid assignment into schedule
insert αepp into S
if |S| < k then
foreach αte ∈ L do
//update assignments
if t = tp and αte is valid then
compute new αte .S;
//by Eq. 4
else if αte is not valid then remove αte from L;
8
9
12
First, we define the assignment score; and then we present
the GRD algorithm.
//by Eq. 4
//initialize assignment list
t
13
III. G REEDY A LGORITHM (GRD)
//generate assignments
return S
Finally, the algorithm terminates when k assignments are
selected.
Complexity Analysis. Initially, the GRD computes the assignments for all event-interval pairs (loop in line 2), which
requires O(|E||T ||U|). Note that, each assignment score
(Eq. 4) is computed in O(|U|). In the next phase (loop in
line 5), the GRD performs k iterations. In each iteration,
the popTopAssgn operation (line 6) traverses the L list of
size |T |(|E| − i), where 0 P
≤ i ≤ k − 1. Thus, in sum, the
k−1
cost for traversing L is O( i=0 |T |(|E| − i)). Additionally,
in each iteration (except the last), |E| − (i + 1) assignment
updates are performed, in the
worst case. Hence, the overall
Pk−2
cost for updates is O(|U| i=0 |E| − i − 1)). Therefore,
the overall computation cost of GRD in the worst case is
O(|U||C| + |E||T ||U| + k|E||T | + k|E||U | − k 2 |T | − k 2 |U|).
Finally, the space complexity is O(|E||T |).
IV. E XPERIMENTAL A NALYSIS
A. Setup
Data. In our experiments we use the largest Meetup dataset
from [9], which contains data from California. Adopting the
same approach as in [11]–[13], [15], in order to define the
interest of a user to an event, we associate the events with
the tags of the group who organize it. Then, we compute the
likeness value using Jaccard similarity over the user-event tags.
After preprocessing, we have the Meetup dataset containing
42,444 users and about 16K events.
Parameters. Adopting the same setting as in the related works
[6], [11]–[13], [15], we set the the default and maximum value
of the of scheduled events k, to 100 and 500, respectively.
Also, we vary the number of time intervals |T |, from k5 up
to 3k, with default value set to 3k
2 . Further, the number of
candidate events |E| is set to 2k.
In order to select the values for the number of competing
events per interval, we analyze the two Meetup datasets from
[9]. From the analysis, we found that, on average, 8.1 events
are taking place during overlapping intervals. Therefore, the
number of competing events per interval is selected by a
uniform distribution having 8.1 as mean value.
(a) Utility vs. k
(b) Time vs. k
(c) Utility vs. |T |
(d) Time vs. |T |
Fig. 1: Varying the number of: (1) scheduled events k; and (2) time intervals |T |
Regarding the value for the number of available events’
locations, we consider the percentage of pairs of events that are
spatio-temporally conflicting, as specified in [11]. As a result,
we set the number of available locations to 25. The social
activity probability σut is defined using a Uniform distribution.
The performance and the effectiveness of the examined
methods are marginally affected by the available/required resources parameters (here as resources we consider organizer’s
staff). Hence, we choose a reasonable number based on our
scenario, setting the number of available resources to 20. Also,
the number of required resources is selected by a uniform
distribution defined over the interval [1, 20
3 ].
Methods. In our evaluation we study our method GRD, as well
as two baselines. The first baseline method, TOP, computes the
assignment scores for all the events and selects the events with
top-k score values. The second, denoted as RAND assigns
events to intervals, randomly. Note that, since the objective,
the solution and the setting of our problem are substantially
different (see Sect. I) from the related works [2], [5], [6],
[11]–[15], the existing methods cannot be used to solve the
SES problem. All algorithms were written in C++ and the
experiments were performed on an 2.67GHz Intel Xeon E5640
with 32GB of RAM.
B. Results
In the first experiment, we study the effect of varying the
number of scheduled events k. In terms of utility (Fig. 1a), we
observe that, in all cases, GRD outperforms significantly both
baselines. The difference between RAND and GRD increases
as k increases. This is expected considering the fact that the
larger the k, the larger the number of “better”, compared to
random, selected assignments. Finally, TOP reports considerably low utility scores in all cases.
The results regarding the execution time are depicted in
Figure 1b. Note that the computations that are performed
due to updates increase with k, while the number of initially
computed scores is the same for all k. Also, TOP performs
only the initial scores’ computations (there are no score
updates). That’s why the difference between the GRD and the
TOP increases with k.
In the next experiment, we vary the number of time intervals
|T |. Regarding utility (Fig. 1c), we observe that, as the number
of intervals increases, the utility of GRD and TOP methods
increases too. This happens because the increase of available
intervals results to a smaller number of events assigned in
the same interval, as well as to a larger number of candidate
assignments. In terms of execution time (Fig. 1d), for the same
reason as in the first experiment, the difference between the
GRD and the TOP increases with |T |.
V. C ONCLUSIONS
This paper introduced the Social Event Scheduling (SES)
problem. The goal of SES is to maximize the overall events’
attendance considering several events’ and users’ factors. We
showed that SES is strongly NP-hard and we developed a
greedy algorithm.
ACKNOWLEDGMENT
This research has been financed by the European Union
through the FP7 ERC IDEAS 308019 NGHCS project, the
Horizon2020 688380 VaVeL project and a 2017 Google Faculty Research Award.
R EFERENCES
[1] I. Boutsis, S. Karanikolaou, and V. Kalogeraki. Personalized event
recommendations using social networks. In MDM, 2015.
[2] Y. Cheng, Y. Yuan, L. Chen, C. G. Giraud-Carrier, G. Wang. Complex
event-participant planning and its incremental variant. In ICDE, 2017.
[3] R. Du, Z. Yu, T. Mei, Z. Wang, Z. Wang, and B. Guo. Predicting
Activity Attendance in Event-based Social Networks: Content, Context
and Social Influence. In UbiComp, 2014.
[4] K. Feng, G. Cong, S. S. Bhowmick, and S. Ma. In Search of Influential
Event Organizers in Online Social Networks. In SIGMOD, 2014.
[5] J. Huang, Y. Zhou, X. Jia, and H. Sun. A novel social event organization
approach for diverse user choices. The Computer Journal, 2016.
[6] K. Li, W. Lu, S. Bhagat, L. V. Lakshmanan, and C. Yu. On Social Event
Organization. In KDD, 2014.
[7] X. Liu, Q. He, Y. Tian, W. Lee, J. McPherson, J. Han. Event-based social
networks: linking the online and offline social worlds. KDD 2012
[8] S. Martello and P. Toth. Knapsack Problems: Algorithms and Computer
Implementations. John Wiley & Sons, 1990.
[9] T. N. Pham, X. Li, G. Cong, Z. Zhang. A General Graph-based Model
for Recommendation in Event-based Social Networks. In ICDE, 2015.
[10] D. Romero, K. Reinecke, L. Robert. The influence of early respondents:
information cascade effects in online event scheduling. WSDM 2017
[11] J. She, Y. Tong, and L. Chen. Utility-Aware Social Event-Participant
Planning. In SIGMOD, 2015.
[12] J. She, Y. Tong, L. Chen, and C. C. Cao. Conflict-aware Eventparticipant Arrangement. In ICDE, 2015.
[13] J. She, Y. Tong, L. Chen, and C. C. Cao. Conflict-aware Eventparticipant Arrangement and Its Variant for Online Setting. TKDE, 2016.
[14] J. She, Y. Tong, L. Chen, and T. Song. Feedback-Aware Social EventParticipant Arrangement. In SIGMOD, 2017.
[15] Y. Tong, J. She, and R. Meng. Bottleneck-aware Arrangement Over
Event-based Social Networks: The Max-Min Approach. WWWJ, 2015.
[16] T.Xu, H.Zhong, H Zhu, H Xiong, E Chen, G Liu. Exploring the impact
of dynamic mutual influence on social event participation. In SDM 2015
[17] W. Zhang, J. Wang, W. Feng. Combining Latent Factor Model with
Location Features for Event-based Group Recommendation. KDD, 2013.
[18] X. Zhang, J. Zhao, and G. Cao. Who Will Attend? - Predicting Event
Attendance in Event-based Social Network. In MDM, 2015.
| 8cs.DS
|
1
Bitwise Retransmission Schemes for Resources
Constrained Uplink Sensor Networks
arXiv:1707.09696v1 [cs.IT] 31 Jul 2017
Mohamed A. M. Hassanien, Pavel Loskot, Salman M. Al-Shehri, Tolga Numanoǧlu, and Mehmet Mert
Abstract—Novel bitwise retransmission schemes are devised
which retransmit only the bits received with small reliability.
The retransmissions are used to accumulate the reliabilities
of individual bits. Unlike the conventional automatic repeat
request (ARQ) schemes, the proposed scheme does not require
a checksum for the error detection. The bits to be retransmitted
are reported as a combination number, or two synchronized
random number generators (RNGs) at the transmitter and
receiver are used to greatly compress the feedback message.
The bitwise retransmission decisions and/or combining can be
performed after the demodulation or after the channel decoding
at the receiver. The bit-error rate (BER) expressions are derived
for the case of one and two retransmissions, and verified by
computer simulations. Assuming three specific retransmission
strategies, the scheme parameters are optimized to minimize
the overall BER. For the same number of retransmissions and
packet length, the proposed schemes always outperform the
frequently used stop-and-wait ARQ. The impact of feedback
errors is also considered. Finally, practical designs of the bitwise
retransmissions for data fusion from sensor nodes in Zigbee, Wifi
and Bluetooth networks are presented.
Index Terms—Automatic repeat request, data fusion protocol,
feedback signaling, performance analysis, retransmissions.
I. I NTRODUCTION
It is well-known that feedback cannot improve the information theoretic capacity of memoryless channels [1]. However, the availability of feedback can greatly simplify the
encoding and decoding complexity [2]. A good example of
such schemes with the reduced implementation complexity
due to feedback are the ARQ retransmission schemes [3],
[4]. The retransmission schemes are optimized to trade-off
the reliability (e.g., the BER), the throughput (or equivalently,
delay) and the implementation complexity [5] or the energy
consumption [6]. For instance, the retransmissions can comprise a smaller number of bits than in the original packet.
The incremental redundancy hybrid ARQ (IR-HARQ) or typeII HARQ schemes progressively reduce the coding rate of
the forward error correction (FEC) code with each additional
retransmission at the expense of increasing the decoding
delay and reducing the throughput [7]. The retransmission
decision delays in HARQ schemes were reduced in [8] by
exploiting the structure of tail-biting convolutional codes. The
permutations of bits in the retransmitted packets are used in
MAH, PL and SMA are with the College of Engineering, Swansea
University, Bay Campus, Swansea SA1 8EN, United Kingdom (email:
{mohmmedali25,salman777881}@hotmail.com, [email protected])
TN and MM are with Aselsan A.S., Communications and IT Division,
Ankara, Turkey (email: {tnumanoglu,mmert}@aselsan.com.tr)
Corresponding author: Pavel Loskot, tel.: +44 1792 602619, fax: +44 1792
295676
[9] and [10] to improve the reliability of ARQ schemes. A
holistic design of the complexity-constrained type-II hybrid
ARQ schemes with turbo codes is considered in [11]. The joint
design of FEC coding for forward data delivery and reverse
feedback signaling is studied in [12] and [13]. The IR diversity
and the time-repetition (TR) diversity are compared in [14].
The time and superposition-coding packet sharing between
two independent information flows is optimized in [15]; the
latter is shown to have a slightly better performance at the
expense of larger design and implementation complexity. Such
transmission sharing strategies can significantly outperform
the conventional HARQ schemes when signal-to-noise ratio
(SNR) is sufficiently large [15]. Assuming Gaussian transmission codebooks, the achievable throughput of the HARQ
schemes are compared to the ergodic channel capacity in
[15], and to the delay-limited channel capacity in [14]. The
transmission powers of ARQ schemes are optimized in [14].
Since the received packets typically contain only a few
transmission errors, the retransmission efficiency can be improved by the partial ARQ schemes [16]. A truncated type II
hybrid ARQ over block fading channels is considered in [17].
In addition to ARQ schemes exploiting a variable number
of retransmissions, the variable rate ARQ schemes optimize
the number of bits in each retransmission [14], [18], [19].
The number of retransmission bits required for a successful
decoding is estimated from the mutual information in [19]
and [15]. The pre-defined retransmission patterns are assumed
in [16]. The reactive rate-adaptive ARQ strategies [7] usually
outperform the proactive strategies [18]. The FEC coding
to recover from the transmission errors may be preferred
to the ARQ retransmissions in case of multimedia [20].
Except the stop-and-wait ARQ schemes, the go-back-N ARQ
retransmissions are optimized, for example, in [21]. A multibit feedback signaling to improve the ARQ performance is
considered in [22] and [15]. Other papers account for more
realistic design constraints such as a noisy feedback [14], [17],
and the channel estimation in [16]. It is shown in [14] that the
errors of 1 and 2-bit feedback messages can be neglected if
their bit error probability is less than 10−3 , or if these errors are
compensated by a non-uniform allocation of the transmission
powers. Furthermore, the repetition diversity appears to be
more robust to feedback errors than the IR-HARQ [14].
In our conference paper [22], we assumed a multi-bit
error-free feedback to evaluate the performance trade-offs
between the throughput and the reliability of the segmentationbased and the bitwise ARQ retransmission schemes. Both
schemes were found to outperform the stop-and-wait ARQ
with the bitwise ARQ providing larger reliability gains than
2
the segmentation-based ARQ, albeit at the expense of a greater
complexity of the feedback signaling. In this paper, we revisit
the selective bitwise retransmission scheme proposed in [22] to
carry out more rigorous performance analysis and to optimize
its design. Recall that the proposed scheme aims to accumulate
the reliabilities of the least reliable individual bits in the
received packet, so it does not require the cyclic redundancy
check (CRC) bits to make the retransmission decisions. It
can be combined with other FEC coding schemes where
the bit reliability is evaluated either before or after the FEC
decoding. From the implementation point of view, it is useful
to keep the number of retransmitted bits as well as the
number of retransmissions constant, for example, to maintain
a constant transmission delay and throughput for each data
packet. However, a variable rate scheme that retransmits all
bits having their reliability below a given threshold is also
investigated. The bits with small reliabilities are reported back
to the sender using either a binomial combination number,
or using a deterministic sequence of bit-permutations. Our
analysis assumes multi-bit error free feedback signaling, however, we also evaluate the conditions when such assumption
is justifiable. Finally, we consider a more specific systemlevel design of the proposed bitwise retransmission scheme
to be employed in an uplink data collection scenario from
the resources-constrained sensors to a data fusion access point
assuming time-sharing of the transmission channel among the
network nodes. The resulting transmission protocol creates
fully occupied packets with the retransmitted bits having a
higher priority than the newly arrived information bits.
The rest of this paper is organized as follows. System model
is introduced in Section II. The mathematical theoretic analysis
of the proposed bitwise retransmission scheme is carried out
in Section III. The retransmission protocols are presented and
optimized in Section IV. The uplink data fusion bitwise ARQ
scheme is designed in Section V. Conclusions are given in
Section VI.
The design and analysis of the proposed bitwise retransmission scheme assumes a point-to-point duplex communication
link between a source node and a destination node. As shown
in [14], the feedback errors in ARQ schemes can be neglected
provided that their probability of error is sufficiently small.
Hence, in our analysis, we assume the error-free feedback,
and in Section IV, we estimate the acceptable probability
of feedback bit errors when such assumption is justified in
practice.
Assuming that M-ary modulation is used for transmission
from the source to the destination node, prior to the FEC
decoding, the reliability of the received bit bi can be calculated
as [23],
Pr(bi = 0|y)
, i = 1, 2, . . . , log2 M
Pr(bi = 1|y)
where h is a residual attenuation, possibly after the (multipath)
equalization, and w is a zero-mean sample of AWGN. The loglikelihood ratio (LLR) (1) can be then rewritten as,
∑s∈S(bi =0) exp −|y − hs|2/N0
(2)
Λ(bi |y) = log
∑s∈S(bi =1) exp (−|y − hs|2/N0 )
where the M-ary modulation constellation is partitioned as,
S = S(bi = 0) ∪ S(bi = 1), depending on the value of the i-th
bit bi in modulation symbols, N0 is the power spectral density
of the AWGN, and | · | denotes the absolute value.
The LLR (1) can be also used as the soft-input decisions
to the FEC decoder. On the other hand, we can also use the
soft-output decisions produced by the decoder to make the
retransmission decisions. In this case, the sums in (2) are
done over the corresponding codewords mapped to sequences
of M-ary modulation symbols, and the sums in (2) are often
approximated using the formula, log ∑i exp(−ai ) ≈ min(ai ).
In order to simplify the analysis and illustrate the main
concepts, in the sequel, we consider a binary antipodal
modulation. Generalization to higher order modulations is
straightforward by scaling the demodulated binary symbols
by appropriate constants computed for a given M-ary modulation√scheme. Hence, √
assume the transmitted binary symbols
S1 = Eb and S2 = − Eb , so the received N-bit packet can
be written as,
ri = S ji + wi , i = 1, 2, . . . , N
where S ji ∈ {S1 , S2 }, and the zero-mean
AWGN samples wi
have the constant variance, E |wi |2 = σ2w = N0 /(2|h|2 ). Then,
the reliabilities of the received bits are [23],
Λ(bi |ri ) ∝ |ri /σw | ≡ |r̄i |.
II. S YSTEM M ODEL
Λ(bi |y) = log
the reception of modulation symbols can be mathematically
modeled as,
y = hs+ w
(1)
where y is the received M-ary modulation symbol. For equally
probable modulation symbols, assuming coherent detection
in a channel with additive white Gaussian noise (AWGN),
Finally, the SNR per binary modulation symbol is defined
as γb = Eb /N0 . If Rf denotes a fraction of information bits
in the sequence of all bits transmitted from the source to the
destination, for a fair comparison of different schemes with
various rates Rf , we assume the SNR per bit, γb = Eb Rf /N0 .
A. Bitwise retransmissions
After the initial transmission of N-bit data packet, the
destination uses the received bit reliabilities to decide which
Wd binary symbols among N received, 0 ≤ Wd ≤ N, should be
retransmitted, where d = 1, 2, . . . , D denotes the retransmission
index. The retransmission request is a feedback message of
Cd ≥ 1 bits sent to the source node from the destination
node via a reverse (feedback) link. Consequently, after D
retransmissions, the total number of bits sent over the forward
link to the destination is,
D
Nf = N +
∑ Wd
d=1
3
C1
N
t0
C2
W1
t1
t1 + T1
CD
W2
t2 + T2
t2
reverse
link
WD
t3
forward
link
t
tD + TD tD+1
Figure 1. The timings of the proposed bitwise retransmission protocol.
whereas the total number of bits sent from the destination to
the source over the reverse link is,
D
Nr =
∑ Cd .
d=1
For instance, the conventional stop-and-wait ARQ scheme
has the parameters Wd = N and Cd = 1 (ACK or NACK).
The retransmitted bits are combined using a maximum ratio
(MRC) or other combining method to improve their reliability.
The timings of the bitwise retransmission protocol are shown
in Fig. 1 where Td represents the delay until the start of
transmission of the window of Wd retransmitted bits. The
forward transmission rate is then defined as,
N
N
Rf (D) =
(3)
=
N
W
N + ∑D
f
d=1 d
and the reverse transmission rate is defined as,
Nr
∑D
d=1 Cd
.
Rr (D) =
=
D
N + Nr
N + ∑d=1 Cd
Note that the feedback delays Td are not included in the
definition of these rates, since during these time intervals,
both forward and reverse links are possibly available for other
transmissions.
In order to simplify the notation, let W1 = W2 = . . . = WD =
W , and C1 = C2 = . . . = CD = C, and we drop the bit index
within the packet unless stated otherwise. The W least reliable
bits requested for the next retransmission can be identified by
sorting the received bits by their reliabilities. A typical sorting
algorithm has the complexity O(N log N) [24]. The locations
of these W bits within the packet of N bits can be reported
back to the sender by the corresponding binomial number. The
number of bits representing such feedback message is,
N
C = log2
W
N
where W
is the binomial number, and ⌈·⌉ is the ceiling
function. More importantly, even though typically C > W , we
always have that, Rf ≫ Rr , as in many other ARQ protocols.
Thus, the multi-bit feedback can be very beneficial to improve
the performance of ARQ protocols [22].
III. P ERFORMANCE A NALYSIS
Given the values of N, W , C, D and γb , we now derive the
average BER of the proposed bitwise retransmission scheme.
We first assume the case of a single retransmission, i.e., D =
1. We optimize the number of feedback message bits C the
forward throughput Rf to minimize the BER, and compare
our scheme to the conventional stop-and-wait ARQ. We then
perform the similar analysis for D = 2 retransmissions before
generalizing the obtained BER expressions to the case of D > 2
retransmissions.
Without any retransmissions (i.e., immediately after the new
data packet of N bits was received), the conditional probability
density function (PDF) of the received bit reliability r̄ can be
written as,
r
√ 2
Eb )
1 No − (r̄No −2
4N
o
e
fr̄ (r̄|S1 , D = 0) =
2r π
√ 2
Eb )
1 No − (r̄No +2
4N
o
fr̄ (r̄|S2 , D = 0) =
e
.
2 π
Since the bits are selected for retransmission based on
their reliability, we first obtain the BER conditioned on the
reliability interval, LD ≤ |r̄| ≤ UD , for some positive constants
UD ≥ LD ≥ 0. This BER is equal to the probability of the error
event ‘ e’ that S1 was transmitted, however, S2 is decided at
the receiver when −UD ≤ r̄ ≤ −LD . With no retransmissions,
the probability of such error event is equal to,
Pr ( e|S1 , D = 0) = Pr −U0 ≤ r̄ ≤ −L0 S1
Z −L0 √Eb
fr̄ (r̄|S1 , D = 0) dr̄
=
√
−U E
0 b
p
p
L0
U0
2γb
2γb
=Q
+1
−Q
+1
2γb
2γb
R
2
where Q(x) = x∞ √12π e−t /2 dt is the Q-function, and due
to symmetry, we have, Pr (e|S1 , D = 0) = Pr (e|S2 , D = 0).
Assuming the a priori transmission probabilities, Pr (S1 ) =
Pr (S2 ) = 1/2, the overall average BER for L0 = 0 becomes,
p
p
U0
2γb − Q
2γb
+1
.
(4)
BER0 = Q
2γb
Note that, for U0 → ∞, the BER (4) corresponds to the
√ BER of
uncoded binary antipodal signaling, i.e., BER = Q( 2γb ) [3].
Subsequently, the probability that the reliability of the received
bit is in the interval LD ≤ |r̄| ≤ UD without any retransmissions
is computed as,
P0|S1 = Pr L0 ≤ |r̄| ≤ U0 S1
Z −L0
Z U0
fr̄ (r̄|S1 , D = 0) dr̄
fr̄ (r̄|S1 , D = 0) dr̄ +
L0
−U0
p
p
L0
U0
2γb
2γb
= Q
+1
−Q
+1
+
2γ
2γ
b
b
p
p
L0
U0
−1
−Q
−1
.
Q
2γb
2γb
2γb
2γb
=
For L0 = 0, this probability is equal to,
p
p
U0
U0
P0|S1 = 1−Q
2γb
2γb
+1 −Q
−1
2γb
2γb
and P0|S1 → 1 for U0 → ∞. Due to symmetry, and for equally
probable symbols Si , we get, P0 = P0|S1 = P0|S2 . Furthermore,
if Z is the (random) number of bits having the reliabilities in
the interval, 0 ≤ |r̄| ≤ U0 , the mean value of Z, E[Z] = N P0 ,
can be used to optimize the retransmission window size. For
example, by letting W = E[Z], we can control the target BER
after the first retransmission.
4
In general, it is convenient to re-scale the received bits
after every retransmission and subsequent MRC combining to
maintain the constant average energy of the received symbols.
In particular, the received symbol after combining the initially
received sample r̄0 with the retransmitted and scaled samples
r̄d , d = 1, 2, . . . , D, can be written as,
!
D
1
r̄MRC =
r̄0 + ∑ r̄d .
D+1
d=1
In the sequel, let Ld = 0 for ∀d ≥ 0, and, without loss of
generality, we assume that the packet of N symbols S1 was
transmitted.
In the appendix, we show that BER1 can be accurately
approximated as,
p
p
U0
+ 1) + Q
2γb (
4γb −
BER1 ≈ Q
2γb
n
2
√
Ak
−α− (U )γ
∑ √1 + 2Bk e k,1 0 b Q β+k,1 (U0) γb +
k=1
+
√ o
e−αk,1 (U0 )γb Q β−
k,1 (U0 ) γb
where the coefficients Ak and Bk , are given in the appendix,
and the auxiliary functions,
α−
k,d (U) =
A. BER with one retransmission
After the first transmission of the packet of N bits, only
those bits that have the reliabilities in the interval 0 ≤ |r̄| ≤ U0
are requested to be retransmitted. After the first retransmission,
the bit samples are combined as,
1
2 (r̄0 + r̄1 ) for |r̄0 | ≤ U0
r̄MRC =
r̄0
for |r̄0 | > U0 .
The random variable r̄0 has the PDF, fr̄ (r̄|S1 , D = 0), whereas
the PDF of the random variable r̄1 can be obtained by
conditioning on |r̄0 | ≤ U0 . Since the random received samples
r̄0 and r̄1 are statistically independent, the PDF of r̄MRC is
given by the convolution (denoted as ∗), i.e.,
fr̄MRC (r̄|S1 , D = 1) =
where
fr̄ (r̄|S1 , d = 0)φ1 (r̄,U0 )
∗ fr̄ (r̄|S1 , d = 0)
Pr |r̄0 | ≤ U0 S1 , D = 0
φ1 (r̄,U0 ) = η(r̄ + U0 ) (1 − η(r̄ − U0))
and η(r̄) is the unit-step function, i.e., η(r̄) = 1 if r̄ ≥ 0, and
0 otherwise. After some manipulations, the PDF of r̄MRC can
be written as,
fr̄MRC (r̄|S1 , D = 1) =
χ1 (r̄,U0 )
Pr |r̄0 | ≤ U0 S1 , D = 0
where we defined the function,
χd (r̄,U0 )
=
r
r
2
γ
r̄−2 Nb
o
−
4
(d+1)No
1 (d + 1)No
e
×
4
!
rπ
n
(d + 1)No
(U0 − r̄) +
erf
4d
!
r
o
(d + 1)No
erf
(U0 + r̄)
4d
and erf(·) denotes the error function [3].
Assuming that all bits with the reliabilities |r̄0 | ≤ U0 are
retransmitted, and then combined using the MRC, the overall
BER1 after the first retransmission is computed as [3],
BER1
=
=
Pr (e|d = 0, S1 ) Pr (d = 0|S1 ) +
Pr (e|d = 1, S1 ) Pr (d = 1|S1 )
Z −U0
−∞
fr̄ (r̄|S1 , d = 0) dr̄ +
Z 0
−∞
χ1 (r̄,U0 ) dr̄.
α+
k,d (U) =
β−
k,d (U) =
Bk (d + 1)(2 − U/γb)2
2d(1 + 2Bd k )
Bk (d + 1)(2 + U/γb)2
1 − BdγkUb
r
,
2B
1+ d k
2(d+1)
2d(1 + 2Bd k )
β+
k,d (U) =
1 + BdγkUb
r
.
2B
1+ d k
2(d+1)
B. BER with two retransmissions
Consider now the case with two retransmissions. After the
second retransmission, the received bits having the reliabilities
|r̄| ≤ U1 are combined with the retransmitted samples r̄2 . The
threshold value U1 is chosen to be greater than U0 due to improvement of the bit reliabilities after the first retransmission.
More generally, we assume that,
UD−1 ≥ UD−2 . . . ≥ U1 ≥ U0 .
Hence, the bit-samples after two retransmissions can be written
as,
1
(r̄0 + r̄1 + r̄2 ) for |r̄0 | ≤ U0 , 21 |r̄0 + r̄1 | ≤ U1
3 1
for |r̄0 | ≤ U0 , 12 |r̄0 + r̄1 | > U1
2 (r̄0 + r̄1 )
r̄MRC =
1
for U0 < |r̄0 | ≤ U1
2 (r̄0 + r̄2 )
r̄0
for |r̄0 | > U1 .
(5)
However, in order to make the analysis mathematically
tractable, we merge the first two conditions in (5), and consider
instead the retransmission scheme,
1
3 (r̄0 + r̄1 + r̄2 ) for |r̄0 | ≤ U0
1
r̄MRC =
(6)
for U0 < |r̄0 | ≤ U1
2 (r̄0 + r̄2 )
r̄0
for |r̄0 | > U1
having the upper-bound performance than the scheme (5). The
scheme (6) can be interpreted as making the decision about the
number of retransmissions for each bit already after receiving
the initial data packet of N bits. Even though the scheme (6)
may unnecessarily retransmit some bits even if their reliability
have already reached the desired threshold, our numerical
results indicate that, for D = 2, the performance difference is
not significant (less than 1 dB). Consequently, the number of
retransmissions for each bit can be determined by quantization
of the initial reliabilities |r̄0 | using the thresholds U0 and U1 .
The bits having the reliabilities |r̄0 | ≤ U0 are always combined
with two other retransmitted bits, whereas the received bits
5
with the reliabilities U0 < |r̄0 | ≤ U1 are combined with exactly
one retransmission.
The PDF fr̄MRC (r̄|S1 , d = 2) of r̄MRC , for |r̄| ≤ U0 , is given
as,
χ1 (r̄,U0 )
fr̄MRC (r̄|S1 , d = 2) =
Pr |r̄0 | ≤ U0 S1 , d = 0
and for U0 < |r̄0 | ≤ U1 , it is given as,
fr̄MRC (r̄|S1 , d = 2) =
where
λd (r̄,Ud ,Ud−1 ) =
( q
erf
1
4
q
λ1 (r̄,U1 ,U0 )
Pr U0 < |r̄0 | ≤ U1 S1 , d = 0
(d+1)No
π
exp −
r̄−2
q
γb
No
4
(d+1)No
2 !
×
q
(d+1)No
(d+1)No
(Ud + r̄) + erf
(Ud − r̄) −
4d
4d
q
)
q
(d+1)No
(d+1)No
(Ud−1 + r̄) −erf
(Ud−1 − r̄) .
erf
4d
4d
For |r̄0 | > U1 , the PDF fr̄MRC (r̄|S1 , d = 2) is given by the PDF
fr̄ (r̄|S1 , d = 0). The overall BER2 is obtained using the law of
the total probability, i.e.,
2
BER2
=
∑ Pr (e|d, S1) Pr (d|S1) =
d=0
+
Z 0
−∞
Z −U1
−∞
λ1 (r̄,U1 ,U0 ) dr̄ +
fr̄ (r̄|S1 , D = 0) dr̄
Z 0
−∞
where the auxiliary functions,
θ+−
k,d (U)
=
D −U)
√
+ √1N + Bk (U
d γb
U
√D
2 γb
1+ d k
(d+1)No
D +U)
√1 + Bk (U√
d γb
No
−
2B
1+ d k
(d+1)No
U
√D
2 γb
D +U)
√
+ √1N + Bk (U
d γb
o
2B
1+ d k
(d+1)No
=
|µ(U)|
2
µ(U) =
UD
√ −
γb
θ−−
k,d (U)
o
2B
θ−+
k,d (U) =
θ++
k,d (U) =
U
√D
2 γb
r
(d + 1)No (1 +
√2
No
2Bk
)
d
√kU
+ 2B
γb
1 + 2Bk
and sign(·) is the sign function.
C. BER with multiple retransmissions
For D ≥ 1 retransmissions, the overall BER is calculated as,
D
BERD =
∑ Pr (e|d, S1) Pr (d|S1) .
d=0
χ2 (r̄,U0 ) dr̄
where the probability of the second retransmission
of the
same bit is, Pr (d = 2|S1 ) = Pr r̄ ≤ −U1 S1 , d = 1 . As for one
retransmission, BER2 can be approximated (see the appendix),
√
√
U1
BER2 ≈ Q 2γb ( 2γ
+
1)
+ Q 6γb
nb −
√
−α (U )γ
Ak
(U
)
e k,2 0 b Q β+
γb
− ∑2k=1 √1+B
0
k,2
k
o
√
−α+
(U
)γ
+ e k,2 0 b Q β−
k,2 (U0 ) γb
n
√
−α−
(U
)γ
2 √ Ak
+ ∑t=1 1+2B e k,1 0 b Q β+
(U
)
γb
0
k
k,1
√
−α+
(U
)γ
+ e k,1 0 b Q β−
(U0 ) γb
k,1
√
−α−
(U
)γ
+
1
b
− e k,1
Q βk,1 (U1 ) γb
√ o
−α+ (U )γ
− e k,1 1 b Q β−
.
k,1 (U1 ) γb
Recall that, for the sake of mathematical tractability, we
assume that the received bits with their initial reliability within
the interval |r̄| ≤ U0 will be retransmitted D times, the bits with
the reliability in the interval U0 < |r̄| ≤ U1 will be retransmitted
D − 1 times and so on. The number of the received copies
for each reliability range after D retransmissions is shown in
Fig. 2. Consequently, the overall BER can be expressed using
the functions χD (r̄,UD−1 ) and λD (r̄,UD ,UD−1 ) as,
R −U
BERD = −∞ D−1 fr̄ (r̄|S1 , D = 0) dr̄
R0
R0
+ ∑D−1
i=1 −∞ λi (r̄,UD−i ,UD−i−1 ) dr̄ + −∞ χD (r̄,U0 ) dr̄
where the middle term is zero for D = 1 retransmission. This
BER can be accurately approximated as,
p
√
UD−1
+
1
+
Q
2γ
(D
+
1)
BERD ≈ Q 2γb 2γ
b
nb −
√
Ak
2
q
−
e−αk,D (U0 )γb Q β+
∑k=1
The probability that the received reliability is in the interval
k,D (U0 ) γb
2Bk
1+ D
+
U0 ≤ |r̄| ≤ U1 is evaluated as,
√ o
γb
+ e−αk,D (U0 )γb Q β−
(U
)
0
k,D
n
P1 = Pr |r̄| ≤ U1 S1 , d = 0 Pr (d = 0|S1 ) +
−
√
−α
(U
D−1 2
R U1
k,i D−i−1 )γb Q β+ (U
q Ak
+
e
)
γb
∑
∑
D−i−1
k=1
i=1
k,i
Pr |r̄| ≤ U1 S1 , d = 1 Pr (d = 1|S1 ) = −U1 fr̄ (r̄|S1 , D = 0) dr̄
2B
1+ i k
R U1
R U0
√
− U0 fr̄ (r̄|S1 , D = 0) dr̄ + −U1 χ1 (r̄,U0 ) dr̄.
−α+ (U
)γ
+ e k,i D−i−1 b Q β−
k,i (UD−i−1 ) γb
This probability can be efficiently approximation as,
√
−α− (U )γ
− e k,i D−i b Q β+
(UD−i ) γb
√
√
k,i
U0
U0
+ 1) + Q 2γb ( 2γ
− 1)
P1 ≈ Q 2γb ( 2γ
√ o
−α+ (U )γ
b
b
− e k,i D−i b Q β−
.
√
√
k,i (UD−i ) γb
U1
U1
+
1)
−
Q
−
1)
−Q 2γb ( 2γ
2γ
(
b
2γ
b
b
The probability PD that the received reliability is in the interval
UD−1 ≤ |r̄| ≤ UD is evaluated as,
√
√
√
√
+Q 2 γb − U1 /n γb − Q 2 γb+ U1 / γb
Z UD
Z U0
+
√
Ak
e−αk,1 (U0 )γb erf θ+−
(U
)
γ
− ∑2k=1 2√1+2B
0
b
P
=
f
(r̄|S
,
D
=
0)
dr̄
−
fr̄ (r̄|S1 , D = 0) dr̄
k,1
D
r̄
1
k
−UD
−U0
++
√
√
−α−
(U
)γ
−+
0
b
+erf θk,1 (U0 ) γb + e k,1
erf θk,1 (U0 ) γb
Z UD
D−1 Z UD
o
√
λ
(r̄,U
,U
)
dr̄
+
χD (r̄,U0 ) dr̄
+
−−
i
D−i
D−i−1
∑
+sign(µ(U0 )) erf θk,1 (U0 ) γb
−UD
−UD
i=1
6
with the third term being zero for D = 1 retransmission. It is
again possible to obtain the approximation of the probability
PD for D > 2. However, the resulting expression is much more
involved than in the case of P1 and P2 , so it is not presented
here.
D. BER for slowly varying fading channels
Since the wireless links are typically subject to time-varying
fading, we can average the BER expressions obtained in the
previous subsection over the time-varying received power, for
example, following the chi-square distribution. In this
case,
γb
1
the received SNR has the PDF, fγb (γb ) = γ¯b exp − γ¯b , where
γ¯b = E[γb ] is the average SNR. Moreover, we assume that the
SNR variations are sufficiently slow (e.g., assuming that the
nodes are nomadic or even stationary), so the SNR can be
considered to be constant during the first packet transmission
as well as during the subsequent D retransmissions. Then, the
long-term average BER is evaluated using the expectation [3],
BERD =
Z ∞
0
BERD (γb ) fγb (γb ) dγb .
The computationally efficient formula for calculating BERD is
displayed below Fig. 2 on the next page. Similarly, the average
probability PD is evaluated as,
P̄D =
Z ∞
0
PD (γb ) fγb (γb ) dγb .
IV. B ITWISE R ETRANSMISSION P ROTOCOLS
We consider three specific bitwise retransmission strategies
and optimize their parameters to maximize the transmission
reliability rather than to maximize their throughput. We also
verify mathematical formulas derived in the previous section
by computer simulations assuming error-free feedback. This
assumption is revisited at the end of this section.
A. Fixed rate technique
In this design, we assume a constant retransmission window
size W determined as,
N 1
−1
W=
D Rf
for a priori given parameters N, D and the forward rate Rf .
Since the window size W is constrained as 1 ≤ W ≤ N, the
1
possible forward rates are restricted to the interval, 1+D
< Rf ≤
1
N
.
Thus,
for
W
=
N,
the
rate
R
=
,
and
the
proposed
f
D+N
1+D
retransmission scheme corresponds to a block repetition code
(BRC) with D repetitions of the original packet. We have the
following proposition.
Proposition 1: For a given SNR, the fixed rate retransmission technique achieves the minimum BER for some specific
value of the rate Rf . The optimum value of Rf minimizing the
BER increases with the SNR.
Prop. 1 can be proved by letting the derivative ( d/ dRf ) BERD
to be equal to zero. For large values of SNR and N, the
forward rate Rf approaches its maximum value of 1, so in such
case, the overhead due to retransmissions can be neglected.
Moreover, when N is large or W = N, the BER of the fixed
rate retransmission scheme approaches the BER of the BRC.
Fig. 3 and Fig. 4 show the BER of the forward link versus
the forward rate Rf for the different number of retransmissions
D at two SNR values γb = 0 dB and γb = 5 dB, respectively.
We observe that the minimum BER value is more pronounced
(i.e., the optimization is more important) for larger values of
SNR. The BER curves in Fig. 3 and Fig. 4 also compare
the simulations with the approximate expressions given in
the previous section; for D > 1, the difference between the
approximate expressions and the simulations is negligible.
The forward rates Rf yielding the minimum BER are shown
in Fig. 5 for the different number of retransmissions D. Finally,
Fig. 6 shows the BER versus the SNR γb for the different
number of retransmissions D assuming that the forward rate
Rf is optimized for each SNR value γb to minimize the
achieved BER. Note again that such optimization becomes
more effective if the SNR is increased.
B. Fixed window technique
As for the fixed rate technique, the retransmission window
size is fixed, and it is determined as,
W = ⌈N P⌋
where the retransmission decision thresholds Ud are set, so
that the probabilities P = P0 = P1 = . . . = PD−1 , which have
been obtained in the previous section, have the same value.
Then, given N, D and W , the forward rate is calculated as,
1
.
Rf =
1 + DW /N
We have the following proposition.
Proposition 2: For a given SNR, the fixed window technique
achieves the minimum BER for some value of the retransmission window size. The optimum window size value decreases
with the SNR.
Prop. 2 can be again proved by letting the derivative,
( d/ dW ) BERD to be equal to zero. When the ratio W /N
approaches unity, the BER of the fixed window technique
approaches the BER of the BRC. On the other hand, when
the ratio W /N approaches zero, the retransmission overhead
can be neglected.
Fig. 7 and Fig. 8 show the BER versus the normalized window size W /N for the different number of retransmissions D
and the SNR values γb = 0 dB and γb = 5 dB, respectively. We
again observe a negligible difference between the approximate
and the simulated BER curves, especially for larger values of
D. In addition, the minimum BER values are more pronounced
when the SNR is increased.
Fig. 9 shows the normalized window size W /N corresponding to the minimum BER for the different number of
retransmissions D. Fig. 10 shows the BER versus the SNR
γb for the different number of retransmissions D assuming
the optimum value of W /N for each γb minimizing the BER.
Note that such minimization of the BER is more effective
when the SNR is large. Finally, for the different number of
retransmission D, Fig. 11 shows the BER for a slowly varying
chi-square distributed, unit-mean received power, provided that
the parameters are optimized to minimize the BER.
7
Figure 2. The quantization levels of the bit reliability with D retransmissions.
BERD ≈ 1 − 12
t
4
γ¯b (UD−1 /γb +2)2
+ ∑ q Ak2B
k=1 1+ Dk
(
×
−1/2
−1/2
1
− 12 1 + γ¯ (D+1)
b
r
1 + γ¯b α−
k,D (U0 ) +
+ 1 + γ¯b(α+
k,D (U0 ) +
(
+1
β+
(U0 ) β+
(U0 )
k,D
√
( k,D√2
2
β−
(U0 ) β−
(U0 )
k,D
√
( k,D√2
2
+
r
1
γ¯b
r
−
β−
k,i (UD−i−1 ) βk,i (UD−i−1 )
√
√
(U
)
+
+ 1 + γ¯b α+
(
+
D−i−1
k,i
2
2
−
β−
D−i ) βk,i (UD−i )
k,i (U
√
− 1 + γ¯b α+
( √2 +
k,i (UD−i ) +
2
C. Fixed threshold technique
Unlike the previous two techniques, the fixed threshold
technique allows for different retransmission window sizes
while assuming the single constant reliability threshold, U0 =
U1 = . . . = UD−1 = U, during each retransmission. Given U, we
obtain the probabilities Pd , d = 1, 2, . . . , D, defined previously,
and calculate the retransmission window sizes as,
Wd = ⌈NPd ⌋ .
The corresponding forward rate is given as,
Rf =
1
.
1 + ∑D
d=1 Wd /N
We have the following proposition.
Proposition 3: For a given SNR, the fixed threshold technique achieves the minimum BER value for some specific
threshold U. This optimum threshold value is increasing with
the SNR.
Prop. 3 can be again proved by letting the derivative,
( d/ dU) BERD to be equal to zero.
1
γ¯b
+ α+
k,D (U0 ) +
+
β+
k,i (UD−i−1 ) βk,i (UD−i−1 )
√
√
(U
)
+
1 + γ¯b α−
(
+
D−i−1
k,i
2
2
+
β+
D−i ) βk,i (UD−i )
k,i (U
√
√
(U
)
+
− 1 + γ¯b α−
(
+
D−i
k,i
2
2
+
2
1
γ¯b
r
+ α−
k,D (U0 ) +
β−
k,D (U0 )
2
2
!−1 )
)
+ α−
k,i (UD−i−1 ) +
1
γ¯b
β+
k,D (U0 )
2
!
−1
D−1 t
+ ∑ ∑
β+
k,i (UD−i−1 )
2
2
β−
k,i (UD−i−1 )
+ α+
(U
)
+
D−i−1
k,i
2
r
β+
k,i (UD−i )
+ α−
(U
)
+
D−i
k,i
2
2
1
γ¯b
r
β−
k,i (UD−i )
+ α+
k,i (UD−i ) +
2
2
1
γ¯b
q Ak
2B
1+ k
!−1 i
i=1 k=1
!−1
!−1 )
2
×
!−1
.
Fig. 12 and√ Fig. 13 show the BER versus the normalized
threshold U/ Eb for the different number of retransmissions
D and the SNR values γb = 0 dB and γb = 5 dB, respectively.
Since the approximate BER expressions were already verified
for the other two techniques considered, the BER curves in
Fig. 12 and Fig. 13 only show these derived expressions.
We again observe that the minimum BER values are more
apparent when the SNR√is increased. Fig. 14 shows the
normalized thresholds U/ Eb having the minimum BER for
the different number of retransmissions D. Finally, Fig. 15
shows the BER versus the √SNR γb for different D assuming
the optimum thresholds U/ Eb for each SNR γb that achieves
the minimum BER. We again find that
√ minimization of the
BER by optimizing the threshold U/ Eb is more effective
when the SNR is increased.
D. Feedback signaling
There are two main issues when designing practical feedback signaling schemes. The first issue is how to constrain the
8
1
10
sim, D=1
math, D=1
sim, D=2
math, D=2
sim, D=3
math, D=3
sim, D=4
math, D=4
-1.2
D =1
D=2
D=3
D=4
0.95
0.9
Rf
BER
0.85
0.8
10
-1.3
0.75
0.7
PSfrag replacements
eplacements
0.55
0.6
0.65
0.7
0.75
Rf
0.8
0.85
0.9
0.95
Figure 3. The BER versus the forward rate R f for SNR = 0 dB and the
different number of retransmissions D.
10
0.65
0
1
3
4
5
γb [dB]
6
7
8
9
10
0
10
-2
binary, theory
BRC
sim, D=1
math, D=1
sim, D=2
math, D=2
sim, D=3
math, D=3
sim, D=4
math, D=4
-1
10
-2
10
BERD
BER
2
Figure 5. The rate R f yielding the minimum BER versus the SNR γb for the
different number of retransmissions D.
sim, D=1
math, D=1
sim, D=2
math, D=2
sim, D=3
math, D=3
sim, D=4
math, D=4
10
1
-3
-3
10
-4
10
PSfrag replacements
eplacements
10
-4
0.55
0.6
0.65
0.7
0.75
Rf
0.8
0.85
0.9
0.95
1
0
1
2
3
4
5
γb [dB]
6
7
8
9
10
Figure 4. The BER versus the forward rate R f for SNR = 5 dB and the
different number of retransmissions D.
Figure 6. The BERD versus the SNR γb for the different number of
retransmissions D.
number of feedback bits. The second issue is the transmission
errors of feedback bits.
Sending a small number of feedback bits in a dedicated
packet is very inefficient due to the associated protocol overheads. In practice, it is common to reserve a few bits within
the packet payload for the feedback signaling, so the packet
overhead is shared by the feedback as well as data. In such
case, it is beneficial to minimize the number of feedback bits in
order to increase the data payload. Here, we reduce the number
of feedback bits sent from the destination to the source over
the reverse link by considering a deterministic sequence of bit
permutations synchronously generated at the transmitter and at
the receiver to encode the feedback message. Such sequence
is conveniently generated as pseudo-random permutations of
N-tuples, (1, 2, . . . , N), using the two synchronized RNGs. In
addition, the transmitter and the receiver RNGs synchronously
advance by exactly 2C1 permutations every symbol period. For
every generated permutation, the receiver checks whether a
sufficient number of bits with small reliabilities fall into a
predefined window of W bits; for instance, we can assume
that the retransmission window is represented by the the first
W bit positions. When such permutation is found, the feedback
message to notify the source is a binary representation of the
permutation order modulo 2C1 . The source then selects the
corresponding W bits in the packet, and retransmits them to the
destination. In particular, for the window size W , the number
of permutations K searched until the desired one has been
found can be expressed as,
K = 2C1 I + (K mod 2C1 )
9
0.35
0.3
W/N,
W/N,
W/N,
W/N,
D
D
D
D
8
9
=1
=2
=3
=4
0.25
-1.2
BER
W /N
10
0.2
0.15
sim, D=1
math, D=1
sim, D=2
math, D=2
sim, D=3
math, D=3
D=4
PSfragsim,
replacements
math, D=4
-1.3
10
replacements
0
0.2
0.4
W /N
0.6
0.8
0.1
0.05
0
0
1
Figure 7. The BER versus the normalized window size W /N for SNR = 0
dB and the different number of retransmissions D.
1
2
3
4
5
γb [dB]
6
7
10
Figure 9. The normalized window size W /N yielding the minimum BER
versus the SNR γb for the different number of retransmissions D.
-2
0
10
10
binary, theory
BRC
sim, D=1
math, D=1
sim, D=2
math, D=2
sim, D=3
math, D=3
sim, D=4
math, D=4
-1
10
-2
BERD
BER
10
-3
10
sim, D=1
math, D=1
sim, D=2
math, D=2
sim, D=3
math, D=3
sim, D=4
PSfrag
replacements
math,
D=4
replacements
-4
10
0
0.1
0.2
0.3
0.4
0.5
W /N
0.6
0.7
0.8
0.9
Figure 8. The BER versus the normalized window size W /N for SNR = 5
dB and the different number of retransmissions D.
where I is an integer number of the idle symbol periods. The
key property is that,
N
C1
(K mod 2 ) ≪
W
N
whereas, on average, E[K] = W
. Hence, the feedback message is only represented by C1 bits. The value of C1 is a design
parameter, and it trade-offs the feedback message size, and the
required delay until the start of the next retransmission [22].
More importantly, given some target BER, the window size
W and the average number of searched permutations E[K] are
decreasing with the SNR. As shown in Fig. 16, E[K] is only
about 30 for all N ≤ 64 when the SNR is at least 10 dB,
and thus, the feedback of C1 = 5 bits can be used to find the
desired permutation during one symbol period.
-3
10
-4
10
0
2
4
γb [dB]
6
8
10
Figure 10. The BERD versus the SNR γb for the different number of
retransmissions D.
The average forward throughput with one retransmission
can be defined as,
N
ζ1 =
.
E[I + 1]
By simulations, we found that there exists an optimum value
C1∗ which minimizes the expected delay E[I + 1], i.e., which
maximizes the throughput ζ1 . This optimum value is given as,
C1∗ = ⌈−0.5 + log2 E[K]⌋
and it is plotted in Fig. 17 for different values of the packet
length N. The corresponding maximum throughput ζ∗1 is
shown in Fig. 18. For comparison, the throughput 1/2 of the
BRC corresponding to the conventional stop-and-wait ARQ
with one retransmission is also shown in Fig. 18. More importantly, we observe that the proposed scheme can achieve better
10
0
-2
10
10
binary, theory
D=1
D=2
D=3
-1
10
math, D=1
math, D=2
math, D=3
math, D=4
-3
10
-2
BER
BER
10
-3
10
-4
10
-4
10
PSfrag replacements
eplacements
-5
10
-5
0
1
2
3
4
5
γb [dB]
6
7
8
9
10
10
0
1
√
3
U/ Eb
2
4
-1.1
15
U/Eb ,
U/Eb ,
U/Eb ,
U/Eb ,
math, D=1
math, D=2
math, D=3
math, D=4
-1.2
10
6
√
Figure 13. The BER versus the normalized threshold U/ Eb for the SNR
γb = 5 dB and the different number of retransmissions D.
Figure 11. The average BER versus the SNR γb over the slowly varying
chi-square distributed transmission power.
10
5
D
D
D
D
=1
=2
=3
=4
10
-1.3
√
U/ Eb
BER
10
-1.4
10
5
-1.5
10
eplacements
-1.6
10
PSfrag replacements
0
0.5
1
√
1.5
U/ Eb
2
2.5
√
Figure 12. The BER versus the normalized threshold U/ Eb for the SNR
γb = 0 dB and the different number of retransmissions D.
throughput than the rate 1/D repetition code for the same
number of retransmissions D, especially at the medium to large
values of SNR. Furthermore, we can show that the throughput
of the fixed window technique in the limit of very large SNR
N
converges to, limγb →∞ ζ1 ≈ N+1
, since limγb →∞ W /N = 0.
Finally, we reconsider the assumption of the error-free feedback which is often adopted in many papers concerning the
ARQ retransmission schemes. Recall that it was shown in [14]
that errors of 1 and 2-bit feedback messages can be neglected
if their bit error probability is less than 10−3. This result can be
readily modified for the case of multi-bit feedback messages
which are utilized in our bitwise retransmission schemes.
In particular, assuming the feedback errors are independent,
and they are occurring with the probability pr , we have the
following proposition.
3
0
0
1
2
3
4
5
γb [dB]
6
7
8
9
10
√
Figure 14. The values of the normalized threshold U/ Eb corresponding
to the minimum BER versus the SNR γb for the different number of
retransmissions D.
Proposition 4: The errors in the feedback messages of C
bits can be tolerated, provided that their bit-error probability
is bounded as,
1/C
pr ≤ 1 − Pmin
where the required minimum probability of the error-free
feedback messages was established in [14] to be, Pmin =
1 − 10−3 = 99.9%.
The proof of Prop. 4 follows from the binomial distribution
of independent and equally probable errors. Hence, the longer
the feedback message, the smaller the feedback bit error
probability pr is required.
11
0
2
10
10
binary, theory
BRC
math, D=1
math, D=2
math, D=3
math, D=4
-1
10
N=8
N=16
N=32
N=64
-2
1
C1∗
BERD
10
10
-3
10
-4
10
replacements
PSfrag replacements
0
0
2
4
γb [dB]
6
8
10
10
Figure 15. The BERD versus the SNR γb for the different number of
retransmissions D.
0
1
2
3
4
5
γb [dB]
6
7
8
9
10
Figure 17. The optimum values C1∗ with D = 1 retransmission versus the
SNR γb and the different packet length N.
9
10
1
N=8
N=16
N=32
N=64
8
10
0.9
7
10
0.8
6
10
5
10
ζ∗1
E[K]
fixed win, N =64
fixed win, N =32
fixed win, N =16
BRC
0.7
4
10
0.6
3
10
0.5
2
10
eplacements
PSfrag replacements
1
10
0
1
2
3
4
5
γb [dB]
6
7
8
9
10
0.4
0
5
γb [dB]
10
15
Figure 16. The expected number of permutations E[K] versus the SNR γb
and the different packet length N.
Figure 18. The throughput ζ∗1 corresponding to the optimum values C1∗ versus
the SNR γb and the different packet length N.
V. DATA F USION A PPLICATION
node technologies: Zigbee 802.15.4, Wifi 802.11b and Bluetooth v. 4.2 802.15.1. The parameters of these three technologies are summarized in Table. I. Their BER expressions have
been obtained by fitting the sum of exponentials (the Prony
method, [25]) to the performance curves reported in [26]. It
is obvious that Zigbee is the most energy efficient technology
for the sensor nodes, and it can operate at small SNR values.
The specific design examples of the bitwise retransmission
schemes with a constant retransmission window size employing the binomial number feedback and the packet segmentation
are listed in Table. II. Therein, p f is the BER of the forward
link, pr is the BER of the reverse link, Nseg is the number
of segments of the disjoint partitioning of the N bit packet,
Wseg is the retransmission window size per segment, and Ctot
is the total length of the feedback message (in bits) required
We now illustrate the use of the proposed bitwise retransmission schemes in a practical scenario of data fusion from
a group of L sensor nodes into a single central access point
(AP). A time-division multiple access (TDMA) protocol with
(L + 1) time slots is used to share the communication channel.
The time-division duplex (TDD) protocol further divides the
available time slots into L uplink time slots to transmit data
from the L sensor nodes, and one time slot is allocated for the
feedback signaling from the AP. Each time slot can carry at
most N bits of information including the protocol overhead.
The AP is assumed not to be battery powered, so the transmit
power in the downlink can be much larger than in the battery
constrained uplink.
More specifically, we consider the following three sensor
12
Table I
T RANSMISSION PARAMETERS OF THE THREE SENSOR NODE TECHNOLOGIES
Zigbee
BER
10−2
10−3
10−4
10−5
10−6
Packet
[bytes]
N [bits]
Segments
Pb (γb ) = 1.5203 e−9.5611γb
γb [dB]
-2.79
-1.16
0.03
0.96
1.73
header
6
payload
127
(6 + 127) × 8 = 1064
14×76, 19×56, 28×38
38×28, 56×19, 76×14
133×8, 152×7, 266×4
532×2, 1064×1
Wifi
Pb (γb )
=
Bluetooth
0.2436 e−0.4997γb
+0.2436 e−0.4997γb
γb [dB]
8.91
10.93
12.30
13.34
14.18
header+CRC
2+2
payload
252
(4 + 252) × 8 = 2048
10.0 e−3.4535γb
Pb (γb )
+1.1066 e−2.0247γb
γb [dB]
3.88
5.43
6.63
7.59
8.37
preamble+header
15–24
payload
1500
(24 + 1500) × 8 = 12192
16×762, 32×381, 48×254
96×127, 127×96, 254×48
381×32, 508×24, 762×16
1016×12, 1524×8, 2032×6
3048×4, 4064×3, 6096×2
12192×1
for the whole packet. Hence, N/Nseg is the segment length,
Wseg Nseg = W is the total number of retransmitted bits, and
Ctot /Nseg is the length of the feedback message per segment.
Furthermore, P̃f is the probability that there are at most Wseg
errors in any segment (i.e., the closer the value of P̃f to 1.0, the
better), and P̃r is the probability that there is at least one error
among the Ctot feedback bits received at the source (i.e., the
probability that the feedback message is received incorrectly
at the source). As shown in the previous section, it is required
that P̃r < Pmin = 10−3 in order to neglect the effect of feedback
errors on the performance. Other system parameters are given
in Table. I including the packet size N and the required SNR
for given values of p f .
We found that, to obtain efficient designs of the proposed
bitwise retransmission schemes, the BER p f of the forward
link should be at most 10−3 , and the BER pr of the reverse
link should be at most 10−5. In terms of the minimum required SNR for bidirectional connections, the Zigbee requires
−1.16 dB and 0.96 dB, the Wifi requires 5.43 dB and 7.59
dB, and the Bluetooth requires 10.93 dB and 13.34 dB,
respectively. Such SNR levels can be satisfied for all three
wireless technologies considered, provided that the reverse link
has 3 dB larger transmit power than the forward link. Such
transmission power unbalances can be readily obtained in the
sensor networks with the centralized mains-powered AP.
Next, we consider scheduling of the packet contents for our
single-cell TDMA/TDD multiple access protocol. We assume
that the parameters W , D, N, L and Ctot are constant, even
though they can be optimized for the uplink and downlink
BERs p f and pr . Recall that all packets have the maximum
length of N bits. In the uplink, the nodes send their data
as well as schedule the retransmitted bits for the previously
transmitted data packets. In the downlink, the AP broadcasts
the retransmission requests to all sensor nodes at once. The
contents in the uplink packets are scheduled following the
following two rules.
1) Include the retransmitted sequences in the order corre-
=
16×128, 32×64, 64×32
128×16, 256×8, 512×4
1024×2, 2048×1
Table II
T HE BITWISE RETRANSMISSION DESIGNS WITH THE CONSTANT WINDOW
SIZE AND THE PACKET SEGMENTATION
pf
pr
Nseg
Wseg
Ctot
P̃f
P̃r
Zigbee
10−3
10−3
10−3
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−5
10−5
10−5
10−5
10−5
10−5
10−5
10−5
10−5
10−5
10−5
2
1
1
4
2
2
2
1
1
1
1
3
4
5
1
1
2
3
1
2
3
4
50
36
44
36
20
36
50
11
20
28
36
0.9978
0.9953
0.9992
0.9997
0.9986
1.0000
1.0000
0.9947
0.9998
1.0000
1.0000
Wifi
10−3
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−6
10−6
10−6
10−6
10−6
10−6
10−6
10−6
10−6
10−6
5.0 · 10−4
3.6 · 10−4
4.4 · 10−4
3.6 · 10−4
2.0 · 10−4
3.6 · 10−4
5.0 · 10−4
1.1 · 10−4
2.0 · 10−4
2.8 · 10−4
3.6 · 10−4
1
4
3
2
2
1
1
1
1
1
21
2
2
3
4
4
5
6
7
8
220
92
69
72
92
50
61
72
83
94
0.9928
0.9962
0.9917
0.9965
0.9996
0.9917
0.9984
0.9997
1.0000
1.0000
Bluetooth
10−2
10−3
10−3
10−3
10−3
10−4
10−4
10−4
10−4
10−4
10−4
10−6
10−6
10−6
10−6
10−6
10−5
10−5
10−5
10−5
10−5
10−5
2
2
1
1
1
4
2
2
1
1
1
18
4
6
7
8
1
1
2
2
3
4
256
72
57
65
73
36
20
38
21
31
40
0.9913
0.9960
0.9949
0.9987
0.9997
0.9987
0.9951
0.9998
0.9988
0.9999
1.0000
2.2 · 10−4
9.2 · 10−5
6.9 · 10−5
7.2 · 10−5
9.2 · 10−5
5.0 · 10−5
6.1 · 10−5
7.2 · 10−5
8.3 · 10−5
9.4 · 10−5
2.6 · 10−4
7.2 · 10−5
5.7 · 10−5
6.5 · 10−5
7.3 · 10−5
3.6 · 10−4
2.0 · 10−4
3.8 · 10−4
2.1 · 10−4
3.1 · 10−4
4.0 · 10−4
sponding to the previously transmitted packets. Only one
retransmitted sequence per each previously transmitted
packet can be scheduled.
2) Add data bits from the buffered information blocks to
13
Table III
T HE UPLINK PACKET CONTENTS FOR N = 1064, W = 4, D = 3, Lpac = 10
#
1
2
3
4
5
6
7
8
9
10
11
12
13
14
packet content
D1 (1064)
R1,1 (4), D2 (1060)
R1,2 (4), D2 (4), D3 (1056)
R1,3 (4), R2,1 (4), D3 (8), D4 (1048)
R2,2 (4), R3,1 (4), D4 (16), D5 (1040)
R2,3 (4), R3,2 (4), R4,1 (4), D5 (24), D6 (1028)
R3,3 (4), R4,2 (4), R5,1 (4), D6 (36), D7 (1016)
R4,3 (4), R5,2 (4), R6,1 (4), D7 (48), D8 (1004)
R5,3 (4), R6,2 (4), R7,1 (4), D8 (60), D9 (992)
R6,3 (4), R7,2 (4), R8,1 (4), D9 (72), D10 (980)
R7,3 (4), R8,2 (4), R9,1 (4), D10 (84)
R8,3 (4), R9,2 (4), R10,1 (4)
R9,3 (4), R10,2 (4)
R10,3 (4)
fill in the whole packet of N bits.
Hence, the transmitted packet can contain retransmitted bits
for multiple previously transmitted packets, and also data bits
from multiple information blocks. As an example, assuming
first-in first-out (FIFO) buffering of the information blocks
of Nbuf = N = 1064 bits (Zigbee protocol) with Nseg = 2
segments, Wseg = 2 bits, i.e., the retransmission window of
W = 2 × 2 = 4 bits, D = 3 retransmissions, and in total Lpac =
10 data packets to be transmitted, the packets content schedule
is shown in Table. III. Therein, we use the notation Dl (n) to
denote a sequence of n bits belonging to the l-th information
block, and Rl,d (m) is the sequence of m retransmitted bits in
the d-th retransmission for the l-th information block where
the sequence indexes, 1 ≤ l ≤ Lpac , 1 ≤ n ≤ N, 1 ≤ d ≤ D and
1 ≤ m ≤ W . We have the following proposition.
Proposition 5: For N ≫ DW , Nbuf = N and Lpac ≥ 1, only
the very first information block is completely transmitted in
the first packet whereas all other information blocks are split
into exactly 2 subsequent packets. The D retransmissions
are scheduled into D subsequent packets immediately after
the transmission of the corresponding information block was
completed. Moreover, only the first Lpac transmitted packets
are fully occupied with N bits. The available transport capacity
in the last (D + 1) transmitted packets can be used to transmit
additional information blocks with the progressively shorter
block lengths and/or smaller number of retransmissions.
The main assumption required in Prop. 5 is that the total
number of retransmitted bits D · W is much smaller than the
block length N. If this condition is not satisfied, the packet
structure of the retransmission protocol is less predictable.
We conclude our discussion about the bitwise retransmission
scheme for the uplink data fusion from L sensor nodes by
considering the packets structure in the downlink. According
to Prop. 5, the AP (the data fusion center) sends the retransmission requests for the l-th information block during the time
slots {1, 2, . . . , D}, if l = 1, and {l + 1, l + 2, . . ., l + D}, if l ≥ 2.
However, since the information blocks with the index l ≥ 2 are
transmitted exactly in 2 subsequent packets with the indexes
l and l + 1, the number of retransmission requests contained
in the downlink packet first raises to the maximum value of
D requests per sensor node. The number of the requests then
remain constant until it is gradually decremented to 1 request
in the last (D − 1) transmissions. Consequently, we have the
last proposition.
Proposition 6: The maximum number of sensor nodes
Lmax which can be supported by the proposed retransmission
scheme is bounded as,
N − Novh
Lmax ≤
D ·Ctot
where Novh is the protocol overhead, and Ctot is the total
number of feedback bits per retransmission and sensor node.
For the example presented in Table. III, assuming Novh = 106
bits (10% of N), we have Ctot = 36 bits, so Lmax ≤ 8. A larger
number of sensor nodes can be supported by trading off the
number of feedback bits Ctot with the number of retransmissions. In addition, it is possible that different sensor nodes
set their retransmission parameters differently, for example, to
match the SNR they experience in the uplink and downlink.
In this case, the value of Lmax in Prop. 6 can be calculated
by assuming the maximum total number of feedback bits,
max(D ·Ctot ), allowed per any sensor node.
VI. C ONCLUSIONS
A novel bitwise retransmission scheme was presented to
selectively retransmit only the bits which were received with
a small reliability. The bitwise retransmission decisions as well
as combining can be done either immediately after demodulating the received symbols, or after the channel decoding. In
case of the turbo decoding, the bit reliabilities are available as
the soft-input (i.e., channel output) or the soft-output values
[23]. The locations of the bits to be retransmitted are reported
to the source as a binomial combination number. Since the
proposed scheme does not involve any complex operations, it
does not limit the processing throughput at the receiver nor at
the transmitter.
The analysis presented in the paper assumes uncoded binary
modulation for mathematical tractability and clarity of the
presentation. However, the proposed scheme can be readily generalized to non-binary modulations by appropriately
scaling the demodulated bits. We derived the overall BER
conditioned on the error-free feedback link. The accurate
closed-form BER expressions as well as their computationally
efficient approximations were presented assuming one and two
retransmissions. The approximations were verified by computer simulations. We showed that the bitwise retransmissions
can be optimized for the given SNR, especially in the forward
link, since the BER curves are convex in the transmission rates
as well as in the reliability threshold. In addition to minimizing
the BER as investigated in this paper, it is possible to maximize
the throughput instead.
We next compared the BER and throughput performances of
the three specific retransmission strategies which are referred
to as the fixed rate technique, the fixed window technique,
and the fixed threshold technique. It was shown that, for
the same number of retransmissions, and the same packet
length, the proposed schemes always outperform the repetition
diversity, and, in some cases including the transmissions over
time-varying channels, the performance improvement can be
14
significant. Furthermore, in order to reduce the number of
feedback bits, we proposed to use two synchronized RNGs at
the transmitter and at the receiver which can greatly compress
the transmitted binary feedback sequences at the expense of
larger delays. We also derived a condition when the impact of
feedback errors to the overall performance can be neglected.
We then considered practical design issues of multi-user
bitwise retransmission schemes for data fusion applications
where the sensor nodes forward data packets in the uplink into
a centralized AP. The retransmission requests are broadcasted
to the sensor nodes from the AP in the downlink. Assuming
TDMA/TDD and the Zigbee, Wifi and Bluetooth protocols, we
presented the design examples of retransmission parameters.
These examples suggest that the efficient designs of the bitwise
retransmission schemes can be obtained provided that the BER
of the forward link is below 10−3 and the BER of the reverse
link is below 10−5 . If the bidirectional link has the same or
similar BERs in both directions, the smaller BER in the reverse
link can be readily achieved by increasing the transmit power
in the downlink (i.e., at the AP). We also devised scheduling
of the information and retransmission bits utilizing the FIFO
buffering of the information blocks in order to fully fill up the
transmitted packets. Finally, we calculated the upper bound
on the number of sensor nodes which can be supported by the
proposed bitwise retransmission schemes.
A PPENDIX
We obtain several approximations to efficiently evaluate
the integrals in the BER expressions. The approximations are
based on the Prony approximation of the Q(x) function [25],
2
Q(x) =
2
∑ Ak e−Bk x
k=1
where A1 = 0.208, A2 =
0.147,B1 = 0.971, and B2 = 0.525.
√
Since erf(x) = 1 − 2Q 2x , assuming positive constants
H > 0, and hi > 0, i = 1, 2, 3, 4, 5, the expressions for BERD
and PD can be approximated using the following expressions.
These approximations are used and verified numerically in the
figures presented in Section IV.
R0
−∞ h1 e
(r̄−h2 )2
h3
≈ ∑tk=1
≈
R0
−
−∞ h1 e
Q(h4 (h5 − r̄)) dr̄
R0
−∞ h1 e
q h1 A k
2 h1 +h4 2 Bk
e
3
−
(r̄−h2 )2
h3
≈ ∑tk=1
A e
dr̄
√
π erfc √
Q(h4 (h5 + r̄)) dr̄
R0
−∞ h1 e
−
(r̄−h2 )2
h3
h1 h3 A k
h3 (1+h3 h4 2 Bk )
≈ √
2
(r̄−h2 )2
−Bk (h4 (h5 −r̄))2
h3
k
2
B h (h −h )2
− k 4 22 5
h2 +h3 h4 2 h5 Bk
1+h3 h4 Bk
h3 (1+h3 h4 2 Bk )
−
2
Ak e−Bk (h4 (h5 +r̄)) dr̄
Bk h4 2 (h2 +h5 )2
−
√
2
2
π erfc √h2 −h3 h4 h52Bk
e 1+h3 h4 Bk
h3 (1+h3 h4 Bk )
RH
−H h1 e
RH
−
(r̄−h2 )2
h3
Q(h4 (h5 − r̄)) dr̄
(r̄−h )2
RH
2
− h2
3
Ak e−Bk (h4 (h5 −r̄)) dr̄
−H h1 e
2
2
B h (h −h )
√ √
− k 4 22 5 n
h2 +H+h3 h4 2 (h5 +H)Bk
h1 A k h3 π
1+h3 h4 Bk
√
e
≈ √
erf
2
h3 (1+h3 h4 2 Bk )
2 1+h
3 h4 B k
h2 +h3 h4 2 h5 Bk
+sign H − 1+h h 2 B
3 4 k
o
q
h2 +h3 h4 2 h5 Bk
2
1
H
−
+
h
B
×erf
4
k
2
h3
1+h3 h4 Bk
≈ ∑tk=1
−H h1 e
−
(r̄−h2 )2
h3
≈ ∑tk=1
Q(h4 (h5 + r̄)) dr̄
RH
−H h1 e
√
−
(r̄−h2 )2
h3
2
Ak e−Bk (h4 (h5 +r̄)) dr̄
B h 2 (h +h )2
− k 4 22 5
1+h3 h4 Bk
≈ √ h1 h3 Ak π2 e
2 h3 (1+h3 h4 Bk )
o
n
2
H−h2 +h3 h4 2 (H+h5 )Bk
5 )Bk
√2 +h3 h4 (H−h
√
+
erf
× erf H+h
2
2
h3 (1+h3 h4 Bk )
h3 (1+h3 h4 Bk )
R EFERENCES
[1] Goldsmith A, Varaiya P. Capacity of fading channels with channel side
information. IEEE Trans. Inform. Theory Nov 1997; 43(6):1986–1992.
[2] Cover TM, Thomas JA. Elements of Information Theory. 2nd edn., John
Wiley & Sons, 1991.
[3] Proakis JG. Digital Communications. 4th edn., McGraw-Hill, 2001.
[4] Lin S, Costello DJ. Error Control Coding: Fundamentals and Applications. Prentice-Hall, 1983.
[5] Beh KC, Doufexi A, Armour SMD. Performance evaluation of hybrid
ARQ schemes of 3GPP LTE OFDMA system. Proc. PIMRC, 2007; 1–5.
[6] Wang G, Wu J, Zheng YR. Optimum energy- and spectral-efficient
transmissions for delay-constrained hybrid ARQ systems. IEEE Trans.
Vehic. Tech. Jul 2016; 65(7):5212–5221.
[7] Ji T, Stark W. Rate-adaptive transmission over correlated fading channels. IEEE Trans. Commun. Oct 2005; 53(10):1663–1670.
[8] Pai HT, Han YS, Chu YJ. New HARQ scheme based on decoding of tailbiting convolutional codes in IEEE 802.16e. IEEE Trans. Vehic. Tech.
Mar 2011; 60(3):912–918.
[9] Wengerter C, von Elbwart AGE, Seidel E, Velev G, Schmitt M.
Advanced hybrid ARQ technique employing a signal constellation
rearrangement. Proc. VTC, vol. 4, 2006; 2002–2006.
[10] Jung J, Park H, Lim J. A novel hybrid ARQ scheme based on shift
column permutation bit interleaving for OFDM systems. Proc. VTC,
2006; 1–5.
[11] Chen H, Maunder RG, Hanzo L. A survey and tutorial on lowcomplexity turbo coding techniques and a holistic hybrid ARQ design
example Nov 2013; 15(4):1546–1566.
[12] Lucani DE, Médard M, Stojanovic M. On coding for delay-network
coding for time-division duplexing. IEEE Trans. Inform. Theory Apr
2012; 58(4):2330–2348.
[13] Moreira A, Lucani DE. Coded schemes for asymmetric wireless interfaces: Theory and practice. IEEE J. Select. Areas Commun. Feb 2015;
33(2):171–184.
[14] Makki B, i Amat AG, Eriksson T. On noisy ARQ in block-fading
channels. IEEE Trans. Vehic. Tech. Feb 2014; 63(2):731–746.
[15] Jabi M, Hamss AE, Szczecinski L, Piantanida P. Multipacket hybrid
ARQ: Closing gap to the ergodic capacity. IEEE Trans. Commun. Dec
2015; 63(12):5191–5205.
[16] Ahmed A, Muhammad Z, Mahmood H, Saqib NA. Partial automatic
repeat request transceiver for bandwidth and power efficient multipleinput-multiple-output orthogonal frequency division multiplexing systems 2015; 9(4):476–486.
[17] Malkamäki E, Leib H. Performance of truncated type-II hybrid ARQ
schemes with noisy feedback over block fading channels. IEEE Trans.
Commun. Sep 2000; 48(9):1477–1487.
[18] Modiano E. An adaptive algorithm for optimizing the packet size used
in wireless arq protocols. Wireless Nets. 1999; 5:279–286.
[19] Ahmad RA, Lacan J, Arnal F, Gineste M, Clarac L. Enhancing satellite
system throughput using adaptive HARQ for delay tolerant services in
mobile communications. WTS, 2015; 1–7.
[20] Shih CH, Kuo CI, Shieh CK, Chou YK. Small-block interleaving for
low-delay cross-packet forward error correction over burst-loss channels.
Int. J. Commun. Syst. 2014; 27:3980–3995.
15
[21] Fujii S, Kaneko T, Hayashida Y. Effectiveness of copy-transmission in
Go-Back-N ARQ system with selective repeat in intra-block and with
limited retransmissions. Proc. COMSNETS, 2010; 187–192.
[22] Hassanien MAM, Loskot P. Improving link reliability complexity tradeoff by exploiting reliable feedback signal. Proc. ISWCS, 2010; 775–779.
[23] Biglieri E. Coding For Wireless Channels. Springer, 2005.
[24] Knuth D. The Art of Computer Programm. Addison-Wesley, 1998.
[25] Loskot P, Beaulieu NC. Prony and polynomial approximations for
evaluation of the average probability of error over slow-fading channels.
IEEE Trans. Vehic. Tech. Mar 2009; 58(3):119–138.
[26] Petrova M, Riihijarvi J, Mahonen P, Labella S. Performance study
of IEEE 802.15.4 using measurements and simulations. Proc. WCNC,
vol. 1, 2006; 487–492.
| 7cs.IT
|
arXiv:1607.02480v1 [cs.AI] 8 Jul 2016
Real-Time Anomaly Detection for Streaming Analytics
Subutai Ahmad
Numenta, Inc., 791 Middlefield Road, Redwood City, CA 94063 USA
SAHMAD @ NUMENTA . COM
Scott Purdy
Numenta, Inc., 791 Middlefield Road, Redwood City, CA 94063 USA
SPURDY @ NUMENTA . COM
Abstract
Much of the worlds data is streaming, time-series
data, where anomalies give significant information in critical situations. Yet detecting anomalies in streaming data is a difficult task, requiring detectors to process data in real-time, and
learn while simultaneously making predictions.
We present a novel anomaly detection technique
based on an on-line sequence memory algorithm
called Hierarchical Temporal Memory (HTM).
We show results from a live application that
detects anomalies in financial metrics in realtime. We also test the algorithm on NAB, a published benchmark for real-time anomaly detection, where our algorithm achieves best-in-class
results.
1. Introduction
Across every industry, we are seeing an exponential increase in the availability of streaming, time-series data.
Largely driven by the rise of the Internet of Things (IoT)
and connected real-time data sources, we now have an
enormous number of applications with sensors that produce
important, continuously changing data.
The detection of anomalies in real-time streaming data has
practical and significant applications across many industries. There are numerous use cases for anomaly detection, including preventative maintenance, fraud prevention,
fault detection, and monitoring. The use cases can be found
throughout numerous industries such as finance, IT, security, medical, energy, e-commerce, and social media.
We define an anomaly as a point in time where the behavior of the system is unusual and significantly different
from past behavior. Under this definition, an anomaly does
not necessarily imply a problem. A change might be for
a negative reason like the temperature sensor on an engine
going up, indicating a possible imminent failure. Or the
change might be for a positive reason like web clicks on
a new product page are abnormally high, showing strong
demand. Either way, the data is unusual and may require
action. Anomalies can be spatial, meaning the value is outside the typical range like the first and third anomalies in
Figure 1. They can also be temporal, where the value isn’t
outside the typical range but the sequence in which it occurs
is unusual. The middle anomaly in Figure 1 is a temporal
anomaly.
Real-time applications impose their own unique constraints
for machine learning. Anomaly detection in streaming applications is particularly challenging. The detector must
process data and output a decision in real-time, rather than
making many passes through batches of files. In most scenarios the number of sensor streams is large and there is
little opportunity for human, let alone expert, intervention.
As such, operating in an unsupervised, automated fashion
(e.g. without manual parameter tweaking) is often a necessity. The underlying system is often non-stationary, and
detectors must continuously learn and adapt to changing
statistics while simultaneously making predictions.
The goal of this paper is to introduce a novel anomaly
detection technique designed for such real-time applications. We show how to use Hierarchical Temporal Memory
(HTM) networks (Hawkins & Ahmad, 2016; Padilla et al.,
2013; Rozado et al., 2012) in a principled way to robustly
detect anomalies in a variety of conditions. The resulting
system is efficient, extremely tolerant to noisy data, continously adapts to changes in the statistics of the data, and
detects very subtle anomalies while minimizing false positives. We show qualitative examples from a real-time financial anomaly detection application. We also report leading
results on an open benchmark for real-time anomaly detection. The algorithm has been deployed commercially, and
we discuss some of the practical lessons learned from these
deployments. We have made the complete source code
(including end application code) available in open source
Real-Time Anomaly Detection
repositories 1 .
Figure 2. CPU utilization (percent) for an Amazon EC2 instance.
A change to the software running on the machine caused the CPU
usage to change. Continuous learning is essential for performing
anomaly detection on streaming data like this.
Figure 1. The figure shows real-world temperature sensor data
from an internal component of a large industrial machine.
Anomalies are labeled with red circles. The first anomaly was
a planned shutdown. The third anomaly is a catastrophic system
failure. The second anomaly, a subtle but observable change in
the behavior, indicated the actual onset of the problem that led to
the eventual system failure.
2. Related Work
Anomaly detection in time-series is a heavily studied
area, dating back to (Fox, 1972). Some techniques,
like classification-based methods, are supervised or semisupervised. While labelled data can be used to improve
results, supervised techniques are typically unsuitable for
anomaly detection (Görnitz et al., 2013). Figure 2 illustrates the need for continuous learning, which is not typically possible with supervised algorithms.
Other techniques, like simple thresholds, clustering, and
exponential smoothing, are only capable of detecting spatial anomalies. Holt-Winters is an example of the latter
that is commonly implemented for commercial applications (Szmit & Szmit, 2012). Also commonly used in practice are change point detection methods, which are capable
of identifying temporal anomalies. The typical approach is
to model the time series in two independent moving windows and detect when there is a significant deviation in the
time series metrics (Basseville & Nikiforov, 1993). These
methods are often extremely fast to compute and have low
memory overhead. The detection performance of these statistical techniques can be sensitive to the size of the windows and thresholds. This sometimes results in many false
positives as the data changes, requiring frequent updates to
the thresholds in order to detect anomalies while minimizing false positives.
1
Please see http://numenta.org
The Skyline project provides an open source implementation of a number of statistical techniques for detecting
anomalies in streaming data (Stanway, 2013). The different
algorithms can be combined into an ensemble. The Skyline
algorithms are included in our results.
There are other algorithms capable of detecting temporal
anomalies in complex scenarios. ARIMA is a general purpose technique for modeling temporal data with seasonality
(Bianco et al., 2001). It is effective at detecting anomalies
in data with regular daily or weekly patterns. It is not capable of dynamically determining the period of seasonality, although extensions have been developed for doing so
(Hyndman & Khandakar, 2008). A technique for applying
ARIMA to multivariate data has also been studied (Tsay,
2000). Bayesian change point detection methods are a natural approach to segmenting time series and can be used for
online anomaly detection (Adams & Mackay, 2007; Tartakovsky et al., 2013). Some additional techniques for general purpose anomaly detection on streaming data include
(Keogh et al., 2005; Rebbapragada et al., 2009).
Yahoo released the open source EGADS framework for
time series anomaly detection that pairs time series forecasting techniques with common anomaly detection algorithms (Laptev et al., 2015). Twitter released its own open
source anomaly detection algorithms for time series data
(Kejariwal, 2015). Both are capable of detecting spatial
and temporal anomalies. Empirical comparison with Twitter’s detection software are included in our results.
There have been a number of model-based approaches applied to specific domains. These tend to be extremely
specific to the domain they are modeling. Examples include anomaly detection in aircraft engine measurements
(Simon & Rinehart, 2015), cloud datacenter temperatures
(Lee et al., 2013), and ATM fraud detection (Klerx et al.,
2014). While these approaches may have success in a spe-
Real-Time Anomaly Detection
cific domain, they are not suitable for general purpose applications.
We have reviewed some of the algorithms most relevant to
our work. A comprehensive literature review is outside the
scope of this paper but there are several thorough reviews
of anomaly detection techniques for further reading (Chandola et al., 2009; Hodge & Austin, 2004; Chandola et al.,
2008).
In this paper we focus on using Hierarchical Temporal
Memory (HTM) for anomaly detection. HTM is a machine
learning algorithm derived from neuroscience that models
spatial and temporal patterns in streaming data (Hawkins
& Ahmad, 2016; Rozado et al., 2012). HTM compares favorably with some state of the art algorithms in sequence
prediction, particularly complex non-Markovian sequences
(Cui et al., 2015; Padilla et al., 2013). HTMs are continuously learning systems that automatically adapt to changing statistics, a property particularly relevant to streaming
analytics.
perform anomaly detection we utilize two different internal
representations available in the HTM. Given an input xt ,
the vector a(xt ) is a sparse binary code representing the
current input. We also utilize an internal state vector π(xt )
which represents a prediction for a(xt+1 ), i.e. a prediction
of the next input xt+1 . The prediction vector incorporates
inferred information about current sequences. In particular,
a given input will lead to different predictions depending on
the current detected sequence and the current inferred position of the input within the sequence. The quality of the
prediction is dependent on how well the HTM is modeling
the current data stream. See (Hawkins & Ahmad, 2016) for
a more detailed explanation of these representations.
Figure 3. The primary functional steps in our algorithm.
3. Anomaly Detection Using HTM
Typical streaming applications involve analyzing a continuous stream of data occurring in real-time. Such applications contain some unique challenges. We formalize this as
follows. Let the vector xt represent the state of a real-time
system at time t. The model receives a continuous stream
of inputs:
. . . , xt−2 , xt−1 , xt , xt+1 , xt+2 , . . .
(1)
a(xt ) and π(xt ) are recomputed at every iteration but do
not directly represent anomalies. In order to create a robust anomaly detection system we introduce two additional
steps. We first compute a raw anomaly score from the two
sparse vectors. We then compute an anomaly likelihood
value which is thresholded to determine whether the system is anomolous. Figure 3 shows a block diagram of our
algorithm. These two steps are detailed below. We then
describe how to robustly handle a larger system consisting
of multiple distinct models.
Consider for example, the task of monitoring a datacenter. Components of xt might include CPU usage for various servers, bandwidth measurements, latencies of servicing requests, etc. At each point in time t we would like
to determine whether the behavior of the system up to that
point is unusual. One of the key challenges is that the determination must be made in real-time, i.e. before time t + 1
and without any look ahead. In practical applications, the
statistics of the system can change dynamically. For example, in a production datacenter, software upgrades might
be installed at any time that alter the behavior of the system (Figure 2). Any retraining of a model must be done
on-line, again before time t + 1. Finally, the individual
measurements are not independent and contain significant
temporal patterns that can be exploited.
The raw anomaly score will be 0 if the current input is perfectly predicted, 1 if it is completely unpredicted, or somewhere in between depending on the similarity between the
input and the prediction.
HTM is a learning algorithm that appears to match the
above constraints. HTM networks are continuously learning and model the spatiotemporal characteristics of their
inputs. HTMs have been shown to work well for prediction
tasks (Cui et al., 2015; Padilla et al., 2013) but HTM networks do not directly output an anomaly score. In order to
An interesting aspect of this score is that branching sequences are handled correctly. In HTMs, multiple predictions are represented in π(xt ) as a binary union of each individual prediction. Similar to Bloom filters, as long as the
vectors are sufficiently sparse and of sufficient dimensionality, a moderate number of predictions can be represented
3.1. Computing the Raw Anomaly Score
We compute a raw anomaly score that measures the deviation between the model’s predicted input and the actual
input. It is computed from the intersection between the predicted and actual sparse vectors. At time t the raw anomaly
score, st , is given as:
st = 1 −
π(xt−1 ) · a(xt )
|a(xt )|
(2)
Real-Time Anomaly Detection
simultaneously with exponentially small chance of error
(Bloom, 1970; Ahmad & Hawkins, 2016). The anomaly
score handles branching sequences gracefully in the following sense. If two completely different inputs are both
possible and predicted, receiving either input will lead to a
0 anomaly score. Any other input will generate a positive
anomaly score.
Changes to the underlying system are also handled gracefully due to the continuous learning nature of HTMs. If
there is a shift in the behavior of the system, the anomaly
score will be high at the point of the shift, but will automatically degrade to zero as the model adapts to the “new normal”. Shifts in the temporal characteristics of the system
are handled in addition to spatial shifts in the underlying
metric values. (See Results section for some examples.)
step. Rather than thresholding the raw score directly, we
model the distribution of anomaly scores and use this distribution to check for the likelihood that the current state
is anomalous. The anomaly likelihood is thus a metric
defining how anomalous the current state is based on the
prediction history of the HTM model. To compute the
anomaly likelihood we maintain a window of the last W
raw anomaly scores. We model the distribution as a rolling
normal distribution where the sample mean and variance
are continuously updated from previous anomaly scores as
follows:
Pi=W −1
i=0
µt =
σt2 =
st−i
(3)
k
Pi=W −1
(st−i − µt )2
k−1
i=0
(4)
We then compute a recent short term average of anomaly
scores, and apply a threshold to the Gaussian tail probability (Q-function, (Karagiannidis & Lioumpas, 2007)) to
decide whether or not to declare an anomaly2 . We define
the anomaly likelihood as the complement of the tail probability:
Lt = 1 − Q
µ̃t − µt
σt
(5)
where:
Figure 4. A very noisy, unpredictable stream. The data shows the
latency (in seconds) of a load balancer on a production website.
The red dot shows the approximate location of an unusual increase
in latencies.
3.2. Computing Anomaly Likelihood
The raw anomaly score described above represents an instantaneous measure of the predictability of the current input stream. This works well for predictable scenarios but in
many practical applications, the underlying system is inherently noisy and unpredictable. In these situations it is often
the change in predictability that is indicative of anamolous
behavior. As an example, consider Figure 4. This data
shows the latency of a load balancer in serving HTTP requests on a production web site. Although the latency is
generally low, it is not unusual to have occasional random jumps, and the corresponding spike in anomaly score.
Thresholding the raw anomaly score directly would lead to
many false positives. However, a sustained increase in the
frequency of high latency requests, as shown in the second half of the figure, is unusual and thus reported as an
anomaly.
To handle this class of scenarios, we introduce a second
Pi=W 0 −1
µ̃t =
i=0
st−i
j
(6)
W 0 here is a window for a short term moving average,
where W 0 W . We threshold Lt and report an anomaly
if it is very close to 1:
anomaly detected ≡ Lt ≥ 1 −
(7)
It is important to note that this test is applied to the distribution of anomaly scores, not to the distribution of underlying metric values xt . As such, it is a measure of how
well the model is able to predict, relative to the recent history. In clean predictable scenarios Lt behaves similarly to
st . In these cases the distribution of scores will have very
small variance and will be centered near 0. Any spike in
st will similarly lead to a corresponding spike in Lt . However in scenarios with some inherent randomness or noise,
the variance will be wider and the mean further from 0. A
2
The Guassian tail probability is the probability that a Gaussian variable will obtain a value larger than x standard deviations
above the mean.
Real-Time Anomaly Detection
single spike in st will not lead to a significant increase in
Lt but a series of spikes will. Interestingly, a scenario that
goes from wildly random to completely predictable will
also trigger an anomaly.
Since thresholding Lt involves thresholding a tail probability, there is an inherent upper limit on the number of alerts.
With very close to 0 it would be unlikely to get alerts with
probability much higher than . This also imposes an upper
bound on the number of false positives. Under the assumption that anomalies themselves are also extremely rare, the
hope is that the ratio of true positives to false positives is
always in a healthy range (see Results below).
Although we use HTM as the underlying temporal model,
the likelihood technique is not specific to HTMs. It could
be used with any other algorithm that outputs a sparse code
or scalar anomaly score. The overall quality of the detector
will be dependent on the ability of the underlying model to
represent the domain.
3.3. Combining Multiple Independent Models
Many industrial or complex environments contain a large
number of sensory data streams. In theory, given sufficient resources, it is possible to create one large complex
model with the entire vector stream as input. In practice,
it is common to decompose a large system into a number
of smaller models. It is easier to train smaller models because the complexity of training and inference grows much
faster than linearly with the size of the input dimensionality (Bishop, 2006). Thus decomposing the solution into a
set of smaller models may lead to improved accuracy and
much faster performance. However, even with such a decomposition, it is important to compute a global measure
that accumulates the results of individual models and indicates whether those portions of the system are in an unusual
state. As an example, consider a datacenter running a production website. An automated alerting system might need
to continually decide whether to generate an alarm and possibly wake up the on-call engineer.
We assume the inputs representing the system are broken
up into M distinct models. Let xm
t be the input at time
t to the m’th model, and sm
t be the raw anomaly scores
associated with each model. We wish to compute a global
metric indicating the overall likelihood of an anomaly in
the system (see Figure 5).
One possible approach is to estimate the joint distribution
−1
P (s0t , . . . , sM
) and apply a threshold to the tail probat
bility. It can be challenging to model the joint distribution,
particularly in a streaming context. If we further assume
the models are independent, we could simplify and instead
estimate:
Figure 5. Functional diagram illustrating a complex system with
multiple independent models.
−1
P (s0t , . . . , sM
)=
t
i=M
Y−1
P (sit )
(8)
i=0
Given this, a version of our anomaly likelihood can be computed as:
1−
i=M
Y−1
i=0
Q
µ̃it − µit
σti
(9)
There is one flaw with the above methodology. In real-time
dynamic scenarios, critical problems in one part of the system can often cascade to other areas. Thus there are often
random temporal delays built in, which can in turn lead to
different temporal delays between anomaly scores in the
various models (Kim et al., 2013). For example, a situation
where multiple unusual events occur close to one another
in different models is far more unlikely and unusual than a
single event in a single model. It is precisely these situations which are valuable to detect and capture in a complex
system.
Ideally we would be able to estimate a joint distribution of anomaly scores that go back in time, i.e.
−2 M −1
P (s0t−j , s1t−j , . . . , sM
, st
). In theory this would capt
ture all the dependencies, but this is even harder to estimate
than the earlier joint probability. Alternatively, in situations
where the system’s topology is relatively clear and under
your control, it is possible to create an explicit graph of
dependencies, monitor expected behavior between pairs of
nodes, and detect anomalies with respect to those expectations. This technique has been shown to enable very precise
determination of anomalies in websites where specific calls
between services are monitored (Kim et al., 2013). However in most applications this technique is also impractical.
It may be difficult to model the various dependencies and it
is often infeasible to instrument arbitrary systems to create
this graph.
Real-Time Anomaly Detection
We would like to have a system that is fast to compute,
makes relatively few assumptions, and is adaptive. We
propose a simple general purpose mechanism for handling
multiple models by modifying Eq. (9) to incorporate a
smooth temporal window. A windowing mechanism allows
the system to incorporate spikes in likelihood that are close
in time but not exactly coincident. Let G be a Gaussian
convolution kernel:
G(x; σ) = √
x2
1
exp− 2σ2
2πσ
(10)
We apply this convolution to each individual model to obtain a final anomaly likelihood score3 :
Lt = 1 −
i=M
Y−1
2(G ∗ Q)(
i=0
µ̃it − µit
)
σti
system is not extremely sensitive to this setting. In all our
experiments below, σ = 6.
Computational efficiency is important for real-time applications. In our implementation each model requires less
than 10 milliseconds per input vector on a current high end
laptop. A server based deployment running multiple models in parallel can run about 5,000 models on a high end
server. This figure assumes that each input arrives once every 5 minutes (thus each of the 5,000 models needs to output one score every 5 minutes). As a service to the community, we have released the complete source code to the algorithms as well as the server application codebase as open
source5 .
(11)
As before, we detect an anomaly if the combined anomaly
likelihood is greater than a threshold Lt ≥ 1−. Eq. 11 represents a principled yet pragmatic approach for detecting
anomalies in complex real time streaming applications4 .
As before, Lt is an indirect measure, computed on top of
the raw anomaly scores from each model. It reflects the underlying predictability of the models at a particular point in
time and does not directly model the sensor measurements
themselves.
Figure 6. Real-time stock anomalies for Feb 4, 2016.
4. Results
3.4. Practical Considerations
There are three parameters in the single model scenario:
W , W 0 , and . W is the duration for computing the distribution of anomaly scores. The system performance is not
sensitive to W as long as it is large enough to compute a reliable distribution. The number W 0 controls the short term
average of anomaly scores. In all our experiments below,
we use a generous value of W = 8000 and W 0 = 10.
The parameter is perhaps the most important parameter. It
controls how often anomalies are reported, and the balance
between false positives and false negatives. In practice we
have found that = 10−5 works well across a large range
of domains. Intuitively, this should represent one false positive about once every 10,000 records.
The muliple model scenario introduces one additional parameter, the window width σ. This is somewhat domain
dependent but due to the soft Gaussian convolution, the
3
Since this is a real-time system with no look-ahead, the kernel
only applies to time ≤ t. As such we use a half-normal distribution as the kernel, hence the additional factor of 2.
4
Due to numerical precision issues with products of probabilities, in our implementation we follow common practice and use
summation of log probabilities.
We first show examples from a deployed application that
qualitatively demonstrates the behavior of our algorithm.
We then show quantitative results on benchmark data.
We have integrated our anomaly detection algorithm into a
live real-time product. The application continuously monitors a large set of financial and social media metrics for a
range of securities and alerts users in real time when significant anomalies occur. Figure 6 shows two screenshots
from our application that demonstrate the value of realtime anomaly detection to end users. In this example, an
anomaly in the volume of Twitter activity (related to a decrease in dividends) preceded a sharp drop in stock price.
The Twitter anomaly occurred well before market opening.
The underlying data streams are extremely noisy and many
of the important anomalies are temporal in nature. Figure 7
shows raw data extracted from our application demonstrating one such anomaly. The figure plots stock trading volume for Facebook for a few hours. Each point represents
average trading volume for a period of five minutes. It is
normal to see a single spike in trading volume but it is extremely unusual to see two consecutive spikes. Two con5
Please see https://github.com/numenta, specifically the nupic and numenta-apps repositories.
Real-Time Anomaly Detection
secutive spikes therefore represents a temporal anomaly in
this stream. The red dashed line indicates the point at which
our algorithm detected the anomaly, i.e. the point in time
where Lt ≥ 1 − 10−5 .
Figure 8. Q-values derived from two Comcast metrics on Feb 4,
2016. The red dashed line shows the point at which our algorithm
detected an anomaly.
Figure 7. Stock trading volume for Facebook on Feb 2, 2016. The
red dashed line shows the point at which our algorithm detected an
anomaly. In this stream two consecutive spikes are highly unusual
and represent an anomaly.
Figure 8 shows a more detailed example demonstrating the
value of combining multiple metrics. The two solid lines
indicate the Q-values in Eq (5) for two separate metrics
related to Comcast shares. The lower the curve, the more
likely it is that the underlying metric is anomalous. When
looking at single models independently, an anomaly would
only be detected if one of these values drops below =
10−5 . In this case no anomaly would be detected since
neither one dips below that value. However, as shown by
the drops, both metrics are behaving unusually. The red
dashed line shows the result of detecting an anomaly based
on our combined metric, Eq. (11). Because the blue curve
is so close to 10−5 , incorporating both metrics correctly
flags an anomaly as soon as the black curve dips below
10−1 . As can be appreciated from the chart, it is difficult to
line up two metrics precisely, hence the value of the smooth
temporal window.
4.1. Results on Real-World Benchmark Data
The above sections demonstrate qualitative results on individual data streams. In this section we provide quantitative benchmark results. NAB is a benchmark containing 58
streams with over 350,000 records of real-time streaming
data taken from a variety of different applications (Lavin
& Ahmad, 2015). The dataset is labeled with anomalies. The first 15% of each data file is reserved for autocalibration. NAB also includes a window around each
anomaly, and incorporates a time-sensitive scoring mech-
A LGORITHM
P ERFECT
HTM
T WITTER ADV EC
E TSY S KYLINE
BAYES . C HANGE P T.
S LIDING T HRESHOLD
R ANDOM
NAB S CORE
L OW FP
L OW FN
100.0
65.3
47.1
35.7
17.7
15.0
11.0
100.0
58.6
33.6
27.1
3.2
0.0
1.2
100.0
69.4
53.5
44.5
32.2
30.1
19.5
Table 1. Performance on the NAB benchmark. The first column
indicates the standard NAB score for each algorithm. The last two
columns indicate the score for the “prefer low FP” and “prefer
low FN” profiles of NAB, representing two specific points on the
ROC curve. For comparison we include a “Perfect” detector, an
idealized detector that makes no mistakes.
anism that favors early detection (as long as detections are
within the anomaly windows). ”Application profiles” define the weighting for false positives and false negatives to
illustrate scenarios where fewer missed detections or fewer
erroneous detections are more valued. The benchmark requires models to use a single set of parameters across all
streams to mimic automation in real-world deployments.
We ran our HTM based anomaly detector on NAB. Table 1
contains our scores as well as scores for several other algorithms, including Etsy Skyline, Twitter AnomalyDetectionVec (ADVec), and a variant of Bayesian online change
point detection (Adams & Mackay, 2007)6 . We include the
score for a “Perfect” detector, i.e. an idealized detector that
6
Please see the NAB repository at https://github.
com/numenta/NAB for specific implementation and parameter
tuning details.
Real-Time Anomaly Detection
Figure 9. Example NAB results for three different data streams. The shapes correspond to different detectors: HTM, Skyline, and ADVec
are diamond, square, and plus respectively. True positives are labeled in black, and false positives are colored red. The pink shaded
regions denote NAB’s anomaly windows; any detection within a window is considered a true positive.
detects every anomaly as early as possible and generates
no false positives. We also include the scores for “sliding
threshold” and “random” detectors as baselines for comparison. The HTM detector achieves the best overall score,
followed by Twitter ADVec and Etsy Skyline. However, as
can be seen by the performance of the Perfect detector, the
benchmark is a challenging one with significant room for
future improvements.
data shown in Figure 1. The anomaly on the left is a somewhat subtle temporal anomaly where the temporal behavior
is unusual but individual readings are within the expected
range. This anomaly (which preceded the catastrophic failure on February 8) is only detected by HTM. All three detectors detect the anomaly on the right, although Skyline
and HTM detect it earlier than ADVec. In this plot HTM
and Skyline also each have a false positive.
The NAB benchmark doesn’t measure an explicit ROC
curve but includes two additional “application profiles” that
tradeoff the weighting between false positives and false
negatives. The “reward low FP” profile applies a higher
cost to false positives. Conversely the “reward low FN”
profile applies a higher cost to false negatives. The last
two columns of Table 1 show our scores for those two profiles. Although there are individual variations (Twitter in
particular improves significantly in the “reward low FN”
column), our detector performs best overall indicating that
it achieves a good tradeoff between false positives and false
negatives.
We chose these examples because they illustrate common
situations that occur in practice. Qualitatively we have
found that temporal changes in behavior often precede a
larger, easily detectable, shift. Temporal and sequence
based anomaly detection techniques may be able to detect
anomalies in streaming data before they are easily visible.
We speculate that the early detection in Figure 9(b) is due
to the temporal modeling in HTMs as the earlier shift is
difficult to detect through purely spatial means. This provides hope that such algorithms can be used in production
to provide early warning and perhaps help avoid problems
far more reliably than spatial techniques.
A closer look at the errors illustrates some interesting situations that arise in real time applications. Figure 9(a)
demonstrates the value of continuous learning. This file
shows CPU usage on a production server over time and
contains two anomalies. The first is a simple spike detected
by all algorithms. The second is a sustained shift in the usage. Skyline and HTM both detect the change but then
adapt to the new normal (with Skyline adapting quickest).
Twitter ADVec however continues to generate anomalies
for several days. Skyline scored highest on this stream.
Low scores for Bayesian change point detection are due
to its strong assumptions regarding a Gaussian or Gamma
distribution of data points. Real-world streaming data is
messy and these assumptions do not hold in many applications. In data where it does hold (e.g. Figs 2, 9a) the
algorithm works well but for most streams (e.g. Figs 1, 4)
it does not. Indeed, it is this lack of any consistent distribution across streaming applications that led us to model the
distribution of anomaly scores rather than the distribution
of metric values.
Figure 9(b) and (c) demonstrate temporal anomalies and
their importance in early detection. Figure 9(b) is an example demonstrating early detection. All three detectors
detect the anomaly but HTM detects it three hours earlier
due to a subtle shift in metric dynamics. Figure 9(c) shows
a close up of the results for the machine temperature sensor
From the standpoint of computational efficiency, on the
first author’s laptop, it takes 48 minutes to run NAB’s complete dataset of 365,558 records. This represents an average
of 8 milliseconds per record.
Real-Time Anomaly Detection
5. Discussion
With the increase in connected real-time sensors, the detection of anomalies in streaming data is becoming increasingly important. The use cases cut across a large number of
industries; anomaly detection might represent the most significant near-term application for machine learning in IoT.
In this paper we have discussed a novel anomaly detection
algorithm for real-time streaming applications. Based on
HTMs, the algorithm is capable of detecting spatial and
temporal anomalies in predictable and noisy domains. A
probabilistic formulation allows the user to control the rate
of false positives, an important consideration in many applications. We have discussed an extension to large systems
with multiple independent models that incorporate temporal windows.
Our results show that the algorithm can achieve best in
class results on a benchmark of real world data sources.
Our system is practical in that it is computationally efficient, automatically adapts to changing statistics, and requires little to no parameter tuning. It is currently in use in
commercial applications, and the complete source code to
the algorithm is available as open source software7 .
There are a number of possible extensions to the algorithm. The error analysis from NAB indicates that the errors from our algorithm are not necessarily correlated with
errors from the other two algorithms. As such an ensemble based approach might provide a significant increase in
accuracy. The assumption of a Gaussian distribution for
anomaly scores is not always correct. Exploring other distributions represents another possible extension that could
potentially improve the results.
Acknowledgements
We thank Alex Lavin, Yuwei Cui, and Jeff Hawkins for
many helpful comments and suggestions.
References
Adams, Ryan Prescott and Mackay, David J. C. Bayesian
Online Changepoint Detection. arXiv.org, pp. 7, 2007.
doi: arXiv:0710.3742v1. URL http://arxiv.org/
abs/0710.3742.
Ahmad, Subutai and Hawkins, Jeff. How do neurons operate on sparse distributed representations? A mathematical theory of sparsity, neurons and active dendrites. pp.
arXiv:1601.00720 [q–bio.NC], jan 2016. URL http:
//arxiv.org/abs/1601.00720.
Basseville, M and Nikiforov, I V. Detection of Abrupt
7
Please see http://numenta.com/nab
Changes, volume 2. 1993.
Bianco, A. M., Garcı́a Ben, M., Martı́nez, E. J., and Yohai,
V. J. Outlier detection in regression models with ARIMA
errors using robust estimates. Journal of Forecasting, 20
(8):565–579, 2001.
Bishop, Christopher M. Pattern Recognition and Machine Learning, volume 4. Springer, 2006. ISBN
9780387310732.
Bloom, Burton H. Space/time trade-offs in hash coding
with allowable errors, 1970. ISSN 00010782.
Chandola, V., Mithal, V., and Kumar, V. Comparative Evaluation of Anomaly Detection Techniques for Sequence
Data. 2008 Eighth IEEE International Conference on
Data Mining, pp. 743–748, 2008. ISSN 1550-4786. doi:
10.1109/ICDM.2008.151.
Chandola, Varun, Banerjee, A, and Kumar, V. Anomaly
detection: A survey. ACM Computing Surveys (CSUR),
(September):1–72, 2009.
Cui, Yuwei, Surpur, Chetan, Ahmad, Subutai, and
Hawkins, Jeff. Continuous online sequence learning with an unsupervised neural network model. pp.
arXiv:1512.05463 [cs.NE], 2015.
URL http://
arxiv.org/abs/1512.05463.
Fox, A J. Outliers in time series. Journal of the Royal Statistical Society, Series B (Methodological), 34(3):350–
363, 1972. ISSN 00359246.
Görnitz, Nico, Kloft, Marius, Rieck, Konrad, and Brefeld,
Ulf. Toward supervised anomaly detection. Journal
of Artificial Intelligence Research, 46:235–262, 2013.
ISSN 10769757. doi: 10.1613/jair.3623.
Hawkins, Jeff and Ahmad, Subutai.
Why Neurons
Have Thousands of Synapses, a Theory of Sequence
Memory in Neocortex.
Frontiers in Neural Circuits, 10(23):1–13, mar 2016.
ISSN 1662-5110.
doi:
10.3389/fncir.2016.00023.
URL http:
//journal.frontiersin.org/article/
10.3389/fncir.2016.00023/abstract.
Hodge, Victoria J. and Austin, Jim. A survey of outlier
detection methodologies, 2004. ISSN 02692821.
Hyndman, Rob J and Khandakar, Yeasmin. Automatic time
series forecasting : the forecast package for R Automatic
time series forecasting : the forecast package for R. Journal Of Statistical Software, 27(3):1–22, 2008.
Karagiannidis, George K. and Lioumpas, Athanaszsios S. An improved approximation for the Gaussian
Q-function. IEEE Communications Letters, 11(8):644–
646, 2007.
Real-Time Anomaly Detection
Kejariwal, Arun. Twitter Engineering: Introducing practical and robust anomaly detection in a time series [Online
blog], 2015. URL http://bit.ly/1xBbX0Z.
Keogh, Eamonn, Lin, Jessica, and Fu, Ada. HOT SAX:
Efficiently finding the most unusual time series subsequence. In Proceedings - IEEE International Conference on Data Mining, ICDM, pp. 226–233, 2005. ISBN
0769522785. doi: 10.1109/ICDM.2005.79.
Kim, Myunghwan, Sumbaly, Roshan, and Shah, Sam. Root
cause detection in a service-oriented architecture. ACM
International Conference on Measurement and Modeling of Computer Systems (SIGMETRICS ’13), pp. 93,
2013. doi: 10.1145/2465529.2465753.
Klerx, Timo, Anderka, Maik, Buning, Hans Kleine, and
Priesterjahn, Steffen. Model-Based Anomaly Detection
for Discrete Event Systems. In 2014 IEEE 26th International Conference on Tools with Artificial Intelligence,
pp. 665–672. IEEE, nov 2014. ISBN 978-1-4799-65724. doi: 10.1109/ICTAI.2014.105.
Laptev, Nikolay, Amizadeh, Saeed, and Flint, Ian. Generic
and Scalable Framework for Automated Time-series
Anomaly Detection. In Proceedings of the 21th ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 1939–1947. ACM, 2015.
Lavin, Alexander and Ahmad, Subutai. Evaluating Realtime Anomaly Detection Algorithms the Numenta
Anomaly Benchmark. In 14th International Conference on Machine Learning and Applications (IEEE
ICMLA’15), Miami, Florida, 2015. IEEE. doi: 10.1109/
ICMLA.2015.141.
Lee, Eun Kyung, Viswanathan, Hariharasudhan, and Pompili, Dario. Model-based thermal anomaly detection in
cloud datacenters. Proceedings - IEEE International
Conference on Distributed Computing in Sensor Systems, DCoSS 2013, pp. 191–198, 2013. doi: 10.1109/
DCOSS.2013.8.
Padilla, Daniel E., Brinkworth, Russell, and McDonnell, Mark D. Performance of a hierarchical temporal
memory network in noisy sequence learning. In 2013
IEEE International Conference on Computational Intelligence and Cybernetics (CYBERNETICSCOM), pp. 45–
51. IEEE, dec 2013. ISBN 978-1-4673-6053-1. doi:
10.1109/CyberneticsCom.2013.6865779.
Rebbapragada, Umaa, Protopapas, Pavlos, Brodley,
Carla E., and Alcock, Charles. Finding anomalous periodic time series : An application to catalogs of periodic
variable stars. Machine Learning, 74(3):281–313, 2009.
ISSN 08856125. doi: 10.1007/s10994-008-5093-3.
Rozado, David, Rodriguez, Francisco B., and Varona,
Pablo. Extending the bioinspired hierarchical temporal
memory paradigm for sign language recognition. Neurocomputing, 79:75–86, mar 2012. ISSN 09252312. doi:
10.1016/j.neucom.2011.10.005.
Simon, Donald L and Rinehart, Aidan W. A Model-Based
Anomaly Detection Approach for Analyzing Streaming
Aircraft Engine Measurement Data. Technical Report
2015-218454, NASA, 2015.
Stanway, A. Etsy Skyline, 2013.
github.com/etsy/skyline.
URL https://
Szmit, MacIej and Szmit, Anna. Usage of modified holtwinters method in the anomaly detection of network traffic: Case studies. Journal of Computer Networks and
Communications, 2012, 2012. ISSN 20907141. doi:
10.1155/2012/192913.
Tartakovsky, Alexander G., Polunchenko, Aleksey S.,
and Sokolov, Grigory. Efficient Computer Network
Anomaly Detection by Changepoint Detection Methods.
IEEE Journal of Selected Topics in Signal
Processing, 7(1):4–11, feb 2013.
ISSN 19324553.
doi: 10.1109/JSTSP.2012.2233713.
URL
http://ieeexplore.ieee.org/lpdocs/
epic03/wrapper.htm?arnumber=6380529.
Tsay, R. S. Outliers in multivariate time series. Biometrika,
87(4):789–804, dec 2000. ISSN 0006-3444. doi: 10.
1093/biomet/87.4.789.
| 3cs.SY
|
SCK: A SPARSE CODING BASED KEY-POINT DETECTOR
Thanh Hong-Phuoc, Yifeng He, Ling Guan
Ryerson Multimedia Research Laboratory, Ryerson University, Toronto, Canada M5B 2K3
ABSTRACT
All current popular hand-crafted key-point detectors such as
Harris corner, MSER, SIFT, SURF… rely on some specific
pre-designed structures for the detection of corners, blobs, or
junctions in an image. In this paper, a novel sparse coding
based key-point detector which requires no particular predesigned structures is presented. The key-point detector is
based on measuring the complexity level of each block in an
image to decide where a key-point should be. The complexity
level of a block is defined as the total number of non-zero
components of a sparse representation of that block.
Generally, a block constructed with more components is
more complex and has greater potential to be a good keypoint. Experimental results on Webcam and EF datasets [1,
2] show that the proposed detector achieves significantly high
repeatability compared to hand-crafted features, and even
outperforms the matching scores of the state-of-the-art
learning based detector.
Index Terms— Key-point, interest point, feature
detector, sparse coding, sparse representation
1. INTRODUCTION
Detection of key-points has been researched for decades, and
a large number of successful key-point detectors have been
proposed, which provide a foundation for a vast number of
applications in computer vision such as image matching,
image registration and object tracking. Currently, there are
two main categories of key-point detectors: hand-crafted
detectors and learning based detectors.
Hand-crafted detectors have attracted much attention since
the introduction of corner detection works by Forstner et al.,
and Harris et al. in the 1980s [3, 4]. Several other outstanding
hand-crafted detectors are SIFT [5], SURF [6], MSER [7], and
SFOP [8]. Although hand-crafted detectors generally produce
good results, they often rely on specific pre-designed
structures. In fact, hand-crafted features are known not to be
very flexible in different contexts [1, 9] probably due to the
nature of pre-designed structures. In practice, there are
situations in which a large number of high quality key-points
are expected, but cannot be easily located by one specific
detector. A common solution is a careful re-selection of an
alternative detector or combination of multiple. However, in
such cases, if multiple structure types could be detected with
a single detector, it would be a simple but very effective
solution.
Learning based detectors have received increased focus with
the recent success of deep learning. The typical examples of
learning based detectors are TIDLE [1] and its following
work [9]. These works demonstrate significant improvements
in the repeatability of the detected key-points in many
challenging datasets. However, to achieve this high
performance, the state-of-the-art learning based feature
detector [9] relies on the output of TIDLE which was trained
on a dataset originating from the results of SIFT detectors.
Thus, the development of hand-crafted detectors or other type
of detectors for creating reliable inputs for the machine
learning algorithms is still an important task.
In this paper, we introduce a key-point detector which
requires no particular pre-designed structures. The proposed
detector is shown to be able to detect all types of structures.
The detector outperforms the existing hand-crafted feature
detectors in terms of the repeatability metric developed by
Mikolajczyk et al. [10] on Webcam and EF datasets [1, 2].
Experimental results show that the proposed detector even
works better than the current top learning based detector on
the matching score metric [10]. The proposed key-point
detector is based on counting the total number of non-zero
components in the sparse representation of each block in an
image. We define the number of non-zero components in a
sparse representation as the complexity measure. A formal
mathematical definition of the measure is presented in
Section 3.2. In general, the more components a block requires
to construct, the more complex the block is. Any blocks with
the complexity measures falling in a specified range can be
considered as good key-points. In the event, the number of
expected key-points is limited, suitable key-points can be
selected based on a proposed novel strength measure. Our
contributions are summarized below:
1.
2.
We propose a novel sparse coding based key-point
detector requiring no particular hand-crafted structure.
We introduce a measure to compare the strength of any
key-points detected by the proposed detector.
The rest of the paper is organized as follows. In Section 2, we
review some popular hand crafted features, and two top
learning based feature detectors. In Section 3, the proposed
key-point detector is presented. The evaluation results of the
proposed detector are reported in Section 4, and the
conclusion is presented in Section 5.
2. RELATED WORK
2.1. Hand-crafted feature detector
The development of hand-crafted detectors often begins with
a human designed structure such as a corner, blob, or junction
with some specially predefined characteristics such that the
structures are stable and easily found in different images
under various transformations. Some typical examples of the
detectors of this type are Harris corners [4] for corner
detection, SIFT [5], SURF [6], MSER [7] for blob detection,
and SFOP [8] for junction detection. Besides the list of the
aforementioned detectors, there are a vast number of
detectors such as SIFER [11], D-SIFER [12], WADE [13],
Edge Foci [2] targeting detection of different structures with
various customizations. Although current detectors rely on
some more or less different pre-designed structures, the
structures share a common factor in that they have some
levels of complexity. If we can measure the complexity of the
structures of all blocks in an image, we do not need a predefined structure for a key-point detector, which is the
motivation for our work.
2.2. Learning based feature detector
Recently, a new trend has emerged regarding the enhanced
repeatability of key-point based learning methods. Two of the
most prominent works are TIDLE [1] and the work in [9]
which is developed based on TIDLE’s output. In TIDLE, the
key-points in overlapped images captured at different times
and seasons with the same viewpoints are first detected by
SIFT detectors. Then only blocks around key-points with
sufficiently high repeatability are selected as positive samples
to train their systems. In [9], the outputs of TIDLE are used
as standard patches to train their systems. Both works deliver
outstanding improvements in repeatability compared to
purely hand-crafted features. However, the works more or
less rely on the selection of hand-crafted features to build the
training datasets. In other words, a good method to detect
reliable key-points as inputs to the training systems is still
necessary. The learning based feature detectors, on the other
hand, are good ways to enhance the performance in terms of
higher repeatability or faster computation.
3. SPARSE CODING BASED KEY-POINT
DETECTOR
Sparse coding has been widely applied in many different
disciplines such as machine learning, neuroscience, and
signal processing [14, 15]. The idea of sparse coding is to
represent an input signal as the sparsest combination of atoms
in a given dictionary. In this section, the proposed sparse
coding based algorithm for the detection of key-points in a
gray-scale image is presented as follows:
3.1. Pre-filtering
A small size low pass filter such as a Gaussian filter is
applied to remove noise and small structures in the image.
3.2. Sparse coding based key-point detection
A loop through every block 𝑿 of size 𝑛 by 𝑛 in the image is
created. For each block, after reshaping 𝑿 to the size 𝑛2 by 1,
the mean is subtracted from 𝑿 . Then the difference is
normalized to unit amplitude which is noted as 𝑿′ . Next, the
sparse representation of the block is found by solving the
following optimization problem:
𝜶 = argmin(‖𝑿′ − 𝑫𝜶‖ + 𝜆‖𝜶‖𝑙1 )
(1)
where 𝜶 is the sparse representation of the block, 𝑫 is a
complete or over-complete dictionary of the sparse coding.
The size of the dictionary is 𝑛2 by 𝑘 where k is the number of
atoms in the dictionary. The size of 𝜶 is k by 1.
The complexity measure (𝑪𝑴) is then defined as follows:
𝑪𝑴 = ‖𝜶‖𝑙0
(2)
where ‖𝜶‖𝑙0 is the total number of non-zero components in
the sparse representation.
If the 𝑪𝑴 of a block falls into a range specified between a
lower and upper limit, the point at the center of the block is
considered as a key-point. The lower limit is used to
guarantee that the detected key-point has a high enough
complexity level to be distinct from other points in the image.
The upper limit is used to avoid the detection of overly
distinctive key-points, which may not be repeatable in
different images of the same scenes. The size of the key-point
is then selected as the square root of 2 multiplied by half of
the block size to cover all pixels (all possible positions of the
key-point) belonging to the block. For example, if a center of
a block of size 9x9 is considered as a key-point, the size of
the key-point is 4.5√2 . The key-point and size selection
techniques are application configurable parameters, which
provide flexibility. Depending on application, the entire
block may be considered as a key-point, or the size of the keypoint may be defined in difference ways.
Since biological evidence has revealed that sparse
representations are similar to the receptive fields of simple
cells in the primary visual cortex [16], sparse coding has been
extensively applied in many saliency detectors such as [16,
17], which look similar to the presented detection method at
the first glance. However, the saliency detection works are
different from ours, since we only consider the structure
based key-points, while the saliency detectors focus on visual
contents that attract human attention.
3.3. Strength measure (𝑺𝑴) of a key-point
Some applications need a limited number of key-points such
as top hundred or top thousand points. In such cases, a metric
to measure the strength of the detected key-points is
necessary, since many key-points given in Section 3.2 have
the same level of complexity, this measure cannot be used to
sort the points. Therefore, we propose a strength measure for
the detector, which is conceptually similar to the corner
response in many corner detectors. After thorough
experiments, the following metric currently generates the best
results in terms of repeatability and matching score:
𝑺𝑴 = 𝑎1 ‖𝜶‖𝑙0 + 𝑎2 ‖𝜶‖𝑙1
(3)
where 𝑎1 , 𝑎2 are any positive parameters, which are selected
as 1 in our experiments. The proposed metric satisfies the
three following conditions to measure the strength of a keypoint given by our detector:
1.
2.
3.
A key-point is stronger if more non-zero components are
required for the construction of the block around it.
A key-point is stronger if larger summation of quantities
(coefficients) of each component is required for the
construction of the block around it.
The metric should provide chance for lower 𝑪𝑴 keypoints to be stronger than higher 𝑪𝑴 key-points. This
condition is to diversify the types of the detected keypoints, which means not only key-points with top 𝑪𝑴
values, but also those with lower 𝑪𝑴s could be detected
even though only a limited number of key-points are
expected.
3.4. Non-maxima suppression
After calculating the 𝑺𝑴s of all detected key-points, a small
window is slid through all key-points in the image. At any
position, a key-point is kept if its 𝑺𝑴 is higher than all of the
𝑺𝑴 s of its neighborhood key-points. Otherwise, it is
suppressed. The suppression process contributes to the
spreading of the detected key-point types, since it diminishes
many nearby key-points which have very similar 𝑪𝑴s. The
suppression also helps improve the distribution of key-points
in the image, which is good for some situations such as
occlusion in which points around a neighborhood are blocked
by other objects and unable to be detected by the detector.
4. EXPERIMENTAL RESULTS
4.1. Experimental setup
Currently, the sparse coding step is implemented using
SPAMs library [14, 15] while other parts are implemented
solely with Matlab. The dictionary used is Haar provided by
Matlab. The block size is 11 by 11. The two benchmark
Webcam and EF datasets [1, 2] are used for the evaluation.
They are challenging datasets since the images are captured
at different scenes in different contexts such as drastic
changes in time, seasons, illumination, or background clutter.
Two metrics for the evaluation are repeatability and matching
scores [10]. The evaluation code is provided by Mikolajczyk
et al. and VLBenchmarks [10, 18]. As defined in [10], the
repeatability is the ratio between the number of
corresponding feature regions and the smaller number of
regions in the common part of an image pair. With a known
transformation between two images, two feature regions are
considered corresponding if the overlap error of one region in
an image and the projected region from another image is less
than 0.4. The matching score, meanwhile, is the ratio between
the number of correct matches and the smaller detected
region number in the image pair. If the distance of a
corresponding pair is the minimum in the descriptor space,
the match is considered a correct one. Since a large number
of features may lead to random matching of features and
unreasonably high performance, 1000 key-points per image
are considered as in [1, 9].
4.2. Quantitative results
As shown in Table 1, we compare our average repeatability
with hand-crafted features such as SIFT [5], SURF [6],
MSER [7], SFOP [8], Hessian Laplace [19], Harris Laplace
[19], Hessian Affine [19], Harris Affine [19]. Except for our
method, the performances of other detectors are collected
from [9]. Since we evaluate our method with the same
benchmarks, the same evaluation methods, and the same
number of key-points (1000/per image), it is a fair
comparison. As demonstrated in Table 1, our detector
outperforms all hand-crafted detectors by a large margin in
Webcam dataset. For EF dataset, it is higher than the top
hand-crafted feature by 0.2%.
For the matching score, we compare our method with SIFT
[5], Hessian Affine (HesAff) [19], and some state-of-the-art
learning based features such as T-P24 (the best version of
TIDLE) [1], Covariant point detector (CovDet) [20], and
state-of-the-art detector [9]. The same evaluation procedure
with SIFT descriptor is used, so the data is again collected
from [9]. Table 2 shows that our detector outperforms even
the top learning based detector on both datasets and has the
highest average score. Our score is, however, lower than
SIFT on EF. One probable reason is that unlike others, our
detector is not currently implemented with the scale invariant
algorithm. EF, in fact, contains drastic changes in scale, and
the selection of a correct pyramid level affects the descriptor
values in the evaluation.
The general high performance of our detector could be
explained as follows:
1.
2.
Since all types of structures such as corners, blobs,
junctions or even more sophisticated structures could be
detected (See Section 5.2), we have the overall benefits
of all different distinctive structures that a detector for a
single type structure does not have.
Since all low quality structures are filtered out, these
structures have no chance to decrease the performance.
Other detectors for detecting a specific structure may
include high and low quality structures of the type, since
other structures cannot be detected to cover the lack of
quality.
Table 1. Repeatability (%) of different methods on the
two benchmark datasets.
Method
SIFT
SURF
MSER
SFOP
Hessian Laplace
Harris Laplace
Hessian Affine
Harris Affine
Ours
Webcam
29.5
46.0
45.1
43.8
51.1
48.2
42.5
38.4
60.4
5.2. The capacity of detecting all types of structures
In basic linear algebra, any vector 𝑿 ∈ 𝑅𝑁𝑥1 can be
represented as a linear combination of a complete basis set of
vectors of the same dimensions. Since the dictionary we used
in the sparse coding step is a complete or over-complete
dictionary, whether a block 𝑿 represents a corner, a blob, a
junction, or any other sophisticated structure, it can be
decomposed into a linear combination of our basis set. If the
number of nonzero components in the combination satisfies
a predefined complexity range, it will be detected as a keypoint by our method. To illustrate the capacity of detecting
all types of structures, Figure 1 is plotted. The x-axis shows
the relative 𝑪𝑴 compared to the maximum 𝑪𝑴 available in
Figure 2. The y-axis shows the percentage of key-points of
other methods that can also be detected by our method. By
decreasing 𝑪𝑴 from 100% to 0%, our method gradually
detects all the points given by other methods.
EF
20.8
39.7
37.1
36.1
38.8
35.7
26.6
22.7
39.9
Table 2. Matching score (%) of different methods on the
two benchmark datasets.
Dataset
Webcam
EF
Average
Detector
SIFT
HesAff
T-P24
CovDet
[9]
Ours
12.9
10.2
11.6
13.8
5.4
9.6
13.4
5.2
9.3
12.0
4.8
8.4
19.4
6.2
12.8
20.4
6.5
13.5
Figure 1. An illustration for the capacity of detection of
all types of structures.
5. DISCUSSION
5.1. The selection of sparse coding technique for the keypoint detector
The purpose of sparse coding is to represent an input block
with the least number of the components. The ideal case is
one or zero component for each block. However, the number
of components required to represent each block often varies
and is larger for more complex blocks even under the pressure
of the sparse coding algorithm. These complex blocks are
considered distinctive blocks and good key-points for the
detector. This point, in fact, separates our work from other
detectors, since we consider the distinctiveness of a key-point
based on the number of primitive components for the
construction of the block around it, while state-of-the-art
detectors consider the distinctiveness as the distinctive shapes
existing in the block.
Figure 2. An image in Webcam dataset (The image was
resized for displaying).
6. CONCLUSIONS
We have developed a new type of key-point detector based
on sparse coding which requires no particular human
designed structures to detect distinctive points in images.
Experimental results have shown that our key-point detector
demonstrates excellent results on two benchmark datasets.
The current work opens a new direction for the researchers in
the key-point detector field, and can potentially be applied in
numerous vision based applications.
7. REFERENCES
[1] Y. Verdie, Kwang Moo Yi, P. Fua and V. Lepetit, "TILDE: A
Temporally Invariant Learned DEtector," 2015 IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), Boston, MA,
2015, pp. 5279-5288.
[2] C. L. Zitnick and K. Ramnath, "Edge foci interest points," 2011
International Conference on Computer Vision, Barcelona, 2011, pp.
359-366.
[3] W. Forstner and E. Gulch, “A Fast Operator for Detection and
Precise Location of Distinct Points, Corners and Centres of Circular
Features,” ISPRS Intercommission Conference on Fast Processing
of Photogrammetric Data, 1987.
[4] C. Harris and M. Stephens, “A Combined Corner and Edge
Detector,” Fourth Alvey Vision Conference, 1988.
[5] David G. Lowe, "Distinctive image features from scale-invariant
keypoints," International Journal of Computer Vision, 60, 2 (2004),
pp. 91-110.
[6] Herbert Bay, Tinne Tuytelaars, Luc Van Gool, "SURF:
Speeded Up Robust Features", Proceedings of the ninth European
Conference on Computer Vision, May 2006.
[7] J. Matas, O. Chum, M. Urban, and T. Pajdla, “Robust wide
baseline stereo from maximally stable extremal regions,” Image and
vision computing, 2004.
[8] W. Förstner, T. Dickscheid and F. Schindler, "Detecting
interpretable and accurate scale-invariant keypoints," 2009 IEEE
12th International Conference on Computer Vision, Kyoto, 2009, pp.
2256-2263.
[9] X. Zhang, F. X. Yu, S. Karaman and S. F. Chang, "Learning
Discriminative and Transformation Covariant Local Feature
Detectors," 2017 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), Honolulu, HI, 2017, pp. 4923-4931.
[10] K. Mikolajczyk, T. Tuytelaars, C. Schmid, A. Zisserman, J.
Matas, F. Schaffalitzky, T. Kadir and L. Van Gool, “A comparison
of affine region detectors,” International Journal of Computer
Vision, 65(1/2):43-72, 2005.
[11] P. Mainali, G. Lafruit, Q. Yang, B. Geelen, L. Van Gool, and
R. Lauwereins, “SIFER: Scale-Invariant Feature Detector with Error
Resilience,” International Journal of Computer Vision, 104(2):172–
197, 2013.
[12] P. Mainali, G. Lafruit, K. Tack, L. Van Gool and R. Lauwereins,
"Derivative-Based Scale Invariant Image Feature Detector With
Error Resilience," in IEEE Transactions on Image Processing, vol.
23, no. 5, pp. 2380-2391, May 2014.
[13] S. Salti, A. Lanza and L. Di Stefano, "Keypoints from
symmetries by wave propagation," 2013 IEEE Conference on
Computer Vision and Pattern Recognition, Portland, OR, 2013, pp.
2898-2905.
[14] J. Mairal, F. Bach and J. Ponce, “Sparse Modeling for Image
and Vision Processing,” Foundations and Trends in Computer
Graphics and Vision. vol 8. number 2-3. pages 85--283. 2014
[15] J. Mairal, F. Bach, J. Ponce and G. Sapiro, “Online Learning
for Matrix Factorization and Sparse Coding,” Journal of Machine
Learning Research, volume 11, pages 19-60. 2010.
[16] J. Han, S. He, X. Qian, D. Wang, L. Guo and T. Liu, "An
Object-Oriented Visual Saliency Detection Framework Based on
Sparse Coding Representations," in IEEE Transactions on Circuits
and Systems for Video Technology, vol. 23, no. 12, pp. 2009-2021,
Dec. 2013.
[17] J. Yan, M. Zhu, H. Liu and Y. Liu, "Visual Saliency Detection
via Sparsity Pursuit," in IEEE Signal Processing Letters, vol. 17, no.
8, pp. 739-742, Aug. 2010.
[18] K. Lenc, V. Gulshan, and A. Vedaldi, VLBenchmarks,
http://www.vlfeat.org/benchmarks/, 2012.
[19] K. Mikolajczyk and C. Schmid, “Scale & affine invariant
interest point detectors,” International Journal of Computer Vision,
60(1):63-86, 2004.
[20] K. Lenc and A. Vedaldi, “Learning covariant feature detectors,”
ECCV Workshop on Geometry Meets Deep Learning, 2016.
| 1cs.CV
|
On Packet Scheduling with Adversarial Jamming and Speedup∗
Martin Böhm1 , Lukasz Jeż2 , Jiřı́ Sgall1 , and Pavel Veselý1
1
Computer Science Institute of Charles University, Prague, Czech Republic,
{bohm,sgall,vesely}@iuuk.mff.cuni.cz
2
Institute of Computer Science, University of Wroclaw, Poland, [email protected]
arXiv:1705.07018v2 [cs.DS] 5 Oct 2017
Abstract
In Packet Scheduling with Adversarial Jamming packets of arbitrary sizes arrive over time to
be transmitted over a channel in which instantaneous jamming errors occur at times chosen by the
adversary and not known to the algorithm. The transmission taking place at the time of jamming is
corrupt, and the algorithm learns this fact immediately. An online algorithm maximizes the total size
of packets it successfully transmits and the goal is to develop an algorithm with the lowest possible
asymptotic competitive ratio, where the additive constant may depend on packet sizes.
Our main contribution is a universal algorithm that works for any speedup and packet sizes and,
unlike previous algorithms for the problem, it does not need to know these properties in advance. We
show that this algorithm guarantees 1-competitiveness with speedup 4, making it the first known
algorithm to maintain 1-competitiveness with a moderate speedup in the general setting of arbitrary
packet sizes. We also prove a lower bound of φ + 1 ≈ 2.618 on the speedup of any 1-competitive
deterministic algorithm, showing that our algorithm is close to the optimum.
Additionally, we formulate a general framework for analyzing our algorithm locally and use it
to show upper bounds on its competitive ratio for speedups in [1, 4) and for several special cases,
recovering some previously known results, each of which had a dedicated proof. In particular, our
algorithm is 3-competitive without speedup, matching the algorithm and the lower bound of Jurdzinski
et al. [7]. We use this framework also for the case of divisible packet sizes in which the size of a
packet divides the size of any larger packet, to show that a slight modification of our algorithm is
1-competitive with speedup 2 and it achieves the optimal competitive ratio of 2 without speedup,
again matching the algorithm and the lower bound of [7].
1
Introduction
We study an online packet scheduling model recently introduced by Anta et al. [1] and extended by
Jurdzinski et al. [7]. In our model, packets of arbitrary sizes arrive over time and they are to be transmitted
over a single communication channel. The algorithm can schedule any packet of its choice at any time,
but cannot interrupt its subsequent transmission; in the scheduling jargon, there is a single machine and
no preemptions. There are, however, instantaneous jamming errors or faults at times chosen by the
adversary, which are not known to the algorithm. A transmission taking place at the time of jamming is
corrupt, and the algorithm learns this fact immediately. The packet whose transmission failed can be
retransmitted immediately or at any later time.
The objective is to maximize the total size of packets successfully transmitted. In particular, the goal
is to develop an online algorithm with the lowest possible competitive ratio. We focus on algorithms with
resource augmentation, namely on online algorithms that transmit packets s ≥ 1 times faster than the
offline optimum solution they are compared against; such algorithm is often said to be speed-s, running
at speed s, or having a speedup of s. We note that this notion has been introduced for a similar job
scheduling problem [8], in which speedup is required for constant ratio. In contrast, our problem allows
constant ratio with no speedup, so we consider the competitive ratio as a function of it. This deviates
∗ Work partially supported by GA ČR project 17-09142S, GAUK project 634217 and Polish National Science Center
grant 2016/21/D/ST6/02402.
1
from previous work, which focused on the case with no speedup or on the speedup sufficient for ratio 1,
ignoring intermediate cases.
The distinguishing feature of the model is that the number of different packet sizes is a constant and
that these packet sizes are considered to be constants. Thus, the additive term in the definition of the
competitive ratio may depend on the number and values of the packet sizes. (It is easy to observe that,
at speed 1, without such additive term, it is impossible to attain constant competitive ratio that does not
depend on packet sizes, even for only two sizes.)
1.1
Previous Results
The model was introduced by Anta et al. [1], who resolved it for two packet sizes: If γ > 1 denotes the
ratio of the two sizes, then the optimum ratio for deterministic algorithms is (γ + bγc)/bγc, which is
always in the range [2, 3). Anta et al. [1] note that their lower bound strategy applies to randomized
algorithms as well, but their argument only applies to the adaptive adversary model. To our best
knowledge, randomized algorithms for the more common oblivious adversary model were never considered
in this setting. Anta et al. [1] demonstrate the necessity of instantaneous error feedback by proving that
discovering errors upon completed transmission rules out constant competitive ratio. They also provide
improved results for a stochastic online setting.
These results were extended by Jurdzinski et al. [7], who proved that the optimum ratio for the
case of multiple (though fixed) packet sizes is given by the same formula for the two packet sizes which
maximize it. Moreover, Jurdzinski et al. [7] gave further results for divisible packet sizes, i.e., instances
in which every packet size divides every larger packet size, for which the optimum ratio is 2. Namely,
they generalized the 2-competitive algorithm to a setting with f independent transmission channels (or
machines in scheduling parlance); jamming errors on each channel are independent, and a packet can
only be transmitted on one channel at a time. Furthermore, for the single channel divisible case, they
proved that speed 2 is sufficient for 1-competitiveness in the resource augmentation setting.
1.2
Our results
Our main new results concern the analysis of the non-divisible case with speedup, where we show that
speed 4 is sufficient for a competitive ratio of 1 (Section 5). However, our major contribution is a uniform
algorithm that works well in every setting, together with a uniform analysis framework (Section 4). This
contrasts with the results of Jurdzinski et al. [7], where each upper bound was attained by a dedicated
algorithm with independently crafted analysis. Moreover, unlike those algorithms, which all require prior
knowledge of all possible packet sizes and speedup, in the sense that there are different algorithms for
speeds 1 and 2 (only divisible case), with none for intermediate speeds, our algorithm is oblivious in
that it requires no such information. Furthermore, our algorithm is more appealing as it is significantly
simpler and “work-conserving” or “busy”, i.e., transmitting some packet whenever there is one pending,
which is desirable in practice. See Section 2 for the description of our algorithm.
To recover the 1-competitiveness at speed 2 and also 2-competitiveness at speed 1 for divisible instances,
we have to modify our algorithm slightly, as otherwise we can guarantee 1-competitiveness for divisible
instances only at speed 2.5. This is to be expected, as divisible instances are a very special case. The
analysis of the modified algorithm for divisible instances is in Section 7.
On the other hand, we prove that our original algorithm is 1-competitive on far broader class of
“well-separated” instances at sufficient speed: If the ratio between two successive packet sizes (in their
sorted list) is no smaller than α ≥ 1, our algorithm is 1-competitive if its speed is at least Sα , which
is a non-increasing function of α such that S1 = 4 and limα→∞ Sα = 2; see Section 4.2 for the precise
definition of Sα .
In Section 3 we demonstrate that the analyses of our algorithm are mostly tight, i.e., that (a) on
general instances, the algorithm is no better than (1 + 2/s)-competitive for s < 2 and no better than
4/s-competitive for s ∈ [2, 4), (b) on divisible instances, it is no better than 4/3-competitive for s < 2.5,
and (c) it is at least 2-competitive for s < 2, even for two divisible packet sizes. See Figure 1 for a graph
of our bounds on the competitive ratio depending on the speedup.
In Section 6 we complement these results with two lower bounds on speed that is necessary to achieve
1-competitiveness by a deterministic algorithm. The first one proves that even for two divisible packet
sizes, speed 2 is required to attain 1-competitiveness, establishing optimality of our modified algorithm
2
3
Local analysis upper bounds
Upper bound from Theorem 4.1
Hard instances for MAIN
Competitive ratio
2.5
2
1.5
1
1
2
3
4
5
6
Speedup
Figure 1: A graph of our upper and lower bounds on the competitive ratio of MAIN (s). The upper
bounds are from Section 4 and the lower bounds are by hard instances from Section 3.
and that of Jurdzinski et al. [7] for the divisible case. The second lower bound strengthens the previous
construction by showing that for non-divisible instances with more packet sizes, speed φ + 1 ≈ 2.618 is
required for 1-competitiveness. Both results hold even if all packets are released together.
1.3
Related Results
We first mention a few results concerning the model with adversarial jamming. Anta et al. [3] consider
essentially the same problem but with a different objective: minimizing the number and/or the total size
of pending (i.e., not yet transmitted) packets. They focus on multiple channels (or machines, as therein
the problem is formulated as job scheduling), and investigate what speedup is necessary and sufficient
for competitiveness. Note that for 1-competitiveness, minimizing the total size of pending packets is
equivalent to our objective. In Sections 4-7 they obtain a tight bound of 2 on speedup necessary for
1-competitiveness for two packet sizes, although in their lower bound of 2 they do not release all packets
together, unlike we do in our lower bound constructions. In Section 8 they claim to have a 1-competitive
algorithm with speedup 7/2, but the proof is incorrect; even the example in Section 3 showing that our
algorithm is no better than 4/s-competitive for s ∈ [2, 4) works the same way for their algorithm.
Georgio et al. [6] consider the same problem without any speedup, restricted to jobs of unit size, under
various efficiency and correctness measures, again focusing on multiple machines/channels. This article is
motivated by distributed computation, and thus distinguishes between different information models and
ways in which the machines communicate. In yet another work, Anta et al. [2] consider latency on top of
the previous two objectives, and study the competitiveness of popular scheduling algorithms for all these
measures. Garncarek et al. [5] consider “synchronized” parallel channels that all suffer errors at the same
time. This work distinguishes between “regular” jamming errors and “crashes”, which also cause the
algorithm’s state to reset, losing any information stored about the past√events and proves that for two
packet sizes the optimum ratio tends to 4/3 for the former and to φ = ( 5 + 1)/2 ≈ 1.618 for the latter
as the number f of channels tends to infinity.
As mentioned before, Kalyanasundaram and Pruhs [8] introduced resource augmentation. Among
other results they proved that a constant competitive ratio is possible with a constant speedup for a
preemptive variant of real-time scheduling where each job has a release time, deadline, processing time
and a weight and the objective is to maximize the weight of jobs completed by their deadlines on a single
machine. Subsequently resource augmentation was applied in various scenarios. Of the most relevant for
3
us are those that considered algorithms with speedup that are 1-competitive, i.e., as good as the optimum.
We mention two models that still contain interesting open problems.
For real-time scheduling, Phillips et al. [12] considered the underloaded case in which there exists
a schedule that completes all the jobs. It is easy to see that on a single machine, the earliest deadline
first (EDF) algorithm is then an optimal online algorithm. Phillips et al. [12] proved that EDF on m
machines is 1-competitive with speedup 2 − 1/m. (Here the weights are not relevant.) Intriguingly, finding
a 1-competitive algorithm with minimal speedup is wide open: It is known that speedup at least 6/5 is
necessary, it has been conjectured that speedup e/(e − 1) is sufficient, but the best upper bound proven
is 2 − 2/(m + 1) from [11]. See Schewior [13] for more on this problem.
Later these results were extended to real-time scheduling of overloaded systems, where for uniform
density (i.e., weight equal to processing time) Lam et al. [10] have shown that a variant of EDF with
admission control is 1-competitive with speedup 2 on a single machine and with speedup 3 on more
machines. For non-uniform densities, the necessary speedup is a constant if each job is tight (its deadline
equals its release time plus its processing time) [9]. Without this restriction it is no longer constant,
depending on the ratio ξ of the maximum and minimum weight. It is known that it is at least Ω(log log ξ)
and at most O(log ξ) [4, 10]; as far as we are aware, closing this gap is still an open problem.
2
Algorithms, preliminaries, notations
The general idea of the algorithm is that after each error, we start by transmitting packets of small sizes,
only increasing the size of packets after a sufficiently long period of uninterrupted transmissions. It turns
out that the right tradeoff is to transmit a packet only if it would have been transmitted successfully if
started just after the last error. It is also crucial that the initial packet after each error has the right
size, namely to ignore small packet sizes if the total size of remaining packets of those sizes is small
compared to a larger packet that can be transmitted. In other words, the size of the first transmitted
packet is larger than the total size of all pending smaller packets and we choose the largest such size.
This guarantees that if no error occurs, all currently pending packets with size equal to or larger than the
size of the initial packet are eventually transmitted before the algorithm starts a smaller packet.
We start by some notations. We assume there are k distinct non-zero packet sizes denoted by `i
and ordered so that `1 < · · · < `k . For convenience, we define `0 = 0. We say that the packet sizes are
divisible if `i divides `i+1 for all i = 1, . . . , k − 1. For a packet p, let `(p) denote the size of p. For a set of
packets A, let `(A) denote the total size of all the packets in A.
During the run of an algorithm, at time t, a packet is pending if it is released before or at t, not
completed before or at t and not started before t and still running. At time t, if no packet is running, the
algorithm may start any pending packet. As a convention of our model, if a fault (jamming error) happens
at time t and this is the completion time of a previously scheduled packet, this packet is considered
completed. Also, at the fault time, the algorithm may start a new packet.
Let LALG (i, Y ) denote the total size of packets of size `i completed by an algorithm ALG during a
time interval Y . Similarly, LALG (≥ i, Y ) (resp. LALG (< i, Y )) denotes the total size of packets of size at
least `i (resp. less than `i ) completed by an algorithm ALG during a time interval Y ; formally we define
Pk
Pi−1
LALG (≥ i, Y ) = j=i LALG (j, Y ) and LALG (< i, Y ) = j=1 LALG (j, Y ). We use the notation LALG (Y )
with a single parameter to denote the size LALG (≥ 1, Y ) of packets of all sizes completed by ALG during
Y and the notation LALG without parameters to denote the size of all packets of all sizes completed by
ALG at any time.
By convention, the schedule starts at time 0 and ends at time T , which is a part of the instance
unknown to an online algorithm until it is reached. (This is similar to the times of jamming errors, one
can also alternatively say that after T the errors are so frequent that no packet is completed.) Algorithm
ALG is called R-competitive, if there exists a constant C, possibly dependent on k and `1 , . . . , `k , such
that for any instance and its optimal schedule OPT we have LOPT ((0, T ]) ≤ R · LALG ((0, T ]) + C. We
remark that in our analyses we show only a crude bound on C.
We denote the algorithm ALG with speedup s ≥ 1 by ALG(s). The meaning is that in ALG(s),
packets of size L need time L/s to process. In the resource-augmentation variant, we are mainly interested
in finding the smallest s such that ALG(s) is 1-competitive, compared to OPT = OPT (1) that runs at
speed 1.
4
2.1
Algorithms MAIN and DIV
We now give a common description of the two algorithms, MAIN for general packet sizes and DIV for
divisible sizes. We divide the run of the algorithm into phases. Each phase starts by an invocation of the
initial step in which we need to carefully select a packet to transmit as discussed above. The phase ends
by a fault, or when there is no pending packet, or when there are pending packets only of sizes larger
than the total size of packets completed in the current phase. The periods of idle time do not belong to
any phase.
Formally, throughout the algorithm, t denotes the current time. The time tB denotes the start of
the current phase; initially tB = 0. We set rel(t) = s · (t − tB ). Since the algorithm does not insert
unnecessary idle time, rel(t) denotes the amount of transmitted packets in the current phase. (Note that
we use rel(t) only when there is no packet running at time t, so there is no partially executed packet.)
Thus rel(t) can be thought as a measure of time relative to the start of the current phase (scaled by the
speed of the algorithm). Note also that the algorithm can evaluate rel(t) without knowing the speedup,
as it can simply observe the total size of the transmitted packets. Let P <i denote the set of pending
packets of sizes `1 , . . . , `i−1 at any given time.
The description of MAIN and DIV differ only in Step (3), where DIV enforces an additional divisibility
condition.
Algorithm MAIN and DIV
(1) If no packet is available, stay idle until the next release time.
(2) Let i be the maximal i ≤ k such that there is a pending packet of size `i and `(P <i ) < `i .
Schedule a packet of size `i and set tB = t.
(3) Choose the maximum i such that
(i) there is a pending packet of size `i ,
(ii) `i ≤ rel(t) and
(iii) in case of DIV , `i divides rel(t).
Schedule a packet of size `i . Repeat Step (3) as long as such i exists.
(4) If no packet satisfies the condition in Step (3), go to Step 1.
We first note that the algorithm is well-defined, i.e., that it is always able to choose a packet in
Step (2) if it has any packets pending, and that if it succeeds in sending it, the length of thus started
phase can be related to the total size of the packets completed in it.
Lemma 2.1. Let ALG be either MAIN or DIV . In Step (2), ALG always chooses some packet if it has any
pending. Moreover, if ALG completes the first packet in the phase, then LALG(s) ((tB , tE ]) > s·(tE −tB )/2,
where tB denotes the start of the phase and tE its end (by a fault or Step (4)).
Proof. For the first property, note that the packet of the smallest size is eligible. For the second property,
note that there is no idle time in the phase, and that only the last packet chosen by ALG in the phase
may not complete due to a jam. By the condition in Step (3), the size of this jammed packet is no larger
than the total size of all the packets ALG previously completed in this phase, which yields the bound.
The following lemma shows a crucial property of the algorithm, namely that if packets of size `i are
pending, the algorithm schedules packets of size at least `i most of the time. Its proof also explains
the reasons behind our choice of the first packet in a phase in Step (2) of the algorithm. We give the
statement and proof of the lemma for MAIN . For DIV , the first part of the lemma still holds, while the
divisibility condition makes the proof a bit more tedious; on the other hand divisibility of packet sizes
ensures that the bound in the second part of the lemma can be improved. We postpone this tightened
analysis to Section 7 where we analyze DIV .
Lemma 2.2. Let u be a start of a phase in MAIN (s) and t = u + `i /s.
(i) If a packet of size `i is pending at time u and no fault occurs in (u, t), then the phase does not end
before t.
(ii) Suppose that v > u is such that any time in [u, v) a packet of size `i is pending and no fault occurs.
Then the phase does not end in (u, v) and LMAIN (s) (< i, (u, v]) < `i + `i−1 . (Recall that `0 = 0.)
5
Proof. (i) Suppose for a contradiction that the phase started at u ends at time t0 < t. We have
rel(t0 ) < rel(t) = `i . Let `j be the smallest packet size among the packets pending at t0 . As there is no
fault, the reason for a new phase has to be that rel(t0 ) < `j , and thus Step (3) did not choose a packet to
be scheduled. Also note that any packet scheduled before t0 was completed. This implies, first, that there
is a pending packet of size `i , as there was one at time u and there was insufficient time to complete it,
so j is well-defined and j ≤ i. Second, all packets of sizes smaller than `j pending at u were completed
before t0 , so their total size is at most rel(t0 ) < `j . However, this contradicts the fact that the phase
started by a packet smaller than `j at time u, as a pending packet of the smallest size equal to or larger
than `j satisfied the condition in Step (2) at time u. (Note that it is possible that no packet of size `j is
pending at u; however, at least a packet of size `i is pending.)
(ii) By (i), the phase that started at u does not end before time t if no fault happens. A packet of size
`i is always pending by the assumption of the lemma, and it is always a valid choice of a packet in Step
(3) from time t on. Thus, the phase that started at u does not end in (u, v), and moreover only packets
of sizes at least `i are started in [t, v). It follows that packets of sizes smaller than `i are started only
before time t and their total size is thus less than rel(t) + `i−1 = `i + `i−1 . The lemma follows.
3
Examples for MAIN and DIV
Before analyzing the algorithms, we show some hard instances obtaining some lower bounds on the
performance of the algorithms, in some cases matching our upper bounds from later sections. At the
same time, the examples are instructive in understanding some of the choices both in the analysis and
the design of algorithms.
3.1
Examples with general packet sizes
First we give an instance which shows that MAIN (s) has competitive ratio at least 1 + 2/s for s < 2,
matching Theorem 4.4.
Choose a large enough integer N . At time 0 the following packets are released: 2N packets of size 1,
one packet of size 2 and N packet of size 4/s − ε for a small enough ε > 0 such that it holds 2 < 4/s − ε.
These are all packets in the instance.
First there are N phases, each of length 4/s − ε and ending by a fault. OPT completes a packet of
size 4/s − ε in each phase, while MAIN (s) completes 2 packets of size 1 and then it starts a packet of size
2 which is not finished.
Then there is a fault every 1 unit of time, so that OPT completes all packets of size 1, while the
algorithm has no pending packet of size 1 and as s < 2 the length of the phase is not sufficient to finish a
longer packet.
Overall, OPT completes packets of total size 2 + 4/s − ε per phase, while the algorithm completes
packets of total size only 2 per phase. The ratio thus tends to 1 + 2/s as ε → 0.
Next we prove that for s ∈ [2, 4), the ratio of MAIN (s) is at least 4/s > 1, which shows that the
speedup of MAIN (s) in Theorem 5.1 cannot be improved.
Choose a large enough integer y. There will be four packet sizes: 1, x, y and z such that 1 < x < y < z,
z = x + y − 1, and x = y · (s − 2)/2 + 2; as s ≥ 2 it holds x > 1 and as s < 4 we have x ≤ y − 1 for a
large enough y.
We will have N phases again. At time 0 the adversary releases all N (y − 1) packets of size 1, all N
packets of size y and a single packet of size z (never completed by either OPT or MAIN (s)), whereas the
packets of size x are released one per phase.
In each phase MAIN (s) completes, in this order: y − 1 packets of size 1 and then a packet of size x,
which has arrived just after the y − 1 packets of size 1 are completed. Next, it will start a packet of size z
and fail due to a jam. We show that OPT will complete a packet of size y. To this end, it is required
that y < 2(x + y − 1)/s, or equivalently x > y · (s − 2)/2 + 1 which holds by the choice of x.
After these N phases, we will have jams every 1 unit of time, so that OPT can complete all the
N (y − 1) packets of size 1, while MAIN (s) will be unable to complete any packet (of size y or larger).
6
The ratio per phase is
OPT
y−1+y
2y − 1
2y − 1
=
=
= y·s
y·(s−2)
MAIN (s)
y−1+x
y−1+ 2 +2
2 +1
which tends to 4/s as y → ∞.
3.2
Example with two packet sizes
Our next example shows that both algorithms have ratio no smaller than 2 when s < 2, even if packet
sizes are only 1 and ` ≥ max{s + , /(2 − s)} for an arbitrarily small > 0. This matches upper bounds
given in Theorem 4.4 for MAIN (2) and an upper bound of 2 for DIV . We denote either algorithm by
ALG.
There will be N phases, that all look the same: In each phase, issue one packet of size ` and ` packets
of size 1, and have the phase end by a fault at time (2` − ε)/s ≥ ` which holds by the bounds on `.
Then ALG will complete all ` packets of size 1 but will not complete the one of size `. By the previous
inequality, OPT can complete the packet of size ` within the phase. Once all N phases are over, the jams
occur every 1 unit of time, which allows OPT completing all N ` remaining packets of size 1. However,
ALG is unable to complete any of the packets of size `. Thus the ratio is 2.
3.3
Example for MAIN in the divisible case
We give an example that shows that MAIN is not very good for divisible instances, in particular it is not
1-competitive for any speed s < 2.5 and thus the bound in Theorem 4.5 is tight. In fact, we show that it
is not better than 4/3-competitive. We use packets of sizes 1, `, and 2` and we take ` sufficiently large
compared to the given speed or competitive ratio. There are many packets of size 1 and 2` available at
the beginning, the packets of size ` arrive at specific times where MAIN schedules them immediately.
The faults occur at times divisible by 2`, so the optimum schedules one packet of size 2` in each
phase between two faults. We have N such phases, N (2` − 1) packets of size 1 and N packets of size 2`
available at the beginning. In each phase, MAIN (s) schedules 2` − 1 packets of size 1, then a packet of
size ` arrives and is scheduled, and then a packet of size 2` is scheduled. The algorithm would need speed
2.5 − 1/(2`) to complete it. So, for ` large, the algorithm completes only packets of total size 3` − 1 per
phase. After these N phases, we have faults after time 1, so the optimum schedules all packets of size 1.
The optimum thus finishes all packets 2` plus all small packets, a total of 4` − 1 per phase. Thus the
ratio tends to 4/3 as ` → ∞.
4
Local analysis and results
In this section we formulate a general method for analyzing our algorithms by comparing locally within
each phase the size of “large” packets completed by the algorithm and by the adversary. This method
simplifies a complicated induction used in [7]; we easily obtain the same upper bounds of 2 and 3 on
competitiveness for divisible and unrestricted packet sizes respectively at no speedup, as well as several
new results for the non-divisible cases included in this section. In Section 5, we use a more complex
charging scheme to obtain our main result. We postpone the use of local analysis for the divisible case to
Section 7.
For the analysis, let ALG denote one of the algorithms MAIN and DIV , and let s ≥ 1 be the speedup.
We fix an instance and its schedules for ALG(s) and OPT .
4.1
Critical times and Master Theorem
The common scheme is the following. We carefully define a sequence of critical times Ck ≤ Ck−1 ≤
· · · ≤ C1 ≤ C0 , where C0 = T is the end of the schedule, satisfying two properties: (1) the algorithm has
completed almost all pending packets of size `i released before Ci and (2) in (Ci , Ci−1 ], a packet of size
`i is always pending. Properties (1) and (2) allow us to relate LOPT (i, (0, Ci ]) and LOPT (≥ i, (Ci , Ci−1 ])
respectively to their “ALG counterparts”. As each packet completed by OPT belongs to exactly one of
7
`k
`k
`i
`i−1
`1
Ck
Ci−1
Ci
C2 C1
C0
Figure 2: An illustration of dividing the schedule of OPT in the local analysis. Rows correspond to
packet sizes and the X-axis to time. Gray horizontal rectangles thus correspond to LOPT (i, (0, Ci ]) and
hatched rectangles correspond to LOPT (≥ i, (Ci , Ci−1 ]).
these sets, summing the bounds gives the desired results. These two facts together imply R-competitiveness
of the algorithm for appropriate R and speed s. See Figure 2 for an illustration.
We first define the notion of i-good times so that they satisfy property (1), and then choose the critical
times among their suprema so that those satisfy property (2) as well.
