Documente Academic
Documente Profesional
Documente Cultură
SIAM Monographs on
Discrete Mathematics and Applications
The series includes advanced monographs reporting on the most recent theoretical, computational, or
applied developments in the field; introductory volumes aimed at mathematicians and other mathematically
motivated readers interested in understanding certain areas of pure or applied combinatorics; and graduate
textbooks. The volumes are devoted to various areas of discrete mathematics and its applications.
Mathematicians, computer scientists, operations researchers, computationally oriented natural and social
scientists, engineers, medical researchers, and other practitioners will find the volumes of interest.
Editor-in-Chief
Peter L. Hammer, RUTCOR, Rutgers, The State University of New Jersey
Editorial Board
A. V. Kostochka, Siberian Branch of the Russian Academy of
M. Aigner, Freie Universitat Berlin, Germany Sciences, Russia
N. Alon, Tel Aviv University, Israel F. T. Leighton, Massachusetts Institute of Technology, USA
E. Balas, Carnegie Mellon University, USA T. Lengauer, Gesellschaft fur Mathematik und
J- C. Bermond, UniversitedeNice-SophiaAntipolis, France Datenverarbeitung mbH, Germany
J. Berstel, Universite Marne-la-Vallee, France S. Martello, DEIS University of Bologna, Italy
N. L. Biggs, The London School of Economics, United Kingdom M. Minoux, Universite Pierre et Marie Curie, France
B. Bollobas, University of Memphis, USA R. Mb'hring, Technische Universitat Berlin, Germany
R. E. Burkard, Technische Universitat Graz, Austria C. L. Monma, Bellcore, USA
D. G. Cornell, University of Toronto, Canada J. Nesetril, Charles University, Czech Republic
I. Gessel, Brandeis University, USA W. R. Pulleyblank, IBM T. J. Watson Research Center, USA
F. Glover, University of Colorado, USA A. Recski, Technical University of Budapest, Hungary
M. C. Golumbic, Bar-Han University, Israel C. C. Ribeiro, Catholic University of Rio de Janeiro, Brazil
R. L. Graham, AT&T Research, USA H. Sachs, Technische Universitat llmenau, Germany
A. J. Hoffman, IBM T. J. Watson Research Center, USA A. Schrijver, CWI, The Netherlands
T. Ibaraki, Kyoto University, Japan R. Shamir, Tel Aviv University, Israel
H. Imai, University of Tokyo, Japan N. J. A. Sloane, AT&T Research, USA
M. Karoriski, Adam Mickiewicz University, Poland, and Emory W. T. Trotter, Arizona State University, USA
University, USA D. J. A. Welsh, University of Oxford, United Kingdom
R. M. Karp, University of Washington, USA D. de Werra, Ecole Polytechnique Federate de Lausanne,
V. Klee, University of Washington, USA Switzerland
K. M. Koh, National University of Singapore, Republic of P. M. Winkler, Bell Labs, Lucent Technologies, USA
Singapore Yue Minyi, Academia Sinica, People's Republic of China
B. Korte, Universitat Bonn, Germany
Series Volumes
Murota, K., Discrete Convex Analysis
Toth, P. and Vigo, D., The Vehicle Routing Problem
Anthony, M., Discrete Mathematics of Neural Networks: Selected Topics
Creignou, N., Khanna, S., and Sudan, M., Complexity Classifications of Boolean Constraint Satisfaction Problems
Hubert, L., Arable, P., and Meulman, J., Combinatorial Data Analysis: Optimization by Dynamic Programming
Peleg, D., Distributed Computing: A Locality-Sensitive Approach
Wegener, I., Branching Programs and Binary Decision Diagrams: Theory and Applications
Brandstadt, A., Le, V. B., and Spinrad, J. P., Graph Classes: A Survey
McKee, T. A. and McMorris, F. R., Topics in Intersection Graph Theory
Grilli di Cortona, P., Manzi, C., Pennisi, A., Ricca, R, and Simeone, B., Evaluation and Optimization of Electoral Systems
DISCRETE CONVEX ANALYSIS
KAZUO MUROTA
University of Tokyo; PRESTO, JST
Tokyo, Japan
10 9 8 7 6 5 4 3 2 1
All rights reserved. Printed in the United States of America. No part of this book may
be reproduced, stored, or transmitted in any manner without the written permission of
the publisher. For information, write to the Society for Industrial and Applied
Mathematics, 3600 University City Science Center, Philadelphia, PA 19104-2688.
QA331.5.M87 2003
515'.8—dc21
2003042468
is a registered trademark.
Contents
List of Figures xi
Notation xiii
Preface xxi
v
vi Contents
10 Algorithms 281
10.1 Minimization of M-Convex Functions 281
10.1.1 Steepest Descent Algorithm 281
10.1.2 Steepest Descent Scaling Algorithm 283
10.1.3 Domain Reduction Algorithm 284
10.1.4 Domain Reduction Scaling Algorithm 286
10.2 Minimization of Submodular Set Functions 288
10.2.1 Basic Framework 288
10.2.2 Schrijver's Algorithm 293
10.2.3 Iwata-Fleischer-Fujishige's Algorithm 296
10.3 Minimization of L-Convex Functions 305
10.3.1 Steepest Descent Algorithm 305
10.3.2 Steepest Descent Scaling Algorithm 308
10.3.3 Reduction to Submodular Function Minimization . . 308
10.4 Algorithms for M-Convex Submodular Flows 308
10.4.1 Two-Stage Algorithm 309
10.4.2 Successive Shortest Path Algorithm 311
10.4.3 Cycle-Canceling Algorithm 312
10.4.4 Primal-Dual Algorithm 313
10.4.5 Conjugate Scaling Algorithm 318
Bibliographical Notes 321
Bibliography 363
Index 379
This page intentionally left blank
List of Figures
xi
xii List of Figures
0: = ( 0 , 0 , . . . , 0 )
1: =(1, !,...,!)
2V: set of all subsets of set V (i.e., power set of V)
V: "for all," "for any," or "for each"
3: "there exists" or "for some"
T
: transpose of a vector or a matrix
+: sum, Minkowski sum (3.21), (3.52)
Q : infimal convolution over R™ (3.20)
DZ : infimal convolution over Z™ (i.e., integer infimal convolution) (6.43)
V: componentwise maximum (1-28)
V: "join" operation in a lattice Note 10.15
A: componentwise minimum (1-28)
A: "meet" operation in a lattice Note 10.15
| • |: cardinality (number of elements) of a set
[•,•]: interval (of reals or integers) (3-1), (3.54)
[-, -]R: interval of real numbers (3-1)
[•, -]z- interval of integers (3.54)
(•,•): inner product, pairing (1-7), (3.18)
|| • ||i: fx-norm of a vector (4.2)
norm a
II • ||oo: ^oo- °f vector (3.60)
"•": convex hull of a set, convex closure of a function (3.56)
[ • 1: rounding up to the nearest integer section 3.4
[ • J: rounding down to the nearest integer section 3.4
&R/(X): subdifferential of (convex) function / at x (3.23), (6.86)
dzf(x): integral subdifferential of (convex) function / at x (6.88)
d'^h(x): subdifferential of (concave) function h at x (8.19)
d'zh(x): integral subdifferential of (concave) function h at x (8.19)
d+a: initial vertex of arc a section 2.2
d~a: terminal vertex of arc a section 2.2
d£: boundary of flow £ (2.27)
<9H: set of boundaries of feasible flows section 9.3
XIII
XIV Notation
xxi
xxii Preface
Introduction to the
Central Concepts
1.1.1 Aim
An optimization problem, or a mathematical programming problem, may be ex-
pressed generically as follows:
This means that we are to find an x that minimizes the value of f ( x ) subject to
the constraint that x should belong to the set 5. Both / and S are given as the
problem data, whereas x is a variable to be determined. The function / is called
the objective function and the set S the feasible set.
In continuous optimization, the variable x typically denotes a finite-dimensional
real vector, say, x 6 R™, and accordingly we have S C R™ and / : R™ —> R (or
/ : S —> R).1 An optimization problem with 5 a convex set and / a convex function
lr
The notation R means the set of all real numbers and Rn the set of n-dimensional real vectors.
1
2 Chapter 1. Introduction to the Central Concepts
is referred to as a convex program, where a set S is convex if the line segment join-
ing any two points in S is contained in S (see Fig. 1.1) and a function / : S —> R
defined on a convex set S is convex if
whenever x, y 6 S and 0 < A < 1 (see Fig. 1.2). Convex programs constitute a class
of optimization problems that are tractable both theoretically and practically, with
a firm theoretical basis provided by "convex analysis." The tractability of convex
programs is largely based on the following properties of convex functions:
2. A flow is optimal if and only if there exists a potential on the vertex set
such that the reduced arc cost with respect to the potential is nonnegative
on every arc. This is a duality statement characterizing the optimality of
a flow in terms of the dual variable (potential). This provides the basis for
primal-dual algorithms.
In more abstract terms, it is accepted that the tractability of the network flow
problems stems from the matroidal structure (or submodularity) inherent therein.
Whereas the meaning of this statement will be substantiated later, it is mentioned
at this point that a matroid is an abstract combinatorial object defined as a pair
of a finite set, say, V, and a family B of subsets of V that satisfies certain abstract
axioms. We refer to V as the ground set, a member of B as a base, and a subset
of a base as an independent set. The matroid is considered to be fundamental in
combinatorial optimization, which is evidenced by the following facts:2
1. A base is optimal with respect to a given weight vector if and only if it
cannot be improved by an elementary exchange, which means a modification
of a base B to another base (B \ {«}) U {v} with u in B and v not in B.
Thus, the local optimality with respect to elementary exchanges guarantees
the global optimality. Moreover, an optimal base can be found by the so-called
greedy algorithm, which may be compared to the steepest descent algorithm
in nonlinear optimization.
functions and convex functions. It was shown by Prank that Edmonds's inter-
section theorem can be rewritten as a separation theorem for a pair of submod-
ular/supermodular functions, with an integrality (discreteness) assertion for the
separating hyperplane in the case of integer-valued functions. Another reformula-
tion of Edmonds's intersection theorem is Fujishige's Fenchel-type min-max duality
theorem for a pair of submodular/supermodular functions, again with an integrality
assertion in the case of integer-valued functions. A precise statement, beyond anal-
ogy, about the relationship between submodular functions and convex functions was
made by Lovasz: A set function is submodular if and only if the so-called Lovasz
extension of that set function is convex. These results led to the recognition that
the essence of the duality for submodular/supermodular functions consists of the
discreteness (integrality) assertion in addition to the duality for convex/concave
functions. Namely,
The theory emphasizes duality and conjugacy with a view to providing a novel
duality framework for nonlinear integer programming. It may be in order to mention
that the present theory extends the direction set forth by J. Edmonds, A. Frank,
S. Fujishige, and L. Lovasz (see section 1.1.2), but it is, rather, independent of the
convexity arguments in the theories of greedoids, antimatroids, convex geometries,
and oriented matroids (Bjorner-Las Vergnas-Sturmfels-White-Ziegler [16], Korte-
Lovasz-Schrader [114]).
Two convexity concepts, called L-convexity and M-convexity, play primary
roles in the present theory. L-convex functions and M-convex functions are both
(extensible to) convex functions and they are conjugate to each other through a
discrete version of the Legendre-Fenchel transformation. L-convex functions and M-
convex functions generalize, respectively, the concepts of submodular set functions
and base polyhedra. It is noted that the "L" in "L-convexity" stands for "lattice"
and the "M" in "M-convexity" for "matrbid."
1.1.2 History
This section is devoted to an account of the history of discrete convex functions
in matroid theory that led to L-convex and M-convex functions (see Table 1.1).
There are, however, many other previous and recent studies on discrete convex-
ity outside the literature of the matroid (Hochbaum-Sharnir-Shanthikumar [92],
Ibaraki-Katoh [93], Kindler [112], Miller [130], and so on).
The concept of matroids was introduced by H. Whitney [218] in 1935, together
with the equivalence between the submodularity of rank functions and the exchange
property of independent sets. This equivalence is the germ of the conjugacy between
L-convex and M-convex functions in the present theory of discrete convex analysis.
In the late 1960s, J. Edmonds found a fundamental duality theorem on the
intersection problem for a pair of (poly)matroids. This theorem, Edmonds's inter-
section theorem, shows a min-max relation between the maximum of a common
independent set and the minimum of a submodular function derived from the rank
functions. The famous article of Edmonds [44] convinced us of the fundamental
role of submodularity in discrete optimization. Analogies of submodular functions
to convex functions and to concave functions were discussed at the same time. The
min-max relation supported the analogy to convex functions, whereas some other
facts pointed to concave functions. No unanimous conclusion was reached at this
point.
The relationship between submodular functions and convex functions, which
was made clear in the early 1980s through the work of A. Frank, S. Fujishige,
and L. Lovasz, was described in section 1.1.1 but is mentioned again in view of
its importance. The fundamental relationship between submodular functions and
convex functions, due to Lovasz [123], says that a set function is submodular if
3
The notation Z means the set of all integers and Zn the set of n-dimensional integer vectors.
6 Chapter 1. Introduction to the Central Concepts
as well as algorithms for the weighted matroid intersection, are generalized to the
valuated matroid intersection (Murota [136]). An analogy of matroid valuations to
concave functions resulted in a Fenchel-type min-max duality theorem for matroid
valuations (Murota [139]). This Fenchel-type duality is neither a generalization nor
a special case of Fujishige's Fenchel-type duality for submodular functions, but these
two can be generalized into a single min-max equation, which is the Fenchel-type
duality theorem in the present theory.
A further analogy of valuated matroids to concave functions led to the concept
of M-convex/concave functions in Murota [137], 1996. M-convexity is a concept of
"convexity" for functions denned on integer lattice points in terms of an exchange
axiom and affords a common generalization of valuated matroids and (integral)
polymatroids. A valuated matroid can be identified with an M-concave function
denned on {0, l}-vectors. The base polyhedron of an integral polymatroid is a syn-
onym for a {0, +oc}-valued M-convex function. The valuated matroid intersection
problem and the polymatroid intersection problem are unified into the M-convex
intersection problem. The Fenchel-type duality theorem for matroid valuations is
generalized for M-convex functions and the submodular flow problem is general-
ized to the M-convex submodular flow problem (Murota [142]), which involves an
M-convex function as a nonlinear cost. The nice optimality criterion using dual
variables survives in this generalization. Thus, M-convex functions yield fruitful
generalizations of many important optimization problems on matroids.
The two independent lines of development, namely, the convexity argument
for submodular functions in the early 1980s and that for valuated matroids and
M-convex functions in the early 1990s, were merged into a unified framework of
discrete convex analysis advocated by Murota [140] in 1998. The concept of L-
convex functions was introduced as a generalization of submodular set functions.
L-convex functions form a conjugate class of M-convex functions with respect to
the Legendre-Fenchel transformation. This completes the picture of conjugacy
advanced by Whitney [218] in 1935 as the equivalence between the submodularity
of the rank function of a matroid and the exchange property of independent sets of
a matroid. The duality theorems carry over to L-convex and M-convex functions.
In particular, the separation theorem for L-convex functions is a generalization of
Frank's separation theorem for submodular functions.
Ramifications of the concepts of L- and M-convexity followed. M*1 -convex func-
tions,4 introduced by Murota-Shioura [151], are essentially equivalent to M-convex
functions, but are sometimes more convenient. For example, a convex function in
one variable, when considered only for integer values of the variable, is an M^-convex
function that is not M-convex. Lfi-convex functions, due to Fujishige-Murota [68],
are an equivalent variant of L-convex functions. It turns out that L''-convex func-
tions are exactly the same as the submodular integrally convex functions that had
been introduced by Favati-Tardella [49] in their study of local vs. global optimality.
The success of polyhedral methods in combinatorial optimization naturally sug-
gests the possibility of polyhedral versions of L- and M-convex functions. This idea
was worked out by Murota-Shioura [152] with the introduction of the concepts of
4
"M^-convex" should be read "M-natural-convex" and similarly for "L^-convex."
1.2. Useful Properties of Convex Functions 9
L- and M-convexity for polyhedral functions (piecewise linear functions in real vari-
ables). These convexity concepts were denned also for quadratic functions (Murota-
Shioura [155]) and for closed convex functions (Murota-Shioura [156], [157]).
We conclude this section with a remark on a subtle point in the relationship
between submodularity and convexity. Prom the discussion in the early 1980s we
have agreed that submodularity should be compared to convexity. This statement
is certainly true for set functions. When it comes to functions on integer points,
however, we need to be careful. As a matter of fact, an M^-concave function is
submodular and concave extensible (Theorems 6.19 and 6.42), whereas an L^-convex
function is submodular and convex extensible (Theorem 7.20). This shows that
submodularity and convexity are mutually independent properties for functions on
integer points. It is undoubtedly true, however, that submodularity is essentially
related to discrete convexity.
for all x,y e R" and for all A with 0 < A < 1, where it is understood that the
inequality is satisfied if f ( x ) or f(y) is equal to +00. The inequality (1.2) implies
that the set
called the effective domain of /, is a convex set. Hence, the present definition of
a convex function coincides with the one in (1.1) that makes explicit reference to
the effective domain S. A special case of inequality (1.2) for A = 1/2 yields the
midpoint convexity
holds for every y and a local optimum if this inequality holds for every y in some
neighborhood of x. Obviously, global optimality implies local optimality. The
converse is not true in general, but it is true for convex functions.
Theorem 1.1. For a convex function, global optimality (or minimality) is guaran-
teed by local optimality.
The above theorem is significant and useful in that it reduces the global prop-
erty to a local one. Still it refers to an infinite number of points or directions around
x for the local optimality. In considering discrete structures on top of convexity we
may hope that a fixed and finite set of directions suffices to guarantee the local
optimality. For example, in the simplest case of a separable convex function
where
by a simple calculation. See Fig. 1.3 for the geometric meaning in the case of n = 1.
The Legendre-Fenchel transformation gives a one-to-one correspondence in
the class of well-behaved convex functions, called closed proper convex functions,
where the precise meaning of this technical terminology (not important here) will
be explained later in section 3.1. The notation /" means (/*)*, the conjugate of
the conjugate function of /.
conditions). If6
Such a theorem, if any, claims a min-max duality relation for integer-valued non-
linear functions, which is not likely to be true for an arbitrary class of discrete
functions. It is emphasized that the definition of convexity itself is left open in the
above generic statement, although h should be called concave when —h is convex.
As for the separation theorem, a possible discrete version would read as fol-
lows, imposing integrality (a* € Z, p* e Zn) on the separating affine function (see
Fig. 1.5).
As is easily verified, the discrete separation theorem, as well as the discrete Fenchel
duality, holds with this definition in the case of n — 1.
When it comes to higher dimensions, the situation is not that simple. The
following examples demonstrate that the discrete separation fails with this naive
definition of convexity.
1.3. Submodular Functions and Base Polyhedra 15
where x = (x(l),x(2)) e Z2. They are integer valued on the integer lattice Z2,
with /(O) = h(0) = 0, and can be extended, respectively, to a convex function
/ : R2 -> R and a concave function h : R2 —> R given by
where x = (x(l), x(2)) e R2. Since f ( x ) > h(x) (Vx e R 2 ), the separation theorem
in convex analysis (Theorem 1.3) applies to the pair (/, h) to yield a (unique)
separating affine function (p*,x), with p* = (1/2,1/2). We have /(x) > (p*,x} >
h(x) for all x € R2 and, a fortiori, /(x) > (p*,x) > /i(x) for all x 6 Z2. However,
there exists no integral vector p* e Z2 such that /(x) > (p*,x) > /i(x) for all
x € Z2. This demonstrates the failure of the desired discreteness in the separating
affine function. •
Example 1.6 (Failure of real-valued separation). This example shows that even
the existence of a separating affine function can be denied. For the discrete functions
for xj= (x(l),x(2)) € R2, since /(1/2,1/2) < h(l/2,1/2). This example also shows
that / > h on R" does not follow from / > h on Z™. •
Similarly, the discrete Fenchel duality fails under the naive definition of con-
vexity. The above two examples serve to demonstrate this.
Thus, the naive approach to discrete convexity does not work, and some deep
combinatorial or discrete-mathematical considerations are needed. We are now
motivated to look at some results in the area of matroids and submodular functions,
which we hope will provide a clue for fruitful definitions of discrete convexity.
Moreover, if p and p, are integer valued, the vector x* can be chosen to be integer
valued.
This inequality for p = xx yields the inequality (1.17) above, where /?* = 0 follows
from 0(0) = p(0) = 0 and fc(0) = /x(0) = 0. Thus, the first half of the discrete sepa-
ration theorem, the existence of a real vector x*, can be proved on the basis of the
separation theorem in convex analysis and the relationship between submodularity
and convexity. The combinatorial essence of the above theorem, therefore, consists
of the second half, claiming the existence of an integer vector for integer-valued
functions. Hence, we have the accepted understanding
Duality for submodular functions = Convexity + Discreteness,
mentioned in section 1.1.1.
We denote by S = S[Z] the class of integer-valued submodular set functions
and by Q£ = QiC[Z —> Z] that of discrete functions obtained as the restriction to Zv
of the Lovasz extensions of some member of S. That is, Q£ consists of functions
g : Zv —> Z U {+00} such that g(p) = p(p] (Vp e Zv) for some p e S. In view of
the above theorems, Q£ is a promising class of discrete convex functions. This is
indeed true, as we will see in section 1.4.1.
18 Chapter 1. Introduction to the Central Concepts
We are particularly interested in the case of integer-valued p, for which the base
polyhedron is integral in the sense of
where the overline designates the convex hull11 in Rv. This integrality means,
in particular, that all the vertices of the polyhedron B(p) are integer points. In this
integral case, we refer to B(/o) as the integral base polyhedron associated with p.
Assuming the integrality of p, we consider a discrete set
the set of integer points contained in integral base polyhedron B(p). If integer-
valued submodular functions can be viewed as well-behaved discrete convex func-
tions, there is a fair chance of such discrete sets B being well-behaved discrete
convex sets. This is indeed the case in many senses, as we will see in Chapter 4.
• Here we focus on an axiomatic characterization of such a B that makes no
explicit reference to the defining submodular function p. Denoting the positive
support and the negative support of a vector x = (x(v) v € V) 6 Zv by
11
The convex hull of a set means the smallest convex set containing the set.
12
The proofs of Theorems 1.9, 1.10, 1.11, and 1.12 are given later when we come to their rigorous
or more general treatments in Theorems 4.15, 8.12, 6.26, and 4.18.
1.3. Submodular Functions and Base Polyhedra 19
Recall also the notation o£[Z —> Z] for the class of the restrictions to Zv of the
Lovasz extensions of integer-valued submodular set functions. Then Theorem 1.9
can be rewritten as follows.
1. the optimal base problem to discuss the issue of local vs. global optimality
and
20 Chapter 1. Introduction to the Central Concepts
Moreover, if pi and p2 are integer valued, the polyhedron P(pi) fl P(/?2) is integral
in the sense of
and there exists an integer-valued vector x* that attains the maximum on the left-
hand side of (1.24).
Note that the submodularity inequality (1.13) for p is a special case of (1.27) with
p — YX and q — ~YV because of the identities
for r = p(V), where 1 = (1,1,..., 1) 6 Ry. This shows the linearity of g with
respect to the translation of p in the direction of 1. The properties (1.27) and
14
Proofs of the claims in this subsection are given in Chapter 7.
22 Chapter 1. Introduction to the Central Concepts
(1.30) of the Lovasz extension of a submodular set function are discretized to the
following definition of L-convex functions.
We say that a function g : 2*v —> RU {+00} with domzg 7^ 0 is L-convex if it
satisfies15-16
for the points p and q, which are discrete approximations to the midpoint (p + q)/2.
This inequality may be thought of as a discrete approximation to the midpoint
convexity (1.3). We return to midpoint convexity in (1.33) below.
It follows from (SBF[Z]) and (TRF[Zj) that the effective domain, say, D, of
an L-convex function satisfies17
15
SBF stands for submodularity for functions and TRF for translation for functions.
16
The notation 3 means "there exists" or "for some" in contrast to V meaning "for all" or "for
any."
17
SBS stands for submodularity for sets and TRS for translation for sets.
1.4. Discrete Convex Functions 23
where [^j2] and [^^J denote, respectively, the integer vectors obtained from E^3-
by componentwise round-up and round-down to the nearest integers. The discrete
midpoint convexity is a natural approximation to the midpoint convexity (1.3) of
ordinary convex functions.
Whereas L^-convex functions are conceptually equivalent to L-convex func-
tions, the class of L^-convex functions is strictly larger than that of L-convex func-
tions. In fact, it is easy to derive the translation submodularity (SBF^Zj) from
(SBF[Z]) and (TRF[Z]) or, more intuitively, a comparison of Figs. 1.9 and 1.8 in-
dicates this. The simplest example of an L^-convex function that is not L-convex is
the one-dimensional discrete convex function depicted in Fig. 1.6 (left).
18
"L^-convex" should be read "L-natural-convex."
24 Chapter 1. Introduction to the Central Concepts
L-convex functions enjoy the following nice properties that are expected of
discrete convex functions.
is an L-convex set. The converse is also true; i.e., any L-convex set has such a
polyhedral description for some 7 satisfying the triangle inequality
The concepts of L'/L^-convexity can also be defined for functions in real vari-
ables through an appropriate adaptation of the conditions (SBF[Zj) and (TRF[Zj).
Namely, we can define a function g : Rv —> R U {+00} with doniRg ^ 0 to be
L-convex if
19
An integer-valued distance function on V means a function 7 : V x V —> Z U {+00} such that
f ( v , v) = 0 for all v e V.
1.4. Discrete Convex Functions 25
More precisely, L-convexity can be defined for closed proper convex functions.20
Instead of dealing with this most general class of functions, this monograph focuses
on polyhedral convex functions21 and quadratic functions. The Lovasz extension
of a submodular set function is a polyhedral L-convex function that has the ad-
ditional property of being positively homogeneous. Quadratic L''-convex functions
are characterized in section 2.1.2 as quadratic forms denned by diagonally dominant
symmetric M-matrices22 and hence they are equivalent to the (finite-dimensional)
Dirichlet forms known in probability theory.
We conclude this section by identifying the four types of L-convex functions
that we are concerned with: real-valued L-convex functions on integers, integer-
valued L-convex functions on integers, real-valued polyhedral L-convex functions
on reals, and quadratic L-convex functions on reals. For the first three classes we
introduce the following notation:
where o£[Z —> Z] is the notation from section 1.3.1 for the class of the restrictions
to Zv of the Lovasz extensions of integer-valued submodular set functions.
for every a with 0 < a < 1. The validity of this inequality can be verified easily
from the definition of a convex function by adding the inequality (1.2) for X — a
and (1.2) for A = 1 - a.
20
The definition of closed proper convex function can be found in section 3.1.
21
A polyhedral convex function is a function that can be represented as the maximum of a finite
number of affine functions on a polyhedral effective domain.
22
Here is an unfortunate conflict of our notation with the standard terminology in matrix theory.
M-matrices do not correspond to M-convex functions but to L-convex functions.
26 Chapter 1. Introduction to the Central Concepts
The inequality (1.39) above shows that the sum of the function values evalu-
ated at two points, x and y, does not increase if the two points approach each other
by the same distance on the line segment connecting them (see Fig. 1.10). For a
function defined on discrete points Z™, we simulate this property by moving two
points along the coordinate axes rather than on the connecting line segment.
We say that a function / : Z ^ - > R U { + o o } with domz/ ^ 0 is M-convex if
it satisfies the following exchange axiom:
(M-EXC[Zj) For x,y e domz/ and u e supp+(x - y), there exists
v e supp~ (x — y) such that
See Fig. 1.11 for an illustration of this exchange property. The inequality (1.40)
implicitly imposes the condition that x~Xu+Xv G domz/ and y+Xu~Xv £ domz/
for the finiteness of the right-hand side. With the use of the notation
A/(z; v, u) = f(z + Xv~ Xu) - f(z) (1.41)
for z e domz/ and u,v e V, the exchange axiom (M-EXC[Zj) can be expressed
alternatively as follows:
where the maximum and the minimum over an empty set are -co and +00, respec-
tively. Naturally, a function h is said to be M-concave if —h is M-convex.
It follows from (M-EXC[Zj) that the effective domain of an M-convex function
satisfies the exchange axiom (B-EXC[Z]) that characterizes the set of integer points
in an integral base polyhedron, since x — Xu + Xv G domz/ and y + Xu ~Xv 6 domz/
1.4. Discrete Convex Functions 27
for x, y £ domz/ in (1.40). In particular, the indicator function SB '• Zv —> {0, +00}
of a set B C Zv is M-convex if and only if B is the set of integer points in an
integral base polyhedron. Accordingly, we refer to a set of integer points satisfying
(B-EXC[Z]) as an M-convex set.
The effective domain of an M-convex function /, being an M-convex set, lies
on a hyperplane {x e Ry x(V) = r} for some integer r and, accordingly, we may
consider the projection of / along a coordinate axis. This means that, instead of
the function / in n = \V\ variables, we consider a function /' in n — 1 variables
defined by
23
"M''-convex" should be read "M-natural-convex."
24
Proofs of the claims in this subsection are given in Chapter 6.
28 Chapter 1. Introduction to the Central Concepts
M-convex functions are closely related to network flow problems, such as the
minimum cost flow problem and the shortest path problem. As an indication of this
connection we mention that the distance 7 on V denned by 7(14, v) = Af(x;v,u)
for u, v G V with a fixed x € domz/ satisfies the triangle inequality (1.35). This
is because the exchange property (M-EXC[Z]) applied to x = x — Xv + Xw and
y = x - Xu + Xv, for which supp + (x - y) = {u,w} and supp~(£ — y) = {v},
yields f(x) + f(y) > f(x - \u + Xv) + f(y + Xu - Xv), which is equivalent to
A/(x; w, v) + A/(z; v, u) > A/(x; w, u).
The concepts of M'/M^-convexity can also be defined for functions in real
variables through an appropriate adaptation of the exchange axiom. Namely, we
can define a function / : Hv —> R U {+00} with doniR/ 7^ 0 to be M-convex if it
satisfies the following exchange property:
1.4. Discrete Convex Functions 29
where A1o[Z] is the notation from section 1.3.2 for the class of indicator functions
of sets of integer points contained in integral base polyhedra.
1.4.3 Conjugacy
The conjugacy relationship between L-convexity and M-convexity is a distinguish-
ing feature of the present theory. Whereas conjugacy in ordinary convex analysis
gives a symmetric one-to-one correspondence within a single class of closed proper
convex functions (Theorem 1.2), conjugacy described in this section establishes a
one-to-one correspondence between two different classes of discrete functions hav-
ing different combinatorial properties denoted by "L" and "M." We describe the
30 Chapter 1. Introduction to the Central Concepts
where <—> above denotes the conjugacy with respect to the discrete Legendre-
Fenchel transformation (1.9). The following theorem25 shows that the conjugacy
extends to a relation between £ and M..
Theorem 1.14. Two classes of discrete functions, £Q = £o[Z] andoM = oM[Z —>
Z], are in one-to-one correspondence under the discrete Legendre-Fenchel transfor-
mation (1.9). That is, for g € £Q and f & ®M, we have g* G oM., /* € £o, g" = g,
and f" = f.
M-convex functions
positively homogeneous (base polyhedra)
M-convex functions M-convex sets
(Theorem 6.59) (Theorem 4.15)
distance functions submodular set functions
(Theorem 5.5) (Theorem 7.40)
L-convex sets positively homogeneous
L-convex functions
L-convex functions
( S ) , distance functions (T), and base polyhedra (Mo)- It is recalled again that the
conjugacy is defined by the (discrete) Legendre-Fenchel transformation (1.9).
The pair of L- and M-convexity prevails in discrete systems.
• In network flow problems, flow and tension are dual objects. Roughly speak-
ing, flow corresponds to M-convexity and tension to L-convexity. Namely,
tension M : flow.
In multiterminal electrical networks consisting of nonlinear resistors, the equi-
librium state can be characterized as a stationary point of a convex function
representing the energy (or power). The function is M-convex when expressed
in terms of the terminal current supplied by current sources. It is L-convex
when expressed in terms of the terminal voltage (or potential) specified by
voltage sources. Network flow problems are discussed in section 2.2 and Chap-
ter 9.
In a matroid, the rank function corresponds to L-convexity and the base family
to M-convexity:
rank function : L <—> M : base family.
In a valuated matroid, the valuation of bases is an M-concave function defined
on the unit cube {0,1} V • Matroids and valuated matroids are explained in
section 2.4.
• The concept of M-matrices corresponds to L-convexity. Specifically, a quadratic
function is L''-convex if and only if it is defined by a diagonally dominant sym-
metric M-matrix. The inverse of such a matrix corresponds to M-convexity.
A diagonally dominant symmetric M-matrix arises, for instance, from a dis-
cretization of the Poisson problem of partial differential equations, where the
matrix is an approximation to the differential operator (Laplacian) and its
inverse corresponds to the Green function. Hence
differential operator : L <—> M : Green function.
32 Chapter 1. Introduction to the Central Concepts
Dirichlet forms in probability theory are exactly the same as L^-convex quadratic
functions. These quadratic forms are discussed in section 2.1.
1.4.4 Duality
Duality theorems for L- and M-convex functions are stated here in the case of
integer-valued functions defined on integer points.26 We explain their significance in
relation to previous results, such as Frank's discrete separation theorem for submod-
ular/supermodular functions, Edmonds's intersection theorem, and Frank's weight-
splitting theorem for the weighted matroid intersection problem.
Recall from section 1.2 the generic form of a Fenchel-type min-max duality
theorem:
[Discrete Fenchel-type duality theorem] Let f : Zv —> Z U {+00} and
h : Zv —> Z U {—00} be convex and concave functions, respectively (in
an appropriate sense). Then
We can now specify the meaning of convexity left open in this generic statement by
L-convexity or M-convexity. Then the following theorems result.
Although the above theorems look different, they are actually the same theo-
rem if we assume the conjugacy between L^-convex functions and M^-convex func-
tions (a variant of Theorem 1.13). In fact, substitution of g = f* and k = h° in
(1.49) yields
26
The proofs of Theorems 1.15, 1.16, 1.17, 1.18, 1.23, and 1.24 are given in Theorems 8.21, 8.21,
8.16, 8.15, 5.9, and 4.21, respectively.
1.4. Discrete Convex Functions 33
Example 1.19. Frank's discrete separation theorem (Theorem 1.8) in the in-
tegral case can be derived from the L-separation theorem (Theorem 1.17). The
submodular and supermodular functions p and fi can be identified, respectively,
with an L''-convex function g : Zv -> Z U {+00} and an L''-concave function
k : Zv -> Z U {-00} by p(X) = g(xx) and n(X) = k(xx) for X C V, where
domz<7 C {0,1}v and dom z fc C {0, l}y. The L-separation theorem applies, since
the first assumption, domzg ndom z fc ^ 0, is met by g(0) = fc(0) = 0, which follows
from p(0) = yu(0) = 0. We see /?* = 0 from the inequality (1-51) for p = 0, and then
the desired inequality (1.17) is obtained from (1.51) with p = xx for X C V. •
34 Chapter 1. Introduction to the Central Concepts
Substituting these expressions into the Fenchel-type min-max relation (1.50) yields
Example 1.21. Frank's weight-splitting theorem for the matroid intersection prob-
lem with integer weights is a special case of the M-separation theorem (Theorem
1.18). Given two matroids (V, BI) and (V, $2) on a common ground set V with
base families BI and B2, as well as an integer-valued weight vector w : V —> Z, the
optimal common base problem is to find B € BI fl B2 that minimizes the weight
w(B) = ^2veBw(v}- Frank's weight-splitting theorem says that a common base
B* (E BI n BI is optimal if and only if there exist integer vectors w\ and w*2 such
that
(i) w = w* + w2,
(ii) B* is a minimum-weight base of (V, BI) with respect to w^, and
(iii) B* is a minimum-weight base of (V,62) with respect to w^-
The "if" part is easy and the content of this theorem lies in the assertion about the
existence of such a weight splitting.
For an optimal common base B*, define
which are M-convex and M-concave, respectively (h is constant on 82)- Noting that
f ( x ) > h(x) (x & Z y ), as well as domz/ n domz/i ^ 0, we apply the M-separation
theorem to obtain a* 6 Z and p* € Zv for which the inequality (1.52) is true. A
weight splitting constructed by
has the desired properties (i) to (iii). In fact, (1.52) with x = XB* reads w(B*) >
a* +p*(B*) > w(B*), which shows a* = w(B*) -p*(B*) = wl(B*). It follows
1.4. Discrete Convex Functions 35
also from (1.52) that w(B) > a* + p*(B) for every B (E BI (namely, (ii)) and
a* +p*(B) > w(B*) for every B e B2 (namely, (iii)).
Moreover, the valuated matroid intersection theorem, a generalization of the
weight-splitting theorem, can be regarded as a special case of the M-separation
theorem (to be explained in Example 8.28). •
Theorem 1.22 (Separation for convex sets). If Si and 82 are disjoint convex sets
36 Chapter 1. Introduction to the Central Concepts
Theorem 1.23 (Discrete separation for L-convex sets). If D\ and D% are disjoint
L-convex sets, there exists x* £ {—1,0,1} V such that
Theorem 1.24 (Discrete separation for M-convex sets). //B\ and B-2 are disjoint
M-convex sets, there exists p* & {0,1}V U {0, — 1}V such that
Let us dwell on the content of these theorems, referring to the latter. The first
implication, explicit in the statement of Theorem 1.24, is that the separating vector
p* is so special that p* or — p* is a {0,1}-vector. The second, less conspicuous and
more subtle, is that B\ D _B2 = 0 follows from B\ n B2 = 0, since (1.55) implies
BI H B2 = 0. The implication
Bibliographical Notes
References for optimization abound in the literature. See, e.g., Nemhauser-Rinnooy
Kan-Todd [166] as a general handbook; Bazaraa-Sherali-Shetty [8], Bertsekas [10],
Fletcher [52], Mangasarian [126], and Nocedal-Wright [169] for nonlinear optimiza-
tion; and Cook-Cunningham-Pulleyblank-Schrijver [26], Du-Pardalos [43], Korte-
Vygen [115], Lawler [119], and Nemhauser-Wolsey [167] for combinatorial optimiza-
tion. References for convex analysis are included in the bibliographical notes at the
1.4. Discrete Convex Functions 37
end of Chapter 3, and those for network flow theory and matroid theory are in
Chapter 2.