Definition 1. Let ALG be one of the algorithms MAIN and DIV , and let s ≥ 1 be the speedup. For
i = 1, . . . k, time t is called i-good if one of the following conditions holds:
(i) At time t, algorithm ALG(s) starts a new phase by scheduling a packet of size larger than `i , or
(ii) at time t, no packet of size `i is pending for ALG(s), or
(iii) t = 0.
We define critical times C0 , C1 , . . . , Ck iteratively as follows:
• C0 = T , i.e., it is the end of the schedule.
• For i = 1, . . . , k, Ci is the supremum of i-good times t such that t ≤ Ci−1 .
Note that all Ci ’s are defined and Ci ≥ 0, as time t = 0 is i-good for all i.
The choice of Ci implies that each Ci is of one of the three types (the types are not disjoint):
• Ci is i-good and a phase starts at Ci (this includes Ci = 0),
• Ci is i-good and Ci = Ci−1 , or
• there exists a packet of size `i pending at Ci , however, any such packet was released at Ci .
If the first two options do not apply, then the last one is the only remaining possibility (as otherwise
some time in a non-empty interval (Ci , Ci−1 ] would be i-good); in this case, Ci is not i-good but only the
supremum of i-good times.
First we bound the total size of packets of size `i completed before Ci ; the proof actually only uses
the fact that each Ci is the supremum of i-good times and justifies the definition above.
Lemma 4.1. Let ALG be one of the algorithms MAIN and DIV , and let s ≥ 1 be the speedup. Then,
for any i, LOPT (i, (0, Ci ]) ≤ LALG(s) (i, (0, Ci ]) + `k .
Proof. If Ci is i-good and satisfies condition (i) in Definition 1, then by the description of Step (2) of the
algorithms, the total size of pending packets of size `i is less than the size of the scheduled packet, which
is at most `k and the lemma follows.
In all the remaining cases it holds that ALG(s) has completed all the jobs of size `i released before Ci .
Thus the inequality holds trivially even without the additive term.
Our remaining goal is to bound LOPT (≥ i, (Ci , Ci−1 ]). We divide (Ci , Ci−1 ] into i-segments by the
faults. We prove the bounds separately for each i-segment. One important fact is that for the first
i-segment we use only a loose bound, as we can use the additive constant. The critical part is then the
bound for i-segments started by a fault, this part determines the competitive ratio and is different for
each case. We summarize the general method by the following definition and master theorem.
Definition 2. The interval (u, v] is called an initial i-segment if u = Ci and v is either Ci−1 or the first
time of a fault after u, whichever comes first.
8
The interval (u, v] is called a proper i-segment if u ∈ (Ci , Ci−1 ) is a time of a fault and v is either
Ci−1 or the first time of a fault after u, whichever comes first.
Note that there is no i-segment if Ci−1 = Ci .
Theorem 4.2 (Master Theorem). Let ALG denote one of the algorithms MAIN and DIV , and let s ≥ 1
be the speedup. Suppose that R ≥ 1 satisfies that for each i = 1, . . . , k and each proper i-segment (u, v]
with v − u ≥ `i we have
(R − 1)LALG(s) ((u, v]) + LALG(s) (≥ i, (u, v]) ≥ LOPT (≥ i, (u, v]) .
(4.1)
Suppose furthermore that for the initial i-segment (u, v] we have
LALG(s) (≥ i, (u, v]) > s(v − u) − 4`k .
(4.2)
Then ALG(s) is R-competitive.
Proof. First note that for a proper i-segment (u, v], u is a fault time. Thus if v − u < `i , then
LOPT (≥ i, (u, v]) = 0 and (4.1) is trivial. It follows that (4.1) holds even without the assumption
v − u ≥ `i .
Now consider an initial i-segment (u, v]. We have LOPT (≥ i, (u, v]) ≤ `k + v − u, as at most a
single packet started before u can be completed. Combining this with (4.2) and using s ≥ 1, we get
LALG(s) (≥ i, (u, v]) > s(v − u) − 4`k ≥ v − u − 4`k ≥ LOPT (≥ i, (u, v]) − 5`k .
Summing this with (4.1) for all proper i-intervals and using R ≥ 1 we get
(R − 1)LALG(s) ((Ci , Ci−1 ]) + LALG(s) (≥ i, (Ci , Ci−1 ]) + 5`k
≥ LOPT (≥ i, (Ci , Ci−1 ]) .
(4.3)
Note that for Ci = Ci−1 , Equation (4.3) holds trivially.
To complete the proof of the theorem, note that each completed job in the optimum contributes to
exactly one among the 2k terms LOPT (≥ i, (Ci , Ci−1 ]) and LOPT (i, (0, Ci ]); similarly for LALG(s) . Thus
by summing both (4.3) and Lemma 4.1 for all i = 1, . . . , k we obtain
LOPT
=
k
X
LOPT (≥ i, (Ci , Ci−1 ]) +
i=1
≤
k
X
k
X
(R − 1)LALG(s) ((Ci , Ci−1 ]) +
i=1
+
LOPT (i, (0, Ci ])
i=1
k
X
k
X
LALG(s) (≥ i, (Ci , Ci−1 ]) + 5`k
i=1
LALG(s) (i, (0, Ci ]) + `k
i=1
≤ (R − 1)LALG(s) + LALG(s) + 6k`k = R · LALG(s) + 6k`k .
The theorem follows.
4.2
Analysis of MAIN
The first part of the following lemma implies the condition (4.2) for the initial i-segments in all cases.
The second part of the lemma is the base of the analysis of a proper i-segment, which is different in each
situation.
Lemma 4.3. (i) If (u, v] is an initial i-segment then LMAIN (s) (≥ i, (u, v]) > s(v − u) − 4`k .
(ii) If (u, v] is a proper i-segment and s(v − u) ≥ `i then LMAIN (s) ((u, v]) > s(v − u)/2 and LMAIN (s) (≥
i, (u, v]) > s(v − u)/2 − `i − `i−1 . (Recall that `0 = 0.)
Proof. (i) Let (u, v] be an initial i-segment. If the phase that starts at u or contains u ends before v, let
u0 be its end; otherwise let u0 = u. We have u0 ≤ u + `i /s, as otherwise any packet of size `i , pending
throughout the i-segment by definition, would be an eligible choice in Step (3) of the algorithm, and the
phase would not end before v. Using Lemma 2.2(ii), we have LMAIN (s) (< i, (u0 , v]) < `i + `i−1 < 2`k .
9
Since at most one packet at the end of the segment is unfinished, we have LMAIN (s) (≥ i, (u, v]) ≥
LMAIN (s) (≥ i, (u0 , v]) > s(v − u0 ) − 3`k > s(v − u) − 4`k .
(ii) Let (u, v] be a proper i-segment. Thus u is a start of a phase that contains at least the whole
interval (u, v] by Lemma 2.2(ii). By the definition of Ci , u is not i-good, so the phase starts by a packet of
size at most `i . If v −u ≥ `i then the first packet finishes (as s ≥ 1) and thus LMAIN (s) ((u, v]) > s(v −u)/2
by Lemma 2.1. The total size of completed packets smaller than `i is at most `i + `i−1 by Lemma 2.2(ii),
and thus LMAIN (s) (≥ i, (u, v]) > s(v − u)/2 − `i − `i−1 .
The next theorem gives a tradeoff of the competitive ratio of MAIN (s) and the speedup s using
our local analysis. While Theorem 5.1 shows that MAIN (s) is 1-competitive for s ≥ 4, here we give a
weaker result that reflects the limits of the local analysis. However, for s = 1 our local analysis is tight as
already the lower bound from [1] shows that no algorithm is better than 3-competitive (for packet sizes 1
and 2 − ε). See Figure 1 for an illustration of our upper and lower bounds on the competitive ratio of
MAIN (s).
Theorem 4.4. MAIN (s) is Rs -competitive where:
Rs = 1 + 2/s for s ∈ [1, 4),
Rs = 2/3 + 2/s for s ∈ [4, 6), and
Rs = 1 for s ≥ 6.
Proof. Lemma 4.3(i) implies the condition (4.2) for the initial i-segments. We now prove (4.1) for any
proper i-segment (u, v] with v − u ≥ `i and appropriate R. The bound then follows by the Master
Theorem.
Since there is a fault at time u, we have LOPT (≥ i, (u, v]) ≤ v − u.
For s ≥ 6, Lemma 4.3(ii) implies
LMAIN (s) (≥ i, (u, v]) > s(v − u)/2 − 2`i
≥ 3(v − u) − 2(v − u) = v − u ≥ LOPT (≥ i, (u, v]) ,
which is (4.1) for R = 1.
For s ∈ [4, 6), by Lemma 4.3(ii) we have LMAIN (s) ((u, v]) > s(v − u)/2 and by multiplying it by
(2/s − 1/3) we obtain
2 1
s
−
(v − u) .
· LMAIN (s) ((u, v]) > 1 −
s 3
6
Thus to prove (4.1) for R = 2/3 + 2/s, it suffices to show that
LMAIN (s) (≥ i, (u, v]) >
s
(v − u) ,
6
as clearly v − u ≥ LOPT (≥ i, (u, v]). The remaining inequality again follows from Lemma 4.3(ii), but we
need to consider two cases:
If (v − u) ≥ 6s `i , then
LMAIN (s) (≥ i, (u, v]) >
s
s
s
s
(v − u) − 2`i ≥ (v − u) − (v − u) = (v − u) .
2
2
3
6
On the other hand, if (v − u) < 6s `i , then using s ≥ 4 as well,
LMAIN (s) (≥ i, (u, v]) >
s
(v − u) − 2`i ≥ 0 ,
2
therefore MAIN (s) completes a packet of size at least `i which implies
LMAIN (s) (≥ i, (u, v]) ≥ `i >
s
(v − u) .
6
For s ∈ [1, 4), by Lemma 4.3(ii) we get (2/s) · LMAIN (s) ((u, v]) > v − u ≥ LOPT (≥ i, (u, v]), where
the last inequality holds as there is a fault at time u. This implies (4.1) for R = 1 + 2/s.
10
5
(4α + 2)/α2
2 + 2/α
3 + 1/α for α < α1
Upper bound from Theorem 4.6
Speedup sufficient for 1-competitiveness
Speedup
4
3
2
1
1
α0
α1
2
α
3
Figure 3: A graph of Sα and the bounds on the speedup that we use in Theorem 4.6.
We can obtain better bounds on the speedup necessary for 1-competitiveness if the packet sizes
are sufficiently different. Namely, we call the packet sizes `1 , . . . , `k α-separated if `i ≥ α`i−1 holds for
i = 2, . . . , k. We start by analyzing MAIN (s) for the case of divisible packet sizes, which is a special case
of 2-separated packet sizes. Note that by examples in Section 3 the bound of 2.5 cannot be improved for
our algorithm.
Theorem 4.5. If the packet sizes are divisible, then MAIN (s) is 1-competitive for s ≥ 2.5.
Proof. Lemma 4.3(i) implies (4.2). We now prove (4.1) for any proper i-segment (u, v] with v − u ≥ `i
and R = 1. The bound then follows by the Master Theorem. Since there is a fault at time u, we have
LOPT (≥ i, (u, v]) ≤ v − u.
By divisibility we have LOPT (≥ i, (u, v]) = n`i for some nonnegative integer n. We distinguish two
cases based on the size of the last packet started by MAIN in the i-segment (u, v], which is possibly
unfinished due to a fault at v.
If the unfinished packet has size at most n`i , then LMAIN (s) (≥ i, (u, v]) > 5(v − u)/2 − 3`i /2 −
n`i ≥ 5n`i /2 − 3`i /2 − n`i ≥ (n − 1)`i by Lemma 2.1 and Lemma 2.2(ii). Divisibility now implies
LMAIN (s) (≥ i, (u, v]) ≥ n`i = LOPT (≥ i, (u, v]).
Otherwise by divisibility the size of the unfinished packet is at least (n + 1)`i and the size of the
completed packets is larger by the condition in Step (3) of the algorithm; here we also use the fact that
MAIN (s) completes the packet started at u, as its size is at most `i ≤ v − u. Thus LMAIN (s) (≥ i, (u, v]) >
(n + 1)`i − 3`i /2 ≥ (n − 1/2)`i . Divisibility again implies LMAIN (s) (≥ i, (u, v]) ≥ n`i = LOPT (≥ i, (u, v]).
The theorem now follows by the Master Theorem.
Next, we show that for α-separated packet sizes, MAIN (Sα ) is 1-competitive for the following Sα .
We define
1 1√
α0 = +
33 ≈ 1.46 , which is the positive root of 3α2 − 3α − 2.
2 6
√
3 + 17
≈ 1.78 , which is the positive root of 2α2 − 3α − 1.
α1 =
4
4α + 2
for α ∈ [1, α0 ],
α2
Sα = 3 + 1
for α ∈ [α0 , α1 ), and
α
2
2+ α
for α ≥ α1 .
11
See Figure 3 for a graph of Sα and all the bounds on it that we use. The value of α0 is chosen as the point
where (4α + 2)/α2 = 2 + 2/α. The value of α1 is chosen as the point from which the argument in case
(viii) of the proof below works, which allows for a better result for α ≥ α1 . If s ≥ Sα then s ≥ (4α + 2)/α2
and s ≥ 2 + 2/α for all α and also s ≥ 3 + 1/α for α < α1 ; these facts follow from inspection of the
functions and are useful for the analysis.
Note that Sα is decreasing in α, with a single discontinuity at α1 . We have S1 = 6, matching the
upper bound for 1-competitiveness using local analysis. We have S2 = 3, i.e., MAIN (3) is 1-competitive
for 2-separated packet sizes, which includes
√ the case of divisible packet sizes studied above. The limit
of Sα for α → +∞ is 2. For α < (1 + 3)/2 ≈ 1.366, we get Sα > 4, while Theorem 5.1 shows that
MAIN (s) is 1-competitive for s ≥ 4; the weaker result of Theorem 4.6 below reflect the limits of the local
analysis.
Theorem 4.6. Let α > 1. If the packet sizes are α-separated, then MAIN (s) is 1-competitive for any
s ≥ Sα .
Proof. Lemma 4.3(i) implies (4.2). We now prove for any proper i-segment (u, v] with v − u ≥ `i that
LMAIN (s) (≥ i, (u, v]) ≥ LOPT (≥ i, (u, v]) ,
(4.4)
which is (4.1) for R = 1. The bound then follows by the Master Theorem.
Let X = LOPT (≥ i, (u, v]). Note that X ≤ v − u.
Lemma 4.3(ii) together with `i−1 ≤ `i /α gives LMAIN (s) (≥ i, (u, v]) > M for M = sX/2 − (1 + 1/α)`i .
As in the case of divisible packet sizes, we use the fact that both X and LMAIN (s) (≥ i, (u, v]) are sums
of some packet sizes `j , j ≥ i, and thus only some of the values are possible. Here the situation is more
complicated, as for example `i+1 , `i+2 , 2`i , `i + `i+1 are possible values, but their ordering may vary.
We distinguish several cases based on X and α. We note in advance that the first five cases suffice for
α < α1 ; only after completing the proof for α < α1 , we analyze the additional cases needed for α ≥ α1 .
Case (i): X = 0. Then (4.4) is trivial.
Case (ii): X = `i . Using s ≥ 2 + 2/α, we obtain M ≥ 0. Thus LMAIN (s) (≥ i, (u, v]) > M ≥ 0 which
implies LMAIN (s) (≥ i, (u, v]) ≥ `i = X and (4.4) holds.
Case (iii): X = `i+1 ≤ 2`i . Using s ≥ (4α + 2)/α2 and X = `i+1 ≥ α`i , we obtain M ≥ (2 +
1/α)`i − (1 + 1/α)`i = `i . Thus LMAIN (s) (≥ i, (u, v]) > `i which together with `i+1 ≤ 2`i implies
LMAIN (s) (≥ i, (u, v]) ≥ `i+1 = X and (4.4) holds.
Case (iv): X ≥ α2 `i . (Note that this includes all cases when a packet of size at least `i+2 contributes to
X.) We first show that s ≥ 2(1 + 1/α2 + 1/α3 ) by straightforward calculations with the golden ratio φ: If
α ≤ φ, we have 1/α + 1/α2 − 1 ≥ 0 and thus s ≥ (4α + 2)/α2 = 2(2/α + 1/α2 ) ≥ 2(1 + 1/α2 + 1/α3 ), as
2/α ≥ 1 + 1/α3 is equivalent to (α − 1)(α2 − α − 1) ≤ 0, which holds for α ∈ (1, φ). If on the other hand
α ≥ φ, then s ≥ 2(1 + 1/α) ≥ 2(1 + 1/α2 + 1/α3 ), as 1/α ≥ 1/α2 + 1/α3 holds for α ≥ φ. Now we obtain
s
1
1
1
1
M −X ≥
−1 X − 1+
`i ≥ 1 + 2 + 3 − 1 X − 1 +
`i
2
α
α
α
α
1
1
1
+ 3 α2 `i − 1 +
`i = 0 ,
≥
2
α
α
α
and (4.4) holds.
Case (v): X ≥ 2`i and α < α1 . (Note that, as long as α < α1 , this includes all cases when at least two
packets contribute to X.) Using s ≥ 3 + 1/α we obtain
1
1
1
1
1
1
M −X ≥
3+
−1 X − 1+
`i ≥
1+
2`i − 1 +
`i = 0 ,
2
α
α
2
α
α
and (4.4) holds.
Proof for α < α1 : We now observe that for α < α1 , we have exhausted all the possible values of X.
Indeed, if (v) does not apply, then at most a single packet contributes to X, and one of the cases (i)-(iv)
applies, as (iv) covers the case when X ≥ `i+2 . Thus (4.4) holds and the proof is complete.
Proof for α ≥ α1 : We now analyze the remaining cases for α ≥ α1 .
12
Case (vi): X ≥ (α + 1)`i . (Note that this includes all cases when two packets not both of size `i
contribute to X.) Using s ≥ 2 + 2/α we obtain
1
1
1
1
`i ≥ (α + 1)`i − 1 +
`i = 0
M −X ≥ 1+ −1 X − 1+
α
α
α
α
and (4.4) holds.
Case (vii): X = n · `i < (α + 1)`i for some n = 2, 3, . . .. Since α > α1 > φ, we have `i+1 > `i + `i−1 .
This implies that the first packet of size at least `i that is scheduled in the phase has size equal to `i by
the condition in Step (3) of the algorithm. Thus, if also a packet of size larger than `i contributes to
LMAIN (s) (≥ i, (u, v]), we have
LMAIN (s) (≥ i, (u, v]) ≥ `i+1 + `i ≥ (α + 1)`i > X
by the case condition and (4.4) holds. Otherwise LMAIN (s) (≥ i, (u, v]) is a multiple of `i . Using
s ≥ 2 + 2/α, we obtain M ≥ (n − 1)(1 + 1/α)`i > (n − 1)`i . This, together with divisibility by `i implies
LMAIN (s) (≥ i, (u, v]) ≥ n · `i = X and (4.4) holds again.
Case (viii): X = `i+1 > 2`i . We distinguish two subcases depending on the size of the unfinished packet
of MAIN (s) in this phase.
If the unfinished packet has size at most `i+1 , the size of the completed packets is bounded by
2
LMAIN (s) ((u, v]) > sX − `i+1 = (s − 1)`i+1 ≥ 1 +
`i+1 ,
α
using s ≥ 2 + 2/α. Since the total size of packets smaller than `i is less then (1 + 1/α)`i by Lemma 2.2(ii),
we obtain
2`i+1
1
LMAIN (s) (≥ i, (u, v]) − X >
− 1+
`i > 0 ,
α
α
as `i+1 /α ≥ `i . Thus (4.4) holds.
Otherwise the unfinished packet has size at least `i+2 and, by Step (3) of the algorithm, also
LMAIN (s) ((u, v]) > `i+2 . We have `i+2 ≥ α`i+1 and by the case condition `i+1 > 2`i we obtain
1
1
`i > 2(α − 1)`i − 1 +
`i ≥ 0 ,
LMAIN (s) (≥ i, (u, v]) − X > (α − 1)`i+1 − 1 +
α
α
as the definition of α1 implies that 2(α − 1) ≥ 1 + 1/α for α ≥ α1 . Thus (4.4) holds.
We now observe that we have exhausted all the possible values of X for α ≥ α1 . Indeed, if at least
two packets contribute to X, either (vi) or (vii) applies. Otherwise, at most a single packet contributes
to X, and one of the cases (i)-(iv) or (viii) applies, as (iv) covers the case when X ≥ `i+2 . Thus (4.4)
holds and the proof is complete.
5
MAIN with speed 4
In this section we prove that speed 4 is sufficient for MAIN to be 1-competitive. Examples in Section 3
show that speed 4 is also necessary for our algorithm.
Theorem 5.1. MAIN (s) is 1-competitive for s ≥ 4.
Intuition. For s ≥ 4 we have that if at the start of a phase MAIN (s) has a packet of size `i pending
and the phase has length at least `i , then MAIN (s) completes a packet of size at least `i . To show this,
assume that the phase starts at time t. Then the first packet p of size at least `i is started before time
t + 2`i /s by Lemma 2.2(ii) and by the condition in Step (3) it has size smaller than 2`i . Thus it completes
before time t + 4`i /s ≤ t + `i , which is before the end of the phase. This property does not hold for s < 4.
It is important in our proof, as it shows that if the optimal schedule completes a job of some size, and
such job is pending for MAIN (s), then MAIN (s) completes a job of the same size or larger. However,
13
this is not sufficient to complete the proof by a local (phase-by-phase) analysis similar to the previous
section, as the next example shows.
Assume that at the beginning, we release N packets of size 1, N packets of size 1.5 − 2ε, one packet
of size 3 − 2ε and a sufficient number of packets of size 1 − ε, for a small ε > 0. Our focus is on packets
of size at least 1. Supposing s = 4 we have the following phases:
• First, there are N phases of length 1. In each phase the optimum completes a packet of size 1,
while among packets of size at least 1, MAIN (s) completes a packet of size 1.5 − 2ε, as it starts
packets of sizes 1 − ε, 1 − ε, 1.5 − 2ε, 3 − 2ε, in this order, and the last packet is jammed.
• Then there are N phases of length 1.5 − 2ε where the optimum completes a packet of size 1.5 − 2ε
while among packets of size at least 1, the algorithm completes only a single packet of size 1, as it
starts packets of sizes 1 − ε, 1 − ε, 1, 3 − 2ε, in this order. The last packet is jammed, since for
s = 4 the phase must have length at least 1.5 − ε to complete it.
In phases of the second type, the algorithm does not complete more (in terms of total size) packets of size
at least 1 than the optimum. Nevertheless, in our example, packets of size 1.5 − 2ε were already finished
by the algorithm, and this is a general rule. The novelty in our proof is a complex charging argument
that exploits such subtle interaction between phases.
Outline of the proof. We define critical times Ci0 similarly as before, but without the condition that
0
they should be ordered, i.e., Ci0 is simply the supremum of i-good times; consequently, either Ci0 ≤ Ci−1
0
0
or Ci > Ci−1 may hold. Our plan is to define, for each i, charges between packets of size `i up to time
Ci0 , since nearly each packet of size `i completed by the adversary before Ci0 must be completed by the
algorithm before Ci0 . Each such charge from an adversary’s packet leads to an algorithm’s packet of the
same size and each algorithm’s packet receives at most one such charge; we thus call these charges 1-to-1
charges.
We want to charge the remaining adversary’s packets of size `i (scheduled after Ci0 ) to some algorithm’s
packets in the same phase. For technical reasons it is actually more convenient to partition the schedule
not into phases but into blocks inbetween successive faults. A block can contain several phases of the
algorithm separated by an execution of Step (4); however, in the most important and tight part of the
analysis the blocks coincide with phases.
After the critical time Ci0 , packets of size `i are always pending for the algorithm, and thus (as we
observed above) the algorithm schedules a packet of size at least `i in blocks in which the adversary
completes a packet of size `i . In the crucial lemma of the proof, based on these observations and their
refinements, we show that we can assign the remaining adversary’s packets (i.e., those without 1-to-1
charges) to algorithm’s packets in the same block so that for each algorithm’s packet q the total size of
packets assigned to it is at most `(q). However, we cannot use this assignment directly to charge the
remaining packets, as some of the algorithm’s big packets may receive 1-to-1 charges, and in this case
the analysis needs to handle the interaction of different blocks. This very issue can be seen even in our
introductory example.
To deal with this, we process blocks in the order of time from the beginning to the end of the schedule,
simultaneously completing the charging to the packets in the current block of MAIN (s) schedule and
possibly modifying ADV (the schedule of the adversary) in the future blocks. In fact, in the assignment
described above, we include not only the packets in ADV without 1-to-1 charge but also packets in ADV
with a 1-to-1 charge to a later block. After creating the assignment, if we have a packet q in ALG that
receives a 1-to-1 charge from a packet p in a later block of ADV , we remove p from ADV in that later
block and replace it there by the packets assigned to q (that are guaranteed to be of smaller total size
than p). After these swaps, the 1-to-1 charges together with the assignment form a valid charging that
charges the remaining not swapped packets in ADV in this block together with the removed packets
from the later blocks in ADV to the packets of MAIN (s) in the current block. This charging is now
independent of the other blocks, so we can continue with the next block.
Furthermore, to analyze blocks in which the algorithm does not complete any packet, we shift the
targets of the 1-to-1 charges backward in time by accounting for the first few packets of each size completed
by the adversary at the beginning of the schedule and after Ci0 in the additive constant of the competitive
ratio.
Preliminaries.
We define the notions of blocks and (modified) critical times.
14
Ci0
ALG:
...
ADV:
Figure 4: An illustration of back, up, and forward 1-to-1 charges for size `i . Crossed packets are included
in the set A (and thus contribute to the additive constant).
Definition 3. Let f1 , f2 , . . . , fN be the times of faults. Let f0 = 0 and fN +1 = T is the end of schedule.
Then the time interval (fi , fi+1 ], i = 0, . . . , N , is called a block.
Definition 4. For i = 1, . . . k, the critical time Ci0 is the supremum of i-good times t ∈ [0, T ], where T is
the end of the schedule and i-good times are as defined in Definition 1.
All Ci0 ’s are defined, as t = 0 is i-good for all i. Similarly to Section 4.1, each Ci0 is of one of the
following types: (i) Ci0 starts a phase and a packet larger than `i is started, (ii) Ci0 = 0, (iii) Ci0 = T , or
(iv) just before time Ci0 no packet of size `i is pending but at time Ci0 one or more packets of size `i are
pending; in this case Ci0 is not i-good but only the supremum of i-good times. We observe that in each
case, at time Ci0 the total size of packets p of size `i pending for MAIN (s) and released before Ci0 is less
than `k .
1-to-1 charges and the additive constant. Next we define the set of packets that contribute to the
additive constant.
Definition 5. Let the set A contain for each i = 1, . . . , k:
(i) the first d4`k /`i e packets of size `i completed by the adversary, and
(ii) the first d4`k /`i e packets of size `i completed by the adversary after Ci0 .
If there are not sufficiently many packets of size `i completed by the adversary in (i) or (ii), we take all
the packets in (i) or all the packets completed after Ci0 in (ii), respectively.
For each i, we put into A packets of size `i of total size at most 10`k . Thus we have `(A) = O(k`k )
which implies that packets in A can be counted in the additive constant.
We define 1-to-1 charges for packets of size `i as follows. Let p1 , p2 , . . . , pn be all the packets of size
`i started by the adversary before Ci0 that are not in A. We claim that MAIN (s) completes at least n
packets of size `i before Ci0 . Indeed, before time Ci0 at least n + d4`k /`i e packets of size `i are started by
the adversary and thus released; by the definition of Ci0 at time Ci0 fewer than `k /`i of them are pending
for MAIN (s), one may be running and the remaining ones must be completed. We now charge each pm
to the mth packet of size `i completed by MAIN (s).
We call a 1-to-1 charge starting and ending in the same block an up charge, a 1-to-1 charge from a
block starting at u to a block ending at v 0 ≤ u a back charge, and a 1-to-1 charge from a block ending at
v to a block starting at u0 ≥ v a forward charge. A charged packet is a packet charged by a 1-to-1 charge.
The definition of A implies the following two important properties. See Figure 4 for an illustration.
Lemma 5.2. Let p be a packet of size `i , started by the adversary at time t, charged by a forward charge
to a packet q started by MAIN (s) at time t0 . Then at any time τ ∈ [t − 3`k , t0 ), more than `k /`i packets
of size `i are pending for MAIN (s).
Proof. Let m be the number of packets of size `i that MAIN (s) completes before q. Then, by the
definition of A, the adversary completes m + d4`k /`i e packets of size `i before p. As fewer than 3`k /`i of
these packets are started in (t − 3`k , t], the remaining more than m + `k /`i packets have been released
before or at time t − 3`k . As only m of them are completed by MAIN (s) before t0 , the remaining more
than `k /`i packets are pending at any time τ ∈ [t − 3`k , t0 ).
Lemma 5.3. Let p 6∈ A be a packet of size `i started by the adversary at time t that is not charged. Then
t − 4`k ≥ Ci0 and thus at any τ ≥ t − 4`k , a packet of size `i is pending for MAIN (s).
Proof. Any packet of size `i started before Ci0 + 4`k is either charged or put in A, thus t − 4`k ≥ Ci0 . After
Ci0 , a packet of size `i is pending by the definition of Ci0 .
15
Processing blocks. Initially, let ADV be an optimal (adversary) schedule. First we remove all packets
in A from ADV . Then we process blocks one by one in the order of time. When we process a block, we
modify ADV so that we (i) remove some packets from ADV , so that the total size of removed packets is
at most the total size of packets completed by MAIN (s) in this block, and (ii) reschedule any remaining
packet in ADV in this block to one of the later blocks, so that the schedule of remaining packets remains
feasible. Summing over all intervals, (i) guarantees that MAIN (s) is 1-competitive with an additive
constant `(A).
When we reschedule a packet in ADV , we keep the packet’s 1-to-1 charge (if it has one), however, its
type may change due to rescheduling. Since we are moving packets only to later times only, the release
times are automatically respected. Also it follows that we can apply Lemmata 5.2 and 5.3 even to ADV
after rescheduling.
After processing of a block, there will remain no charges to or from it. For the charges from the
block this is automatic, as there remain no packets to be charged. For the charges to the block, this is
guaranteed as in the process we remove from ADV all the packets in later blocks charged by back charges
to the current block.
From now on, let (u, v] be the current block that we are processing; all previous blocks ending at
v 0 ≤ u are processed. As there are no charges to the previous blocks, any packet scheduled in ADV in
(u, v] is charged by an up charge or a forward charge, or else it is not charged at all. We distinguish two
main cases of the proof, depending on whether MAIN (s) finishes any packet in the current block.
Main case 1: Empty block. The algorithm does not finish any packet in (u, v]. We claim that ADV
does not finish any packet. The processing of the block is then trivial.
For a contradiction, assume that ADV starts a packet p of size `i at time t and completes it. The
packet p cannot be charged by an up charge, as MAIN (s) completes no packet in this block. Thus p is
either charged by a forward charge or not charged. Lemma 5.2 or 5.3 implies that at time t some packet
of size `i is pending for MAIN (s).
Since MAIN does not idle unnecessarily, this means that some packet q of size `j for some j is started
in MAIN (s) at time τ ≤ t and running at t. As MAIN (s) does not complete any packet in (u, v], the
packet q is jammed by the fault at time v. This implies that j > i, as `j > s(v − τ ) ≥ v − t ≥ `i ; we also
have t − τ < `j . Moreover, q is the only packet started by MAIN (s) in this block, thus it starts a phase.
As this phase is started by packet q of size `j > `i , the time τ is i-good and Ci0 ≥ τ . All packets ADV
started before time Ci0 + 4`k /s are charged, as the packets in A are removed from ADV and packets in
ADV are rescheduled only to later times. Packet p is started before v < τ + `j /s < Ci0 + `k /s, thus it is
charged. It follows that p is charged by a forward charge. We now apply Lemma 5.2 again and observe
that it implies that at τ > t − `j there are more than `k /`i packets of size `i pending for MAIN (s). This
is in contradiction with the fact that at τ , MAIN (s) started a phase by q of size `j > `i .
Main case 2: Non-empty block. Otherwise, MAIN (s) completes a packet in the current block (u, v].
Let Q be the set of packets completed by MAIN (s) in (u, v] that do not receive an up charge. Note
that no packet in Q receives a forward charge, as the modified ADV contains no packets before u, so
packets in Q either get a back charge or no charge at all. Let P be the set of packets completed in ADV
in (u, v] that are not charged by an up charge. Note that P includes packets charged by a forward charge
and uncharged packets, as no packets are charged to a previous block.
We first assign packets in P to packets in Q so that for each packet q ∈ Q the total size of packets
assigned to q is at most `(q). Formally, we iteratively define a provisional assignment f : P → Q such
that `(f −1 (q)) ≤ `(q) for each q ∈ Q.
Provisional assignment. We maintain a set O ⊆ Q of occupied packets that we do not use for a
future assignment. Whenever we assign a packet p to q ∈ Q and `(q) − `(f −1 (q)) < `(p), we add q to O.
This rule guarantees that each packet q ∈ O has `(f −1 (q)) > `(q)/2.
We process packets in P in the order of decreasing sizes as follows. We take the largest unassigned
packet p ∈ P of size `(p) (if there are more unassigned packets of size `(p), we take an arbitrary one)
and choose an arbitrary packet q ∈ Q \ O such that `(q) ≥ `(p); we prove in Lemma 5.4 below that
such a q exists. We assign p to q, that is, we set f (p) = q. Furthermore, as described above, if
`(q) − `(f −1 (q)) < `(p), we add q to O. We continue until all packets are assigned.
16
If a packet p is assigned to q and q is not put in O, it follows that `(q) − `(f −1 (q)) ≥ `(p). This implies
that after the next packet p0 is assigned to q, we have `(q) ≥ `(f −1 (q)), as the packets are processed from
the largest one and thus `(p0 ) ≤ `(p). If follows that at the end we obtain a valid provisional assignment.
Lemma 5.4. The assignment process above assigns all packets in P .
Proof. For each size `j we show that all packets of size `j in P are assigned, which is clearly sufficient.
We fix the size `j and define a few quantities.
Let n denote
P the number of packets of size `j in P . Let o denote the total occupied size, defined as
o = `(O) + q∈Q\O `(f −1 (q)) at the time just before we start assigning the packets of size `j . Note that
the rule for adding P
packets to O implies that `(f −1 (Q)) ≥ o/2. Let a denote the current total available
size defined as a = q∈Q\O:`(q)≥`j (`(q) − `(f −1 (q))).
First we claim that it is sufficient to show that a > (2n − 2)`j before we start assigning the packets of
size `j . As long as a > 0, there is a packet q ∈ Q \ O of size at least `j and thus we may assign the next
packet (and, as noted before, actually a ≥ `j , as otherwise q ∈ O). Furthermore, assigning a packet p of
size `j to q decreases a by `j if q is not added to O and by less than 2`j if q is added to O. Altogether,
after assigning the first n − 1 packets, a decreases by less than (2n − 2)`j , thus we still have a > 0, and
we can assign the last packet. The claim follows.
We now split the analysis into two cases, in each we prove that the available space a is sufficiently
large before assigning the packets of size `j .
In the first case, we suppose that a packet of size `j is pending for MAIN (s) at all times in [u, v). Let
z be the total size of packets of size at least `j charged by up charges in this block. The size of packets in
P already assigned is at least `(f −1 (Q)) ≥ o/2 and we have n yet unassigned packets of size `j in P . As
ADV has to schedule all these packets and the packets with up charges in this block, its size satisfies
v − u ≥ `(P ) + z ≥ n`j + o/2 + z. Now consider the schedule of MAIN (s) in this block. By Lemma 2.2,
there is no end of phase in (u, v) and jobs smaller than `j scheduled by MAIN (s) have total size less
than 2`j . All the other completed packets contribute to one of a, o, or z. Using Lemma 2.1, the previous
bound on v − u and s ≥ 4, the total size of completed packets is at least s(v − u)/2 ≥ 2n`j + o + 2z.
Hence a > (2n`j + o + 2z) − 2`j − o − z ≥ (2n − 2)`j , which completes the proof of the lemma in this case.
Otherwise, in the second case, there is a time in [u, v) when no packet of size `j is pending for
MAIN (s). Let τ be the supremum of times τ 0 ∈ [u, v] such that MAIN (s) has no pending packet of size
at least `j at time τ 0 ; if no such τ 0 exists we set τ = u. Let t be the time when the adversary starts the
first packet p of size `j from P .
Since p is charged using a forward charge or p is not charged, we can apply Lemma 5.2 or 5.3, which
implies that packets of size `j are pending for MAIN (s) from time t − 3`k till at least v. By the case
condition, there is a time in [u, v) when no packet of size `j is pending, and this time is thus before
t − 3`k , implying u < t − 3`k . The definition of τ now implies that τ ≤ t − 3`k .
Towards bounding a, we show that (i) MAIN (s) runs a limited amount of small packets after τ and
thus a + o is large, and that (ii) f −1 (Q) contains only packets run by ADV from τ on, and thus o is small.
We claim that the total size of packets smaller than `j completed in MAIN (s) in (τ, v] is less than 3`k .
This claim is similar to Lemma 2.2 and we also argue similarly. Let τ1 < τ2 < . . . < τα be all the ends
of phases in (τ, v) (possibly there is none, then α = 0); also let τ0 = τ . For i = 1, . . . , α, let ri denote
the packet started by MAIN (s) at τi ; note that ri exists since after τ there is a pending packet at any
time in [τ, v] by the definition of τ . First note that any packet started at or after time τα + `k /s has size
at least `j , as such a packet is pending and satisfies the condition in Step (3) of the algorithm. Thus
the total amount of the small packets completed in (τα , v]) is less than `k + `k−1 < 2`k . The claim now
follows for α = 0. Otherwise, as there is no fault in (u, v), at τi , i = 1, . . . , α, Step (4) of the algorithm
is reached and thus no packet of size at most s(τi − τi−1 ) is pending. In particular, this implies that
`(ri ) > s(τi − τi−1 ) for i = 1, . . . , α. This also implies that the amount of the small packets completed
in (τ0 , τ1 ] is less than `k and the claim for α = 1 follows. For α ≥ 2 first note that by Lemma 2.2(i),
s(τi − τi−1 ) ≥ `j for all i = 2, . . . , α and thus ri is not a small packet. Thus for i = 3, . . . , α, the amount
of small packets in (τi−1 , τi ] is at most s(τi−1 − τi ) − `(ri−1 ) < `(ri ) − `(ri−1 ). The amount of small
packets completed in (τ1 , τ2 ] is at most s(τ2 − τ1 ) < `(r2 ) and the amount of small packets completed in
(τα , v] is at most 2`k − `(rα ). Summing this together, the amount of small packets completed in (τ1 , v] is
at most 2`k and the claim follows.
Let z be the total size of packets of size at least `j charged by up charges in this block and completed
by MAIN (s) after τ . After τ , MAIN (s) processes packets of total size more than s(v − τ ) − `k and all of
17
`i
2`i
≥ `j
`j
`i
2`i
≥ `j
`j
ALG:
ADV:
`i
`i
`j
`j
Figure 5: An illustration of the provisional assignment on the left; note that a packet of size `j with a
forward charge is also assigned. Full arcs depict 1-to-1 charges and dashed arcs depict the provisional
assignment. The result of modifying the adversary schedule on the right.
these packets contribute to one of a, o, z, or the volume of less than 3`k of small packets from the claim
above. Thus, using s ≥ 4, we get
a > 4(v − τ ) − o − z − 4`k .
(5.1)
Now we derive two lower bounds on v − τ using ADV schedule.
Observe that no packet contributing to z except for possibly one (the one possibly started by MAIN (s)
before τ ) is started by ADV before τ , as otherwise it would be pending for MAIN (s) just before τ ,
contradicting the definition of τ .
Also, observe that in (u, τ ], ADV runs no packet p ∈ P with `(p) > `j : For a contradiction, assume
that such a p exists. As τ ≤ Cj 0 for any j 0 > j, such a p is charged. As p ∈ P , it is charged by a forward
charge. However, then Lemma 5.2 implies that at all times between the start of p in ADV and v a
packet of size `(p) is pending for MAIN (s); in particular such a packet is pending in the interval before
τ , contradicting the definition of τ .
These two observations imply that in [τ, v], ADV starts and completes all the assigned packets from
P , the n packets of size `j from P , and all packets except possibly one contributing to z. This gives
v − τ ≥ `(f −1 (Q)) + n`j + z − `k ≥ o/2 + n`j + z − `k .
To obtain the second bound, we observe that the n packets of size `j from P are scheduled in [t, v]
and together with t ≥ τ + 3`k we obtain v − τ = v − t + t − τ ≥ n`j + 3`k .
Summing the two bounds on v − τ and multiplying by two we get 4(v − τ ) ≥ 4n`j + 4`k + o + 2z.
Summing with (5.1) we get a > 4n`j + z ≥ 4n`j . This completes the proof of the second case.
As a remark, note that in the previous proof, the first case deals with blocks after Cj , it is the typical
and tight case. The second case deals mainly with the block containing Cj . (The blocks that are before
Cj may fall in either of the cases.) This brings some technical difficulties, but it is a case which has slack.
This is similar to the situation in the local analysis using Master Theorem.
Modifying the adversary schedule. Now all the packets from P are provisionally assigned by f and
for each q ∈ Q we have that `(f −1 (q)) ≤ `(q).
We process all packets q completed by MAIN (s) in (u, v] one by one according to one of the following
three cases; in each case we remove from ADV one or more packets with total size at most `(q).
If q 6∈ Q then the definition of P and Q implies that q is charged by an up charge from some packet
p 6∈ P of the same size. We remove p from ADV .
If q ∈ Q does not receive a charge, we remove f −1 (q) from ADV . We have `(f −1 (q)) ≤ `(q), so the
size is as required. If any packet p ∈ f −1 (q) is charged (necessarily by a forward charge), we remove this
charge.
If q ∈ Q receives a charge, it is a back charge from some packet p of the same size. We remove p from
ADV and in the interval where p was scheduled, we schedule packets from f −1 (q) in an arbitrary order.
As `(f −1 (q)) ≤ `(q), this is feasible. If any packet p ∈ f −1 (q) is charged, we keep its charge to the same
packet in MAIN (s); the charge was necessarily a forward charge, so it leads to some later block. See
Figure 5 for an illustration.
After we have processed all the packets q, we have modified ADV by removing an allowed total size of
packets and rescheduling the remaining packets in (u, v] so that any remaining charges go to later blocks.
This completes processing of the block (u, v] and thus also the proof of 1-competitiveness.
18
τ ≥ t + `/s − 2
t
ALG:
ADV:
`
`
Figure 6: An illustration of Case (D3).
6
Lower Bounds
In this section we study lower bounds on the speed necessary to achieve 1-competitiveness. We start with
a lower bound of 2 which holds even for the divisible case. It follows that our algorithm DIV and the
algorithm in Jurdzinski et al. [7] are optimal. Note that this lower bound follows from results of Anta et
al. [3] by a similar construction, although the packets in their construction are not released together.
Theorem 6.1. There is no 1-competitive deterministic online algorithm running with speed s < 2, even
if packets have sizes only 1 and ` for ` > 2s/(2 − s) and all of them are released at time 0.
Proof. For a contradiction, consider an algorithm ALG running with speed s < 2 that is claimed to be
1-competitive with additive constant A where A may depend on `. At time 0 the adversary releases
N1 = dA/`e + 1 packets of size ` and N0 = d(2/s) · N1 · (s − 1) · `2 + (2/s) · ` · (A + 1)e packets of size 1.
These are all packets in the instance.
The adversary’s strategy works by blocks where a block is a time interval between two faults, where
the first block begins at time 0. The adversary ensures that in each such block ALG completes no packet
of size ` and moreover ADV either completes an `-sized packet or completes more 1’s (packets of size 1)
than ALG.
Let t be the time of the last fault; initially t = 0. Let τ ≥ t be the time when ALG starts the
first `-sized packet after t; we set τ = ∞ if it does not happen. Note that we use here that ALG is
deterministic. In a block beginning at time t, the adversary proceeds according to the first case below
that applies.
(D1) If ADV has less than 2`/s pending packets of size 1, then the end of the schedule is at t.
(D2) If ADV has all packets of size ` completed, then it stops the current process and issues a fault at
times t + 1, t + 2, . . . Between each two consecutive faults after t it completes one packet of size 1
and it continues issuing faults until it has no pending packet of size 1. Then there is the end of the
schedule. Clearly, ALG may complete only packets of size 1 after t as ` > 2s/(2 − s) > s.
(D3) If τ ≥ t + `/s − 2, then the next fault is at time t + `. In the current block, the adversary completes
a packet `. ALG completes at most s · ` packets of size 1 and then it possibly starts ` at τ (if
τ < t + `) which is jammed, since it would be completed at
2`
2
`
−2=t+`+
−1 `−2>t+`
τ + ≥t+
s
s
s
where the last inequality follows from 2s − 1 ` > 2 which is equivalent to ` > 2s/(2 − s). Thus the
`-sized packet would be completed after the fault. See Figure 6 for an illustration.
(D4) Otherwise, if τ < t + `/s − 2, then the next fault is at time τ + `/s − ε for a small enough ε > 0. In
the current block, ADV completes as many packets of size 1 as it can, that is bτ + `/s − ε − tc
packets of size 1; note that by Case (D1), ADV has enough 1’s pending. Again, the algorithm does
not complete the packet of size ` started at τ , because it would be finished at τ + `/s. See Figure 7
for an illustration.
First notice that the process above ends, since in each block the adversary completes a packet. We
now show LADV > LALG + A which contradicts the claimed 1-competitiveness of ALG.
19
t
ALG:
ADV:
τ < t + `/s − 2
`
Gain = 7
Gain = 8
Figure 7: An illustration of Case (D4).
If the adversary’s strategy ends in Case (D2), then ADV has all `’s completed and then it schedules
all 1’s, thus LADV = N1 · ` + N0 > A + N0 . However, ALG does not complete any `-sized packet and
hence LALG ≤ N0 which concludes this case.
Otherwise, the adversary’s strategy ends in Case (D1). We first claim that in a block (t, t0 ] created in
Case (D4), ADV finishes more 1’s than ALG. Indeed, let o be the number of 1’s completed by ALG in
(t, t0 ]. Then τ ≥ t + o/s where τ is from the adversary’s strategy in (t, t0 ], and we also have o < ` − 2s or
equivalently ` > o + 2s, because τ < t + `/s − 2 in Case (D4). The number of 1’s scheduled by ADV is
`
o `
o o + 2s
2
τ + −ε−t ≥ t+ + −ε−t ≥
+
−ε =
o+2−ε
s
s s
s
s
s
2
o+1 ≥o+1
≥
s
and we proved the claim.
Let α be the number of blocks created in Case (D3); note that α ≤ N1 , since in each such block
ADV finishes one `-sized packet. ALG completes at most s` packets of size 1 in such a block, thus
LADV ((u, v]) − LALG ((u, v]) ≥ (1 − s) · ` for a block (u, v] created in Case (D3).
Let β be the number of blocks created in Case (D4). We have
s
2`
s · N0
β>
· N0 −
=
− 1,
2`
s
2`
because in each such block ADV schedules less than 2`/s packets of size 1 and less than 2`/s of these
packets are pending at the end. By the claim above, we have LADV ((u, v]) − LALG ((u, v]) ≥ 1 for a block
(u, v] created in Case (D4).
Summing over all blocks and using the value of N0 we get
LADV − LALG ≥ α · (1 − s) · ` + β
s · N0
−1
2`
≥ N1 · (1 − s) · ` + N1 · (s − 1) · ` + A + 1 − 1 = A
> N1 · (1 − s) · ` +
where we used s ≥ 1 which we may suppose w.l.o.g. This concludes the proof.
Our main lower bound of φ + 1 = φ2 ≈ 2.618 generalizes the construction of Theorem 6.1 for more
packet sizes, which are no longer divisible. Still, we make no use of release times.
Theorem 6.2. There is no 1-competitive deterministic online algorithm running with speed s < φ + 1,
even if all packets are released at time 0.
Proof. The adversary chooses ε > 0 small enough and k ∈ N large enough so that s < φ + 1 − 1/φk−1 .
For convenience, the smallest size in the instance is ε instead of 1. There will be k + 1 packet sizes in the
instance, namely `0 = ε, and `i = φi−1 for i = 1, . . . , k.
Suppose for a contradiction that there is an algorithm ALG running with speed s < φ + 1 that is
claimed to be 1-competitive with additive constant A where A may depend on `i ’s, in particular on ε
and k. The adversary issues Ni packets of size `i at time 0, for i = 0, . . . , k; Ni ’s are chosen so that
N0 N1 · · · Nk . These are all packets in the instance.
20
ALG:
ADV:
ε 1
φ3
φ
φ2
Figure 8: An illustration of Case (B5).
Pk
More precisely, Ni ’s are defined inductively so that it holds Nk > A/`k , Ni > `k φs j=i+1 Nj + A/`i
Pk
for i = k − 1, . . . , 1, and finally N0 > φ`k (A + 1 + sφ`k i=1 Ni )/ε2 .
The adversary’s strategy works by blocks where a block is again a time interval between two faults
and the first block begins at time 0. Let t be the time of the last fault; initially t = 0. Let τi ≥ t be the
time when ALG starts the first packet of size `i after t; we set τi = ∞ if it does not happen. Again, we
use here that ALG is deterministic. Let τ≥i = minj≥i τj be the time when ALG starts the first packet of
size at least `i after t. Let PADV (i) be the total size of `i ’s (packets of size `i ) pending for the adversary
at time t.
In a block beginning at time t, the adversary proceeds according to the first case below that applies.
Each case has an intuitive explanation which we make precise later.
(B1) If there are less than φ`k /ε packets of size ε pending for ADV , then the end of the schedule is at
time t.
Lemma 6.3 below shows that in blocks in which ADV schedules ε’s it completes more than ALG in
terms of total size. It follows that the schedule of ADV has much bigger total size for N0 large
enough, since the adversary scheduled nearly all packets of size ε; see Lemma 6.8.
(B2) If there is i ≥ 1 such that PADV (i) = 0, then ADV stops the current process and continues by
Strategy Finish described below.
(B3) If τ1 < t + `1 /(φ · s), then the next fault occurs at time τ1 + `1 /s − ε, so that ALG does not finish
the first `1 -sized packet. ADV schedules as many ε’s as it can.
In this case, ALG schedules `1 too early and in Lemma 6.3 we show that the total size of packets
completed by ADV is larger than the total size of packets completed by ALG.
(B4) If τ≥2 < t + `2 /(φ · s), then the next fault is at time τ≥2 + `2 /s − ε, so that ALG does not finish
the first packet of size at least `2 . ADV again schedules as many ε’s as it can. Similarly as in the
previous case, ALG starts `2 or a larger packet too early and we show that ADV completes more
in terms of size than ALG, again using Lemma 6.3.
(B5) If there is 1 ≤ i < k such that τ≥i+1 < τi , then we choose the smallest such i and the next fault is
at time t + `i . ADV schedules a packet of size `i . See Figure 8 for an illustration.
Intuitively, this case means that ALG skips `i and schedules `i+1 (or a larger packet) earlier.
Lemma 6.5 shows that the algorithm cannot finish its first packet of size at least `i+1 (thus it also
does not schedule `i ) provided that this case is not triggered for a smaller i, or previous cases are
not triggered.
(B6) Otherwise, the next fault occurs at t + `k and ADV schedules a packet of size `k in this block.
Lemma 6.6 shows that ALG cannot complete an `k -sized packet in this block. See Figure 9 for an
illustration.
We remark that the process above eventually ends either in Case (B1), or in Case (B2), since in each
block ADV schedules a packet. Also note that the length of each block is at most φ`k .
We describe Strategy Finish, started in Case (B2). Let i be the smallest index i0 ≥ 1 such that
PADV (i0 ) = 0. For brevity, we call a packet of size at least `i long, and a packet of size `j with 1 ≤ j < i
short; note that ε’s are not short packets. In a nutshell, ADV tries to schedule all short packets, while
preventing the algorithm from completing any long packet. Similarly to Cases (B3) and (B4), if ALG is
starting a long packet too early, ADV schedules ε’s and gains in terms of total size.
21
ALG:
ADV:
ε 1
φ
φ2
φk
...
φk
Figure 9: An illustration of Case (B6).
Adversary’s Strategy Finish works again by blocks. Let t be the time of the last fault. Let τ ≥ t
be the time when ALG starts the first long packet after t; we set τ = ∞ if it does not happen. The
adversary proceeds according to the first case below that applies:
(F1) If PADV (0) < φ`k , then the end of the schedule is at time t.
(F2) If ADV has no pending short packet, then the strategy Finish ends and ADV issues a fault at times
t + ε, t + 2ε, . . . Between each two consecutive faults after t it completes one packet of size ε and it
continues issuing faults until it has no pending ε. Then there is the end of the schedule. Clearly,
ALG may complete only ε’s after t if ε is small enough. Note that for i = 1 this case is immediately
triggered, as `0 -sized packets are not short, hence there are no short packets whatsoever.
(F3) If τ < t + `i /(φ · s), then the next fault is at time τ + `i /s − ε, so that ALG does not finish the first
long packet. ADV schedules as many ε’s as it can. Note that the length of this block is less than
`i /(φ · s) + `i /s ≤ φ`k . Again, we show that ADV completes more in terms of size using Lemma 6.3.
(F4) Otherwise, τ ≥ t + `i /(φ · s). ADV issues the next fault at time t + `i−1 . Let j be the largest
j 0 < i such that PADV (j 0 ) > 0. ADV schedules a packet of size `j which is completed as j ≤ i − 1.
Lemma 6.7 shows that ALG does not complete the long packet started at τ .
Again, in each block ADV completes a packet, thus Strategy Finish eventually ends. Note that the
length of each block is less than φ`k .
We now prove the lemmata mentioned above. In the following, t is the beginning of the considered
block and t0 is the end of the block, i.e., the next fault after t. Recall that LALG ((t, t0 ]) is the total size
of packets completed by ALG in (t, t0 ]. We start with a general lemma that covers all cases in which
ADV schedules many ε’s.
Lemma 6.3. In Cases (B3), (B4), and (F3) it holds LADV ((t, t0 ]) ≥ LALG ((t, t0 ]) + ε.
Proof. Let i and τ be as in Case (F3); we set i = 1 and τ = τ1 in Case (B3), and i = 2 and τ = τ≥2 in
Case (B4). Note that the first packet of size (at least) `i is started at τ with τ < t + `i /(φ · s) and that
the next fault occurs at time τ + `i /s − ε. Furthermore, PADV (0, t) ≥ φ`k by Cases (B1) and (F1). As
t0 − t ≤ φ`k it follows that ADV has enough ε’s to fill nearly the whole block with them, so in particular
LADV ((t, t0 ]) > t0 − t − ε.
Let a = LALG ((t, t0 ]). Since ALG does not complete the `i -sized packet we have τ ≥ t + a/s and thus
also a < `i /φ as τ < t + `i /(φ · s).
If a < `i /φ − 3sε/φ which is equivalent to `i > φ · a + 3sε, then we show the required inequality by
the following calculation:
LADV ((t, t0 ]) + ε > t0 − t = τ +
`i
a `i
a + φ · a + 3sε
−ε−t≥ + −ε>
− ε > a + 2ε .
s
s
s
s
Otherwise, a is nearly `i /φ and thus large enough. Then we get
LADV ((t, t0 ]) + ε > t0 − t = τ +
`i
a `i
a φa
−ε−t≥ + −ε> +
− ε > a + 2ε
s
s
s
s
s
where the penultimate inequality follows by `i > φa, and the last inequality holds as (1 + φ)a/s > a + 3ε
for ε small enough and a ≥ `i /φ − 3sε/φ.
22
− 1 and Si = Si−1 + `i for i = 1, . . . , k. Thus Si =
Pi For brevity, we inductively define S0 = φ i+1
`
+
φ
−
1
and
a
calculation
shows
S
=
φ
− 1. We prove a useful observation.
i
i
j=1
Lemma 6.4. Fix j ≥ 2. If Case (B3) and Case (B5) for i < j are not triggered in the block, then
τi+1 ≥ t + Si /s for each i < j.
Proof. We have τ1 ≥ t + `1 /(φ · s) = t + (φ − 1)/s by Case (B3) and τi+1 ≥ τi + `i /s for any i < j, since
Case (B5) was not triggered for i < j and the first `i -sized packet needs to be finished before starting the
next packet. Summing the bounds gives the inequalities in the lemma.
Lemma 6.5. In Case (B5) the algorithm does not complete any packet of size `i or larger.
Proof. Recall that we have τ≥i+1 < τi , thus the first packet p of size at least `i has size at least `i+1 . It
suffices to prove
`i+1
τ≥i+1 +
− t > `i ,
(6.1)
s
which means that p would be completed after the next fault at time t + `i .
We start with the case i = 1 in which τ≥2 < τ1 . Since Case (B4) was not triggered, we have
τ≥2 ≥ t + `2 /(φ · s) = t + 1/s. We show (6.1) by the following calculation:
τ≥2 +
`2
1 `2
1+φ
−t≥ +
=
> 1 = `1 ,
s
s
s
s
where the strict inequality holds by s < φ + 1.
Now consider the case i ≥ 2. By the minimality of i satisfying the condition of Case (B5) we
use Lemma 6.4 for j = i to get τi−1 ≥ t + Si−2 /s. Since a packet `i−1 is started at τi−1 , it holds
τ≥i+1 ≥ t + (Si−2 + `i−1 )/s = t + Si−1 /s. Thus
τ≥i+1 +
`i+1
Si−1 + `i+1
φi − 1 + φi
−t≥
=
s
s
s
φi+1 + φi−2 − 1
φi+1
=
≥
≥ φi−1 = `i ,
s
s
where the penultimate inequality holds by i ≥ 2 and the last inequality by s < φ + 1. (We remark that
the penultimate inequality has a significant slack for i > 2.)
Lemma 6.6. In Case (B6) ALG does not complete a packet of size `k .
Proof. It suffices to prove
1
τk > t + 1 −
s
since then ALG completes the first `k -sized packet
`k
>t+ 1−
τk +
s
`k ,
(6.2)
at
1
`k
= t + `k ,
`k +
s
s
thus after the fault.
Since Case (B3) and Case (B5) are not triggered, we use Lemma 6.4 for j = k to get τk ≥
t + Sk−1 /s = t + (φk − 1)/s. Recall that k is large enough so that s < φ + 1 − 1/φk−1 which is equivalent
to (φk − 1)/s > (1 − 1/s)φk−1 . This shows (6.2), because τk − t ≥ (φk − 1)/s and `k = φk−1 .
Lemma 6.7. In Case (F4) ALG does not complete any long packet.
Proof. Recall that the first long packet p is started at τ and it has size of at least `i , thus it would be
completed at least at τ + `i /s. We show τ + `i /s − t > `i−1 by the following calculation:
τ+
`i
`i
`i
φ`i
`i
−t≥
+ =
>
= `i−1 ,
s
φ·s
s
s
φ
where the strict inequality holds by s < φ + 1. This implies that the long packet p would be completed
after the next fault at time t + `i−1 .
23
We are ready to prove that at the end of the schedule LADV > LALG + A holds, which contradicts
the claimed 1-competitiveness of ALG. We inspect all the cases in which the instances may end, starting
with Cases (B1) and (F1). We remark that we use only crude bounds to keep the analysis simple.
Lemma 6.8. If the schedule ends in Case (B1) or (F1), we have LADV > LALG + A.
Proof. Recall that each block (t, t0 ] has length of at most φ`k , thus LALG ((t, t0 ]) ≤ sφ`k and LADV ((t, t0 ]) ≤
φ`k .
We call a block in which ADV schedules many ε’s small, other blocks are big. Note that Cases (B3),
(B4), and (F3) concern small blocks, whereas Cases (B5), (B6), and (F4) concern big blocks.
Using Lemma 6.3, we have that in each small block (t, t0 ] it holds LADV ((t, t0 ]) ≥ LALG ((t, t0 ]) + ε.
Let β be the number of small blocks. We observe that
N0 − φ`εk ε
β≥
,
φ`k
because in each such block ADV schedules at most φ`k /ε packets of size ε and PADV (0) < φ`k at the
end in Cases (B1) and (F1).
Pk
The number of big blocks is at most i=1 Ni , since in each such block ADV schedules a packet of
size at least `1 . For each such block we have LADV ((t, t0 ]) − LALG ((t, t0 ]) ≥ −sφ`k which is only a crude
bound, but it suffices for N0 large enough.
Summing over all blocks we obtain
k
k
N0 − φ`εk ε2
X
X
− sφ`k
Ni
LADV − LALG ≥ βε − sφ`k
Ni ≥
φ`k
i=1
i=1
> A + sφ`k
k
X
Ni − sφ`k
i=1
where (6.3) follows from N0 > φ`k (A + 1 + sφ`k
k
X
Ni = A ,
(6.3)
i=1
Pk
i=1
Ni )/ε2 .
It remains to prove the same for termination by Case (F2).
Lemma 6.9. If Strategy Finish ends in Case (F2), then LADV > LALG + A.
Proof. Note that ADV schedules all short packets and all ε’s, i.e., those of size less than `i . In particular,
we have LADV (< i) ≥ LALG (< i).
Call a block in which ALG completes a packet of size at least `i bad. As the length of any block is
at most φ`k we get that LALG (≥ i, (t, t0 ]) ≤ sφ`k for a bad block (t, t0 ]. Bad blocks are created only in
Cases (B5) and (B6), but in each bad block ADV finishes a packet strictly larger than `i ; note that
Pk
here we use Lemmata 6.5 and 6.6. Hence the number of bad blocks is bounded by j=i+1 Nj . As ADV
completes all `i ’s we obtain
LADV (≥ i) − LALG (≥ i) ≥ `i Ni − `k φs
k
X
Nj
j=i+1
> `i `k φs
k
X
Nj + A − `k φs
j=i+1
k
X
Nj ≥ A ,
j=i+1
Pk
where in the strict inequality we used Nk > A/`k for i = k and Ni > `k φs j=i+1 Nj + A/`i for i < k.
By summing it with LADV (< i) ≥ LALG (< i) we conclude that LADV > LALG + A.
Using Lemmata 6.8 and 6.9 we conclude the proof of the theorem.
24
7
Analysis of DIV
In this section we analyze DIV , giving a simple proof that it matches the performance of the algorithms
from [7] on divisible instances. We use the framework of local analysis from Section 4.
Throughout the section we assume that the packet sizes are divisible. We start by a few observations
on divisibility that lead to a strengthening of Lemma 2.2. Recall that rel(t) = s · (t − tB ) is the relative
time after the start of the current phase tB , scaled by the speed of the algorithm.
Lemma 7.1. (i) If DIV starts or completes a packet of size `i at time t, then `i divides rel(t).
(ii) Let t be a time with rel(t) divisible by `i and rel(t) > 0. If a packet of size `i is pending at time t,
then DIV starts or continues running a packet of size at least `i at time t.
(iii) If at the beginning of phase at time u a packet of size `i is pending and no fault occurs before time
t = u + `i /s, then the phase does not end before t.
Proof. (i) follows trivially from the description of the algorithm.
(ii): If DIV continues running some packet at t, it cannot be a packet smaller than `i by (i) and
the claim follows. If DIV starts a new packet, then a packet of size `i is pending by the assumption.
Furthermore, it satisfies all the conditions from Step 3 of the algorithm, as rel(t) is divisible by `i and
rel(t) ≥ `i (from rel(t) > 0 and divisibility).
(iii): We proceed by induction on i. Assume that no fault happens before t. If the phase starts by
a packet of size at least `i , the claim holds trivially, as the packet is not completed before t. This also
proves the base of the induction for i = 1.
It remains to handle the case when the phase starts by a packet smaller than `i . Let P <i be the set
of all packets of size smaller than `i pending at time u. By the Step (2) of the algorithm, `(P <i ) ≥ `i .
We show that all packets of P <i are completed if no fault happens, which implies that the phase does not
end before t.
Let j be such that `j is the maximum size of a packet in P <i ; note that j exists, as the phase starts by
a packet smaller than `i . By the induction assumption, the phase does not end before time t0 = u + `j /s.
From time t0 on, the conditions in Step (3) guarantee that the remaining packets from P <i are processed
from the largest ones, possibly interleaved with some of the newly arriving packets of larger sizes, as
rel(τ ) for the current time τ is always divisible by the size of the largest pending packet from P <i . This
shows that the phase cannot end before all packets from P <i are completed, if no fault happens.
Now we prove a stronger analogue of Lemma 4.3.
Lemma 7.2. (i) If (u, v] is an initial i-segment then LDIV (s) (≥ i, (u, v]) > s(v − u) − 3`k .
(ii) If (u, v] is a proper i-segment and s(v − u) ≥ `i then LDIV (s) (≥ i, (u, v]) > s(v − u)/2 − `i ;
furthermore LDIV (s) ((u, v]) > s(v − u)/2 and LDIV (s) ((u, v]) is divisible by `i .
Proof. Suppose that time t ∈ [u, v) satisfies that rel(t) divisible by `i and rel(t) > 0. Then observe that
Lemma 7.1(ii) together with the assumption that a packet of size `i is always pending in [u, v) implies
that from time t on only packets of size at least `i are scheduled, and thus the current phase does not
end before v.
For a proper i-segment (u, v], the previous observation for t = u + `i /s immediately implies (ii):
Observe that t ≤ v by the assumption of (ii). Now LDIV (s) (< i, (u, v]) is either equal to 0 (if the phase
starts by a packet of size `i at time u) or equal to `i (if the phase starts by a smaller packet). In both
cases `i divides LDIV (s) (< i, (u, v]) and thus also LDIV (s) ((u, v]). As in the analysis of MAIN , the total
size of completed packets is more than s(v − u)/2 and (ii) follows.
For an initial i-segment (u, v] we first observe that the claim is trivial if s(v − u) ≤ 2`i . So we may
assume that u + 2`i /s ≤ v. Now we distinguish two cases:
1. The phase of u ends at some time u0 ≤ u + `i /s: Then, by Lemma 7.1(iii) and the initial observation,
the phase that immediately follows the one of u does not end in (u0 , v) and from time u0 + `i /s on,
only packets of size at least `i are scheduled. Thus LDIV (s) (< i, (u, v]) ≤ 2`i .
2. The phase of u does not end by time u + `i /s: Thus there exists t ∈ (u, u + `i /s] such that `i
divides rel(t) and also rel(t) > 0 as t > u. Using the initial observation for this t we obtain that the
phase does not end in (u, v) and from time t on only packets of size at least `i are scheduled. Thus
LDIV (s) (< i, (u, v]) ≤ `i .
25
In both cases LDIV (s) (< i, (u, v]) ≤ 2`i , furthermore only a single packet is possibly unfinished at time v.
Thus LDIV (s) (≥ i, (u, v]) > s(v − u) − 2`i − `k and (i) follows.
Theorem 7.3. Let the packet sizes be divisible. Then DIV (1) is 2-competitive. Also, for any speed s ≥ 2,
DIV (s) is 1-competitive.
Proof. Lemma 7.2(i) implies (4.2). We now prove (4.1) for any proper i-segment (u, v] with v − u ≥ `i
and appropriate R. The theorem then follows by the Master Theorem.
Since u is a time of a fault, we have LOPT (≥ i, (u, v]) ≤ v − u. If LOPT (≥ i, (u, v]) = 0, (4.1) is trivial.
Otherwise LOPT (≥ i, (u, v]) ≥ `i , thus v − u ≥ `i ≥ `i /s and the assumption of Lemma 7.2(ii) holds.
For s ≥ 2, Lemma 7.2(ii) implies
LDIV (s) (≥ i, (u, v]) > s(v − u)/2 − `i ≥ v − u − `i ≥ LOPT (≥ i, (u, v]) − `i .
Since both LDIV (s) (≥ i, (u, v]) and LOPT (≥ i, (u, v]) are divisible by `i , this implies LDIV (s) (≥ i, (u, v]) ≥
LOPT (≥ i, (u, v]), i.e., (4.1) holds for R = 1.
For s = 1, Lemma 7.2(ii) implies
LDIV ((u, v]) + LDIV (≥ i, (u, v]) > (v − u)/2 + (v − u)/2 − `i
≥ v − u − `i ≥ LOPT (≥ i, (u, v]) − `i .
The divisibility of LDIV ((u, v]), LDIV (≥ i, (u, v]), and LOPT (≥ i, (u, v]) by `i implies LDIV ((u, v]) +
LDIV (≥ i, (u, v]) ≥ LOPT (≥ i, (u, v]), i.e., (4.1) holds for R = 1.
References
[1] Antonio Fernández Anta, Chryssis Georgiou, Dariusz R. Kowalski, Joerg Widmer, and Elli Zavou.
Measuring the impact of adversarial errors on packet scheduling strategies. J. Scheduling, 19(2):135–
152, 2016. Also appeared in Proc. of SIROCCO 2013: 261–273. URL: http://dx.doi.org/10.
1007/s10951-015-0451-z.
[2] Antonio Fernández Anta, Chryssis Georgiou, Dariusz R. Kowalski, and Elli Zavou. Competitive analysis of task scheduling algorithms on a fault-prone machine and the impact of resource augmentation. In Proc. of the 2nd Int. Workshop on Adaptive Resource Management
and Scheduling for Cloud Computing (ARMS-CC@PODC), LNCS 9438, pages 1–16, 2015. URL:
http://dx.doi.org/10.1007/978-3-319-28448-4_1.
[3] Antonio Fernández Anta, Chryssis Georgiou, Dariusz R. Kowalski, and Elli Zavou. Online parallel
scheduling of non-uniform tasks: Trading failures for energy. Theor. Comput. Sci., 590:129–146,
2015. Also appeared in Proc. of FCT 2013: 145-158. URL: http://dx.doi.org/10.1016/j.tcs.
2015.01.027.
[4] Marek Chrobak, Leah Epstein, John Noga, Jiřı́ Sgall, Rob van Stee, Tomáš Tichý, and Nodari
Vakhania. Preemptive scheduling in overloaded systems. J. Comput. Syst. Sci., 67:183–197, 2003.
URL: http://dx.doi.org/10.1016/S0022-0000(03)00070-9.
[5] P. Garncarek, T. Jurdziński, and K. Loryś. Fault-tolerant online packet scheduling on parallel
channels. In 2017 IEEE International Parallel and Distributed Processing Symposium (IPDPS),
pages 347–356, May 2017. doi:10.1109/IPDPS.2017.105.
[6] Chryssis Georgiou and Dariusz R. Kowalski. On the competitiveness of scheduling dynamically
injected tasks on processes prone to crashes and restarts. J. Parallel Distrib. Comput., 84:94–107,
2015. URL: http://dx.doi.org/10.1016/j.jpdc.2015.07.007.
[7] Tomasz Jurdzinski, Dariusz R. Kowalski, and Krzysztof Loryś. Online packet scheduling under
adversarial jamming. In Proc. of the 12th Workshop on Approximation and Online Algorithms
(WAOA), LNCS 8952, pages 193–206, 2014. See http://arxiv.org/abs/1310.4935 for missing
proofs. URL: http://dx.doi.org/10.1007/978-3-319-18263-6_17.
26
[8] Bala Kalyanasundaram and Kirk Pruhs. Speed is as powerful as clairvoyance. J. of the ACM,
47(4):617–643, 2000. Also appeared in Proc. of the 36nd IEEE Symp. on Foundations of Comput.
Sci. (FOCS), pages 214–221, 1995. URL: http://doi.acm.org/10.1145/347476.347479.
[9] Chiu-Yuen Koo, Tak Wah Lam, Tsuen-Wan Ngan, Kunihiko Sadakane, and Kar-Keung To. On-line
scheduling with tight deadlines. Theor. Comput. Sci., 295:251–261, 2003. URL: http://dx.doi.
org/10.1016/S0304-3975(02)00407-3.
[10] Tak Wah Lam, Tsuen-Wan Ngan, and Kar-Keung To. Performance guarantee for EDF under overload.
J. Algorithms, 52:193–206, 2004. URL: http://dx.doi.org/10.1016/j.jalgor.2003.10.004.
[11] Tak Wah Lam and Kar-Keung To. Trade-offs between speed and processor in hard-deadline scheduling.
In Proc. of the 10th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 623–632.
ACM/SIAM, 1999. URL: http://dl.acm.org/citation.cfm?id=314500.314884.
[12] Cynthia A. Phillips, Clifford Stein, Eric Torng, and Joel Wein. Optimal time-critical scheduling
via resource augmentation. Algorithmica, 32:163–200, 2002. URL: http://dx.doi.org/10.1007/
s00453-001-0068-9.
[13] Kevin Schewior. Deadline Scheduling and Convex-Body Chasing. PhD dissertation, TU Berlin, 2016.
URL: http://dx.doi.org/10.14279/depositonce-5427.
27
| 8cs.DS
|
Isotropic Dynamic Hierarchical Clustering
Victor Sadikov
Oliver Rutishauser
at&t
200 S Laurel Av. A5-2D20
Middletown, NJ 07748
(732) 420-7453
Atlantic Style
P.O.Box 9
Oakhurst, NJ 07755
(732) 455-2081
[email protected]
[email protected]
ABSTRACT
We face a business need of discovering a pattern in
locations of a great number of points in a high-dimensional
space. We assume that there should be a certain structure,
so that in some locations the points are close while in other
locations the points are more dispersed. Our goal is to
group the close points together. The process of grouping
close objects is known under the name of clustering.
1. We are particularly interested in a hierarchical
structure. A plain structure may reduce the number of
objects, but the data are still difficult to manage or present.
2. The classical technique suited for the task at hand is a
B-Tree. The key properties of the B-Tree are that it is
hierarchical and balanced, and it can be dynamically
constructed from the input data. In these terms, B-Tree has
certain advantages over other clustering algorithms, where
the number of clusters needs to be defined a priori. The BTree approach allows to hope that the structure of input
data will be well determine without any supervised
learning.
3. The space is Euclidean and isotropic. This is the most
challenging part of the project, because currently there are
no B-Tree implementations processing indices in a
symmetrical and isotropical way. Some known
implementations are based on constructing compound
asymmetrical indices from point coordinates, where the
main index works as a key, while the function of other
(999!) indices is lost; and the other known implementations
split the nodes along the coordinate hyper-planes,
sacrificing the isotropy of the original space. In the latter
case the clusters become coordinate parallelepipeds, which
is a rather artificial and unnecessary assumption. Our
implementation of a B-Tree for a high-dimensional space is
based directly on concepts of factor analysis.
4. We need to process a great deal of data, something
like tens of millions of points in a thousand-dimensional
space. The application has to be scalable, even though,
technically, out task is not considered a true Big Data
problem. We use dispersed data structures, and optimized
algorithms. Ideally, a cluster should be an ellipsoid in a
high-dimensional space, but such implementation would
require to store O(n2) ellipse axes, which is impractical. So,
we are using multi-dimensional balls defined by the centers
and radii. On the other hand, calculation of statistical
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that
copies bear this notice and the full citation on the first page. To copy
otherwise, or republish, to post on servers or to redistribute to lists,
requires prior specific permission and/or a fee.
The only reasonable numbers
are zero, one, and infinity.
Willem Louis van der Poel
values like the mean and the average deviation, can be done
in an incremental way. This mean that when adding a
point to a tree, the statistical values for nodes of all levels
may be recalculated in O(1) time. The node statistical
values are used to split the overloaded nodes in an optimal
way. We support both, brute force O(2n) and greedy O(n2)
split algorithms. Statistical and aggregated node
information also allows to manipulate (to search, to delete)
aggregated sets of closely located points.
5. Hierarchical information retrieval. When searching,
the user is provided with the highest appropriate nodes in
the tree hierarchy, with the most important clusters
emerging in the hierarchy automatically. Then, if
interested, the user may navigate down the tree to more
specific points.
The system is implemented as a library of Java classes
representing Points in multi-dimensional space, Sets of
points with aggregated statistical information (mean,
standard deviation,) B-tree, and Nodes with a support of
serialization and storage in a MySQL database.
CCS Concepts
➝
• Theory of computation
Theory and algorithms for application
domains
Machine learning theory
Unsupervised learning
Mathematical
and clustering • Mathematics of computing
software
Statistical software • Information systems
Information retrieval
Retrieval tasks and goals
Clustering
and classification
➝
➝
➝
➝
➝
➝
➝
Keywords
Clustering; Hierarchical clustering; Dynamic clustering; Isotropic
clustering; Multi-dimensional space; B-tree; Factor analysis.
1. ( Points, Implementation )
In a high-dimensional space we assume that a considerable
number of coordinates will contain zero values. To
optimize the memory and storage space we would like to
keep non-zero coordinates only. Thus, a Point object
contains 3 fields: the number of non-zero coordinates, the
array of sorted coordinate indices, and the array of
corresponding coordinate values.
class Point
{
int N = 0;
int[] key;
float[] val;
}
// # of non-zero coordinates
// array of coordinate indices
// array of coordinate values
The Point class provides methods for calculating the
Euclidean length of the point vector, getting a particular
coordinate (or zero,) adding another point to the given
point, calculating a few useful functions like distance to
another point, dot product, and serializing to a base-64
string.
Some functions, e.g. adding a point, may change the
number of non-zero coordinates. The main loop for adding
two points looks like the following.
int Ix = 0; int Ip = 0; int Iz = 0;
for(; Ix != this.N && Ip != p.N ;)
{
if(this.key[Ix] == p.key[Ip])
{
z.key[Iz] = this.key[Ix];
z.val[Iz] = this.val[Ix] + p.val[Ip];
Ix++; Ip++; Iz++;
}
else if(this.key[Ix] < p.key[Ip])
{
z.key[Iz] = this.key[Ix];
z.val[Iz] = this.val[Ix];
Ix++; Iz++;
}
else
{
z.key[Iz] = p.key[Ip];
z.val[Iz] = p.val[Ip];
Ip++; Iz++;
}
}
2. ( Sets, Mathematics )
The next step of our approach is the introduction of Sets of
Points. The Sets allow calculation of aggregated statistical
values. The most important value is the number of points
(N) in the Set. It needs to be corrected each time a new
point is added to the Set. The obvious way of calculating
the new number of points is to increment the current
number by one.
N = N + 1;
Other important statistical values of the set of points are
arithmetic mean and the standard deviation. These values
should also be adjusted every time a point is added to the
Set. We could recalculate the arithmetical mean (M) from
scratch, but we would like to follow the incremental
approach and move it towards the newly added point (P) by
the 1/N of the distance.
M = M + (P - M) /N;
As for standard deviation, at first sight, it seems to be a
value that requires the full recalculation. Fortunately, this is
not the case. We can store and adjust the standard
deviation in an incremental way too, based on the following
formula.
E[X – E(X)]2 = E[X2] – (E[X]) 2
This means that to calculate the standard deviation it is
enough to store the sum of the squares of point coordinates
(S,) which can be adjusted incrementally.
S = S + |P|2;
And when we need to calculate the standard deviation we
will do the following.
D = sqrt( S/N - |M|2 );
3. ( Clustering, Example )
Clustering basically means grouping similar objects
together. If the objects have a number of numerical
attributes they may be represented as points in a multidimensional space. The clustering will mean to partition
the whole set of points into a number of disjoint sub-sets.
Let's consider an example in a one-dimensional space. The
example is free of the challenges related to multidimensional clustering and is easy to comprehend. Assume
we are given the set of five numbers { 0. 4. 5. 9. 13.}
Figure 1
We can see that initially the points occupy the segment [0,
13.] Assume we need to split the set of the given points on
the number line into two clusters then the resulting clusters
will be sub-segments. Even in this simple example two
different splits are possible. The first splits makes
segments [0, 5] and [9, 13] while the other makes segments
[0, 4] and [5, 13] (see Figure 1.)
In terms of segments, the first split looks better, because it
finds two compact segments rich of points, while the
second split covers almost the whole initial segment. In
terms of statistical variables, the first split is better too,
because the sum of the deviations of the resulting sets is
minimal.
4. ( Dynamic Clustering, Approach )
If we continue adding new points, then we need to decide
which sub-segment each new point belongs to. Points
below 5 we can add to the first sub-segment, and points
above 9 we add to the second sub-segment.
We may place points between 5 and 9 into either subsegment. Our criterion here is to avoid big segments, or
(which is the same) to keep the sum of deviations minimal.
But we always need to update the boundaries of the subsegments so that we can exactly know in which segment a
particular point is to be found. If the points are well spread,
knowing exact boundaries of the segments may also
improve unsuccessful search.
After adding a certain number of points to a sub-segment,
we will need to split this sub-segment into two subsegments with a smaller number of points. This can be
done exactly in the same manner as we split the initial
segment. After that we will be adding new points to the set
of three sub-segments. Then we will need to split another
sub-segment. And the number of sub-segments will
increase again.
Unlike the static approach where the set of all objects exists
before the procedure of clustering starts; dynamic approach
assumes that clusters are incrementally adjusted each time a
new object gets added into the set. This eliminates the
dedicated step of clustering for the price of a longer time
needed to include objects. Dynamic clustering provides
better flexibility and can be performed with less a priori
known information about the data, e.g. when the total
number of target clusters is unknown.
5. ( Hierarchical Clustering, Approach )
The bigger the number of points in our data set, the bigger
the number segments. Soon it gets big enough, and we
may need to introduce the next level of hierarchy, when
smaller clusters are, in their turn, grouped into clusters of
the higher level. The approach of adding points, splitting
segments, and adding new levels when necessary is quite
similar to adding objects to a B-tree [1.] The tree starts
with the root node, responsible for all points stored in the
tree. At each level, the parent node consist of a number of
sub-nodes. The points in each sub-node are close one to
another.
A classical one-dimensional B-tree design focuses on
minimization of the information stored at the node level.
Namely, a parent node stores a number of adjacent values
in ascending order, with the sub-nodes being placed
between adjacent values, plus one at the beginning and one
at each end. Thus we know that the elements of each subtree are greater than the left adjacent value and less than the
right one.
On the contrary, our design prefers to store excessive
boundary and statistical information about the sub-nodes.
In the case of one-dimensional space, each node occupies a
segment and sub-nodes of one parent do not intersect. The
boundary segment can be defined by its two ends, or by the
middle point (C) and the distance (R) to the ends. The
letter way will occupy less memory in a general multidimensional case. We also keep the number of points in
each sub-tree, their arithmetic mean and standard deviation.
As we stated above, statistical values facilitate splitting the
nodes in the optimal way, while the boundary information
allows us in some cases to terminate an unsuccessful search
heuristically.
6. ( Isotropic B-tree, Approach )
B-trees do their work great, as long as the attributes of the
objects are one-dimensional. Unfortunately, there are some
problem with direct extension of B-tree to a multidimensional case. The two common approaches are the
following.
The mix approach assumes composing the compound index
based on the component indices, and then making use of a
one-dimensional B-tree. The problem with this approach is
that the component indices are treated by far not equally.
One index plays the main role, while the role of the others
is insignificant.
The other approach is more complicated. R-tree [2] is a
variant of B-tree where the nodes are bound by coordinate
rectangles. This approach is more symmetrical in terms of
using indices. But the directions along the coordinate axes
are still different from arbitrary ones.
We would like to build a variant of B-tree, where the nodes
are bound with circles or ellipses. This decision ensures
that the essential property of isotropy of the physical space
is not ignored.
7. ( Isotropic Dynamic Hierarchical
Clustering, Two-Dimensional Case )
The circles are defined by the center point (C) and the
radius (R.) It the 2-dimensional space, the center is defined
by the two coordinates, and so we need to store 3 real
values. Alternatively, if we decided to present clusters as
ellipses in general orientation, we will need to store the
semi-principal axes and the angles, which would require
O(n2) memory, with n being the number of dimensions in
the space.
In the picture to the right, the
circles corresponding to the
sub-nodes of one and the same
parent do not intersect. The
biggest circle (P) corresponds
to the parent node, while C1,
C2, and C3 correspond to the
sub-nodes. Keeping on adding
new points to the tree we
cannot avoid the situation where the sub-nodes start to
intersect. We will dwell on the intersecting areas later in
this paper.
7.1 Selecting a Sub-Node
As mentioned in section 4, while adding a new point
outside any bounding circle, we need to select the most
appropriate sub-node. To do this, we will try to add the
new point to each sub-node, and calculate the new
bounding radii. Then we will select a sub-node so that the
sum of the new squared radii should be minimal, because
our goal is to make eventually all sub-node circles of about
the same absolute size in the given space.
Let’s assume that the old radius was Ri, then the new radius
will be Ri+Hi/2, where Hi is the distance from the new point
to the circle Ci. If we select the i-th sub-node, the sum of
new squared radii will grow by (Ri+Hi/2)2 – Ri2, i.e. by
RiHi+Hi2/4, so we need to select the sub-node where
Ti= 4RiHi+Hi2 is minimal.
7.2 Splitting the Overloaded Node
When a node gets overloaded, i.e. it includes more than the
maximum number of sub-nodes or points; the node needs
to be split into two nodes at the same level. In the classical
B-tree the node is split into two nodes with the equal
number of elements.
In our case, we can split the node taking into account the
following criteria:
• the maximum radius of the two new circles is minimal;
• the new nodes intersect with the minimal area;
• the sum of standard deviations of the new nodes is
minimal.
To this goal, first we will find
the pair of points at the
longest distance.
In the
figure to the right, such a pair
consists of points a and b. If
these points both belonged to
one and the same bounding
circle, the radius of this circle would be greater than the
distance Da,b/2, which, as we assume, is the maximum
distance. So, to minimize the maximum radius we need to
distribute points a and b to the different bounding circles Ca
and Cb. Now we will find point c, so that the distance from
c to either of circles Ca and Cb is the longest. In the picture
above, it is the distance between point c and circle Ca
(which now consists of just one point a.) Once again, to
minimize the would-be radii, we need to distribute point c
to the other circle, Cb. Continuing in the same manner, we
will ultimately get circles Ca and Cb, as shown in the
picture.
7.3 Adjusting Bounding Circles
Let’s assume that we need to add a new point (Q) to the
tree. We start from the root and go down the tree, level by
level. At each level we need to select the most appropriate
sub-node. E.g. at the parent level P, we need to select one
of the sub-nodes, C1, C2, or C3. Logically, there are two
different cases. If the new point belongs to a particular
bounding circle, no adjustment is needed. But if the new
point is outside of any bounding circle, we need to select
the most appropriate sub-node; to add the new point to the
selected sub-node; and to adjust the corresponding
bounding circle, so that it should include the new point as
well as all old points. See the figure below.
points.
In the 2-dimensional case, the
new minimal bounding circle can
be exactly calculated. In the
figure above, the minimal circle
is the circle circumscribed about
the triangle of the first two points
and point Q. It can be calculated
from the coordinates of these
E.g. the radius of the circumscribed circle is
L1*L2*L3/sqrt( (L1+L2+L3)*(L2+L3-L1)*(L1-L2+L3)*(L1+L2-L3) ),
where L1, L2 and L3 are the lengths of the sides.
In a multi-dimensional space, the situation is much more
complicated. Fortunately, we do not need the exact
minimal bounding circles. The bounding circles are very
useful in many procedures where they heuristically allow to
reduce the amount of calculation, but fortunately they are
not critical. So, we would recommend to use a less exact
but easier to calculate approximation.
We can easily construct the new
bounding circle (C’) about the old
bounding circle (C) and the newly
added point (Q) as shown in the
figure to the right. First, we calculate
the distance (H) from the point Q to
the circle C. Then we move the center of the circle C
towards Q by H/2. This will be the center of C’. The
radius of the new circle will be the old radius (R) plus H/2.
Thus, the circle C’ will surely contain all old points as well
as the new point.
Moreover, the new bounding circle
(C’) can be easily optimized. By
construction, the new circle has to
lie through the point Q, but it
doesn’t probably lie through any
other actual points. So, we can
shrink the circle, so that it lies
through yet another point.
Let’s shift the center (O) of the circle towards the point Q,
so that the new circle with the center Oa lies both through
points Q and A. See the picture above. The value of the
shift can be calculated based on the points Q, O and A.
When we shift the center for the point A, this doesn’t
necessarily mean that any other point B will belong to the
new shrunken circle. But we can repeat this procedure for
all points of the set, and find the largest shrunken circle,
corresponding to the shortest allowed shift. That circle will
work for all the points of the set.
7.4 Calculation of a Quasi-Minimal Bounding
Circle
Let’s assume that point T is the foot
of the perpendicular dropped from
point A on the line (Q,O.)
Then |A T|2 + |T Q|2 = |A Q|2. and
|A T|2 + |T Oa|2 = |A Oa|2.
But |T Oa| = |T Q| - |Oa Q| and
|Oa Q| = |Oa A|. So, finally,
|Oa Q| should be |A Q|2 / 2 |T Q|,
where |T Q| is the projection of the
vector [Q A] on line (Q,O) and can be calculated by means
of the dot product.
7.5 Exact Minimal Bounding Circle
Please notice that, in some cases, the exact minimal bounding
circle may be slightly smaller than the quasi-minimal circle
constructed above. Moreover, the shrunken circle (C’) depends
on the initial circle C.
If the set consists of just one point, the exact minimal circle is the
circle with the center in that point and the radius of zero. If the set
contains two points, the exact minimal circle is the circle from the
center of mass and the radius of half the distance between the
points.
If the set contains three points, there are two cases. Either, the
exact minimal circle is the circle circumscribed around these three
points. Or, the circle built on the two of the three points as a
diameter, provided that the third point lies inside it. If we want to
build the circle for the second case, we need first to find the two
(out of the three) points with the longest distance between them.
Then, we need to check that the third point (X) will make an
obtuse-angled triangle. I.e. |A X|2 + |X B|2 < |A B|2. Sets of more
than three points are quite similar.
So, to find whether point x belongs to the circle C1, we need to
calculate |O1 - x|2 - |O2 - x|2 - R12 + R22, and to compare it with
zero.
If we do not want to build the exact minimal circle, we may use
the shrinking technique. But we will need a good first
approximation for the minimal circle. As shown above, the circle
build on the longest segment plays an essential role in
construction of minimal circles; besides, and it is easy to
calculate. This makes it a good first approximation for building
quasi-minimal bounding areas.
While splitting the node, we expect that the
results nodes will have bounding areas with
smaller radii. Let’s assumes that a and b
are the points with the greatest distance
between them, as in the picture above. If we ultimately put these
points into one and the same node, the radius of the bounding
circle for that node can not be less than |a b|, which is almost the
radius of the bounding circle for the original node. To avoid this,
we have to put a and b into different nodes. Now we have nodes
Ca and Cb, consists of points a and b, correspondingly.
8. ( Multi-Dimensional S-Tree, Full Details)
So far we have clearly described what we would like to achieve in
our multi-dimensional S-Tree. A real implementation is not so
smooth, and requires solutions to a number of complicated issues.
8.1 Overlapped Circle
As mentioned in Section 7, some sub-nodes of a given node may
overlap. Such cases may occur when a new point is added to a
sub-node, which results in adjusting the bounding circle of this
sub-node, or when a sub-node is split into two sub-nodes at the
same level, as discussed in Section 7.2. In this case the actual
areas of sub-nodes are not circle, but they are rather “cut” circles,
as shown in the picture below.
The reason for the sub-node areas to be
nothing but the “cut” circles is that the
areas need to be convex. Now, the more
complicated form of the areas makes us
change the way we calculate the
appropriate sub-node when the given
point x belongs to both bounding circles.
In the two-dimensional space we use line L, to determine where
point x belongs to. All points at one side of line L belong to one
sub-node, while all the points at the other side belong to the other
sub-node. Analogically, in the multi-dimensional space, the
border line L will become a plane. All points at one side of the
plane will belong to one sub-node, and all points at the other side
of the plane L will belong to the other sub-node.
8.2 Splitting the Node
When a node gets overloaded, i.e. it includes more than the
maximum number of sub-nodes or points;
the node needs to be split into two nodes at
the same level.
At the next step let’s consider, say, the point c. We can either put
it into Ca or Cb. And we need to estimate how good or bad it
would be to put it into a particular node. E.g. we can try to
optimize (to keep minimal) the maximum radius of Ca and Cb.
But, now we know that there is something more unpleasant than
just big radii; it is overlapped circles. So we may want to keep the
circles overlapped in the minimal possible measure.
In the previous section we have introduced the expression
L2 – (R2 – r2). It shows in what manner the circles are overlapped,
and should be greater than zero. We will try to optimize (to keep
maximal) this expression. It gives the same rules for selecting
points a, b, and c, but gives better results at next steps.
9. ( Example, Two-Dimensional Case )
The most challenging task of implementing a B-tree is a split of
overloaded nodes. When a node is split into two nodes, there
appears a new boundary. All sub-nodes of the given node may
need to be recursively split by the new boundary. The picture
below illustrates the result of a split for a two-dimensional case.
Every plane in the multi-dimensional space can be defined by its
normal vector and the distance from the point of origin. The only
drawback is that theoretically we will need a border plane for each
pair of sub-nodes. E.g. if a parent node has, say, 5 sub-nodes, we
will need to store 10 border plains. A more memory-effective
way would be to calculate the equation of the plane L on the fly,
based on the bounding circles.
Namely, as we discussed, the circle C1 can be defined by its
center, O1, and the radius R1. Analogically, we define the circle
C2 by the center O2 and the radius R2. It is easy to see that the
border plane L is the set of all the point such as the difference
between their squared distances to points O1 and O2 is constant.
|O1 - x|2 - |O2 - x|2 = F1,2
To be precise, the constant F1,2 is the difference between the
squared radii of the circles in question, i.e. F1,2 = R12 – R22
Here the new boundary is highlighted in red. The new sub-trees
are painted olive and violet. The higher levels are darker than
lower levels, so that it would be easier to trace the centers of
clusters.
10. ( Illustration, Three-Dimensional Case )
As we discussed in Section 8.1, ideally, the clusters of a particular
level should form isotropic ellipsoids. The result we would like to
get should look like soap foam bubbles.
Now we can define a target point in a 1024-dimensional space
where each dimension corresponds to a meaning. The coordinates
of the target point are calculated by accumulating all weights
corresponding to particular meanings. It also seems reasonable to
divide the coordinates by the number of the words in the text, so
that repletion of sentences or words does not affect the meaning of
the text.
12. CONCLUSIONS
Arbitrary points in multi-dimensional space can be isotropically
clustered into a balanced hierarchical structure, similar to a B-tree.
Clustering into a multi-dimensional B-tree does not require any
supervision or any a priori given information, like the number of
clusters.
Clustering into a multi-dimensional tree can be done dynamically
and efficiently. Adding new points to the tree requires only
incremental updates of statistical values associated with nodes.
Text pages can be mapped to points into 1000-dimentional
semantic space.
The search of pages close to a given semantic point can return a
hierarchically ordered results, allowing the user to select more
general or more specific topics.
11. ( Example, Multi-Dimensional Case )
As an example we will cluster wiki pages.
1. The first step is to parse an HTML page and to extract pure
text. There are several tools available for this purpose. We use
JavaCC, an open source parser and lexical analyzer generator.
The generator accepts a formal grammar definition, written in .JJ
file, which also allows to define additional custom code. The
input to the lexical analyzer is a sequence of characters; the output
is a sequence of tokens.
In our example, we are interested in skipping HTML tags and
parsing the text further, so that it become a list of words. At this
stage we are dropping everything what is not a word, i.e. numbers,
email addresses, references to web pages, expressions, identifiers,
etc.
2. Then, we analyze the text and map it to a point in a semantic
space. For each word in text we will find the root. Basically, for
verbs we drop endings as -s, -ed, -ing; for nouns we drop ending –
s. Actually, the procedure is a bit more complicated due to
language exceptions. Secondly, we calculate the weight of the
word. We assume that more frequent words should have a lighter
weight than infrequent words. So, we distributed all the words to
256 sets with about equal frequencies. Thirdly, we find the
meaning of the word in question. We have split all words to 1024
groups with similar meanings. Please notice, that one word can
have more than one meanings, with only one of them being
actualized in the text. Without knowing what the actual meaning
is, we have to add all meaning with the same weights depending
on the frequency of the word.
13. ACKNOWLEDGMENTS
This research was not sponsored by National Science Foundation
or any other financial source.
14. REFERENCES
[1] Donald E. Knuth, 1973, Sorting and Searching, volume 3,
The Art of Computer Programming, Addison-Wesley.
[2] Antonin Guttman, 1984, R-trees: a dynamic index structure
for spatial searching, ACM, New York, USA. (1984-r-treeguttman.pdf)
[3] Sean Owen et al. 2011, Mahout in Action, Manning
Publications Co., New York, USA. (ISBN 9781935182689,
Mahout.in.Action.pdf)
[4] Richard A. Reyment, K. G. Joereskog, 1993, Applied Factor
Analysis in the Natural Sciences, Cambridge University
Press, UK.
[5] George A. Miller, 2003, WordNet Lexical Database of
English Language, Cognitive Science Laboratory of
Princeton University
[6] Roget's Thesaurus, 2006, Electronic Lexical Knowledge
Base (ELKB) http://www.nzdl.org/ELKB
[7] Adam Kilgarriff, 1995, BNC Database and Word Frequency
Lists, http://www.kilgarriff.co.uk/bnc-readme.html
[8] Brian Goetz's, 2003, HTML Parser.
| 8cs.DS
|
arXiv:1211.6555v1 [math.AC] 28 Nov 2012
Vertex covers and sensor networks
MAURIZIO IMBESI
Faculty of Sciences, Department of Mathematics,
University of Messina
Viale Ferdinando Stagno d’Alcontres, 31 − I 98166 Messina, Italy
e-mail: [email protected]
MONICA LA BARBIERA
Faculty of Sciences, Department of Mathematics,
University of Messina
Viale Ferdinando Stagno d’Alcontres, 31 − I 98166 Messina, Italy
e-mail: [email protected]
Abstract
We consider algebraic developments of graph theory through suitable applications in real
connection problems. We investigate ideals of vertex covers for the edge ideals associated
to a significative class of connected graphs. It is shown that algebraic procedures linked to
minimal vertex covering of such graphs are good instruments concerned about how sensor
networks can be encoded and their properties measured. Moreover algebraic properties of
such ideals are studied. Using the notion of linear quotients, standard algebraic invariants
are computed and these ideals are stated to be Cohen-Macaulay.
AMS 2010 Subject Classification: 05C25, 05E40, 13C15.
Key words and phrases: Ideals of vertex covers. Sensor network. Linear quotients and
standard invariants.
Introduction
In this paper we are interested in studying standard properties of monomial
ideals arising from a class of graphs in order to explain how algebraic models,
built through graph theory, are useful for setting suitable solutions of several
problems concerning network areas (infrastructure nets, circuits electrical
workers, urban and territorial analysis, wireless sensor networks, etc...).
In the last years, Commutative and Combinatorial Algebra have paid particular attention to large classes of monomial ideals derived from graphs in
order to describe algebraic properties of them. Such ideals turn out to be
very interesting from a combinatorial point of view too. Classical topics include graph algebraic theory, in particular the edge ideals, namely monomial
ideals generated in degree two which descend from the edges of a graph.
Moreover, using the notion of minimal covering for the nodes of a graph,
recent results show that it is possible to attach to graphs the so-called ideals of vertex covers; an intensive study of them carries to solve interesting
and unthinkable problems related to these fields. More precisely, the ideals
of vertex covers of a wide class of simple graphs are examined and useful
properties are studied. Moreover, we explain how algebraic methods arising
from graph theory are suitable solutions for several problems concerning the
wireless sensor networks.
Wireless sensor networks (WSNs) are an important technology for large-scale
monitoring, providing sensor measurements at high temporal and spatial resolution. A sensor network is composed of a large number of sensor nodes,
which are densely deployed either inside the phenomenon or very close to it.
These sensor nodes, which consist of sensing, data processing, and communicating components, leverage the idea of sensor networks based on collaborative effort of a large number of nodes. Sensor nodes are fitted with an
on-board processor. Instead of sending the raw data to the nodes responsible
for the fusion, sensor nodes use their processing abilities to locally carry out
simple computations and transmit only the required and partially processed
data. The above described features ensure a wide range of monitoring applications for sensor networks. Some of the application areas are environmental,
military, security and commercial ([1]).
We consider the problem of the placement of as few as possible leader nodes
which collect the information received by the sensors and send them to a base
station, so that each sensor is near to at least one of them. This problem is
related to the energy saving of the sensors. The suitable model to represent
this situation is a simple graph G and the problem is linked to the study
of the minimal vertex covers of G. We introduce a class of simple graphs
and some algebraic aspects linked to the minimal vertex covers are analyzed.
We show that the algebraic procedures introduced are good instruments concerned how this problem related to wireless sensor networks can be encoded
and solved.
Let G be a graph on n vertices v1 , . . . , vn . An algebraic object attached to G
is the edge ideal I(G), a monomial ideal of the polynomial ring in n variables
R = K[X1 , . . . , Xn ], K a field.
When G is a simple graph, I(G) is generated by squarefree
monomials of
degree two in R, I(G) = {Xi Xj | {vi , vj } is an edge of G } .
The ideal of vertex covers of I(G), denoted by Ic (G) is the ideal of R generated by all the monomials Xi1 · · · Xik such that (Xi1 , . . . , Xik ) is an associated
2
prime ideal of I(G) ([9]). Under special situations, I(G) reflects properties of
Ic (G) and the study of I(G) states some properties of Ic (G). In particular, the
Cohen-Macaulay property of the ideals Ic (G) was studied by G. Lyubeznik
in [8] using graph theoretical methods; the central notion is that of chordal
graph, that is the right notion to express some homological properties of edge
ideals and the Cohen-Macaulay property of the ideal of vertex covers.
We examine a wide class of squarefree edge ideals associated to connected
graphs H consisting of the union of a complete graph Km and star graphs
with centers the vertices of Km . The ideals of vertex covers for the edge
ideals associated to the these graphs were described in [6]. The generators
of these ideals give all the minimal vertex covers and their cardinality. If H
is considered to be the model of a WSN, then the vertex covers with minimum cardinality give the position of the leaders in the sensor network that
H represents.
Moreover we study some algebraic aspects of the ideals Ic (H). We prove that
Ic (H) has linear quotients. In general, if a monomial ideal of R generated in
one degree has linear quotients, then it has a linear resolution ([2], Lemma
4.1). In our case, Ic (H) has linear quotients, but in general Ic (H) has not
a linear resolution because it is not generated in the same degree. Using
some properties of the graph H and of its edge ideal we can obtain conditions such that Ic (H) has a linear resolution and some algebraic invariants
are computed. Starting from theoretic properties of the graph H and of its
edge ideal we can obtain information about the Cohen-Macaulay property of
I(H) and of its ideal of vertex covers.
The work is subdivided as follows. In section 1 we describe the developments
of graph theory to networks concerning of problems of minimal vertex covers.
The class of graphs H and the ideals of vertex covers Ic (H) are analyzed as
a geometric model of a WSN using computational and algebraic methods.
In section 2 we examine algebraic properties of Ic (H). We prove that these
ideals have linear quotients and we investigate conditions such that they
have a linear resolution. Moreover using the technique of studying the linear
quotients of such ideals, as previously employed in [7], standard algebraic
invariants are calculated.
1
Developments of graph theory to networks
The aim of this section is to explain how algebraic and geometric models,
built through graph theory, are useful for setting suitable solutions of several
problems concerning the wireless sensor networks.
A wireless sensor network (WSN) consists of spatially distributed autonomous
3
sensors (called sensor nodes) to monitor physical or environmental conditions
(such as temperature, sound, vibration, pressure, humidity,...) and to pass
their data through the network to a main location. The sensors collect data
of interest and send them to special gateway nodes, which collect the information received by the sensors and send them to a base station. The sensor
networks can improve the quality of information: for example, can guarantee
a high fidelity, can provide information in real time and can reduce the cost
of transmitting the same information.
Current research indicates highly developments for this type of wireless networks which are already in use in many fields. Useful applications of sensor
networks are environmental, military, and commercial. Area monitoring is
a common application of WSNs that are deployed over a region where some
phenomenons are to be monitored: the urban traffic or natural phenomena
such as glaciers, volcanoes and earthquakes. Some environmental applications of sensor networks include tracking the movements of small animals;
detecting foreign chemical agents in the air and the water, irrigation; precision agriculture; environmental monitoring in marine; forest fire detection;
meteorological or geophysical research. Some of the commercial applications
are monitoring product quality; interactive museums; monitoring disaster
area; transportation; vehicle tracking and detection ([1]).
Consider a sensor network in which each sensor must send information it
has collected to the nearest gateway node. The problem is the placement
of as few as possible gateway nodes so that each sensor is near to at least
one of them. This problem is related to the energy saving of the sensors.
The formalization of the model to represent the sensor network is given by
a graph G = (V (G), E(G)), V (G) = {v1 , . . . , vn } is the set of nodes and
E(G) = {{vi , vj }|vi 6= vj , vi , vj ∈ V (G)} is the set of edges (if two nodes v1
and v2 are adjacent then {v1 , v2 } ∈ E(G)).
Our problem consists of finding a set of nodes leaders (gateway nodes) to be
responsible to send messages to nodes belonging to the higher level. Such
leaders must have the feature of being near to all nodes. This means that we
want to find a subset of nodes C ⊂ V (G) such that each node in V (G) \ C
is joined to a node in C. This is a problem of minimum vertex cover. The
optimization problem of minimum vertex cover is looking for a covering of
vertices of a graph that has minimum cardinality. The corresponding decision problem of vertex cover is, given a graph G and an integer k, if there is
a vertex covering of G of minimal cardinality k or of cardinality at most k.
The basic network topologies are represented by complete graphs, meshes,
star graphs, trees. A single network may consist of several interconnected
subnets of different topologies. We consider as a network topology a class of
simple graphs union of a complete graph and star graphs, and some algebraic
4
aspects linked to this problem of minimal vertex covers are analyzed.
Let’s introduce some preliminary notions.
Let G be a graph. A cycle of G of length q is a subgraph Cq of G such that
E(Cq ) = {{vi1 , vi2 }, {vi2 , vi3 }, . . . , {viq−1 , viq }, {viq , vi1 }}, where {vi1 , . . . , viq } ∈
V (G) and vij 6= vik if ij 6= ik . A chord of a cycle Cq is an edge {vij , vik } of G
such that vij and vik are vertices of Cq with {vij , vik } ∈
/ E(Cq ). A graph G is
said chordal or triangulated if every cycle Cq in G, q > 3, has a chord.
A graph G on n vertices v1 , . . . , vn is complete if there exists an edge for all
pairs {vi , vj } of vertices of G.
It is denoted by Kn .
A star graph on vertex set {vi }, {v1 , . . . , vi−1 , vi+1 , vn } with center vi , denoted by stari (n), i = 1, . . . , n , is a complete bipartite graph of the form
K1, n−1 . The complementary graph of a graph G on {v1 , . . . , vn } is the graph
G on {v1 , . . . , vn } whose edge set E(G) consists of those 2-element subset
{vi , vj } for which {vi , vj } ∈
/ E(G).
If V (G) = {v1 , . . . , vn } and R = K[X1 , . . . , Xn ] is the polynomial ring
over a field K such that each variable Xi corresponds to the vertex vi , the
edge ideal I(G) associated to G is the ideal {Xi Xj | {vi , vj } ∈ E(G)} ⊂ R .
Definition 1.1. Let G be a graph with vertex set V (G) = {v1 , . . . , vn }. A
subset C of V (G) is said a minimal vertex cover for G if:
(1) every edge of G is incident with one vertex in C;
(2) there is no proper subset of C with this property.
If C satisfies condition (1) only, then C is called a vertex cover of G and
C is said to cover all the edges of G. The smallest number of vertices in any
minimal vertex cover of G is said vertex covering number. We denote it by
α0 (G).
We consider some algebraic aspects linked to the minimal vertex covers of
a graph G with set of edges E(G). More precisely, starting by the notion
of minimal vertex cover, recent results show that is possible to attach to
graphs the so-called ideals of vertex covers ([9]). There exists a one to one
correspondence between the minimal vertex covers of G and the minimal
prime ideals of I(G). In fact, ℘ is a minimal prime ideal of I(G) if and only
if ℘ = (C), for some minimal vertex cover C of G ([9], Prop. 6.1.16). Hence
ht(I(G)) = α0 (G). Thus the primary decomposition of the edge ideal of G is
given by I(G) = (C1 ) ∩ · · · ∩ (Cp ), where C1 , . . . , Cp are the minimal vertex
covers of G.
Definition 1.2. The ideal of vertex covers of I(G), denoted by Ic (G) , is the
ideal of R generated by all monomials Xi1 · · · Xik such that (Xi1 , . . . , Xik ) is
an associated prime ideal of I(G).
5
Hence Ic (G) = ({Xi1 · · · Xik | {vi1 , . . . , vik } is a vertex cover of G}) and
the minimal generators of Ic (G) correspond to the minimal vertex covers.
Now we study this topic for a wide class of connected graphs and we
show how it is possible to apply the results to the problem related to wireless sensor networks. More precisely, H is the connected graph on n vertices
v1 , . . . , vn consisting of the union of a complete graph Km and star graphs
in the vertices of Km : H = Km ∪ starj (k), where Km is the complete graph
on m vertices v1 , . . . , vm , m < n, and starj (k) is the star graph on k vertices with center vj , for some j = 1, . . . , m, k ≤ n − m. We observe that
V (starj (k)) ∩ V (Km ) = {vj } for j = 1, . . . , m. We denote |E(starj (k))| = ij ,
for j = 1, . . . , m. The edge ideal I(H) ⊂ R = K[X1 , . . . , Xn ] is stated in [6]:
I(H) = (X1 Xm+1 , X1 Xm+2 , . . . , X1 Xm+i1 , X1 X2 , X2 Xm+i1 +1 , X2 Xm+i1 +2 , . . . ,
X2 Xm+i1 +i2 , X2 X3 , X1 X3 , X3 Xm+i1 +i2 +1 , X3 Xm+i1 +i2 +2 , . . . , X3 Xm+i1 +i2 +i3 ,
X3 X4 , X2 X4 , X1 X4 , . . . , Xm Xm+i1 +i2 +···+im−1 +1 , Xm Xm+i1 +i2 +···+im−1 +2 , . . . ,
Xm Xm+i1 +i2 +···+im−1 +im ).
Remark 1.1. I(H) has a linear resolution ([6], Prop 1.1). R. Fröberg proved
that the edge ideal of a simple graph has a linear resolution if and only if its
complementary graph is chordal ([3]; [4], Th. 9.2.3). Hence H is chordal.
In [6] it is given the structure of the ideal of vertex cover of I(H). In
order to describe Ic (H) we rename the vertices of the complete graph Km
vα1 , . . . , vαm , 1 6 α1 < α2 < . . . < αm = n, and α1 = 1 if there is not the
star graph with center vα1 .
The following proposition gives explicitly the generators of the ideals of
vertex covers for the edge ideals I(H) when at least a vertex of Km has degree
m − 1.
Proposition 1.1. ([6], Prop. 2.1)
Let H = Km ∪ starj (k) with n vertices v1 , . . . , vn and Km , m < n, with
vertices vα1 , . . . , vαm , 1 6 α1 < α2 < . . . < αm = n. Let starj (k) be of
the type starα1 (α1 ) on vertices v1 , . . . , vα1 , or starαi (αi − αi−1 ), for some
i = 2, . . . , m. Then Ic (H) has m monomial squarefree generators and it is:
(X1 ··· Xα1 −1 Xα2 ··· Xαm , Xα1 Xα3 ··· Xαm , . . . , Xα1 ··· Xαm−1 ), if i = 1,
or
(Xα2 ··· Xαm , . . . , Xα1 ··· Xαi−1 Xαi−1 +1··· Xαi+1−1 Xαi+1 ··· Xαm , . . . , Xα1 ··· Xαm−1 ),
otherwise.
More in general, it is given the structure of the ideal of vertex covers of
I(H) when all the vertices of Km have degree at least m.
6
Theorem 1.1. ([6], Th. 2.1)
Let H = Km ∪ starj (k) with m star graphs with centers each of the vertices
of Km . The ideal of vertex covers of I(H) has m + 1 monomial squarefree
generators and Ic (H) is
(Xα1 Xα2 ··· Xαm , X1··· Xα1 −1 Xα2 ··· Xαm , Xα1 Xα1 +1··· Xα2 −1 Xα3 ··· Xαm , . . . , Xα1 ·
·· Xαm−2 Xαm−2 +1 ··· Xαm−1 −1 Xαm , Xα1 ··· Xαm−1 Xαm−1 +1 ··· Xαm −1 ) .
The monomial ideal Ic (H) is generated in degree ≥ m. Hence using the
previous results we give the answer to the decision problem of vertex cover :
given the graph H there is a vertex covering of H of minimum cardinality m,
where m is the number of vertices of the complete graph Km . Moreover the
generators of Ic (H) give the structure of each vertex covering of H.
Example 1.1. Let H be the simple graph with V (H) = {v1 , . . . , v11 }
H = K4 ∪ star2 (2) ∪ star6 (4) ∪ star8 (2) ∪ star11 (3).
This graph is the model of a wireless sensor network.
Ic (H) is generated by {X1 X6 X8 X11 , X2 X6 X8 X11 , X2 X6 X7 X11 , X2 X6 X8 X9 X10 ,
X2 X3 X4 X5 X8 X11 }. The minimal vertex covers of H are:
C1 = {v1 , v6 , v8 , v11 }, C2 = {v2 , v6 , v8 , v11 }, C3 = {v2 , v6 , v7 , v11 },
C4 = {v2 , v6 , v8 , v9 , v10 }, C5 = {v2 , v3 , v4 , v5 , v8 , v11 }.
Choose C1 = {v1 , v6 , v8 , v11 } among the three vertex covers with minimum
cardinality. C1 gives the position of the leaders in the sensor network that
H represents:
7
2
Algebraic properties of Ic(H)
The class of monomial ideals Ic (H) is a suitable algebraic model to solve
problems of minimal vertex covers. Moreover Ic (H) is an interesting class of
monomial ideals for its good algebraic properties. In this section we study
some of them.
Let R = K[X1 , . . . , Xn ] be the polynomial ring in n variables over a field K
with deg Xi = 1, for all i = 1, . . . , n. For a monomial ideal I ⊂ R we denote
by G(I) its unique minimal set of monomial generators.
Definition 2.1. A monomial ideal I ⊂ R has linear quotients if there is an
ordering u1 , . . . , ut of monomials belonging to G(I) such that the colon ideal
(u1 , . . . , uj−1) : (uj ) is generated by a subset of {X1 , . . . , Xn }, for 2 6 j 6 t .
For a monomial ideal I of R having linear quotients with respect to the
ordering u1 , . . . , ut of the monomials of G(I), let qj (I) denote the number of
the variables which is required to generate the ideal (u1 , . . . , uj−1) : (uj ).
Set q(I) = max26j6t qj (I). The integer q(I) is independent on the choice of
the ordering of the generators that gives linear quotients ([5]).
Proposition 2.1. Ic (H) has linear quotients.
Proof. 1) If at least a vertex of Km has degree m − 1 (Proposition 1.1), then
Ic (H) has m monomial squarefree generators. We have the following cases:
a) H is formed by the union of a complete graph Km , m < n, with vertices vα1 , . . . , vαm , 1 6 α1 < α2 < . . . < αm = n, and of a star graph
starα1 (α1 ) on vertices v1 , . . . , vα1 . Then we set the following order for the
monomial generators of Ic (H): f1 = Xα1 ··· Xαm−1 , f2 = Xα1 ··· Xαm−2 Xαm ,
. . ., fm−1 = Xα1 Xα3 ··· Xαm , fm = X1 ··· Xα1 −1 Xα2 ··· Xαm .
It results: (f1 ) : (f2 ) = (Xαm−1 ), (f1 , f2 ) : (f3 ) = (Xαm−2 ), . . ., and so on up
to (f1 , . . . , fm−1 ) : (fm ) = (Xα1 ).
b) H is formed by the union of a complete graph Km , m < n, with vertices vα1 , . . . , vαm , 1 6 α1 < α2 < . . . < αm = n, and of a star graph
starαi (αi − αi−1 ), for some i = 2, . . . , m . Then we set the following order for
the monomial generators of Ic (H): f1 = Xα1 ··· Xαm−1 , f2 = Xα1 ··· Xαm−2 Xαm ,
. . ., fm−1 = Xα1 Xα2 ··· Xαm , fm = Xα2 ··· Xαm . Hence: (f1 ) : (f2 ) = (Xαm−1 ),
(f1 , f2 ) : (f3 ) = (Xαm−2 ), . . ., and so on up to (f1 , . . . , fm−1 ) : (fm ) = (Xα1 ).
2) If all the vertices of Km have degree at least m (Theorem 1.1), then we set
the following order for the monomial generators of Ic (H): f1 = Xα1 Xα2···Xαm ,
f2 = Xα1 ··· Xαm−1 Xαm−1 +1··· Xαm −1 , f3 = Xα1 ··· Xαm−2 Xαm−2 +1··· Xαm−1 −1 Xαm ,
. . . , fm = Xα1 Xα1 +1···Xα2 −1 Xα3···Xαm , fm+1 = X1···Xα1 −1 Xα2···Xαm . It results:
(f1 ) : (f2 ) = (Xαm ), (f1 , f2 ) : (f3 ) = (Xαm−1 ), (f1 , f2 , f3 ) : (f4 ) = (Xαm−2 ),
8
. . ., and so on up to (f1 , . . . , fm ) : (fm+1 ) = (Xα1 ).
Hence Ic (H) has linear quotients.
Remark 2.1. If a monomial ideal I of R generated in one degree has linear
quotients, then I has a linear resolution ([2], Lemma 4.1). Ic (H) has linear
quotients, but in general it has not a linear resolution because it is not
generated in the same degree.
Using some properties of the graph H and of its edge ideal I(H) we can
obtain the conditions for which Ic (H) has a linear resolution.
Theorem 2.1. (([6], Th. 1.1)
Let R = K[X1 , . . . , Xn ] and I(H) ⊂ R, H = Km ∪ starj (k). Let |E(starj (k))| =
ij , for j = 1, . . . , m. Then:
1) dimR (R/I(H)) = n − m .
2) pdR (R/I(H)) = m + max16j6m ij − 1 .
3) depthR (R/I(H)) = n − m − max16j6m ij + 1 .
4) regR (R/I(H)) = 1 .
Proposition 2.2. Let I(H) ⊂ R = K[X1 , . . . , Xn ]. I(H) is Cohen Macaulay
if and only if k = 2.
Proof. Let I(H) be the edge ideal of the graph H as described in section 1.
By Theorem 2.1 one has dimR (R/I(H)) = n−m and depthR (R/I(H)) = n−
m−max16j6m ij +1. I(H) is Cohen Macaulay if and only if dimR (R/I(H)) =
depthR (R/I(H)). Hence the equality holds if and only if max16j6m ij = 1.
This is equivalent to say that |E(starj (k))| = 1 for all j = 1, . . . , m, i.e.
starj (k) has k = 2 vertices for all j = 1, . . . , m.
Proposition 2.3. Let H = Km ∪ starj (k). Then Ic (H) has linear resolution
if and only if k = 2.
Proof. I(H) has linear resolution ([6], Prop. 1.1), that is equivalent to say
that the complementary graph of H is chordal ([4], Th. 9.2.3). By Proposition 2.2, k = 2 if and only if I(H) is Cohen-Macaulay. Hence by [9] (Prop.
6.7.15) this is equivalent to say that Ic (H) has linear resolution.
\
(Xi , Xj ).
In general for a graph G one has Ic (G) =
Xi Xj ∈I(G)
i6=j
Hence Ic (G) is an unmixed ideal of height two ([9]). But in general it is
not Cohen-Macaulay. The Cohen-Macaulay property of the ideal Ic (G) was
9
studied in [8]. The notion of chordal graph is used to express this algebraic
property and to study an interesting duality between I(G) and Ic (G) ([3]).
For the class of graphs H starting by some properties of the graph H and of
its edge ideal, we state information about the Cohen-Macaulay property of
Ic (H).
Proposition 2.4. Ic (H) is Cohen-Macaulay.
Proof. I(H) has linear resolution ([6], Prop. 1.1). Hence the complementary
graph H of H is chordal ([4], Th. 9.2.3). This is equivalent to say that Ic (H)
is Cohen-Macaulay ([8]).
The study of the linear quotients is useful in order to investigate algebraic invariants of R/Ic (H): the dimension, dimR (R/Ic (H)), the depth,
depth(R/Ic (H)), the projective dimension, pdR (R/Ic (H)).
Lemma 2.1. Let R = K[X1 , . . . , Xn ] and Ic (H) ⊂ R. Then q(Ic (H)) = 1.
Proof. By the computation of the linear quotients (Proposition 2.1), the
number of the variables which is required to generate the ideal (f1 , . . . , fh−1 ) :
(fh ) is 1, for all h = 1, . . . , t, t ≤ m + 1.
Theorem 2.2. Let R = K[X1 , . . . , Xn ] and Ic (H) ⊂ R. Then:
1) pdR (R/Ic (H)) = 2
2) depthR (R/Ic (H)) = n − 2
3) dimR (R/Ic (H)) = n − 2
4) regR (R/Ic (H)) = max{deg(vαi )} − 1, i = 1, . . . , m.
Proof. 1) The length of the minimal free resolution of R/Ic (H) over R is
equal to q(Ic (H)) + 1 ([5], Corollary 1.6). Then pdR (R/Ic (H)) = 2 .
2) As a consequence of 1), by Auslander-Buchsbaum formula, one has
depthR (R/Ic (H)) = n − pdR (R/Ic (H)) = n − 2.
3) By Proposition 2.4 Ic (H) is Cohen-Macaulay, hence dimR (R/Ic (H)) =
depthR (R/Ic (H)) = n − 2.
4) By [2] (Lemma 4.1) the Castelnuovo-Mumford regularity is regR (R/Ic (H)) =
max{deg(f )| f minimal generator of Ic (H)} − 1. By the structure of generators of Ic (H) it follows that max{deg(f )| f minimal generator of Ic (H)} =
m − 1 − max{deg(vαi )} − (m − 1) = max{deg(vαi )}, i = 1, . . . , m. Hence
regR (R/Ic (H)) = max{deg(vαi )} − 1, i = 1, . . . , m.
10
Example 2.1. Let H = K3 ∪ star1 (2) ∪ star2 (2) ∪ star3 (2) on vertices
{v1 , . . . , v6 }.
R = K[X1 , . . . , X6 ], I(H) = (X1 X2 , X2 X6 , X2 X4 , X4 X6 , X3 X4 , X5 X6 ).
Ic (H) = (X2 X4 X6 , X2 X4 X5 , X2 X3 X6 , X1 X4 X6 ) is generated in the same degree 3. Ic (H) has linear quotients, hence linear resolution:
0 → R3 (−4) → R4 (−3) → R → R/Ic (H) → 0
One has:
1) pdR (R/Ic (H)) = 2 .
2) dimR (R/Ic (H)) = depthR (R/Ic (H)) = 4 .
3) regR (R/Ic (H)) = 2
References
[1] Akyildiz I.F., Su W., Sankarasubramaniam Y., Cayirci E. - Wireless
sensor networks: a survey. Computer Networks 38 (2002), 393-422.
[2] Conca, A., and Herzog, J. - Castelnuovo-Mumford regularity of product
of ideals. Collect. Math. 54 (2003), 137–152.
[3] Fröberg, R. - On Stanley-Reisner rings. Topics in Algebra (1990), 57–70.
[4] Herzog, J., and Hibi, T. - Monomial Ideals. Springer, Graduate Texts
in Mathematics 260 (2011).
[5] Herzog, J., and Takayama, Y. - Resolutions by mapping cones. Homology, Homotopy Appl. 4, 2 (2002), 277–294.
[6] Imbesi, M., and La Barbiera M. - Monomial ideals of graphs with loops.
Anal. Sţiinţ. Univ. Al. I Cuza din Iaşi, Sect. Mat., In press.
[7] La Barbiera M. - On a class of monomial ideals generated by s-sequences.
Math. Reports, 12(62), 3 (2010), 201–216.
[8] Lyubeznik, G. - The minimal non-Cohen-Macaulay monomial ideals. J.
Pure and Applied Algebra, 51 (1988), 261–266.
[9] Villarreal, R.H. - Monomial Algebras. M. Dekker Inc., New York, Pure
and Appl. Math. 238 (2001).
11
| 0math.AC
|
Swift: Compiled Inference for Probabilistic Programming Languages
arXiv:1606.09242v1 [cs.AI] 30 Jun 2016
Yi Wu
UC Berkeley
[email protected]
Lei Li
Toutiao.com
[email protected]
Abstract
A probabilistic program defines a probability measure over its semantic structures. One common goal
of probabilistic programming languages (PPLs)
is to compute posterior probabilities for arbitrary
models and queries, given observed evidence, using a generic inference engine. Most PPL inference engines—even the compiled ones—incur significant runtime interpretation overhead, especially
for contingent and open-universe models. This
paper describes Swift, a compiler for the BLOG
PPL. Swift-generated code incorporates optimizations that eliminate interpretation overhead, maintain dynamic dependencies efficiently, and handle
memory management for possible worlds of varying sizes. Experiments comparing Swift with other
PPL engines on a variety of inference problems
demonstrate speedups ranging from 12x to 326x.
1
Introduction
Probabilistic programming languages (PPLs) aim to combine
sufficient expressive power for writing real-world probability
models with efficient, general-purpose inference algorithms
that can answer arbitrary queries with respect to those models. One underlying motive is to relieve the user of the obligation to carry out machine learning research and implement
new algorithms for each problem that comes along. Another
is to support a wide range of cognitive functions in AI systems and to model those functions in humans.
General-purpose inference for PPLs is very challenging;
they may include unbounded numbers of discrete and continuous variables, a rich library of distributions, and the ability
to describe uncertainty over functions, relations, and the existence and identity of objects (so-called open-universe models). Existing PPL inference algorithms include likelihood
weighting (LW) [Milch et al., 2005b], parental Metropolis–
Hastings (PMH) [Milch and Russell, 2006; Goodman et
al., 2008], generalized Gibbs sampling (Gibbs) [Arora et
al., 2010], generalized sequential Monte Carlo [Wood et
al., 2014], Hamiltonian Monte Carlo (HMC) [Stan Development Team, 2014], variational methods [Minka et al., 2014;
Kucukelbir et al., 2015] and a form of approximate Bayesian
Stuart Russell
Rastislav Bodik
UC Berkeley
University of Washington
[email protected] [email protected]
computation [Mansinghka et al., 2013]. While better algorithms are certainly possible, our focus in this paper is on
achieving orders-of-magnitude improvement in the execution
efficiency of a given algorithmic process.
A PPL system takes a probabilistic program (PP) specifying a probabilistic model as its input and performs inference to compute the posterior distribution of a query given
some observed evidence. The inference process does not (in
general) execute the PP, but instead executes the steps of an
inference algorithm (e.g., Gibbs) guided by the dependency
structures implicit in the PP. In many PPL systems the PP exists as an internal data structure consulted by the inference
algorithm at each step [Pfeffer, 2001; Lunn et al., 2000;
Plummer, 2003; Milch et al., 2005a; Pfeffer, 2009]. This process is in essence an interpreter for the PP, similar to early
Prolog systems that interpreted the logic program. Particularly when running sampling algorithms that involve millions of repetitive steps, the overhead can be enormous. A
natural solution is to produce model-specific compiled inference code, but, as we show in Sec. 2, existing compilers for general open-universe models [Wingate et al., 2011;
Yang et al., 2014; Hur et al., 2014; Chaganty et al., 2013;
Nori et al., 2014] miss out on optimization opportunities and
often produce inefficient inference code.
The paper analyzes the optimization opportunities for PPL
compilers and describes the Swift compiler, which takes as
input a BLOG program [Milch et al., 2005a] and one of three
inference algorithms (LW, PMH, Gibbs) and generates target
code for answering queries. Swift includes three main contributions: (1) elimination of interpretative overhead by joint
analysis of the model structure and inference algorithm; (2) a
dynamic slice maintenance method (FIDS) for incremental
computation of the current dependency structure as sampling
proceeds; and (3) efficient runtime memory management for
maintaining the current-possible-world data structure as the
number of objects changes. Comparisons between Swift and
other PPLs on a variety of models demonstrate speedups
ranging from 12x to 326x, leading in some cases to performance comparable to that of hand-built model-specific code.
To the extent possible, we also analyze the contributions of
each optimization technique to the overall speedup.
Although Swift is developed for the BLOG language, the
overall design and the choices of optimizations can be applied
to other PPLs and may bring useful insights to similar AI
Classical Probabilistic
Program Optimizations
(insensitive to inference
algorithm)
Classical Algorithm
Optimizations
(insensitive to model)
Inference Code 𝑷𝑰
Probabilistic
Program
(Model)
𝑷𝑴
𝓕𝑰 𝑷𝑴 → 𝑷𝑰
Algorithm-Specific
Code
𝓕cg 𝑷𝑰 → 𝑷𝑻
Target Code
(C++)
𝑷𝑻
Model-Specific Code
Our Optimization: FIDS
(combine model and algorithm)
Figure 1: PPL compilers and optimization opportunities.
systems for real-world applications.
2
Figure 2: The urn-ball model
Our Optimization:
Data Structures for FIDS
Existing PPL Compilation Approaches
In a general purpose programming language (e.g., C++), the
compiler compiles exactly what the user writes (the program).
By contrast, a PPL compiler essentially compiles the inference algorithm, which is written by the PPL developer, as
applied to a PP. This means that different implementations of
the same inference algorithm for the same PP result in completely different target code.
As shown in Fig. 1, a PPL compiler first produces an intermediate representation combining the inference algorithm (I)
and the input model (PM ) as the inference code (PI ), and then
compiles PI to the target code (PT ). Swift focuses on optimizing the inference code PI and the target code PT given a
fixed input model PM .
Although there have been successful PPL compilers, these
compilers are all designed for a restricted class of models
with fixed dependency structures: see work by Tristan [2014]
and the Stan Development Team [2014] for closed-world
Bayes nets, Minka [2014] for factor graphs, and Kazemi and
Poole [2016] for Markov logic networks.
Church [Goodman et al., 2008] and its variants [Wingate
et al., 2011; Yang et al., 2014; Ritchie et al., 2016] provide a lightweight compilation framework for performing
the Metropolis–Hastings algorithm (MH) over general openuniverse probability models (OUPMs). However, these approaches (1) are based on an inefficient implementation of
MH, which results in overhead in PI , and (2) rely on inefficient data structures in the target code PT .
For the first point, consider an MH iteration where we are
proposing a new possible world w0 from w accoding to the
proposal g(·) by resampling random variable X from v to v 0 .
The computation of acceptance ratio α follows
g(v 0 → v) Pr[w0 ]
α = min 1,
.
(1)
g(v → v 0 ) Pr[w]
Since only X is resampled, it is sufficient to compute α using merely the Markov blanket of X, which leads to a much
simplified formula for α from Eq.(1) by cancelling terms in
Pr[w] and Pr[w0 ]. However, when applying MH for a contingent model, the Markov blanket of a random variable cannot be determined at compile time since the model dependencies vary during inference. This introduces tremendous runtime overhead for interpretive systems (e.g., BLOG, Figaro),
which track all the dependencies at runtime even though typically only a tiny portion of the dependency structure may
change per iteration. Church simply avoids keeping track of
dependencies and uses Eq.(1), including a potentially huge
overhead for redundant probability computations.
For the second point, in order to track the existence of
the variables in an open-universe model, similar to BLOG,
Church also maintains a complicated dynamic string-based
hashing scheme in the target code, which again causes interpretation overhead.
Lastly, techniques proposed by Hur et al. [2014], Chaganty
et al. [2013] and Nori et al. [2014] primarily focus on optimizing PM by analyzing the static properties of the input PP,
which are complementary to Swift.
3
Background
This paper focuses on the BLOG language, although our approach also applies to other PPLs with equivalent semantics [McAllester et al., 2008; Wu et al., 2014]. Other PPLs
can be also converted to BLOG via static single assignment form (SSA form) transformation [Cytron et al., 1989;
Hur et al., 2014].
3.1
The BLOG Language
The BLOG language [Milch et al., 2005a] defines probability measures over first-order (relational) possible worlds; in
this sense it is a probabilistic analogue of first-order logic. A
BLOG program declares types for objects and defines distributions over their numbers, as well as defining distributions
for the values of random functions applied to objects. Observation statements supply evidence, and a query statement
specifies the posterior probability of interest. A random variable in BLOG corresponds to the application of a random
function to specific objects in a possible world. BLOG naturally supports open universe probability models (OUPMs)
and context-specific dependencies.
Fig. 2 demonstrates the open-universe urn-ball model. In
this version the query asks for the color of the next random
pick from an urn given the colors of balls drawn previously.
Line 4 is a number statement stating that the number variable #Ball, corresponding to the total number of balls, is
uniformly distributed between 1 and 20. Lines 5–6 declare a
random function color(·), applied to balls, picking Blue or
Green with a biased probability. Lines 7–8 state that each
draw chooses a ball at random from the urn, with replacement.
Fig. 3 describes another OUPM, the infinite Gaussian mixture model (∞-GMM). The model includes an unknown
• Provide efficient memory management and fast data
structures for Monte Carlo sampling algorithms to avoid
interpretive overhead at runtime.
Figure 3: The infinite Gaussian mixture model
Algorithm 1: Metropolis–Hastings algorithm (MH)
1
Input: M, E, Q, N ; Output: samples H
initialize a possible world w(0) with E satisfied;
2
for i ← 1 to N do
6
randomly pick a variable X from w(i−1) ;
w(i) ← w(i−1) and v ← X(w(i−1) );
propose a value v 0 for X via proposal g ;
X(w(i) ) ← v 0 and ensure w(i) is self-supporting;
7
α ← min 1,
3
4
5
g(v 0 →v) Pr[w(i) ]
g(v→v 0 ) Pr[w(i−1) ]
(i)
if rand(0, 1) ≥ α then w
H ← H + (Q(w(i) ));
8
;
← w(i−1)
number of clusters as stated in line 3, which is to be inferred
from the data.
3.2
Generic Inference Algorithms
All PPLs need to infer the posterior distribution of a query
given the observed evidence. The great majority of PPLs use
Monte Carlo sampling methods such as likelihood weighting
(LW) and Metropolis–Hastings MCMC (MH). LW samples
unobserved random variables sequentially in topological order, conditioned on evidence and sampled values earlier in
the sequence; each complete sample is weighted by the likelihood of the evidence variables appearing in the sequence.
The MH algorithm is summarized in Alg. 1. M denotes a
BLOG model, E its evidence, Q a query, N the number of
samples, w a possible world, X(w) the value of random variable X in w, Pr[X(w)|w-X ] denotes the conditional probability of X in w, and Pr[w] denotes the likelihood of possible world w. In particular, when the proposal distribution
g in Alg.1 samples a variable conditioned on its parents, the
algorithm becomes the parental MH algorithm (PMH). Our
discussion in the paper focuses on LW and PMH but our proposed solution applies to other Monte Carlo methods as well.
4
For the first goal, we propose a novel framework, FIDS
(Framework for Incrementally updating Dynamic Slices), for
generating the inference code (PI in Fig. 1). For the second
goal, we carefully choose data structures in the target C++
code (PT ).
For convenience, r.v. is short for random variable. We
demonstrate examples of compiled code in the following
discussion: the inference code (PI ) generated by FIDS in
Sec. 4.1 is in pseudo-code while the target code (PT ) by Swift
shown in Sec. 4.2 is in C++. Due to limited space, we show
the detailed transformation rules only for the adaptive contingency updating (ACU) technique and omit the others.
4.1
Optimizations in FIDS
Our discussion in this section focuses on LW and PMH, although FIDS can be also applied to other algorithms. FIDS
includes three optimizations: dynamic backchaining (DB),
adaptive contingency updating (ACU), and reference counting (RC). DB is applied to both LW and PMH while ACU
and RC are specific to PMH.
Dynamic backchaining
Dynamic backchaining (DB) [Milch et al., 2005a] constructs
a dynamic slice incrementally in each LW iteration and samples only those variables necessary at runtime. DB in Swift
is an example of compiling lazy evaluation: in each iteration,
DB backtracks from the evidence and query and samples an
r.v. only when its value is required during inference.
For every r.v. declaration random T X ∼ CX ; in the
input PP, FIDS generates a getter function get X(), which
is used to (1) sample a value for X from its declaration CX
and (2) memoize the sampled value for later references in
the current iteration. Whenever an r.v. is referred to during
inference, its getter function will be evoked.
Compiling DB is the fundamental technique for Swift. The
key insight is to replace dependency look-ups and method invocations from some internal PP data structure with direct
machine address accessing and branching in the target executable file.
For example, the inference code for the getter function of
x(d) in the ∞-GMM model (Fig. 3) is shown below.
The Swift Compiler
Swift has the following design goals for handling openuniverse probability models.
• Provide a general framework to automatically and efficiently (1) track the exact Markov blanket for each
random variable and (2) maintain the minimum set of
random variables necessary to evaluate the query and
the evidence (the dynamic slice [Agrawal and Horgan,
1990], or equivalently the minimum self-supporting partial world [Milch et al., 2005a]).
memoization denotes some pseudo-code snippet for
performing memoization. Since get_z(·) is called before
get_mu(·), only those mu(c) corresponding to non-empty
clusters will be sampled. Notably, with the inference code
above, no explicit dependency look-up is ever required at runtime to discover those non-empty clusters.
When sampling a number variable, we need to allocate
memory for the associated variables. For example, in the urnball model (Fig. 2), we need to allocate memory for color(b)
after sampling #Ball. The corresponding generated inference code is shown below.
allocate_memory_color(val) denotes some pseudocode segment for allocating val chunks of memory for the
values of color(b).
Reference Counting
Reference counting (RC) generalize the idea of DB to incrementally maintain the dynamic slice in PMH. RC is an efficient compilation strategy for the interpretive BLOG to dynamically maintain references to variables and exclude those
without any references from the current possible world with
minimal runtime overhead. RC is also similar to dynamic
garbage collection in programming language community.
For an r.v. being tracked, say X, RC maintains a reference counter cnt(X) defined by cnt(X) = |Chw (X)|.
Chw (X) denote the children of X in the possible world w.
When cnt(X) becomes zero, X is removed from w; when
cnt(X) become positive, X is instantiated and added back
to w. This procedure is performed recursively.
Tracking references to every r.v. might cause unnecessary
overhead. For example, for classical Bayes nets, RC never excludes any variables. Hence, as a trade-off, Swift only counts
references in open-universe models, particularly, to those
variables associated with number variables (e.g., color(b) in
the urn-ball model).
Take the urn-ball model as an example. When resampling
drawn(d) and accepting the proposed value v, the generated
code for accepting the proposal will be
w[X ← v] may require instantiating new variables not existing in w due to dependency changes.
Computing P arw (X) for X is straightforward by executing its generation process, CX , within w, which is often inexpensive. Hence, the principal challenge for maintaining
the Markov blanket is to efficiently maintain a children set
Ch(X) for each r.v. X with the aim of keeping identical to
the true set Chw (X) for the current possible world w.
With a proposed possible world (PW) w0 = w[X ← v],
it is easy to add all missing dependencies from a particular
r.v. U . That is, for an r.v. U and every r.v. V ∈ P arw0 (U ),
since U must be in Chw0 (V ), we can keep Ch(V) up-to-date
by adding U to Ch(V) if U 6∈ Ch(V). Therefore, the key
step is tracking the set of variables, ∆(w[X ← v]), which
will have a set of parents in w[X ← v] different from w.
∆(W [X ← v]) = {Y : P arw (Y ) 6= P arw[X←v] (Y )}
Precisely computing ∆(w[X ← v]) is very expensive at
runtime but computing an upper bound for ∆(w[X ← v])
does not influence the correctness of the inference code.
Therefore, we proceed to find an upper bound that holds for
any value v. We call this over-approximation the contingent
set Contw (X).
Note that for every r.v. U with dependency that changes in
w(X ← v), X must be reached in the condition of a control
statement on the execution path of CU within w. This implies
∆(w[X ← v]) ⊆ Chw (X). One straightforward idea is set
Contw (X) = Chw (X). However, this leads to too much
overhead. For example, ∆(·) should be always empty for
models with fixed dependencies.
Another approach is to first find out the set of switching
variables SX for every r.v. X, which is defined by
SX = {Y : ∃w, v P arw (X) 6= P arw[Y ←v] (X)},
This brings an immediate advantage: SX can be statically
computed at compile time. However, computing SX is NPComplete1 .
Our solution is to derive the approximation SbX by taking
the union of free variables of if/case conditions, function arguments, and set expression conditions in CX , and then set
Contw (X) = {Y : Y ∈ Chw (X) ∧ X ∈ SbY }.
The function inc_cnt(X) and dec_cnt(X) update the
references to X. The function accept_value_drawn() is
specialized to r.v. drawn(d): Swift analyzes the input program and generates specialized codes for different variables.
Adaptive contingency updating
The goal of ACU is to incrementally maintain the Markov
Blanket for every r.v. with minimal efforts.
CX denotes the declaration of r.v. X in the input PP;
P arw (X) denote the parents of X in the possible world w;
w[X ← v] denotes the new possible world derived from w
by only changing the value of X to v. Note that deriving
Contw (X) is a function of the sampling variable X and the
current PW w. Likewise, for every r.v. X, we maintain a
runtime contingent set Cont(X) identical to the true set
Contw (X) under the current PW w. Cont(X) can be incrementally updated as well.
Back to the original focus of ACU, suppose we are adding
new the dependencies in the new PW w0 = w[X ← v]. There
are three steps to accomplish the goal: (1) enumerating all the
variables U ∈ Cont(X), (2) for all V ∈ P arw0 (U ), add U
to Ch(V), and (3) for all V ∈ P arw0 (U ) ∩ SbU , add U to
Cont(V). These steps can be also repeated in a similar way
to remove the vanished dependencies.
Take the ∞-GMM model (Fig. 3) as an example , when resampling z(d), we need to change the dependency of r.v. x(d)
1
Since the declaration CX may contain arbitrary boolean formulas, one can reduce the 3-SAT problem to computing SX .
Fc (M = random T Id([T Id ,]∗ ) ∼ C) =
void Id::add_to_Ch()
{ Fcc (C, Id, {}) }
void Id::accept_value(T v)
{ for(u in Cont(Id)) u.del_from_Ch();
Id = v;
for(u in Cont(Id)) u.add_to_Ch(); }
Here |w| denotes the total number of random variables existing in w, which can be maintained via RC.
Finally, the computation time of ACU is strictly shorter
than that of acceptance ratio α (Eq. 2).
4.2
Fcc (C = Exp, X, seen) = Fcc (Exp, X, seen)
Fcc (C = Dist(Exp), X, seen) = Fcc (Exp, X, seen)
Fcc (C = if (Exp) then C1 else C2 , X, seen) =
Fcc (Exp, X, seen);
if (Exp)
{ Fcc (C1 , X, seen ∪ F V (Exp)); }
else
{ Fcc (C2 , X, seen ∪ F V (Exp)); }
Fcc (Exp, X, seen) =
/* emit a line of code for every r.v. U in the corresponding set */
∀ U ∈ ((F V (Exp) ∩ SbX )\seen) : Cont(U) += X;
∀ U ∈ (F V (Exp)\seen) : Ch(U) += X;
Figure 4: Transformation rules for outputting inference code
with ACU. F V (Expr) denote the free variables in Expr.
We assume the switching variables SbX have already been
computed. The rules for number statements and case statements are not shown for conciseness—they are analogous to
the rules for random variables and if statements respectively.
since Contw (z(d)) = {x(d)} for any w. The generated code
for this process is shown below.
We omit the code for del_from_Ch() for conciseness,
which is essentially the same as add_to_Ch().
The formal transformation rules for ACU are demonstrated
in Fig. 4. Fc takes in a random variable declaration statement
in the BLOG program and outputs the inference code containing methods accept value() and add to Ch(). Fcc
takes in a BLOG expression and generates the inference code
inside method add to Ch(). It keeps track of already seen
variables in seen, which makes sure that a variable will be
added to the contingent set or the children set at most once.
Code for removing variables can be generated similarly.
Since we maintain the exact children for each r.v. with
ACU, the computation of acceptance ratio α in Eq. 1 for PMH
can be simplified to
min 1,
0
|w| Pr[X = v|w-X
]
Q
|w0 | Pr[X = v 0 |w-X ]
U ∈Chw0 (X)
Q
0
Pr[U (w0 )|w-U
]
V ∈Chw (X)
!
Pr[V (w)|w-V ]
(2)
Implementation of Swift
Lightweight memoization
Here we introduce the implementation details of the memoization code in the getter function mentioned in section 4.1.
Objects will be converted to integers in the target code.
Swift analyzes the input PP and allocates static memory for
memoization in the target code. For open-universe models
where the number of random variables are unknown, the dynamic table data structure (e.g., vector in C++) is used to
reduce the amount of dynamic memory allocations: we only
increase the length of the array when the number becomes
larger than the capacity.
One potential weakness of directly applying a dynamic table is that, for models with multiple nested number statements
(e.g. the aircraft tracking model with multiple aircrafts and
blips for each one), the memory consumption can be large.
In Swift, the user can force the target code to clear all the
unused memory every fixed number of iterations via a compilation option, which is turned off by default.
The following code demonstrates the target code for the
number variable #Ball and its associated color(b)’s in the
urn-ball model (Fig. 2) where iter is the current iteration
number.
Note that when generating a new PW, by simply increasing the counter iter, all the memoization flags (e.g.,
mark color for color(·)) will be automatically cleared.
Lastly, in PMH, we need to randomly select an r.v. to sample per iteration. In the target C++ code, this process is accomplished via polymorphism by declaring an abstract class
with a resample() method and a derived class for every
r.v. in the PP. An example code snippet for the ∞-GMM
model is shown below.
Efficient proposal manipulation
Although one PMH iteration only samples a single variable,
generating a new PW may still involve (de-)instantiating an
arbitrary number of random variables due to RC or sampling
a number variable. The general approach commonly adopted
by many PPL systems is to construct the proposed possible
world in dynamically allocated memory and then copy it to
the current world [Milch et al., 2005a; Wingate et al., 2011],
which suffers from significant overhead.
In order to generate and accept new PWs in PMH with
negligible memory management overhead, we extend the
lightweight memoization to manipulate proposals: Swift statically allocates an extra memoized cache for each random
variable, which is dedicated to storing the proposed value for
that variable. During resampling, all the intermediate results
are stored in the cache. When accepting the proposal, the proposed value is directly loaded from the cache; when rejection,
no action is needed due to our memoization mechanism.
Here is an example target code fragment for mu(c) in the
∞-GMM model. proposed vars is an array storing all
the variables to be updated if the proposal gets accepted.
Supporting new algorithms
We demonstrate that FIDS can be applied to new algorithms
by implementing a translator for the Gibbs sampling [Arora
et al., 2010] (Gibbs) in Swift.
Gibbs is a variant of MH with the proposal distribution g(·)
set to be the posterior distribution. The acceptance ratio α is
always 1 while the disadvantage is that it is only possible to
explicitly construct the posterior distribution with conjugate
priors. In Gibbs, the proposal distribution is still constructed
from the Markov blanket, which again requires maintaining
the children set. Hence, FIDS can be fully utilized.
However, different conjugate priors yield different forms
of posterior distributions. In order to support a variety of
proposal distributions, we need to (1) implement a conjugacy
checker to check the form of posterior distribution for each
r.v. in the PP at compile time (the checker has nothing to do
with FIDS); (2) implement a posterior sampler for every conjugate prior in the runtime library of Swift.
5
Experiments
In the experiments, Swift generates target code in C++ with
C++ standard <random> library for random number generation and the armadillo package [Sanderson, 2010] for matrix computation. The baseline systems include BLOG (version 0.9.1), Figaro (version 3.3.0), Church (webchurch), Infer.NET (version 2.6), BUGS (winBUGS 1.4.3) and Stan
(cmdStan 2.6.0).
5.1
Benchmark models
We collect a set of benchmark models2 which exhibit various capabilities of a PPL (Tab. 1), including the burglary
model (Bur), the hurricane model (Hur), the urn-ball model
(U-B(x,y) denotes the urn-ball model with at most x balls and
y draws), the TrueSkill model [Herbrich et al., 2007] (T-K),
1-dimensional Gaussian mixture model (GMM, 100 points
with 4 clusters) and the infinite GMM model (∞-GMM). We
also include a real-world dataset: handwritten digits [LeCun
et al., 1998] using the PPCA model [Tipping and Bishop,
1999]. All the models can be downloaded from the GitHub
repository of BLOG.
5.2
Speedup by FIDS within Swift
We first evaluate the speedup promoted by each of the three
optimizations in FIDS individually.
Dynamic Backchaining for LW
We compare the running time of the following versions of
code: (1) the code generated by Swift (“Swift”); (2) the modified compiled code with DB manually turned off (“No DB”),
which sequentially samples all the variables in the PP; (3)
the hand-optimized code without unnecessary memoization
or recursive function calls (“Hand-Opt”).
2
Most models are simplified to ensure that all the benchmark
systems can produce an output in reasonably short running time. All
of these models can be enlarged to handle more data without adding
extra lines of BLOG code.
Bur
R
CT
CC
OU
CG
Hur
X
X
X
X
U-B
X
X
X
T-K
GMM
∞-GMM
PPCA
X
X
X
X
X
X
X
X
X
30
25
20
2.3x
No RC
Swift
15
10
5
X
Table 1: Features of the benchmark models. R: continuous
scalar or vector variables. CT: context-specific dependencies
(contingency). CC: cyclic dependencies in the PP (while in
any particular PW the dependency structure remains acyclic).
OU: open-universe. CG: conjugate priors.
Alg.
Bur
Hur
U-B(20,2) U-B(20,8)
Running Time (s): Swift v.s. No DB
No DB
0.768
1.288
2.952
5.040
Swift
0.782
1.115
1.755
4.601
Speedup
0.982
1.155
1.682
1.10
Running Time (s): Swift v.s. Hand-Optimized
Hand-Opt
0.768
1.099
1.723
4.492
Swift
0.782
1.115
1.755
4.601
Overhead
1.8%
1.4%
1.8%
2.4%
Calls to Random Number Generator
No DB
3 ∗ 107
3 ∗ 107
1.35 ∗ 108 1.95 ∗ 108
Swift
3 ∗ 107 2.0 ∗ 107 4.82 ∗ 107 1.42 ∗ 108
Calls Saved
0%
33.3%
64.3%
27.1%
Table 2: Performance on LW with 107 samples for Swift, the
version without DB and the hand-optimized version
We measure the running time for all the 3 versions and the
number of calls to the random number generator for “Swift”
and “No DB”. The result is concluded in Tab. 2.
The overhead due to memoization compared against the
hand-optimized code is less than 2.4%. We can further notice
that the speedup is proportional to the number of calls saved.
Reference Counting for PMH
RC only applies to open-universe models in Swift. Hence we
focus on the urn-ball model with various model parameters.
The urn-ball model represents a common model structure appearing in many applications with open-universe uncertainty.
This experiment reveals the potential speedup by RC for realworld problems with similar structures. RC achieves greater
speedup when the number of balls and the number of observations become larger.
We compare the code produced by Swift with RC (“Swift”)
and RC manually turned off (“No RC”). For “No RC”, we
traverse the whole dependency structure and reconstruct the
dynamic slice every iteration. Fig. 5 shows the running time
of both versions. RC is indeed effective – leading to up to
2.3x speedup.
Adaptive Contingency Updating for PMH
We compare the code generated by Swift with ACU (“Swift”)
and the manually modified version without ACU (“Static
Dep”) and measure the number of calls to the likelihood function in Tab. 3.
The version without ACU (“Static Dep”) demonstrates the
0
U-B(10,5)
U-B(20,5)
U-B(40,5)
U-B(40,10)
U-B(40,20)
U-B(40,40)
Figure 5: Running time (s) of PMH in Swift with and without
RC on urn-ball models for 20 million samples.
Alg.
Bur
Hur
U-B(20,10) U-B(40,20)
Running Time (s): Swift v.s. Static Dep
Static Dep
0.986
1.642
4.433
7.722
Swift
0.988
1.492
3.891
4.514
Speedup
0.998
1.100
1.139
1.711
Calls to Likelihood Functions
Static Dep
1.8 ∗ 107 2.7 ∗ 107 1.5 ∗ 108
3.0 ∗ 108
7
7
7
Swift
1.8 ∗ 10
1.7 ∗ 10
4.1 ∗ 10
5.5 ∗ 107
Calls Saved
0%
37.6%
72.8%
81.7%
Table 3: Performance of PMH in Swift and the version utilizing static dependencies on benchmark models.
best efficiency that can be achieved via compile-time analysis without maintaining dependencies at runtime. This version statically computes an upper bound of the children set
c
by Ch(X)
= {Y : X appears in CY } and again uses Eq.(2)
to compute the acceptance ratio. Thus, for models with fixed
dependencies, “Static Dep” should be faster than “Swift”.
In Tab. 3, “Swift” is up to 1.7x faster than “Static Dep”,
thanks to up to 81% reduction of likelihood function evaluations. Note that for the model with fixed dependencies
(Bur), the overhead by ACU is almost negligible: 0.2% due
to traversing the hashmap to access to child variables.
5.3
Swift against other PPL systems
We compare Swift with other systems using LW, PMH and
Gibbs respectively on the benchmark models. The running
time is presented in Tab. 4. The speedup is measured between
Swift and the fastest one among the rest. An empty entry
means that the corresponding PPL fails to perform inference
on that model. Though these PPLs have different host languages (C++, Java, Scala, etc.), the performance difference
resulting from host languages is within a factor of 23 .
5.4
Experiment on real-world dataset
We use Swift with Gibbs to handle the probabilistic principal component analysis (PPCA) [Tipping and Bishop, 1999],
with real-world data: 5958 images of digit “2” from the handwritten digits dataset (MNIST) [LeCun et al., 1998]. We
compute 10 principal components for the digits. The training
and testing sets include 5958 and 1032 images respectively,
each with 28x28 pixels and pixel value rescaled to [0, 1].
Since most benchmark PPL systems are too slow to handle this amount of data, we compare Swift against other two
3
http://benchmarksgame.alioth.debian.org/
PPL
Bur
Church
Figaro
BLOG
Swift
Speedup
9.6
15.8
8.4
0.04
196
Church
Figaro
BLOG
Swift
Speedup
12.7
10.6
6.7
0.11
61
BUGS
Infer.NET
Swift
Speedup
87.7
1.5
0.12
12
Hur
U-B
(20,10)
T-K
LW with 106 samples
22.9
179
57.4
24.7
176
48.6
11.9
189
49.5
0.08
0.54
0.24
145
326
202
PMH with 106 samples
25.2
246
173
59.6
17.4
18.5
30.4
38.7
0.15
0.4
0.32
121
75
54
Gibbs with 106 samples
0.19
0.34
∞
∞
-
GMM
∞-GMM
1627
997
998
6.8
147
1038
235
261
1.1
214
3703
151
72.5
0.76
95
1057
62.2
68.3
0.60
103
84.4
77.8
0.42
185
0.86
∞
Table 4: Running time (s) of Swift and other PPLs on the
benchmark models using LW, PMH and Gibbs.
scalable PPLs, Infer.NET and Stan, on this model. Both PPLs
have compiled inference and are widely used for real-world
applications. Stan uses HMC as its inference algorithm. For
Infer.NET, we select variational message passing algorithm
(VMP), which is Infer.NET’s primary focus. Note that HMC
and VMP are usually favored for fast convergence.
Stan requires a tuning process before it can produce samples. We ran Stan with 0, 5 and 9 tuning steps respectively4 .
We measure the perplexity of the generated samples over test
images w.r.t. the running time in Fig. 6, where we also visualize the produced principal components. For Infer.NET, we
consider the mean of the approximation distribution.
Swift quickly generates visually meaningful outputs with
around 105 Gibbs iterations in 5 seconds. Infer.NET takes
13.4 seconds to finish the first iteration5 and converges to a
result with the same perplexity after 25 iterations and 150
seconds. The overall convergence of Gibbs w.r.t. the running
time significantly benefits from the speedup by Swift.
For Stan, its no-U-turn sampler [Homan and Gelman,
2014] suffers from significant parameter tuning issues. We
also tried to manually tune the parameters, which does not
help much. Nevertheless, Stan does work for the simplified
PPCA model with 1-dimensional data. Although further investigation is still needed, we conjecture that Stan is very sensitive to its parameters in high-dimensional cases. Lastly, this
experiment also suggests that those parameter-robust inference engines, such as Swift with Gibbs and Infer.NET with
VMP, would be preferred at practice when possible.
4
We also ran Stan with 50 and 100 tuning steps, which took more
than 3 days to finish 130 iterations (including tuning). However, the
results are almost the same as that with 9 tuning steps.
5
Gibbs by Swift samples a single variable while Infer.NET processes all the variables per iteration.
Figure 6: Log-perplexity w.r.t running time(s) on the PPCA
model with visualized principal components. Swift converges
faster.
6
Conclusion and Future Work
We have developed Swift, a PPL compiler that generates
model-specific target code for performing inference. Swift
uses a dynamic slicing framework for open-universe models
to incrementally maintain the dependency structure at runtime. In addition, we carefully design data structures in the
target code to avoid dynamic memory allocations when possible. Our experiments show that Swift achieves orders of
magnitudes speedup against other PPL engines.
The next step for Swift is to support more algorithms, such
as SMC [Wood et al., 2014], as well as more samplers, such
as the block sampler for (nearly) deterministic dependencies [Li et al., 2013]. Although FIDS is a general framework
that can be naturally extended to these cases, there are still
implementation details to be studied.
Another direction is partial evaluation, which allows the
compiler to reason about the input program to simplify it.