Pujishige [65] is a standard reference for submodular functions, and Narayanan
[165] and Topkis [203] cover some other topics related to electrical networks and
economics, respectively. Theorem 1.7, connecting submodularity and convexity, is
due to Lovasz [123], and the name "Lovasz extension" was coined by Fujishige [63],
[65]. The discrete separation for submodular functions, Theorem 1.8, is due to Frank
[55]. Theorem 1.9, the equivalence between submodularity and exchangeability, is
folklore from the 1980s. Seeing that no explicit and rigorous proof can be found in
the literature, we will provide a proof in Theorem 4.15 in this book. The recasting
into Theorem 1.10 is by Murota [140]. The local optimality criterion for the linearly
weighted base problem in a matroid (Theorem 1.11) is a standard result (see, e.g.,
Corollary 8.7 of [26]). The intersection theorem, Theorem 1.12, is due to Edmonds
[44]. The weight-splitting theorem described in Example 1.21 is due to Frank [54].
M-convex functions are introduced in Murota [137], followed by L-convex func-
tions in Murota [140]. Their fundamental properties are established in Murota
[137], [140], [141], [142]; the discrete conjugacy theorem (Theorem 1.13) and the
L-separation theorem (Theorem 1.17) in [140]; the M-separation theorem (Theo-
rem 1.18) in [137], [140], [142]; and the Fenchel-type duality theorem for M-convex
functions (Theorem 1.16) in [137], [140]. The separation theorems for L-convex
sets and M-convex sets (Theorems 1.23 and 1.24) are due to [140]. M-convexity
38 Chapter 1. Introduction to the Central Concepts
Table 1.2. Operations for discrete convex sets and functions (/: function,
S: set; Q: Yes [cf. Theorem, Prop.], x: No).
Miller's convex integrally separable
discrete convex extensible convex convex
/1+/2 X 0 X 0
SinS 2 X o X o
/+ sep-conv X o O [3-24] o
S H [a, b] o o O o
f+ affine X o O [3-25] o
/iQz/2 X X X o
Si+S 2 X X X o
/' II * o X o
dom/ 0 o O [3.28] o
arg min / 0 o O [3.28] o
1 1 Mg - convex L2 -convex M^-convex L^-convex
/1 + /2 X x x (Mj-conv) O [7.11]
Sin 6-2 X x x (M2-conv) O [5.7]
/+ sep-conv O x O [6.15] O [7.H]
5 n [a, 6] o x O O
/+ affine o O O [6.15] O [7.H]
/lQz/2 X x O [6-15] x (L2-conv)
Si + S2 X X O [4-23] x (L2-conv)
/* x (L2-conv) x (M2-conv) x (L^-conv) x (M^-conv)
dom/ O [8.29] O [8.39] O [6.7] 0 [7.8]
arg min / O [8-30] O [8.40] O [6-29] O [7.16]
and L-convexity are investigated also for functions in real variables for polyhedral,
quadratic, and closed convex functions in Murota-Shioura [152], [155], [156], [157].
M^-convex functions are introduced by Murota-Shioura [151] and L''-convex
functions by Pujishige—Murota [68]. The concept of submodular integrally convex
functions, together with a characterization by discrete midpoint convexity, is due
to Favati-Tardella [49]. The equivalence of this concept to L''-convexity is shown in
[68]. Table 1.2 is taken from Murota-Shioura [153].
Chapter 2
The objective of this chapter is to demonstrate how convex functions with combi-
natorial structures arise naturally from a variety of discrete systems, such as (i)
discretizations of the Poisson equation, (ii) electrical networks consisting of linear
(ohmic) and nonlinear resistors, and (iii) matrices (matroids) and polynomial matri-
ces (valuated matroids). It is emphasized that such functions are always equipped
with a pair of combinatorial properties, namely, submodularity (L-convexity) and
exchangeability (M-convexity).
39
40 Chapter 2. Convex Functions with Combinatorial Structures
well known, the convexity (resp., strict convexity) of / is equivalent to the positive
semidefiniteness (resp., positive definiteness) of A:
and
The criterion (2.8) compares favorably with (2.7) in that there are only n leading
principal minors as opposed to 2™ principal minors. Positive (semi)definiteness can
be checked with O(n3) arithmetic operations by an algorithm similar to Gaussian
elimination.
A change of the variable in (2.1), x = Sy with a nonsingular matrix S, re-
sults in another quadratic form fs(y] = f ( S y ) , which is associated with another
symmetric matrix S^AS. The convexity of a quadratic form is preserved under
such linear transformations of the variable, and the positive semidefiniteness of a
symmetric matrix also remains invariant.
The change of the variable x = Sy with a general nonsingular S, rotating the
coordinate axes, does not respect any special coordinate directions. It would be
reasonable to expect that such a general transformation should not be compatible
with any combinatorial properties relevant to discreteness in direction. Conversely,
we may regard properties of a quadratic form or of a symmetric matrix as being
combinatorial with discreteness in direction if they are not invariant with respect to
the entire class of transformations but are invariant with respect to some restricted
subclass thereof (the class of diagonal scalings, for example). A typical combinato-
rial property of this kind is the sign pattern of the entries of a matrix. This is what
we will study in the following subsection.
2.1. Quadratic Functions 41
Note that the second condition (2.10) can also be expressed, under (2.9), as
Example 2.1. Consider the Poisson equation —Au = cr, where A is the Laplacian
Si=i d2/do;?, a denotes the source term, and d is the dimension of the space.28 A
standard discretization scheme for this differential equation, where we assume d = 1
for illustration purposes, gives rise to a system of linear equations described by a
matrix like
This matrix satisfies the two conditions (2.9) and (2.10) above.
Example 2.2. Consider the simple electrical network depicted in Fig. 2.1. It
consists of five branches (linear resistors) connected at four nodes. We denote the
conductance (the reciprocal of resistance) of branch j by g$ > 0 (j = 1,...,5),
the potential at node i by Pi (i = 1 , . . . , 4 ) , the voltage across branch j by j]j
(j = 1 , . . . , 5), and the current in branch j by ^ (j = 1 , . . . , 5). The underlying
graph can be represented by the incidence matrix
whose rows and columns correspond, respectively, to the nodes and branches; the
jth column has entry 1 at its initial node and —1 at its terminal node. The voltage
28
We assume the Dirichlet boundary condition.
42 Chapter 2. Convex Functions with Combinatorial Structures
which satisfies the two conditions (2.9) and (2.10) above. The matrix L is called
the node admittance matrix. •
Proposition 2.4. A symmetric matrix L with properties (2.9) and (2.10) is positive
semidefinite.
where L\\ is of order n — 1 and £22 = 4m- If fnn = 0; then lni = i^n — 0
(i = 1, . . . , n — 1) and therefore detL = 0. Suppose £nn > 0 and put LH =
£11 - Li2L22~lL2i. An off-diagonal entry i^ - iintnn^(-nj of Ln, where 1 < i ^
j < n — 1, is nonpositive by lij,lin,lnj < 0 and inn > 0. For the ith row sum of
LH we have
Thus, LH satisfies (2.9) and (2.10), and, by the induction hypothesis, its determi-
nant is nonnegative. Hence, detL = lnn • det-Ln > 0. D
which is convex by Proposition 2.4 and (2.2). Our goal here is to reveal a key
combinatorial property of g(p) that reflects the combinatorial properties (2.9) and
(2.10) of the matrix L.
represents the power (energy) consumed in the network. The Poisson equation
—Aw = a (with d = 1) in Example 2.1 can be translated into a variational problem
of minimizing a functional
44 Chapter 2. Convex Functions with Combinatorial Structures
Proof. The inequality (2.17) for p — Xi (*th unit vector) and q = Xj (jth unit
vector) yields (2.9). For the converse, put a — p/\q, p = a + p, and q = a + q. Then
pV q = a + p + q. Substitution of these into (2.14) shows that the right-hand side
of (2.17) minus the left-hand side of (2.17) is represented as
Theorem 2.7. For a symmetric matrix L, conditions (a) and (b) are equivalent.
(a) L has off-diagonal nonpositivity (2.9) and diagonal dominance (2.10).
(b) g(p) has translation submodularity (SBF^Rj).
Proof, (b) => (a): Proposition 2.6 shows (2.9). (SBF^R]) with p = Xi, 1 = -1,
and a = 1 yields (2.10), since
2.1. Quadratic Functions 45
(a) =>• (b): Put I = {i \ a < Pi - qi}, and let J be the complement of I.
We have
wmcn
With the use of X^fce/ ^ — ~ SjeJ ^'' is a consequence of (2.10), as well as
(2.9), we obtain
Note 2.8. The quadratic form considered in this section coincides with what
is known as the Dirichlet form (in finite dimension) in the theory of the Markov
process and potential theory; see Fukushima-Oshima-Takeda [71]. We mention here
the equivalence among the five conditions (a), (b), (c), (d), and (e) below.29 The
five conditions are equivalent, by Theorem 2.7, to the translation submodularity
(SBF"[B.])offl(p).
(a) L has off-diagonal nonpositivity (2.9) and diagonal dominance (2.10).
(b) The normal contraction operates on g(p) = |pTLp; i.e., for p, q € R™,
29 l
In the terminology of the theory of Dirichlet forms, — L corresponds to the generator, a Sa
to the resolvent, and Tt to the semigroup.
46 Chapter 2. Convex Functions with Combinatorial Structures
takes the unique minimum at p = po, where po = Sax. For go = (0 Vpo) A1 we have
g(Po) > g(qo) by (c) and (p0 -x)~r(p0 -x) > (q0 - x)1 (qQ - x) by 0 < x < 1. Hence
V'(Po) > ^(9o); which implies po = qo = (0 Vpo) A 1. This shows 0 < po = Sax < 1.
[(d) => (c)]: Since Sa = (sij) is Markovian, we have s^ > 0 (1 < i, j < n) and
T™_, sa <l(l<i<n). Define
for a > 0. Then g^(p) tends to g(p) as a —> +00. On the other hand, the
expression
shows that every normal contraction operates on g^a\p). The limit of a -^ +00
establishes (c).
[(d) =4> (e)]: This is due to the formula
2.1. Quadratic Functions 47
Example 2.10. Recall the Poisson equation —Au = a in Example 2.1. The inverse
of the matrix L in (2.11) is given by
It should be clear that Xi designates the iih unit vector for 1 < i < n while xo is
the zero vector and
for x — (xi i = 1,... ,n) & Rn. Recall that such an exchange property can be
viewed as a combinatorial analogue of the basic inequality (1.39) valid for a general
convex function.
The following is the main theorem of this section, stating the conjugacy rela-
tionship between L''-convexity and M^-convexity for strictly convex quadratic func-
tions.
Theorem 2.11. Suppose that strictly convex quadratic forms g and f are conjugate
to each other. Then g satisfies translation submodularity (SBF^R]) if and only if
f has exchange property (M^-EXCfR]).
The conjugacy relationship between (SBF11 [R]) and (M^-EXC[R]) stated above
for strictly convex quadratic forms is in fact valid for a more general class of func-
tions, as is fully developed in Chapter 8. This particular case, however, deserves
separate consideration, in that it admits a matrix-algebraic proof using the Farkas
lemma and thereby provides a new insight into (SBF^Rj) vs. (M^-EXC[R]) conju-
gacy.
In what follows we prove Theorem 2.11 by establishing Theorem 2.12 below.30
As variants of (M^-EXC^R]) we consider the following:
(M^-EXC + [Rj) For x,y G domR/ and i e supp+(o; - y), there exist
j e supp~(x - y) U {0} and a positive number a0 G R-++ such that
valid for the directional derivative f ' ( x ; d ) and sufficiently small a > 0. We also
denote by (M^-EXC^R]) the property (M^-EXC^R]) with < replaced by strict
inequality <.
30
The reader may go on to section 2.2, skipping the technical matters given in the rest of
section 2.1.
2.1. Quadratic Functions 49
The implications indicated by <— or j are easy to see and those by <&• or f|- are
proved below.
(a) <=» (b+): Denoting M~l by L = (iij), we have ML = /; i.e., X)?=i ^jimj =
Xi, which can be rewritten as
50 Chapter 2. Convex Functions with Combinatorial Structures
The condition (a), L G £, is equivalent to all the coefficients in this expression being
nonnegative, whereas the latter is equivalent, by the Farkas lemma (Proposition 2.13
below), to
This is nothing but (b + ), since xTXi > 0 is the same as i & supp + (x).
(b) =>• (b+): This follows from the above argument and the latter half of
Proposition 2.13 below.
(a) => (c+): Fix x £ Rn. It suffices to show
Let i e supp+(x) attain the minimum on the left-hand side. Put S = supp+(:c) U
supp~(x) and let x € Rs denote the restriction of x to S. The submatrix of M
with row and column indices in S is denoted by M = (Wij \ j & S), where WLj e Rs.
Then we have supp + (x) = supp + (x), supp~(x) — supp~(o;), i & supp+(;r), Xj =£ Q
(Vj 6 5), and xTmj = xrmj (Vj 6 S). Since M e £~1 by Proposition 2.14 below,
M satisfies (b+). Hence
in which
as well as a similar expression for f ' ( x ; —\i + Xj) + f ' ( y \ Xi — Xj}- We then replace
x — y with x.
(d+) => (e+): This follows easily from (2.22). D
Proposition 2.13. For a matrix A and a vector b, the conditions (a) and (b) below
are equivalent (Farkas lemma}:31
(a) Ax = b for some nonnegative x > 0.
31
Inequality between vectors means componentwise inequality; e.g., x > 0 for x = (xi)™=l means
Xi > 0 for i = 1,... ,n.
2.1. Quadratic Functions 51
Proof, (a) => (b) => (c) is obvious, (b) => (a) is proved later in Theorem 3.9. For
(c) => (b), there exists z such that ZTA — 1T by the assumed nonsingularity of A.
Then (y + e;z) T A > 0T for any e > 0, and (c) yields (y + ez)Tb > 0, which implies
yT& > 0, since £ > 0 is arbitrary. D
l 1
Proposition 2.14. Any principal submatrix of M & £ belongs to £ .
To prove MH e C { by induction on the size of MH, we may assume M22 and £22
are 1x1. Since L22 = inn > 0, we have Mn"1 = Lu - L^I^"1^21 (= LU),
which shows the nonsingularity of M\I. Then the proof of Proposition 2.4 shows
MII e £~l. D
Note 2.15. It is worth noting that conditions (c) and (c+) in Theorem 2.12
immediately imply positive semidefiniteness and positive definiteness, respectively.
The proof for the former reads as follows, while a similar proof works for the latter.
Let /i be an eigenvalue of the matrix M and x be the corresponding eigenvector
with supp+(x) ^ 0. Then (c) shows
for i e supp+ (x). This implies JJL > 0, since, otherwise, the left-hand side is negative
and the right-hand side is zero. Hence M is positive semidefinite by (2.4). It is
also noted that ma > 0 follows from (c) with x = Xi and fn%j > 0 from (c) with
x = Xi + o-Xj with a > 0 large. •
where
The conjugacy stated in Theorem 2.11 for strictly convex quadratic forms
is generalized as follows. See Murota-Shioura [155] for the proof as well as the
structure of the coefficient matrices L and M.
entering v, respectively. For each arc a G A, d+a designates the initial vertex of a
and d~a the terminal vertex of a. We consider here a minimum cost flow problem,
in which each arc is associated with a nonlinear convex cost function and the supply
(or demand) of flow is specified at terminal vertices.
The physical model we have in mind is a multiterminal electrical network
that consists of nonlinear resistors and is driven by a (current or voltage) source
applied to the terminal vertices. To reinforce physical intuition, we sometimes
use terminology such as current and voltage instead of flow and tension, but no
physics is really involved in our arguments. A reader who is more comfortable with
combinatorial optimization terminology may replace the terminology as follows:
electrical network network
current flow
voltage tension
current source supply of flow
potential potential (dual variable)
current potential cost function in flow
voltage potential cost function in tension
characteristic curve kilter diagram
Each arc a € A is associated with a flow (or current) £(a) and a tension (or
voltage) r/(a) and each vertex v e V with a potential p ( v ) . The boundary of flow £
is defined to be32
which represents the net flow leaving vertex v. The coboundary of potential p is
which expresses the difference in the potentials at the end vertices of an arc a.
32
Rockafellar's notations div and A in [178] are related to ours by div = d and A = —S.
54 Chapter 2. Convex Functions with Combinatorial Structures
For each terminal vertex v G T, let x(v) denote the amount of flow going out
of the network at v and p(v) be the potential at v. We have structural equations
In physical terms a characteristic curve shows the constitutive equation for a non-
linear resistor, describing the possible pairs of current and voltage. In the linear
case (as in Example 2.2) we have
for some Ra > 0, which represents the resistance of an ohmic resistor. We consider
here a nonlinear case, where monotonicity
which means that /a(£) is the area below F0 in Fig. 2.3 and ga(f]) is the area above
Fa. The functions /a(£) and ffa(^) are both convex as a consequence of the assumed
2.2. Nonlinear Networks 55
Hence follows
That is, fa and ga are conjugate to each other with respect to the Legendre-Fenchel
transformation (1.6).
In the theory of electrical networks, the function /<,(£) is sometimes called the
current potential (or content) and ga(r]) the voltage potential (or cocontent). In the
case of a linear resistor, the functions fa and ga are quadratic, i.e.,
and they are both equal to half the power consumed in the resistor.
When a current source described by a; G RT with X^eT x(v^ = 0 is applied
to the terminal vertices of the network, the equilibrium state of (£(a) a G A)
and (77(0) | a & A) is determined as a solution to the structural equations (2.29)
and the constitutive equations (2.31). A variational formulation of this problem is
to minimize the total current potential ]CoeA/a(£( a )) among au possible current
distributions £ subject to the conservation law (2.29), with the current vector at
the equilibrium state being characterized as a minimizer of this problem (see Note
2.18 in section 2.2.3). We define f ( x ) to be the minimum value of the total current
potential in this variational problem; i.e.,
The functions / and g introduced above are both convex (see Note 2.17 in
section 2.2.3) and they are conjugate to each other (see Note 2.18). In this sense
56 Chapter 2. Convex Functions with Combinatorial Structures
they stand on equal footing and there seems to be no concept in convex analysis
that distinguishes between / and g.
When it comes to combinatorial properties, however, these functions have
distinctive features. As is shown in Notes 2.19 and 2.20 in section 2.2.3, the function
/ : RT —> R U {+00} is endowed with an exchange property:
(M-EXC[R]) For x,y e doniR./ and u 6 supp+(z - y), there exist
v € supp~(x — y) and a positive number a$ e R++ such that
and
It should be clear that (2.41) is a discrete analogue of the conjugacy (2.36), the dis-
crete Legendre-Fenchel transformation (1.9) for univariate functions. An example
of such a conjugate pair of cost functions is demonstrated in Fig. 2.4.
The characteristic curve Ta in this discrete setting is defined to be a subset
of Z2 induced from (fa,ga) by (2.35). It can be characterized as a subset of Z2
with the monotonicity property (2.32). Figure 2.5 shows the characteristic curve
Ta induced from (fa,9a) in Fig. 2.4.
In parallel with (2.37) and (2.38) we define functions / : ZT -»• ZU {+00} and
58 Chapter 2. Convex Functions with Combinatorial Structures
Note that these expressions are identical to (2.37) and (2.38) except that the vectors
are now integer valued and, in particular, the infima are taken over integer vectors.