Shah et al. [2016] proposes a partial evaluation framework
for the inference code (PI in Fig. 1). It is very interesting to
extend this work to the whole Swift pipeline.
Acknowledgement
We would like to thank our anonymous reviewers, as well
as Rohin Shah for valuable discussions. This work is supported by the DARPA PPAML program, contract FA875014-C-0011.
References
[Agrawal and Horgan, 1990] Hiralal Agrawal and Joseph R Horgan. Dynamic program slicing. In ACM SIGPLAN Notices, volume 25, pages 246–256. ACM, 1990.
[Arora et al., 2010] Nimar S. Arora, Rodrigo de Salvo Braz, Erik B.
Sudderth, and Stuart J. Russell. Gibbs sampling in open-universe
stochastic languages. In UAI, pages 30–39. AUAI Press, 2010.
[Chaganty et al., 2013] Arun Chaganty, Aditya Nori, and Sriram
Rajamani. Efficiently sampling probabilistic programs via program analysis. In AISTATS, pages 153–160, 2013.
[Cytron et al., 1989] Ron Cytron, Jeanne Ferrante, Barry K Rosen,
Mark N Wegman, and F Kenneth Zadeck. An efficient method
of computing static single assignment form. In Proceedings of
the 16th ACM SIGPLAN-SIGACT symposium on Principles of
programming languages, pages 25–35. ACM, 1989.
[Goodman et al., 2008] Noah D. Goodman, Vikash K. Mansinghka, Daniel M. Roy, Keith Bonawitz, and Joshua B. Tenenbaum. Church: A language for generative models. In UAI, pages
220–229, 2008.
[Herbrich et al., 2007] Ralf Herbrich, Tom Minka, and Thore Graepel. Trueskill(tm): A bayesian skill rating system. In Advances in
Neural Information Processing Systems 20, pages 569–576. MIT
Press, January 2007.
[Homan and Gelman, 2014] Matthew D Homan and Andrew Gelman. The no-u-turn sampler: Adaptively setting path lengths in
hamiltonian monte carlo. The Journal of Machine Learning Research, 15(1):1593–1623, 2014.
[Hur et al., 2014] Chung-Kil Hur, Aditya V. Nori, Sriram K. Rajamani, and Selva Samuel. Slicing probabilistic programs. In Proceedings of the 35th ACM SIGPLAN Conference on Programming Language Design and Implementation, PLDI ’14, pages
133–144, New York, NY, USA, 2014. ACM.
[Kazemi and Poole, 2016] Seyed Mehran Kazemi and David Poole.
Knowledge compilation for lifted probabilistic inference: Compiling to a low-level language. 2016.
[Kucukelbir et al., 2015] Alp Kucukelbir, Rajesh Ranganath, Andrew Gelman, and David Blei. Automatic variational inference
in Stan. In NIPS, pages 568–576, 2015.
[LeCun et al., 1998] Yann LeCun, Léon Bottou, Yoshua Bengio,
and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324,
1998.
[Li et al., 2013] Lei Li, Bharath Ramsundar, and Stuart Russell.
Dynamic scaled sampling for deterministic constraints. In AISTATS, pages 397–405, 2013.
[Lunn et al., 2000] David J. Lunn, Andrew Thomas, Nicky Best,
and David Spiegelhalter. WinBUGS - a Bayesian modelling
framework: Concepts, structure, and extensibility. Statistics and
Computing, 10(4):325–337, October 2000.
[Mansinghka et al., 2013] Vikash K. Mansinghka, Tejas D. Kulkarni, Yura N. Perov, and Joshua B. Tenenbaum. Approximate Bayesian image interpretation using generative probabilistic
graphics programs. In NIPS, pages 1520–1528, 2013.
[McAllester et al., 2008] David McAllester, Brian Milch, and
Noah D Goodman. Random-world semantics and syntactic independence for expressive languages. Technical report, 2008.
[Milch and Russell, 2006] Brian Milch and Stuart J. Russell.
General-purpose MCMC inference over relational structures. In
UAI. AUAI Press, 2006.
[Milch et al., 2005a] Brian Milch, Bhaskara Marthi, Stuart Russell,
David Sontag, Daniel L. Ong, and Andrey Kolobov. BLOG:
Probabilistic models with unknown objects. In IJCAI, pages
1352–1359, 2005.
[Milch et al., 2005b] Brian Milch, Bhaskara Marthi, David Sontag, Stuart Russell, Daniel L. Ong, and Andrey Kolobov. Approximate inference for infinite contingent Bayesian networks.
In Tenth International Workshop on Artificial Intelligence and
Statistics, Barbados, 2005.
[Minka et al., 2014] T. Minka, J.M. Winn, J.P. Guiver, S. Webster,
Y. Zaykov, B. Yangel, A. Spengler, and J. Bronskill. Infer.NET
2.6, 2014.
[Nori et al., 2014] Aditya V Nori, Chung-Kil Hur, Sriram K Rajamani, and Selva Samuel. R2: An efficient MCMC sampler for
probabilistic programs. In AAAI Conference on Artificial Intelligence, 2014.
[Pfeffer, 2001] Avi Pfeffer. IBAL: A probabilistic rational programming language. In In Proc. 17th IJCAI, pages 733–740. Morgan
Kaufmann Publishers, 2001.
[Pfeffer, 2009] Avi Pfeffer. Figaro: An object-oriented probabilistic programming language. Charles River Analytics Technical
Report, 2009.
[Plummer, 2003] Martyn Plummer. JAGS: A program for analysis
of Bayesian graphical models using Gibbs sampling. In Proceedings of the 3rd International Workshop on Distributed Statistical
Computing, volume 124, page 125. Vienna, 2003.
[Ritchie et al., 2016] Daniel Ritchie, Andreas Stuhlmüller, and
Noah D. Goodman. C3: Lightweight incrementalized MCMC for
probabilistic programs using continuations and callsite caching.
In AISTATS 2016, 2016.
[Sanderson, 2010] Conrad Sanderson. Armadillo: An open source
c++ linear algebra library for fast prototyping and computationally intensive experiments. 2010.
[Shah et al., 2016] Rohin Shah, Emina Torlak, and Rastislav Bodik.
SIMPL: A DSL for automatic specialization of inference algorithms. arXiv preprint arXiv:1604.04729, 2016.
[Stan Development Team, 2014] Stan Development Team. Stan
Modeling Language Users Guide and Reference Manual, Version
2.5.0, 2014.
[Tipping and Bishop, 1999] Michael E. Tipping and Chris M.
Bishop. Probabilistic principal component analysis. Journal of
the Royal Statistical Society, Series B, 61:611–622, 1999.
[Tristan et al., 2014] Jean-Baptiste Tristan, Daniel Huang, Joseph
Tassarotti, Adam C Pocock, Stephen Green, and Guy L Steele.
Augur: Data-parallel probabilistic modeling. In NIPS, pages
2600–2608. 2014.
[Wingate et al., 2011] David Wingate, Andreas Stuhlmueller, and
Noah D Goodman. Lightweight implementations of probabilistic programming languages via transformational compilation. In
International Conference on Artificial Intelligence and Statistics,
pages 770–778, 2011.
[Wood et al., 2014] Frank Wood, Jan Willem van de Meent, and
Vikash Mansinghka. A new approach to probabilistic programming inference. In Proceedings of the 17th International conference on Artificial Intelligence and Statistics, pages 2–46, 2014.
[Wu et al., 2014] Yi Wu, Lei Li, and Stuart J. Russell. BFiT:
From possible-world semantics to random-evaluation semantics
in open universe. In Neural Information Processing Systems,
Probabilistic Programming workshop, 2014.
[Yang et al., 2014] Lingfeng Yang, Pat Hanrahan, and Noah D.
Goodman. Generating efficient MCMC kernels from probabilistic programs. In AISTATS, pages 1068–1076, 2014.
| 6cs.PL
|
Proceedings of Machine Learning Research vol 75:1–15, 2018
Stable recovery of deep linear networks under sparsity constraints
arXiv:1706.00342v2 [math.OC] 20 Feb 2018
François Malgouyres
MALGOUYRES @ MATH . UNIV- TOULOUSE . FR
Institut de Mathématiques de Toulouse ; UMR5219
Université de Toulouse ; CNRS
UPS IMT F-31062 Toulouse Cedex 9, France Address 1
Joseph Landsberg
Department of Mathematics
Mailstop 3368
Texas A& M University
JML @ MATH . TAMU . EDU
Abstract
We study a deep linear network expressed under the form of a matrix factorization problem. It
takes as input a matrix X obtained by multiplying K matrices (called factors and corresponding
to the action of a layer). Each factor is obtained by applying a fixed linear operator to a vector of
parameters satisfying a sparsity constraint. In machine learning, the error between the product of
the estimated factors and X (i.e. the reconstruction error) relates to the statistical risk. The stable
recovery of the parameters defining the factors is required in order to interpret the factors and the
intermediate layers of the network.
In this paper, we provide sharp conditions on the network topology under which the error on the
parameters defining the factors (i.e. the stability of the recovered parameters) scales linearly with
the reconstruction error (i.e. the risk). Therefore, under these conditions on the network topology,
any successful learning tasks leads to robust and therefore interpretable layers.
The analysis is based on the recently proposed Tensorial Lifting. The particularity of this
paper is to consider a sparse prior. As an illustration, we detail the analysis and provide sharp
guarantees for the stable recovery of convolutional linear network under sparsity prior. As expected,
the condition are rather strong.
Keywords: Stable recovery, deep linear networks, convolutional linear networks, feature robustess.
1. Introduction
Let K ∈ N∗ , m1 . . . mK+1 ∈ N, write m1 = m, mK+1 = n. We impose the factors to be structured
matrices defined by a number S of unknown parameters. More precisely, for k = 1 . . . K, let
Mk : RS
−→ Rmk ×mk+1 ,
h 7−→ Mk (h)
be a linear map. We assume that we know the matrix X ∈ Rm×n which is provided by
X = M1 (h1 ) · · · MK (hK ) + e,
(1)
for an unknown error term e, such that kek ≤ δ, and parameters h = (hk )k=1..K ∈ RS×K .
Moreover, considering a family of possible supports M (e.g., all the supports of size S ′ , for a
c 2018 J. Landsberg.
given S ′ ≤ S), we assume that theh satisfy a sparsity constraint
of the form : there exists S =
(S k )k=1..K ∈ M such that supp h ⊂ S (i.e.: ∀k, supp hk ⊂ S k ).
This work investigates necessary and sufficient conditions imposed on the constituents of (1) for
which we can (up to obvious scale rearrangement) stably recover the parameters h from X. Beside
such that
these conditions, we assume that we have a way to find S ∗ ∈ M and h∗ ∈ RS×K
S∗
η = kM1 (h∗1 ) . . . MK (h∗K ) − Xk
, is small.
(2)
As we will discuss later, at the writing, the success of algorithms for constructing h∗ is mostly
supported by empirical evidence and lack theoretical justifications. These aspects of the problem
are out of the scope of the present paper. However, in machine learning problems, the reconstruction
error η represents the risk. There is therefore no point in analyzing the properties of h∗ , if η is large.
The established upper-bound on the recovery error of the parameters linearly depends on δ + η.
Therefore, when the learning algorithm is successful (i.e. the sum of the risk η and noise δ is
sufficiently small), if the deep linear network satisfies the conditions established in this paper the
estimation of the parameters is stable. The latter property is required if one wants to interpret the
features provided by the machine learning algorithm. That is the main interest of the proposed
analysis. Notice that we also establish that the conditions are sharp.
Also, the study considers deep linear networks instead of deep neural networks. As can be
deduced from Eldan and Shamir (2016), this significantly diminishes the expressiveness of the network. The main argument for studying deep linear networks (as is done in the present paper)
comes from a remark in Safran and Shamir (2016). For the rectified linear unit activation function
(ReLU)1 , between each layer every entry is multiplied by an element of the discrete set {0, 1}. As a
consequence, the parameter space RS×K can be partitioned into subsets such that, on every subset,
the action of the non-linear network is the same deep linear network (i.e. the activation function has
a constant action when h varies in the subset). Therefore, the objective function optimized in deep
learning is made of pieces and on every piece it is the objective function of a deep linear network.
As a consequence, properties of the objective function for deep neural networks generalize properties of the objective function for deep linear networks. Restricting the analysis to linear networks is
legitimate as a step towards the study of deep neural networks.
Notice, that the authors of Choromanska et al. (2015a,b); Kawaguchi (2016) use a different
argument but also end-up studying deep linear networks. The simplifying assumption assumes the
independence of the activation to the input. Taking the expectation then leads to linear networks
that the authors analyse. As explained by the same authors in Choromanska et al. (2015b), this is
however a moderatly convincing argument. We prefer to say clearly that we consider deep linear
networks.
Finally, S ∗ and h∗ are typically found by an algorithm (most often a heuristic) that tries to lower
kM1 (h1 ) . . . MK (hK ) − Xk2
(3)
while avoiding overfit. A classical strategy is the dropout of Srivastava et al. (2014). This is perfectly compatible with the assumption (2). However, even if we ignore the overfit issue and restrict
the analysis to the minimization of (3), we see that it is non-convex. Again, we do not address this
minimization issue but there is significant empirical evidence suggesting that (3) can be minimized
efficiently in a surprisingly large number of situations. Despite an increasing theoretical activity
1. ReLU is the most common activation function.
2
S TABLE
RECOVERY OF DEEP LINEAR NETWORKS UNDER SPARSITY CONSTRAINTS
related to that question the theory explaining this phenomenon is still far from satisfactory when
K ≥ 3 (see Livni et al. (2014); Haeffele and Vidal (2015); Kawaguchi (2016); Choromanska et al.
(2015a,b); Safran and Shamir (2016) ).
The approach developed in this paper extends to K ≥ 3 existing results for K ≤ 2. In particular,
when K = 1, the considered problems boils down to a compressed sensing problem Elad (2010).
When K = 2 and when extended to other constraints on the parameters h, the statements apply
to already studied problems such as: low rank approximation Candes et al. (2013), Non-negative
matrix factorization Lee and Seung (1999); Donoho and Stodden (2003); Laurberg et al. (2008);
Arora et al. (2012), dictionary learning Jenatton et al. (2012), phase retrieval Candes et al. (2013),
blind deconvolution Ahmed et al. (2014); Choudhary and Mitra (2014); Li et al. (2016). Most of
these papers use the same lifting property we are using. They further propose to convexify the problem. A more general bilinear framework is considered in Choudhary and Mitra (2014). The only
existing statements when K ≥ 3 are very recent Malgouyres and Landsberg (2017). They are also
applied to deep linear networks but do not include sparsity constraint.
The present work describes an alternative analysis, specialized to sparsity constraints, of the
results exposed in Malgouyres and Landsberg (2017). Doing so, we obtain better bounds (defined
with an analogue of the lower-RIP) and weaker constraints on the model. Its application to sparse
convolutional linear networks leads to simple necessary and sufficient conditions of stable recovery,
for a large class of solvers. The stability inequality (see Theorem 5) only involves explicit and
simple ingredients of the problem. The condition on the network topology is rather strong but
takes an simple format. Implementing a test checking if the condition is met is easy and the test
only requires to apply the networks as many times as the network has leaves, for every couple of
supports.
2. Notations and preliminaries on Tensorial Lifting
= {h ∈ RS×K , ∀k = 1..K, khk k =
6 0}. Define an equivalence
Set NK = {1, . . . , K} and RS×K
∗
S×K
S×K
, h ∼ g if and only if there exists (λk )k=1..K ∈ RK such
relation in R∗ : for any h, g ∈ R
that
K
Y
λk = 1
and
∀k = 1..K, hk = λk gk .
k=1
Denote the equivalence class of h ∈ RS×K
by [h]. For any p ∈ [1,∞], we denote the usual ℓp norm
∗
S×K
by k.kp and define the mapping dp : (R∗ / ∼) × (RS×K
/ ∼) → R by
∗
dp ([h], [g]) =
inf
S×K
h′ ∈[h]∩Rdiag
kh′ − g′ kp
,
, ∀h, g ∈ RS×K
∗
(4)
S×K
g′ ∈[g]∩Rdiag
where
RS×K
= {h ∈ RS×K
, ∀k = 1..K, khk k∞ = kh1 k∞ }.
diag
∗
/ ∼.
It is proved in Malgouyres and Landsberg (2017) that dp is a metric on RS×K
∗
The real valued tensors of order K whose axes are of size S are denoted by T ∈ RS×...×S . The
K
K
space of tensors is abbreviated RS . We say that a tensor T ∈ RS is of rank 1 if and only if there
exists a collection of vectors h ∈ RS×K such that, for any i = (i1 , . . . , iK ) ∈ NK
S,
Ti = h1,i1 . . . hK,iK .
3
The set of all the tensors of rank 1 is denoted by Σ1 . Moreover, we parametrize Σ1 ⊂ RS
the Segre embedding
K
using
K
P : RS×K −→ Σ1 ⊂ RS
h 7−→ (h1,i1 h2,i2 . . . hK,iK )i∈NK
(5)
S
As stated in the two next theorems, we can control the distortion of the distance induced by P
and its inverse.
Theorem 1 Stability of [h] from P (h), see Malgouyres and Landsberg (2017)
Let h and g ∈ RS×K
be such that kP (g) − P (h)k∞ ≤ 21 max (kP (h)k∞ , kP (g)k∞ ). For all
∗
p, q ∈ [1, ∞],
1
1
1
−1
−1
K
K
, kP (g)k∞
kP (h) − P (g)kq .
(6)
dp ([h], [g]) ≤ 7(KS) p min kP (h)k∞
Theorem 2 Lipschitz continuity of P , see Malgouyres and Landsberg (2017)
We have for any q ∈ [1, ∞] and any h and g ∈ RS×K
,
∗
1
1
K−1
1
1− K
1− K
1−
kP (h) − P (g)kq ≤ S q K q max kP (h)k∞ , kP (g)k∞
dq ([h], [g]).
(7)
The Tensorial Lifting (see Malgouyres and Landsberg (2017)) states that there exists a unique
linear map
K
A : RS −→ Rm×n ,
such that for all h ∈ RS×K
M1 (h1 )M2 (h2 ) . . . MK (hK ) = AP (h).
(8)
The intuition leading to this equality is that every entry in M1 (h1 )M2 (h2 ) . . . MK (hK ) is a multivariate polynomial whose variables are in h. Moreover, every monomial of the polynomials is
of the form ai P (h)i for i ∈ NK
S , where ai is a coefficient depending on M1 , . . . , MK . The great
property of the Tensorial Lifting is to express any deep linear network using the Segre Embedding
and a linear operator A. The Segre embedding is non-linear and might seem difficult to deal with at
the first sight, but it is always the same whatever the network topology, the sparsity pattern, the action of the ReLU activation function. . . These constituents of the problem only influence the lifting
linear operator A.
In the next section, we study what properties of A are required to obtain the stable recovery. In
Section 4, we study these properties when A corresponds to a sparse convolutional linear network.
3. General conditions for the stable recovery under sparsity constraint
From now on, the analysis differs from the one presented in Malgouyres and Landsberg (2017). It is
dedicated to models that enforce sparsity. In this particular situation, we can indeed have a different
view of the geometry of the problem. In order to describe it, we first establish some notation.
4
S TABLE
RECOVERY OF DEEP LINEAR NETWORKS UNDER SPARSITY CONSTRAINTS
We define a support by S = (Sk )k=1..K , with Sk ⊂ NS , and denote the set of all supports by
K
K
P(NK
S ) (the parts of NS ). For a given support S ∈ P(NS ), we denote
= {h ∈ RS×K | hk,i = 0, for all k = 1..K and i 6∈ Sk }
RS×K
S
(i.e., for all k, supp (hk ) ⊂ Sk ) and
RSS
K
K
= {T ∈ RS | Ti = 0, if ∃k = 1..K, such that ik 6∈ Sk }.
K
K
We also denote by PS the orthogonal projection from RS onto RSS . We trivially have for all
K
T ∈ RS and all i ∈ NK
S
Ti , if i ∈ S,
(PS T )i =
0 , otherwise.
As explained in the introduction, we assume that there exists a known family of admissible
S×K
supports M ⊂ P(NK
that we
S ), an unknown support S ∈ M and unknown parameters h ∈ RS
would like to estimate from the noisy matrix product
X = M1 (h1 ) . . . MK (hK ) + e.
(9)
We assume that there exists δ ≥ 0 such that the error satisfies
kek ≤ δ.
(10)
Also, we consider an inexact minimization and assume that we have a way to find S ∗ ∈ M and
h∗ ∈ RS×K
S∗
η = kM1 (h∗1 ) . . . MK (h∗K ) − Xk
is small.
We remind that, in machine learning problems, η represents the risk.
In the geometrical view described in the sequel, we consider different linear operators AS , with
S×K
S ∈ P(NK
S ), such that for all h ∈ RS
AS P (h) = M1 (h1 ) . . . MK (hK ).
In order to achieve that, considering (8), we simply define for any S ∈ P(NK
S)
AS = APS .
(11)
The following property will turn out to be necessary and sufficient to guarantee the stable recovery property.
Definition 1 Deep-M-Null Space Property
Let γ ≥ 1 and ρ > 0, we say that A satisfies the deep-M-Null Space Property (deep-M-NSP )
)+P (RS×K
) satisfying
with constants (γ, ρ) if and only if for all S and S ′ ∈ M, any T ∈ P (RS×K
S
S′
′
kAS∪S ′ T k ≤ ρ and any T ∈ Ker (AS∪S ′ ), we have
kT k ≤ γkT − PS∪S ′ T ′ k.
5
(12)
)+
Geometrically, the deep-M-NSP does not hold when PS∪S ′ Ker (AS∪S ′ ) intersects P (RS×K
S
P (RS×K
)
away
from
the
origin
or
tangentially
at
0.
It
holds
when
the
two
sets
intersect
”transverS′
sally” at 0. Despite an apparent abstract nature, we will be able to characterize precisely when
the lifting operator corresponding to a convolutional linear network satisfies the deep-M-NSP (see
Section 4). We will also be able to calculate the constants (γ, ρ).
Proposition 1 Sufficient condition for deep-M-NSP
K
If Ker (A) ∩ RSS∪S ′ = {0}, for all S and S ′ ∈ M, then A satisfies the deep-M-NSP with
constants (γ, ρ) = (1, +∞).
Proof In order to prove the proposition, let us consider S and S ′ ∈ M, T ′ ∈ Ker (AS∪S ′ ). We have
K
APS∪S ′ T ′ = 0 and therefore PS∪S ′ T ′ ∈ Ker (A). Moreover, by definition, PS∪S ′ T ′ ∈ RSS∪S ′ .
Therefore, applying the hypothesis of the proposition, we obtain PS∪S ′ T ′ = 0 and (12) holds for
any T , when γ = 1. Therefore, A satisfies the deep-M-NSP with constants (γ, ρ) = (1, +∞).
If NK
S ∈ M, the condition becomes Ker (A) = {0}, which is sufficient but obviously not necessary
for the deep-M-NSP to hold. However, when M truly imposes sparsity, the condition Ker (A) ∩
K
RSS∪S ′ = {0} says that the elements of Ker (A) shall not be sparse in some (tensorial) way. This
nicely generalizes the case K = 1.
Definition 2 Deep-lower-RIP constant
There exists a constant σM > 0 such that for any S and S ′ ∈ M and any T in the orthogonal
complement of Ker (AS∪S ′ )
(13)
σM kPS∪S ′ T k ≤ kAS∪S ′ T k.
We call σM Deep-lower-RIP constant of A with regard to M.
Proof The existence of σM is a straightforward consequence of the fact that the restriction of AS∪S ′
on the orthogonal complement of Ker (AS∪S ′ ) is injective. We therefore have for all T in the
orthogonal complement of Ker (AS∪S ′ )
kAS∪S ′ T k ≥ σS∪S ′ kT k ≥ σS∪S ′ kPS∪S ′ T k,
where σS∪S ′ > 0 is the smallest non-zero singular value of AS∪S ′ . We obtain the existence of σM
by taking the minimum of the constants σS∪S ′ over the finite family of S and S ′ ∈ M.
Theorem 3 Sufficient condition for stable recovery
Assume A satisfies the deep-M-NSP with the constants γ ≥ 1, ρ > 0. For any S ∗ ∈ M and
∗
as in (2) with η + δ ≤ ρ, we have
h ∈ RS×K
S∗
kP (h∗ ) − P (h)k ≤
γ
(δ + η),
σM
where σM is the Deep-lower-RIP constant of A with regard to M.
Moreover, if σγM (δ + η) ≤ 12 max kP (h∗ )k∞ , kP (h)k∞ , then
1
1
1
γ
−1
∗
∗ K −1
K
p
dp ([h ], [h]) ≤ 7(KS) min kP (h)k∞ , kP (h )k∞
(δ + η).
σM
6
S TABLE
RECOVERY OF DEEP LINEAR NETWORKS UNDER SPARSITY CONSTRAINTS
Proof We have
kAS ∗ ∪S (P (h∗ ) − P (h))k = kAS ∗ ∪S P (h∗ ) − AS ∗ ∪S P (h)k
= kAP (h∗ ) − AP (h)k
≤ kAP (h∗ ) − Xk + kAP (h) − Xk
≤ δ+η
If we further decompose (the decomposition is unique)
where T ′ ∈ Ker AS ∗ ∪S
P (h∗ ) − P (h) = T + T ′ ,
and T is orthogonal to Ker AS ∗ ∪S , we have
kAS ∗ ∪S (P (h∗ ) − P (h))k = kAS ∗ ∪S T k ≥ σM kPS ∗ ∪S T k,
where σM is the Deep-lower-RIP constant of A with regard to M. We finally obtain, since
PS ∗ ∪S P (h∗ ) = P (h∗ ) and PS ∗ ∪S P (h) = P (h),
kP (h∗ ) − P (h) − PS ∗ ∪S T ′ k = kPS ∗ ∪S T k ≤
δ+η
.
σM
Since A satisfies the deep-M-NSP with constants (γ, ρ) and δ + η ≤ ρ, we have
kP (h∗ ) − P (h)k ≤ γkP (h∗ ) − P (h) − PS ∗ ∪S T ′ k
δ+η
≤ γ
σM
When δ + η satisfy the condition in the theorem, we can apply Theorem 1 and obtain the last inequality.
Theorem 3 differs from the analogous theorem in Malgouyres and Landsberg (2017). In particular, it is dedicated to sparsity constraint with much weaker hypotheses on A. The constant of the
upper bound is also different.
One might again ask whether the condition “A satisfies the deep-M-NSP ” is sharp or not. As
stated in the following proposition, the answer is affirmative.
Theorem 4 Necessary condition for stable recovery
Assume the stable recovery property holds: There exists M, C and δ > 0 such that for any
, any X = AP (h) + e, with kek ≤ δ, and any S ∗ ∈ M and
S ∈ M and any h ∈ RS×K
S
S×K
∗
h ∈ RS ∗ such that
kAP (h∗ ) − Xk ≤ kek
we have
∗
1
−1
K
d2 ([h ], [h]) ≤ C min kP (h)k∞
Then, A satisfies the deep-M-NSP with constants
K−1 √
K σmax
γ = CS 2
∗
1
−1
K
, kP (h )k∞
and
kek.
ρ = δ,
where σmax is the spectral radius of A.
The proof is very similar to the proof of the Theorem 6, in Malgouyres and Landsberg (2017) and
the proof of the analogous converse statement in Cohen et al. (2009). It is provided in Appendix A.
7
edges of depth 3
2
1
RN
RN
RN root r
leaves
RN
RN
Figure 1: Example of the considered convolutional linear network. To every edge/neuron is attached a convolution kernel. The network does not involve non-linearities or sampling.
4. Application to convolutional linear network under sparsity prior
We consider a sparse convolutional linear network as depicted in Figure 1. The network typically
aims at performing a linear analysis or synthesis of a signal living in RN . The considered convolutional linear network is defined from a rooted directed acyclic graph G(E, N ) composed of nodes
N and edges E. Each edge connects two nodes. The root of the graph is denoted by r and the set
containing all its leaves is denoted by F. We denote by Pa the set of all paths connecting the leaves
and the root. We assume, without loss of generality, that the length of any path between any leaf
and the root is independent of the considered leaf and equal to some constant K ≥ 0. We also
assume that, for any edge e ∈ E, the number of edges separating e and the root is the same for all
paths between e and r. This length is called the depth of e. For any k = 1..K, we denote the set
containing all the edges of depth k, by E(k). For e ∈ E(k), we also say that e belongs to the layer
k.
Moreover, to any edge e is attached
Pa convolution kernel of maximal support Se ⊂ NN . We assume (without loss of generality) that e∈E(k) |Se | is independent of k (|Se | denotes the cardinality
of Se ). We take
X
S=
|Se |.
e∈E(1)
For any edge e, we consider the mapping Te : RS −→ RN that maps any h ∈ RS into the convolution kernel he , attached to the edge e, whose support is Se . It simply writes at the right location
(i.e. those in Se ) the entries of h defining the kernel on the edge e. As in the previous section, we
assume a sparsity constraint and will only consider a family M of possible supports S ⊂ NK
S.
At each layer k, the convolutional linear network computes, for all e ∈ E(k), the convolution
between the signal at the origin of e; then, it attaches to any ending node the sum of all the convolutions arriving at that node. Examples of such convolutional linear networks includes wavelets,
wavelet packets Mallat (1998) or the fast transforms optimized in Chabiron et al. (2014, 2016). It
is similar to the usual convolutional neural network except that the linear network does not involve
any non-linearity and the supports are not fixed. It is clear that the operation performed at any layer
depends linearly on the parameters h ∈ RS and that its results serves as inputs for the next layer.
8
S TABLE
RECOVERY OF DEEP LINEAR NETWORKS UNDER SPARSITY CONSTRAINTS
The convolutional linear network therefore depends on parameters h ∈ RS×K and takes the form
X = M1 (h1 ) . . . MK (hK ),
where the operators Mk satisfy the hypothesis of the present paper.
This section applies the results of the preceding section
in order to identify conditions such that
any unknown parameters h ∈ RS×K satisfying supp h ⊂ S, for a given S ∈ M, can be stably
recovered from X = M1 (h1 ) . . . MK (hK ) (possibly corrupted by an error).
In order to do so, let us define a few notations. Notice first that, we apply the convolutional
linear network to an input x ∈ RN |F | , where x is the concatenation of the signals xf ∈ RN for
f ∈ F. Therefore, X is the (horizontal) concatenation of |F| matrices X f ∈ RN ×N such that
X
Xx =
X f xf
, for all x ∈ RN |F | .
f ∈F
Let us consider the convolutional linear network defined by h ∈ RS×K as well as f ∈ F and
n = 1..N . The column of X corresponding to the entry n in the leaf f is the translation by n of
X
T p (h)
(14)
p∈Pa (f )
where Pa (f ) contains all the paths of Pa starting from the leaf f and
T p (h) = Te1 (h1 ) ∗ . . . ∗ TeK (hK )
, where p = (e1 , . . . , eK ).
Moreover, we define for any k = 1..K the mapping ek : NS −→ E(k) which provides for any
i = 1..S the unique edge of E(k) such that the ith entry of h ∈ RS contributes to Tek (i) (h). Also,
for any i ∈ NK
S , we denote pi = (e1 (i1 ), . . . , eK (iK )) and, for any S ∈ M,
IS = i ∈ NK
S |i ∈ S and pi ∈ Pa .
The latter contains all the indices of S corresponding to a valid path in the network. For any set of
parameters h ∈ RS×K and any path p ∈ Pa , we also denote by hp the restriction of h to its indices
i
S×K by
contributing to the kernels on the path p. We also define, for any i ∈ NK
S,h ∈R
1 , if j = ik
i
hk,j =
, for all k = 1..K and j = 1..S.
(15)
0 otherwise
We can deduce from (14) that, when i ∈ IS , AP (hi ) simply convolves the entries at one leaf with a
dirac delta function. Thefore, all the entries of AP (hi ) are in {0, 1} and we denote Di = {(i, j) ∈
NN × NN |F | |AP (hi )i,j = 1}.
We also denote 1 ∈ RS a vector of size S with all its entries equal to 1. For any edge e ∈ E,
1e ∈ RS consists of zeroes except for the entries corresponding to the edge e which are equal to 1.
For any S ⊂ NS , we define 1S ∈ RS which consists of zeroes except for the entries corresponding
to the indexes in S. For any p = (e1 , . . . , eK ) ∈ Pa , the support of M1 (1e1 ) . . . MK (1eK ) is
denoted by D p .
Finally, we remind that because of (8), there exists a unique mapping
A : RS
K
−→ RN ×N |F |
9
such that
, for all h ∈ RS×K ,
AP (h) = M1 (h1 ) . . . MK (hK )
where P is the Segre embedding defined in (5).
Proposition 2 Necessary condition of identifiability of a sparse network
′
′
Either RS×K is not identifiable or, for any S and S ′ ∈ M, all the entries of M1 (1S∪S ) . . . MK (1S∪S )
belong to {0, 1}. When the latter holds :
′
1. For any distinct p and p′ ∈ Pa , we have D p ∩ D p = ∅.
K
2. Ker (AS∪S ′ ) = {T ∈ RS |∀i ∈ IS∪S ′ , Ti = 0}.
Proof
′
′
Let us assume that: There exist S and S ′ ∈ M and an entry of M1 (1S∪S ) . . . MK (1S∪S )
that does not belong to {0, 1}.
Using (14), we know that there is f ∈ F and n = 1..N such that
X
T p (1) n ≥ 2.
p∈Pa (f )
As a consequence, there is i and j ∈ S ∪ S ′ with i 6= j and
T pi (hi ) n = T pj (hj ) n = 1.
Therefore,
AP (hi ) = AP (hj )
and the network is not identifiable. This proves the first statement.
′
′
Let us assume that: For any S and S ′ ∈ M, all the entries of M1 (1S∪S ) . . . MK (1S∪S )
belong to {0, 1}.
We immediately observe that (14) leads to the item 1 of the Proposition.
To prove the second item, we can easily check that (P (hi ))i6∈IS∪S ′ forms a basis of {T ∈
K
RS |∀i ∈ IS∪S ′ , Ti = 0}. We can also easily check using (14) and (11) that, for any i 6∈ IS∪S ′ ,
0
, if i 6∈ S ∪ S ′
i
AS∪S ′ P (h ) =
M1 (hi1 ) . . . MK (hiK ) = 0 , if i ∈ S ∪ S ′ and pi 6∈ Pa
K
As a consequence, {T ∈ RS |∀i ∈ IS∪S ′ , Ti = 0} ⊂ Ker (AS∪S ′ ).
To prove the converse inclusion, we observe that for any distinct i and j ∈ IS∪S ′ , we have
Di ∩ Dj = ∅. This implies that
K
rk (AS∪S ′ ) ≥ |IS∪S ′ | = S K − dim({T ∈ RS |∀i ∈ IS∪S ′ , Ti = 0}).
K
Finally, we deduce that dim(Ker (AS∪S ′ )) ≤ dim({T ∈ RS |∀i ∈ IS∪S ′ , Ti = 0}) and therefore
K
Ker (AS∪S ′ ) = {T ∈ RS |∀i ∈ IS∪S ′ , Ti = 0}.
10
S TABLE
RECOVERY OF DEEP LINEAR NETWORKS UNDER SPARSITY CONSTRAINTS
Proposition 2 extends Proposition 8 of Malgouyres and Landsberg (2017) by considering several
possible supports. Said differently, Proposition 8 of Malgouyres and Landsberg (2017) corresponds
to Proposition 2 when M = {NK
S }.
The interest of the condition in Proposition 2 is that it can easily be computed by applying the network to dirac delta functions, when |M| is not too large. Notice that, beside the
known examples in blind-deconvolution (i.e. when K = 2 and |Pa | = 1) Ahmed et al. (2014);
Bahmani and Romberg (2015), there are known convolutional linear networks (with K ≥ 2) that
satisfy the condition of the first statement of Proposition 2. For instance, the convolutional linear
network corresponding to the un-decimated Haar wavelet2 transform is a tree and for any of its
leaves f ∈ F, |Pa (f )| = 1. Moreover, the support of the kernel living on the edge e, of depth
k, on this path is {0, 2k }. It is not difficult to check that the first condition of Proposition 2 holds.
Otherwise, it is clear that the necessary condition will be rarely satisfied.
Proposition 3 If |Pa | = 1 and if, for any S and S ′ ∈ M, all the entries of M1 (1S∪S ) . . . MK (1S∪S )
K
belong to {0, 1}, then Ker (AS∪S ′ ) is the orthogonal complement of RSS∪S ′ and A satisfies the deepM-NSP√with constants (γ, ρ) = (1, +∞). Moreover, the deep-lower-RIP of A with regard to M is
σM = N .
′
K
′
Proof The fact that, Ker (AS∪S ′ ) is the orthogonal complement of RSS∪S ′ is a direct consequence
of Proposition 2 and the fact that, when |Pa | = 1, IS∪S ′ = S ∪ S ′ . We then trivialy deduce that,
for any T ′ ∈ Ker (AS∪S ′ ), PS∪S ′ T ′ = 0. A straightforward consequence is that A satisfies the
deep-M-NSP with constants (γ, ρ) = (1, +∞).
To calculate σM , let us considerP
S, S ′ ∈ M and T in the orthogonal complement of Ker (AS∪S ′ ).
We express T under the form T = i∈I ′ Ti P (hi ), where hi is defined (15). Let us also remind
S∪S
that, applying Proposition 2, the supports of AP (hi ) and AP (hj ) are disjoint, when i 6= j. Let us
finally add that, since AP (hj ) is the matrix of a convolution with a Dirac mass, we have |Dj | = N .
We finally have
X
kAT k2 = k
Ti AP (hi )k2 ,
i∈I
= N
X
i∈I
Ti2 = N kT k2 ,
from which we deduce the value of σM .
In the sequel, we establish stability results for a convolutional linear network estimator. In order
to do so, we consider a convolutional linear network of known structure G(E, N ), (Se )e∈E and
M. The convolutional
linear network is defined by unknown parameters h ∈ RS×K satisfying a
constraint supp h ⊂ S for an unknown support S ∈ M. We consider the noisy situation where
X = M1 (h1 ) . . . MK (hK ) + e,
2. Un-decimated means computed with the ”Algorithme à trous”, Mallat (1998), Section 5.5.2 and 6.3.2. The Haar
wavelet is described in Mallat (1998), Section 7.2.2, p. 247 and Example 7.7, p. 235
11
with kek ≤ δ and an estimate h∗ ∈ RS×K such that
kM1 (h∗1 ) . . . MK (h∗K ) − Xk ≤ η.
The equivalence relationship ∼ does not suffise to group parameters leading to the same network
action. Indeed, with networks, we can rescale the kernels on different path differently. Therefore,
we say that two networks sharing the same topology and defined by the parameters h and g ∈ RS×K
are equivalent if and only if
Y
∀p ∈ Pa , ∃(λe )e∈p ∈ Rp , such that
λe = 1 and ∀e ∈ p, Te (g) = λe Te (h).
e∈p
We trivially observe that applying the networks defined by equivalent parameters lead to the same
result. The equivalence class of h ∈ RS×K is denoted by {h}. For any p ∈ [1, +∞], we define
δp ({h}, {g}) =
1
p
X
p p
p
dp ([h ], [g ])
p∈Pa
,
where we remind that hp (resp gp ) denotes the restriction of h (resp g) to the path p and dp is
defined in (4). Since dp is a metric, we easily prove that δp is a metric between network classes.
Theorem 5 If for any S and S ′ ∈ M, all the entries of M1 (1S∪S ) . . . MK (1S∪S ) belong to
{0, 1} and if there exists ε > 0 such that for all e ∈ E, kTe (h)k∞ ≥ ε then
′
′
1
(KS) p
δp ({h }, {h}) ≤ 7 √
(δ + η).
N εK−1
∗
Proof Let us consider a path p ∈ Pa , using (14), since all the entries of M1 (1S∪S ) . . . MK (1S∪S )
belong to {0, 1}, the restriction of the network to p satisfy the same property. Therefore, we can
apply Proposition 3 and Theorem 3 to the restriction of the convolutional linear network to p and
obtain for any p ∈ [1, ∞]
′
1
(KS) p 1−K p
dp ([(h ) ], [h ]) ≤ 7 √
ε
(δ + η p ),
N
∗ p
p
where δp and η p are the restrictions of the errors on D p . Finally, using item 1 of Proposition 2
1
p
1
p
(KS) 1−K
δp ({h }, {h}) ≤ 7 √
ε
(δp + η p )p ,
N
p∈Pa
∗
1
X
(KS) p
≤ 7√
(δ + η).
N εK−1
12
′
S TABLE
RECOVERY OF DEEP LINEAR NETWORKS UNDER SPARSITY CONSTRAINTS
Acknowledgments
Joseph Landsberg is supported by NSF DMS-1405348.
References
Arif Ahmed, Benjamin Recht, and Justin Romberg. Blind deconvolution using convex programming. IEEE Transactions on Information Theory, 60(3):1711–1732, 2014.
Sanjeev Arora, Rong Ge, Ravindran Kannan, and Ankur Moitra. Computing a nonnegative matrix
factorization–provably. In Proceedings of the forty-fourth annual ACM symposium on Theory of
computing, pages 145–162. ACM, 2012.
Sohail Bahmani and Justin Romberg. Lifting for blind deconvolution in random mask imaging:
Identifiability and convex relaxation. SIAM Journal on Imaging Sciences, 8(4):2203–2238, 2015.
Emmanuel J Candes, Thomas Strohmer, and Vladislav Voroninski. Phaselift: Exact and stable
signal recovery from magnitude measurements via convex programming. Communications on
Pure and Applied Mathematics, 66(8):1241–1274, 2013.
Olivier Chabiron, François Malgouyres, Jean-Yves Tourneret, and Nicolas Dobigeon. Toward fast
transform learning. International Journal of Computer Vision, pages 1–22, 2014.
Olivier Chabiron, François Malgouyres, Herwig Wendt, and Jean-Yves Tourneret. Optimization of
a fast transform structured as a convolutional tree. preprint HAL, (hal-01258514), 2016.
Anna Choromanska, Mikael Henaff, Michael Mathieu, Gérard Ben Arous, and Yann LeCun. The
loss surfaces of multilayer networks. In Artificial Intelligence and Statistics, pages 192–204,
2015a.
Anna Choromanska, Yann LeCun, and Gérard Ben Arous. Open problem: The landscape of the loss
surfaces of multilayer networks. In Conference on Learning Theory, pages 1756–1760, 2015b.
Sunav Choudhary and Urbashi Mitra. Identifiability scaling laws in bilinear inverse problems. arXiv
preprint arXiv:1402.2637, 2014.
Albert Cohen, Wolfgang Dahmen, and Ronald DeVore. Compressed sensing and best -term approximation. Journal of the American mathematical society, 22(1):211–231, 2009.
David L Donoho and Victoria Stodden. When does non-negative matrix factorization give a correct
decomposition into parts? 2003.
Michael Elad. Sparse and Redundant Representations: From Theory to Applications in Signal and
Image Processing. Springer, 2010.
Ronen Eldan and Ohad Shamir. The power of depth for feedforward neural networks. In Conference
on Learning Theory, pages 907–940, 2016.
Benjamin D Haeffele and René Vidal. Global optimality in tensor factorization, deep learning, and
beyond. arXiv preprint arXiv:1506.07540, 2015.
13
Rodolphe Jenatton, Rémi Gribonval, and Francis Bach. Local stability and robustness of sparse
dictionary learning in the presence of noise. arxiv, 2012.
Kenji Kawaguchi. Deep learning without poor local minima. In Advances in Neural Information
Processing Systems, pages 586–594, 2016.
Hans Laurberg, Mads Græsbøll Christensen, Mark D Plumbley, Lars Kai Hansen, and Søren Holdt
Jensen. Theorems on positive data: On the uniqueness of nmf. Computational intelligence and
neuroscience, 2008, 2008.
Daniel D Lee and H Sebastian Seung. Learning the parts of objects by non-negative matrix factorization. Nature, 401(6755):788–791, 1999.
Xiaodong Li, Shuyang Ling, Thomas Strohmer, and Ke Wei. Rapid, robust, and reliable
blind deconvolution via nonconvex optimization. CoRR, abs/1606.04933, 2016. URL
http://arxiv.org/abs/1606.04933.
Roi Livni, Shai Shalev-Shwartz, and Ohad Shamir. On the computational efficiency of training
neural networks. In Advances in Neural Information Processing Systems, pages 855–863, 2014.
François Malgouyres and Joseph Landsberg. Stable recovery of the factors from a deep matrix
product and application to convolutional network. arXiv preprint arXiv:1703.08044, 2017.
Stéphane Mallat. A Wavelet Tour of Signal Processing. Academic Press, Boston, 1998.
Itay Safran and Ohad Shamir. On the quality of the initial basin in overspecified neural networks.
In International Conference on Machine Learning, pages 774–782, 2016.
Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.
Dropout: a simple way to prevent neural networks from overfitting. Journal of machine learning
research, 15(1):1929–1958, 2014.
Appendix A. Proof of Theorem 4
′
′
S×K
and
Proof Let S and S ′ ∈ M. Let h ∈ RS×K
h
∈
R
be
such
that
kA
P
(h)
−
P
(h
)
k≤
′
S
S
δ. Throughout the proof, we also consider T ′ ∈ Ker (AS∪S ′ ). We assume that kP (h)k∞ ≤
′
′
kP (h )k∞ . If it is not the case, we simply switch the role of h and h in the definition of X and e,
below. We denote
′
X = AP (h)
and
e = AP (h) − AP (h ).
′
We have X = AP (h ) + e with kek ≤ δ. Moreover, when S and h play the role of S ∗ and h∗ in
the hypothesis, since h ∈ RS×K
and kek ≤ δ, we have
S
′
′
1
K
d2 ([h ], [h]) ≤ CkP (h )k∞
14
−1
kek.
S TABLE
RECOVERY OF DEEP LINEAR NETWORKS UNDER SPARSITY CONSTRAINTS
′
Using the fact that e = AS∪S ′ (P (h) − P (h )), for any T ′ ∈ Ker (AS∪S ′ )
′
kek = kAS∪S ′ (P (h) − P (h ) − T ′ )k,
′
≤ σmax kPS∪S ′ (P (h) − P (h ) − T ′ )k,
′
= σmax kP (h) − P (h ) − PS∪S ′ T ′ k,
where σmax is the spectral radius of A. Therefore,
′
′
1
K
d2 ([h ], [h]) ≤ CkP (h )k∞
−1
′
σmax kP (h) − P (h ) − PS∪S ′ T ′ k,
′
Finally, using Theorem 2 and the fact that kP (h)k∞ ≤ kP (h )k∞ , we obtain
′
K−1
2
′
1
1−
1
′
K 1− 2 kP (h )k∞ K d2 ([h ], [h])
K−1 √
′
≤ CS 2
K σmax kP (h) − P (h ) − PS∪S ′ T ′ k
kP (h ) − P (h)k ≤ S
′
= γkP (h) − P (h ) − PS∪S ′ T ′ k
K−1 √
for γ = CS 2
K σmax .
Summarizing, we conclude that under the hypothesis of the theorem: For any S and S ′ ∈ M
) + P (RS×K
) such that kAT k = kAS∪S ′ T k ≤ δ, we have for any T ′ ∈
and any T ∈ P (RS×K
S
S′
Ker (AS∪S ′ )
kT k ≤ γkT − PS∪S ′ T ′ k.
In words, A satisfies the deep-M-NSP with the constants of Theorem 4.
15
| 10math.ST
|
Learning Deep Similarity Models with Focus Ranking for Fabric Image Retrieval
Daiguo Denga , Ruomei Wanga , Hefeng Wub,∗, Huayong Hea , Qi Lic , Xiaonan Luod
a Sun
Yat-sen University, Guangzhou 510006, China
University of Foreign Studies, Guangzhou 510006, China
c Western Kentucky University, Bowling Green, KY, 42101, USA
d Guilin University of Electronic Technology, Guilin 541004, China
b Guangdong
arXiv:1712.10211v1 [cs.CV] 29 Dec 2017
Abstract
Fabric image retrieval is beneficial to many applications including clothing searching, online shopping and cloth modeling. Learning
pairwise image similarity is of great importance to an image retrieval task. With the resurgence of Convolutional Neural Networks
(CNNs), recent works have achieved significant progresses via deep representation learning with metric embedding, which drives
similar examples close to each other in a feature space, and dissimilar ones apart from each other. In this paper, we propose a
novel embedding method termed focus ranking that can be easily unified into a CNN for jointly learning image representations and
metrics in the context of fine-grained fabric image retrieval. Focus ranking aims to rank similar examples higher than all dissimilar
ones by penalizing ranking disorders via the minimization of the overall cost attributed to similar samples being ranked below
dissimilar ones. At the training stage, training samples are organized into focus ranking units for efficient optimization. We build
a large-scale fabric image retrieval dataset (FIRD) with about 25,000 images of 4,300 fabrics, and test the proposed model on the
FIRD dataset. Experimental results show the superiority of the proposed model over existing metric embedding models.
Keywords: convolutional neural network, fabric image retrieval, metric embedding, focus ranking
2010 MSC: 00-01, 99-00
1. Introduction
Fabric image retrieval, as a special case of generic image
retrieval, benefits a wide range of applications, e.g., large-scale
clothing searching, online shopping and cloth modeling. Given
a query fabric image, a retrieval algorithm is expected to output
fabric images that are identical or similar to the query image. It
remains a challenging task due to severe variations in illuminations, orientations, scales and wrinkles among fabric images.
Traditional methods for image retrieval mainly involve two
critical components: i) design a robust and discriminative image representation, and ii) determine an effective distance or
similarity metric for a given image representation. Image representations used in a traditional method are usually hand-crafted,
e.g., SIFT [1], GIST [2, 3], Bag of Words (BoW) [4], Fisher
Vector (FV) [5, 6], and VLAD [7]. Although achieving reasonable successes in image retrieval, these methods depend heavily on feature engineering. More seriously, the two components
are designed or learned separately, leading to a sub-optimal solution.
Recently, methods were proposed to learn an image representation and a distance or similarity metric jointly [8, 9] based
on Convolutional Neural Networks (CNN) [10, 11], which can
be used seamlessly in image retrieval. Specifically, these methods trained a CNN with metric learning embedding. Two simple yet effective metric learning embedding methods are pair
∗ Corresponding
author.
Email address: [email protected] (Hefeng Wu)
Preprint submitted to Image and Vision Computing
[12, 13] and triplet embedding [14, 15]. These two embedding
methods are optimized to pull samples of different labels apart
from each other and push samples with the same labels close to
each other. The most important advantage of these discriminative models is that they can jointly learn an image representation and semantically meaningful metric, which is more robust
against intra-class variations and inter-class confusions.
An image retrieval system aims to find out samples with the
same labels against a great many negative ones. But pair and
triplet embedding methods model a metric with no more than
one negative image as reference, which are extremely rough approximation to a real setting. In this paper, we propose a novel
embedding method named focus ranking, which can easily be
unified into a CNN for a joint optimization. In particular, the
proposed model aims to rank a sample with the same label (i.e.,
the matched sample) top over all negative ones. Hence, we penalize any ranking disorder by minimizing the overall cost of
the matched samples ranked below any negative ones. At the
training stage, we organize training samples into focus ranking
units, each of which consists of a probe sample, a matched sample, and a reference set, for an efficient optimization. It learns
to rank the matched sample top over all the negative ones in the
reference set.
To the best of our knowledge, there are not fabric image retrieval datasets publicly available. We build a large-scale fabric
image retrieval datasets (FIRD). It contains 4,300 fabrics, each
of which has five to ten instances. We divide the FIRD dataset
into two partitions, via randomly selecting half of the fabrics
as the training set and the rest as the test set. We randomly
January 1, 2018
select 2/5 images as queries for each fabric in the test set and
the rest form the retrieval set. Given a query image, we aim to
find out the ones of the same fabric at the retrieval set. Extensive experiments on FIRD demonstrate that the proposed model
outperforms existing ones by a large margin.
In summary, this paper makes four main contributions to the
community.
[16] hierarchically quantized local descriptors to a hierarchical vocabulary, which improves not only the retrieval efficiency
but also its quality. Jegou et al. [17] presented a contextual
dissimilarity measure for accurate and efficient image search,
while Fraundorfer et al. [18] proposed a binning scheme for
fast hard drive based image search. Jegou et al. [19, 20] developed a more precise representation by integrating hamming
embedding and weak geometric consistency within the inverted
file. Some works also aggregated a set of local descriptors into
global ones for large scale image retrieval, e.g., Fisher Vector (FV) [5, 6], and Vector of Locally Aggregated Descriptors
(VLAD) [7]. Perronnin et al. [21] represented an image as a
Fisher Vector, and compressed Fisher Vectors using the binarization technique to reduce their memory footprint and speed
up the retrieval. VLAD can jointly optimize dimensionality reduction and indexing algorithms, and thus produce a more compact representation for image retrieval. Although achieving significant progress in visual retrieval, these works depended heavily on hand-crafted features, which were not always optimized
for particular tasks.
• We propose a novel focus ranking embedding method,
which aims to rank the matched sample top over any negative ones. This method can be easily unified into a CNN
to learn a deep similarity model by optimizing the image
representation and metric jointly. Compared to existing
embedding methods, it can learn more robust and discriminative representation to reduce intra-class variations
and enhance inter-class discrimination.
• We apply focus ranking unit generation to model training,
which highlights the focus by the negative-positive ratio
in the process of training. It can also help increase the
diversity of training samples, which is greatly important
to learning.
2.2. Deep representation learning
Recently, deep representation learning has been success• We build a large-scale fabric image retrieval dataset (FIRD),
fully applied to various computer vision areas, such as image
which contains 4,300 fabrics and about 25,000 images for
classification [10, 22, 23], object detection [24, 25, 26], pixeltraining and test. To the best of our knowledge, it is the
wise image labeling [27, 28] and human centric analysis [29,
first large-scale dataset for this task.
30]. We review some recent works that applied deep learning
• Extensive experiments demonstrate the superiority of the
to image retrieval [31, 32, 8, 33, 34, 35]. Babenko et al. [31]
proposed method over existing metric embedding methextracted global features for image retrieval from the fully conods. We also carefully evaluate and discuss key componected layer of CNN pre-trained with an image classification
nents of our model that improve the performance.
dataset [36], and demonstrated fine-tuning the network with
annotated target images could boost the retrieval performance.
The rest of this paper is organized as follows. Related works
Their succeeding work [32] simply aggregated the average and
are reviewed in Section 2. The proposed model is presented in
max pooling of the last convolutional layer as image repreSection 3, and the FIRD dataset is described in Section 4. Secsentation and achieved better results than using fully convotion 5 presents an extensive evaluation and comparison between
lutional layer. Perronnin and Larlus [34] proposed a hybrid
the proposed method and several state-of-the-art methods, in
image retrieval architecture that combined the strengths of suaddition to the component analysis of the proposed method.
pervised deep representation and unsupervised Fisher Vectors
Section 6 concludes the paper.
[5, 6]. However, these methods extracted deep features using
the CNN trained with a classification objective function, which
merely focused on feature learning and ignored metric learning.
2. Related Works
In this section, we review three topics most related to this
work according to its objective and the technical components.
These three topics are traditional descriptors, deep representation learning and distance learning.
2.3. Distance learning
Distance metric learning (DML) plays an important role in
many computer vision tasks such as face recognition, person
re-identification and image retrieval. Xing et al. [37] proposed
to learn a distance metric from given examples of similar pairs
and demonstrated the learned distance metric can improve clustering performance significantly. Zheng et al. [38] formulated
distance learning as a probabilistic relative distance comparison model to maximize the likelihood in which the examples
from a matched pair have smaller distance than those from a
mismatched pair. Mignon and Jurie [39] proposed a new distance learning method with sparse pairwise constraints. The
success of deep learning also inspired recent works that applied
neural network models to address the distance learning problem
[40, 41, 42].
2.1. Traditional descriptors
Traditional image retrieval systems mainly focus on designing discriminative hand-crafted image representations for matching two images. Scale-Invariant Feature Transform (SIFT) [1]
was used widely in image retrieval, due to its distinct invariance to image translation, scaling, and rotation. Many works
applied Bag of Features (BOF) using SIFT features for large
scale image retrieval, such as [4]. They first quantized local
feature vectors into visual words, and represented the image by
the frequency histogram of visual words. Nister and Stewenius
2
the scenario that has K categories of fabrics, in which the jth category has nj images. Suppose there are N training images X = {xi |i = 1, 2, 3, ..., N }, so the training set is T =
{(xi , yi )|i = 1, 2, 3, ..., N }, where xi is the i-th training image
and yi ∈ {1, 2, 3, ..., K} denotes the fabric category the image
xi belongs to. Given an image x, there exist some images with
the same label, which are the positive ones and denoted by X + .
All the others are considered to be negative ones, denoted by
X − . A CNN is applied to model the transformation f (·) from
a raw image to its feature vector.
Given an image x, it is expected that the distances between
f (x) and f (x+ ) should be smaller than those between f (x) and
f (x− ), for ∀x+ ∈ X + and ∀x− ∈ X − , i.e.,
The following works are related to our work in spirit of deep
metric embedding. Hu et al. [43] presented a new discriminative deep metric learning (DDML) method for face verification
in the wild. Bell and Bala [41] learned deep metric for visual
search in interior design using contrastive embedding. In [44],
Rippel et al. proposed to maintain an explicit model of the distributions of different classes in representation space and employ it to achieve local discrimination in DML by penalizing
class distribution overlap. Wang et al. [8] used a multi-scale
CNN trained using triplet embedding for learning fine-grained
image similarity directly from the image pixels. Similar ideas
were employed in [14] and [45] for face verification and hash
learning, respectively.
Two common embedding methods are closely related to our
focus ranking embedding method, i.e., pair [12, 13] and triplet
[14, 15] embedding. The goal of all the three embedding methods are to push the samples with the same labels close to each
other and pull the samples of different labels apart from each
other. However, they are quite different in training and optimization. The pair method builds its energy function on a training set composed of image pairs and optimize to assign small
distance to similar pairs and large distance to dissimilar pairs.
The training set of the triplet method is arranged into image
triplets (a probe image, a matched one and a reference one). The
resulting optimization aims to make the distance of the matched
image to the probe image smaller than that of the negative reference one. However, these two methods do not address the challenging fine-grained fabric image retrieval problem well. The
proposed focus ranking method arranges the training set into focus ranking units (see Section 3.3), with much larger number of
reference images in contrast to the matched one. Then our optimization penalizes any ranking disorder that negative images in
the reference set rank over the matched one. Extensive experimental validation demonstrates that the deep similarity model
optimized with our focus ranking embedding method outperforms previous models favorably.
D(f (x), f (x+ )) < D(f (x), f (x− )), ∀x+ ∈ X + , ∀x− ∈ X − ,
(1)
where D(·, ·) denotes the distance between two feature vectors. In this paper, we apply L2 -norm to measure the distance,
i.e., D(f (x), f (y)) = kf (x) − f (y)k22 . Intuitively, we penalize the ranking disorder cases that are D(f (x), f (x+ )) >
D(f (x), f (x− )). Formally, we apply a 0-1 function to measure
the penalty, so the ranking loss of x with respective to X + and
X − can be expressed as
loss(x|X + , X − ) =
X
X
x+ ∈X +
x− ∈X −
1(D(f (x), f (x− ))−D(f (x), f (x+ )) < 0),
(2)
where 1(·) is the indictor function whose value is 1 if the expression is true, and 0 otherwise. The loss function can be formulated
P by summing up the ranking loss over the training set:
L = x∈X loss(x|X + , X − ).
Optimizing the above loss function can minimize the ranking disorder cases and benefit for top ranking the positive samples. However, the 0-1 loss function is non-differential in optimization, so it is difficult to be jointly optimized with the CNN.
One common solution for this problem is to apply a differential upper-bound function of the 0-1 function. Inspired by
[46], we replace the 1(x < 0) by the logistic function σ(x) =
log2 (1 + 2−x ). The loss function can then be expressed as
X X
X
L=
σ(D(f (x), f (x− ))−D(f (x), f (x+ ))).
3. Deep Similarity Model with Focus Ranking
In this section, we describe the proposed deep focus ranking
model in detail. In the training stage, we first arrange training
images into focus ranking units, each of which contains one
probe image, one image of the same fabric (i.e., positive image), and a number of images of different fabrics (i.e., negative images). Then, images in each focus ranking unit are fed
into a CNN to derive features, based on which the “probe-topositive” and “probe-to-negative” distances will be computed,
respectively. Recall that the probe-to-positive distance is expected to be constantly smaller than the probe-to-negative distance. Thus, we penalize ranking disorders in all units. For
each pair of input images, the trained CNN will be applied to
extract their features separately and then compute their distance
directly in the test stage.
x∈X x+ ∈X + x− ∈X −
(3)
It is worth noting that the proposed model successfully integrates feature extraction and metric learning. In a traditional
method, features are usually hand-crafed, and the selection of a
metric is done separately from feature extraction. In this work,
we take advantage of the CNN for feature learning and optimize
the two components jointly. We will introduce the detailed network architecture and the optimization in the following subsections.
3.2. Network architecture
We build our deep focus ranking model on the popular CNN
architecture Alex-Net [10]. The first seven layers of Alex-Net
are used as the early layers of our network architecture, and then
we add auxiliary structure to fulfill similarity evaluation. Our
network architecture is illustrated in Fig. 1. Specifically, the
3.1. Deep metric formulation
We first present some notations that would be used throughout this article. Without loss of generalization, we consider
3
fc
3
fc
conv
96
conv
384
Stride 4
27
227 Max-pooling
3x3 kernel, stride 2
Normalization
Input image
Layer1
Kernel 5x5x96
Stride 1
256
13
Kernel 3x3x256
Stride 1
13
13
13
27
55
conv
384
13
55
227
conv
conv
256
Kernel
11x11x3
Max-pooling
3x3 kernel, stride 2
Layer3
1
Loss layer
Kernel 3x3x384 Kernel 3x3x384
Stride 1
Stride 1
Max-pooling
3x3 kernel, stride 2
Normalization
Layer2
13
Layer4
Layer5
4096
Layer6
4096
Layer7
Figure 1: Illustration of our network architecture.
mentioned σ(x) = log2 (1 + 2−x ), we first obtain
first five of the seven layers are convolutional layers and the remaining two are fully connected layers. The first convolutional
layer takes an image of size 227 × 227 × 3 as input and has 96
kernels of size 11 × 11 × 3 with a stride of 4 pixels, resulting
in 96 feature maps of size 55 × 55. The second convolutional
layer has 256 kernels of size 5 × 5 × 96 and outputs 256 feature
maps of size 27 × 27. The output feature maps of the first two
convolutional layers are max-pooled (3 × 3 kernel with stride
2) and normalized before being fed into the next convolutional
layer. The third, fourth, and fifth convolutional layers have 384
kernels of size 3 × 3 × 256, 384 kernels of size 3 × 3 × 384, and
256 kernels of size 3 × 3 × 384, respectively. The output feature
maps of the third and fourth convolutional layers are not intervened by any pooling or normalization, while the output of the
fifth convolutional layer is max-pooled (3 × 3 kernel with stride
2) and results in 256 feature maps of size 6 × 6. The sixth and
seventh fully-connected layers have 4096 neurons each. Moreover, the outputs of all the seven layers are immediately filtered
by a rectified linear unit (ReLU) before any pooling or normalization operation. Finally, we add an auxiliary layer to calculate
the loss function (3).
"
#
X X
−δ(x, x+ , x− )
∂D(f (x), f (x− ))
∂L
∂D(f (x), f (x+ ))
=
−
∂f (x)
1 + δ(x, x+ , x− )
∂f (x)
∂f (x)
x∈X x− ∈Ux
"
#
+
−
−
X X
∂L
−δ(x, x , x )
∂D(f (x), f (x ))
∂D(f (x), f (x+ ))
=
−
∂f (x+ )
1 + δ(x, x+ , x− )
∂f (x+ )
∂f (x+ )
x∈X x− ∈Ux
"
#
X X
∂L
−δ(x, x+ , x− )
∂D(f (x), f (x− ))
∂D(f (x), f (x+ ))
=
−
∂f (x− )
1 + δ(x, x+ , x− )
∂f (x− )
∂f (x− )
x∈X −
x
∈Ux
where
δ(x, x+ , x− ) = 2−[D(f (x),f (x
−
))−D(f (x),f (x+ ))]
.
(5)
Remembering that D(·, ·) is the L2 norm, then we can obtain the derivatives as follows:
X X −δ(x, x+ , x− )
∂L = 2
f (x+ ) − f (x− )
+
−
∂f (x)
1 + δ(x, x , x )
x∈X x− ∈Ux
X X −δ(x, x+ , x− )
∂L
=
2
f (x) − f (x+ )
+)
+ , x− )
∂f
(x
1
+
δ(x,
x
x∈X x− ∈Ux
X X −δ(x, x+ , x− )
∂L
=
2
f (x− ) − f (x)
∂f (x− )
+
−
1 + δ(x, x , x )
−
3.3. Optimization with focus ranking unit generation
We optimize the proposed model using stochastic gradient
descent (SGD) algorithm with momentum [47]. It need to feed
the whole dataset into the GPU memory if directly using loss
function (3), which is almost intractable for large scale optimization. To tackle this problem, similar to [14, 30], we organize the training data into ranking units to optimize the model.
Specifically, for a probe image x, we take one positive image
x+ ∈ X + and a small subset of X − , i.e., Ux ⊂ X − , into consideration. Note that x+ , randomly sampled from X + , is the
positive match and Ux randomly sampled from X − , is the reference set, for probe image x. So a focus ranking unit is defined
to contain a probe image x, a positive match image x+ and the
reference set Ux . Then the loss can be expressed as
X X
L=
σ(D(f (x), f (x− )) − D(f (x), f (x+ ))). (4)
x∈X x ∈Ux
(6)
We employ the unit sampling generation to optimize the
model for the following serval reasons. First, taking a random subset in each iteration and applying sufficient iterations
is approximately equivalent to using the whole reference set.
Second, it is more efficient and practical to take a subset for
training, especially for large-scale learning. Third, training the
model with a random subset introduces a high degree of diversity, which is of great importance to learning.
4. Fabric Image Retrieval Dataset (FIRD)
To the best of our knowledge, there are not public datasets
for the fabric image retrieval problem. Since fabric image retrieval benefits a wide range of application, e.g., large-scale
clothing searching and online shopping, we construct a largescale fabric image retrieval dataset (FIRD). Specifically, we collect 4,300 fabrics, and take different photos for each fabric. In
x∈X x− ∈Ux
The gradients of the loss with respective to the feature vector can be computed according to the chain rule. With the afore4
Figure 2: Samples from the FIRD dataset under different photographing conditions.
(the ith fabric labelled i). We randomly select half of the 4,300
fabrics as training set and the rest as test set. So both training
and test sets contain 2,150 categories of fabric. During testing,
we randomly select 2/5 images of each category in the test set
as query images, and arrange the rest images as a retrieval set.
Given a query image, we aim to find the ones with the same
label.
order to imitate real application scenarios, we vary the collecting conditions from the following perspectives.
Photographing device: Without loss of generalization, we use
cell phones for collecting images as most of clothing searching
and online shopping applications are used in cell phones. We
use varied cell phones, from Android to iOS, from low- to highend, to increase the diversity of collected images.
Photographing circumstance: We take photos of a fabric in
varied circumstances, including different scenes (indoor and
outdoor), time (day and night), and illumination conditions.
Photographing view: We change viewing conditions as follows: (1) we rotate the fabric to collect images with different
angles; (2) we take photos in near and far views to collect images with different scales; (3) we wrinkle the fabric to collect
fabric images with various degrees of wrinkle.
Figure 2 illustrates some samples from the FIRD dataset.
It is a challenging dataset due to the complex variations in illumination, rotation, scale and wrinkle. All the 4,300 fabrics
are incorporated in the dataset and, for each fabric, 5–10 images are collected under different above-mentioned conditions.
The images of the same fabric are annotated by the same label
5. Experiments
In this section, we will present experimental comparions between the proposal method and existing metric learning methods. We will also present intra-model comparisons to analyze
the contribution of focus ranking unit generation.
5.1. Evaluation protocols
We utilize the following measures to give an extensive evaluation and analysis of the tested methods’ performance.
Recall@K. One of the critical measures for evaluating the
retrieval performance is the recall at particular ranks, i.e., Recall@K [42, 48]. In this measure, given a query image, the
5
images with top K smallest distances to it are retrieved, and
Recall@K is the ratio of retrieved relevant images over the total
relevant images. The Recall@K on the dataset is then computed
by averaging the scores of all query images.
mAP. Another commonly used measure is the mean average precision (mAP) [49]. For each query image, a precision/recall curve can be obtained, and its average precision (the
area under the curve) is computed. The mAP is obtained as the
mean over all the queries.
F1 and NMI scores. To further demonstrate the superiority of the proposed model, we also apply F1 and NMI scores
[50] to evaluate the performance. The Normalized Mutual Information (NMI) score is calculated by conducting clustering
experiments on the test set with the learned features.
5.3. Comparison results
Note that the original feature dimensions of all models are
4,096, and they are too large for an efficient retrieval system.
Therefore, we follow Babenko et al. [31] to apply the Principal
Component Analysis (PCA) [52] to reduce the dimension to
smaller ones (i.e., 512, 64, 16, 8), and perform comparisons on
various feature dimensions.
We first perform comparisons on the FIRD test set. For
each fabric, we randomly select 2/5 images of each category
as queries and use the rest as retrieval ones, making up a retrieval set of about 7,500 images. The comparison results of
Recall@K are presented in Figure 3. It can be observed that all
the other tested models perform well beyond the Alex baseline
one. In addition, Figure 3(a) reveals that, as the feature dimension is reduced from 4,096 to 16, the recall rates of the four
metric embedding models are nearly fixed but the Alex baseline and Softmax models suffer from a severe drop, suggesting
that models with metric embedding can generate more compact
and discriminative features. Among the four metric embedding
models, our model can still outperform the other three by a large
margin. For example, with the feature dimension of 4096, although the pair, Magnet and triplet models can achieve the high
recall rates of 83.9%, 85.6% and 88.5% respectively, our model
further improves them to 93.9%. Figure 3(b) reports the Recall@K scores when changing the number of retrieved images
K from 1 to 32 and fixing the the feature dimension as 256.
Since more relevant images can be retrieved when K increases,
so the scores of all the tested methods will also monotonously
increase. The figure shows that our method has a superior performance over the other methods at each value of K.
Figure 4 shows the comparison results measured in mAP
with respect to the feature dimension. It reveals again the other
tested models all perform better than the Alex baseline one. It
should be noted that the Softmax model with fine-tuning on
the FIRD training set has significant performance gain from the
Alex one. The evaluation results in mAP further demonstrate
the superiority of our model. It suggests that our model can
perform much better than the existing ones in fabric image retrieval systems.
We also present the comparison results of F1 and NMI scores
in Figure 5. The F1 score takes both accuracy and recall into
consideration, and the results depicted in Figure 5(a) give further approval in the effectiveness of our model. We also conduct clustering experiments on the FIRD test set using the features extracted by all the tested models, and Figure 5(b) reports the comparison results measured in NMI. Our model performs much better than all the others, suggesting that it can also
achieve much better clustering quality.
In summary, observed from the quantitative performance of
all the tested models in different measures, the effectiveness of
the proposed method is sufficiently verified. We also observe
that the pair model’s performance is the least sensitive to feature dimension changes, while the triplet model’s performance
ranks second in most measures. The Softmax model exhibits
impressive performance improvement from the Alex baseline
model and can outperform the pair model in some measures,
5.2. Compared methods and implementation details
We compare our model with five methods, i.e., Alex baseline, Softmax, pair, triplet and Magnet. Among them, the Alex
baseline method applies the Alex-net model [10], which is trained
on the ImageNet classification dataset [36] using the Caffe framework [51], and extracts the features of the first fully connected
layer (i.e., fc6), so that the best performance can be ensured,
as suggested in [31]. The Softmax method uses the Alex-net
model pre-trained on ImageNet and then fine-tunes it on our
FIRD training set. Afterwards, the feature extracted on the fc6
layer is used for evaluation. We also note that, for fair comparison, all the other tested methods are based on the Alex-net
model and use the feature from the fc6 layer.
The Magnet model applies the Magnet loss [44]. It is a
recent deep metric embedding method and sculpts the feature
representation by identifying and handling intra-class variation
and inter-class similarity. We use its publicly available TensorFlow code1 and the Alex-net tensorflow model2 pre-trained
on ImageNet to train on our FIRD dataset. We initialize the
learning rate as 0.01, with momentum of 0.9, weight decay of
0.0005, and batch size of 80 (each batch randomly selects 16
classes and 5 images for each class).
The two common deep metric embedding methods (i.e.,
pair and triplet) and the proposed method are implemented in
the Caffe framework. These models are initialized with the parameters of the Alex-net model pre-trained on ImageNet and
fine-tuned on our training set using the SGD method with backpropagation. For pair and triplet models, we initialize the learning rate as 0.001, with momentum of 0.9, weight decay of 0.0005,
and batch sizes of 80 and 120 respectively. For our focus ranking model, we fix the positive and negative ratio as 1:32 in the
focus ranking unit, which can well approximate the proposed
model.
During training, the images are resized to 256 × 256. In
addition, the images are randomly cropped and randomly mirrored horizontally for data augmentation. During inference, all
images are resized to 256 × 256 and cropped 227 × 227 at the
image center for feature extraction. All experiments were tested
on a single NVIDIA Tesla K40 GPU.
1 https://github.com/pumpikano/tf-magnet-loss
2 http://www.cs.toronto.edu/˜guerzhoy/tf
alexnet/
6
100
87
92
Recall@K (%)
Recall@K (%)
95
79
Alex
Softmax
71
84
Alex
Softmax
76
Pair
Pair
magnet
63
magnet
68
Triplet
Triplet
Ours
Ours
55
60
8
16
64
512
4096
1
2
4
Feature Dimension
8
16
32
Top K
(a)
(b)
Figure 3: Comparison on the FIRD dataset using the Recall@K measure. (a) Recall@K vs. feature dimension with K=16. (b) Recall@K vs. K with the feature
dimension of 256. Best viewed in color.
75
mAP score (%)
65
55
Alex
45
Softmax
Pair
magnet
35
Triplet
Ours
25
8
16
64
512
4096
Feature Dimension
42
95
37
91
NMI score (%)
F1 Score (%)
Figure 4: Comparison on the FIRD dataset using the mAP measure. Best viewed in color.
32
Alex
27
Softmax
87
Alex
83
Softmax
Pair
Pair
magnet
22
magnet
79
Triplet
Triplet
Ours
17
Ours
75
8
16
64
512
4096
8
Feature Dimension
16
64
512
4096
Feature Dimension
(a)
(b)
Figure 5: Comparison on the FIRD dataset with F1 and NMI scores. (a) F1 score vs. feature dimension. (b) NMI score vs. feature dimension. Best viewed in color.
which shows that fine-tuning on the target dataset can boost the
baseline model’s performance. For the Magnet model, it does
not work quite well in our experiments. We should note that it is
implemented in TensorFlow, different from the others, and better explicit models of distributions of classes and more careful
fine-tuning may boost its performance further.
To give a qualitative analysis of our method, Figure 6 presents
the top 8 retrieved images of some random query examples and
Figure 7 presents those of some failure cases. We find that, in
most cases, our model can successfully retrieve the images be-
longing to the same fabric, despite of the complex variations in
illumination, rotation, scale, and wrinkle. Moreover, most images of the same fabric rank top, and other retrieved ones also
have very similar textural pattern with the query ones, which is
also very essential for real-world systems. On the other hand,
for the failure examples, most retrieved images come from the
fabrics with subtle differences to the query.
In order to better mimic the real-world scenarios, we further
do experiments with a much larger dataset. We collect 100,000
fabrics and take one image for each fabric under the varied pho7
Figure 6: Retrieval results of query images randomly selected from the FIRD dataset. The first column are the query images, and the rest are the retrieved ones,
arranged according to the distance to the corresponding query images. The retrieved images of the same class to the query image are marked in red boxes.
Figure 7: Retrieval results of failure query images selected from the FIRD dataset. The first column are the query images, and the rest are the retrieved ones,
arranged according to the distance to the corresponding query images.
8
95
81
83
Recall@K (%)
Recall@K (%)
93
69
Alex
57
Softmax
71
Alex
Softmax
59
Pair
Pair
magnet
45
magnet
47
Triplet
Triplet
Ours
Ours
33
35
8
16
64
512
4096
1
2
4
Feature Dimension
8
16
32
Top K
(a)
(b)
95
98
92
93
Recall@K (%)
Recall@K (%)
Figure 8: Comparison on the FIRD dataset with additional 100,000 fabric images. (a) Recall@K vs. feature dimension with K=16. (b) Recall@K vs. K with the
feature dimension of 256. Best viewed in color.
89
Ours (1:1)
Ours (2:1)
86
88
Ours (1:1)
Ours (2:1)
83
Ours (4:1)
Ours (4:1)
Ours (8:1)
83
Ours (8:1)
78
Ours (16:1)
Ours (16:1)
Ours (32:1)
Ours (32:1)
80
73
8
16
64
512
4096
1
2
4
Feature Dimension
8
16
32
Top K
(a)
(b)
Figure 9: Analysis of focus ranking unit generation. (a) Recall@K vs. feature dimension with K=16. (b) Recall@K vs. K with the feature dimension of 256. Best
viewed in color.
tographing conditions, making up a dataset of 100,000 fabric
images. The original retrieval set of about 7,500 images and
these 100,000 images are mixed to construct a new retrieval
set. The retrieval results using the new retrieval set are presented in Figure 8 with the Recall@K measure. It shows that
our model again suppresses all other methods by an even larger
margin, suggesting our model is more suitable for real-world
large scale fabric retrieval systems.
increasing the ratio further can hardly obtain more improvement on the recall rate. These experimental results show that
increasing the ratio may boost the performance by quite a large
margin, but it becomes near saturation when the ratio reaches
32:1. It also suggests that a ratio of 32:1 can well approximate
our ranking model.
5.4. Analysis of focus ranking unit generation
In this paper, we propose a novel focus ranking embedding
method, which aims to rank similar examples top over all dissimilar ones. The focus ranking method can easily be unified
into a deep CNN to jointly learn an image representation and a
metric. At the training stage, we apply focus ranking unit generation to approximate the proposed model for efficient training, and at the inference stage, a test image can be directly fed
into the learnt CNN to extract the features in one shot. Extensive experimental results suggest that our deep focus ranking
model shows superior performance over existing metric embedding methods.
6. Conclusion
In this work, we generate focus ranking units by sampling
to approximate our ranking model in the process of model optimization, and the negative-positive ratio is a very critical parameter. In this section, we analyze the ranking unit generation algorithm by evaluating the performance when training the
model with the negative-positive ratio of 1:1, 2:1, 4:1, 8:1, 16:1
and 32:1 on the FIRD dataset. Figure 9 presents the quantitative
evaluation results measured in Recall@K. It shows that there
are obvious improvement consistently when increasing the ratio from 1:1 to 16:1. However, the improvement of recall rate
when increasing the ratio from 16:1 to 32:1 is negligible, and
9
Acknowledgment
[20] H. Jégou, M. Douze, C. Schmid, Improving bag-of-features for large scale
image search, International Journal of Computer Vision 87 (3) (2010)
316–336.
[21] F. Perronnin, Y. Liu, J. Sánchez, H. Poirier, Large-scale image retrieval
with compressed fisher vectors, in: Proceedings of IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), 2010, pp. 3384–3391.
[22] J. Zhu, S. Liao, Z. Lei, S. Z. Li, Multi-label convolutional neural network
based pedestrian attribute classification, Image and Vision Computing 58
(2017) 224–229.
[23] Z. Wang, T. Chen, G. Li, R. Xu, L. Lin, Multi-label image recognition
by recurrently discovering attentional regions, in: Proceedings of IEEE
International Conference on Computer Vision (ICCV), 2017.
[24] R. Girshick, J. Donahue, T. Darrell, J. Malik, Rich feature hierarchies for
accurate object detection and semantic segmentation, in: Proceedings of
IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
2014, pp. 580–587.
[25] R. Girshick, Fast R-CNN, in: Proceedings of IEEE International Conference on Computer Vision (ICCV), 2015, pp. 1440–1448.
[26] T. Chen, L. Lin, W. Zuo, X. Luo, L. Zhang, Learning a wavelet-like autoencoder to accelerate deep neural networks, in: Proceedings of AAAI
Conference on Artificial Intelligence, 2018.
[27] J. Long, E. Shelhamer, T. Darrell, Fully convolutional networks for semantic segmentation, in: Proceedings of IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), 2015, pp. 3431–3440.
[28] T. Chen, L. Lin, L. Liu, X. Luo, X. Li, DISC: deep image saliency
computing via progressive representation learning, IEEE Transactions on
Neural Networks and Learning Systems 27 (6) (2016) 1135–1149.
[29] L. Lin, K. Wang, W. Zuo, M. Wang, J. Luo, L. Zhang, A deep structured
model with radius-margin bound for 3D human activity recognition, International Journal of Computer Vision 118 (2) (2016) 256–273.
[30] S.-Z. Chen, C.-C. Guo, J.-H. Lai, Deep ranking for person reidentification via joint representation learning, IEEE Transactions on Image Processing 25 (5) (2016) 2353–2367.
[31] A. Babenko, A. Slesarev, A. Chigorin, V. Lempitsky, Neural codes for
image retrieval, in: Proceedings of European Conference on Computer
Vision (ECCV), 2014, pp. 584–599.
[32] A. Babenko, V. Lempitsky, Aggregating local deep features for image
retrieval, in: Proceedings of IEEE International Conference on Computer
Vision (ICCV), 2015, pp. 1269–1277.
[33] M. Paulin, M. Douze, Z. Harchaoui, J. Mairal, F. Perronin, C. Schmid, Local convolutional features with unsupervised training for image retrieval,
in: Proceedings of IEEE International Conference on Computer Vision
(ICCV), 2015, pp. 91–99.
[34] F. Perronnin, D. Larlus, Fisher vectors meet neural networks: A hybrid
classification architecture, in: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015, pp. 3743–3752.
[35] N. Jammalamadaka, A. Zisserman, C. V. Jawahar, Human pose search
using deep networks, Image and Vision Computing 59 (2017) 31–43.
[36] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang,
A. Karpathy, A. Khosla, M. Bernstein, et al., Imagenet large scale visual
recognition challenge, International Journal of Computer Vision 115 (3)
(2015) 211–252.
[37] E. P. Xing, A. Y. Ng, M. I. Jordan, S. J. Russell, Distance metric learning
with application to clustering with side-information, in: Proceedings of
Advances in Neural Information Processing Systems (NIPS), 2002, pp.
505–512.
[38] W. Zheng, S. Gong, T. Xiang, Person re-identification by probabilistic relative distance comparison, in: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2011, pp. 649–656.
[39] A. Mignon, F. Jurie, PCCA: A new approach for distance learning from
sparse pairwise constraints, in: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012, pp. 2666–2672.
[40] H. Liu, B. Ma, L. Qin, J. Pang, C. Zhang, Q. Huang, Set-label modeling and deep metric learning on person re-identification, Neurocomputing
151 (2015) 1283–1292.
[41] S. Bell, K. Bala, Learning visual similarity for product design with convolutional neural networks, ACM Transactions on Graphics 34 (4) (2015)
98.
[42] H. O. Song, Y. Xiang, S. Jegelka, S. Savarese, Deep metric learning via
lifted structured feature embedding, in: Proceedings of IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 4004–
This work was supported in part by the National Natural
Science Foundation of China (61379112, 61402120, 61672547,
61472455, 61370186). The authors would like to thank Shenzhen Micro Vision Technology Co., Ltd. for providing the
large-scale dataset of 100,000 fabric images for evaluation.
References
[1] D. G. Lowe, Distinctive image features from scale-invariant keypoints,
International Journal of Computer Vision 60 (2) (2004) 91–110.
[2] A. Oliva, A. Torralba, Modeling the shape of the scene: A holistic representation of the spatial envelope, International Journal of Computer Vision 42 (3) (2001) 145–175.
[3] M. Douze, H. Jégou, H. Sandhawalia, L. Amsaleg, C. Schmid, Evaluation
of GIST descriptors for web-scale image search, in: Proceedings of the
ACM International Conference on Image and Video Retrieval, 2009, pp.
19:1–8.
[4] J. Sivic, A. Zisserman, Video google: A text retrieval approach to object
matching in videos, in: Proceedings of IEEE International Conference on
Computer Vision (ICCV), 2003, pp. 1470–1477.
[5] F. Perronnin, C. Dance, Fisher kernels on visual vocabularies for image
categorization, in: Proceedings of IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), 2007, pp. 1–8.
[6] F. Perronnin, J. Sánchez, T. Mensink, Improving the fisher kernel for
large-scale image classification, in: Proceedings of European Conference
on Computer Vision (ECCV), 2010, pp. 143–156.
[7] H. Jégou, M. Douze, C. Schmid, P. Pérez, Aggregating local descriptors
into a compact image representation, in: Proceedings of IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), 2010, pp. 3304–
3311.
[8] J. Wang, Y. Song, T. Leung, C. Rosenberg, J. Wang, J. Philbin, B. Chen,
Y. Wu, Learning fine-grained image similarity with deep ranking, in: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014, pp. 1386–1393.
[9] E. Hoffer, N. Ailon, Deep metric learning using triplet network, in: Proceedings of International Workshop on Similarity-Based Pattern Recognition, 2015, pp. 84–92.
[10] A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with
deep convolutional neural networks, in: Proceedings of Advances in Neural Information Processing Systems (NIPS), 2012, pp. 1097–1105.
[11] M. D. Zeiler, R. Fergus, Visualizing and understanding convolutional
networks, in: Proceedings of European Conference on Computer Vision
(ECCV), 2014, pp. 818–833.
[12] R. Hadsell, S. Chopra, Y. LeCun, Dimensionality reduction by learning
an invariant mapping, in: Proceedings of IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), Vol. 2, 2006, pp. 1735–1742.
[13] S. Chopra, R. Hadsell, Y. LeCun, Learning a similarity metric discriminatively, with application to face verification, in: Proceedings of IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), Vol. 1,
IEEE, 2005, pp. 539–546.
[14] F. Schroff, D. Kalenichenko, J. Philbin, Facenet: A unified embedding for
face recognition and clustering, in: Proceedings of IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), 2015, pp. 815–823.
[15] K. Q. Weinberger, L. K. Saul, Distance metric learning for large margin
nearest neighbor classification, Journal of Machine Learning Research 10
(2009) 207–244.
[16] D. Nister, H. Stewenius, Scalable recognition with a vocabulary tree, in:
Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2006, pp. 2161–2168.
[17] H. Jegou, H. Harzallah, C. Schmid, A contextual dissimilarity measure for
accurate and efficient image search, in: Proceedings of IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), 2007, pp. 1–8.
[18] F. Fraundorfer, H. Stewénius, D. Nistér, A binning scheme for fast hard
drive based image search, in: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2007, pp. 1–6.
[19] H. Jegou, M. Douze, C. Schmid, Hamming embedding and weak geometric consistency for large scale image search, in: Proceedings of European
Conference on Computer Vision (ECCV), 2008, pp. 304–317.
10
4012.
[43] J. Hu, J. Lu, Y.-P. Tan, Discriminative deep metric learning for face verification in the wild, in: Proceedings of IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), 2014, pp. 1875–1882.
[44] O. Rippel, M. Paluri, P. Dollár, L. D. Bourdev, Metric learning with adaptive density discrimination, Proceedings of International Conference on
Learning Representations (ICLR) (2016) 1–15.
[45] R. Zhang, L. Lin, R. Zhang, W. Zuo, L. Zhang, Bit-scalable deep hashing with regularized similarity learning for image retrieval and person
re-identification, IEEE Transactions on Image Processing 24 (12) (2015)
4766–4779.
[46] H. Yun, P. Raman, S. Vishwanathan, Ranking via robust binary classification, in: Proceedings of Advances in Neural Information Processing
Systems (NIPS), 2014, pp. 2582–2590.
[47] L. Bottou, Stochastic gradient descent tricks, in: Neural Networks: Tricks
of the Trade, 2012, pp. 421–436.
[48] H. Jegou, M. Douze, C. Schmid, Product quantization for nearest neighbor search, IEEE Transactions on Pattern Analysis and Machine Intelligence 33 (1) (2011) 117–128.
[49] J. Philbin, O. Chum, M. Isard, J. Sivic, A. Zisserman, Object retrieval
with large vocabularies and fast spatial matching, in: Proceedings of IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), 2007,
pp. 1–8.
[50] C. D. Manning, P. Raghavan, H. Schütze, Introduction to information retrieval, Cambridge University Press, Cambridge, 2008.
[51] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick,
S. Guadarrama, T. Darrell, Caffe: Convolutional architecture for fast feature embedding, in: Proceedings of the ACM International Conference on
Multimedia, 2014, pp. 675–678.
[52] I. Jolliffe, Principal component analysis, Springer-Verlag, New York,
2002.
11
| 1cs.CV
|
arXiv:1504.07697v2 [cs.CC] 11 Jul 2016
POLYNOMIAL FACTORIZATION OVER FINITE FIELDS BY COMPUTING
EULER-POINCARE CHARACTERISTICS OF DRINFELD MODULES
ANAND KUMAR NARAYANAN
[email protected]
Abstract. We propose and rigorously analyze two randomized algorithms to factor univariate polynomials over finite fields using rank 2 Drinfeld modules. The first algorithm estimates the degree of
an irreducible factor of a polynomial from Euler-Poincare characteristics of random Drinfeld modules.
Knowledge of a factor degree allows one to rapidly extract all factors of that degree. As a consequence,
the problem of factoring polynomials over finite fields in time nearly linear in the degree is reduced to
finding Euler-Poincare characteristics of random Drinfeld modules with high probability. The second
algorithm is a random Drinfeld module analogue of Berlekamp’s algorithm. During the course of its
analysis, we prove a new bound on degree distributions in factorization patterns of polynomials over
finite fields in certain short intervals.
1. Introduction
1.1. Current State of the Art. Let Fq denote the finite field with q elements and Fq [t] the polynomial
ring in one indeterminate. The fastest known randomized algorithm for factorization in Fq [t] is the
Kaltofen-Shoup algorithm [KS98, § 2] implemented by Kedlaya-Umans fast modular composition [KU08].
It belongs in the Cantor-Zassenhaus [CZ81] framework and to factor a polynomial of degree n takes
e 3/2 log q + n log2 q) expected time by employing the following sequence of steps1. The first is square
O(n
free factorization where the polynomial in question is written as a product of square free polynomials. A
square free polynomial is one that does not contain a square of an irreducible polynomial as a factor. The
second step known as distinct degree factorization takes a monic square free polynomial and decomposes
it into factors each of which is a product of irreducible polynomials of the same degree. The final step
is equal degree factorization which splits a polynomial all of whose irreducible factors are of the same
degree into irreducible factors. The bottleneck is distinct degree factorization and currently the difficulty
appears to be in finding the factor degrees. Given the degree of a factor, one can extract all factors of
e log2 q) expected time [KU08].
that degree in O(n
1.2. Polynomial Factorization and Euler-Poincare Characteristic of Drinfeld Modules. We
propose a novel procedure to read off the smallest factor degree of a monic square free polynomial
h ∈ Fq [t] from the Euler-Poincare characteristic χφ,h of a random rank 2 Drinfeld module φ reduced
at h. The reduction of φ at h is a finite Fq [t]-module and its Euler-Poincare characteristic χφ,h ∈ Fq [t]
(see Q
Definition 2.1) is an
Q Fq [t] valued cardinality measure. If h factors into monic irreducible factors as
h = i pi , then χφ,h = i χφ,pi . A Drinfeld module analogue of Hasse’s theorem for elliptic curves, due to
Gekeler [Gek91], asserts for each pi that χφ,pi = pi + cφ,i for some cφ,i ∈ Fq [t] with deg(cφ,i ) ≤ deg(pi )/2.
Hence, the leading coefficients of h and χφ,h agree and the number of agreements reveals information
about the degree of the smallest degree factor of h. In § 3, we prove,
The author was partially supported by NSF grant CCF 1423544.
1The soft O
e notation suppresses no(1) and logo(1) q terms for ease of exposition.
1
2
ANAND KUMAR NARAYANAN
√
Theorem 1.1. For n ≤ q/2, the smallest factor degree of a degree n square free h ∈ Fq [t] can be
inferred in O(n log q) time from χφ,h with probability at least 1/4 for a randomly chosen φ.
Consider an algorithm B that takes a square free h ∈ Fq [t] and a random Drinfeld module φ as inputs
such that with constant probability the output B(h, φ) is χφ,h . That is, B is a montecarlo algorithm to
compute Euler-Poincare characteristics. By choosing φ at random and invoking Theorem 1.1, we establish
that a non trivial factor can be found in nearly linear time with oracle access to B.
e log2 q) expected time algorithm (with oracle access to B) to find an
Corollary 1.2. There exists an O(n
√
irreducible factor of a square free polynomial h ∈ Fq [t] of degree n ≤ q/2 with only O(1) queries to B,
each of the form B(h).
√
The requirement q ≥ 2n in Theorem 1.1 and Corollary 1.2 is without loss of generality since if q were
√
smaller, we could choose a small power q ′ of q that satisfies q ′ ≥ 2n and obtain the factorization over
Fq′ [t] with the running time unchanged up to polylogarithmic factors in n (see Remark 3.2).
Given oracle access to B, obtaining the complete factorization by naively extracting one factor at a
time using Corollary 1.2 leads to a 3/2 running time exponent on n. For instance, a polynomial of degree
n with one irreducible factor each of degree 1, 2, 3, . . . , m − 1, m where m = Θ(n1/2 ) requires Θ(n1/2 )
extractions. In [GNU15], an algorithm to quickly obtain the complete factorization given a procedure
to extract a non trivial factor as a subroutine is described. As a consequence, an algorithm for implementing B with exponent less than 3/2 would lead to a polynomial factorization algorithm with exponent
less than 3/2. As an illustration, in § 3.2, we obtain the following corollary of Theorem 1.1 describing
an implication of a nearly linear time algorithm for computing Euler-Poincare characteristic of random
Drinfeld modules.
e logO(1) q) expected time for
Corollary 1.3. An implementation of the oracle function B that takes O(n
O(1)
4/3
e
inputs of degree n yields an O(n
log
q) expected time polynomial factorization algorithm.
We describe our first complete polynomial factorization algorithm by presenting an implementation of B.
Assuming the matrix multiplication exponent ω is 2, our implementation of B and hence the polynomial
factorization algorithm both have running time exponent 3/2 in the input degree. A faster implementation of B would break the 3/2 exponent barrier in polynomial factorization. Thus, the problem of
computing Euler-Poincare characteristics of random Drinfeld modules warrants a thorough investigation.
The problem is analogous to point counting on elliptic curves over finite fields and the question as to if
there is a Drinfeld module analogue of Schoof’s algorithm [Sch95] is immediate.
We next briefly sketch our implementation of B. The obvious procedure to compute χφ,h for a given
a square free h ∈ A and a rank 2 Drinfeld module φ, is to compute the characteristic polynomial of the
(Fq linear) φ action on Fq [t]/(h). However, the complexity of such generic linear algebraic techniques is
equivalent to inverting square matrices of dimension deg(h). We devise a faster implementation of B by
exploiting the fact that the input φ to B is chosen at random. For q ≥ 2 deg(h)4 , which we may assume
without loss of generality, we prove it is likely that the reduction of a random φ at h is a cyclic Fq [t]module and further that χφ,h coincides with the order (that is, the monic generator of the annihilator)
of a random element in φ reduced at h. Implementing B is thus reduced to the finding the order of a
random element in a random Drinfeld module φ reduced at h with constant probability. In § 3.1, we
solve the order finding problem by posing it as an instance of the automorphism projection problem of
e (1+ω)/2 log q + n log2 q) expected time algorithm to find
Kaltofen-Shoup [KS98] and thereby obtain an O(n
a factor. In § 3.2, we describe how to obtain the complete factorization.
POLYNOMIAL FACTORIZATION AND EULER-POINCARE CHARACTERISTICS OF DRINFELD MODULES
3
1.3. Drinfeld Module Analog of Berlekemp’s Algorithm. Our second algorithm is a randomized
Drinfeld analogue of Berlekamp’s algorithm [Ber67] wherein Fq [t]-modules twisted by the Frobenius action is replaced with reductions of random rank-2 Drinfeld modules. It has the distinction of being the
only polynomial factorization algorithm over finite fields that does not involve a quadratic residuosity
like map. To factor h ∈ Fq [t], Berlekamp’s algorithm proceeds by finding a basis for the Berlekamp
subalgebra, which is the fixed space of the q th power Frobenius τ acting on Fq [t]/(h). Then (as in the
Cantor-Zassenhaus [CZ81] variant) a random element β in the Berlekamp subalgebra is generated by
taking a random Fq linear combination of the basis elements. When q is odd, with probability at least
half, the greatest common divisor of h and a lift of β (q−1)/2 − 1 is a non trivial factor of h. In place of
the Fq [t]-module Fq [t]/(h) and its Berlekamp subalgebra (which is the τ − 1 torsion in Fq [t]/(h)), our
second algorithm works over a random Drinfeld module reduced at h and its torsion corresponding to
low degree polynomials in the Drinfeld action. These low degree polynomials are precisely the low degree
factors of χφ,h . In terms of implementation, the algorithm closely resembles the fast black-box algorithm
e (1+ω)/2+o(1) log q + n log2 q) expected running time.
of Kaltofen-Shoup [KS98, § 3] and shares its O(n
Our analysis of the Drinfeld analogue of Berlekamp’s algorithm relies on bounds on the degree distribution of factorization patterns of polynomials in short intervals in Fq [t]. We prove the required
bounds, but only when q is very large compared to n. When q is not large enough, the claimed running
times hold under a widely believed conjecture (see Conjecture 4.2) and a slower running time bound is
proven unconditionally (see Remark 5.3). We next state the bounds for large q since they might be of
independent interest.
1.4. Factorization Patterns of Polynomials in Short Intervals. For a partition λ of a positive
integer d, let P (λ) denote the fraction of permutations on d letters whose cycle decompositions correspond
to λ. When q is large enough compared to d, a random polynomial in Fq [t] of degree d has a factorization
pattern corresponding to a partition λ of d with probability about P (λ) [Coh]. In § 4, we prove that the
degree distribution of a random polynomial in the interval If,m := {f + a|a ∈ Fq [t], deg(a) ≤ m} around
f ∈ Fq [t] is not far from the degree distribution of a random polynomial of degree d.
Theorem 1.4. For every f ∈ Fq [t] of degree d bounded by log q ≥ 3d log d, for every m ≥ 2 and for every
partition λ of d,
|{g ∈ If,m |λg = λ}|
1
1
P (λ) ≤
≤ 1+ √
1− √
P (λ)
q
|If,m |
q
where λg denotes the partition of deg(g) induced by the degrees of the irreducible factors of g.
1.5. Related Work. Our algorithms draw inspiration from Lenstra’s elliptic curve integer factorization
[Len87] where the role of multiplicative groups modulo primes in Pollard’s p − 1 algorithm [Pol74] was
recast with the group of rational points on random Elliptic curves modulo primes. Our first algorithm
for degree estimation using Euler-Poincare characteristic is a random Drinfeld module analogue of an
algorithm described in the author’s Ph.D thesis [Nar14, Chap 7] using Carlitz modules. To our knowledge,
the use of Drinfeld modules to factor polynomials over finite fields originated with Panchishkin and
Potemine [PP89] whose algorithm was rediscovered by van der Heiden [vdH04] (see also [vdH04-1]). Our
Drinfeld module analogue of Berlekamp’s algorithm shares some similarities with the algorithm in [vdH04]
but they differ in the following aspects. In contrast to our algorithm, the algorithm in [vdH04] only works
for equal degree factorization and targets torsion corresponding to large degree polynomials to aid in the
splitting; resulting in a slower running time. Further, its analysis was merely supported by heuristics.
Using Theorem 1.4 and Lemma 2.3, the proof of Lemma 5.2 can be extended to rigorously analyze the
algorithm in [vdH04] for large q when restricted to rank 2 Drinfeld modules.
4
ANAND KUMAR NARAYANAN
1.6. Organization. The analysis of our algorithms rely critically on the distribution of the characteristic polynomial (of the Frobenius endomorphism in the representation of the endomorphism ring) of a
random rank 2 Drinfeld module (on its ℓ-adic Tate modules). In § 2, we prove Lemma 2.3, the required
equidistribution lemma on the aforementioned characteristic polynomial. We begin § 2 by recounting the
theory of rank 2 Drinfeld modules followed by § 2.1, where the structure of rank 2 Drinfeld modules in
terms of its characteristic polynomial is described. In § 2.2, we state a weighted class number formula for
isomorphism classes of Drinfeld modules with a given characteristic polynomial (due to Gekeler [Gek08])
and from it derive Lemma 2.3. Gekeler’s weighted class number formula is obtained through complex
multiplication theory, that is, a correspondence between isomorphism classes of Drinfeld modules and
Gauss class numbers in certain imaginary quadratic orders. While they hold in even characteristic, for
ease of exposition, we assume in § 2.2 and consequently in the analysis of our algorithms that q is odd.
In § 3, we state and analyze our first algorithm, namely to estimate factor degrees by computing EulerPoincare characteristic and prove Theorem 1.1 and Corollaries 1.2 and 1.3. Theorem 1.4 concerning
factorization patterns of polynomials in short intervals is proven in § 4, which can be read independent
of the rest of the paper. The Drinfeld analogue of Berlekamp’s algorithm is described and analyzed in § 5.
A remarkable
feature of our algorithms andQreductions is that we address the factorization of a square
Q
free h = i pi ∈ Fq [t] by looking at χφ,h = i χφ,pi for random φ. We may view χφ,h as a perturbation
of h, since χφ,pi is the irreducible factor pi of h perturbed within a half degree (of pi ) interval. These
intervals widen with Drinfeld modules of increasing rank. Theorem 1.4 and Lemma 2.3 together imply
that even for rank 2, the intervals are large enough for χφ,pi to exhibit random factorization patterns.
This allowed us to pose the worst case complexity of polynomial factorization in terms of an average
complexity statement (Corollary 1.3) concerning Drinfeld modules. It would be interesting to see if this
perturbation leads to worst case to average case reductions for other multiplicative problems.
2. Finite Rank-2 Drinfeld Modules
Let A = Fq [t] denote the polynomial ring in the indeterminate t and let K be a field with a non zero ring
homomorphism γ : A → K. Necessarily, K contains Fq as a subfield. Fix an algebraic closure K̄ of K
and let τ : K̄ −→ K̄ denote the q th power Frobenius endomorphism. The ring of endomorphisms of the
additive group scheme Ga over K can be identified with the skew polynomial ring Khτ i where τ satisfies
the commutation rule ∀u ∈ K, τ u = uq τ . A rank-2 Drinfeld module over K is (the A-module structure
on Ga given by) a ring homomorphism
φ : A −→ Khτ i
t 7−→ γ(t) + gφ τ + ∆φ τ 2
for some gφ ∈ K and ∆φ ∈ K × . For a ∈ A, let φa denote the image of a under φ. We will concern
ourselves primarily with rank 2 Drinfeld modules and unless otherwise noted, a Drinfeld module will
mean a rank 2 Drinfeld module.
To every A-algebra L over K̄, the Drinfeld module φ endows a new A-module structure (which, we
denote by φ(L)) through the A-action
∀f ∈ L, ∀a ∈ A, a ∗ f = φa (f ).
For every A-algebra homomorphism ρ : L −→ L, ∀a ∈ A and ∀f ∈ L, ρ(φ(f )) = φ(ρ(f )). Thus ρ, when
thought of as a map from φ(L) −→ φ(L), is an A-module homomorphism. For every direct product
L × L′ of A-algebras over K̄, we hence have the corresponding direct sum of A-modules
∼ φ(L) ⊕ φ(L′ ).
φ(L × L′ ) =
POLYNOMIAL FACTORIZATION AND EULER-POINCARE CHARACTERISTICS OF DRINFELD MODULES
5
Henceforth, we restrict our attention to Drinfeld modules φ : A −→ Fq (t) over Fq (t) (with γ : A → Fq (t)
being the inclusion (identity map), gφ ∈ A and ∆φ ∈ A× ) and their reductions.
For a proper ideal f ⊂ A, let Ff denote A/f. For a prime ideal p ⊂ A, if ∆φ is non zero modulo p,
then the reduction φ/p := φ ⊗ Fp of φ at p is defined through the ring homomorphism
φ/p : A −→ Fp hτ i
t 7−→ t + (gφ mod p)τ + (∆φ mod p)τ 2
and the image of a ∈ A under φ/p is denoted by (φ/p)a . Even if ∆φ is zero modulo p, one can still obtain
the reduction (φ/p) of φ at p through minimal models of φ (c.f. [Gek91]). We refrain from addressing
this case since our algorithms do not require it.
As before, the Drinfeld module φ/p endows a new A-module structure (denoted by (φ/p)(L)) to every A-algebra L over the algebraic closure of Fp through the A-action
∀f ∈ L, ∀a ∈ A, a ∗ f = (φ/p)a (f )
′
and for every direct product L × L of A-algebras over the algebraic closure of Fp ,
(φ/p)(L × L′ ) ∼
= (φ/p)(L) ⊕ (φ/p)(L′ ).
Further, for every A-algebra L over the algebraic closure of Fp , φ(L) ∼
= (φ/p)(L).
Define the annihilator Ann(L) of a finite A-module L to be the monic generator of the annihilator
ideal {a ∈ A|aL = 0} of L. Define the A-order Ord(α) of an element α in a finite A-module L to be
the monic generator of the annihilator ideal {a ∈ A|aα = 0} of α. For f ∈ A, denote by (f ) the ideal
generated by f and by deg(f ) the degree of f . For a non zero ideal f ⊂ A, let deg(f) denote the degree
of its monic generator. For f, g ∈ A, by gcd(f, g) we mean the monic generator of the ideal generated by
f and g.
Definition 2.1. Following [Gek91], the Euler-Poincare characteristic of a finite A-module L is defined
as the unique monic polynomial χ(L) ∈ A such that
(1) If L ∼
= A/p for a prime ideal p ⊆ A, then (χ(L)) = p,
(2) If 0 → L1 → L → L2 → 0 is exact, then χ(L) = χ(L1 )χ(L2 ).
The above definition is a minor departure from [Gek91], where χ(L) was defined as an ideal, not a monic
generator. For a non zero ideal f ⊂ A and a Drinfeld module φ, let χφ,f denote χ(φ(Ff )). By definition,
Ann(φ(Ff )) divides χφ,f.
2.1. Frobenius Distributions and the Structure of Rank 2 Drinfeld Modules. In this subsection,
we recount the characterization of the A-module structure of reductions of Drinfeld modules at primes
due to Gekeler. Unless stated otherwise, proofs of claims made here are in [Gek91]. For Drinfeld modules
φ, ψ with reduction at a prime p ⊂ A, a µ ∈ Fp hτ i such that
∀a ∈ A, µ(φ/p)a = (ψ/p)a µ
is called as a morphism from φ/p to ψ/p. Let EndFp (φ) denote the endomorphism ring of φ/p. The
Frobenius at p, τ deg(p) is in EndFp (φ) and there exists a polynomial
Pφ,p (X) = X 2 − aφ,p X + bφ,p ∈ A[X]
(called the characteristic polynomial of the Frobenius at p) such that Pφ,p (τ deg(p) ) = 0 in EndFp (φ). The
polynomial Pφ,p is called as the characteristic polynomial because it is the characteristic polynomial of
6
ANAND KUMAR NARAYANAN
τ deg(p) in the representations of EndFp (φ) on the ℓ-adic Tate modules for prime ideals ℓ ⊂ A.
Furthermore, a rank-2 Drinfeld module analogue of Hasse’s theorem for elliptic curves states that
(bφ,p ) = p , deg(aφ,p ) ≤ deg(p)/2.
To be precise, the coefficient bφ,p = ǫφ,p p where p is the monic generator of p and
ǫφ,p = 1/((−1)deg p NFp /Fq (∆φ )) ∈ F×
q
where NFp /Fq is the norm from Fp to Fq . As in the elliptic curve case, aφ,p is referred to as the Frobenius
trace; a consequence of the aforementioned connection to Tate modules. Since τ deg(p) acts as the identity
on φ(Fp ), Pφ,p (1) kills φ(Fp ) (that is, (φ/p)Pφ,p (1) is the zero element in EndFp (φ)). In fact,
(χφ,p ) = (Pφ,p (1)) ⇒ χφ,p = p − (aφ,p − 1)/ǫφ,p.
As a consequence of φ being of rank 2, φ(Fp ) is either a cyclic A-module or a direct sum of two cyclic
A-modules. That is, there exists monic polynomials mφ,p , nφ,p ∈ A (not necessarily relatively prime)
such that as A-modules
φ(Fp ) ∼
= A/(mφ,p ) ⊕ A/(mφ,p nφ,p ).
In particular,
χφ,p = p − (aφ,p − 1)/ǫφ,p = m2φ,p nφ,p , Ann(φ(Fp )) = lcm(mφ,p , mφ,p nφ,p ) = mφ,p nφ,p .
Further still, when q is odd, mφ,p and nφ,p are completely determined by Pφ,p and a precise characterization of mφ,p and nφ,p in terms of aφ,p and ǫφ,p is described by Cojocaru and Papikian [CP14, Cor3].
Seeing that the characteristic polynomial completely determines the A-module structure of a Drinfeld
module reduced at a prime p, the question as to which polynomials can arise as such characteristic polynomials is immediate and is addressed in [Dri77], [Gek91] and [Yu95]. Yu [Yu95] relates the number of
isomorphism classes of Drinfeld modules (φ/p) with a given characteristic polynomial X 2 − aX + ǫp to
class numbers of orders in imaginary quadratic extensions over K. This relation arises from a theory that
bears likeness to the complex multiplication theory of elliptic curves. Gekeler [Gek08] using the complex
multiplication theory proved a precise characterization of the number of isomorphism classes of Drinfeld
modules (φ/p) with a given characteristic polynomial X 2 − aX + ǫp. Gekeler’s characterization implies
a certain equidistribution of the probability of X 2 − aX + ǫp being Pφ,p for a randomly chosen φ and is
discussed in the subsequent subsection.
2.2. Frobenius Distributions of Rank 2 Drinfeld Modules. Let p ⊂ A be a prime of degree d and
let p be its monic generator. Analysis of our algorithms will involve counting Drinfeld modules (φ/p) with
characteristic polynomial Pφ,p (X) = X 2 − aX + ǫp for a given ǫ ∈ F×
q and a ∈ A of degree at most d/2.
Such precise counts were proven by Gekeler [Gek08] by building on the connection between isomorphism
classes of Drinfeld modules over Fp and class numbers of imaginary quadratic orders established by Yu
[Yu95]. We begin the section by stating the result of Gekeler in equation 2.1 and later prove lemmas that
will find use in the analysis of our algorithms.
We identify a Drinfeld module (φ/p) with the tuple (gφ mod p, ∆φ mod p). The number of Drinfeld
×
modules (φ/p) is |Fp ||F×
p | since we get to pick gφ mod p from Fp and ∆φ mod p from Fp . Two Drinfeld
×
modules (φ/p) and (ψ/p) are isomorphic over Fp if and only if there is a c ∈ Fp such that gψ = cq−1 gφ
2
mod p and ∆ψ = cq −1 ∆φ mod p.
As a consequence, the automorphism group AutFp (φ) depends on if gφ mod p = 0. If gφ mod p = 0 and
POLYNOMIAL FACTORIZATION AND EULER-POINCARE CHARACTERISTICS OF DRINFELD MODULES
7
∼ F×2 . Else, AutF (φ) ∼
if further Fp contains a quadratic extension (call Fq2 ) of Fq , then AutFp (φ) =
= F×
p
q .
q
×
∼
The former case AutFp (φ) = Fq2 corresponds to φ having complex multiplication by Fq2 [t] and is rare.
For ǫ ∈ F×
q and a ∈ A not necessarily monic and of degree at most deg(p)/2, denote by H(a, ǫ, p) a
set of representatives of isomorphism classes of Drinfeld modules (φ/p) with Pφ,p (X) = X 2 − aX + ǫp.
Define
X
q−1
h∗ (a, ǫ, p) :=
#AutFp (φ)
(φ/p)∈H(a,ǫ,p)
which might be thought of as a weighted count of the isomorphism classes of Drinfeld modules with
Pφ,p (X) = X 2 − aX + ǫp where the weight (q − 1)/AutFp (φ) is 1 except in rare cases. We next describe
the connection between class numbers of certain imaginary quadratic orders and h∗ (a, ǫ, p).
Fix an ǫ ∈ F×
q and an a ∈ A of degree at most deg(p)/2. Let C be the A-algebra generated by a
root of X 2 − aX + ǫp and let E be the quotient field of C. It turns out that E is an imaginary quadratic
extension of k.
Yu [Yu95] proved that two Drinfeld modules are isogenous if and only if they have the same characteristic polynomial and further established that the number of isomorphism classes of Drinfeld modules
with characteristic polynomial X 2 − aX + ǫp equals the Gauss class number of C. This connection is
analogous to a similar statement concerning elliptic curves due to Deuring [Deu41]. Further, the weighted
count h∗ (a, ǫ, p) equals a certain appropriately weighted Gauss class number of C which was explicitly
computed by Gekeler [Gek08] using an analytic class number formula. We next summarize this result of
Gekeler assuming for ease of exposition that Fq is of odd characteristic throughout this section.
Let B be the integral closure of A in E. Let D denote the discriminant a2 − 4ǫp and f the largest
monic square factor of D. Let D0 = D/f . Then C = A + f B.
Let ξ denote the Dirichlet character associated with E and for ℜ(s) > 1 define the L-function
Y
−1
L(s, χ) :=
.
1 − ξ(ℓ)|ℓ|−s
ℓ prime of k
The unique prime at infinity ∞ is ramified in E/k if deg(D0 ) is odd and is inert if deg(D0 ) is even. Let
η denote the ramification index of ∞ in E/k (that is, η = 2 if deg(D0 ) is odd and η = 1 otherwise). Let
g denote the genus of the algebraic curve associated with E. Then
h∗ (a, ǫ, p) = ηq g S(f )L(1, ξ)
where
S(f ) :=
X
f′ |f
|Ff′ |
Y
(1 − ξ(ℓ)|Fℓ |−1 ).
ℓ|f′
Here f ⊆ A is the ideal generated by f , the summation is over proper ideals f′ dividing f and the product
is over prime ideals ℓ dividing f′ . When f is not A, we have
The conductor cond(ξ) of ξ is
S(f ) ≥ |Ff |.
(
(D0 )
if deg(D0 ) is even
cond(ξ) =
(D0 ).∞ if deg(D0 ) is odd
8
ANAND KUMAR NARAYANAN
But for a couple of exceptional cases, the genus g is determined by cond(ξ) as
g = deg(cond(ξ))/2 − 1.
The inequality 2.1 we arrive at for h∗ (a, ǫ, p) will be accurate in those exceptional cases as well. Thus we
refrain from mentioning the exceptional cases referring instead the interested reader to [Gek08]. Thus
(
q deg(D0 )/2−1
if deg(D0 ) is even
qg =
q deg(D0 )/2−1/2 if deg(D0 ) is odd
and
(
|Ff |q g =
q deg(D)/2−1
q deg(D)/2−1/2
if deg(D) is even
if deg(D) is odd.
In summary,
(2.1)
∗
h (a, ǫ, p) ≥
(
1 deg(D)/2
L(1, ξ)
qq
√2 q deg(D)/2 L(1, ξ)
q
if deg(D) is even
if deg(D) is odd
Lemma 2.2. Let p ⊂ A be a prime ideal and p its monic generator. For every S ⊆ {(a, ǫ) ∈ A ×
2
F×
q | deg(a − 4ǫp) = deg(p)},
( p
X
√
if deg(p) is even
|S| |Fp |(1 − deg(p)
q )/q
∗
p
h (a, ǫ, p) ≥
deg(p) √
2|S| |Fp |(1 − √q )/ q if deg(p) is odd.
(a,ǫ)∈S
Proof. We first lower bound L(1, ξ) for a ξ corresponding to an arbitrary (a, ǫ).
Y
−1
1
L(1, ξ) =
1 − ξ(ℓ)|ℓ|−1
−
(1 − ξ(∞)|∞| 1)
ℓ prime of A
≥
Y
ℓ prime of A
1 − ξ(ℓ)|ℓ|−1
−1
since ξ(∞) ∈ {0, 1} and |∞| = q. From Weil’s proof of the Riemann hypothesis for curves over finite
√
fields, there exists wi ∈ C with |wi | = q such that
Y
ℓ prime of A
1 − ξ(ℓ)|ℓ|
Since deg(cond(ξ)) − 1 ≤ deg(p),
deg(cond(ξ))−1
Y
i=1
1−
wi
q
−1 −1
≥
=
deg(cond(ξ))−1
Y
i=1
wi
1−
q
.
deg(p)
1
deg p
1− √
≥1− √ .
q
q
For all (a, ǫ) ∈ S, since deg(a2 − 4ǫp) = deg(p), inequality 2.1 implies
( p
X
√
|S| |Fp |(1 − deg(p)
if deg(p) is even
q )/q
∗
p
h (a, ǫ, p) ≥
deg(p) √
√
2|S| |Fp |(1 −
q )/ q if deg(p) is odd.
(a,ǫ)∈S
POLYNOMIAL FACTORIZATION AND EULER-POINCARE CHARACTERISTICS OF DRINFELD MODULES
9
Lemma 2.3. Let p ⊂ A be a prime ideal and p its monic generator. For every S ⊆ {(a, ǫ) ∈ A ×
2
F×
q | deg(a − 4ǫp) = deg(p)}, the number of Drinfeld modules (φ/p) over Fp with (aφ,p , ǫφ,p ) ∈ S is lower
bounded by
( ×
p
√
if deg(p) is even
|Fp ||S| |Fp |(1 − deg(p)
q )/(q(q − 1))
p
|{(φ/p)|(aφ,p , ǫφ,p ) ∈ S}| ≥
√
deg(p)
×
√
2|Fp ||S| |Fp |(1 −
q )/( q(q − 1)) if deg(p) is odd.
2
Proof. Fix an S ⊆ {(a, ǫ) ∈ A × F×
q | deg(a − 4ǫp) = deg(p)}.
|{(φ/p)|(aφ,p , ǫφ,p ) ∈ S}| =
=
X
X
(a,ǫ)∈S (φ/p)∈H(a,ǫ,p)
|F×
p|
AutFp (φ)
X
|F×
p|
h∗ (a, ǫ, p).
q−1
(a,ǫ)∈S
Applying Lemma 2.2, we get
|{(φ/p)|(aφ,p , ǫφ,p ) ∈ S}| ≥
(
p
deg(p)
|F×
if deg(p) is even
p ||S| |Fp |(1 − √q )/(q(q − 1))
p
√
deg(p)
×
2|Fp ||S| |Fp |(1 − √q )/( q(q − 1)) if deg(p) is odd.
3. Degree Estimation and Euler-Poincare Characteristic of Drinfeld Modules
Q
Henceforth, let h ∈ A denote the monic square free reducible polynomial whose factorization h =Q i pi
into monic irreducible polynomials pi ∈ A we seek. By the chinese reminder theorem, Fh = i Fpi
where h and pi s are the principal ideals generated by h and the pi s respectively. We next present a novel
algorithm to compute the degree (call sh ) of the smallest degree factor of h using Drinfeld modules.
For a Drinfeld module φ that has reduction at each prime dividing h,
Y
Y
M
M
(pi − (aφ,pi − 1)/ǫφ,pi ) .
χφ,pi =
(φ/pi )(Fpi ) ⇒ χφ,h =
φ(Fpi ) ∼
φ(Fh ) ∼
=
=
i
i
i
i
Since ∀i, deg((aφ,pi − 1)/ǫφ,pi ) ≤ deg(pi )/2,
χφ,h = h + terms of smaller degree.
In fact,
h − χφ,h =
X
i:deg(pi )=sh
Y
−
1
a
φ,pi
pj + (terms of degree < (deg(h) − ⌈sh /2⌉))
ǫφ,pi
j6=i
⇒ deg(h − χφ,h ) ≤ deg(h) − ⌈sh /2⌉.
When φ is chosen at random, the equidistribution theorem of Gekeler suggests, with high probability,
Y
X
−
1
a
φ,p
i
pj = deg(h) − ⌈sh /2⌉ ⇒ deg(h − χφ,h ) = deg(h) − ⌈sh /2⌉
deg
ǫφ,pi
i:deg(pi )=sh
j6=i
leading to the following algorithm to compute sh .
Algorithm 3.1.
Input : Monic square free reducible polynomial h ∈ A of degree n.
10
ANAND KUMAR NARAYANAN
(1) Choose a Drinfeld module φ by picking gφ ∈ A and ∆φ ∈ A× each of degree less than deg(h)
independently and uniformly at random.
(2) If gcd(∆φ , h) 6= 1, output it as a factor. Else φ has reduction at primes dividing h and we proceed.
(3) Compute χφ,h .
(4) Output: n − deg(h − χφ,h ).
The running time of the algorithm is dominated by step (3). One way to compute χφ,h is as the characteristic polynomial of φt viewed as a linear transformation on Fh . Computing characteristic polynomials
of linear transformations over finite fields can be performed in polynomial time [Sto01].
The output is at least ⌈sh /2⌉. We prove in the ensuing lemma that when q is large enough compared
to n, the output is ⌈sh /2⌉ with probability at least 1/4. We are thus ensured of finding ⌈sh /2⌉ with
probability 1 − δ with only O(log(1/δ)) repetitions of the algorithm. From ⌈sh /2⌉, we infer that sh is
s
either 2⌈sh /2⌉ − 1 or 2⌈sh /2⌉. We can test which one is correct by checking if gcd(tq h − t, h) is non
trivial and further extract the product of factors of degree sh .
Remark 3.2. In the analysis of our algorithms, we may assume without loss of generality that q ≥ c1 nc2
for some absolute positive constants c1 and c2 . If q were smaller, we could choose the smallest prime c
that satisfies q c ≥ c1 nc2 and obtain the factorization in Fqc [t] with the running time unchanged up to
polylogarithmic factors in n. Factors of h irreducible over Fq and of degree prime to c remain irreducible
over Fqc . Factors of h that are irreducible over Fq of degree (say d) divisible by c will split into c
c
c
distinct
Q irreducible factors over Fq . From the factorization over Fq thus obtained, express h as h =
Q
h
where
h
∈
F
[t]
are
irreducible
factors
that
remained
irreducible
over Fqc and hd ∈ Fqc [t]
h
d
i
q
i
d
i
is the product of all irreducible factors of h in Fqc [t] (but not in Fq [t]) of degree d/c. In fact, hd is the
product of all Fq [t] irreducible factors of h of degree d and hence hd ∈ Fq [t]. We may perform equal
degree factorization on hd to obtain all Fq [t] irreducible degree d factors of h. Since c is bounded by
an absolute constant, the post processing steps after obtaining the factorization over Fqc take at most
O(n1+o(1) (log q)1+o(1) + n(log q)2+o(1) ) expected time.
√
Lemma 3.3. If q is odd and q ≥ 2n, algorithm 3.1 outputs ⌈sh /2⌉ with probability at least 1/4.
Proof. For the output deg(h) − deg(h − χφ,h ) of algorithm 3.1 to be ⌈sh /2⌉, it suffices for
Y
X
−
1
a
φ,pi
pj = deg(h) − ⌈sh /2⌉
deg
ǫφ,pi
i:deg(pi )=sh
j6=i
to hold. Since pi are all monic, this is equivalent to
X
aφ,pi ,⌊sh /2⌋
(3.1)
6= 0
ǫφ,pi
i:deg(pi )=sh
where aφ,pi ,⌊sh /2⌋ ∈ Fq denotes the coefficient of the t⌊sh /2⌋ term in aφ,pi .
Fix a factor pj of h of degree sh .
Since gφ ∈ A and ∆φ ∈ A× are each chosen of degree less than deg(h) independently and uniformly at
random (with gcd(∆φ , h) =Q
1), by the chinese remainder theorem, the tuple {(gφ mod pi , ∆φ mod pi )}i
is distributed uniformly in i (Fpi × F×
pi ).
POLYNOMIAL FACTORIZATION AND EULER-POINCARE CHARACTERISTICS OF DRINFELD MODULES
11
In particular, ∀i 6= j, (aφ,pi , ǫφ,pi ) and (aφ,pj , ǫφ,pj ) are independent and
X
X
X
aφ,pj ,⌊sh /2⌋
aφ,pi ,⌊sh /2⌋
aφ,pi ,⌊sh /2⌋
P r
6= 0 =
= −θ P r
6= θ
Pr
ǫφ,pi
ǫφ,pi
ǫφ,pj
θ∈Fq
i:deg(pi )=sh
⇒ Pr
(3.2)
Fix a θ ∈ Fq and let
X
i:deg(pi )=sh
i:deg(pi )=sh ,i6=j
aφ,pi ,⌊sh /2⌋
6= 0 ≥ min P r
θ∈Fq
ǫφ,pi
aφ,pj ,⌊sh /2⌋
6= θ .
ǫφ,pj
2
Sθ := {(a, ǫ) ∈ A × F×
q | deg(a − 4ǫpj ) = deg(pj ), a⌊sh /2⌋ /ǫ 6= θ}
where a⌊sh /2⌋ ∈ Fq denotes the coefficient of the t⌊sh /2⌋ term in a.
We next count elements in Sθ to lower bound its size. The condition deg(a2 −4ǫpj ) = deg(pj ) implies that
we are only allowed to pick a ∈ A of degree at most deg(pj )/2. If deg(pj ) is odd, deg(a2 − 4ǫpj ) = deg(pj )
is always satisfied for every a ∈ A of degree at most deg(pj )/2. To pick an element in Sθ , but for the
coefficient a⌊sh /2⌋ , we may choose the coefficients of a ∈ A arbitrarily of degree at most deg(pj )/2 and
2
arbitrarily choose ǫ ∈ F×
q . For each such choice, to satisfy deg(a − 4ǫpj ) = deg(pj ) and a⌊sh /2⌋ /ǫ 6= θ,
we need to exclude at most two choices for a⌊sh /2⌋ if deg(pj ) is even and at most one choice if deg(pj ) is
odd. Thus
(
p
(q − 2) |Fpj |q
if deg(pj ) is even
p
|Sθ | ≥
√
(q − 1) |Fpj | q if deg(pj ) is odd.
Applying Lemma 2.3 for Sθ , we get
Thus, for
√
1
deg(pj )
1−
|Fpj ||F×
{(φ/pj )|(aφ,pj , ǫφ,pj ) ∈ Sθ } ≥ 1 − √
pj |.
q
q−1
q ≥ 2 deg(h) ≥ 2 deg(pj ),
aφ,pj ,⌊sh /2⌋
1
1
deg(pj )
1−
≥
6= θ ≥ 1 − √
Pr
ǫφ,pj
q
q−1
4
and by equation 3.2 the lemma follows.
Proof of Theorem 1.1 and Corollary 1.2: Theorem 1.1 follows from the proof of correctness (Lemma
3.3) of Algorithm 3.1. Corollary 1.2 follows by considering lines one and three in algorithm 3.1 to be
performed by the black box B in Corollary 1.2. The case gcd(∆φ , h) 6= 1 for a random φ is unlikely to
happen and hence step 2 of algorithm 3.1 may be ignored in the reduction.
When q is large enough (say q ≥ 2 deg(h)4 ), it is likely for a randomly chosen φ that φ(Fh ) is a cyclic
A-module. Further, for a random α ∈ φ(Fh ), it is likely that Ord(α) = Ann(φ(Fh )). Since φ(Fh ) being
cyclic implies χφ,h = Ann(φ(Fh )), it is likely that Ord(α) = χφ,h . Thus, instead of computing χφ,h
in Algorithm 3.1, we could compute Ord(α) for a random α ∈ φ(Fh ) and be assured that the output
deg(h) − deg(h − Ord(α)) likely is ⌈sh /2⌉.
Algorithm 3.4.
Input : Monic square free reducible polynomial h ∈ A of degree n.
(1) Choose a Drinfeld module φ by picking gφ ∈ A and ∆φ ∈ A× each of degree less than deg(h)
independently and uniformly at random.
12
ANAND KUMAR NARAYANAN
(2) If gcd(∆φ , h) 6= 1, output it as a factor. Else φ has reduction at primes dividing h and we proceed.
(3) Choose α ∈ φ(Fh ) at random and compute Ord(α) with constant probability.
(4) If deg(Ord(α)) = deg(h), Output: n − deg(h − Ord(α)).
Every step except for (3) can be performed in O(n log q) time. In § 3.1, we show that the order of
an element in φ(Fh ) (and hence step (3)) can be computed with probability arbitrarily close to 1 in
O(n(1+ω)/2 + o(1)(log q)1+o(1) + n1+o(1) (log q)2+o(1) ) expected time. In the subsequent lemma, we prove a
lower bound on the probability that reductions of Drinfeld modules are cyclic and use it in Theorem 3.6
to prove that Algorithm 3.4 outputs sh with constant probability. Consequently, we have an O(n(1+ω)/2 +
o(1)(log q)1+o(1) + n1+o(1) (log q)2+o(1) ) expected time algorithm to extract a non trivial factor.
Lemma 3.5. For odd q, for everyprime ideal p ⊂ A, the probability that φ(Fp ) is a cyclic A-module for
a randomly chosen φ/p is at least 1 − deg(p)+1/2
.
2(q−1)
Proof. Let p ⊂ A be a prime ideal and p its monic generator. Cojocaru and Papikian [CP14, Cor 3]
determined the following precise characterization of the A-module structure of finite Drinfeld modules
when Fq is of odd characteristic. For a Drinfeld module φ with reduction at p, let fφ,p ∈ A denote the
largest monic square factor of the discriminant a2φ,p − 4ǫφ,p p of Pφ,p . As A-modules
φ(Fp ) ∼
= A/(mφ,p ) ⊕ A/(mφ,p nφ,p )
where
mφ,p = gcd(fφ,p , aφ,p − 2).
In particular, φ(Fp ) is A-cyclic if and only if gcd(fφ,p , aφ,p − 2) = 1. Let
2
2
Sp := {(a, ǫ) ∈ A × F×
q | deg(a − 4ǫp) = deg(p), gcd(a − 4ǫp, a − 2) = 1}.
We next estimate the size of Sp . An element in Sp can be chosen as follows. Pick a ∈ A arbitrarily of
degree at most deg(p)/2. For such a chosen a, to satisfy gcd(a2 − 4ǫp, a − 2) = 1, pick ǫ such that for
all monic irreducible polynomials ℓ dividing a − 2, a2 − 4ǫp 6= 0 mod ℓ. For a fixed monic irreducible
2
ℓ dividing a − 2, there is at most one ǫ ∈ F×
q such that a − 4ǫp = 0 mod ℓ for if there were two, then
that would imply ℓ divides p which contradicts the fact that p is irreducible and of degree higher than
ℓ. Thus, for a chosen a, to ensure a2 − 4ǫp and a − 2 are relatively prime, we need to exclude at most
deg(p)/2 values for ǫ. When deg(p) is even, for a chosen a, to ensure deg(a2 − 4ǫp) = deg(p), we need to
exclude at most one choice for ǫ.
(
p
if deg(p) is even
(q − 2 − deg(p)/2) |Fpj |q
p
⇒ |Sp | ≥
√
(q − 1 − deg(p)/2) |Fpj | q if deg(p) is odd.
Applying Lemma 2.3 for Sp , we get
deg(p) + 1/2
|Fp ||F×
1−
p |.
2(q − 1)
Thus φ(Fp ) is A-cyclic with probability at least 1 − deg(p)+1/2
.
2(q−1)
|{(φ/p)|(aφ,p , ǫφ,p ) ∈ Sp }| ≥
Theorem 3.6. There exists a positive constant c such that for q odd and at least 2n4 , algorithm 3.4
outputs ⌈sh /2⌉ with probability at least c.
Proof. Assume q is odd and q ≥ 2n4 . For a choice of α and φ made in algorithm 3.4, if the following
three conditions hold, then clearly the output is ⌈sh /2⌉.
• Ord(α) = Ann(φ(Fh )),
• Ann(φ(Fh )) = χφ,h ,
POLYNOMIAL FACTORIZATION AND EULER-POINCARE CHARACTERISTICS OF DRINFELD MODULES
13
• deg(h) − deg(h − χφ,h ) = ⌈sh /2⌉.
For a fixed φ and a random α ∈ φ(Fh ), from the A-module decomposition of φ(Fh ) into invariant factors,
we infer that Ord(α) = Ann(φ(Fh )) with probability at least
1
|{a ∈ A| deg(a) < deg(Ann(φ(Fh ))), gcd(a, Ann(φ(Fh ))) = 1}|
≥ (1 − 1/q)deg(h) ≥ 1 − deg(h)/q ≥ .
2
q deg(Ann(φ(Fh )))
The last inequality is a consequence of q ≥ 2 deg(h)4 . If φ(Fh ) is a cyclic A-module, then by definition
χφ,h = Ann(φ(Fh )). Hence, to claim the theorem, it suffices to prove that for a Drinfeld module φ chosen
at random as in algorithm 3.4, the following two conditions hold with constant probability
• φ(Fh ) is a cyclic A-module,
• deg(h) − deg(h − χφ,h ) = ⌈sh /2⌉.
The proof proceeds by induction on the factors of h. Let m denote the number of irreducible
facQ
tors of h. Without loss of generality relabel the irreducible factors of h such that h = m
p
where
i
i=1
Qi
deg(p1 ) ≥ deg(p2 ) ≥ . . . ≥ deg(pm ). In particular, deg(pm ) = sh . Let hi := j=1 pj and hi := (hi ).
Induction Hypothesis: For i < m, assume φ(Fhi ) is a cyclic A-module with probability at least 1 −
The initial case i = 1 of the induction hypothesis (that is, h1 is prime), follows from Lemma 3.5.
deg(hi )
√
q
i−1
We next lower bound the probability that φ(Fh ) (= φ(Fhm )) is A-cyclic and deg(h) − deg(h − χφ,h )) =
⌈sh /2⌉ conditioned on φ(Fhm−1 ) being A-cyclic. Since gφ ∈ A and ∆φ ∈ A× are each chosen of degree less
than deg(h) independently and uniformly at random (with gcd(∆φ , h) = 1),Qby the chinese remainder
theorem, the tuple {(gφ mod pi , ∆φ mod pi )}i is distributed uniformly in i (Fpi × F×
pi ). In particular, (aφ,pm , ǫφ,pm ) is independent of the structure of φ(Fhm−1 ). Consequently, instead of conditioning on
φ(Fhm−1 ) being A-cyclic, we fix a tuple
Y
(gφ mod pi , ∆φ mod pi )
i<m
such that φ(Fhm−1 ) is A-cyclic. In particular, χφ,hm−1 is fixed. In the remainder of the proof, the only
randomness in φ comes from choosing (gφ mod pm , ∆φ mod pm ) uniformly at random from Fpm × F×
pm .
As reasoned in the proof of Lemma 3.3, for deg(h) − deg(h − χφ,h ) to be ⌈sh /2⌉, it suffices for
X
(3.3)
(aφ,pi ,⌊sh /2⌋ /ǫφ,pi ) 6= 0
i:deg(pi )=sh
to hold where aφ,pi ,⌊sh /2⌋ ∈ Fq denotes the coefficient of the t⌊sh /2⌋ term in aφ,pi .
Since ∀i 6= j, (gφ mod pi , ∆φ mod pi ) is fixed,
X
θ :=
(aφ,pi ,⌊sh /2⌋ /ǫφ,pi )
i:deg(pi )=sh ,i6=m
is fixed. Clearly equation 3.3 holds if and only if (aφ,pm ,⌊sh /2⌋ /ǫφ,pm ) 6= θ.
Since φ(Fhm−1 ) is A-cyclic, if φ(Fpm ) is A-cyclic and χφ,pm is relative prime to χφ,hm−1 , then φ(Fh )
is A-cyclic. Thus, to ensure φ(Fh ) is A-cyclic, it suffices if gcd(pm − (aφ,pm − 1)ǫφ,pm , χφ,hm−1 ) = 1 and
.
14
ANAND KUMAR NARAYANAN
gcd(a2φ,pm − 4ǫφ,pm pm , aφ,pm − 2) = 1. The argument as to why is identical to the discussion in the proof
of Lemma 3.5.
For a ∈ A, denote by a⌊sh /2⌋ ∈ Fq the coefficient of t⌊sh /2⌋ in a. To summarize, the set Spm of tuples (a, ǫ) ∈ A × F×
q that satisfy the four conditions
(i)
(ii)
(iii)
(iv)
deg(a2 − 4ǫpm ) = deg(pm ),
a⌊sh /2⌋ 6= θǫ,
gcd(ǫpm − (a − 1), χφ,bh ) = 1,
gcd(a2 − 4ǫpm , a − 2) = 1,
has the property that if (aφ,pm , ǫφ,pm ) ∈ Spm then φ(Fh ) is A-cyclic and deg(h) − deg(h − χφ,h ) = ⌈sh /2⌉.
We estimate a lower bound on the size of Spm by choosing a ∈ A arbitrarily of degree at most deg(pm )/2
and for each choice of a, picking only those ǫ such that the four conditions are satisfied. To ensure the
first condition, we need to exclude at most one choice each for ǫ. To satisfy the fourth condition, for a
fixed choice of a, we need to exclude at most deg(a − 2) ≤ deg(pm )/2 choices for ǫ. This is because for
2
each monic irreducible polynomial ℓ dividing a − 2, there is at most one ǫ ∈ F×
q such that a − 4ǫpm = 0
mod ℓ. For if there were two, then ℓ would divide pm , which is an irreducible polynomial of degree higher
than deg(ℓ). To satisfy the third condition, for a fixed choice of a 6= 1, we need to exclude at most
deg(hm−1 ) choices for ǫ. This is because for each monic irreducible polynomial ℓ dividing χφ,hm−1 , there
is at most one ǫ ∈ F×
q such that ǫpm − (a − 1) = 0 mod ℓ. For if there were two, then ℓ would divide pm
in which case restricting to a 6= 1 assures that ǫpm − (a − 1) 6= 0 mod ℓ. If θ 6= 0, to satisfy the second
condition we need to exclude at most one choice for ǫ which implies
(p
if deg(pm ) is even
|Fp | − 1(q − deg(hm−1 ) − deg(pm )/2 − 3)q
θ 6= 0 ⇒ |Spm | ≥ p m
√
|Fpm | − 1(q − deg(hm−1 ) − deg(pm )/2 − 3) q if deg(pm ) is odd.
For θ = 0, the second condition can be satisfied by restricting the count to non zero a⌊sh /2⌋ and we get
(p
if deg(pm ) is even
|Fp | − 1(q − deg(hm−1 ) − deg(pm )/2 − 2)(q − 1)/q
θ = 0 ⇒ |Spm | ≥ p m
√
|Fpm | − 1(q − deg(hm−1 ) − deg(pm )/2 − 2)(q − 1)/ q if deg(pm ) is odd.
Applying Lemma 2.3 for Spm and assuming q ≥ 2 deg(h)4 , we get
deg(h)
|{(φ/pm )|(aφ,pm , ǫφ,pm ) ∈ Spm }| ≥ 1 − √
|Fpm ||F×
pm |.
q
The probability that φ(Fh ) is A-cyclic and deg(h) − deg(h − Ann(φ(Fh ))) = ⌈sh /2⌉ conditioned on
φ(Fhm−1 ) being A-cyclic is hence at least
1−
deg(h)
√ .
q
By induction φ(Fh ) is A-cyclic and deg(h) − deg(h(t) − Ann(φ(Fh ))) = ⌈sh /2⌉ with probability at least
m
deg(h)
m deg(h)
deg(h)2
1− √
≥1−
≥1− √
√
q
q
q
which is lower bounded by a constant since q ≥ 2 deg(h)4 and the theorem follows.
POLYNOMIAL FACTORIZATION AND EULER-POINCARE CHARACTERISTICS OF DRINFELD MODULES
15
3.1. Order Finding in Finite Drinfeld Modules. We sketch a Montecarlo randomized algorithm to
compute Ord(α) with probability arbitrarily close to 1 in O(n(1+ω)/2+o(1) (log q)1+o(1) + n1+o(1) (log q)2 )
time2. The algorithm works for every φ with reduction at h and every α ∈ φ(Fh ). Fix a φ with reduction
at h and an α ∈ φ(Fh ). Compute the minimal polynomial of the linear sequence {U(φjt (α)), j ∈ Z≥0 }
where U : Fh −→ Fq is a random Fq linear map. The minimal polynomial of the linear sequence divides
Ord(α) and with probability at least half equals Ord(α). Hence the least common multiple of the minimal
polynomials of the resulting linear sequences of O(log(δ)) independent trials is Ord(α) with probability
at least 1 − δ. For a trial, the minimal polynomial of a sequence can be computed in O(n1+o(1) log q) time
using the fast Berlekamp Massey algorithm given the first 2 deg(h) − 1 elements in the sequence. Hence
the critical step is the computation of
2 deg(h)−2
{U(α), U(φt (α)), U(φ2t (α)), . . . , U(φt
(3.4)
(α))}
for a randomly chosen U. This is virtually identical to the automorphism projection problem of KaltofenShoup. The difference being that the Frobenius endomorphism modulo h is replaced by the Drinfeld
endomophism φt modulo h. In Kaltofen-Shoup apart from being an Fq linear endomorphism of Fh , the
e 1+o(1) (log q)2 ) time
only property of the Frobenius exploited is that τ (t mod h) can be computed in O(n
using the vonzur Gathen-Shoup algorithm. To adapt the automorphism projection algorithm of Kaltofen
and Shoup [KS98][§ 3.2] to apply in our setting, we merely have to demonstrate how to efficiently compute
φt (t mod h) given h, gφ mod h and ∆φ mod h. Since
φt (t
mod h) = t2
mod h + τ (t
mod h) + τ 2 (t
mod h),
we can compute φt (t mod h) in O(n1+o(1) (log q)2 ) time with three Frobenius powers and two additions
modulo h thereby making the Katofen-Shoup algorithm applicable to our setting.
3.2. Obtaining the Complete Factorization from a Factor Finding Procedure. In this subsection, we prove that given access to a blackbox D that takes as input a square free f ∈ A and outputs an
irreducible factor, there is an O(n4/3+o(1) (log q)1+o(1) ) expected time algorithm F to factor a polynomial
of degree n over Fq into its irreducible factors. Further, this algorithm makes at most n1/3 calls to D.
Thereby, Corollary 1.3 would follow from Theorem 1.1.
Without loss of generality, assume that the input h ∈ A to F is square free and of degree n. Obtaining the factorization
of h by extracting one irreducible factor at a time using D could in the worst
√
case take Θ( n) calls to D. A faster alternative is to use the Kaltofen-Shoup algorithm with fast modular
composition to extract small degree factors of h and then invoke D to extract the large degree factors
one at a time. In particular, using [KU08, Lem 8.4, Thm 8.5], extract all the irreducible factors of h of
degree at most n2/3 in O(n4/3+o(1) (log q)1+o(1) ) expected time. The remaining irreducible factors of h
each have degree at least ⌈n2/3 ⌉. Hence there are at most n1/3 irreducible factors of h remaining and the
complete factorization of h can be obtained by extracting a factor at a time with at most n1/3 calls to D.
Remark 3.7. Kaltofen and Shoup [KS98, § 3.1] through the blackbox Berlekamp algorithm [KL94] reduced
polynomial factorization in time nearly linear in degree to two problems that are transposes of each
other, namely automorphism projection and automorphism evaluation. Being transposes, a straight
line program that computes Fq linear forms in the input for one would in linear time yield a straight
line program for the other of the same complexity. In particular, there is nearly linear polynomial
factorization algorithm if there is a nearly linear time Fq linear solution to the automorphism projection
problem. Our order finding problem is no harder than automorphism projection. We hence arrive at the
stronger assertion that polynomial factorization is reducible to automorphism projection. In particular,
2We may replace (1 + ω)/2 with ω /2, where ω is the exponent of n × n by n × n2 matrix multiplication (see [KU08]).
2
2
16
ANAND KUMAR NARAYANAN
no assumptions on the Fq linearity of automorphism projection algorithm is made. We must however
remark that the automorphism projection we consider (see equation 3.4) is broader than that stated in
[KS98, § 3.2] where only the Frobenius automorphism is considered.
3.3. Degree Estimation Using Carlitz Modules. The degree estimation algorithm framework also
gives rise to variants where there is no randomization with respect to the choice of Drinfeld modules. In
fact, the following deterministic example from the author’s Ph.D thesis [Nar14] using Carlitz modules
(rank 1 Drinfeld modules) partly motivated the randomized version. The Carlitz module based algorithm
is suited to the case when the characteristic of Fq does not divide the number of factors of the smallest
degree.
Example 3.8. Factor Degree Estimation using Carlitz Modules.
Input : Monic square free reducible polynomial h ∈ A.
(1) Choose the Carlitz module φ (the rank 1 Drinfeld module φ with gφ = 1 and ∆φ = 0).
(2) Compute χφ,h .
(3) Output: deg(h) − deg(h − χφ,h ).
In [Nar14] it is shown that the output is exactly sh provided the number of factors of degree sh is not
divisible by the characteristic of Fq . This is true since for the Carlitz module φ, for all prime ideals p ⊂ A,
χφ,p = p − 1 where p is the monic generator of p.
Curiously for the Carlitz module φ, finding χφ,h is easily seen to be no harder than factoring h. Computing χφ,h is linear time reducible to factoring h since given the factorization of h, it is trivial to write
down χφ,h in O(deg(h) log q) time.
4. Factorization Patterns of Polynomials in Small Intervals
Our analysis of the Drinfeld module analogue of the black-box Berlekamp algorithm relies on the degree
distribution in factorization patterns of polynomials in short intervals which we study in this section.
For a partition λ of a positive integer e, let Cλ := {σ ∈ Se |λσ = λ} denote its conjugacy class where Se
is the symmetric group on e elements and λσ is the partition of e induced by the factorization of σ into
disjoint cycles. Let P (λ) := |Cλ |/|Se |. For a polynomial f ∈ A, let λ(f ) denote the partition of deg(f )
induced by the degrees of the irreducible factors in the factorization of f in A.
For f ∈ A and a positive integer m < deg(f ), define the interval around f corresponding to the degree bound m as
Im (f ) := {f + a|a ∈ A, deg(a) ≤ m}.
For I ⊂ A where each polynomial in I is of degree exactly d > 1 and a partition λ of d, define
Bq (I, λ) := {a ∈ I|λ(a) = λ}
and πq (I, λ) := |Bq (I, λ)|.
Bank, Bary-Soroker and Rosenzweig [BBR14] recently proved the following theorem when the field size
q tends to infinity while d is fixed.
Theorem 4.1. ([BBR14, Thm 1]) For all monic f ∈ A of fixed degree d, for all positive integers
2 < m < d and for all partitions λ of d,
πq (Im (f ), λ) ∼ P (λ)|Im (f )|
It is widely conjectured (see [BBR14]) that
as
q → ∞.
POLYNOMIAL FACTORIZATION AND EULER-POINCARE CHARACTERISTICS OF DRINFELD MODULES
Conjecture 4.2. For all monic f ∈ A of degree d such that 3 < d <
πq (Im (f ), λ) ∼ P (λ)|Im (f )|
as
17
√
q/2 and for all partitions λ of d,
q d → ∞.
In the next subsection § 4.1, by applying an effective Lang-Weil bound to the argument in [BBR14], we
prove an effective version of Theorem 4.1 that holds for log q > 5d log(d).
4.1. A High Dimensional Variant of the Function Field Chebotarev Density Theorem. Let E
denote the rational function field Fq (t1 , . . . , tm ) in the m indeterminates t1 , . . . , tm . Let F/E be a finite
Galois extension of E. Fix an algebraic closure Fq of Fq and let
α : Gal(F/E) −→ Gal((Fq ∩ F )/Fq )
denote the restriction map. Let V = Spec(Fq [t1 , . . . , tm ]) and let Vur (Fq ) ⊂ V (Fq ) denote the subset
of Fq rational places in V that are etale in the extension F/E. Let OF denote the integral closure of
Fq [t1 , . . . , tm ] in F and let W = Spec(OF ). For a place B ∈ W lying above a place p ∈ V that is etale in
F/E, let σB ∈ Gal(F/E) denote its Artin symbol. For a place p ∈ V that is etale in F/E, let
Θp := {σB |B ∈ W, B|p} ⊆ ker(α)
denote the conjugacy class of Artin symbols above p.
Lemma 4.3. If q ≥ 2(m + 1)[F : E]2 , for every conjugacy class Θ ⊆ ker(α),
|Θ|
|Θ|
qm
m
13/3 m−1
|{p ∈ Pur (Fq )|Θp = Θ}| −
q ≤
q
([F : E] − 1)([F : E] − 2) √ + 5[F : E]
.
|ker(α)|
|ker(α)|
q
Proof. Fix a conjugacy class Θ ⊆ ker(α) and let U := {p ∈ Pur (Fq )|Θp = Θ}.
Let ρ : W −→ V denote the norm map from W down to V . Applying [Bar12][Prop. 2.2] to (V,W,ρ, Θ)
c and a finite separable morphism
implies the existence of a smooth irreducible affine Fq -variety W
c
π : W −→ V such that
c) = U ,
(i) π(W
(ii) deg(π) = |ker(α)|,
c (Fq )| = |ker(α)|/|Θ|.
(iii) ∀p ∈ U, |π −1 (p) ∩ W
c −→ V is finite, W
c and V have the same dimension, namely m. Further, W and W
c are twists
Since π : W
c
of each other [Bar12]. As a consequence, W and W have the same degree, namely [F : E].
c (Fq ) using an effective Lang-Weil bound [CM06],
Bounding the size of W
qm
c (Fq )| − q m ≤ ([F : E] − 1)([F : E] − 2) √
+ 5[F : E]13/3 q m−1 .
|W
q
c ) = U and ∀p ∈ U, |π −1 (p) ∩ W
c (Fq )| = |ker(α)|/|Θ|, |W
c (Fq )| = |ker(α)||U|
Since π(W
|Θ|
m
|Θ|
q
|Θ|
⇒ |U | −
([F : E] − 1)([F : E] − 2) √ + 5[F : E]13/3 q m−1 .
qm ≤
|ker(α)|
|ker(α)|
q
Theorem 4.4. For every positive integer m ≥ 2, for every monic f ∈ A of degree greater than m and
for every partition λ of deg(f ), if log q ≥ 5 deg(f ) log(deg(f )) then
1
|πq (Im (f ), λ) − P (λ)|Im (f )|| ≤ P (λ)|Im (f )|.
2
18
ANAND KUMAR NARAYANAN
Proof. Fix a monic non constant polynomial f ∈ A of degree at least m and let
m
X
xi ti−1 .
Ff := f (t) +
i=1
Since the indeterminate x1 only appears in Ff as the linear term x1 , Ff is absolutely irreducible and
separable in t. Thus the splitting field Ff of Ff over E = Fq (x1 , . . . , xm ) is Galois. We will shortly apply
Lemma 4.3 to the extension Ff /E. Before doing so, we argue that Ff /E is a geometric extension.
The splitting field of Ff over Fq (x1 , . . . , xm ) is the composite Ff .Fq and we have
Gal(Ff .Fq /Fq (x1 , . . . , xm )) = Gal(Ff .Fq /Fq (x1 , . . . , xm )) ≤ Gal(Ff /E) ≤ Sdeg(f ) .
By [BBR14][Prop 3.6], Gal(Ff .Fq /Fq (x1 , . . . , xm )) ∼
= Sdeg(f )
⇒ Gal(Ff .Fq /Fq (x1 , . . . , xm )) ∼
= Gal(Ff /E) ∼
= Sdeg(f ) ⇒ Ff ∩ Fq = Fq .
Hence Ff /E is a geometric extension. Since Gal(Ff ∩ Fq /Fq ) is trivial, the restriction map
αf : Gal(Ff /E) −→ Gal(Ff ∩ Fq /Fq )
has kernel
ker(αf ) = Gal(Ff /E) ∼
= Sdeg(f ) .
Since Gal(Fq /Fq ) = hτ i ∼
= Z where τ is the q th power Frobenius, homomorphisms from Gal(Fq /Fq ) to
Sdeg(f ) are parametrized by the permutations in Sdeg(f ) they map τ to. That is, σ ∈ Sdeg(f ) corresponds
to θσ ∈ Hom(Gal(Fq /Fq ), Sdeg(f ) ) that takes τ to σ.
Fix a partition λ of deg(f ). For the conjugacy class Θλ := {θσ |σ ∈ Cλ } ⊆ ker(αf ),
|Cλ |
|Θλ |
=
= P (λ)
| ker(α)|
|Sdeg(f ) |
∼ Sdeg(f ) and |Θλ | = |Cλ | and the second
where the first equality follows from the fact that ker(α) =
equality follows from the definition of P (λ).
(4.1)
Equation 4.1 together with Lemma 4.3 applied to the extension Ff /E yields
||{p ∈ Vur,f (Fq )|Θp = Θλ }| − P (λ)q m | ≤ 2P (λ)[Ff : E]m+1 q m/2
where Vur,f (Fq ) ⊆ V (Fq ) is the set of Fq -rational places in E that are etale in Ff . Identifying V (Fq )
with Am (Fq ), a prime p = (a1 , . . . , am ) ∈ Vur,f (Fq ) has Θp = Θλ if and only if λ(Ff (a1 , . . . , am , t)) ∈ Cλ
([Bar12, Lem 2.1]). Thus
For q > 20 deg(f )5 ,
||{(a1 , . . . , am ) ∈ Vur,f (Fq )|λ(Ff (a1 , . . . , am , x)) = λ}| − P (λ)q m |
qm
13/3 m−1
≤ P (λ) (deg(f ) − 1)(deg(f ) − 2) √ + 5 deg(f )
.
q
q
1
(deg(f ) − 1)(deg(f ) − 2) 5 deg(f )13/3
+
≤ ,
√
q
q
2
3P (λ)
P (λ)
|Im (f )| ≤ πq (Im (f ), λ) ≤
|Im (f )| + πqra (Im (f ), λ)
⇒
2
2
where
πqra (Im (f ), λ) := {(a1 , . . . , am ) ∈ V (Fq ) \ Vur,f (Fq )|λ(Ff (a1 , . . . , am , x)) = λ}|
POLYNOMIAL FACTORIZATION AND EULER-POINCARE CHARACTERISTICS OF DRINFELD MODULES
19
accounts for ramified primes with the factorization pattern corresponding to λ. The ramified part
πqra (Im (f ), λ) is bounded by the number of Fq points in the variety defined by the discriminant ∆t (Ff ) ∈
Fq [x1 , . . . , xm ] of Ff with respect to t. By expressing ∆t (Ff ) as the resultant of Ff and its derivative
with respect to t, we see, deg(∆t (Ff )) ≤ 2[Ff : E] − 1. Applying an effective version of the Lang-Weil
bound [CM06], πqra (Im (f ), λ) turns out to be negligible in our computation.
Since [Ff : E] = deg(f )!, for log q > 5 deg(f ) log(deg(f )),
1
3
P (λ)|Im (f )| ≤ πq (Im (f ), λ) ≤ P (λ)|Im (f )|
2
2
and the theorem follows.
Remark 4.5. Recent break through algorithms for discrete logarithm computation [Jou][BGJT] over a
small characteristic finite field (say Frd ) have the following initial polynomial search step. Given r and
d, search for h0 , h1 ∈ Fr2 [t], each of degree 2 such that the factorization of h1 tr − h0 over Fr2 [t] has
an irreducible factor of degree d. The search is known to succeed only under heuristic assumptions. If
Theorem 4.4 were true for q ≥ (n − 1)2 , then as a corollary (by setting q = r2 , f = tr+1 and m = 2), the
search provably succeeds (even when h1 is fixed as h1 = t) without making any heuristic assumptions.
More generally, if Theorem 4.4 holds for q ≥ c1 nc2 for some positive absolute constants c1 , c2 , then the
heuristic assumptions in the polynomial selection step (with appropriate modifications) may be removed.
5. Drinfeld Module Analog of Berlekamp’s Algorithm
We motivate the Drinfeld module analog of Berlekamp’s algorithm with a brief description of Lenstra’s
algorithm for integer factorization. Pollard’s p-1 algorithm [Pol74] is designed to factor an integer that
has a prime factor modulo which the multiplicative group has smooth order. Say for instance that a
positive integer n has a prime factor p such that every prime power factor of p − 1 is bounded by b.
The algorithm proceeds by choosing a positive integer B as the smoothness bound and computes m, the
product of all prime powers bounded by B. A positive integer a < n is then chosen at random. Assume
a is prime to n for otherwise gcd(a, n) is a non trivial factor of n. If B ≥ b, since p − 1 divides m,
am − 1 = (ap−1 )m/(p−1) − 1 ∼
= 0 mod p ⇒ p | am − 1
and gcd(am − 1, n) is likely a non trivial factor of n.
The running time is at least exponential in the size of B. For typical n, B needs to be as big as the
smallest factor of n and thus the running time is typically exponential in the size of the smallest factor of n.
Lenstra’s elliptic curve factorization algorithm [Len87] factors every n in (heuristic) expected time subexponential in the size of the smallest factor p of n. A key insight of Lenstra was to substitute the
multiplicative group (Z/pZ)× in Pollard’s p-1 algorithm with the group E(Fp ) of Fp rational points of a
random elliptic curve E over Fp . The running time depends on the smoothness of the group order |E(Fp )|
√
for a randomly chosen E. The Hasse-Weil bound guarantees that ||E(Fp )| − (p + 1)| ≤ 2 p and Lenstra
proved that his algorithm runs in expected time subexponential in the size of p assuming a heuristic on
√
√
the probability that a random integer in the interval [p + 1 − 2 p, p + 1 + 2 p] is smooth.