Fortunately, such a discretization in the definitions of / and g does not destroy
the combinatorial properties discussed above. On the contrary, the discretization
turns out to be compatible with natural discretizations of the combinatorial prop-
erties. Namely, it can be shown (see Note 2.19) that the function / has a discrete
version of the exchange property:
(M-EXC[Zj) For x,y € domz/ and u e supp+(:r — y), there exists
v 6 supp~ (x — y) such that
This is essentially the same as (M-EXC[Rj) with cto = a = 1. On the other hand,
the function g satisfies
with r = 0 (see Note 2.20). Furthermore, these functions / and g are conjugate to
each other with respect to the discrete Legendre-Fenchel transformation (1.9), as
is proved later in section 9.6.
We have thus seen that the transition from R to Z is quite smooth in the
network flow problem. The combinatorial properties are discretized compatibly in
the discretization of the problem data. We emphasize that this is by no means
a general phenomenon but is an outstanding characteristic of the network flow
problem.
Note 2.17. We prove that / in (2.44) and g in (2.45) are convex functions under
the assumption that / > — oo and g > — oo.
To show the convexity of /, fix x, y G doniR,/. For any e > 0 there exist £x
and £v such that
where • \T denotes the restriction of a vector to T. For A € [0, I]R we have d(A£x +
(1 - A)£j,) = \d£x + (1 - X)d£y and, therefore,
This implies Ag(p) + (1 — A)g(g) > ^(Ap + (1 — A)g), since e > 0 is arbitrary. •
Note 2.18. The variational formulations of the network equilibrium are derived
here under the assumption of the existence of an equilibrium. Also shown is the
conjugacy between / in (2.44) and g in (2.45). It follows from (2.34) that
for any £, rj, p, x, and p satisfying the conservation laws expressed by (2.29) and
(2.30). By (2.35), the inequality above is an equality if and only if (£(a), r)(a)) 6 Fa
for each a e A. Suppose that an equilibrium state exists when a current source
x = x* is applied to the network and let £*, 77*, p*, p* be the vectors at the
equilibrium. The inequality (2.50) with 77 = 77*, p = p*, p — p*, and x = x* yields
60 Chapter 2. Convex Functions with Combinatorial Structures
which shows that the minimum of the current potential Y^aeA fa(£,(a)) is attained
by £ = £*. A similar result occurs for the variational formulation using the voltage
potential X^oeA 9a(rj(a)) when a voltage source described by p = p* is applied. For
the conjugacy of / and g, note that
follows from (2.50) and the definitions of / and g in (2.44) and (2.45). This in-
equality for x — x* turns into an equality for p = p*, which shows that f ( x * ) =
supp{(p,:c*} — g(p)} = g'(x*). A similar argument works for g = f*. The argument
here is admittedly lacking in mathematical rigor, for which the reader is referred to
Rockafellar [178]. •
Note 2.19. We prove that / in (2.44) satisfies (M-EXC[R]) under the assumption
that / > — oo. Fix x,y € doniR,/. For any e > 0 there exist £x and £y satisfying
(2.46) and (2.47). Consider the difference in the flows, £y — £x & R/4, for which we
have
Since u & supp+(x — y), there exists a path compatible with £y — £x that connects u
to some vertex in supp~(:r — y) (i.e., an augmenting path with respect to the pair of
flows £x and £ y ). More formally, there exist n : A —> {0, ±1} and v G supp~(x — y)
such that
where Xu, Xv G Rv are the characteristic vectors of u and v. For two flows £x + air
and £y — air with 0 < a < a0, where a0 = min |£y(a) - £s(a)| (> 0), we have
a:|ir(o)| = l
This implies (M-EXC[Rj) if a0 = a0(s) does not tend to zero as £ —> 0 and v = v(e)
remains the same as e —> 0. For the former property, we can take an augmenting
path TT such that UQ > (x(u) — y(u))/\A\, and for the latter we may take a subse-
quence of e that corresponds to a single v. In the discrete case of (2.39), UQ is a
positive integer and a = I is a valid choice. Hence follows (M-EXC[Zj). •
2.3. Substitutes and Complements in Network Flows 61
Note 2.20. We prove that g in (2.45) satisfies (SBF[R]) and (TRF[R]) under the
assumption that g > —oo. First, (TRF[R]) is obvious from 5(p + al) = dp. Fix
p, q € doniR^. For any e > 0 there exist r]p, rjq, p, and q satisfying (2.48) and (2.49).
For ?7V = — 8{p V q) and rj^ = ~S(p A q), we have
Therefore, we obtain
which implies (SBF[R]), since e > 0 is arbitrary. The discrete case (2.43) with ga
in (2.40) can be treated similarly. •
and supermodular if
Theorem 2.22. Let P be a parallel arc set and S a series arc set.
(1) F is submodular in WP and in cp.
(2) F is supermodular in u>s and in c$.
33
Formally, a simple cycle is an alternating sequence (VQ, ai,vi, 0,2,..., v^—i, ak,Vk) of vertices
Vi (i = 0 , 1 , . . . , fc) and arcs a, (i = 1,..., k) such that {9+aj, 9~a;} = {i>;~i, i>j} (i = 1 , . . . , k),
v
o — v/f, and Vi ^ Vj (1 < i < j < k).
2.3. Substitutes and Complements in Network Flows 63
Theorem 2.23. Let P be a parallel arc set and S a series arc set.
(1) F is l)- -convex in WP and f$-concave in cp.
(2) F is M^-convex in ws and L^-concave in eg.
Proposition 2.25. Let S be a series arc set and -K\ and 7T2 be circuits. If
supp+(7Ti) n supp+(7T2) (~l S ^ 0, there exists a circuit TT such that supp+(7r) C
supp+(7Ti) U supp+(?T2), supp~(?r) C supp~(7Ti) U supp"^), and supp + (7r) n S =
(sUpp+(7Ti) U SUpp+(7T2)) n S.
64 Chapter 2. Convex Functions with Combinatorial Structures
Proof. Suppose a e (supp + (7r 2 )\supp + (7ri))n5 < . By an elementary graph argument
we can find a circuit TT' such that supp + (7r') C supp+(?r1) Usupp + (7r 2 ), supp~(7r') C
supp~ (TTI ) U supp~ (7T2), and supp+ (TT' ) n S D (supp+ (TTI ) D S) U {a}. Repeating this
we can find TT. D
where fa > 0 and TTJ : A —> {0, ±1} is a circuit with supp+(7Ti) C supp + (£) and
supp~(7Tj) C supp~(£) for i — 1 , . . . , m.
where the left-hand side is bounded by F(w + \Xa,c) + F(w + nXb,c) and the
right-hand side is equal to_F(w,c) + F(w + \Xa + HXb,c). If £(a) < £(a), we
can take £a = £ and £& = ? to meet (2.60). If £(6) < £(6), we can take £a — £
and £;, = £ to meet (2.60). Otherwise, we make use of the conformal decomposition
£ —£ = Y^iL\ fii^i. Since a e supp + (J-£), we may assume 7Tj(a) = 1 for i = 1,..., I
and 7Tj(a) = 0 for i = 1+1,..., m. We have 7Tj(6) = 0 for i = 1 , . . . , (. by Proposition
2.24 (1), since P is parallel and {a, 6} C supp+(£ - £). Then £a = £ + £)Li AT»
and ^6 = £ + Y^iLe+i A7r» are feaslble circulations that satisfy (2.60).
To show (2.59) let £ and £ be optimal circulations for w and w + \xa — A*XP-
We can establish (2.59) by constructing feasible circulations £a and £p such that
34
More generally, the conformal decomposition is defined for a vector in a subspace in terms of
elementary vectors of the subspace; see Iri [94] and Rockafellar [178].
2.3. Substitutes and Complements in Network Flows 65
Proof of M^-Concavity in CP
We prove the M^-concavity of F in CP by establishing (M^-EXC[R]) for — F as a
function in cp. In our notation this reads as follows:
Let ci,c 2 e R+ be capacities with ci(a') = c2(a') for all a' € A\P.
For each a e supp+(ci — 02), there exist 6 e supp~(ci — c2) U {0} and a
positive number ao such that
If £i(a) < ci(a), we can take a0 = ci(a) — £i(a), & = 0, £{ = £1, and £2 = £2
to meet (2.62). Suppose £i(a) = ci(a). We have ^(a) = Ci(a) > c 2 (a) > £ 2 (a).
Let TT be a circuit such that a 6 supp+(?r) C supp+(£i — £2) and supp~"(?r) C
supp~(£i -£ 2 ). Since P is parallel and a e supp + (7r), we have supp+(?r) DP = {a}
and |supp~(7r) n P| < 1 by Proposition 2.24 (1). If |supp~(?r) n P = 1, define
b by {6} = supp~(?r) n P; otherwise put 6 = 0. We can take a0 > 0 such that
«o < ICi( a/ ) — £2(0')! for all a' e supp+(7r) U supp~(?r). Then £i = £1 - air and
£, = £2 + CKTT satisfy (2.62) if 0 < a < a0.
Proof of M"-Convexity in w§
We prove the M11-convexity of F in ws by establishing (M^-EXCfR]). In our notation
this reads as follows:
Let Wi,w 2 G R.A be weight vectors with Wi(a') = w2(a') for all a' €
A\S'. For each a e supp + (w;i-w 2 ), there exist 6 € supp~(wi— w 2 )U{0}
and a positive number a0 such that
Let £1 and £2 be optimal circulations for w\ and W2, respectively, with £i(a) mini-
mum and £2(0) maximum.
Proposition 2.26. There exists O.Q > 0 such that £1 is optimal for w\ — a\a and
£2 is optimal for w^ + aXa for all a e [0, O;O]R-
Proof. For any circuit TT such that IT (a) = —1 and 0 < £1 + f3ir < c for some
/3 > 0, we have (toi,£i + /?7r) < (wi,£i) by the choice of £1- Let a\ > 0 be the
minimum of — (wi, TT) over all such circuits TT. Then £1 is optimal for w\ — a\a for
all a e [0. O!I]R, since (u>i -ax a > Ci +/?7r) < (wi ~aXa, £1} for any /? > 0 and circuit
TT such that 0 < £1 + /?TT < c. Similarly, let 0:2 > 0 be the minimum of —(w2,n)
over all circuits TT such that 7r(a) = 1 and 0 < £2 + 0ir < c for some (3 > 0. Then £2
is optimal for 102 + QXa for all a G [0, O^JR. Put ao = min(ai, a^)- D
where the last equality is by Proposition 2.26. In what follows we assume £i(a) <
6 (a).
By Proposition 2.24 (2), we can impose further conditions on £1 and £2 that,
for each /> 6 S \ {a}, £1(6) is maximum among all optimal £1 for w\ with £i(a)
minimum, and &(b) is minimum among all optimal £2 for w-2 with £2(0) maximum.
Proposition 2.27. There exists a® > 0 SMC/I iftorf £1 zs optimal for w\ — a(xa — Xb)
and £2 is optimal for w^ + a(Xa — Xb) for all b £ S\ {a} and for all a & [0, O;O]R-
Proof. For any circuit TT such that TT(<Z) — 7r(6) = —1 for some b & S \ {a} and
0 < £1 + /^TT < c for some (3 > 0, we have (tyi, £1 + /571"} < (wi > £1} by the choice of £1.
Let cti > 0 be the minimum of —(WI,TT) over all such circuits TT. Then £1 is optimal
for Wi — a(xo — Xb) f°r an a G [0, «I]R- Similarly, let a2 > 0 be the minimum of
— (w2,Tr) over all circuits TT such that ?r(a) — TT(&) = 1 for some b e S* \ {a} and
0 < £2 + /371" < c f°r some /3 > 0. Then £2 is optimal for w^ + a(x<z ~ Xb) f°r a^
a e [0, CC2JR- Put a0 = min(ai, a 2 ). D
Proof. We have (U>I,TTI} < 0, since £1 is optimal for u>i and 0 < £1 + /3i7Ti < c.
Similarly, we have (w?, — TTI) < 0. Hence,
Since u>i(a) — 1112(0.) > 0 in this summation, we must have Wi(b) — W2(b) < 0 for
some b e supp+(7Ti) n S. D
Proof of L"-Concavity in GS
If £ a (a) < c (a), we can take £ = £a and £ = £5 to meet (2.66). If £&(&) < c(6), we
can take £ = £b and £ = £a to meet (2.66). Otherwise, we have £ 0 (a) > c(a) > £b(a)
and £ 0 (6) < c(6) < £b(6), and therefore a e supp+(£a — £b) and fc e supp~(£a — £;,).
We make use of the conformal decomposition £a — £f, = J]iii A71"*) where we assume
7Tj(a) = 1 for i = 1 , . . . , I and 7Tj(a) = 0 for i = £ + 1,..., m. We have 7Tj(&) = 0 for
i = 1,... ,1 by Proposition 2.24 (2), since S is series and a € supp+(£a — &,) and
b e supp-(£a - &). Then £ = £a - £-=i frm and f = 6 + Eti A^ satisfy ( 2 - 66 )-
To show (2.65), let £a and £s be optimal circulations for c+ \Xa and c — ^xs-
We can establish (2.65) by constructing circulations £ and £ such that
and 7Ti(a) = 0 for i = i + 1 , . . . , m. Then supp (TTJ) n S = 0 for i = I,... ,1. Noting
Z)i=i A = £a( a ) ~£s(a) > £a(a) — c(a), let k be the smallest integer with £i=1 A >
£a(a)_-c(a) and define /?' - K«(<0-<=(<»)]-E^ 1 A- Then £ = £a-Ei=i A^-A^fc
and £ = £s + Y^=l fc*i + P'^k satisfy (2.67), since £(o) = £ 0 (a) - £^1' & - P =
c(a), £(a) = £ s (a) + ^=1 A + @' = £s(a) + Ca(a) - c(a) < c(a) + X - p,, and, for
any b € supp + (7Tfc) n S, we have
2.4 Matroids
In section 1.3.2 we introduced the concept of base polyhedra in terms of an abstract
exchange axiom and mentioned that a matroid can be identified with a base polyhe-
dron having vertices of {0, l}-vectors. To compensate for such an abstract definition
of matroids, we explain here some linear-algebraic facts behind the abstract axioms.
The key is the Grassmann—Pliicker relation for determinants, qualitative analyses
of which lead to the concepts of matroids and valuated matroids.
where a i , . . . , < Z 5 6 R3 denote the column vectors. Let V denote the set of its
column indices; we have V = {1,..., 5} in our example.
The concept of matroids is derived from a combinatorial consideration of linear
dependence among column vectors. We say that a subset J of V is independent if
the corresponding column vectors {a, | j <E J} are linearly independent. Since a
subset of an independent set is obviously independent, we may focus on maximal
independent sets (maximal with respect to set inclusion). A maximal independent
2.4. Matroids 69
set is called a base and the family of bases (or base family) is denoted by B. In our
example we have
valid for any J,J'<^V and any i € J\J' (the proof is sketched in Note 2.30). The
notation detA[J — i+j] here means the determinant of A[J] with column i replaced
with column j.
To prove (B), suppose that J, J' e B. Then the left-hand side of (2.69) is
distinct from zero, and therefore, there exists a nonzero term, say, indexed by
j € J' \ J, in the summation on the right-hand side of (2.69). Then we have
det A[J -i+j]^Q and det A[J' + i - j] ^ 0; i.e., J-i+jeB and J' + i - j e B.
This proves (B). We emphasize that the exchange property (B) is derived from
the Grassmann-Pliicker relation by a qualitative consideration that distinguishes
between zero and nonzero, while disregarding the numerical information.
An alternative representation of linear independence among column vectors is
given by the rank function p : 2V —> Z denned by
The rank function has the following properties (proved in Note 2.31):
35
We consider J such that A[J] is square and implicitly assume an ordering of the elements of
J, which affects the sign of the determinant.
70 Chapter 2. Convex Functions with Combinatorial Structures
The third property (R3) shows that p is a submodular function. The rank function
and the base family determine each other by
From a given matrix we have thus derived a set family B C 2V with property
(B) and a set function p : 2V —> Z with property (R) (i.e., (Rl) to (R3)), where
V is the set of columns of the given matrix. It is all-important to realize that the
properties (B) and (R) are stated without reference to the given matrix and, as such,
they make sense as properties of a set family and a set function in general. Matroid
theory adopts these properties as abstract axioms and studies the combinatorial
structure implied by these axioms (and nothing else).
It turns out that a set family B satisfying axiom (B) and a set function p
satisfying axiom (R) are equivalent to each other. More precisely, we have the
following statement.
Theorem 2.29. The class of set functions p : 2V —> Z satisfying (Rl), (R2), and
(R3) and the class of nonempty families B C 2V satisfying (B) are in one-to-one
correspondence through the mutually inverse mappings (2.71) and (2.72).
In this sense, the two objects B and p represent one and the same combinatorial
structure, which is called a matroid. That is, a matroid is a pair (V,B) of a finite
set V and a family B of subsets of V that satisfies (B), or, alternatively, a matroid
is a pair (V,p) of a finite set V and a set function p on V that satisfies (R). We
may also denote a matroid by a triple (V, B, p). The set V is called the ground set,
B is the base family, and p is the rank function of the matroid. A member of B is
a base and a subset of a base is an independent set.
Though defined by such simple axioms, the concept of matroids is fundamental
and fruitful in studying combinatorial structures. The exchange property (B) above
is the germ of (B-EXC[Zj) treated in section 1.3.2 and (R3) is the submodularity
featured in section 1.3.1. The one-to-one correspondence between B and p stated
in Theorem 2.29 above is the germ of Theorem 1.9 that establishes the equivalence
between exchangeability and submodularity.
Note 2.30. The idea of the proof of the Grassmann-Plucker relation (2.69) is
indicated here for a 3 x 5 matrix
2.4. Matroids 71
where a'j = a,j (j = 1,3,4,5). The generalized Laplace expansion applied to det A
yields
On the other hand, subtracting the lower half (three rows) from the upper half
(three rows) of A yields
Note 2.31. We prove (Rl), (R2), and (R3) for the rank function (2.70) associated
with a matrix. (Rl) and (R2) are obvious. To prove (R3), let {a,,- | j G JXY} (where
JXY C XC\Y) be a base of {a,- | j € Xr\Y}. There exists some Jx C X\Y such that
{flj j & JXY U Jx} is a base of {aj | j 6 X}, since any set of independent vectors
can be augmented to a base. For the same reason, there exists some Jy C Y \ X
such that {dj | j G JXY U Jx U Jy} is a base of {a,,- j & X U Y}. Then we have
\JXY = p(X n Y), \JXY\ + \Jx = p ( X ) , \JXY\ + \JX\ + \JY = p(X U y), and
I JXY I + \JY\ < p(Y)> where the last inequality is due to the independence of the
vectors indexed by JXY U JY- Hence follows (R3). •
where we put uj(J) = -oo if det A[J] = 0 or A[J] is nonsquare (when det A[J] is
not denned). Using the notation B for the family of bases we have
72 Chapter 2. Convex Functions with Combinatorial Structures
(VM) For any J,J'eB and any i e J \ J', there exists j e J' \ J such
that J - i + j € B, J' + i - j e B, and
The inequality can be strict due to possible cancellations of the highest degree terms
on the right-hand side of (2.69).
In our example (2.73) we have detA[{l,2}] = detA[{3,4}] = 1, and hence
u>(J) = w(J') = 0 for J = {1,2} and J' = {3,4}. For i = I e J \ J' we can choose
j = 3e J' \ J, for which w( J - i + j) + u(J' + i - j) = 2.
The concept of valuated matroids is obtained by adopting (VM) as an axiom.
Namely, a valuated matroid is a pair (V,u>) of a finite set V and a set function
ui : 2V —* RU {—00} that satisfies (VM), where it is assumed that
is nonempty.
Not surprisingly, valuated matroids are closely related to matroids. First,
(VM) for u implies (B) for B. This means that (V, B) is a matroid if (V, ui) is a
valuated matroid. Accordingly, B is called the base family of the valuated matroid
(V, u). It is also said that a; is a valuation of the matroid (V, B).
Next, the maximizers of u> form the base family of a matroid. This is because,
for two maximizers J and J' with u(J) = u(J'} = max u, we have w( J — i + j) =
w(J' + i — j) = maxo; in (VM), which means (B) for the family of maximizers of w.
Furthermore, for any p — (p(i) i 6 V) (E Ry, the function u>[—p] : 2V —> RU{—00}
defined bv
is a valuated matroid, and therefore, the maximizers of w[—p] form the base family
of a matroid for each p 6 Rv. This property, in turn, characterizes a valuated
matroid as follows.
Theorem 2.32. Let (V, B) be a matroid with ground set V and base family B. A
function LU : B —> R is a valuation if and only if for any p : V —> R the maximizers
of uj[—p] form the base family of a matroid.
where domz/ = {\j | J e B} with B in (2.75). It is easy to observe that (VM) for
(jj is equivalent to (M-EXC[Zj) for /. That is, a; is a valuated matroid if and only
if / is an M-convex function. For instance, the valuated matroid (V, a;) associated
with the polynomial matrix in (2.73) can be identified with an M-convex function
( 0 ( a r e Bo),
/(a:) =4 -1 (a:eB\Bo))
[ +00 (x£Zv\B),
As is shown in Note 2.34, the function g is submodular over the integer lattice:
by (2.78) and (2.71), where p is the rank function of the underlying matroid (V, B).
Note 2.33. We have started with a polynomial matrix to define a function u> with
property (VM). As is evident from the proof, the same construction works for a
matrix over a non-Archimedian valuated field (van der Waerden [205]). The name
"valuated matroid" comes from this fact. •
74 Chapter 2. Convex Functions with Combinatorial Structures
Note 2.34. The submodularity (2.79) of the function g in (2.78) can be proved as
follows. First we show
which is a special case of (2.79) for p = Xi and </ = Xj- Take /, J e B with
g(0) = u(I) and g(Xi + Xj) = w(J) + \J n {ij}\. If \J n {i,j}\ < 1, we have
9(Xi + Xj) = max(s(x»),5(Xj)), which implies (2.80). The case of | J n {i,j}\ =
\I H {*> j}| = 2 is also easy. Suppose that \J n {«, j}| = 2 and |J n {i,j}\ < 1, and
assume j € J\/ without loss of generality. By (VM), there exists k & I\ J for which
u(I)+u(J) < u}(I+j-k)+u}(J-j + k). This establishes (2.80), since w(7) = 0(0),
u(J) = <?(X* + Xj) - 2, w(/ + j - fc) <ff(xj-)- 1, and w(J - j +fc)< 5 (Xi) - 1.
For p = (p A q) + Xi and 9 = (p A g) + Xj (* 7^ j)i the same argument applies to
u/(J) = w(J) + ^j€ j(pA 9)0) to prove (2.79). The general case can be treated by
induction on ||p — <?||i; we may assume supp + (p — q) ^ 0 and add the inequalities
(2.79) for (p - Xi, q) and for (p, (p V q) - Xi) with i 6 supp+(p - q). •
Bibliographical Notes
M-matrices are well-studied objects in applied mathematics and a comprehensive
treatment of their mathematical properties can be found in Berman-Plemmons
[9]. The connection of symmetric M-matrices to L-/M-convex quadratic functions
presented in section 2.1 is mostly based on Murota [147], whereas the general case
described in section 2.1.4 is due to Murota-Shioura [155]. See Fukushima-Oshima-
Takeda [71] and Doyle-Snell [39] for connections to probability theory.
For network flow problems in combinatorial optimization, Ford-Fulkerson [53]
is the classic, whereas Ahuja-Magnanti-Orlin [1] describes recent algorithmic devel-
opments. Thorough treatments of the network flow problem on the basis of convex
analysis can be found in Iri [94] and Rockafellar [178], the former putting more
emphasis on physical issues and the latter more mathematical. In particular, the
functions / and g in (2.37) and (2.38) are considered in the case of |T| = 2 in
[94] and [178]. The variational formulations are also discussed in Brayton-Moser
[19] and Clay [25]. The terminologies of current potential and voltage potential are
taken from [19], though they seem to be more often called content and cocontent.
M-convexity and L-convexity in the network flow problem are pointed out in Murota
[140], [141], [142].
Substitutes and complements in network flows are discussed in Gale-Politof
[73], Granot-Veinott [79], and Shapley [184], [185]. In particular, Theorem 2.22 is
due to Gale-Politof [73]. The connection to M-convexity and L-convexity (Theorem
2.23) is due to Murota-Shioura [158].
A number of books on matroids are available: Oxley [170], Welsh [211], and
White [212], [213], [214] are standard mathematical textbooks; Recski [175] realizes
a successful balance between theory and application; and Murota [146] empha-
sizes linear-algebraic motivations. For optimization on matroids, see, e.g., Cook-
Cunningham-Pulleyblank-Schrijver [26], Faigle [48], Korte-Vygen [115], Lawler
[119], and Schrijver [183]. Key papers in the development of matroid theory, in-
cluding Whitney [218], are collected in Kung [116]. Nakasawa [164] gives simple
2.4. Matroids 75
exchange axioms for matroids. The simultaneous exchange property for matroids
is due to Brualdi [20]. The concept of valuated matroids is due to Dress-Wenzel
[41], [42]. Chapter 5 of [146] is a systematic presentation of the theory of valuated
matroids including duality. Circuit axioms are investigated in Murota-Tamura
[159] and constrained optimizations in Althofer-Wenzel [2]. Oriented matroids are
another ramification of matroids, for which a comprehensive monograph of Bjorner-
Las Vergnas-Sturmfels-White-Ziegler [16] is available.
This page intentionally left blank
Chapter 3
Convex Analysis,
Linear Programming,
and Integrality
This chapter provides technical elements that are needed in subsequent chapters.
Some basic facts in convex analysis and linear programming are given in the first
two sections. The following two sections address integrality issues, i.e., integrality
for a pair of integral polyhedra and the concept of integrally convex functions.
where, if a(i) = —oo, for example, a(i) < x ( i ) is to be understood as —oo < x(i).
Sets such as [a, 6] and (a, 6) are referred to as closed intervals and open intervals,
respectively.
For a function / : R™ —> R U {±00} we define
Note that — oo is excluded from the possible function values of a convex function
and that the inequality (3.4) is satisfied, by convention, if it is in the form of
+00 < +00. A convex function with a nonempty effective domain is called a
proper convex function. A function is strictly convex if it satisfies (3.4) with strict
77
78 Chapter 3. Convex Analysis, Linear Programming, and Integrality
inequalities, i.e., if
A convex cone is a cone that is convex and a set S is a convex cone if and only if
it satisfies the condition
with nonnegative coefficients A; (1 < i < m) with unit sum (Y^iLi \ — 1) is called
a convex combination of those points. The convex closure of 5*, denoted as S, is
defined to be the set of all possible convex combinations of a finite number of points
of S. If 5" is convex, any convex combination of any finite set of points of S belongs
to S, and vice versa, and therefore S is convex if and only if S = S.
For a set 5, the intersection of all the convex sets containing S is the smallest
convex set containing S, which is called the convex hull of S. The convex hull of S
coincides with the convex closure of S. The convex hull of a set S is not necessarily
closed (in the topological sense). The smallest closed convex set containing S is
called the closed convex hull of S. For a finite set S, the convex hull is always
closed.
The affine hull of a set S is defined to be the smallest affine set (a translation
of a linear space) containing S and is denoted by aff S. The relative interior of S,
denoted as riS, is the set of points x € S such that {y e Rra | \\y - x\\ < e} n aff 5
3.1. Convex Analysis 79
is contained in S for some > 0. In other words, the relative interior of S is the
set of the interior points of S with respect to the topology induced from aff S.
We have so far defined convex functions and convex sets independently, but
they are actually closely related to each other. The indicator function of a set
S C R™ is a function 6S : Rn -> {0, +00} denned by
which shows that the convexity concept for functions can be induced from that for
sets. In passing, we mention that a function / is said to be closed convex if epi / is
a closed convex set in Rn+1.
A (global) minimizer of / is a point x such that f ( x ) < f(y) for all y. The
set of the minimizers of /, denoted by
The sum of two convex functions is convex, and the infimal convolution of two
convex functions is convex if it does not take the value of — oo. If /i and /2 are
the indicator functions of sets Si and 82, then /i + /2 and /id /2 are the indicator
functions of the intersection Si fl 82 and the Minkowski sum Si +82, respectively,
where
when this limit (finite or infinite) exists, where a j 0 means that a tends to 0 from
the positive side (a > 0). For convex /, the limit exists for all d, and f ' ( x ; d ) is a
convex function in d. For polyhedral convex /, there exists £ > 0, independent of
x 6 dom /, such that
3.1. Convex Analysis 81
This is a convex function in p, being the maximum of (infinitely many) affine func-
tions in p indexed by x. The function /• is also called the (convex) Legendre-Fenchel
transform of /, and the mapping / i—> /* is referred to as the (convex) Legendre-
Fenchel transformation.
In the favorable situation where / is a smooth convex function and the supre-
mum in (3.26) is attained by a unique x = x(p) for each p, we have
the conjugate is given by f'(p) = exp(p — 1). This can be verified by a simple
calculation based on (3.27). •
82 Chapter 3. Convex Analysis, Linear Programming, and Integrality
holds for any A > 0 and p 6 Rra (this condition yields g(0) = 0 if dom g ^ 0).
Theorem 3.2 implies a one-to-one correspondence between closed convex sets and
positively homogeneous closed proper convex functions. In this sense, positively
homogeneous convex functions are convex sets in disguise. For a closed convex
function / and a point x in the relative interior of dom /, for example, the directional
derivative f ' ( x ; d ) is a positively homogeneous closed proper convex function in d
and it coincides with the support function of the subdifferential &R,f(x):
The support function of a convex cone S C R™ agrees with the indicator function
of another convex cone,
called the polar cone of S. By Theorem 3.2, (S*)* = S for a closed convex cone S.
When S is represented as
3.1. Convex Analysis 83
The duality principle constitutes the core of convex analysis. It can be stated
in many different forms, but we focus here on separation theorems (for sets and for
functions) and the Fenchel duality theorem.
The following is the separation theorem for convex sets (see Fig. 3.2).
Theorem 3.4 (Separation for convex sets). Let 81,82 C R™ be nonempty convex
sets.
84 Chapter 3. Convex Analysis, Linear Programming, and Integrality
If, in addition, Si and 82 are closed and at least one of them is bounded, then the
inequality > above can be replaced with strict inequality >.
(2) ri S\ n ri 62 — 0 if and only if there exists a vector p* € R™ such that (3.37)
holds and
Proof. See Theorems 11.3 and 20.2 and Corollary 11.4.2 of Rockafellar [176]. D
The separation theorem for convex functions, illustrated in Fig. 3.3, asserts
the existence of an affine function that lies between a convex function and a concave
function.
Proof. The proof is based on Theorem 3.4 applied to epigraphs. See Corollary 5.1.6
in Stoer-Witzgall [194] and the proof of Theorem 31.1 in Rockafellar [176]. D
Another expression of the duality principle is in the form of the Fenchel duality.
This is a min-max relation between a pair of convex function / and concave function
h and their conjugate functions /* and h°. We include a proof to demonstrate the
equivalence of the Fenchel duality and the separation for functions.
Proof. By the definitions (3.26) and (3.28) of the conjugate functions, we have
for any x and p. This shows inf > sup in (3.41). Hence (3.41) holds if inf = — oo or
sup = +00. In what follows we assume inf > —oo and sup < +00.
Suppose that (al) or (a2) is satisfied. Then inf in (3.41) is of finite value, say,
A, and Theorem 3.5 applies to (/ — A, /i) and yields some a* G R and p* G R™
such that
This means that f ' ( p * ) < -a* -A and h°(p*) > -a*, implying inf = A < h°(p*)-
f*(p*} < sup and hence (3.41). This also shows that p* attains the supremum.
In the remaining case where (bl) or (b2) is satisfied, we can use a similar
argument for (/*, h°) on the basis of the identities (/*)* = / and (/i°)° = h shown
in Theorem 3.2. In the case (b2) note that the conjugate function of a polyhedral
convex function is again polyhedral. D
37
By this we mean that / and —h are closed convex functions.
86 Chapter 3. Convex Analysis, Linear Programming, and Integrality
Example 3.7. The separation theorem and the Fenchel duality theorem are illus-
trated for the convex function
and the concave function h(x) = —f(—x). The graphs of Y = f ( x ) and Y = h(x) are
tangent to the F-axis at the origin (0,0), and therefore there exists no separating
affine function, although f ( x ) > h(x) (Vx). This does not contradict Theorem
3.5, since neither (al) nor (a2) is satisfied. This shows the importance of the
conditions (al) and (a2) in Theorem 3.5. The conjugate functions are given by
f*(p) — exp(p — 1) and h°(p) = — exp(p — 1), and hence
Therefore, the infimum and the supremum in the Fenchel duality (3.41) are both
equal to 0; the infimum is attained by x = 0, whereas the supremum is not attained.
Note that the condition (bl) in Theorem 3.6 is met. •
Example 3.8. The infimum and the supremum in the Fenchel duality (3.41) can
be distinct if none of the conditions (al)-(b2) in Theorem 3.6 is satisfied. For the
convex function / and the concave function h in x = (x(l),a;(2)) defined by
we have inf = 0 and sup = —1 in (3.41). Note that dom/* = {p \ p(2) < 0} and
dornh0 = {p p ( l ) > 0,p(2) > 0}, which shows that ri (dom/*) n ri (domft 0 ) = 0,
the failure of condition (bl) in Theorem 3.6. •
The addition (3.19) and the infimal convolution (3.20) are conjugate oper-
ations with respect to the Legendre-Fenchel transformation. For proper convex
functions /i and /2 we have
where the latter is true under the assumption that ri (dom/i) n ri (dom/g) ^ 0.
Theorem 3.9 (Farkas lemma). For a matrix A and a vector b, the conditions (a)
and (b) below are equivalent:38
(a) Ax = b for some nonnegative x > 0.
(b) j/ T 6 > 0 for any y such that y1 A > 0 T .
Proof, [(a) => (b)]: It follows from Ax = b, x > 0, and y1 A > 0T that j/ T 6 =
yTAx > 0.
[(b) => (a)]: Let 5 be the convex cone generated by the column vectors a,j
(j = l,...,n)ofA- i.e., S = {£"=1 x(j)aj \ x(j) > 0}. If (a) fails, then b £ S, which
implies, by the separation theorem for convex sets (Theorem 3.4), that y1 a^ > 0
(j = 1 , . . . , n) and y T 6 < 0 for some y. D
The LPs in such a pair are said to be dual to each other. For convenience of
reference, we call the problem on the left the primal problem and the one on the
right the dual problem. We denote the feasible regions of the above problems by
This common value is finite if and only if both P and D are nonempty, and in
that case, the infimum and the supremum are attained by some x & P and y e D,
respectively.
38
Inequality between vectors means componentwise inequality; e.g., x > 0 for x = (x(j))™=l
means x(j) > 0 for j = 1 , . . . , n.
39
By convention, inf xe p = +00 if P = 0 and sup yeD = —oo if D = 0.
88 Chapter 3. Convex Analysis, Linear Programming, and Integrality
Proof. (1) is easy to see. The essence of this theorem lies in (2), which can be
derived from the Farkas lemma. Then (3) follows. See, e.g., Chvatal [24], Dantzig
[36], Schrijver [181], and Vanderbei [206]. D
Note that the z'th row of C is the characteristic vector of Xi. For two chains C1 and
C 2 , with incidence matrices C1 and C2, the matrix A— [^ 2 ] is totally unimodular.
To prove this, it suffices to assume that A is square and to show detA e
{0, ±1}. Let Ck : X? <^X$ ^ • • • ^ X^k (k = 1,2) be the chains and, for k = 1,2,
define Dk to be the matrix with the zth row of Ck replaced with the characteristic
vector of X? \ Xk_{ for i = 1,.. .,mk, where Xfi = 0. Put A = [ J^ 2 ]. Then
det A = ± det A by the construction. We also have det A £ {0, ±1}, since A, having
at most one entry of 1 and at most one entry of —1 in each column, can be regarded
as a submatrix of the incidence matrix of a graph (see Example 3.11). •
The following theorem relates the total unimodularity of the coefficient matrix
to the integrality of optimal solutions of LPs.
3.2. Linear Programming 89
Proposition 3.14. Let G = (V+, V~; E) be a bipartite graph with a perfect match-
ing, and let c : V+ x V~ —> R U {+00} be a (weight or cost) vector such that
c(u,v) < +00 •& (u,v) G E. Then there exist a vector*0 p : V+ U V~ —> R and
orderings of vertices V+ = {HI, ..., um} and V~ = {v\,... ,vm} such that
The set of arcs {(ui, Vi) i = 1,..., m} is a perfect matching of minimum weight,
and, therefore,
m
Minimum weight of a perfect matching — Jj (p(vi) — p(ui)). (3.49)
z=l
and c is the vector of weights (c(u, v) \ (u,v) G E). Since A is totally unimodular
by Example 3.11, the optimal solution x may be chosen, by Theorem 3.13, to
be an integer vector, which, being a {0, l}-vector because of the constraints, can
be interpreted as the incidence vector of an optimal matching. The dual optimal
solution can be identified with a vector p : V+ U V~ —> R, and the condition (3.48)
follows from the dual feasibility and the complementarity. D
4i:)
This p is called a potential or an optimal potential.
90 Chapter 3. Convex Analysis, Linear Programming, and Integrality
which means that all the integer points contained in the convex hull of S belong to
S itself. A finite set of integer points is hole free if and only if it is the set of integer
points in some integral polytope.41
The hole-free property (3.50) seems to be a natural requirement for a discrete
set to be qualified as being convex. This is indeed compatible with our previous
naive idea of convexity for discrete functions in terms of the extensibility to convex
functions formulated in (1.11). We define the indicator function 63 '• Z™ —> {0, +00}
of a discrete set S by
Then a discrete set is hole free if and only if its indicator function is extensible to
a convex function.
We are now interested in the compatibility of the hole-free property with the
Minkowski addition (3.21), which is one of the fundamental operations in convex
analysis. We define the Minkowski sum Si + 63 of two discrete sets Si, 62 C Zn by
which we also call the discrete Minkowski sum or the integral Minkowski sum to
emphasize the discreteness. If the hole-free property can be qualified as a discrete
version of convexity, this property should be preserved in Minkowski addition. Con-
trary to this optimistic expectation, the Minkowski sum of hole-free sets can have
a hole, as is demonstrated in Example 3.15 below.
are hole free with Si = S^"n Z2 and S2 = S2 n Z2 (see Fig. 3.4). Nevertheless, the
discrete Minkowski sum
41
A polytope is a bounded polyhedron.