Our algorithm can be thought of as an analogue of Berlekamp’s algorithm wherein the Frobenius action is replaced with a random rank-2 Drinfeld action; much like Lenstra’s algorithm is an analogue
of Pollard’s p − 1 obtained through replacing the multiplicative group modulo a prime with a random
elliptic curve group. Before outlining the algorithm, a few remarks regarding notation are in order. For a
20
ANAND KUMAR NARAYANAN
positive integer b, we call a polynomial b-smooth if all its irreducible factors are of degree at most b. For
a Drinfeld module φ (with reduction at primes dividing h) and β ∈ φ(Fh ), by gcd(β, h) we really mean
the gcd of h and a lift of β to A.
Drinfeld Module Analogue of the Black-box Berlekamp Algorithm.
Input : Monic square free reducible polynomial h ∈ A.
(1) Pick a smoothness bound b ≥ 1.
(2) Choose a Drinfeld module φ at random by picking gφ ∈ A and ∆φ ∈ A× each of degree less than
deg(h) independently and uniformly at random.
(3) Choose a random non zero α ∈ φ(Fh ) and compute Ord(α).
(4) Find a monic b-smooth factor f of Ord(α) (if one exists).
(5) Output: gcd(φOrd(α)/f (α), h) is likely a non trivial factor of h.
There is flexibility on how the smooth factor f is determined once b is chosen. One extreme is to set f
to be the largest b-smooth factor of Ord(α). The other, is to further factor the largest b-smooth factor
of Ord(α) (recursively or by other means) and to set f to one of the b-smooth irreducible factors of
Ord(α). A rigorous analysis of the former choice with b = 1 is in § 5.1. An informal discussion of why
the algorithm is likely to succeed with the latter choice follows keeping in mind that
M
φ(Fpi ) , χφ,pi = pi − (aφ,pi − 1)/ǫφ,pi , ∀i.
φ(Fh ) ∼
=
i
For d > 0, a random polynomial of degree d has a linear factor with at least constant probability. Assume
the plausible hypothesis (which is for large q true by Theorem 1.4) that for every i, the probability
of a polynomial in the interval Ipi := {pi + a | a ∈ A, deg(a) ≤ deg(pi )/2} around pi possessing a
b-smooth factor roughly equals the probability of a random polynomial of degree deg(pi ) possessing a bsmooth factor. This smoothness hypothesis along with the equidistribution of characteristic polynomials
(equation 2.1) suggests it is likely for every b > 0 that there exists a j such that χφ,pj = pj −(aφ,pj −1)/ǫφ,pj
has a b-smooth factor. Since Ann(φ(Fpj )) divides χ(φ(Fpj )), it is thus likely that Ann(φ(Fh )) (which is
the least common multiple of {Ann(φ(Fpi ))}i ) has a b-smooth factor. Assuming that is the case, since
α is chosen at random, with probability at least 1 − 1/q there is a monic b-smooth polynomial dividing
Ord(α). The algorithm picks one such monic irreducible factor f of Ord(α). The fact that the reduction
of φ at h is random and equidistribution of characteristic polynomials (equation 2.1) imply the likely
existence of k such that f (t) does not divide χφ,pk . Consequently
Y
φOrd(α)/f (α) ∼
pi
= 0 mod pk , φOrd(α)/f (α) ≇ 0 mod
i6=k
and thus gcd(φOrd(α)/f (α), h) is likely a non trivial factor of h.
The computation of Ord(α) can be performed efficiently through linear algebra as discussed in § 3.1.
This is in stark contrast to the integer analog, where finding the order of an element in the multiplicative
group modulo a composite appears hard. A consequence is that unlike Lenstra’s algorithm, our success
probability is reliant not on χφ,p being smooth but merely on it possessing a smooth factor. The running
times are thus bounded by a polynomial in the problem size.
5.1. Drinfeld Module Analogue of Berlekamp’s Algorithm With Linear Smoothness. In this
section we formally state and analyze the version of the Drinfeld analog of blackbox Berlekamp algorithm
where the smooth factor chosen is the product of all linear factors of the order of a randomly chosen
element in φ(Fh ).
POLYNOMIAL FACTORIZATION AND EULER-POINCARE CHARACTERISTICS OF DRINFELD MODULES
21
Algorithm 5.1.
Input : Monic square free reducible polynomial h ∈ A of degree n.
(1) Choose a Drinfeld module φ by picking gφ ∈ A and ∆φ ∈ A× each of degree less than deg(h)
independently and uniformly at random.
(2) If gcd(∆φ , h) 6= 1, output it as a factor. Else φ has reduction at primes dividing h and we proceed.
(3) Choose α ∈ φ(Fh ) at random and compute Ord(α).
(4) Compute f = gcd(tq − t, Ord(α)).
(5) Compute φOrd(α)/f (α).
(6) Output: gcd(h, φOrd(α)/f (α)).
The running time of the algorithm is dominated by steps (3) and (5). As in § 3.1 , step (3) can be performed with O(n(1+w)/2+o(1) (log q)o(1) + n1+o(1) (log q)2 ) expected time by adapting the automorphism
projection algorithm of Kaltofen-Shoup. Step (5) poses the transpose problem of step (3) and can be
performed in identical expected time as step 3 by the transposition principle (see [KS98, § 3.2]).
The rest of the section is devoted to showing that algorithm 5.1 outputs a non trivial factor of h with
constant probability. In fact, we prove something stronger in Lemma 5.2 by showing that there exists positive constants c1 and c2 such that for every factor pi of h, pi divides gcd(h, φOrd(α)/f (α)) with probability
at least c1 and pi does not divide gcd(h, φOrd(α)/f (α)) with probability at least c2 . As a consequence,
not merely a factor but the complete factorization of h can be obtained by recursing the algorithm with
recursion depth bounded by O((log n)2 ) (see [KS98, § 3]).
Lemma 5.2. There exists positive a constant c such that, for q odd and log q ≥ 5n log n, at the termination of algorithm 5.1, for every prime factor p of h with monic generator p, p divides gcd(h, φOrd(α)/f (α))
with probability at least c and p does not divide gcd(h, φOrd(α)/f (α)) with probability at least c.
Proof. Fix a prime factor p of h with monic generator p. Assume q is odd, log q ≥ 3n log n and let
2
q
Sin := {(a, ǫ) ∈ A × F×
q | deg(a − 4ǫp) = deg(p), gcd(t − t, p − (a − 1)/ǫ) = 1},
2
q
Sout := {(a, ǫ) ∈ A × F×
q | deg(a − 4ǫp) = deg(p), gcd(t − t, p − (a − 1)/ǫ) 6= 1}.
b denote the set of partitions of deg(p)
Let Λ denote the set of partitions of deg(p) that contain 1 and let Λ
that do not contain 1.
Since deg(a2 − 4ǫp) = deg(p) is always true when deg(p) is odd and deg(a) ≤ 2 deg(p), by Theorem
4.4 it follows for deg(p) odd that
!√
X
q(q − 1) q
|Fp |,
|Sin | ≥
P (λ)
2
λ∈Λ
√
X
q(q − 1) q
|Sout | ≥
P (λ)
|Fp |.
2
b
λ∈Λ
When deg(p) is even, since the characteristic of Fq is assumed odd, we can enforce deg(a2 − 4ǫp) = deg(p)
by restricting the choice of ǫ to ensure 4ǫ is not a square in F×
q and picking a ∈ A arbitrarily of degree at
most deg(p)/2. There are at least (q − 1)/2 such choices for ǫ and applying Theorem 4.4 once for each
such choice we get for deg(p) even and Fq of odd characteristic,
!
q
X
q(q − 1)
|Fp |,
2|Sin | ≥
P (λ)
2
λ∈Λ
22
ANAND KUMAR NARAYANAN
2|Sout | ≥
X
b
λ∈Λ
q(q − 1)
P (λ)
2
q
|Fp |.
The number of permutations in Sdeg(p) with no fixed points is ⌈deg(p)!/e⌉ if deg(p) is even and
⌊deg(p)!/e⌋ otherwise. Thus
X
X
⌊deg(p)!/e⌋
⌈deg(p)!/e⌉
P (λ) ≥
and
P (λ) ≥ 1 −
deg(p)!
deg(p)!
λ∈Λ
b
λ∈Λ
and there exists positive constants b1 , b2 such that
X
X
P (λ) ≥ b1 and
P (λ) ≥ b2 .
λ∈Λ
b
λ∈Λ
Applying Lemma 2.3 once each to Sin and Sout , there exists positive constants d1 and d2 such that,
{(φ/p)|(aφ,p , ǫφ,p ) ∈ Sin } ≥ d1 |Fp ||F×
p |,
{(φ/p)|(aφ,p , ǫφ,p ) ∈ Sout } ≥ d2 |Fp ||F×
p |.
Since gφ ∈ A and ∆φ ∈ A× are each chosen of degree less than deg(h) independently and uniformly at
random and gcd(∆φ , h) = 1, by the chinese remainder theorem (gφ mod p, ∆φ mod p) is distributed
uniformly in Fpi × F×
pi . Thus the probability that Ann(φ(Fp )) has a linear factor is at least d1 and
the probability that Ann(φ(Fp )) does not have a linear factor is at least d2 . Since the projection of a
random α ∈ φ(Fh ) into φ(Fp ) has order Ann(φ(Fp )) with probability at least (1 − 1/q) ≥ 1/2, the lemma
follows.
Remark 5.3. If q < 5n log n, we may work over a finite extension F′q /Fq such that q ′ ≥ 3n log n and by
Lemma 5.2 be assured that the algorithm 5.1 succeeds, however, we incur an extra factor of n in the
expected running time. If Conjecture 4.2 is true, then Lemma 5.2 holds with only the requirements that
√
√
q is odd and q ≥ 2n. The assumption q ≥ 2n may be made without loss in generality (Remark 3.2).
Acknowledgement
I thank Lior Bary-Soroker, Zeyo Guo, Ming-Deh Huang and Chris Umans for valuable discussions.
References
[BBR14] E. Bank, L. Bary-Soroker and L Rosenzweig, Prime polynomials in short intervals and in arithmetic progressions,
http://arxiv.org/abs/1302.0625
[BGJT] R. Barbulescu, P. Gaudry, A. Joux , E. Thome, “A quasi-polynomial algorithm for discrete logarithm in finite
fields of small characteristic”, http://arxiv.org/abs/1306.4244
[Bar12] L. Bary-Soroker, Irreducible values of polynomials, Adv. Math., 229 (2), 854-874 (2012).
[Ber67] E. R. Berlekamp, Factoring Polynomials Over Finite Fields, Bell System Tech. J., 46:1853-1849. 1967.
[CM06] A. Cafure and G. Matera, Improved explicit estimates on the number of solutions of equations over a finite field,
Finite Fields and Their Applications Volume 12, Issue 2, April 2006, Pages 155185
[Coh] S. D. Cohen, The Distribution of Polynomials over Finite Fields, Acta Arith., 17 (1970), 255-271.
[CZ81] D. G. Cantor and H. Zassenhaus, A new algorithm for factoring polynomials over finite fields, Math. Comp., vol.
36, 587592, 1981.
[CP14] A. C. Cojocaru and M. Papikian, Drinfeld Modules, Frobenius Endomorphisms, and CM-Liftings, Int Math Res
Notices (2014)
[Cor99] G. Cornellison, Deligne’s congruence and supersingular reduction of Drinfeld modules, Arch. Math. 72 (1999)
346-353.
[Del74] P. Deligne, “La conjecture de Weil. I”, Publications Mathematiques de l’IHES (43): 273-307.
[Deu41] M. Deuring, Die Typen der Multiplikatorenringe elliptischer Funktionenkorper, Abh. Math. Sem. Hamburg 14
(1941), 197272. MR0005125.
POLYNOMIAL FACTORIZATION AND EULER-POINCARE CHARACTERISTICS OF DRINFELD MODULES
23
[Dor] D. R. Dorman, On singular moduli for rank 2 Drinfeld modules, Compositio Mathematica (1991) Volume: 80, Issue:
3, page 235-256.
[Dri74] V. G. Drinfeld, Elliptic modules, Mat. Sb. (N.S.), 1974, Volume 94(136), Number 4(8), Pages 594627
[Dri77] V. G. Drinfeld, Elliptic modules. II, Mat. Sb. (N.S.), 102(144):2 (1977), 182-194.
[vzGS92] J. von zur Gathen and V. Shoup, Computing Frobenius maps and factoring polynomials, Comput. Complexity,
vol. 2, 187-224, 1992.
[Gek08] E-U Gekeler, Frobenius distributions of Drinfeld modules over finite fields, Trans. Amer. Math. Soc. 360 (2008),
1695-1721.
[Gek91] E-U Gekeler, On finite Drinfeld modules, Journal of Algebra, Volume 141, Issue 1, 1 August 1991, Pages 187-203.
[Gos96] D. Goss, Basic Structures of Function Field Arithmetic, Ergeb. Math. Grenzgeb. (3), vol. 35, Springer, Berlin,
1996.
[GNU15] Z. Guo, A. Narayanan and C. Umans, Algebraic problems equivalent to beating the 3/2 exponent for polynomial
factorization over finite fields, In preparation.
[vdH04] G. J. van der Heiden, Factoring polynomials over finite fields with Drinfeld modules, Math. Comp. 73 (2004),
317-322.
[vdH04-1] G. J. van der Heiden, Addendum to “Factoring polynomials over finite fields with Drinfeld modules”,Math.
Comp. 73(2004), Number 248, Page 2109
[Jou] Antoine Joux. A new index calculus algorithm with complexity L(1/4 + o(1)) in very small characteristic. Cryptology
ePrint Archive, Report 2013/095, 2013.
[KL94] E. Kaltofen and A. Lobo, Factoring high-degree polynomials by the black box Berlekamp algorithm, ISSAC ’94
Proceedings of the international symposium on Symbolic and algebraic computation Pages 90 - 98.
[KS98] E. Kaltofen and V. Shoup, Subquadratic-time factoring of polynomials over finite fields, Math. Comput.,
67(223):1179-1197, July 1998.
[KU08] K. Kedlaya and C. Umans, Fast modular composition in any characteristic, Proceedings of the 49th Annual IEEE
Symposium on Foundations of Computer Science (FOCS). pages 146-155. 2008.
[Len87] H. W. Lenstra Jr, Factoring integers with elliptic curves, Annals of Mathematics 126 (3): 649673. (1987).
[Nar14] A. K. Narayanan, Computation of Class Groups and Residue Class Rings of Function Fields over Finite Fields,
Ph.d Dissertation, Computer Science Department, University of Southern California 2014.
[PP89] A. Panchishkin and I. Potemine, An algorithm for the factorization of polynomials using elliptic modules, In Proceedings of the Conference, Constructive methods and algorithms in number theory, p. 117. Mathematical Institute of
AN BSSR, Minsk, 1989 (Russian).
[Pol74] J. M. Pollard, Theorems of factorization and primality testing, Proceedings of the Cambridge Philosophical Society
76 (3): 521-528. (1974).
[Sch95] R. Schoof: Counting Points on Elliptic Curves over Finite Fields. J. Theor. Nombres Bordeaux 7:219254, 1995.
[Sto01] A. Storjohann, Deterministic computation of the Frobenius form, In Proc. 42nd Annual Symp. Foundations of
Comp. Sci., pages 368-377, 2001.
[VS] G. D. Villa Salvador, Topics in the Theory of Algebraic Function Fields, 2006. Birkhauser, 2006.
[Yu95] J-K Yu, Isogenies of Drinfeld modules over finite fields, J. Number Th. 54 (1995), 161-171.
[Yun76] D. Y. Y. Yun, On square-free decomposition algorithms, Proc. 1976 ACM Symp. on Symbolic and Algebraic
Computation ISSAC 76.
Anand Kumar Narayanan, Department of Computing and Mathematical Sciences, California Institute of
Technology, Pasadena, California, USA, 91125
E-mail address: [email protected]
| 8cs.DS
|
On geodesic ray bundles in hyperbolic groups
Nicholas Touikan
arXiv:1706.01979v1 [math.GR] 6 Jun 2017
June 8, 2017
Abstract
We construct a Cayley graph CayS (Γ) of a hyperbolic group Γ such that
there are elements g, h ∈ Γ and a point γ ∈ ∂∞ Γ = ∂∞ CayS (Γ) such that
the sets RB (g, γ) and RB (h, γ) in CayS (Γ) of vertices along geodesic rays
from g, h to γ have infinite symmetric difference; thus answering a question of
Huang, Sabok and Shinko.
1
Introduction
To every infinite finite valence tree T we can associate a boundary at infinity
∂∞ T corresponding to ends of infinite rays. ∂∞ T is homeomorphic to a Cantor
set. A metric space is called δ-hyperbolic if, roughly speaking, up to an error
term δ it has a tree-like structure. Analogously to a tree, to a δ-hyperbolic
space X, one can assign a Gromov boundary ∂∞ X which is a compact, metrizable, yet oftentimes exotic set, corresponding to equivalence classes of ends of
infinite geodesic rays. A group Γ is called hyperbolic if one of its Cayley graphs
is is a δ-hyperbolic metric space for some δ ≥ 0. In this case to Γ we can assign
a canonical Gromov boundary ∂∞ Γ on which Γ acts non-trivially. The deep
connections between the properties of ∂∞ Γ and the group Γ makes it highly a
structured, and therefore fascinating, object to study.
In [HSS17] Huang, Sabok and Shinko investigate Borel equivalence relations on ∂∞ Γ. They show that if Γ is a hyperbolic group with the additional
property that Γ acts properly discontinuously and cocompactly on a CAT(0)
cube complex, i.e. Γ is cubulated, then the action of Γ on its boundary ∂∞ Γ
is hyperfinite. This generalizes a result of Dougherty, Jackson and Kechris
[DJK94, Corollary 8.2] from the class of free groups to the larger class of
cubulated hyperbolic groups.
Although the result of [HSS17] feels like it should be true for all hyperbolic groups, an additional cubulation requirement is needed to prove a key
lemma, [HSS17, Lemma 1.3], which states that for any two vertices x, y of a
δ-hyperbolic CAT(0) cube complex C and for any point γ ∈ ∂∞ C the sets,
called ray bundles, RB (x, γ) and RB (y, γ) of vertices of C that occur along
geodesic rays from x and y (respectively) to γ ∈ ∂∞ C have finite symmetric
difference.
The authors pose [HSS17, Question 1.4] which asks if [HSS17, Lemma 1.3]
holds for any Cayley graph of a hyperbolic groups. Not only would a positive
1
answer immediately imply that the action of any hyperbolic group Γ on ∂∞ Γ
is hyperfinite, but this is also a very natural question to ask from the point of
view of geometric group theory. This paper gives a negative answer by giving
examples of Cayley graphs CayS (Γ) of hyperbolic groups Γ with vertices x, y
and some γ ∈ ∂∞ CayS (Γ) such that the ray bundles RB (x, γ) and RB (y, γ)
have infinite symmetric difference. This example, if anything, reinforces the
relevance of [HSS17, Lemma 1.3].
The methods of this paper will be familiar to geometric group theorists,
but, since this paper is aimed at a broader audience, necessary background is
included to make it self-contained. That being said, the reader is expected to
know the following notions from topology: group presentations, fundamental
groups, the Seifert-van Kampen theorem, and universal covering spaces.
1.1
Acknowledgements
I first wish to thank Michael Hull and Jindrich Zapletal for the invitation to
the South Eastern Logic Symposium 2017, which greatly increased my appreciation of the contemporary work of descriptive set theorists. I also wish to
thank Marcin Sabok for posing this question about symmetric differences of
ray bundles, specifically about the embedability of bad ladders into Cayley
graphs, and for an interesting discussion, encouragement and feedback. Finally I am grateful to Bob Gilman and Paul Schupp for conversations that
confirmed that the main result of this paper is probably not a trivial consequence of what is known about the automaticity of the language of geodesics
in hyperbolic groups.
2
Hyperbolic groups and their boundary
The author recommends [Aea91] for an accessible yet thorough treatment of
the topics in this section. Given a group Γ and a generating set S of Γ we can
construct a Cayley graph CayS (Γ) which is a directed graph whose vertices
are the elements of Γ and for each g ∈ Γ and s ∈ S we draw the edge
g
s
I
gs
.
By declaring each edge to be an isometric copy of the closed unit interval,
we make graphs into connected metric spaces via the path metric. If X is a
graph we say that a path starting at a vertex v and ending at a vertex u is
geodesic if it is the shortest possible path between u, v. Typically there will be
multiple geodesics between a pair of vertices. A metric space is δ-hyperbolic if
it has the following property: for any three vertices u, v, w if α, β, and γ are
geodesics from u to v, v to w, and w to u respectively then α is contained in a
δ-neighbourhood of β ∪ γ. If a group Γ has a δ-hyperbolic Cayley graph with
respect to one finite generating set, then for any other finite generating set the
corresponding Cayley graph will also be δ 0 -hyperbolic, though possibly with
δ 0 6= δ. Such a group will therefore be called a hyperbolic group.
For example, if A is a finite set of symbols and F(A) is the free group on
A, then, taking A as a generating set of F(A), the Cayley graph CayA (F(A))
2
is a regular tree with valence |A| and in particular for any geodesics α, β, and
γ as above, α ⊂ β ∪ γ so that CayA (F(A)) is in fact 0-hyperbolic.
Let us now give a precise definition of the Gromov boundary ∂∞ Γ. Let S
be a finite generating set of Γ. A geodesic ray is a continuous map
ρ : [0, ∞) → CayS (Γ)
such that for every pair of positive integers m < n, ρ(m) is a vertex and the
segment ρ([n, m]) is a geodesic. ∂∞ Γ is the set of geodesic rays of CayS (Γ)
modulo the relation: ρ ∼ ρ0 ⇔ there is some R ≥ 0 such that ρ ([0, ∞)) is
contained in an R−neighbourhood of ρ0 ([0, ∞)) and ρ0 ([0, ∞)) is contained in
an R−neighbourhood of ρ ([0, ∞)).
We recommend the following exercises:
• If Γ = F(A) is a free group as above, then ∂Γ is naturally identified with
a Cantor set.
• If Γ = hai ⊕ hbi, the free abelian group of rank two (which is not a
hyperbolic group), then ∂∞ CayA (Γ) can be identified with the circle at
infinity for R2 , but the action of Γ (induced by translating rays) yields a
trivial action on ∂∞ Γ.
That ∂∞ Γ, thus given, is well-defined, non-trivial, canonical for Γ, and
admits a non-trivial Γ action is a consequence of δ-hyperbolicity. The reader
may consult [GdlH90, §6-§8] or [KB02] for a complete treatment of the topic.
3
The bad ladder
Consider the infinite graph L consisting of two sides, copies of R, with a vertex
at each integer point, and countably many rungs, edges connecting vertices at
corresponding integral vertices on each side. Add a vertex to the middle of
each rung. The resulting graph L is shown in Figure 1.
Figure 1: A ladder with a vertex x on a side and a vertex y in the middle of a rung.
We note that any two geodesic rays either go to the left or to the right,
and if they go in the same direction, they remain at a bounded distance. It
follows that ∂∞ L consists of two points.
Proposition 3.1. Let x be a vertex on a side of L, let y be a vertex in the
middle of a rung and let γ ∈ ∂∞ L = {±∞} correspond to one of the ends
of the ladder. Then the sets RB (x, γ) and RB (y, γ) have infinite symmetric
difference.
3
Proof. Without loss of generality we may assume that γ corresponds to +∞.
As any geodesic ρ travels towards γ it must eventually stay within one of the
sides of L. If ρ originates at x then it is allowed to travel once through a rung
to reach the other side. It follows that every vertex on a rung that is “greater”
than x is in RB (x, γ). If ρ originates at y in the middle of a rung, then once
it enters a side s1 it is no longer able to switch because if that happens then
there is some initial segment ρ0 of ρ whose length does not realize the distance
between y and the first point it encounters in s2 6= s1 . See Figure 1. It follows
that RB (y, γ) doesn’t contain any vertices contained in rungs; thus the two
sets have infinite symmetric difference.
Although L is a hyperbolic graph, due to its nonhomogeniety, it cannot be
the Cayley graph of a group. We will now construct the Cayley graph of a
group, in fact a free group, which contains a ladder L as a convex subgraph.
That is to say any geodesic connecting two points on the ladder inside this
larger graphs must stay within the ladder. To show this we must reach a sufficiently complete understanding of the geometry of a Cayley graph. Although
it is not invoked explicitly, the proof is informed by the Bass-Serre theory of
groups acting on trees and corresponding decompositions into graphs of spaces
[SW79, Ser03].
4
Embedding bad ladders into Cayley graphs
We will take some liberties with notation and identify group presentations with
the groups they present. First consider the presentation
Γ0 = p, q, t | t−1 ptq −1 ≈ F2 .
For any group presentation, there is a standard construction known as a presentation complex, which is a CW-complex P (Γ0 ) obtained by gluing polygons (corresponding to relations) to graphs (edges correspond to generators)
in such a way (as a consequence of the Seifert-van Kampen Theorem) that
π1 (P (Γ0 )) ≈ Γ0 .
In this case presentation complex P (Γ0 ) consists of a graph with one vertex,
three directed edges labelled p, q, t, and a square along whose boundary the
word t−1 ptq −1 can be read. This word specifies the identifying map between
the boundary of the square and a closed loop in the graph, making the latter
nullhomotopic. As a topological space P (Γ0 ) can also be obtained by taking
a cylinder A = [−1, 1] × S 1 , picking a point on each boundary component and
identifying them. This is shown on Figure 2.
Remark 4.1. The 1-skeleton of the universal cover P^
(Γ0 ) corresponds to the
Cayley graph Cay{p,q,t} (F2 ), i.e. the Cayley graph relative to the generating
set explicitly given by the group presentation. This is true for any presentation
complex.
The universal cover P^
(Γ0 ) is a tree of spaces obtained by taking an infinite
collection of copies of strips corresponding to connected components of the lift
A ⊂ P (Γ0 ) in P^
(Γ0 ) attached by points. We call these pq-strips. This is
4
Figure 2: The presentation complex P (Γ0 )
shown if Figure 3. There is also a collection of bi-infinite lines in P^
(Γ0 ) along
which we read . . . ttt . . ., we call these t-lines. Consider now the amalgamated
Figure 3: The universal cover of P (Γ0 ). Decorated edges are labelled t. pq-strips
are shaded grey.
free product:
Γ1 = Γ0 ∗t=s2 hsi = p, q, t, s | t−1 ptq −1 , s2 t−1 ≈ F2
corresponding to adjoining a square root s to the basis element t ∈ F2 . By the
Seifert-van Kampen Theorem, it can be realized as the fundamental group of
a space P1 , which is not a presentation complex, obtained by taking a copy
of P (Γ0 ), a circle C = S 1 , and attaching another cylinder D = [−1, 1] × S 1
so that the attaching map {−1} × S 1 → P (Γ0 ) wraps with degree 1 around
the loop corresponding to the edge with label t and the other attaching map
{1} × S 1 → C wraps with degree 2. See Figure 4.
5
f1 . Below, how the CW-complex
Figure 4: On top, a portion of the universal cover P
P1 is obtained from P0 .
f1 , P (Γ0 ) ⊂ P1 lifts to a countable collecIn the universal covering space P
tion of disjoint copies of P^
(Γ0 ) called Γ0 -pieces and the circle C ⊂ P1 lifts to a
countable collection of disjoint lines called C-lines. The connected components
of lifts of the cylinder D are called D−strips, copies of [−1, 1] × R connecting
t-lines in Γ0 -pieces to C-lines. In particular each C-line is attached to two
D-strips. Globally, the universal cover has the structure of a tree of spaces.
See Figure 5.
f1 depicted as a tree of spaces obtained by attaching Γ0 -pieces
Figure 5: A portion of P
to D-strips (shown in grey) along t-lines. Note that although drawn as “pancakes”
the Γ0 -pieces are actually copies of the space shown in Figure 3.
Our final presentation Γ is obtained via the following Tietze transformation:
Γ1 = p, q, t, s | t−1 ptq −1 , s2 t−1 ≈ p, q, s | s−2 ps2 q = Γ.
This Tietze transformation corresponds to the fact that, since s2 = t, we can
remove t from the generating set. Geometrically the resulting presentation
complex is obtained by collapsing the [−1, 1] factor in the cylinder D = S 1 ×
[−1, 1] ⊂ P1 to a point. See Figure 6. The universal cover of the presentation
^
f1 , subdividing each
complex P
(Γ) can be obtained by taking the Γ0 -pieces in P
t-labelled edge into a length 2 edge path labelled ss, replacing t−lines with
6
Figure 6: Collapsing D-strips (shaded grey) onto lines as seen from the universal
cover, and the resulting pq-ladders, contained in a Γ0 -piece.
s-lines, and then identifying two s-lines in different Γ0 -pieces if they were both
f1 has a large scale
connected by D-strips to the same C-line. In this way P
tree of spaces structure obtained taking resulting Γ0 -pieces and attaching them
along s−lines. Furthermore we observe that each Γ0 -piece contains a ladder L
obtained by gluing together squares labelled s−2 ps2 q along segments labelled
s2 . We call such a ladder a pq-ladder. See Figure 6.
^
In this way P
(Γ) admits a depth 2 hierarchical decomposition as a tree of
spaces. At the top level we have Γ0 -pieces connected along s-lines as a tree
of spaces, then the Γ0 -pieces themselves are trees of pq-ladders, connected by
vertices.
^
Proposition 4.2. A pq-ladder L is convex in the 1-skeleton of P
(Γ). Furthermore any geodesic ray starting in L and going to one of the ends of L must
stay in L.
Proof. Let p be the Γ0 -piece containing a pq-ladder L. Let u, v ∈ L be vertices
and let ρ be a geodesic connecting u and v.
Claim 1: ρ cannot exit p. Suppose towards the contrary that this was
the case then, by the tree of spaces structure, ρ must exit p at some point a
contained in some s-line s, and then re-enter p by at some other point b ∈ s
in the same s-line. It follows that if ρ is geodesic it cannot exit p because
the subsegment ρ([na , nb ]), where ρ(na ) = a, ρ(nb ) = b, can be replaced the
strictly shorter segment from a to b contained within s.
Claim 2: If ρ stays in the Γ0 -piece p, it cannot exit L. Indeed each piece
consists of a tree of pq-ladders connected by points; since it is the same space
as the one shown in Figure 3 except with each edge labelled t replaced by a
path of length 2 labelled ss. If ρ leaves L at some vertex p, then to re-enter L
it must pass through p again, contradicting that it is geodesic.
The convexity of L now follows. This implies that any infinite path that
stays in the p or q side of L is a geodesic ray. It remains to show that any
geodesic ray starting at x ∈ L going to γ ∈ ∂∞ L stays in L. Let ρ be one such
^
geodesic ray and let β : [0, ∞) → P
(Γ) be another arc-length parameterized
7
geodesic ray from x to γ. Suppose that β exits L at the point β(N ).
By convexity of L, β cannot re-enter L, but it could still travel close to
it. By definition of the Gromov boundary there must be some bound R such
that for all z, d(β(z), ρ) ≤ R. However, since β is geodesic and arc-length
parameterized, d(β(N + M ), β(N )) = M and since the shortest path from
β(N + M ) to L must pass through β(N ) we conclude that
d(β(N + M ), ρ) ≥ d(β(N + M ), L) = M.
Since β is an infinite ray we may take M > R which yields a contradiction.
Proposition 3.1 and 4.2 immediately imply the main result:
Corollary 4.3. Let Γ = p, q, s | s−2 ps2 q ≈ F2 and let X = Cay{p,q,s} (Γ).
If γ ∈ ∂∞ X = ∂∞ Γ corresponds to an end of a pq-ladder L, x is a vertex
contained in a side of L, and y is a vertex contained in a rung of L, then
RB (x, γ) and RB (y, γ) have infinite symmetric difference.
4.1
A one-ended example
The example we just gave is somewhat unsatisfying since it is a free group. We
will outline another construction, which was the original example found by the
author. This group is not free since it is one-ended, which in the torsion-free
case means it does not decompose as a non-trivial free product. Consider first
the presentation
Σ0 = a, b, p, q, t | abpa−1 b−1 q, ptqt−1 .
This is an explicit decomposition of Σ0 as an HNN extension of a free group of
rank 3 and the presentation complex is homeomorphic to an orientable closed
surface of genus 2. We then repeat the construction in the previous section
Σ1 = a, b, p, q, t | abpa−1 b−1 q, ptqt−1 ∗t=s2 s2
≈ a, b, p, q, s | abpa−1 b−1 q, ps2 qs−2 = Σ
to embed a bad ladder into a the Cayley graph corresponding to the presentation Σ. Since Σ0 is a closed surface group, therefore one-ended, and hsi cannot
act with an infinite orbit on a tree if s2 fixes a point, [Tou15, Theorem 3.1]
implies that Σ is one-ended. In particular Σ is not free. Hyperbolicity follows
from the combination theorems [BF92, BF96, KM98].
Again the universal cover is a tree of spaces obtained by gluing hyperbolic
planes H2 along s-lines and the proof goes similarly to Proposition 4.2. The
first claim goes through as is, we leave the proof of Claim 2 (convexity of
pq-ladders) as an exercise in small cancellation theory (one can use [MW02,
Theorem 9.4].)
4.2
Cubulating bad ladders
A bad ladder consists of a chain of hexagons glued along edges. As an illustration of [HSS17, Lemma 1.3], observe that if we cubulate a bad ladder, i.e.
8
Figure 7: A cubulated bad ladder
make it into a cube complex, (see Figure 7) the conclusion of Proposition 3.1
no longer holds. In fact both groups shown in this paper can be cubulated
and therefore do not give counterexamples to the conjecture that the action of
every hyperbolic group Γ on ∂∞ Γ is hyperfinite, a conjecture that this author
believes to be true.
References
[Aea91]
J. M. Alonso and et al. Notes on word hyperbolic groups. In Group
theory from a geometrical viewpoint (Trieste, 1990), pages 3–63.
World Sci. Publ., River Edge, NJ, 1991. Edited by H. Short.
[BF92]
M. Bestvina and M. Feighn. A combination theorem for negatively
curved groups. J. Differential Geom., 35(1):85–101, 1992.
[BF96]
Mladen Bestvina and Mark Feighn. Addendum and correction to:
“A combination theorem for negatively curved groups” [J. Differential Geom. 35 (1992), no. 1, 85–101; MR1152226 (93d:53053)]. J.
Differential Geom., 43(4):783–788, 1996.
[DJK94] Randall Dougherty, Steve Jackson, and Alexander S Kechris. The
structure of hyperfinite borel equivalence relations. Transactions of
the American Mathematical Society, 341(1):193–225, 1994.
[GdlH90] Étienne Ghys and Pierre de la Harpe. Sur les groupes hyperboliques
d’apres Mikhael Gromov, volume 83 of Progress in Mathematics.
Birkhäuser Boston, Inc., Boston, MA, 1990.
[HSS17]
Jingyin Huang, Marcin Sabok, and Forte Shinko. Hyperfiniteness
of boundary actions of cubulated hyperbolic groups. arXiv preprint
arXiv:1701.03969, 2017.
[KB02]
Ilya Kapovich and Nadia Benakli. Boundaries of hyperbolic groups.
Combinatorial and geometric group theory (New York, 2000/Hoboken, NJ, 2001), 296:39–93, 2002.
[KM98]
O. Kharlampovich and A. Myasnikov. Hyperbolic groups and free
constructions. Trans. Amer. Math. Soc., 350(2):571–613, 1998.
[MW02] Jonathan P. McCammond and Daniel T. Wise. Fans and ladders in
small cancellation theory. Proc. London Math. Soc. (3), 84(3):599–
644, 2002.
[Ser03]
Jean-Pierre Serre. Trees. Springer Monographs in Mathematics.
Springer-Verlag, Berlin, 2003. Translated from the French original
9
by John Stillwell, Corrected 2nd printing of the 1980 English translation.
[SW79]
Peter Scott and Terry Wall. Topological methods in group theory.
In Homological group theory (Proc. Sympos., Durham, 1977), volume 36 of London Math. Soc. Lecture Note Ser., pages 137–203.
Cambridge Univ. Press, Cambridge, 1979.
[Tou15]
Nicholas Touikan. On the one-endedness of graphs of groups. Pacific
Journal of Mathematics, 278(2):463–478, 2015.
10
| 4math.GR
|
arXiv:1302.4766v1 [math.AC] 19 Feb 2013
ELASTICITY IN POLYNOMIAL-TYPE EXTENSIONS
MARK BATELL AND JIM COYKENDALL
Abstract. The elasticity of an atomic integral domain is, in some sense, a
measure of how far the domain is from being a unique factorization domain (or,
more properly, a half-factorial domain). We consider the relationship between
the elasticity of a domain, R, and the elasticity of its polynomial ring R[x]. For
example, if R has at least one atom, a sufficient condition for the polynomial
ring R[x] to have elasticity 1 is that every nonconstant irreducible polynomial
f ∈ R[x] be irreducible in K[x]. We will determine the integral domains R
whose polynomial rings satisfy this condition.
1. Introduction and Motivation
In this paper, R will always be an integral domain with quotient field K. The
notation Irr(R) will stand for the irreducible elements of R, A(R) will be the elements of R that can be expressed as a product of atoms, U (R) and Cl(R) will
respectively denote the unit group and the class group of R. For a nonzero nonunit
element x ∈ A(R), we define the elasticity of x to be
n
|x = π1 π2 · · · πn = ξ1 ξ2 · · · ξm }
m
where each πi , ξj is an irreducible element of R. For example, if x is a product of
primes or is an element of a half-factorial domain, then ρ(x) = 1. If x ∈
/ A(R) then
ρ(x) is undefined.
For an integral domain R, the elasticity is defined by
ρ(x) = sup{
ρ(R) = sup{ρ(x)|x ∈ A(R)}.
As previously, we say that the elasticity of a domain without any atoms is undefined. It is well-known that if R is atomic, then ρ(R) = 1 if and only if R is an HFD,
but in the nonatomic case, the situation can be more exotic. For example in [CZ] a
domain was constructed with a unique (up to associates) irreducible element. Such
a domain, R, is necessarily nonatomic, but ρ(R) = 1.
More generally in [CM] it is shown that any atomic monoid can be realized as
the “atomic part” of an integral domain (again, usually non-atomic). Hence, one
can construct nonatomic domains that display any prescribed elasticity.
Given an integral domain, R, it is natural to ask what is the relationship between ρ(R) and ρ(R[x]). Since any factorization of a constant in R[x] must be a
factorization in R, it is easy to see that, in general, ρ(R[x]) ≥ ρ(R). More generally,
one can ask if we have the sequence of integral domains
R = R0 ⊆ R1 ⊆ R2 ⊆ · · · ⊆ K
2000 Mathematics Subject Classification. Primary: 13F15.
Key words and phrases. Factorization, unique factorization.
1
2
MARK BATELL AND JIM COYKENDALL
what is the relationship between ρ(R0 + R1 x + R2 x2 + · · · ) and the collection of
data ρ(Ri )? Some special cases of this general construction worth noting are the
polynomial ring (Ri = R for all i ≥ 0), the construction R + xK[x] (R0 = R and
Ri = K for all i ≥ 1), and R + Rx + x2 K[x] (R0 = R1 = R and Ri = K for all
i ≥ 2).
2. Some Polynomial-Type Constructions
We begin this section with some preliminary lemmata, but first we recall the
notion of an AP-domain.
Definition 2.1. We say that the integral domain, R, is an AP-domain if every
irreducible (atom) in R is prime.
Lemma 2.2. Let R be an AP-domain with at least one irreducible element, then
ρ(R) = 1.
Proof. Of course, if R is an AP-domain vacuously (that is, in the case that R has
no atoms), then ρ(R) is undefined. Suppose, on the other hand, that Irr(R) is
nonempty. Since all atoms in an AP-domain, are prime, any irreducible factorization is a prime factorization and therefore is unique. Hence ρ(x) = 1 for all
x ∈ A(R) and so ρ(R) = 1.
Lemma 2.3. Let R be a domain, p ∈ R be a nonzero prime element, and a ∈ A(R).
Then ρ(a) ≥ ρ(ap).
Proof. Suppose we have the following irreducible factorization of ap:
ap = ξ1 ξ2 · · · ξn ,
where each ξi ∈ Irr(R) for all 1 ≤ i ≤ n. Since p ∈ R is prime, p must be associated
with one of the ξi ; we will say, without loss of generality, that ξn = up for some
u ∈ U (R). Since R is an integral domain, we cancel the factor of p to obtain
a = ξ1 ξ2 · · · ξn−1 u.
The upshot is that an arbitrary irreducible factorization of ap is (up to associates)
p times an irreducible factorization of a. Hence there is a factorization of ap of
length m + 1, if and only if there is a corresponding factorization of a of length m.
m+1
Since m ≥ k ≥ 1 implies that m
k ≥ k+1 , we have that ρ(a) ≥ ρ(ap).
Remark 2.4. To tie up a loose end, we note √that the inequality in the previous
result can be strict. For example, in the ring Z[ −14] the element 81 has precisely
two irreducible factorizations (up to associates and reordering):
√
√
81 = (3)(3)(3)(3) = (5 + 2 −14)(5 − 2 −14).
√
So ρ(81) = 2. Now consider the element 81 as an element of Z[ −14][x]. As
before 81 has only two irreducible factorizations (the ones mentioned previously)
and x is a prime element. An easy check shows that ρ(81x) = 35 < 2.
We now present the following theorem.
ELASTICITY IN POLYNOMIAL-TYPE EXTENSIONS
3
Theorem 2.5. Let R be an integral domain with quotient field K. If R contains
at least one atom, then ρ(R) = ρ(R + xK[x]). If R has no atoms (that is, R is an
antimatter domain) then ρ(R) is undefined and ρ(R + xK[x]) = 1.
Proof. Let g(x) be a nonconstant polynomial in R + xK[x]. We claim that if g(x)
is irreducible, then g(x) is (up to associates) either x or of the form 1 + xf (x)
where 1 + xf (x) ∈ Irr(K[x]). To see this, note that if g(x) is nonconstant, then
g(x) = r + xk(x) with k(x) ∈ K[x] \ {0}. If r = 0 then the stipulation that g(x) is
irreducible forces the condition k(x) ∈ U (R). On the other hand, if r 6= 0, then the
factorization g(x) = r(1 + r1 xk(x)) shows that if g(x) is irreducible, then r ∈ U (R)
and 1 + r1 xk(x) ∈ Irr(K[x]). This establishes the claim.
We also note that the elements x and 1 + xf (x) ∈ Irr(K[x]) are, in fact, prime
elements of R+xK[x]. The fact that x is prime is straightforward. For an irreducible
of the form 1 + xf (x), note that if 1 + xf (x) divides the product h(x)k(x) (with
h(x), k(x) ∈ R + xK[x]) then without loss of generality, 1 + xf (x) divides h(x) in
K[x]. We say that h(x) = (1 + xf (x))q(x), and comparing constant terms, we see
that q(x) ∈ R + xK[x]. Hence x and irreducibles of the form 1 + xf (x) are prime
in R + xK[x].
From the previous observations, we see that if R is an antimatter domain, then
R + xK[x] is an AP-domain (with Irr(R + xK[x]) nonempty) and hence ρ(R +
xK[x]) = 1 by Lemma 2.2.
Now suppose that R has at least one irreducible element. Since any element of
R, factored as an element of R[x], has only factors from R (and any irreducible in
R remains irreducible in R[x]), we have that ρ(R + xK[x]) ≥ ρ(R). On the other
hand, let f (x) ∈ A(R + xK[x]). We factor f (x) into irreducibles as follows:
f (x) = π1 π2 · · · πm g1 (x)g2 (x) · · · gk (x)
with πi ∈ Irr(R) and gi (x) ∈ Irr(R + xK[x]) of degree at least 1. By our previous
remarks, each gi (x) is prime. Hence by Lemma 2.3, ρ(f (x)) ≤ ρ(π1 π2 · · · πm ).
Hence ρ(R + xK[x]) ≤ ρ(R) and so, we have equality.
In stark contrast, the next result shows that a minor tweaking of the previous
construction can yield a domain with infinite elasticity. This also gives a strong
indication of the level of difficulty of determining the elasticity of R0 +R1 x+R2 x2 +
· · · in terms of the elasticities ρ(Ri ).
Proposition 2.6. Let R be a domain that contains at least one atom, then
ρ(R + Rx + x2 K[x]) = ∞.
Proof. Let π ∈ Irr(R). For all n ∈ N0 the polynomial (π n ± x) ∈ Irr(R + Rx +
x2 K[x]). The irreducible factorizations
(π n + x)(π n − x) = π 2n (1 −
1 2
x )
π 2n
have lengths 2 and 2n + 1 respectively. Hence we see that ρ(R + Rx + x2 K[x]) =
∞.
4
MARK BATELL AND JIM COYKENDALL
We now specialize to the case R[x]. In comparing the elasticities ρ(R) and
ρ(R[x]), there are two dynamics to consider. The first is the factorization of constants (which is reflected in ρ(R)) and the different factorizations that may result
from the polynomial structure. To illustrate we consider the following examples.
√
Example 2.7. It is well-known (see for example [C1]) that √
Z[ −3] is a halffactorial domain (and hence has elasticity 1). The domain Z[ −3][x] is not an
HFD. The irreducible factorizations
√
√
(2x + (1 + −3))(2x + (1 − −3)) = (2)(2)(x2 + x + 1)
demonstrates that the elasticity of the polynomial extension exceeds 1.
A close look at√the mechanics of the previous example reveals that the failure
of the domain Z[ −3] to be integrally closed allowed the creation of this “bad
factorization.” In the proof of the main theorem in [C2] it is shown that if R is not
integrally closed, a simliar effect occurs.
It is known (see [Za]) that if R is a Krull domain, then R[x] is an HFD if and
only if |Cl(R)| ≤ 2. It is also known from [Ca] that if R is a ring of algebraic
integers (and hence, certainly a Krull domain), then R is an HFD if and only if
|Cl(R)| ≤ 2. Hence if R is a ring of algebraic integers with |Cl(R)| ≤ 2, then R
is an HFD and so is R[x]. In this case ρ(R) = ρ(R[x]), but the equality can be
delicate as we will demonstrate in the following example. The following example
should be constrasted with the previous as it is integrally closed.
√
Example 2.8. The integral domain R := Z[ −5] is a ring of integers of class
number precisely 2 (see, for example, the tables in [Co]) and hence is an HFD that
does not have unique factorization. But although ρ(R[x])
√ = 1, the factorizations
√
can be exotic. The elements 2x2 + 2x + 3, 2, 2x + 1 + −5, and 2x + 1 − −5 are
all elements of Irr(R[x]). Consider the factorizations
√
√
(2)(2x2 + 2x + 3) = (2x + 1 + −5)(2x + 1 − −5).
The upshot is that even in this relatively “nice” domain, the factorizations of
elements can depend on how the polynomials break down (with respect to degree) in
a nontrivial way.
It is well-known that if R is a UFD with quotient field K, then any irreducible
polynomial over R[x] remains irreducible over K[x]. More generally, domains, R,
for which every irreducible polynomial of degree at least one remains irreducible
in K[x] would seem to be the basic case to solve. For these domains, it would
seem likely that there is a more direct correlation between ρ(R) and ρ(R[x]), since
there must be a one to one correspondence between irreducible factors of degree at
least 1 for any two irreducible factorizations of the same element. Certainly, bad
factorizations of the ilk of the previous two examples are precluded (although we
feel obligated to point out again that the second example is an HFD).
Although it may seem reasonable to consider domains where irreducibles of degree at least one in R[x] remain irreducible in K[x], it is not obvious that in this case
ρ(R) = ρ(R[x]). To illustrate the problem, consider the irreducible factorizations
π1 π2 · · · πk f1 (x)f2 (x) · · · fm (x) = ξ1 ξ2 · · · ξt g1 (x)g2 (x) · · · gn (x)
with each πi , ξj ∈ Irr(R) and fi (x), gj (x) all irreducible of degree at least one.
ELASTICITY IN POLYNOMIAL-TYPE EXTENSIONS
5
Even if we have that m = n and each fi (x) and gj (x) pair off (up to units in
K), there is no guarantee that the ratio of k and t are within the elasticity bounds
of R (precisely because there is ambiguity up to units in K).
That being said, we present the following theorem. The rest of the paper will
be devoted to establishing this result. The theorem will follow quickly from our
classification of domains with this property.
Theorem 2.9. Let R be a domain such that every irreducible of R[x] of degree
greater than or equal to 1 is irreducible in K[x]. Then if ρ(R) is defined, then
ρ(R) = ρ(R[x]).
As noted before, these conditions are not necessary as Example 2.8 shows.
3. Irreducibles of R[x] versus irreducibles of K[x]
In this section, we use the following facts and definitions many times without
further mention.
a) If F is a nonzero fractional ideal of R, then Fv = (F −1 )−1 .
b) The elements a1 , a2 , · · · , an ∈ R have a greatest common divisor provided
that (a1 , a2 , · · · , an )v is principal.
c) If f ∈ K[x], the ideal generated by the coefficients of f is denoted Af .
d) The greatest common divisor of a1 , a2 , · · · , an ∈ R will be denoted [a1 , a2 , · · · , an ]
if it exists.
Our goal in this last section is to characterize those domains R having the following property:
(P) every nonconstant irreducible polynomial f ∈ R[x] is irreducible in K[x]
The techniques used in the proof we shall give strongly resemble those used in
the classical proof that every polynomial ring R[x] over a UFD is again a UFD.
The proof of this classical result boils down to showing that any UFD, R, satisfies
property (P) above. Gauss’s Lemma, which states that the product of two primitive
polynomials is primitive, is the key which allows this proof to go through in the
UFD case.
Conditions under which the product of two primitive polynomials remains primitive has been studied in more general domains (see for instance, [T]). It turns
out that the domains satisfying property (P) must satisfy a condition somewhat
stronger than Gauss’s Lemma; they must satisfy what is called the PSP-property.
Definition 3.1. A domain R has the PSP-property if whenever a0 +a1 x+· · ·+an xn
is a primitive polynomial over R and z ∈ (a0 , a1 , · · · , an )−1 , then z ∈ R.
Polynomials a0 + a1 x + · · · + an xn satisfying the above definition are called
superprimitive. Thus a domain has the PSP-property if every primitive polynomial
is superprimitive.
For integral domains, the following implications are well-known
UFD =⇒ GCD =⇒ PSP-property =⇒ GL-property =⇒ AP-property.
and in [AQ] it is shown that all of these types are equivalent for atomic domains.
Arnold and Sheldon [AS, Example 2.5] gave an example of a domain satisfying
Gauss’s Lemma (such domains are said to have the GL-property), but failing to
have the PSP-property. The domain they considered was the domain F [{xα : α >
6
MARK BATELL AND JIM COYKENDALL
0}, {y α : α > 0}, {z αxβ : α, β > 0}, {z αy β : α, β > 0}]. Here, all exponents α
and β are understood to be taken from the field Q of rational numbers. This is an
example of a monoid domain, and can intuitively be thought of as the ring of all
formal polynomials in the given ”indeterminates” with coefficients in F , the field
of two elements. We note that yt + x is a primitive polynomial in t that fails to be
superprimitive, as z ∈ (x, y)−1 . This leads us to the following theorem.
Proposition 3.2. Assume every nonconstant irreducible f ∈ R[x] is irreducible in
K[x]. Then R is integrally closed and has the PSP-property.
Proof. Assume R is not integrally closed. Choose an element ω ∈ K that satisfies
a monic irreducible polynomial f ∈ R[x] of degree ≥ 2. Since ω is a root of f , the
division algorithm in K[x] implies that f = (x − ω)g, where g is a polynomial in
K[x] of degree ≥ 1. Hence f is irreducible over R but reducible over K. This is a
contradiction.
Next, we assume R does not have the PSP-property. Let y0 + y1 x + · · · + yn xn
be a primitive polynomial and let z ∈ K be such that z ∈ (y0 , y1 , · · · , yn )−1 but
z∈
/ R. In the collection of all primitive polynomials that are not superprimitive, we
assume that I := y0 + y1 x + · · · + yn xn is one of minimal degree. In K[x] we have
the factorization yn xn+1 + (yn−1 + zyn)xn + · · · + (y0 + zy1 )x + zy0 = (x + z)(yn xn +
yn−1 xn−1 + · · · + y1 x + y0 ) where the polynomial f on the left hand side belongs to
R[x]. We claim that f is irreducible over R. If f = gh for some g, h ∈ R[x] then x+z
divides g or h in K[x], say g = (x + z)p(x). Since R is integrally closed, p(x) ∈ R[x]
(see [G, Theorem 10.4]), say p(x) = ak xk + ak−1 xk−1 + · · · + a1 x + a0 . Then g(x) =
ak xk+1 +(ak−1 +zak )xk +· · ·+(a0 +za1 )x+za0 , so that za0 , za1 , · · · , zak ∈ R, and
hence p(x) is not superprimitive. But p(x) is a factor of the primitive polynomial
I. Hence p(x) is primitive, so that the minimality assumption on I implies that
k = n. It follows that h is a unit so that f is irreducible over R, but not over K,
the desired contradiction.
Thus in our search for the domains satisfying property (P), we may restrict our
attention to integrally closed domains having the PSP-property.
To show that a particular domain actually has property (P), one would probably
employ the following strategy: Suppose f ∈ R[x] is a nonconstant polynomial that
is irreducible over R, but fails to be irreducible over the field of fractions K, say
f = gh in K[x]. Now “clear the denominators,” that is, choose nonzero b, d ∈ R
such that bdf = (bg)(dh) and bg, dh ∈ R[x]. At this point in the proof, we would
probably need to find some way to cancel out b and d to get a contradiction, namely,
that f = g ′ h′ for some g ′ , h′ ∈ R[x]. It turns out that if R has the PSP-property,
then we can assert, after clearing denominators, that the greatest common divisor
of the coefficients of bdf exists and is equal to bd, as shown by the following.
Proposition 3.3. Let R be a domain. The following are equivalent.
a) R has the PSP-property.
b) Whenever the elements a1 , a2 , · · · , an ∈ R have a greatest common divisor
and 0 6= b ∈ R, then [ba1 , ba2 , · · · , ban ] = b[a1 , a2 , · · · , an ].
Proof. Assume R has the PSP-property and [a1 , a2 , · · · , an ] = g. Given b ∈ R, it is
clear that bg is a common divisor of ba1 , ba2 , · · · , ban . If x is some other common
a1 a2
an −1
. This implies bg
divisor, then bg
x ∈ ( g , g ,··· , g )
x ∈ R since R is PSP. In other
words, x|bg so that [ba1 , ba2 , · · · , ban ] = bg
ELASTICITY IN POLYNOMIAL-TYPE EXTENSIONS
7
Conversely, assume b) holds and rs ∈ (a1 , a2 , · · · , an )−1 , where a0 + a1 x + · · · +
an xn is some primitive polynomial over R. Then s|rai for all i, so b) implies that
s|r. Thus rs ∈ R, so that R is PSP.
Proposition 3.4. Let R be an integrally closed PSP-domain. The following are
equivalent.
a) Every nonconstant irreducible polynomial f ∈ R[x] is irreducible in K[x]
b) Whenever f = gh in R[x] and the greatest common divisor of the coefficients
of f exists, then the greastest common divisor of the coefficients of g exists
Proof. b) =⇒ a). Assume b) holds. Let f ∈ R[x] be a nonconstant irreducible
polynomial (hence the greatest common divisor of the coefficients is 1). Suppose
f = gh, where g, h ∈ K[x] have degrees > 1. Choose nonzero b, d ∈ R such that
bg, dh ∈ R[x]. Then we have the equation bdf = (bg)(dh) in R[x], and since R is
PSP, Proposition 3.3 implies that the greatest common divisor of the coefficients of
bdf exists and is equal to bd. Hence the greatest common divisor of the coefficients
of bg exists, say u, and the greatest common divisor of the coefficients of dh exists,
bd
say v. Note that uv divides the coefficients of bdf . Hence uv
f = g1 h1 , where g1 , h1
bd
bd
is a unit. But
are primitive. Since R has the GL-property, uv f is primitive, so uv
then f is reducible over R, a contradiction.
a) =⇒ b). Assume a) holds. Suppose f = gh in R[x] and the greatest common
divisor of the coefficients of f exists, say s. We can assume deg f > 1. Then f = sf ′ ,
where f ′ is primitive. Since f ′ is primitive, f ′ is the product of irreducibles, say
f ′ = f1 f2 · · · fk . By unique factorization in K[x], g = uf1 f2 · · · fr for some r 6 k
(without loss of generality) and some u ∈ K. Since R is PSP, u ∈ R and the
greatest common divisor of the coefficients of g equals u.
In the paper [AS], Arnold and Sheldon proved the following theorem.
Theorem 3.5. Let R be a domain with quotient field K. The following are equivalent.
1) R[x] is an AP-Domain
2) R[x] is a GL-Domain
3) Each of the following holds:
(α) R has the PSP-property
(β) R is integrally closed, and
(γ) Whenever B, C are finitely generated fractional ideals of R such that
(BC)v = R, then Bv is principal
Condition (γ) clearly has a resemblance to condition b) of the proposition we
just proved. In fact, we have the following theorem.
Proposition 3.6. Let R be an integrally closed PSP-domain. The following are
equivalent.
a) Every nonconstant irreducible polynomial f ∈ R[x] is irreducible in K[x]
b) Whenever f = gh in R[x] and the greatest common divisor of the coefficients
of f exists, then the greatest common divisor of the coefficients of g exists
c) Whenever B, C are finitely generated fractional ideals of R such that (BC)v =
R, then Bv is principal
Proof. We already proved the equivalence of a) and b).
8
MARK BATELL AND JIM COYKENDALL
b) =⇒ c). Suppose (BC)v = R. Let g be a polynomial whose coefficients are
the generators of B (if generators of B are chosen in order B = (b0 , b1 , · · · , bn ) we
will define, g = b0 + b1 x + · · · + bn xn ) and let h be defined similarly with respect to
chosen generators of C. Choose nonzero b, c ∈ R such that bg, ch ∈ R[x]. Since R
is integrally closed, (Agh )v = (Ag Ah )v [G, Proposition 34.8], so that (Agh )v = R.
Since (bcAgh )v = bc((Agh )v ) [G, Proposition 32.1(1)], we therefore have (bcAgh )v =
bcR. This implies that the greatest common divisor of the coefficients of bcgh equals
bc. Since bcgh = (bg)(ch), it follows by assumption that the greatest common
divisor of the coefficients of bg exists, so that (Abg )v = (bB)v is principal [AQ,
Theorem 3.3]. Hence Bv is principal.
c) =⇒ b). Suppose f = gh in R[x] and the greatest common divisor of the
coefficients of f exists, say s. Let f1 = fs and h1 = hs . Then (Ag Ah1 )v = (Agh1 )v =
(Af1 )v , so (Ag Ah1 )v = R. Hence (Ag )v is principal. Hence the greatest common
divisor of the coefficients of g exists.
Putting together the results of this section we obtain our main result, which is
the following.
Theorem 3.7. Let R be a domain with quotient field K. The following are equivalent.
1) Every nonconstant irreducible polynomial f ∈ R[x] is irreducible in K[x]
2) R[x] is an AP-Domain
3) R[x] is a GL-Domain
4) Each of the following holds:
(α) R has the PSP-property
(β) R is integrally closed, and
(γ) Whenever B, C are finitely generated fractional ideals of R such that
(BC)v = R, then Bv is principal
We close with a few observations. First, it is of note that if R[x] is atomic,
then R[x] is an AP-domain if and only if R is a UFD. Also we note that our main
theorem of the previous section has its resolution in this stronger result. Indeed, if
we have the hypothesis of Theorem 2.9, then R[x] is an AP-domain. Hence R is an
AP-domain. If R has at least one atom then ρ(R) = ρ(R[x]) = 1.
Finally, if R is a Prüfer domain satisfying property (P) then since every finitely
generated ideal is invertible we must have Bv principal for each B. Hence R is
a GCD-domain. And if [a0 , · · · , an ] = 1, then (a0 , · · · , an )−1 = R. Hence there
exist r0 , · · · , rn ∈ R such that r0 a0 + · · · + rn an = 1. We conclude that the
greatest common divisor of a finite set of elements is a linear combination of that
set, and so R is a Bezout domain. Thus we obtain [G, Theorem 28.8], which says
(paraphrasing) that a Prüfer domain R has property (P) iff R is Bezout.
References
[AQ]
[AS]
[Ca]
[Co]
[C1]
D.D. Anderson and R.O. Quintero, Some generalizations of GCD-domains, Lecture Notes
in Pure and Appl. Math., 189, Marcel Dekker, 1997, 189–195.
J. Arnold and P. Sheldon, Integral domains that satisfy Gauss’s lemma, Michigan Math.
J. 22 (1975), 39–51.
L. Carlitz, A characterization of algebraic number fields with class number two, Proc.
Amer. Math. Soc. 11 (1960), 391–392.
H. Cohn, “Advanced Number Theory,” Dover Publications, New York, 1980.
J. Coykendall, Half-factorial domains in quadratic fields, J. Algebra, 235 (2001), 417–430.
ELASTICITY IN POLYNOMIAL-TYPE EXTENSIONS
[C2]
[CM]
[CZ]
[G]
[T]
[Za]
9
J. Coykendall, A characterization of polynomial rings with the half-factorial property,
Lecture Notes in Pure and Appl. Math., 189, Marcel Dekker, 1997, 291–294.
J. Coykendall and B. Mammenga, An embedding theorem, J. Algebra, 325 (2011), 177–
185.
J. Coykendall and M. Zafrullah, AP-domains and unique factorization, J. Pure Appl.
Algebra, 189 (2004), 27–35.
R. Gilmer, Multiplicative Ideal Theory, Marcel Dekker, 1972.
H. Tang, Gauss’ lemma, Proc. Amer. Math. Soc. 35 (1972), 372–376.
A. Zaks, Half-factorial domains, Israel J. Math., 37 (1980), 281–302.
Department of Mathematics, North Dakota State University, Fargo, ND 58108
E-mail address: [email protected]
Department of Mathematics, North Dakota State University, Fargo, ND 58108
E-mail address: [email protected]
| 0math.AC
|
Towards Practical, Precise and Parametric
Energy Analysis of IT Controlled Systems
Bernard van Gastel
Marko van Eekelen
Faculty of Management, Science and Technology,
Open University of the Netherlands, Heerlen, The Netherlands
Institute for Computing and Information Sciences,
Radboud University, Nijmegen, The Netherlands
{Bernard.vanGastel,Marko.vanEekelen}@ou.nl
Energy consumption analysis of IT-controlled systems can play a major role in minimising the overall
energy consumption of such IT systems, during the development phase, or for optimisation in the
field. Recently, a precise energy analysis was developed, with the property of being parametric
in the hardware. In principle, this creates the opportunity to analyse which is the best software
implementation for given hardware, or the other way around: choose the best hardware for a given
algorithm.
The precise analysis was introduced for a very limited language: ECA. In this paper, several
important steps are taken towards practical energy analysis. The ECA language is extended with
common programming language features. The application domain is further explored, and threats
to the validity are identified and discussed. Altogether, this constitutes an important step towards
analysing energy consumption of IT-controlled systems in practice.
1
Introduction
Energy analysis of IT systems is an important emerging field. Its focus is on analysing the software that
controls the IT system using models of the components of the system under analysis. Components can
vary from small components such as a sensor in the Internet of Things to large subsystems as present in
self-driving cars.
As traditionally many savings did occur on the hardware side of a computer, energy consumption is
almost a blind spot when developing software. Each next hardware generation consumed less energy to
perform the same amount of work. However, recently this development has lost its pace. At the same
time, it becomes more and more clear that software has a huge impact on the behaviour and the properties
of devices it runs on. A recent example of software influencing the working of a device is the Volkswagen
scandal. The car manufacturer used software to detect if the car was being tested. If this was found to
be the case, the diesel motor was programmed to operate in such a way that it exhausted less toxic gases
and fumes. In [13] it is calculated that 44,000 years of human life are lost in Europe because of the
fraud, which lasted at least six years. Another example is fridges from Panasonic, which could detect if
a test was going on and suppressed energy intensive defrost cycles during this test. These are negative
examples, but they do make clear that the software is in control of the device and its (energy) behaviour.
G. Bonfante, G. Moser (Eds.): 8th Workshop on Developments in
Implicit Computational complExity and 5th Workshop on Foundational
and Practical Aspects of Resource Analysis (DICE-FOPARA 2017)
EPTCS 248, 2017, pp. 24–37, doi:10.4204/EPTCS.248.7
c B. van Gastel and M. van Eekelen
This work is licensed under the
Creative Commons Attribution License.
B. van Gastel and M. van Eekelen
25
Although the software is evidently in control of the devices, there is almost no time dedicated in most
computer science curricula to the energy efficiency of software. This is peculiar since energy is of vital
importance to the modern (software) industry. For years, data centres have been located at places where
the energy is cheap, and since the rise of the smartphone more software engineers recognise that to get
good user reviews, their software should not rapidly deplete the battery charge of the user’s phone. Due
to this lack of educational attention to energy-aware programming, most aspiring programmers never
learn to produce energy efficient code. Software engineers have trouble assessing how much energy will
be consumed by their software on a target device, especially when the software is run on a multitude of
different systems. With the advent of the internet of things, where software is increasingly embedded in
our daily life, the software industry should become aware of their energy footprint, and methods must be
developed to assist in reducing this footprint.
Furthermore, the combination of many individual negative effects can also affect our society at large.
Although this effect is less direct, it is no less essential. If devices that are present in large quantities
in our society all exhibit the same negative behaviour, such as incurring needlessly a too high energy
consumption, they can impact public utilities and our economy and will consume the finite resources
of Earth even faster. Governments increasingly recognise this societal effect, as indicated by the new
laws in the European Union issuing ecodesign requirements for many kinds of devices. One of the aims
of these requirements is to make devices more energy efficient. Examples of product categories with
ecodesign requirements include vacuum cleaners, electrical motors, lighting, heaters, cooking appliances,
televisions and coffee machines. Even requirements leading to relative small improvements in energy
efficiency can yield large results at scale, even in the case of devices of which one would expect no
significant electricity savings to be possible.
Modern devices and appliances are controlled by software, which makes analysing the energy consumption challenging of these devices, as the behaviour of its software is difficult to predict. To analyse the
consumption of hardware, the software controlling the hardware needs to be analysed together with the
hardware.
Our approach To this end, we proposed in [5] a hybrid approach, joining energy behaviour models of
the hardware with the energy-aware semantics of software and a program transformation. The interface
between hardware and software is made explicit and has to be well defined, allowing for exchanging
of hardware or software components. Using this parametric approach, multiple implementations can be
analysed. Such an approach can be used on design level or for optimisation. One can e.g. choose the best
software implementation for given hardware, or the other way around: choose the best hardware for a
given algorithm.
The described approach derives energy consumption functions. These energy functions signify the exact
energy behaviour when the software is executing and controlling the modelled hardware. Hardware is
modelled as a finite state machine, with both the states and transitions labelled with energy consumption.
The programs that can be analysed are written in the software language ECA, which is an imperative
language inspired by C and Java. Currently, only a limited set of program constructs is supported.
The most important contributions of this article are:
• extended support for common language features in the software to be analysed: adding types, data
structures, global variables, and recursion;
26
Towards Practical, Precise and Parametric energy Analysis of IT Controlled Systems
• description of application domain of the ECA energy analysis method;
• identified threats to the validity of proposed approach and a discussion of how to deal with these
threats.
Overview Section 2 introduces the ECA energy analysis. In section 3 extensions of the ECA language
are defined including the new derivation rules that are needed for the analysis. Section 4 explores the
application domain of ECA. The validity of the results of the analysis is discussed in section 5. Finally,
we conclude with related work, future work and conclusions in sections 6 and 7.
2
Introduction to energy analysis with ECA
Energy analysis combines hardware modelling with the energy-aware semantics of software. To this
end the language ECA is specified, on which our analysis is targeted. Based on this, a semantics of this
language can be defined, which includes energy consumption. Using this energy-aware semantics, a
program transformation is given. This transformation generates an executable model, using a parametric
function. If both a concrete input and one or multiple hardware models are specified, the parametric
function will result in the energy consumption occurring when running the software on the given hardware.
To illustrate this process, we start with the hardware modelling and continue with describing for some
program constructs the semantic rules, the program transformations and the effect on hardware. This is
an introductory overview. For further details, the reader is referred to [5, 6, 10].
The hardware conceptually consists of a component state and a set of component functions which operate
on the component state. We use finite state models to model these hardware components, with the
transitions constituting function calls on the hardware. Energy usage is expressed by labelling both the
vertices and edges with energy consumption, which can be in any unit. Labels on vertices constitute
time-bound energy consumption, i.e. power draw. Edges are labelled with the consumption of a certain
amount of energy, not time-bound but corresponding to the transition. Depending on your needs, you
can model energy consumption in one way or the other. Besides the ones above described, there are
no additional requirements. We use the power draw function φ which translates a component state to a
power draw. The result of this function is used to calculate energy consumption for the time spent in a
specific state.
All transitions in a component model are explicit in the ECA source code. We use the notation C:: f to
refer to a function f() operating on a component C. Multiple different hardware components can be used
simaltaniously from the same program, the components are differentiated by a unique name (substituted
in the rules for C). Besides this addition, the ECA language is a fairly default imperative language sporting
functions, a single signed integer type, conditionals, looping constructs and expressions that can be used
as statements.
Next, we move on to the semantics, for now without energy added. Besides the function environment
∆ and the program state σ , we have the (hardware) component states Γ. This makes the effect on the
hardware explicit. The effect of the component function C:: f is split into two: the effect function δ C:: f and
a function rvC:: f calculating the return value of the component call. Both are working on the component
B. van Gastel and M. van Eekelen
27
8J
1 J/s
b
4J
start
a
3J
c
10 J
8 J/s
0 J/s
1J
d
4 J/s
Figure 1: Hardware model with the energy consumption expressed in Joule.
state retrieved from Γ (by Γ(C)). The straightforward semantic definition of the component function
sCmp, not taking energy into account, is given below.
e
∆ ` he1 , σ , Γi →
− ha, σ 0 , Γ0 i
e
∆,C:: f = (δ C:: f , rvC:: f ) ` hC:: f (e1 ), σ , Γi →
− hrvC:: f (Γ0 (C), a), σ 0 , Γ0 [C ← δ C:: f (Γ0 (C), a)]i
(sCmp)
Adding to it, the energy cost of a component function call consists of including the time taken to execute
this function and the explicit energy cost attributed to this call resulting in rule esCmp.
e
∆; Φ ` he1 , σ , Γi →
− ha, σ 0 , Γ0 , E 0 i
e
∆,C:: f = (. . . , . . . ,t f , E f ); Φ ` hC:: f (e1 ), σ , Γi →
− h. . ., . . ., . . ., E 0 + Φ(Γ0 ) · t f + E f i
(esCmp)
The approach works by transforming these semantic rules into higher order expressions. When executed
on a concrete program state PState and component state CState, this expression yields the energy consumption (and new states and possible value of an ECA expression). Expressions from the ECA language
are transformed into rules that result in a tuple of three elements: a value function V , a state update function Σ (for both program state and the hardware state), and an energy consumption function E. Statements
from the ECA language are only transformed into the latter two.
These compositional expressions are composed with higher order combinators. One of these combinators
is the composition operator >>>, which first applies the left-hand side, and on the resulting state applies
the right-hand side. Another is the + operator, which is a higher order addition operator: when executed,
it calculates the energy consumptions of the two operands based on the input states and adds them together.
To explain the component call rule btCmp and function call rule btCall, we need an operator for higher
order scoping. This operator creates a new program environment but retains the component state. It can
even update the component state given a Σ function, which is needed because this Σ needs to be evaluated
28
Towards Practical, Precise and Parametric energy Analysis of IT Controlled Systems
using the original program state. The definition is as follows:
[x 7→ V, Σ] : Var ×(PState × CState → Value)
× (PState × CState → PState × CState)
→ (PState × CState → PState × CState)
[x 7→ V, Σ](ps, cs) = ([x 7→ V (ps, cs)], cs0 ) where ( , cs0 ) = Σ(ps, cs)
We also need an additional operator split, because the program state is isolated, but the component state
is not. The split function forks the evaluation into two state update functions and joins the results together.
The first argument defines the resulting program state; the second defines the resulting component state.
∆ ` e : hVex , Σex , . . .i
∆, C:: f = (xf ,Vf , Σf , . . . , . . .) `
C:: f (e) : h[xf 7→ Vex , Σex ] >>> Vf , split(Σex , [xf 7→ Vex , Σex ] >>> Σf ), . . .i
(btCmp)
The environment ∆ is extended for each component function C :: f with two elements: an energy judgment
E f and a run-time t f . Time independent energy usage can be encoded into this E f function. For functions
defined in the language, the derived energy judgement is inserted into the environment. Using the patterns
described above the component function call is expressed as:
∆ ` e : hVex , Σex , Eex i
∆, C:: f = (xf , . . . , . . . , E f ,t f ) `
C:: f (e) : h. . ., . . ., Eex + (Σex >>> (tdec (t f ) + E f ))i
(etCmp)
This concludes the short introduction to energy analysis with ECA. For a more thorough coverage,
see [5, 6, 10].
3
Increasing the expressivity of ECA
To bridge the gap between practical programming languages and ECA, several extensions to ECA are
introduced in this section: adding data structures and types, global variables and recursion.
3.1
Adding data structures and types
The only supported type in the ECA language was a signed integer. There were no explicit Booleans,
floating point numbers or data structures. To add those, types of variables need to be supported. We need
multiple modifications for this change: modifying the grammar, and adding type distinction to both the
semantic environments and program transformations.
We consider variables to be passed by-value. Functions can have side effects on the components and,
as introduced in section 3.2, on the global variables. Functions are statically scoped. Recursion is now
supported, and the changes needed to the semantic rules and program transformations are discussed in
section 3.3.
B. van Gastel and M. van Eekelen
29
The extended BNF grammar for the ECA language is defined in listing 2. We presume there is a way to
differentiate between identifiers that represent variables hvari, function names hfun-namei, components
hcomponenti, and constants hconsti.
Functions on components can now have a variable number of arguments and optionally return a value
(if not, the type void should be used). A constructor for data structures is included, with a syntax like
a function call with as name the type of the data structure), and a construct to access fields of a data
structure (with the . operator). A type checking phase is now needed to detect typing errors, like using
a data structure as a condition in the if construct. Only correctly typed programs are considered. The
language retains an explicit construct for operations on hardware components (e.g. memory, storage or
network devices). The notation C:: f refers to a function f() operating on a component C. This allows us
to reason about components in a straightforward manner.
hprogrami
::= hstruct-def i hprogrami | hfun-def i hprogrami | htypei hvari ‘=’ hexpri | ε
hstruct-def i
::= ‘struct’ hstruct-namei ‘begin’ hstruct-fieldsi ‘end’
hstruct-fieldsi ::= htypei hfield-namei ‘;’ hstruct-fieldsi | ε
htypei
::= ‘void’ | ‘bool’ | ‘int’ | ‘float’ | hstruct-namei
hfun-def i
::= htypei hfun-namei ‘(’ [hfun-argsi] ‘)’ ‘begin’ hexpri ‘end’
hfun-argsi
::= htypei hnamei ‘,’ hfun-argsi | htypei hnamei
hbin-opi
::= ‘+’ | ‘-’ | ‘*’ | ‘>’ | ‘>=’ | ‘==’ | ‘!=’ | ‘<=’ | ‘<’ | ‘and’ | ‘or’
hexpri
::=
|
|
|
|
|
|
|
hargsi
::= hexpri ‘,’ hargsi | hexpri
hstmti
::=
|
|
|
hconsti | hvari | hexpri hbin-opi hexpri
hstruct-namei ‘(’ hargsi ‘)’
hexpri ‘.’ hfield-namei
htypei hvari ‘=’ hexpri
hvari ‘=’ hexpri
hcomponenti ‘::’ hfun-namei ‘(’ [hargsi] ‘)’
hfun-namei ‘(’ [hargsi] ‘)’
hstmti ‘,’ hexpri
‘skip’ | hstmti ‘;’ hstmti | hexpri
‘if’ hexpri ‘then’ hstmti [‘else’ hstmti] ‘end’
‘repeat’ hexpri ‘begin’ hstmti ‘end’
‘while’ hexpri ‘begin’ hstmti ‘end’
Listing 2: Extended BNF grammar for the ECA language, with types and data structures added, as well
as one construct in hprogrami for global variable support (see next section).
30
Towards Practical, Precise and Parametric energy Analysis of IT Controlled Systems
A typical (predictive recursive descent) parser of this extended language is in the LL(2) class of parsers,
with a small second pass. This second pass is needed, to avoid a possible infinite lookahead that is needed
to differentiate between expressions and statements. During the first phase expressions and statements
are combined into one construct. The small post-processing step differentiates between the two. In this
way, the language can still be efficiently parsed in a simple manner.
Next are the adjustments to the semantic rules. Because a type checking phase was added, no typing
error can occur when applying the semantic rules. Although the meaning differs, the syntax of the rules
remains largely the same. Likewise, we adjust the program transformation rules. To support the new
grammar rules, we add additional rules to the existing body of rules. Below is the rule for field access on
a variable listed.
e
∆; Φ ` he, σ , Γi →
− hv, σ 0 , Γ0 , E 0 i
e
∆; Φ ` he.a, σ , Γi →
− hv(a), σ 0 , Γ0 , E 0 + Φ(Γ0 ) · tfieldaccess i
3.2
(esField)
Global variables
Control software often works with global variables. To support analysis of this control software we
consequently need support in our ECA language for global variables. Hardware components are already
handled as global state. To also support global variables, we need to introduce an additional global
program state environment in addition to the local program state environment as it is currently used.
In the semantics, an additional global program state G is added to all tuples in every rule. Lookups
are first performed in the local scope, the already existing σ . Although for scoping a layered program
state can be preferred, or one based on indirections, we use a different approach. Because of special
handling of the, by definition global, hardware components, the global program state is handled in the
same manner as the global hardware component states. We add rules in the semantics for global variable
definitions. The assignment rule is split depending if you assign a global or local variable. The variable
loop is adjusted to first look in the local scope and if nothing is found, continue in the global scope. We
start with introducing the global assignment rule.
e
∆; Φ ` he1 , σ , G, Γi →
− hn, σ 0 , G0 , Γ0 , E 0 i
e
∆; Φ ` hx := e1 , σ , G, Γi →
− hn, σ 0 , G0 [x ← n], Γ0 , E 0 + Φ(Γ0 ) · tassign i
(esGlobAssign)
Next, we adjust the variable lookup rule, with a ∪ defined on two program environments. This ∪ creates
one environment, according to the scoping rules. It a variable is defined in both, the left-hand argument
to the ∪ is normative.
e
∆; Φ ` hx, σ , GiΓ →
− h(σ ∪ G)(x), σ , G, Γ, Φ(Γ) · tvar i
(esVar)
For the program transformation rules, more extensive changes are needed. We extend the (global)
component states CState with a global program state of type PState. For clarity of presentation, we
introduce the type GState (global state) which is a combination of PState × CState. In this way, only
the higher order combinators need to be changed, as the rules themselves remain intact. Most changes
are only to the signature of those combinators. As a result, these combinators retain their compositional
B. van Gastel and M. van Eekelen
31
properties. As an example, the higher order scoping rule of section 2 is redefined below.
[x 7→ V, Σ] : Var ×(PState × GState → Value)
× (PState × GState → PState × GState)
→ (PState × GState → PState × GState)
[x 7→ V, Σ](ps, gs) = ([x 7→ V (ps, gs)], gs0 ) where ( , gs0 ) = Σ(ps, gs)
The lookup function that the variable lookup rule depends on is redefined as follows:
lookupx : PState × GState → Value
lookupx (ps, gs) = ps(x)
if x exists in local program scope ps
lookupx (ps, gs) = Variables(gs)(x)
3.3
if x exists in the global variables part of gs
Adding recursion
We can define the function call in a similar way as the component call, which was introduced in section 2.
However, to support recursion, a special subst higher order function is introduced to unfold the function
definition once, just before it is executed on a concrete environment. The subst is defined as follows, with
PState signifying a program state, GState signifying the global state (extended in section 3.2 to be both
the component states and the global variables) and T a type variable (depending on whether a state update
or a value function is substituted):
subst : (PState × GState → T )
× (PState × GState → T )
→ (PState × GState → T )
subst(T, R)(ps, gs) = (T [R ← subst(T, R)])(ps, gs)
A recursive call is represented by the abstract higher-order function rec, which is a placeholder for
applying substitution on. There are multiple variants, depending on the resulting type, with the recV one
for a resulting value function, and the recΣ one for a resulting state update function.
recV : PState × GState → Value
recΣ : PState × GState → PState × GState
If there is a function body B computing a Value with for example recV in it, the value of the recursive
function can be computed by executing subst(B, recV ). As long as the original function terminates on the
given input environment, this analysis will terminate on the same input. This is the essential difference
from the btCmp rule, as can be seen in the definition of btCall below:
∆, f = (xf ,Vf , Σf ) ` e : hVex , Σex i
∆v ,
f = (xf ,Vf , Σf ) ` f (e) : h[xf 7→ Vex , Σex ] >>> subst(Vf , recV ( f )),
split(Σex , [xf 7→ Vex , Σex ] >>> subst(Σf , recΣ ( f ))) i
(btCall)
32
Towards Practical, Precise and Parametric energy Analysis of IT Controlled Systems
For each language function, a definition is placed in ∆v using the btFuncDef rule. The body of the
function is analysed, and recursive calls to the function are replaced with rec placeholders using the btRec
rule. To support this, the function definition rule inserts a special definition in the function environment
∆v , on which the btRec rule works. This leads to the following definition of btFuncDef, with P the
remaining program definition:
∆v , f = (x) ` e : hVex , Σex i
∆v , f = (x,Vex , Σex ) ` P : Σst
(btFuncDef)
∆v ` function f (x) begin e end P : Σst
The placeholders are inserted using the btRec rule. This rule analyses the expression used as the argument,
like the component and function call rules do. The definition is in fact very similar to those definitions:
∆v , f = (xf ) ` e : hVex , Σex i
∆v , f = (xf ) `
f (e) : h[xf 7→ Vex , Σex ] >>> recV ( f ), split(Σex , [xf 7→ Vex , Σex ] >>> recΣ ( f ))i
4
(btRec)
Exploring the application domain of ECA
The foreseen application area of the proposed analysis is in predicting the energy consumption of control
systems, where software controls peripherals. This includes control systems in factories, cars, aeroplanes,
smart-home applications, etc. Examples of hardware components range from heaters to engines, motors
and urban lighting. Depending on the target device energy consumption can be electricity, gas, water, or
any other resources where the consumption increases monotonically. The proposed analysis can predict
the energy consumption of multiple algorithms and different hardware configurations. The choice of
algorithm or configuration may depend on the expected workload. This makes the proposed technique
useful for both programmers and operators. Below, we discuss the application domain of ECA in a way
which is partly and informally published in the lecture notes for the TACLe PhD summer school in 2016
in Yspertal, Austria [7].
The possibility to abstract from the actual hardware specification makes the proposed approach still
applicable even when no final hardware component is available for basing the hardware model on, or
when such a model is not yet created. We observe that many decisions are based on relative properties
between systems. Abstracting hardware models can be used to focus e.g. on the relative differences
between component methods and component states.
Compared to the Hoare logic in [10], many restrictions are not present in ECA. Foremost, this type system
does not have the limitation that state change cannot depend on the argument of a component function nor
that the return value of a component function cannot depend on the state of the component. More realistic
models can, therefore, be used. This widens the number of applications, as behaviour of hardware can be
modelled that previously could not be expressed in the modelling.
However, there are still certain properties hardware models must satisfy for ECA to be applicable. Foremost, the models have to be discrete. Energy consumption that gradually increases or decreases over time
can therefore not be modelled directly. However, discrete approximations may be used. Secondly, every
state change has to be the consequence of an explicit application of a component function. So, implicit
state changes by hardware components cannot be expressed.
B. van Gastel and M. van Eekelen
33
The quality of the derived energy expressions is directly related to the quality of the used hardware models.
Depending on the goal, it is possible to use multiple models for one and the same hardware component.
For instance, if the hardware model is constructed as a worst-case model, this approach will produce
worst-case information. Similarly one can use average-case models to derive average case information.
It can be difficult to obtain detailed hardware models, sometimes for the simple reason that the hardware
is yet to be developed and not ready. We expect that, in cases where multiple software implementations
are to be compared, the relative consumption information will be sufficient to support design decisions.
This allows for constructing abstract models, with not much detail but including the relevant information
which is needed to make a proper comparison. However, this abstraction could impact the validity of the
results (as a realistic model could yield different results).
A class of applications where the approach described in this article could be useful is a company that
produces many variations of the same device. Variations occur based on local requirements, or on
regional differences in the electric grid, or on different requirements set by integrators or consumers, or
any combination thereof. The ECA approach allows for quickly designing those variations, and having a
clear view on how changes will impact all those variations.
5
Validity
The analysis is sound and complete as can be proven by induction on the syntactic structure of the
program in a similar way as in an earlier version of our analysis described in [14], in which the proof is
more complex due to the presence of approximations. An analysis method may be in itself sound and
complete but the validity of applying the method in practice can not automatically be inferred from that.
There are several validity constraints to the technique as it is discussed in this article. The quality of
the results depends directly on the quality of the component models used. There are severe restrictions
on component models, e.g. the power draw is assumed to be constant in every state of the component.
This is in practice not true for most devices, e.g. the power draw can be a function of time. This has
to be modelled in an abstract discretised component model. It is to be seen whether with discrete
approximations for real world hardware component models can be created with a level of precision that
is suited to make accurate energy estimates. It is hard to construct and validation such component models
on the right level of abstraction, as there are several real world practical issues. If basing the component
model on specifications from hardware vendors, all kinds of errors in the specification are transferred to
the component model. Production errors in the hardware, and eventually the degradation of hardware,
can induce erratic energy consumption behaviour that does not conform to the specification/component
model.
Validating a component model with a test setup is hard, as energy is hard to measure. Small differences in
energy consumption are hard to measure correctly, and outside conditions like temperature can influence
the results greatly. Energy differs significantly from other kinds of resources (e.g. memory and time),
which are measurable with great precision within a computer by the computer itself. Introducing a
standard energy consumption measurement interval might help in making measurements more uniform.
Validating if the number of states of a component model is the same as the actual number of states of
34
Towards Practical, Precise and Parametric energy Analysis of IT Controlled Systems
an actual hardware component, is a hard problem by itself. With powerful models, the actual validation
process with real hardware might just take too much time forcing the user to settle for a feasible but not
fully validated model.
There is another potential source of not matching the actual energy consumption of a realistic situation.
Compiler errors and optimisations can impact the (energy) behaviour of a source program greatly. The
compiler has influence on the timing of high-level language constructs. The timing constants used for
these language constructs should match the time it takes to execute those language constructs. Such a
match could be guaranteed by creating a resource consumption certified compiler in a similar way as was
done in the CompCert certified compiler project [11]. Of course this would require the availability of
energy aware semantics both on the source and the target level. An even more complicating matter may
be the complex design of modern processors executing the software. Even relatively small embedded
microprocessors have features (register bypass e.g.), which impact the execution timing of statements
significantly. Proper documentation of such features may be hard to find since e.g. the inner details of the
pipeline of modern CPU’s can often not be found in the documentation.
These constraints on modelling hardware components and validity implications should be lifted and
further investigated to make the technique discussed applicable to general, real-world problems. However,
depending on the context and the precision needed, the current technique can already be applicable
now. If the hardware component is relatively simple, a suited component model can be constructed.
Another valid area for the techniques discussed is to give feedback to a prospective programmer, such that
during construction of software the developer can optimise the energy consumption for various hardware
configurations.
6
Related work and future work
A few options are available to a programmer who wants to write energy-efficient code. The programmer
can look for programming guidelines and design patterns, which in most cases produce more energyefficient programs, e.g. [15, 3]. Then, he/she might make use of a compiler that optimises for energyefficiency, e.g. [20]. If the programmer is lucky, there is an energy analysis available for the specific
platform at hand, such as [9] in which the energy consumption of a processor is modelled in SimpleScalar.
However, for most platforms, this is not a viable option. In that case, the programmer might use dynamic
analysis with a measurement set-up. This, however, is not a trivial task and requires a complex set-up [8, 4].
Moreover, it only yields information for a specific benchmark [12]. Nevertheless, these approaches are
always applicable. A programmer might, however, prefer an approach that yields additional insight in a
more predictive manner.
In future work, we aim to fully implement this precise analysis to evaluate its suitability for larger systems
and further explore practical applicability. We intend to experiment with additional implementations of
various derived approximating analyses to evaluate which techniques/approximations work best in which
context.
A current limitation of the analysis is that it allows only one control process (processor). Actual systems
often consist of a network of interacting systems. Therefore, incorporating interacting systems would
increase the applicability of the approach. Such systems can be seen as hybrid automata. Theoretical and
practical results in modelling hybrid automata [1, 2] might be a useful starting point for further research.
B. van Gastel and M. van Eekelen
35
To make the presentation more concise, it might be useful to use as a subject language a first-order strictevaluation functional programming language. One can expect that this will alleviates the need to have a
separate basic dependent type system which transforms all variables into expressions over input variables.
However, expressions would still need to be expressed in terms over input variables. To support the
analysis of data types, a size analysis of data types might be useful to enable iteration over data structures,
e.g. using techniques similar to [16, 17].
On the language level, the type system is precise. However, it does not take into account optimisations
and transformations below the language level. This can be achieved by analysing the software on a lower
level, for example, the intermediate representation of a modern compiler, or even the binary level. For
increased accuracy, this may certainly be worthwhile. Another motivation to use such an intermediate
representation as the language that is analysed is the ability to support (combinations of) many higher
level languages. In this way, programs written in and consisting of multiple languages can be analysed.
It can also account for optimisations (such as common subexpression elimination, inlining, statically
evaluating expressions), which in general reduce the execution time of the program and therefore impact
the time-dependent energy usage (calls with side effects like component function calls are not optimised).
Compared with [10], the ECA energy consumption in this paper is precise instead of over-approximated.
Future research can show if the expressions derived by this type system can be transformed in such a way
that also upper bound expressions are derived. To support this, recursion and loops should be transformed
in a Cost Relation System (CRS), a special case of recurrence relations. Solving this CRS, one can
acquire a direct formula expressing the energy consumption of the recursive function or loop.
Another approach to providing more precise estimates is described in [19]. In this approach suitable
program inputs are identified through which, by measurement, more precise results can be achieved in
combination with an auxiliary energy model taking into account the energy consumption of instructions
in relation to each other.
Finally, a systematic approach to constructing component models should be looked into. One can create a
model from the specifications given by the vendor. Another way is using model learning [18] techniques,
which create a finite state model from black box testing and measuring. All the states should have a time
dependent energy consumption assigned to them, and all the transitions should be assigned incidental
energy consumption. Such a model can then be used as a component model.
7
Conclusion
Energy analysis consists of a hybrid approach in analysing both hardware and software together, to
derive energy consumptions when executing the software on the hardware. This can be used during
the development phase, or for optimisation. One can e.g. choose the best software implementation for
given hardware, or the other way around: choose the best hardware for a given algorithm. The key
to analysing larger systems is compositionality. Many programs encountered in the real world feature
language constructs such as global variables and recursion. To analyse these programs, support in the
ECA language and program transformations is required. This article extends ECA with language support
for multiple types, support for recursive functions and global variables. Important properties are retained:
it remains a composable, precise, and parametric energy analysis.
36
Towards Practical, Precise and Parametric energy Analysis of IT Controlled Systems
Furthermore, to gain additional insights in the feasibility of the approach, the article explores the
(im)possibilities of using the approach in practice by discussing validity and applications.
All in all, this constitutes a practical step towards the application of the proposed energy analysis on real
world problems.
References
[1] Rajeev Alur, Costas Courcoubetis, Thomas A. Henzinger & Pei-Hsin Ho (1992): Hybrid Automata: An
Algorithmic Approach to the Specification and Verification of Hybrid Systems. In Robert L. Grossman, Anil
Nerode, Anders P. Ravn & Hans Rischel, editors: Hybrid Systems, Lecture Notes in Computer Science 736,
Springer, pp. 209–229, doi:10.1007/3-540-57318-6 30.
[2] Rajeev Alur, Salar Moarref & Ufuk Topcu (2016): Compositional Synthesis with Parametric Reactive Controllers. In Alessandro Abate & Georgios E. Fainekos, editors: Proceedings of the 19th International Conference on Hybrid Systems: Computation and Control, HSCC 2016, Vienna, Austria, April 12-14, 2016, ACM,
pp. 215–224, doi:10.1145/2883817.2883842.
[3] Steven te Brinke, Somayeh Malakuti, Christoph Bockisch, Lodewijk Bergmans & Mehmet Aksit (2013):
A design method for modular energy-aware software. In Sung Y. Shin & José Carlos Maldonado, editors:
Proceedings of the 28th Annual ACM Symposium on Applied Computing, SAC ’13, Coimbra, Portugal,
March 18-22, 2013, ACM, pp. 1180–1182, doi:10.1145/2480362.2480584.
[4] Miguel A Ferreira, Eric Hoekstra, Bo Merkus, Bram Visser & Joost Visser (2013): SEFLab: A lab for
measuring software energy footprints. In: 2nd International Workshop on Green and Sustainable Software
(GREENS), 2013, IEEE, pp. 30–37, doi:10.1109/GREENS.2013.6606419.
[5] Bernard van Gastel, Rody Kersten & Marko C. J. D. van Eekelen (2015): Using Dependent Types to Define Energy Augmented Semantics of Programs. In Marko C. J. D. van Eekelen & Ugo Dal Lago, editors:
Foundational and Practical Aspects of Resource Analysis - 4th International Workshop, FOPARA 2015, London, UK, April 11, 2015, Revised Selected Papers, Lecture Notes in Computer Science 9964, pp. 20–39,
doi:10.1007/978-3-319-46559-3 2.
[6] Bernard van Gastel (2016): Assessing sustainability of software; analysing correctness, memory and energy consumption. Ph.D. thesis, Open Universiteit. Available at http://sustainablesoftware.info/
download/thesis-met-cover.pdf.
[7] Bernard van Gastel & Marko van Eekelen (2016): Lecture notes on ‘Analysing energy consumption
by software’, for the TACLe Summerschool 2016 in Yspertal, Austria. Technical Report, Radboud
University Nijmegen. Available at https://pms.cs.ru.nl/iris-diglib/src/getContent.php?id=
2016-van-Eekelen-EnergyAnalysing.
[8] Erik A. Jagroep, Jan Martijn van der Werf, Sjaak Brinkkemper, Giuseppe Procaccianti, Patricia Lago, Leen
Blom & Rob van Vliet (2016): Software Energy Profiling: Comparing Releases of a Software Product. In:
Proceedings of the 38th International Conference on Software Engineering Companion, ICSE ’16, ACM, New
York, NY, USA, pp. 523–532, doi:10.1145/2889160.2889216.
[9] Ramkumar Jayaseelan, Tulika Mitra & Xianfeng Li (2006): Estimating the Worst-Case Energy Consumption of Embedded Software. In: Proceedings of the 12th IEEE Real-Time and Embedded Technology and
Applications Symposium, IEEE, pp. 81–90, doi:10.1109/RTAS.2006.17.
[10] Rody Kersten, Paolo Parisen Toldin, Bernard van Gastel & Marko van Eekelen (2014): A Hoare Logic for
Energy Consumption Analysis. In: Proceedings of the Third International Workshop on Foundational and
Practical Aspects of Resource Analysis (FOPARA’13), LNCS 8552, Springer, pp. 93–109, doi:10.1007/9783-319-12466-7 6.
B. van Gastel and M. van Eekelen
37
[11] Xavier Leroy (2009): Formal verification of a realistic compiler. Communications of the ACM 52(7),
pp. 107–115, doi:10.1145/1538788.1538814. Available at http://gallium.inria.fr/~xleroy/publi/
compcert-CACM.pdf.
[12] F. A. Moghaddam, T. Geenen, P. Lago & P. Grosso (2015): A user perspective on energy profiling tools in
large scale computing environments. In: Sustainable Internet and ICT for Sustainability (SustainIT), 2015,
pp. 1–5, doi:10.1109/SustainIT.2015.7101364.
[13] Rik Oldenkamp, Rosalie van Zelm & Mark A.J. Huijbregts (2016): Valuing the human health damage caused
by the fraud of Volkswagen. Environmental Pollution 212, pp. 121 – 127, doi:10.1016/j.envpol.2016.01.053.
Available at http://www.sciencedirect.com/science/article/pii/S0269749116300537.
[14] Paolo Parisen Toldin, Rody Kersten, Bernard van Gastel & Marko van Eekelen (2013): Soundness
proof for a Hoare logic for energy consumption analysis. Technical Report ICIS–R13009, Radboud
University Nijmegen. Available at https://pms.cs.ru.nl/iris-diglib/src/getContent.php?id=
2013-ParisenToldin-SoundnessLogic.
[15] Eric Saxe (2010): Power-efficient software.
doi:10.1145/1646353.1646370.
Communications of the ACM 53(2), pp. 44–48,
[16] Olha Shkaravska, Marko C. J. D. van Eekelen & Alejandro Tamalet (2013): Collected Size Semantics for
Strict Functional Programs over General Polymorphic Lists. In Ugo Dal Lago & Ricardo Peña, editors:
Foundational and Practical Aspects of Resource Analysis - Third International Workshop, FOPARA 2013,
Bertinoro, Italy, August 29-31, 2013, Revised Selected Papers, Lecture Notes in Computer Science 8552,
Springer, pp. 143–159, doi:10.1007/978-3-319-12466-7 9.
[17] Alejandro Tamalet, Olha Shkaravska & Marko C.J.D. van Eekelen (2009): Size Analysis of Algebraic Data
Types. In Peter Achten, Pieter Koopman & Marco T. Morazán, editors: Trends in Functional Programming,
Trends in Functional Programming 9, Intellect, pp. 33–48. Available at http://www.intellectbooks.co.
uk/books/view-Book,id=4648/. ISBN 978-1-84150-277-9.
[18] Frits W. Vaandrager (2017): Model learning. Commun. ACM 60(2), pp. 86–95, doi:10.1145/2967606.
[19] Peter Wagemann, Tobias Distler, Timo Honig, Heiko Janker, Rudiger Kapitza & Wolfgang SchroderPreikschat (2015): Worst-Case Energy Consumption Analysis for Energy-Constrained Embedded
Systems.
2015 27th Euromicro Conference on Real-Time Systems (ECRTS) 00, pp. 105–114,
doi:10.1109/ECRTS.2015.17.
[20] Dmitry Zhurikhin, Andrey Belevantsev, Arutyun Avetisyan, Kirill Batuzov & Semun Lee (2009): Evaluating power aware optimizations within GCC compiler. In: GROW-2009: International Workshop on GCC
Research Opportunities, pp. 1–9, doi:10.1.1.470.8078.
| 6cs.PL
|
VARIABLE SELECTION IN MULTIVARIATE LINEAR MODELS WITH
HIGH-DIMENSIONAL COVARIANCE MATRIX ESTIMATION
arXiv:1707.04145v1 [math.ST] 13 Jul 2017
M. PERROT-DOCKÈS, C. LÉVY-LEDUC, L. SANSONNET, AND J. CHIQUET
Abstract. In this paper, we propose a novel variable selection approach in the framework
of multivariate linear models taking into account the dependence that may exist between
the responses. It consists in estimating beforehand the covariance matrix Σ of the responses
and to plug this estimator in a Lasso criterion, in order to obtain a sparse estimator of the
coefficient matrix. The properties of our approach are investigated both from a theoretical
and a numerical point of view. More precisely, we give general conditions that the estimators
of the covariance matrix and its inverse have to satisfy in order to recover the positions of
the null and non null entries of the coefficient matrix when the size of Σ is not fixed and
can tend to infinity. We prove that these conditions are satisfied in the particular case of
some Toeplitz matrices. Our approach is implemented in the R package MultiVarSel available
from the Comprehensive R Archive Network (CRAN) and is very attractive since it benefits
from a low computational load. We also assess the performance of our methodology using
synthetic data and compare it with alternative approaches. Our numerical experiments show
that including the estimation of the covariance matrix in the Lasso criterion dramatically
improves the variable selection performance in many cases.
1. Introduction
The multivariate linear model consists in generalizing the classical linear model, in which
a single response is explained by p variables, to the case where the number q of responses
is larger than 1. Such a general modeling can be used in a wide variety of applications
ranging from econometrics (Lütkepohl (2005)) to bioinformatics (Meng et al. (2014)). In the
latter field, for instance, multivariate models have been used to gain insight into complex
biological mechanisms like metabolism or gene regulation. This has been made possible
thanks to recently developed sequencing technologies. For further details, we refer the reader
to Mehmood et al. (2012). However, the downside of such a technological expansion is to
include irrelevant variables in the statistical models. To circumvent this, devising efficient
variable selection approaches in the multivariate setting has become a growing concern.
A first naive approach to deal with the variable selection issue in the multivariate setting
consists in applying classical univariate variable selection strategies to each response separately. Some well-known variable selection methods include the least absolute shrinkage and
selection operator (LASSO) proposed by Tibshirani (1996) and the smoothly clipped absolute
deviation (SCAD) approach devised by Fan and Li (2001). However, such a strategy does
not take into account the dependence that may exist between the different responses.
In this paper, we shall consider the following multivariate linear model:
(1)
Y = XB + E,
Date: July 14, 2017.
1
2
M. PERROT-DOCKÈS, C. LÉVY-LEDUC, L. SANSONNET, AND J. CHIQUET
where Y = (Yi,j )1≤i≤n,1≤j≤q denotes the n × q random response matrix, X denotes the n × p
design matrix, B denotes a p×q coefficient matrix and E = (Ei,j )1≤i≤n,1≤j≤q denotes the n×q
random error matrix, where n is the sample size. In order to model the potential dependence
that may exist between the columns of E, we shall assume that for each i in {1, . . . , n},
(2)
(Ei,1 , . . . , Ei,q ) ∼ N (0, Σ),
where Σ denotes the covariance matrix of the ith row of the error matrix E. We shall moreover
assume that the different rows of E are independent. With such assumptions, there is some
dependence between the columns of E but not between the rows. Our goal is here to design
a variable selection approach which is able to identify the positions of the null and non null
entries in the sparse matrix B by taking into account the dependence between the columns
of E.
This issue has recently been considered by Lee and Liu (2012) who extended the approach
of Rothman et al. (2010). More precisely, Lee and Liu (2012) proposed three approaches for
dealing with this issue based on penalized maximum likelihood with a weighted `1 regularization. In their first approach B is estimated by using a plug-in estimator of Σ−1 , in the second
one, Σ−1 is estimated by using a plug-in estimator of B and in the third one, Σ−1 and B are
estimated simultaneously. Lee and Liu (2012) also investigate the asymptotic properties of
their methods when the sample size n tends to infinity and the number of rows and columns
q of Σ is fixed.
In this paper, we propose to estimate Σ beforehand and to plug this estimator in a Lasso
criterion, in order to obtain a sparse estimator of B. Hence, our methodology is close to
the first approach of Lee and Liu (2012). However, there are two main differences: The
first one is the asymptotic framework in which our theoretical results are established and the
second one is the strategy that we use for estimating Σ. More precisely, in our asymptotic
framework, q is allowed to depend on n and thus to tend to infinity as n tends to infinity at a
polynomial rate. Moreover, in Lee and Liu (2012), Σ−1 is estimated by using an adaptation of
the Graphical Lasso (GLASSO) proposed by Friedman et al. (2008). This technique has also
been considered by Yuan and Lin (2007), Banerjee et al. (2008) and Rothman et al. (2008).
In this paper, we give general conditions that the estimators of Σ and Σ−1 have to satisfy in
order to be able to recover the support of B that is to find the positions of the null and non
null entries of the matrix B. We prove that when Σ is a particular Toeplitz matrix, namely
the covariance matrix of an AR(1) process, the assumptions of the theorem are satisfied.
Let us now describe more precisely our methodology. We start by “whitening” the observations Y by applying the following transformation to Model (1):
(3)
Y Σ−1/2 = XB Σ−1/2 + E Σ−1/2 .
The goal of such a transformation is to remove the dependence between the columns of Y .
Then, for estimating B, we proceed as follows. Let us observe that (3) can be rewritten as:
(4)
Y = X B + E,
with
(5)
Y = vec(Y Σ−1/2 ), X = (Σ−1/2 )0 ⊗ X, B = vec(B) and E = vec(E Σ−1/2 ),
where vec denotes the vectorization operator and ⊗ the Kronecker product.
VARIABLE SELECTION IN MULTIVARIATE LINEAR MODELS
3
With Model (4), estimating B is equivalent to estimate B since B = vec(B). Then, for
estimating B, we use the classical LASSO criterion defined as follows for a nonnegative λ:
b
(6)
B(λ)
= ArgminB kY − X Bk22 + λkBk1 ,
where k.k1 and k.k2 denote the classical `1 -norm and `2 -norm, respectively. Inspired by Zhao
and Yu (2006), Theorem 1 established some conditions under which the positions of the null
b
and non null entries of B can be recovered by using B.
In practical situations, the covariance matrix Σ is generally unknown and has thus to be
b denote an estimator of Σ. Then, the estimator Σ
b −1/2 of Σ−1/2 is such that
estimated. Let Σ
b −1 = Σ
b −1/2 (Σ
b −1/2 )0 .
Σ
b −1/2 , (3) becomes
When Σ−1/2 is replaced by Σ
b −1/2 = XB Σ
b −1/2 + E Σ
b −1/2 ,
YΣ
(7)
which can be rewritten as follows:
e
Ye = XeB + E,
(8)
where
(9)
b −1/2 ), Xe = (Σ
b −1/2 )0 ⊗ X, B = vec(B) and Ee = vec(E Σ
b −1/2 ).
Ye = vec(Y Σ
In Model (8), B is estimated by
(10)
n
o
e
B(λ)
= ArgminB kYe − XeBk22 + λkBk1 .
By extending Theorem 1, Theorem 5 gives some conditions on the eigenvalues of Σ−1 and
b and its inverse to Σ and Σ−1 , respectively, under which the
on the convergence rate of Σ
e
positions of the null and non null entries of B can be recovered by using B.
We prove in Section 2.3 that when Σ is a particular Toeplitz matrix, namely the covariance
matrix of an AR(1) process, the assumptions of Theorem 5 are satisfied. This strategy has
been implemented in the R package MultiVarSel, which is available on the Comprehensive
R Archive Network (CRAN), for more general Toeplitz matrices Σ such as the covariance
matrix of ARMA processes or general stationary processes. For a successful application of
this methodology to particular “-omic” data, namely metabolomic data, we refer the reader
to Perrot-Dockès et al. (2017). For a review of the most recent methods for estimating highdimensional covariance matrices, we refer the reader to Pourahmadi (2013).
The paper is organized as follows. Section 2 is devoted to the theoretical results of the
paper. The assumptions under which the positions of the non null and null entries of B can
be recovered are established in Theorem 1 when Σ is known and in Theorem 5 when Σ is
unknown. Section 2.3 studies the specific case of the AR(1) model. We present in Section
3 some numerical experiments in order to support our theoretical results. The proofs of our
main theoretical results are given in Section 4.
2. Theoretical results
2.1. Case where Σ is known. Let us first introduce some notations. Let
(11)
C=
1 0
X X and J = {1 ≤ j ≤ pq, Bj 6= 0},
nq
4
M. PERROT-DOCKÈS, C. LÉVY-LEDUC, L. SANSONNET, AND J. CHIQUET
where X is defined in (5) and where Bj denotes the jth component of the vector B defined in
(5).
Let also define
1
1
(12)
CJ,J =
(X•,J )0 X•,J and CJ c ,J =
(X•,J c )0 X•,J ,
nq
nq
where X•,J and X•,J c denote the columns of X belonging to the set J defined in (11) and to
its complement J c , respectively.
More generally, for any matrix A, AI,J denotes the partitioned matrix extracted from A
by considering the rows of A belonging to the set I and the columns of A belonging to the
set J, with • indicating all the rows or all the columns.
The following theorem gives some conditions under which the estimator Bb defined in (6) is
sign-consistent as defined by Zhao and Yu (2006), namely,
b = sign(B) → 1, as n → ∞,
P sign(B)
where the sign function maps positive entries to 1, negative entries to -1 and zero to 0.
Theorem 1. Assume that Y = (Y1 , Y2 , . . . , Ynq )0 satisfies Model (4). Assume also that there
exist some positive constants M1 , M2 , M3 and positive numbers c1 , c2 such that 0 < c1 + c2 <
1/2 satisfying:
(A1) for all n ≥ 1, for all j ∈ {1, . . . , pq}, n1 (X•,j )0 X•,j ≤ M1 , where X•,j is the jth column
of X defined in (5),
(A2) for all n ≥ 1, n1 λmin ((X 0 X )J,J ) ≥ M2 , where λmin (A) denotes the smallest eigenvalue
of A,
(A3) |J| = O(q c1 ), where J is defined in (11) and |J| is the cardinality of the set J,
(A4) q c2 minj∈J |Bj | ≥ M3 .
Assume also that the following strong Irrepresentable Condition holds:
(IC) There exists a positive constant vector η such that
(X 0 X )J c ,J ((X 0 X )J,J )−1 sign(BJ ) ≤ 1 − η,
where 1 is a (pq − |J|) vector of 1 and the inequality holds element-wise.
Then, for all λ that satisfies
1
λ
λ
= o q −(c1 +c2 ) , as n → ∞,
(L) q = qn = o n 2(c1 +c2 ) , √ → ∞ and
n
n
we have
b
P sign(B(λ))
= sign(B) → 1, as n → ∞,
b
where B(λ)
is defined by (6).
Remark 1. Observe that if c1 + c2 < (2k)−1 , for some positive k, then the first condition of
(L) becomes q = o(nk ). Hence for large values of k, the size q of Σ is much larger than n.
The proof of Theorem 1 is given in Section 4. It is based on Proposition 2 which is an
adaptation to the multivariate case of Proposition 1 in Zhao and Yu (2006).
b
Proposition 2. Let B(λ)
be defined by (6). Then
b
P sign(B(λ))
= sign(B) ≥ P(An ∩ Bn ),
VARIABLE SELECTION IN MULTIVARIATE LINEAR MODELS
5
where
(13)
An =
−1
(CJ,J )
WJ <
√
λ
−1
nq |BJ | −
|(CJ,J ) sign(BJ )|
2nq
and
(14)
Bn =
−1
CJ c ,J (CJ,J )
WJ − WJ c
λ
−1
,
1 − CJ c ,J (CJ,J ) sign(BJ )
≤ √
2 nq
√
with W = X 0 E/ nq. In (13) and (14), CJ,J and CJ c ,J are defined in (12) and WJ and
WJ c denote the components of W being in J and J c , respectively. Note that the previous
inequalities hold element-wise.
The proof of Proposition 2 is given in Section 4.
We give in the following proposition which is proved in Section 4 some conditions on X
and Σ under which Assumptions (A1) and (A2) of Theorem 1 hold.
Proposition 3. If there exist some positive constants M10 , M20 , m1 , m2 such that, for all
n ≥ 1,
(C1) for all j ∈ {1, . . . , p}, n1 (X•,j )0 X•,j ≤ M10 ,
(C2) n1 λmin (X 0 X) ≥ M20 ,
(C3) λmax (Σ−1 ) ≤ m1 ,
(C4) λmin (Σ−1 ) ≥ m2 ,
then Assumptions (A1) and (A2) of Theorem 1 are satisfied.
Remark 2. Observe that (C1) and (C2) hold in the case where the columns of the matrix X
are orthogonal.
We give in Proposition 6 in Section 2.3 some conditions under which Condition (IC) holds
in the specific case where Σ is the covariance matrix of an AR(1) process.
2.2. Case where Σ is unknown. Similarly as in (11) and (12), we introduce the following
notations:
e = 1 Xe0 Xe
(15)
C
nq
and
eJ,J = 1 (Xe•,J )0 Xe•,J and C
eJ c ,J = 1 (Xe•,J c )0 Xe•,J ,
C
nq
nq
(16)
where Xe•,J and Xe•,J c denote the columns of Xe belonging to the set J defined in (11) and to
its complement J c , respectively.
A straightforward extension of Proposition 2 leads to the following proposition for Model (8).
e
Proposition 4. Let B(λ)
be defined by (10). Then
e
fn ∩ B
fn ),
P sign(B(λ))
= sign(B) ≥ P(A
where
(17)
fn =
A
λ e
√
−1 f
−1
e
(CJ,J ) WJ < nq |BJ | −
|(CJ,J ) sign(BJ )|
2nq
6
M. PERROT-DOCKÈS, C. LÉVY-LEDUC, L. SANSONNET, AND J. CHIQUET
and
(18)
fn =
B
eJ c ,J (C
eJ,J )−1 W
fJ
C
fJ c
−W
λ
−1
eJ c ,J (C
eJ,J ) sign(BJ )
,
≤ √
1− C
2 nq
e √nq. In (17) and (18), C
eJ,J and C
eJ c ,J are defined in (16) and W
fJ and
with W = Xe0 E/
fJ c denote the components of W
f being in J and J c , respectively. Note that the previous
W
inequalities hold element-wise.
The following theorem extends Theorem 1 to the case where Σ is unknown and gives
some conditions under which the estimator Be defined in (10) is sign-consistent. The proof of
Theorem 5 is given in Section 4 and is based on Proposition 4.
Theorem 5. Assume that Assumptions (A1), (A2), (A3), (A4), (IC) and (L) of Theorem 1
hold. Assume also that, there exist some positive constants M4 , M5 , M6 and M7 , such that
for all n ≥ 1,
(A5) k(X 0 X)/nk∞ ≤ M4 ,
(A6) λmin ((X 0 X)/n) ≥ M5 ,
(A7) λmax (Σ−1 ) ≤ M6 ,
(A8) λmin (Σ−1 ) ≥ M7 .
Suppose also that
b −1 k∞ = OP ((nq)−1/2 ), as n tends to infinity,
(A9) kΣ−1 − Σ
b = OP ((nq)−1/2 ), as n tends to infinity.
(A10) ρ(Σ − Σ)
e
Let B(λ)
be defined by (10), then
e
P sign(B(λ))
= sign(B) → 1, as n → ∞.
In the previous assumptions, λmax (A), λmin (A), ρ(A) and kAk∞ denote the largest eigenvalue,
the smallest eigenvalue, the spectral radius and the infinite norm (induced by the associated
vector norm) of the matrix A.
Remark 3. Observe that Assumptions (A5) and (A6) hold in the case where the columns of
the matrix X are orthogonal. Note also that (A7) and (A8) are the same as (C3) and (C4)
in Proposition 3.
In order to estimate Σ, we propose the following strategy:
• Fitting a classical linear model to each column of the matrix Y in order to have access
b of the random error matrix E. It is possible since p is assumed to
to an estimation E
be fixed and smaller than n.
b by assuming that Σ has a particular structure, Toeplitz for
• Estimating Σ from E
instance.
b defined in the first step is such that:
More precisely, E
b = IdRn − X(X 0 X)−1 X 0 E =: ΠE,
(19)
E
which implies that
(20)
where E is defined in (5).
b = [IdRq ⊗ Π] E,
Eb = vec(E)
VARIABLE SELECTION IN MULTIVARIATE LINEAR MODELS
7
We prove in Proposition 7 below that our strategy for estimating Σ provides an estimator
satisfying the assumptions of Theorem 5 in the case where (E1,t )t , (E2,t )t , ..., (En,t )t are
assumed to be independent AR(1) processes.
2.3. The AR(1) case.
2.3.1. Sufficient conditions for Assumption (IC) of Theorem 1. The following proposition
gives some conditions under which the strong Irrepresentable Condition (IC) of Theorem 1
holds.
Proposition 6. Assume that (E1,t )t , (E2,t )t , ..., (En,t )t in Model (1) are independent AR(1)
processes satisfying:
Ei,t − φ1 Ei,t−1 = Zi,t , ∀i ∈ {1, . . . , n},
where the Zi,t ’s are zero-mean i.i.d. Gaussian random variables with variance σ 2 and |φ1 | < 1.
Assume also that X defined in (1) is such that X 0 X = νIdRp , where ν is a positive constant.
Moreover, suppose that if j ∈ J, then j > p and j < pq − p. Suppose also that for all j, j − p
or j + p is not in J. Then, the strong Irrepresentable Condition (IC) of Theorem 1 holds.
The proof of Proposition 6 is given in Section 4.
2.3.2. Sufficient conditions for Assumptions (A7), (A8), (A9) and (A10) of Theorem 5. The
following proposition establishes that in the particular case where the (E1,t )t , (E2,t )t , ...,
(En,t )t are independent AR(1) processes, our strategy for estimating Σ provides an estimator
satisfying the assumptions of Theorem 5.
Proposition 7. Assume that (E1,t )t , (E2,t )t , ..., (En,t )t in Model (1) are independent AR(1)
processes satisfying:
Ei,t − φ1 Ei,t−1 = Zi,t , ∀i ∈ {1, . . . , n},
where the Zi,t ’s are zero-mean i.i.d. Gaussian random variables with variance σ 2 and |φ1 | < 1.
Let
1
φb1 φb21 . . . φbq−1
1
φb
1 φb1 . . . φbq−2
1
1
1
..
..
..
..
..
b
.
.
.
Σ=
.
,
.
1 − φb21 .
.
.
.
.
..
..
..
..
..
q−1
... ... ...
1
φb1
where
Pn Pq b b
`=2 Ei,` Ei,`−1
b
(21)
φ1 = i=1
,
Pn Pq−1 b 2
E
i=1
`=1
i,`
b = (E
bi,` )1≤i≤n,1≤`≤q is defined in (19). Then, Assumptions (A7), (A8), (A9) and
where E
(A10) of Theorem 5 are valid.
The proof of Proposition 7 is given in Section 4. It is based on the following lemma.
Lemma 8. Assume that (E1,t )t , (E2,t )t , ..., (En,t )t in Model (1) are independent AR(1)
processes satisfying:
Ei,t − φ1 Ei,t−1 = Zi,t , ∀i ∈ {1, . . . , n},
8
M. PERROT-DOCKÈS, C. LÉVY-LEDUC, L. SANSONNET, AND J. CHIQUET
where the Zi,t ’s are zero-mean i.i.d. Gaussian random variables with variance σ 2 and |φ1 | < 1.
Let
Pn Pq b b
`=2 Ei,` Ei,`−1
b
,
φ1 = i=1
Pn Pq−1 b 2
i=1
`=1 Ei,`
b = (E
bi,` )1≤i≤n,1≤`≤q is defined in (19). Then,
where E
√
nqn (φb1 − φ1 ) = Op (1), as n → ∞.
Lemma 8 is proved in Section 4. Its proof is based on Lemma 10 in Section 5.
3. Numerical experiments
The goal of this section is twofold: i) to provide sanity checks for our theoretical results in
a well-controlled framework; and ii) to investigate the robustness of our estimator to some
violations of the assumptions of our theoretical results. The latter may reveal a broader scope
of applicability for our method than the one guaranteed by the theoretical results.
We investigate i) in the AR(1) framework presented in Section 2.3. Indeed, all assumptions
made in Theorems 1 and 5 can be specified with well-controllable simulation parameters in
the AR(1) case with balanced design matrix X.
Point ii) aims to explore the limitations of our theoretical framework and assess its robustness. To this end, we propose two numerical studies relaxing some of the assumptions of our
theorems: first, we study the effect of an unbalanced design – which violates the sufficient condition of the irrepresentability condition (IC) given in Proposition 6 – on the sign-consistency;
and second, we study the effect of other types of dependence than an AR(1).
In all experiments, the performance are assessed in terms of sign-consistency. In other
words, we evaluate the probability for the sign of various estimators to be equal to sign(B).
We compare the performance of three different estimators:
• Bb defined in (6), which corresponds to the LASSO criterion applied to the data
whitened with the true covariance matrix Σ; we call this estimator oracle. Its theoretical properties are established in Theorem 1.
• Be defined in (10), which corresponds to the LASSO criterion applied to the data
b we refer to this estimator as
whitened with an estimator of the covariance matrix Σ;
whitened-lasso. Its theoretical properties are established in Theorem 5.
• the LASSO criterion applied to the raw data, which we call raw-lasso hereafter.
Its theoretical properties are established only in the univariate case in Alquier and
Doukhan (2011).
3.1. AR(1) dependence structure with balanced one-way ANOVA. In this section,
we consider Model (1) where X is the design matrix of a one-way ANOVA with two balanced
groups. Each row of the random error matrix E is distributed as a centered Gaussian random
vector as in Equation (2) where the matrix Σ is the covariance matrix of an AR(1) process
defined in Section 2.3.
In this setting, Assumptions (A1), (A2) and Condition (IC) of Theorem 1 are satisfied, see
Propositions 3 and 6. The three remaining assumptions (A3), (A4) and (L) are related to
more practical quantities: (A3) controls the sparsity level of the problem, involving c1 ; (A4)
basically controls the signal-to-noise ratio, involving c2 and (L) links the sample size n, q and
the two constants c1 , c2 , so that an appropriate range of penalty λ exists for having a large
probability of support recovery. This latter assumption is used in our experiments to tune
VARIABLE SELECTION IN MULTIVARIATE LINEAR MODELS
9
the difficulty of the support recovery as follows: we consider different values of n, q, c1 , c2
and we choose a sparsity level |J| and a minimal magnitude in B such that Assumptions (A3)
and (A4) are fulfilled. Hence, the problem difficulty is essentially driven by the validity of
Assumption (L) where q = o(nk ) with c1 + c2 = 1/2k, and so by the relationship between n,
q and k.
We consider a large range of sample sizes n varying from 10 to 1000 and three different
values for q in {10, 50, 1000}. The constants c1 , c2 are chosen such that c1 + c2 = 1/2k with
c1 = c2 and k in {1, 2, 4}. Additional values of c1 and c2 have also been considered and
the corresponding results are available upon request. Finally, we consider two values for the
parameter φ1 appearing in the definition of the AR(1) process: φ1 ∈ {0.5, 0.95}.
Note that in this AR(1) setting with the estimator φb1 of φ1 defined in (21) , all the
assumptions of Theorem 5 are fulfilled, see Proposition 7.
The frequencies of support recovery for the three estimators averaged over 1000 replications
is displayed in Figure 1.
k= 1
1.00
●
●
●
●
●
● ●
● ●●
k= 2
●
●
● ● ● ● ●●●●
●
● ● ● ● ●●●
● ●●
● ●●
k= 4
●
●
● ● ● ● ●●●●
● ●●●
● ● ● ●
●
●
● ●
● ●
● ●●
●
● ●●
●
● ● ● ● ●●●●
●●●●
● ● ●●
●
0.75
●
q= 10
●
●
●
0.50
● ●
● ●
●
●
●
●
0.25
φ1
0.00
1.00
●
● ● ● ● ●●●●
● ● ● ● ●●●●
●
● ●
●●
0.75
●
●
●
●●
●●
●●
●●●
● ●
● ●
● ●
●
0.5
● ● ● ● ●●●●
0.95
●
●
●
●
●
●
●
●
●
●
●
●●
● ●
●
●
●
●
●
●
0.25
●
●
q= 50
0.50
●
● ●●
●
●
●
●
oracle
●
0.00
●
●
●
raw−lasso
1.00
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
0.00
10
● ●●
100
●
●
●
●
●
●
●
● ●
●
●
●
●
whithened−lasso
● ● ● ● ●●●●
●
●
● ●
●
●
●
●
●
●
●
0.25
●●
● ●
●
●
q= 1000
●
0.50
●
●
●
0.75
● ● ● ● ●●●●
●
1000 10
●
●
100
1000 10
100
1000
n
Figure 1. Frequencies of support recovery in a multivariate one-way ANOVA
model with two balanced groups and an AR(1) dependence.
10
M. PERROT-DOCKÈS, C. LÉVY-LEDUC, L. SANSONNET, AND J. CHIQUET
We observe from Figure 1 that whitened-lasso and oracle have similar performance
since φ1 is well estimated. These two approaches always exhibit better performance than
raw-lasso, especially when φ1 = 0.95. In this case, the sample size n required to reach the
same performance is indeed ten time larger for raw-lasso than for oracle and whitened-lasso.
Finally, the performance of all estimators are altered when n is too small, especially in
situations where the signal to noise ratio (SNR) is small and the signal is not sparse enough,
these two characteristics corresponding to small values of k.
3.2. Robustness to unbalanced designs and correlated features. The goal of this
section is to study some particular design matrices X in Model (1) that may lead to violation
of the Irrepresentability Condition (IC).
To this end, we consider the multivariate linear model (1) with the same AR(1) dependence
as the one considered in Section 3.1. Then, two different matrices X are considered: First, an
one-way ANOVA model with two unbalanced groups with respective sizes n1 and n2 such that
n1 + n2 = n; and second, a multiple regression model with p correlated Gaussian predictors
such that the rows of X are i.i.d. N (0, ΣX ).
For the one-way ANOVA, violation of (IC) may occur when r = n1 /n is too different from
1/2, as stated in Proposition 6. For the regression model, we choose for ΣX a 9 × 9 matrix
X
(p = 9) such that ΣX
i,i = 1, Σi,j = ρ, when i 6= j. The other simulation parameters are fixed
as in Section 3.1.
We report in Figure 2 the results for the case where q = 1000 and k = 2 both for unbalanced
one-way ANOVA (top panels) and regression with correlated predictors (bottom panels).
For the one-way ANOVA, r varies in {0.4, 0.2, 0.1}. For the regression case, ρ varies in
{0.2, 0.6, 0.9}. In both cases, the gray lines correspond to the ideal situation (that is, either
unbalanced or uncorrelated) denoted Ideal in the legend of Figure 2. The probability of
support recovery is estimated over 1000 runs.
From this figure, we note that correlated features or unbalanced designs deteriorate the
support recovery of all estimators. This was expected for these LASSO-based methods which
all suffer from the violation of the irrepresentability condition (IC). However, we also note
that whitened-lasso and oracle have similar performance, which means that the estimation
of Σ is not altered, and that whitening always improves the support recovery.
3.3. Robustness to more general autoregressive processes. In this section, we consider
the case where X is the design matrix of a one-way ANOVA with two balanced groups
and where Σ is the covariance matrix of an AR(m) process with m in {5, 10}. Figure 3
displays the performance of the different estimators when q = 500. Here, for computing
b in whitened-lasso, the parameters φ1 , . . . , φm of the AR(m) process are estimated as
Σ
b defined in (19) the estimations
follows. They are obtained by averaging over the n rows of E
(i)
(i)
b by using standard estimation approaches for AR
φb1 , . . . , φbm obtained for the ith row of E
processes described in Brockwell and Davis (1990). As previously, we observe from this figure
that whitened-lasso and oracle have better performance than raw-lasso.
4. Proofs
Proof of Proposition 2. For a fixed nonnegative λ, by (6),
b
Bb = B(λ)
= ArgminB kY − X Bk22 + λkBk1 .
VARIABLE SELECTION IN MULTIVARIATE LINEAR MODELS
r = 0.4
1.00
●
●
●
●
●
●
r = 0.2
●
●
●
●
●
●
●
●
r = 0.1
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.50
●
●
●
0.25
0.00
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
0
250
500
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
q = 1000
0.75
11
750 1000 0
250
500
750 1000 0
250
500
750 1000
n
ρ = 0.2
1.00
●
●
●
●
●●
●
●
●
●
ρ = 0.6
●
●
●
●
●
●
●●
0.50
0.25
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●●
●
●●
0.00
0
250
500
750 1000 0
250
500
●
oracle
●
raw−lasso
●
whithened−lasso
●
oracle
raw−lasso
●
●
●●
●
Ideal
ρ = 0.8
q = 1000
0.75
●
●
750 1000 0
●
●
250
●
●
●
500
●
●
●
whithened−lasso
●
750 1000
n
Figure 2. Frequencies of support recovery in general linear models with unbalanced designs: one-way ANOVA and regression.
m= 5
1.00
●
m= 10
●
●
●
●
●
0.75
●
●
●
●
q= 500
●
0.50
●
●
●
●
●
●
●
0.25
●
0.00
0
●
100
200
300
400
500 0
raw−lasso
whithened−lasso
●
●
●
oracle
100
200
300
400
500
n
Figure 3. Frequencies of support recovery in one-way ANOVA with AR(m)
covariance matrix.
Denoting u
b = Bb − B, we get
b 2 + λkBk
b 1 = kX B + E − X Bk
b 2 + λkb
kY − X Bk
u + Bk1 = kE − X u
bk22 + λkb
u + Bk1
2
2
= kEk22 − 2b
u0 X 0 E + u
b0 X 0 X u
b + λkb
u + Bk1 .
12
M. PERROT-DOCKÈS, C. LÉVY-LEDUC, L. SANSONNET, AND J. CHIQUET
Thus,
u
b = Argminu V (u),
where
√
√
√
V (u) = −2( nqu)0 W + ( nqu)0 C( nqu) + λku + Bk1 .
Since the first derivative of V with respect to u is equal to
√
√
2 nq C( nqu) − W + λ sign(u + B),
u
b satisfies
λ
λ
√
CJ,J ( nqb
uJ ) − WJ = − √ sign(b
uJ + BJ ) = − √ sign(BbJ ), if b
uJ + BJ = BbJ 6= 0
2 nq
2 nq
and
λ
√
CJ c ,J ( nqb
uJ ) − WJ c ≤ √ .
2 nq
Note that, if |b
uJ | < |BJ |, then BbJ 6= 0 and sign(BbJ ) = sign(BJ ).
Let us now prove that when An and Bn , defined in (13) and (14), are satisfied then there
exists u
b satisfying:
(22)
(23)
(24)
λ
√
CJ,J ( nqb
uJ ) − WJ = − √ sign(BJ ),
2 nq
|b
uJ | < |BJ |,
λ
√
CJ c ,J ( nqb
uJ ) − WJ c ≤ √ .
2 nq
Note that An implies:
(25)
λ
λ
√
√
−1
−1
−1
(CJ,J ) sign(BJ ) < (CJ,J ) WJ < nq |BJ | +
(CJ,J ) sign(BJ ) .
nq −|BJ | +
2nq
2nq
By denoting
(26)
1
λ
u
bJ = √ (CJ,J )−1 WJ −
(CJ,J )−1 sign(BJ ),
nq
2nq
we obtain from (25) that (22) and (23) hold. Note that Bn implies:
λ
− √
1 − CJ c ,J (CJ,J )−1 sign(BJ )
2 nq
λ
≤ CJ c ,J (CJ,J )−1 WJ − WJ c ≤ √
1 + CJ c ,J (CJ,J )−1 sign(BJ ) .
2 nq
Hence,
λ
λ
−1
−1
CJ c ,J (CJ,J ) WJ − √ (CJ,J ) sign(BJ ) − WJ c ≤ √ ,
2 nq
2 nq
which is (24) by (26). This concludes the proof.
Proof of Theorem 1. By Proposition 2,
b
P sign(B(λ))
= sign(B) ≥ P(An ∩ Bn ) = 1 − P(Acn ∪ Bnc ) ≥ 1 − P(Acn ) − P(Bnc ),
VARIABLE SELECTION IN MULTIVARIATE LINEAR MODELS
13
where An and Bn are defined in (13) and (14). It is thus enough to prove that P(Acn ) and
P(Bnc ) tend to zero as n tends to infinity.
By definition of An ,
λ
√
c
−1
−1
P(An ) = P (CJ,J ) WJ ≥ nq |BJ | −
|(CJ,J ) sign(BJ )|
2nq
λ
√
(27)
|bj |
,
≤ sup P |ξj | ≥ nq |Bj | −
2nq
j∈J
where
1
ξ = (ξj )j∈J = (CJ,J )−1 WJ = √ (CJ,J )−1 (X•,J )0 E =: HA E,
nq
and
b = (bj )j∈J = (CJ,J )−1 sign(BJ ).
By definition of Bn and (IC),
λ
−1
c
−1
1 − CJ c ,J (CJ,J ) sign(BJ )
P(Bn ) = P CJ c ,J (CJ,J ) WJ − WJ c > √
2 nq
λ
−1
c
c
≤ P CJ ,J (CJ,J ) WJ − WJ > √ η
2 nq
λ
(28)
≤ sup P |ζj | > √ η ,
c
2
nq
j∈J
where
1
ζ = (ζj )j∈J c = CJ c ,J (CJ,J )−1 WJ − WJ c = √
CJ c ,J (CJ,J )−1 (X•,J )0 − (X•,J c )0 E =: HB E.
nq
Note that, for all j in J,
|bj | ≤
X
1/2
X
p
p
|bj | ≤ |J|
b2j = |J|kbk2 .
j∈J
j∈J
Moreover,
kbk2 = k(CJ,J )−1 sign(BJ )k2 ≤ k(CJ,J )−1 k2
p
p
|J| := λmax ((CJ,J )−1 ) |J|,
where λmax (A) denotes the largest eigenvalue of the matrix A. Observe that
(29)
λmax ((CJ,J )−1 ) =
1
q
q
=
≤
,
0
λmin (CJ,J )
λmin ((X X )J,J )/n
M2
by Assumption (A2) of Theorem 1. Thus, for all j in J,
(30)
|bj | ≤
q|J|
.
M2
By Assumption (A4) of Theorem 1, we get thus that for all j in J,
(31)
λ
λ
λq|J|
√
√
√
−1
−c2
nq |Bj | −
(CJ,J ) sign(BJ ) j = nq |Bj | −
|bj | ≥ nq M3 q
−
.
2nq
2nq
2nqM2
14
M. PERROT-DOCKÈS, C. LÉVY-LEDUC, L. SANSONNET, AND J. CHIQUET
Thus,
P(Acn )
(32)
≤ sup P |ξj | ≥
√
j∈J
λq|J|
−c2
.
nq M3 q
−
2nqM2
Since E is a centered Gaussian random vector having a covariance matrix equal to identity,
ξ = HA E is a centered Gaussian random vector with a covariance matrix equal to:
0
HA HA
=
1
(CJ,J )−1 (X•,J )0 X•,J (CJ,J )−1 = (CJ,J )−1 .
nq
Hence, by (29), we get that for all j in J,
Var(ξj ) = (CJ,J )−1
jj
−1
≤ λmax (CJ,J
)≤
q
.
M2
Thus,
√
λq|J|
M2
λq|J|
√
−c2
−c2 √
P |ξj | ≥ nq M3 q
−
≤ P |Z| ≥ √
M3 q
nq − √
,
2nqM2
q
2 nqM2
where Z is a standard Gaussian random variable. By Chernoff inequality, we thus obtain that
for all j in J,
2 !
λq|J|
M
λq|J|
√
√
2
P |ξj | ≥ nq M3 q −c2 −
M3 q −c2 nq − √
≤ 2 exp −
.
2nqM2
2q
2 nqM2
By Assumption (A3) of Theorem 1, we get that under the last condition of (L),
λq|J|
√
= o q −c2 nq , as n → ∞.
√
nq
(33)
Thus,
P(Acn ) → 0, as n → ∞.
(34)
Let us now bound P(Bnc ). Observe that ζ = HB E is a centered Gaussian random vector
with a covariance matrix equal to:
1
0
−1
(CJ c ,J (CJ,J )−1 (X•,J )0 − X•,J
CJ,J c − X•,J c )
c )(X•,J (CJ,J )
nq
1
= CJ c ,J c − CJ c ,J (CJ,J )−1 CJ,J c =
(X•,J c )0 IdRnq − X•,J ((X•,J )0 X•,J )−1 (X•,J )0 X•,J c
nq
1
(X•,J c )0 IdRnq − ΠIm(X•,J ) X•,J c ,
=
nq
0
HB HB
=
where ΠIm(X•,J ) denotes the orthogonal projection onto the column space of X•,J . Note that,
for all j in J c ,
1
Var(ζj ) =
(X•,J c )0 IdRnq − ΠIm(X•,J ) X•,J c
nq
jj
1
1
(X•,J c )0 X•,J c jj −
=
(X•,J c )0 ΠIm(X•,J ) X•,J c
nq
nq
jj
M
1
1
(X•,J c )0 X•,J c jj ≤
,
≤
nq
q
VARIABLE SELECTION IN MULTIVARIATE LINEAR MODELS
15
where the inequalities come from Lemma 9 and Assumption (A1) of Theorem 1. Thus, for
all j in J c ,
√
λ q
λ
P |ζj | > √ η ≤ P |Z| > √ √ η ,
2 nq
2 M1 nq
where Z is a standard Gaussian random variable. By Chernoff inequality, for all j in J c ,
(
2 )
λ
λ
1
√ √ η
.
P |ζj | > √ η ≤ 2 exp −
2 nq
2 2 M1 n
Hence, under the following assumption
λ
√ → ∞,
n
which is the second condition of (L),
(35)
P(Bnc ) → 0, as n → ∞.
Proof of Proposition 3. Let us first prove that (C1) and (C3) imply (A1). For j ∈ {1, . . . , pq},
by considering the Euclidian division of j − 1 by p given by (j − 1) = pkj + rj , we observe
that
(X•,j )0 X•,j = (((Σ−1/2 )0 ⊗ X)•,j )0 ((Σ−1/2 )0 ⊗ X)•,j
= ((Σ−1/2 ) ⊗ X 0 )j,• )((Σ−1/2 )0 ⊗ X)•,j
= ((Σ−1/2 )kj +1,• ⊗ (X•,rj +1 )0 )(((Σ−1/2 )•,kj +1 )0 ⊗ X•,rj +1 )
= (Σ−1/2 )kj +1,• ((Σ−1/2 )•,kj +1 )0 ⊗ (X•,rj +1 )0 X•,rj +1
= (Σ−1 )kj +1,kj +1 ⊗ (X•,rj +1 )0 X•,rj +1
= (Σ−1 )kj +1,kj +1 (X•,rj +1 )0 X•,rj +1 .
Hence, using (C1), we get that for all j in {1, . . . , pq},
1
(X•,j )0 X•,j ≤ M10 (Σ−1 )kj +1,kj +1 ≤ M10
sup
((Σ−1 )k+1,k+1 )
n
k∈{0,...,q−1}
≤ M10 λmax (Σ−1 ) ≤ M10 m1 ,
where the last inequality comes from (C3), which gives (A1).
Let us now prove that (C2) and (C4) imply (A2). Note that
(X 0 X )J,J = (((Σ−1/2 )0 ⊗ X)0 ((Σ−1/2 )0 ⊗ X))J,J
= (Σ−1/2 (Σ−1/2 )0 ⊗ X 0 X)J,J
= (Σ−1 ⊗ X 0 X)J,J .
Then, by Theorem 4.3.15 of Horn and Johnson (1986),
λmin ((X 0 X )J,J ) = λmin ((Σ−1 ⊗ X 0 X)J,J )
≥ λmin (Σ−1 ⊗ X 0 X)
(36)
= λmin (X 0 X)λmin (Σ−1 ).
16
M. PERROT-DOCKÈS, C. LÉVY-LEDUC, L. SANSONNET, AND J. CHIQUET
Finally, by using Conditions (C2) and (C4), we obtain
1
1
λmin (X 0 X )J,J ≥ λmin (X 0 X)λmin (Σ−1 ) ≥ M20 m2 ,
n
n
which gives (A2).
Proof of Theorem 5. By Proposition 4,
e
en ∩ B
en ) = 1 − P(A
ec ∪ B
e c ) ≥ 1 − P(A
ec ) − P(B
e c ),
P sign(B(λ))
= sign(B) ≥ P(A
n
n
n
n
en and B
en are defined in (17) and (18). By definition of A
en , we get
where A
λ e
√
c
−1 f
−1
e
e
P(An ) = P
(CJ,J ) WJ ≥ nq |BJ | −
|(CJ,J ) sign(BJ )|
.
2nq
Observing that
eJ,J )−1 W
fJ = (CJ,J )−1 WJ + (CJ,J )−1 W
fJ − WJ
(C
eJ,J )−1 − (CJ,J )−1 WJ
+ (C
fJ − WJ ,
eJ,J )−1 − (CJ,J )−1 W
+ (C
eJ,J )−1 sign(BJ ) = (CJ,J )−1 sign(BJ ) + (C
eJ,J )−1 − (CJ,J )−1 sign(BJ ),
(C
and using the triangle inequality, we obtain that
√
nq
λ
−1
−1
c
e
(CJ,J ) WJ ≥
(CJ,J ) sign(BJ )
P(An ) ≤ P
|BJ | −
5
2nq
√
nq
λ
−1 f
−1
WJ − W J ≥
+P
(CJ,J )
(CJ,J ) sign(BJ )
|BJ | −
5
2nq
√
nq
λ
−1
−1
−1
eJ,J ) − (CJ,J )
WJ ≥
(C
+P
|BJ | −
(CJ,J ) sign(BJ )
5
2nq
√
nq
λ
−1
−1
−1
e
f
WJ − WJ ≥
+P
(CJ,J ) − (CJ,J )
|BJ | −
(CJ,J ) sign(BJ )
5
2nq
√
nq
λ
λ
−1
−1
−1
e
(
C
)
−
(C
)
sign(B
)
≥
|B
|
−
(C
)
sign(B
)
.
+P
√
J,J
J,J
J
J
J,J
J
2 nq
5
2nq
(37)
The first term in the r.h.s of (37) tends to 0 by the definition of Acn and (34). By (31), the
last term of (37) satisfies, for all j ∈ J:
2nq
λ
−1
−1
−1
eJ,J ) − (CJ,J )
P
(C
sign(BJ ) ≥
|BJ | −
(CJ,J ) sign(BJ )
5λ
2nq
2nq
λq|J|
−1
−1
−c2
e
(CJ,J ) − (CJ,J )
≥
≤P
sign(BJ )
M3 q
−
.
5λ
2nqM2
j
VARIABLE SELECTION IN MULTIVARIATE LINEAR MODELS
17
eJ,J )−1 − (CJ,J )−1 and s = sign(BJ ) then for all j in J:
Let U = (C
|(U s)j | =
(38)
X
Ujk sk ≤
p
|J|kU k2 .
k∈J
We focus on
eJ,J )−1 − (CJ,J )−1 k2 = k(C
eJ,J )−1 (CJ,J − C
eJ,J )(CJ,J )−1 k2 ≤ k(C
eJ,J )−1 k2 kCJ,J − C
eJ,J k2 k(CJ,J )−1 k2
k(C
≤
eJ,J )
eJ,J )
ρ(CJ,J − C
ρ(CJ,J − C
≤
,
eJ,J )λmin (CJJ )
eJ,J )(M2 /q)
λmin (C
λmin (C
where the last inequality comes from Assumption (A2) of Theorem 1, which gives that
q
(39)
k(CJ,J )−1 k2 ≤
.
M2
Using Theorem 4.3.15 of Horn and Johnson (1986), we get
eJ,J )−1 − (CJ,J )−1 k2 ≤
k(C
e
qρ(C − C)
.
e 2
λmin (C)M
e given in (11) and (15), respectively, we get
By definition of C and C
(40)
C=
0
b −1
Σ−1 ⊗ (X 0 X)
e = Σ ⊗ (X X) .
and C
nq
nq
By using that the eigenvalues of the Kronecker product of two matrices is equal to the product
of the eigenvalues of the two matrices, we obtain
b −1 )λmax ((X 0 X)/n)q
b −1 )λmax (Σ)λ
b max ((X 0 X)/n)q
ρ(Σ−1 − Σ
ρ(Σ−1 − Σ
≤
b −1 )λmin ((X 0 X)/n)M2
λmin ((X 0 X)/n)M2
λmin (Σ
b −1 ) ρ(Σ
b − Σ) + λmax (Σ) λmax ((X 0 X)/n)q
ρ(Σ−1 − Σ
≤
,
λmin ((X 0 X)/n)M2
eJ,J )−1 − (CJ,J )−1 k2 ≤
k(C
(41)
where the last inequality follows from Theorem 4.3.1 of Horn and Johnson (1986). Thus, by
Assumptions (A5), (A6), (A8), (A9) and (A10), we get that
(42)
eJ,J )−1 − (CJ,J )−1 k2 = OP (q(nq)−1/2 ), as n → ∞.
k(C
Hence, by (38), we get for all j in J that
eJ,J )−1 − (CJ,J )−1 sign(BJ )
P
(C
λq|J|
2nq
M3 q −c2 −
≥
5λ
2nqM2
j
√
p
2 nq
λq|J|
−1
−1
−c2 √
e
≤P
|J| k(CJ,J ) − (CJ,J ) k2 ≥
M3 q
nq − √
.
5λ
2 nqM2
By (33), (42) and (A3), it is enough to prove that
nq −c2
P q c1 /2 q(nq)−1/2 ≥
q
→ 0, as n → ∞.
λ
By the last condition of (L),
nq −c2
λ q
→ ∞, as n → ∞,
q 1+c1
18
M. PERROT-DOCKÈS, C. LÉVY-LEDUC, L. SANSONNET, AND J. CHIQUET
and the result follows since n tends to infinity. Hence, the last term of (37) tends to zero as
n tends to infinity.
Let us now study the second term in the r.h.s of (37).
1 e0 e
fJ − WJ = √1
W
Xe0 Ee − X 0 E J = √
X E − X 0E
nq
nq
J
J
i
1 h b −1/2
b −1/2 )0 ⊗ IdRn Vec(E) − Σ−1/2 ⊗ X 0 (Σ−1/2 )0 ⊗ IdRn Vec(E)
=√
Σ
⊗ X 0 (Σ
nq
J
(43)
o
i
1 hn b −1
d
=√
Σ − Σ−1 ⊗ X 0 Vec(E) = AZ,
nq
J
where Z is a centered Gaussian random vector having a covariance matrix equal to identity
and
on
oi
1 hn b −1
(44)
A= √
.
Σ − Σ−1 ⊗ X 0
(Σ1/2 )0 ⊗ IdRn
nq
J,•
By Cauchy-Schwarz inequality, we get for all K × nq matrix B, and all nq × 1 vector U that
for all k in {1, . . . , K},
|(BU )k | =
(45)
nq
X
Bk,` U` ≤ kBk2 kU k2 .
`=1
Thus, for all j in J, for all γ in R and all |J| × |J| matrix D,
fJ − WJ
(46)
P
D W
≥ γ = P (DAZ)j ≥ γ ≤ P (kDk2 kAk2 kZk2 ≥ γ) ,
j
where A is defined in (44) and Z is a centered Gaussian random vector having a covariance
matrix equal to identity. Hence, for all j in J,
√
nq
λ
−1 f
−1
WJ − WJ
P
(CJ,J )
≥
|Bj | −
(CJ,J ) sign(BJ ) j
5
2nq
j
√
nq
λq|J|
−c2
−1
M3 q
−
≤ P k(CJ,J ) k2 kAk2 kZk2 ≥
.
5
2nqM2
Let us bound kAk2 . Observe that
on
oi
hn
b −1 − Σ−1 ⊗ X 0
Σ
(Σ1/2 )0 ⊗ Id
=ρ
h
J,• 2
i
b −1 − Σ−1 Σ Σ
b −1 − Σ−1 ⊗ (X 0 X)
Σ
1/2
1/2
J,J
b −1 − Σ−1 Σ Σ
b −1 − Σ−1
Σ
λmax X 0 X
b −1 − Σ−1 λmax (Σ)1/2 λmax X 0 X 1/2 ,
≤ρ Σ
≤ρ
(47)
1/2
VARIABLE SELECTION IN MULTIVARIATE LINEAR MODELS
19
where the first inequality comes from Theorem 4.3.15 of Horn and Johnson (1986). Hence,
by (A5), (A8) and (A9)
on
oi
hn
1
b −1 − Σ−1 ⊗ X 0
= OP (q −1/2 (nq)−1/2 ).
Σ
(Σ1/2 )0 ⊗ Id
(48)
kAk2 = √
nq
J,• 2
By (33), (39) and (48), it is enough to prove that
!
nq
X
P
Zk2 ≥ nq n q −2c2 → 0, as n → ∞.
k=1
The result follows from the Markov inequality and the first condition of (L).
Let us now study the third term in the r.h.s of (37). Observe that
i
1 h −1/2
WJ = √
Σ
⊗ X 0 (Σ−1/2 )0 ⊗ IdRn Vec(E)
nq
J
h
i
1
d
(49)
Z =: A1 Z,
=√
Σ−1 ⊗ X 0 (Σ1/2 )0 ⊗ IdRn
nq
J,•
where Z is a centered Gaussian random vector having a covariance matrix equal to identity
and
i
1 h −1
(50)
A1 = √
.
Σ ⊗ X 0 (Σ1/2 )0 ⊗ IdRn
nq
J,•
Using (45), we get for all j in J, for all γ in R and all |J| × |J| matrix D,
(51)
P (D WJ )j ≥ γ = P (DA1 Z)j ≥ γ ≤ P (kDk2 kA1 k2 kZk2 ≥ γ) ,
where A1 is defined in (50) and Z is a centered Gaussian random vector having a covariance
matrix equal to identity. Hence, for all j in J,
√
nq
λ
−1
−1
−1
e
WJ
P
(CJ,J ) − (CJ,J )
≥
|Bj | −
(CJ,J ) sign(BJ ) j
5
2nq
j
√
nq
λq|J|
−1
−1
−c2
e
≤ P (CJ,J ) − (CJ,J )
kA1 k2 kZk2 ≥
M3 q
−
.
5
2nqM2
2
Let us now bound kA1 k2 . Note that
h
i
Σ−1 ⊗ X 0 (Σ1/2 )0 ⊗ IdRn
J,• 2
=
h
i
Σ−1/2 ⊗ X 0
J,• 2
1/2
1/2
= ρ Σ−1 ⊗ (X 0 X) J,J
≤ ρ Σ−1 ⊗ (X 0 X)
≤ λmax (Σ−1 )1/2 λmax (X 0 X)1/2 ,
where the first inequality comes from Theorem 4.3.15 of Horn and Johnson (1986). Hence,
by (A5) and (A7),
1
(52)
kA1 k2 ≤
λmax (Σ−1 )1/2 λmax (X 0 X)1/2 = OP (q −1/2 ).
nq
By (33), (42) and (52) it is thus enough to prove that
!
nq
X
2
−2c2
P
Zk ≥ nq n q
→ 0, as n → ∞.
k=1
The result follows from the Markov inequality and the first condition of (L).
20
M. PERROT-DOCKÈS, C. LÉVY-LEDUC, L. SANSONNET, AND J. CHIQUET
Let us now study the fourth term in the r.h.s of (37). By (46), for all j in J,
P
≤P
eJ,J )−1 − CJ,J )−1
(C
√
nq
λ
≥
|Bj | −
(CJ,J )−1 sign(BJ ) j
5
2nq
j
√
nq
λq|J|
−c2
M3 q
−
,
kAk2 kZk2 ≥
5
2nqM2
fJ − WJ
W
eJ,J )−1 − CJ,J )−1
(C
2
where A is defined in (44).
By (33), (42) and (48), it is thus enough to prove that
P
nq
X
!
Zk2 ≥ (nq) n2 q 1−2c2
→ 0, as n → ∞.
k=1
The result follows from the Markov inequality and the fact that c2 < 1/2.
en ). By definition of B
en , we get that
Let us now study P(B
enc ) = P
P(B
λ
−1 f
−1
e
f
e
e
e
c
c
c
CJ ,J (CJ,J ) WJ − WJ ≥ √
1 − |CJ ,J (CJ,J ) sign(BJ )|
.
2 nq
Observe that
eJ c ,J (C
eJ,J )−1 W
fJ − W
fJ c = CJ c ,J (CJ,J )−1 WJ − WJ c
C
fJ − WJ
+ CJ c ,J (CJ,J )−1 W
eJ,J )−1 − (CJ,J )−1 WJ
+ CJ c ,J (C
eJ,J )−1 − (CJ,J )−1 W
fJ − WJ
+ CJ c ,J (C
eJ c ,J − CJ c ,J (CJ,J )−1 WJ
+ C
fJ − WJ
eJ c ,J − CJ c ,J (CJ,J )−1 W
+ C
eJ c ,J − CJ c ,J (C
eJ,J )−1 − (CJ,J )−1 WJ
+ C
fJ − WJ
eJ c ,J − CJ c ,J (C
eJ,J )−1 − (CJ,J )−1 W
+ C
fJ c .
+ WJ c − W
Moreover,
eJ c ,J (C
eJ,J )−1 sign(BJ ) = CJ c ,J (CJ,J )−1 sign(BJ )
C
eJ,J )−1 − (CJ,J )−1 sign(BJ )
+ CJ c ,J (C
eJ c ,J − CJ c ,J (CJ,J )−1 sign(BJ )
+ C
eJ c ,J − CJ c ,J (C
eJ,J )−1 − (CJ,J )−1 sign(BJ ).
+ C
VARIABLE SELECTION IN MULTIVARIATE LINEAR MODELS
By (IC) and the triangle inequality, we obtain that
λ
c
−1
e ) ≤ P CJ c ,J (CJ,J ) WJ − WJ c ≥ √ η
P(B
n
24 nq
λ
−1
f
WJ − WJ ≥ √ η
+ P CJ c ,J (CJ,J )
24 nq
λ
−1
−1
e
+P
CJ c ,J (CJ,J ) − (CJ,J )
WJ ≥ √ η
24 nq
λ
−1
−1
e
f
+P
CJ c ,J (CJ,J ) − (CJ,J )
WJ − WJ ≥ √ η
24 nq
λ
eJ c ,J − CJ c ,J (CJ,J )−1 WJ ≥ √
+P
C
η
24 nq
λ
−1
fJ − WJ ≥ √ η
eJ c ,J − CJ c ,J (CJ,J )
W
+P
C
24 nq
λ
−1
−1
e
e
+P
CJ c ,J − CJ c ,J (CJ,J ) − (CJ,J )
WJ ≥ √ η
24 nq
λ
−1
−1
f
e
e
WJ − WJ ≥ √ η
+P
CJ c ,J − CJ c ,J (CJ,J ) − (CJ,J )
24 nq
λ
fJ c ≥ √
η
+P
WJ c − W
24 nq
n
o
eJ,J )−1 − (CJ,J )−1 sign(BJ ) ≥ η
+P
CJ c ,J (C
12
n
η o
−1
e
+P
CJ c ,J − CJ c ,J (CJ,J ) sign(BJ ) ≥
12
n
η o
−1
−1
e
e
+P
sign(BJ ) ≥
.
CJ c ,J − CJ c ,J (CJ,J ) − (CJ,J )
12
(53)
The first term in the r.h.s of (53) tends to 0 by (35).
Let us now study the second term of (53). By (46), we get that for all j in J c ,
λ
−1
f
P
CJ c ,J (CJ,J )
WJ − WJ
≥ √ η
24 nq
j
λ
−1
≤ P kCJ c ,J k2 k(CJ,J ) k2 kAk2 kZk2 ≥ √ η .
24 nq
Observe that
k(X•,J c )0 k2 kX•,J k2
(X•,J c )0 X•,J (X•,J )0 X•,J c 1/2
1
kCJ c ,J k2 = ρ
=
(X•,J c )0 X•,J 2 ≤
√
√
nq
nq
nq
nq
nq
1/2
1/2
0
0
(X•,J c ) X•,J c
(X•,J ) X•,J
ρ
≤ρ
nq
nq
λmax (Σ−1 )
(54)
= ρ(CJ c J c )1/2 ρ(CJ,J )1/2 ≤ ρ(C) =
λmax (X 0 X/n) = OP (q −1 ).
q
21
22
M. PERROT-DOCKÈS, C. LÉVY-LEDUC, L. SANSONNET, AND J. CHIQUET
In (54) the last inequality and the fourth equality come from Theorem 4.3.15 of Horn and
Johnson (1986) and (40), respectively. The last equality comes from (A5) and (A7).
By (39), (48) and (54), it is thus enough to prove that
2 !
2 !
nq
nq
X
X
λ
λ
√
P
Zk2 ≥ (nq)1/2 q √
→ 0, as n → ∞,
=P
Zk2 ≥ (nq) √
nq
n
k=1
k=1
which holds true by the second condition of (L) and Markov inequality. Hence, the second
term of (53) tends to zero as n tends to infinity.
Let us now study the third term of (53). By (51), we get that for all j in J c ,
λ
−1
−1
e
P
CJ c ,J (CJ,J ) − (CJ,J )
WJ
≥ √ η
24 nq
j
λ
−1
−1
e
≤ P kCJ c ,J k2 k(CJ,J ) − (CJ,J ) k2 kA1 k2 kZk2 ≥ √ η .
24 nq
By (42), (52) and (54), it is thus enough to prove that
2 !
!
nq
nq
X
X
λ
λ 2
2
2
1/2 √
q√
=P
Zk ≥ (nq) √
→ 0, as n → ∞,
P
Zk ≥ (nq)
nq
n
k=1
k=1
which holds true by the second condition of (L) and Markov inequality. Hence, the third term
of (53) tends to zero as n tends to infinity.
Let us now study the fourth term of (53). By (46), it amounts to prove that
λ
−1
−1
e
P kCJ c ,J k2 k(CJ,J ) − (CJ,J ) k2 kAk2 kZk2 ≥ √ η → 0, as n → ∞.
24 nq
By (54), (42) and (48) it is enough tho prove that
!
nq
X
λ 2
2
P
Zk ≥ (nq) (nq) √
→ 0, as n → ∞,
n
k=1
which holds true by the second condition of (L). Hence, the fourth term of (53) tends to zero
as n tends to infinity.