3.3. Integrality for a Pair of Integral Polyhedra 91
The above example issues the following warnings to us about the integrality
for a pair of hole-free discrete sets Si and £2 with Sk = Sk n Z™ (k — 1,2).
The above facts suggest that the hole-free property (3.50) alone is not appro-
priate as the condition of discrete convexity for sets. Some deeper combinatorial
properties are needed.
The first two properties above will turn out to be critical in many situations,
and, in fact, they are essentially equivalent to each other.
Proposition 3.16. Suppose that a family T of sets of integer points has the
property
where x — S = {x — y \ y £ S}. Then conditions (a) and (b) below are equivalent
for T. __ _
(a) VSi, S2 e f: Sl n S2 = 0 => Sl n S2 = 0.
(b) VSi, S2 € f: Sl+S2=S1 + S2n Z™.
Proof, (a) =» (b): For_£ e_S_i + S2 n Z" we have xe(Sl+ S2) n Zn by Proposition
3.17 (4) below. Hence S[ nS2 ^ 0 for S[ = Si and S'2 = x-S2. By (a), there exists
y e S[ n S2. Then y € Si and 3z & S2 : y = x - z. Therefore, x € Si + S2.
92 Chapter 3. Convex Analysis, Linear Programming, and Integrality
Proof. (1), (2), and (3) are obvious. We prove (4). First, note that Si + S2 2
Si + S2, which follows from 51 + 52 2 Si + 62 and the convexity of Si + S2. To
show the reverse inclusion, take
where A, > 0, ]T\ Aj = 1, y, € Si, /u.,,- > 0, J^ • /Zj = 1, and Zj € ^2 (the summations
being finite sums). With i>ij = \i/j,j we obtain
where, if a(i) = — oo, for example, a(i) < x(i) is to be understood as —oo < x(i).
The restriction of a function / : Z n —> R U {+00} to an interval [a, b] is denned as
the function /[0,6] : Z" —-> R U {+00} given by
3.4. Integrally Convex Functions 93
we say that / is convex extensible and call / the convex extension of /.42 The
following fact is easy to see.
where, for z € R in general, \z\ denotes the smallest integer not smaller than z
(rounding up to the nearest integer) and \_z] the largest integer not larger than z
(rounding down to the nearest integer). Note an alternative expression
42
We say that / is concave extensible if —/ is convex extensible and then — (—/) is the concave
extension of /.
94 Chapter 3. Convex Analysis, Linear Programming, and Integrality
with
with an integral point z 6 Z™, but is not necessarily convex in the entire space R n .
If / is convex on R n , the function / is said to be integrally convex. Alternatively,
we can define
Proof. It suffices to show <*=. Put NI(X) = {y 6 R™ \\y - x\\x < 1} for
x e domz/. By (3.63) and (3.65) we have f ( x ) < f(y]_ for all y € 7Vi(z). Combining
this with integral convexity (3.64) shows f ( x ) < f(y) (Vy 6 N\(x)), the local
minimality of / at x. Since / is convex, a; is a global minimizer of / by Theorem
1.1 and, a fortiori, a global minimizer of /. D
Proof. It suffices to show that the latter condition in (3.66) implies f ( x ) < f ( x +
XY — Xz) for any disjoint Y and Z. On putting U — {1,... ,n} \ (Y U Z) and
x° = x — 1 we have f(x) — f ( x ° ) and
Here we have f ( x ° + \Xu + Xv) > f ( x ° ) , since f ( x ° + ^xu + Xv) can be represented,
by integral convexity, as a convex combination of f ( x ° + xw + XY) with W C U
and f ( x ° + xw + XY) > f ( x ° ) for any W C U by the assumption. D
Note 3.23. The optimality criterion in Theorem 3.21 is certainly local, but not
satisfactory from the computational complexity viewpoint. We need O(3n) function
evaluations to verify the local optimality condition in (3.65). •
is the set of univariate discrete convex functions. Similarly, we denote by C[Z —*• Z]
the set of integer-valued univariate discrete convex functions.
Proof. Put f(x) = / 0 (z) + Y^i=i /i( x (*))) where /0 is integrally convex and /j e
C[Z —> R] for i = 1,..., n. For any (A y ) e A with X]yeAr(z) ^yV =x
> we ^ave
Proposition 3.25.
(1) A separable convex function is integrally convex.
(2) /[—p] is integrally convex for integrally convex f and vector p € R™.
Proof. This follows from (3.72), since N ( x ) = {x} for an integer point x. D
Note 3.27. The family of integrally convex sets has neither convexity in intersec-
tion nor convexity in Minkowski sum. Example 3.15 shows this. •
Proof. The implication => was shown in Proposition 3.28. For the converse we are
to show f ( x ) = f ( x ) for x 6 R™. If x £ dom/, we have f ( x ) = f ( x ) = +00 by
dom/ = dom/ and (3.62). Assume x e dom/, and consider a pair of (mutually
dual) LPs:
Here (p,a) and (A y | y £ dom/) are the variables of (P) and (D), respectively.
Problem (P) is obviously feasible, and so is (D) by a; € dom/. Let (p*,a*) and
A* = (A* y e dom/) be optimal solutions of (P) and (D), respectively. Then
98 Chapters. Convex Analysis, Linear Programming, and Integrality
Proof. The proof is not difficult; see Lemma 6.13 in Murota-Shioura [152]. D
Note 3.31. The intersection of integrally convex sets is not necessarily integrally
convex; e.g., Si = {(0,0,0), (0,1,1), (1,1, 0), (1,2,1)} and S2 = {(0,0,0), (0,1,0),
(1,1,1), (1,2,1)} are integrally convex, but their intersection Si n 62 = {(0,0,0),
(1,2,1)} is not, since (Si n S2) n N(x) = 0 for x = (1/2,1,1/2) e Si n S2. This
implies also that the sum of integrally convex functions is not necessarily integrally
convex. The discrete separation theorem (see section 1.2) does not hold for integrally
convex functions; Example 1.5 shows this.
holds for any x,y € domz/ and any a e [0, I]R. An integrally convex function
satisfies this condition. The optimality criterion (3.65) stated for integrally convex
functions in Theorem 3.21 is in fact valid for Miller's discrete convex functions. •
3.4. Integrally Convex Functions 99
Bibliographical Notes
The introduction to convex analysis in section 3.1 is kept to the minimum needed
for later developments in this book. For a systematic and comprehensive account,
see Borwein-Lewis [17], Hiriart-Urruty-Lemarechal [89], Rockafellar [176], Rock-
afellar [177], Rockafellar-Wets [179], and Stoer-Witzgall [194]. In particular, see
Theorems 11.3 and 20.2 and Corollary 11.4.2 of [176] for separation for convex sets
(Theorem 3.4); Corollary 5.1.6 in [194] and the proof of Theorem 31.1 in [176]
for separation for convex functions (Theorem 3.5); and Theorem 31.1 in [176] and
Corollary 5.1.4 in [194] for Fenchel duality (Theorem 3.6). Example 3.8 is taken
from [194].
References on linear programming abound in the literature; see, e.g., Chvatal
[24], Dantzig [36], Schrijver [181], and Vanderbei [206]. Matching is one of the
central topics in graph theory, the standard reference being Lovasz-Plummer [125].
Matching is also fundamental in combinatorial optimization; see Cook-Cunningham-
Pulleyblank-Schrijver [26], Du-Pardalos [43], Korte-Vygen [115], Lawler [119], and
Nemhauser-Wolsey [167].
Section 3.3 is a collection of basic facts, as presented in Murota [147]. The
terms convexity in intersection and convexity in Minkowski sum are coined here.
Proposition 3.16 is explicit in Danilov-Koshevoy [32], where a general framework
for convexity in intersection and convexity in Minkowski sum is provided.
The concept of integrally convex functions was introduced by Favati-Tardella
[49], where the effective domains are assumed to be integer intervals. The optimality
criterion (Theorem 3.21) is in [49]. Propositions 3.24 and 3.28 are due to Murota-
Shioura [153], and Theorem 3.29 (implicit in [153]) is taken from Murota [147].
Miller's discrete convex functions are introduced by Miller [130] along with the
optimality criterion (3.65).
This page intentionally left blank
Chapter 4
M-convex sets form a class of well-behaved discrete convex sets. They are defined
in terms of an exchange axiom and correspond one-to-one to integer-valued sub-
modular set functions. An M-convex set is exactly the same as the integer points
contained in the base polyhedron associated with some integral submodular func-
tion. This chapter, accordingly, is a systematic presentation of known results in the
theory of matroids and submodular functions from the viewpoint of discrete convex
analysis.
4.1 Definition
Let V be a finite set, say, V = {!,..., n}. A nonempty set of integer points B C Zv
is defined to be an M-convex set if it satisfies the following exchange axiom:
Here supp+(x — y) and supp~(x — y) are the positive support and the negative
support of x — y defined in (1.19) and \u is the characteristic vector of u 6 V. We
denote by A^o[Z] the set of M-convex sets.
M-convexity thus defined for a set B C Zv is equivalent to the M-convexity
of the indicator function SB '• Zv —> {0,+00} (defined in (3.51)). Namely, B is
an M-convex set satisfying (B-EXC[Z]) if and only if SB is an M-convex function
satisfying (M-EXC[Zj) introduced in section 1.4.2.
Recall that we encountered (B-EXC[Zj) in section 1.3.2 as the exchange prop-
erty that characterizes the sets of integer points associated with a submodular
function (see Theorem 1.9). Hence, an M-convex set is exactly the same as the
set of integer points contained in the base polyhedron defined by an integer-valued
submodular set function.
An immediate consequence of the exchange axiom (B-EXC[Zj) is that an M-
convex set lies on a hyperplane {x € Rv x(V) = r} for some r € Z, where we use
101
102 Chapter 4. M-Convex Sets and Submodular Set Functions
the notation
Proposition 4.1. For an M-convex set B we have x(V) — y(V) for any x,y € B.
is nonempty. Take a pair (x,y) G V with minimum \\x — y\\i\ we have \\x — y\\i > 4.
Fix u* € supp+(o; — y ) as above, take any UQ 6 supp+(x — y — x«, )> and put
Put y" = y' + Xu* ~ Xv = y + Xu0 + Xu. ~ Xv0 - Xv Note that y + Xu. - Xv £ B,
since (a;, y) e T> and x - Xu* + Xv € B. If X D Y ^ 0, we have y + Xu, ~Xv$-B and
y + Xu.~ Xv0 i B, and therefore, y" £ B by (B-EXC+[Z]). If X n y = 0, we have
V + Xu, ~ Xvi B and y + x«0 ~Xv<£B, and therefore, y" ^ B by (B-EXC_i oc [Z]).
In either case we have (x, y') G T>. D
Throughout this book we assume, for a set function p in general, that p(0) = 0,
p(V) is finite, and either p : 2V -> R U {+00} or p : 2V -> R U {-oo}.
The Lovdsz extension*3 of a set function p is a function p : R y —> R U {±00}
defined as follows. Given a vector p e R^, we denote by p\ > p2 > • • • > pm the
distinct values of its components and put
Since pi — pi+i > 0 ( l < z < m — 1) and p(Um) = p(V) is finite on the right-hand
side of (4.6), we have
We denote by <S[R] the set of submodular set functions p with p(0) = 0 and p(V) <
+00 and by <S[Z] the set of integer valued such submodular set functions; i.e.,
For p € <S[R], the effective domain T> = domp forms a ring family (sublattice of the
Boolean lattice 2V), which means, by definition, that
43
The Lovasz extension is also called the Choquet integral or the linear extension. Often p(p)
is denned only for nonnegative p, although it is more convenient for us to define it over the entire
space R^.
4.3. Submodular Functions and Base Polyhedra 105
the base polyhedron associated with p. A point (element) of B(/o) is called a base
and an extreme point of B(p) is an extreme base.
Proof. For simplicity of description we assume that dom p has a maximal chain
of length n = \V\. On suitably indexing the elements of V, V = {v\,V2, • • • ,vn},
we have Vj = {v-\_, v%,..., Vj} € domp for j = 1,..., n. Define a vector x £ R^ by
x(vj) = p(Vj) - p(Vj-i) for j = 1,... ,n with V0 = 0. We show x(X) < p(X) by
induction on \X\. When \X\ = 0, this is obviously true by p(0) = 0. When |X| > 1,
let j be the maximum index such that u, £ X. Then we have
Proof. For simplicity we assume that dom p has a maximal chain of length n = \V\.
Consider a pair of linear programs (LPs):
Here x €. Rv and y = (yx \ X e domp) are the variables of (A) and (B), respec-
tively, and p 6 R^ is regarded as a parameter. Note that the equality constraints
in (B) can be written as
106 Chapter 4. M-Convex Sets and Submodular Set Functions
is a feasible solution for (B). For the proof of =>, take a feasible y that maximizes
r = E{yx\x\2 l^edomp}.
Claim 2: C = [X G domp \ {V} yx > 0} U {V} forms a chain.
(Proof of Claim 2) For Y, Z G C \ {V} with yY>Vz> 0, we have the identity
By LP duality (Theorem 3.10 (1)), this shows the optimality of y* for (B); hence
^xy*xp(X) = min(B). On the other hand, £x V*xp(X) = p(p) by (4.6) and
(4.17). Combining these with (4.16) shows (4.14).
4.3. Submodular Functions and Base Polyhedra 107
In the case of infeasible (B), we have min (B) = +00 in (4.16), whereas p(p) —
+00 by Claim 1 and (4.8). Hence (4.14) follows. D
Proof. For each p, the optimal base (4.18) is an integer vector for p e S[Z]. D
Note 4.7. A ring family (4.12) typically arises from a graph, and conversely, any
ring family can be represented by a graph. For a directed graph G = (V, A) with
vertex set V and arc set A, we call a subset X of V an ideal if (u, v) £ A and u & X
imply v 6 X. That is, X is an ideal if and only if no arc leaves X. The set of ideals
is a ring family with {0, V} C T>. Conversely, for a ring family T> on a set V, we
consider a directed graph G = (V, .A) with
Denote by minP the minimum element of T) and by max I? the maximum element.
Then T> coincides with the family of ideals of G that contain min T> and are contained
in max£>. In particular, V = V(G) if {0, V] C V. The graph G given by (4.20) is
transitive; i.e., (u,v) 6 A, (v,w) 6 A => (u,w) 6 A. The constructions (4.19) and
(4.20) establish a one-to-one correspondence between the set of ring families on V
including {0, V} and the set of transitive directed graphs with vertex set V.
An acyclic graph44 G = (V, A) represents a partial order ^ on V defined by
[v X u &• v is reachable from u by a directed path], and the corresponding ring
family T> has the property that a maximal chain of V is of length \V\. In this
particular case, (4.19) reads
Note 4.8. The minimizers of a submodular set function p form a ring family. Let
a denote the minimum of p. If p(X) = p(Y) = a, we have
44
A directed graph is called acyclic if it does not contain directed cycles.
108 Chapter 4. M-Convex Sets and Submodular Set Functions
This follows from Note 4.8 applied to p(X) - x ( X ) . Note that {0, V} C V(x). •
Note 4.10. LP (A) in the proof of Proposition 4.5 is the problem of maximizing
the weight of a base x € B(/o) with respect to a given weight vector p. An optimum
solution is given by (4.18) for an ordering of the elements of V satisfying p(v\) >
• • • > p(vn), where domp is assumed to have a maximal chain of length n = \V\.
Often we refer to this fact by saying that the greedy algorithm works for finding an
optimal base. •
Note 4.11. A partial order on V is associated with each extreme base. Assume
that T> = dom/9 has a maximal chain of length n = |V| and denote by ^ the partial
order on V associated with T>, as in (4.21). A linear order < on V, or an ordering
of elements of V, is said to be an extension of ^ if [v ^ u => v < u]. A linear
extension of ^ generates an extreme base in (4.18). Conversely, any extreme base
x is generated in this way, but there can be several linear orders that generate x.
We define a partial order -<x on V associated with x by
The partially ordered set P(x) = (V, ^x) thus defined corresponds to the family of
tight sets T>(x) in the sense of Note 4.7. In particular,
with distinct Xi (1 < i < m). We may assume that there is a positive integer N
such that N\i e Z + (1 < i < m). For a representation of the form (4.24), we define
gives another representation of the form (4.24), for which <J> is smaller at least by
4min(Aj,Afc) (> 4/AT). The condition NXi € Z+ is preserved in this modification.
The process of modification ends with m = 1, showing x £ B. D
(l)peS[Z].
( 2 ) B = B(p).
Proof. (1) First we show the submodularity inequality (4.9) for X and Y with
p(X U y) and p(X n y) both finite. Take y,z&B with p(A" U y) = y(A" U y) and
p(X n y) = z(X n y); we choose such (y, z) with | \y — z\ \i minimum. Then we have
y(v) = z(v) (v G X n y), since, otherwise, 3u G (X n y) n supp + (z — y) and, by
(B-EXC+[Z]), 3v 6 supp^(z — y) such that y' = y + \u — Xv G B, for which we
have y'(X U y) > y(X U y) and \\y' — z\\\ < \\y - z\\i — 2, a contradiction to our
choice of ( y , z ) . Therefore,
(2) The inclusion B C ~B(p) is obvious. For the converse we may assume that
p(X) is finite for all X C V (by the same argument as in the latter half of the proof
of (1)). Let z e Rv be an extreme point of B(p). As we have seen in (4.18), there
is an ordering of the elements of V, say, V = {i>i,..., vn}, such that z(Vj) = p(Vj)
with Vj = {vi,... ,Vj} for j = 1 , . . . , n , where n = \V\. For each j = 1 , . . . , n , there
exists Xj G B with p(Vj) = Xj(Vj). By repeated applications of (B-EXC + [Zj), as in
the proof of (1) above, we can show the existence of x & B such that x(Vj) = Xj(Vj)
for j = 1 , . . . , n. We then have z(Vj) = p(Vj) — Xj(Vj) = x(Vj) for j = 1 , . . . , n,
which means z = x s B. Since any extreme point of B(p) is contained in _B, we
must have B(/o) C B. 0
Proof. (1) First, B is nonempty by Propositions 4.4 and 4.6. By Proposition 4.2
it suffices to show (B-EXC+[Z]). Suppose, to the contrary, that (B-EXC+[Z]) fails
for some x, y e B and some u & supp+(x — y ) . For each v £ supp~(x — y), we have
y + X u ~ X v $. B, which, together with the integrality of p, implies the existence of a
tight set Xv e V(y) with u € Xv and v £ Xv. For Z — r\vesuPP-(x-y) Xv-, we have
y(Z) = p(Z) by (4.23), whereas x(Z) > y ( Z ) by u e Z and Z n supp~(a; - y) = 0.
It then follows that x(Z) > p ( Z ) , a contradiction to x & B(p).
(2) This follows from (4.14) with p = xx and (4.7). An alternative proof is as
follows. Since p(X) > sup{x(X) \ x € B(p)} is obvious, it suffices to establish an
equality here in the case of sup < +00. Let x 6 B(p) attain the supremum. Then
for each u 6 X and u e V" \ X there exists Xuv e £>(£) with u e JsTuw and w ^ X uu .