Let us now study the fifth term of (53). By (51), proving that the fifth term of (53) tends
to 0 amounts to proving that
λ
−1
e
c
c
P CJ ,J − CJ ,J k(CJ,J ) k2 kA1 k2 kZk2 ≥ √ η → 0, as n → ∞.
24 nq
2
eJ c ,J k2 .
Let us now bound kCJ c ,J − C
CJ c ,J
≤
(55)
=
eJ c ,J
−C
e
C −C
2
=
J c ,J
1
b −1
Σ−1 − Σ
q
e
C −C
e
C −C
J c ,J 2
1/2
J c ,J ∞
X 0X
∞
n
≤
=ρ
e
C −C
e
C −C
J c ,J
e
C −C
e
C −C
1/2
∞
1/2
J c ,J
e
≤ C −C
= OP (q −1 (nq)−1/2 ),
∞
as n tends to infinity, where the last equality comes from (A5) and (A9).
∞
VARIABLE SELECTION IN MULTIVARIATE LINEAR MODELS
23
By (39), (52) and (55), to prove that the fifth term of (53) tends to zero as n tends to
infinity, it is enough to prove that
!
nq
X
λ 2
2
→ 0, as n → ∞,
P
Zk ≥ nq √
n
k=1
which holds using Markov’s inequality and the second condition of (L).
Using similar arguments as those used for proving that the second, third and fourth terms
of (53) tend to zero, we get that the sixth, seventh and eighth terms of (53) tend to zero, as
n tends to infinity, by replacing (54) by (55).
Let us now study the ninth term of (53). Replacing J by J c in (43), (44), (46, (47) and
(48) in order to prove that the ninth term of (53) tends to 0 it is enough to prove that
2 !
nq
X
λ
P
Zk2 ≥ nq √
→ 0, as n → ∞,
n
k=1
which holds using Markov’s inequality and the second condition of (L).
Let us now study the tenth term of (53). Using the same idea as the one used for proving
(38), we get that
n
o
eJ,J )−1 − (CJ,J )−1 sign(BJ ) ≥ η
P
CJ c ,J (C
12
p
η
−1
−1
e
≤P
|J| kCJ c ,J k2 (CJ,J ) − (CJ,J )
≥
,
12
2
which tends to zero as n tends to infinity by (A3), (42), (54) and the fact that c1 < 1/2.
Let us now study the eleventh term of (53). Using the same idea as the one used for proving
(38), we get that
o
n
eJ c ,J − CJ c ,J (CJ,J )−1 sign(BJ ) ≥ η
P
C
12
p
−1
eJ c ,J − CJ c ,J k(CJ,J ) k2 ≥ η ,
≤P
|J| C
12
2
which tends to zero as n tends to infinity by (A3), (39) and (55) and the fact that c1 < 1/2.
Finally, the twelfth term of (53) can be bounded as follows:
n
o
eJ c ,J − CJ c ,J (C
eJ,J )−1 − (CJ,J )−1 sign(BJ ) ≥ η
C
P
12
p
η
−1
−1
eJ c ,J − CJ c ,J
eJ,J ) − (CJ,J )
≤P
|J| C
(C
≥
,
12
2
2
which tends to zero as n tends to infinity by (A3), (42) and (55) and the fact that c1 < 1/2.
Proof of Proposition 6. Observe that
1
−φ1
0
···
0
2
···
0
−φ1 1 + φ1 −φ1
..
..
..
−1
.
.
−φ1
.
(56)
Σ = 0
.
.
.
..
..
. . 1 + φ2 −φ
1
1
0
0
···
−φ1
1
.
24
M. PERROT-DOCKÈS, C. LÉVY-LEDUC, L. SANSONNET, AND J. CHIQUET
Let S = X 0 X = Σ−1 ⊗ X 0 X. Then,
nri +1
(1 + φ21 )nri +1
Si,j =
−φ1 nri +1
0
if j = i and ki ∈ {0, q − 1}
if j = i and ki ∈
/ {0, q − 1}
,
if j = i + p or if j = i − p
otherwise
where i − 1 = (p − 1)ki + ri corresponds to the Euclidean division of (i − 1) by (p − 1).
In order to prove (IC), it is enough to prove that
kSJ c ,J (SJ,J )−1 k∞ ≤ 1 − η,
where η ∈ (0, 1).
Since for all j, (j − p) ∈ J c or (j + p) ∈ J c ,
kSJ c ,J k∞ = ν|φ1 |.
Let A = SJ,J . Since A = (ai,j ) is a diagonally dominant matrix, then, by Theorem 1 of Varah
(1975),
1
X
kA−1 k∞ ≤
.
mink (ak,k −
ak,j )
1≤j≤|J|
j6=k
Using that for all j, (j − p) ∈
Jc
J c,
or (j + p) ∈
X
ak,j ≤ ν|φ1 |.
1≤j≤|J|
j6=k
If k ∈ J then k > p and k < pq − p. Thus,
ak,k ≥ ν(1 + φ21 ).
Hence,
kA−1 k∞ ≤
1
ν(1 + φ21 − |φ1 |)
and
|φ1 |
.
1 + φ21 − |φ1 |
Since |φ1 | < 1, the strong Irrepresentability Condition holds when
kSJ c ,J (SJ,J )−1 k∞ ≤ kSJ c ,J k∞ k(SJ,J )−1 k∞ ≤
|φ1 | ≤ (1 − η)(1 + |φ1 |2 − |φ1 |),
which is true for a small enough η.
Proof of Proposition 7. Since |φ1 | < 1,
kΣ−1 k∞ ≤ |φ1 | + |1 + φ21 | ≤ 3,
which gives (A7) by Theorem 5.6.9 of Horn and Johnson (1986).
Observe that
!
q−1
X
3 − |φ1 |
1
1
2
3
h
kΣk∞ ≤
1+2
|φ1 |
≤
=
≤
,
1+
2
2
2
1 − |φ1 |
1 − φ1
1 − φ1
1 − φ1
1 − φ21
h=1
which gives (A8) by Theorem 5.6.9 of Horn and Johnson (1986).
VARIABLE SELECTION IN MULTIVARIATE LINEAR MODELS
25
c1
b −1 has the same expression as Σ−1 defined in (56) except that φ1 is replaced by φ
Since Σ
defined in (21), we get that
b −1
Σ−1 − Σ
∞
2
≤ 2 φ1 − φb1 + φ1 − φb1 ,
which implies Assumption (A9) of Theorem 5 by Lemma 8.
Let us now check Assumption (A10) of Theorem 5. Since, by Theorem 5.6.9 of Horn and
b ≤ kΣ − Σk
b ∞ , it is enough to prove that
Johnson (1986), ρ(Σ − Σ)
b
Σ−Σ
∞
= OP ((nq)−1/2 ), as n → ∞.
Observe that
b
Σ−Σ
∞
≤
q−1
X
1
φh1
φbh1
1
−
+
2
−
1 − φ21 1 − φb21
1 − φ21 1 − φb21
h=1
q−1
q−1
X
X
φh1 − φbh1
≤
+2
φbh1
+2
1 − φ21
(1 − φ21 )(1 − φb21 )
h=1
h=1
φ21 − φb21
1
1
−
2
1 − φ1 1 − φb21
!
!
q−1
q−1
X
X
φh1 − φbh1
1
(φ1 − φb1 )(φ1 + φb1 )
1
≤
+2
+2
φbh1 − φh1
−
1 − φ21
1 − φ21 1 − φb21
(1 − φ21 )(1 − φb21 )
h=1
h=1
!
q−1
X
1
1
h
+2
φ1
−
1 − φ21 1 − φb21
h=1
! q−1
(φ1 − φb1 )(φ1 + φb1 )
(φ1 − φb1 )(φ1 + φb1 ) X bh
2
1
≤
1+
+2
φ1 − φh1 .
2| +
2
2
2
2
b
b
1
−
|φ
|
|1
−
φ
1
(1 − φ1 )(1 − φ1 )
(1 − φ1 )(1 − φ1 )
1
h=1
Moreover,
q−1
X
φbh1 − φh1 ≤ φb1 − φ1
h=1
≤ φb1 − φ1
q−1 X
h−1
X
k
|φ1 | |φb1 |h−k−1 ≤ φb1 − φ1
h=1 k=0
1
1 − |φb1 |
!
1
1 − |φ1 |
1 − |φb1 |q−1
1 − |φb1 |
!
1 − |φ1 |q−1
1 − |φ1 |
.
Thus, by Lemma 8,
b
Σ−Σ
∞
= OP ((nq)−1/2 ),
which implies Assumption (A10) of Theorem 5.
Proof of Lemma 8. In the following, for notational simplicity, q = qn . Observe that
Pn Pq b b
√1
i=1
√ b
`=2 Ei,` Ei,`−1
nq
nq φ1 =
.
P
P
q−1 b 2
n
1
i=1
`=1 Ei,`
nq
26
M. PERROT-DOCKÈS, C. LÉVY-LEDUC, L. SANSONNET, AND J. CHIQUET
By (19),
q
n X
X
bi,` E
bi,`−1 =
E
i=1 `=2
q
X
b•,`−1 =
b•,` )0 E
(E
`=2
`=2
(57)
=
q
X
q
X
(ΠE•,` )0 (ΠE•,`−1 )
(φ1 ΠE•,`−1 + ΠZ•,` )0 (ΠE•,`−1 )
`=2
q−1
X
= φ1
0
(ΠE•,` ) (ΠE•,` ) +
`=1
q
X
(ΠZ•,` )0 (ΠE•,`−1 ),
`=2
where (57) comes from the definition of (Ei,t ).
Hence,
Pq
0
√1
√
`=2 (ΠZ•,` ) (ΠE•,`−1 )
nq
b
nq(φ1 − φ1 ) =
.
1 Pn Pq−1 b 2
i=1
`=1 Ei,`
nq
√
In order to prove that nq(φb1 − φ1 ) = OP (1), it is enough to prove that
n q−1
n q−1
1 XX 2
1 X X b2
Ei,` −
Ei,` = oP (1), as n → ∞,
nq
nq
(58)
i=1 `=1
i=1 `=1
by Lemma 10 and
q
1 X
(ΠZ•,` )0 (ΠE•,`−1 ) = OP (1), as n → ∞.
√
nq
(59)
`=2
Let us first prove (58). By (20),
Eb = [IdRq ⊗ Π] E := AE.
Note that
b = A(Σ ⊗ IdRn )A0 = Σ ⊗ Π.
Cov(E)
Hence, for all i
Var(Ebi ) ≤ λmax (Σ).
Since the covariance matrix of E is equal to Σ ⊗ IdRn , for all i
Var(Ei ) ≤ λmax (Σ).
By Markov inequality,
q
q
n q−1
n q−1
n
n
1 X X b2
1 XX 2
1 X X b2
1 XX 2
Ei,` −
Ei,` =
Ei,` −
Ei,` + oP (1)
nq
nq
nq
nq
i=1 `=1
i=1 `=1
i=1 `=1
i=1 `=1
1
b 2 + oP (1).
=
kEk22 − kEk
2
nq
Observe that
b 2 = kEk2 − kAEk2 = E 0 E − E 0 A0 AE = E 0 (IdRnq − IdRq ⊗ Π) E
kEk2 − kEk
2
0
2
2
2
= E (IdRq ⊗ (IdRn − Π)) E =
pq
X
i=1
Eei2 ,
VARIABLE SELECTION IN MULTIVARIATE LINEAR MODELS
27
where Ee = OE, where O is an orthogonal matrix. Using that
e i,i ≤ λmax (Σ),
E(Eei2 ) = Cov(E)
and Markov inequality, we get (58).
Let us now prove (59). By definition of (Ei,t ) and since |φ1 | < 1, E[(ΠZ•,` )0 (ΠE•,`−1 )] = 0.
Moreover,
2
!2
! n
q X
q
n
n
X
X
X
X
(ΠZ•,` )0 (ΠE•,`−1 ) = E
Πi,k Zk,`
E
Πi,j Ej,`−1
`=2 i=1
`=2
=
X
X
k=1
j=1
Πi,k Πi0 ,k0 Πi,j Πi0 ,j 0 E Zk,` Zk0 ,`0 Ej,`−1 Ej 0 ,`0 −1
2≤`,`0 ≤q 1≤i,j,k,i0 ,j 0 ,k0 ≤n
=
X
X
Πi,k Πi0 ,k0 Πi,j Πi0 ,j 0
2≤`,`0 ≤q 1≤i,j,k,i0 ,j 0 ,k0 ≤n
X
φr1 φs1 E Zk,` Zk0 ,`0 Zj,`−1−r Zj 0 ,`0 −1−s ,
r,s≥0
since the (Ei,t ) are AR(1) processes with |φ1 | < 1. Note that E(Zk,` Zk0 ,`0 Zj,`−1−r Zj 0 ,`0 −1−s ) =
0 except when ` = `0 , k = k 0 , j = j 0 and r = s.
Thus,
!2
q
q
X
X
X
X
0
4
2r
(ΠZ•,` ) (ΠE•,`−1 )
=σ
E
φ1
Πi,k Πi0 ,k Πi,j Πi0 ,j
r≥0
`=2
=
`=2 1≤i,j,k,i0 ≤n
qσ 4
nqσ 4
Tr(Π) ≤
,
2
1 − φ1
1 − φ21
where Tr(Π) denotes the trace of Π, which concludes the proof of (59) by Markov inequality.
5. Technical lemmas
Lemma 9. Let A ∈ Mn (R) and Π an orthogonal projection matrix. For any j in {1, . . . , n}
(A0 ΠA)jj ≥ 0.
Proof of Lemma 9. Observe that
(A0 ΠA) = A0 Π0 ΠA = (ΠA)0 (ΠA),
since Π is an orthogonal projection matrix. Moreover,
(A0 ΠA)jj = e0j (ΠA)0 (ΠA)ej ≥ 0,
since (ΠA)0 (ΠA) is a positive semidefinite symmetric matrix, where ej is a vector containing
null entries except the jth entry which is equal to 1.
Lemma 10. Assume that (E1,t )t , (E2,t )t , ..., (En,t )t are independent AR(1) processes satisfying:
Ei,t − φ1 Ei,t−1 = Zi,t , ∀i ∈ {1, . . . , n},
where the Zi,t ’s are zero-mean i.i.d. Gaussian random variables with variance σ 2 and |φ1 | < 1.
Then,
n qn −1
σ2
1 XX 2 P
Ei,` −→
, as n → ∞.
nqn
1 − φ21
i=1
`=1
28
M. PERROT-DOCKÈS, C. LÉVY-LEDUC, L. SANSONNET, AND J. CHIQUET
2 ) =
Proof of Lemma 10. In the following, for notational simplicity, q = qn . Since E(Ei,`
σ 2 /(1 − φ21 ), it is enough to prove that
n q−1
P
1 XX 2
2
Ei,` − E(Ei,`
) −→ 0, as n → ∞.
nq
i=1 `=1
Since
2
2
Ei,`
=
X
φj1 Zi,`−j =
(60)
=
n q−1
1 XX 2
2
Ei,` − E(Ei,`
)
nq
1
(nq)2
i=1 `=1
n
X
X
X
0
φj1 φj1 Zi,`−j Zi,`−j 0 ,
j,j 0 ≥0
j≥0
Var
X
!
n
1 X
=
(nq)2
X
i=1
0
X
2
2
Cov(Ei,`
; Ei,`
0)
1≤`,`0 ≤q−1
0
φj1 φj1 φk1 φk1 Cov(Zi,`−j Zi,`−j 0 ; Zi,`0 −k Zi,`0 −k0 ).
i=1 1≤`,`0 ≤q−1 j,j 0 ≥0 k,k0 ≥0
By Cauchy-Schwarz P
inequality | Cov(Zi,`−j Zi,`−j 0 ; Zi,`0 −k Zi,`0 −k0 )| is bounded by a positive
constant. Moreover j≥0 |φ1 |j < ∞, hence (60) tends to zero as n tend to infinity, which
concludes the proof of the lemma.
References
Alquier, P. and P. Doukhan (2011). Sparsity considerations for dependent variables. Electron.
J. Statist. 5, 750–774.
Banerjee, O., L. E. Ghaoui, and A. D’aspremont (2008). Model selection through sparse maximum likelihood estimation for multivariate gaussian or binary data. Journal of Machine
Learning Research 9, 485–516.
Brockwell, P. J. and R. A. Davis (1990). Time Series: Theory and Methods. New York, NY,
USA: Springer-Verlag New York, Inc.
Fan, J. and R. Li (2001). Variable selection via nonconcave penalized likelihood and its oracle
properties. Journal of the American Statistical Association 96 (456), 1348–1360.
Friedman, J., T. Hastie, and R. Tibshirani (2008). Sparse inverse covariance estimation with
the graphical lasso. Biostatistics 9 (3), 432.
Horn, R. A. and C. R. Johnson (1986). Matrix Analysis. New York, NY, USA: Cambridge
University Press.
Lee, W. and Y. Liu (2012).
Simultaneous Multiple Response Regression and Inverse Covariance Matrix Estimation via Penalized Gaussian Maximum Likelihood.
J. Multivar. Anal. 111, 241–255.
Lütkepohl, H. (2005). New introduction to multiple time series analysis. Berlin: Springer.
Mehmood, T., K. H. Liland, L. Snipen, and S. Saebo (2012). A review of variable selection methods in partial least squares regression. Chemometrics and Intelligent Laboratory
Systems 118, 62 – 69.
Meng, C., B. Kuster, A. C. Culhane, and A. M. Gholami (2014). A multivariate approach to
the integration of multi-omics datasets. BMC Bioinformatics 15 (1), 162.
Perrot-Dockès, M., C. Lévy-Leduc, J. Chiquet, L. Sansonnet, M. Brégère, M. P. Étienne,
S. Robin, and G. Genta-Jouve (2017). A multivariate variable selection approach for analyzing LC-MS metabolomics data. arXiv:1704.00076.
VARIABLE SELECTION IN MULTIVARIATE LINEAR MODELS
29
Pourahmadi, M. (2013). High-Dimensional Covariance Estimation. Wiley Series in Probability and Statistics.
Rothman, A. J., P. J. Bickel, E. Levina, and J. Zhu (2008). Sparse permutation invariant
covariance estimation. Electron. J. Statist. 2, 494–515.
Rothman, A. J., E. Levina, and J. Zhu (2010). Sparse multivariate regression with covariance
estimation. Journal of Computational and Graphical Statistics 19 (4), 947–962.
Tibshirani, R. (1996). Regression shrinkage and selection via the Lasso. J. Royal. Statist. Soc
B. 58 (1), 267–288.
Varah, J. (1975). A lower bound for the smallest singular value of a matrix. Linear Algebra
and its Applications 11 (1), 3 – 5.
Yuan, M. and Y. Lin (2007). Model selection and estimation in the gaussian graphical model.
Biometrika 94 (1), 19–35.
Zhao, P. and B. Yu (2006). On model selection consistency of lasso. Journal of Machine
Learning Research 7, 2541–2563.
UMR MIA-Paris, AgroParisTech, INRA, Université Paris-Saclay, 75005, Paris, France
E-mail address: [email protected]
UMR MIA-Paris, AgroParisTech, INRA, Université Paris-Saclay, 75005, Paris, France
E-mail address: [email protected]
UMR MIA-Paris, AgroParisTech, INRA, Université Paris-Saclay, 75005, Paris, France
E-mail address: [email protected]
UMR MIA-Paris, AgroParisTech, INRA, Université Paris-Saclay, 75005, Paris, France
E-mail address: [email protected]
| 10math.ST
|
A Survey Of Methods
For Explaining Black Box Models
Riccardo Guidotti1,2 , Anna Monreale1 , Franco Turini1 , Dino Pedreschi1 , Fosca
Giannotti2
1
University of Pisa, {name.surname}@di.unipi.it
ISTI-CNR, Pisa, {name.surname}@isti.cnr.it
arXiv:1802.01933v2 [cs.CY] 19 Feb 2018
2
Abstract. In the last years many accurate decision support systems
have been constructed as black boxes, that is as systems that hide their
internal logic to the user. This lack of explanation constitutes both a
practical and an ethical issue. The literature reports many approaches
aimed at overcoming this crucial weakness sometimes at the cost of scarifying accuracy for interpretability. The applications in which black box
decision systems can be used are various, and each approach is typically developed to provide a solution for a specific problem and, as a
consequence, delineating explicitly or implicitly its own definition of interpretability and explanation. The aim of this paper is to provide a
classification of the main problems addressed in the literature with respect to the notion of explanation and the type of black box system.
Given a problem definition, a black box type, and a desired explanation
this survey should help the researcher to find the proposals more useful for his own work. The proposed classification of approaches to open
black box models should also be useful for putting the many research
open questions in perspective.
Keywords: Open The Black Box, Explanations, Interpretability, Transparent
Models
1
Introduction
The last decade has witnessed the rise of ubiquitous opaque decision systems.
These black box systems exploit sophisticated machine learning models to predict individual information that may also be sensitive. We can consider credit
score, insurance risk, health status, as examples. Machine learning algorithms
build predictive models which are able to map user features into a class (outcome
or decision) thanks to a learning phase. This learning process is made possible
by the digital traces that people leave behind them while performing everyday
activities (e.g., movements, purchases, comments in social networks, etc.). This
enormous amount of data may contain human biases and prejudices. Thus, decision models learned on them may inherit such biases, possibly leading to unfair
and wrong decisions.
2
R. Guidotti et al.
The European Parliament recently adopted the General Data Protection Regulation (GDPR), which will become law in May 2018. An innovative aspect of
the GDPR, which has been much debated, are the clauses on automated (algorithmic) individual decision-making, including profiling, which for the first time
introduce, to some extent, a right of explanation for all individuals to obtain
“meaningful explanations of the logic involved” when automated decision making takes place. Despite divergent opinions among legal scholars regarding the
real scope of these clauses [31,101,15], everybody agrees that the need for the
implementation of such a principle is urgent and that it represents today a huge
open scientific challenge. Without an enabling technology capable of explaining
the logic of black boxes, the right to an explanation will remain a “dead letter”.
By relying on sophisticated machine learning models trained on massive
datasets thanks to scalable, high-performance infrastructures, we risk to create and use decision systems that we do not really understand. This impacts not
only information on ethics, but also on safety and on industrial liability. Companies increasingly market services and products by embedding machine learning
components, often in safety-critical industries such as self-driving cars, robotic
assistants, and personalized medicine. Another inherent risk of these components
is the possibility of inadvertently making wrong decisions, learned from artifacts
or spurious correlations in the training data, such as recognizing an object in
a picture by the properties of the background or lighting, due to a systematic
bias in training data collection. How can companies trust their products without
understanding and validating the underlying rationale of their machine learning
components? Gartner predicts that “by 2018 half of business ethics violations
will occur through the improper use of Big Data analytics”. Explanation technologies are an immense help to companies for creating safer, more trustable
products, and better managing any possible liability they may have. Likewise,
the use of machine learning models in scientific research, for example in medicine,
biology, socio-economic sciences, requires an explanation not only for trust and
acceptance of results, but also for the sake of the openness of scientific discovery
and the progress of research.
As a consequence, explanation is at the heart of a responsible, open data
science, across multiple industry sectors and scientific disciplines. Different scientific communities studied the problem of explaining machine learning decision
models. However, each community addresses the problem from a different perspective and provides a different meaning to explanation. Most of the works in
the literature come from the machine learning and data mining communities.
The first one is mostly focused on describing how black boxes work, while the
second one is more interested in explaining the decisions even without understanding the details on how the opaque decision systems work in general.
Despite the fact that interpretable machine learning has been a topic for
quite some time and received recently much attention, today there are many
ad-hoc scattered results, and a systematic organization and classification of
these methodologies is missing. Many questions feed the papers in the literature proposing methodologies for interpreting black box systems [106,34]: What
Open The Black Box Survey
3
does it mean that a model is interpretable or transparent? What is an explanation? When a model or an explanation is comprehensible? Which is the best
way to provide an explanation and which kind of model is more interpretable?
Which are the problems requiring interpretable models/predictions? What kind
of decision data are affected? Which type of data records is more comprehensible? How much are we willing to lose in prediction accuracy to gain any form of
interpretability?
We believe that a clear classification considering simultaneously all these
aspects is needed to organize the body of knowledge about research investigating methodologies for opening and understanding the black box. Existing works
tend to provide just a general overview of the problem [59] highlighting unanswered questions and problems [24]. On the other hand, other works focus on
particular aspects like the impact of representation formats on comprehensibility [36], or the interpretability issues in term of advantages and disadvantages of
selected predictive models [27]. Consequently, after recognizing four categories
of problems and a set of ways to provide an explanation, we have chosen to
group the methodologies for opening and understanding black box predictors
by considering simultaneously the problem they are facing, the class of solutions
proposed for the explanation, the kind of data analyzed and the type of predictor
explained.
The rest of the paper is organized as follows. Firstly, in Section 3 we discuss
what interpretability is. Section 2 show which are the motivations for requiring
explanation for black box systems by illustrating some real cases. In Section 4
we formalize our problem definitions used to categorize the state of the art
works. Details of the classification and crucial points distinguishing the various
approaches and papers are discussed in Section 5. Sections 6, 7, 8 and 9 present
the details of the solutions proposed. Finally, Section 10 summarizes the crucial
aspects emerged from the analysis of the state of the art and discusses which are
the open research questions and future research directions.
2
Needs for Interpretable Models
Which are the real problems requiring interpretable models and explainable predictions? In this section, we briefly report some cases showing how and why
black boxes can be dangerous. Indeed, delegating decisions to black boxes without the possibility of an interpretation may be critical, can create discrimination
and trust issues.
Training a classifier on historical datasets, reporting human decisions, could
lead to the discovery of endemic preconceptions [76]. Moreover, since these rules
can be deeply concealed within the trained classifier, we risk to consider, maybe
unconsciously, such practices and prejudices as general rules. We are warned
about a growing “black box society” [74], governed by “secret algorithms protected by industrial secrecy, legal protections, obfuscation, so that intentional or
unintentional discrimination becomes invisible and mitigation becomes impossible.”
4
R. Guidotti et al.
Automated discrimination is not new and is not necessarily due to “black
box” models. A computer program for screening job applicants were used during the 1970s and 1980s in St. George’s Hospital Medical School (London). The
program used information from applicants’ forms, without any reference to ethnicity. However, the program was found to unfairly discriminate against ethnic
minorities and women by inferring this information from surnames and place of
birth, and lowering their chances of being selected for interview [62].
More recently, the journalists of propublica.org have shown that the COMPAS
score, a predictive model for the “risk of crime recidivism” (proprietary secret of
Northpointe), has a strong ethnic bias. Indeed, according to this score, a black
who did not re-offend were classified as high risk twice as much as whites who
did not re-offend, and white repeat offenders were classified as low risk twice as
much as black repeat offenders3 .
Similarly, a study at Princeton [11] shows how text and web corpora contain
human biases: names that are associated with black people are found to be significantly more associated with unpleasant than with pleasant terms, compared
to names associated with whites. As a consequence, the models learned on such
text data for opinion or sentiment mining have a possibility of inheriting the
prejudices reflected in the data.
Another example is related to Amazon.com. In 2016, the software used to
determine the areas of the US to which Amazon would offer free same-day delivery, unintentionally restricted minority neighborhoods from participating in the
program (often when every surrounding neighborhood was allowed)4 .
With respect to credit bureaus, it is shown in [12] that banks providing credit
scoring for millions of individuals, are often discordant: in a study of 500, 000
records, 29% of consumers received credit scores that differed by at least fifty
points among three major US banks (Experian, TransUnion, and Equifax). Such
a difference might mean tens of thousands of dollars over the life of a mortgage.
So much variability implies that the three scoring systems either have a very
different and undisclosed bias, or are highly arbitrary.
As example of bias we can consider [27] and [84]. In these works, the authors
show how accurate black box classifiers may result from an accidental artifact in
the training data. In [27] the military trained a classifier to recognize enemy tanks
from friendly tanks. The classifier resulted in a high accuracy on the test set, but
when it was used in the field had very poor performance. Later was discovered
that enemy photos were taken on overcast days, while friendly photos on sunny
days. Similarly, in [84] is shown that a classifier trained to recognize wolves and
husky dogs were basing its predictions to classify a wolf solely on the presence
of snow in the background.
Nowadays, Deep Neural Networks (DNNs) have been reaching very good
performances on different pattern-recognition tasks such as visual and text classification which are easily performed by humans: e.g., saying that a tomato is
displaced in a picture or that a text is about a certain topic. Thus, what differ3
4
http://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
http://www.techinsider.io/how-algorithms-can-be-racist-2016-4
Open The Black Box Survey
5
ences remain between DNNs and humans? Despite the excellent performance of
DNNs it seems to be a lot. In [93] it is shown the alteration of an image (e.g.
of a tomato) such that the change is undetectable for humans can lead a DNN
to tag the image as something else (e.g., labeling a tomato as a dog). In [69]
a related result is shown. It is easy to produce images that DNNs believe to
be recognizable with 99.99% confidence, but which are completely unrecognizable to humans (e.g., labeling white static noise as a tomato). Similarly in [46]
visually-indistinguishable training-set are created using DNNs and linear models. With respect to text, in [58] effective methods to attack DNN text classifiers
are presented. Experiments show that the perturbations introduced in the text
are difficult to be perceived by a human but are still able to fool a state-of-the-art
DNN to misclassify a text as any desirable class. These results show interesting
differences between humans and DNNs, and raise reasonable doubts about trusting such black boxes. In [112] it is shown how conventional regularization and
small generalization error fail to explain why DNNs generalize well in practice.
Specifically, they prove that established state-of-the-art CNN trained for image
classification easily fits a random labeling of the training data. This phenomenon
occurs even if the true images are replaced by unstructured random noise.
3
Interpretable, Explainable and Comprehensible Models
Before presenting the classification of the problems addressed in the literature
with respect to black box predictors, and the corresponding solutions and models
categorization, it is crucial to understand what interpretability is. Thus, in this
section, we discuss what an interpretable model is, and we analyze the various
dimensions of interpretability as well as the desiderata for an interpretable model.
Moreover, we also discuss the meaning of words like interpretability, explainability
and comprehensibility which are largely used in the literature.
To interpret means to give or provide the meaning or to explain and present
in understandable terms some concept5 . Therefore, in data mining and machine
learning, interpretability is defined as the ability to explain or to provide the
meaning in understandable terms to a human [24]. These definitions assume
implicitly that the concepts expressed in the understandable terms composing an
explanation are self-contained and do not need further explanations. Essentially,
an explanation is an “interface” between humans and a decision maker that is at
the same time both an accurate proxy of the decision maker and comprehensible
to humans.
As shown in the previous section another significant aspect to mention about
interpretability is the reason why a system, a service or a method should be
interpretable. On the other hand, an explanation could be not required if there
are no decisions that have to be made on the outcome of the prediction. For
example, if we want to know if an image contains a cat or not and this information
is not required to take any sort of crucial decision, or there are no consequences
5
https://www.merriam-webster.com/
6
R. Guidotti et al.
for unacceptable results, then we do not need an interpretable model, and we
can accept any black box.
3.1
Dimensions of Interpretability
In the analysis of the interpretability of predictive models, we can identify a set
of dimensions to be taken into consideration, and that characterize the interpretability of the model [24].
Global and Local Interpretability: A model may be completely interpretable,
i.e., we are able to understand the whole logic of a model and follow the entire reasoning leading to all the different possible outcomes. In this case, we
are speaking about global interpretability. Instead, we indicate with local interpretability the situation in which it is possible to understand only the reasons
for a specific decision: only the single prediction/decision is interpretable.
Time Limitation: An important aspect is the time that the user is available or
is allowed to spend on understanding an explanation. The user time availability
is strictly related to the scenario where the predictive model has to be used.
Therefore, in some contexts where the user needs to quickly take the decision
(e.g., a disaster is imminent), it is preferable to have an explanation simple to
understand. While in contexts where the decision time is not a constraint (e.g.,
during a procedure to release a loan) one might prefer a more complex and
exhaustive explanation.
Nature of User Expertise: Users of a predictive model may have different
background knowledge and experience in the task: decision-makers, scientists,
compliance and safety engineers, data scientists, etc. Knowing the user experience in the task is a key aspect of the perception of interpretability of a model.
Domain experts may prefer a larger and more sophisticated model over a smaller
and sometimes more opaque one.
The works reviewed in the literature only implicitly specify if their proposal
is global or local. Just a few of them take into account the nature of user expertise [29,84,87], and no one provides real experiments about the time required to
understand an explanation. Instead, some of the works consider the “complexity” of an explanation through an approximation. For example, they define the
model complexity as the model’s size (e.g. tree depth, number of rules, number
of conjunctive terms) [22,32,40,84]. In the following, we further discuss issues
related to the complexity of an explanation.
3.2
Desiderata of an Interpretable Model
An interpretable model is required to provide an explanation. Thus, to realize
an interpretable model it is necessary to take into account the following list of
desiderata which are mentioned by a set of papers in the state of art [5,24,27,38]:
– Interpretability: to which extent the model and or the prediction are human understandable. The most addressed discussion is related to how the
Open The Black Box Survey
7
interpretability can be measured. In [27] a component for measuring the interpretability is the complexity of the predictive model in terms of the model
size. According to the literature, we refer to interpretability also with the
name comprehensibility.
– Accuracy: to which extent the model accurately predict unseen instances.
The accuracy of a model can be measured using various evaluation measures
like the accuracy score, the F1-score [95], etc. Producing an interpretable
model maintaining competitive levels of accuracy is the most common target
among the papers in the literature.
– Fidelity: to which extent the model is able to accurately imitate a black-box
predictor. The fidelity captures how much is good an interpretable model
in the mimic of the behavior of a black-box. Similarly to the accuracy, the
fidelity is measured in terms of accuracy score, F1-score, etc. but with respect
to the outcome of the black box which is considered as an oracle.
Moreover, besides these features strictly related to interpretability, yet according to [5,24,27,38] a data mining and machine learning model should have
other important desiderata. Some of these desiderata are related to ethical aspects such as fairness and privacy. The first principle requires that the model
guarantees the protection of groups against (implicit or explicit) discrimination
[85]; while the second one requires that the model does not reveal sensitive information about people [4]. The level of interpretability of a model together with
the standards of privacy and non-discrimination which are guaranteed may impact on how much human users trust that model. The degree of trust on a model
increases if the model is built by respecting constraints of monotonicity given by
the users [66,75,99]. A predictor respecting the monotonicity principle is, for example, a predictor where the increase of the values of a numerical attribute tends
to either increase or decrease in a monotonic way the probability of a record of
being member of a class [27]. Another property that influences the trust level
of a model is usability: people tend to trust more models providing information
that assist them to accomplish a task with awareness. In this line, an interactive
and queryable explanation results to be more usable than a textual and fixed
explanation.
Data mining and machine learning models should have other important
desiderata such as reliability, robustness, causality, scalability and generality.
This means that a model should have the ability to maintain certain levels of
performance independently from the parameters or from the input data (reliability/robustness) and that controlled changes in the input due to a perturbation
affect the model behavior (causality). Moreover, since we are in the Big Data
era, it is opportune to have models able to scale to large input data with large
input spaces. Finally, since often in different application scenarios one might use
the same model with different data, it is preferable to have portable models that
do not require special training regimes or restrictions (generality).
8
R. Guidotti et al.
Fig. 1. Example of decision tree.
3.3
Recognized Interpretable Models
In the state of the art a small set of existing interpretable models is recognized:
decision tree, rules, linear models [27,36,84]. These models are considered easily
understandable and interpretable for humans.
A decision system based on a decision tree exploits a graph structured like a
tree and composed of internal nodes representing tests on features or attributes
(e.g., whether a variable has a value lower than, equals to or grater than a
threshold, see Figure 1), and leaf nodes representing a class label. Each branch
represents a possible outcome [79]. The paths from the root to the leaves represent the classification rules. Indeed, a decision tree can be linearized into a set
of decision rules with the if-then form [77,78,26]:
if condition1 ∧ condition2 ∧ condition3 then outcome.
Here, the outcome corresponds to the class label of a leaf node while the conjunctions of conditions in the if clause correspond to the different conditions in
the path from the root to that leaf node.
More generally, a decision rule is a function which maps an observation to
an appropriate action. Decision rules can be extracted by generating the socalled classification rules, i.e., association rules that in the consequence have
the class label [3]. The most common rules are if-then rules where the if clause
is a combination of conditions on the input variables. In particular, it may be
formed by conjunctions, negations and disjunctions. However, methods for rule
extraction typically take into consideration only rules with conjunctions. Other
types of rules are: m-of-n rules where given a set of n conditions if m of them are
verified then the consequence of the rule is considered true [68]; list of rules where
given an ordered set of rules is considered true the consequent of the first rule
which is verified [109]; falling rule lists consists of a list of if-then rules ordered
with respect to the probability of a specific outcome and the order identifies the
example to be classified by that rule [102]; decision sets where an unordered set
of classification rules is provided such that the rules are not connected by else
statements, but each rule is an independent classifier that can assign its label
without regard for any other rules [53].
The interpretation of rules and decision trees is different with respect to
different aspects [27]. Decision trees are widely adopted for their graphical representation, while rules have a textual representation. The main difference is
Open The Black Box Survey
9
Fig. 2. Example of feature importance for a linear model.
that textual representation does not provide immediately information about the
more relevant attributes of a rule. On the other hand, the hierarchical position
of the features in a tree gives this kind of clue.
Attributes’ relative importance could be added to rules by means of positional information. Specifically, rule conditions are shown by following the order
in which the rule extraction algorithm added them to the rule. Even though
the representation of rules causes some difficulties in understanding the whole
model, it enables the study of single rules representing partial parts of the whole
knowledge (“local patterns”) which are composable. Also in a decision tree, the
analysis of each path separately from the leaf node to the root, enables users to
focus on such local patterns. However, if the tree is very deep in this case it is a
much more complex task. A further crucial difference between rules and decision
trees is that in a decision tree each record is classified by only one leaf node,
i.e., the class predicted are represented in a mutually exclusive and exhaustive
way by the set of leaves and their paths to the root node. On the other hand,
a certain record can satisfy the antecedent of rules having as consequent a different class for that record. Indeed, rule based classifiers have the disadvantage
of requiring an additional approach for resolving such situations of conflicting
outcome [107]. Many rule based classifiers deal with this issue by returning an
ordered rule list, instead of an unordered rule set. In this way it is returned the
outcome corresponding to the first rule matching the test record and ignoring
the other rules in the list. We notice that ordered rule lists may be harder to
interpret than classical rules. In fact, in this model a given rule cannot be considered independently from the precedent rules in the list [107]. Another widely
used approach consists in considering the top-k rules satisfying the test record
where the ordering is given by a certain weight (e.g. accuracy, Laplace accuracy,
etc.). Then, the outcome of the rules with the average highest weight among the
top-k is returned as predicted class [109].
Finally, explanations can also be provided through linear models [47,84]. This
can be done by considering and visualizing both the sign and the magnitude of
the contribution of the attributes for a given prediction (see Figure 2). If the
contribution of an attribute-value is positive, then it contributes by increasing
the model’s output. Instead, if the sign is negative then the attribute-value decreases the output of the model. If an attribute-value has an higher contribution
than another, then it means that it has an higher influence on the prediction
of the model. The produced contributions summarize the performance of the
model, thus the difference between the predictions of the model and expected
10
R. Guidotti et al.
predictions, providing the opportunity of quantifying the changes of the model
prediction for each test record. In particular, it is possible to identify the attributes leading to this change and for each attribute how much it contributed
to the change.
As last remark we point out that in general, when an explanation for a prediction is provided, it is often useful to analyze besides the explanation (satisfied
rules, branch of the tree, set of weights, etc.), also instances which are exceptions
with respect to the “boundaries” provided by the explanation, or with very few
differences with respect to the prototypes returned as explanation. For example,
instances covered by the rule body but with an outcome label different from the
class of the outcome predicted. Even though this sort of exception analysis is
hardly performed, it can be more informative than the direct explanation, and
it can also provide clues about the application domain [73].
3.4
Explanations and Interpretable Models Complexity
In the literature, very little space is dedicated to a crucial aspect: the model
complexity. The evaluation of the model complexity is generally tied to the model
comprehensibility, and this is a very hard task to address. As a consequence,
this evaluation is generally estimated with a rough approximation related to the
size of the model. Moreover, complexity is often used as an opposed term to
interpretability.
In [32] the complexity is identified by the number of regions, i.e., the parts
of the model, for which the boundaries are defined. In [84] as complexity for
linear models is adopted the number of non-zero weights, while for decision
trees the depth of the tree. In [22] the complexity of a rule (and thus of an
explanation) is measured by the length of the rule condition, defined as the
number of attribute-value pairs in the condition. Given two rules with similar
frequency and accuracy, the rule with a smaller length may be preferred as it is
more interpretable. Similarly, in case of lists of rules the complexity is typically
measured considering the total number of attribute-value pairs in the whole set of
rules. However, this could be a suitable way for measuring the model complexity,
since in an ordered rule list different test records need distinct numbers of rules
to be evaluated [27]. In this kind of model, a more honest measure could be
the average number of conditions evaluated to classify a set of test records [72].
However, this is more a “measure of the explanation” of a list of rules.
Differently from the not flexible representation of decision tree where the
prediction of a single record is mutually exhaustive and exclusive, rules characterization contains only significant clauses. As a consequence, an optimal set
of rules does not contain any duplicated information, given the fact that an
outcome label can appear only one time in the consequent of a set of rules,
while in a decision tree it typically comes out more than once. Moreover, rules
do not capture insignificant clauses, while decision trees can also have insignificant branches. This happens because rule based classifier generally select one
attribute-value while expanding a rule, whereas decision tree algorithms usually
select one attribute while expanding the tree [27]. Considering these aspects to
Open The Black Box Survey
11
estimate the complexity is very difficult. Consequently, even though a model
equivalence exists, the estimation of the fact that a different representation for
the same model (or explanation) is more complex than another when using decision trees or rules can be very subjective with respect to the interpreter.
3.5
Interpretable Data for Interpretable Models
The types of data used for classification may have diverse nature. Different types
of data present a different level of interpretability for a human. The most understandable data format for humans is the table [36]. Since matrices and vectors
are the typical data representation used by the vast majority of data mining and
machine learning techniques, tables are also easily managed by these algorithms
without requiring specific transformations.
Other forms of data which are very common in human daily life are images
and texts. They are perhaps for human brain even more easily understandable
than tables. On the other hand, the processing of these data for predictive models
requires their transformation into vectors that make them easier to process by
algorithms but less interpretable for humans. Indeed, on images and texts, the
state of art techniques typically apply predictive models based on super vector
machine, neural networks or deep neural networks that are usually hard to be
interpreted. As a consequence, certain recognized interpretable models cannot be
directly employed for this type of data in order to obtain an interpretable model
or a human understandable explanation. Transformations using equivalences,
approximations or heuristics are required in such a way that images and texts
can be employed by prediction systems and used for providing the interpretation
of the model and/or the prediction at the same time.
Finally, there exist other forms of data such as sequence data, spatio-temporal
data and complex network data that may be used by data mining and machine
learning algorithms. However in the literature, to the best of our knowledge,
there is no work addressing the interpretability of models for data different from
images, texts, and tabular data.
4
Open The Black Box Problems
An accurate analysis and review of the literature lead to the identification of
different categories of problems.
At a very high level, we can distinguish between reverse engineering and
design of explanations. In the first case, given the decision records produced by
a black box decision maker the problem consists in reconstructing an explanation
for it. The original dataset upon which the black box is trained is generally not
known in real life. Details about reverse engineering approaches are discussed at
the end of this section. On the other hand, it is used and exploited to build the
explanations by most of the works presented in this survey. In the second case,
given a dataset of training decision records the task consists in developing an
interpretable predictor model together with its explanations.
12
R. Guidotti et al.
Fig. 3. Classification Problem.
Through a deep analysis of the state of the art we are able to further refine
the first category obtaining three different problems. We name them black box
model explanation problem, black box outcome explanation problem, and black
box inspection problem. We name the second category transparent box design
problem. All these problems can be formalized as specific cases of the general
classification problems with the common target of providing an interpretable
and accurate predictive model. Details of the formalization are provided in the
following sections. Other important variants are generally not treated in the literature making the problem of discovering an explanation increasingly difficult:
(i) Is it allowed to query the black box at will to obtain new decision examples,
or only a fixed dataset of decision records is available? (ii) Is the complete set of
features used by the decision model known, or instead only part of these features
is known? In this survey we do not address these issues as in the literature there
is not sufficient material.
4.1
Problem Formulation
In the following, we generalize the classification problem (see Figure 3).
A predictor, also named model or classifier, is a function b : X m → Y where
m
X is the feature space with m corresponding to the number of features, and
Y is the target space. The feature space X can correspond to any basic data
type like the set of integers X = Im , reals X = Rm , booleans X = {0, 1}m , and
strings X = S m , where S = Σ ∗ and Σ = {a, b, c, . . . , } is the alphabet (a finite
non-empty set of symbols). The feature space X can also be a complex data type
composed of different basic data type. For example, X = I × R2 × S contains an
integer feature, two real features and a string feature. On the other hand, the
target space Y (with dimensionality equals to one) contains the different labels
(classes or outcomes) and identifies a semantic concept where Y can be a set of
booleans, integers or strings.
A predictor b is the output of a learner function Lb such that Lb : (X n×m ×
n
Y ) → (X m → Y). The learner Lb takes as input a dataset D = {X, Y } with
n samples where X ∈ X n×m and Y ∈ Y n and returns the predictor b. Given a
data record in the feature space x ∈ X m , the predictor b can be employed to
predict the target value ŷ, i.e., b(x) = ŷ.
Typically, in supervised learning [95], a training dataset Dtrain is used for
training the learner Lb (Dtrain ) which builds the predictor b, and a test dataset
Dtest is used for evaluating the performance of b. Given Dtest = {X, Y }, the
Open The Black Box Survey
13
evaluation is performed by observing for each couple of data record and target
value (x, y) ∈ Dtest the number of correspondences between y and b(x) = ŷ.
In the following we indicate with b a black box predictor belonging to the
set of uninterpretable data mining and machine learning models. According to
Section 3, b is a black box because the reasoning behind the function is not
understandable by humans and the outcome returned does not provide any clue
for its choice. In real-world applications, b is an opaque classifier resulting from
a learning Lb . Similarly, we indicate with c a comprehensible predictor for which
is available a global or a local explanation.
The performance of the comprehensible predictor c is generally evaluated by
two measures. The accuracy is used to evaluate how good are the performance of
both the black box predictor b and the comprehensible predictor c. The fidelity
is employed to evaluate how good is the comprehensible predictor c in mimicking
the black box predictor b. Indeed, given a data set D = {X, Y } we can apply to
each record x ∈ X
S both the predictors: (i) for the black box b we get the set of
predictions Ŷ = x∈X b(x),Swhile (ii) for the comprehensible predictor c we get
the set of predictions Ȳ = x∈X c(x).
Thus, we can evaluate the accuracy of the black box b and of the comprehensible predictor c by comparing the real target values Y against the predicted
target values Ŷ , and Ȳ with accuracy(Ŷ , Y ) and accuracy(Ȳ , Y ), respectively.
Moreover, we can evaluate the behavior of the predictor c with respect to b evaluating the fidelity of c by means of the function fidelity(Ŷ , Ȳ ). Note that the
fidelity score can be calculated by applying the same calculus of the accuracy
function where as target value is used the prediction Ȳ of the black box b instead
of the real values Y .
Black Box Model Explanation
Given a black box model solving a classification problem, the black box explanation problem consists in providing an interpretable and transparent model which
is able to mimic the behavior of the black box and which is also understandable
by humans (see Figure 4). In other words, the interpretable model approximating the black box must be globally interpretable. As consequence, we define the
black box model explanation problem as follows:
Definition 1 (Black Box Model Explanation). Given a black box predictor
b and a dataset D = {X, Y }, the black box model explanation problem consists
in finding a function f : (X m → Y) × (X n×m × Y n ) → (X m → Y) which takes
as input a black box b and a dataset D, and returns a comprehensible global
predictor cg , i.e., f (b, D) = cg , such that cg is able to mimic the behavior of b,
and exists a global explanator function εg : (X m → Y) → E that can derive from
cg a set of explanations E ∈ E modeling in a human understandable way the
logic behind cg , i.e., εg (cg ) = E.
A large set of the papers reviewed in this survey describe various designs for
the function f to solve the black box explanation problem. The set of explanations E can be modeled for example by a decision tree or by a set of rules [36],
14
R. Guidotti et al.
Fig. 4. Black Box Model Explanation Problem.
Fig. 5. Black Box Outcome Explanation Problem.
while the comprehensible global predictor cg is the predictor returning as global
explanation εg the decision tree or the set of rules.
Black Box Outcome Explanation
Given a black box model solving a classification problem, the black box outcome
explanation problem consists in providing an interpretable outcome, that is a
method for providing an explanation for the outcome of the black box. In other
words, the interpretable model must return the prediction together with an explanation about the reasons for that prediction, i.e., the prediction is only locally
interpretable. It is not required to explain the whole logic behind the black box
but only the reasons for the choice of a particular instance. Consequently, we
define the black box outcome explanation problem as:
Definition 2 (Black Box Outcome Explanation). Given a black box predictor b and a dataset D = {X, Y }, the black box outcome explanation problem
consists in finding a function f : (X m → Y) × (X n×m × Y n ) → (X m → Y) which
takes as input a black box b and a dataset D, and returns a comprehensible local
predictor cl , i.e., f (b, D) = cl , such that cl is able to mimic the behavior of b, and
exists a local explanator function εl : ((X m → Y) × (X m → Y) × X m ) → E which
takes as input the black box b, the comprehensible local predictor cl , and a data
record x with features in X m , and returns a human understandable explanation
e ∈ E for the data record x, i.e., εl (b, cl , x) = e.
We report in this survey recent works describing very diversified approaches
to implement function f , overcoming the limitations of explaining the whole
model (illustrated in Section 6). As an example, in this view of the problem, we
can consider that the explanation el may be either a path of a decision tree or
an association rule [27].
Open The Black Box Survey
15
Fig. 6. Black Box Inspection Problem.
Black Box Inspection Problem
Given a black box model solving a classification problem, the black box inspection
problem consists in providing a representation (visual or textual) for understanding either how the black box model works or why the black box returns certain
predictions more likely than others.
Definition 3 (Black Box Inspection Problem). Given a black box predictor
b and a dataset D = {X, Y }, the black box inspection problem consists in finding
a function f : (X → Y) × (X n × Y n ) → V which takes as input a black box b
and a dataset D, and returns a visual representation of the behavior of the black
box, f (b, D) = v with V being the set of all possible representations.
For example, the function f may be a technique based on sensitivity analysis
that, by observing the changing occurring in the predictions when varying the
input of b, returns a set of visualizations (e.g, partial dependence plots [48], or
variable effect characteristic curve [17]) highlighting the feature importance for
the predictions.
Transparent Box Design Problem
Given a classification problem the transparent box design problem consists in
providing a model which is locally or globally interpretable on its own.
Definition 4 (Transparent Box Design Problem). Given a dataset D =
{X, Y }, the transparent box design problem consists in finding a learning function Lc : (X n×m ×Y) → (X m → Y) which takes as input the dataset D = {X, Y }
and returns a (locally or globally) comprehensible predictor c, i.e., Lc (D) = c.
This implies that there exists a local explanator function εl or a global explanator
function εg (defined as before) that takes as input the comprehensible predictor c
and returns a human understandable explanation e ∈ E or a set of explanations
E.
For example, the functions Lc and c may be the decision tree learner and
predictor respectively, while the global explanator εg may return as explanation
a system for following the choices taken along the various branches of the tree,
and εl may return a textual representation of the path followed according to the
decision suggested by the predictor.
16
R. Guidotti et al.
Fig. 7. Transparent Box Design Problem.
Thus, according to our problem definitions, in this survey, when we say that
a method is able to open the black box, we are referring to one of the following
statements: (i) it explains the model, (ii) it explains the outcome, (iii) it can
inspect the black box internally, (iv) it provides a transparent solution.
5
Problem And Explanator Based Classification
In this survey, we propose a classification based on the type of problem faced
and on the explanator adopted to open the black box. In particular, in our
classification we take into account the following features:
–
–
–
–
the
the
the
the
type
type
type
type
of
of
of
of
problem faced (according to the definitions in Section 4);
explanator adopted to open the black box;
black box model that the explanator is able to open;
data used as input by the black box model.
In each section we group together all the papers that share the same problem
definition, while the subsections correspond to the different solutions adopted to
develop the explanators. In turn, in each subsection, we group the papers that
try to explain the same type of black box. Finally, we keep the type of data used
by the black box as a feature which is specified for each work analyzed.
We organize the sections discussing the different problems as follows. In Section 6 we analyze the papers presenting approaches to solve the black box model
explanation problem. These approaches provide a globally interpretable predictor which is able to mimic the black box. On the other hand, in Section 7 are
reviewed the methods solving the black box outcome explanation problem: the
predictor returned is locally interpretable and provides an explanation only for
a given record. In Section 8 we discuss the papers proposing methodologies for
inspecting black boxes, i.e., not providing a comprehensible predictor but a visualization tool for studying how the black box work internally, and what can
happen when a certain input is provided. Finally, in Section 9 we report the
papers designing a transparent predictor to overcome the “obscure” limitation
of black boxes. These approaches try to provide a global or local interpretable
model without sacrificing the accuracy of a black box learned to solve the same
task.
Open The Black Box Survey
17
For each of the sections above, we propose a further categorization with
respect to the type of explanator adopted. This categorization reflects on the
papers grouped into the various subsections:
– Decision Tree (DT) or Single Tree. It is commonly recognized that decision
tree is one of the more interpretable and easily understandable models, primarily for global, but also for local, explanations. Indeed, a very widespread
technique for opening the black box is the so-called “single tree approximation”.
– Decion Rules (DR) or Rule Based Explanator. Decision rules are among the
more human understandable techniques. There exist various types of rules
(illustrated in Section 3.3). They are used to explain the model, the outcome
and also for the transparent design. We remark the existence of techniques
for transforming a tree into a set of rules.
– Features Importance (FI). A very simple but effective solution acting as
either global or local explanation consists in returning as explanation the set
of features used by the black box together with their weight.
– Salient Mask (SM). An efficient way of pointing out what causes a certain outcome, especially when images or texts are treated, consists in using
“masks” visually highlighting the determining aspects of the record analyzed.
They are generally used to explain deep neural networks.
– Sensitivity Analysis (SA). It consists of evaluating the uncertainty in the
outcome of a black box with respect to different sources of uncertainty in its
inputs. It is generally used to develop visual tools for black box inspection.
– Partial Dependence Plot (PDP). These plots help in visualizing and understanding the relationship between the outcome of a black box and the input
in a reduced feature space.
– Prototype Selection (PS). This explanator consists in returning, together
with the outcome, an example very similar to the classified record, in order
to make clear which criteria the prediction was returned. A prototype is
an object that is representative of a set of similar instances and is part of
the observed points, or it is an artifact summarizing a subset of them with
similar characteristics.
– Neurons Activation (NA). The inspection of neural networks and deep neural
network can be carried out also by observing which are the fundamental
neurons activated with respect to particular input records.
In the following, we list all the black boxes opened in the reviewed papers.
These black boxes are all supervised learning algorithm designed to solve a classification problem [95].
– Neural Network (NN). Inspired by biological neural networks, artificial neural networks learn to do tasks by considering examples. A NN is formed by a
set of connected neurons. Each link between neurons can transmit a signal.
The receiving neuron can process the signal and then transmit to downstream neurons connected to it. Typically, neurons are organized in layers.
Different layers perform different transformations on their inputs. Signals
18
R. Guidotti et al.
travel from the input layer, to the output layer, passing through the hidden
layer(s) in the middle multiple times. Neurons and connections may also have
a weight that varies as learning proceeds, which can increase or decrease the
strength of the signal that it sends.
– Tree Ensemble (TE). Ensemble methods combine more than one learning
algorithm to improve the predictive power of any of the single learning algorithms that they combines. Random forests, boosted trees and tree bagging
are examples of TEs. They combine the predictions of different decision trees
each one trained on an independent subset of the input data.
– Support Vector Machine (SVM). Support Vector Machines utilize a subset of
the training data, called support vectors, to represent the decision boundary.
A SVN is a classifier that searches for hyperplanes with the largest margin
for the decision boundary.
– Deep Neural Network (DNN). A DNN is a NN that can model complex nonlinear relationship with multiple hidden layers. A DNN architecture is formed
by a composition of models expressed as a layered combination of basic units.
In DNNs the data typically flows from the input to the output layer without
looping back. The most used DNN are Recurrent Neural Networks (RNNs).
A peculiar component of RNNs are Long Short-Term Memory (LSTM) nodes
which are particularly effective for language modeling. On the other hand, in
image processing Convolutional Neural Networks (CNNs) are typically used.
Moreover, recently agnostic approaches for explaining black boxes are being
developed. An Agnostic Explanator (AGN) is a comprehensible predictor which
is not tied to a particular type of black box, explanation or data type. In other
words, in theory, an agnostic predictor can explain indifferently a neural network
or a tree ensemble using a single tree or a set of rules. Since only a few approaches
in the literature describe themselves to be fully agnostic, and since the principal
task is to explain a black box predictor, in this paper, if not differently specified,
we term agnostic the approaches defined to explain any type of black box.
The types of data used as input of black boxes analyzed in this survey are
the following:
– Tabular (TAB). With tabular data, we indicate any classical dataset in which
every record shares the same set of features and each feature is either numerical, categorical or boolean.
– Image (IMG). Many black boxes work with labeled images. These images
can be treated as they are by the black box or can be preprocessed (e.g,
re-sized in order to have all the same dimensions).
– Text (TXT). As language modeling is one of the tasks most widely assessed
nowadays together with image recognition, labeled datasets of text are generally used for tasks like spam detection or topic classification.
In data mining and machine learning many other types of data are also used
like sequences, networks, mobility trajectories, etc. However, they are not used
as input in the methods of the papers proposing a solution for opening the black
box.
Open The Black Box Survey
19
Fig. 8. Reverse Engineering approach: the learned black box predictor is queried with
a test dataset to produce an oracle which associate to each record a label which is not
real but assigned by the black box.
Table 1 lists the methods for opening and explaining black boxes and summarizes the various fundamental features and characteristics listed so far, together
with additional information that we believe could be useful for the reader. The
columns Examples, Code and Dataset indicates if any kind of example of explanation is shown in the paper, and if the source code and the dataset used in the
experiments are publicly available, respectively. The columns General and Random are discussed in the following section. We point out that Table 1 reports the
main references only, while existing extensions or derived works are discussed in
the survey. Table 2 reports the legend of Table 1, i.e., the expanded acronym and
the meaning of the features in Table 1. Moreover, in order to provide the reader
with a useful tool to find a particular set of papers with determined characteristics, Appendix A provides Tables 3, 4 and 5, in which are reported the list of
the papers with respect to each problem, explanator and black box, respectively.
Reverse Engineering: A Common Approach For Understanding The
Black Box
Before proceeding in the detailed analysis and classification of papers proposing
method f for understanding black boxes b, we present in this section the most
largely used approach to solve the black box model and outcome explanation
problems and the black box inspection problem. We refer to this approach as
reverse engineering because the black box predictor b is queried with a certain
test dataset in order to create an oracle dataset that in turn will be used to
train the comprehensible predictor (see Figure 8). The name reverse engineering
comes from the fact that we can only observe the input and output of the black
box.
With respect to the black box model and outcome explanation problems, the
possibility of action tied with this approach relies on the choice of adopting a
particular type of comprehensible predictor, and in the possibility of querying
the black box with input records created in a controlled way and/or by using
random perturbations of the initial train or test dataset. Regarding the random
perturbations of the input used to feed the black box, it is important to recall
that recent studies discovered that DNN built for classification problems on texts
and images can be easily fooled (see Section 2). Not human perceptible changes in
an image can lead a DNN to label the record as something else. Thus, according
20
R. Guidotti et al.
Trepan
DecText
GPDT
Tree Metrics
CCM
STA
CDT
TSP
Conj Rules
G-REX
REFNE
RxREN
SVM+P
inTrees
GoldenEye
PALM
CAM
Grad-CAM
LIME
MES
NID
GDP
IG
VEC
VIN
ICE
Prospector
Auditing
OPIA
TreeView
IP
CPAR
FRL
BRL
TLBR
IDS
Rule Set
1Rule
PS
BCM
OT-SpAMs
[20]
Craven et al.
[50]
Krishnan et al.
[9]
Boz
[39] Johansson et al.
[14]
Chipman et al.
[23] Domingos et al.
[29]
Gibbons et al.
[114]
Zhou et al.
[87]
Schetinin et al.
[32]
Hara et al.
[94]
Tan et al.
[19]
Craven et al.
[37] Johansson et al.
[115]
Zhou et al.
[6]
Augasta et al.
[70]
Nunez et al.
[28]
Fung et al.
[22]
Deng
[61]
Lou et al.
[33]
Henelius et al.
[51]
Krishnan et al.
[97]
Tolomei et al.
[108]
Xu et al.
[25]
Fong et al.
[113]
Zhou et al.
[89]
Selvaraju et al.
[56]
Lei et al.
[83]
Ribeiro et al.
[98]
Turner et al.
[71]
Olden et al.
[7]
Baehrens
[92]
Sundararajan
[16]
Cortez et al.
[35]
Hooker
[30] Goldstein et al.
[48]
Krause et al.
[2]
Adler et al.
[1]
Adebayo et al.
[110]
Yosinski et al.
[96] Thiagarajan et al.
[90]
Shwartz et al.
[81]
Radford
[109]
Yin et al.
[102]
Wang et al.
[57]
Letham et al.
[91]
Su et al.
[53] Lakkaraju et al.
[104]
Wang et al.
[64] Malioutov et al.
[8]
Bien et al.
[44]
Kim et al.
[63] Mahendran et al.
[47] Kononenko et al.
[103]
Wang et al.
1996
1999
2002
2009
1998
1998
2013
2016
2007
2016
2016
1994
2003
2003
2012
2002
2005
2014
2013
2014
2017
2017
2015
2017
2016
2016
2016
2016
2016
2002
2010
2017
2011
2004
2015
2016
2016
2016
2015
2016
2017
2017
2003
2015
2015
2015
2016
2016
2017
2011
2014
2015
2010
2015
Model Expl.
Model Expl.
Model Expl.
Model Expl.
Model Expl.
Model Expl.
Model Expl.
Model Expl.
Model Expl.
Model Expl.
Model Expl.
Model Expl.
Model Expl.
Model Expl.
Model Expl.
Model Expl.
Model Expl.
Model Expl.
Model Expl.
Model Expl.
Model Expl.
Model Expl.
Outcome Expl.
Outcome Expl.
Outcome Expl.
Outcome Expl.
Outcome Expl.
Outcome Expl.
Outcome Expl.
Inspection
Inspection
Inspection
Inspection
Inspection
Inspection
Inspection
Inspection
Inspection
Inspection
Inspection
Inspection
Inspection
Transp. Design
Transp. Design
Transp. Design
Transp. Design
Transp. Design
Transp. Design
Transp. Design
Transp. Design
Transp. Design
Transp. Design
Transp. Design
Transp. Design
DT
DT
DT
DT
DT
DT
DT
DT
DT
DT
DT
DR
DR
DR
DR
DR
DR
DR
FI
FI
DT
FI
SM
SM
SM
SM
SM
FI
DR
SA
SA
SA
SA
PDP
PDP
PDP
PDP
PDP
NA
DT
NA
NA
DR
DR
DR
DR
DR
DR
DR
PS
PS
PS
FI
DT
NN
NN
NN
NN
TE
TE
TE
TE
TE
TE
TE
NN
NN
NN
NN
SVM
SVM
TE
AGN
AGN
AGN
TE
DNN
DNN
DNN
DNN
DNN
AGN
AGN
NN
AGN
DNN
AGN
AGN
AGN
AGN
AGN
AGN
DNN
DNN
DNN
DNN
-
TAB
TAB
TAB
TAB
TAB
TAB
TAB
TAB
TAB
TAB
TAB
TAB
TAB
TAB
TAB
TAB
TAB
TAB
TAB
TAB
ANY
TAB
IMG
IMG
IMG
IMG
TXT
ANY
ANY
TAB
TAB
ANY
TAB
TAB
TAB
TAB
TAB
TAB
IMG
TAB
TAB
TXT
TAB
TAB
TAB
TAB
TAB
TAB
TAB
ANY
ANY
IMG
TAB
TAB
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
at
as
e
t
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
D
de
Co
pl
es
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
Ex
am
G
en
er
al
R
an
do
m
Ty
pe
D
at
a
Bo
x
Bl
ac
k
Ex
pl
an
at
or
lem
Pr
ob
r
Ye
a
ho
rs
A
ut
ef
.
R
N
am
e
Table 1. Summary of methods for opening and explaining black boxes.
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
Open The Black Box Survey
21
Table 2. Legend of Table 1. In the following are described the features reported and
the abbreviations adopted.
Feature
Problem
Description
Model Explanation, Outcome Explanation, Black Box Inspection, Transparent
Design
Explanator DT - Decision Tree, DR - Decision Rules, FI - Features Importance, SM Saliency Masks, SA - Sensitivity Analysis, PDP - Partial Dependence Plot,
NA - Neurons Activation, PS - Prototype Selection
Black Box NN - Neural Network, TE - Tree Ensemble, SVM - Support Vector Machines,
DNN - Deep Neural Network, AGN - AGNostic black box
Data Type TAB - TABular, IMG - IMaGe, TXT - TeXT, ANY - ANY type of data
General Indicates if an explanatory approach can be generalized for every black box,
i.e., it does not consider peculiarities of the black box to produce the explanation
Random Indicates if any kind of random perturbation or permutation of the original
dataset is required for the explanation
Examples Indicates if example of explanations are shown in the paper
Code
Indicates if the source code is available
Dataset Indicates if the datasets used in the experiments are available
to these discoveries, the methods treating images or text, in theory, should not
be enabled to use completely random perturbations of their input. However, this
is not always the case in practice [84].
Such reverse engineering approach can be classified as generalizable or not
(or pedagocial vs. decompsitional as described in [65]). We say that an approach
is generalizable when a purely reverse engineering procedure is followed, i.e.,
the black box is only queried with different input records to obtain an oracle
used for learning the comprehensible predictor (see Figure 9-(left)). In other
words, internal peculiarities of the black box are not exploited to build the
comprehensible predictor. Thus, if an approach is generalizable, even though it
is presented to explain a particular type of black box, in reality, it can be used
to interpret any kind of black box predictor. That is, it is an agnostic approach
for interpreting black boxes. On the other hand, we say that an approach is not
generalizable if it can be used to open only that particular type of black box
for which it was designed for (see Figure 9-(right)). For example, if an approach
is designed to interpret random forest and internally use a concept of distance
between trees, then such approach can not be utilized to explain predictions of
a NN. A not generalizable approach can not be black box agnostic.
In Table 1 we keep track of these aspects with the two features General and
Random. With General we indicate if an explanatory approach can be generalized for every black box, while with Random we indicate if any kind of random
perturbation or permutation of the original dataset is used by the explanatory
approach.
In light of these concepts, as the reader will discover below, a further classification not explicitly indicated emerges from the analysis of these papers. This
22
R. Guidotti et al.
Fig. 9. (Left) Generalizable reverse engineering approach: internal peculiarities of the
black box are not exploited to build the comprehensible predictor. (Right) Not Generalizable reverse engineering approach: the comprehensible predictor is the result of a
procedure involving internal characteristics of the black box.
fact can be at the same time a strong point or a weakness of the current state
of the art. Indeed, we highlight that the works for opening the black box are
realized for two cases. The first (larger) group contains approaches proposed to
tackle a particular problem (e.g., medical cases) or to explain a particular type
of black box, that is, the solutions are specific for the problem instance. The
second group contains general purpose solutions that try to be general as much
as possible and propose agnostic and generalizable solutions.
6
Solving the Black Box Model Explanation Problem
In this section we review the methods for opening the black box facing the black
box model explanation problem (see Section 4.1). That is, the proposed methods provide globally interpretable models which are able to mimic the behavior
of black boxes and which are also understandable by humans. We recognized
different groups of approaches. In Section 6.1 we analyze the proposals using
a decision tree as explanator, while in Section 6.2 they use rules. Section 6.3
describes the methods which are designed to work with any type of black box.
Finally, Section 6.4 contains the remaining ones.
6.1
Explanation via Single Tree Approximation
In this section we present a set of works addressing the black box model explanation problem by implementing in different ways the function f . All the following
works adopt a decision tree as comprehensible global predictor cg , and consequently represent the explanation εg with the decision tree itself. Moreover, we
point out that all the methods presented in this section work on tabular data.
Explanation of Neural Networks. The following papers describe the implementation of functions f which are able to interpret a black box b consisting in
a Neural Network (NN) [95] with a comprehensible global predictor cg consisting in a decision tree. In these works, the NNs are considered black-boxes, i.e.,
Open The Black Box Survey
23
the only interface permitted is presenting an input x to the neural network b
and obtaining the outcome ŷ. The final goal is to comprehend how the neural
networks behave by submitting to it a large set of instances and analyzing their
different predictions.
Single tree approximations for NNs were first presented in 1996 by Craven et
al. [20]. The comprehensible representations of the neural network b is returned
by Trepan which is the implementation of function f . Trepan queries the neural
network b to induce a decision tree cg approximating the concepts represented
by the networks by maximizing the gain ratio [95] together with an estimation of
the current model fidelity. Another advantage of Trepan with respect to common
tree classifiers like ID3 or C4.5 [95] is that, thanks to the black box b, it can use
as many instances as desired for each split, so that also the node splits near to
the bottom of the tree are realized using a considerable amount of training data.
In [50], Krishnan et al. present a three step method f . The first step generates
a sort of “prototype” for each target class in Y by using genetic programming
to query the trained neural network b. The input features dataset X is exploited
for constraining the prototypes. The second step selects the best prototypes for
inducing the learning of the decision tree cg in the third step. This approach leads
to get more understandable and smaller decision trees starting from smaller data
sets.
In [9], Boz describes DecText, another procedure that uses a decision tree c
to explain neural network black boxes b. The overall procedure recalls Trepan
[20] with the innovation of four splitting methods aimed at finding the most
relevant features during the tree construction. Moreover, since one of the main
purposes of the tree is to maximize the fidelity while keeping the model simple,
a fidelity pruning strategy to reduce the tree size is defined. A set of random
instances are generated. Then, starting from the bottom of the tree, for each
internal node a leaf is created with the majority label using the labeling of the
random instances. If the fidelity of the new tree overtakes the old one, than the
maximum fidelity and the tree are updated.
In [39] Johansson et al. use Genetic Programming to evolve Decision Trees
(the comprehensible global predictor cg ), in order to mimic the behavior of a
neural network ensemble (the black box b). The dataset D used by genetic programming (implementing function f ) consists of a lot of different combinations
of the original data and oracle data labeled by b. The paper shows that trees
based only on original training data have the worst performance in terms of
accuracy in the test data, while the trees evolved using both the oracle guide
and the original data produce significantly more accurate trees cg .
We underline that, even though these approaches are developed to explain
neural networks, since peculiarities of the neural networks are not used by f ,
which uses b only as an oracle, these approaches can be potentially adopted as
agnostic explanators, i.e., they can be used to open any kind of black box and
represent it with a single tree.
24
R. Guidotti et al.
Explanation of Tree Ensembles. Richer collections of trees provide higher
performance and less uncertainty in the prediction. On the other hand, it is
generally difficult to make sense of the resultant forests. The papers presented in
this section describe functions f for approximating a black box model b consisting
in Tree Ensembles (TE) [95] (e.g. random forests) with a global comprehensible
predictor cg in the form of a decision tree, and explanation εg as a the decision
tree as before.
Unlike previous works, the tree ensembles are not only viewed as black boxes,
but also some of their internal features are used to derive the global comprehensible model cg . For example, Chipman et al., in [14] observe that although
hundreds of distinct trees are identified by random forests, in practice, many
of them generally differ only by few nodes. In addition, some trees may differ
only in the topology, but use the same partitioning of the feature space X . The
paper proposes several measures of dissimilarity for trees. Such measures are
used to summarize forest of trees through clustering, and finally use archetypes
of the associated clusters as model explanation. Here, f corresponds to the clustering procedure, and the global comprehensible predictor cg is the set of tree
archetypes minimizing the distance among all the trees in each cluster. In this
approach, f does not extend the input dataset D with random data.
On the other hand, random data enrichment and model combination are the
basis for the Combined Multiple Model (CCM) procedure f presented in [23].
Given the tree ensemble black box b, it first modifies n times the input dataset D
and learns a set of n black boxes bi ∀i = 1, . . . , n, and then it randomly generates
data record x which are labeled using a combination (e.g. bagging) of the n black
boxes bi , i.e., Cb1 ,...,bn (x) = ŷ. In this way, the training dataset D = D ∪ {x, ŷ}
is increased. Finally, it builds the global comprehensible model cg as a decision
tree (C4.5 [79]) on the enriched dataset D. Since it is not exploiting particular
features of the tree ensemble b, also this approach can be generalized with respect
to the black box b. In line with [23], the authors of [29] generate a very large
artificial dataset D using the prediction of the random forest b, then explain
b by training a decision tree cg on this artificial dataset in order to mime the
behavior of the random forest. Finally, they improve the comprehensibility of cg
by cutting the decision tree with respect to a human understandable depth (i.e.,
from 6 to 11 nodes of depth). [114] proposes Single Tree Approximation (STA),
an extension of [29] which empowers the construction of the final decision tree
cg by using test hypothesis to understand which are the best splits observing
the Gini indexes on the trees of the random forest b.
Schetinin et al. in [87] present an approach for the probabilistic interpretation
of the black box b Bayesian decision trees ensembles [10] through a quantitative
evaluation of uncertainty of a Confident Decision Tree (CDT) cg . The methodology f for interpreting b is summarized as follows: (i) the classification confidence
for each tree in the ensemble is calculated using the training data D, (ii) the
decision tree cg that covers the maximal number of correct training examples
is selected, keeping minimal the amount of misclassifications on the remaining
examples by sub-sequentially refining the training dataset D. Similarly to [14],
Open The Black Box Survey
25
also this explanation method f does not extend the input dataset D with random data and cannot be generalized to other black boxes but can be used only
with Bayesian decision tree ensembles.
In [32], Hara et al. reinterpret Additive Tree Models (ATM) (the black box b)
using a probabilistic generative model interpretable by humans. An interpretable
ATM has a sufficiently small number of regions. Therefore, their aim is to reduce
the number of regions in an ATM while minimizing the model error. To satisfy
these requirements, they propose a post processing method f that works as
follows. First, it learns an ATM b generating a number of regions. Then, it mimics
b using a simpler model (the comprehensible global predictor cg ) where the
number of regions is fixed as small, e.g., ten. In particular, to obtain the simpler
model an Expectation Maximization algorithm is adopted [95] minimizing the
Kullback-Leibler divergence from the ensemble b.
The authors of [94] propose Tree Space Prototype (TSP), an approach f for
interpretating tree ensembles (the black box b) by finding tree prototypes (the
comprehensible global predictor cg ) in the tree space. The main contributions for
f are: (i) the definition of the random forest proximity between trees, and (ii)
the design of the procedure to extract the tree prototypes used for classification.
6.2
Explanation via Rule Extraction
Another commonly used state of the art interpretable and easily understandable
model is the set of rules. When a set of rules describing the logic behind the
black box model is returned the interpretability is provided at a global level.
In the following, we present a set of reference works solving the black box model
explanation problem by implementing in different ways function f , and by adopting any kind of decision rules as comprehensible global predictor cg . Hence, the
global explanation εg change accordingly to the type of rules extracted by cg .
Similarly to the previous section, also all the methods presented in this section
work on tabular data.
Explanation of Neural Networks. The following papers describe the implementation of functions f which are able to interpret a black box b consisting in
a Neural Network (NN) [95]. In the literature already exists a survey specialized
on techniques extracting rules from neural networks [5]. It provides an overview
of mechanisms designed to (i) insert knowledge into neural networks (knowledge
initialization), (ii) extract rules from trained NNs (rule extraction), and (iii) use
NNs to refine existing rules (rule refinement). The approaches presented in [5]
are strongly dependent on the black box b and on the specific type of decision
rules cg . Thus, they are not generalizable and can not be employed to solve other
instances of the black box model explanation problem. The survey [5] classifies
the methods according to the following criteria:
– Expressive power of the extracted rules.
– Translucency: that is decompositional, pedagogical and eclectic properties.
– Portability of the rule extraction technique.
26
R. Guidotti et al.
– Quality of the rules extracted. Quality includes accuracy, fidelity, consistency,
i.e., different training of the NN extract the rules that lead to the same
classification of unseen examples.
– Algorithmic complexity.
A typical paper analyzed in [5] is [19] where Craven et al. present a method
f to explain the behavior of a neural network b by transforming rule extraction
(which is a search problem) into a learning problem. The original training data
D and a randomized extension of it are provided as input to the black box b.
If the input x ∈ D with outcome ŷ is not covered by the set of rules, then a
conjunctive (or m-of-n) rule is formed from {x, ŷ} considering all the possible
antecedents. The procedure ends when all the target classes have been processed.
In [40] Johansson et al. exploit G-REX [37], an algorithm for rule extraction, as function f to explain a neural network b. They use the classical reverse
engineering schema where random permutations of the original dataset D are
annotated by b, and such dataset is used as input by G-REX, which corresponds
with cg in this case. In particular, G-REX extracts rules by exploiting genetic
programming as a key concept. In subsequent works, the authors show that the
proposed methodology f can be also employed to interpret trees ensembles. [38]
extends G-REX for handling regression problems by generating regression trees,
and classification problems by generating fuzzy rules.
In [115] the authors present REFNE, an approach f to explain neural network ensembles b. REFNE uses ensembles for generating instances and then, extracts symbolic rules cg from those instances. REFNE avoids useless discretizations of continuous attributes, by applying a particular discretization leading
to discretize different continuous attributes using different intervals. Moreover,
REFNE can also be used as a rule learning approach, i.e., it solves the transparent box design problem (see Section 4.1). Also in [6] Augasta et al. propose
RxREN a rule extraction algorithm cg which returns the explanation of a trained
NN b. The method f works as follows. First, it prunes the insignificant input
neurons from trained NNs and identifies the data range necessary to classify
the given test instance with a specific class. Second, using a reverse engineering
technique, through RxREN generates the classification rules for each class label
exploiting the data ranges previously identified, and improve the initial set of
rules by a process that prunes and updates the rules.
Explanation of Support Vector Machines. The following papers show implementations of functions f for explaining a black box b consisting in a Support
Vector Machine (SVM) [95] still returning a comprehensible global predictor cg
consisting in a set of decision rules.
The authors of [70] propose the SVM+Prototypes (SVM+P) procedure f for
rule extraction cg from support vector machines b. It works as follows: it first
determines the decision function by means of a SVM, then a clustering algorithm
is used to find out a prototype vector for each class. By using geometric methods,
these points are joined with the support vectors for defining ellipsoids in the input
space that can be transformed into if-then rules.
Open The Black Box Survey
27
Fig. 10. From [65]: pedagogical (a) and decompositional (b) rule extraction techniques.
Fung et al., in [28], describe as function f an algorithm based on constraint
programming for converting linear SVM b (and other hyperplane-based linear
classifiers) into a set of non overlapping and interpretable rules cg . These rules
are asymptotically equivalent to the original linear SVM. Each iteration of the
algorithm for extracting the rules is designed to solve a constrained optimization
problem having a low computational cost. We underline that this black box
explanation solution f is not generalizable and can be employed only for Linear
SVM-like black boxes.
In [65] the authors propose a qualitative comparison of the explanations returned by techniques for extraction of rules from SVM black boxes (e.g. SVM+P
[70], Fung method [28]) against the redefining of methods designed for explaining neural networks, i.e., C4.5 [95], Trepan [20] and G-REX [37]. How we anticipated in the previous section, the authors delineate the existence of two type of
approaches to extract rules: pedagogical and decompositional. Pedagogical techniques f directly extract rules which relate the inputs and outputs of the predictor (e.g. [20,37]), while decompositional approaches are closely intertwined with
the internal structure of the SVM (e.g. [70,28]). We recall that, in Table 1 we
identify with the term generalizable the pedagogical approaches.
Explanation of Tree Ensembles. Finally, in [22], Deng proposes the inTrees
framework f to explain black boxes b defined as Tree Ensembles (TE) by returning a set of decision rules cg . In particular, InTrees extracts, measures, prunes and
selects rules from tree ensembles, and calculates frequent variable interactions.
The set of black boxes b that inTrees can explain is represented by any kind of
tree ensemble like random forests, regularized random forests and boosted trees.
InTrees framework can be used for both classification and regression problems.
The technique described by InTrees is also known as Simplified Tree Ensamble
Learner (STEL): it extracts the most supported and simplest rules form the
trees ensemble.
28
6.3
R. Guidotti et al.
Agnostic Explanator
Recent approaches for interpretation are agnostic (AGN) with respect to the
black box to be explained. In this section, we present a set of works solving the
black box model explanation problem by implementing function f such that any
type of black box b can be explained. These approaches do not return a specific
comprehensible global predictor cg , thus the type of explanation εg change with
respect to f and cg . By definition all these approaches are generalizable.
Probably the first attempt of an agnostic solution was proposed in [60]. Lou
et al. propose a method f which exploits Generalized Additive Models (GAMs)
and it is able to interpret regression splines (linear and logistics), single trees and
tree ensembles (bagged trees, boosted trees, boosted bagged trees and random
forests). GAMs are presented as the gold standard for intelligibility when only
univariate terms are considered. Indeed, the explanation εc is returned as the
importance of the contribution of the individual features in b together with their
shape function, such that the impact of each predictor can be quantified. A shape
function is the plot of a function capturing the linearities and nonlinearities
together with its shape. It works on tabular data. A refinement of the GAM
approach is proposed by the same authors in [61]. A case study on health care
showing the application of the GAM the refinement is presented in [13]. In
particular, this approach is used for the prediction of the pneumonia risk and
hospital 30-day readmission.
In [33] the authors present an iterative algorithm f that allows finding features and dependencies exploited by a classifier when producing a prediction. The
attributes and the dependencies among the grouped attributes depict the global
explanation εg . The proposed approach f named GoldenEye is based on tabular
data randomization (within class permutation, dataset permutation, etc.) and
on grouping attributes with interactions have an impact on the predictive power.
In [51], PALM is presented (Partition Aware Local Model ) to implement
f . In particular, PALM is a method that is able to learn and summarize the
structure of the training dataset to help the machine learning debugging. PALM
mimes a black box b using a meta-model for partitioning the training dataset,
and a set of sub-models for approximating and miming the patterns within each
partition. As meta-model it uses a decision tree (cg ) so that the user can examine
its structure and determine if the rules detected follow the intuition or not, and
link efficiently problematic test records to the responsible train data. The submodels linked to the leaves of the tree can be a arbitrarily complex model able
to catch elaborate local patterns, but yet interpretable by humans. Thus, with
respect to the final sub-models PALM is not only black box agnostic but also
explanator agnostic. Moreover, PALM is also data agnostic; i.e., it can work on
any kind of data.
6.4
Explanation via Other Approaches
In [97] a solution for the black box model explanation problem is presented. It
adopts an approach that can not be classified as one of the previous. The proposed approach f uses the internals of a random forest model b to produce
Open The Black Box Survey
29
recommendations on the transformation of true negative examples into positively predicted examples. These recommendations, which are strictly related to
the feature importance, corresponds to the comprehensible global predictor cg .
In particular, the function f aims at transforming a negative instance into a
positive instance by analyzing the path on the trees in the forest predicting such
instance as positive or negative. The explanation of b is provided by means of
the helpfulness of the features in the paths adopted for changing the instance
outcome from negative to positive.
7
Solving the Black Box Outcome Explanation Problem
In this section we review the methods solving the black box outcome explanation
problem (see Section 4.1). These methods provide a locally interpretable model
which is able to explain the prediction of the black box in understandable terms
for humans. This category of approaches using a local point of view with respect
to the prediction is becoming the most studied in the last years. Section 7.1
describes the methods providing the salient parts of the record for which a
prediction is required using Deep Neural Networks (DNNs), while Section 7.2
analyzes the methods which are able to provide a local explanation for any type
of black box.
7.1
Explanation of Deep Neural Network via Saliency Masks
In the following works the opened black box b is a DNN and the explanation
is provided by using a Saliency Mask (SM), i.e. a subset of the original record
which is mainly responsible for the prediction. For example, as salient mask we
can consider the part of an image or a sentence in a text. A saliency image
summarizes where a DNN looks into an image for recognizing their predictions.
The function f to extract the local explanation εl is always not generalizable
and often strictly tied with the particular type of network, i.e., convolutional,
recursive, etc.
The work [108] introduces an attention based model f which automatically
identifies the contents of an image. The black box is a neural network which
consists of a combination of a Convolutional NN (CNN) for the features extraction and a Recursive NN (RNN) containing Long Short Term Memory (LSTM),
nodes producing the image caption by generating a single word for each iteration. The explanation εl of the prediction is provided through a visualization of
the attention (area of an image, see Figure 11-left) for each word in the caption. A similar result is obtained by Fong et al. in [25]. In this work the authors
propose a framework f of explanations cl as meta-predictors. In their view, an
explanation εl , and thus a meta-predictor, is a rule that predicts the response
of a black box b to certain inputs. Moreover, they propose to use saliency maps
as explanations for black boxes to highlight the salient part of the images (see
Figure 11-right).
30
R. Guidotti et al.
Fig. 11. Saliency Masks for explanation of deep neural network. (Left) From [108] the
elements of the image highlighted. (Right) From [25] the mask and the level of accuracy
on the image considering and not considering the learned mask.
Similarly, another set of works produce saliency masks incorporating network
activations into their visualizations. This kind of approaches f are named Class
Activation Mapping (CAM). In [113], global average pooling in CNN (the black
box b) is used for generating the CAM. A CAM (the local explanation εl ) for a
particular outcome label indicates the discriminative active region that identifies
that label. [89] defines its relaxed generalization Grad-CAM which visualizes
the linear combination of a late layer’s activations and label-specific weights
(or gradients for [113]). All these approaches arbitrarily invoke different back
propagation and/or activation, which results in aesthetically pleasing, heuristic
explanations of image saliency. Their solution is not black box agnostic limited
to NN, but it requires specific architectural modifications [113] or access to
intermediate layers [89].
With respect to texts, in [56] the authors develop an approach f which incorporates rationales as part of the learning process of b. A rationale is a simple
subset of words representing a short and coherent piece of text (e.g., phrases),
and alone must be sufficient for the prediction of the original text. A rational
is the local explanator εl and provides the saliency of the text analyzed, i.e.,
indicates the reason for a certain outcome.
7.2
Agnostic Explanator
In this section we present the agnostic solutions proposed for the black box
outcome explanation problem implementing function f such that any type of
black box b can be explained. All these approaches are generalizable by definition
and return a comprehensible local predictor cl . Thus, they can be employed for
diversified data types.
In [84], Ribeiro et al. present the Local Interpretable Model-agnostic Explanations (LIME) approach f which does not depend on the type of data, nor on
the type of black box b to be opened, nor on a particular type of comprehensible local predictor cl or explanation εl . In other words, LIME can return an
understandable explanation for the prediction obtained by any black box. The
main intuition of LIME is that the explanation may be derived locally from the
records generated randomly in the neighborhood of the record to be explained,
and weighted according to their proximity to it. In their experiments, the authors
adopt linear models as comprehensible local predictor cl returning the importance of the features as explanation εl . As black box b the following classifiers are
Open The Black Box Survey
31
tested: decision trees, logistic regression, nearest neighbors, SVM and random
forest. A weak point of this approach is the required transformation of any type
of data in a binary format which is claimed to be human interpretable. [83] and
[82] propose extensions of LIME with an analysis of particular aspects and cases.
A similar approach is presented in [98], where Turner et al. design the Model
Explanation System (MES) f that augments black box predictions with explanations by using a Monte Carlo algorithm. In practice, they derive a scoring
system for finding the best explanation based on formal requirements and consider that the explanations εl are simple logical statements, i.e., decision rules.
The authors test logistic regression and SVMs as black box b.
8
Solving the Black Box Inspection Problem
In this section we review the methods for opening the black box facing the black
box inspection problem (see Section 4.1). Given a black box solving a classification problem, the inspection problem consists in providing a representation
for understanding either how the black box model works or why the black box
returns certain predictions more likely than others. In [88], Seifet et al. provide
a survey of visualizations of DNNs by defining a classification scheme describing
visualization goals and methods. They found that most papers use pixel displays to show neuron activations. As in the previous sections, in the following
we propose a classification based on the type of technique f used to provide the
visual explanation of how the black box works. Most papers in this section try
to inspect NNs and DNNs.
8.1
Inspection via Sensitivity Analysis
In this section we review the works solving the black box inspection problem by
implementing function f using Sensitivity Analysis (SA). Sensitivity analysis
studies the correlation between the uncertainty in the output of a predictor and
that one in its inputs [86]. All the following methods work on tabular datasets.
Sensitivity analysis for “illuminating” the black box was first proposed by
Olden in [71] where a visual method for understanding the mechanism of NN
is described. In particular, they propose to assess the importance of axon connections and the contribution of input variables by means of sensitivity analysis
and Neural Interpretation Diagram (NID) to remove not significant connections
and improve the network interpretability.
In [7] the authors propose a procedure based on Gaussian Process Classification (GDP) which allows explaining the decisions of any classification method
through an explanation vector. That is, the procedure f is black box agnostic.
The explanation vectors are visualized to highlight the features that were most
influential for the decision of a particular instance. Thus, we are dealing with an
inspection for outcome explanation εl .
In [21], Datta et al. introduce a set of Quantitative Input Influence (QII)
measures f capturing how much inputs influence the outputs of black box predictors. These measures provide a foundation for transparency reports of black
32
R. Guidotti et al.
Fig. 12. (Left). From [16] VEC curve and histogram for the pH input feature (x-axis)
and the respective high quality wine probability outcome (left of y-axis) and frequency
(right of y-axis). (Right). From [48] Age at enrollment shown as line plot (top) and
partial dependence bar (middle). Color denotes the predicted risk of the outcome.
box predictors. In practice, the output consists in the feature importance for
outcome predictions.
[92] studies the problem of attributing the prediction of a DNN (the black
box b) to its input features. Two fundamental axioms are identified: sensitivity
and implementation invariance. These axioms guide the design of an attribution
method f , called Integrated Gradients (IG), that requires no modification to the
original network. Differently from the previous work, this approach is tested on
different types of data.
Finally, Cortez in [16,17] uses sensitivity analysis based and visualization
techniques f to explain black boxes b. The sensitivity measures are variables
calculated as the range, gradient, variance of the prediction. Then, the visualizations realized are barplots for the features importance, and Variable Effect
Characteristic curve (VEC) [18] plotting the input values (x-axis) versus the
(average) outcome responses (see Figure 12 - (left)).
8.2
Inspection via Partial Dependence
In this section we report a set of approaches solving the black box inspection
problem by implementing a function f which returns a Partial Dependence Plot
(PDP). Partial dependence plot is a tool for visualizing the relationship between
the response variable and predictor variables in a reduced feature space. All the
approaches presented in this section are black box agnostic and are tested on
tabular datasets.
In [35], the authors present an approach f aimed at evaluating the importance of non-additive interactions between any set of features. The implementa-
Open The Black Box Survey
33
tion uses the Variable Interaction Network (VIN) visualization generated from
the use of ANOVA statistical methodology (a technique to calculate partial dependence plots). VIN allows to visualize the importance of the features together
with their interdependences. Goldstein et al. provide in [30] a technique f which
extends classical PDP named Individual Conditional Expectation (ICE) to visualize the model approximated by a black box b that help in visualizing the
average partial relationship between the outcome and some features. ICE plots
improves PDP by highlighting the variation in the fitted values across the range
of covariates. In [48], Krause et al. introduce random perturbations on the black
box b input values to understand to which extent every feature impact the prediction through a visual inspection using the PDPs f . The main idea of Prospector
is to observe how the output varies by varying the input changing one variable at
a time. It provides an effective way to understand which are the most important
features for a certain prediction εl so that it can help in providing a valuable
interpretation (see Figure 12 - (right)). In [2] the authors propose a method f
for auditing (i.e., inspecting) black box predictors b, studying to which extent
existing models benefit of specific features in the data. This method does not
assume any knowledge on the models behavior. In particular, the method f focuses on indirect influence and visualizes the global inspection εg through an
obscurity vs. accuracy plot (the features are obscured one after the other). Yet,
the dependence of a black box b on its input features is relatively quantified by
the procedure f proposed in [1], where the authors present an iterative procedure based on Orthogonal Projection of Input Attributes (OPIA), for enabling
the interpretability of black box predictors.
8.3
Inspection via Other Approaches
In the following, we present solutions for the black box inspection problem that
adopt an approach f which can be categorized as none of the previous ones.
They all refer to DNNs as black box b and are not generalizable.
[110] proposes two tools for visualizing and interpreting DNNs and for understanding what computations DNNs perform at intermediate layers and which
neurons activate. These tools visualize the activations of each layer of a trained
CNN during the process of images or videos. Moreover, they visualize the features of the different layers by regularized optimization in image space. Yosinski
et al. found that by analyzing the live activations, changing in correspondence
of different inputs, helps to generate an explanation on the DNNs behave. [96]
shows the extraction of a visual interpretation of a DNN using a decision tree.
The method TreeView f works as follows. Given the black box b as a DNN, it
first decomposes the feature space into K (user defined) potentially overlapping
factors. Then, it builds a meta feature for each of the K clusters and a random
forest that predicts the cluster labels. Finally, it generates and shows a surrogate
decision tree from the forest as an approximation of the black box.
Shwartz-Ziv et al. in [90] showed the effectiveness of the Information Plane
f visualization of DNNs highlighting that the empirical error minimization of
34
R. Guidotti et al.
each stochastic gradient descent phase epoch is always followed by a slow representation compression.
Finally, it is worth mentioning that [81] presents the discovery that a single
neuron unit of a DNN can perform alone a sentiment analysis task after the
training of the network reaching the same level of performance of strong baselines. Also in [111], Zeiler et al. backtrack the network computations to identify
which image patches are responsible for certain neural activations.
9
Solving the Transparent Box Design Problem
In this section we review the approaches designed to solve the classification
problem using a transparent method which is locally or globally interpretable
on its own, i.e., solving the transparent box design problem (see Section 4.1).
9.1
Explanation via Rule Extraction
In this section we present the most relevant state of the art works solving the
transparent box design problem by means of comprehensible predictors c based
on rules. In these cases, cg is a comprehensible global predictor providing the
whole set of rules leading to any possible decision: a global explanator εg is made
available by cg . All the methods presented in this section work on tabular data.
In [109] the authors propose the approach f named CPAR (Classification
based on Predictive Association Rules) combining the positive aspects of both
associative classification and traditional rule-based classification. Indeed, following the basic idea of FOIL [80], CPAR does not generate a large set of candidates
as in associative classification, and applies a greedy approach for generating rules
cg directly from training data.
Wang and Rudin, in [102] propose a method f to extract falling rule lists
cg (see Section 3.3) instead of classical rules. The falling rule lists extraction
method f relies on a Bayesian framework.
In [57], the authors tackle the problem to build a system for medical scoring which is interpretable and characterized by high accuracy. To this end, they
propose Bayesian Rule Lists (BRL) f to extract the comprehensible global predictor cg as a decision list. A decision list consists of a series of if-then statements
discretizing the whole feature space into a set of simple and directly interpretable
decision statements.
A Bayesian approach is followed also in [91]. The authors propose algorithms
f for learning Two-Level Boolean Rules (TLBR) in Conjunctive Normal Form
or Disjunctive Normal Form cg . Two formulations are proposed. The first one
is an integer program whose objective function combines the total number of
errors and the total number of features used in the rule. The second formulation
replaces the 0-1 classification error with the Hamming distance from the current
two-level rule to the closest rule that correctly classifies a sample. In [54] the
authors propose a method f exploiting a two-level boolean rule predictor to
Open The Black Box Survey
35
solve the black box model explanation, i.e., the transparent approach is used in
the reverse engineering approach to explain the black box.
Yet another type of rule is exploited in [53]. Here, Lakkaraju et al. propose a
framework f for generating prediction models, which are both interpretable and
accuratem, by extracting Interpretable Decision Sets (IDS) cg , i.e., independent
if-then rules. Since each rule is independently applicable, decision sets are simple,
succinct, and easily to be interpreted. In particular, this approach can learn
accurate, short, and non-overlapping rules covering the whole feature space.
Rule Sets are adopted in [104] as comprehensible global predictor cg . The
authors present a Bayesian framework f for learning Rule Sets. A set of parameters is provided to the user to encourage the model to have a desired size and
shape in order to conform with a domain-specific definition of interpretability. A
Rule Set consists of a small number of short rules where an instance is classified
as positive if it satisfies at least one of the rules. The rule set provides reasons
for predictions, and also descriptions of a particular class.
Finally, in [64] an approach f is designed to learn both sparse conjunctive
and disjunctive clause rules from training data through a linear programming
solution. The optimization formulation leads the resulting rule-based global predictor cg (1Rule) to automatically balance accuracy and interpretability.
9.2
Explanation via Prototype Selection
In this section we present the design of a set of approaches f for solving the
transparent box design problem returning a comprehensible predictor cg equipped
with a human understandable global explanator function εg . A prototype, also
referred to with the name artifact or archetype, is an object that is representative
of a set of similar instances. A prototype can be an instance x part of the training
set D = {X, Y }, or can lie anywhere in the space X m ×Y of the dataset D. Having
only prototypes among the observed points is desirable for interpretability, but
it can also improve the classification error. As an example of a prototype we
can consider the record minimizing the sum of the distances with all the other
points of a set (like in K-Medoids) or the record generated averaging the value
of the features of a set of points (like in K-Means) [95]. Different definitions and
requirements to find a prototype are specified in each work using the prototypes
to explain the black box.
In [8], Bien et al. design the transparent Prototype Selection (PS) approach
f that first seeks for the best prototype (two strategies are proposed), and then
assigns the points in D to the label corresponding to the prototype. In particular, they face the problem of recognizing hand written digits. In this approach,
every instance can be described by more than one prototype, and more than a
prototype can refer to the same label (e.g., there can be more than one prototype
for digit zero, more than one for digit one, etc.). The comprehensible predictor
cg provides a global explanation in which every instance must have a prototype
corresponding to its label in its neighborhood; no instances should have a prototype with a different label in its neighborhood, and there should be as few
prototypes as possible.
36
R. Guidotti et al.
Kim et al. in [44,45] design the Bayesian Case Model (BCM) comprehensible predictor cl able to learn prototypes by clustering the data and to learn
subspaces. Each prototype is the representative sample of a given cluster, while
the subspaces are set of features which are important in identifying the cluster
prototype. That is, the global explanator εg returns a set of prototypes together
with their fundamental features. Possible drawbacks of this approach are the
high number of parameters (e.g., number of clusters) and various types of probability distributions which are assumed to be correct for each type of data. [42]
proposes an extension of BCM which exploits humans interaction to improve
the prototypes. Finally, in [43] the approach is further expanded to include criticisms, where a criticism is an instance that does not fit the model very well,
i.e., a counter-example part of the cluster of a prototype.
With respect to prototypes and DNN, [63] proposes a method b to change the
image representations in order to use only information from the original image
representation and from a generic natural image prior. This task is mainly related
to image reconstruction rather than black box explanation, but it is realized
with the aim of understanding the example to which the DNN b is related to
producing a certain prediction by realizing a sort of artificial image prototype.
There is a significant amount of work in understanding the representation of
DNN by means of artifact images, [41,100,105,111].
We conclude this section presenting how [25] deals with artifacts in DNNs.
Finding a single representative prototype by perturbation, deletion, preservation, and similar approaches has the risk of triggering artifacts of the black box.
As discussed in Section 8.3, NN and DNN are known to be affected by surprising
artifacts. For example, [52] shows that a nearly-invisible image perturbation can
lead a NN to classify an object for another; [69] constructs abstract synthetic
images that are classified arbitrarily; [63] finds deconstructed versions of an image which are indistinguishable from the viewpoint of the DNN from the original
image, and also with respect to texts [58] inserts typos and random sentences in
real texts that are classified arbitrarily. These examples demonstrate that it is
possible to find particular inputs that can drive the DNN to generate nonsensical or unexpected outputs. While not all artifacts look “unnatural”, nevertheless
they form a subset of images that are sampled with negligible probability when
the network is normally operated. In our opinion, two guidelines should be followed to avoid such artifacts in generating explanations for DNNs, and for every
black box in general. The first one is that powerful explanations should, just like
any predictor, generalize as much as possible. Second, the artifacts should not
be representative of natural perturbations.
9.3
Explanation via Other Approaches
In the following we present solutions for the transparent box design problem
adopting approaches f that can not be categorized as the previous ones. [47]
describes a method f based on Naive Bayes aimed to explain individual predictions εl of black boxes b. The proposed approach exploits notions from coalitional
game theory, and explains predictions utilizing the contribution of the value of
Open The Black Box Survey
37
different individual features εl (see Figure 2). The method is agnostic with respect to the black box used and is tested only on tabular data. Finally, in [103]
Wang et al. propose a method f named OT-SpAMs based on oblique tree sparse
additive models for obtaining a global interpretable predictor cg as a decision
tree. OT-SpAMs divides the feature space into regions using a sparse oblique tree
splitting and assigns local sparse additive experts (leaf of the tree) to individual
regions. Basically, OT-SpAMs passes from complicated trees/linear models to
an explainable tree εg .
10
Conclusion
In this paper we have presented a comprehensive overview of methods proposed
in the literature for explaining decision systems based on opaque and obscure
machine learning models. First, we have identified the different components of
the family of the explanation problems. In particular, we have provided a formal
definition of each problem belonging to that family capturing for each one the
proper peculiarity. We have named these problems: black box model explanation
problem, black box outcome explanation problem, black box inspection problem
and transparent box design problem. Then, we have proposed a classification
of methods studied in the literature which take into account the following dimensions: the specific explanation problem addressed, the type of explanator
adopted, the black box model opened, and the type of data used as input by the
black box model.
As shown in this paper, a considerable amount of work has already been
done in different scientific communities and especially in the machine learning
and data mining communities. The first one is mostly focused on describing how
the black boxes work, while the second one is more interested into explaining
the decisions even without understanding the details on how the opaque decision
systems work in general.
The analysis of the literature conducted in this paper has led to the conclusion that despite many approaches have been proposed to explain black boxes,
some important scientific questions still remain unanswered. One of the most
important open problems is that, until now, there is no agreement on what an
explanation is. Indeed, some works provide as explanation a set of rules, others
a decision tree, others a prototype (especially in the context of images). It is
evident that the research activity in this field completely ignored the importance of studying a general and common formalism for defining an explanation,
identifying which are the properties that an explanation should guarantee, e.g.,
soundness, completeness, compactness and comprehensibility. Concerning this
last property, there is no work that seriously addresses the problem of quantifying the grade of comprehensibility of an explanation for humans, although it is
of fundamental importance. The study of measures able to capture this aspect
is challenging because it also consider also aspects like the expertise of the user
or the amount of time available to understand the explanation. The definition
of a (mathematical) formalism for explanations and of tools for measuring how
38
R. Guidotti et al.
much an explanation is comprehensible for humans would improve the practical
applicability of most of the approaches presented in this paper.
Moreover, there are other open research questions related to black boxes and
explanations that are starting to be treated by the scientific community and that
deserve attention and more investigation.
A common assumption of all categories of works presented in this paper is
that the features used by the black box decision system are completely known.
However, a black box might use additional information besides that explicitly
asked to the user. For example, it might link the user’s information with different
data sources for augmenting the data to be exploited for the prediction. Therefore, an important aspect to be investigated is to understand how an explanation
might also be derived in cases where black box systems make decisions in presence of latent features. An interesting starting point for this research direction
is the framework proposed in [55] by Lakkaraju et al. for the evaluation of the
prediction models performances on labeled data where the decision of decisionmakers (either humans or black-boxes) is taken in the presence of unobserved
features.
Another open research question is related to providing explanations in the
field of recommender systems. When a suggestion is provided to a customer,
it should come together with the reasons for this recommendation. In [67] the
authors define a case-based reasoning approach to generate recommendations
with the opportunity of obtaining both the explanation of the recommendation
process and of the produced recommendations.
Lastly, a further interesting point is the fact that explanations are important
on their own and predictors might be learned directly from explanations. A
starting study of this aspect is [49] that presents a software agent learned to
simulate the Mario Bros game only utilizing explanations rather than the logs
of previous plays.
Table 3. Summary of methods for opening and explaining black boxes with respect
to the problem faced.
Problem
Model Explanation
Outcome Explanation
Black Box Inspection
Transparent Design
References
[20], [50], [9], [39], [14], [29], [114], [87], [32], [94], [5], [19], [37], [115],
[6], [70], [28], [65], [22], [61], [33], [51], [97]
[108], [25], [113], [89], [56], [83], [98]
[71], [7], [92], [16], [35], [30], [48], [2], [1], [110], [96], [90], [81]
[109], [102], [57], [91], [53], [104], [64], [8], [44], [63], [47], [103]
Open The Black Box Survey
A
39
Supplementary Materials
Acknowledgement
This work is partially supported by the European Community’s H2020 Program
under the funding scheme “INFRAIA-1-2014-2015: Research Infrastructures”,
grant agreement 654024, SoBigData, http://www.sobigdata.eu.
References
1. J. Adebayo and L. Kagal. Iterative orthogonal feature projection for diagnosing
bias in black-box models. arXiv preprint arXiv:1611.04967, 2016.
2. P. Adler, C. Falk, S. A. Friedler, G. Rybeck, C. Scheidegger, B. Smith, and
S. Venkatasubramanian. Auditing black-box models for indirect influence. In
Data Mining (ICDM), 2016 IEEE 16th International Conference on, pages 1–10.
IEEE, 2016.
3. R. Agrawal, R. Srikant, et al. Fast algorithms for mining association rules. In
Proc. 20th int. conf. very large data bases, VLDB, volume 1215, pages 487–499,
1994.
4. Y. A. A. S. Aldeen, M. Salleh, and M. A. Razzaque. A comprehensive review on
privacy preserving data mining. SpringerPlus, 4(1):694, 2015.
5. R. Andrews, J. Diederich, and A. B. Tickle. Survey and critique of techniques for
extracting rules from trained artificial neural networks. Knowledge-based systems,
8(6):373–389, 1995.
6. M. G. Augasta and T. Kathirvalavakumar. Reverse engineering the neural networks for rule extraction in classification problems. Neural processing letters,
35(2):131–150, 2012.
7. D. Baehrens, T. Schroeter, S. Harmeling, M. Kawanabe, K. Hansen, and K.-R.
MÞller. How to explain individual classification decisions. Journal of Machine
Learning Research, 11(Jun):1803–1831, 2010.
8. J. Bien and R. Tibshirani. Prototype selection for interpretable classification.
The Annals of Applied Statistics, pages 2403–2424, 2011.
Table 4. Summary of methods for opening and explaining black boxes with respect
to the explanator adopted.
Explanator
Decition Tree (DT)
Decision Rules (DR)
Features Importance (FI)
Saliency Mask (SM)
Sensitivity Analysis (SA)
Partial Dependence Plot (PDP)
Neurons Activation (NA)
Prototype Selection (PS)
References
[20], [50], [9], [39], [14], [23], [29], [114],
[87], [32], [94], [51], [96], [103]
[5], [19], [37], [115], [6], [70], [28], [65], [22],
[98], [109], [102], [57], [91], [53], [104], [64]
[61], [33], [97], [83]
[108], [25], [113], [89], [56]
[71], [7], [92], [16]
[35], [30], [48], [2], [1]
[110], [90], [81]
[8], [44], [63]
40
R. Guidotti et al.
9. O. Boz. Extracting decision trees from trained neural networks. In Proceedings of
the eighth ACM SIGKDD international conference on Knowledge discovery and
data mining, pages 456–461. ACM, 2002.
10. L. Breiman, J. Friedman, C. J. Stone, and R. A. Olshen. Classification and
regression trees. CRC press, 1984.
11. A. Caliskan-Islam, J. J. Bryson, and A. Narayanan. Semantics derived automatically from language corpora necessarily contain human biases. arXiv preprint
arXiv:1608.07187, 2016.
12. C. Carter, E. Renuart, M. Saunders, and C. C. Wu. The credit card market and
regulation: In need of repair. NC Banking Inst., 10:23, 2006.
13. R. Caruana, Y. Lou, J. Gehrke, P. Koch, M. Sturm, and N. Elhadad. Intelligible
models for healthcare: Predicting pneumonia risk and hospital 30-day readmission. In Proceedings of the 21th ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining, pages 1721–1730. ACM, 2015.
14. H. Chipman, E. George, and R. McCulloh. Making sense of a forest of trees.
Computing Science and Statistics, pages 84–92, 1998.
15. G. Comandè. Regulating algorithms regulation? first ethico-legal principles, problems, and opportunities of algorithms. In Transparent Data Mining for Big and
Small Data, pages 169–206. Springer, 2017.
16. P. Cortez and M. J. Embrechts. Opening black box data mining models using
sensitivity analysis. In Computational Intelligence and Data Mining (CIDM),
2011 IEEE Symposium on, pages 341–348. IEEE, 2011.
17. P. Cortez and M. J. Embrechts. Using sensitivity analysis and visualization techniques to open black box data mining models. Information Sciences, 225:1–17,
2013.
18. P. Cortez, J. Teixeira, A. Cerdeira, F. Almeida, T. Matos, and J. Reis. Using
data mining for wine quality assessment. In Discovery Science, volume 5808,
pages 66–79. Springer, 2009.
19. M. Craven and J. W. Shavlik. Using sampling and queries to extract rules from
trained neural networks. In ICML, pages 37–45, 1994.
20. M. Craven and J. W. Shavlik. Extracting tree-structured representations of
trained networks. In Advances in neural information processing systems, pages
24–30, 1996.
21. A. Datta, S. Sen, and Y. Zick. Algorithmic transparency via quantitative input
influence: Theory and experiments with learning systems. In Security and Privacy
(SP), 2016 IEEE Symposium on, pages 598–617. IEEE, 2016.
22. H. Deng.
Interpreting tree ensembles with intrees.
arXiv preprint
arXiv:1408.5456, 2014.
Table 5. Summary of methods for opening and explaining black boxes with respect
to the black box explained.
Black Box
Neural Network (NN)
Tree Ensemble (TE)
Support Vector Machines (SVM)
Deep Neural Network (DNN)
Agnostic black box (AGN)
References
[20], [50], [9], [39], [23], [5], [19], [37], [115], [6], [71]
[14], [23], [29], [114], [87], [32], [94], [22], [97]
[70], [28], [65]
[108], [25], [113], [89], [56], [92], [110], [96], [90], [81]
[61], [33], [51], [83], [98], [7], [16], [35], [30], [48], [2], [1]
Open The Black Box Survey
41
23. P. Domingos. Knowledge discovery via multiple models. Intelligent Data Analysis,
2(1-4):187–202, 1998.
24. F. Doshi-Velez and B. Kim. Towards a rigorous science of interpretable machine
learning. 2017.
25. R. Fong and A. Vedaldi. Interpretable explanations of black boxes by meaningful
perturbation. arXiv preprint arXiv:1704.03296, 2017.
26. E. Frank and I. H. Witten. Generating accurate rule sets without global optimization. 1998.
27. A. A. Freitas. Comprehensible classification models: a position paper. ACM
SIGKDD explorations newsletter, 15(1):1–10, 2014.
28. G. Fung, S. Sandilya, and R. B. Rao. Rule extraction from linear support vector
machines. In Proceedings of the eleventh ACM SIGKDD international conference
on Knowledge discovery in data mining, pages 32–40. ACM, 2005.
29. R. D. Gibbons, G. Hooker, M. D. Finkelman, D. J. Weiss, P. A. Pilkonis, E. Frank,
T. Moore, and D. J. Kupfer. The cad-mdd: a computerized adaptive diagnostic
screening tool for depression. The Journal of clinical psychiatry, 74(7):669, 2013.
30. A. Goldstein, A. Kapelner, J. Bleich, and E. Pitkin. Peeking inside the black box:
Visualizing statistical learning with plots of individual conditional expectation.
Journal of Computational and Graphical Statistics, 24(1):44–65, 2015.
31. B. Goodman and S. Flaxman. Eu regulations on algorithmic decision-making and
a right to explanation. In ICML workshop on human interpretability in machine
learning (WHI 2016), New York, NY. http://arxiv. org/abs/1606.08813 v1, 2016.
32. S. Hara and K. Hayashi. Making tree ensembles interpretable. arXiv preprint
arXiv:1606.05390, 2016.
33. A. Henelius, K. Puolamäki, H. Boström, L. Asker, and P. Papapetrou. A peek into
the black box: exploring classifiers by randomization. Data mining and knowledge
discovery, 28(5-6):1503–1529, 2014.
34. J. M. Hofman, A. Sharma, and D. J. Watts. Prediction and explanation in social
systems. Science, 355(6324):486–488, 2017.
35. G. Hooker. Discovering additive structure in black box functions. In Proceedings
of the tenth ACM SIGKDD international conference on Knowledge discovery and
data mining, pages 575–580. ACM, 2004.
36. J. Huysmans, K. Dejaeger, C. Mues, J. Vanthienen, and B. Baesens. An empirical
evaluation of the comprehensibility of decision table, tree and rule based predictive
models. Decision Support Systems, 51(1):141–154, 2011.
37. U. Johansson, R. König, and L. Niklasson. Rule extraction from trained neural networks using genetic programming. In 13th International Conference on
Artificial Neural Networks, pages 13–16, 2003.
38. U. Johansson, R. König, and L. Niklasson. The truth is in there-rule extraction
from opaque models using genetic programming. In FLAIRS Conference, pages
658–663, 2004.
39. U. Johansson and L. Niklasson. Evolving decision trees using oracle guides. In
Computational Intelligence and Data Mining, 2009. CIDM’09. IEEE Symposium
on, pages 238–244. IEEE, 2009.
40. U. Johansson, L. Niklasson, and R. König. Accuracy vs. comprehensibility in
data mining models. In Proceedings of the seventh international conference on
information fusion, volume 1, pages 295–300, 2004.
41. H. Kato and T. Harada. Image reconstruction from bag-of-visual-words. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages
955–962, 2014.
42
R. Guidotti et al.
42. B. Kim, E. Glassman, B. Johnson, and J. Shah. ibcm: Interactive bayesian case
model empowering humans via intuitive interaction. 2015.
43. B. Kim, O. O. Koyejo, and R. Khanna. Examples are not enough, learn to criticize! criticism for interpretability. In Advances In Neural Information Processing
Systems, pages 2280–2288, 2016.
44. B. Kim, C. Rudin, and J. A. Shah. The bayesian case model: A generative
approach for case-based reasoning and prototype classification. In Advances in
Neural Information Processing Systems, pages 1952–1960, 2014.
45. B. Kim, J. A. Shah, and F. Doshi-Velez. Mind the gap: A generative approach to
interpretable feature selection and extraction. In Advances in Neural Information
Processing Systems, pages 2260–2268, 2015.
46. P. W. Koh and P. Liang. Understanding black-box predictions via influence
functions. arXiv preprint arXiv:1703.04730, 2017.
47. I. Kononenko et al. An efficient explanation of individual classifications using
game theory. Journal of Machine Learning Research, 11(Jan):1–18, 2010.
48. J. Krause, A. Perer, and K. Ng. Interacting with predictions: Visual inspection of
black-box machine learning models. In Proceedings of the 2016 CHI Conference
on Human Factors in Computing Systems, pages 5686–5697. ACM, 2016.
49. S. Krening, B. Harrison, K. M. Feigh, C. L. Isbell, M. Riedl, and A. Thomaz.
Learning from explanations using sentiment and advice in rl. IEEE Transactions
on Cognitive and Developmental Systems, 9(1):44–55, 2017.
50. R. Krishnan, G. Sivakumar, and P. Bhattacharya. Extracting decision trees from
trained neural networks. Pattern recognition, 32(12), 1999.
51. S. Krishnan and E. Wu. Palm: Machine learning explanations for iterative debugging. In Proceedings of the 2nd Workshop on Human-In-the-Loop Data Analytics,
page 4. ACM, 2017.
52. A. Kurakin, I. Goodfellow, and S. Bengio. Adversarial examples in the physical
world. arXiv preprint arXiv:1607.02533, 2016.
53. H. Lakkaraju, S. H. Bach, and J. Leskovec. Interpretable decision sets: A joint
framework for description and prediction. In Proceedings of the 22nd ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining,
pages 1675–1684. ACM, 2016.
54. H. Lakkaraju, E. Kamar, R. Caruana, and J. Leskovec. Interpretable & explorable
approximations of black box models. arXiv preprint arXiv:1707.01154, 2017.
55. H. Lakkaraju, J. Kleinberg, J. Leskovec, J. Ludwig, and S. Mullainathan. The
selective labels problem: Evaluating algorithmic predictions in the presence of unobservables. In Proceedings of the 23rd ACM SIGKDD International Conference
on Knowledge Discovery and Data Mining, pages 275–284. ACM, 2017.
56. T. Lei, R. Barzilay, and T. Jaakkola. Rationalizing neural predictions. arXiv
preprint arXiv:1606.04155, 2016.
57. B. Letham, C. Rudin, T. H. McCormick, D. Madigan, et al. Interpretable classifiers using rules and bayesian analysis: Building a better stroke prediction model.
The Annals of Applied Statistics, 9(3):1350–1371, 2015.
58. B. Liang, H. Li, M. Su, P. Bian, X. Li, and W. Shi. Deep text classification can
be fooled. arXiv preprint arXiv:1704.08006, 2017.
59. Z. C. Lipton.
The mythos of model interpretability.
arXiv preprint
arXiv:1606.03490, 2016.
60. Y. Lou, R. Caruana, and J. Gehrke. Intelligible models for classification and
regression. In Proceedings of the 18th ACM SIGKDD international conference on
Knowledge discovery and data mining, pages 150–158. ACM, 2012.
Open The Black Box Survey
43
61. Y. Lou, R. Caruana, J. Gehrke, and G. Hooker. Accurate intelligible models with
pairwise interactions. In Proceedings of the 19th ACM SIGKDD international
conference on Knowledge discovery and data mining, pages 623–631. ACM, 2013.
62. S. Lowry and G. Macpherson. A blot on the profession. British medical journal
(Clinical research ed.), 296(6623):657, 1988.
63. A. Mahendran and A. Vedaldi. Understanding deep image representations by
inverting them. In Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pages 5188–5196, 2015.
64. D. M. Malioutov, K. R. Varshney, A. Emad, and S. Dash. Learning interpretable
classification rules with boolean compressed sensing. In Transparent Data Mining
for Big and Small Data, pages 95–121. Springer, 2017.
65. D. Martens, B. Baesens, T. Van Gestel, and J. Vanthienen. Comprehensible credit
scoring models using rule extraction from support vector machines. European
journal of operational research, 183(3):1466–1476, 2007.
66. D. Martens, J. Vanthienen, W. Verbeke, and B. Baesens. Performance of classification models from a user perspective. Decision Support Systems, 51(4):782–793,
2011.
67. D. McSherry. Explanation in recommender systems. Artificial Intelligence Review,
24(2):179–197, 2005.
68. P. M. Murphy and M. J. Pazzani. Id2-of-3: Constructive induction of m-of-n concepts for discriminators in decision trees. In Proceedings of the eighth international
workshop on machine learning, pages 183–187, 1991.
69. A. Nguyen, J. Yosinski, and J. Clune. Deep neural networks are easily fooled:
High confidence predictions for unrecognizable images. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, pages 427–436, 2015.
70. H. Núñez, C. Angulo, and A. Català. Rule extraction from support vector machines. In Esann, pages 107–112, 2002.
71. J. D. Olden and D. A. Jackson. Illuminating the black box: a randomization
approach for understanding variable contributions in artificial neural networks.
Ecological modelling, 154(1):135–150, 2002.
72. F. E. Otero and A. A. Freitas. Improving the interpretability of classification
rules discovered by an ant colony algorithm. In Proceedings of the 15th annual
conference on Genetic and evolutionary computation, pages 73–80. ACM, 2013.
73. G. L. Pappa, A. J. Baines, and A. A. Freitas. Predicting post-synaptic activity
in proteins with data mining. Bioinformatics, 21(suppl 2):ii19–ii25, 2005.
74. F. Pasquale. The black box society: The secret algorithms that control money and
information. Harvard University Press, 2015.
75. M. J. Pazzani, S. Mani, W. R. Shankle, et al. Acceptance of rules generated by
machine learning among medical experts. Methods of information in medicine,
40(5):380–385, 2001.
76. D. Pedreshi, S. Ruggieri, and F. Turini. Discrimination-aware data mining. In
Proceedings of the 14th ACM SIGKDD international conference on Knowledge
discovery and data mining, pages 560–568. ACM, 2008.
77. J. R. Quinlan. Generating production rules from decision trees. In ijcai, volume 87, pages 304–307, 1987.
78. J. R. Quinlan. Simplifying decision trees. International journal of man-machine
studies, 27(3):221–234, 1987.
79. J. R. Quinlan. C4. 5: Programs for Machine Learning. Elsevier, 1993.
80. J. R. Quinlan and R. M. Cameron-Jones. Foil: A midterm report. In European
conference on machine learning, pages 1–20. Springer, 1993.
44
R. Guidotti et al.
81. A. Radford, R. Jozefowicz, and I. Sutskever. Learning to generate reviews and
discovering sentiment. arXiv preprint arXiv:1704.01444, 2017.
82. M. T. Ribeiro, S. Singh, and C. Guestrin. Model-agnostic interpretability of
machine learning. arXiv preprint arXiv:1606.05386, 2016.
83. M. T. Ribeiro, S. Singh, and C. Guestrin. Nothing else matters: Modelagnostic explanations by identifying prediction invariance.
arXiv preprint
arXiv:1611.05817, 2016.
84. M. T. Ribeiro, S. Singh, and C. Guestrin. Why should i trust you?: Explaining
the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 1135–1144.
ACM, 2016.
85. A. Romei and S. Ruggieri. A multidisciplinary survey on discrimination analysis.
The Knowledge Engineering Review, 29(5):582–638, 2014.
86. A. Saltelli. Sensitivity analysis for importance assessment. Risk analysis,
22(3):579–590, 2002.
87. V. Schetinin, J. E. Fieldsend, D. Partridge, T. J. Coats, W. J. Krzanowski, R. M.
Everson, T. C. Bailey, and A. Hernandez. Confident interpretation of bayesian decision tree ensembles for clinical applications. IEEE Transactions on Information
Technology in Biomedicine, 11(3):312–319, 2007.
88. C. Seifert, A. Aamir, A. Balagopalan, D. Jain, A. Sharma, S. Grottel, and
S. Gumhold. Visualizations of deep neural networks in computer vision: A survey.
In Transparent Data Mining for Big and Small Data, pages 123–144. Springer,
2017.
89. R. R. Selvaraju, A. Das, R. Vedantam, M. Cogswell, D. Parikh, and D. Batra.
Grad-cam: Why did you say that? visual explanations from deep networks via
gradient-based localization. arXiv preprint arXiv:1610.02391, 2016.
90. R. Shwartz-Ziv and N. Tishby. Opening the black box of deep neural networks
via information. arXiv preprint arXiv:1703.00810, 2017.
91. G. Su, D. Wei, K. R. Varshney, and D. M. Malioutov. Interpretable two-level
boolean rule learning for classification. arXiv preprint arXiv:1511.07361, 2015.
92. M. Sundararajan, A. Taly, and Q. Yan. Axiomatic attribution for deep networks.
arXiv preprint arXiv:1703.01365, 2017.
93. C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow,
and R. Fergus. Intriguing properties of neural networks. arXiv preprint
arXiv:1312.6199, 2013.
94. H. F. Tan, G. Hooker, and M. T. Wells. Tree space prototypes: Another look at
making tree ensembles interpretable. arXiv preprint arXiv:1611.07115, 2016.
95. P.-N. Tan et al. Introduction to data mining. Pearson Education India, 2006.
96. J. J. Thiagarajan, B. Kailkhura, P. Sattigeri, and K. N. Ramamurthy. Treeview:
Peeking into deep neural networks via feature-space partitioning. arXiv preprint
arXiv:1611.07429, 2016.
97. G. Tolomei, F. Silvestri, A. Haines, and M. Lalmas. Interpretable predictions
of tree-based ensembles via actionable feature tweaking. In Proceedings of the
23rd ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining, pages 465–474. ACM, 2017.
98. R. Turner. A model explanation system. In Machine Learning for Signal Processing (MLSP), 2016 IEEE 26th International Workshop on, pages 1–6. IEEE,
2016.
99. W. Verbeke, D. Martens, C. Mues, and B. Baesens. Building comprehensible customer churn prediction models with advanced rule induction techniques. Expert
Systems with Applications, 38(3):2354–2364, 2011.
Open The Black Box Survey
45
100. C. Vondrick, A. Khosla, T. Malisiewicz, and A. Torralba. Hoggles: Visualizing
object detection features. In Proceedings of the IEEE International Conference
on Computer Vision, pages 1–8, 2013.
101. S. Wachter, B. Mittelstadt, and L. Floridi. Why a right to explanation of automated decision-making does not exist in the general data protection regulation.
International Data Privacy Law, 7(2):76–99, 2017.
102. F. Wang and C. Rudin. Falling rule lists. In Artificial Intelligence and Statistics,
pages 1013–1022, 2015.
103. J. Wang, R. Fujimaki, and Y. Motohashi. Trading interpretability for accuracy:
Oblique treed sparse additive models. In Proceedings of the 21th ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, pages 1245–
1254. ACM, 2015.
104. T. Wang, C. Rudin, F. Velez-Doshi, Y. Liu, E. Klampfl, and P. MacNeille.
Bayesian rule sets for interpretable classification. In Data Mining (ICDM), 2016
IEEE 16th International Conference on, pages 1269–1274. IEEE, 2016.
105. P. Weinzaepfel, H. Jégou, and P. Pérez. Reconstructing an image from its local
descriptors. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE
Conference on, pages 337–344. IEEE, 2011.
106. A. Weller. Challenges for transparency. arXiv preprint arXiv:1708.01870, 2017.
107. D. Wettschereck, D. W. Aha, and T. Mohri. A review and empirical evaluation
of feature weighting methods for a class of lazy learning algorithms. In Lazy
learning, pages 273–314. Springer, 1997.
108. K. Xu, J. Ba, R. Kiros, K. Cho, A. Courville, R. Salakhudinov, R. Zemel, and
Y. Bengio. Show, attend and tell: Neural image caption generation with visual
attention. In International Conference on Machine Learning, pages 2048–2057,
2015.
109. X. Yin and J. Han. Cpar: Classification based on predictive association rules. In
Proceedings of the 2003 SIAM International Conference on Data Mining, pages
331–335. SIAM, 2003.
110. J. Yosinski, J. Clune, A. Nguyen, T. Fuchs, and H. Lipson. Understanding neural
networks through deep visualization. arXiv preprint arXiv:1506.06579, 2015.
111. M. D. Zeiler and R. Fergus. Visualizing and understanding convolutional networks. In European conference on computer vision, pages 818–833. Springer,
2014.
112. C. Zhang, S. Bengio, M. Hardt, B. Recht, and O. Vinyals. Understanding deep
learning requires rethinking generalization. arXiv preprint arXiv:1611.03530,
2016.
113. B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba. Learning deep
features for discriminative localization. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, pages 2921–2929, 2016.
114. Y. Zhou and G. Hooker. Interpreting models via single tree approximation. arXiv
preprint arXiv:1610.09036, 2016.
115. Z.-H. Zhou, Y. Jiang, and S.-F. Chen. Extracting symbolic rules from trained
neural network ensembles. Ai Communications, 16(1):3–15, 2003.
| 2cs.AI
|
1
Phase Noise Compensation for OFDM Systems
arXiv:1708.05443v1 [cs.IT] 17 Aug 2017
Amir Leshem and Michal Yemini
Abstract—We describe a low complexity method for time
domain compensation of phase noise in OFDM systems. We
extend existing methods in several respects. First we suggest using
the Karhunen-Loève representation of the phase noise process
to estimate the phase noise. We then derive an improved datadirected choice of basis elements for LS phase noise estimation
and present its total least square counterpart problem. The
proposed method helps overcome one of the major weaknesses of
OFDM systems. We also generalize the time domain phase noise
compensation to the multiuser MIMO context. Finally we present
simulation results using both simulated and measured phased
noise. We quantify the tracking performance in the presence of
residual carrier offset.
I. I NTRODUCTION
OFDM has become a prominent technology that is utilized
in many modern communication systems including cellular
systems such as 3GPP LTE [2], wireless LAN (WLAN) [3]
and WiMax [4]. In future LTE systems (see 802.11ac) multicarrier modulations such as Generalized frequency division
multiplex (GFDM), Filter bank multi-carrier (FBMC), Universal filtered multi-carrier (UFMC) and Filtered OFDM (fOFDM) [5]–[8] may be implemented. In spite of its popularity
and robustness to multipath propagation, OFDM is known to
be extremely affected by phase noise (PN) and frequency
offset [9]–[13]. MIMO OFDM receivers are very sensitive
to the phase noise coming from the difference between the
carrier frequency and the local oscillator (LO). In the case
of very high data rates this is actually the limiting factor on
performance. High order modulations such as 256 QAM are
also severely affected by phase noise. Phase noise is typically
modeled as a multiplicative noise. When the LO is locked to
the carrier frequency, the phase noise is lower and is modeled
as a finite power random process. When the LO is not locked
the phase noise is modeled as a Wiener process with infinite
power, the effect of such phase noise is analyzed from an
information theoretic perspective in [14]. In all typical 802.11
and LTE implementations the LO is locked; hence, we will
concentrate on the first model. Phase noise can be considered
to have two components: a common phase error (CPE) that
is common to all carriers, and a time varying part that is
frequency dependent. This part is typically weaker than the
CPE and generates the undesirable and harmful ICI.
A popular approach to ICI mitigation is the use of an MMSE
equalizer in the frequency domain that balances the AWGN
and the colored ICI. This approach can be found in many
The phase noise compensation scheme presented in this paper is patented
[1].
The authors are with the Faculty of Engineering, Bar-Ilan University, Ramat Gan 52900, Israel (email: [email protected],
[email protected]).
This work was supported by the Israel Science Foundation under Grant
903/2013.
articles including [13], [15]–[18]. The alternative approach of
PN cancellation/mitigation aims at jointly detecting the data
and cancelling the PN as described in [19], [20], or jointly
estimating the channel and PN [21]–[25] and even jointly
estimating the channel, detecting the data and compensating
for the PN as in [26]–[29]. These schemes suffer from high
complexity that is impractical to implement at high rates
and in high spectral efficiency communication systems. Yet
a third approach to reducing ICI is time domain processing.
Casas et al. [30] proposed a LS approach in the time domain
where they represented the noise using one of two fixed bases:
DFT or DCT. The dominant phase noise components are then
estimated using LS fitting of a few basis vectors (typically
the low frequency components due to phase noise properties).
When a single basis vector is used, the method reduces to that
of [15]–[17]. The main drawback of the method in [30] is that
typically phase noise cannot be compactly represented in the
fixed basis. Another LS PN compensation scheme estimates
both the channel coefficients and the PN with assumed low
complexity; however, the assumption that the number of pilot
subcarriers is larger than the number of transmitting antennas
leads to high complexity in massive MIMO communication
systems [31]–[33]. Another issue we consider is time variation
of the statistical properties of the PN process. Even when
the LO is locked, the statistical properties of the PN are
time varying due to external conditions such as temperature;
e.g., heat from the mobile device, thus making real time
basis selection very desirable for practical implementations.
Therefore, statistical knowledge of the PN covariance matrix
should be acquired.
Phase noise is present in many communication systems such
as WLAN, millimeter wave systems, full-duplex systems and
massive MIMO systems. Full-duplex systems are especially
affected by phase noise since self-interference is performed in
order to extract the received signals, see for example [34]–
[36]. The phase noise prevents transmitters from properly
subtract their self-interference, this is especially harmful in
high order modulations which demand high values of received
SNR which are translated to low values of error vector
magnitudes (EVM). OFDM millimeter wave systems are also
affected by phase noise, see for example [37], [38] ; here,
as in other systems the effect of phase noise is most severe
when using high order modulations, and is a key issue in
implementing these systems. Phase noise is also present in
optical communication systems using coherent optical OFDM
[39]. In such networks the phase is estimated digitally without
using a optical phase-locked loop, however, this estimation
is not perfect and thus phase noise is cause this imperfect
estimation. The phase compensation scheme that we present
is adaptable to the transmission of OFDM symbols in all of
these systems.
2
In this paper we replace the fixed basis proposed in previous
works with an adaptive basis which is the best representation
of the noise with respect to the L2 norm for a random noise
process [1]. Since in locked systems the phase noise behavior
is quite stationary we can either pre-calibrate the phase noise
PSD and then use an eigen-decomposition of the covariance
matrix or estimate in real time the basis elements as well as the
LS coefficients of the phase noise. The latter is more robust to
environmental changes such as temperature which might affect
the statistical properties of the phase noise. We can also replace
the LS estimation of the coefficient with a total least square
estimation [40] that considers imperfection of the model.
The main contributions of this article are: 1) Utilizing the
Karhunen-Loève representation of the phase noise process covariance matrix as basis elements. This dramatically improves
the results of [30]. 2) The introduction of the implementation
of the total least square estimator for phase noise mitigation
schemes. 3) We efficiently track of the subspace of the
covariance matrix of the phase noise process for system with
no information regarding the covariance matrix of the phase
noise process. This is performed utilizing the PAST algorithm
[41] and behaves well even in the presence of carrier frequency
offset. 4) We extended the above contributions for multiuser
uplink beamforming OFDM systems.
This paper is organized as follows: in section II we present
the phase noise model. Section III presents our results on
phase noise compensation in SISO systems and discusses
the computational aspects of the compensation scheme. In
Section IV we discuss two enhancements to the compensation
method. Section V is dedicated to tracking the dominant
subspace of the Karhunen-Loève (KL) representation presented in Section III. In Section VI we present simulation and
measured results of the proposed phase noise compensation
method. Section VI-A covers simulated phase noise whereas
Section VI-B discusses measured phase noise. Section VI-C
analyzes simulations of the tracking algorithm proposed in
Section V for the measured phase noise. Finally, Section VII
concludes the paper.
Notations: We denote the convolution between two continuous time signals x and y at time t by x ∗ y(t); for discrete
time we denote the convolution between the two discrete time
signals x and y at time k by x ∗ y(k). Vectors and matrices
appear in bold. Let a be a vector, we denote by aT the
transpose vector of a and by a∗ the conjugate transpose of a;
note that if a is of dimension 1; i.e., scalar, then a∗ is equal
to the conjugate of a. Moreover, M † is the Moore-Penrose
pseudoinverse of a matrix M . Finally, k · k denotes the L2
norm and k · kF denotes the Frobenius norm.
where ωk = ω0 + k∆ω is the frequency of the k’th channel,
k = − N2 , . . . , N2 , ω0 is the carrier frequency and
2π
Ts
∆ω =
(2)
is the angular sampling frequency. Additionally, s(k) is the
symbol transmitted of the k’th channel and is independent of
symbols transmitted over other channels. The OFDM symbol
passes through a time invariant channel (we assume a quasistationary fading process) and the received signal y(t) is given
by
y(t) = h ∗ x(t) + n(t).
(3)
Phase noise is multiplicative noise resulting from the jitter
of the LO of the OFDM system. We can model the received
signal with the effect of the phase noise as
z(t) = y(t)ejφ(t) ,
(4)
where φ(t) is a random process that can be considered to
be a filtered Gaussian process with PSD Pφ (f ). The process
ψ (t) = ejφ(t) is the multiplicative noise process that can
also include the residual frequency offset (a linear phase
component) and the common phase error that is constant
across frequencies. We want to estimate this and remove
its effect, since it introduces inter-channel interference (ICI).
We assume that ψ (t) is a stationary process with a known
covariance rψ (τ ) = E [ψ (t) ψ ∗ (t − τ )]. This assumption is
very reasonable when the LO is locked to a stable frequency
source through a Phase-Locked Loop (PLL).
T
Let ψ = (ψ1 , . . . , ψN ) be a vector of N consecutive
samples of the phase noise process ψm = ψ (mTs ). We define
the covariance matrix of the phase noise process by
Rψψ
E (ψ1 ψ1∗ )
..
=
.
E (ψN ψ1∗ )
...
..
.
...
∗
E (ψ1 ψN
)
..
.
.
∗
E (ψN ψN )
(5)
When trying to represent the phase noise along a single
OFDM symbol it is natural to use the basis of the eigenvectors
of Rψψ . We decompose Rψψ using an eigen-decomposition
as
Rψψ =
N
−1
X
µi ui u∗i ,
(6)
i=0
II. P HASE NOISE MODEL
In this section we describe a mathematical model for
the phase noise process and its effects on OFDM systems.
Consider an OFDM system described by
N −1
1 X
s(k)ejωk t ,
x(t) = √
N k=0
0 ≤ t ≤ Ts
(1)
where µ0 , . . . , µN −1 denote the eigenvalues of Rψψ and
u0 , . . . , uN −1 denote their respective eigenvectors.
We now describe the received signal and channel. We begin
with a SISO model and then extend it to a SIMO model. Note
that we are only interested in the SIMO case since the phase
noise is identical on all spatial channels. We assume that the
OFDM symbols are synchronized and that the cyclic prefix
3
has been removed, so that the channel matrix can be assumed
to be circulant, and thus be given by
h0
h1 · · ·
···
hN −1
hN −1 h0 . . .
hN −2
..
..
.. ..
..
(7)
H=
.
.
.
.
.
.
..
..
..
..
.
.
.
.
h1
· · · · · · hN −1
h0
Furthermore, we consider a single OFDM symbol. The time
domain OFDM symbol is given by
x = F ∗N s
(8)
where
s = [s0 , . . . , sN −1 ]
T
(9)
III. T IME DOMAIN COMPENSATION IN SISO SYSTEMS
We now describe a time domain method for reducing the
phase noise. The idea is to use available pilot data to estimate
the coefficients of a representation of the phase noise. We first
present the phase compensation algorithm in [30]. Then we
discuss the choice of basis for the phase noise compensation.
We show that using a fixed basis such as Fourier vectors or
discrete cosine transform vectors does not yield large gains
in terms of ICI cancellation. We then propose the use of data
directed basis selection and show the improvement achieved
through this approach. This is important for ICI cancellation
since common phase noise removal involves choosing the first
basis vector to be the N dimensional all ones vectors 1N =
T
[1, . . . , 1] .
Let v 0 , . . . , v N −1 be a basis for CN . Denote the phase
T
and let γ =
noise realization by ejϕ = ejφ0 , . . . , ejφN−1
T
[γ0 , . . . , γN−1 ] satisfy
N
−1
X
−jϕ
is the frequency domain OFDM symbol, and
γk v k ,
(16)
e
=
k=0
1
1
...
1
−2π·1·1/N
−2π·1·(N −1)/N
or equivalently
1
e
.
.
.
e
1
F N = √ .
..
..
.
.
.
e−jϕ = V γ,
(17)
N .
.
.
.
−2π(N −1)·1/N
−2π·(N −1)·(N −1)/N
1 e
... e
where V = [v 0 , . . . , v N −1 ]. If we allow only d basis vectors
(10) V (d) = [v , . . . , v
0
d−1 ] we can pose the problem as a least
squares problem. Had we known the vector e−jϕ our objective
is the DFT matrix. Let n = [n0 , . . . , nN −1 ] be the additive
would have been finding γ̂ ∈ CN such that V (d) γ cancels the
white Gaussian noise. It follows that
phase noise optimally (in LS sense), i.e.
y = HF ∗N s + n = Hx + n
(11)
γ̂ = arg min e−jϕ − V (d) γ
γ
is the received signal when no phase noise is present.
T
The received OFDM symbol z = [z0 , . . . , zN −1 ] is given
by
z = ejΦ (Hx + n)
2
.
(18)
However, since e−jϕ is the vector which we want to estimate,
we cannot use this naive approach. We discuss ways to
overcome this issue below.
(12)
A. LS compensation based on [30]
where
Φ = diag (φ0 , . . . , φN−1 )
(13)
is the phase noise vector.
Define a received data matrix Z; this matrix is given by
z0
jΦ
..
Z = diag(z) =
(14)
=Ye
.
zN −1
where
Y = diag(y) =
y0
..
.
yN −1
.
(15)
Our problem is to estimate the phase noise and construct a
time domain vector that cancels out the harmful effect of the
phase noise.
Since we do not know the phase noise e−jϕ we rely on
known OFDM pilot tones. In this case we need to modify (18)
T
assuming that we have known values sp = [si1 , . . . , sir ] ∈
npilot ×1
C
. Let ŷ be an estimate of the time domain symbol with
the phase noise removed:
ŷ = ZV (d) γ̂ ' Hx + n = HF ∗N s + n
(19)
Since H is diagonalized by the DFT matrix F N ; i.e., H =
F ∗N ΛF N , we obtain that
ŝ = Λ−1 F N ZV (d) γ̂.
(20)
is an estimate of the received OFDM frequency domain
symbol. Defining
W = Λ−1 F N ZV (d) ,
(21)
we obtain that our LS estimate of γ is given by
2
γ̂ = arg min ks − W γk .
γ
(22)
4
Therefore, we obtain that
γ̂ =
C. The basis vectors
W †p sp ,
(23)
where W p is obtained by choosing the rows that correspond to
pilot tones alone. The estimate of the phase noise cancellation
vector is now given by
e−jϕ = V γ̂.
(24)
Figure 1 depicts this phase noise compensation scheme.
Note that the components of Z are affected by noise and
the noise is multiplied by Λ−1 F N . This suggests that the
estimation of γ can be improved using total least squares
(TLS) instead of the LS described above. We discuss this in
the next section.
B. TLS compensation
As discussed above, in the training period, the training
symbols spilot are sent over the channel. Since these training
symbols are predefined and known by the receiver, our compensation problem can be represented as a data least squares
(DLS) problem [42]. However, since we use the basis V (d) of
size d ≤ N we also consider discrepancies in sp .
By Eq. (20) we have that
s ≈ Λ−1 F N ZV (d) γ̂.
(25)
The uncertainty in the model is caused by the estimation of
the channel matrix H, the additive noise of the channel and
the reduced basis dimensions. It it therefore natural to consider
the TLS estimation of γ. Let ∆W and ∆s be such that
(W p + ∆W )γ = sp + ∆s
(26)
where as before, W p is obtained by choosing the rows that
correspond to pilot tones alone.
The TLS problem is then
arg
s.t.:
min
γ,∆W ,∆s
k[∆W , ∆s]kF
(W p + ∆W )γ = sp + ∆s
(27)
where as stated above (Section I) the notationk · kF denotes
the Frobenius norm. The solution in terms of γ̂ is obtained
by following Algorithm 1 in [40] which we describe next. Let
W p and sp as stated above and let
[W p , sp ] = U ΣQT
(28)
be the singular value decomposition (SVD) of [W p , sp ] ∈
Cnpilot ×(d+1) . Denote
q 11 q 12
Q,
(29)
q 21 q22
where q 11 ∈ Cd×d , q 12 ∈ Cd×1 , q 21 ∈ C1×d and q22 ∈ C. If
q22 6= 0 , note that q22 is scalar, then
γ̂ TLS = −q 12 /q22 .
(30)
When q22 = 0, there is no solution and we set γ̂ TLS = γ̂ LS .
Note that in some practical implementations we trade off accuracy for lower complexity. However, if an SVD computation
engine is available because of beamforming, for example, then
this SVD engine can be used for solving the TLS problem in
(28).
We now turn to the question of the choice of the basis
vectors v 0 , . . . , v N −1 . The authors of [30] proposed using
either the columns of the DFT matrix F N or the columns
of the DCT matrix. As will be seen in the simulations, this
typically leads to minor improvement over simply cancelling
the common phase but the ICI is still significant. We suggest
a different approach and choose the basis elements using the
properties of the phase noise process. We assume that the
phase noise process has covariance
Rψψ =
N
X
µk uk u∗k
(31)
k=1
where u0 , . . . , uN −1 are the eigenvectors corresponding to
eigenvalues µ0 > . . . > µN−1 , respectively. This basis is
the best choice for representing random realizations of a
random process with covariance Rψψ (this is basically a KL
representation of the process). Since the statistical properties
of the phase noise process are stationary for quite long periods
they can be calibrated in advance.
D. Computational aspects
The compensation of the phase noise involves O((N log N ) ×
d + npilot × d2 ) complex multiplications where d is the number
of the basis elements. This is due to the FFT complexity
adding d multiplications per symbol and the additional complexity of the LS itself. We next explain these values in more
detail. Forming the matrix W involves computing
W = Λ−1 F N ZV (d) .
(32)
Matrices Λ and Z are diagonal matrices and matrix F n
is symmetric. Thus, calculating of the multiplication ZV (d)
requires at most N × d complex multiplication operations.
Using the fast Fourier transform (FFT), we have that the
complexity of calculating F N ZV (d) is O((N log N ) × d).
The, additional multiplication by Λ−1 does not increase the
complexity.
Solving the LS problem (22) only involves matrices of size
npilot × d resulting in O(npilot × d2 ) operations.
It follows that the complexity depends on d and npilot and
is O((N log N ) × d + npilot × d2 ).
We note that the complexity of calculating the eigenvectors of the covariance matrix is O(N 3 ). This is done once
and thus the complexity over time vanishes. However, since
phase noise statistics can slowly vary due to frequency offset
variations, we also provide a tracking solution. In Section V
we demonstrate that we can track the d leading eigenvectors
with a complexity of O(N d) operations per update. We can
conclude that the complexity of our compensation scheme is
O((N log N ) × d + npilot × d2 ) even if tracking the d leading
eigenvectors is performed. This is on par with the complexity
of the compensation scheme presented in [30] while providing
superior results due to the better and adaptive basis selection.
5
V̂ 1
Rψψ
Compute
SVD of
Rψψ
Select
V̂ 1 , . . . , V̂ d .
..
as
basis elements
V̂ 1
...
V̂ d
V̂ d
...
V̂11
pilot 1
γ̂1
..
.
pilot kp
null tone 1
..
.
LS
(TLS)
solver
⊗
V̂1N
...
...
V̂d1
⊗
..
.
γ̂d
⊗
null tone kn
⊗
P
z1
V̂dN
P
⊗
zN
y1
⊗
yN
...
FFT
...
To FEQ
Fig. 1: Phase noise compensation process without subspace tracking.
IV. E XTENSIONS
We now describe two extensions to the proposed method.
These extensions can contribute substantially to the performance of the proposed scheme by increasing the number of
available equations for PN mitigation.
A. Using null tones
While channel estimation can only use tones in which
energy has been transmitted, null tones can also provide
information on ISI. This depends, of course, on the amount
of adjacent channel suppression; however, when the adjacent
channel suppression is good, the ICI can be estimated based
on the null tones as well.
B. Phase estimation in MIMO system
When a MIMO system is used, typically all transceiver
chains use the same LO. Hence the phase noise can be jointly
estimated based on the pilot symbols from all the receive
antennas. This extra information substantially enhances the
applicability of the proposed method and improves the quality
of the LS fitting of the coefficients, especially in modern
802.11ac and massive MIMO systems.
Next we provide a detailed example for adapting the algorithm presented in this paper to multiuser uplink beamforming
OFDM. The example assumes transmitter phase compensation
[43], [44], for this reason it focuses on compensating for
the receiver phase noise. Under the assumption of proper
phase noise compensation at the transmitters, the phase noise
at the receiver is the dominating noise compared with the
compensated phase noise at the transmitters.
Suppose that a receiver with Nr antennas serves Nu single
antenna users. Each user ku ∈ {1, . . . , Nu } sends the vector of
symbols sku = [sku ,0 , . . . , sku ,N −1 ]T , by aiming to transmit
the signal
N −1
1 X
sku ,k ejωk t ,
xku (t) = √
N k=0
0 ≤ t ≤ Ts .
(33)
Denote by φku (t) the residual phase noise at transmitter ku
after performing phase noise compensation at the transmitter.
The transmitted signal with the residual phase noise is given
by1
x̃ku (t) = ejφku (t) xku (t) ≈ xku (t)
(34)
Denote by hku ,kr (t) the channel between user ku and the
receiver antenna kr . The receiving signal at antenna kr is
ykr (t) =
Nu
X
ku =1
hku ,kr ∗ x̃ku (t) + nkr (t).
(35)
Assuming a common LO for all the receiving antennas (which
is the case in many communication systems), we can model
the received signal with the effect of the phase noise at the
receiver as
zkr (t) = ejφ(t) ykr (t);
(36)
1 Note that (34) displays our assumption that adequate phase noise compensation was performed by the transmitters, thus the remaining phase noise
at the transmitters is negligible.
6
an example of multiple LOs at the receiving end is discussed
for example in [45]. Since we discuss in this example systems
with transmitter phase noise compensation, we can reasonably
assume that ejφ(t) is considerably larger than ejφku (t) . Also, as
before, we assume that the OFDM symbols are synchronized
and that the cyclic prefix has been removed, so that the channel
matrix can be assumed to be circulant, and thus for every
antenna the channel is given by
H ku ,kr =
hku ,kr ,0
hku ,kr ,N −1
..
.
..
.
···
..
.
..
.
..
.
···
hku ,kr ,1
hku ,kr ,0
..
.
···
hku ,kr ,1
···
..
.
..
.
hku ,kr ,N −1
hku ,kr ,N −2
..
.
..
.
hku ,kr ,0
hku ,kr ,N −1
.
(47)
where 1Nr is a column vector of ones of size Nr . Define
W = (I Nu ⊗ F N )BZ(I Nr ⊗ V )(1Nr ⊗ I N ).
(48)
Since
(1Nr ⊗ (V γ)) = (I Nr ⊗ V )(1Nr ⊗ I N )γ
(49)
it follows from (47) and from our assumption that adequate
phase noise compensation was performed by the transmitters
s ≈ W γ.
(50)
We can obtain the linear LS estimation by solving
2
γ̂ = arg min ks − W γk .
γ
[xT1 , . . . , xTNu ]T
(51)
Therefore, as before, we obtain that
F ∗N sku .
(38)
γ̂ = W †p sp ,
x̃ku = ejΦku xku
(39)
where W p is obtained by choosing the rows that correspond
to pilot tones alone of the different users. Alternatively, we
can solve the respective TLS problem.
xku =
Further, let
where Φku = diag (φku ,0 , . . . , φku ,N −1 ) and φku ,m =
φku (mTs ). Let x̃ = [x̃T1 , . . . , x̃TNu ]T , we have that
Nu
X
H ku ,kr x̃ku + nkr
(52)
V. T RACKING THE DOMINANT SUBSPACE OF THE
K ARHUNEN L O ÈVE REPRESENTATION
jΦ
z1
e y1
..
..
jΦ
z= . =
= (I Nr ⊗ e )y.
.
z Nr
ejΦ y Nr
(43)
Y = diag(y)
(44)
There are several methods to obtain the correlation matrix
Rψψ . The first is to pre-calibrate it and generate fixed basis
vectors that are either measured or computed from the LO
design. Hence, one can estimate Rψψ from the data and
apply an eigen-decomposition to obtain the basis vectors.
These alternatives are hard to implement since they depend on
component variability in manufacturing and therefore needs to
be performed for each chip separately. This might also lead to
performance degradation due to environmental changes such
as temperature or vendor dependent behavior. Furthermore,
residual carrier offset also affects the optimal basis. A better
choice that makes it possible to overcome the non-stationarity
of the phase noise process is to track a basis for the phase
noise subspace. To track the phase noise vectors we propose
using the PAST algorithm [41]. Since we do not require our
basis elements to be orthogonal we do not need the deflation
approach of PAST-d (see [41]). Alternatively, the subspace can
be tracked using traditional methods such that appear in [46].
The PAST algorithm is implemented as follows.
Z , diag(z) = Y (I Nr ⊗ ejΦ ),
(45)
A. The PAST algorithm for subspace tracking
ku =1
= H kr x̃ + nkr .
(40)
Defining Φ as in (13), we write
z kr = ejΦ y kr = ejΦ (H kr x̃ + nkr ).
Now, let H = [H T1 , . . . , H TNr ]T and
define y and z as follows
H 1 x̃ + n1
y1
..
..
y= . =
.
y Nr
H Nr x̃ + nNr
Denote
s ≈ (I Nu ⊗ F N )BZ(1Nr ⊗ (V γ))
(37)
Let H kr = [H 1,kr , . . . , H Nu ,kr ] and x =
where
y kr =
Choosing a ZF beamforming matrix B and assuming that
Nr > Nu , we have that
(41)
n = [nT1 , . . . , nTNr ]T ;
= H x̃ + n
(42)
it follows that
2
where ⊗ is the Kronecker product . Let V be defined to be a
basis of Rψψ (see (5)), and define γ as in (16) and (17). We
can conclude that
(I Nr ⊗ ejΦ )(I Nr ⊗ (V γ)) = I N ·Nr
(46)
2 We note that in Figure 1 and Figure 2 the symbol ⊗ denotes the scalar
multiplication. However, this is a special case of the Kronecker product.
(d)
Let V 0 be a matrix composed of the d low frequency
vectors of the DFT matrix or any other a-priori estimate of
the phase noise dominant eigenvectors (i.e., eigenvectors of
Rψψ corresponding to higher eigenvalues). Let P 0 = I d×d .
(d)
At OFDM symbol m we use V m−1 to perform the phase noise
removal in the time domain as described above. The estimated
symbols in the frequency domain ŝ(m) (k) are used to remove
7
Let λ be an eigenvalue of Rψψ and v its respective eigenvec∗
−1
tor. By the definition of c, diag (c) = diag (c) , thus
Initialize β,
(d)
y m = (V m−1 )∗ ϕ̂m ,
∗
Rψ̃ψ̃ diag (c) v = diag (c) Rψψ diag (c) diag (c) v
hm = P m−1 y m ,
1
gm =
β+y ∗
m hm
Pm =
1
β
= diag (c) Rψψ diag (c)
hm ,
−1
diag (c) v
= diag (c) Rψψ v
(P m−1 − g m h∗m ) ,
= λdiag (c) v
(58)
(d)
em = xm − V m−1 y m ,
(d)
(d)
V m = V m−1 + em g ∗m .
TABLE I: PAST algorithm for subspace tracking.
the desired signal from the received time domain signal and
estimate the phase noise process at each time by
ϕ̂m (t) = e
P −1 (m)
−1
j] zm (t)( N
(k)ĥ(m) (k)ejωk t )
k=0 ŝ
, 0 ≤ t ≤ Ts .
(53)
Rψψ is updated by
m−1
∗
Rm
ψψ = (1 − α)Rψψ + αϕm ϕm .
(54)
(d)
Vm
is updated using the PAST algorithm as described in
Table I (see [41]).
Note that to track the subspace we do not need to compute
Rψψ and that is optional. Figure 2 depicts this phase noise
compensation scheme with subspace tracking.
and we conclude that diag (c) v is an eigenvector of Rψ̃ψ̃
for the eigenvalue λ. It follows that the basis elements of
Rψψ ; i.e., the covariance matrix of the PN process without the
residual frequency offset, are multiplied by exponentials of the
form c(t) = ej2π∆f t . This is especially appealing when we
implement the adaptive tracking of the KLT basis elements, as
will be demonstrated in the simulations. Note that the subspace
tracking algorithm, does not require ∆f to be known, but it
is affected by it implicitly.
C. The asymptotic behavior of the PAST algorithm
The asymptotic behavior of the PAST and PASTd algorithms is discussed in [47]–[49] for real valued independent
identically distributed (i.i.d.) Gaussian random vectors.
Under the reasonable assumption that the random process ψ
is M -dependent, we can sample ψ every M + 1 samples. This
will slow the convergence by a linear factor of M + 1. To use
the results of [47]–[49] the problem can be represented as a
real problem, using standard transformation. Let,
Re(V ) −Im(V )
V 2rep =
(59)
Im(V ) Re(V )
it follows that
B. Residual carrier offset
In the presence of residual carrier offset there is another time
varying multiplicative noise; namely, the slowly varying residual carrier. While for small values (e.g., 1 ppm) the residual
carrier does not affect the decoding, it does have a detrimental
effect on the common phase error estimation since the phase is
no longer fixed at the pilots. Interestingly, this residual carrier
can be incorporated into the KLT representation of the phase
noise process by replacing ψ(t) with the random process ψ̃(t)
ψ̃(t) = ej2π∆f t ψ(t)
= ej(φ(t)+2π∆f t)
(55)
and then tracking the subspace of Rψ̃ψ̃ instead of Rψψ .
We note that ψ̃(mTs ) and ψ̃ are defined similarly to the
scenario in which there is no residual carrier offset; that is,
ψ̃m = ψ̃(mTs ) and ψ̃ = [ψ̃1 , . . . , ψ̃N ]T . Let c(t) = ej2π∆f t
and cm = c(mTs ). We can represent Rψ̃ψ̃ by the following
identity
∗
Rψ̃ψ̃ = diag (c) Rψψ diag (c)
(56)
c = [c1 , . . . , cN ]T .
(57)
where
Re(e−jϕ )
Re(γ)
= V 2rep
Im(e−jϕ )
Im(γ)
(60)
and we can track the subspace of the matrix V 2rep . Thus,
assuming that the vectors ψ are circularly symmetric complex
normal vector, we can utilize the results of [48], [49] and
have that under mild conditions, the subspace V (d)
m converges
to V (d) with probability 1 as m → ∞.
VI. E XPERIMENTAL A NALYSIS
In this section we present the simulated and experimental
analysis of the proposed PN compensation scheme in subsection III-A. This section is divided into three parts: Section VI-A covers simulation tests of the PN compensation
scheme with no tracking. Section VI-B is dedicated to the
analysis of the compensation scheme with no tracking performed on the measured PN. Last, Section VI-C is dedicated
to the analysis of the tracking algorithm that was proposed in
Section V. This analysis is carried out on the measured PN of
Section VI-B. A detailed description of the measurements that
were used to produce the figure in this sections is included
in [50]. The SNR was chosen such that the phase noise is
the dominant noise, and limits the reception of 256 QAM.
Still, a noise figure of 7 dB is quite high and even with this
strong noise the phase noise is still dominant. Additionally, PN
compensation complicates the design of the receiver, therefore
8
ϕ1
..
.
PAST algorithm
Table I
..
.
ϕN
V̂ 1
...
...
V̂ d
V̂ 1
...
V̂11
pilot 1
γ̂1
null tone 1
..
.
LS
(TLS)
solver
...
V̂1N
⊗
..
.
pilot kp
V̂ d
...
V̂d1
V̂dN
⊗
..
.
γ̂d
⊗
null tone kn
⊗
P
z1
P
⊗
⊗
zN
y1
yN
...
FFT
...
To FEQ
Fig. 2: Phase noise compensation process utilizing the PAST algorithm.
Channels frequency response
it is the most cost effective when the PN is the dominant
noise which typically occurs when transmitting high order
constellations (i.e., high SNR) together with multi-antenna
receiver.
We now present a simulated experiment testing the performance of the algorithm on real WLAN channels. This
set of simulations assumes that the actual estimate of the
eigenvectors of Rψψ is given. An example of a measured
channel is depicted in Figure 3. The transmitted power was
10 dBm, the assumed noise figure was 7 dB, the thermal
noise was −174 dBm/Hz and the bandwidth was 20 MHz.
The phase noise process was generated using a second order
Chebychev type I and a PSD of the phase noise process is
depicted in Figure 4. The standard deviation of the phase noise
was σφ = 3◦ . At each time we used two receive channels and
tone numbers 1−7, 21, 43, 58−64 at each of the two receivers
for pilot symbols np = 16. The OFDM had 64 tones; i.e.,
N = 64 and the modulation was 256 QAM.
Figure 5 depicts the dependence of the residual Error Vector
Magnitude (EVM) on the number of basis elements using KL
basis vectors and DFT basis vectors which the algorithm that
was presented in [30]. The KL eigenvectors were computed
based on 10 OFDM symbols. We averaged the EVM over
300 measured pairs of channels (1 × 2 systems, i.e., one
transmitting antenna and two receiving antennas). For each
channel we averaged the EVM over 100 OFDM symbols.
Note that for a large number of basis elements, the EVM may
be larger due to the insufficient number of equations when
WLAN channel 1
WLAN channel 2
−70
Amplitude [dB]
A. Simulation tests
−65
−75
−80
−85
−90
−95
−100
0
2
4
6
8
10
12
14
16
18
20
Freq [MHz]
Fig. 3: Two 20 MHz receive channels.
more basis elements are used than pilot symbols. The large
gain of our compensation scheme compared to the DFT basis
presented in [30] is clearly visible. Furthermore, the maximal
tolerated EVM for 256 QAM for 5G is -32dB for a SISO
transmission. The typical values for EVM are between -40dB
and -35dB for MIMO channels. As can be seen, the method of
[30] narrowly meets the requirements while our compensation
scheme is well in the desired region.
To obtain performance under good phase noise and channel
conditions we repeated the experiment with a simulated phase
noise with standard deviation of σφ = 0.7◦ and with the
channel attenuation reduced by 5dB compared to Figure 3.
9
Simulated phase noise PSD
0
−20
−20
Power Spectrum Magnitude [dB]
Power Spectrum Magnitude [dB]
Simulated phase noise PSD
0
−40
−60
−80
−100
−120
−140
−160
−40
−60
−80
−100
−120
−140
−160
0
2
4
6
8
10
12
14
16
18
−180
20
0
2
4
6
Frequency [MHz]
Fig. 4: Simulated phase noise process PSD. σφ = 3◦ .
12
14
16
18
20
EVM vs. estimation order
−25
KL based compensation
DFT based compensation [30]
−15
10
Fig. 6: Simulated phase noise PSD. σφ = 0.7◦ .
EVM vs. estimation order
−10
8
Frequency [MHz]
KL based compensation
DFT based compensation [30]
−30
−35
−25
EVM [dB]
EVM [dB]
−20
−30
−35
−40
−45
−40
−50
−45
−50
0
2
4
6
8
10
12
14
16
18
20
No. of basis elements
Fig. 5: EVM vs. number of basis elements for DFT (which
is the compensation scheme in [30]) and KL based methods.
Phase noise 3◦ . 0 - No compensation. 1 - CPE compensation.
σφ = 3◦
The results are presented in Figures 6-7. Even in this case
there was a substantial gain achieved by cancelling the phase
noise although the phase noise performance was reasonable
even with CPE compensation alone.
We also simulated the performance of our phase noise compensation scheme and that of Casas et al. [30], as a function
of the of the standard deviation of the phase noise. That is, we
chose a fixed number of 8 basis elements, and simulated the
EVM as a function of the phase noise standard deviation. The
results of this simulation are depicted in Figure 8. It is clear
that our scheme outperformed the scheme presented in [30].
Specifically, using DFT basis, a phase noise with standard
deviation of 3◦ is the maximal phase noise which can be
tolerated for a constellation of 256QAM, whereas our scheme
can tolerate up to 8◦ standard deviation phase noise. As before
we used in this simulation 300 pairs of measured channels. For
each pair channel we used 100 OFDM symbols; the phase
noise process was produced for each of these channels as
−55
0
2
4
6
8
10
12
14
16
18
20
No. of basis elements
Fig. 7: EVM vs. number of basis elements for DFT (see
[30]) and KL based methods. 0 - No compensation. 1 - CPE
compensation. σφ = 0.7◦
described at the beginning of Section VI-A.
In addition, we performed a coded simulation using a
convolutional code with rate 12 , interleaver and soft-decision
Viterbi decoding; its results are depicted in Figure 9. As before
we used in this simulation 300 pairs of measured channels.
For each pair channel we used 100 OFDM symbols; in total
we used 7.68 × 106 uncoded bits. The phase noise process
was produced for each of these channels as described at the
beginning of Section VI-A. As can be seen from Figure 9
phase noise compensation is essential for correct decoding at
the receiver. Further, the KL compensation scheme continued
to outperform the DFT compensation scheme with bit error
rate (BER) of less than 10−6 .
The last simulation of this section depicts the performance
of the phase noise compensation scheme for multi user MIMO
system (see Section IV-B) with two single antenna transmitters
and two antenna receiver (2 × 2 system); Figure 10 depicts an
example of the channels of a 2×2 communication system. The
transmitted power of each transmitting antenna was 10 dBm.
10
EVM vs. phase noise standard deviation
Channels frequency response
−10
−65
−15
−70
Amplitude [dB]
EVM [dB]
−20
−25
−30
−35
−75
−80
−85
−90
−40
−45
KL based compensation
DFT based compensation [30]
−50
1
2
3
4
5
6
7
8
9
10
11
−100
12
2
4
6
8
10
channel
channel
channel
channel
12
14
11
22
12
21
16
18
20
Freq [MHz]
Fig. 8: EVM vs. phase noise standard deviation. 8 basis
elements.
Fig. 10: Four 20 MHz receive channels of 2×2 communication
system.
EVM vs. receiver phase noise standard deviation◦
BER vs. phase noise standard deviationcirc
−10
KL based compensation
DFT based compensation [30]
No PN compensation
10−1
0
◦
Standard deviation
100
WLAN
WLAN
WLAN
WLAN
−95
−15
−20
EVM [dB]
BER
10−2
10−3
10−4
−25
−30
−35
−40
10−5
10−6
KL based compensation
DFT based compensation [30]
KL based compensation - TX PN
DFT based compensation - TX PN [30]
−45
1
1.5
2
2.5
3
3.5
4
4.5
5
5.5
6
◦
Standard deviation
−50
1
1.5
2
2.5
3
3.5
4
4.5
5
5.5
6
Receiver’s phase noise standard deviation◦
Fig. 9: BER vs. phase noise standard deviation. 8 basis
elements.
Fig. 11: EVM vs. phase noise standard deviation (at the
receiver). 8 basis elements.
For each of the transmitting antennas we generated a phase
noise process with standard deviation of 1◦ which represents
the residual phase noise process at the transmitter. The two
phase noise processes at the transmitters were generated independently, and were also independent of the receiver phase
noise process. We compared the performance of the system
with and without transmitter phase noise. The results of the
simulation are presented in Figure 11, the receiver phase noise
standard deviation was varied from 1◦ to 6◦ . We can clearly
see from Figure 11 that when the receiver phase noise standard
deviation is greater than 1◦ , the performance of the phase noise
compensation schemes with or without transmitter phase noise
are very close to one another. Clearly, the method presented in
this paper does not break down in the presence of independent
residual transmitter phase noise. Note that as before we used in
this simulation 300 pairs of measured channels. For each pair
channel we used 100 OFDM symbols; the receiver phase noise
process was produced for each of these channels as described
at the beginning of Section VI-A.
B. Measured phase noise analysis
In the second set of simulated experiments we used measured phase noise samples. The measured signal comprised a
sine wave at 5 MHz, sampled at 40 MHz and then filtered to
remove the original sine wave of 5 MHz, thus only the phase
noise remained. The PSD of the phase noise is depicted in
Figure 12. We repeated the experiment above with samples
of the measured phase noise. The results are depicted in
Figure 13. There was a clear gain of 5dB for 5 basis elements
and above.
C. Simulations of the tracking algorithm for measured phase
noise
In this section we analyze the tracking capability of the
PAST algorithm combined with decision direction. We used
3 We note that the two spikes in the PSD of the measured phase noise
process are due to residual of the sine and its harmony of the sine wave at 5
MHz.
11
EVM with phase noise compensation and tracking
−28
−20
−30
Smoothed EVM (64 symbols) [dB]
Power Spectrum Magnitude [dB]
Measured phase noise PSD
0
−40
−60
−80
−100
−120
−140
−160
−180
−200
0
2
4
6
8
10
12
14
16
18
−32
−34
−36
−38
−40
−42
−44
−48
20
CPE compensation
DFT 4 basis elements [30]
Tracking using PAST
−46
0
100
200
EVM vs. estimation order
Smoothed EVM (64 symbols) [dB]
EVM [dB]
−25
−30
−35
−40
−45
0
1
2
3
4
5
6
600
700
800
900
1,000
−30
−20
−50
500
EVM with phase noise compensation and tracking
KL based compensation
DFT based compensation [30]
−15
400
Fig. 14: Tracking performance of PAST. σφ = 3.5◦ , residual
carrier 1 ppm, 4 basis vectors.
.
Fig. 12: PSD of measured phase noise.3
−10
300
OFDM symbol no.
Frequency [MHz]
7
8
9
10
No. of basis elements
Fig. 13: EVM vs. number of basis elements for DFT and KL
based methods. No. of basis element: 0 - No compensation. 1
- CPE compensation. Measured phase noise.
the same system as in previous sections with the measured
phased noise (scaled to 3.5◦ total phase noise) and 4 basis
elements. However, we also added a 1ppm residual carrier to
model carrier offset for a carrier frequency of 5 GHz. It is
clear (see Figure 14) that our tracking scheme yields better
phase and residual carrier compensation of 5dB, compared to
the other methods, i.e., the DFT of [30] and a simple CPE
compensation.
To test the stationarity of the basis elements after convergence was achieved; we repeated the experiment when no
residual carrier was present and limited the training phase
to 300 symbols. This is important for the quality of the
initial basis, based on previous estimates. We performed two
simulations whose results appear in Figure 15 and Figure
16. Figure 15 depicts the results of the simulation for the
two receive channels that are depicted in Figure 3. Figure
16 depicts the results of the simulation averaged over 300
pairs of receive channels. It is clear that the basis vectors were
relatively stationary since the EVM was fixed for the duration
−32
−34
−36
−38
−40
−42
−44
−46
CPE compensation
DFT 4 basis elements [30]
Tracking using PAST
−48
−50
0
500
1,000
1,500
2,000
2,500
3,000
OFDM symbol no.
Fig. 15: Limited training of 300 symbols. No residual carrier,
σφ = 3.5◦ and β = 0.9, one pair of receive channels
.
of the next 2700 symbols. Further, the estimates were quite
good and the mean EVM remained constant. This suggests that
our stationarity assumption was sufficiently good. Note that
both common phase removal and DFT based compensation
were not as good and led to a 2 − 4dB loss.
To depict the advantages of tracking we performed the
following simulation over the measured channels, we set the
residual carrier to be 5 ppm and simulated phase noise of
3.5◦ , we averaged the results over 300 pairs of channels. We
stopped the tracking of the limited tracking after 250 OFDM
symbols and performed the tracking for the “Tracking using
PAST” line during the whole simulation. As can be seen from
Figure 17, tracking dramatically improves the behavior of the
phase noise compensation scheme presented in this paper.
VII. C ONCLUSION
We presented a novel phase noise estimation technique. This
technique leads to considerable reduction in phase noise and in
12
EVM with phase noise compensation and tracking
−30
Smoothed EVM (64 symbols) [dB]
−32
−34
−36
−38
−40
−42
−44
−46
CPE compensation
DFT 4 basis elements [30]
Tracking using PAST
−48
−50
0
500
1,000
1,500
2,000
2,500
3,000
OFDM symbol no.
Fig. 16: Limited training of 300 symbols. No residual carrier,
σφ = 3.5◦ and β = 0.9, averaged over 300 pairs of receive
channels
.
EVM with phase noise compensation and tracking
Smoothed EVM (64 symbols) [dB]
−15
−20
−25
−30
−35
CPE compensation
DFT 4 basis elements [30]
Limited Tracking using PAST
Tracking using PAST
−40
−45
0
100
200
300
400
500
600
700
800
900
1,000
OFDM symbol no.
Fig. 17: Comparison of the performance of phase noise
compensation schemes with or without tracking.
particular works very well for strong phase noise. The specific
basis proposed in this paper (using the eigenvectors of the
noise correlation process) accounts primarily for its success
compared to previous work. We identified several possible
extensions of the method to multi user MIMO systems as well
as online calibration and exploitation of null tones. Finally, we
tested the possibility of decision directed tracking of the basis
vectors. The tracking results using measured phase noise and
residual carrier offset suggest that our stationarity assumption
also holds.
ACKNOWLEDGEMENT
The authors would like to thank the anonymous reviewers
and the associate editor for their helpful and constructive
comments which helped improve the content of this paper.
R EFERENCES
[1] A. Leshem, Patent US 7 609 789 B2, October, 2009.
[2] 3GPP, “3GPP Home Page.” [Online]. Available: http://www.3gpp.org/
technologies/keywords-acronyms/98-lte
[3] IEEE, “IEEE Standards Association.” [Online]. Available: http:
//standards.ieee.org/about/get/802/802.11.html
[4] ——, “IEEE Standards Association.” [Online]. Available: https:
//standards.ieee.org/about/get/802/802.16.html
[5] P. Banelli, S. Buzzi, G. Colavolpe, A. Modenini, F. Rusek, and
A. Ugolini, “Modulation formats and waveforms for 5G networks: Who
will be the heir of OFDM?: An overview of alternative modulation
schemes for improved spectral efficiency,” IEEE Signal Processing
Magazine, vol. 31, no. 6, pp. 80–93, Nov 2014.
[6] F. Schaich and T. Wild, “Waveform contenders for 5G - OFDM vs.
FBMC vs. UFMC,” in Communications, Control and Signal Processing
(ISCCSP), 2014 6th International Symposium on, May 2014, pp. 457–
460.
[7] F. Schaich, T. Wild, and Y. Chen, “Waveform contenders for 5G suitability for short packet and low latency transmissions,” in 2014 IEEE
79th Vehicular Technology Conference (VTC Spring), May 2014, pp. 1–
5.
[8] C. Ibars, U. Kumar, H. Niu, H. Jung, and S. Pawar, “A comparison of
waveform candidates for 5G millimeter wave systems,” in 2015 49th
Asilomar Conference on Signals, Systems and Computers, Nov 2015,
pp. 1747–1751.
[9] L. Tomba, “On the effect of wiener phase noise in OFDM systems,”
IEEE Transactions on Communications, vol. 46, no. 5, pp. 580–583,
May 1998.
[10] A. G. Armada and M. Calvo, “Phase noise and sub-carrier spacing
effects on the performance of an OFDM communication system,” IEEE
Communications Letters, vol. 2, no. 1, pp. 11–13, Jan 1998.
[11] L. Piazzo and P. Mandarini, “Analysis of phase noise effects in OFDM
modems,” IEEE Transactions on Communications, vol. 50, no. 10, pp.
1696–1705, Oct 2002.
[12] A. G. Armada, “Understanding the effects of phase noise in orthogonal frequency division multiplexing (OFDM),” IEEE Transactions on
Broadcasting, vol. 47, no. 2, pp. 153–159, Jun 2001.
[13] D. Petrovic, W. Rave, and G. Fettweis, “Effects of phase noise on OFDM
systems with and without PLL: Characterization and compensation,”
IEEE Transactions on Communications, vol. 55, no. 8, pp. 1607–1616,
Aug 2007.
[14] L. Barletta, M. Magarini, and A. Spalvieri, “The information rate
transferred through the discrete-time wiener’s phase noise channel,”
Journal of Lightwave Technology, vol. 30, no. 10, pp. 1480–1486, May
2012.
[15] S. Wu and Y. Bar-Ness, “A phase noise suppression algorithm for
OFDM-based WLANs,” IEEE Communications Letters, vol. 6, no. 12,
pp. 535–537, Dec 2002.
[16] ——, “OFDM systems in the presence of phase noise: consequences
and solutions,” IEEE Transactions on Communications, vol. 52, no. 11,
pp. 1988–1996, Nov 2004.
[17] S. Wu, P. Liu, and Y. Bar-Ness, “Phase noise estimation and mitigation
for OFDM systems,” IEEE Transactions on Wireless Communications,
vol. 5, no. 12, pp. 3616–3625, December 2006.
[18] K. Nikitopoulos, S. Stefanatos, and A. K. Katsaggelos, “Decision-aided
compensation of severe phase-impairment-induced inter-carrier interference in frequency-selective OFDM,” IEEE Transactions on Wireless
Communications, vol. 8, no. 4, pp. 1614–1619, April 2009.
[19] Y. Gong and X. Hong, “A new algorithm for OFDM joint data detection
and phase noise cancellation,” in 2008 IEEE International Conference
on Communications, May 2008, pp. 636–640.
[20] D. D. Lin and T. J. Lim, “The variational inference approach to joint
data detection and phase noise estimation in OFDM,” IEEE Transactions
on Signal Processing, vol. 55, no. 5, pp. 1862–1874, May 2007.
[21] S. Suyama, H. Suzuki, K. Fukawa, and J. Izumi, “Iterative receiver employing phase noise compensation and channel estimation for millimeterwave OFDM systems,” IEEE Journal on Selected Areas in Communications, vol. 27, no. 8, pp. 1358–1366, October 2009.
[22] F. Munier, T. Eriksson, and A. Svensson, “An ICI reduction scheme
for OFDM system with phase noise over fading channels,” IEEE
Transactions on Communications, vol. 56, no. 7, pp. 1119–1126, July
2008.
[23] F. Septier, Y. Delignon, A. Menhaj-Rivenq, and C. Garnier, “Monte
carlo methods for channel, phase noise, and frequency offset estimation
with unknown noise variances in OFDM systems,” IEEE Transactions
on Signal Processing, vol. 56, no. 8, pp. 3613–3626, Aug 2008.
[24] J. Tao, J. Wu, and C. Xiao, “Estimation of channel transfer function
and carrier frequency offset for OFDM systems with phase noise,” IEEE
13
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
Transactions on Vehicular Technology, vol. 58, no. 8, pp. 4380–4387,
Oct 2009.
D. D. Lin, R. A. Pacheco, T. J. Lim, and D. Hatzinakos, “Joint estimation
of channel response, frequency offset, and phase noise in OFDM,” IEEE
Transactions on Signal Processing, vol. 54, no. 9, pp. 3542–3554, Sept
2006.
H. Huang, W. G. J. Wang, and J. He, “Phase noise and frequency
offset compensation in high frequency MIMO-OFDM system,” in 2015
IEEE International Conference on Communications (ICC), June 2015,
pp. 1280–1285.
Q. Zou, A. Tarighat, and A. H. Sayed, “Compensation of phase noise
in OFDM wireless systems,” IEEE Transactions on Signal Processing,
vol. 55, no. 11, pp. 5407–5424, Nov 2007.
O. H. Salim, A. A. Nasir, H. Mehrpouyan, W. Xiang, S. Durrani, and
R. A. Kennedy, “Channel, phase noise, and frequency offset in OFDM
systems: Joint estimation, data detection, and hybrid Cramér-Rao lower
bound,” IEEE Transactions on Communications, vol. 62, no. 9, pp.
3311–3325, Sept 2014.
R. Wang, H. Mehrpouyan, M. Tao, and Y. Hua, “Channel estimation,
carrier recovery, and data detection in the presence of phase noise in
OFDM relay systems,” IEEE Transactions on Wireless Communications,
vol. 15, no. 2, pp. 1186–1205, Feb 2016.
R. Casas, S. Biracree, and A. Youtz, “Time domain phase noise
correction for OFDM signals,” IEEE trans. on Broadcasting, vol. 48,
no. 3, pp. 230–236, September 2002.
F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta,
O. Edfors, and F. Tufvesson, “Scaling up MIMO: Opportunities and
challenges with very large arrays,” IEEE Signal Processing Magazine,
vol. 30, no. 1, pp. 40–60, Jan 2013.
J. Hoydis, S. ten Brink, and M. Debbah, “Massive MIMO in the UL/DL
of cellular networks: How many antennas do we need?” IEEE Journal on
Selected Areas in Communications, vol. 31, no. 2, pp. 160–171, February
2013.
E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive
MIMO for next generation wireless systems,” IEEE Communications
Magazine, vol. 52, no. 2, pp. 186–195, February 2014.
A. Sahai, G. Patel, C. Dick, and A. Sabharwal, “Understanding the
impact of phase noise on active cancellation in wireless full-duplex,”
in 2012 Conference Record of the Forty Sixth Asilomar Conference on
Signals, Systems and Computers (ASILOMAR), 2012, pp. 29–33.
L. Samara, M. Mokhtar, . zdemir, R. Hamila, and T. Khattab, “Residual
self-interference analysis for full-duplex OFDM transceivers under phase
noise and I/Q imbalance,” IEEE Communications Letters, vol. 21, no. 2,
pp. 314–317, Feb 2017.
X. Quan, Y. Liu, S. Shao, C. Huang, and Y. Tang, “Impacts of phase
noise on digital self-interference cancellation in full-duplex communications,” IEEE Transactions on Signal Processing, vol. 65, no. 7, pp.
1881–1893, April 2017.
R. C. Daniels and R. W. H. Jr., “60 GHz wireless communications:
emerging requirements and design recommendations,” IEEE Vehicular
Technology Magazine, vol. 2, no. 3, pp. 41–50, Sept 2007.
P. Smulders, “Exploiting the 60 GHz band for local wireless multimedia
access: prospects and future directions,” IEEE Communications Magazine, vol. 40, no. 1, pp. 140–147, Jan 2002.
X. Yi, W. Shieh, and Y. Ma, “Phase noise effects on high spectral
efficiency coherent optical OFDM transmission,” Journal of Lightwave
Technology, vol. 26, no. 10, pp. 1309–1316, May 2008.
I. Markovsky and S. V. Huffel, “Overview of total least-squares methods,” Signal Processing, vol. 87, no. 10, pp. 2283 – 2302, 2007, special
Section: Total Least Squares and Errors-in-Variables Modeling.
B. Yang, “Projection approximation subspace tracking,” IEEE trans. on
Signal processing, vol. 49, no. 1, pp. 95–107, January 1995.
R. D. DeGroat and E. M. Dowling, “The data least squares problem and
channel equalization,” IEEE trans. on Signal processing, vol. 41, no. 1,
pp. 407–411, January 1993.
W. Peng, Q. Gao, X. Zhang, S. Zhou, and X. Xu, “Analysis of transmitter
phase noise in ofdm system and its mitigation,” in VTC Spring 2009 IEEE 69th Vehicular Technology Conference, April 2009, pp. 1–5.
Z. Xu, G. Ren, and Y. Zhang, “Phase noise pre-correction scheme
for SC-FDMA signals in LTE-uplink,” Chinese Journal of Electronics,
vol. 26, no. 3, pp. 634–639, 2017.
M. R. Khanzadi, G. Durisi, and T. Eriksson, “Capacity of SIMO and
MISO phase-noise channels with common/separate oscillators,” IEEE
Transactions on Communications, vol. 63, no. 9, pp. 3218–3231, Sept
2015.
[46] P. Comon and G. H. Golub, “Tracking a few extreme singular values and
vectors in signal processing,” Proceedings of the IEEE, vol. 78, no. 8,
pp. 1327–1343, Aug 1990.
[47] B. Yang and F. Gersemsky, “Asymptotic distribution of recursive subspace estimators,” in 1996 IEEE International Conference on Acoustics,
Speech, and Signal Processing Conference Proceedings, vol. 5, May
1996, pp. 2856–2859.
[48] B. Yang, “Convergence analysis of the subspace tracking algorithms
PAST and PASTd,” in 1996 IEEE International Conference on Acoustics, Speech, and Signal Processing Conference Proceedings, vol. 3, May
1996, pp. 1759–1762.
[49] ——, “Asymptotic convergence analysis of the projection approximation
subspace tracking algorithms,” Signal Processing, vol. 50, no. 1-2, pp.
123–136, April 1996.
[50] A. Leshem, N. Tal, L. Kravitz, and E. Gerson, “Super-resolution
technique for estimating MIMO WLAN channels with application to
5 GHz channel measurements,” in 2004 IEEE International Conference
on Acoustics, Speech, and Signal Processing, vol. 4, May 2004, pp.
IV–945–948.
Amir Leshem (M’98-SM’06) received the B.Sc. degree (cum laude) in
mathematics and physics and the M.Sc. (cum laude) and Ph.D. degrees in
mathematics from the Hebrew University, Jerusalem, Israel, in 1986, 1990,
and 1998, respectively. He is a Professor and one of the founders of the Faculty
of Engineering at Bar-Ilan University, Ramat-Gan, Israel, where he heads the
Communications track. He held visiting positions with Delft University, Delft,
The Netherlands, and Stanford University, Stanford, CA, USA, in 2008 to
2009. From 2000 to 2003, he was Director of Advanced Technologies with
Metalink Broadband, Yakum, Israel, where he was responsible for research
and development of new DSL and wireless MIMO modem technologies and
served as a member of ITU-T SG15, ETSI TM06, NIPP-NAI, and IEEE 802.3
and 802.11. From 2000 to 2002, he was also a Visiting Researcher with
Delft University of Technology. From 2003 to 2005, he was the Technical
Manager of the U-BROAD consortium developing technologies to provide
100 Mb/s and beyond over copper lines. From 1998 to 2000, he was with
the Faculty of Information Technology and Systems, Delft University of
Technology, as a Postdoctoral Fellow working on algorithms for the reduction of terrestrial electromagnetic interference in radio-astronomical radiotelescope antenna arrays and signal processing for communication. His main
research interests include multichannel wireless and wireline communication,
applications of game theory to dynamic and adaptive spectrum management
of communication networks, software defined networks, array and statistical
signal processing with applications to multiple-element sensor arrays and
networks, wireless communications, radio-astronomical imaging and brain
research, Information processing in social networks, set theory, logic, and
foundations of mathematics.
Prof. Leshem served as an Associate Editor for the IEEE TRANSACTIONS ON SIGNAL PROCESSING from 2008 to 2011 and as a Guest Editor
for several special issues of the IEEE Signal Processing Magazine and the
IIEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING. Since
2010, he has been a member of the IEEE technical committee on Signal
Processing for Communications and Networking.
Michal Yemini received the B.Sc. in computer engineering from the
Technion-Israel Institute of Technology, Haifa, Israel, in 2011. She is currently
completing her Ph.D. degree in the joint M.Sc.-Ph.D. program in the Faculty
of Engineering, Bar-Ilan University, Ramat-Gan, Israel. Her main research
interests include information theory, connectivity of large heterogeneous
networks, dynamic spectrum allocation and resource management for wireless
networks.
| 7cs.IT
|
arXiv:0907.0749v1 [cs.PL] 4 Jul 2009
Function Interface Models for Hardware
Compilation: Types, Signatures, Protocols
Dan R. Ghica
University of Birmingham
[email protected]
February 24, 2013
Abstract
The problem of synthesis of gate-level descriptions of digital circuits
from behavioural specifications written in higher-level programming languages (hardware compilation) has been studied for a long time yet a
definitive solution has not been forthcoming. The argument of this essay
is mainly methodological, bringing a perspective that is informed by recent developments in programming-language theory. We argue that one
of the major obstacles in the way of hardware compilation becoming a
useful and mature technology is the lack of a well defined function interface model, i.e. a canonical way in which functions communicate with
arguments. We discuss the consequences of this problem and propose a
solution based on new developments in programming language theory. We
conclude by presenting a prototype implementation and some examples
illustrating our principles.
1
Introduction
The problem of hardware compilation turned out to be surprisingly difficult.
Although the pioneering work of van Berkel [31], Page, Luk and the Oxford
hardware compilation group [23, 18, 13] yielded promising initial results, more
than a decade later this technology has yet to enter the mainstream of digital
design. Several C-to-gates hardware compilers, such as Roccc [7], Cash [6]
and HandelC [8] have been developed, but their take-up was limited. Recently,
Mentor has introduced its own C-to-gates compiler, Catapult-C1 but it is too
early to evaluate its impact. In this paper, when we refer to existing hardware
compilers we mean these above. We will not refer to design flows based on
SystemC2 or CoWare3 , hardware compilers based on process calculi (e.g. [33,
1 http://www.mentor.com/products/c-based_design/catapult_c_synthesis
2 http://www.systemc.org
3 http://www.coware.com
1
30]), or higher-order structural languages such as Lava [5]; these are interesting
and useful, but conceptually different ways of approaching VLSI design.
So, why did hardware compilers not meet expectations?
Part of the answer has to do with performance, as higher-level behavioural
design is unlikely to be as efficient as structural or low-level, hand-crafted designs. However, the economics of reconfigurable architectures such as fieldprogrammable gate arrays (FPGAs) or other complex programmable logic devices (CPLDs) are such that the design costs can often become the overriding
concern, allowing for weaker performance as a trade-off, especially since their
competition is often not custom-design application-specific circuits (ASIC) but
(embedded) software. This is a well known consideration, so hardware compilers
often target reconfigurable architectures.
Another performance-related issue is that of concurrency; the performance
advantage of hardware comes from its potential for massive parallelism, rather
than high clock rates. Conventional programming languages that can serve as
candidates for compilation into hardware (C, Java, etc.) have either no built-in
support for it or offer unsuitable concurrency models such as threads or processes. This is a serious dilemma, because a concurrency model needs to be both
low-level enough to reflect the underlying capabilities of the platform (“closeto-metal ”) and thus give predictable performance, but also high-level enough
to unburden the programmer from detailed management and book-keeping of
resources. An example of a successful such trade-off is Nvidia’s CUDA platform4 . However, the importance of this argument must not be overstated. A
recent study showed that a large class of non-trivial structural designs can be recreated automatically by a hardware compiler from behavioural specifications,
using standard optimisation techniques such a parallelisation [27].
While the arguments above carry some strength and reflect the conventional
wisdom on the topic, we will argue that we must look deeper in order to discover
the key problems that beset hardware compilation, and to find ways to address
them. In this essay we propose that one such key problem is the lack of canonical
function interface models (FIM). We will briefly examine their traditional role
in software compilation and operating systems, and explain the possible reasons
for their absence in hardware compilers. We also propose a simple but nontrivial FIM, illustrating it with several typical examples and explain how such
FIMs can be canonically designed.
§
In modern programming-language theory the notion of type is paramount. Beyond just classifying data in various categories such as integers, floating-point,
strings, etc., types fill a much more fundamental role. It is difficult to give
a better summary that does justice to the importance of types than Robert
Harper’s5 :
4 http://www.nvidia.com/cuda
5 http://www.cs.cmu.edu/
~rwh/research.htm
2
Over the last two decades type theory has emerged as the central
organizing framework for the design and implementation of programming languages. Type theory (the study of type systems) provides the
theoretical foundation for safe component integration. In the words
of John Reynolds, “a type system is a syntactic discipline for enforcing levels of abstraction”. By syntactic we mean that the type
system is part of the program, rather than purely in the mind of
the implementer. By discipline we mean that type restrictions are
enforced; ill-typed combinations are ruled out by the type system.
By levels of abstraction we mean the clean separation between conceptually distinct data objects that may, in a particular program or
compiler, have the same or similar representations.
Implicit in this definition is an important principle: the power
of a type system lies as much in what it precludes as what it allows.
The most powerful type system of all is the one that rules out all programs. However, such a type system, while surely preventing safety
violations, is hardly very useful. The goal of type system design is to
increase expressiveness by admitting useful programs, but without
compromising safety.
We will see in the following how type systems are indeed essential in hardware
compilation of programming languages because of the special challenge of “safe
component integration” in hardware, as opposed to software.
2
Function interface models
Functions and related concepts (procedures, methods, subroutines, etc.) are
the main mechanism for implementing abstraction in a programming language.
The importance of using functional abstractions lies at the core of any good
programming methodology and its benefits have long been established. Functions play a fundamental role in the operation of a conventional stored-program
computer and they were in fact a feature of the first such computer, the EDSAC [34]. Except for very early models such as the HP 2100, microprocessors
always supported function call in their instruction set directly (e.g. Intel’s x86)
or at least provided instructions for stack management, meant mainly to implement function calls (RISC architectures).
All three initial major modern programming languages (Fortran, Lisp,
Cobol) provided support for functions, albeit sometimes in a clumsy or limited way. Of the three, the most advanced support for functions is found in Lisp.
Not only does it support higher-order functions (functions that take functions
as arguments) but it also introduced the new concept of a Foreign Function Interface: a way for a program written in Lisp to call functions written in another
programming language. This idea was adopted by all subsequent programming
languages that had mature compilers, under one name or another6 . A special
6 In
Java it is called the Java Native Interface (JNI).
3
and privileged position is that of the C programming language; because of the
close relationship between C and the operating system its own calling convention is usually implemented directly by the OS and is called the Application
binary interface (ABI).
One of the decisions of the C and Unix designers with the farthest-reaching
consequences was to make the details of the calling convention public [16]. The
positive effects of this decision cannot be overstated, as it massively improved
the interoperability of applications by opening the OS functionality to other
programming languages and by facilitating support for stand-alone languageindependent libraries.
In this paper, to avoid ambiguity, we introduce the terminology of function
interface model (FIM) to encompass the closely-related issues of the FFI and
ABI.
2.1
FIMs and hardware compilation
It is taken as a given that stored-program computers must offer well-defined
FIMs. It is inconceivable to design a modern operating system or compiler if
this fundamental requirement is not met. On the other hand, in the world of
hardware the concept of a FIM simply did not arise. The net-lists (boxes and
wires) that are the underlying computational models of hardware languages are
not structured in a way that suggests any obvious native FIM.
The abstraction mechanism common in hardware languages such as Verilog
or VHDL, the module 7 , is a form of placing a design inside a conceptual box with
a specified interface. This does serve the purpose of hiding the implementation
details of the module from its user, which is one of the main advantages of
abstraction. However, a module is a less powerful than functional abstraction
as found in programming languages in several significant ways:
1. Modules are always first order. A module can only take data as input
and not other modules; in contrast, functions can take other functions as
input.
2. Modules must be explicitly instantiated. The hardware designer must
manually keep track of the names of each module occurrence and its ports,
which must be connected explicitly by wires to other elements of the design. In contrast, the run-time management of various instances of functions is managed by the OS (using activation records) in a way that is
transparent to the programmer.
3. Sharing a module from various points in a design is not supported by
hardware design languages. The designer must design ad hoc circuitry,
such as multiplexers or de-multiplexers and ad hoc protocols to achieve
this. The lack of language support for sharing makes this a delicate and
error-prone task.
7 Not
to be confused with “modules” as encountered in programming languages such as
ML.
4
These limitations may seem inconsequential, and in a certain sense they are so.
They are not limitations on the expressiveness of hardware design languages
or the performance of the circuits synthesised. However, the impact of these
limitations becomes more serious as designs become more complex. For the
design of, say, an adder the limitations above are irrelevant. But if the design is
an implementation of a complex algorithm that uses many instances of the same
kind of modules, interacting in non-trivial ways the burden of micro-managing
modules and ports and the inability to use generic algorithms such as mapreduce, which are inherently higher-order, can become extremely taxing. The
burden of managing sharing is also substantial, and the conventional solutions to
this problem, such as bus or network-on-chip (NOC) architectures, are complex,
heavy-duty and not provided with language-level support.
Generally, existent hardware compilers are not built around well-defined
FIMs. What would be the unthinkable in the world of software compilation
is the norm in the world of hardware compilation. But this is not entirely
surprising, considering the fact that hardware does not have a “native” FIM to
offer. Its absence damages the usability and the performance of the compilers,
imposes restrictions on interoperability and prevents library support.
Let us briefly consider how functions are managed in the hardware compilers
mentioned earlier.
Roccc only supports functions via inlining, i.e. the textual substitution of the
body of the function at the point of call. This on-the-cheap solution to the
problem of functional abstraction has a serious problem: it makes sharing
of resources impossible. Inlining is a compiler optimisation that targeted
to a conventional computer is sometimes justified, trading off better execution time for increased program size. However, from a hardware compilation perspective it is inefficient as it amounts to the re-instantiating the
module implementing the function every time it is called. The example
that illustrates the failure of this approach best is compilation of floatingpoint programs: the circuits implementing floating-point operations are
expensive and the requirement to re-instantiate multiplication or division
every time they are used cannot be satisfied. The hardware-compilation of
floating-point programs is impossible if these elementary circuits cannot
be shared in the program.
Cash proposes a token-based mechanism that allows the implementation of a
function to be shared, therefore avoiding the need for re-instantiation. The
token-based mechanism is powerful enough to support recursive functions,
which is expressive but inefficient. This is a promising approach, and the
token-based mechanism represents a genuine FIM. The problem with this
approach, inspired by tagged-token data-flow [3], is that it is unnecessarily
heavy-duty. The management of the tokens require sophisticated circuitry
and access to an external RAM for each function call. This is likely to be
expensive in both time and foot-print, and frequent RAM access represents
a bottleneck for concurrency. A beta version of the compiler has been
announced but is not available yet, so we can only speculate regarding the
5
performance penalty that must be paid in overheads for this token-based
mechanism.
Handel C seems to provide static support for functions. The details of the FIM
are not disclosed by Celoxica but evidence can be seen in the generated
HDL. Unfortunately, the simultaneous support for sharing and concurrency in the absence of a type system that would control such behaviour
leads to some surprising race conditions. For example the execution of
the concurrent statement par {x=f(1); y=f(0);} always stores the same
values in registers x,y, regardless of what function f() computes. The reason is that the simple-minded run-time sharing of the implementation for
f() causes a race condition on its shared inputs. Having a race condition
even in the absence of side-effects is too surprising for a conventional programmer and it must be considered a serious defect in the design of the
language. It is not a defect in the implementation of the language because
the Handel C manual warns that such race conditions may occur. But,
concurrent programming, difficult enough as is, becomes almost impossible when confronted with such recondite sources of racing behaviour.
Catapult C and a score of other commercial compilers are difficult to asses.
The system is a closed one, and Mentor does not publish the details of
the support it provides for functions in its technical literature8 . Even
if such languages do have internally a well-defined FIM the fact that it
is undisclosed still represents a significant barrier to interoperability and
run-time support.
The closest correspondence to a FIM in hardware is the concept of bus, a mechanism for establishing logical connections between several components using the
same set of wires. High-performance bus architectures such as ARM’s AMBA9
or IBM’s CoreConnect10 are marketed as especially relevant for “systemson-chip” (SoC) designs, as a way of managing complexity and reduce design
cost and time. The aims of using a bus architecture are similar to those of
using FIMs, but the techniques involved are substantially different. Whereas a
bus is a global mechanism of communication between components, a FIM is a
local mechanism for providing access to a component. Therefore, bus protocols
are complex and layered, more similar to network or distributed-communication
protocols. They can handle transmission delay, transmission errors and transactions. In comparison, the FIMs assume perfect, instantaneous (synchronous)
communication between components. As a result, interfacing components using
a bus requires high overhead and is potential bottleneck for parallelism, whereas
managing the FIMs requires typically very low overhead and has some support
for parallel access. Buses work best at the architectural level, to connect large
components over long distances (e.g. across the boundaries of a chip or of a
8 http://www.mentor.com/products/esl/techpubs/
9 http://www.arm.com/products/solutions/AMBAHomePage.html
10 http://www-306.ibm.com/chips/products/coreconnect/
6
board) whereas FIMs are meant to handle local connections, with low overhead, low latency (at most one clock cycle in our prototype compiler) and high
throughput.
A further elaboration of the concept of bus is the so-called “network on
chip” architecture, which makes communication between components even more
network-like. For complex applications such as SOC it can be a notable improvement in functionality over the simpler bus technology, but at a correspondingly
increased cost [17].
3
The design of a hardware FIM
The previous section looked at the state-of-play in hardware compilation. Hopefully the reader will have accepted the basic point that well-defined FIMs are
important and useful and it is next to impossible to imagine a mature compiler
without them. From here on we will propose a way to derive hardware-friendly
FIMs in a canonical way using some recent developments from programming
language theory.
It is a well-known fact that relying on the primitive concepts of channel,
event and communication makes process calculi good intermediate abstractions
for hardware, and refining processes into hardware-level representation has been
extensively studied [32]. On the other hand, there has been significant research work regarding the encoding of (prototypical) functional programming
languages into process calculi [19].
Game Semantics [1, 15] is a process-calculus-like model for programming
languages introduced in the mid-90s, which proved to be extremely successful
at giving precise interpretations for a variety of languages, thus solving longstanding open problems in the theory of programming languages. Process calculi
are versatile but have little structure, whereas game semantics encapsulates the
right mathematical structure needed to interpret programming languages. Like
process calculi, it is event-oriented and can be refined into hardware.
Therefore we propose the following methodological principle:
Principle 1 Hardware compilation of programming languages via
game semantics is a natural approach based on solid foundational
results.
In the design of the FIM we will consider the following key topics: the static
aspects of the FIM, the dynamic aspects of the FIM, and language-level support
for FIM compliance. To give focus to our presentation we will use a particular language as a case-study, with the following defining features: a functional
fragment based on the affine lambda calculus, combined with the simple imperative language (locally-scoped block variables, iteration, branching), with boolean
variables only11 . This language is a simplified version of Syntactic Control of Interference (SCI), a language which has been studied extensively [24, 25, 21, 20].
11 Adding other finite data-types such as integers or floating point is conceptually straightforward, but the technical complications would detract from the main point.
7
The primitive types of the language are commands, memory cells and (boolean)
expressions: σ ::= com | cell | exp. Additionally, the language contains function
types and products:
θ ::= σ | θ × θ0 | θ → θ.
What is peculiar about the types above is that pairs of terms may share identifiers but functions may not share identifiers with their arguments. Terms
have types, described by typing judgments of the form Γ ` M : θ, where
Γ = x1 : θ1 , . . . xn : θn is a variable type assignment, M is a term and θ
the type of the term.
Identity
x:θ`x:θ
Γ`M :θ
Weakening
Γ, x : θ0 ` M : θ
Γ, x : θ0 ` M : θ
→ Introduction
Γ ` λx.M : θ0 → θ
Γ ` F : θ0 → θ
∆ ` M : θ0
→ Elimination
Γ, ∆ ` F M : θ
Γ ` M : θ0
Γ`N :θ
× Introduction
Γ ` hM, N i : θ0 × θ
The language contains the standard constructs for structured state manipulation
and control. It is convenient to present these constructs in a functional form:
1 : exp
constant
0 : exp
constant
skip : com
no-op
asg : cell × exp → com
assignment
der : cell → exp
dereferencing
seq : com × com → com
sequencing
par : com → com → com
parallel execution
op : exp × exp → exp
logical operations
if : exp × com × com → com
while : exp × com → com
branching
iteration
newvar : (cell → com) → com
local variable.
Product has syntactic precedence over arrow, which associates to the right.
For example, a term with conventional syntax, such as
integer x;
x = 0;
if (x<>y) x=!x;
would be written as
newvar(λx.
8
seq(asg(x, 0),
if(neq, asg(x, not x), skip))).
The interesting structural feature of the type system SCI is allowing sharing of
identifiers (contraction) in product-formation but disallowing it in function application. Reynolds, the inventor of SCI, was interested in this rule to eliminate
covert interference between terms that ostensibly do not share identifiers, hence
the name. This type system facilitates reasoning about program correctness
and it has eventually led to the development of bunched [22] and separation
logic [26].
This restriction can be exploited in several ways, as noticed in the different type signature on sequential (uncurried) versus parallel (curried) composition. This makes terms such as λx.x; x are legal, while λx.x || x is illegal.
Another consequence of this restriction is that nested function application as in
λf λx.f (f (x)) is also banned. We note that the type-system SCI is an instance
of the type-system Syntactic Control of Concurrency (SCC), which has been
studied using game semantics [12]12 .
3.1
Static interfaces for FIM in hardware
As one would perhaps expect, there will be a strong connection between the
interface of a hardware module and the signature of the function or, more generally, the term, it implements.
The game-semantic model interprets types as so-called arenas, which are
structured sets of basic actions called moves.
Definition 1 An arena A is a triple hMA , λA , `A i where
• MA is a set of moves,
• λA : MA → {O, P } × {Q, A} is a function determining for each m ∈ MA
whether it is an Opponent or a Proponent move, and a question or an
answer.
• `A is a binary relation on MA , called enabling, satisfying
– if m `A n for no m then λA (n) = (O, Q),
OP
– if m `A n then λOP
A (m) 6= λA (n),
– if m `A n then λQA
A (m) = Q.
QA
We write λOP
A , λA for the composite of λA with respectively the first and second
projections. If m `A n we say that m enables n. We shall write IA for the set of
all moves of A which have no enabler; such moves are called initial. Note that
an initial move must be an Opponent question.
The arenas for the basic types are as follows:
12 SCI
is SCC with all concurrency bounds set to the unit value.
9
• JcomK = h{q, a}, {q 7→ (O, Q), a 7→ (P, A)}, {(q, a)}i.
This tells us that a command has two observable actions, corresponding
to a request for execution and an acknowledgement that the execution
completed successfully.
• JexpK = h{q, t, f }, {q 7→ (O, Q), t 7→ (P, A), f 7→ (P, A)}, {(q, t), (q, f )}i.
A boolean expression has three observable actions, a request for execution
and two possible outcomes.
• JcellK = h{q, t, f, wt, wf, a}, {q 7→ (O, Q), t 7→ (P, A), f 7→ (P, A), wt 7→
(O, Q), wf 7→ (O, Q), a 7→ (P, A)}, {(q, t), (q, f ), (wt, a), (wf, a)}i.
A memory cell has three actions for reading, the same as in the case of a
boolean expression, and three actions for writing: requests to write true
or write false and an acknowledgement.
The enabling relation reflects, intuitively, the basic causal connection between
moves: for a base type, an answer must be ’enabled’ by a question.
Composite types are interpreted by composite arenas constructed as follows:
• MA×B = MA + MB , [λA , λB ], `A + `B .
The product arena is simply the structure-preserving disjoint union of the
two arenas.
• MA→B = MA + MB , [λA , λB ], `A + `B + IB × IA , where λm = (O, x) if
and only if λm = (P, x).
The arrow arena is similar to the product arena but it involves the switching of polarities Proponent-Opponent for the arena in the contra-variant
position, and extending the enabling relation so that initial moves of the
argument are justified by the initial moves of the function.
The computational intuition for reversing polarities is that, in the case of
an argument the “inputs” become “outputs” and vice-versa. The extension of the enabling relation is motivated by establishing a causal connection between computations executed by the argument and those executed
by the main body of the function.
For example, the arena for type com → com is:
• M = {q1, q2, a1, a2}
• λ = {q1 7→ OQ, q2 7→ OQ, a1 7→ P A, a2 7→ P A}
• ` = {(q1, q2), (q1, a1), (q2, a2)}.
In the definition of the set of moves, Mcom→com = Mcom + Mcom we defined the
two injections associated to the disjoint sum as ini (x) = xi.
There is a natural and elegant connection between an arena and an interface:
10
Principle 2 There is an immediate connection between the following game-semantic and hardware concepts:
• An arena corresponds to an interface.
• A move corresponds to a port in the interface.
• A Proponent move is an output port.
• An Opponent move is an input port.
• A question is a request.
• An answer is an acknowledgment.
A term of type Γ ` M : T is interpreted in arena JΓK → JT K. This means that
this term will be compiled into a circuit with interface defined by its arena. For
example, for a term x : com ` M : com, the interface will be given by the arena
com → com discussed above:
q2
q1
a2
M
a1
What is not obvious in the conventional interface depiction above is the
enabling relation, which plays an essential role in the definition of the semantics
and of the FIM. In fact the concept of enabling does not have any obvious
hardware equivalent.
4
The sequentially reentrant access protocol
In the previous section we gave a method for defining an interface for any type
as a structured set of ports. In this section we give a protocol which governs
access to these interfaces, rules which will be formulated in an assume-guarantee
way: assuming the environment is well behaved in the way it provides input
actions, a circuit must guarantee that its outputs will behave in a certain way.
This protocol will be derived directly from the game-semantic model.
Let us start again with some basic game-semantic definitions.
A justified sequence in arena A is a finite sequence of moves of A so that the
first move is initial and each subsequent occurrence of a move n must have a
uniquely associated earlier occurrence of a move m such that m `A n. We say
that n is (explicitly) justified by m or, when n is an answer, that n answers m.
If a question does not have an answer in a justified sequence, we say that it
is pending in that sequence. In what follows we use the letters q and a to refer
to question- and answer-moves respectively, m will be used for arbitrary moves
and mA will be a move from MA .
A justified sequence is a legal move if it satisfies:
y
Fork : In any prefix s0 = · · · q · · · m of s, the question q must be pending
before m is played.
11
Wait : In any prefix s0 = · · ·
be answered.
q
z
···
a of s, all questions justified by q must
Serial : In any prefix s0 = · · · q · · · q the first occurrence of q cannot be pending
when the second one is played.
The simplest sequences of moves that violate Fork, Wait and Serial respectively
are:
su
m
Fork : q1
a1
Wait : q1
sw
Serial : q1
q2
sw
q2
a1
q2
For an arena A, the set of all plays satisfying these conditions is PA , its set of
(legal) plays.
The intuition behind these game rules is intimately connected to the language SCI. We can think of a question as the start of a process and an answer
as its termination. The Fork rule says essentially that only a “live” process
can spawn children; Join stipulates that a process can only terminate if all
its children processes have terminated. This is essentially the static nature of
concurrency in our language, where the only concurrent construct is par. The
final rule, Serial, restricts any process to only one active instance at any given
time, which is a consequence of our banning of sharing in concurrent contexts;
in the context of hardware it means that each module may receive a new initial
request only when the previous initial request has been acknowledged.
The consequence of the Serial rule is the following:
Theorem 2 ([12]) The representation of the game model for any term Γ `
M : T of SCI is a regular language.
This justifies our choice of SCI as an interesting target language, as it has both
higher-order functional features, imperative features, and a finite-state model
at any term which recommends it for efficient hardware compilation.
Principle 3 There is an immediate connection between the following game-semantic and hardware concepts:
• An occurrence of a move is a signal on a port.
• A justified sequence in an arena is a waveform on an interface.
• The set of plays on an interface is a protocol of access to the
interface.
In the above we abstract the definition of a “signal,” i.e. its precise phase encoding. Below, to keep the presentation focused we assume a standard low-high-low
encoding, but other encodings can be equally used. This also affects the exact
shape of the waveform.
12
Let us call the protocol corresponding to the SCI game model the Sequentially Reentrant Access Protocol (SRAP).
For instance, any communication mediated by interface Jcom → comK must
behave following the following pattern:
1. the first action must be an input request on q1;
2. the second action may be an output request on q2 which indicates that
the function is requesting an evaluation of its procedural argument
3. then, the third action must be an input acknowledgment on a2 indicating
successful completion of the procedural argument’s execution;
4. the protocol continues from Step 2;
5. the second action may also be an output acknowledgment on a1 indicating
that the main procedure has completed execution.
If Steps 2 and 3 are omitted then we have a function that is ignoring its argument
(i.e. non-strict). According to this protocol, all SRAP-compliant traces for this
interface are captured by the regular expression q1(q2 · a2)∗ a1.
§
The game-semantic model is denotational, i.e. inductively defined on the syntax
of the language. The model of a term is a strategy, i.e. a set of plays indicating
how the Proponent will respond to any Opponent action, given the play up
to that point. A strategy is subject to certain closure conditions, which are
not relevant for our discussion here. More relevant is a property of strategies
(thread-independence [11, Sec. 2.6.2]) that says that repeated plays of the same
strategy will be essentially the same, i.e. the strategy has no hidden state information preserved between its initial question and its (final) answer. We call
this property “Reset” noting that it is not a property of the protocol, but it is
a property that circuits that participate in the protocol must satisfy. Note that
it is known that circuits that are meant to be used compositionally must satisfy
this self-resetting property [29].
One of the important properties of strategies is that they compose in a way
that preserves both the legality of the resulting composite plays and the closure
conditions of the resulting composite strategy. The details of composition are
not relevant for our current discussion, but the following consequence is:
Theorem 3 (Compositionality) If A and B are two SRAP-compliant circuits with interfaces JT1 → T2 K and JT2 → T3 K respectively, then the circuit
B ◦ A of interface JT1 → T3 K formed by connecting the ports T2 in the two
circuits will be SRAP-compliant:
T1
T2
13
T3
B
...
...
...
...
A
T2
q1
q1,q2,a2,a1
q2
q1,q2,a2,a1
a1/q2,a2,a1
q1/q1,q2,a2
vs.
,q2
q1
a1
a2
2
,q
1
,q
1
/a
q2
a1
q1
1,
,a2
/a
a2
a
2,
q1,q2,a2,a1
Figure 1: Asynchronous vs. synchronous model for com → com
The compositionality of a protocol rarely receives the emphasis it deserves.
This property reduces substantially the burden of proof in creating complex
composite circuits. It is enough to show that each individual module is protocolcompliant, then using compositionality we can safely assume that any interconnection of the sub-circuits that is compatible with the signature of the interface
will result in a protocol-compliant circuit.
Principle 4 Interface-access protocols should be compositional in
order to simplify the design of complex circuits.
Note that the Reset property is also compositional in the same sense.
4.1
Synchronous versus asynchronous SRAP and related
optimisations
Conventional game-semantic models exhibit an asynchronous view of concurrency. This makes them naturally suitable for the design of asynchronous hardware. When using game models to target synchronous (clocked) hardware one
must take into account the possibility of several moves/signals occurring simultaneously. Note that this is level of synchronicity is simply used to to enhance
performance, processing several inputs at once and issuing outputs instantaneously is faster, but has no bearing on correctness. A synchronous version of a
protocol can be immediately derived using round abstraction [2]. For example,
the (asynchronous) set of plays for com → com and its round-abstracted (synchronous) version can be seen in Fig. 1 (simultaneous signals are separated by
commas, alternate labellings for a transition are separated by a slash).
The use of a asynchronous versus synchronous implementation has a serious
impact on performance. The round-abstraction algorithm guarantees that the
synchronous version of state machine has fewer (or equal) states, but may have
more transitions. It is difficult to assess apriori the impact of round abstraction
on the logical complexity of a circuit, but the synchronous version has an obvious
14
speed advantage. Whereas the asynchronous model alternates between receiving
inputs and producing outputs the synchronous model responds instantly (i.e.
in the same clock cycle) to an input, and can handle several inputs at once.
This makes implementation of constants and simple and ubiquitous functional
constants (such as sequential composition) far more efficient, because they no
longer introduce any delays at all, but respond instantly to input; in fact for
both constants (true, false) and sequential composition the implementation is
simply wires connected the right input and output ports.
The fact that our circuits work under a set protocol (SRAP) raises new
opportunities for optimisation through automaton minimisation. Minimisation
algorithms such as Hopcroft’s [14] work by identifying bisimulation-equivalent
states; however, when the protocol is set, it means that the automaton can only
receive inputs from a restricted language, that of the protocol, which raises new
possibilities for optimisation. Note that the more restrictive the protocol the
greater the possibility of optimisation. The trivially extreme case of an empty
protocol would allow any automaton to be reduced to the one-state automaton accepting all strings. Because we are working with a fully abstract gamesemantic model, it means that the underlying protocol is the most restricted
possible which still allows the compilation of any program. In other words, any
finite interaction which is SRAP-compliant can be realised by a program [12].
This point is illustrate in Figs. 2–4. In it we show three basic programming language constructs: the true constant, sequential composition and iteration. In the figure we show the resulting configurations on an FPGA (Xilinx
XC5VLX50) for the original asynchronous game model, the round-abstracted
synchronous version and its optimisation under the assumed SRAP-compliant
behaviour. These circuits are intended only to give a concrete feel for the effectiveness of the algorithms.
The reduction of true and seq to simply connectors is a remarkable optimisation which has a tremendous impact on the overall efficiency (time and space)
of the circuits generated. Also, we should emphasise that the synchronous versions of the circuits we obtain are well-behaved under instant feedback, a wellknown problem in conventional synchronous languages such as Esterel [4].
4.2
Activation managers
The development up to this point addresses gives a FIM consisting of a static definition (the interface) and a dynamic set of rules (the protocol) which solves the
first problem we identified earlier: designing interfaces that correspond to (possibly higher-order) functions. In this section we shall see how re-instantiation
can be avoided by using sharing instead. This also solves the problem of bookkeeping of names for modules and ports.
Suppose that we have a (slave) circuit P with interface JT K, and suppose it
is used twice by a (master) circuit Q. In order to avoid instantiating P twice
we will use a special circuit called an activation manager (AM):
15
Asynchronous:
Synchronous:
Minimised:
Figure 2: Circuit for the constant true
P
P
T
T
T1
Q
P
vs.
T
Am
T2
Q
If Q makes the initial request on a root port of JT1 K then AM will establish a
logical connection between P and the ports of T1 ; it can be seen that AM works
just like a local simple bus for P . If P is large and AM small this can vastly
improve the total footprint. We will see an example in the next section.
The question to ask at this point is: How can we implement an AM? Or
even a more basic question: What does it mean for an AM to be correct?
Note that the AM is not one circuit, but a family of circuits indexed by any
type in the language.
The correctness of the AM is precisely captured by the diagram above, which
can be written equationally as
Q ◦ hP, P i = Q ◦ Am ◦ P.
In order for the composition to type-check, we need the circuits involved to have
the following interfaces: P : X → T , Q : T × T → Y and Am : T → T × T . In
fact the mention of Q is irrelevant, what we need is simply
hP, P i = Am ◦ P.
These above are circuits, but now if we trace our intuitions back through the
game-semantic model to the programming language it is quite obvious that we
16
Asynchronous:
Synchronous:
Minimised:
Figure 3: Circuit for sequential composition
17
Asynchronous:
Synchronous:
Minimised:
Figure 4: Circuit for iteration
18
want Am to represent the denotation of the diagonal function λx.hx, xi that
given an argument returns a pairing of the argument with itself.
Note that this need for circuits to behave similarly under sharing versus
replication is a main reason why they need to have the self-resetting property.
Clearly, if P above has internal state that is changes in a way that survives a
round of usage—let us say that P has an internal counter for the number of
times it is activated—then it is not possible to expect the same behaviour in
the shared scenario as in the replicated scenario, as the shared P will be used
twice as many times as the replicated instances.
Fortunately, we can extract a representation for the diagonal, at any type
directly from the game-semantic model [12]: this is the Application Manager!
From Thm. 2 we know that this is a finite-state model which can be mapped
into hardware using standard methods (the synchronous version of the diagonal
must undergo round abstraction, as explained before).
In Fig. 5 we show the asynchronous model for the diagonal com → com × com,
the synchronous model and the circuits resulting from mapping the two diagonals on a Xilinx Virtex-5 (XC5VLX50) FPGA. The asynchronous AM requires
7 flip-flops (FF) and 22 look-up tables (LUT), while the synchronous version
requires 3 FFs and 6 LUTs.
Finally, note that the AM for com → com × com has both the same signature and the same behaviour as the CALL logical module in the micro-pipelines
framework [28]. It is reasonable to claim that application managers are a generalisation across all type signatures of the concept of a CALL module.
4.3
Protocol compliance through type systems
As seen from the previous section, the proper operation of an AM is premised on
it being used to connect SRAP-compliant, self-resetting components. The AMs
themselves satisfy these conditions and, using the compositionality property
(Thm. 3) the resulting circuits are also SRAP-compliant and can participate
in other larger circuits structured using AMs. This is perfectly consistent with
Harper’s earlier quote:
Principle 5 In hardware, type systems must be realised as compositional protocols.
In hardware compilation the compositionality of SRAP simplifies the compilation process in the following way. Each basic construct of the language is compiled into a simple, SRAP-compliant circuit. Composition of sub-programs with
disjoint sets of identifiers can be realised by wiring the corresponding circuits
in a way that is consistent with the typing of their interfaces (cf. Theorem 3).
Sharing of identifiers is then realised through AMs. The type system of the
language will automatically allow sharing of identifiers in pairing but disallow
it in function application. As already mentioned, according to the type system it is legal to form terms such as hf (x); f (x)i but illegal to form f (f (x)) or
f (x) || f (x).
19
?r2.0,!r.0,?d.0,!d2.0
?r1.0,!r.0,?d.0,!d1.0
?r1.0 ?r2.0
?r2.0,!r.0 ?d.0,!d2.0
!d1.0
!r.0
!r.0
?d.0,!d1.0
!d2.0
?d.0,!d2.0,?r2.0,!r.0
?r1.0,!r.0
?d.0,!d1.0,?r2.0,!r.0
?d.0
?d.0
?d.0,!d2.0,?r1.0,!r.0
?d.0,!d1.0,?r1.0,!r.0
Figure 5: Asynchronous and synchronous AMs for com → com × com
20
Skip
Skip
Seq
Q'_1
A'_1
A_1
CK
A'_2
Q'_0
A'_0 Am
A_0
Q'_2
Q_1
A_2
Q_0
Q_2
Figure 6: Sequential function call
Because the circuits are a direct representation of the semantics of the type
system and the language, a compiler can be constructed denotationally. Function application is simply circuit composition, contraction is implemented by
AMs of the right type, and lambda abstraction is, through the currying isomorphism, only a relabelling of ports. The implementations for the constants
is a hardware representation of their game model. Additionally, the fact that
all circuits are SRAP-compliant means that the FSM representation for their
game model can be optimized beyond conventional minimisation [14], taking
into account that certain input combinations, which violate the access protocol,
are not possible [10].
4.4
Working with application managers
First lets look at an example of SRAP-compliant use of the activation manager. The representation of the game-semantic model (synchronous and optimized) of skip and seq are simply connectors, so the (legal) compiled program f (skip); f (skip) has schematic in Fig. 6, where AM is the diagonal for
(com → com) → (com → com) × (com → com).
Its behaviour is as follows: it will call the interface for f twice, and each
time f asks for an argument it will provide skip, which is implemented as a
wire between its own Q and A ports. A typical (legal) trace on this AM is
Q02 Q00 Q0 Q2 A2 A0 A00 A02 Q01 Q00 Q0 Q1 A1 A0 A00 A01 .
If we attempt to synthesise the invalid program f (f (skip)) we obtain the
schematic in Fig. 7, with skip implemented as a connector and function application implemented as wiring.
21
Skip
Application
Q'_1
A'_1
A_1
A'_2
CK
Q'_0
A'_0 Am
A_0
Q_1
Q'_2
A_2
Q_2
Q_0
Figure 7: Nested function call
To be more precise, the circuit above is
λf.(λf 0 .(π1 f 0 )((π2 f 0 )(skip)))(δf )).
Where δ = λx.hx, xi is the diagonal, implemented as Am.
The AM, same as before, must handle now traces of the form
Q02 Q00 Q0 Q2 Q01 Q00 Q0 Q1 A1 A0 A00 A01 A2 A0
Note the nesting of request and acknowledgements Q0 and A0 , which breaks
serialization. Simulating the circuit in a tool such as Xilinx ISE we can see that
the activation manager simply deadlocks after this trace and it never produces
the A0 . The reason is that after A01 the AM has established a communication
channel with the first projection, so the input on A2 is unexpected and cannot
be handled. Keeping track of nested function calls requires more than static
resources and is substantially more expensive to implement.
§
The second typical way to violate SRAP is by trying to use concurrent access to interface components, as in the the program for par(f (skip))(f (skip)),
synthesised as in Fig. 8, where par is part of the implementation of parallel
composition. The problems here are obvious and no trace analysis is required.
AM is not equipped to arbitrate the simultaneous initial requests on Q01 , Q02 ,
but even if it did, they are forwarded at the same time to the same instance of
f via Q00 ! This is precisely the source of the bug in HandelC.
22
Par
Q'_1
A'_1
A_1
CK
A'_2
Q'_0
par
A_0
A'_0 Am
Q'_2
Q_1
A_2
Q_0
Q_2
Skip
Figure 8: Concurrent function call
5
Conclusion
The effectiveness of this approach can be easily argued, as all the weaknesses
enumerated before, arising form the lack of a FIM are resolved using the gamesbased approach. We are able to define interfaces that implement higher-order
functions and thanks to the well-defined SRA protocol we can construct (automatically) light-weight AMs, at least as compared to alternatives such as buses
or NOC, that can efficiently (in time and space) share (self-resetting) modules
with any SRAP-compliant signature.
An earlier version of this work was the Geometry of Synthesis (GOS) [9], a
direct circuit-based semantics for a similar language (lacking parallel composition). The difference between GOS and the simpler method we propose here is
that the former is inspired by game semantics, whereas the latter is simply game
semantics. Using the game semantic model has certain advantages, such as eliminating the need for re-proving soundness. Using the game-semantic model also
has some other, subtler advantages that take advantage of the full-abstraction
property of the model to realize further optimizations of the concrete representation of the game model [10].
Also, establishing a methodology for the semantic-directed compilation of
game models into hardware, preserving correctness, allows us to take advantage
of the rich game-semantic literature, which provides fully abstract models for a
large variety of programming languages.
A proof-of-concept compiler was implemented using this approach. One of
the test applications was a back-propagation neural-network with 12 4-bit neurons. The project was carried out by a student with no training in digital design.
Using a standard textbook algorithm of 167 lines of source code, the compiler
produced a design that fitted comfortably on a Spartan XC3S200 FPGA, using
10% of available FFs, 13% of available LUTs and 19% of available slices, clocked
23
at maximum frequency.
The argument of this paper is a methodological one: FIMs, access protocols,
application managers and language-support through type systems, all fit naturally in the game-semantic framework and can be immediately and naturally
translated into relevant hardware concepts. These concepts can overcome the
absence of FIMs in hardware compilation, which is a significant obstacle in the
way of producing mature and usable hardware compilers. Without well defined
FIMs it is not possible to support separate compilation units, which greatly
restricts library support and run-time inter-operability of components. Another
immediate consequence of having a well defined FIM is the ability to implement
activation managers, circuits that can share the functionality of other circuits.
We have presented a simple (but non-trivial) example for a FIM and an associated access protocol that guarantees the correct usage of circuits that implement
the interface. The SRA Protocol supports a rich type system and has built-in
support for safe concurrency, which is an important step forward compared to
the state-of-the-art in hardware compilation.
Principle 6 Lessons learned decades ago concerning the design and implementation of compilers must not be ignored, but must be adapted to hardware compilation.
References
[1] S. Abramsky, R. Jagadeesan, and P. Malacaria. Full abstraction for PCF. Inf.
Comput., 163(2):409–470, 2000.
[2] R. Alur and T. A. Henzinger. Reactive modules. Formal Methods in System
Design, 15(1):7–48, 1999.
[3] Arvind and R. S. Nikhil. Executing a program on the mit tagged-token dataflow
architecture. IEEE Trans. Computers, 39(3):300–318, 1990.
[4] G. Berry and G. Gonthier. The Esterel synchronous programming language:
Design, semantics, implementation. Sci. Comput. Program., 19(2):87–152, 1992.
[5] P. Bjesse, K. Claessen, M. Sheeran, and S. Singh. Lava: hardware design in
haskell. In ICFP, pages 174–184, 1998.
[6] M. Budiu and S. C. Goldstein. Compiling application-specific hardware. In FPL,
pages 853–863, 2002.
[7] B. Buyukkurt, Z. Guo, and W. A. Najjar. Impact of loop unrolling on area,
throughput and clock frequency in Roccc: C to VHDL compiler for FPGAs. In
ARC, pages 401–412, 2006.
[8] Celoxica. Handel-C Reference Manual. http://www.celoxica.com.
[9] D. R. Ghica. Geometry of Synthesis: a structured approach to VLSI design. In
POPL, pages 363–375, 2007.
[10] D. R. Ghica. Geometry of Synthesis 2: Compiling finite-bound concurrency into
hardware. Workshop on Games for Logic and Programming Languages IV, March
2009.
[11] D. R. Ghica and A. Murawski. Angelic semantics of fine-grained concurrency.
Annals of Pure and Applied Logic, 151(2-3):89–114, 2008.
[12] D. R. Ghica, A. S. Murawski, and C.-H. L. Ong. Syntactic control of concurrency.
Theor. Comput. Sci., 350(2-3):234–251, 2006.
24
[13] S. Guo and W. Luk. Compiling Ruby into FPGAs. In FPL, pages 188–197, 1995.
[14] J. E. Hopcroft. An nlogn algorithm for minimizing the states in a finite automaton. In Z. Kohave, editor, The Theory of Machines and Computations. Academic
Press, New York, 1971.
[15] J. M. E. Hyland and C.-H. L. Ong. On full abstraction for PCF: I, II, and III.
Inf. Comput., 163(2):285–408, 2000.
[16] S. C. Johnson and D. M. Ritchie. The C language calling sequence. Technical
Report Computer Science 102, Bell Laboratories, September 1981.
[17] S. Kumar, A. Jantsch, J.-P. Soininen, M. Forsell, M. Millberg, J. Oberg, K. Tiensyrja, and A. Hemani. A network on chip architecture and design methodology.
In Proceedings. IEEE Computer Society Annual Symposium on VLSI., pages 105–
112, 2002.
[18] W. Luk, D. Ferguson, and I. Page. Structured hardware compilation of parallel
programs. In W. Moore and W. Luk, editors, More FPGAs. Abingdon EE&CS
Books, 1994.
[19] R. Milner. Functions as processes. In ICALP 17, volume 443 of LNCS, pages
167–180, 1990.
[20] P. W. O’Hearn. On bunched typing. J. Funct. Program., 13(4):747–796, 2003.
[21] P. W. O’Hearn, J. Power, M. Takeyama, and R. D. Tennent. Syntactic control
of interference revisited. Theor. Comput. Sci., 228(1-2):211–252, 1999.
[22] P. W. O’Hearn and D. J. Pym. The logic of bunched implications. Bulletin of
Symbolic Logic, 5(2):215–244, 1999.
[23] I. Page and W. Luk. Compiling Occam into FPGAs. In W. Moore and W. Luk,
editors, FPGAs, pages 271–283. Abingdon EE&CS Books, 1991.
[24] J. C. Reynolds. Syntactic control of interference. In POPL, pages 39–46, 1978.
[25] J. C. Reynolds. Syntactic control of inference, part 2. In ICALP, pages 704–722,
1989.
[26] J. C. Reynolds. Separation logic: A logic for shared mutable data structures. In
LICS, pages 55–74, 2002.
[27] S. Sirowy. C is for circuits: Capturing FPGA circuits as sequential code for
portability. In FPGA, 2008.
[28] I. E. Sutherland. Micropipelines. Commun. ACM, 32(6):720–738, 1989. Turing
Award Paper.
[29] I. E. Sutherland and J. K. Lexau. Designing fast asynchronous circuits. In Seventh
International Symposium on Asynchronous Circuits and Systems, pages 184–194,
2001.
[30] The University of Manchester Advanced Processor Technologies. BALSA Reference Manual. http://www.cs.manchester.ac.uk/apt/projects/tools/balsa/.
[31] C. H. van Berkel and R. W. J. J. Saeijs. Compilation of communicating processes
into delay-insensitive circuits. In Proceedings of ICCD, 1988.
[32] K. van Berkel. Handshake circuits: An intermediary between communicating
processes and VLSI. PhD thesis, Technische Univ., Eindhoven (Netherlands),
1992.
[33] K. van Berkel, J. Kessels, M. Roncken, R. Saeijs, and F. Schalij. The VLSIprogramming language Tangram and its translation into handshake circuits. In
EURO-DAC, pages 384–389, 1991.
[34] M. V. Wilkes, D. J. Wheeler, and S. Gill. The Preparation of Programs for an
Electronic Digital Computer. Addison Wesley, 1951.
25
| 6cs.PL
|
A Practical and Worst-Case Efficient
Algorithm for Divisor Methods of
Apportionment
Raphael Reitzig∗
Sebastian Wild∗
arXiv:1504.06475v4 [cs.DS] 5 Dec 2017
December 6, 2017
Proportional apportionment is the problem of assigning seats to parties
according to their relative share of votes. Divisor methods are the de-facto
standard solution, used in many countries.
In recent literature, there are two algorithms that implement divisor methods: one by Cheng and Eppstein [CE14] has worst-case optimal running time
but is complex, while the other [Puk14] is relatively simple and fast in practice
but does not offer worst-case guarantees.
We demonstrate that the former algorithm is much slower than the other
in practice and propose a novel algorithm that avoids the shortcomings of
both. We investigate the running-time behavior of the three contenders in
order to determine which is most useful in practice.
1. Introduction
The problem of proportional apportionment arises whenever we have a finite supply
of k indivisible, identical resource units which we have to distribute across n parties
fairly, that is according to the proportional share of publicly known and agreed-upon
P
values v1 , . . . , vn (of the sum V = vi of these values). We elaborate in this section on
applications of and solutions for this problem.
∗
Department of Computer Science, University of Kaiserslautern; {reitzig, wild} @ cs.uni-kl.de
1
1. Introduction
Apportionment arises naturally in politics. Here are two prominent examples:
• In a proportional-representation electoral system we have to assign seats in parliament to political parties according to their share of all votes.
The resources are seats, and the values are vote counts.
• In federal states the number of representatives from each component state often
reflects the population of that state, even though there will typically be at least
one representative for any state no matter how small it is.
Resources are again seats, values are the numbers of residents.
In order to use consistent language throughout this article, we will stick to the first
metaphor. That is, we assign k seats to parties [1..n] proportionally to their respective
votes vi , and we call k the house size.
A fair allocation should assign vi /V seats to party i, where V = v1 + · · · + vn is the
total vote count of all parties. In case of electoral systems which exclude parties below a
certain threshold of overall votes from seat allocation altogether, we assume they have
already been removed from our list of n parties.
As seats are indivisible, this is only possible if, by chance, all vi /V are integers; otherwise
we have to come up with some rounding scheme. This is where apportionment methods
come into play. The books by Balinski and Young [BY01] and Pukelsheim [Puk14] give
comprehensive introductions into the topic with its historical, political and mathematical
dimensions.
Mathematically speaking, an apportionment method is a function f : Rn>0 × N → Nn0 that
maps vote counts v = (v1 , . . . , vn ) and house size k to a seat allocation s = (s1 , . . . , sn ) :=
f (v, k) so that s1 + · · · + sn = k. We interpret s as party i getting si seats.
There are many conceivable such methods, but there are at least three natural properties
one would like apportionment systems to have:
(P1) Pairwise vote monotonicity: When votes change, f should not take away seats
from a party that has gained votes while at the same time awarding seats to one
that has lost votes.
(P2) House monotonicity: f should not take seats away from any party when the house
grows (in number of seats) but votes do not change.
(P3) Quota rule: The number of seats of each party should be its proportional share,
rounded either up or down.
Balinski and Young have shown that
• (P1) implies (P2) [BY01, Cor. 4.3.1],
• no method can always guarantee (P1) and (P3) [BY01, Thm. 6.1], and
2
1. Introduction
Method
Divisor Sequence
δ(x)
Sandwich
Smallest divisors
0, 1, 2, 3, . . .
x
—
Greatest divisors
1, 2, 3, 4, . . .
x+1
—
Sainte-Laguë
1, 3, 5, 7, . . .
2x + 1
—
Modified Sainte-Laguë
Equal Proportions
1.4, 3, 5, 7, . . .
√ √ √
0, 2, 6, 12, . . .
Harmonic Mean
0, 43 ,
12 24
5 , 7 ,
...
Imperiali
2, 3, 4, 5, . . .
Danish
1, 4, 7, 10, . . .
2x+1
x(x + 1)
x+
2x(x+1)
2x+1
6
5
1
4
x+
1
4
x≥1
1.6x+1.4 x<1
p
2x +
x+2
—
3x + 1
—
±
±
±
1
5
1
4
1
4
Table 1: Commonly used divisor methods [CE14, Table 1]. For each of the methods, we
give a possible continuation δ of the respective divisor sequence (cf. Section 2)
as well as linear sandwich bounds on δ, if non-trivial (cf. Lemma 2).
• (P1) holds exactly for divisor methods [BY01, Thm. 4.3].
Property (P1) is essential for upholding the principle of “one-person, one-vote”, an ideal
pursued by electoral systems around the globe and occasionally enforced by law [Puk14,
Section 2.4]. Therefore, divisor a. k. a. Huntington methods can be the only choice, for
the price of (P3). Other choices can be made, of course; the aforementioned books [BY01;
Puk14] discuss different trade-offs.
Divisor methods are characterized by divisor sequences which control the notion of
“fairness” implemented by the respective method. There are many popular choices (cf.
Table 1). It is not per se clear which divisor sequence is the best; there still seems to be
active discussion, e. g., for the U. S. House of Representatives. One reason is that no-one
has yet been able to propose a convincing, universally agreed-upon mathematical criterion
that would single out one method as superior to the others. In fact, there are competing
notions of fairness, each favoring a different divisor method [BY01, Section A.3]. A
reasonable approach is therefore to run computer simulations of different methods and
compare their outcomes empirically, for example w. r. t. the distribution of final average
votes per seat vi /si . For this purpose, many apportionments have to be computed, so
efficient algorithms can become an issue.
We thus study the problem of computing the final seat allocation by divisor methods
(given by their divisor sequences) according to vote counts and house size.
For the case of almost linear divisor sequences, the problem can be solved in time O(n);
this has been shown by Cheng and Eppstein [CE14] who propose a worst-case runningtime-optimal algorithm which we call ChengEppsteinSelect. It is quite involved and
rather difficult to implement (cf. Appendix C.3).
3
2. Divisor Methods Formalized
Pukelsheim [Puk14], on the other hand, proposes algorithm JumpAndStep whose
running time is not asymptotically optimal in the worst case but tends to perform well in
practice, at least if some insight about the used divisor sequence is available and inputs
are good-natured (cf. Appendix C.2).
After introducing divisor methods formally in Section 2, we propose a new algorithm in
Section 3 that also attains the O(n) worst-case running time bound but is straight-forward
to implement and efficient in practice as well. It is based on a generalization of our
solution for the envy-free stick-division problem [RW15b].
We finally compare the performance of the three contending algorithms with extensive
running time experiments, an executive summary of which we give in Section 4.
Additional material includes an index of notation in Appendix F.
2. Divisor Methods Formalized
Let d = (dj )∞
j=0 be an arbitrary divisor sequence, i. e. a nonnegative, strictly increasing
and unbounded sequence of real numbers. We formally set d−1 := −∞.
We require that there is a smooth continuation of d on the reals which is easy to invert.
That is, we assume a function δ : R≥0 → R≥d0 with
i) δ is continuous and strictly increasing,
ii) δ −1 (x) for x ≥ d0 can be computed with a constant number of arithmetic operations,
and
iii) δ(j) = dj (and thus δ −1 (dj ) = j) for all j ∈ N0 .
All the divisor sequences used in practice fulfill these requirements; cf. Table 1. For
convenience, we continue δ −1 on the complete real line requiring
iv) δ −1 (x) ∈ [−1, 0) for x < d0 .
Corollary 1: Assuming i) to iv), δ −1 (x) is continuous and strictly increasing on R≥d0 .
Furthermore, it is the inverse of j 7→ dj in the sense that
bδ −1 (x)c = max{j ∈ Z≥−1 | dj ≤ x}
for all x ∈ R.
In particular, bδ −1 (x)c = j for dj ≤ x < dj+1 so the floored δ −1 is the (zero-based) rank
function for the set of all dj as long as x ≥ d0 .
Note how this reproduces what is called d-rounding in the literature [BY01; Puk14]; we
obtain an efficient way of calculating this function via δ −1 .
4
2. Divisor Methods Formalized
Now the set of all seat assignments that are valid w. r. t. d is given by [BY01]
S(v, k) =
n
s ∈ Nn0
n
X
i=1
si = k ∧ ∃ a > 0. ∀ i ∈ [1..n]. si ∈ bδ −1 (vi · a)c + {0, 1} .
o
We call a realization of a proportionality constant a∗ ; intuitively, every seat corresponds
to roughly 1/a∗ votes.
An equivalent definition is by the set of possible results of the following algorithm [BY01,
Prop. 3.3].
Algorithm 1: IterativeMethodd (v, k) :
Step 1 Initialize s = 0n .
Step 2 While k > 0,
Step 2.1 Determine I = arg minni=1 dsi /vi .
Step 2.2 Update sI ← sI + 1 and k ← k − 1.
Step 3 Return s.
We can obtain a proportionality constant [Puk14, 59f] by
a∗ = max{dsi −1 /vi | 1 ≤ i ≤ n},
(1)
which in turn defines the set S(v, k).
Note that we work with dj /vi instead of vi /dj in the classical literature; Cheng and
Eppstein [CE14] and we prefer the reciprocals because the case d0 = 0 then handles
gracefully and without special treatment. Therefore, our a∗ is also the reciprocal of the
proportionality constant as e. g. Pukelsheim [Puk14] defines it, we multiply by a in the
definition of S and we take the minimum in IterativeMethod. It is important to note
that the defined set S remains unchanged by this switch.
Following the notation of Cheng and Eppstein [CE14], we furthermore define for given
votes v = (v1 , . . . , vn ) ∈ Qn>0 the sets
Ai :=
ai,j j = 0, 1, 2, . . .
n
o
with ai,j :=
dj
vi
and their multiset union
A :=
n
]
Ai .
i=1
As we will see later, the relative rank of elements in A turns out to be of interest; we
therefore define the rank function r(x, A) which denotes the number of elements in
5
3. Fast Apportionment by Rank Selection
multiset A that are no larger than x, that is
r(x, A) :=
A ∩ (−∞, x]
=
n
X
i=1
(2)
|{ai,j ∈ A | ai,j ≤ x}|.
We write r(x) instead of r(x, A) when A is clear from context.
We need two more convenient shorthands: Assuming we have a∗ ≤ x, we denote with
Ix :=
(3)
i ∈ {1, . . . , n} | vi > d0 /x
the set of parties that can hope for a seat, and with
x
A
:=
dj
∈A
<x
vi
vi
] dj
i∈Ix
=
n
]
dj
i=1
dj
∈A
<x
vi
vi
= A ∩ (−∞, x) (4)
the multiset of elements from sequences of these parties that are smaller than x, i. e.
reasonable candidates for a∗ .
3. Fast Apportionment by Rank Selection
From (1) together with strict monotonicity of d, we obtain immediately that a∗ = A(k) ,
i. e. the kth smallest element of A (counting duplicates) is a suitable proportionality
constant. This allows us to switch gears from the iteration-based world of Pukelsheim
[Puk14] to selection-based algorithms, as previously seen by Cheng and Eppstein [CE14].
Note that even though A is infinite, A(k) always exists because the terms ai,j = dj/vi are
strictly increasing in j for all i ∈ {1, . . . , n}.
Borrowing terminology from the field of mathematical optimization, we call a feasible if
r(a) ≥ k, otherwise it is infeasible. Feasible a 6= a∗ are called suboptimal. Our goal is to
find a subset of A that contains a∗ but as few infeasible or suboptimal a as possible; we
can then apply a rank-selection algorithm on this subset and obtain (via a∗ ) the solution
to the apportionment problem.
Now since d is unbounded, setting any upper bound x on the ai,j yields a finite search
space Ax . By choosing any such bound that maintains |Ax | ≥ k, we retain the property
that a∗ is the kth smallest element under consideration.
One naive way is to make sure that the party with the most votes (which should get
the most seats) contributes at least k values to A. This can be achieved by letting
x = dk−1 / max v + ε (cf. the proof of Theorem 3). This alone, however, leads only
to an algorithm with worst-case running time in Θ(kn), which is worse than even
IterativeMethod (with priority queues).
We can actually not improve this upper bound x; it is tight for the case that one party
has many more votes than all others and gets (almost) all of the seats. We can, however,
6
3. Fast Apportionment by Rank Selection
exclude many individual elements in Ax because they are too small to be feasible or too
large to be optimal.
Towards finding suitable upper and lower bounds on a∗ , we investigate its rank in the
multiset A of all candidates. All we know is that
k ≤ r(a∗ ) ≤ k + |Ix |
since we may have any number between one and |Ix | parties that tie for the last seat.
We can still make an ansatz with r(a) ≥ k + |Ix | and r(a) < k, express rank function r
in terms of δ −1 (cf. Lemma 4 in Appendix B) and derive that
X
i∈Ix
and
δ −1 (vi · a) ≤ k − |Ix |
X
i∈Ix
δ −1 (vi · a) ≥ k.
(5)
This pair of inequalities is indeed a sufficient condition for admissible pairs of bounds
(a, a); we can conclude that a ≤ a∗ ≤ a. For a formal proof, see Lemma 5 in Appendix B.
We now want to derive a sandwich on a∗ by fulfilling the inequalities in (5) as tightly as
possible. Depending on δ −1 , this may be hard to do analytically. However, we can make
the same assumption as Cheng and Eppstein [CE14] and explicitly compute suitable
bounds for divisor sequences which behave roughly linearly. This does not limit the scope
of our investigation by much; see Appendix A for more on this.
Lemma 2: Assume the continuation δ of divisor sequence d fulfills
αx + β ≤ δ(x) ≤ αx + β
for all x ∈ R≥0 with α > 0, β ∈ [0, α] and β ≥ 0. Let further some x > a∗ be given.
Then, the pair (a, a) defined by
a := max 0,
with Vx :=
P
i∈Ix
αk − (α − β) · |Ix |
Vx
and a :=
αk + β · |Ix |
Vx
vi fulfills the conditions of Lemma 5, that is a ≤ a∗ ≤ a. Moreover,
A ∩ [a, a] ≤ 2 1 +
β−β
α
· |Ix |.
The proof consists mostly of rote calculation towards applying Lemma 5; see Appendix B
for the details.
We have now derived our main improvement over the work by Cheng and Eppstein
[CE14]; where they have only a one-sided bound on a∗ and thus have to employ an
involved search on A, we have sandwiched a∗ from both sides, and so tightly that the
remaining search space is small enough for a simple rank selection to be efficient.
7
3. Fast Apportionment by Rank Selection
Building on the bounds from Lemma 2, we can improve upon the naive idea using only
x by excluding also many more elements from A which are for sure not a∗ . Since we
remove in particular too small elements, this means that we also have to modify the rank
we select; we will see that our bounds are chosen so that we can use δ −1 to count the
number of elements we discard exactly.
Recall that we assume a fixed apportionment scheme, that is fixed d with known α, β
and β as per Lemma 2.
Algorithm 2: SandwichSelect(v, k)d :
Step 1 Find the v (1) = max{v1 , . . . , vn }.
Step 2 Set x := dk−1 /v (1) + ε for suitable1 constant ε > 0.
Step 3 Compute Ix as per (3).
Step 4 Compute a and a as per Lemma 2.
Step 5 Initialize  := ∅ and k̂ := k.
Step 6 For all i ∈ Ix , do:
Step 6.1 Compute j := max 0, δ −1 (vi · a)
and j := bδ −1 (vi · a)c.
Step 6.2 Add all dj /vi to  for which j ≤ j ≤ j.
Step 6.3 Update k̂ ← k̂ − j.
Step 7 Select and return Â(k̂) .
Theorem 3:
Algorithm 2 computes a∗ in time O(n) for any divisor sequence d that fulfills the
requirements of Lemma 2.
Proof: First, we have to show that Ix as we compute it in Steps 1-3 is correct. We
have x > a∗ = Ax as already r(x − ε) = r(dk−1 /v (1) ) ≥ k; at least the k elements
d
d0
(1) . We thus never need to consider elements
, . . . , vk−1
(1) ∈ A are no larger than dk−1 /v
v (1)
a ≥ x, and in particular A(k) = Ax(k) as Ax = A ∩ (−∞, x).
So far, we have needed no additional restriction on ε in Step 2; we only need it to be
positive so we do not discard a∗ by accident if it is exactly dk−1 /v (1) . However, the size
of Ax can be arbitrarily large – depending on the input values vi which we do not want.
Therefore, we require
0<ε<
1
dk − dk−1
;
v (1)
(6)
Neither correctness nor Θ-running-time is affected by the choice of ε here since it affects only the size
of Ix , which is bounded by n in any case. In particular, the size of  is affected only up to a constant
factor. For tweaking performance in practice, see the proof of Theorem 3.
8
4. Comparison of Algorithms
such exists because d is strictly increasing. Note how then x < dk /v (1) so we do not keep
any additional suboptimal values.
From Step 4 on, we then construct multiset  ⊆ A as the subsequent union of Ai ∩ [a, a],
that is
 =
] dj
i∈Ix
vi
] dj
=
i∈Ix
vi
] dj
=
i∈Ix
]
=
i∈Ix
vi
j(i) ≤ j ≤ j(i)
∈A δ
−1
(vi · a) ≤ j ≤ δ
∈ A vi · a ≤ dj ≤ vi · a
dj
dj
∈A a≤
≤a
vi
vi
−1
(vi · a)
= A ∩ [a, a].
In particular, the last step follows from (4) with x > a∗ . By Lemma 2, we know that
a ≤ a∗ ≤ a for the bounds computed in Step 4, so we get in particular that a∗ ∈ Â.
It remains to show that we calculate k̂ correctly. Clearly, we discard with (ai,0 , . . . , ai,j−1 )
exactly j elements in Step 6.2, that is |Ai ∩ (−∞, a)| = j(i). Therefore, we compute with
k̂ = k −
X
i∈Ix
Ai ∩ (−∞, a) = r(a∗ , A) − A ∩ (−∞, a) = r(a∗ , Â)
the correct rank of a∗ in Â.
For the running time, we observe that the computations in steps 1 to 5 are easily
done with O(n) primitive instructions. The loop in Step 6 and therewith steps 6.1
and 6.3 are executed |Ix | ≤ n times. The overall number of set operations in Step 6.2 is
|Â| ∈ O(|Ix |) ⊆ O(n) (cf. Lemma 2). Finally, Step 7 runs in time O(|Â|) ⊆ O(n) when
using a (worst-case) linear-time rank selection algorithm (e. g., the median-of-medians
algorithm [Blu+73]).
We have obtained a relatively simple algorithm that implements many divisor methods
and has optimal asymptotic running time in the worst case. It remains to be seen if it is
also efficient in practice.
4. Comparison of Algorithms
We have implemented all algorithms mentioned above in Java [RW15a] with a focus on clarity and performance. Reviewing the algorithms resp. implementations (cf. Appendix C),
9
4. Comparison of Algorithms
Average running time in µs/n
(α, β) = (2, 1) and k = 100n
(α, β) = (1, 3/4) and k = 10n
4
0.6
3
0.4
2
0.2
1
0
2
4
6
n
8
0
10
0
50
100
n
150
200
Figure 1: This figure shows average running times of SandwichSelect , ChengEppsteinSelect , JumpAndStep with naive resp. priority-queue minimum
selection, and IterativeMethod with naive resp. priority-queue minimum
selection, normalized by the number of parties n. The inputs are random apportionment instances with vote counts vi drawn i. i. d. uniformly from [1, 3]. The
numbers of parties n, house size k and method parameters (α, β) have been
chosen to resemble national parliaments in Europe (left) and the U. S. House of
Representatives (right), respectively.
we observe that neither IterativeMethod nor JumpAndStep are asymptotically
worst-case efficient whereas ChengEppsteinSelect does not seem to be practical
regarding implementability. SandwichSelect does not have either deficiency and is
still the shortest of the non-trivial algorithms.
We evaluate relative practical efficiency by performing running time experiments on
artificial instances; we fix the number of parties n, house size k and the used divisor
method and draw multiple vote vectors v at random according to different distributions.
Where possible, we draw votes from a continuous distribution with fixed expectation;
this ensures that vote proportions do not devolve to trivial situations as n grows.
In order to keep the parameter space manageable, we use n as free variable and fix k to
a multiple of n. For ease of implementation, we restrict ourselves to divisor sequences of
the form (αj + β)j∈N0 ; this still allows us to cover a range of relevant divisor methods at
least approximately (cf. Table 1). We describe the machine configuration used for the
experiments and further details of the setup in Appendix D.
Figure 1 shows the results of two experiments with practical parameter choices. It is clear
that JumpAndStep dominates the field; of the other algorithms, only SandwichSelect
comes close in performance. These observations are stable across many parameter
choices; see also Appendix E. We will therefore restrict ourselves to JumpAndStep and
SandwichSelect in the sequel.
Towards understanding what influences the performance of these algorithms the most,
10
4. Comparison of Algorithms
Running time in ms
8
10
6
4
5
2
0
0
0
1,000
2,000
0
3,000
0.5
1
|Â|
|∆a |
1.5
·105
Figure 2: Running times on individual inputs plotted against |∆a | for JumpAndStep (left)
resp. |Â| for SandwichSelect (right). Inputs are random with exponentially
distributed vi for n ∈ {1 , 5 , 10 , 20 , 30 , 40 , 50 , 75 , 100 }·103 and k = 5n;
they have been apportioned w. r. t. (α, β) = (2, 1).
we have investigated how ∆a (the number of seats JumpAndStep assigns too much, i. e.
P
k − si ) resp. |Â| (the number of candidates SandwichSelect selects from) relate to
the measured running times. While the connection is clear for SandwichSelect, we
need to look at cases where Pukelsheim’s estimators are bad; as long as |∆a | n, the
Θ(n) portions of JumpAndStep dominate. Figure 2 exhibits such a setting.
While JumpAndStep is faster than SandwichSelect in the experiments of Figure 1
and similar ones, we observe that SandwichSelect is more robust against changing
parameters. Figure 3 exhibits this for switching between different vote distributions:
the average running times of SandwichSelect are close to each other where those of
JumpAndStep spread out quite a bit. It may be noteworthy that each algorithm has
one “outlier” distribution but they are not the same.
JumpAndStep does indeed seem to outperform SandwichSelect consistently so far,
if not by much in some cases. We have found a parameterization which, even though
it is admittedly rather artificial, clearly suggests that JumpAndStep does indeed have
ω(n) worst-case behavior and that SandwichSelect can be faster; see Figure 4. The
question after realistic settings for which this is the case remains open.
In summary, we have seen that SandwichSelect provides good performance in a reliable
way, i. e., its efficiency does not depend much on divisor sequence or input. On the other
hand, JumpAndStep is faster on average when good estimators are available, but can
be slower in certain settings.
11
Average running time in µs/n
4. Comparison of Algorithms
0.1
0.1
8 · 10−2
8 · 10−2
6 · 10−2
6 · 10−2
50
100
n
150
4 · 10−2
200
0
50
100
n
150
200
0.12
0.26
0.1
0.24
8 · 10−2
|∆a |/n
Average running time in µs/n
Figure 3: Normalized average runtimes of SandwichSelect (left) and JumpAndStep
(right) on vi drawn randomly from uniform , exponential , Poisson and Pareto
distributions, respectively, and with k = 5n and (α, β) = (2, 1).
6 · 10−2
0.22
4 · 10−2
2 · 10−2
0.2
0
0
0.2
0.4
0.6
n
0.8
1
0
0.2
0.4
0.6
n
·105
0.8
1
·105
Figure 4: The left plot shows normalized running times of SandwichSelect and
JumpAndStep on instances with k = 2n and Pareto-distributed vi for
(α, β) = (1.0, 0.001). The right plot shows that the average of |∆a | seems to
converge towards a constant fraction of n in this case.
12
5. Conclusion
5. Conclusion
We have derived an algorithm implementing divisor methods of apportionment that is
worst-case efficient, simple and practicable. As such, it does not have the shortcomings of
previously known algorithms. Even though it can not usually outperform JumpAndStep,
its robustness against changing parameters makes it a viable candidate for use in practice.
Acknowledgments
We thank Chao Xu for pointing us towards the work by Cheng and Eppstein [CE14] and
noting that the problem of envy-free stick-division [RW15b] is related to proportional
apportionment as discussed there. He also observed that our approach for cutting sticks –
the core ideas of which turned out to carry over to this article – could be improved to
run in linear time.
Furthermore, we owe thanks to an anonymous reviewer whose constructive feedback
sparked broad changes which have greatly improved the article over its first incarnation.
References
[Blu+73]
Manuel Blum et al. “Time Bounds for Selection.” In: Journal of Computer
and System Sciences 7.4 (Aug. 1973), pp. 448–461. doi: 10.1016/S00220000(73)80033-9.
[BY01]
Michel L. Balinski and H. Peyton Young. Fair Representation. Meeting the
Ideal of One Man, One Vote. 2nd. Brookings Institution Press, 2001. isbn:
978-0-8157-0111-8.
[CE14]
Zhanpeng Cheng and David Eppstein. “Linear-time Algorithms for Proportional Apportionment.” In: International Symposium on Algorithms and
Computation (ISAAC) 2014. Springer, 2014. doi: 10 . 1007 / 978 - 3 - 319 13075-0_46.
[GKP94]
Ronald L. Graham, Donald E. Knuth, and Oren Patashnik. Concrete mathematics: a foundation for computer science. Addison-Wesley, 1994. isbn:
978-0-20-155802-9.
[Puk14]
Friedrich Pukelsheim. Proportional Representation. Apportionment Methods
and Their Applications. 1st ed. Springer, 2014. isbn: 978-3-319-03855-1. doi:
10.1007/978-3-319-03856-8.
[RW15a]
Raphael Reitzig and Sebastian Wild. Companion Source Code. revision
db43ee7f05. 2015. url: https://github.com/reitzig/2015_apportionment.
[RW15b]
Raphael Reitzig and Sebastian Wild. Efficient Algorithms for Envy-Free Stick
Division With Fewest Cuts. 2015. arXiv: 1502.04048.
13
References
[SW11]
Robert Sedgewick and Kevin Wayne. Algorithms. 4th. Addison-Wesley, 2011.
isbn: 978-0-321-57351-3. url: http://algs4.cs.princeton.edu.
14
A. Our Scope of different Methods of Apportionment
A. Our Scope of different Methods of Apportionment
As we have seen in Section 2 there are many possible divisor sequences. For our main
result (cf. page 7) we follow Cheng and Eppstein [CE14] and require the sequences to
be “almost” linear; we should check that we do not unduly restrict the scope of our
investigation.
We refer to the recent reference work by Pukelsheim [Puk14] and, by extension, to
Balinski and Young [BY01] who classify different divisor methods of apportionment in
terms of signpost sequences, a concept equivalent to the divisor sequences we use. They
distinguish these classes of such sequences (cf. [Puk14, Sections 3.11-12]):
• stationary sign-posts of the form s(n) = n − 1 + r with r ∈ (0, 1);
• power-mean sign-posts defined by
s̃p (0) = 0,
s̃p (n) =
(n − 1)p + np
2
1/p
,
for p 6= −∞, 0, ∞;
• and special cases s̃−∞ (n) = n − 1, s̃0 (n) =
(n − 1)n, and s̃∞ (n) = n.
p
It is easy to see that stationary sign-posts correspond do divisor sequences dj = j + β
with β ∈ (0, 1) (up to a shift by one); as such, Lemma 2 applies immediateley with α = 1
and β = β = β, and yields a particularly nice (and tight, for our choices of a and a)
upper bound on the size of the candidate set A. We cover the special cases as well; see
Table 1 for the corresponding sandwich bounds.
As for the remaining power-mean sign-posts, the trivial bounds β = 0 and β = 1 already
work. One can apply the power-mean inequality and use the slightly better bounds
for p ∈ {−∞, −1, 0, 1, ∞} as given in Table 1. Even better bounds can be gleaned
from observing that s̃p (n) converges to n − 1/2 from one side, and quickly so; s̃p (1) thus
determines either β or β and the other can be chosen as 1/2.
In summary, our algorithm SandwichSelect applies to all divisor methods treated by
Pukelsheim [Puk14] and Balinski and Young [BY01]
B. Lemmata and Proofs
Lemma 4: For rank function r(x, A),
r(x, A) =
n
X
i=1
bδ −1 (vi · x)c + 1.
15
B. Lemmata and Proofs
Moreover, for x < x we have
r(x, A) =
X
bδ −1 (vi · x)c + 1
i∈Ix
with Ix = i ∈ {1, . . . , n} | vi > d0 /x .
B.1. Proof of Lemma 4
By eq. (2) on page 6, it suffices to show that
{ai,j | ai,j ≤ x} = bδ −1 (vi · x)c + 1
for each i ∈ {1, . . . , n}. Now, if x ≥ ai,j = dj/vi for some j, then vi · x ≥ dj , and so
bδ −1 (vi · x)c is the largest index j 0 for which ai,j 0 = dj 0 /vi ≤ x. As dj is zero-based, there
are j 0 + 1 ≥ 1 such elements ai,j ≤ x and the equation follows.
Otherwise, that is ai,j > x for all j, we have j 0 = bδ −1 (vi · x)c = −1 by iv) and Corollary 1
and the equality holds with 0 on both sides.
For the second equality, we only have to show that the omitted summands are zero. So
let i ∈
/ Ix be given, that is vi ≤ d0 /x. For x < x, we have
vi · x ≤
d0
d0
·x <
· x = d0 ,
x
x
and hence bδ −1 (vi · x)c = −1 by iv).
Lemma 5: Let x > a∗ and assume a and a are chosen so that they fulfill
X
i∈Ix
δ −1 (vi · a) ≤ k − |Ix |
X
and
i∈Ix
δ −1 (vi · a) ≥ k.
Then, a ≤ a∗ ≤ a.
The lemma follows more or less directly; one uses the sandwich bounds on r to show
that a < a are infeasible, i. e., r(a) < k, and that a is feasible, and thus all a > a are
suboptimal since a∗ is the smallest feasible element in A.
B.2. Proof of Lemma 5
As a direct consequence of Lemma 4 together with the fundamental bounds y−1 < byc ≤ y
on floors, we find that
X
i∈Ix
δ −1 (vi · x) < r(x, A) ≤
X
i∈Ix
δ −1 (vi · x) + 1
16
= |Ix | +
X
i∈Ix
δ −1 (vi · x) (7)
B. Lemmata and Proofs
for any x and all x < x. We can therewith pin down the value of r to an interval of
width |Ix | using only δ −1 . We can use this to derive upper and lower bounds on a∗ .
We show that smaller a are infeasible and larger a are clearly suboptimal, so the optimal
a∗ must lie in between. Let us first consider a < a. There are two cases: if there is a vi ,
such that vi a ≥ d0 , we get by strict monotonicity of δ −1
r(a) ≤ |Ix | +
(7)
< |Ix | +
X
i∈Ix
X
i∈Ix
δ −1 (vi · a)
δ −1 (vi · a)
≤ k
and a is infeasible. If otherwise vi a < d0 , i. e., a < d0 /vi , for all i, a must clearly have
rank r(a) = 0 as it is smaller than any element ai,j ∈ A. In both cases we found that
a < a has rank r(a) < k.
Now consider the upper bound, i. e., we have a > a. In case a ≥ x, we have a > x > a∗
by assumption and any such a cannot be optimal. Otherwise, for a < x, we have
r(a) >
(7)
X
i∈Ix
δ −1 (vi · a) ≥ k,
so a is feasible. Any element a > a can thus not be the optimal solution a∗ , which is the
minimal a with r(a) ≥ k.
B.3. Proof of Lemma 2
We consider the linear divisor sequence continuations
δ(j) = αj + β
and
δ(j) = αj + β
for all j ∈ R≥0 and start by noting that the inverses are
δ −1 (x) = x/α − β/α
and
δ −1 (x) = x/α − β/α
for x ≥ δ(0) = β and x ≥ δ(0) = β, respectively. For smaller x, we are free to choose the
value of the continuation from [−1, 0) (cf. iv)); noting that x/α − β/α < 0 for x < β, a
choice that will turn out convenient is
δ
−1
x β
(x) := max
− , −1
α α
resp.
δ
−1
x β
(x) := max
− , −1 .
α α
(8)
We state the following simple property for reference; it follows from δ(j) ≤ δ(j) ≤ δ(j)
and the definition of the inverses (recall that β ≤ α):
x β
x β
−
≤ δ −1 (x) ≤ δ −1 (x) ≤ δ −1 (x) ≤
− ,
α α
α α
17
for x ≥ 0.
(9)
B. Lemmata and Proofs
Equipped with these preliminaries, we compute
a =
⇐⇒
αk + β|Ix |
.
Vx
a X
β
·
vi = k +
· |Ix |,
α i∈I
α
x
⇐⇒ k =
X vi · a
α
i∈Ix
β
−
α
X
≤
(9) i∈I
x
δ −1 (vi · a),
so a satisfies the condition of Lemma 5. Similarly, we find
a =
αk − (α − β) · |Ix |
,
Vx
a
· Vx = k − (1 − β/α) · |Ix |,
α
X vi · a
X
β
⇐⇒ k = |Ix | +
≥ |Ix | +
−
δ −1 (vi · a),
α
α
(9)
i∈I
i∈I
⇐⇒
x
x
that is a also fulfills the conditions of Lemma 5.
For the bound on the number of elements falling between a and a, we compute
A ∩ [a, a]
=
X
i∈Ix
=
Ai ∩ [a, a]
X
i∈Ix
=
≤
j ∈ N0 δ −1 (vi · a) ≤ j ≤ δ −1 (vi · a)
X
X
i∈Ix
X
(9) i∈I
x
≤
j ∈ N0
j ∈ N0 vi · a ≤ dj ≤ vi · a
i∈Ix
≤
δ −1 (vi · a) − δ −1 (vi · a) + 1
δ −1 (vi · a) − δ
X vi · a − β
α
(9) i∈I
x
=
X
1+
i∈Ix
=
X
i∈Ix
=
dj
a≤
≤a
vi
1+
−
−1
(vi · a) + 1
vi · a − β
+1
α
β−β
vi · a − vi · a
+
α
α
β−β
α
· |Ix | + (a − a) ·
18
Vx
α
C. Implementing the Algorithms
=
β−β
1+
α
β−β
= 2 1+
α
· |Ix | +
(α + β − β) · |Ix | Vx
·
Vx
α
· |Ix |.
C. Implementing the Algorithms
In this section, we review existing algorithms for divisor methods. In particular, we
elaborate on how we have implemented them for our experiments [RW15a], and on
problems we have encountered in this process.
We have taken care not to render the algorithm unnecessarily inefficient in order to
perform a fair comparison of running times; the result is to the best of our abilities
conditioned on a limited time budget. In particular, all of our implementations have
been refined on the programming level to roughly the same degree.
For the purpose of a fair comparison, all implementation have to conform to the same
interface.
Parameters: A pair (α, β) ∈ R2 with α > 0 and β > 0.
Input: Votes v and house size k.
Output: A (symbolic) representation of all seat assignments valid w. r. t. divisor sequence
(αj + β)j≥0 , as well as proportionality constant a∗ .
More specifically, the output is encoded as a vector of undisputed seats and a binary
vector indicating which parties are tied for the remaining seats. We skip the step from
a∗ resp. a valid seat assignment to this representation in the pseudo code since it is
elementary: all parties with “current” resp. “next” value vi /dsi −1 resp. vi /dsi equal a∗
are tied. A simple Θ(n)-time post-processing identifies these in all cases.
We have established confidence in the correctness of our implementations by extensive
random testing [RW15a, TestMain.java]; every implementation has been run on thousands of instances. The correctness of the results has been confirmed, besides rudimentary
sanity checks such as matching vector dimensions, by checking Pukelsheim’s Max-Min
Inequality [Puk14, Theorem 4.5].
All implementations share the same numerical weakness, though: using fixed-precision
arithmetics, two computations that should lead to the same result (say, a∗ ) yield different
numbers. We compensate for that by using fuzzy comparisons: we identify numbers if
they are within some constant of each other. Thus, we can reliably identify tied parties,
for instance.
There is a drawback, though: if distinct values vi /dj are closer than (or, even without
the adaption, the resolution of the chosen fixed-precision number representation), we
may identify them and thus compute wrong seat assignments.
19
C. Implementing the Algorithms
This issue can not be circumvented on the algorithmic level. The only robust resort
we know of is using arbitrary-precision arithmetics, inevitably slowing down all the
algorithms.
C.1. Iterative Divisor Method
Implementing IterativeMethod is straight-forward. An implementation using a priority
queue implementation from the standard library runs in time Θ(n + k log n). Since we
expect overhead for the queue to be significant for small n, we also implement a variant
which determines I using a simple linear scan, resulting in a total running time in Θ(kn).
Shared code aside, IterativeMethod takes about 50 resp. 65 lines of code with resp.
without priority queues.
C.2. Jump-and-Step
The jump-and-step algorithm [Puk14, Section 4.6] can be formulated using our notation
as follows:
Algorithm 3: JumpAndStepd (v, k) :
Step 1 Compute an estimator a for a∗ .
Step 2 Initialize si = bδ −1 (vi · a)c + 1.
Step 3 Iterate similarly to IterativeMethod until
I=
arg maxn
i=1 vi /dsi ,
P
si < k;
arg minn
P
si > k.
i=1 vi /dsi −1 ,
P
si = k with
The performance of this algorithm clearly depends on ∆a := si − k after Step 2; the
running time is in Θ(n + |∆a | · log n) when using priority queues for Step 3 (which may
not be advisable in practice if |∆a | can be expected to be very small). As such, the
running time is not per se bounded in n and k.
P
We follow the recommendations of Pukelsheim and use the estimator [Puk14, Section 6.1]
α k + n · (β/α − 1/2),
a :=
·
V k + n · bβ/αc,
0 ≤ β/α ≤ 1;
else.
The first case corresponds to Pukelsheim’s recommended estimator for stationary signpost
sequences, the second to his good universal estimator generalized to divisor sequences
that are not signpost sequences in the strict sense. The additional factor α rescales the
value appropriately; Pukelsheim only considers α = 1.
20
C. Implementing the Algorithms
Given that these estimators guarantee |∆a | ≤ n in the worst case, we can assume
that JumpAndStep runs in time O(n log n). Furthermore, Pukelsheim claims that the
recommended estimator is good in practice in the sense that |∆a | ∈ O(1) in expectation,
so JumpAndStep may be efficient in practice for large n as well. Since their proof is
limited to uniformly distributed votes and k → ∞, we investigate this in Section 4.
Shared code aside, JumpAndStep takes about 120 lines of code, with or without priority
queues.
C.3. The Algorithm of Cheng and Eppstein
Cheng and Eppstein [CE14] do not give pseudocode for the main procedure of their
algorithm which would combine the individual steps to compute A(k) . For the reader’s
convenience and for clarity concerning our running-time comparisons we give this top-level
procedure as we have inferred it.
Algorithm 4: ChengEppsteinSelectd (v, k) :
Step 1 Compute a suitable finite representation of A.
Step 2 C := FindContributingSequences(A, k).
Step 3 ξ := s−1 (k) [CE14, (3)].
Step 4 If r(ξ, A) ≥ k then
ξ := LowerRankCoarseSolution(A, k, ξ).
Step 5 Return CoarseToExact(A, k, ξ).
The subroutines are given in sufficient detail in their Algorithms 1 to 3, respectively.
The pseudo code given uses some high-level set operations which we did not implement
naively due to performance concerns; we compute several steps during a single iteration
over the respective sets of sequences.
Note that we have (hopefully) fixed an off-by-one mistake in the text. The definition of
rank r(x, A) is, “the number of elements of A less than or equal to x”; that is, the rank
of A(j) is j + 1 since A is zero-based (the first element is A(0)). However, the authors
continue to say that r(x, A) “is the index j such that A(j) ≤ x < A(j + 1).”
Regarding performance, Cheng and Eppstein show that their algorithm runs in time Θ(n)
in the worst case. Since ChengEppsteinSelect computes a linear number of medians
and requires a linear number of evaluations of rank function r(x, A) (with geometrically
shrinking |A| – otherwise the algorithm would not run in linear time), it is unclear
whether the algorithm is efficient in practice.
Shared code aside, ChengEppsteinSelect take about 300 lines of code. By this
measure, it is the most complex of the algorithms we consider.
21
C. Implementing the Algorithms
Additional Issues with Numerics
In addition to the concerns expressed above, there are additional numerical issues when
implementing ChengEppsteinSelect using fixed-precision floating-point arithmetics.
In short, we have to compute certain floors and ceilings of real numbers exactly or we
may compute a wrong result.
More specifically, we evaluate r(x, A) several times by computing terms of the form
bδ −1 (_)c (cf. Lemma 4). The problem is that the result of δ −1 (_) is non-integral in
general, but is integral when the argument evaluates exactly to a dj . With the usual
floating-point arithmetic the result might be slightly smaller, though. We then erroneously
round down to the next smaller integer – a critical error!
In practice, we can add a small constant to the mantissa before taking the floor. This
constant has to be chosen large enough to cover potential rounding errors, but also
small enough so as to not change subsequent calculations; ChengEppsteinSelect may
compute a wrong answer otherwise. This is a very delicate requirement we do not know
how to fulfill in general.
C.4. SandwichSelect
We already discuss our algorithm at length in Section 3. Since we want to investigate
practical performance, we implement rank-selection using average-case efficient Quickselect
as opposed to using a linear-time algorithm with large constant factors.
We want to emphasize that our final algorithm SandwichSelect is conceptually simple
in the sense that there is little hidden complexity. We need exactly one call to a rank
selection algorithm on a linear-size list which takes five additional linear-time operations
to come up with: finding the maximal value v (1) , constructing index set Ix , computing Vx ,
constructing multiset  and computing k̂. These are all quite elementary tasks in that
they use one for-loop each which run for at most n iterations with only few operations
in each. We therefore think that we can outperform ChengEppsteinSelect in practice,
and should not be far behind JumpAndStep, either.
Regarding implementation, the delicate part was to get the bounds on j (cf. Step 6.1)
right. We use floor and ceiling functions on real numbers, so rounding errors that occur
in fixed-precision floating-point arithmetic can cause harm. We can circumvent this by
adding (subtracting) a conservatively large constant to the mantissa of the floats before
taking floors (ceilings). If this constant is larger than necessary for covering rounding
errors, we might add slightly more candidates to  (at most two per party) which
would slightly degrade performance. Correctness, however, is not affected (in contrast to
ChengEppsteinSelect).
We also remark here that the code [RW15a] for the experimental results discussed in
Section 4 is based on an earlier version of Lemma 2 with slightly weaker bounds (cf.
22
D. Experimental Setup
Appendix G). Experiments with the updated code are to follow, and might yield slight
improvements for SandwichSelect.
Shared code aside, SandwichSelect takes about 100 lines of code. By this measure, it
is the least complex of the non-trivial algorithms we consider.
D. Experimental Setup
We have run the experiments with Java 7 on Ubuntu 14.04 LTS running kernel 3.13.0-34generic x86_64 GNU/Linux. The hardware platform is a ThinkPad T430s Tablet with
the following core parameters according to lshw.
CPU: Intel R CoreTM i5-3320M CPU @ 2.60GHz
Cache: L1 32KiB, L2 256KiB, L3 3MiB
RAM: 4+4GiB SODIMM DDR3 Synchronous 1600 MHz (0.6 ns)
As our code is written in Java, we include a warm-up phase to trigger just-in-time
compilation of the relevant methods. All times are measured using the built-in method
System.nanoTime(). We use the same set of inputs for all algorithms, all of which have
to construct the full set S(v, k) for each input (v, k) during the measurement.
In order to increase accuracy, we repeat the execution of each algorithm on each input
several times and measure the total time; we then report the average time per execution.
For the selection-based algorithms, we use the randomized Quickselect-based implementation by Sedgewick and Wayne [SW11] as published on the book website. We use
the (pseudo) random number generators for several distributions from the same library
(download of stdlib-package.jar on August 11th, 2015).
For reproducing our running time experiments, make sure you have working GNU/Linux2
installation with Ruby, Java 7 and Ant; then execute
ruby run_experiments.rb arxiv.experiment
for the data represented in Section 4 and Appendix E. Be warned: this may run for a
long time, and it will create lots of images (provided you have gnuplot installed).
E. More Running-Time Experiments
We apologize to only offer draft graphics without commentary for the time being.
2
Our framework may work on other platforms, maybe with small adjustments to the Ruby code, but we
have not tried. See README.md for a workaround.
23
E. More Running-Time Experiments
Average normalized runtimes for several input distributions and across several orders of
magnitudes of n.
24
E. More Running-Time Experiments
Normalized runtimes of ChengEppsteinSelect for several input distributions and
across several orders of magnitudes of n.
25
E. More Running-Time Experiments
Normalized runtimes of JumpAndStep for several input distributions and across several
orders of magnitudes of n.
26
E. More Running-Time Experiments
Normalized runtimes of SandwichSelect for several input distributions and across
several orders of magnitudes of n.
27
E. More Running-Time Experiments
Normalized ∆a of JumpAndStep for several input distributions and across several orders
of magnitudes of n.
28
E. More Running-Time Experiments
Normalized |Â| of SandwichSelect for several input distributions and across several
orders of magnitudes of n.
29
E. More Running-Time Experiments
Runtimes against ∆a of JumpAndStep for several input distributions and across several
orders of magnitudes of n. Each color stands for one n.
30
E. More Running-Time Experiments
Runtimes against |Â| of SandwichSelect for several input distributions and across
several orders of magnitudes of n. Each color stands for one n.
31
F. Index of Used Notation
F. Index of Used Notation
In this section, we collect the notation used in this paper. Some might be seen as
“standard”, but we think including them here hurts less than a potential misunderstanding
caused by omitting them.
Generic Mathematical Notation
bxc, dxe . . . . . . . . floor and ceiling functions, as used in [GKP94].
M(k) . . . . . . . . . . The kth smallest element of (multi)set/vector M (assuming it exists);
if the elements of M can be written in non-decreasing order, M is given
by M(1) ≤ M(2) ≤ M(3) ≤ · · · .
Example: For M = {5, 8, 8, 8, 10, 10}, we have M(1) = 5,
M(2) = M(3) = M(4) = 8, and M(5) = M(6) = 10.
M (k) . . . . . . . . . . Similar to M(k) , but M (k) denotes the kth largest element.
x = (x1 , . . . , xd ) . . . to emphasize that x is a vector, it is written in bold;
components of the vector are written in regular type.
M . . . . . . . . . . . to emphasize that M is a multiset, it is written in calligraphic type.
M1 ] M2 . . . . . . . multiset union; multiplicities add up.
Notation Specific to the Problem
party, seat, vote (count), chamber size
Parties are assigned seats (in parliament), so that the number of seats
si that party i is assigned is (roughly) proportional to that party’s vote
count vi and the overall number of assigned seats equals the chamber
size k.
d = (dj )∞
j=0 . . . . . . the divisor sequence used in the highest averages method; d must be a
nonnegative, (strictly) increasing and unbounded sequence.
δ, δ −1 . . . . . . . . . a continuation of j 7→ dj on the reals and its inverse, both of which can
be evaluated in constant time.
n . . . . . . . . . . . . number of parties in the input.
v, vi . . . . . . . . . . v = (v1 , . . . , vn ) ∈ Qn>0 , vote counts of the parties in the input.
V . . . . . . . . . . . . the sum v1 + · · · + vn of all vote counts.
k . . . . . . . . . . . . k ∈ N, the number of seats to be assigned; also called house size.
s, si . . . . . . . . . . s = (s1 , . . . , sn ) ∈ N0 , the number of seats assigned to the respective
parties; the result.
ai,j . . . . . . . . . . . ai,j := dj /vi , the ratio used to define divisor methods; i is the party, j
is the number of seats i has already been assigned.
32
G. Changelog
Ai . . . . . . . . . . . For party i, Ai := {ai,0 , ai,1 , ai,2 , . . .} is the list of (reciprocals of)
party i’s ratios.
a . . . . . . . . . . . . We use a as a free variable when an arbitrary ai,j is meant.
A . . . . . . . . . . . . A := A1 ] · · · ] An is the multiset of all averages.
r(x, A) . . . . . . . . . the rank of x in A, that is the number of elements in multiset A that
are no larger than x; r(x) for short if A is clear from context.
a∗ . . . . . . . . . . . . the ratio a∗ = ai∗,j ∗ selected for assigning the last (i. e. the kth) seat;
corresponds to s by si = r(a∗ , Ai ); a∗ = A(k) (cf. Section 2 and
Section 3).
x . . . . . . . . . . . . an upper bound x > a∗ ; we use x = dk−1 /v1 + ε, where ε > 0 is a
suitable constant.
Ix . . . . . . . . . . . . Ix := {i | vi > d0 /x}; the set of parties i whose vote count is large
enough, so that ai,0 < x, i. e. so that they contribute to the rank of x in
A.
Vx . . . . . . . . . . . the sum of the vote counts of all parties in Ix .
Ax . . . . . . . . . . . the elements in A that are smaller than x, i. e., A ∩ (−∞, x).
a, a . . . . . . . . . . . lower and upper bounds on candidates a ≤ a ≤ a such that still
a∗ ∈ A ∩ [a, a].
G. Changelog
The following (substantial) changes have been made from arXiv version 2 to 3.
• Lemma 2 has been strengthened; both a and the upper bound on |A ∩ [a, a]| have
been improved. Both changes are due to the observation that we could require
β ≤ α without loss of generality.
Related notation update: (β̌, β)
(β, β).
• We have added Appendix A in order to clarify that the assumptions we make for
our main result do restrict the scope of divisor methods we cover by too much.
33
| 8cs.DS
|
Histogram of Oriented Depth Gradients for Action Recognition
Nachwa ABOU BAKR1
James CROWLEY2
Univ. Grenoble Alpes, INRIA, Grenoble INP, CNRS, Laboratoire LIG
1
[email protected]
2
[email protected]
arXiv:1801.09477v1 [cs.CV] 29 Jan 2018
Résumé
Cet article présente des expériences menées sur la perception de la profondeur de mouvements dans le cadre de la
reconnaissance visuelle d’actions à partir de mesures locales extraites de séquences vidéo RGBD encodées selon la
norme MPEG. Nous démontrons que ces mesures peuvent
être combinées avec des descripteurs RGB spatio-temporels
locaux pour l’élaboration d’une méthode de calcul efficace
pour la reconnaissance d’actions. Des vecteurs de Fisher
sont utilisés pour encoder et concaténer un descripteur de
profondeur avec les descripteurs existants RGB locaux. Ces
vecteurs sont ensuite utilisés pour la reconnaissance d’actions manuelles à l’aide d’un classifieur linéaire SVM. Nous
comparons l’efficacité de ces mesures à l’état de l’art sur
deux jeux de données récents pour la reconnaissance d’actions dans des environnements culinaires.
Mots Clef
Vision, Profondeur, Reconnaissance d’Actions, Codage
MPEG.
Abstract
In this paper, we report on experiments with the use of local measures for depth motion for visual action recognition
from MPEG encoded RGBD video sequences. We show that
such measures can be combined with local space-time video
descriptors for appearance to provide a computationally
efficient method for recognition of actions. Fisher vectors
are used for encoding and concatenating a depth descriptor
with existing RGB local descriptors. We then employ a linear SVM for recognizing manipulation actions using such
vectors. We evaluate the effectiveness of such measures by
comparison to the state-of-the-art using two recent datasets
for action recognition in kitchen environments.
Keywords
Vision, Depth, Action Recognition, MPEG encoding.
1 Introduction
Visual action recognition is the process of labelling image
sequences with action labels. In this work we investigate
the use of dense local descriptors for action recognition.
Dense sampling has been shown to improve results when
compared to sparse interest points for action recognition
[3]. Requirements for reduced computational time have lead
some authors to investigate methods that extract motion vectors directly from compressed video signals. For example,
methods have been proposed to extract flow directly from
MPEG encoded video [4]. The resulting method has been
shown to provide useful description of Dense Trajectories
(DT) [5], at the cost of some loss of recognition accuracy.
This has led us to ask if this approach can be extended to
provide fast computation of dense local descriptors from the
depth channel of MPEG encoded RGBD image sequences.
In the following, we report the results on experiments in
recognition of action labels using features extracted from
depth channel of Kinect-like sensors using MPEG encoded
motion flow. We describe a new descriptor based on oriented gradients of depth extracted from the motion channel
of RGBD image sequences. This descriptor is used in addition to the MPEG Flow (MF) descriptors based on the RGB
channels from the same sequence. Local video descriptors
using Dense Trajectories [5] and MPEG Flow [4] are employed as baseline descriptors for comparison.
DT, as a video descriptor for action recognition, starts by
sampling interest points densely at regular fixed grids. Detected interest points are then tracked over time and points
for which no motion is detected are eliminated. This set of
trajectories are then used as space-time interest points and
descriptors are computed around those points. Four descriptors have been used in DT : Histogram of Oriented Gaussian
(HOG), Histogram of Optical Flow (HOF), Motion Boundary Histogram in x,y directions (MBHx , MBHy ). An analysis on computational requirements of DT shows that 61%
of the running time is spent on the computation of optical
flows. Reducing this computational burden is an interesting
challenge.
The MF method exploits the estimated motion vectors used
for video compression and uses them as interest points. This
change provides a significant improvement in the time required for extraction of features at the cost of a small decrease
in recognition accuracy [4].
Both DT and MF are video descriptors computed around
motion trajectories. The primary differences between these
two methods are that the interest points to be described in
MF are at sparse positions of motion fields. In DT, points
are sampled at the positions of a regular fixed grid, and they
are tracked over time to build trajectories. Motion is encoded
in DT using Gunnar-Farnebäck optical flows while MF uses
estimated motion fields during video compression. Our goal
is to determine if it is possible to improve MF results by
exploiting data from depth channel of Kinect-like sensors.
2
Histogram of Oriented Depth Gradients
Depth can capture the structural features of objects. In addition to encoding the position of objects in space. Thus it is
reasonable to ask if we can improve the results of MF video
descriptor by including the depth channel to better describe
the scene. To do so, we compute gradients of depth which
encodes the structure of the depth image. As in MF, interest
points are a sparse set of points extracted while compressing
video. This ensures describing area around motion points.
Practically, a volume of 32 × 32 × 15 is computed around
interest points. This volume is divided by a grid of 2 which
produces a grid cell of 16 × 5. For each grid cell, we extract the gradients by applying the Sobel filter [1, 0, −1]
in a 16 × 16 pixels. Then, an 8-orientation bins discretized histogram is built to encode the gradient orientations.
Gradient magnitudes are used to weight the gradient orientations. For each temporal slice the gradient is normalized
using l2-normalization.
After that descriptors are encoded by transforming the collection of local video descriptors {xi , . . . , xN } into a fixedsize vector representation. Following [4], we chose to use
Fisher Vectors (FV) [6] for descriptor encoding. In particular, only first-order differences of FV are used. Action categories are then classified by a linear one-vs-rest SVM classifier.
3 Experiments
Experiments are performed on two publicly available datasets : 50 salads [1] and Actions of Cooking Eggs (ACE)
[2]. For 50 salads dataset ; we used provided annotations
of core actions only as verbs without taking into account
their corresponding objects ; in total, we used 6 different action labels. ACE also provides 6 different annotated actions
but only one level of granularity which includes pre- and
post movements of performed actions. For the evaluation,
we measured the mean of Average Precision (mAP).
We consider two image sequences ; one provides an RGB
view of the scene and the other encodes scene depth. To
study the impact of motion depth image sequences in describing actions, we extracted a Histogram of Oriented Depth
Gradients (HODG) from depth image sequences. We extracted as well RGB descriptors of DT 1 and MF 2 . Combining
both descriptors improves the results of MF method on two
datasets as shown in Table 1 where RGB+HODG is a concatenation of encoded descriptors by FV. We compare the results to DT since MF follows the same structure and set of
descriptors.
DT approach achieved the best results in this experiment at
the cost of the computation time. In the other hand, adding
HODG to MF descriptors achieves better results than [4]
while preserving the computational cost 3 . Table 2 shows the
number of processed frames per second while computing
both RGB-based and HODG using the implementation of
DT and MF. From the tables, we can also see that HODG
alone performs relatively good in action recognition.
DT(RGB)
MF(RGB)
MF(HODG)
MF(RGB+HODG)
50salads
99.60%
90.56%
86.29%
94.73%
ACE
94.12%
92.04%
89.94%
99.15%
TABLE 1 – The results of our experiments using mAP.
RGB (HOG, HOF, MBH)
Depth (HODG)
DT (fps)
3.87
6.41
MF (fps)
27.78
111.13
TABLE 2 – The number of processed frames per second
while descriptors extraction for 50 salads dataset.
that can be treated in parallel with no additional cost or sequentially while preserving a close to real-time capability.
The idea is to combine the gradients of depth image sequences (HODG) extracted at sparse motion fields which
are estimated in video compression to the set of RGB descriptors of MF [4]. Both RGB and Depth descriptors are encoded using FV.
We found that trajectories of depth image sequences, described by HODG, can describe manipulation actions with an
acceptable recognition precision. To this end, we evaluated
our proposed method on two recent datasets ; 50 salads, and
ACE. We recorded the computational time during extraction
of descriptors using both methods. Results show a significant improvement in recognition precision with almost no
significant additional cost.
Références
[1] Stein, S. and McKenna, S.J. "Combining embedded accelerometers with computer vision for recognizing food
preparation activities." Proceedings of the 2013 ACM
international joint conference on Pervasive and ubiquitous computing. ACM, 2013.
[2] Shimada, A., Kondo, K., Deguchi, D., Morin, G. and
Stern, H. "Kitchen scene context based gesture recognition : A contest in ICPR2012". Advances in depth image
analysis and applications. Springer Berlin Heidelberg,
2013. 168-185.
[3] Wang, H., Ullah, M.M., Klaser, A., Laptev, I. and
Schmid. "Evaluation of local spatio-temporal features
for action recognition". BMVC 2009-British Machine
Vision Conference. BMVA Press, 2009.
[4] Kantorov, V. and Laptev, I. "Efficient feature extraction, encoding and classification for action recognition". In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2014.
[5] Wang, H., Kläser, A., Schmid, C. and Liu, C.L. "Action
recognition by dense trajectories". Computer Vision and
Pattern Recognition (CVPR), 2011 IEEE Conference
4 Conclusion
on. IEEE, 2011.
[6] Sánchez, J., Perronnin, F., Mensink, T. and Verbeek,
In this work, we reported an improvement of the results of
J. "Image classification with the fisher vector : Theory
MF method by using the benefit of depth channel in Kinectand practice". International journal of computer vision
like sensors. Since depth channels provide separate images
105.3 (2013) : 222-245.
1. https://lear.inrialpes.fr/people/wang/dense_trajectories
[7]
Pedregosa,
Fabian, et al. "Scikit-learn : Machine lear2. https://github.com/vadimkantorov/cvpr2014
ning
in
Python."
Journal of Machine Learning Research
3. Each channel of the sensor can be treated individually in the descriptor extraction process
12.Oct (2011) : 2825-2830.
| 1cs.CV
|
Classifying Time-Varying Complex Networks on the Tensor Manifold
Romeil Sandhu
[email protected]
arXiv:1708.07904v1 [cs.SY] 25 Aug 2017
Departments of Biomedical Informatics, Computer Science and Applied Mathematics/Statistics
Stony Brook University, Stony Brook, NY 11794
Abstract
At the core of understanding dynamical systems is the
ability to maintain and control the systems behavior that includes notions of robustness, heterogeneity, and/or regimeshift detection. Recently, to explore such functional properties, a convenient representation has been to model such
dynamical systems as a weighted graph consisting of a finite, but very large number of interacting agents. This said,
there exists very limited relevant statistical theory that is
able cope with real-life data, i.e., how does perform simple
analysis and/or statistics over a “family” of networks as
opposed to a specific network or network-to-network variation. Here, we are interested in the analysis of network
families whereby each network represents a “point” on an
underlying statistical manifold. From this, we explore the
Riemannian structure of the statistical (tensor) manifold in
order to define notions of “geodesics” or shortest distance
amongst such points as well as a statistical framework for
time-varying complex networks for which we can utilize in
higher order classification tasks.
1. Introduction
Notions of robustness, heterogeneity, and phase changes
are ubiquitous concepts employed in understanding complex dynamical systems with a variety of applications [1, 2,
3, 4]. This is seen in Figure 1. While recent advancements
in network analysis has arisen through the usage of spectral techniques [5], expander graphs [6], percolation theory
[7], information theoretic approaches [8, 9], scale-free networks [4], and a myriad of graph measures reliant on the
underlying discrete graph space (e.g., degree distribution
[10], shortest path [8], centrality [11, 12, 13]), such measures rarely incorporate the underlying dynamics in its construction. Limitations of such static measures can be more
aptly seen in multilayer networks whereby standard network measures are immediately ambiguous; shortest paths
Figure 1. This work focuses on the exploiting the underlying geometry in a statistical framework to elucidate regime-shifts and
bifurcation detection in dynamical systems. We apply our analysis
towards broader cell differentiation and signaling promiscuity in
biological systems with cancer applications (e.g., inducing terminal homogeneous state from a pluripotent stem-cell like state)
(geodesics) are non-trivial [14] while centrality measures
[15], clustering methods [16], and diffusion models [17]
require simplifications to make sense of real-world configurations. Moreover, static properties have been unable to
model time-varying nonlinear higher order effects (e.g., targeted therapy [18], cellular reprogramming [19]) which can
be seen pictorially in Figure 1 and Figure 2. This is akin to
waterbed theory which refers to the observation of when a
complex behavior is “pushed” down, the invariable effect of
complexity causes for that behavior to “pop-up” elsewhere
[20].
To this end, we have previously developed fundamental
relationships between network functionality [8, 21] and certain topological and geometric properties of the corresponding graph [22, 23]. In particular, recent work of ours has
shown the geometric notion of curvature (a measure of “flatness”) is positively correlated with a system’s robustness or
its ability to adapt to dynamic changes [24, 25]. This can
be seen in Figure 3A. In this regard, network curvature may
relate to anomaly detection [26], congestion in communication [27, 28], drug resistance [29], supply chain risk [30], to
Figure 2. Current modeling of static network properties leads to ill-posed “whack-a-mole”-based solutions whereby to alter functionality,
one “whacks” a vulnerabilities of a system at a given time t. However, due to heterogeneity, this leads to new vulnerabilities and further
(perhaps accelerated) breakdown via feedback.
even systemic risk in banking systems [31, 32]. Moreover,
this is particularly compelling since curvature can be readily
computed for large-scale networks using theory from optimal mass transport [33], and more importantly, highlights
the significance of employing geometry in network analysis. This said, there exists very limited relevant statistical
theory that is able cope with a family of networks and one
often focuses on a given network or is limited to measures
that exploit network-to-network variations as it pertains to
functionality. Here, we are interested in expounding upon
previous work [24] in order to exploit statistics over a family of networks such that higher order classification tasks
can be performed to better understand regime shifts.
To do so, we are interested in studying geometry of
networks as a dynamical system that evolves over time in
which each given network or observation may represent
a “point” (encoded by some matrix-based positive definite
model) on an underlying Riemannian manifold. From this,
we are interested in defining geodesics such that distances
between “points” (networks) in a given a family can be measured and proper statistical analysis can be achieved. For
a broad range of problems, such information may include,
but not limited to, vector-valued time series, power spectral
density [34], to density matrices from quantum mechanics
that aptly describes the mixed state of the system [35]. In
general, the set of Hermitian positive definite (PD) matrices
plays a fundamental role in a variety of engineering applications with the most recent attention renewed on developing
lower order statistics [36, 37, 38, 39]. In relationship to this
mathematical structure, significant related work on employing statistical analysis for Riemannian manifolds has focused on directional statistics, statistics on shape spaces, as
well as specific manifolds - we refer the reader [40, 41, 42]
and references therein. This being said, while these methods are mostly extrinsic whereby one embeds the manifold
in an ambient Euclidean space, we are interested in studying intrinsic Riemannian metrics that can be naturally applied to the network setting without restrictions on the network topology. For example, from a network geometric per-
spective, there has been very interesting work modeling the
space of tree-like structures in understanding phylogenetic
trees for cancer and evolutionary biology [43, 44]. Recently,
phylogenetic trees were shown to live in a negatively curved
space [43] resulting in a new push to develop an appropriate statistical framework (i.e., if a manifold is negatively
curved, then one can define a unique geodesic). This has
led to methods in hierarchical learning and clustering [45].
Nevertheless, as seen in Figure 4, non tree-like networks
still “live” upon some Riemannian (statistical) manifold for
which we need to construct a framework capable of resolving unique geodesics (shortest path) amongst such points.
In turn, we can then perform statistics on this space allowing
us to classify certain behaviors and operating regimes which
will be key to maintaining control and avoiding shotgunbased solutions during “black swan” events [46] whereby
the continuous failing of interacting agents may result in
catastrophic failure.
The remainder of the present paper is outlined as follows: In the next section, we first provide important preliminaries in motivating the theoretical need of understanding
geometry by revisiting curvature and entropy as it pertains
to functionality. Here, we see that changes in the geometry provides intrinsic information about the dynamic system and thus, its exploration fits towards a larger program
developed in this paper and current related work [24, 25].
From this, Section 3 lays the foundation of a Riemannian
framework previously employed for Diffusion Tensor Imaging whereby we now adapt such notions towards complex
networks. Then, Section 4 presents preliminary results on
synthetic and biological data. We conclude with a summary
and future work in Section 5.
2. Geometry and System Behavior
To illustrate how geometry elucidates the functional behavior of a dynamical system and how such ingredients are
ever-important the analysis and control of time-varying networks, let us revisit optimal mass transport (OMT) [33].
Figure 3. Visual illustration of curvature. (A) The thematic and motivating vision of this work is based on the above-illustrated result
[24, 25]. (B) Other forms of curvature that play important roles in (projective) geometry and machine learning.
The first notion of OMT was proposed by Gaspar Monge
in 1781 with the concern of finding the minimal transportation cost for moving a pile of soil from one site to another.
The modern formulation, given by Kantorovich, has been
ubiquitously used in fields of econometrics, fluid dynamics, to shape analysis [33, 47, 48] and recently, has received
a renewed mathematical interest [49, 50]. More formally,
let (X, µ0 ) and (Y, µ1 ) be two probability spaces and let
π(µ0 , µ1 ) denote the set of all couplings on X × Y whose
marginals are µ0 and
R µ1 . As such, the Kantorovich costs
seeks to minimize c(x, y)dπ(x, y)∀π ∈ π(µ0 , µ1 ) where
c(x, y) is the cost for transporting one unit of mass from x to
y. The cost originally defined in a distance form on a metric
space leads to the Lp Wasserstein distance as follows:
Wp (µ1 , µ2 ) :=
Z Z
inf
p1
d(x, y) dµ(x, y))
p
µ∈π(µ0 ,µ1 )
(1)
From this, let us begin considering M to be a Riemannian
manifold for which “points” are now defined
Z
P := µ ≥ 0 : µ dvol(M ) = 1
Z
(2)
Tµ P := η : η dvol(M ) = 0
as the space of probability densities and the tangent space at
a given point µ, respectively. Due to the work of Benamou
and Brenier [48], one can naturally compute the geodesic
(in the Wasserstein sense) between two densities µ0 , µ1 ∈
P as the below optimal control problem:
Z Z 1
inf
µ(t, x)k∇g(t, x)kdtdvol(M )
µ,g
0
subject to
∂u
+ div(µ∇g) = 0
∂t
µ(0, .) = µ0 ,
(3)
µ(1, .) = µ1
which leads us to give P a Riemannian structure due to
some very nice work of Jordan et. al [51]. Armed with
this, we are now able to see how changes to the underlying geometry (namely curvature) may serve as a proxy for
system behavior (namely functional robustness). As such,
let
Z
H(µt ) :=
log µt dvol(M )
(4)
M
represent Boltzmann entropy where the dependency on x
has been dropped for convenience and we consider a family
of densities evolving over time. Taking the second variation
with respect to time t in the Wasserstein sense (i.e., our gradient steps are taken to minimize or maximize the Wasserstein distance rather than the classical Euclidean norm) and
noting that, by construction, η := ∂µ
∂t |t = 0, we have that
d2
H(µt )|t=0 = hHess(H)(η), ηiW
dt2 Z
1
=−
h∇gη , ∇∆gη i + ∆ k∇gη k2 µ0 dvol(M ) (5)
2
M
where µ0 and gη satisfy equation (3). Using the Bochner
formula [54], which relates harmonic functions on a Riemannian manifold to Ricci curvature (a geometric measure
of “flatness” and herein denoted as “Ric”), we can further
assume Ric ≥ kI as quadratic forms where k is a constant
and I is the identity matrix. Then, due to a beautiful result of Sturm [53] as well as Lott and Villani [33], one
can show that the Hess(H) is k-convex with respect to the
Wasserstein norm:
H(µt )≤ tH(µ0 )+(1 − t)H(µ1 ) − φ(k, t, µ0 , µ1 )∀t ∈ [0, 1]
(6)
where the right hand portion of the above equation denoted as φ(.) can be shown to be φ(k, t, µ0 , µ1 ) = k2 t(1 −
t)W2 (µ0 , µ1 )2 allowing for k-convexity. That is, changes in
entropy and curvature are positively correlated, i.e., ∆H ×
∆Ric ≥ 0. Now, let p (t) denote that probability that the
mean of a given observable deviates by more than from
the original (unperturbed) value at time t. Then from large
deviations theory, one can define a rate function as:
1
R := lim
− log p (t)
(7)
t→∞,→0
t
Therefore, a large R means a fast return to the original state
(robustness), and a small R corresponds to a slow return
(fragility). In thermodynamics, it is well-known that entropy and rate functions from large deviations are closely
related [21]; the Fluctuation Theorem [8] is an expression
of this fact for complex networks and may be expressed as
∆H × ∆R ≥ 0. With the result of equation (6), we have
∆Ric × ∆R ≥ 0
(8)
which states that changes in Ricci curvature are positively
correlated to changes in robustness. Perhaps more importantly, the manner in which the underlying geometry
changes elucidates certain behaviors of a complex dynamical system and this stylized fact, sets the motivation
for this note in which we attempt to generalize statistics to
a family of networks that live on a certain manifold.
2.1. Scalar Model: Network-to-Network Analysis
As highlighted in the previous section, geometry of networks plays an integral role in understanding functionality. In particular, in the scalar case with the structure given
by equation (2), one can discretely defined Ricci Curvature
(due to Ollivier [22]) between any two “points” x and y as:
κ(x, y) := 1 −
W1 (µx , µy )
d(x, y)
(9)
This definition, motivated by coarse geometry, is applicable
to the graph setting whereby the geodesic distance d(x, y)
is given by the hop metric and W1 can be computed simply
via linear programming [55]. However, this measure is limited to local network-to-network variation as opposed to a
family of networks and as such, we require a generalization
to the matrix valued setting; a non-trivial problem that has
been recently studied [50]. Therefore, this note and latter
(future) work will seek to construct matrix-valued curvature
based measures. To do so, we first need notion of matrixvalued distance dM (x, y).
3. Manifold of Time-Varying Networks
This section will introduce a framework capable of defining statistics on a family of networks as seen in Figure 4.
3.1. Geodesics on the Network Manifold
Consider a dynamical system that evolves over time and
whose information is encapsulated by a positive definite
matrix (tensor). In this regard, it is well-known that the
space of tensors is not a vector space, but instead forms a
Riemannian manifold M [56]. In particular, we will leverage the theory of symmetric spaces which has been extensively studied since the seminal work of Nomizu [57]; a
comprehensive work on tensor manifolds can be found in
[58, 59]. Riemannian manifolds are endowed with a metric
that smoothly assign to each point ρ ∈ M an inner product
on Tρ , the tangent space to M at ρ. More formally, let us
denote Υ and Υ+ and as the set of all Hermitian matrices
and the cone of positive-definite matrices, respectively. We
may then define an analogous structure to equation (2) as:
Λ+ := ρ ∈ Υ+ | tr(ρ) = 1
Tρ := χ ∈ Υ | tr(χ) = 0
(10)
where Λ+ represents our space of networks and Tρ is
the corresponding tangent space. From this, we let
the observations of our dynamic system be recorded as
{ρ0 , ρ1 , ρ2 , ..., ρt } and where each “point” (network) ρi ∈
Υ+ is given by a tensor. As seen in Figure 4, we need a
sensible notion of distance that is dependent on the underlying geometry to define statistics for classification. To do,
one defines the exponential map as the function that maps
→
+
to each vector −
ρ−
0 ρ1 ∈ Tρ0 , the point ρ1 ∈ Λ on the manifold M that is reached after unit time by the geodesic starting at ρ0 with this tangent vector. The exponential map,
herein denoted as expρ : Tρ 7→ M at point ρ is defined on
the whole tangent space and for the particular space of tensors, this map is also one-to-one. Moreover, one can also
define a unique inverse map denoted as the logarithm map
Logρ0 : M 7→ Tρ0 that maps a point ρ1 ∈ M to the unique
tangent vector χ ∈ Tρ0 at ρ0 whose initial velocity is that of
the unique geodesic γ with γ(0) = ρ0 and γ(1) = ρ1 .
Thus, the problem of classifying time-varying networks
amounts to being able to properly computing geodesics on
this space and for which, we must now consider a family of
curves γ(t) on the manifold and its speed vector γ̇(t). Then,
in the general setting, our geodesic will be a curve that realizes the minimum distance (length) between any two points,
e.g., ρ0 and ρ1 . That is, we want to compute:
Lρρ10 = min
γ
Z
ρ1
ρ0
Z ρ1
= min
γ
kγ̇(t)kγ(t) dt
hγ̇(t), γ̇(t)iγ(t)
(11)
ρ0
where γ(0) = ρ0 and γ(1) = ρ1 . Given this, we seek to
exploit the Riemannian structure of the tensor space proposed by Pennec [59] towards the network setting; most recently used in Diffusion Tensor Imaging [60, 61]. Here,
authors utilize a result from differential geometry regarding
geodesics for invariant metrics on affine symmetric spaces
[62, 63] together with the Sylvester equation from control,
to propose an affine invariant metric whereby the unique
geodesic curve from point ρ (and at the origin) on our network manifold with a tangent vector χ can be shown to be
1
1
1
1
γ(t)(ρ,χ) = ρ 2 exp(tρ− 2 χρ− 2 )ρ 2
γ(t)(I,χ) = exp(tχ)
(12)
Figure 4. (A) To understand more global functionality, we treat time-varying networks as “points” on a manifold. (B) The study geodesics
is needed in order to compute a statistical framework for a family of networks. (C) Manifold of Positive Definite Matrices
P+∞
k
where exp(χ) = k=0 = χk! is the usual matrix exponential. From this, the speed vector γ̇(t) can be computed:
dγ(t)
d
=
exp(tχ)
dt
dt
tχ
e 1,1
d
0
Q
=
dt
0
0
etχ2,2
0
= QDIAG χi exp(tχi ) QT
1
!
...
0
T
...
0
Q
... etχn,n
1
Σ=
= exp(tχ) 2 χ exp(tχ) 2 )
1
2
= γ(t)(I,χ)
?χ
(13)
T
where the A ? B = ABA . Noting the geodesic γ(t)(ρ,χ)
between any two points ρ0 , ρ1 ∈ M where ρ1 = expρ0 (χ)
at t = 1, the logarithm map can be seen as
1
−1
−1
1
γ(1)(ρ0 ,χ) = ρ1 = ρ02 exp(ρ0 2 χρ0 2 )ρ02
−1
−1
−1
It has been noted that although no closed form expression
for the mean exists, the above converges towards a steadystate solution in a few iterations. As in the traditional Euclidean distance and similar to the Karcher or Frechet mean,
the above mean is constructed on the premise that we seek
a “point”
the sum of squared distance.e.,
PN that minimizes
2
ρ =
dist
(ρ,
ρ
).
In
a similar fashion, the variance
i
i=1
can be computed as follows:
− 12
log(ρ0 2 ρ1 ρ0 2 ) = ρ0 2 χρ0
1
1
−1
−1
→
ρ02 log(ρ0 2 ρ1 ρ0 2 )ρ02 = χ = −
ρ−
0 ρ1 = Logρ0 (ρ1 )
(14)
From this, the distance between networks can finally be
computed as
dist2 (ρ0 , ρ1 ) = kLogρ0 (ρ1 )k2ρ0
−1
−1
= klog(ρ0 2 ρ1 ρ0 2 )k22 .
(15)
We now have a natural distance that measures the length of
the geodesic curve that connects any two “points” on our
network manifold. This is particularly important as we can
now begin to perform statistics over a family of networks.
3.2. Mean, Variance, Mahalanobis Distance
Given the above geodesic distance, one can formulate necessary statistics on our network manifold. Using
the framework established by Pennec [59] with the above
affine invariant metric, the unique mean ρ̄ of N “points”
{ρ0 , ρ1 , ρ2 , ..., ρN } on the tensor manifold can be computed
via gradient descent:
X
N
1
1
1
1
−
−1
ρ̄t+1 = ρ̄t2 exp
log ρ̄t 2 ρi ρ̄t 2
ρ̄t2
(16)
N i=1
N
T
1 X
Vecρ̄ Logρ̄ (ρi ) Vecρ̄ Logρ̄ (ρi )
(17)
N − 1 i=1
where Vecρ̄ (ρi ) = VecI log(ρ̄ ? ρi ) and where the operation VecI (A) is a vectorized projection of the independent coefficients of our
√ tensor in√the above√formulation, i.e.,
VecI (A) = (a1,1 , 2a1,2 , ... 2a1,n , ... 2an−1,n , an,n ).
From this, one obtains a generalized form of the Gaussian
distribution on the tensor manifold as:
1
N(ρ̄,Σ) (ρi ) = k exp
ξ(ρ̄,Σ) (ρi )
(18)
2
where ξ(ρ̄,Σ) represents the classic Mahalanobis distance
now defined in the Riemannian setting as seen below
T
ξ(ρ̄,Σ) (ρi ) = Vecρ̄ Logρ̄ (ρi ) Σ Vecρ̄ Logρ̄ (ρi ) . (19)
While the above provides a few basic tools, we have omitted, for the sake of brevity varying other advantages that
can be employed towards the network setting through the
framework established above. This will be a subject of future work in which powerful tools, previously employed for
DTI imaging, will now be used to exploit time-varying networks. Of course, to do so, we need to define our matrixbased model that will allow for networks to be represented
on the tensor manifold. This is discussed next.
3.3. Matrix-Based Model: Graph Laplacian
Given that we are focused network control, we employ an approximated graph Laplacian as our matrix-based
model whose structure provides a global representation of a
given network with many useful properties that are intrinsically tied to system functionality. While a complete review
Figure 5. (A) Graphical view of the three synthetic “toy” graphs that can be representative of varying types of communication networks,
each with 200 nodes and 400edges (i.e., only topology changes). (B) Distances of network-to-network under varying weights or “noise.”
We see that the proposed framework is able to properly quantify difference as compared to Euclidean measure, i.e., the star-like network is
“closest” to the random network as supported in [8, 24].
is beyond the scope of this note (we refer the reader to [5]),
a few areas that have garnered attention have included construction of expander graphs [6] to Cheegers inequality with
increasing attention on connections between spectral graph
theory and its respective geometry [64]. Mathematically,
one can define the Laplacian
P operator for a specific network
as ∆f (x) = f (x) − y f (y)µx (y) with f being a realvalued function which coincides with the usual normalized
graph Laplacian given by:
1
1
L = I − D− 2 AD− 2
(20)
It is interesting to note in this connection that if k ≤ κ(x, y)
is a lower bound for Ricci curvature defined in equation (9),
then the eigenvalues of Laplacian is bounded k ≤ λ2 ≤
... ≤ λN ≤ 2 − k [64]. This relationship is important
since 2 − λN measures the deviation of the graph from being bipartite, i.e., a graph whose vertices can be divided into
two disjoints sets U and V such that every edge connects a
vertex in U to one in V . As such ideas appear in resource
allocation and control, the study of such structures in the
time-varying setting motivate our focus on the Laplacian
matrix. Lastly, while one can form a normalized Laplacian
with unitary trace, we note that the Laplacian matrix by
definition is a positive semi-definite matrix. Therefore, to
enforce positive definiteness, we utilized an approximated
Laplacian L̂ := L + L I where 0 < L << 1 such that
xT L̂x > 0 for all nonzero x ∈ Rn .
4. Results
We now present preliminary results. While these experiments are by no means exhaustive (and will be extended),
these results highlight potential of the above framework for
time-varying network analysis.
4.1. Toy Network Configurations
To provide initial intuition of the proposed framework,
we constructed three toy “communication” based networks.
As one can see in Figure 5, the networks are composed of
200 nodes and 400 edges with varying topology that represents a chain, a random, and a star-like network. Similar
to our earlier work [24], we are interested in using the proposed framework to measure distances amongst these structures under varying “noise”. As highlighted by network
curvature [24] and network entropy[8], one should expect
from a communication perspective the star-like network is
“closer” to a “random” network. That is, given that their
functional properties such as robustness (a measure of “flatness”) are similar compared towards the chain-like network,
we should expect that such networks on the manifold lie
closer to one another. This is exactly what is shown in Figure 5B. Moreover, as we move from an unweighted graph
to a weighted graph in which weights are chosen randomly
in a uniform manner, this relationship still holds. On the
other hand, if we utilize the classical Euclidean distance
(Frobenius norm) to measure network-to-network distance,
we find that the chain-like network is in fact “closer” to the
star-like network. From a communications perspective, this
does not make sense and highlights the proposed framework. Lastly, it is interestingly to see that as we increase
the support of weight selection, the distances between such
networks increases as to be expected.
4.2. Scale-Free and Random Networks
This section presents clustering results and the importance of exploiting the underlying geometry as it relates
scale-free networks. In particular, we generate 50 scale-free
networks (i.e., degree distribution follows power law) [1]
and 50 random (Erdos-Renyi) networks with each network
composed 200 nodes and 1164 edges (differences are only
in topology). Such scale-free networks stylistically could
represent “business as usual” behavior while such random
networks may represent a particular crisis in which structure (correlation of agents) breaks down. From this, one
can generate an approximated normalized graph Laplacian
Figure 6. Heat map of distances between scale-free and random networks measured by the proposed Riemannian metric to that of the
Frobenius norm. (B) Average distance of network-to-network under varying weights or “noise.” We see that the proposed framework is
able to properly quantify difference as compared to Euclidean measure. Note: Each random and scale-free network possess same number
of nodes and edges, only topology changed.
as defined in Section 3.3 whereby each network now represents a point on our tensor manifold. In doing so, we compute pairwise distances amongst all potential network pairs
based on the Euclidean distance, defined by the Frobenius
norm, as well as the Riemannian distance, defined by equation (15). Figure 6A presents a “heat map” of the distances.
As one can see, Euclidean distance fails to cluster networks
as the distance between scale-free to random networks are
“closer” than random to random networks. The framework proposed here is able to properly cluster the aforementioned networks highlighting the thematic importance
of using Riemannian geometry when exploiting distances
and/or similarity measures. Figure 6B also presents average distances amongst scale-free-to-scale-free, randomto-scale-free, and random-to-random networks. Lastly, we
note that if the above experiments were repeated and/or the
number of nodes/edges were changed, the clustering result
would still hold.
4.3. Waddingtons Landscape: Cell Differentiation
In this section, we now turn our attention in utilizing the
proposed framework towards a regime shifts of cellular differentiation of Waddington’s Landscape as depicted in Figure 1. This epigenetic landscape provides a stylistic hierarchal view cellular differentiation [65], i.e., landscape elevation associates to varying degrees of pluripotency or the
ability for a cell to differentiate to multiple function specific tissues. That is, pluripotent and progenitor cells are
“higher” compared to a further differentiated state whether
that be another function specific stem-cell or tissue. Nevertheless, at a high level, one can consider the intricate and
complex processes of cellular differentiation as a interconnected dynamical system (network) whereby such bifurcations or regime-shift relate to differentiated states that lie on
Waddington’s landscape [66].