Since P(x) is a ring family (see (4.23)), we have X = \Ju€X C\vev\x xuv 6 T>(x],
which means x ( X ) = p ( X ) . D
Proposition 4.13 and Theorem 4.12, respectively. For p 6 <5[Z], we have ^>(p) €
.A/fo[Z] and $ o *&(p) = p by Propositions 4.6 and 4.14. D
Theorem 4.16 (Lovasz). A set function p : 2V —> R U {+00} with p(0) = 0 and
p(V) < +00 is submodular if and only if its Lovasz extension, p : Hv —> Ru{+oo},
defined in (4.6) is convex; i.e.,
Proof. It suffices to prove •£= only. Since p is positively homogeneous, the convexity
of p implies
an
This shows the submodularity of p, since p(xx) — p(X)i P(XY) — pW)i d P(XX +
XY) = p(X U Y) + p(X n y) by (4.7) and (4.6). D
Moreover, if p and /j, are integer valued (namely, p, —/j, 6 «5[Z]), the vector x* can
be chosen to be integer valued (namely, x* € Zv).
The combinatorial essence of the above theorem lies in the second half, claim-
ing the existence of an integer vector for integer-valued functions, whereas the exis-
tence of a real vector x* alone can be proved on the basis of the separation theorem
112 Chapter 4. M-Convex Sets and Submodular Set Functions
in convex analysis and the relationship between submodularity and convexity stated
in Theorem 4.16 (see Note 4.20).
The proof of Theorem 4.17 is based on Edmonds's intersection theorem below,
which is the most important duality theorem in the theory of submodular functions.
For a submodular set function p £ <S[R], we define a polyhedron
Theorem 4.18 (Edmonds's intersection theorem). Let pi,p2 '• 2V —> Ru{+oo} be
submodular set functions with pi(0) = p2($) = 0, pi(V) < +00, and p2(V) < +00
(namely, pi,pz £ S[R}). Then
Moreover, if pi and p2 are integer valued (namely, pi,p2 & <5[Z]), the polyhedron
P(pi) C}P(p2) is integral in the sense of
and there exists an integer-valued vector x* that attains the maximum on the left-
hand side of (4.29).
Here a; e Ry and (yix \ X e £>i,z = 1,2) are the variables of (A) and (B),
respectively, and p 6 Hv is a parameter. Note that the equality constraints in (B)
can be written as
Problem (A) is feasible. Let p be such that (A) has an optimal solution. Then
Problem (B) also has an optimal solution by LP duality (Theorem 3.10).
4.5. Submodular Functions as Discrete Convex Functions 113
There exists an optimal solution to (B) such that Ci = {X e T>i \ yix > 0}
forms a chain for i — 1,2. To prove this, take an optimal solution that maximizes
If ViY > Viz > 0 for some i e {1,2} and Y, Z € T>j, we have
where the latter is due to the submodularity of p^. This means that the modification
of (j/ix) to (y'ix) defined by
for pi,p2 G «S[Z], the integrality of the intersection of integral base polyhedra.
We are now in the position to prove Frank's discrete separation theorem (The-
orem 4.17). Consider Edmonds's intersection theorem (Theorem 4.18) for
114 Chapter 4. M-Convex Sets and Submodular Set Functions
It follows from (4.26) that the minimum on the right-hand side of (4.29) is equal
to (j,(V). Hence, there exists x* e P(pi) n P(p2) such that x*(V) = fj,(V). The
condition x* e P(pi) is equivalent to x*(X] < p(X) (\/X C V), and x* e P(pz)
is equivalent to x*(V \X) < p2(V \ X) (VX C V), which is equivalent further to
x*(X) > A*PO (VX C V) by x*(V) = n(V). Hence, p(X) > x*(X) > fi(X) for all
X C V. For integer-valued p and /u, we can take an integral x* by the integrality
assertion in Theorem 4.18. This completes the proof of Theorem 4.17.
Note 4.20. The first half of the discrete separation theorem (Theorem 4.17) is
derived here from the separation theorem in convex analysis and the relationship
between submodularity and convexity given in Theorem 4.16. Let p and /} be the
Lovasz extensions of p and /z, respectively. We have p(p) > £t,(p) (\/p e R+) by the
assumption p > n as well as the definition (4.6) of the Lovasz extension. Define
functions g : R7 —> R U {+00} and k : Rv —> R U {-00} by
Then g is convex and k is concave by Theorem 4.16; these functions are polyhedral,
and domg PI domfc ^ 0 by <?(0) = fc(0) = 0. The separation theorem in convex
analysis (Theorem 3.5) applies to the pair of g and k, yielding j3* 6 R and x* e R K
such that
Theorem 4.21 (Discrete separation for M-convex sets). Let BI and B2 (C Zv) be
M-convex sets. If they are disjoint (BinB2 = 0), there exists p* e {0, Ij^UfO, -1}V
such that
4.6. M-Convex Sets as Discrete Convex Sets 115
The content of Theorem 4.21 consists of two claims. The first, explicit in the
statement, is that the separating vector p* is so special that p* or — p* is a {0,1}-
vector. The second, less conspicuous and more subtle, is that BI r\B2 = 0 is implied
by B\r\B2 — 0, since otherwise the inequality (4.33) is impossible. The implication
We now turn to the Minkowski sum of M-convex sets. The claim (3) in the
theorem below says that the Minkowski sum of M-convex sets is again an M-convex
set. An important consequence of this is that the family of M-convex sets has the
property of convexity in Minkowski sum considered in section 3.3.
Theorem 4.23.
(1) For submodular set functions pi,p2 6 «5[R], we have
Proof. (1) The proof of B(pi) + B(p 2 ) C B(pi + p2) is easy: A vector x e
B(pi) + B(p2) can be decomposed as x = Xi + x2 with £j € B(pi) (i = 1, 2), which
implies x ( X ) = xi(X) + x2(X) < pi(X)+ p2(X) (equality for X = V). Conversely,
for x e B(pi + p2), we have x ( X ) — p2(X) < p i ( X ) , and by the discrete separation
theorem (Theorem 4.17), there exists y 6 Rv such that x(X) - p2(X) < y(X) <
pi(X), with equality for X = V. For z = x — y, we have y e B(pi) and z e B(p2).
Hence x = y + z e B(pi) + B(p 2 ).
(2) This is because the vectors x, y, z above can be taken to be integral.
(3) We can represent 5; = B(PJ) n Z y with p{ 6 «S[Z] (z = 1,2). Then the
left-hand side of (2) is B\ + B2, whereas for the right-hand side of (2) we have
Note 4.25. The intersection of two M-convex sets is not necessarily M-convex,
though it is integrally convex (see Theorem 8.31). Such a set is referred to as an
M2-convex set. An example of an M2-convex set that is not M-convex is given by
for some M-convex set B C Z^uv. It turns out that an M^-convex set Q can be
characterized by an exchange axiom:
(B^-EXCfZ]) For x,y £ Q and u e supp+(x - y),
(i) x - Xu e Q and y + Xu € Q, or
(ii) there exists v € supp~(x - y) such that x - Xu + Xv £ Q and
y + Xu - Xv e Q.
It is required in (B^-EXCfZ]) that at least one of (i) and (ii) be satisfied, depending
on a given triple (x, y,u). Examples of M^-convex sets are given in Fig. 4.1.
Whereas M^-convex sets are conceptually equivalent to M-convex sets, the
class of M^-convex sets is strictly larger than that of M-convex sets. This follows
from the implication (B-EXC[Z]) => (B^-EXCfZ]), as well as from the example of an
integer interval [a, b]z that is not M-convex but M^-convex. We denote by .Mg[Z]
the set of M^-convex sets.
The projection of a base polyhedron is known to coincide with what is called a
generalized polymatroid (or g-polymatroid for short); see Theorem 3.58 of Fujishige
[65]. Hence, an M^-convex set is precisely the set of integer points of an integral
g-polymatroid and the convex hull of an M^-convex set (Afi -convex polyhedron) is
represented as
for the pair (p, JJL) of integer-valued submodular function p € <S[Z] and supermodular
function n (i.e., —// 6 <S[Z]) such that
where Q(/o, /u) is defined in (4.36), and the following exchange axiom:
(B^-EXCfRj) For x,y e Q and u e supp+(x - y), there exist v e
supp^(a: — y) U {0} and a positive number a0 e R-++ such that x -
a(Xu - Xv) e Q and y + a(xu - Xv) 6 Q for all a e [0, O<O]R.
We denote by A^glR] and .MQ[Z|R] the sets of M^-convex polyhedra and integral
M^-convex polyhedra, respectively.
4.8. M-Convex Polyhedra 119
where we may assume that A is transitive; i.e., (u, v) & A, (v, w) € A => (u, w) e A.
See Theorem 3.26 of Fujishige [65] for the extreme rays of an M-convex cone.
An M-convex polyhedron is characterized as a polyhedron such that the tan-
gent cone at each point is an M-convex cone (by (a) •& (b) in Theorem 6.63).
Combining this with (4.42) yields a characterization of an M-convex polyhedron in
terms of the direction of edges:
B is an M-convex polyhedron
Similarly,
Q is an M^-convex polyhedron
where xo — fl-
it is noted again that M-convex polyhedra and M^-convex polyhedra are syn-
onyms of base polyhedra and g-polymatroids.
Bibliographical Notes
As remarked already, this chapter is a reorganization of known results in the the-
ory of submodular functions; see Fujishige [65] and Schrijver [183]. The proof of
Theorem 4.12 (hole-free property) is taken from Murota [141]. The equivalence
of exchangeability and submodularity (Theorem 4.15) is well known, but neither
precise statement nor proof can be found in the literature; Theorem 4.15 and the
proof are taken from [141]. The name "Lovasz extension" was introduced by Fu-
jishige [63], [65]. Theorem 4.16 (submodularity vs. convexity) is by Lovasz [123]
and Theorem 4.17 (discrete separation theorem) by Frank [55]. The intersection
theorem (Theorem 4.18) and the related statements (Theorem 4.22 (convexity in
intersection) and Theorem 4.23 (convexity in Minkowski sum)) are due to Edmonds
[44]. Theorem 4.21 (separation of M-convex sets) is observed in Murota [140]. In-
tegral convexity of M-convex sets (Theorem 4.24) is due to Murota-Shioura [153].
The example in Note 4.25 is an adaptation of Example 3.7 of [153]. The ter-
minology of M^-convex sets as well as that of the exchange axiom (B^-EXCfZ])
is introduced by Murota-Shioura [151]. The concept of g-polymatroids is due to
Frank [57] (see also Frank-Tardos [58]), whereas the characterization as the pro-
jection of base polyhedra is by Fujishige [64]. Proofs of (4.38) and (4.39) can be
120 Chapter 4. M-Convex Sets and Submodular Set Functions
L-convex sets form another class of well-behaved discrete convex sets. They are
defined in terms of an abstract axiom and correspond one-to-one to integer-valued
distance functions satisfying the triangle inequality. L-convex sets (or their convex
hull) are, in fact, a familiar object in the theory of network flows, though the ter-
minology of L-convexity is not used there. Emphasis here is placed on a systematic
presentation of various properties of L-convex sets from the viewpoint of discrete
convex analysis.
5.1 Definition
A nonempty set of integer points D C Zv is denned to be an L-convex set if it
satisfies the following two conditions:
(SBS[Z]) p,qeD => PVq,pKqGD,
(TRS[Z]) peD=>p±l&D.
We denote by CQ [Z] the set of L-convex sets.
L-convexity thus defined for a set D C Zv is equivalent to the L-convexity of
the indicator function SD : Zv —> {0, +00} of D, defined in (3.51). Namely, D is
an L-convex set, satisfying (SBS[Zj) and (TRS[Z]), if and only if SD is an L-convex
function, satisfying (SBF[Z]) and (TRF[Zj) introduced in section 1.4.1.
Since an L-convex set is homogeneous in the direction of 1 by (TRS[Z]), we
may consider the restriction of an L-convex set to the coordinate plane defined
by P(VO) = 0 for an arbitrary VQ € V. A set derived from an L-convex set by
such a restriction (intersection with a coordinate plane) is called an L^-convex set.
Whereas L''-convex sets are conceptually equivalent to L-convex sets, the class of
L^-convex sets is strictly larger than that of L-convex sets. The simplest example
of an L''-convex set that is not L-convex is an integer interval [a, b]z-
We focus on L-convex sets in the development of the theory and deal with
L''-convex sets in section 5.5.
121
122 Chapter 5. L-Convex Sets and Distance Functions
where ~f(u,v) represents the length of arc (u,v). We denote by 7(11, v) the shortest
length of a path from u to v in G7. The function 7 is well defined if there exists
no negative cycle in G7, where a negative cycle means a directed cycle of negative
length.
The triangle inequality
and~D(-y) =0(7).
(3) For 7 6 ^"[R.], D(7) is nonempty and
Here A = (Xuv \ (u, v) G A*/) and p are the variables of (P) and (D), respectively. The
coefficient matrix is totally unimodular, being the negative of the incidence matrix
of graph G7 (see Example 3.11). The set of feasible solutions to (D) coincides with
D(7).
(1) If (D) is feasible, the sum of the inequalities (5.3) for arcs in a directed
cycle shows the nonnegativity of the cycle. Conversely, if no negative cycle exists,
we can define p(v) to be the shortest path length from a fixed starting vertex to v
to obtain a feasible solution p to (D) (with an obvious modification for vertices v
not reachable from the starting vertex).
(2) In the particular case of x = Xv0 ~Xu0 for distinct UQ, VQ G V, the objective
function of (D) is equal to (p,x) = P(VQ) — P(UQ), and hence the optimal value of
(D) equals the right-hand side of (5.5) for (u, v) = (UQ,VQ). By Theorem 3.13, the
optimal solution A to (P) can be chosen to be an integer vector, which is in fact a
{0, l}-vector. Such an optimal solution to (P) represents a shortest path from UQ to
i>o, and therefore, the optimal value of (P) is equal to J(UQ,VQ). By the feasibility
of (D), LP duality (Theorem 3.10 (2)) applies to show (5.5). By 7 > 7 we have
D(T) 2 0(7). For p G 0(7), adding the inequalities (5.3) for arcs in the shortest
path from UQ to VQ yields p(i>o) — P(WO) 5- T"(UO> ^0)1 which shows 0(7) C D(T").
(3) The condition 7 G T[R] implies the nonexistence of negative cycles and
7 = 7 in (5.5).
(4) By Theorem 3.13, the integrality of 0(7) follows from the total unimod-
ularity of the coefficient matrix. D
with distinct pi (1 < i < m). The representation (5.7) with TO = 1 means p 6 D
(we are done). When TO > 2, repeated modifications of (5.7) as
A sort of converse of the above proposition is true. Note that the triangle
inequality (5.2) is not assumed in the proposition below.
Proof. (TRS[Z]) is obvious. For (SBS[Zj) we can easily show that p(v) - p(u) <
j(u,v) and q(v) — q(u) < ^(u,v) imply (p V q)(v) — (p V q)(u) < j(u,v) and (p A
q)(v) -(pAg)(u) <^(u,v). D
Proof. (1), (2) It follows from 0(71) n 0(72) = D(7i 2 ) that DI n D2 = (D(li) H
Zv) n (D(72) n Zv) = (D(7i) n D(72)) n Zv = D( 7l2 ) n Zv. Since D(_7i2JJs an
integral polyhedron, we obtain DI n D2 = D(7i 2 ) = D(7i) n D(72) = DI n D2.
(3) This is by (2) and Proposition 5.1 (1).
(4) This follows from (2) and Proposition 5.4. D
Proof. T = £o[Z meets the condition (3.53) in Proposition 3.16 and has the
property (a) there by (5.9). D
The following discrete separation theorem holds for two L-convex sets.
Theorem 5.9 (Discrete separation for L-convex sets). Let DI and D2 (C Zv) be
L-convex sets. If they are disjoint (Di C\D2 = 0), there exists x* € { — 1,0,1} V such
that
and, therefore,
5.4. L-Convex Sets as Discrete Convex Sets 127
The content of Theorem 5.9 consists of two claims. The first, explicit in the
statement, is that the separating vector x* is so special that it is a {0, ±l}-vector.
The second, less conspicuous and more subtle, is the implication (5.9), convexity in
intersection, since otherwise the inequality (5.10) is impossible.
Finally, we show the integral convexity of an L-convex set by deriving an
expression of the convex hull of an L-convex set.
For a vector p e Rv, let a\ > a2 > • • • > am be the distinct values of the
nonzero components of vector a = p — \p\, and define
Proof. The expression (5.11) shows the inclusion D in (5.12). To show the converse,
take p e D and put p0 = \p\, a = p - \p\, and & = \j>\ + XUi(P) (i - 0 , 1 , . . . , m).
In the representation D = D(7) with an integer-valued distance function 7 (by
Theorem 5.5), we have
Since \a(v} — a(u)\ < 1, po(v) —po(u) e Z, and 7(1*, v) e Z, we havepo(^) ~Po(u) <
j(u,v) for any u,v, and furthermore po(v) — Po(u) + 1 < 7(1*, v) if a(i>) > a(u).
Hence follows
128 Chapter 5. L-Convex Sets and Distance Functions
Note 5.11. The Minkowski sum of two L-convex sets, to be called an L^-convex
set, is not necessarily L-convex, though it is integrally convex (see Theorem 8.42).
An example of an L2-convex set that is not L-convex is given by
for some L-convex set D C Zv. It turns out that an L''-convex set P can be
characterized by the property
(SBS"[Z]) p,q&P =» ( p - a l ) V g , p/\(q + al)eP (Va e Z+)
(see Note 5.12 for the proof). This condition for a = 0 agrees with (SBS[Zj).
Examples of L^-convex sets are given in Fig. 5.1.
Whereas L^-convex sets are conceptually equivalent to L-convex sets, the class
of L^-convex sets is strictly larger than that of L-convex sets. This follows from the
implication [(SBS[Z]) and (TRS[Z])] =>• (SBS^Z]), as well as from the example of
an integer interval [a, b]z that is not L-convex but L^-convex. We denote by jCg[Z]
the set of L^-convex sets.
For a set P C Zv, (SBS^Zj) above is equivalent to either of the following
conditions:
5.5. iJ-Convex Sets 129
(see Note 5.12 for the proof). The property (5.15) is called discrete midpoint con-
vexity (see Fig. 5.2), where \z\ and \z\ denote, respectively, the integer vectors
obtained from z G R^ by componentwise round-up and round-down to the nearest
integer. Thus, L''-convex sets can be characterized by one of the three equivalent
conditions (SBS^Z]), (5.14), and (5.15).
The convex hull of an L''-convex set (called an L^-convex polyhedron) can be
represented as
(as well as by 7(1;, v) = 0 (Vu € V)) satisfies the triangle inequality. A set P C Zv
satisfies (SBS^Zj) if and only if it can be represented as P = P(7,7,7) n Zv.
The Minkowski sum of two L^-convex sets is called an L^-convex set. An L^-
convex set is a restriction of an L2-convex set, and the class of L^-convex sets is
strictly larger than that of L2-convex sets.
130 Chapter 5. L-Convex Sets and Distance Functions
Note 5.12. We prove the equivalence among L''-convexity (as induced from L-
convexity by restriction), (SBS^Z]), (5.14), and (5.15) for P C Zv.
[lAconvexity <^> (SBS^[Z])]: By (5.13) and (TRS[Z]) we have
Assuming a = q0 — p0 > 0, put p' = p — PQ\ and q' = q - q0l. Then (p V q) — (po V
</o)l = (p' - al) V q1 and (p A q) — (p0 A qo)l = p' A (q1 + al). Hence, the above
condition is equivalent to (SBS^[Zj).