To illustrate this and from a classification perspective, we
accessed the Gene Expression Omnibus (GEO). In particular, we selected two public datasets, GSE11508 and GSE
30652, that possess 59 pluripotent / 160 non-pluripotent and
159 pluripotent / 32 non-pluripotent cell samples, respectively. Within each pluripotent samples of both datasets,
there contains varying types that include human embryonic
stem cells (hESCs) and induced pluripotent stem cells (iPSCs). This results in a potential mix of differentiation,
i.e., iPSCs pluripotency may be much lower than hESCs
yet are deemed pluripotent. Here, we are primarily interested to examine how “far” apart differentiated samples
are from their counterparts and would be key in devising
a necessary quantitive framework characterizing Waddingtons Landscape. As such, through the use of the KEGG
database downloaded from Pathway Commons, we construct a single pluripotent network whereby the weights are
the gene-to-gene correlation of twenty random pluripotent
samples selected from each of the respective data sets. This
was done similarly for a single non-pluripotent network.
Together, this was repeated to generate 50 pluripotent and
non-pluripotent networks from each GSE11508 and GSE
30652. Figure 7 presents heat map results as well as the average network distance between such networks. As one can
see, we see a marked increase in distance between pluro
and non-pluro samples. We note the Frobenius norm follows this same pattern and can primarily be attributed to the
fact that we fixed the topology between pluro/non-pluro networks, the small number of samples for correlation, as well
as the dependency of our underlying interactome KEGG
network. This said, we have presented such results here
to motivate the need for future work.
5. Summary & Conclusion
This note introduces a statistical geometric framework
[59] towards time-varying network analysis. While such
Figure 7. Heat map of Riemannian distances between pluripotent and non-pluripotent networks under two public datasets. (B) Average
numerical distance between the pluro and non-pluro networks. We see the proposed measure is able to classify such samples as well as the
Euclidean measure. We note the topology here is fixed across each networks unlike previous synthetic results.
mathematical concepts have been successfully employed
for DTI Imaging [61], the underlying structure is equally
relevant in the classification of regime-shifts in networks.
To illustrate the importance of geometry in this setting, we
presented preliminary results that show classical Euclidean
distances are unable to properly differentiate between not
only toy network configurations, but that of scale-free and
random networks. We concluded with an example of how
the proposed framework may be utilized on real-life biological stem-cell data. Putting this all together, these preliminary results and methodology set the foundation for a
more expansive study on global notions of robustness, heterogeneity, and/or regime-shift detection.
References
[1] A. Barabasi. “The Network Takeover,” Nature
Physics. 2012. 1, 6
[2] H. Kitano. “Cancer as a robust system: implications
for anticancer therapy.” Nature Reviews Cancer. 2004.
1
[3] M. Csete and J. Doyle. “Reverse engineering of biological complexity.” Science. 2002. 1
[4] A. Barbasi and R. Albert. “Emergence of Scaling Random Networks.” Science. 1999. 1
[5] R.K. Fan Chung. “Spectral Graph Theory.” American
Mathematical Society. 1997. 1, 6
[6] S. Hoory, N. Linial, and A. Widgerson. “Expander
Graphs and Their Applications.” American Mathematical Society. 2006. 1, 6
[9] J. West, G. Bianconi, S. Severini, and A. Teschendorff. “Differential Network Entropy Reveals Cancer
System Hallmarks.” Nature (Scientific reports) 2012.
1
[10] I. Taylor, R. Lindling, D. Warde-Farley, Y. Liu, C.
Pesquita, D. Faria, S. Bull, T. Pawson, Q. Morris,
and J. Wrana. “Dynamic Modularity in Protein Interactions Networks Predicts Breast Cancer Outcome.”
Nature Biotechnology. 2009 1
[11] S. Borgatti. “Centrality and Network Flow.” Social
Networks. (2005). 1
[12] S. Borgatti and M. Everett. “A Graph-Theoretic Perspective on Centrality.” Social Networks. (2006). 1
[13] G. Ghoshal and A. Barabasi. “Ranking Stability and
Super-Stable Nodes in Complex Networks.” Nature
Communications. 2011. 1
[14] P. Brodka, P. Stawiak, and P. Kazienko. “Shortest Path
Discovery in the Multilayered Social Network.? International Conference on Advances in Social Networks
Analysis. 2011. 1
[15] M. Domenico, A. Ribalta, E. Omodei, S. Gomez,
and A. Arenas. “Ranking in Interconnected Multilayer
Networks Reveals Versatile Nodes.” Nature Communications. 2015. 1
[16] P. Mucha, T. Richardson, K. Macon, M. Porter, and
J.P. Onnela. “Community Structure in Time- Dependent, Multiscale, and Multiplex Networks.” Science.
2010. 1
[7] D. Stauffer and A. Aharony. “Introduction to Percolation Theory.” Taylor and Francis. 1994 1
[17] M. Domenico, C. Granell, M. Porter, A. Arenas. “The
Physics of Spreading Processes in Multilayer Networks,” Nature Physics. 2016. 1
[8] L. Demetrius and T. Manke. “Robustness and Network
Evolution: Entropic principle.” Physica A: Statistical
Mechanics and its Applications. 2005. 1, 4, 6
[18] L. Kwong, T. Heffernan, L. Chin. “A systems biology
approach to personalizing therapeutic combinations.”
Cancer Discovery. 2013. 1
[19] J. Utikal, J. Polo, M. Stadtfeld, N. Maherali, W.
Kulalert, R. Walsh, A. Khalil, J. Rheinwald, and
K Hochedlinger. “Immortalization eliminates a roadblock during cellular reprogramming into iPS cells.”
Nature. 2009. 1
[33] C. Villani. Optimal Transport, Old and New. SpringerVerlag. 2008. 2, 3
[20] E. Nicola. “When Bode meets Shannon: Controloriented feedback communication schemes.” IEEE
transactions on Automatic Control. 2004. 1
[35] S. Gustafson and I. M. Sigal. Mathematical Concepts
of Quantum Mechanics. Springer. 2011. 2
[21] S. R. S. Varadhan. “Large Deviations and Applications.” SIAM. 1984. 1, 4
[36] Z. Chebb and M. Moahker. “Means of Hermitian positive-definite matricies based on the logdeterminant alpha-divergence function.” Linear Algebra and Applications. 2012. 2
[22] Y. Ollivier. “Ricci Curvature of Markov Chains on
Metric spaces.” Journal of Functional Analysis. 2009.
1, 4
[23] Y. Ollivier. “Ricci Curvature of Metric Spaces.” C. R.
Math. Acad. Sci. Paris. 2007. 1
[24] R. Sandhu, T. Georgiou, E. Reznik, L. Zhu, I. Kolesov,
Y. Senbabaoglu, and A. Tannenbaum. “Graph Curvature for Differentiating Cancer Networks.” Nature
(Scientific Reports). 2015. 1, 2, 3, 6
[25] R. Sandhu, T. Georgiou, and A. Tannenbaum.
“Ricci Curvature: An Economic Indicator for Market Fragility and Systemic Risk.” Science (Science Advances). 2016. 1, 2, 3
[26] B. Anderson, D. Quist, J. Neil, C. Storlie, and T.
Laine. “Graph-Based Malware Detection Using Dynamic Analysis.” Journal in Computer Virology. 2011.
1
[27] C. Wang, E. Jonckheere, and R. Banirazi. “Wireless
network capacity versus Ollivier-Ricci curvature under Heat Diffusion (HD) protocol.” IEEE American
Control Conference. 2014. 1
[28] P. Ranjan and E.H. Abed. “Nonlinear dynamics of
TCP with RED control.” Workshop on Large-Scale
Communication Networks: Topology, Routing, Traffic,
and Control. Institute for Pure and Applied Mathematics (IPAM). 2002. 1
[29] R. Sandhu, S. Lamhamedi-Cherradi, S. Tannenbaum,
J. Ludwig, and A. Tannenbaum. “An Analytical Approach for Insulin-like Growth Factor Receptor 1 and
Mammalian Target of Rapamycin Blockades in Ewing
Sarcoma,” arXiv. 2016. 1
[30] S. Wagner and N. Neshat. “Assessing the Vulnerability
of Supply Chains Using Graph Theory.” International
Journal of Production Economics. 2010. 1
[34] T. Georgiou. “An Instrinisc metric for Power Spectral
Density Functions.” IEEE Signal Processing Letters.
2007. 2
[37] F. Hai and D. Petz. “Riemannian geometry on positive
definite matricies related to means.” Linear Algebra
and Applications. 2009. 2
[38] S. Bonnabel and R. Sepulchre. “Geometric distance
and mean for positive semi-definite matrices of fixed
rank.” SIAM Journal of Matrix Analysis and Applications. 2009. 2
[39] H. Karcher. “Riemannian center of mass and mollifer smoothing.” Communications on Pure and Applied Mathematics.1977. 2
[40] P. Jupp and K. Mardia. “A unfied view of the theory of
directional statistics.” International Statistical Review.
1989. 2
[41] M. Moakher. “Means and averaging in the group of
rotations.” SIAM Journal of Matrix Analysis and Applications. 2002. 2
[42] A. Edelman, T. Arias, and S Smith. “The Geometry
of Algorithms with Orthonganlity Constraints.” SIAM
Journal of Matrix Analysis and Applications. 2002. 2
[43] L. Billera, S. Holmes, K. Vogtmann. “Geometry of
the space of phylogenetic trees.” Advances in Applied
Mathematics. 2001. 2
[44] C. Whidden and F. Matsen. “Ricci-Ollivier curvature of the rooted phylogenetic subtree-prune-regraft
graph.” Theoretical Computer Science. 2017. 2
[45] S. Zairis, H. Khiabanian, A. Blumberg, R. Rabadan.
“Moduli spaces of phylogenetic trees describing tumor evolutionary patterns.” Lectures Notes in Computer Science. 2014. 2
[46] N. Taleb. “The Black Swan: The Impact of the Highly
Improbable.” Random House. 2007. 2
[31] A. Haldane and R. May. “Systemic risk in banking
ecosystems.” Nature 2011. 1
[47] S. Haker, L. Zhu, and A. Tannenbaum. “Optimal
Mass Transport for Registration and Warping,” International Journal of Computer Vision. 2004. 3
[32] S. Battiston, G. Caldarelli, C. Georg, R. May, and J.
Stiglitz. “Complex derivatives.” Nature Physics. 2013.
1
[48] J. Benamou and Y. Brenier. “A Computational Fluid
Mechanics Solution to the Monge-Kantrovich Mass
Transfer Problem.” Numerische Mathematik. 2000. 3
[49] W. Li, S. Osher, and W. Gangbo. “A Fast Algorithm
for Earth Mover?s Distance based on Optimal Transport and L1 Type Regularization.” arXiv. 2016. 3
[50] Y. Chen, T. Georgiou, and A. Tannenbaum. “Matrix
Optimal Mass Transport: A Quantum Mechanical Approach.” arXiv. 2016. 3, 4
[51] R. Jordan, D. Kinderlehrer, and F. Otto. “The Variational Formulation of the Fokker-Planck Equation.”
SIAM Journal on Mathematical Analysis. 1998. 3
[52] J. Lott and C. Villani. “Ricci Curvature for MetricMeasure Spaces via Optimal Transport.” Annals of
Mathematics. 2009.
[53] K. Sturm. “On the geometry of metric measure
spaces.” Acta Math. 2006. 3
[54] B. Chow, P. Lu, L. Ni. “Hamilton?s Ricci Flow.”
American Mathematical Society. 2006. 3
[55] Y. Rubner, C. Tomasi, and L. Guibas. “The Earth
Mover’s Distance as a Metric for Image Retrieval.” International Journal on Computer Vision. 2000. 4
[56] M. DoCarmo. Riemannian Geometry. Birkhauser.
1992. 4
[57] K. Nomizu. “Invariant affine connections on homogeneous spaces.” American Journal of Math. 1954. 4
[58] P. Fletcher and S. Joshi. “Principal geodesic analysis
on symmetric spaces: Statistics of diffusion tensors.”
ECCV. 2004 4
[59] X. Pennec, P. Fillard, and N. Ayache. “A Riemannian
Framework for Tensor Computing.” IJCV. 2005. 4, 5,
7
[60] V. Arsingy, P. Fillard, X. Pennec, and N. Ayache.
“Log-Euclidean metrics for fast and simple calculus on diffusion tensors.” Magnetic Resonance in
Medicine. 2006. 4
[61] Y. Rathi, A. Tannenbaum, O. Michailovich. “Segmenting Images on the Tensor Manifold.” CVPR. 2007
4, 8
[62] S. Helgason. Differential Geometry, Lie Groups, and
Symmetric Spaces. Academic Press. 1978 4
[63] S. Kobayashi and K. Nomizu. Foundations of Differential Geometry. Interscience Tracts in Pure and Applied Mathematics. 1969 4
[64] F. Bauer, J. Jost, and S. Liu. “Ollivier-Ricci Curvature
and the Spectrum of the Normalized Graph Laplace
Operator.” arXiv. 2013. 6
[65] G Keller. “Embryonic stem cell differentiation: emergence of a new era in biology and medicine.” Genes
Dev 19, 1129?1155 2005. 7
[66] C. Banerji, D. Saavedra, S. Severini, M. Widshwendter, T. Enver, J. Zhou, and A. Teschendorff.
“Cellular network entropy as the energy potential in
Waddington differentiation landscape.” Nature (Scientific Reports). 2013. 7
| 3cs.SY
|
Robustly Learning a Gaussian: Getting Optimal Error, Efficiently
Ilias Diakonikolas∗
CS, USC
[email protected]
arXiv:1704.03866v2 [cs.DS] 5 Nov 2017
Jerry Li §
EECS & CSAIL, MIT
[email protected]
Gautam Kamath†
EECS & CSAIL, MIT
[email protected]
Ankur Moitra¶
Math & CSAIL, MIT
[email protected]
Daniel M. Kane‡
CSE & Math, UCSD
[email protected]
Alistair Stewartk
CS, USC
[email protected]
November 7, 2017
Abstract
We study the fundamental problem of learning the parameters of a high-dimensional
Gaussian in the presence of noise — where an ε-fraction of our samples were chosen by
an adversary. We give robust estimators that achieve estimation error O(ε) in the total
variation distance, which is optimal up to a universal constant that is independent of
the dimension.
In the case√where just the mean is unknown, our robustness guarantee is optimal up
to a factor of 2 and the running time is polynomial in d and 1/ε. When both the mean
and covariance are unknown, the running time is polynomial in d and quasipolynomial
in 1/ε. Moreover all of our algorithms require only a polynomial number of samples.
Our work shows that the same sorts of error guarantees that were established over fifty
years ago in the one-dimensional setting can also be achieved by efficient algorithms in
high-dimensional settings.
∗
Supported by NSF CAREER Award CCF-1652862, a Sloan Research Fellowship, and a Google Faculty Research
Award.
†
Supported by NSF CCF-1551875, CCF-1617730, CCF-1650733, and ONR N00014-12-1-0999.
‡
Supported by NSF CAREER Award CCF-1553288 and a Sloan Research Fellowship.
§
Supported by NSF CAREER Award CCF-1453261, CCF-1565235, a Google Faculty Research Award, and an
NSF Graduate Research Fellowship.
¶
Supported by NSF CAREER Award CCF-1453261, CCF-1565235, a Packard Fellowship, a Sloan Research Fellowship, a grant from the MIT NEC Corporation, and a Google Faculty Research Award.
k
Supported by a USC startup grant.
1
Introduction
1.1
Background
The most popular and widely used modeling assumption is that data is approximately Gaussian. This is a convenient simplification to make when modeling velocities of particles in an ideal
gas [Goo15], measuring physical characteristics across a population (after controlling for gender),
and even modeling fluctuations in a stock price on a logarithmic scale. However, real data is not
actually Gaussian and is at best crudely approximated by a Gaussian (e.g., with heavier tails).
What’s worse is that estimators designed under this assumption can perform poorly in practice
and be heavily biased by just a few errant samples that do not fit the model.
For over fifty years, the field of robust statistics [HR09, HRRS86, RL05] has studied exactly
this phenomenon — the sensitivity or insensitivity of estimators to small deviations in the model.
Unsurprisingly, one of the central questions that shaped its development was the problem of learning
the parameters of a one-dimensional Gaussian distribution when a small fraction of the samples
are arbitrarily corrupted. More precisely, in 1964, Huber [Hub64] introduced the following model:
Definition 1. In Huber’s contamination model, we are given samples from a distribution
D = (1 − ε)N (µ, σ 2 ) + εZ ,
where N (µ, σ 2 ) is a Gaussian of mean µ and variance σ 2 , and Z is an arbitrary distribution chosen
by an adversary.
Intuitively, among our samples, about a (1−ε) fraction will have been generated from a Gaussian
and are called inliers, and the rest are called outliers or gross corruptions. We will work with an
even more challenging1 model — called the strong contamination model (Definition 2) — where the
adversary is allowed to look at the inliers and then decide on the outliers. The literature on robust
statistics has given numerous explanations and empirical investigations [GCSR14, Ham01] into how
such outliers might arise as the result of equipment failure, data being entered incorrectly, or even
from a subpopulation that was not accounted for in a medical study. These types of errors are
erratic and difficult to model, so instead our goal is to design a procedure that accurately estimates
µ and σ 2 without making any assumptions about them.
In one dimension, the median and median absolute deviation are well-known robust estimators
for the mean and variance respectively. In particular, given samples X1 , X2 , . . . , Xn , we can compute
µ
b = median(X1 , X2 , . . . , Xn ) and σ
b=
median(|Xi − µ
b|)
,
−1
Φ (3/4)
where Φ is the cumulative distribution of the standard Gaussian. (This scaling constant is needed
to ensure that σ
b is an unbiased estimator when there is no noise.) If n ≥ C logε21/δ , then with
probability at least 1 − δ we have that dT V (N (µ, σ 2 ), N (b
µ, σ
b2 )) ≤ Cε. In Huber’s contamination
model, this is the strongest type of error guarantee we could hope for2 and captures both the task
of learning the underlying parameters µ and σ 2 , and finding the approximately best fit to the
observed distribution within the family of one-dimensional Gaussians. In fact there are plentifully
many other estimators — such as the trimmed mean, winsorized mean, Tukey’s biweight function,
1
None of the results in our paper were previously known in Huber’s contamination model either. The reason we
work with this stronger model is because we can — nothing in our analysis relies on the inliers and outliers being
independent.
2
See Lemma 17.
1
and the interquartile range — that achieve the same sorts of error guarantees, up to constant factors.
The design of robust estimators for location (e.g., estimating µ) and scale (e.g., estimating σ 2 ) is
guided by certain overarching principles, such as the notion of the influence curve [HRRS86] or the
notion of breakdown point [RL05]. In some cases, it is even possible to design robust estimators
that are minimax optimal [Hub64].
These days, much of modern data analysis revolves around high-dimensional data — for example, when we model documents [BNJ03], images [OF96], and genomes [NJB+ 08] as vectors in a very
high-dimensional space. The need for robust estimators is even more pressing in these applications,
since it is infeasible to remove obvious outliers by inspection. However, adapting robust statistics
to high-dimensional settings is fraught with challenges. The principles that guided the design of
robust estimators in one dimension seem to inherently lead to high-dimensional estimators that are
hard to compute [Ber06, HM13].
In this paper, we focus on the central problem of learning the parameters of a multivariate
Gaussian N (µ, Σ) in the strong contamination model. The textbook estimators for the mean and
covariance – such as the Tukey median [Tuk75] and minimum volume enclosing ellipsoid [Rou85]
– essentially search for directions where the projection of D is suitably non-Gaussian. However,
trying to find a direction where the projection is non-Gaussian can be like looking for a needle
in an exponentially-large haystack – these statistics are not efficiently computable, in general.
Furthermore, a random projection will look Gaussian with high probability [Kla07].
In this paper, our main result is an efficiently computable estimator for a high-dimensional
Gaussian that achieves error
b ≤ Cε
dT V (N (µ, Σ), N (b
µ, Σ))
in the strong contamination model, for a universal constant C that is independent of the dimension.
For a Gaussian distribution, we consider estimation in terms of total variation distance, which is
equivalent to estimating the parameters under the natural measures. Our main idea is to use various
regularity conditions satisfied by the inliers to make the problem of searching for non-Gaussian
projections easier. When just the mean µ is unknown, our algorithm runs in time polynomial in
the dimension d and 1/ε. When both the mean and covariance are unknown, our algorithm runs
in time polynomial in d and quasi-polynomial in 1/ε. All of our algorithms achieve polynomial
sample complexity.
Prior to our work, the best known algorithm of Diakonikolas et al. [DKK+ 16] achieved estimation error O(ε log 1/ε) for this problem3 , again with respect to total variation distance. Concurrently, Lai, Rao and Vempala [LRV16] gave an algorithm which achieves estimation error roughly
O(ε1/2 log1/2 d). In fact, the algorithm of Diakonikolas et al. [DKK+ 16] works in a stronger model
than what we consider here, where an adversary gets to look at the samples and then decides on
an ε-fraction to move arbitrarily. Such errors are both additive and subtractive (because inliers
are removed). Interestingly, Diakonikolas, Kane and Stewart [DKS17] proved that any Statistical
Query learning algorithm that works in such an additive and subtractive model and achieves an
error guarantee asymptotically better than O(ε log1/2 1/ε) must make a super-polynomial number
of statistical queries. Our work shows a natural conclusion that in an additive only model it is
possible to algorithmically achieve the same error guarantees as are possible in the one-dimensional
case, up to a universal constant.
3
We note that, as stated, the results in [DKK+ 16] give estimation error O(ε log3/2 1/ε). However, combining the
techniques in [DKK+ 16] with the arguments in Section 7 of this paper gives the stated bound. This argument will
be included in the full version of [DKK+ 16].
2
1.2
Our Results and Techniques
In what follows, we will explain both our work as well as prior work through the following lens:
At the core of any robust estimator is some procedure to certify that the estimates have
not been moved too far away from the true parameters by a small number of corruptions.
First, we consider the subproblem where the covariance Σ = I is known and only the mean µ is
unknown. In the terminology of robust statistics, this is called robust estimation of location. If we
could compute the Tukey median, we would have an estimate that satisfies dT V (N (µ, I), N (b
µ, I)) ≤
Cε. The way that the Tukey median guarantees that it is close to the true mean is that along every
direction u it is close to the median of the projection of the samples. More precisely, at least a
1−ε
T
T b, and at least a 1−ε fraction of the samples satisfy
2 fraction of the samples satisfy u Xi ≥ u µ
2
T
T
u µ
b ≥ u Xi . However, if we have a candidate µ
b, finding a direction u that violates this condition
is again like searching for a needle in an exponentially large haystack.
The approach of Diakonikolas et al. [DKK+ 16] was essentially a data-dependent way to search
for appropriate directions u, by looking for directions where the empirical variance is larger than
it should be (if there were no corruptions). However, because their approach considers only a
single direction at a time, it naturally gets stuck at error Θ(ε log1/2 1/ε). This is because along the
direction u, only when a point is Ω(log1/2 1/ε) away from most of the rest of the samples can we be
relatively confident that it is an outlier. Thus, an adversary could safely place all the corruptions in
the tails and move the mean by as much as Θ(ε log1/2 1/ε). This would not affect the Tukey median
by as much, but would affect an estimate based on the empirical mean (because the algorithm could
find no other outliers to remove) by considerably more.
Our approach is to consider logarithmically many directions at once. Even though an inlier
can be logarithmically many standard deviations away from the mean along a single direction u
with reasonable probability, it is unlikely to be that many standard deviations away simultaneously
across many orthogonal directions. Essentially, this allows us to remove the influence of outliers on
all but a logarithmic dimensional subspace. Combining this with an algorithm for robustly learning
the mean in time exponential in the dimension (but polynomial in the number of samples), we obtain
our first main result:
Theorem 1. Suppose we are given a set of n = poly(d, 1/ε) samples from the strong contamination
model, where the underlying d-dimensional Gaussian is N (µ, I). Let ε ≤ ε0 , where ε0 is a positive
universal constant. For any β > 0, there is an algorithm to learn an estimate N (b
µ, I) that with
high probability satisfies
1
1
2
ε.
dT V (N (µ, I), N (b
µ, I)) ≤ √ + O √ + ε
β
2
Moreover, the algorithm runs in time poly(n, (1/ε)β ).
We prove an almost matching lower bound of 2ε + Ω(ε2 ) on the estimation error. Thus, our
√
robustness guarantee is optimal up to a factor√of 2, even among computationally inefficient robust
estimators. Interestingly, our extra factor of 2 comes from the following geometric fact which we
make crucial
√ use of: Any convex body of diameter D in any dimension can be covered by a ball of
radius D/ 2, and moreover such a ball can be (approximately) found in time exponential in the
dimension. Suppose that along some direction u we have an estimate p that is guaranteed to be
within ε/2 of the projection of the true mean µ. We can now confine µ to a slab of width ε, and by
taking the intersection of all such slabs we get a convex body that contains µ and has diameter of
3
√
at most ε. By covering
√ the body with a ball of radius ε/ 2, we are guaranteed that the center of
the ball is within ε/ 2 of the true mean. This gives us a general way to combine one-dimensional
robust estimates along a net of directions.
We note that, for general isotropic sub-Gaussian distributions, the bound of O(ε log1/2 1/ε)
of [DKK+ 17] is optimal for robust mean estimation, even in one dimension. See Section A for
a proof of this fact. However, our results can be seen to hold more generally than stated above
– indeed, the same arguments work for a class of symmetric isotropic sub-Gaussian distributions
which are sufficiently smooth near their mean. More precisely, we require that along any univariate
projection, the mean is robustly estimated by the median.
We next consider the subproblem where the mean µ = 0 is known and only the covariance Σ is
unknown. In the terminology of robust statistics, this is called robust estimation of scale. In this
b that satisfies4 kΣ − Σk
b F ≤ Cε. When Σ
b does not satisfy
case, we want to compute an estimate Σ
this condition, it can be shown (in Section 6.2.3) that there is a degree-two polynomial p(X), where
E
[p(X)] = 1 and
X∼N (0,Σ)
E
[p(X)] = 1 + C ′ ε .
b
X∼N (0,Σ)
It turns out that, even given the polynomial p(X), deciding whether or not the above conditions
b we can certainly compute E
approximately hold is challenging. Given p(X) and Σ,
b [p(X)].
X∼N (0,Σ)
But given only contaminated samples from N (0, Σ) and without knowing what Σ is, can we estimate
EX∼N (0,Σ) [p(X)]?
Often, univariate robust estimation problems are considered easy, with a simple recipe: Construct an unbiased estimator for the statistic for which each sample point has low influence. However, in our setting, it is highly non-trivial to construct such an estimator. The naive attempt in
this case would be the median – this immediately fails since the distribution of p(X) is asymmetric.
Even if there were no noise, that would not necessarily be an unbiased estimator. So how can we
dampen the influence of outliers, if there is no natural symmetry in the distribution? We construct
a robust estimator crucially using the fact that p(X) is the weighted sum of chi-squared random
variables when there is no noise. The key structural fact we exploit is the following: Given two
sums of chi-squared random variables, if the random variables are far in total variation distance,
most of their difference must lie close to their means. We use this fact to show how, given a weak
estimate of the mean (i.e., one which is only accurate up ω(ε)), one can improve the estimate by a
constant factor. Our result follows by an iterative application of this technique.
However, there is still a major complication in utilizing our low-dimensional estimator to obtain
a high-dimensional estimator. In the unknown mean case, we knew the higher-order moments (since
we assumed that the covariance is the identity). Here, we do not have control over the higher-order
moments of the unknown Gaussian. Overcoming this difficulty requires several new techniques,
which are quite complicated, and we defer the full details to Section 6. Our second main result is:
Theorem 2. Suppose we are given a set of n = poly(d, 1/ε) samples from the strong contamination
model, where the underlying d-dimensional Gaussian is N (0, Σ). There is an algorithm to learn an
b that runs in time poly(n, (1/ε)O(log 4 1/ε) ) and with high probability satisfies
estimate N (0, Σ)
b ≤ Cε ,
dT V (N (0, Σ), N (0, Σ))
for a universal constant C that is independent of the dimension.
4
More precisely, to obtain O(ε) error guarantee with respect to the total variation distance, we need to robustly
approximate Σ within O(ε) in Mahalanobis distance, which is a stronger metric than the Frobenius norm. As part
of our approach, we are able to efficiently reduce to the case that Σ is close to the identity matrix, in which case the
Frobenius error suffices.
4
A key technical problem arises when we attempt to combine estimates for the covariance restricted to a subspace and its orthogonal complement. We refer to this as a stitching problem,
where if we write Σ as
ΣV AT
,
Σ=
A ΣV ⊥
and have accurate estimates for ΣV and ΣV ⊥ , we still need to accurately estimate A. Our algorithm
utilizes an unexpected connection to the unknown mean case: We show that, under a carefully
chosen projection scheme, we can simulate noisy samples from a Gaussian with identity covariance,
where the mean of this distribution encodes the information needed to recover A. We defer the full
details to Section 6.4.
It turns out that we can solve the general case when both µ and Σ are unknown, by directly
reducing to the previous subproblems, exactly as was done in [DKK+ 16] (with some caveats, addressed in Section 4.4). Since all of our error guarantees are optimal up to constant factors, there
is only a constant factor loss in this reduction. Finally, we obtain the following corollary:
Corollary 1. Suppose we are given a set of n = poly(d, 1/ε) samples from the strong contamination
model, where the underlying d-dimensional Gaussian is N (µ, Σ). There is an algorithm to learn an
b that runs in time poly(n, (1/ε)O(log 4 1/ε) ) and with high probability satisfies
estimate N (b
µ, Σ)
b ≤ Cε ,
dT V (N (µ, Σ), N (b
µ, Σ))
for a universal constant C that is independent of the dimension.
This essentially settles the complexity of robustly learning a high-dimensional Gaussian. The
sample complexity of our algorithm depends polynomially on d and 1/ε, and the running time
depends polynomially on d and quasi-polynomially on 1/ε. Up to a constant factor, ours is the first
high-dimensional algorithm that achieves the same error guarantees as in the one-dimensional case,
where results were known for more than fifty years! It is an interesting open problem to reduce the
running time to polynomial in 1/ε (while still being polynomial in d). As we explain in Section 6.6,
this seems to require fundamentally new ideas.
More Related Work
In addition to the works mentioned above, there has been an exciting flurry of recent work on
robust high-dimensional estimation. This includes studying graphical models in the presence of
noise [DKS16], tolerating much more noise by allowing the algorithm to output a list of candidate hypotheses [CSV17], formulating general conditions under which robust estimation is possible [SCV18], developing robust algorithms under sparsity assumptions [Li17, DBS17, BDLS17]
where the number of samples is sublinear in the dimension, and leveraging theoretical insights to
give practical algorithms that can be applied to genomic data [DKK+ 17]. We note that, in comparison to all these other works, ours is the only to efficiently achieve the information-theoretically
optimal error guarantee (up to constant factors). Despite all of this rapid progress, there are still
many interesting theoretical and practical questions left to explore.
1.3
Organization
In Section 2, we go over preliminaries and notation that we will use throughout the paper. In Section 3, we describe an algorithm for robustly estimating the mean of a Gaussian in low-dimensional
settings, and crucially apply it in the design of an algorithm for mean-estimation in high dimensions, described in Section 4. Similarly, in Section 5, we give an algorithm for robustly estimating
5
the mean of degree-two polynomials in certain settings, which is applied in the context of our
covariance-estimation algorithm in Section 6. Finally, we put these tools together and describe our
general algorithm for robustly estimating a Gaussian in Section 7.
2
Preliminaries
In this section, we give various definitions and lemmata we will require throughout the paper. First,
given a distribution F , we let EF [f (X)] = EX∼F [f (X)] denote the expectation of f (X) under F .
If S is a finite set, we let ES [f (X)] = EX∼unif(S) [f (X)] denote the expectation of f (X) under the
uniform distribution over points in S (i.e., the empirical mean of f under S). Given any subspace
V ⊆ Rd , we let ΠV : Rd → Rd be the projection operator onto V . If V = span(v) is 1-dimensional,
we will denote this projection as Πv .
2.1
The Strong Contamination Model
Here we formally define the strong contamination model.
Definition 2. Fix ε > 0. We say a set of samples X1 , . . . , Xn was generated from the strong
contamination model on a distribution F , if it was generated via the following process:
1. We produce (1 − ε)n i.i.d. samples G from F .
2. An adversary is allowed to observe these samples and add εn points E arbitrarily.
We are then given the set of samples G ∪ E in random order. Also, we will say that the samples
X1 , . . . , Xn are ε-corrupted. Moreover given an ε-corrupted set of samples S, we will write S =
(G, E) where G is the set of uncorrupted points and E is the set of corrupted points. Moreover,
given a subset S ′ ⊂ S, we will also write S ′ = (G′ , E ′ ), where G′ = S ′ ∩ G and E ′ = S ′ ∩ E denote
the set of uncorrupted points and corrupted points remaining in S ′ . L will denote G \ G′ , which is
the set of “lost” uncorrupted points.
Given a contaminated set S ′ = (G′ , E ′ ) and a set G so that G′ ⊆ G, define the following
quantities
φ(S ′ , G) =
|G \ G′ |
,
|S ′ |
ψ(S ′ , G) =
|E ′ |
,
|S ′ |
∆(S ′ , G) = ψ(S ′ , G) + φ(S ′ , G) log
1
.
φ(S ′ , G)
(1)
In particular, observe that if ∆(S ′ , G) < O(ε), then a simple calculation implies that φ(S ′ , G) ≤
O(ε/ log 1/ε). Equivalently, we have removed at most an O(ε/ log 1/ε) fraction of good points from
G. This is crucial, as if we throw out an ε-fraction of good points then we essentially put ourselves
in the subtractive model, and there our guarantees no longer hold.
There are two differences between the strong contamination model and Huber’s contamination
model. First, the number of corrupted points is fixed to be εn instead of being a random variable.
However, this difference is negligible. It follows from basic Chernoff bounds that n samples from
Huber’s contamination model with parameter ε (for n sufficiently large) can be simulated by a
(1 + o(1))ε-corrupted set of samples, except with negligible failure probability. Hence, we lose
only an additive o(ε) term when translating from Huber’s contamination model to the strong
contamination model, which will not change any of the guarantees in our paper. The second
difference is that the adversary is allowed to inspect the uncorrupted points before deciding on the
corrupted points. This makes the model genuinely stronger since the samples we are given are no
longer completely independent of each other.
6
2.2
Deterministic Regularity Conditions
In analyzing our algorithms, we only need certain deterministic regularity conditions to hold on the
uncorrupted points. In this subsection, we formally state what these conditions are. It follows from
known concentration bounds that these conditions all hold with high probability given a polynomial
number of samples. Now with these regularity conditions defined once and for all, we will be able to
streamline our proofs in the sense that each step in the analysis will only ever use one of these fixed
set of conditions and will not use the randomness in the sampling procedure. We remark that some
subroutines in our algorithm only need a subset of these conditions to hold, so we could improve the
sample complexity by changing the regularity conditions we need at each step. However, since we
will not be concerned with optimizing the sample complexity beyond showing that it is polynomial,
we choose not to complicate our proofs in this manner.
2.2.1
Regularity Conditions for Unknown Mean
In the unknown mean case, we will require the following condition:
Definition 3. Let G be a multiset of points in Rd and η, δ > 0. We say that G is (η, δ)-good with
respect to N (µ, I) if the following hold:
p
• For all x ∈ G we have kx − µk2 ≤ O( d log(|G|/δ)).
• For every affine function L : Rd → R we have | PrG (L(X) ≥ 0) − PrN (µ,I) (L(X) ≥ 0)| ≤
η/(d log(d/ηδ)) .
• We have that k EG [X] − EN (µ,I) [X]k2 ≤ η.
• We have that kCovG [X] − Ik2 ≤ η/d.
• For any even degree-2 polynomial p : Rd → R we have that
E[p(X)] −
G
E[p2 (X)] −
G
E [p(X)] ≤ η
N (µ,I)
E [p2 (X)] ≤ η
N (µ,I)
E [p2 (X)]1/2 ,
N (µ,I)
E [p2 (X)], and
N (µ,I)
Pr[p(X) ≥ 0] ≤ Pr [p(X) ≥ 0] +
G
N (µ,I)
η
.
d log(|G|/δ)
It is easy to show (see Lemma 6) that given enough samples from N (µ, I), the empirical data set
will satisfy these conditions with high probability.
2.2.2
Regularity Conditions for Unknown Covariance
In the unknown covariance case, we will require the following condition:
Definition 4. Let G be a set of n points of Rd , and η, δ > 0. We say that G is (η, δ)-good with
respect to N (0, Σ) if the following hold:
• For all x ∈ G we have that xT Σ−1 x = O(d log(|G|/δ)).
7
• For any even degree-2 polynomial p : Rd → R we have
E[p(X)] −
G
E[p2 (X)] −
G
E [p(X)] ≤ η
N (0,Σ)
E [p2 (X)] ≤ η
N (0,Σ)
E [p2 (X)]1/2 ,
N (0,Σ)
E [p2 (X)] and
N (0,Σ)
Pr [p(X) ≥ 0] ≤ Pr [p(X) ≥ 0] +
X∼G
N (0,Σ)
η2
.
d log(|G|/δ)
• For any even degree-4 polynomial p : Rd → R we have
E[p(X)] −
G
E [p(X)] ≤ η Var [p(X)]1/2 ,
N (0,Σ)
N (0,Σ)
Pr[p(X) ≥ 0] ≤ Pr [p(X) ≥ 0] +
G
N (µ,I)
η2
.
2 log(1/ε)(d log(|G|/δ))2
As before, it is easy to show (see Lemma 14) that given enough samples from N (0, Σ), the empirical
data set will satisfy these conditions with high probability.
2.3
Bounds on the Total Variation Distance
We will require some simple bounds on the total variation distance between two Gaussians. These
bounds are well-known. Roughly speaking, they say that the total variation distance between
two Gaussians with identity covariance is governed by the ℓ2 norm between their means, and the
total variation distance between two Gaussians with mean zero is governed by the Frobenius norm
between their covariance matrices, provided that the matrices are close to the identity.
Lemma 1. Let µ1 , µ2 ∈ Rd be such that kµ1 − µ2 k2 = ε for ε < 1. Then
1
dTV (N (µ1 , I), N (µ2 , I)) = √ + o(1) ε .
2π
For clarity of exposition we defer this calculation to the Appendix.
We also need to bound the total variation distance between two Gaussians with zero mean and
different covariance matrices. The natural norm to use is the Mahalanobis distance. But in our
setting, we will be able to use the more convenient Frobenius norm instead (because we effectively
reduce to the case that the covariance matrices will be close to the identity):
b be such that kΣ−IkF ≤ O(ε log 1/ε), and kΣ− Σk
b F ≤
Lemma 2 (Cor. 2.14 in [DKK+ 16]). Let Σ, Σ
b ≤ O(ε).
Cε. Then dTV (N (0, Σ), N (0, Σ))
These lemmata show that parameter estimation and approximation in total variation distance
are essentially equivalent. Indeed, in this paper, we achieve both guarantees, but state our results
in terms of total variation estimation.
3
Robustly Learning the Mean in Low Dimensions
This section is dedicated to the proof of the following theorem:
8
Theorem 3. Fix µ ∈ Rd , and let ε, γ, δ > 0. Let S0 = (G0 , E0 ) be such that G0 is a (γε, δ)-good
set with respect to N (µ, I), and |E0 |/|S0 | ≤ ε. Let S = (G, E) be another set with ∆(S, S0 ) < ε.
Let V ⊆ Rd be a subspace. For all 0 < ρ < 1, the algorithm LearnMeanLowD(V, γ, ε, δ, S, ρ)
runs in time poly(d, |S|, (1/ρ)O(dim(V )) , log(ρε/(1 − ρ)), log(1/ρ)) and returns a µ
e so that
kΠV (µ − µ
e)k2 =
γ
1 + 2ρ √
ε.
π+O
1−ρ
d
√
In particular, as we let ρ, γ → 0, the √
parameter estimation error approaches πε (corresponding to
a total variation approximation of ε/ 2). In Lemma 17
p in the Appendix we show√that no algorithm
can achieve parameter estimation error better than π2 ε. Thus, we achieve a 2 approximation
to the optimal error.
For simplicity, in the rest of this section, we will let V = Rd , that is, we assume there is no
projection. It should be clear that this can be done without loss of generality. Our algorithm
proceeds as follows: First, we show that in one dimension, the median produces an estimate which
is optimal, up to lower order terms, if the sample set is (γε, δ)-good with respect to the underlying
d
Gaussian. Then, we show that
p π by using a net argument, we can produce a convex body in R with
diameter at most 2R = 2( 2 + o(1))ε which must contain the true mean. Finally, we use an old
√
result of Jung [Jun01] that such a set can be circumscribed by a ball of radius 2R (see [BW41]
for an English language version of the result). We use the center of the ball as our estimate µ
e.
3.1
Robustness of the Median
First we show that if we projectponto one dimension, then the median of the corrupted data differs
from the true mean by at most π2 ε + o(ε). Our proof will rely only on the notion of a (γε, δ)-good
set with respect to N (µ, I) and thus it works even in the strong contamination model. Formally,
we show:
Lemma 3. Fix any v ∈ Rd . Fix µ ∈ Rd , and let δ > 0. Let S0 = (G0 , E0 ) be so that G0 be a
set
with
(γε, δ)-good set with respect to N (µ, I), and |E0 |/|S0 | ≤ ε. Let S = (G, E) be another
pπ
γ
ε.
∆(S, S0 ) < ε. Let b be the median of S when projected onto v. Then, |b − Πv µ| ≤
2 +O d
Proof. For any a ∈ R, we have
Pr [hv, Xi > a] =
X∼S
Observe that we have
|E|
|S|
|G|
|E|
Pr [hv, Xi > a] +
Pr [hv, Xi > a] .
|S| X∼G
|S| X∼E
PrX∼E [hv, Xi > a] ≤ ψ(S, G). Moreover, by simple calculation we have
Pr [hv, Xi > a] − Pr [hv, Xi > a] ≤ 2φ(S, G) .
X∼G
X∼G0
Hence, we have
Pr [hv, Xi > a] −
X∼S
|G|
Pr [hv, Xi > a] ≤ ψ(S, G) + o(ε) ,
|S| X∼G0
since by assumption ∆(S, G) ≤ ε. Similarly, we have that for all a ∈ R,
Pr [hv, Xi < a] −
X∼S
|G|
Pr [hv, Xi < a] ≤ ψ(S, G) + o(ε) .
|S| X∼G0
9
For |a| = O(ε) we have that PrN (0,I) [X > a] = 12 − √12π a + O(ε3 ). Thus, by (γε, δ)-goodness of G0 ,
this implies that for |a| = O(ε), we have
γε
|G| 1
1
Pr [hv, Xi > Πy µ + a] −
.
(2)
− √ a ≤ ψ(S, G) + O
X∼S
|S| 2
d
2π
Thus, for |a| = O(ε) we have
r
γε
|G| 1
1
|E|
π
+ o(ε)
a] ≤
−√ a +
+O
Pr [hv, Xi > Πy µ +
X∼S
2
|S| 2
|S|
d
2π
γε
1 |G| 1
|E|
√ a+
≤ −
+ o(ε) .
+O
2
|S| 2π
2|S|
d
pπ
p
In particular, we see that if a > π2 ε + O γε
d + o(ε), then PrX∼S
p π hv, Xi > Πy µ + 2 ε < 1/2.
By symmetric logic, we also have that Pr
hv, Xi > Πy µ − 2 ε > 1/2. Thus, the median in
pX∼S
π
direction v differs from Πy µ by at most 2 ε + O γε
d + o(ε).
3.2
Finding a Minimum Radius Circumscribing Ball
For any x ∈ Rd and r > 0, let B(x, r) = {y ∈ Rd : kx − yk2 ≤ r} denote the closed ball of
radius r centered at x. The following classical result of Gale gives a bound on the radius of the
circumscribing ball of any convex set in terms of its diameter:
Theorem 4 (see [Jun01, BW41]). Fix R > 0. Let C ⊆ Rd be a convex √
body so that for all x, y ∈ C,
we have kx − yk2 ≤ 2R. Then C is contained within a ball of radius R 2.
The bound is asymptotically achieved for the standard simplex as we increase its dimension.
The goal of this subsection is to show that the (approximately) minimum radius circumscribing
ball can be found efficiently. We will assume we are given an approximate projection oracle for the
convex body that given a point y ∈ Rd , outputs a point which is almost the closest point in C to x:
Definition 5. A ρ-projection oracle for a convex body C is a function O : Rd → Rd , which, given
a point y ∈ Rd , outputs a point x ∈ C so that kx − yk2 ≤ inf x′ ∈C kx′ − yk2 + ρ.
Our first step is to use such an oracle to construct a net for C. First, we need the following
well-known bound on the size of the net.
Claim 1. Fix r > 0. Then, for any β > 0, there is a β-net F for the sphere of radius r around 0
in Rd of size (r/β)O(d) . Moreover, this net can be constructed in time poly(d, |F|).
With this, we can show:
Lemma 4. Fix R, C as in Theorem 4, and let 1 > ρ > 0. Let x ∈ C be arbitrary. Let O be a
(ρR/3)-projection oracle for C. Suppose a call to O runs in time T . Then, there is an algorithm
CircumscribeNet(R, ρ, O, x) which runs in time poly((R/ρ)O(d) , T ) and outputs a set X ⊆ Rd
so that X is a (ρR)-net for C, and moreover, |X | ≤ (R/ρ)O(d) .
The algorithm is fairly straightforward. First, we observe that C is contained within B(x, 2R).
We then form a (ρR)/3-net of B(x, 2R) using Claim 1. We then iterate over every element v of this
net, and use our projection oracle to (approximately) find the closest point in C to v. If this point
is too far away, we throw it out, otherwise, we add this projected point into the net. The formal
pseudocode for CircumscribeNet is given in Algorithm 1.
10
Algorithm 1 Generating a net of C
1: function CircumscribeNet(R, ρ, O, x)
2:
Form an ρ/3-net F ′ of the sphere of radius 2 of size (1/ρ)O(d) as in Claim 1.
3:
Let F = R · F ′ + x.
4:
Let X ← ∅
5:
for each v ∈ F do
6:
Let uv ← O(v)
7:
if kv − uv k2 ≤ 2ρR/3 then
8:
Add uv to X
9:
end if
10:
end for
11:
return X
12: end function
Proof. The runtime bound follows from Claim 1. We now turn our attention to correctness. By
Claim 1, and rescaling and shifting, the set F is clearly a (ρR)/3-net for a ball B of radius 2R
containing C. We now claim that the set X is indeed a (ρR)/3-net for C. Fix y ∈ C. Since C ⊆ B,
this implies there is some v ∈ F so that ky − vk2 ≤ ρR/3. Thus, in Line 7, when processing v,
we must find some uv ∈ C so that kuv − vk2 ≤ 2ρR/3. The claim then follows from the triangle
inequality.
With this, we obtain:
Corollary 2. Fix R, C, ρ, O, x as in Lemma 4. Suppose a call to O runs in time T . Then, there is
O(d) , T ) and returns a point
an algorithm Circumscribe(R, ρ, O, x) which runs
√ in time poly((R/ρ)
yb so that C is contained within a ball of radius 2(1 + 2ρ)R.
The algorithm at this point is very simple. Using the output of CircumscribeNet, we iterate
over all points √
in a net over B(x, 2R), find an x in this net so that the distance to all points in the
net is at most 2(1 + ρ)R, and output any such point. The formal pseudocode for Circumscribe
is given in Algorithm 2.
Algorithm 2 Finding a circumscribing ball of small radius
1: function Circumscribe(R, ρ, O, x)
2:
Form an ρ/3-net F ′ of B(0, 2) of size (1/ρ)O(d) as in Claim 1.
3:
Let F = R · F ′ + x.
4:
Let X ← CircumscribeNet(R, ρ, O, x).
5:
for each v ∈ F do
√
6:
if for all u ∈ X , we have ku − vk2 ≤ 2(1 + ρ)R then
7:
return u
8:
end if
9:
end for
10: end function
Proof. The√runtime bound is immediate. By Theorem 4, there is some y ∈ B(x, √
2R) so that
′
C ⊆ B(y, R 2). Thus, by the triangle inequality, there is some y ∈ F so that C ⊆ B(y, 2(1+ρ)R).
Thus, the algorithm will output some point y ′′ ∈ F. By an additional
application of the triangle
√
′′
inequality, since X is a ρR-net for C, this implies that C ⊆ B(y , 2(1 + 2ρ)R), as claimed.
11
3.3
The Full Low-Dimensional Algorithm
We now have all the tools to describe the full algorithm in low-dimensions. Let S be our corrupted
dataset as in Theorem 3. Fix ρ > 0. We first produce a ρ-net F over the unit sphere in Rd .
By (a slight modification of) Claim 1, this net has size (1/ρ)O(d) and can be constructed in time
poly(d, |F|). For each v ∈ F, we project all points in S onto v, and take the median of these points
to produce bv . We then construct the following set:
\
C=
{y ∈ Rd : hv, yi ∈ [bv − β, bv + β]} ,
(3)
v∈F
pπ
γε
where β = 2 ε + O d + o(ε) is as in Lemma 3. We now show two properties of this set, which
in conjunction with the machinery above, allows us to prove Theorem 3. The first shows that C
has small diameter:
Claim 2. For all x, y ∈ C, we have kx − yk2 ≤ 2β/(1 − ρ).
Proof. Fix any x, y ∈ C. By definition of C, it follows that for all v ∈ F, we have |hx − y, vi| ≤ 2β.
For any u with kuk2 = 1, there is some v ∈ F with ku − vk2 ≤ ε, and so we have
|hx − y, ui| ≤ |hx − y, vi| + |hx − y, u − vi|
≤ 2β + ρkx − yk2 .
Taking the supremum over all unit vectors u and simplifying yields that kx − yk2 ≤ 2β/(1 − ρ), as
claimed.
The second property shows that we may find an α-projection oracle for C efficiently.
Claim 3. Fix ρ′ > 0. There is a ρ′ -projection oracle ProjOracle(y, ρ′ , C) for C which runs in
time poly((1/ρ)O(d) , log(γε/(1 − ρ)), log(1/ρ′ )).
Proof. The projection problem may be stated as
min kx − yk2 s.t.hv, yi ∈ [bv − β, bv + β], ∀v ∈ F .
This is a convex minimization problem with linear constraints. By the classical theory of optimization [GLS88], finding a ρ-approximate y can be done in poly(d, log(vol(C)/ρ′ )) queries to a
separation oracle for C. Since the separation oracle must only consider the constraints induced by
F, this can be done in time (1/ρ)O(d) . Since by Claim 2 we have vol(C) ≤ (2β/(1 − ρ))O(d) , the
desired runtime follows immediately.
We now finally describe LearnMeanLowD. Using convex optimization, we first find an arbitrary x ∈ C. By Lemma 3 we know µ ∈ C and so this step succeeds. After constructing C,
we run Circumscribe with appropriate parameters, and return the outputted point. The formal
pseudocode for LearnMeanLowD is given in Algorithm 3.
Proof. The runtime claim follows from the runtime claims for Circumscribe and ProjOracle.
Thus, it suffices to prove correctness of this algorithm.
C. By
√ 1+2ρBy Lemma 3, we know that µ√∈ 1+2ρ
Claim 2 and Corollary 2, the output y satisfies B(y, 2 1−ρ β). Thus, we have kµ−yk2 ≤ 2 1−ρ β,
as claimed.
12
Algorithm 3 Finding a circumscribing ball of small radius
1: function LearnMeanLowD(ε, δ, S, ρ)
2:
Form a ρ-net F of B(0, 1) of size (1/ρ)O(d) as in Claim 1.
3:
for each v ∈ F do
4:
Let bv be the median of S projected onto v.
5:
end for
6:
Form C as in Equation (3).
7:
Find an p
x ∈ C using convex
optimization.
+
o(ε)
8:
Let β = π2 ε + O γε
d
9:
Let R = β/(1 − ρ)
10:
Let O(·) = ProjOracle(·, (ρR)/3, C)
11:
return the output of Circumscribe(R, ρ, O, x)
12: end function
4
Robustly Learning the Mean in High Dimensions
In this section, we prove the following theorem, which is our first main result:
Theorem 5. Fix ε, γ, δ > 0, and let X1 , . . . , Xn be an ε-corrupted set of points from N (µ, I), where
kµk2 ≤ O(ε log 1/ε), and where
(d log(d/γεδ))6
.
n=Ω
γ 2 ε2
Then, for every α, β > 0, there is an algorithm RecoverMean(X1 , . . . , Xn , ε, δ, γ, α, β) which
runs in timepoly(d,
1/γ, 1/εβ , 1/α, log 1/δ) and outputs a µ
b so that with probability 1 − δ, we have
√
π+O(γ)
1
kb
µ − µk2 ≤
+ √β ε.
1−α
In particular, observe that Theorem 5, in conjunction with Lemma 1, gives us Theorem 1, if we set
γ = o(1). With this, we may state our primary algorithmic contribution:
Theorem 6. Fix ε, γ, α, δ, β > 0, and let S0 = (G0 , E0 ) be an ε-corrupted set of samples of size
n from N (µ, I), where kµk2 ≤ O(ε log 1/ε), and where n = poly(d, 1/(γε), log 1/δ). Suppose that
G0 is (γε, δ)-good with respect to N (µ, I). Let S ⊆ S0 be a set so that ∆(S, G0 ) ≤ ε. Then, there
exists an algorithm FilterMeanOpt that given S, ε, γ, α, β outputs one of two possible outcomes:
√
π+O(γ)
√1
+
ε.
(i) A µ
b, so that kb
µ − µk2 ≤
1−α
β
(ii) A set S ′ ⊂ S so that ∆(S ′ , G0 ) < ∆(S, G0 ).
Moreover, FilterMeanOpt runs in time poly(d, 1/γ, 1/εβ , 1/α, log 1/δ).
+
By
p first running the algorithm of [DKK 16] to obtain an estimate of the mean to error
O(ε log 1/ε), then running FilterMeanOpt at most polynomially many times, we clearly recover the guarantee in Theorem 5. Thus, the rest of the section is dedicated to the proof of Theorem
6.
At a high level, the structure of the argument is as follows: We first show that if there is a
subspace of eigenvectors of dimension at least O(log 1/ε) of the empirical covariance matrix with
large associated eigenvalues, then we can produce a filter using a degree-2 polynomial (Section 4.1).
Otherwise, we know that there are at most O(log 1/ε) eigenvectors of the empirical covariance
13
with a large eigenvalue. We can learn the mean in this small dimensional subspace using our
learning algorithm from the previous section, and then we can argue that the empirical mean on
the remaining subspace is close to the true mean (Section 4.2).
This outline largely follows the structure of the filter arguments given in [DKK+ 16], however,
the filtering algorithm we use here requires a couple of crucial new ideas. First, to produce the
filter, instead of using a generic degree-2 polynomial over this subspace, we construct an explicit,
structured, degree-2 polynomial which produces such a filter. Crucially, we can exploit the structure
of this polynomial to obtain very tight tail bounds, e.g., via the Hanson-Wright inequality. This is
critical to avoid a quasi-polynomial runtime. If instead we used arbitrary degree-2 polynomials in
this subspace, it would need to be of dimension O(log2 1/ε) and the low-dimensional algorithm in
the second step would take quasi-polynomial time.
Second, we must be careful to throw out far fewer good points than corrupted points. In particular, by our definition of ∆ (which gives an additional logarithmic penalty to discarding good
points) and our guarantee that ∆ decreases, our filter can only afford to throw out an ε/ log(1/ε)
fraction of good points in total, since ∆ is initially ε. This is critical, as if we threw away an
ε-fraction of good points, then proving that the problem remains efficiently solvable becomes problematic. In particular, if these points were thrown away arbitrarily, then this becomes the full
additive and subtractive model, for which a statistical query lower bound prevents us from getting
an O(ε)-approximate answer in polynomial time [DKS17]. To avoid discarding too many good
points, we exploit tight exponential tail bounds of Gaussians, and observe that by slightly increasing the threshold at which we filter away points, we decrease the fraction of good points thrown
away dramatically.
4.1
Making Progress with Many Large Eigenvalues
We now give an algorithm for the case when there are many eigenvalues which are somewhat large.
Formally, we show:
Theorem 7. Fix ε, γ, δ, α, β > 0, and let S0 = (G0 , E0 ) be an ε-corrupted set of samples of size
n from N (µ, I), where kµk2 ≤ O(ε log 1/ε), and where n = poly(d, 1/(γε), log 1/δ). Suppose that
b be the
G0 is (γε, δ)-good with respect to N (µ, I). Let S ⊆ S0 be a set so that ∆(S, G0 ) ≤ ε. Let Σ
sample covariance of S, let µ
b be the sample mean of S, and let V be the subspace of all eigenvectors
b − I with eigenvalue more than 1 ε. Then, there exists an algorithm FilterMeanManyEig
of Σ
β
that given S, ε, γ, δ, α, β outputs one of two possible outcomes:
1. If dim(V ) ≥ C1 β log(1/ε), then it outputs an S ′ so that ∆(S ′ , G0 ) < ∆(S, G0 ).
2. Otherwise, the algorithm outputs “OK”, and outputs an orthonormal basis for V .
b − I, and if there are too
Our algorithm works as follows: It finds all large eigenvalues of Σ
many, produces an explicit degree-2 polynomial which, as we will argue, produces a valid filter.
The formal pseudocode for our algorithm is in Algorithm 4.
For clarity of exposition, we defer the proof of Theorem 7 to Appendix C.
4.2
Returning an Estimate When There are Few Large Eigenvalues
At this point, we have run the filter of Algorithm 4 until there are few large eigenvalues. In the
subspace with large eigenvalues, we again run the low dimensional algorithm to obtain an estimate
for the mean in this subspace. Recall that Lemma 3 guarantees the accuracy of this estimator
within this subspace. In the complement of this subspace, where the empirical covariance is very
14
Algorithm 4 Filter if there are many large eigenvalues of the covariance
1: function FilterMeanManyEig(S, ε, γ, δ, α, β)
2:
Let C1 , C2 , C3 > 0 be sufficiently large constants.
b be the empirical mean and covariance of S, respectively.
3:
Let µ
b and Σ
b − I with eigenvalue more than
4:
Let V be the subspace of Rd spanned by eigenvectors of Σ
1
β ε.
5:
if dim(V ) ≥ C1 β log(1/ε) then
6:
Let V ′ be a subspace of V of dimension C1 β log(1/ε).
√
7:
Let µ̃ be an approximation to ΠV ′ (µ) with ℓ2 -error π+O(γ)
1−α ε, computed using
LearnMeanLowD(V, γ, ε, δ, S, γ).
8:
Let p(x) be the quadratic polynomial p(x) = kΠV ′ (x) − µ̃k22 − dim(V ′ ).
9:
Find a value T > 0 so that either:
(a) T > C2 d log(|S|/δ) and p(x) > T for at least one x ∈ S, or
(b) T > 2C3 log(1/ε)/c0 and PrS (p(x) > T ) > exp(−c0 T /(2C3 )) + γε/(d log(|S|/δ)).
10:
11:
12:
13:
14:
return S ′ = {x ∈ S : p(x) ≤ T }
else
return an orthonormal basis for V .
end if
end function
close to the identity, Lemma 5 (stated below) shows that the empirical mean is close to the true
mean. This leads to a simple algorithm which outputs an estimate for the mean, described in
Algorithm 5.
Algorithm 5 Return a mean if there are few large eigenvalues of the covariance
1: function FilterMeanFewEig(S, ε, γ, δ, α, β, V )
√
π+O(γ)
2:
Let µ̃V be an approximation to ΠV (µ) with ℓ2 -error
1−α ε, computed using
LearnMeanLowD(V, γ, ε, δ, S, γ).
3:
Let µ̃V ⊥ be the empirical mean on V ⊥ , ΠV ⊥ µ̂.
4:
return µ̃V + µ̃V ⊥ .
5: end function
Lemma 5. Let µ, η, G0 , S be as in Theorem 7. Let µ
b be the sample mean of S, and let v be a unit
ε
vector. Suppose that hv, µ − µ
bi > β 1/2
. Then VarS [hv, Xi] > 1 + βε .
For clarity of exposition, we defer the proof of Lemma 5 to Appendix C.
4.3
The Full High-Dimensional Algorithm
We now have almost all the pieces needed to prove the full result. The last ingredient is the
fact that, given enough samples, the good set condition is satisfied by the samples from the true
distribution. Formally,
Lemma 6. Fix η, δ > 0. Let X1 , . . . , Xn be independent samples from N (µ, I), where n =
Ω((d log(d/ηδ))6 /η 2 ). Then, S = {X1 , . . . , Xn } is (η, δ)-good with respect to N (µ, I) with probability at least 1 − δ.
15
Proof. This follows from Lemmas 8.3 and 8.16 of [DKK+ 16].
At this point, we conclude with the proof of Theorem
5. Within the subspace V , Lemma
√
π+O(γ)
3 guarantees that the mean is accurate up to ℓ2 -error 1−α ε. Within the subspace V ⊥ , the
ε
contrapositive of the statement of Lemma 5 guarantees the mean is accurate up to ℓ2 -error β 1/2
.
The desired result follows from the Pythagorean theorem.
4.4
An Extension, with Small Spectral Noise
For learning of arbitrary Gaussians, we will need a simple extension that allows us to learn the mean
even in the presence of some spectral norm error in the covariance matrix. Since the algorithms
and proofs are almost identical to the techniques above, we omit them for conciseness. Formally,
we require:
Theorem 8. Fix χ, ε, δ > 0, and let X1 , . . . , Xn be an ε-corrupted set of points from N (µ, Σ),
where kΣ − Ik2 ≤ O(χ), kµk2 ≤ O(ε log 1/ε), and where n = poly(d, 1/χ, 1/ε, log 1/δ). For
any γ > 0, there is an algorithm RecoverMeanNoisy(X1 , . . . , Xn , ε, δ, γ, χ) which runs in time
poly(d, 1/χ, 1/ε, log 1/δ) and outputs a µ
b so that with probability 1 − δ, we have kb
µ − µk2 ≤
(C + γ)ε + O(χ).
This extension follows from two elementary observations:
1. For the learning in low dimensions, observe that the median is naturally robust to error in the
covariance, and in general, by the same calculation we did, the error of the median becomes
O(ε + α).
2. For the filter, observe that we only need concentration of squares of linear functions, and
whatever error we have in this concentration goes directly into our error guarantee. Thus, by
the same calculations that we had above, if we filtered for eigenvalues above 1 + O(ε + α), we
would immediately get the desired bound.
5
Robustly Estimating the Mean of Degree Two Polynomials
In this section, we give robust estimates of E[p2 (X)] for degree-2 polynomials p in subspaces of
small dimension, which is an important prerequisite to learning the covariance in high-dimensions.
A crucial ingredient in our algorithm is the following improvement theorem (stated and proved in
the next section) which shows how to take any weak high-dimensional estimate for the covariance
and use it to get an even better robust estimate for E[p2 (X)].
5.1
Additional Preliminaries
Here we give some additional preliminaries we require for the low-dimensional learning algorithm
we present here. We will need the following well-known tail bound for degree-2 polynomials:
Lemma 7 (Hanson-Wright Inequality [LM00, Ver10]). Let X ∼ N (0, I) ∈ Rd and A be a d × d
matrix. Then for some absolute constant c0 , for every t ≥ 0,
2
t
t
T
T
,
.
Pr X AX − E[X AX] > t ≤ 2 exp −c0 · min
kAk2F kAk2
We will also require the following lemmata:
16
Lemma 8 (Hölder’s inequality for Schatten norms). Let A, B be matrices. Then, for all p, q so
that 1p + 1q = 1, we have kABkS 1 ≤ kAkS p kBkS q .
This implies the following corollary:
b M be so that kΣ − Σk
b F ≤ O(δ), and so that kM kF = 1. Then, we have
Corollary 3. Let Σ, Σ,
1/2
1/2
1/2
1/2
b
b
kΣ M Σ − Σ M Σ kS 1 ≤ 5δ.
Proof. We have
b 1/2 M Σ
b 1/2 kS 1 ≤ kΣ1/2 M Σ1/2 − M kS 1 + kM − Σ
b 1/2 M Σ
b 1/2 kS 1 .
kΣ1/2 M Σ1/2 − Σ
We will bound the first term on the RHS by 5δ/2; the second term is bounded symmetrically. We
have
kΣ1/2 M Σ1/2 − M kS 1 ≤ kΣ1/2 M Σ1/2 − M Σ1/2 kS 1 + kM Σ1/2 − M kS 1
kΣ1/2 − IkS 2 kM Σ1/2 kS 2 + kΣ1/2 − IkS 2 kM kS 2
≤ 5δ/2 ,
where the last line follows from Hölder’s inequality for Schatten norms.
5.2
An Improvement Theorem
Here we state and prove one of the main technical ingredients in our algorithm for robustly learning
the covariance.
Theorem 9. Fix ε, δ, τ > 0. Let Σ be so that kΣ − IkF ≤ O(ε log 1/ε), and fix a p ∈ P2 , where P2
denotes the set of even degree-2 polynomials in d variables. Let G0 be an (ε, δ)-good set of samples
from N (0, Σ), and let S = {X1 , . . . , Xn } be so that ∆(S, G0 ) ≤ ε. Then, for any C > 0 there is an
algorithm LearnMeanChiSquared which, given p, X1 , . . . , Xn , and ε, outputs a µ
b so that with
probability 1 − τ over the randomness of the algorithm,
µ
b−
E
X∼N (0,Σ)
[p(X)] ≤ kΣ − IkF /C + O(log(C)ε) .
Moreover, the algorithm runs in time O(|S| + log(1/τ )/ε2 ).
The way to think about how this result fits into the overall strategy is that robustly estimating
the covariance is equivalent to robustly estimating the mean of every (normalized) degree-two
polynomial p. The above theorem shows how a weak estimate in high-dimensions can be used
to obtain stronger estimates in one dimension, which ultimately we will use to improve the highdimensional estimate as well. The above theorem is the workhorse in our proof.
Our algorithm itself is simple, however, its correctness is quite non-trivial. We define some
threshold T . Given our corrupted set of samples from N (0, Σ), we use our corrupted data set to
estimate the mean of p(X) conditioned on the event that |p(X)| ≤ T . Then, to estimate the contribution of the mean from points X so that |p(X)| > T , we estimate this by EX∼N (0,I) [p(X)1|p(X)|>T ].
In other words, we are replacing the contribution of the true tail by an estimate of the contribution
of p(X) when X ∼ N (0, I) on this tail. The formal pseudocode is given in Algorithm 6.
Intuitively, this algorithm works because of two reasons. First, it is not hard to show that the
influence of points p(X) within the threshold T on the estimator are bounded by at most T . Hence,
the adversary cannot add corrupted points within this threshold and cause our estimator to deviate
17
too much. Secondly, because we know that kΣ − IkF is small, by carefully utilizing smoothness
properties of sums of chi-squared random variables, we are able to show that our estimate for the
contribution of the tail is not too large. At a high level, this is because “most” of the distance
between two chi-squared random variables must remain close to the means, so the difference in the
tails is much smaller. Proving that this holds in a formal sense is the majority of the technical
work of this section.
Proof. We know the distribution of p(X ′ ) for X ′ ∼ N (0, I) explicitly and wish to use this to get
a better estimate for the mean of p(X) for X ∼ N (0, Σ) than might be given by the mean of the
ε-corrupted set of samples.
Algorithm 6 Approximating E[p(X)] for X ∼ N (0, Σ) with corrupted samples.
1: function LearnMeanChiSquared(X1 , . . . , Xn , p(x), ε, τ )
2:
Let T = O(log
C).
x − T, for x ≥ T
3:
Let f (x) =
0,
for |x| ≤ T .
x + T, for x ≤ −T
P
4:
Compute α = ni=1 (p(Xi ) − f (p(Xi )))/n.
′ from X ′ ∼ N (0, I).
2
5:
Simulate m = O((ln
X1′ , . . . , Xm
Pmτ )/ε ) samples
′
6:
Return µ
b = α + i=1 f (p(Xi ))/n.
7: end function
It follows from (ε, δ)-goodness that | PrX ′ ∼N (0,Σ) [p(X ′ ) > t] − #{Xi : p(Xi ) > t}/n| ≤ 2ε for all
t. We need to express the expectation in terms that we can use this to bound. For Z = p(X ′ ), we
have that
Z T
Z T
Pr[Z < −t]dt .
Pr[Z > t]dt −
E[Z − f (Z)] = E[max{T, min{−T, Z}}] =
0
0
Similarly, the samples have
α=
n
X
i=1
(p(Xi ) − f (p(Xi ))/n =
Z
T
0
#{Xi : p(Xi ) > t}
dt −
n
Z
t
0
#{Xi : p(Xi ) < −t}
dt .
n
Thus, we have | E[Z − f (Z)] − α| ≤ 2T ε.
Since p ∈ P2 , we have E[p(X ′ )] = 1 for X ′ ∼ N (0, I). Thus, we have Var[f (p(X ′ ))] ≤
E[f (p(X ′ ))2 ] ≤ E[p(X ′ )2 ] = 1. It follows by
results that the empirical after
Pstandard concentration
′ )/n − E[p(X ′ )]| ≤ ε with probability 1 − τ .
taking m = O(ln(1 − τ )/ε2 ) samples has | m
f
(p(X
)
i
i=1
When this holds, we have
µ
b−
E
X∼N (0,Σ)
[p(X)] ≤ (2T + 1)ε +
E
[f (p(X))] −
X∼N (0,I)
E
[f (p(X))] .
X∼N (0,Σ)
To prove the correctness of the algorithm it remains to show that:
Lemma 9. For any constant C > 0, for T = O(log C), we obtain
E
[f (p(X))] −
X∼N (0,I)
E
[f (p(X ′ ))] ≤
X ′ ∼N (0,Σ)
18
kΣ − IkF
.
C
Proof. Let M be the symmetric matrix with kM kF = 1 such that p(x) = xT M x for x ∈ Rd . We
can write p(X) with X ∼ N (0, I) as
p(X) + tr(M ) = X T M X = X T O T DOX = Y T DY =
d
X
ai Yi2 ,
i=1
where O is orthogonal, D is diagonal, ai are the eigenvalues
of M , and Y ∼ N (0, I) hence Yi are
P 2
i.i.d. from N (0, 1). Since kM kF = 1, here we have
ai = 1. If instead we express p(X ′ ) in terms
of X ′ ∼ N (0, σ 2 ), we obtain:
′
p(X ) + tr(M ) = Y
′T
Σ
1/2
MΣ
1/2
′
Y =Y
′T
′T
′
′
′
T
′
O DOY =Y DY =
d
X
bi Yi2 ,
i=1
where O′ is orthogonal, D ′ is diagonal, bi are the eigenvalues of Σ1/2 M Σ1/2 , and Y ′ , Y ∼ N (0, I)
hence Yi are i.i.d. from N (0, 1). P
By Corollary 3, we have that i |bi − ai | ≤ (5/2)kΣ − IkF . Now consider the random variables
Zi,λ = −tr(M ) +
where
d
X
j=1
cj (Yi2 − 1) ,
bj ,
cj = (1 − λ)ai + λbi
aj
for j < i
for j = i ,
for j > i
for 1 ≤ i ≤ d and 0 ≤ λ ≤ 1. Note that Zi,1 = Zi+1,0 , for 1 ≤ i ≤ d − 1. Note that, to prove the
lemma, it suffices to show that
X
2|bi − ai |/5C .
| E[f (Z1,0 )] − E[f (Zd,1 )]| ≤
i
To this end,
Pconsider how
P E[f (Zi,λ )] varies with λ. We can write Zi,λ = Z−i + Zi , where Z−i =
−tr(M ) + j<i bj Yj2 + j>i aj Yj2 ; and Zi = ci Yi2 , where ci = ((1 − λ)ai + λbi ). We assume for
now that ci 6= 0. Since only Zi depends on λ, we have
d E[f (Zi,λ )]
d
=
E[f (Z−i + Zi )]
dλ
dλ
d
E [f (Z−i + Zi )]
= E
Z−i dλ Zi
Z ∞
d
= E
f (Z−i + x)Pi (x)dx
Z−i dλ −∞
Z ∞
f (Z−i + x)dPi (x)/dλdx ,
= E
Z−i
−∞
where Pi (x) is the probability density function of the random variable Zi . Standard results about
the χ2 distribution give that:
Fact 1. Let Y1 , Y2 , Y3 ∼ N (0, 1). Then the probability density function of Y12 is
Y12
+
Y22
is
1 −x/2
,
2e
and of
Y12
+
Y22
+
Y32
is
√
√ x e−x/2 .
2π
19
√ 1 e−x/2 ,
2πx
of
This gives that Pi (x) =
√ 1
e−x/2ci .
2πxci
Now consider the derivative:
1
1
/dλ · e−x/2ci + √
· de−x/2ci /dλ
2πxci
2πxci
1
dci
x
1
dci −x/2ci
=− √
· 2 · e−x/2ci ·
·e
+√
·
3/2 dλ
dλ
2πxci 2ci
2 2πxci
√
(bi − ai ) x −x/2ci
bi − a i
e−x/2ci + √
e
=− √
3/2
5/2
2 2πxci
2πci
dPi (x)/dλ = d √
(3)
= ((bi − ai )/2ci )(Pi (x) − Pi (x)) ,
(3)
where Pi (x) is the distribution of Zi + Zi′ + Zi′′ , where Zi′ and Zi′′ are i.i.d. copies of Zi . We thus
have
Z ∞
d E[f (Zi,λ )]
f (Z−i + x)dPi (x)/dλdx
= E
Z−i
dλ
−∞
= ((bi − ai )/2ci ) E [E[f (Z−i + Zi )] − E[f (Z−i + Zi + Zi′ + Zi′′ )]]
Z−i
= ((bi − ai )/2ci ) E[f (Zi,λ ) − f (Zi,λ + Zi′ + Zi′′ )]
= ((bi − ai )/2ci ) E [f (Zi,λ ) − ′E ′′ [f (Zi,λ + Zi′ + Zi′′ )]] .
Zi,λ
Zi ,Zi
Since f has Lipschitz constant 1, |f (Zi,λ ) − EZi′ ,Zi′′ [f (Zi,λ + Zi′ + Zi′′ ]| ≤ E[Zi′ + Zi′′ ] = 2ci whatever
value Zi,λ takes.
Using the probability distribution of Zi′ + Zi′′ , in the case ci > 0, we have for all |z| ≤ T ,
Z ∞
′′
′
e−x/2ci /2ci · (z + x − T )dx
E ′′ [f (z + Zi + Zi )] =
′
Zi ,Zi
T −z
−(T −z)/2ci
=e
(z − T + 2ci + (T − z))
= 2ci e−(T −z)/2ci .
Since ci ≤ max ai , bi ≤ 1 + O(ε log(1/ε)) ≤ 2, for large enough T = O(log C), we have e−(T /2)/2ci ≤
1/C 2 T . We assume that this holds.
For −T ≤ z ≤ T /2, we have 2ci e−(T −z)/2ci ≤ 2ci /C 2 T . A similar argument when ci < 0 gives
that for −T /2 ≤ z ≤ T , we have 2|ci |e−(T −z)/2ci ≤ 2|ci |/C 2 T . Now we have enough to show that
d E[f (Zi,λ )]
≤ |(bi − ai )/2ci || E [f (Zi,λ ) − ′E ′′ [f (Zi,λ ) + Zi′ + Zi′′ )]|
Zi,λ
dλ
Zi ,Zi
≤ Pr [|Zi,λ − tr(M )| ≥ T /2]|(bi − ai )| + |(bi − ai )|/C 2 .
Zi,λ
By the Hanson-Wright inequality, we have that for any x,
√
Pr [|Zi,λ − E[Zi,λ ]| ≥ 2kck2 x + kck∞ x] ≤ 2e−x ,
Zi,λ
where c = (b1 , . . . , bi − 1, ci , ai+1 , . . . , ad ). Note that kck2 ≤ 1 + O(kΣ − IkF ) ≤ 2. Also, E[Zi,λ ] +
P
tr(M ) is the sum of coordinates of c, and so | E[Zi,λ ]| ≤ ij=1 |bj − aj | ≤ O(kΣ − IkF ) ≤ 1. Putting
this together, we obtain
Pr [|Zi,λ − tr(M )| ≥ T /2] ≤ 2 exp(−((T − 1)/8)2 ) ≤ 1/C 2 .
Zi,λ
20
Finally, we have that, assuming ci 6= 0,
d E[f (Zi,λ )]
≤ 2|(bi − ai )|/C 2 .
dλ
Now we need to deal with the special case ci = 0. Note that since f is Lipschitz, for any β ∈ R,
we have | E[f (Z−i + βYi2 ) − f (Z−i )]|/|β| ≤ | E[Yi2 ]| ≤ 1. By considering the limit as β tends to
d E[f (Zi,λ )]
≤ 1.
zero from above or below, we get that when ci = 0, the derivative still exists and
dλ
Since the limit 2|(bi − ai )|/C 2 only does not apply at one point where E[f (Zi,λ )] is still continuous
as a function of λ, this is not an issue. We still obtain that
| E[f (Zi,0 )] − E[f (Zi,1 )]| ≤ 2|(bi − ai )|/C 2 .
Recalling that Zi,1 = Zi+1,0 with Z1,0 = p(X ′ ) and Zd,1 = p(X) for X ′ ∼ N (0, 1) and X ∼ N (0, Σ),
we have by the Mean Value Theorem, that
X
| E[f (p(X ′ ))] − E[f (p(X)]| ≤
| E[f (Zi,0 )] − E[f (Zi,1 )]|
i
≤ 2|(bi − ai )|/C 2 ≤ O(kΣ − IkF /C 2 )
≤ kΣ − IkF /C ,
as required.
This completes the proof of the theorem.
5.3
Working in a Low-Dimensional Space of Degree-Two Polynomials
We now show that via similar techniques as before, we can patch our estimates together to find
a matrix which agrees with the ground truth on all degree-two polynomials in a fixed subspace of
low dimension. Formally, we show:
Theorem 10. Fix ε, τ > 0. Let Σ be so that kΣ−IkF ≤ O(ε log 1/ε). Let G0 be an (ε, δ)-good set of
samples from N (0, Σ), and let S = {X1 , . . . , Xn } be so that ∆(S, G0 ) ≤ ε. Let W1 be a subspace of
degree-2 polynomials, and let W2 be an orthogonal subspace of degree-2 polynomials, so that we have
b so that EX∼N (0,Σ) [p(X)] − E
aΣ
b [p(X)] ≤ ξ for all p ∈ W2 . Then there is an algorithm
X∼N (0,Σ)
b runs in time poly(d, |S|, 2O(dim(W1 )) , log 1/τ ),
LearnMeanPolyLowD which given ε, S, W1 , W2 , Σ
and returns a Σ′ so that
E
N (0,Σ)
[p(X)] −
E [p(X)] ≤ 4 (kΣ − IkF /C + O(log(C)ε) + ξ ,
N (0,Σ′ )
for all p ∈ span(W1 ∪ W2 ) ∩ P2 , with probability 1 − τ .
In particular, this implies:
Corollary 4. Fix ε, τ > 0. Let Σ be so that kΣ − IkF ≤ O(ε log 1/ε). Let G0 be an (ε, δ)good set of samples from N (0, Σ). Let S = {X1 , . . . , Xn } be so that ∆(S, G0 ) ≤ ε. Let V be a
subspace of Rd . Then there is an algorithm LearnCovLowDim which given S, ε, ξ, τ, V runs in
2
time poly(|S|, 2O(dim(V ) ) , log 1/τ ) and returns a Σ′ so that
kΠV Σ − Σ′ ΠV k2 ≤ 4 (kΣ − IkF /C + O(log(C)ε) ,
with probability 1 − τ .
21
Proof. Observe that the dimension of the space of degree-2 polynomials W in V is O(dim(V )2 ).
Run the algorithm in Theorem 10 with the same parameters as before, with W1 = W and W2 = ∅
(so that we may take ξ = 0), and then the guarantee of that algorithm, along with Lemma 10,
gives our desired guarantee.
We now describe the algorithm for Theorem 10. Essentially, we do the same thing as we did
for low-dimensional learning in the unknown mean case: we take a constant net over V ∩ P2 , l earn
the mean over every polynomial in the net, and then find a Σ′ which is close in each direction to
the learned mean. Since we will not attempt to optimize the constant factor here, will will use a
naive LP-based approach to find a point which is close to optimal. The formal pseudocode is given
in Algorithm 7.
Algorithm 7 Filter if there are many large eigenvalues of the covariance
1: function LowDimCovLearning(S, ε, ξ, τ, W1 , W2 )
2:
Generate a 1/2-cover C for W1 ∩ P2 .
3:
Let τ ′ = 2−|C| τ
4:
for p ∈ C do
5:
Compute mp = LearnMeanChiSquared(S, p, ε, τ ′ ).
6:
Generate a linear constraint cp (Σ′ ): EN (0,Σ′ ) [p(X)] − mp ≤ kΣ − IkF /C + O(log(C))ε.
7:
end for
8:
Generate the convex constraint that EN (0,Σ′ ) [p(X)] − EN (0,Σ′ ) [p(X)] ≤ ξ for all p ∈ W2 .
9:
Using a convex program, return any matrix Σ′ which obeys cp (Σ′ ) for all p ∈ C.
10: end function
Observe that every constraint for each polynomial in W1 is indeed linear in Σ′ , by Lemma 10.
Moreover, the constraint for W2 has an explicit separation oracle, since it induces a norm, and
for any p ∈ W2 , we may explicitly compute EN (0,Σ′ ) [p(X)] − EN (0,Σ′ ) [p(X)]. Thus, we may use
separating hyperplane techniques to solve this convex program in the claimed running time.
Proof of Theorem 10. Let us condition on the event that LearnMeanChSquared succeeds for
each p ∈ C. By a union bound, this occurs with probability at least 1 − τ . Thus, in each p ∈ C, we
have that |mp − EX∼N (0,Σ) [p(X)]| ≤ β, where β = kΣ − IkF /C + O(log(C))ε. Let Σ′ be the matrix
we find. By the triangle inequality, we then have that for every p ∈ C, that | EN (0,Σ′ ) [p(X)] −
EN (0,Σ) [p(X)]| ≤ 2β. Hence, by the usual net arguments, we know that for every p ∈ V ∩ P2 ,
|
E
[p(X)] −
N (0,Σ′ )
E [p(X)]| ≤ 4β .
N (0,Σ)
Moreover, by triangle inequality, for every p ∈ W2 , we have EN (0,Σ′ ) [p(X)] − EN (0,Σ′ ) [p(X)] ≤ 2ξ.
The result then follows from the Pythagorean theorem.
6
Robustly Learning the Covariance in High-Dimensions
In this section, we show how to robustly estimate the covariance of a mean-zero Gaussian in highdimensions up to error O(ε). We use our low-dimensional learning algorithm from the previous
section as a crucial subroutine in what follows.
Our main algorithmic contribution is as follows:
Theorem 11. Fix ε, δ > 0, and let S0 = (G0 , E0 ) be an ε-corrupted set of samples of size n from
N (0, Σ), where kΣ − IkF ≤ ξ where ξ = O(ε log 1/ε), and where n = poly(d, 1/ε, log 1/δ). Suppose
22
that G0 is (ε, δ)-good with respect to N (0, Σ). Let S ⊆ S0 be a set so that ∆(S, G0 ) ≤ ε. Then, there
exists an algorithm ImproveCov that given S, ξ, ε, fails with probability at most poly(ε, 1/d, δ),
and otherwise outputs one of two possible outcomes:
b so that kΣ
b − ΣkF ≤ kΣ − IkF /2.
(i) A matrix Σ,
(ii) A set S ′ ⊂ S so that ∆(S ′ , G0 ) < ∆(S, G0 ).
Moreover, ImproveCov runs in time poly(d, (1/ε)O(log
4
1/ε) , log 1/δ).
By first applying the algorithm in [DKK+ 16] to produce an initial estimate for Σ, and then
iterating the above algorithm polynomially many times, this immediately yields:
Corollary 5. Fix ε, δ > 0, and let G0 be a set of i.i.d. samples from N (0, Σ), where n =
poly(d, 1/ε, log 1/δ). Let S be so that ∆(S, G0 ) ≤ ε. There is an universal constant C and an
b so that with probability 1 − δ, we have kΣ
b −1/2 ΣΣ
b −1/2 − IkF ≤ Cε. In
algorithm which outputs a Σ
particular, this implies that
b ≤ 2Cε .
dTV N (0, Σ), N (0, Σ)
6.1
Technical Overview
Our strategy for obtaining a high-dimensional estimate for the covariance based on solving lowdimensional subproblems will be substantially more challenging than it was for the unknown mean
case. The natural approach is to take the poly log(1/ε)-dimensional subspace of degree-2 polynomials of largest empirical variance and construct a filter. However, this fails because, unlike in the
mean case, we do not know the variance of these degree-2 polynomials to small error. For the unknown mean case, because we assumed that we knew the covariance was the identity (or spectrally
close to the identity), this was not an issue. Now, the variance of our polynomials depends on the
(unknown) covariance of the true Gaussian, which may be more than O(ε)-far from our current
estimate. Indeed, it is not difficult to come up with counterexamples where there are many large
eigenvalues of the empirical covariance matrix, but no filter can make progress.
We overcome this hurdle in several steps. First, in Section 6.3, we show how to find a filter
if there are many medium-sized eigenvalues of the empirical covariance matrix. This will proceed
roughly in the same way that the filter for the unknown mean does. If no filter is created, then
we know there are at most logarithmically large eigenvalues of the empirical covariance. In the
subspace V ⊆ Rd spanned by their eigenvectors, we can then learn the covariance to high accuracy
using our low-dimensional estimator.
Then, in Section 6.3, we show that if we restrict to the orthogonal subspace, i.e., the subspace where the empirical covariance matrix does not have large eigenvalues, we can indeed either
produce a filter or improve our estimate of the covariance restricted to this subspace using our
low-dimensional estimator. While the blueprint is similar to the filter for the unknown mean, the
techniques are much more involved and subtle.
Supposing we have not yet created a filter, we have now estimated the covariance on a polylogarithmic dimensional subspace V , and on V ⊥ . This does not in general imply that we have
learned the covariance in Frobenius norm. In block form, if we write
ΣV AT
,
Σ=
A ΣV ⊥
23
where here Rd is written as V ⊕ V ⊥ , this implies we have learned ΣV and ΣV ⊥ to high accuracy.
Thus, it remains to estimate the cross term A.
In Section 6.4, we show, given a polylogarithmically sized subspace V , and a good estimate of
the covariance matrix on V and V ⊥ , how to fill in the entire covariance matrix. Roughly, we do
this by randomly fixing directions in V , and performing rejection sampling based on the correlation
in the direction in V , and showing that the problem reduces to one of robustly learning the mean
of a Gaussian, which (conveniently) we have already solved. These steps together yield our overall
algorithm ImproveCov. Finally, in Section 6.6 we explain why there is a natural barrier that
makes reducing the running time from quasi-polynomial to polynomial (in 1/ε) difficult.
6.2
Additional Preliminaries
Here we give some additional preliminaries we will require in this Section.
6.2.1
The Agnostic Tournament
We also require the following classical result, which allows us to do agnostic hypothesis selection
with corrupted samples (see e.g., [DL01, DDS12, DK14, DDS15]).
Theorem 12. Fix ε, δ > 0. Let D1 , . . . , Dk , D be a set of distributions where mini dTV (Di , D) = γ.
1/δ
. There is an algorithm Tournament which given oracles for evaluating
Set n = Ω log k+log
ε2
the pdfs of D1 , . . . , Dk along with n independent samples X1 , . . . Xn from D, outputs a Di so that
dTV (Di , D) ≤ 3γ + ε with probability 1 − δ. Moreover, the running time and number of oracle calls
needed is at most O(n2 /ε2 ).
Remark 1. As a simple corollary of the agnostic tournament, observe that this allows us to do
agnostic learning without knowing the precise error rate ε. Throughout the paper, we assume the
algorithm knows ε. However, if the algorithm is not given this information, and instead given an
η and asked to return something with error at most O(ε + η), we may simply grid over {η, (1 +
γ)η, (1 + γ)2 η, . . . , 1} (here γ is some arbitrary constant that governs a tradeoff between runtime
and accuracy), run our algorithm with ε set to each element in this set, and perform hypothesis
selection via Tournament. Then it is not hard to see that we are guaranteed to output something
which has error at most O(ε + (1 + γ)η).
6.2.2
The Fourth Moment Tensor of a Gaussian
As in [DKK+ 16], it will be crucial for us to understand the behavior of the fourth moment tensor
of a Gaussian. Let ⊗ denote the Kronecker product on matrices. We will make crucial use of the
following definition:
2
Definition 6. For any matrix M ∈ Rd×d , let M ♭ ∈ Rd denote its canonical flattening into a vector
2
2
in Rd , and for any vector v ∈ Rd , let v ♯ denote the unique matrix M ∈ Rd×d so that M ♭ = v.
We will also require the following definition:
2
Definition 7. Let Ssym = {M ♭ ∈ Rd : M is symmetric}.
The following result was proven in [DKK+ 16]:
Theorem 13 (Theorem 4.15 in [DKK+ 16]). Let X ∼ N (0, Σ). Let M be the d2 × d2 matrix given
by M = E[(X ⊗ X)(X ⊗ X)T ]. Then, as an operator on Ssym , we have
T
.
M = 2Σ⊗2 + Σ♭ Σ♭
24
6.2.3
Polynomials in Gaussian Space
Here we review some basic facts about polynomials under Gaussian measure, which will be crucial
for our algorithm for learning Gaussians with unknown covariance. We equip the set of polynomials
over Rd with the Gaussian inner product, defined by hf, gi = EX∼N (0,I) [f (X)g(X)], and we let
kf k22 = hf, f i.
For any symmetric M with kM kF = 1, define the degree-2 polynomial p(x) = √12 (xT M x −
tr(M )). We call p the polynomial associated to M . Observe that p is even (i.e., has no degree-1
terms). We will use the following properties of such polynomials:
Lemma 10. Let M be symmetric, so that kM kF = 1. Let p be its associated polynomial. Then,
we have:
(i) EX∼N (0,I) [p(X)] = 0.
(ii) More generally, for any positive definite matrix Σ, we have EX∼N (0,Σ) [p(X)] = hM, Σ − Ii.
(iii) VarX∼N (0,I) [p(X)] = EX∼N (0,I) [p2 (X)] = hp, pi = 1.
(iv) More generally, for any positive definite matrix Σ, we have
2
T
1
E
[p2 (X)] = M ♭ Σ⊗2 M ♭ +
hΣ − I, M ♭ i .
2
X∼N (0,Σ)
Proof. The first three properties are a straightforward calculation. We show the last one here. By
definition, we have
E
X∼N (0,Σ)
1
E
[(X T M X − tr(M ))2 ]
2 X∼N (0,Σ)
1
E
[(X T M X)2 − 2(X T M X)tr(M ) + tr(M )2 ]
=
2 X∼N (0,Σ)
1
T
♭
2
♭T
=
E
[M (X ⊗ X)(X ⊗ X) M ] − 2hM, Σitr(M ) + tr(M )
2 X∼N (0,Σ)
T
T
(a) 1
M ♭ 2Σ⊗2 + Σ♭ Σ♭ M ♭ − 2hM, Σitr(M ) + tr(M )2
=
2
T
1 ♭
= M ♭ Σ⊗2 M ♭ +
hM , Σi2 − 2hM, Σitr(M ) + tr(M )2
2
2
1
♭
♭ T ⊗2
=M Σ M +
hΣ − I, M ♭ i ,
2
[p2 (X)] =
as claimed, where (a) follows from Theorem 13.
Observe that Lemma 10(iv) implies that if we take the top eigenvector of the d2 × d2 matrix
Σ⊗2 +
on the linear subspace
1 ♭ ♭ ⊤
M
M
2
V = {M ♭ : M is a symmetric d × d matrix} ,
then the associated polynomial maximizes EX∼N (0,Σ) [p2 (X)], and so we can find these polynomials
efficiently. More generally, if we take any linear subspace of degree two polynomials with associated
25
matrix subspace V ′ , so that V ′ ⊆ V , then the top eigenvector of the same matrix restricted to V ′
allows us to find the polynomial in this subspace which maximizes EX∼N (0,Σ) [p2 (X)] efficiently.
We have the following tail bound for degree-2 polynomials in Gaussian space: We will use ΠV (x)
and ΠV (S) to denote projection to a subspace V , of a point x and a set of points S, respectively.
We will also need the following hypercontractivity theorem for low-degree polynomials in Gaussian
space:
Theorem 14 (Hypercontractivity in Gaussian space, see e.g. [O’D14]). Let p : Rd → R be a degree
√
m polynomial, and let q ≥ 2 be even. Then EX∼N (0,I) [p(X)q ]1/q ≤ ( q − 1)m kpk2 .
We need the following definition:
Definition 8. Let Pk denote the set of even degree-k polynomials over d variables satisfying
VarX∼N (0,1) [p(X)] = 1. Moreover, for any subspace W ⊆ Rd , let Pk (W ) denote the set of even
polynomials over d variables which only depend on the coordinates in W .
b
Then by the arguments above, we have that for any two matrices Σ, Σ,
!
b F = sup
kΣ − Σk
E
[p(X)] −
E
[p(X)] .
p∈P2
X∼N (0,Σ)
b
X∼N (0,Σ)
In particular, by Lemma 2, this implies that when kΣ − Ik2 is small, then learning a Gaussian with
unknown covariance in total variation distance is equivalent to learning the expectation of every
even degree-2 polynomial.
Theorem 14 implies the following concentration for degree-4 (more generally, low-degree) polynomials of Gaussians:
Corollary 6. Let p be a degree-4 polynomial. Then there is some A, C ≥ 0 so that for all t ≥ C,
we have
Pr [|p(X) − E [p(X)]| ≥ tkpk2 ] ≤ exp −At1/2 .
N (0,I)
N (0,I)
Proof. Hypercontractivity in particular implies the following moment bound: for all q ≥ 2 even,
we have
E [(p(X) − E [p(X)])q ] ≤ (q − 1)qm/2 kp(X) − E [p(X)]kq2 .
N (0,I)
N (0,I)
N (0,I)
By a typical moment argument, and optimizing the choice of q, this gives the desired bound.
Hermite polynomials Hermite polynomials are what arise by Gram-Schmidt orthogonalization
applied with respect to this inner product. For a vector of non-negative integers a = (a1 , . . . , ad ),
we let Ha (x) : Rd → R be the Hermite
Pd polynomial associated with multi-index a. It is well-known
that the degree of Ha is |a| =
i=1 ai , and moreover, hHa , Hb i = δa,b . In particular, for any
r ≥ 1, the Hermite polynomials of degree at most r form an orthonormal basis with respect to the
Gaussian inner product for all polynomials with degree at most r.
Therefore, given any polynomial p : Rd → R with degree r, we may write it uniquely as
X
p(x) =
ca (p)Ha (x), where ca (p) = hp, Ha i .
|a|≤r
We define the kth harmonic component of p to be
X
p[k](x) =
ca (p)Ha (x) ,
|a|=k
and we say p is harmonic of degree k if it equals its kth part.
26
6.3
Working with Many Large Eigenvalues of the Second and Fourth Moment
As in the unknown mean case, we will need a filter to detect if there are many directions of the
empirical covariance which have too large an eigenvalue. Formally, we need:
Theorem 15. Fix ε, δ > 0. Assume kΣ − IkF ≤ ξ, where ξ = O(ε log 1/ε). Suppose that G0
b = ES [XX T ].
is (ε, δ)-good with respect to N (0, Σ). Let S be a set so that ∆(S, G0 ) ≤ ε. Let Σ
Then there is an algorithm FilterCovManyDeg2Eig and a universal constant C such that the
following guarantee holds:
b − I has more than O(log 1/ε) eigenvalues larger than Cξ, then the algorithm outputs a
1. If Σ
′
S so that ∆(S ′ , G0 ) < ∆(S, G0 ).
2. Otherwise, the algorithm outputs “OK”, and outputs an orthonormal basis v1 , . . . , vk for the
b − I with eigenvalue larger than Cξ.
subspace V of vectors spanned by all eigenvectors of Σ
The filter developed here is almost identical to the one developed for unknown mean. Thus, for
conciseness we describe and prove the theorem in Appendix D.1.
We will also need a subroutine to enforce the condition that not only does the fourth moment
tensor have spectral norm which is at most O(ε log2 1/ε) (restricted to a certain subspace of polynomials), but there can only be at most O(poly log 1/ε) directions in which the eigenvalue is large.
However, the techniques here are a bit more complicated, for a number of reasons. Intuitively,
the main complication comes from the fact that we do not know what the fourth moment tensor
looks like, whereas in the unknown mean case, we knew that the covariance was the identity by
assumption. Our main result in this subsection is the following subroutine:
Theorem 16 (Filtering when there are many large eigenvalues). Fix ε, δ > 0. Assume kΣ − IkF ≤
ξ, where ξ = O(ε log 1/ε). Let C be the universal constant in FilterCovManyDeg2Eig. Let
W ⊆ Rd be a subspace, so that for all v ∈ W with kvk2 = 1, we have v T ES [XX T ]v ≤ 1 + Cξ.
Suppose that G0 is (ε, δ)-good with respect to N (0, Σ). Let S be a set so that ∆(S, G0 ) ≤ ε. Let
k = O(log4 1/ε). Then there is an algorithm FilterCovManyDeg4Eig and universal constants
C1 , C2 such that the following guarantee holds:
1. If there exist p1 , . . . , pk ∈ P2 (W ) so that hpj , pℓ i = δjℓ for all j, ℓ, and so that ES [p2j (Y )] − 1 ≥
C1 ε for all j, then the algorithm outputs an S ′ so that ∆(S ′ , G0 ) < ∆(S, G0 ).
2. Otherwise, the algorithm outputs “OK”, and outputs an orthonormal basis p1 , . . . , pk′ for a
subspace V of degree-2 polynomials in P2 (W ) with k′ ≤ k so that for all p ∈ V ⊥ ∩ P2 , we
have ES [p2 (X)] − 1 ≤ C2 ε.
Moreover, FilterCovManyEig runs in time poly(d, 1/ε, log 1/δ).
Roughly, we will show that if there are many polynomials with large empirical variance, this
implies that there is a degree-four polynomial whose value is much larger than it could be if w
were the set of uniform weights over the uncorrupted points. Moreover, we can explicitly construct
this polynomial, and it has a certain low-rank structure which allows us to use the concentration
bounds we have previously derived.
6.4
Stitching Together Two Subspaces
This section is dedicated to giving an algorithm which allows us to fully reconstruct the covariance
matrix given that we know it up to small error on a low-dimensional subspace V and on W = V ⊥ .
27
Algorithm 8 Filter if there are many large eigenvalues of the fourth moment tensor
1: function FilterCovManyEig(S, ε, ξ, δ, W )
b = ES [XX T ]
2:
Let Σ
3:
Let C1 , C2 , C3 be some universal constants sufficiently large
4:
Let A be the constant in Corollary 6
5:
Let B be the constant in Claim 20
6:
Let m = 0
7:
Let k = O(log4 1/ε)
8:
while there exists p ∈ P2 (W ) so that p ∈ V ⊥ and ES [p2 (X)] − 1 > C1 ξ do
9:
Let Vm+1 = span(Vm ∪ p)
10:
Let m ← m + 1
11:
end while
12:
Let p1 , . . . , pm be an orthonormal basis for Vm
13:
if m ≥ k then
[4]
14:
Let qi = (p2i ) be the 4th harmonic component of p2i
15:
Let ri = p2i − qi be the degree-2 component of p2i
P
16:
Let Q(x) = ki=1 qi
17:
Find a T so that either:
√
• T > C3 d2 k log(|S|) and p(X) > T for at least one x ∈ S ′ , OR
√
• T > 4A2 C2 B k log2 (1/ε) and
√
Pr [Q(X) > T ] > exp(−A(T /4B k)1/2 ) + ε2 /(d log(|S|/δ))2 .
X∈u S
18:
19:
20:
21:
22:
return the set S ′ = {X ∈ S : Q(X) ≤ T }
else
return “OK”, and output p1 , . . . , pm
end if
end function
Theorem 17. Let 1 > ξ > η > ε > 0, and let τ > 0. Let Σ so that kΣ − IkF ≤ ξ. Suppose that
Rd is written as V ⊕ W for orthogonal subspaces V and W with dim(V ) = O(log(1/ε)). Suppose
furthermore that
ΣV AT
,
Σ=
A ΣW
with kΣV − IV kF , kΣW − IW kF = O(η). Let S0 = (G0 , E0 ) be an ε-corrupted set of samples from
N (0, Σ), and let S ⊆ S0 with ∆(S, G) ≤ O(ε) of size poly(d, 1/η, log 1/δ).
Then, there exists a universal constant C5 and an algorithm Stitching that given V, W, ξ, η, ε, τ
and S runs in polynomial time and with probability at least 1 − τ returns a matrix Σ0 with kΣ0 −
ΣkF = C5 η + O(ξ 2 ).
In the latter, we will show the algorithm works when τ = 2/3. As usual the probability of
success can be boosted by repeating it independently.5 The basic idea of the proof is as follows.
Since we already know good approximations to ΣV and ΣW , it suffices to find an approximation to
5
Observe the only randomness at this point is in the random choices made by the algorithm. Thus, one can just
(j)
(ℓ)
(1)
run this algorithm O(log 1/δ) times to obtain Σ0 , . . . , Σ0 and find any Σ0 which is O(η + ξ 2 ) close to at least a
2/3 + o(1) fraction of the other outputs.
28
A. In order to do this, we note that if we take a sample x from G conditioned on its projection to V
being some vector v, we find that the distribution over W is a Gaussian with mean approximately
Av. Running our algorithm for approximating the mean of a noisy-Gaussian, we can then compute
the mapping v → Av, which will allow us to compute A.
There are three main technical obstacles to this approach. The first is that we cannot condition
on xV taking a particular value, as we will likely see no samples from X with exactly that projection.
Instead, what we will do is given samples from X we will reject them with probabilities depending
on their projections to V in such a way to approximate the conditioning we require. The second
obstacle is that the errors in X may well be concentrated around some particular projection to
V . Therefore, some of these conditional distributions may have a much larger percentage of errors
than ε. To circumvent this, we will show that by carefully choosing how we do our conditioning
and by carefully picking the correct distribution over vectors v, that on average these errors are
only O(ε). Finally, we need to be able to reconstruct A from a collection of noisy approximations
to Av. We show that this can be done by computing these approximations at a suitably large
random set of v’s, and finding the matrix A that minimizes the average ℓ2 error between Av and
its approximation.
Our algorithm is given in Algorithm 9:
Algorithm 9 Stitching the two subspaces together
1: function Stitching(V, W, δ, ε, τ, S)
2:
Given a vector x, let xV and xW be the projections onto V and W , respectively.
3:
Let C be a sufficiently large constant (where C may depend on the constants in the big-O
terms in the guarantee that dim(V ) = O(log(1/ε))).
4:
Generate a set V = {v1 , . . . , vm } of (n/ε)C independent random samples from N (0, 2IV ).
5:
for v ∈ V do
6:
For each sample x ∈ S, add xW to a new set T independently with probability
exp(−kxV − vk2 /2) .
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
Treat T as a collection of independent samples from a noisy Gaussian with covariance
matrix IW + O(η).
Set av equal to 0 if T did not contain enough samples for our algorithm or if kµ̃k2 >
C log(1/ε).
for ε ∈ {1, 1/2, 1/4, 1/8, . . . , η} do
Let µ
e be the output of RecoverMeanNoisy(T, ε, (ε/n)2C , o(1), O(η)).
end for
Run Tournament with the output hypotheses.
Set av = µ̃, where µ
e is the winning hypothesis.
end for
Use linear programming to find the dim(W ) × dim(V )-matrix B that minimizes the convex
function Ev∈u S [|av − Bv|].
return
IV 2B T
.
Σ0 =
2B IW
end function
Before we prove Theorem 17, we will need the following definition.
29
Definition
9. A function f : Rd → R is a positive measure if f ≥ 0 and
R
|f |1 = f , and we say f1 ≤ f2 if f1 (x) ≤ f2 (x) pointwise.
R
f ≤ 1. We write
Proof of Theorem 17. Throughout this proof, let G = N (0, Σ). It is clear that this algorithm has
polynomial runtime and sample complexity. We have yet to show correctness. The first thing
that we need to understand is the procedure of rejection sampling, where we reject x except with
probability exp(−kxV − vk2 /2). Therefore, given a distribution D, we let the positive measure
Dv be what is obtained by sampling from D and accepting a sample x only with probability
exp(−kxV − vk2 /2). We need to understand the distributions Gv and Xv .
Note that the pdf of G is given by
p
(2π)−n/2 det(Σ−1 ) exp(−xΣ−1 x/2)dx.
Therefore, the density of Gv is
p
(2π)−n/2 det(Σ−1 ) exp(−xΣ−1 x/2 − kxV − vk22 /2)dx
p
=(2π)−n/2 det(Σ−1 ) exp(−x(Σ−1 + IV )x/2 + x · v − kvk22 /2)dx.
Note that letting µv := (Σ−1 + IV )−1 v = v/2 + O(δkvk2 ) that this equals
p
=(2π)−n/2 det(Σ−1 ) exp(−(x − µv )(Σ−1 + IV )(x − µv )/2 − kvk22 /2 + µv · v/2)dx
p
=(2π)−n/2 det(Σ−1 ) exp(−(x − µv )(Σ−1 + IV )(x − µv )/2 − kvk22 (1 + O(δ))/4)dx.
Note that this is a Gaussian with mean µv weighted by
s
det(Σ−1 )
exp(−kvk22 (1 + O(δ))/4) = Θ(2− dim(V )/2 exp(−kvk22 /4)) ,
det(Σ−1 + IV )
so long as kvk2 ≪ δ−1/2 . Therefore, if this condition holds, a random sample from G is accepted
by this procedure with probability Θ(2− dim(V ) exp(−kvk22 /4).
We also need to understand the fraction of samples from Xv that are erroneous. In a slight
abuse of notation, let E also denote the distribution which is uniform over the points in E, and let
L be the distribution which is uniform over the points in L. Therefore, we define
εv := 2n/2 exp(kvk22 /4)(|Ev |1 + |Lv |1 log(1/ε)) ,
that is approximately the fraction of samples from Xv that are errors (where subtractive errors are
weighted more heavily).
p
Note that so long as kvk2 < C log(1/ε) that so long as |Lv |1 = o(1) that T in Step 5 will
have sufficiently many samples with high probability. Furthermore, if this is the case, we have
(assuming our mean estimation algorithm succeeds) that kµ̃ − πW (µv )k2 ≤ O(εv + η). Therefore,
unless kµv k2 = Ω(C log(1/ε)) (which can only happen in kvk2 = Ω(C log(1/ε))), we have with high
probability that kav − πW (µv )k2 = O(εv + η). In fact, we assume that this holds for all v ∈ S with
kvk2 ≪ (C log(1/ε)).
In order to show that this is generally a good approximation, we need to know that εv is not
too large on average. In particular, we show:
Lemma 11. If v ∼ N (0, 2IV ), then Ev [εv ] = O(ε).
30
Proof. Letting F = E + log(1/ε)L, we have that
Z
− dim(V )/2
E[εv ] = (2π)
exp(kvk22 /4) exp(−kvk22 /4)dFv dv
v
Z
Z
− dim(V )/2
2
= (2π)
exp(−kx − vk /2)dF (x)dv = dF = O(ε).
Note that πW (µv ) = M v, where M = πW (Σ−1 + IV )−1 πVT . Let B be any dim(W ) × dim(V )
matrix. Note that
E
[kBv − M vk22 ] = 2kB − M k2F .
v∼N (0,2IV )
Therefore, Ev∼N (0,2IV ) [kBv − M vk2 ] = O(kB − M kF ). Since kBv − M vk22 is a degree-2 polynomial
in v, by Corollary 4 in [Kan12],
#
"
r
E
[kBv − M vk42 ]/2 > 1/81.
Pr
[kBv − M vk2 > kB − M kF /2] >
Pr
kBv − M vk22 >
v∼N (0,2IV )
v∼N (0,2IV )
v∼N (0,2IV )
Combining these we find that
E
[kBv − M vk2 ] = Θ(kB − M kF ) .
v∼N (0,2IV )
Next we show that our choice of S derandomizes this result.
Lemma 12. With high probability over the choice of S, we have that
E [kBv − M vk2 ] = Θ(kB − M kF ).
v∈u S
for all dim(W ) × dim(V ) matrices B.
Proof. This is equivalent to showing that for all matrices U = B − M it holds
E [kU vk] = Θ(kU kF ).
v∈u S
By the standard scaling laws, it suffices to show this only for U with kU kF = 1.
We also note that it suffices to show this only for U in an ε-net for all such matrices. This is
because if kU − U ′ kF < ε and if U ′ satisfies the desired condition, then
E [kU vk2 ] = E [kU ′ vk2 ] + O( E [k(U − U ′ )vk2 ]) = Θ(1) + O(ε) E [kvk2 ] ,
v∈u S
v∈u S
v∈u S
v∈u S
and with high probability Ev∈u S [kvk2 ] = O(log(1/ε)).
Note that such nets exist with size exp(poly(n/ε)). Therefore, it suffices to show that this
condition holds for each such U with probability exp(−(n/ε)Ω(C) ).
As noted above, it suffices that
E [kU vk2 ] = O( E [kU vk2 ]) = O(1)
v∈u S
v∈u S
and that
Pr [kU vk2 > 1/4] ≥ Pr [kU vk2 ≥ 1/4] − 1/100.
v∈u S
v∈u S
The first follows because if S = {v1 , . . . , vm } then Ev∈u S [kU vk2 ] is a degree-2 polynomial in the
√
vi with mean O(1) and variance O(1/ m), so by standard concentration results, is O(1) with 1 −
exp(−Ω(|S|)) probability. The latter follows from standard concentration bounds. This completes
the proof.
31
We also note that the random choice of S has another nice property:
Lemma 13. With S and av as above, with high probability we have that
E [kav − M vk2 ] = O(η) .
v∈u S
Proof. Let S = {v1 , . . . , vm }. Then kf (vi ) − M vi k2 are independent random variables with mean
E[O(εv + η)] = O(η) and variance at most
E[kav k22 + kM vk22 ] ≤ O(log2 (1/ε)) + E[kvk22 ] = O(log2 (1/ε)).
The result follows by Chernoff bounds.
Now by Lemmas 12 and 13 with high probability over the choice of S in Step 4, and the av in
Step 5 we have that for all dim(W ) × dim(V )-matrices B that
E [kBv − M vk2 ] = Θ(kB − M kF ),
v∈u S
and have that
E [kav − M vk2 ] = O(η).
v∈u S
Combining these statements, we have that for all dim(W ) × dim(V )-matrices B that
E [kav − Bk2 ] = Θ(kB − M kF ) + O(η).
v∈u S
Note that by taking B = M , this quantity is O(η). Therefore, the B found in Step 15 satisfies
O(ε) = E [kav − Bk2 ] = Θ(kB − M kF ) + O(η) ,
v∈u S
and therefore, kB − M kF = O(η).
The rest of the proof is a simple computation of the matrices involved. In particular, recall that
IV AT
ΣV AT
+ O(η) ,
=
Σ=
A IW
A ΣW
where the O(η) denotes a matrix with Frobenius norm O(η) and where kAkF = O(δ). It is easy to
see that
I
−AT
+ O(η + δ2 ).
Σ−1 = V
−A IW
Therefore,
Σ
−1
Hence,
(Σ
−1
2IV −AT
+ O(η + δ2 ).
+ IV =
−A IW
−1
+ IV )
IV /2 AT /2
=
+ O(η + δ2 ).
A/2
IW
Therefore, M = A/2 + O(η + δ2 ). Therefore, A = 2M + O(η + δ2 ) = 2B + O(η + δ2 ). And finally,
we conclude that
IV 2B T
+ O(η + δ2 ) = Σ0 + O(η + δ2 ).
Σ=
2B IW
32
6.5
The Full High-Dimensional Algorithm
We now show how to prove Theorem 11, given the pieces we have. We first show that given enough
samples from N (0, Σ), the empirical data set without corruptions satisfies the regularity conditions
in Section 2.2.2 with high probability. For clarity of exposition, the proof of this lemma is deferred
to Appendix D.3.
Lemma 14. Fix η, δ > 0. Let X1 , . . . , Xn be independent samples from N (µ, I), where n =
poly(d, 1/η, log 1/δ). Then, S = {X1 , . . . , Xn } is (η, δ)-good with respect to N (µ, I) with probability
at least 1 − δ.
Finally, we require the following guarantee, which states that if there is a degree-2 polynomial
whose expectation under S and the truth differs by a lot (equivalently, if the empirical covariance
differs from the true covariance in Frobenius norm substantially), then it must also have very large
variance under S.
Lemma 15. Fix ε, δ > 0. Assume kΣ − IkF ≤ ξ, where ξ = O(ε log 1/ε). Suppose that G0 is
(ε, δ)-good with respect to N (0, Σ), and let S ⊆ S0 be a set so that ∆(S, G0 ) ≤ ε. There is some
√
absolute constant C5 so that if p ∈ P2 is a polynomial so that ES [p(X)] − EN (0,Σ) [p(X)] > C5 ξε,
then ES [p2 (X)] − 1 > C1 ξ.
We defer the proof of this lemma to the Appendix.
We are now ready to present the full algorithm as Algorithm 10.
Algorithm 10 Filter if there are many large eigenvalues of the covariance
1: function ImproveCov(S, ξ, ε, δ)
2:
Let C be the universal constant in FilterCovManyDeg2Eig
3:
Let τ = poly(ε, 1/d, δ).
4:
Run FilterCovManyDeg2Eig(S, ε, ξ)
5:
if FilterCovManyDeg2Eig outputs S ′ then
6:
return S ′
7:
else
8:
Let V be the subspace returned by FilterCovManyDeg2Eig
9:
Let W = V ⊥ .
10:
Run FilterCovManyDeg4Eig(S, ε, ξ, δ, W )
11:
if FilterCovManyDeg4Eig outputs S ′ then
12:
return S ′
13:
else
14:
Let U1 be the subspace of degree 2 polynomials over W it returns
15:
Let U2 be the perpendicular subspace of degree 2 polynomials over W
b = ES [XX T ]
16:
Let Σ
17:
Let ΣV = LearnCovLowDim(S, ε, ξ, τ, V )
b
18:
Let ΣW = LearnMeanPolyLowD(S, ε, ξ, τ, U1 , U2 , Σ)
′
19:
Take poly(n, 1/ε) fresh ε-corrupted samples S
20:
return Stitching(V, W, ΣV , ΣW , ξ, ε, S ′ ).
21:
end if
22:
end if
23: end function
33
Proof of Theorem 11. Condition on the events that neither LearnCovLowDim nor Stitching
fail. This happens with probability at least poly(ε, 1/d, δ). Observe that if we pass the “if”
statement in Line 5, then by the guarantee of FilterCovManyDeg2Eig this is indeed an S ′
satisfying the desired properties. Otherwise, by the guarantees of FilterCovManyDeg2Eig, we
have that W satisfies the conditions needed by FilterCovManyDeg4Eig. Hence, if we pass
the “if” statement in Line 11, then the guarantee of FilterCovManyDeg4Eig this is indeed a
S ′ satisfying the desired properties. Otherwise, by Lemma 15, we know
√ that for all polynomials
p ∈ P2 over W orthogonal to U1 , we have | EN (0,Σ) − ES [XX T ]| ≤ C5 ξε. Thus, ΣW satisfies the
conditions needed by Stitching.
By Corollary 4, we know that ΣV satisfies the conditions for Stitching, and so the correctness
of the algorithm follows from Theorem 17.
6.6
The Barrier at Quasi-Polynomial
Here we explain why improving the running time from quasi-polynomial to polynomial in 1/ε will
likely be rather difficult. Recall that our strategy is to project the problem onto lower dimensional
subproblems and stitch together the answer. We need the dimension of the subspace to be large
enough that we can find a polynomial Q that is itself the sum of squares of k orthogonal degree
two polynomials pi so that the value of Q on the corrupted points is considerably larger than the
value on the uncorrupted points. More precisely, if we let S = (G, E) denote our corrupted set of
samples then we want EE [Q(X)] to be larger than Q(X) for all but a poly(ε) fraction of X ∈ G.
We then remove all points X ∈ S with large Q(X) and by the properties of Q we are guaranteed
that we throw out mostly corrupted points. It
√ turns out that the most aggressive we could be
is removing points where Q(X) is more than k standard deviations away from its expectation
under the true Gaussian. But since Q is a degree-four polynomial and we want Q(X)
√ to be smaller
than our cutoff for all but a poly(ε) fraction of X ∈ G, we are forced to choose k = Ω(log 1/ε),
which means that we need to reduce to k = Ω(log2 1/ε) dimensional subproblems. Thus, if we
solve low-dimensional subproblems in time exponential in the dimension, we naturally arrive at
a quasi-polynomial running time. It seems that any approach for reducing the running time to
polynomial would require fundamentally new ideas.
7
The General Algorithm
We now have all the tools to robustly learn the mean and covariance of an arbitrary high-dimensional
Gaussian. We first show how to reduce the problem of robustly learning the covariance of N (µ, Σ) to
learning the covariance of N (0, Σ), by at most doubling error, a trick previously used in [DKK+ 16]
and [LRV16]. Given an
√ ε-corrupted set of samples X1 , . . . , X2n of size 2n from N (µ, Σ), we may
let Yi = (Xi − Xn+i )/ 2. Then we see that if Xi and Xn+i are uncorrupted, then Yi ∼ N (0, Σ).
Moreover, at most 2εn of the Yi can be corrupted, since there are at most 2εn corrupted Xi .
Therefore, by doubling the error rate, we may assume that µ = 0. We may then apply the algorithm
b so that with high probability, we have kΣ
b −1/2 ΣΣ
b −1/2 − IkF ≤ O(ε)
in Corollary 5 to obtain a Σ
4
O(log
1/ε)
with polynomially many samples, and in poly(d, (1/ε)
) time.
b −1/2 X ′ .
We may then take an additional set of ε-corrupted samples {Xi′ , . . . , Xn′ }, and let Yi′ = Σ
i
′
′
b we have that if X is uncorrupted, then Y ∼ N (0, Σ)
e where
Then, by our guarantee on Σ,
i
i
e − IkF ≤ O(ε). We then run RecoverMeanNoisy with the Y ′ to obtain a µ
kΣ
b so that
i
−1/2
−1/2
b
e
b
b
kb
µ−Σ
µk2 ≤ O(ε). This guarantees that dTV (N (b
µ, Σ), N (Σ
µ, Σ)) ≤ O(ε), which in turn
b N (µ, Σ)) ≤ O(ε), as claimed.
implies that dTV (N (b
µ, Σ),
34
Therefore, we have shown:
Theorem 18. Fix ε, δ > 0. Given an ε-corrupted set of samples S from N (µ, Σ), where n =
poly(d, 1/ε, log 1/δ), there is an algorithm RecoverGaussian which takes as input S, ε, δ, and
b so that
outputs a µ
b, Σ
b ≤ O(ε) .
dTV (N (µ, Σ), N (b
µ, Σ))
Moreover, the algorithm runs in time poly(d, (1/ε)O(log
4
1/ε) , log 1/δ).
References
[BDLS17] S. Balakrishnan, S. S. Du, J. Li, and A. Singh. Computationally efficient robust sparse
estimation in high dimensions. In Proceedings of the 30th Annual Conference on Learning Theory, COLT ’17, 2017.
[Ber06]
T. Bernholt. Robust estimators are hard to compute. Technical report, University of
Dortmund, Germany, 2006.
[BNJ03]
D. M Blei, A. Y. Ng, and M. I. Jordan. Latent dirichlet allocation. Journal of machine
Learning research, 3(Jan):993–1022, 2003.
[BW41]
L. M. Blumenthal and G. E. Wahlin. On the spherical surface of smallest radius enclosing a bounded subset of n-dimensional euclidean space. Bull. Amer. Math. Soc.,
47:771–777, 1941.
[CSV17]
M. Charikar, J. Steinhardt, and G. Valiant. Learning from untrusted data. In Proc.
49th Annual ACM Symposium on Theory of Computing (STOC), pages 47–60. ACM
Press, 2017.
[CW01]
A. Carbery and J. Wright. Distributional and Lq norm inequalities for polynomials over
convex bodies in Rn . Mathematical Research Letters, 8(3):233–248, 2001.
[DBS17]
S. S. Du, S. Balakrishnan, and A. Singh. Computationally efficient robust estimation
of sparse functionals. CoRR, abs/1702.07709, 2017.
[DDS12]
C. Daskalakis, I. Diakonikolas, and R.A. Servedio. Learning Poisson Binomial Distributions. In Proceedings of the 44th Symposium on Theory of Computing, pages 709–728,
2012.
[DDS15]
A. De, I. Diakonikolas, and R. Servedio. Learning from satisfying assignments. In
Proceedings of the 26th Annual ACM-SIAM Symposium on Discrete Algorithms, SODA
2015, pages 478–497, 2015.
[DK14]
C. Daskalakis and G. Kamath. Faster and sample near-optimal algorithms for proper
learning mixtures of gaussians. In Proceedings of The 27th Conference on Learning
Theory, COLT 2014, pages 1183–1213, 2014.
[DKK+ 16] I. Diakonikolas, G. Kamath, D. M. Kane, J. Li, A. Moitra, and A. Stewart. Robust
estimators in high dimensions without the computational intractability. In Proceedings
of FOCS’16, 2016.
35
[DKK+ 17] I. Diakonikolas, G. Kamath, D. M. Kane, J. Li, A. Moitra, and A. Stewart. Being
robust (in high dimensions) can be practical. In Proceedings of the 34th International
Conference on Machine Learning, ICML ’17, pages 999–1008. JMLR, Inc., 2017.
[DKS16]
I. Diakonikolas, D. M. Kane, and A. Stewart. Robust learning of fixed-structure bayesian
networks. CoRR, abs/1606.07384, 2016.
[DKS17]
I. Diakonikolas, D. M. Kane, and A. Stewart. Statistical query lower bounds for robust
estimation of high-dimensional gaussians and gaussian mixtures. In Proceedings of
the 58th Annual IEEE Symposium on Foundations of Computer Science, FOCS ’17,
Washington, DC, USA, 2017. IEEE Computer Society.
[DL01]
L. Devroye and G. Lugosi. Combinatorial methods in density estimation. Springer Series
in Statistics, Springer, 2001.
[GCSR14] A. Gelman, J. B. Carlin, H. S. Stern, and D. B. Rubin. Bayesian data analysis, volume 2.
Chapman & Hall/CRC Boca Raton, FL, USA, 2014.
[GLS88]
M. Grötschel, L. Lovász, and A. Schrijver. Geometric Algorithms and Combinatorial
Optimization, volume 2. Springer, 1988.
[Goo15]
R. Goodell. Personal communication, 2015.
[Ham01]
F. Hampel. Robust statistics: A brief introduction and overview. In First International
Symposium on Robust Statistics and Fuzzy Techniques in Geodesy and GIS. A. Carosio,
H. Kutterer (editors), Swiss Federal Institute of Technology Zurich (ETH), Institute of
Geodesy and Photogrammetry, IGP-Bericht, number 295, pages 13–17, 2001.
[HM13]
M. Hardt and A. Moitra. Algorithms and hardness for robust subspace recovery. In
COLT 2013, pages 354–375, 2013.
[HR09]
P.J. Huber and E. M. Ronchetti. Robust statistics. Wiley New York, 2009.
[HRRS86] F. R. Hampel, E. M. Ronchetti, P. J. Rousseeuw, and W. A. Stahel. Robust statistics.
The approach based on influence functions. Wiley New York, 1986.
[Hub64]
P. J. Huber. Robust estimation of a location parameter. The Annals of Mathematical
Statistics, 35(1):73–101, 1964.
[Jun01]
H. W. E. Jung. Über die kleinste kugel, die eine räumliche figur einschliesst. J. Reine
Angew. Math, 123:241–257, 1901.
[Kan12]
D. M. Kane. A structure theorem for poorly anticoncentrated gaussian chaoses and
applications to the study of polynomial threshold functions. In FOCS, pages 91–100,
2012.
[Kla07]
B. Klartag. A central limit theorem for convex sets.
168(1):91–131, 2007.
[Li17]
J. Li. Robust sparse estimation tasks in high dimensions. CoRR, abs/1702.05860, 2017.
[LM00]
B. Laurent and P. Massart. Adaptive estimation of a quadratic functional by model
selection. Ann. Statist., 28(5):1302–1338, 2000.
36
Inventiones mathematicae,
[LRV16]
K. A. Lai, A. B. Rao, and S. Vempala. Agnostic estimation of mean and covariance. In
Proceedings of FOCS’16, 2016.
[NJB+ 08] J. Novembre, T. Johnson, K. Bryc, Z. Kutalik, A. R. Boyko, A. Auton, A. Indap, K. S.
King, S. Bergmann, M. R. Nelson, et al. Genes mirror geography within europe. Nature,
456(7218):98–101, 2008.
[O’D14]
R. O’Donnell. Analysis of Boolean Functions. Cambridge University Press, 2014.
[OF96]
B. A. Olshausen and D. J. Field. Emergence of simple-cell receptive field properties by
learning a sparse code for natural images. Nature, 381(6583):607, 1996.
[RL05]
P. J. Rousseeuw and A. M. Leroy. Robust regression and outlier detection, volume 589.
John Wiley & Sons, 2005.
[Rou85]
P. Rousseeuw. Multivariate estimation with high breakdown point. Mathematical Statistics and Applications, pages 283–297, 1985.
[SCV18]
J. Steinhardt, M. Charikar, and G. Valiant. Resilience: A criterion for learning in
the presence of arbitrary outliers. In Proc. of the 9th Conference on Innovations in
Theoretical Computer Science, 2018. to appear.
[Tuk75]
J. W. Tukey. Mathematics and picturing of data. In Proceedings of ICM, volume 6,
pages 523–531, 1975.
[Ver10]
R. Vershynin. Introduction to the non-asymptotic analysis of random matrices, 2010.
A
Lower Bounds on Agnostic Learning
In this section, we prove information theoretic lower bounds for robust distribution learning. In
particular, we will prove that no algorithm (efficient or inefficient) can learn a general distribution
ε
. We note that our lower
from an ε-corrupted set of samples to total variation distance less than 1−ε
bound applies in more general settings than we consider in this paper in couple of ways:
1. Our construction is univariate;
2. Our construction works for any pair of distributions which are
Gaussian distributions; and
ε
1−ε -close,
and not just for
3. Our construction holds for Huber’s ε-contamination model, which is weaker than the noise
model studied in this paper.
In Huber’s ε-contamination model, data is drawn from a mixture distribution (1 − ε)P + εQ,
where P is some distribution that we wish to estimate and Q is arbitrary (in particular, it might
ε
-close can be made indistindepend on P ). We will show that any two distributions which are 1−ε
guishable under this contamination model.
ε
Lemma 16. Let p1 and p2 be two distributions such that dTV (p1 , p2 ) = 1−ε
. Then there exist
Huber ε-contaminations p1′ and p2′ of p1 and p2 respectively, such that p1′ = p2′ . Therefore, no
ε
in Huber’s ε-contamination model.
algorithm can learn a distribution p up to accuracy < 1−ε
37
Proof. Since dTV (p1 , p2 ) =
ε
1−ε ,
we can write
ε
ε
p1 = 1 −
q1 ,
pc +
1−ε
1−ε
ε
ε
p2 = 1 −
q2 ,
pc +
1−ε
1−ε
where pc , q1 , q2 are distributions. Let p1′ be the Huber ε-contamination of p1 with q2 , p2′ be the
Huber ε-contamination of p2 with q1 . In other words,
p1′ = (1 − 2ε) pc + εq1 + εq2 ,
p2′ = (1 − 2ε) pc + εq2 + εq1 ,
and thus p1′ = p2′ as desired.
We note that a similar lower bound holds when p1 and p2 are required to be Gaussians, but
weakened by a factor of 2. This is due to the geometry of the space of Gaussian distributions, and
we note that in the case where the variance is known, this is achieved by the median.
Lemma 17.pNo algorithm
can output an estimate for the mean of a unit-variance Gaussian at
π
accuracy <
2 − o(1) ε with probability > 1/2 in Huber’s ε-contamination model. Consequently,
by Lemma 1, no algorithm can learn a Gaussian to total variation distance < 12 − o(1) ε in Huber’s
ε-contamination model.
Proof. We will consider the distributions p1 = N (−α, 1) and p2 = N (α, 1), where α is to be
specified later. We note that if p1 and p2 can be ε-corrupted into the same distribution, then the
best estimate for the mean is to output 0 (by symmetry), and the result holds.
2 (x)}
and
We will show that p1 can be ε-corrupted into a distribution f , where f (x) = max{p1 (x),p
η
η is a normalizing constant (the case of p2 follows similarly). In other words, f can be written as
(1 − ε)p1 + εq1 for some distribution q1 . Since q1 is a distribution, it is non-negative, and thus we
2 (x)}
require that (1 − ε)p1 (x) ≤ f (x) for all x ∈ R. Note that f (x) = max{p1 (x),p
≥ p1η(x) .
η
We can compute η as follows:
Z ∞
Z ∞
p2 (x)dx
max {p1 (x), p2 (x)} dx = 2
η=
0
−∞
r
α
α
1 1
2
+ erf √
α − O(α3 )
= 1 + erf √
=1+
=2
2 2
π
2
2
Thus,
p (x)
p1 (x)
q 1
=
f (x) ≥
≥
η
1 + π2 α − O(α3 )
1−
r
!
2
α + O(α2 ) p1 (x).
π
q
Again, since we require that (1 − ε)p1 (x) ≤ f (x), it suffices that 1 − ε ≤ 1 − π2 α + O(α2 ) and thus
q
pπ
that π2 α − O(α2 ) ≤ ε. This corresponds to α ≤
2 − o(1) ε, and the proof is complete.
Observe that this lower bound can be strengthened by a factor of two in the subtractive adversary model, where the adversary is able to both remove εN samples and add εN samples. In this
38
infinite sample regime, this means that a distribution can be corrupted to any distribution which is
ε-far in total
√ variation
distance. By noting that N (−α, 1) and N (α, 1) are both ε-far from N (0, 1)
for α =
2π + o(1) ε, we obtain the following lemma. Note that this lower bound is also achieved
by the median in this model.
Lemma 18.√No algorithm
can output an estimate for the mean of a unit-variance Gaussian at
2π + o(1) ε with probability > 1/2 in the subtractive adversary model. Consequently,
accuracy <
by Lemma 1, no algorithm can learn a Gaussian to total variation distance < ε in the subtractive
adversary model.
Finally, we conclude by sketching a lower bound for mean estimation of sub-Gaussian distributions.
Lemma 19. No algorithm can output an estimate for the mean of a sub-Gaussian distribution at
accuracy o(ε log1/2 (1/ε)) with probability > 1/2 in the Huber’s ε-contamination model.
Proof. We start with the distribution q = N (0, 1). We construct p1 by truncating the right tail of
q at the point xr = c1 log1/2 (1/ε) for some constant c1 , and rescaling the rest of the distribution
appropriately. Observe that, for an appropriate choice of c1 :
1
;
• p1 is sub-Gaussian with a constant of O 1−ε
• The mean of p1 is −c2 ε log1/2 (1/ε) for some constant c2 ;
• p1 can be corrupted in Huber’s ε-contamination model to be q.
We can similarly consider p2 , which is constructed by truncating the left tail of q at the point
xl = −c1 log1/2 (1/ε). Since p1 and p2 are indistinguishable when they are both ε-corrupted to q,
and the mean of all three distributions are separated by ≥ c2 ε log1/2 (1/ε), the lemma follows.
B
B.1
Omitted Proofs from Section 2
Proof of Lemma 1
Proof. Observe that by rotational and translational invariance, it suffices to consider the problem
when µ1 = −εe1 /2 and µ2 = εe1 /2, where e1 is the first standard basis vector. By the decomposability of TV distance, we have that the TV distance can in fact be written as a 1 dimensional
integral:
Z ∞
1
1
2
2
dTV (N (µ1 , I), N (µ2 , I)) = · √
e−(x−ε/2) /2 − e−(x+ε/2) /2 dx .
2
2π −∞
2 /2
The value of the function f (x) = e−(x−ε/2)
when x > 0, hence this integral becomes
2 /2
− e−(x+ε/2)
is negative when x < 0 and positive
Z ∞
1
2
2
e−(x−ε/2) /2 − e−(x+ε/2) /2 dx
dTV (N (µ1 , I), N (µ2 , I)) = √
2π 0
= F (ε/2) − F (−ε/2) ,
39
Rx
2
where F (x) = √12π −∞ e−t /2 dt is the CDF of the standard normal Gaussian. By Taylor’s theorem,
and since F ′′ (x) is bounded when x ∈ [−1, 1], we have
F (ε/2) − F (−ε/2) = F ′ (−ε/2)ε + O(ε3 )
1
2
= √ e−(ε/2) /2 ε + O(ε3 )
2π
1
= √ + o(1) ε ,
2π
which proves the claim.
C
Omitted Proofs from Section 4
C.1
Proof of Theorem 7
First, we require the following several claims about p(x) under various distributions, including the
true Gaussian and from choosing points uniformly at random from the sets of interest. Note that
throughout this section we will use Lemma 3, which implies that kΠV ′ µ − µ̃k2 < 2ε when the
parameters α, γ are chosen appropriately.
Claim 4. EN (µ,I) [p(X)] = kΠV ′ µ − µ̃k22 .
Proof. Letting v1 , . . . , vC1 β log(1/ε) be an orthonormal basis of V ′ , we have
C1 β log(1/ε)
E [p(X)] =
N (µ,I)
X
i=1
C1 β log(1/ε)
E
N (viT µ,1)
2
[hvi , X − µ̃i − 1] =
X
i=1
(viT (µ − µ̃))2 = kΠV ′ µ − µ̃k22 .
Claim 5. For some absolute constant c0 and all t ≥ kΠV ′ µ − µ̃k22 ,
(t − kΠV ′ µ − µ̃k22 )2
2
Pr (p(X) > t) ≤ 2 exp −c0 · min
, t − kΠV ′ µ − µ̃k2
.
C1 β log(1/ε)
N (µ,I)
Proof. This follows from Lemma 7, after re-centering the polynomial using Claim 4 and noting that
the spectral norm and squared Frobenius norm of the corresponding A matrix are at most 1 and
C1 β log(1/ε), respectively.
Claim 6. ES [p(X)] ≥ C1 ε log(1/ε).
Proof. Recall that µ̂ is the empirical mean of the point set.
E[p(X)] = E[kΠV ′ X − µ̃k22 ] − dim(V ′ )
S
S
≥ E[kΠV ′ X − ΠV ′ µ̂k22 ] − dim(V ′ )
S
1
≥ 1 + ε dim(V ′ ) − dim(V ′ ) = C1 ε log(1/ε) .
β
Claim 7. EG0 [p(X)] ≤ kµ − µ̃k22 + O(γ)ε.
40
Proof.
E [p(X)] ≤
G0
≤
≤
Z
0
Z
∞
Pr[p(X) ≥ t]dt
G0
O(d log(d/εδ))
Pr[p(X) ≥ t]dt
G0
0
Z
O(d log(d/εδ))
Pr [p(X) ≥ t]dt + O(γ)ε
N (µ,I)
0
≤ kµ − µ̃k22 + O(γ)ε
The inequalities follow from (γε, δ)-goodness and Claim 4.
Claim 8. φ(S, G0 ) EL [p(X)] = (C1 β + O(γ) + o(1))ε.
Proof.
φ(S, G0 ) E[p(X)] =
L
Z
∞
φ(S, G0 ) Pr[p(X) > t]dt
L
0
=
Z
O(d log(|G|/δ))
φ(S, G0 ) Pr[p(X) > t]dt
=
Z
(4)
L
0
C1 β log(1/ε)+4ε2
φ(S, G0 ) Pr[p(X) > t]dt
0
+
Z
L
O(d log(|G|/δ))
C1 β log(1/ε)+4ε2
φ(S, G0 ) Pr[p(X) > t]dt
L
2
≤ (C1 β log(1/ε) + 4ε )φ(S, G0 ) + 2
Z
O(d log(|G|/δ))
Pr[p(X) > t]dt
C1 β log(1/ε)+4ε2 G0
≤ (C1 β log(1/ε) + 4ε2 )φ(S, G0 )
Z O(d log(|G|/δ))
Pr [p(X) > t] +
+2
C1 β log(1/ε)+4ε2 N (µ,I)
2
γε
dt
d log(|G|/δ)
≤ (C1 β log(1/ε) + 4ε )φ(S, G0 ) + O(γε) + 8εc0 C1 β
≤ (C1 β + O(γ) + o(1))ε.
(5)
(6)
(7)
(4) and (5) follow from G0 being (γε, δ)-good, (6) is from Claim 5, and (7) is because
φ(S, G0 ) log(1/φ(S, G0 )) ≤ (1 + o(1))ε .
This gives:
Claim 9. ψ EE [p(X)] ≥ C1 ε log 1/ε − (C1 β + 1 + O(γ) + o(1))ε.
Proof. This immediately follows from Claims 6, 7, and 8.
We now show that in the case that there are many large eigenvalues, there will be a T satisfying
the conditions of the filter.
Claim 10. Suppose dim(V ) ≥ C1 β log(1/ε). Then there is a T satisfying the conditions in the
algorithm.
41
Proof. Suppose not. Then, we have
ψ E[p(X)] ≤
E
Z
2C3 log(1/ε)/c0
0
ψ Pr[p(X) ≥ t]dt +
E
≤ 2(C3 /c0 )ψ log 1/ε +
≤ 2(C3 /c0 )ψ log 1/ε +
≤ (2C3 /c0 )ε log 1/ε +
Z
Z
Z
C2 d log |S|/δ
Z
2(C3 /c0 ) log(1/ε)
C2 d log |S|/δ
∞
2C3 log(1/ε)/c0
ψ Pr[p(X) ≥ t]dt
E
ψ Pr[p(X) ≥ t]dt
E
Pr[p(X) ≥ t]dt
2(C3 /c0 ) log(1/ε) S
C2 d log |S|/δ
Pr[p(X) ≥ t]dt .
2(C3 /c0 ) log(1/ε) S
By applying the contradiction assumption, we have
c0 T
γε
Pr[p(X) ≥ t]dt ≤
exp −
dt
+
2C3
d log |S|/δ
2(C3 /c0 ) log(1/ε) S
2(C3 /c0 ) log(1/ε)
Z C2 d log |S|/δ
c0 T
exp −
≤
dt + C2 γε
2C3
2(C3 /c0 ) log(1/ε)
Z ∞
c0 T
dt + C2 γε
exp −
≤
2C3
2(C3 /c0 ) log(1/ε)
Z ∞
exp (−t) dt + C2 γε
≤ (C3 /c0 )
Z
Z
C2 d log |S|/δ
C2 d log |S|/δ
2 log(1/ε)
2
≤ (C3 /c0 ) · O(ε ) + C2 γε ,
and thus, we have
ψ E[p(X)] ≪ C1 ε log 1/ε − (C1 β + 1 + O(γ) + o(1))ε ,
E
which contradicts Claim 9.
It now suffices to prove that if we construct a filter, then the invariant that ∆ decreases is
preserved. Formally, we show:
Claim 11. Suppose dim(V ) ≥ C1 β log(1/ε). Let S ′ be the set of points we return. Then ∆(S ′ , G) <
∆(S, G).
Proof. Let T be the threshold we pick. If T > C2 d log(|S|/δ) then the invariant is satisfied since we
remove no good points, by (γε, δ)-goodness. It suffices to show that in the other case, we remove
log(1/ε) times many more bad points than good points. By definition we remove at least
γε
c0 T
+
.
|S| · exp −
2C3
d log |S|/δ
points. On the other hand, by (γε, δ)-goodness, Claim 5, we know that we throw away at most
|G| · Pr[p(X) > T ] ≤ |G0 | · Pr[p(X) > T ]
G
G0
γε
(T − kµ − µ̃k22 )2
2
, (T − kµ − µ̃k2 )
+
≤ |G0 | exp −c0 · min
.
C1 β log(1/ε)
2 log(1/ε)(d log(|S|/δ))
42
By our choice of T , and since kµ − µ
ek2 ≤ O(δ) ≪ 1, the first term is upper bounded by
(C3 − 1)2
(T − kµ − µ̃k22 )2
2
, (T − kµ − µ̃k2 )
≤ exp −c0 · min
T, (C3 − 1)T )
exp −c0 · min
C1 β log(1/ε)
C1 β
and so
exp −c0 · min
(C3 − 1)2
c0 T
T, (C3 − 1)T )
≤ exp −
C1 β
C3
c0 T 2
= exp −
2C3
c0 T
·ε
≤ exp −
2C3
1
c0 T
≪
.
exp −
2C3
log2 1/ε
Since we have maintained this invariant so far, in particular, we have thrown away more bad points
than good points, and so |G0 | ≤ (1 + ε)|S|.
γε
c0 T
Therefore, we have log(1/ε) · |G| · PrG [p(X) > T ] ≤ |S| · exp − 2C
+
d log |S|/δ , and hence,
3
the invariant is satisfied.
Claims 10 and 11 together imply the correctness of Theorem 7.
C.2
Proof of Lemma 5
We require the following basic fact of good sets (see, e.g., Fact 8.6 in [DKK+ 16]):
Fact 2. Let v ∈ Rd be any unit vector. Let G be (η, δ)-good with respect to N (µ, I). Then for any
T > 0,
Pr(|v · (X − µ)| > T ) ≤ 2 exp(−T 2 /2) + η/(d log(d/ηδ)) ,
G
and
Pr (|v · (X − µ)| > T ) ≤ 2 exp(−T 2 /2).
N (µ,I)
ε
Proof. Let hv, µ − µ
bi = R > β 1/2
. Observe that by direct calculation, we have ES [hv, X −
2
2
2
µi ] − ES [hv,
bi ] ≤ 3R . Hence, it suffices to show that ES [hv, X − µi2 ] ≥ 1 + Ω(R2 /ε) −
X − µ
ε.
γ + O β1
We write S = (G0 \ L, E) where |E| = ψ|S| and |L| = φ|S|. First we consider the expectation
of v · X over X in S. This is
µ0 − φµL + ψµE
.
1−φ+ψ
where µ0 is the mean over G0 , µL the mean over L and µE the mean over E. By (γε, δ)-goodness,
we have that |hv, µ0 − µi| ≤ γε. We also have that
Z ∞
Pr [|hv, X − µi| ≥ t] dt
|hv, µL − µi| ≤
L
0
Z √
O(
d log(d/εδ))
Pr [|hv, X − µi| ≥ t] dt
=
L
0
43
Since we have
|G0 |
Pr [|hv, X − µi| > t] ,
Pr [|hv, (X − µ)i| > t] ≤ min 1,
L
|L| G0
(8)
by Fact 2 and the calculations done in the proof of Corollary 8.8 in [DKK+ 16], we have that
ε
,
|hv, µL − µi| ≤ O(log |S|/|L|) ≤ O(log 1/φ). Since by assumption we have hv, µ − µ
bi = R > β 1/2
this implies that hv, µE − µi = Ω(R/ε). In particular, this implies that
E hv, X − µi2 ≥ E [hv, X − µi]2
E
E
2
R
≥Ω
.
ε2
Next we consider the expectation of hv, X −µi2 . By (γε, δ)-goodness, we have EG0 [hv, X −µi2 ] =
1 + O(ε). We also have that
Z ∞
2
φ Pr |hv, X − µi|2 ≥ t dt
φ E hv, X − µi ≤
L
0
L
5
≤ log(1/ε)φ +
β
≤
5
log(1/ε)φ +
β
Z
O(d log d/εδ)
(5/β) log(1/ε)
Z
O(d log d/εδ)
φ Pr |hv, X − µi|2 ≥ t dt
L
Pr |hv, X − µi|2 ≥ t dt
(5/β) log(1/ε) G0
Z
O(d log d/εδ)
5
Pr |hv, X − µi|2 ≥ t dt + γε
log(1/ε)φ +
β
(5/β) log(1/ε) N (µ,I)
1
= O
+ γ + o(1) ε.
β
=
Since
EG0 [hv, (X − µ)i2 ] − φ EL [hv, (X − µ)i2 ] + ψ EE [hv, (X − µ)i2 ]
,
S
1−φ+ψ
ε.
this is at least 1 + Ω(R2 /ε) − γ + O β1
E[hv, (X − µ)i2 ] =
D
D.1
Omitted Proofs from Section 6
Proof of Theorem 15
b − I, and
Our algorithm works as follows, just as for Algorithm 4. It finds all large eigenvalues of Σ
if there are too many, produces an explicit degree-2 polynomial which, as we will argue, produces
a valid filter. The formal pseudocode for our algorithm is in Algorithm 11.
The proofs of the following claims are identical to the proofs of Claim 6-9, by applying the
corresponding property of (γε, δ)-goodness for this setting, and so we omit them.
Claim 12. ES [p(x)] ≥ Cξ.
Claim 13. EG0 [p(X)] ≤
ε
.
(d log 1/ε)2
Claim 14. φ(S, G0 ) EL [p(X)] = (C + O(γ) + o(1))ε.
44
Algorithm 11 Filter if there are many large eigenvalues of the covariance
1: function FilterCovManyDeg2Eig(S, ε, ξ, δ)
b be the empirical second moment of S, respectively.
2:
Let Σ
b − I with eigenvalue more than
3:
Let V be the subspace of Rd spanned by eigenvectors of Σ
Cξ.
4:
if dim(V ) ≥ C1 log(1/ε) then
5:
Let V ′ be a subspace of V of dimension C1 log(1/ε).
6:
Let p(x) be the quadratic polynomial
p(x) = kΠV ′ (x)k22 − dim(V ′ ).
7:
Find a value T > 0 so that either:
• T > C2 d log(|S|/δ) and p(x) > T for at least one x ∈ S, or
• T > 2C3 log(1/ε)/c0 and PrS (p(x) > T ) > exp(−c0 T /(2C3 )) + ε/(d log(|S|/δ)).
8:
return
S ′ = {x ∈ S : p(x) ≤ T } .
else
return an orthonormal basis for V .
11:
end if
12: end function
9:
10:
Claim 15. ψ EE [p(X)] ≥ Cξ − (C1 + 1 + O(γ) + o(1))ε.
These claims imply just as before that there is a T with the desired properties. Again, the proof
is identical. Formally:
Claim 16. Suppose dim(V ) ≥ C log(1/ε). Then there is a T satisfying the conditions in the
algorithm.
Finally, we show the invariant that ∆ decreases. This is almost identical to the proof of Claim
11, however, we need to slightly change our application of the Hanson-Wright inequality. Formally,
we show:
Claim 17. Suppose dim(V ) ≥ C1 log 1/ε. Let S ′ be the set of points we return. Then ∆(S ′ , G) <
∆(S, G).
Proof. Let T be the threshold we pick. If T > C2 d log(|S|/δ) then the invariant is satisfied since we
remove no good points, by (γε, δ)-goodness. It suffices to show that in the other case, we remove
log 1/ε times many more bad points than good points. By definition we remove at least
c0 T
ε
|S| · exp −
+
2C3
d log |S|/δ
points. On the other hand, observe that if v1 , . . . , vk is an orthonormal basis for V , we have
X
kΠV xk22 =
hvi , Xi2
X
vi viT Σ1/2 X ,
= yΣ1/2
45
so that ifPX ∼ N (0, Σ), we have Y ∼ N (0, I). Let M = Σ1/2
kM kF ≤ ni=1 kΣk2 ≤ (1 + ξ)k, and kM k2 ≤ kΣk2 ≤ (1 + ξ). Since
E [kΠV xk22 ] − k =
N (0,Σ)
≤
k
X
i=1
k
X
i=1
P
vi viT Σ1/2 . We have that
viT (Σ − I)vi
kΣ − Ik2
≤ (1 + ξ)k ,
we have by Hanson-Wright that
Pr
N (0,Σ)
(T − (1 + ξ)k)2
, (T − kµ − (1 + ξ)kk22 )
,
kΠV xk22 − dim(V ) > T ≤ exp −c0 · min
C1 log(1/ε)
so by our choice of T , we have
(T − kµ − µ̃k22 )2
(C3 − 1)2
2
exp −c0 · min
, (T − kµ − µ̃k2 )
≤ exp −c0 · min
T, (C3 − 1)T )
.
C1 log(1/ε)
C1
The remaining proof now proceeds identically to the proof of Claim 17.
D.2
Proof of Theorem 16
Clearly, the only non-trivial condition to certify for Theorem 16 is that if we are in Case (1), the
returned set satisfies the desired properties.
Our proof will roughly follow the same structure as the proof of Theorem 7. We will first show
that the empirical average of the polynomial Q can only be large because of the contribution of the
points in E (Claim 21). We will then show that this implies that there exists a threshold T which
the algorithm will find in this case (Claim 22). Finally, we will show that for any such T we find,
the returned set of points will indeed satisfy ∆(S ′ , G0 ) < ∆(S, G0 ), which implies the correctness
of the algorithm (Claim 23).
We first show the following claim:
Claim 18. ES [ri ] − EN (0,Σ) [ri ] ≤ (4C + 1)ξ
√
Proof. Let us suppose that pi corresponds to the matrix Ai , given by (Ai )a,b = ∇a ∇b pi / 2!, so
that the Ai are orthonormal with respect to the Frobenius norm. Then the constant harmonic
2
2
2
part
√ of2 pi corresponds to kAi kF ≤ 1. The degree-2 harmonic part of p2i corresponds to the matrix
2 2Ai . This is because if we let Bi be the matrix corresponding to pi , we have
1
∇a ∇b p2i (X)
E
(Bi )a,b = √
2 X∼N (0,I)
2
=√
E
[∇b (pi (X)∇a pi (X))]
2 X∼N (0,I)
2
E
[∇a pi (X)∇b pi (X) + pi (X)∇a ∇b pi (X)]
=√
2 X∼N (0,I)
2
=√
E
[∇a pi (X)∇b pi (X)]
2 X∼N (0,I)
√
= 2 2h(Ai )a , (Ai )b i ,
46
where the second to last line follows since ∇a ∇b pi (X) is a constant, and the last line follows from
explicit computation. In particular,√this implies that the non-constant component of
Pri corresponds
to matrix with trace norm at most
2
2
≤
4.
Therefore,
r
can
be
written
as
r
(x)
=
αi hvi , xi2 +C0
i
i
P
for some constant C0 , where
|αi | ≤ 4. Thus, by our assumption, we have ES [ri ] ≤ 1 + Cξ. The
claim then follows since Σ and I are differ in Frobenius norm by at most ξ.
Claim 19. ES [Q(X)] − EX∼N (0,Σ) [Q(X)] ≥ (C1 − 6)ξk
Proof. Observe that since kΣ − IkF ≤ ξ, in particular we have kΣ⊗2 − I ⊗2 k2 ≤ ξ, and hence
by Lemma 10 we have that EN (0,Σ) [p2 (X)] − EX∼N (0,I) [p2 (X)] ≤ 2ξ for all p ∈ P2 . Hence, in
particular, we have
E[p2i (X)] − E [p2i (X)] ≥ (C1 − 2)ξ .
S
By Claim 18, we have
E
S
"
In particular, this implies that
N (0,Σ)
k
X
i=1
#
ri −
E [Q(X)] −
S
E
N (0,Σ)
"
k
X
i=1
#
ri ≤ (4C + 1)k .
E [Q(X)] ≥ (C1 − 4C − 5)ξk ,
N (0,I)
as claimed.
We now show:
Claim 20. There is some universal constant B so that
E
√
[Q2 (X)]1/2 ≤ B k .
X∼N (0,Σ)
Before we prove this, we need the following lemma:
Lemma 20. For any degree-4 polynomial p, and any Σ, if we let F (y, p, Σ) denote the yth percentile
of p under Σ, then we have F (1/4, p2 , Σ), F (3/4, p2 , Σ) = Θ(EX∼N (0,Σ) [p2 (X)]).
Proof. Let µ′ = EX∼N (0,Σ) [p2 (X)]. First, we note that Pr(p2 (X) > 4µ′ ) ≤ 1/4, so F (3/4, p2 , Σ) ≤
4µ′ . On the other hand,
anti-concentration bounds [CW01], we have that Pr(p2 (X) ≤
√ ′ by known
1/8
′
εµ ) = Pr(|p(X)| ≤ εµ ) = O(ε ). So, for ε a sufficiently small constant, Pr(p2 (X) ≤ εµ′ ) < 1/4,
and therefore, F (1/4, p2 , Σ) ≥ εµ′ . Since εµ′ ≤ F (1/4, p2 , Σ) ≤ F (3/4, p2 , Σ) ≤ 4µ′ , this completes
our proof.
Proof of Claim 20. Since dTV (N (0, Σ), N (0, I)) = O(ε log 1/ε) = o(1), we have that F (3/4, Q2 , Σ) ≤
F (1/4, Q2 , Σ), so by the above lemma, we have that
E
[Q2 (X)]1/2 ≤ O(
X∼N (0,Σ)
E
[Q2 (X)]1/2 ) .
X∼N (0,I)
Hence, it suffices to bound kQk22 .
47
√
Let us again suppose that pi corresponds to the matrix Ai , given by (Ai )a,b = ∇a ∇b pi / 2!.
Note that the Ai are symmetric matrices that form an orthonormal set. We note that qi is the
harmonic degree-4 polynomial corresponding to the rank-4 tensor
√
Ta,b,c,d = ∇a ∇b ∇c ∇d qi / 24
√
= ∇a ∇b ∇c ∇d (p2i )/ 24
√
= (2(∇a ∇b pi )(∇c ∇d pi ) + 2(∇a ∇c pi )(∇b ∇d pi ) + 2(∇a ∇d pi )(∇b ∇c pi ))/ 24
√
= (4(Ai )a,b (Ai )c,d + 4(Ai )a,c (Ai )b,d + 4(Ai )a,d (Ai )b,c )/ 24
√
= 6Sym(Ai ⊗ Ai ).
√ P
By linearity, Q corresponds to the rank-4 tensor T = 6 ki=1 Sym(Ai ⊗ Ai ). It thus suffices to
show that kT k2 = O(k), where here k · k2 denotes the square root of the sum of the squares of the
entries of the tensor. In order to show this, we note that kT k2 = supkV k2 =1 hV, T i. Therefore, it
√
suffices to show that for all 4-tensors V with kV k2 ≤ 1, that hV, T i = O( k). We note that
+
+
*
*
k
k
X
X
√
√
hV, T i = 6 V,
Sym(Ai ) ⊗ Ai = 6 Sym(V ),
Ai ⊗ Ai .
i=1
i=1
Note that Sym(V ) is a symmetric 4-tensor of ℓ2 norm at most 1. Thinking of Sym(V ) as a symmetric
Pn 2
matrix over 2-Tensors, we can write it as Sym(V ) =
j=1 λj Bj ⊗ Bj , where the Bj ’s are an
Pn 2 2
orthonormal basis for the set of 2-tensors and j=1 λj = kSym(V )k22 ≤ 1. Then, if ci,j := hAi , Bj i,
P
and cj = ki=1 c2i,j , we have that
2
2
k
n
n
X
√ X
√ X
c2i,j = 6
λj
λj cj .
hV, T i = 6
i=1
j=1
j=1
Since the Ai are orthonormal, cj ≤ 1 for all j. Furthermore,
2
2
n
X
cj =
c2i,j =
k
X
i=1
i=1 j=1
j=1
Therefore,
n
k X
X
kAi k2F = k.
v
v
u n2
u n2
n2
uX
√
√
√ X
√ uX
hV, T i = 6
λj cj ≤ 6t
λ2j t
c2j ≤ 6 · 1 · k = O( k) .
j=1
j=1
j=1
Let µ′ = EX∼N (0,Σ) [Q(X)]. We now show that since G0 is ε-good, then almost all of the
difference in Claim 19 must be because of the points in E.
Claim 21. EE [Q(Y )] − µ′ ≥
C1 −6
2
·
ξk
φ
Proof. It suffices to show that
E[Q(X)] − µ′ ≤
G
48
C1 − 6
ξk ,
2
(9)
that is, the good points do not contribute
√ much to the difference. By (ε, δ)-goodness of G0 and
′
Claim 20 we have E0 [Q(Y )] − µ ≤ O ε k . Moreover, we have
φ(E[Q(X)] −
L
E
[Q(X)]) ≤
X∼N (0,Σ)
(a)
=
Z
∞
0
Z
φ Pr Q(X) − µ′ ≥ t dt
L
√ 2
O( kd log |S|/δ)
φ Pr Q(X) − µ′ ≥ t dt
L
0
Z
√
2
≤ φ · O( k log 1/ξ) +
√
≤ φ · O( k log2 1/ξ) +
√
≤ φ · O( k log2 1/ξ) +
(b)
√
O( kd2 log |S|/δ)
φ Pr
√
L
O( k log2 1/ξ)
Z O(√kd2 log |S|/δ)
Pr
√
G0
O( k log2 1/ξ)
Z O(√kd2 log |S|/δ)
Q(X) − µ′ ≥ t dt
Q(X) − µ′ ≥ t dt
Pr [Q(X) − µ′ ≥ t]dt + O(ξ)
√
O( k log2 1/ξ)
N (0,Σ)
√
≤ O(ξ k) + exp(−Ω(log 1/ξ))
(c)
≪ 5ξk ,
where (a) follows from the boundedness condition of (ε, δ)-goodness, (b) follows from the last
condition of goodness, and (c) follows from hypercontractivity. This shows (9), which completes
the proof.
We now show that this implies that there must be a T satisfying the conditions in Algorithm 8.
Claim 22. If dim(Vm ) ≥ k, then Algorithm 8 returns a T satisfying the conditions in the algorithm.
Proof. Suppose not. By the assumption that kI − ΣkF ≤ ξ, we have |µ′ | = | EX∼N (0,Σ) [Q(X)]| ≤
P
2ξkQk2 ≤ ki=1 kpi k2 + kri k2 ≤ 5ξk. Thus, we have
C1 − 6
C1 − 6
− 5 ξk ≫
E[Q(X)] ≥
ξk .
E
2ψ
3ψ
But since we assume there is no T satisfying the conditions in Algorithm 8, we have
Z ∞
Pr[p(X) > t]dt
ψ E [p(X)] ≤ ψ
X∈E
0
E
√
2
≤ ψ · 4B k log (1/ξ) +
≤ 4Bψ ·
√
k log2 (1/ξ) +
≤ 4C2 Bψ ·
√
Z
√
O(d2 k log(|S|/δ))
ψ Pr[p(X)
√
E
4B k log2 1/ξ
Z O(d2 √k log(|S|/δ))
Pr[p(X)
√
S
4B k log2 1/ξ
Z O(d2 √k log(|S|/δ))
k log2 (1/ξ) +
√
4A2 C2 B k log2 1/ξ
√ Z ∞
√
≤ 4A2 C2 Bψ k log2 (1/ξ) + 4B k
√
e 2 ) ≪ 10ξk ,
≤ O( kξ) + O(ε
A2 log2 1/ξ
> t]dt
> t]dt
t
√
exp −A
4B k
1/2 !
ε2
+ 2
d log |G|/δ
e 2)
exp(−Ω(t1/2 ))dt + O(ε
which is a contradiction, for our choice of C1 , C2 , C3 , and since we chose k = O(log4 1/ε).
49
!
dt
The final thing we must verify is that the number of good points we remove is much smaller
than the number of bad points we remove. Formally, we show:
Claim 23. If dim(Vm ) ≥ k, then Algorithm 8 returns a S ′ satisfying ∆(S ′ , G0 ) < ∆(S, G0 ).
√
Proof. Let T be the threshold we pick. If T > C3 d2 k log(|S|) then the invariant is satisfied since
we remove no good points, by (ε, δ)-goodness. It suffices to show that we remove log 1/ε times
many more bad points than good points. Otherwise, by definition we remove a total of
!
1/2 !
ε2
T
√
+ 2
|S| · exp −A
d log2 |S|/δ
2B k
points. On the other hand, by (ε, δ)-goodness, hypercontractivity, and Claim 20, we know that the
total number of points we throw away is at most
|G| · Pr[Q(X) > T ] ≤ |G0 | · Pr[Q(X) > T ]
G
G0
−1/2 !
!
ε2
+
≤ |G0 | exp −A
2 log(1/ε)(d log(|G|/δ))2
!
1/2 !
T
ε2
√
≤ |G0 | exp −A
.
+
2 log(1/ε)(d log(|G|/δ))2
B k
T
kQk2
Since we have maintained this invariant so far, in particular, we have√
thrown away more bad points
than good points, and so |G0 | ≤ (1 + ε)|S|. Moreover, since T ≥ 4B k log2 1/ε, we have
exp −A
T
√
B k
1/2 !
= exp −A
≤ exp −A
≤ exp −A
≤ exp −A
T
√
4B k
T
√
4B k
T
√
4B k
T
√
4B k
1/2 !2
1/2 !
1/2 !
1/2 !
· exp −A
· ε−1
·
!1/2
√
2
4B k log 1/ε
√
4B k
1
.
log 1/ε
1/2
Therefore, we have log(1/ε) · |G| · PrG [Q(X) > T ] ≤ |S| · exp −A 4BT√k
+
hence, the invariant is satisfied.
ε2
d2 log |G|/δ
, and
Claims 22 and 23 together prove the theorem.
D.3
Proof of Lemma 6
By Lemma 8.16 in [DKK+ 16], the first two items hold together with probability 1 − O(δ) after
taking O(poly(d, 1/η, log 1/δ)) samples. Thus, it suffices to show that the last property holds with
probability 1 − O(δ) given O(poly(d, 1/η, log 1/δ)) samples. We may clearly WLOG take Σ = I.
Moreover, these properties are clearly invariant under scaling, and hence, it suffices to prove them
50
for degree-four polynomials with VarN (0,I) [p(X)] = 1. For any fixed degree-4 polynomial p, by
P
hypercontractivity, since P (X) = n1 ni=1 p(Xi ) has variance Var[P (X)] = n1 , we have that
"
#
n
1X
Pr
p(Xi ) − E [p(X)] ≥ t ≤ exp −At1/2 n1/4 .
n
N (0,I)
i=1
In particular, if we take
n=Ω
then
Pr
"
log4 1/δ
η2
,
#
n
1X
p(Xi ) − E [p(X)] ≥ η ≤ O(δ) .
n
N (0,I)
i=1
4
Since there is a 1/3-net over all degree-4polynomials
with unit variance of size (1/3)O(d ) , by union
d4 log1/δ
samples, then the second to last property holds
bounding, we obtain that if we take Ω
η2
with probability 1 − δ.
Finally, the same net technique used to prove Lemma 8.16, along with hypercontractivity, may
be used to show the last property holds when given poly(d, log 1/η, log 1/δ) samples.
D.4
Proof of Lemma 15
Proof. Let µ′ = EN (0,Σ) [p(X)], s = EN (0,Σ) [p2 (X)], and let ES [p(X)] − EN (0,Σ) [p(X)] = K. By
(ε, δ)-goodness and Lemma 10, we have
E [p(X)] − µ′ ≤ ε
E [p2 (X)]1/2 ≤ 2ε .
G0
N (0,Σ)
(10)
Moreover, for some appropriate choice of β1 , we have
Z ∞
′
φ Pr [|p(X) − µ| ≥ t] dt
φ E[p(X)] − φµ ≤
L
0
L
≤ β1 φ log 1/ε +
(a)
Z
≤ β1 φ log 1/ε +
(b)
≤ β1 φ log 1/ε +
(c)
O(d log |S|/δ)
β1 log 1/ε
|G0 |
|S|
|G0 |
|S|
Z
Z
φ Pr [|p(X) − µ| ≥ t] dt
L
O(d log |S|/δ)
β1 log 1/ε
O(d log |S|/δ)
β1 log 1/ε
Pr [|p(X) − µ| ≥ t] dt
G0
Pr [|p(X) − µ| ≥ t] dt + O(ε)
N (0,Σ)
≤ β1 φ log 1/ε + εO(β1 ) + O(ε)
= O(ε) ,
(11)
where (a), (b), follows from the definition of goodness, and (c) follows from standard Guassian
concentration bounds. In particular, (10) and (11) together imply that
2K
K
≥ E[p(X) − µ′ ] ≥
,
E
ψ
2ψ
51
so this implies
K2
.
4ψ 2
E[(p(X) − µ′ )2 ] ≥
E
Expanding this yields that
E[(p(X) − µ′ )2 ] = E[p2 (X)] − 2µ′ E[p(X)] + (µ′ )2 ,
E
E
E
so since |µ′ | ≤ kΣ − IkF ≤ ξ, we have
E[p2 (X)] ≥
E
2K
K2
−
+ O(ξ 2 ) .
4ψ 2
ψ
(12)
On the other hand, by Lemma 10, we have
1
|s − 1| ≤ kΣ⊗2 − I ⊗2 k2 + kΣ − Ik2F
2
≤ ξ + O(ξ 2 ) .
(13)
Also, by (ε, δ)-goodness and Lemma 10, we have
E [p2 (X)] − s ≤ ε(1 + O(ξ)) ,
(14)
G0
and, for A as in Corollary 6, and β appropriately chosen,
Z ∞
2
φ(E[p (X)] − s) ≤
φ Pr[|p2 (X) − s| ≥ t]dt
L
0
L
2
≤ βAφ log 1/ε +
Z
O((d log |S|/δ)2 )
|G0 |
≤ βAφ log 1/ε +
|S|
2
(a)
≤ βAφ log2 1/ε +
(b)
≤ βAφ log2 1/ε +
φ Pr[|p2 (X) − s| ≥ t]dt
L
2A log2 1/ε
|G0 |
|S|
Z ∞
Z
O((d log |S|/δ)2 )
βA log 1/ε
Z
Pr[|p2 (X) − s| ≥ t]dt
G0
2
O((d log |S|/δ)2 )
2
βA log 1/ε
1/2 /2
e−At
2
Pr [|p2 (X) − s| ≥ t]dt + O(ε2 )
N (0,Σ)
dt + O(ε2 )
βA log 1/ε
2
≪ 2βAξ + O(ε) + O(ε ) ,
(15)
where (a) follows from goodness, and (b) follows from Corollary 6 and since kΣ − IkF = o(1). Thus,
(12), (13), (14), and (15) together imply that
E[p2 (X)] − 1 ≥ 1 − (ξ + O(ξ 2 )) − O(ε) − 2βAξO(ε) +
S
and so since K ≥ Ω
√
εξ , this gives the desired bound.
52
K2
− 2K + O(ξ 2 ) ,
4ψ
| 7cs.IT
|
POSTNIKOV-SHAPIRO ALGEBRAS, GRAPHICAL
MATROIDS AND THEIR GENERALIZATIONS
arXiv:1509.08736v4 [math.CO] 5 Mar 2017
GLEB NENASHEV
Abstract. A.Postnikov and B.Shapiro introduced a class of commutative algebras which enumerate forests and trees of graphs.
Our main result is that the algebra counting forests depends only
on graphical matroid and the converse.
By these algebras we motivate generalization of the definition of
spanning forests and trees for hypergraphs and the corresponding
"hypergraphical" matroid. We present 3 different equivalent definitions, which can be read independently from other parts of the
paper.
Contents
1. Introduction
1
2. Algebras and Matroids
4
3. Algebras associated with t-labelled forests
10
4. Vector configurations and Hypergraphs
13
5. Algebras corresponding to spanning trees, Problems 18
References
20
1. Introduction
There are a lot of results about enumeration of spanning trees and
forests of graphs. The famous matrix-tree theorem of Kirchhoff (see [16]
and p. 138 in [26]) claims that the number of spanning trees of a given
graph G equals the determinant of the Laplacian matrix of G. Many
generalizations of the classical matrix-tree theorem were constructed
in a long period, e.g. for directed graphs, matrix-forest theorems, etc
(see e.g. [8], see [6] and references therein). It is also well known that
the number of maximal spanning forests of G (or equivalently trees for
connected G) equals TG (1, 1) while the number of all spanning forests
of G equals TG (2, 1), where TG is the Tutte polynomial of G (see p. 237
in [26]).
We focus here on algebras introduced by A. Postnikov and B. Shapiro
in [22] and on graphical matroids, both objects enumerate spanning
Key words and phrases. Commutative algebra, Graph, Hypergraph, Matroid,
Spanning trees and forests, Tutte polynomial.
1
2
G. NENASHEV
trees and forests. Before our paper there was no direct relation between them, only across the Tutte polynomial, Critical group (Sandpile
model). Also with the motivation given by these algebras we introduce
3 equivalent definitions of spanning trees (forests) of a hypergraph and
present corresponding "hypergraphical" matroid. If you are interested
only in the latter, scroll to section 4.2.
The field K of zero characteristic is fixed throughout this paper, for
example C or R. By a graph we always mean an undirected graph
without loops (multiple edges are allowed). We use the standard notation: E(G) and e(G) are the set and the number of edges of graph
G; V (G) and v(G) are the set and the number of vertices of G; c(G)
is the number of connected components of G and TG (x, y) is the Tutte
polynomial of G.
At first, we present two definitions of Postnikov-Shapiro algebras
counting spanning forests.
Notation 1.1. Take an undirected graph G on n vertices.
(I) Let ΦFG be the commutative algebra over K generated by {φe : e ∈
E(G)} satisfying the relations φ2e = 0, for any e ∈ E(G).
Fix any linear order of vertices of G. For i = 1, . . . , n, set
X
Xi =
ci,e φe ,
e∈G
where
ci,e
1 if e = (i, j), i < j;
= −1 if e = (i, j), i > j;
0 otherwise.
Denote by CGF the subalgebra of ΦFG generated by X1 , . . . , Xn .
(II) Consider the ideal JGF in the ring K[x1 , · · · , xn ] generated by
!DI +1
X
pFI =
xi
,
i∈I
where I ranges over all nonempty subsets of vertices, and DI is the total
number of edges between vertices in I and vertices outside I. Define
F
the algebra BG
as the quotient K[x1 , . . . , xn ]/JGF .
Remark 1.1. In the first definition, the order of vertices is not important. Easy to see that all such algebras are isomorphic to the above
definition of CGF , even though we can separately choose the "smallest"
vertex for each edge. For an orientation G of graph G, we denote by CGF
the subalgebra of {φe : e ∈ E(G)}, where ci,e = 1 for the end of e and
−1 for the beginning of e.
The original motivation of these algebras is that, for the complete
graph Kn , the algebra is isomorphic to the cohomology ring of the flag
POSTNIKOV-SHAPIRO ALGEBRAS AND MATROIDS
3
manifold F ln , see [23] and [24]. Algebras of complete graphs are also
related to Fomin-Kirillov and Orlik-Terao algebras, see their definition
in [11], [20] and relations between the last two in [18]. In paper [4] a big
class of algebras was considered, which includes our case, Orlik-Terao
and others. Postnikov-Shapiro algebras have direct connections with
G-parking functions (see [22], [9]), whose are stable configurations of
the sandpile model introduced in [10].
F
T
There are a few generalizations of BG
and BG
in the literature (the
latter algebra counts trees in G instead of forests, see notation in section 5), see the most important generalizations in papers [2] and [13].
In both papers the main object of the generalizations is the algebra
of the second definition (quotient algebra), they defined the so called
Zonotopal algebras, where dimensions of algebras are numbers of lat∆
tice points of corresponding zonotops. In [14] the definitions of BG
and
∆
CG were given in terms of some simplex complex ∆ on the set of verF
T
tices of G. With this definition algebras BG
and BG
become particular
cases corresponding to different simplexes.
We need the following classic notation, which was constructed by
Tutte for his original definition of the polynomial.
Notation 1.2. Fix some linear order on the set E(G) of edges of G.
Let F be any spanning forest in G. By actG (F ) denote the number of
all externally active edges of F , i.e. the number of edges e ∈ E(G) \ F
such that (i) subgraph F + e has a cycle; (ii) e is the minimal edge in
this cycle in the above linear order.
Denote by F + the set of edges of the forest F together with all externally active edges, and denote by F − = E(G) \ F + the set of externally
nonactive edges.
The following theorem is the main result of [22], which shows that
both definitions are equivalent and why it is called algebra counting
spanning forests.
F
Theorem 1.1 (cf. [22]). For any graph G, algebras BG
and CGF are
isomorphic, their total dimension over K is equal to the number of
spanning forests in G.
Moreover, the dimension of the k-th graded component of these algebras equals the number of spanning forests F of G with external activity
e(G) − e(F ) − k.
In fact, the second part of Theorem 1.1 claims that the above Hilbert
polynomial is a specialization of the Tutte polynomial of G.
Corollary 1.1. Given a graph G, the Hilbert series of the algebra CGF
is given by
1
HCGF (t) = TG 1 + t,
· te(G)−v(G)+c(G) ,
t
4
G. NENASHEV
where c(G) is the number of the connected components of G.
Both definitions of algebras are important, and sometimes it is easier
to use the first and sometimes the second. For example: if you want
to use computer algebra, then the first is better, because it has less
relations, but in fact, the first definition is subalgebra of the bigger
algebra ΦFG .
The structure of the paper is the following:
In § 2 we show that the Postnikov-Shapiro algebras counting spanning forests and graphical matroids are in one to one correspondence,
see Theorem 2.1. It means that these algebras store almost all information about the graphs.
In § 3 we consider t-labelled analouge of algebras, namely we use the
condition φt+1
= 0 instead φ2e = 0. The similar properties hold and a
e
new property that we can reconstruct the Tutte polynomial from the
Hilbert series of the algebra corresponding to large t, see Theorem 3.3.
In § 4 we construct a family of algebras corresponding to a given
hypergraph, and we present a natural definition of a matroid of a hypergraph such the Hilbert series of algebras is a specialization of the
Tutte polynomial corresponding to this matroid.
In § 5 we discuss similar problems for algebras counting spanning
trees and formulate Conjecture 1.
Acknowledgement. I am grateful to my supervisor Boris Shapiro
for introducing me to this area, for his comments and editorial help
with this text.
2. Algebras and Matroids
Obviously, the original Postnikov-Shapiro algebra corresponding to
a disconnected graph G is the Cartesian product of the algebras corresponding to the connected components of G. In particular, it is also
true for 2-connected components (maximal connected subgraphs such
that they remain connected, after removal of any vertex). The same
fact is also true for matroids. In this section we prove the following
result.
F
F
Theorem 2.1. Algebras BG
and BG
for graphs G1 and G2 are iso1
2
morphic if and only if the graphical matroids of these graphs coincide.
P.S. The algebraic isomorphism can be thought of either as graded or
as non-graded, the statement holds in both cases.
In the subsequent paper [19] filtered algebras are considered, that
distinguish graphs. The proof is based on some ideas from this paper
and on another trick, that makes a proof easier. Here we should proof
both sides and in the proof of Theorem 2.1 we use the following theorem
of H. Whitney.
POSTNIKOV-SHAPIRO ALGEBRAS AND MATROIDS
5
Theorem 2.2 (Whitney’s 2-isomorphism theorem, see [27], [21]). Let
G1 and G2 be two graphs. Then their graphical matroids are isomorphic
if and only if G1 can be transformed to a graph, which is isomorphic
to G2 by a sequence of operations of vertex identification, cleaving and
twisting.
These three operations are defined below.
1a) Identification: Let v and v ′ be vertices from different connected
components of the graph. We modify the graph by identifying v and
v ′ as a new vertex v ′′ .
1b) Cleaving (the inverse of identification): A graph can only be cleft
at a cut-vertex.
2) Twisting: Suppose that the graph G is obtained from two disjoint
graphs G1 and G2 by identifying vertices u1 of G1 and u2 of G2 as the
vertex u of G and additionally identifying vertices v1 of G1 and v2 of
G2 as the vertex v of G. In a twisting of G about {u, v}, we identify
u1 with v2 and u2 with v1 to get a new graph G′ .
We split our proof of Theorem 2.1 in two parts presented in § 2.1 and
in § 2.2.
2.1. Algebras are isomorphic if their matroids are isomorphic.
F
Because algebras BG
and CGF are isomorphic for any graph by Theorem 1.1, it suffices to prove Theorem 2.1 for algebras CGF 1 and CGF 2 .
Lemma 2.1. If graphs G and G′ differ by a sequence of Whitney’s
deformations, then the algebras CGF and CGF ′ are isomorphic.
Proof. It is sufficient to check the claim for each deformation separately.
1◦ Identification and Cleaving. We need to prove our fact only for
cleaving, because identification is the inverse of cleaving.
In this case algebras doesn’t change, because the linear subspace
defined by Xi for vertices doesn’t change. This holds, because if we
split a vertex k into k ′ and k ′′ , then in the new graph, Xk′ equals to
the minus sum of Xi corresponding to the vertices from its component
except k ′ (sum of all X. from one connected component is zero), i.e.
Xk′ belongs to the linear space <X1 , . . . , Xk−1 , Xk+1, . . . , Xn >. Similarly Xk′′ belongs to the linear space <X1 , . . . , Xk−1 , Xk+1, . . . , Xn >.
Hence, <X1 , . . . , Xk′ , Xk′′ , . . . , Xn > is a subspace of the linear space
<X1 , . . . , Xn >. The equation, Xk = Xk′ + Xk′′ implies that these linear spaces coincide.
2◦ Whitney’s deformation of the second kind. Define the digraph
G as the orientation of the graph G, where each arrow goes to the
"smallest" vertex (see Remark 1.1).
Let us make a twist of the vertices u and v. Let G1 and G2 be the
orientations of G1 and G2 corresponding to G. Let G′ be the orientation
of G′ corresponding to the gluing G1 and G2 with reversing each arrow
6
G. NENASHEV
from G2 . Vertex u′ in G′ is obtained by gluing of u1 and v2 ; v ′ is
obtained by gluing of v1 and u2 .
Let Xk , X1,k , X2,k and Xk′ be the sums with signs of edges incident
to vertex k in graphs G, G1 , G2 and G′ .
For a vertex k of G1 except u1 and v1 , we get
Xk = X1,k = Xk′ .
For a vertex k of G2 except u2 and v2 , we get
Xk = X2,k = −Xk′ , because we reverse the orientation in the second
part of twisting.
For other vertices we have:
Xu = X1,u1 + X2,u2 ;
Xv = X1,v1 + X2,v2 ;
Xu′ ′ = X1,u1 − X2,v2 = Xu − (X2,u2 + X2,v2 );
Xv′ ′ = X1,v1 − X2,u2 = Xv − (X2,u2 + X2,v2 ).
We know that the sum of variables corresponding to the vertices of any
graph is zero, because each edge goes with a plus to one vertex and
with a minus to another. We have
X
X2,k = 0,
X
k∈G2
X2,k + X2,u2 + X2,v2 = 0,
k∈G2 \u2 ,v2
X2,u2 + X2,v2 = −
X
X2,k = −
k∈G2 \{u2 ,v2 }
Xu′ ′
Xv′ ′
X
Xk .
k∈G2 \{u2 ,v2 }
belong to the linear space generated by Xk , where
and
Hence,
k ∈ G. In other words, the linear space for G′ is a linear subspace of
the space for G. Similarly we can prove the converse. Then, the linear
spaces coincide, and since we have the same relations (φ2e = 0 for any
edge), the algebras in these bases coincide.
We have proved the first part of Theorem 2.1, because if the corresponding graphical matroids are isomorphic, then there exists such a
sequence of Whitney’s operations.
Corollary 2.1. The algebra corresponding to a graph G is the Cartesian product of the algebras corresponding to the 2-connected components of G.
2.2. Reconstruction of the matroid.
Lemma 2.2. It is possible to reconstruct a matroid of a graph G from
the algebra CGF .
Remark 2.1. We assume that we only know CGF as an algebra. I.e.
we assume that we do not know the basis corresponding to the vertices
of G, and that we have no information about the graded components
of CGF .
POSTNIKOV-SHAPIRO ALGEBRAS AND MATROIDS
7
Proof. For an element Y ∈ CGF , we define its length ℓ(Y ) as the minimal
number such that Y ℓ+1 is zero (the length can be infinite).
WePcall an element Y ∈ CGF irreducible if there is no representation
Y = i Z2i−1 Z2i such that ℓ(Zj ) is finite for any j.
Consider a basis {Y1 , ..., Ym } of the algebra CGF with the following
conditions:
• Each Yj is irreducible;
• For any k 6 m, different nonzero numbers r1 , . . . , rk ∈ K, any
r1′ , . . . , rk′ ∈ K and different i1 , . . . , ik ∈ [m], we have
ℓ(r1 Yi1 + · · · + rk Yik ) > ℓ(r1′ Yi1 + · · · + rk′ Yik );
• For any linear combination Y of {Y1 , ..., Ym } and a reducible Z,
P
ℓ(Y ) 6 ℓ(Y + Z);
• i ℓ(Yi ) is minimal.
Such a basis of CGF always exists. For example, the basis X1 , . . . , Xn
(corresponding to the vertices) satisfies the first three conditions. However, the sum of lengths of Xi is not minimal.
For an element Y ∈ CGF , we define its linear part as Y and its remainder as Yb , where Y = Y + Yb , Y belongs to the 1-graded component of
CGF and Yb belongs to the linear span of the other graded components.
Observe that we do not know this decomposition explicitly, because we
do not know the 1-graded component. We say that an edge e belongs
to Y if Y includes the variable φe corresponding to the edge e with a
nonzero coefficient.
Proposition 2.3. A basis {Y1 , ..., Ym } as above satisfies additionally
the following conditions:
(1) The set {Y 1 , ..., Y m } is a basis;
(2) For any linear combination Y of {Y1 , ..., Ym },
ℓ(Y ) = ℓ(Y );
(3) Each edge belongs to one or two Y i and Y j . If it belongs to two,
then with the opposite coefficients;
(4) Each Y i has edges only from one 2-connected component.
Proof. (1). We know that {Y1 , ..., Ym } is a basis. Hence, the 1-graded
component of CGF coincides with a linear span of <Y 1 , ..., Y m >. Then,
<Y 1 , ..., Y m > is also a basis.
(2). For any i ∈ [m], the part Ybi is reducible, otherwise ℓ(Yi ) is
infinite, and the sum in the last condition is infinite. However for
the basis corresponding to the vertices this sum is finite. Then for
any linear combination Y of {Y1 , ..., Ym }, we have ℓ(Y ) 6 ℓ(Y ) < ∞
(by third condition of choice). However, it is clear that ℓ(Y ) > ℓ(Y ).
Hence, ℓ(Y ) = ℓ(Y ) for any Y from the linear space <Y1 , ..., Ym >.
8
G. NENASHEV
(3). Obviously, each edge belongs to at least one Y j , because
any edge belongs to some Xi and this Xi is a linear combination of
Y 1 , ..., Y m .
Assume that there is an edge e, which belongs to Y i1 , Y i2 and Y i3 .
Then there are different nonzero numbers r1 , r2 and r3 such that, for
Y = r1 Y1 + r2 Y2 + r3 Y3 , Y without e. Then ℓ(Y ) is at most the number
of all edges in Y i1 , Y i2 and Y i3 minus one.
Consider r1′ , r2′ and r3′ in general position, then, for Y ′ = r1′ Y1 +
r2′ Y2 + r3′ Y3 , ℓ(Y ′ ) is the number of all edges in Y i1 , Y i2 and Y i3 . We
have
ℓ(r1 Y1 + r2 Y2 + r3 Y3 ) < ℓ(r1′ Y1 + r2′ Y2 + r3′ Y3 ).
Using (2) we also have
ℓ(r1 Y1 + r2 Y2 + r3 Y3 ) < ℓ(r1′ Y1 + r2′ Y2 + r3′ Y3 ),
which contradicts our choice of the basis.
We proved the first part of condition (3); the proof of the second
part is the same, but only for two different r1 and r2 .
(4). Assume the opposite, i.e. that there is Y i which has edges
belonging to two different 2-connected components.
We know that the algebra is the Cartesian product of the subalgebras corresponding to 2-connected components. Then there are Z1
and Z2 from the 1-graded component, such that Y i = Z1 + Z2 and
ℓ(Z1 ), ℓ(Z2 ) < ℓ(Y i ).(For example, Z1 is a part corresponding to some
2-connected component and Z2 is another part).
Because Z1 and Z2 belong to the linear space <Y 1 , ..., Y m >, we
can change Y i to Z1 or Z2 in the basis {Y 1 , ..., Y m }. (Indeed if we
can not do it, then Z1 and Z2 belong to <Y 1 , . . . Y i−1 , Y i−1 , . . . Y m >.
Therefore Y i = Z1 + Z2 also belongs to the latter space). Let us
{Y 1 , . . . Y i−1 , Z1 , Y i−1 , . . . Y m } is also a basis.
We have a new basis {Y 1 , . . . Y i−1 , Z1 , Y i−1 , . . . Y m }, whose sum of
lengths is less than the sum of lengths of {Y 1 , . . . , Y m }, which is equal
to the sum of lengths of {Y1 , . . . , Ym }. And for this basis, the first three
conditions of a choice of basis hold, and the sum of lengths is smaller.
Then we should choose the basis {Y 1 , . . . Y i−1 , Z1 , Y i−1 , . . . Y m } instead
of {Y1 , . . . , Ym }. Contradiction.
Let us now construct the cut space of G. This will finish the proof,
because we can define the graphical matroid in terms of the cut space
of a graph. By a cut we mean a set C of edges such that the subgraph
G \ C has more connected components than G. By an elementary cut
we mean a minimal cut, i.e. a cut, whose arbitrary subset is not a cut.
POSTNIKOV-SHAPIRO ALGEBRAS AND MATROIDS
9
P i
P i
The sum
2 Yi =
2 Yi has each edge with a nonzero coefficient
by (2) of Proposition 2.3. Hence,
!
!
m
m
X
X
e(G) = ℓ
2i Y i = ℓ
2i Y i .
i=1
i=1
Therefore we know the number of edges in the graph.
Consider the set τ = {ψ1 , . . . , ψe(G) } consisting of e(G) elements and
a family of subsets K1 , . . . , Km constructed by the following rules.
ℓ(Y )+ℓ(Y )−ℓ(Y +Y )
j
i
j
own ele• For each pair i and j, we prescribe i
2
ments from τ and add them to both Zi and Zj ;
P
ℓ(Y )+ℓ(Yj )−ℓ(Yi +Yj )
• For every i, we choose ℓ(Yi ) − j6=i ( i
) own el2
ements from τ and add them to Zi .
In fact, for any edge e from G, we choose the corresponding element
ψk(e) and add it to Zi if and only if e belongs to Y i .
Consider the space Γ of subsets in τ with the operation ∆ (symmetric difference) generated by Z1 , . . . , Zm . We want to prove that Γ is
isomorphic to the cut space of G.
Let C be an elementary cut of G. Let XC be the sum of Xi corresponding to the vertices, which belong to some new component
of G \ C (hence, XC is in 1-graded component). Then XC has
an edge with a nonzero coefficient if and only if the edge belongs
to C. Consider the minimal t such that there is a linear combination XC = a1 Y i1 + . . . + at Y it ; consider the sum XC′ = Y i1 + . . . + Y it .
Obviously, XC′ is nonzero and has edges only from the cut C, because
an edge belongs to the sum XC′ if and only if it is exactly in one of
Y i1 , . . . , Y it .
Assume that XC′ has not all edges from XC . Let C ′ be a subset of
the edges of C belonging to XC′ . The set of edges C ′ is not a cut of
G, because C is an elementary cut. Hence, for any edge e ∈ C ′ , there
is a path e1 , . . . , ek in G \ C ′ , such that e, e1 , . . . , ek is a simple cycle.
Let the edge e connect vertices bk and b0 , and the edge ei connect bi−1
and bi P
for any i ∈ [k]. Because XC′ belongs to the 1-graded component,
′
XC = ni=1 abi Xi , where abi ∈ K and Xi are the elements corresponding
to the vertices of G. For i ∈ [k], the edge ei does not belong to XC′ ,
however variable φei belongs only to Xbi−1 and Xbi from {X1 , . . . , Xn }.
Furthermore it belongs to one of {Xbi−1 , Xbi } with coefficient 1 and to
another with −1, hence, b
abi−1 = b
abi . Then, we also have b
ab0 = b
abk .
′
Hence the variable φe belongs to XC with a zero coefficient. Contradiction.
We concluded that a subset of τ corresponding to an elementary cut
belongs to the space Γ. To finish the proof, we need to show, that if a
subset of τ belongs to Γ, then it either corresponds to a cut or to the
empty set.
10
G. NENASHEV
Assume the contrary, i.e. assume that there is Zi1 ∆Zi2 ∆ · · · Zis is
not a cut. Let C be a set of edges from Y i1 + Y i2 + · · · + Y is , then C is
not a cut in G. By Proposition 2.3 we can split the summation into the
summations inside individual connected components (and even inside
2-connected component). Then, for any connected component it is not
a cut.
Let Y ij1 + Y ij2 + · · · + Y ijr corresponds to a connected component
′
G , and has the set of edges C ′ . Then
X
av Xv .
Y ij1 + Y ij2 + · · · + Y ijr =
v∈V (G′ )
We know that G′ \ C ′ is connected. Therefore there is a spanning tree
T in G′ \ C ′ . For any edge P
vi vj from T , we have avi = avj ; otherwise the edge vi vj belongs to v∈V (G′ ) av Xv with a nonzero coefficient.
′
Since
P T is a spanningPtree of G , all coefficients av are the same. Thus
v∈V (G′ ) av Xv = a(
v∈V (G′ ) Xv ) = 0; the last sum is zero, because
the sum of variables corresponding to vertices from a connected component is zero. Then we also have Y i1 + Y i2 + · · · + Y is = 0, hence,
Zi1 ∆Zi2 ∆ · · · Zis is the empty set.
Therefore the space Γ is isomorphic to the cut space of G, i.e. there
is a unique graphical matroid corresponding to CGF .
3. Algebras associated with t-labelled forests
In this section we substitute the square-free algebra ΦFG for the (t+1)free algebra ΦFGt .
Notation 3.1. Take an undirected graph G on n vertices. Let t > 0
be a positive integer.
(I) Let ΦFGt be a commutative algebra over K generated by {φe : e ∈
E(G)} satisfying the relations φt+1
= 0, for any e ∈ E(G).
e
Fix any linear order on the vertices of G. For i = 1, . . . , n, set
X
Xi =
ci,e φe ,
e∈G
where ci,e as in the notation 1.1. Denote by CGFt the subalgebra of ΦFGt
generated by X1 , . . . , Xn .
(II) Consider the ideal JGFt in the ring K[x1 , · · · , xn ] generated by
!tDI +1
X
xi
pFI t =
,
i∈I
where I ranges over all nonempty subsets of vertices, and DI is the total
number of edges between vertices in I and vertices outside I. Define
Ft
as the quotient K[x1 , . . . , xn ]/JGFt .
the algebra BG
POSTNIKOV-SHAPIRO ALGEBRAS AND MATROIDS
11
It enumerates the so-called t-labelled spanning forests.
Consider a finite labelling set {1, 2, . . . , t} containing t different labels.
Definition 3.2. A spanning forest of a graph G with a label from
{1, 2, . . . , t} on each edge is called a t-labelled forest. The weight of
a t-labelled forest F , denoted by ω(F ), is the sum of the labels of all its
edges.
Ft
Theorem 3.1. For any graph G and a positive integer t, algebras BG
and CGFt are isomorphic. Their total dimension over K is equal to the
number of t-labelled forests in G.
Ft
is
The dimension of the k-th graded component of the algebra BG
equal to the number of t-labelled forests F of G with the weight t ·
(e(G) − actG (F )) − k.
b the graph on n vertices and with t · e(G) edges
Proof. Denote by G
b In
such that each edge of G corresponds to its t clones in the graph G.
other words, each edge of G is substituted by its t copies with labels
b are
1, 2, . . . , t. For each edge e ∈ E(G), its clones e1 , . . . , et ∈ E(G)
ordered according to their numbers; clones of different edges have the
same linear order as the original edges. Thus we obtain a linear order
b
of the edges of G.
Consider the following bijection between t-labelled forests in G and
b each t-labelled forest F ⊂ G corresponds to the
usual forests in G:
′
b such that for each edge e ∈ E(F ), the forest F ′ has the
forest F ⊂ G,
clone of the edge e whose number is identical to the label of the edge
e in the forest F .
Obviously,
actGb (F ′ ) = t · actG (F ) + ω(F ) − e(F ),
b = t · e(G). Since BFt and BF are the same, the Hilbert series
and e(G)
G
b
G
Ft
F
of the algebra BG coincides with the Hilbert series of the algebra BG
b,
which settles the second part of Theorem 3.1.
Ft
F
and BG
To prove the first part of this Theorem, observe that BG
b are
F
F
isomorphic, and algebras CGb and BGb are isomorphic. Thus we must
show that algebras CGFb and CGFt are isomorphic. This is indeed true,
because for every edge e ∈ E(G), the elements φe , . . . , φte are linearly
independent in the algebra ΦFGt with coefficients containing no φe . Also
elements (φe1 + · · · + φet ), . . . , (φe1 + · · · + φet )t are linearly independent
in the algebra ΦFGb with coefficients containing no φe1 , . . . , φet , and (φe1 +
· · · + φet )t+1 = 0. Moreover elements φei only occur in the sum (φe1 +
· · · + φet ) in the algebra ΦFGb .
12
G. NENASHEV
Ft
was calculated in papers [2] and [17].
In fact Hilbert Series of BG
b where each edge
Furthermore, the Hilbert Series was computed for G,
is replaced by its own number of edges. The Hilbert series is a specialization of the multivariate Tutte polynomial (see definition in [25]).
When "t" is the same for every edge, the multivariate Tutte polynomial is calculated from the usual Tutte polynomial. So in our case the
Ft
Hilbert Series of BG
is a specialization of the Tutte polynomial of G.
Ft
is
Theorem 3.2. The dimension of the k-th graded component of BG
equal to the coefficient of the monomial y t·e(G)−v(G)+c(G)−k in the polynomial
t
v(G)−c(G)
t+1
y −1
y −1 t
,y .
· TG
y−1
y t+1 − y
b constructed in the proof of Theorem 3.1. We
Consider the graph G
need the following technical lemma which was proved in [5].
Lemma 3.1 (Lemma 6.3.24 in [5]).
v(G)−c(G)
t
t
y − y + x(y − 1) t
y −1
,y .
· TG
TGb (x, y) =
y−1
yt − 1
After the substitution x → 1 + y1 , we get the following equality.
Corollary 3.1.
v(G)−c(G)
t
t+1
y −1
y −1 t
1
,y .
· TG
TGb 1 + , y =
y
y−1
y t+1 − y
Ft
F
is isomorphic to the algebra BG
Proof of Theorem 3.2. Algebra BG
b
(which was shown in the proof of Theorem 3.1), furthermore they are
isomorphic as graded algebras. So it is enough to show that dimenF
sion of the k-th graded component of BG
b is equal to the coefficient of
b
b
b
the monomial y t·e(G)−v(G)+c(G)−k = y e(G)−v(G)+c(G)−k in the polynomial
TGb (1 + y1 , y). This fact is true by Corollary 1.1.
Theorem 3.3. For any positive integer t > n, it is possible to restore
the Tutte polynomial of any connected graph G on n vertices knowing
Ft
.
only the dimensions of each graded component of the algebra BG
Proof. By Theorem 3.1 we know that the degree of the maximal
Ft
nonempty graded component of BG
is equal to the maximum of
t · (e(G) − actG (F )) − ω(F ) taken over F . It attains its maximal value
for the empty forest (i.e. F = ∅). Then we know the value of t · e(G),
and hence, we know the number of edges of the graph G.
Since we know that t · e(G) − v(G) + c(G) = t · e(G) − n + 1 (G
is connected, i.e., c(G) = 1), by Theorem 3.2 we can calculate the
polynomial
t
v(G)−c(G)
t+1
y −1
y −1 t
,y
· TG
y−1
y t+1 − y
POSTNIKOV-SHAPIRO ALGEBRAS AND MATROIDS
13
t+1
−1
from the Hilbert series. Then we can also calculate TG yyt+1 −y
, yt .
It is well P
known that for any graph G, its Tutte polynomial TG (x, y)
is equal to F (x − 1)c(F )−c(G) y actG (F ) , where the summation is taken
over all spanning forests of G. Then we obtain
TG
y t+1 − 1 t
,y
y t+1 − y
=
X
F
=
X y t+1 − 1
F
y t+1 − y
1
t−1
y(y + · · · + 1)
−1
n−1−e(F )
n−1−e(F )
y t·actG (F ) =
y t·actG (F ) .
Hence, we can restore the polynomial
X
e(F ) t·act (F )+e(F )
y t−1 + · · · + 1
y G
.
(∗)
F
Since e(F ) < t, we can compute the number of usual spanning forests
with a fixed pair of parameters e(F ) and actG (F ). Indeed, consider the
monomial of the minimal degree in the polynomial (∗), and represent
it in the form s · y m . Observe that s is the number
of spanning forests
F such that F ≡ m (mod t) and actG (F ) = mt . Remove from the
polynomial (∗) all summands for these spanning forests, and repeat
this operation until we get 0. P
Note again that TG (x, y) = a,b #{F : e(F ) = a, act(F ) = b} · (x −
n−1−a
1)
· y b. Therefore we know the whole Tutte polynomial of G, since
we know the number of usual spanning forests with any fixed number
of edges and any fixed external activity.
4. Vector configurations and Hypergraphs
4.1. Algebra corresponding to vector configuration. The following algebra was introduced by A. Postnikov, B. Shapiro and M. Shapiro
in [23].
Notation 4.1. Given a finite set A = {a1 , . . . , am } of vectors in Kn ,
let ΦFm be the commutative algebra over K generated by {φi : i ∈ [m]}
with relations φ2i = 0, for each
P i ∈ [m].
For i = 1, . . . , n, set Xi = k∈[m] ak,i φk . Denote by CA the subalgebra
of ΦFm generated by X1 , . . . , Xn .
The Hilbert series of CA also corresponds to a specialization of the
Tutte polynomial of the corresponding vector matroid, see Theorem 3
in [23].
Theorem 4.1 (cf. [23]). The dimension of algebra CA is equal to the
number of independent subsets in V . Moreover, the dimension of the
14
G. NENASHEV
k-th graded component is equal to the number of independent subsets
S ⊂ A such that k = m − |S| − act(S).
Corollary 4.1. Given a vector configuration A in Kn , the Hilbert series
of algebra CA is
1
HCA (t) = TA 1 + t,
· t|A|−rk(A),
t
where TA is the Tutte polynomial corresponding to A, |A| is the number
of vectors in the configuration and rk(A) is the dimension of the linear
span of these vectors.
The set of different vector configurations depends on continuous parameters. Additionally, there are uncountably many non-isomorphic
algebras each corresponding to its vector configuration. At the same
time the number of matroids is countable, and it is finite for a fixed
number of vectors. It means that there are many different vector configurations with the same corresponding matroid. In other words, it
is in principle impossible to reconstruct a vector configuration and its
algebra from the corresponding matroid.
4.2. Hypergraphs. In this subsection we present a family of algebras
corresponding to a hypergraph. Almost all algebras from this family
(generic algebras) have the same Hilbert series and this generic Hilbert
series counts forests of this hypergraph. There are many definitions of
spanning trees of a hypergraph, for example: a spanning cacti in [1]; a
hypertree in [7] (also known as an arboreal hypergraph in [3]). However all these definitions allow trees to have different number of edges,
whence spanning tree of a usual graph should have the same number
of edges. We define spanning trees such that this property holds and
also other natural properties hold. Also we define the hypergraphical
matroid and the corresponding Tutte polynomial, whose points T (2, 1)
and T (1, 1) calculate the numbers of spanning forests and of spanning
trees, resp. Similar definition of spanning trees and forests was presented in [15], for that definition there is also Tutte polynomial for a
hypergraph, however, there is no matroid.
First we define the family of algebras.
Given a hypergraph H on n vertices, let us associate commuting
variables φe , e ∈ H to all edges of H.
Set ΦH be the algebra generated by {φe : e ∈ H} with relations
2
φe = 0, for any e ∈ H.
Define C = {ci,e ∈ K : i ∈ [n], e ∈ H} as a set of parameters of
H, for any edge e ∈ H, ci,e = 0 for vertices non-incident to e, and
P
n
i=1 ci,e = 0.
POSTNIKOV-SHAPIRO ALGEBRAS AND MATROIDS
For i = 1, . . . , n, set
Xi =
X
15
ci,e φe ,
e∈H
F
CH(C)
Denote by
the subalgebra of ΦH generated by X1 , . . . , Xn , and
F
b
denote by CH the family of such subalgebras.
The following trivial properties hold for this family of algebras.
Proposition 4.2.
P(I) For a hypergraph H, the dimension of the space
of parameters is e∈E (|e| − 1).
(II) Given a set of parameters C and non-zero numbers ae , e ∈ E,
let C ′ be the set of parameters such that c′i,e = ae ci,e for any i ∈ [n] and
e ∈ E. Then the subalgebras for C and for C ′ are isomorphic.
Corollary 4.2. For a usual graph G, almost all algebras from CbG are
isomorphic to CG .
We define a hypergraphical matroid using the definition of an independent set of edges of a hypergraph.
Definition 4.2. Let H be a hypergraph on n vertices. A set F of
edges is called independent if there is a set of parameters C of H, such
that vectors corresponding to edges from F are linearly independent. In
other words, F is independent if, for a generic set of parameters of H,
vectors are linearly independent. Define the hypergraphical matroid of
H as the matroid with the ground set E(H).
There is a combinatorial definition of an independent set of edges.
First we need to define a cycle of H.
Definition 4.3. A subset of edges C ⊂ E is called a cycle if
• |C| = | ∪e∈C e|
• There is no subset |C ′ ⊂ C|, such that the first property holds
for C ′ .
Definitions of dependence and of a cycle are similar.
Theorem 4.3. A subset of edges X ⊂ E is dependent if and only if
there is a cycle C ⊂ X.
We present a proof of this theorem after Theorem 4.6.
Definition 4.4. A set of edges F is called a spanning forest if F has
no cycles, in other words, F is forest if and only if F is an independent
set (by Theorem 4.3). A set of edges T ⊂ H is called a spanning tree
if it is a forest and T has exactly v(H) − 1 edges.
A hypergraph H is called strongly connected if it has at least one
spanning tree.
Proposition 4.4. Maximal spanning forests of a hypergraph have the
same number of edges. In fact, if H = (V, E) is a strongly connected
16
G. NENASHEV
hypergraph, then for any spanning forest F ⊂ E there is a spanning
tree T which contains F (i.e. F ⊂ T ⊂ E).
Proof. By Theorem 4.3 we know that a spanning forest is the same as
an independent set of edges. Then we can add edges to a forest until the
number of edges is less than the dimension of the linear space. Then all
maximal spanning forests have the same size. Hence, if a hypergraph
is strongly connected, then any spanning forest is contained in some
spanning tree.
F
The Hilbert series of algebras in CbH
are also counting forests of H.
F
Theorem 4.5. For a hypergraph H, generic algebras from CbH
have the
same Hilbert series. The dimension of the k-th graded component of a
generic algebra equals the number of spanning forests F in H with the
external activity e(H) − e(F ) − k.
Proof. By Theorem 4.3 we can change the definition of spanning forests
to independent sets. Consider a generic set of parameters C. By TheF
and it is the same for all
orem 4.1 we know the Hilbert series of CH(C)
generic sets of parameters.
We define the Tutte polynomial of H as the Tutte polynomial of
the corresponding hypergraphical matroid. By the theorems above we
know that
• TH (2, 1) is the number of spanning forests;
• TH (1, 1) is the number of maximal spanning forests. In fact,
TH (1, 1) is the number of spanning trees if H is srongly connected.
By Theorem 4.5, we get that a generic Hilbert series is a specialization of the Tutte polynomial of H.
Corollary 4.3. Given a hypergraph H and its generic set of parameters
F
C, the Hilbert series of the algebra CH(C)
is given by
1
HCH (C)F (t) = TG 1 + t,
· te(H)−rkH ,
t
where rkH is the size of a maximal spanning forest of H.
There is another definition of forests/trees of H, which again shows
that it is a generalization of forests/trees of a usual graph.
Theorem 4.6. A subset of edges X ⊂ E is a forest (tree) if and only
if there is map from edges to pairs: ek → (i, j), where vi , vj ∈ ek , such
that these pairs form a forest (tree) in the complete graph Kn .
Proof. Consider our forest F and the hypergraph H, add to them n −
1 − e(F ) full edges, i.e. edges of type V . We get a new hypergraph H ′
and a subset of edges F ′ . It is clear that F ′ is a tree, because there is
POSTNIKOV-SHAPIRO ALGEBRAS AND MATROIDS
17
no cycle without new edges, and with at least one new edge we need
to cover all vertices and their number is bigger than number of edges.
So let F ′ = T and we will prove our Theorem for a spanning tree.
Consider the bipartite graph B with two sets of vertices: the first
set are edges of T and the second are vertices V \ v1 , where there is an
edge between ei and vj if and only if vj ∈ ei .
There is a perfect matching in B, because we can use Hall’s marriage
theorem (see [12] or any classical book). We know that e(T ) = |V \
v1 | and, for any X ⊂ T , they cover at least e(X) vertices (otherwise
these edges cover these vertices and may be v1 and then X has cycle).
Consider a bijection f from T to V \ v1 constructed by this perfect
matching. Now we construct by recursion a map g from T to pairs of
vertices:
(1) A := {v1 } and B := T
(2) repeat until A 6= V :
• choose the minimal edge ei from B such that ei ∩ A 6= ∅
• g(ei ) = (u, f (ei )), where u ∈ ei ∩ A
• A := A ∪ {f (ei )} and B := B \ {ei }
It works, otherwise we can not chose such an edge ei , then either B = ∅
or B 6= ∅ at this moment. We know that |A| + |B| = n, then in the
first case we already have A = V ; in the second case edges from B have
vertices only from V \ A, then there is a cycle on these edges, i.e. T is
not a tree. Then this algorithm gives some usual tree.
Proof of Theorem 4.3. Assume the contrary, then there is a subset X ⊂
E, which is dependent and without cycles, i.e. X is a spanning forest.
By Theorem 4.6 we know that there is map g from X to the pairs of
vertices, which gives a usual forest. Consider the vector set
{ae := zg1 (e) − zg2 (e) , e ∈ X},
where zv = (0, . . . , 0, 1, 0, . . . , 0) is the unit vertex correspinding to the
vertex v. Since g gives the usual forest, these vectors are independent.
Hence, generic vectors {be , e ∈ X} are also linear independent. We
get that the edges are independent, contradiction.
By the induced subgraph on vertices V ′ ⊂ V , we assume a hypergraph (V ′ , E ′ ), where E ′ are all edges of E, which have vertices only
from V ′ (i.e. e ∈ E ′ if e ⊂ V ′ ). This definition works well with colorings of hypergraphs, because if we want to color a hypergraph such
a way that there are no monochromatic edges, then it is the same as
splitting vertices into sets with empty induced subgraphs. Also this
definition works well with standard sense of connectivity.
Proposition 4.7. Let V1 and V2 be the subsets of vertices such that the
induced subgraphs of H on Vj are strongly connected and V1 ∩ V2 6= ∅.
Then the induced subgraph of H on V1 ∪ V2 is also strongly connected
18
G. NENASHEV
Proof. For any vertex vi ∈ V we will consider the corresponding unit
vector zvi = (0, . . . , 0, 1, 0, . . . , 0). Fix a vertex u, which lie in the intersection V1 ∩ V2 . Let H1 and H2 be induced subgraphs on V1 and V2 ,
resp.
Consider vectors bi corresponding to ei ∈ E1 . We know that there is
a spanning tree of the graph H1 , then the dimension of the linear space
of such vectors is |V1 | − 1, furthermore any sum of coordinates of any
vector is zero. Hence, for any v ∈ V1 , zv − zu ∈ span{bi : ei ∈ E1 }.
Similarly we get the same for H2 and, hence, we have the same for
H1 ∪ H2 .
We get that for any v ∈ V1 ∪ V2 , zv − zu ∈ span{bi : ei ∈ E1 ∪ E2 }.
Hence, the hypergraph H1 ∪ H2 has |V1 ∪ V2 | − 1 independent edges,
then has a spanning tree. We have that the induced subgraph of H
on vertices V1 ∪ V2 is strongly connected, since it has all edges from
H1 ∪ H2 .
5. Algebras corresponding to spanning trees, Problems
In this section we discuss analogous algebras counting spanning trees.
T
Recall the definition of algebras BG
and CGT borrowed from [22].
Notation 5.1. Take an undirected graph G with n vertices.
(I) Let ΦTG be the commutative algebra over Q
K generated by {φe : e ∈
2
G} with relations φe = 0, for any e ∈ G, and e∈H φe = 0, for any cut
H ⊂ E(G).
Fix a linear order of vertices of G. For i = 1, . . . , n, set
X
Xi =
ci,e φe ,
e∈E(G)
where ci,e as in Notation 1.1. Denote by CGT the subalgebra of ΦTG generated by X1 , . . . , Xn .
(II) Consider the ideal JGT in the ring K[x1 , · · · , xn ] generated by
p[n] = x1 + . . . + xn
and by
pTI =
X
i∈I
xi
! DI
,
where I ranges over all nonempty proper subsets of vertices, and DI is
the total number of edges between vertices in I and vertices outside I.
T
Define the algebra BG
as the quotient K[x1 , . . . , xn ]/JGT .
The case of disconnected graphs is not interesting, because both
algebras are trivial. In the paper [22] the following result was proved:
POSTNIKOV-SHAPIRO ALGEBRAS AND MATROIDS
19
T
Theorem 5.1 (cf. [22]). For any connected graph G, algebras BG
and
T
CG are isomorphic; their total dimension over K is equal to the number
of spanning trees in G.
Moreover, the dimension of the k-th graded component of these algebras equals the number of spanning trees of G with external activity
e(G) − v(G) + 1 − k.
Corollary 5.1. Given a connected graph G, the Hilbert series of the
algebra CGT is given by
1
HCGT (t) = TG 1,
· te(G)−v(G)+c(G) ,
t
where c(G) is the number of connected components of G.
5.1. Algebras and matroids. For graph G, we define its bridge-free
matroid as the usual graphical matroid of the graph G′ which is obtained from G after removing all its bridges.
Proposition 5.2. For any pair of connected graphs G1 and G2 with
T
T
isomorphic bridge-free matroids, their algebras BG
and BG
are iso1
2
morphic.
Proof. Notice that, if we add an edge e and a vertex v to G, such that
v is an endpoint of e and another endpoint of e is some vertex of G,
then algebra BT does not change (this is obvious because e is a bridge,
and hence, φe is one of the generators of the ideal). This operation
doesn’t change bridge-free matroid. Therefore it is enough to prove
Proposition 5.2 only for graphs with the same number of edges.
Assume that |E(G1 )| = |E(G2 )|. In this case an isomorphism of
bridge-free matroids is equivalent to an isomorphism of matroids of
graphs G1 and G2 .
In fact, in Lemma 2.1 we construct orientations G1 and G2 of graphs
G1 and G2 on the same set of edges E(G1 ) = E(G2 ) (it was constructed
for graphs differ in one Whitney’s deformation, so we can extend it to
a sequence of deformations), such that they give the same graphical
matroid on edges and with these orientations the algebras CGF 1 and CGF 2
coincide as subalgebras of ΦFGi (ΦFG1 and ΦFG2 are the same, because
graphs have common set of edges).
Let I be the ideal generated by the products of edges from the cuts
of G1 in ΦFG1 . Because the variables on edges in G1 and G2 are the
same and C is a cut in G1 if and only if C is a cut in G2 , then I is also
the ideal generated by the cuts of G2 .
Thus we have ΦTG1 = ΦFG1 /I, hence,
CGT 1 = CGF 1 /I,
similarly
CGT 2 = CGF 2 /I.
20
G. NENASHEV
It means that the algebras CGT 1 and CGT 2 are also the same in orientations
G1 and G2 .
We formulate the following converse conjecture.
T
T
Conjecture 1. Algebras BG
and BG
for the connected graphs G1
1
2
and G2 are isomorphic if and only if their bridge-free matroids are
isomorphic.
5.2. t-labelled trees. It is possible to introduce similar algebras which
Tt
enumerate t-labelled trees, but it is not veryPexciting. Let BG
be an
DI
algebra in which we change the generators ( i∈I xi ) of the ideal by
P
( i∈I xi )tDI . The definition of CGTt will change in a more complicated
way.
However, there is no result about a reconstruction of the Tutte polynomial from the Hilbert series, because all trees have the same number
of edges and then HBTt (x) = (1 + x)n−1 HBGT (xt ). In other words, the
G
Tt
T
and of BG
contain the same information about the
Hilbert series of BG
graph.
5.3. Algebras for hypergraphs. The main problem is to construct
T
a family CbH
of algebras, which count spanning trees of H.
By paper [23], for a hypergraph H and a set of parameters C, we can
F
F
. We can consider the
as a quotient algebra, i.e, as BH(C)
present CH(C)
F
T
algebra BH(C) , which is obtained from BH(C) by changing the powers of
the generators of the ideal (writing always one less). By paper [2] algeT
bra BH(C)
should count spanning trees of H. However at this moment,
T
we can not present ΦTH such that its generic subalgebra CH(C)
counts
spanning trees of H.
Probably, we need to add to ΦFH relations corresponding to cuts,
where a cut is a subset of edges such that without it H has no spanning
trees. However, we need to prove it and if we want to do something
similar to the proof of Theorem 5.1, then we need to define H-parking
functions for a hypergraph.
References
[1] A. Abdesselam, The Grassmann-Berezin calculus and theorems of the matrixtree type, Advances in Applied Mathematics 33(1) (2004), pp. 51-70.
[2] F. Ardila, A. Postnikov, Combinatorics and geometry of power ideals, Trans.
Amer. Math. Soc., 362(8) (2010), pp. 4357-4384.
[3] C. Berge, Hypergraphs, North-Holland Mathematical Library 45, 1989.
[4] A. Berget, Products of linear forms and Tutte polynomials, European J. Combin.
31 no. 7 (2010), pp. 1924-1935.
[5] T. Brylawski, J. Oxley, The Tutte polynomial and its applications, chapter in
Matroid Applications, Encyclopedia of Mathematics and its Applications 40
(1992).
POSTNIKOV-SHAPIRO ALGEBRAS AND MATROIDS
21
[6] Y. Burman, A. Ploskonosov, A. Trofimova, Matrix-tree theorems and discrete
path integration, Linear Algebra and its Applications, 466 (2014), pp. 64-82.
[7] A. Brandstadt, F. Dragan, V. Chepoi, V. Voloshin, Dually chordal graphs, SIAM
Journal on Discrete Mathematics 11(3) (1998), pp. 437-455.
[8] S. Chaiken, D. J. Kleitman, Matrix Tree Theorems, Journal of Combinatorial
Theory, Series A 24 (1978), pp. 377-381.
[9] D. Chebikin, P. Pylyavskyy, A family of bijections between G-parking functions
and spanning trees, Journal of Combinatorial Theory, Series A 110 (2005), pp.
31-41.
[10] D. Dhar, Self-organised critical state of the sandpile automaton models, Physical Review Letters 64 no. 14 (1990), 1613-1616.
[11] S. Fomin, A. N. Kirillov, Quadratic algebras, Dunkl elements, and Schubert calculus, Advances in geometry, volume 172 of Progr. Math. (1999), pp. 147-182.
[12] P. Hall, On Representatives of Subsets, J.London Math. Soc. 10(1) (1935), pp.
26-30.
[13] O. Holtz, A. Ron, Zonotopal algebra, Advances in Mathematics, 227 (2011), pp.
847-894.
[14] B. Huang, Monomization of Power Ideals and Generalized Parking Functions,
https://math.mit.edu/research/highschool/primes/materials/2014/Huang.pdf.
[15] T. Kálmán, A version of Tutte’s polynomial for hypergraphs, Advances in Mathematics 244 (2013), pp. 823-873.
[16] G. Kirchhoff, Uber die Auflösung der Gleichungen, auf welche man bei der
untersuchung der linearen verteilung galvanischer Ströme geführt wird, Ann.
Phys. Chem. 72 (1847), pp. 497-508.
[17] M. Lenz, Hierarchical Zonotopal Power Ideals, European Journal of Combinatorics 33 (2012), pp. 1120-1141.
[18] R. I. Liu, On the commutative quotient of Fomin-Kirillov algebras, European
Journal of Combinatorics Volume 54 (2016), pp. 65-75.
[19] G. Nenashev, B. Shapiro, "K-theoretic" analog of Postnikov-Shapiro algebra
distinguishes graphs, Journal of Combinatorial Theory, Series A, 148 (2017),
pp. 316-332.
[20] P. Orlik, H. Terao, Commutative algebras for arrangements, Nagoya Math. J.
134 (1994), pp. 65-73.
[21] J. G. Oxley, Matroid theory, Oxford University Press, New York (1992).
[22] A. Postnikov, B. Shapiro, Trees, parking functions, syzygies, and deformations
of monomial ideals, Trans. Amer. Math. Soc. 356 (2004), pp. 3109-3142.
[23] A. Postnikov, B. Shapiro, M. Shapiro, Algebra of curvature forms on homogeneous manifolds, AMS Trans. Ser 2, vol. 194 (1999), pp. 227-235.
[24] B. Shapiro, M. Shapiro, On ring generated by Chern 2-forms on SLn /B, C. R.
Acad. Sci. Paris Sér. I Math. 326(1) (1998), pp. 75-80.
[25] A. Sokal, The multivariate Tutte polynomial (alias Potts model) for graphs and
matroids, In Surveys in Combinatorics (2005), pp. 173-226.
[26] W. T. Tutte, Graph Theory, Cambridge University Press (2001).
[27] H. Whitney, 2-isomorphic graphs, American Journal of Mathematics 55 (1933),
pp. 245-254.
Department of Mathematics, Stockholm University, SE-106 91 Stockholm, Sweden,
E-mail address: [email protected]
| 0math.AC
|
Efficient Black-Box Reductions for Separable Cost Sharing
Tobias Harks∗
Martin Hoefer†
Anja Huber‡
Manuel Surek§
Abstract
arXiv:1802.10351v1 [cs.GT] 28 Feb 2018
In cost sharing games with delays, a set of agents jointly allocates a finite subset of resources.
Each resource has a fixed cost that has to be shared by the players, and each agent has a nonshareable player-specific delay for each resource. A prominent example is uncapacitated facility
location (UFL), where facilities need to be opened (at a shareable cost) and clients want to
connect to opened facilities. Each client pays a cost share and his non-shareable physical connection cost. Given any profile of subsets allocated by the agents, a separable cost sharing protocol
determines cost shares that satisfy budget balance on every resource and separability over the
resources. Moreover, a separable protocol guarantees existence of pure Nash equilibria in the
induced strategic game for the agents.
In this paper, we study separable cost sharing protocols in several general combinatorial domains. We provide black-box reductions to reduce the design of a separable cost-sharing protocol
to the design of an approximation algorithm for the underlying cost minimization problem. In
this way, we obtain new separable cost-sharing protocols in games based on arbitrary playerspecific matroids, single-source connection games without delays, and connection games on nseries-parallel graphs with delays. All these reductions are efficiently computable – given an
initial allocation profile, we obtain a cheaper profile and separable cost shares turning the profile
into a pure Nash equilibrium. Hence, in these domains any approximation algorithm can be used
to obtain a separable cost sharing protocol with a price of stability bounded by the approximation
factor.
∗
Universität Augsburg, Institut für Mathematik, Germany. [email protected]
Goethe University Frankfurt, Institute of Computer Science, Germany. [email protected]
‡
Universität Augsburg, Institut für Mathematik, Germany. [email protected]
§
Universität Augsburg, Institut für Mathematik, Germany. [email protected]
†
1
1
Introduction
Cost sharing is a fundamental task in networks with strategic agents and has attracted a large amount
of interest in algorithmic game theory. Traditionally, cost sharing has been studied in a cooperative
sense, i.e., in the form of cooperative games or mechanism design. Many of these approaches treat
cost in a non-separable way and return a single, global cost share for each agent. In contrast,
when agents jointly design a resource infrastructure in large networks, it is much more desirable
to provide algorithms and protocols for separable cost sharing that specify which agent needs to
pay how much to each resource. Here the natural approach are strategic cost sharing games with
n players that allocate subsets of m resources. Each resource generates a cost depending on the
subset of players allocating it. A protocol determines a cost share for each resource and each player
using it. In addition to separability, there are further natural desiderata for such protocols, such as
budget-balance (distribute exactly the arising cost of each resource) and existence of a pure Nash
equilibrium (PNE), i.e., allow the resulting game to stabilize.
Perhaps the most prominent such protocol is the fair-share protocol, in which the cost of each
resource is allocated in equal shares to the players using it. This approach has been studied intensively
(see our discussion below), but there are several significant drawbacks. It can be PLS-hard to find [50]
a PNE, even in connection games on undirected networks. The price of stability (PoS), i.e., the total
cost of the best Nash equilibrium compared to the cost of the optimal allocation, can be as large as
Ω(log n) [5, 17], even though much better solutions can often be found in polynomial time.
In this paper, we study a slight generalization of cost sharing games, where every resource has a
shareable cost component and a non-shareable player-specific delay component. The shareable cost
needs to be shared by the players using it, the non-shareable player-specific delay represents, e.g.,
a physical delay and is thus unavoidable. This setting arises in several relevant scenarios, such as
uncapacitated facility location (UFL) [36]. Here players share the monetary cost of opened facilities
but additionally experience delays measured by the distance to the closest open facility. Another
important example appears in network design, where players jointly buy edges of a graph to connect
their terminals. Besides the monetary cost for buying edges, each player experiences player-specific
delays on the chosen paths. In such a distributed network environment, it is not clear a priori if an
optimal solution can be stable – i.e., if the shareable costs can be distributed among the players in
a separable way so that players do not want to deviate from it. This question leads directly to the
design of protocols that distribute the costs in order to induce stable and good-quality solutions of
the resulting strategic game.
Our results are three polynomial-time black-box reductions for the price of stability of separable
cost sharing protocols in combinatorial resource allocation problems. Our domains represent broad
generalizations of UFL – arbitrary, player-specific matroids, single-source connection games without
delays, and connection games on undirected n-series-parallel graphs with delays. In each of these
domains, we take as input an arbitrary profile and efficiently turn it into a cheaper profile and a
sharing of the shareable costs such that it is a Nash equilibrium. Our protocols are polynomial-time
in several ways. Firstly, the games we study are succinctly represented. In matroids, we assume that
strategies are represented implicitly via an independence oracle. For connection games on graphs,
the strategies of each player are a set of paths, which is implicitly specified by terminal vertices of
the player and the graph structure. The cost sharing protocol is represented by a strategy profile S
and a sharing of the shareable costs arising in S on each resource. While in principle the protocol
must specify a sharing of the costs for all of the other (possibly exponentially many) strategy profiles,
one can do so implicitly by a simple lexicographic assignment rule. It guarantees that the profile S
becomes a PNE. As such, starting from an arbitrary initial profile S ′ , we can give in polynomial time
the Nash equilibrium profile S, the cost shares for S, and the assignment rule for cost shares in the
other profiles. Hence, if S ′ is polynomial-time computable, then both protocol and Nash equilibrium
S are both polynomial-time computable and polynomial-space representable.
2
1.1
Our Results
We present several new polynomial-time black-box reductions for separable cost sharing protocols
with small price of stability (PoS). We study three domains that represent broad generalizations of the
uncapacitated facility location problem. In each domain, we devise an efficient black-box reduction
that takes as input an arbitrary strategy profile and computes a new profile of lower cost together
with a separable cost sharing protocol inducing the cheaper profile as a PNE. Thus, any polynomialtime α-approximation of the social cost can be turned into a separable cost sharing protocol with
PoS at most α.
Matroidal Set Systems. In Section 3 we provide a black-box reduction for matroidal set systems.
Our results even apply to the broad class of subadditive cost functions that include fixed costs and
discrete concave costs even with weighted players as a special case. Here we assume access to a value
oracle for the subadditive cost function for each resource. Matroidal set systems with player-specific
delays include uncapacitated facility location as a special case, since these correspond to matroid
games, where each player has a uniform rank 1 matroid. For metric UFL, there is for instance a
1.488-approximation algorithm [44] using ideas of a previous 1.5-approximation algorithm [11]. This
leads to a separable cost sharing protocol with PoS of 1.488. Also, the existing hardness results for
UFL carry over to the design of separable cost sharing protocols, and for metric UFL there is a lower
bound of 1.46 [31].
Connection Games with Fixed Costs. In Section 4 we consider cost sharing games on graphs,
where the set systems correspond to paths connecting a player-specific source with a player-specific
terminal. The underlying optimization problem is Steiner forest. For multi-terminal connection
games without delays, we observe that a simple greedy algorithm for the underlying Steiner forest
problem combined with the idea of Prim-Sharing [17] yields a separable protocol in polynomial time.
Since the greedy algorithm has recently been shown to provide a constant-factor approximation [32],
the protocol yields a constant PoS.
For single-source multi-terminal connection games we again provide a polynomial-time black-box
reduction. Our result improves significantly over the existing Prim-Sharing [17] with a price of stability of 2. We obtain separable protocols based on any approximation algorithm for Steiner tree, such
as, e.g., the classic 1.55-approximation algorithm [48], or the celebrated recent 1.39-approximation
algorithm [12]. Our black-box reduction continues to hold even for directed graphs, where we can use
any algorithm for the Directed Steiner Tree problem [15], or games based on the (directed or undirected) Group Steiner Tree problem [16, 26]. Similarly, all lower bounds on approximation hardness
translate to the price of stability of polynomial-time computable separable protocols.
Connection Games with Delays. Finally, in Section 5 we study multi-terminal connection games
with delays and fixed costs. For directed graphs, an optimal Steiner forest is not enforceable by a
separable cost sharing protocol, even for two players [17]. Very recently, a similar result was shown
even for two-player games on undirected graphs [34]. Thus, for general graphs, we cannot expect
separable protocols with optimal or close-to-optimal equilibria, or (efficient) black-box reductions. We
introduce a class of so-called n-series-parallel graphs, which allows to obtain a black-box reduction in
polynomial time. The transformation directly implies that the n-series-parallel graphs always admit
a separable cost sharing protocol inducing an optimal Steiner forest as an equilibrium.
The reduction also applies to discrete-concave cost functions and player-specific delays, however,
we do not know if polynomial running time is guaranteed. n-series-parallel graphs have treewidth
at most 2, thus, for fixed edge costs and no delays, it is possible to compute efficiently even an
optimal Steiner forest [7]. Hence, in this case we obtain a separable protocol with PoS of 1 in
polynomial time. We finally demonstrate that the specific setting of n-series-parallel graphs is in
3
some sense necessary: Even for generalized series-parallel graphs we give a counterexample showing
that a black-box reduction is impossible to achieve.
1.2
Preliminaries and Related Work
Cooperative cost sharing games have been studied over the last decades for a variety of combinatorial
optimization problems, such as minimum spanning tree [10], Steiner tree [29, 30, 45, 51], facility
location [28], vertex cover [22], and many more. Cooperative cost sharing games have interesting
implications for (group-)strategyproof cost sharing mechanisms [40, 41, 46, 47]. For Bayesian costsharing mechanisms there even exist efficient black-box reductions from algorithm to mechanism
design [27]. A major difference to our work is that cooperative cost sharing is not separable.
The most prominent example of a separable cost sharing protocol is the fair-share protocol, in
which the cost of each resource is divided in equal shares among the players that allocate it. This
protocol is also anonymous, and it implies that the resulting game is a congestion game [49]. It
guarantees the smallest price of stability within a class of anonymous protocols [17]. The fair-share
protocol has attracted a serious amount of research interest over the last decade [1, 5, 8, 33], especially the notorious open problem of a constant price of stability for connection games in undirected
graphs [9, 23, 25, 42, 43]. However, as a significant drawback, outside of the domain of undirected
connection games the price of stability is often as large as Ω(log n). Moreover, computing a PNE is
PLS-hard, even for undirected connection games [50].
More general separable protocols have been studied mostly in terms of the price of anarchy, e.g.,
for scheduling (or matroid games) [6, 13, 19, 24, 52] or single-source network design with [20, 21] and
without uncertainty [17]. The best result here is a price of anarchy (and stability) of 2 via PrimSharing [17], a protocol inspired by Prim’s MST algorithm. A protocol with logarithmic price of
stability was shown for capacitated UFL games [36].
We note here that separable protocols with low PoS can be obtained using results for cost sharing
games with so-called arbitrary sharing. In cost sharing games with arbitrary sharing, each agent
i ∈ N specifies as strategy a set Si of allocated resources and a payment
P pi,e for every resource e ∈ E.
A resource e ∈ E is bought if the total P
payments exceed the costs i pi,e ≥ ce (S). The private cost
of player i is the sum of all payments e pi,e if all resources e ∈ Si are bought, and ∞ otherwise.
Note that for games with fixed costs ce (S) = ce , one usually drops the explicit allocation Si from
the strategy of a player. Instead, each player
P i ∈ N simply specifies strategic payments pi,e for each
e ∈ E. Then the private cost of player i is e pi,e if payments suffice to buy at least one feasible set
in Si , and ∞ otherwise. The following proposition is an interesting, straightforward insight. It has
been observed before in the special case of single-source connection games [17, Proposition 6.5].
Proposition 1. If for a cost sharing model, the non-cooperative game with arbitrary sharing has
a pure Nash equilibrium, then there is a separable cost sharing protocol with the same pure Nash
equilibrium.
Proof. It is easy to see that in a PNE (S, p) for a game with arbitrary sharing, every player i ∈ N
contributes only to resources from one feasible set Si ∈ Si . Moreover, the cost of every resource is
exactly paid for. Finally, if a player i deviates to a different feasible set Si′ , then for each e ∈ Si′ \ Si
she only needs to contribute the marginal costs that arise due to her presence. In particular, for fixed
costs, she can use all resources bought by others for free.
Hence, given a PNE (S, p) for the game with arbitrary sharing, we obtain a basic and separable
protocol Ξ as follows. If in profile S ′ a resource e is allocated by the set Ne (S), then we assign
ξe,i (S ′ ) = pi,e for every i ∈ Ne (S). For a profile S ′ , in which at least one other player i ∈ Ne (S ′ )\Ne (S)
allocates e, we pick one of these players i, and she has to pay the full cost ξi,e (S ′ ) = ce (S ′ ). If players
from a strict subset Ne (S ′ ) ⊂ Ne (S) allocate e in S ′ , we can use an arbitrary budget-balanced sharing
4
of ce (S ′ ). It is straightforward to verify that such a protocol is basic and separable, and the state S
is a PNE.
This implies existence of separable protocols with optimal PNE and price of stability 1 for a variety
of classes of games, including matroid games with uniform discrete-concave costs [35], uncapacitated
facility location with fixed [14] and discrete-concave costs [38], connection games (single-source [2,
37] and other classes [3, 4, 39]) with fixed costs, and more. However, the large majority of these
results are inefficient, i.e., there is no polynomial-time algorithm that computes the required optimal
equilibrium.
Alternatively, one may resort to approximate equilibria in games with arbitrary sharing that
are efficiently computable. The most prominent technique works via reducing costs by an additive
value ε to ensure polynomial running time (put forward for single-source connection games in [2]
and used in much of the follow-up work [3, 4, 14, 37]). This approach does not translate to separable
protocols, since a player must eventually contribute to all resources. This is impossible for the model
we consider here.
2
Separable Cost Sharing Protocols
We are given a finite set N of players and a finite set E of resources. Each player i ∈ N is associated
with a predefined family of subsets Si ⊆ 2E from which player i needs to pick at least one. The space
of strategy profiles is denoted by S := ×i∈N Si . For S ∈ S we denote by Ne (S) = {i ∈ N : e ∈ Si }
the set of players that allocate resource e. Every resource e ∈ E has a fixed cost ce ≥ 0, e ∈ E that is
assumed to be shareable by the players. In addition to the shareable costs, there are player-specific
constant costs di,e ≥P
0, i ∈ N, e ∈ E that are not shareable. If player i chooses subset Si , then the
player-specific costsP e∈Si di,e must
Pbe paid
P completely by player i. The total cost of a profile S is
defined as C(S) = e∈∪i∈N Si ce + i∈N e∈Si di,e .
A cost sharing protocol Ξ assigns cost share functions ξi,e : S → R≥0 for all i ∈ N and e ∈ E
and thus induces theP
strategic game (N, S, ξ). For a player i, her total private cost of strategy Si in
profile S is ξi (S) := e∈Si (ξi,e (S) + di,e ). We assume that every player picks a strategy in order to
minimize her private cost. A prominent solution concept in non-cooperative game theory are pure
Nash equilibria. Using standard notation in game theory, for a strategy profile S ∈ S we denote
by (Si′ , S−i ) := (S1 , . . . , Si−1 , Si′ , Si+1 , . . . , Sn ) ∈ S the profile that arises if only player i deviates to
strategy Si′ ∈ Si . A profile is a pure Nash equilibrium (PNE) if for all i ∈ N it holds ξi (S) ≤ ξi (Si′ , S−i )
for all Si′ ∈ Si .
In order to be practically relevant, cost sharing protocols need to satisfy several desiderata. In
this regard, separable cost sharing protocols are defined as follows [17].
Definition 2 (Cost Sharing Protocols and Enforceability). A cost sharing protocol Ξ is
1. stable if it induces only games that admit at least one pure Nash equilibrium.
2. budget balanced, if for all e ∈ E with Ne (S) 6= ∅
X
ce =
ξi,e (S) and ξi,e (S) = 0 for all i 6∈ Ne (S).
i∈Ne (S)
3. separable if it is stable, budget-balanced and induces only games for which in any two profiles
S, S ′ ∈ S for every resource e ∈ E,
Ne (S) = Ne (S ′ ) ⇒ ξi,e (S) = ξi,e (S ′ ) for all i ∈ Ne (S).
5
4. polynomial time computable, if the cost sharing functions ξ can be computed in polynomial
time in the encoding length of the cost sharing game.
We say that a strategy profile S is enforceable, if there is a separable protocol inducing S as a pure
Nash equilibrium.
Separability means that for any two profiles S, S ′ the cost shares on e are the same if the set of
players using e remains unchanged. Still, separable protocols can assign cost share functions that
are specifically tailored to a given congestion model, for example based on an optimal profile. In
this paper, we are additionally interested in polynomial-time computable protocols that we introduce
here.
3
Matroid Games
In this section, we consider matroid games. As usual in matroid theory, we will write Bi instead of Si ,
and B instead of S, when considering
S matroid games. The tupel M = (N, E, B, (ce )e∈E , (di,e )e∈E,i∈N )
is called a matroid game if E = i∈N Ei , and each set system Bi ⊆ 2Ei forms the base set of some
matroid Mi = (Ei , Bi ). While seemingly abstract, the class includes several prominent application
domains, such as UFL games. In a UFL game, the resources are facilities (e.g. common transport
hubs) and the players incur delay di,e in addition to their cost shares for opening used facilities.
Every player i chooses exactly one resource, that is |Bi | = 1 for all Bi ∈ Bi and i ∈ N and hence Bi
corresponds to a uniform matroid of rank one. Recall that every base B of a matroid Mi = (Ei , Bi )
has the same cardinality which we denote with rki (the rank of Mi ).
In the following, instead of fixed costs on the resource, we allow for general subadditive cost
functions ce : 2N → R+ , e ∈ E. ce is called subadditive, if it satisfies (1) ce (S) ≤ ce (T ) for all
S ⊆ T ⊆ N , and (2) ce (S + {i}) ≤ ce (S) + ce ({i}) for all S ⊂ N, i ∈ N . Note that subadditive
functions include fixed costs and discrete concave costs as a special case including the possibility of
weighted demands as in weighted congestion games.
Let us denote the cost of the cheapest alternative of player i to resource e for profile B ∈ B
(cf (Bi + f − e, B−i ) + di,f ). Here we use the intuitive notation ce (B) :=
by ∆ei (B) := min f ∈E
Bi +f −e∈Bi
ce (Ne (B)). We recapitulate a characterization of enforceable strategy profiles obtained in [36].1
Lemma 3. A collection of bases B = (B1 , . . . , Bn ) is enforceable by a separable protocol if and only
if the following two properties are satisfied. Note that (D1) implies that each summand ∆ei (B) − di,e
in (D2) is nonnegative.
di,e ≤ ∆ei (B) for all i ∈ N, e ∈ Bi
X
(∆ei (B) − di,e ) for all e ∈ E.
ce (B) ≤
(D1)
(D2)
i∈Ne (B)
Remark 4. The characterization was used in [36] to prove that an optimal collection of bases is
enforceable. This implies a PoS of 1 for a separable cost sharing protocol that relies on the optimal
profile. As such, the protocol is not efficiently computable (unless P = N P ).
In the following, we devise a black-box reduction in Algorithm 1. It takes as input an arbitrary
collection of bases B and transforms them in polynomial time into an enforceable set of bases B ′ of
lower cost C(B ′ ) ≤ C(B). We define for each i ∈ N, e ∈ E a virtual cost value πie = ce ({i})+ di,e , and
1
The original characterization in [36] was proven for weighted players and load-dependent non-decreasing cost
functions but the proof also works for subadditive cost functions.
6
Algorithm 1: Transforming any profile B into an enforceable profile B ′
Input: Congestion model (N, B, c, d) and profile B ∈ B
Output: Enforceable profile B ′ with C(B ′ ) ≤ C(B).
1
2
Set B ′ ← B
while there is e ∈ E that satisfies at least one of the following conditions:
¯ e (B ′ ) for some i ∈ Ne (B ′ )
di,e > ∆
(1) or
i
X
¯ e (B ′ ) − di,e
(2) do
∆
ce (B ′ ) >
i
i∈Ne (B ′ )
3
4
5
6
7
8
9
10
if (1) holds true for some i ∈ Ne (B ′ ) then
Let fi ∈ arg min πif
Update
f ∈E
Bi′ +f −e∈Bi
Bi′ ← Bi′ + fi −
e
else if (2) holds true then
while (2) holds true on e do
¯ e (B ′ )
Pick i ∈ Ne (B ′ ) with πei > ∆
i
Let fi ∈ arg min πif
Update
f ∈E
Bi′ +f −e∈Bi
Bi′ ← Bi′ + fi −
e
¯ e (B) = min
for each B ∈ B, i ∈ N, e ∈ E a virtual deviation cost ∆
i
f ∈E
Bi +f −e∈Bi
πif . The algorithm now
iteratively checks whether (D1) and (D2) from Lemma 3 hold true (in fact it checks this condition for
smaller values on the right hand side given by the virtual values), and if not, exchanges one element
of some player. We show that the algorithm terminates with an enforceable profile after polynomially
many steps.
Theorem 5. Let B be a strategy profile for a matroid congestion model with subadditive costs. There
is an enforceable strategy B ′ with C(B ′ ) ≤ C(B) that can be computed in at most n·m·rk(B) iterations
of the while-loop in Algorithm 1, where rk(B) = maxi∈N rki .
¯ e (B) ≤
Proof. First, observe that if (D1) and (D2) from Lemma 3 hold true for smaller values 0 ≤ ∆
i
∆ei (B), i ∈ N, e ∈ E, then the profile B is also enforceable. Hence, if the algorithm terminates, the
resulting strategy profile B ′ will be enforceable.
To show
8. By subadditivity
P that the algorithm ′ is well-defined, we only′ need
Pto check ¯Line
e
′
we get i∈Ne (B ′ ) ce ({i}) ≥ ce (B ). Thus, whenever ce (B ) > i∈Ne (B ′ ) ∆i (B ) − di,e , there is an
¯ e (B ′ ).
i ∈ Ne (B ′ ) with ce ({i}) + di,e > ∆
i
It is left to bound the running time. For this we consider player i and the matroid bases Bi . We
interpret a basis Bi ∈ Bi as distributing exactly rki unit sized packets over the resources in E. This
way, we can interpret the algorithm as iteratively moving packets away from those resources e ∈ E
for which either (1) or (2) holds true. We give each packet a unique ID ik , k = 1, . . . , rki . For Bi ∈ Bi ,
let eik denote the resource on which packet ik is located. We now analyze the two types of packet
movements during the execution of the algorithm. For a packet movement executed in Line 5 of
¯ e (B ′ ), thus, when packet ik located on e = ei is moved to fi , it holds
Algorithm 1, we have di,e > ∆
i
k
eik
¯ e (B ′ ) = π fi . For packet movements executed in Line 10, then by the choice
that πi = πie ≥ di,e > ∆
i
i
ei
¯ e (B ′ ) = π fi .
of player i ∈ Ne (B ′ ) (see Line 8) for the corresponding packet ik it holds πi k = πie > ∆
i
i
In both cases we obtain πie > πifi . Hence, every movement of a single packet ik is in strictly decreasing
7
order of virtual value of the resource. Note that the virtual cost value πie does not depend on the
profile B. Thus, there are at most m different virtual cost values that a packet ik of player i can
experience, and thus packet ik can move at most m −P
1 times. The following is an upper bound on
the total number of packet movements for all players i∈N rki ·(m − 1) ≤ n · m · rk(B).
It is left to argue that the final output B ′ has lower cost. We prove this inductively by the
different types of packet movements. Consider first a packet movement of type (1). Let B and B ′ be
the profiles before and after packet ik has been moved from e to fi , respectively. We obtain
C(B ′ ) − C(B) = (cfi (B ′ ) − cfi (B) + di,fi ) − (ce (B) − ce (B ′ ) + di,e )
≤ cfi ({i}) + di,fi − (ce (B) − ce (B ′ ) + di,e )
¯ ei (B) − di,e + (ce (B ′ ) − ce (B))
=∆
¯ e (B) − di,e
≤∆
i
<
0.
The first inequality follows from subadditivity, the second inequality from monotonicity of costs ce .
The last strict inequality follows from assumption (1).
Now consider packet movements of type (2). We treat all movements occurring in one run of
the while loop in Line 7. Let B denote the profile before and B ′ after all these movements. Let
Te (B) ⊆ Ne (B) denote
S the set of those players whose packet ik on e is moved to fi during the while
loop. Let Fe (B) = i∈Te (B) {fi } and for i ∈ Te (B) define Tfi (B) = {j ∈ Te (B) | fj = fi }. We derive
some useful observations. Before entering the while loop, it holds
X
X
X
¯ ei (B) − di,e =
¯ ei (B) − di,e +
¯ ei (B) − di,e . (3)
ce (B) >
∆
∆
∆
i∈Ne (B)
i∈Ne (B)\Te (B)
Moreover, after exiting the while loop it holds
X
ce (B ′ ) ≤
i∈Te (B)
¯ e (B) − di,e .
∆
(4)
i
i∈Ne (B)\Te (B)
Thus, combining (3) and (4) we get
ce (B) − ce (B ′ ) >
X
¯ ei (B) − di,e .
∆
(5)
i∈Te (B)
Putting everything together, we obtain
X
X
X
cfi (B ′ ) − cfi (B) +
di,fi − ce (B) − ce (B ′ ) +
di,e
C(B ′ ) − C(B) =
fi ∈Fe (B)
≤
X
i∈Te (B)
X
cfi ({j}) +
fi ∈Fe (B) j∈Tfi (B)
=
X
i∈Te (B)
X
i∈Te (B)
di,fi − ce (B) − ce (B ′ ) +
i∈Te (B)
¯ e (B) − ce (B) − ce (B ′ ) +
∆
i
X
i∈Te (B)
X
i∈Te (B)
di,e
<
di,e
0,
where the first inequality follows from subadditivity and the last inequality follows from (5).
4
Connection Games without Delays
In this section, we study connection games in an undirected graph G = (V, E) with a common
source vertex s ∈ V . Every player i wants to connect a player-specific terminal node ti ∈ V to s.
Consequently, every strategy Pi of player i is an (s, ti )-path in G. We denote the set of paths for
player i by Pi and the set of profiles by P.
8
Note that when each edge cost contains a player-specific delay component di,e , we can take any
multi-source multi-terminal connection game and introduce a new auxiliary source vertex s. Then
connect s to each si with an auxiliary edge ei , which has cost di,ei = 0 and dj,ei = M , for some
prohibitively large constant M . Now in any equilibrium and any optimal state of the resulting
game, player i will choose an (s, ti )-path which begins with edge ei . Moreover, ei does not generate
additional cost for player i. As such, the optimal solutions, the Nash equilibria, and their total
costs correspond exactly to the ones of the original multi-source multi-terminal game. Hence, in
games with non-shareable player-specific delays, the assumption of a common source is without loss
of generality and existing lower bounds on the price of stability apply [17, 34].
In this section, we instead focus on connection games with fixed shareable costs ce ≥ 0 and no
player-specific delays di,e = 0, for all players i and all edges e ∈ E. For the general multi-terminal
multi-source case with such costs, it is straightforward to observe that the greedy algorithm analyzed
by Gupta and Kumar [32] can be turned into a separable protocol via the Prim-Sharing idea [17].
This implies that we can obtain separable cost sharing protocols with a constant price of stability in
polynomial time.
Proposition 6. For every connection game in undirected graphs with fixed costs, there is an enforceable profile that can be computed in polynomial time and yields a separable cost sharing protocol with
constant price of stability.
For single-source games with fixed costs, existing results for cost sharing games with arbitrary
sharing imply that an optimal profile is always enforceable [2, 17]. We here provide a significantly
stronger result for polynomial-time computation of cheap enforceable profiles.
Theorem 7. Let P be a strategy profile for a single-source connection game with fixed costs. There
is an enforceable strategy P ′ with C(P ′ ) ≤ C(P ) that can be computed by Algorithm 2 in polynomial
time.
It is straightforward that for fixed costs we can transform each profile P into a cheaper tree
profile P̂ , in which the union of player paths constitute a tree T . Over the course of the algorithm,
we adjust this tree and construct a cost sharing for it in a bottom-up fashion. The approach has
similarities to an approach for obtaining approximate equilibria for single-source cost sharing games
with arbitrary sharing [2]. However, our algorithm exploits crucial properties of separable protocols,
thereby providing an exact Nash equilibrium and polynomial running time.
When designing a separable protocol based on a state P̂ , we can always assume that when a
player i deviates unilaterally to one or more edges e ∈ G \ P̂i , she needs to pay all of ce . As such,
player i always picks a collection of shortest paths with respect to ce between pairs of nodes on her
current path P̂i . All these paths in G are concisely represented in the algorithm as “auxiliary edges”.
The algorithm initially sets up an auxiliary graph Ĝ given by T and the set of auxiliary edges based
on P̂ . It adjusts the tree T by removing edges of T and adding auxiliary edges in a structured fashion.
We first show in the following lemma that this adjustment procedure improves the total cost
of the tree, and that the final tree T̂ is enforceable in Ĝ. In the corresponding cost sharing, every
auxiliary edge contained in T̂ is completely paid for by a single player that uses it. In the subsequent
proof of the theorem, we only need to show that for the auxiliary edges in T̂ , the edge costs of
the corresponding shortest paths in G can be assigned to the players such that we obtain a Nash
equilibrium in G. The proof shows that the profile P ′ evolving in this way is enforceable in G and
only cheaper than P .
Lemma 8. Algorithm 2 computes a cost sharing of a feasible tree T̂ in the graph Ĝ. The total cost
C(T̂ ) ≤ C(T ), every auxiliary edge in T̂ is paid for by a single player, and the corresponding profile
P̂ is enforceable in Ĝ.
9
Algorithm 2: Transforming any profile P into an enforceable profile P ′
Input: Connection game (N, G, (t1 , . . . , tn ), s, c) and profile P ∈ P
Output: Enforceable profile P ′ with C(P ′ ) ≤ C(P ).
S
1 Transform P into a tree profile P̂ and let T ←
i P̂i
2 ĉe (i) ← 0, for all e ∈ T, i ∈ N
3 Insert T into empty graph G′ , root T in s, number vertices of T in BFS order from s
4 foreach i ∈ N and each vk , vk ′ ∈ P̂i with k > k ′ do
5
Add to Ĝ an auxiliary edge e = (vk , vk′ )
in G from vk to vk′
6
P (vk , vk′ ) ← shortest path P
7
For all j ∈ N , set ĉe (j) ← f ∈P (vk ,v ′ ) cf
k
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
Label every e ∈ T as “open”
foreach open e ∈ T in bottom-up order do
ĉe (i) ← ce , for all i ∈ Ne (P̂ )
Pd (i) ←
P in Ĝ for edge costs ĉe (i)
Pshortest (s, ti )-path
∆ei ← e′ ∈Pd (i) ĉe′ (i) − e′ ∈P̂i ,e′ 6=e ĉe′ (i)
P
if ce ≤ i∈Ne (P̂ ) ∆ei then
P
For all i ∈ Ne (P̂ ), assign ĉe (i) ∈ [0, ∆ei ] such that i∈Ne (P̂ ) ĉe (i) = ce
Label e as “closed”
else
D ← set of highest deviation vertices
PT (e, v) ← path between higher node of e and v in T
Remove all paths PT (e, v) from T and Ĝ
foreach v ∈ D do
Pick one auxiliary edge e′ ∈ Pd (i), for some i ∈ Ne (P̂ ), such that e′ = (v, u) with u
a node above e in T , and add e′ to T
P̂i ← Pd (i)
foreach player j 6= i with tj below v in T do
P̂j ← (P̂j from tj to v) ∪ (Pd (i) from v to s)
ĉe′ (j) ← 0
Label e′ as “closed”
For every i ∈ N , compute Pi′ by replacing in P̂i every auxiliary edge e = (u, v) by the
corresponding shortest path P (u, v) in G
Proof. After building Ĝ, the algorithm considers T rooted in the source s. Initially, all edges of T
are assumed to have zero cost for all players. All edges of T are labelled “open”. Our proof works
by induction. We assume that players are happy with their strategies P̂i if all open edges of T have
cost 0, all open edges outside T have cost ĉe for every player, and the closed edges e ∈ T are shared
as determined by ĉe .
The algorithm proceeds in a bottom-up fashion. In an iteration, it restores the cost of an open
edge e to its original value. It then considers how much each player i ∈ Ne (P̂ ) is willing to contribute
to e. The maximum contribution ∆ei is given by the difference in the cheapest costs to buy an (s, ti )path for i when (1) e has cost 0 and (2) e has cost ce . By induction, for case (1) we can assume that
i is happy with P̂i when e has cost 0. In case (2), suppose i deviates from (parts of) his current path
P̂i and buys auxiliary edges.
10
Since by induction i is happy with P̂i when e has cost 0, there is no incentive to deviate from
P̂i between two vertices of P̂i below e. Moreover, clearly, there is no incentive to deviate from P̂i
between two vertices above e (since edges of T above e are assumed to have zero cost). Hence, if in
case (2) the path Pd (i) includes e, then Pd (i) = P̂i , so ∆ei = ce . Otherwise, player i finds a path that
avoids e. By the observations so far, Pd (i) can be assumed to follow P̂i from ti up to a vertex v, then
picks a single auxiliary edge (v, u) to node u above e, and then follows P̂i to s. We call the vertex v
the deviation vertex of Pd (i).
Based on Pd (i), the algorithm computes a maximum contribution ∆ei for each player i ∈ Ne (P̂ ),
which i is willing to pay for edge e currently under consideration. If in total these contributions
suffice to pay for e, then we determine an arbitrary cost sharing of ce such that each player i ∈ Ne
pays at most ∆ei . Thereby, every player i ∈ Ne (P̂ ) remains happy with his path P̂i , and the inductive
assumptions used above remain true. We can label e as closed and proceed to work on the next open
edge in the tree T .
Otherwise, if the contributions ∆ei do not suffice to pay for ce , then for every i ∈ Ne (P̂ ) the path
Pd (i) avoids e and contains a deviation vertex. The algorithm needs to drop e and change the strategy
of every such player. It considers the “highest” subset D of deviation vertices, i.e., the unique subset
such that D contains exactly one deviation vertex above each terminal ti . The algorithm removes
all edges from T that lie below and including e and above any v ∈ D. For each v ∈ D, it then adds
one auxiliary edge from the corresponding Pd (i) to T . As observed above, these edges connect v to
some node u above e, and thereby yield a new feasible tree T in Ĝ.
Since Pd (i) is a best response for player i, we assign i to pay for the cost of the auxiliary edge.
After this update, i is clearly happy with P̂i . Moreover, every other player j ∈ Ne (P̂ ) that now uses
the auxiliary edge paid by player i is happy with his new strategy P̂j . The auxiliary edge has cost
zero for player j, and the path from ti to the deviation vertex v has not changed. By induction j
was happy with this path after we finished paying for the last edge below v. Thus, we can label all
auxiliary edges added to T as closed and proceed to work on the next open edge in the tree T .
By induction, this proves that the algorithm computes a cost sharing that induces a separable
protocol with the final tree T ′ being a Nash equilibrium in Ĝ. Moreover, if we change the tree
during the iteration for edge e, it is straightforward to verify that the total cost of the tree strictly
decreases.
Proof of Theorem 7. The previous lemma shows that the algorithm computes a cost sharing of a
tree T̂ in Ĝ, such that every player is happy with the path P̂i and every auxiliary edge in T̂ is paid
for completely by a single player. We now transform P̂ into P ′ by replacing each auxiliary edge
e = (u, v) ∈ P̂i by the corresponding shortest path P (u, v) in G. We denote by Ei the set of edges
introduced in the shortest paths for auxiliary edges in P̂i . For the total cost of the resulting profile we
have that C(P ′ ) ≤ C(P̂ ) ≤ C(P ), since the sets Ei can overlap with each other or the non-auxiliary
edges of T̂ .
We show that P ′ is enforceable by transforming the cost sharing constructed in function ĉ into
separable cost sharing functions as follows. Initially, set ξi,e (P ′ ) = 0 for all e ∈ E and i ∈ N . Then,
for each non-auxiliary edge e ∈ T̂ we assign ξi,e (P ′ ) = ĉe (i) if e ∈ P̂i and ξi,e (P ′ ) = 0 otherwise.
Finally, number players arbitrarily
P from 1 to n and proceed in that order. For player i, consider the
edges in Ei . For every e ∈ Ei , if j<i ξj,e (P ′ ) = 0, then set ξi,e (P ′ ) = ce .
This yields a budget-balanced assignment for state P ′ . As usual, if a player i deviates in P ′ from
′
Pi to Pi′′ , we can assume player i is assigned to pay the full cost ce for every edge e ∈ Pi′′ \ Pi′ . To
show that there is no profitable deviation from P ′ , we first consider a thought experiment, where
every edge in Ei comes as a separate edge bought by player i. Then, clearly P ′ is enforceable – the
cost of Pi′ with ξ is exactly the same as the cost of P̂ with ĉ in Ĝ. Moreover, any deviation Pi′′ can
be interpreted as an (s, ti )-path in Ĝ by replacing all subpaths consisting of non-auxiliary edges in
P ′′ by the corresponding auxiliary edge of Ĝ. As such, the cost of Pi′′ is exactly the same as the cost
11
of the corresponding deviation in Ĝ. Now, there is not a separate copy for every edge in Ei . The
set Ei can overlap with other sets Ej and/or non-auxiliary edges. Then player i might not need to
pay the full cost on some e ∈ Ei . Note, however, every edge for which player i pays less than ce is
present in Pi′ as well. Hence, Pi′′ cannot improve over Pi′ due to this property.
The result continues to hold for various generalizations. For example, we can immediately apply
the arguments in directed graphs, where every player i seeks to establish a directed path between
ti and s. Moreover, the proof can also be applied readily for a group-connection game, where each
player wants to establish a directed path to s from at least one node of a set Vi ⊂ V . For this game,
we simply add a separate super-terminal ti for every player i and draw a directed edge of cost 0 from
ti to every node in Vi .
Corollary 9. Let P be a strategy profile for a single-source group-connection game in directed graphs
with fixed costs. There is an enforceable profile P ′ with C(P ′ ) ≤ C(P ) that can be computed by
Algorithm 2 in polynomial time.
5
Connection Games and Graph Structure
In this section, we consider connection games played in undirected graphs G = (V, E) with playerspecific source-terminal pairs. Each player i ∈ N has a source-terminal-pair (si , ti ). Note that we
can assume w.l.o.g. that (G, (s1 , t1 ), . . . , (sn , tn )) is irredundant, meaning that each edge and each
vertex of G is contained in at least one (si , ti )-path for some player i ∈ N (nodes and edges which
are not used by any player can easily be recognized (and then deleted) by Algorithm Irredundant
at the end of the section; adapted from Algorithm 1 in [18]).
Harks et al. [34] characterized enforceability for the special case with di,e = 0 for all i ∈ N, e ∈ E via
an LP. We can directly adapt this characterization as follows:
X
LP(P ) max
ξi,e
i∈N,e∈Pi
X
s.t.:
ξi,e ≤ ce
∀e ∈ E with Ne (P ) 6= ∅
i∈Ne (P )
X
(ξi,e + di,e ) ≤
e∈Pi \Pi′
X
(ce + di,e )
∀Pi′ ∈ Pi ∀i ∈ N
(NE)
e∈Pi′ \Pi
ξi,e ≥ 0
∀e ∈ Pi ∀i ∈ N
Theorem 10. The strategy profile P = (P1 , . . . , Pn ) is enforceable if and only if there is an optimal
solution (ξi,e )i∈N,e∈Pi for LP(P ) with
X
ξi,e = ce ∀e ∈ E with Ne (P ) 6= ∅.
(BB)
i∈Ne (P )
Given an optimal solution (ξi,e )i∈N,e∈Pi for LP(P ) with the property (BB), the profile P becomes
a PNE in the game induced by ξ, which assigns for each i ∈ N and e ∈ E and each strategy profile
P ′ = (P1′ , . . . , Pn′ ) the following cost shares (these cost shares resemble those introduced in [52]):
ξi,e , if i ∈ Se (P ′ ) = Se (P ),
c , if i ∈ (S (P ′ ) \ S (P )) and i = min(S (P ′ ) \ S (P )),
e
e
e
e
e
ξi,e (P ′ ) =
′
′
ce , if i ∈ Se (P ) ( Se (P ) and i = min Se (P ),
0,
else.
12
We now introduce a subclass of generalized series-parallel graphs for which we design a polynomial
time black-box reduction that computes, for a given strategy profile P , an enforceable strategy profile
with smaller cost.
Definition 11 (n-series-parallel graph). An irredundant graph (G, (s1 , t1 ), . . . , (sn , tn )) is n-seriesparallel if, for all i ∈ N , the subgraph Gi (induced by Pi ) is created by a sequence of series and/or
parallel operations starting from the edge si − ti . For an edge e = u − v, a series operation replaces
it by a new vertex w and two edges u − w, w − v; A parallel operation adds to e = u − v a parallel
edge e′ = u − v.
The following theorem summarizes our results for n-series-parallel graphs.
Theorem 12. If (G, (s1 , t1 ), . . . , (sn , tn )) is n-series-parallel, the following holds:
(1) Given an arbitrary strategy profile P , an enforceable strategy profile P ′ with cost C(P ′ ) ≤ C(P ),
and corresponding cost share functions ξ, can be computed in polynomial time.
(2) For all cost functions c, d, every optimal strategy profile of (G, (s1 , t1 ), . . . , (sn , tn ), c, d) is enforceable.
(3) For all edge costs c, an optimal Steiner forest of (G, (s1 , t1 ), . . . , (sn , tn ), c) can be computed in
polynomial time.
To prove Theorem 12, we need to introduce some notation. Let (ξi,e )i∈N,e∈Pi be an optimal
solution for LP(P ). For i ∈ N and P
f ∈ Pi , we consider allPpaths Pi′ ∈ Pi with f ∈
/ Pi′ , Pi ∪ Pi′
′
contains a unique cycle C(Pi ) and
e∈Pi′ \Pi (ce + di,e ). Among all these
e∈Pi \Pi′ (ξi,e + di,e ) =
′
paths, choose one for which the number of edges in C(Pi ) ∩ Pi is minimal. The corresponding path
Ai,f := C(Pi′ ) ∩ Pi′ is called a smallest tight alternative of player i for f . If we say that player i
substitutes f by using Ai,f , we mean that the current path Pi of player i is changed by using Ai,f
instead of the subpath C(Pi′ ) ∩ Pi (which contains f ). Figure 1 illustrates the described concepts.
Note that Ai,f is also smallest in the sense that every other (tight) alternative for f substitutes
a superset of the edges substituted by Ai,f .
We are now able to prove Theorem 12.
Proof of Theorem 12. We first describe how to compute, given an arbitrary strategy profile P =
(P1 , . . . , Pn ), an enforceable strategy profile with cost at most C(P ).
Assume that P is not enforceable (otherwise there is nothing to do). Let (ξi,e )i∈N,e∈Pi be an
optimal solution for LP(P ). In the following, we denote the variables (ξi,e )i∈N,e∈Pi as cost shares,
although they do not correspond to a budget-balanced cost sharing protocol (since
PP is not enforceable). There is at least one edge f which is not completely paid, i.e. for which i∈Nf (P ) ξi,f < cf
holds. The optimality of the cost shares (ξi,e )i∈N,e∈Pi for LP(P ) implies that each player i ∈ Nf (P ),
i.e. each user of f , has an alternative path Pi′ with f ∈
/ Pi′ , for which equality holds in the corresponding LP(P )-inequality (otherwise increasing ξi,f by a small amount, while all other cost shares remain
unchanged, would yield a feasible LP-solution with higher objective function value). Using the notation introduced above, each user i of f has a smallest tight alternative Ai,f for f . Furthermore,
if Pi contains more than one edge which is not completely paid, there is a combination of smallest
tight alternatives so that all edges which are not completely paid are substituted (see Figure 2, where
f, g, h are not completely paid and we substitute all these edges by combining Ai,g and Ai,h ).
We now consider the strategy profile P ′ = (P1′ , . . . , Pn′ ) which results from P if all players with
unpaid edges in their paths substitute all these edges by a combination of smallest tight alternatives.
13
si
ti
f
(a) Example for Pi (thick) and all alternative paths with tight inequality in LP(P ).
si
ti
f
(b) Dashed path
Pi′
with f ∈
/
Pi′ ,
with f ∈
/
Pi′ ,
but no unique cycle with Pi .
si
ti
f
(c) Dashed path
Pi′
unique cycle
C(Pi′ )
with Pi , but not smallest for f .
Ai,f
f
si
ti
(d) Substituting f by using smallest tight alternative Ai,f .
Figure 1: Illustration of the introduced concepts.
si
g
f
h
ti
Figure 2: Illustration of combining smallest tight alternatives.
Furthermore we define cost shares (again not necessarily budget-balanced) for P ′ as follows: For each
player i and each edge e ∈ Pi′ :
(
ξi,e , for e ∈ Pi′ ∩ Pi ,
ξi,e (P ′ ) =
ce , for e ∈ Pi′ \ Pi .
Note that the private cost of player i under P P
equals the private cost of i under P ′ since the players
use tight alternatives. Furthermore note that i∈Ne (P ′ ) ξi,e (P ′ ) > ce is possible (for example if there
are two players which did not
use it in P ′ and
P use an edge ′e with ce > 0 in their paths under P , but
′
therefore both pay ce ). If i∈Ne (P ′ ) ξi,e (P ) ≥ ce holds for all edges e with Ne (P ) 6= ∅, we found a
strategy profile with the desired properties: P ′ is cheaper than P since
X
XX
C(P ) =
ce +
di,e
e∈E:Ne (P )6=∅
>
XX
i∈N e∈Pi
(ξi,e + di,e )
i∈N e∈Pi
=
XX
ξi,e (P ′ ) + di,e
i∈N e∈Pi′
=
X
X
ξi,e (P ′ ) +
e∈E:Ne (P ′ )6=∅ i∈Ne (P ′ )
≥
X
e∈E:Ne (P ′ )6=∅
ce +
XX
i∈N e∈Pi′
14
XX
i∈N e∈Pi′
di,e = C(P ′ ).
di,e
The strict inequality holds since P is not enforceable, the following equality because the private costs
remain unchanged, and the last inequality because of our assumption above. Furthermore, P ′ is
enforceable since the cost shares (ξi,e (P ′ ))i∈N,e∈Pi′ induce a feasible solution of LP(P ′ ) with (BB) if
we decrease the cost shares for overpaid edges arbitrarily until P
we reach budget-balance.
Thus assume that there is at least one edge f for which i∈Nf (P ′ ) ξi,f (P ′ ) < cf holds. First
note, for each player i ∈ Nf (P ′ ), that f ∈ Pi has to hold, since all edges in Pi′ \ Pi are completely
paid (player i pays ce for e ∈ Pi′ \ Pi ). As we will show below, all i ∈ Nf (P ′ ) have a smallest tight
alternative Ai,f for f . We can therefore again update the strategy profile (resulting in P ′′ ) by letting
all players deviate from all nonpaid edges using a combination of smallest tight alternatives. Figure 3
illustrates this second phase of deviation, where the edges r and s are now not completely paid. Note
that Ai,r is not unique in this example, and to use Ai,s , we need to deviate from Ai,h (which player
i uses in Pi′ ).
si
r
s
h
ti
si
r
s
h
ti
Figure 3: Illustration for the second phase of deviation.
The cost shares are again adapted, that means for each player i and each edge e ∈ Pi′′ :
(
ξi,e , for e ∈ Pi′′ ∩ Pi ,
ξi,e (P ′′ ) =
ce , for e ∈ Pi′′ \ Pi .
It is clear that the private costs of the players again remain unchanged and therefore, if all edges are
now completely paid, the cost of P ′′ is smaller than the cost of P and P ′′ is enforceable.
We now show that the tight alternatives used in the second phase of deviation exist. Assume, by
contradiction, that there is a player j, an edge f ∈ Pj′ which is not completely paid according to P ′ ,
and player j has no tight alternative for f . Now recall that, whenever an edge f is not completely
paid in P ′ , all users i ∈ Nf (P ′ ) already used f in P and therefore ξi,f (P ′ ) = ξi,f holds for all
i ∈ Nf (P ′ ) ⊆ Nf (P ). Furthermore f was completely paid according to the cost shares of P since we
substituted all unpaid edges in the first phase of deviation from P to P ′ . We get
X
X
X
X
ξi,f +
ξi,f = cf >
ξi,f (P ′ ) =
ξi,f ,
i∈Nf (P ′ )
i∈Nf (P )\Nf (P ′ )
i∈Nf (P ′ )
i∈Nf (P ′ )
thus there has to be at least one player k which used f in Pk , but not in Pk′ , and with ξk,f > 0.
Let Ak,g be the smallest tight alternative that player k used (to substitute the edge g which was
not completely paid in P ), and also substituted f . The situation is illustrated in Figure 4. We
now show that the LP(P )-solution cannot be optimal. Since player j has no tight alternative for
f , increasing ξj,f by some suitably small amount, and decreasing ξk,f by the same amount, yields
a feasible LP(P )-solution. But now player k has no tight alternative for g anymore, since all tight
alternatives for g also substitute f . Therefore we can increase ξk,g by some small amount, leading to
a feasible LP(P )-solution with higher objective function value, contradiction. Thus we showed that
the tight alternatives used in the second phase of deviation exist.
15
g
sk
f
sj
tk
tj
(a) Illustration of the paths Pj and Pk (given by the thick edges).
Ak,g
g
sk
f
sj
tk
tj
(b) Situation after the first phase of deviation (player k used Ak,g ).
Figure 4: Illustration for the proof that tight alternatives exist.
As already mentioned above, if all edges in P ′′ are completely paid, P ′′ is enforceable and cheaper
than P and we are finished. Thus we again assume that there is at least one edge which is not
completely paid. Analogously as for P ′ we can show that, for each such edge f and each player
i ∈ Nf (P ′′ ), f ∈ Pi has to hold. Furthermore all users of a nonpaid edge have a tight alternative
for this edge (the proof that this holds is a little bit more complicated as above, we possibly need
to involve three players now): Assume that a player i does not have a tight alternative for an edge
f ∈ Pi which is not completely paid according to P ′′ , but was completely paid before (i.e. according
to P ′ and also according to P ). Thus there has to be a player j ∈ Nf (P ) with ξj,f > 0 who deviated
from f by using Aj,g in some phase before. If player j did this in the first phase of deviation, the edge
g was not completely paid according to P and we can change the LP(P )-solution as described above
to get a contradiction. If the deviation happened in the second phase, the edge g was completely
paid according to P . Thus there has to be a third player k (but k = i possible) with ξk,g > 0 which
used some Ak,h in the first phase of deviation that also substituted g. We are now able to change
the cost shares of P to get a contradiction: First, player i increases ξi,f , while player j decreases ξj,f .
Now player j increases ξj,g , while player k decreases ξk,g . Finally player k increases ξk,h . By suitably
small changes, we get a feasible solution for LP(P ) with higher objective function value than the
original optimal cost shares; contradiction. Therefore, in a third phase of deviation, all players with
nonpaid edges deviate from all those edges by a combination of smallest tight alternatives.
If we proceed in this manner, we finally have to reach a strategy profile for which all edges
are completely paid (and thus it is enforceable and cheaper than the profile P ): In each phase of
deviation, at least one edge is substituted by all players which use this edge in P . Furthermore,
players never return to substituted edges. Therefore, after at most |P | phases of deviation, we reach
a strategy profile with the desired property (where |P | denotes the number of edges in the union of
the paths P1 , . . . , Pn ). The existence of the needed tight alternatives in the kth phase of deviation
can be shown as follows: Assume that P (k) is the current strategy profile, f an edge which is not
completely paid according to P (k) , and there is a player i which uses an edge f , but has no tight
alternative for it. Then there has to be a player j with ξj,f > 0 who deviated from f in some phase
ℓ ≤ k − 1 by using Aj,g , where g was not completely paid in the corresponding strategy profile P (ℓ) .
If ℓ = 1 holds, we can decrease ξj,f and increase ξi,f , ξj,g ; contradiction. For ℓ ≥ 2, the edge g
was completely paid in P and therefore, there has to be a player p with ξp,g > 0 which substituted
g in some phase ≤ ℓ − 1 by using Ap,h , and so on. This yields a sequence of players and edges
(i, f ), (j, g), (p, h), . . . , (q, s), where the edge s was not completely paid according to P . We can now
change the cost shares along this sequence (as described above for the third phase of deviation) to
16
get a contradiction.
Algorithm n-SePa summarizes the described procedure for computing an enforceable strategy
profile P ′ with cost C(P ′ ) ≤ C(P ) and corresponding cost share functions ξ. To complete the proof of
the first statement of Theorem 12, it remains to show that P ′ and ξ can be computed in polynomial
time, i.e. Algorithm n-SePa has polynomial running time. As a first step, we show how to compute
an optimal solution for LP(P ) in polynomial time. To this end we show that, for every player i, we
do not need to consider all paths Pi′ ∈ Pi in (NE) of LP(P ), which can be exponentially many paths,
but only a set of alternatives Ai of polynomial cardinality. Recall that the graph Gi (induced by Pi )
essentially looks like displayed in Figure 5, and we can w.l.o.g. assume that Pi is given by the thick
edges.
si
ti
Figure 5: Structure of Gi , where Pi is given by the thick edges.
An arbitrary (si , ti )-path Pi′ consists of subpaths of Pi together with some of the “arcs”. We call
these arcs alternatives (according to Pi ), and formally, an alternative is a path A which connects
two nodes of Pi , but is edge-disjoint with Pi . The subpath of Pi with the same endnodes as A is
denoted by PiA , and we say that this subpath is substituted by A (cf. Figure 6 for illustration). Note
that there can be different alternatives which substitute the same subpath of Pi (in Figure 6, this
holds for example for the two arcs on the left which both substitute the second and third edge of
Pi ). Whenever this is the case, we choose such an alternative with smallest sum of edge costs plus
player i’s delays, and denote this alternative A as a cheapest alternative for PiA . Let Ai be the set
of all cheapest alternatives according to Pi . It is clear that
X
X
(ξi,e + di,e ) ≤
(ce + di,e ) ∀Pi′ ∈ Pi
e∈Pi \Pi′
e∈Pi′ \Pi
holds if and only if
X
e∈PiA
(ξi,e + di,e ) ≤
X
(ce + di,e )
∀A ∈ Ai
e∈A
holds. Since the paths in Ai are edge-disjoint, |Ai | is bounded by |E|. Algorithm Alternatives(i)
computes Ai in polynomial time. Thus, we can solve LP(P ) in polynomial time.
A
si
ti
Figure 6: Alternative A (dashed); substituted subpath PiA (thick).
To complete the proof that Algorithm n-SePa has polynomial running time, it remains to show
that the combination of smallest tight alternatives in Line 7 of Algorithm n-SePa can be found in
polynomial time (recall that there will be at most |P | ≤ |E| calls of the repeat-loop; and all other
steps are obviously polynomial). Since |Ai | ≤ |E| holds for all i ∈ N , we can find, for each edge f
which is not completely paid and each user i of f , a smallest tight alternative Ai,f in polynomial time.
If player i uses more than one edge which is not completely paid, a combination of the corresponding
smallest tight alternatives can also easily be found; thus step 7 is polynomial.
17
Overall we showed that the first statement of Theorem 12 holds. The second statement, i.e. that
every optimal strategy profile of (G, (s1 , t1 ), . . . , (sn , tn ), c, d) is enforceable, follows very easily from
the proof of statement (1). Note that, if P is not enforceable, Algorithm n-SePa computes a strategy
profile with strictly smaller cost. Since this would lead to a contradiction if P is an optimal, but not
enforceable strategy profile, every optimal strategy profile has to be enforceable.
We finally want to show the last statement of Theorem 12, i.e. that an optimal Steiner forest can
be computed in polynomial time (for di,e = 0 for all i ∈ N, e ∈ E). To this end we want to use the
result of Bateni et al. [7] that the Steiner forest problem can be solved in polynomial time on graphs
with treewidth at most 2. Thus it is sufficient to show that G has treewidth at most 2, or, since
generalized series-parallel graphs have treewidth at most 2 (what can easily be seen by induction on
the number of operations), that G is generalized series-parallel (note that we can assume w.l.o.g. that
G is connected, otherwise we can obviously treat each connected component separately). Recall that
generalized series-parallel graphs are created by a sequence of series, parallel, and/or add operations
starting from a single edge, where an add-operation adds a new vertex w and connects it to a given
vertex v by the edge w − v. We show that G can be created like this. It is clear that this holds for
each Gi since they are series-parallel; but since the Gi s are (in general) neither equal nor disjoint, it
is not completely obvious that this also holds for their union G.
We now show, starting with the subgraph G1 which is generalized series-parallel, that we can
consecutively choose one player and add the vertices and edges of her paths which are not already
contained in the subgraph constructed so far by add, series, and parallel operations. Since this again
yields a generalized series-parallel graph, we finally conclude that G is generalized series-parallel.
Let G′ 6= G be the generalized series-parallel subgraph constructed so far. Choose a player i so
that Gi is not node-disjoint with G′ (exists since G is connected). Let Pi be an (si , ti )-path which is
not node-disjoint with G′ and subdivide Pi into the following three subpaths Pi1 , Pi2 , Pi3 (where some
of the subpaths may consist of only one node): Pi1 starts in si and ends in the first node u which is
contained in G′ ; Pi2 starts in u and ends in the last node v which is in G′′ , and Pi3 starts in v and
ends in ti . Note that Gi consists of Pi together with all alternatives of player i (according to Pi ).
The following points show that Gi \ G′ can be added to G′ by series, parallel and add operations:
1. Pi1 (Pi3 ) can obviously be added by an add operation at u (v) and series operations.
2. Pi2 is completely contained in G′′ . Thus Pi2 does not need to be added.
3. Any alternatives where both endnodes are in Pi1 or Pi3 are internal node-disjoint with G′′ and
can therefore be added by parallel and series operations during the addition of Pi1 and Pi3 .
4. Alternatives with both endnodes in Pi2 are already contained in G′′ .
5. There are no alternatives with endnodes in different subpaths.
Note that 2.-5. holds since otherwise there would be a new (sj , tj )-path for a player j already added;
contradiction.
This completes the proof of statement (3). Hence, Theorem 12 is shown.
Remark 13. The first two results of Theorem 12 can be generalized to nonnegative, nondecreasing and discrete-concave shareable edge cost functions. However, we do not know whether or not
polynomial running time can be guaranteed.
We now demonstrate that the assumption of n-sepa graphs is in some sense well justified.
Theorem 14. For n ≥ 3 players, there is a generalized series-parallel graph with fixed edge costs and
no player-specific delays, so that the unique optimal Steiner forest is not enforceable. Therefore, a
black-box reduction as for n-series-parallel graphs is impossible for generalized series-parallel graphs
(even without player-specific delays).
18
Proof. To prove Theorem 14, consider Figure 7. The displayed graph G is generalized series-parallel
since it can be created from a K2 by a sequence of series- and parallel-operations (as executed in
Figure 8). But the unique optimal Steiner forest OPT of (G, (s1 , t1 ), (s2 , t2 ), (s3 , t3 ), c), given by the
solid edges, is not enforceable. To see this, note that the cost of OPT is C(OPT) = 346. Furthermore,
we can upper-bound the sum of cost shares that the players will pay for using their paths in OPT
by 100 + 69 + 170 = 339 < C(OPT), thus showing that OPT is not enforceable: Player 1 will pay
at most 100, because she can use the edge s1 − t1 with cost 100. Player 3 could use the edge s3 − t3
with cost 69, thus she will pay at most 69. It remains to analyze the cost shares of Player 2. Instead
of using the subpath from s2 to s1 of her path in OPT, Player 2 could use the edge s2 − s1 with
cost 84. Furthermore, she could use the edge t3 − t2 with cost 86 instead of her subpath from t3 to
t2 . Since the mentioned subpaths cover the complete path of Player 2 in OPT, she will pay at most
84 + 86 = 170.
For n ≥ 4, we obviously get an instance with the properties stated in Theorem 14 by choosing an
arbitrary node v of G and setting si = ti = v for all i ∈ {4, . . . , n}.
s1
s3
60
84
100
s2
57
t1
71
38
38
69
t3
82
86
t2
Figure 7: Graph G with three source-terminal pairs (s1 , t1 ), (s2 , t2 ), (s3 , t3 ), fixed edge costs c given
on the edges, and no player-specific delays d.
Figure 8: Verification that G is generalized series-parallel.
Algorithm 3: Irredundant
Data: Undirected graph G, source-terminal pairs (s1 , t1 ), . . . , (sn , tn )
Result: Maximum irredundant subgraph G′
1 C ← set of cut vertices of G;
2 for each i ∈ {1, . . . , n} do
3
Gi ← G;
4
for each c ∈ C do
5
Remove from Gi all vertices in the components of G − c containing neither si nor ti (if
any)
6
7
G′ ← G;
Delete from G′ all nodes and edges which are not contained in any Gi .
19
Algorithm 4: n-SePa
Data: Connection Game (N, G, (s1 , t1 ), . . . , (sn , tn ), c, d) with n-series-parallel graph
(G, (s1 , t1 ), . . . , (sn , tn )); strategy profile P = (P1 , . . . , Pn )
Result: Enforceable strategy profile P ′ with cost C(P ′ ) ≤ C(P ); cost share functions ξ so
that P ′ is PNE
1 Solve LP(P ); let (ξi,e )i∈N,e∈Pi be the computed optimal solution;
2 P′ ← P;
3 ξi,e (P ′ ) ← ξi,e for all i ∈ N , e ∈ Pi′ ;
4 if (BB) does not hold then
5
Let S be the set of edges which are not completely paid according to P ′ ;
6
repeat
7
All players with edges in S deviate from all unpaid edges by using a combination of
smallest tight alternatives;
8
Update P ′ accordingly;
9
Update the cost shares (for each player i and each edge e ∈ Pi′ ):
′
ξi,e (P ) =
10
11
12
(
ξi,e , for e ∈ Pi′ ∩ Pi ,
ce , for e ∈ Pi′ \ Pi .
Update S (set of edges which are not completely paid according to P ′ );
until S = ∅;
If there are overpaid edges, decrease the corresponding cost shares arbitrarily until these
edges are exactly paid;
output P ′ and ξ (induced by (ξi,e (P ′ ))i∈N,e∈Pi′ );
Algorithm 5: Alternatives(i)
Data: Connection Game (N, G, (s1 , t1 ), . . . , (sn , tn ), c, d) with n-series-parallel graph
(G, (s1 , t1 ), . . . , (sn , tn )); player i ∈ N ; path Pi ∈ Pi
Result: Set of cheapest alternatives Ai according to Pi
1 Define c̃e := ce + di,e for all e ∈ Gi ;
2 Delete all edges of Pi ;
3 repeat
4
Let u be the first node of Pi which is not yet deleted;
5
repeat
6
Starting in u, execute a (partial) BFS in Gi , stop if a new node of Pi is reached;
7
if a new node v of Pi is reached then
8
Compute a shortest u − v-path Ai (shortest according to c̃);
9
Insert Ai in Ai ;
10
Delete all nodes (different from u, v) and edges of Ai from Gi ;
11
Mark v as visited;
12
else
13
Delete all nodes (∈
/ Pi ) and edges of the connected component of u;
14
Delete u;
15
Remove all marks.
16
17
until u has been deleted;
until all nodes of Pi have been deleted;
20
References
[1] N. Andelman, M. Feldman, and Y. Mansour. Strong price of anarchy. Games Econom. Behav.,
65(2):289–317, 2009.
[2] E. Anshelevich, A. Dasgupta, É. Tardos, and T. Wexler. Near-optimal network design with
selfish agents. Theory of Computing, 4(1):77–109, 2008.
[3] E. Anshelevich and B. Caskurlu. Exact and approximate equilibria for optimal group network
formation. Theor. Comput. Sci., 412(39):5298–5314, 2011.
[4] E. Anshelevich and B. Caskurlu. Price of stability in survivable network design. Theory Comput.
Syst., 49(1):98–138, 2011.
[5] E. Anshelevich, A. Dasgupta, J. Kleinberg, T. Roughgarden, É. Tardos, and T. Wexler. The
price of stability for network design with fair cost allocation. SIAM J. Comput., 38(4):1602–1623,
2008.
[6] G. Avni and T. Tamir. Cost-sharing scheduling games on restricted unrelated machines. Theor.
Comput. Sci., 646:26–39, 2016.
[7] M. Bateni, M. Hajiaghayi, and D. Marx. Approximation schemes for steiner forest on planar
graphs and graphs of bounded treewidth. JACM, 58(5):21:1–21:37, 2011.
[8] V. Biló, A. Fanelli, M. Flammini, and L. Moscardelli. When ignorance helps: Graphical multicast
cost sharing games. Theoretical Computer Science, 411(3):660 – 671, 2010.
[9] V. Bilò, M. Flammini, and L. Moscardelli. The price of stability for undirected broadcast network
design with fair cost allocation is constant. In Proc. 54th Symp. Foundations of Computer Science
(FOCS), pages 638–647, 2013.
[10] C. Bird. On cost allocation for a spanning tree: A game theoretic approach. Networks, 6:335–350,
1976.
[11] J. Byrka and K. Aardal. An optimal bifactor approximation algorithm for the metric uncapacitated facility location problem. SIAM Journal on Computing, 39(6):2212–2231, 2010.
[12] J. Byrka, F. Grandoni, T. Rothvoß, and L. Sanità. Steiner tree approximation via iterative
randomized rounding. J. ACM, 60(1):6:1–6:33, 2013.
[13] I. Caragiannis, V. Gkatzelis, and C. Vinci. Coordination mechanisms, cost-sharing, and approximation algorithms for scheduling. In N. R. Devanur and P. Lu, editors, Web and Internet
Economics, pages 74–87, 2017.
[14] J. Cardinal and M. Hoefer. Non-cooperative facility location and covering games. Theor. Comput. Sci., 411:1855–1876, March 2010.
[15] M. Charikar, C. Chekuri, T.-Y. Cheung, Z. Dai, A. Goel, and S. Guha. Approximation algorithms for directed Steiner problems. J. Algorithms, 33(1):192–200, 1999.
[16] C. Chekuri and M. Pál. A recursive greedy algorithm for walks in directed graphs. In Proc. 46th
Symp. Foundations of Computer Science (FOCS), pages 245–253, 2005.
[17] H.-L. Chen, T. Roughgarden, and G. Valiant. Designing network protocols for good equilibria.
SIAM J. Comput., 39(5):1799–1832, 2010.
21
[18] X. Chen, Z. Diao, and X. Hu. Network characterizations for excluding Braess’s paradox. Theory
Comput. Syst., 59(4):747–780, 2016.
[19] G. Christodoulou, V. Gkatzelis, and A. Sgouritsa. Cost-sharing methods for scheduling games
under uncertainty. In Proc. 18th ACM Conf. Economics and Computation (EC), pages 441–458,
2017.
[20] G. Christodoulou, S. Leonardi, and A. Sgouritsa. Designing cost-sharing methods for bayesian
games. In Proc. 9th Intl. Symp. Algorithmic Game Theory (SAGT), pages 327–339, 2016.
[21] G. Christodoulou and A. Sgouritsa. Designing networks with good equilibria under uncertainty.
In Proc. 27th Symp. Discrete Algorithms (SODA), pages 72–89, 2016.
[22] X. Deng, T. Ibaraki, and H. Nagamochi. Algorithmic aspects of the core of combinatorial
optimization games. Math. Oper. Res., 24(3):751–766, 1999.
[23] Y. Disser, A. Feldmann, M. Klimm, and M. Mihalák. Improving the Hk -bound on the price
of stability in undirected shapley network design games. Theoret. Comput. Sci., 562:557–564,
2015.
[24] M. Feldman and T. Tamir. Conflicting congestion effects in resource allocation games. Oper.
Res., 60(3):529–540, 2012.
[25] A. Fiat, H. Kaplan, M. Levy, S. Olonetzky, and R. Shabo. On the price of stability for designing
undirected networks with fair cost allocations. In Proc. 33rd Intl. Coll. Automata, Languages
and Programming (ICALP), volume 1, pages 608–618, 2006.
[26] N. Garg, G. Konjevod, and R. Ravi. A polylogarithmic approximation algorithm for the Group
Steiner tree problem. J. Algorithms, 37:66–84, 2000.
[27] K. Georgiou and C. Swamy. Black-box reductions for cost-sharing mechanism design. In Proc.
23rd Symp. Discrete Algorithms (SODA), pages 896–913, 2012.
[28] M. X. Goemans and M. Skutella. Cooperative facility location games. J. Algorithms, 50(2):194–
214, 2004.
[29] D. Granot and G. Huberman. On minimum cost spanning tree games. Math. Prog., 21:1–18,
1981.
[30] D. Granot and M. Maschler. Spanning network games. Int. J. Game Theory, 27:467–500, 1998.
[31] S. Guha and S. Khuller. Greedy strikes back: Improved facility location algorithms. J. Algorithms, 31:228–248, 1999.
[32] A. Gupta and A. Kumar. Greedy algorithms for Steiner forest. In Proc. 47th Symp. Theory of
Computing (STOC), pages 871–878, 2015.
[33] T. D. Hansen and O. Telelis. Improved bounds for facility location games with fair cost allocation.
In Proc. 3rd Intl. Conf. Combinatorial Optimization and Applications (COCOA), pages 174–185,
2009.
[34] T. Harks, A. Huber, and M. Surek. A characterization of undirected graphs admitting optimal
cost shares. In N. R. Devanur and P. Lu, editors, Web and Internet Economics, pages 237–251,
Cham, 2017. Springer International Publishing.
[35] T. Harks and B. Peis. Resource buying games. Algorithmica, 70(3):493–512, 2014.
22
[36] T. Harks and P. von Falkenhausen. Optimal cost sharing for capacitated facility location games.
European Journal of Operational Research, 239(1):187–198, 2014.
[37] M. Hoefer. Non-cooperative tree creation. Algorithmica, 53:104–131, 2009.
[38] M. Hoefer. Competitive cost sharing with economies of scale. Algorithmica, 60:743–765, 2011.
[39] M. Hoefer. Strategic cooperation in cost sharing games. Internat. J. Game Theory, 42(1):29–53,
2013.
[40] K. Jain and V. Vazirani. Applications of approximation algorithms to cooperative games. In
Proc. 33rd Symp. Theory of Computing (STOC), pages 364–372, 2001.
[41] J. Könemann, S. Leonardi, G. Schäfer, and S. van Zwam. A group-strategyproof cost sharing
mechanism for the steiner forest game. SIAM J. Comput., 37(5):1319–1341, 2008.
[42] E. Lee and K. Ligett. Improved bounds on the price of stability in network cost sharing games.
In Proc. 14th Conf. Electronic Commerce (EC), pages 607–620, 2013.
[43] J. Li. An o(log(n)/log(log(n))) upper bound on the price of stability for undirected shapley
network design games. Inf. Process. Lett., 109(15):876–878, 2009.
[44] S. Li. A 1.488 approximation algorithm for the uncapacitated facility location problem. Inf.
Comput., 222:45–58, 2013.
[45] N. Megiddo. Cost allocation for Steiner trees. Networks, 8(1):1–6, 1978.
[46] H. Moulin and S. Shenker. Strategyproof sharing of submodular costs: budget balance versus
efficiency. Econom. Theory, 18(3):511–533, 2001.
[47] M. Pál and É. Tardos. Group strategyproof mechanisms via primal-dual algorithms. In FOCS,
pages 584–593, 2003.
[48] G. Robins and A. Zelikovsky. Tighter bounds for graph Steiner tree approximation. SIAM J.
Disc. Math., 19(1):122–134, 2005.
[49] R. Rosenthal. A class of games possessing pure-strategy Nash equilibria. Int. J. Game Theory,
2:65–67, 1973.
[50] V. Syrgkanis. The complexity of equilibria in cost sharing games. In A. Saberi, editor, Proc. 6th
Internat. Workshop on Internet and Network Econom., LNCS, pages 366–377, 2010.
[51] A. Tamir. On the core of network synthesis games. Math. Prog., 50:123–135, 1991.
[52] P. von Falkenhausen and T. Harks. Optimal cost sharing for resource selection games. Math.
Oper. Res., 38(1):184–208, 2013.
23
| 8cs.DS
|
arXiv:1609.07613v1 [math.RA] 24 Sep 2016
Higher-dimensional absolute versions of
symmetric, Frobenius, and quasi-Frobenius
algebras
Mitsuyasu Hashimoto∗
Department of Mathematics, Okayama University
Okayama 700–8530, JAPAN
[email protected]
Abstract
In this paper, we define and discuss higher-dimensional and absolute
versions of symmetric, Frobenius, and quasi-Frobenius algebras. In
particular, we compare these with the relative notions defined by Scheja
and Storch. We also prove the validity of codimension two-argument for
modules over a coherent sheaf of algebras with a 2-canonical module,
generalizing a result of the author.
1.
Introduction
(1.1) Let (R, m) be a semilocal Noetherian commutative ring, and Λ a
module-finite R-algebra. In [6], we defined the canonical module KΛ of Λ.
The purpose of this paper is two fold, each of which is deeply related to KΛ .
This work was supported by JSPS KAKENHI Grant Number 26400045.
2010 Mathematics Subject Classification. Primary 16E65; Secondary 14A15. Key Words
and Phrases. canonical module; symmetric algebra; Frobenius algebra; quasi-Frobenius
algebra; n-canonical module.
∗
1
(1.2) In the first part, we define and discuss higher-dimensional and absolute
notions of symmetric, Frobenius, and quasi-Frobenius algebras and their nonCohen–Macaulay versions. In commutative algebra, the non-Cohen–Macaulay
version of Gorenstein ring is known as quasi-Gorenstein rings. What we discuss here is a non-commutative version of such rings. Scheja and Storch [7]
discussed a relative notion, and our definition is absolute in the sense that it
depends only on Λ and is independent of the choice of R. If R is local, our
quasi-Frobenius property agrees with Gorensteinness discussed by Goto and
Nishida [1], see Proposition 3.7 and Corollary 3.8.
(1.3) In the second part, we show that the codimension-two argument using
the existence of 2-canonical modules in [4] is still valid in non-commutative
settings. For the definition of an n-canonical module, see (2.8). Codimensiontwo argument, which states (roughly speaking) that removing a closed subset
of codimension two or more does not change the category of coherent sheaves
which satisfy Serre’s (S2′ ) condition, is sometimes used in algebraic geometry,
commutative algebra and invariant theory. For example, information on the
canonical sheaf and the class group is retained when we remove the singular
locus of a normal variety over an algebraically closed field, and then these
objects are respectively grasped as the top exterior power of the cotangent
bundle and the Picard group of a smooth variety. In [4], almost principal
bundles are studied. They are principal bundles after removing closed subsets
of codimension two or more.
We prove the following. Let X be a locally Noetherian scheme, U an open
subset of X such that codimX (X \ U) ≥ 2. Let i : U → X be the inclusion.
Let Λ be a coherent OX -algebra. If X possesses a 2-canonical module ω, then
the inverse image i∗ induces the equivalence between the category of coherent
right Λ-modules which satisfy the (S2′ ) condition and the category of coherent
right i∗ Λ-modules which satisfy the (S2′ ) condition. The quasi-inverse is given
by the direct image i∗ . What was proved in [4] was the case that Λ = OX .
If, moreover, ω = OX (that is to say, X satisfy the (S2 ) and (G1 ) condition),
then the assertion has been well-known, see [3].
(1.4) 2-canonical modules are ubiquitous in algebraic geometry. If I is a
dualizing complex of a Noetherian scheme X, then the lowest non-vanishing
cohomology group of I is semicanonical. A rank-one reflexive sheaf over a
normal variety is 2-canonical.
2
(1.5) Section 2 is for preliminaries. Section 3 is devoted to the discussion
of the first theme mentioned in the paragraph (1.2), while Section 4 is for the
second theme mentioned in (1.3).
(1.6) Acknowledgments: Special thanks are due to Professor Osamu Iyama
for valuable advice and discussion.
The essential part of this paper has first appeared as [5, sections 9–10].
When it is published as [6], they have been removed after the requirement to
shorten the paper (also, the title has been changed slightly). Here we revive
them as an independent paper.
2.
Preliminaries
(2.1) Throughout this paper, R denotes a Noetherian commutative ring.
For a module-finite R-algebra Λ, a Λ-module means a left Λ-module. Λop
denotes the opposite algebra of Λ, and thus a Λop -module is identified with a
right Λ-module. A Λ-bimodule means a Λ ⊗R Λop -module. The category of
finite Λ-modules is denoted by Λ mod. The category Λop mod is also denoted
by mod Λ.
(2.2) Let (R, m) be semilocal and Λ be a module-finite R-algebra. For an
R-module M, the m-adic completion of M is denoted by M̂. For a finite
Λ-module M, by dim M or dimΛ M we mean dimR M, which is independent
of the choice of R. By depth M or depthΛ M we mean depthR (m, M), which
is independent of R. We say that M is globally Cohen–Macaulay (GCM for
short) if dim M = depth M. We say that M is globally maximal Cohen–
Macaulay (GMCM for short) if dim Λ = depth M. If R happens to be local,
then M is GCM (resp. GMCM) if and only if M is Cohen–Macaulay (resp.
maximal Cohen–Macaulay) as an R-module.
(2.3) For M ∈ Λ mod, we say that M satisfies (Sn′ )Λ,R , (Sn′ )R or (Sn′ ) if
depthRP MP ≥ min(n, htR P ) for every P ∈ Spec R (this notion depends on
R).
(2.4) Let X be a locally Noetherian scheme and Λ a coherent OX -algebra.
For a coherent Λ-module M, we say that M satisfies (Sn′ ) or (Sn′ )Λ,X , or
sometimes M ∈ (Sn′ )Λ,X , if depthOX,x Mx ≥ min(n, dim OX,x ) for every x ∈
X.
3
(2.5) Assume that (R, m) is complete semilocal, and Λ 6= 0 a module-finite
R-algebra. Let I be a normalized dualizing complex of R. The lowest noni
vanishing cohomology group Ext−s
R (Λ, I) (ExtR (Λ, I) = 0 for i < −s) is denoted by KΛ , and is called the canonical module of Λ. If Λ = 0, then we define
that KΛ = 0. For basics on the canonical modules, we refer the reader to [6].
Note that KΛ depends only on Λ, and is independent of R.
(2.6) Assume that (R, m) is semilocal which may not be complete. We say
that a finitely generated Λ-bimodule K is a canonical module of Λ if K̂ is
isomorphic to the canonical module KΛ̂ as a Λ̂-bimodule. It is unique up to
isomorphisms, and denoted by KΛ . We say that K ∈ mod Λ is a right (resp.
left) canonical module of Λ if K̂ is isomorphic to KΛ̂ in mod Λ̂ (resp. Λ̂ mod).
If KΛ exists, then K is a right canonical module if and only if K ∼
= KΛ in
mod Λ.
(2.7) We say that ω is an R-semicanonical right Λ-module if for any P ∈
Spec R, RP ⊗R ω is the right canonical module RP ⊗R Λ for any P ∈ suppR ω.
(2.8) Let C ∈ mod Λ. We say that C is an n-canonical right Λ-module over
R if C ∈ (Sn′ )R , and for each P ∈ Spec R with ht P < n, we have that CP is
an RP -semicanonical right ΛP -module.
3.
Symmetric and Frobenius algebras
(3.1) Let (R, m) be a Noetherian semilocal ring, and Λ a module-finite Ralgebra. Let KΛ denote the canonical module of Λ, see [6].
We say that Λ is quasi-symmetric if Λ is the canonical module of Λ. That
is, Λ ∼
= KΛ as Λ-bimodules. It is called symmetric if it is quasi-symmetric
and GCM. Note that Λ is quasi-symmetric (resp. symmetric) if and only if
Λ̂ is so. Note also that quasi-symmetric and symmetric are absolute notion,
and is independent of the choice of R in the sense that the definition does not
change when we replace R by the center of Λ, because KΛ is independent of
the choice of R.
(3.2) For (non-semilocal) Noetherian ring R, we say that Λ is locally quasisymmetric (resp. locally symmetric) over R if for any P ∈ Spec R, ΛP is a
quasi-symmetric (resp. symmetric) RP -algebra. This is equivalent to say that
for any maximal ideal m of R, Λm is quasi-symmetric (resp. symmetric), see
[6, (7.6)].
4
In the case that (R, m) is semilocal, Λ is locally quasi-symmetric (resp.
locally symmetric) over R if it is quasi-symmetric (resp. symmetric), but the
converse is not true in general.
Lemma 3.3. Let (R, m) be a Noetherian semilocal ring, and Λ a module-finite
R-algebra. Then the following are equivalent.
1 ΛΛ is the right canonical module of Λ.
2
ΛΛ
is the left canonical module of Λ.
Proof. We may assume that R is complete. Then replacing R by a Noether
normalization of R/ annR Λ, we may assume that R is regular and Λ is a
faithful R-module.
We prove 1⇒2. By [6, Lemma 5.10], KΛ satisfies (S2′ )R . By assumption, ΛΛ satisfies (S2′ )R . As R is regular and dim R = dim Λ, KΛ = Λ∗ =
HomR (Λ, R). So we get an R-linear map
ϕ : Λ ⊗R Λ → R
such that ϕ(ab ⊗ c) = ϕ(a ⊗ bc) and that the induced map h : Λ → Λ∗
given by h(a)(c) = ϕ(a ⊗ c) is an isomorphism (in mod Λ). Now ϕ induces a
homomorphism h′ : Λ → Λ∗ in Λ mod given by h′ (c)(a) = ϕ(a ⊗ c). To verify
that this is an isomorphism, as Λ and Λ∗ are reflexive R-modules, we may
localize at a prime P of R of height at most one, and then take a completion,
and hence we may further assume that dim R ≤ 1. Then Λ is a finite free
R-module, and the matrices of h and h′ are transpose each other. As the
matrix of h is invertible, so is that of h′ , and h′ is an isomorphism.
2⇒1 follows from 1⇒2, considering the opposite ring.
Definition 3.4. Let (R, m) be semilocal. We say that Λ is a pseudo-Frobenius
R-algebra if the equivalent conditions of Lemma 3.3 are satisfied. If Λ is GCM
in addition, then it is called a Frobenius R-algebra. Note that these definitions
are independent of the choice of R. Moreover, Λ is pseudo-Frobenius (resp.
Frobenius) if and only if Λ̂ is so. For a general R, we say that Λ is locally
pseudo-Frobenius (resp. locally Frobenius) over R if ΛP is pseudo-Frobenius
(resp. Frobenius) for P ∈ Spec R.
Lemma 3.5. Let (R, m) be semilocal. Then the following are equivalent.
1 (KΛ̂ )Λ̂ is projective in mod Λ̂.
5
2
Λ̂ (KΛ̂ )
is projective in Λ̂ mod,
where ˆ? denotes the m-adic completion.
Proof. We may assume that (R, m, k) is complete regular local and Λ is a
faithful R-module. Let ¯? denote the functor k⊗R ?. Then Λ̄ is a finite dimensional k-algebra. So mod Λ̄ and Λ̄ mod have the same number of simple
modules, say n. An indecomposable projective module in mod Λ is nothing
but the projective cover of a simple module in mod Λ̄. So mod Λ and Λ mod
have n indecomposable projectives. Now HomR (?, R) is an equivalence between add(KΛ )Λ and add Λ Λ. It is also an equivalence between add Λ (KΛ ) and
add ΛΛ . So both add(KΛ )Λ and add Λ (KΛ ) also have n indecomposables. So 1
is equivalent to add(KΛ )Λ = add ΛΛ . 2 is equivalent to add Λ (KΛ ) = add Λ Λ.
So 1⇔2 is proved simply applying the duality HomR (?, R).
(3.6) Let (R, m) be semilocal. If the equivalent conditions in Lemma 3.5
are satisfied, then we say that Λ is pseudo-quasi-Frobenius. If it is GCM
in addition, then we say that it is quasi-Frobenius. These definitions are
independent of the choice of R. Note that Λ is pseudo-quasi-Frobenius (resp.
quasi-Frobenius) if and only if Λ̂ is so.
Proposition 3.7. Let (R, m) be semilocal. Then the following are equivalent.
1 Λ is quasi-Frobenius.
2 Λ is GCM, and dim Λ = idim Λ Λ, where idim denotes the injective dimension.
3 Λ is GCM, and dim Λ = idim ΛΛ .
Proof. 1⇒2. By definition, Λ is GCM. To prove that dim Λ = idim Λ Λ, we
may assume that R is local. Then by [1, (3.5)], we may assume that R
is complete. Replacing R by the Noetherian normalization of R/ annR Λ,
we may assume that R is a complete regular local ring of dimension d, and
Λ its maximal Cohen–Macaulay (that is, finite free) module. As add Λ Λ =
add Λ (KΛ ) by the proof of Lemma 3.5, it suffices to prove idim Λ (KΛ ) = d.
Let IR be the minimal injective resolution of the R-module R. Then J =
HomR (Λ, IR ) is an injective resolution of KΛ = HomR (Λ, R) as a left Λ-module.
As the length of J is d and
ExtdΛ (Λ/mΛ, KΛ ) ∼
= ExtdR (Λ/mΛ, R) 6= 0,
6
we have that idim Λ (KΛ ) = d.
2⇒1. We may assume that R is complete regular local and Λ is maximal
Cohen–Macaulay. By [1, (3.6)], we may further assume that R is a field.
Then Λ Λ is injective. So (KΛ )Λ = HomR (Λ, R) is projective, and Λ is quasiFrobenius, see [8, (IV.3.7)].
1⇔3 is proved similarly.
Corollary 3.8. Let R be arbitrary. Then the following are equivalent.
1 For any P ∈ Spec R, ΛP is quasi-Frobenius.
2 For any maximal ideal m of R, Λm is quasi-Frobenius.
3 Λ is a Gorenstein R-algebra in the sense that Λ is a Cohen–Macaulay
R-module, and idimΛP ΛP ΛP = dim ΛP for any P ∈ Spec R.
Proof. 1⇒2 is trivial.
2⇒3. By Proposition 3.7, we have idim Λm Λm = dim Λm for each m. Then
by [1, (4.7)], Λ is a Gorenstein R-algebra.
3⇒1 follows from Proposition 3.7.
(3.9) Let R be arbitrary. We say that Λ is a quasi-Gorenstein R-algebra if
ΛP is pseudo-quasi-Frobenius for each P ∈ Spec R.
Definition 3.10 (Scheja–Storch [7]). Let R be general. We say that Λ is
symmetric (resp. Frobenius) relative to R if Λ is R-projective, and Λ∗ :=
HomR (Λ, R) is isomorphic to Λ as a Λ-bimodule (resp. as a right Λ-module).
It is called quasi-Frobenius relative to R if the right Λ-module Λ∗ is projective.
Lemma 3.11. Let (R, m) be local.
1 If dim Λ = dim R, R is quasi-Gorenstein, and Λ∗ ∼
= Λ as Λ-bimodules
∗ ∼
∗
(resp. Λ = Λ as right Λ-modules, Λ is projective as a right Λ-module),
then Λ is quasi-symmetric (resp. pseudo-Frobenius, pseudo-quasi-Frobenius).
2 If R is Gorenstein and Λ is symmetric (resp. Frobenius, quasi-Frobenius)
relative to R, then Λ is symmetric (resp. Frobenius, quasi-Frobenius).
3 If Λ is nonzero and R-projective, then Λ is quasi-symmetric (resp. pseudoFrobenius, pseudo-quasi-Frobenius) if and only if R is quasi-Gorenstein
and Λ is symmetric (resp. Frobenius, quasi-Frobenius) relative to R.
7
4 If Λ is nonzero and R-projective, then Λ is symmetric (resp. Frobenius,
quasi-Frobenius) if and only if R is Gorenstein and Λ is symmetric (resp.
Frobenius, quasi-Frobenius) relative to R.
Proof. We can take the completion, and we may assume that R is complete
local.
1. Let d = dim Λ = dim R, and let I be the normalized dualizing complex
(see [6, (5.2)]) of R. Then
−d
∼
KΛ = Ext−d
(I)) ∼
= Hom(Λ, KR ) ∼
= Hom(Λ, R) = Λ∗
R (Λ, I) = HomR (Λ, H
as Λ-bimodules, and the result follows.
2. We may assume that Λ is nonzero. As R is Cohen–Macaulay and Λ is
a finite projective R-module, Λ is a maximal Cohen–Macaulay R-module. By
1, the result follows.
3. The ‘if’ part follows from 1. We prove the ‘only if’ part. As Λ
is R-projective and nonzero, dim Λ = dim R. As Λ is R-finite free, KΛ ∼
=
∗
∗
Λ
⊗
K
.
As
K
is
R-free
and
Λ
⊗
K
is
nonzero
and
is
HomR (Λ, KR ) ∼
=
R
R
Λ
R
R
isomorphic to a direct sum of copies of KR , we have that KR is R-projective,
and hence R is quasi-Gorenstein, and KR ∼
= Λ∗ , and the
= R. Hence KΛ ∼
result follows.
4 follows from 3 easily.
(3.12) Let (R, m) be semilocal. Let a finite group G act on Λ by R-algebra
automorphisms.
L Let Ω = Λ ∗ G, the twisted group algebra. That is, Ω =
Λ ⊗R RG =
g∈G Λg as an R-module, and the product of Ω is given by
′ ′
(ag)(a g ) = (a(ga′ ))(gg ′) for a, a′ ∈ Λ and g, g ′ ∈ G. This makes Ω a modulefinite R-algebra.
(3.13) We simply call an RG-module a G-module. We say that M is a
(G, Λ)-module if M is a G-module, Λ-module, the R-module structures coming
from that of the G-module structure and the Λ-module structure agree, and
g(am) = (ga)(gm) for g ∈ G, a ∈ Λ, and m ∈ M. A (G, Λ)-module and an
Ω-module are one and the same thing.
(3.14) By the action ((a ⊗ a′ )g)a1 = a(ga1 )a′ , we have that Λ is a (Λ ⊗
Λop ) ∗ G-module in a natural way. So it is an Ω-module by the action (ag)a1 =
a(ga1 ). It is also a right Ω-module by the action a1 (ag) = g −1 (a1 a). If the
action of G on Λ is trivial, then these actions make an Ω-bimodule.
8
(3.15) Given an Ω-module M and an RG-module V , M ⊗R V is an Ωmodule by (ag)(m ⊗ v) = (ag)m ⊗ gv. HomR (M, V ) is a right Ω-module by
(ϕ(ag))(m) = g −1 ϕ(a(gm)). It is easy to see that the standard isomorphism
HomR (M ⊗R V, W ) → HomR (M, HomR (V, W ))
is an isomorphism of right Ω-modules for a left Ω-module M and G-modules
V and W .
(3.16) Now consider the case Λ = R. Then the pairing φ : RG ⊗R RG → R
given by φ(g ⊗ g ′ ) = δgg′ ,e (Kronecker’s delta) is non-degenerate, and induces
an RG-bimodule isomorphism Ω = RG → (RG)∗ = Ω∗ . As Ω = RG is a finite
free R-module, we have that Ω = RG is symmetric relative to R.
Lemma 3.17. If Λ is quasi-symmetric (resp. symmetric) and the action of G
on Λ is trivial, then Ω is quasi-symmetric (resp. symmetric).
Proof. Taking the completion, we may assume that R is complete. Then
replacing R by a Noether normalization of R/ annR Λ, we may assume that R
is a regular local ring, and Λ is a faithful R-module. As the action of G on Λ
is trivial, Ω = Λ ⊗R RG is quasi-symmetric (resp. symmetric), as can be seen
easily.
(3.18) In particular, if Λ is commutative quasi-Gorenstein (resp. Gorenstein)
and the action of G on Λ is trivial, then Ω = ΛG is quasi-symmetric (resp.
symmetric).
(3.19) In general, Ω Ω ∼
= Λ ⊗R RG as Ω-modules.
Lemma 3.20. Let M and N be right Ω-modules, and let ϕ : M → N be a
homomorphism of right Λ-modules. Then ψ : M ⊗ RG → N ⊗ RG given by
ψ(m ⊗ g) = g(ϕ(g −1m)) ⊗ g is an Ω-homomorphism. In particular,
1 If ϕ is a Λ-isomorphism, then ψ is an Ω-isomorphism.
2 If ϕ is a split monomorphism in mod Λ, then ψ is a split monomorphism
in mod Ω.
Proof. Straightforward.
Proposition 3.21. Let G be a finite group acting on Λ. Set Ω := Λ ∗ G.
1 If the action of G on Λ is trivial and Λ is quasi-symmetric (resp. symmetric), then so is Ω.
9
2 If Λ is pseudo-Frobenius (resp. Frobenius), then so is Ω.
3 If Λ is pseudo-quasi-Frobenius (resp. quasi-Frobenius), then so is Ω.
Proof. 1 is Lemma 3.17. To prove 2 and 3, we may assume that (R, m) is
complete regular local and Λ is a faithful module.
2.
(KΩ )Ω ∼
= HomR (Λ, R) ⊗ (RG)∗ ∼
= KΛ ⊗ RG
= HomR (Λ ⊗R RG, R) ∼
as right Ω-modules. It is isomorphic to ΛΩ ⊗ RG ∼
= ΩΩ by Lemma 3.20, 1,
∼
since KΛ = Λ in mod Λ. Hence Ω is pseudo-Frobenius. If, in addition, Λ is
Cohen–Macaulay, then Ω is also Cohen–Macaulay, and hence Ω is Frobenius.
3 is proved similarly, using Lemma 3.20, 2.
Note that the assertions for Frobenius and quasi-Frobenius properties also
follow easily from Lemma 3.11 and [7, (3.2)].
4.
Codimension-two argument
(4.1) This section is the second part of this paper. In this section, we show
that the codimension-two argument using the existence of 2-canonical modules
in [4] is still valid in non-commutative settings, as announced in (1.3).
(4.2) Let X be a locally Noetherian scheme, U its open subscheme, and Λ
a coherent OX -algebra. Let i : U ֒→ X be the inclusion.
(4.3) Let M ∈ mod Λ. That is, M is a coherent right Λ-module. Then by
restriction, i∗ M ∈ mod i∗ Λ.
(4.4) For a quasi-coherent i∗ Λ-module N , we have an action
1⊗u
a
i∗ N ⊗OX Λ −−→ i∗ N ⊗OX i∗ i∗ Λ → i∗ (N ⊗OU i∗ Λ) −
→ i∗ N ,
where u is the unit map for the adjoint pair (i∗ , i∗ ). So we get a functor
i∗ : Mod i∗ Λ → Mod Λ, where Mod i∗ Λ (resp. Mod Λ) denote the category of
quasi-coherent i∗ Λ-modules (resp. Λ-modules).
Lemma 4.5. Let the notation be as above. Assume that U is large in X
op
(that is, codimX (X \ U) ≥ 2). If M ∈ (S2′ )Λ ,X , then the canonical map
u : M → i∗ i∗ M is an isomorphism.
10
Proof. Follows immediately from [4, (7.31)].
Proposition 4.6. Let the notation be as above, and let U be large in X. Assume that there is a 2-canonical right Λ-module. Then we have the following.
1 If N ∈ (S2′ )i
∗ Λop ,U
, then i∗ N ∈ (S2′ )Λ
2 i∗ : (S2′ )Λ ,X → (S2′ )i
inverse each other.
op
∗ Λop ,U
op ,X
.
and i∗ : (S2′ )i
∗ Λop ,U
→ (S2′ )Λ
op ,X
are quasi-
Proof. The question is local, and we may assume that X is affine.
1. There is a coherent subsheaf Q of i∗ N such that i∗ Q = i∗ i∗ N = N by
[2, Exercise II.5.15]. Let V be the Λ-submodule of i∗ N generated by Q. That
is, the image of the composite
Q ⊗OX Λ → i∗ N ⊗OX Λ → i∗ N .
Note that V is coherent, and i∗ Q ⊂ i∗ V ⊂ i∗ i∗ N = i∗ Q = N .
Let C be a 2-canonical right Λ-module. Let ?† := HomΛop (?, C), Γ =
EndΛ C, and ?‡ := HomΓ (?, C). Let M be the double dual V †‡ . Then M ∈
op
(S2′ )Λ ,X , and hence
M∼
= i∗ i∗ M ∼
= i∗ i∗ (V †‡ ) ∼
= i∗ (i∗ V)†‡ ∼
= i∗ (N †‡ ) ∼
= i∗ N .
op
So i∗ N ∼
= M lies in (S2′ )Λ ,X .
2 follows from 1 and Lemma 4.5 immediately.
References
[1] S. Goto and K. Nishida, Towards a theory of Bass numbers with
applications to Gorenstein algebras, Colloq. Math. 91 (2002), 191–
253.
[2] R. Hartshorne, Algebraic Geometry, Graduate Texts in Math. 52,
Springer Verlag (1977).
[3] R. Hartshorne, Generalized divisors on Gorenstein schemes, KTheory 8 (1994), 287–339.
[4] M. Hashimoto, Equivariant class group. III. Almost principal fiber
bundles, arXiv:1503.02133v1
11
[5] M. Hashimoto, Canonical and n-canonical modules on a Noetherian
algebra, arXiv:1508.07552v1
[6] M. Hashimoto, Canonical and n-canonical modules of a Noetherian
algebra, to appear in Nagoya Math. J.
[7] G. Scheja and U. Storch, Quasi-Frobenius-Algebren und lokal
vollständige Durchschnitte, Manuscripta Math. 19 (1976), 75–104.
Addendum: 20 (1977), 99–100.
[8] A. Skowroński and K. Yamagata, Frobenius Algebras. I., Basic Representation Theory, European Mathematical Society (2011).
12
| 0math.AC
|
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.