[(SBSll[Z])^>(5.14)]: For a = maxvev{p(v) - q(v)} - 1, we have a > 0, (p -
al) V q = q + xx, and p A (q + al) = p - xx- Then (SBS"[Z]) implies (5.14).
[(5.14)^(5.15)]: Put p" = [2±2] and q" = [2±2j, and define p',q' £ Zv by
Note that \p'(v)—q'(v)\ < 1 (Vu e V), supp+(p'-g') C supp + (p—q), andsupp~(p' —
q') C supp^(p - q). Repeated applications of (5.14) to (p,q) yield p',q' 6 P, and
an application of (5.14) to (p',qr) gives p",q" e P.
[(5.15)^(SBSll[Z])]: For p,q £ P, define a sequence (q(0\q(1\ ...) of integer
points as follows:
It follows that there exists some positive integer JV such that q^ = q(N>) for any
integer k > N. Because of (i)-(iii) such a q(N^ is equal to (p— 1) V (p/\q) and hence
we have (p — 1) V (p A q) € P. Replacing p with (p — 1) V (p A q) and repeating the
above argument, we also have (p — 2 • 1) V (p A q) e P. Repeating this argument
(or more rigorously by induction), we have (p — al) V (p A g) e P for a e Z+. In
particular, we have p A q £ P. By symmetry we also have (p V </) A (q + al) e P
for a € Z+ and, in particular, p V q £ P. Now, replacing p with p V Q in the above
argument from the beginning, we have (p — al) V q 6 P for a e Z+. By symmetry
we also have p A (g + al) G P for a 6 Z+. •
5.6. L-Convex Polyhedra 131
We also have
Bibliographical Notes
The concept of L-convex sets was introduced by Murota [140] and that of L^-convex
sets by Pujishige-Murota [68]. The polyhedron D(7) associated with a distance
132 Chapter 5. L-Convex Sets and Distance Functions
M-Convex Functions
M-convex functions form a class of well-behaved discrete convex functions. They are
defined in terms of an exchange axiom and are characterized as functions obtained
by piecing together M-convex sets in a consistent way or as collections of distance
functions with some consistency. Fundamental properties of M-convex functions
are established in this chapter, including the local optimality criterion for global
optimality, the proximity theorem for minimizers, integral convexity, and extensi-
bility to convex functions. Duality and conjugacy issues are treated in Chapter 8
and algorithms in Chapter 10.
133
134 Chapter 6. M-Convex Functions
where the maximum and the minimum over an empty set are — oo and +00, respec-
tively. We denote by M [Z —» R] the set of M-convex functions and by M [Z —> Z]
the set of integer-valued M-convex functions.
with the notation V' = V \ {WQ} and (XQ,X') & Z x Zv . A function derived from
an M-convex function by such a projection is called an M^-convex function.
More formally, let 0 denote a new element not in V, and put V — {0} U V. A
function / : Zv -> RU{+oo} is called M*-convex if the function / : Zv -> Ru{+oo}
defined by
Proof. The first half follows from the obvious implication (M-EXC[Z])=^(M1'-
EXC[Z]) and Theorem 6.2. The second half is from the equivalence of (M-EXC[Zj)
and (M^-EXCfZj) under the condition on the effective domain. D
where Mn and M\ denote, respectively, the sets of M-convex functions and M11-
convex functions in n variables, and the expression M\ ^ Mn+i means a correspon-
dence of their elements (functions) up to a translation of the effective domain along
a coordinate axis, where (6.4) gives the correspondence under the normalization of
r = 0.
By the equivalence between M-convex functions and M^-convex functions, all
theorems stated for M-convex functions can be rephrased for M^-convex functions,
and vice versa. In this book we primarily work with M-convex functions, making
explicit statements for M^-convex functions when appropriate.
On expressing y = x — \Ul - Xu2 + Xv^ + Xv2 with 111,1(2,^1,^2 6 V and {ui, 1(2} n
{^1,^2} = 0, we see that (M-EXC]OC[Z]) is written as
is nonempty, where B = dom/. Take a pair (x,y) G T> with minimum \\x — y\\i,
where ||o; — y||i > 4 by (M-EXCioc[Z]), and fix u* & supp + (x — y) appearing in the
definition of T>. For a fixed e > 0, define p : V —> R by
Claim 1:
(Proof of Claim 1) The equality (6.12) is obvious from the definition of p, and
(6.13) can be seen as follows. If x — Xu* +Xv G B, (6.13) follows from Afp(x; v, w*) =
0 (by (6.12)) and
(by the definition of u»). If x - Xu, + Xv £ B, (6.13) follows from the fact that
A f p ( y ; u f , v) = e or +00 depending on whether y + Xu, — Xv 6 B or not.
Claim 2: There exist UQ 6 supp+(x — y) and VQ e supp~~(x — y) such that
y + Xu0 ~ Xv0 e B, U* e supp+(a; - (y + Xu0 ~ XvJ), and
6.2. Local Exchange Axiom 137
Note the difference in the two axioms: "Vu, 3u" in (M-EXC[Zj) and "3u, 3u" in
(M-EXC W [Z]).
Proof. It suffices to show <= when dom/ / 0. (M-EXCW[Z]) for / implies (B-
EXCW[Z]) for dom/, and therefore, dom/ is an M-convex set by Theorem 4.3.
Then Theorem 6.4 establishes the claim. D
Note 6.6. The proof of Theorem 6.2 is completed here. It remains to show that
(M^-EXC[Z]) for / implies (M-EXC[Z]) for / denned by (6.4). First, dom/ is an
M-convex set, since (M^-EXCfZ]) for / implies (B^-EXCtZj) for dom/ and dom/
is the projection of dom/ (see section 4.7 and (4.35) in particular). By Theorem
6.4, (M-EXC[Zj) is equivalent to (M-EXCioc[Z]), and therefore, it suffices to show
(6.7), (6.8), and (6.9) for x, y such that x - y — Xm + Xu2 - Xvt - Xv2 with
{ui,U2,vi,V2} C V U {0} and {ui,u2} PI {vi,v2} = 0. Since (6.7) is obvious from
(6.5), it remains to show (6.8) for four cases:
138 Chapter 6. M-Convex Functions
where HI, u?, vi, v% are all distinct. We deal with (al) and (b4) below; the other
cases are left to the reader.
Case (al): We abbreviate z
(ceiazPifa)', for instance, We are to derive We are to derive.
This, however, is derived from (6.5) for (02), (10), and u = v\.
6.3 Examples
We have already seen M-convexity in network flows and in matroids (section 2.2,
section 2.4). In this section we see some other examples of M-convex functions, such
as linear functions, quadratic functions, and separable convex functions.
We start by recalling the following facts from Proposition 6.1.
Proposition 6.7.
(1) The effective domain of an M-convex function is an M-convex set.
(2) The effective domain of an M^-convex function is an M^-convex set.
6.3. Examples 139
The proof is given in Note 6.10. Note that cross-terms in (6.23) have a common
coefficient b.
The following proposition shows a necessary and sufficient condition for a
quadratic form to have M-convexity.
45
In this section, V = {1,..., n}, f denotes a real-valued function in integer variables, i.e., / :
Z —> Ru{+oo}, and x(i) is the ith component of an integer vector x = (x(i) i = 1,... ,n) 6 Z n .
n
46
As a special case of (6.25) and (6.26) with i — j and k = I, we have an +dkk ^ la-ik- Similarly,
we have an > an, as a special case of (6.27).
140 Chapter 6. M-Convex Functions
for some w\ > w^ > • • • > wn satisfies (6.25), and the associated quadratic form
f ( x ) = j x Ax defined on {x 6 Z™ | Y^t=i x(^) = r} f°r some r e Z is M-convex
by Proposition 6.8 (1) above. For n = 4, such a matrix looks like
with univariate discrete convex functions fi e C[Z —> R] (i = 1,... ,n) is M-convex
or M^-convex according to whether dom/ is M-convex or M^-convex; see (3.68) for
the notation C[Z —> R]. As special cases we see the following.
is M-convex.
Note 6.10. The M11-convexity of the quadratic function in (6.23) with (6.24) is
proved here. Let i e supp+(x — y ) . If x > y, we have
If x(X) > y(X) for all X e T containing u, we can take v = 0 to meet (6.38) and
(6.39). Otherwise, let XQ be the unique minimal element of T such that u G XQ
and X(XQ) < y(Xo). By the minimality of XQ and (6.37) we have
(i) 3v e XQ \ Urer(Xo) ^ such that x(v) < y(v) or
(ii) 3 X l e T(X0) such that x(Xl) < y ( X l ) .
In case (i), this v is valid for (6.38) and (6.39). In case (ii), from (6.37) follows
(i) 3v € Xi \ UysT(Xi) ^ sucn tnat x(v) < v(v} or
(ii) 3X2 e T(Xl) such that a;(X2) < y ( X 2 ) .
Repeating this argument we eventually arrive at case (i). •
where Oy\{/ means the zero vector in Zv\u. For a pair of functions /$ : Zv —>
R U {+00} (i — 1, 2), the integer infimal convolution is a function /iD z /2 : Z v —>
R U {±00} denned by
where the right-hand side means the discrete Minkowski sum (3.52). The projection
fu can be represented as
which says that fu coincides with the restriction to U of the integer infimal con-
volution of / with the indicator function 5fj of U — {x e Zy | x(v) = 0 (v 6 U)}.
We continue to use the notation f[—p] for p € Ry denned in (3.69) and /[Qifc] for
an integer interval [a, b] defined in (3.55).
M-convex functions admit the following operations.
Proof. (1), (2), (5), and (6) are obvious and (3) is a special case of (4).
144 Chapter 6. M-Convex Functions
(4) For x, y G dom/ C dom/ and u € supp + (x — y), use (M-EXC[Zj) for / to
obtain v € supp~(o; — y) satisfying (6.1) for /. Then
The operations in Theorem 6.13 are also valid for M^-convex functions. In
addition, the projection is allowed for M^-convex functions.
Proof. (2) This follows from (6.45) as well as the M^-versions of (6) and (8) of
Theorem 6.13. D
Note 6.16. The sum of two M-convex functions is not necessarily M-convex. For
example, recall the M-convex sets £?i and B% in Note 4.25 such that B\ D B% is
not M-convex. Their indicator functions are M-convex, but their sum, which is the
indicator function of B\ n #2 > is not M-convex. The sum of two M-convex functions
is studied under the name M^-convex function in section 8.3. A similar argument
applies to the sum of two M^-convex functions. •
Note 6.17. The proviso fu* > —oo in Theorem 6.13 (7) can be weakened to
fu*(xo) > —oo for some XQ. A similar weakening holds for / > —oo in Theorem
6.13 (8) and fu > -oo in Theorem 6.15 (2). •
Note 6.18. For a function / : Zv —> RU {+00} and a positive integer a, we define
a function /" : Zv -> R U {+00} by
6.5. Supermodularity 145
6.5 Supermodularity
M^-convex functions are supermodular on the integer lattice.
\\x — y\\i > 3, we may assume x V y,x A y e dom/ and also ^{x(u) — y(u) u e
supp+(x — y)}> 2, by symmetry. Take u e supp+(o; — y) and put x' = (x A y) + Xu
and y' = y + Xu- Since dom/ is an M^-convex set, it includes the integer interval
[x A y, x V y]z and, in particular, x',y' e dom/. By ||x' — y||i < ||x - y\\i — I and
\\x — y'\\i — \\x — y\\i — 1, the induction hypothesis yields
Example 6.20. The converse of Theorem 6.19 is not true. For instance, a function
/ : Z3 -> RU{+oo} denned by dom/ = {0,1}3 and /(!,!,!) = 2, /(I,1,0) =
/(1,0,1) = 1, /(0,0,0) = /(1,0,0) = /(0,1,0) = /(0,0,1) = /(0,1,1) = 0 is
supermodular and not M^-convex; (M^-EXC^]) fails for x = (0,1,1), y = (1,0,0),
and u = 2. •
Note 6.22. The supermodular inequality (6.48) is void for an M-convex function
/ because xVy,x/\y£ dom/ occurs only when x = y. For an M-convex function
/, the property corresponding to (6.49) is expressed as
where xo — 0) as usual.
Proof. It suffices to prove (1). The implication => is immediate from Theorem
6.13 (3) and Proposition 6.23. The converse <£= follows from Claims 1 and 2 below
by Theorem 6.4.
Claim 1: B — dom/ is an M-convex set.
(Proof of Claim 1) For x,y e B and u € supp+(x — y), take a sufficiently large
M > 0 and define p : V —> R by
148 Chapter 6. M-Convex Functions
Then f\p](x) > f\p](y), and by (M-SI[Zj) there exist w & supp+(x - y) and v 6
supp~(x-y) such that f\p](x)-f\p](x-xw+Xv) = f(x)~ f(x~Xw + Xv)+p(w)-
M > 0. This is possible only if w = it, which shows (B-EXC_[Z]) for B. Then B
is an M-convex set by Theorem 4.3.
Claim 2: / satisfies the local exchange axiom (M-EXCioc[Z]).
(Proof of Claim 2) Take x,y e B with ||x - y||i = 4 and put y = x —
Xui ~ Xu2 + Xvi + Xv2 with Ui,u2,vi,v2 e V and {wi,M 2 } n {vi,v2} = 0. In the
following we assume u\ ^ u2 and v\ ^ ^2 (the other cases can be treated similarly).
Consider a bipartite graph G = (V+, V~; E) with vertex bipartition V+ = {ui,u2},
V~ = {^1,^2} and arc set E — {(ui,Vj) A . f ( x ; v j , U i ) < +00 (i,j = 1,2)}. The
graph G has a perfect matching as a consequence of (B-EXC[Zj) for B. We think
of A f ( x ; V j , U i ) as the weight of arc (ui,Vj) and apply Proposition 3.14 to obtain
p : V —» R such that
and p(ui) + p(u2) — p(v\) — p(v2) is equal to the right-hand side of (6.11) (and
p(v) = 0 for v G V \ {ui,U2,vi,V2})- Failure of inequality (6.11) would imply
a contradiction to (M-SI[Zj). D
Proposition 6.25. For an M-convex function f & M.\L —> R] and x,y € dom/,
we have
where
6.7 Minimizers
Global optimality for an M-convex function is characterized by local optimality.
Proof. It suffices to prove 4= in (1), but this follows from Proposition 6.23. D
Example 6.27. The minimum spanning tree problem serves as a canonical exam-
ple to illustrate the M-optimality criterion. Let G = (U, E) be a graph with vertex
set U and arc set E. A set T of arcs is called a spanning tree if it forms a con-
nected subgraph that contains no circuit and covers all the vertices. The minimum
spanning tree problem is to find a spanning tree T that has the minimum weight
with respect to a given weight w : E —> R, where the weight of T is defined as
Seer w(e)- It is well known that a spanning tree T has the minimum weight if and
only if w(e) < w(e') for any e e T and e' G E \ T such that T - e + e' is a spanning
tree (see Fig. 6.2). This well-known optimality criterion is a special case of Theorem
6.26 (1) applied to an M-convex function / : ZE —> R U {+00} defined by
Note that, for a spanning tree T and arcs e € T and e' £ E \ T, we have /(XT) <
f(XT — Xe + Xe') if and only if (i) w(e) < w(e') or (ii) T — e + e' is not a spanning
tree. In this connection it is noted that T - e + e' is a spanning tree if and only if e
belongs to the unique circuit contained in T U {e'}, called the fundamental circuit
with respect to (T, e'). •
Theorem 6.26 above shows how to verify the optimality of a given point with
O(n 2 ) function evaluations. The next theorem suggests how to find a minimizer.
Stating that a given point can be easily separated from some minimizer, it serves
as a basis of the domain reduction algorithm for M-convex function minimization,
to be explained in section 10.1.3.
shows how the concept of M-convex functions can be denned from that of M-
convex sets.
Proof. It suffices to prove (1). The implication =>• is immediate from Theorem 6.13
(3) and Proposition 6.29. For 4=, it suffices, by Theorem 6.4, to show that B —
dom/ is an M-convex set and / satisfies the local exchange axiom (M-EXCioc[Z]).
Claim 1: B is an M-convex set.
(Proof of Claim 1) Put Bp = argmin/[— p] for each p. Then we have B =
|Jp Bp for the convex hulls of B and Bp. For x, y G B, there exists p such that y G Bp
and z = tx + ( l - % e £ ^ for some t > 0. It follows from (B-EXC+[R]) of ~B~P that,
for u G supp+ (x — y) = supp+ (z — y ) , there exists v G supp~(z — y) = supp~ (x — y)
such that y + a(xu - Xv) G Bp C B for all sufficiently small a > 0. This shows
(B-EXC+[R]) for B. Therefore, B is an M-convex set.
For (M-EXCioc[Z]), take x, y € B with ||o;-y||i = 4. Let / : Rv -> RU{+oo}
be the convex closure of /, where it is noted that / is not assumed to be convex
extensible. Let p G R^ be a subgradient of / at c = (x + y)/2 £ R y . We have
c e argmin/[—p] = Bp, where Bp is an integral M-convex polyhedron. Hence,
the intersection of Bp with the interval / = [a; A y,x V J/]R is an integral M-convex
polyhedron, in which c is contained. This means that c can be represented as a
convex combination of some integral vectors, say, zi,..., zm G (lr\Bp)r\Zv = If~\Bp:
Translating this into G, we see that for each vertex v^ there is an edge that covers
(is incident to) Vi and also there is another edge that avoids (is not incident to) DJ.
This condition implies the existence of a perfect matching in G.
152 Chapter 6. M-Convex Functions
Finally we derive (M-EXCioc[Z]) from Claim 2. We divide into two cases, (i) If
{{vi,Vz}, {i>3, ^4}} C EQ, both x and y appear among the z/c's, and hence x, y 6 Bp.
By (B-EXC[Z]) for Bp, we have x - Xvf + Xv, e Bp and y + \Vi - xVj e Bp for some
i € {1, 2} and j € {3,4}. Hence,
which shows (6.11) with equality, (ii) If {{^1,^2}, {^3, ^4}} £ •f'O, it follows from
Claim 2 that {{ui,i>j}, {^2, Vj}} C J50 for some i, j with {z, j} = {3,4}. Then
Hence,
Note 6.31. The boundedness assumption on dom/ in Theorem 6.30 is not re-
strictive substantially, since we know from Proposition 6.14 that / is M-convex if
and only if its restriction f[a,b] to every bounded integer interval [a, b] is M-convex
(as long as dom/[ a>b ] ^ 0). On the other hand, the boundedness assumption seems
inevitable. The function
A particular case of this inequality with q = p + axu for u € V and a > 0 yields
y(u) < x(u). Namely, we have
and hence
This shows that (6.61), (6.62), and (6.63) are satisfied in equalities. In particular,
we have y + \u — \v € argmin/[g], a contradiction to our choice of y. D
For a function / let / be the function given by (6.4). It is easy to see that /
satisfies (M-GS[Z]) if and only if / satisfies the following:
154 Chapter 6. M-Convex Functions
Proof. The implications => in (1) and (2) have been shown in Propositions 6.32
and 6.33. We give a proof of <= for (1) by using Theorem 6.30. It suffices to
show that B = argmin/ is an M-convex set, since (M-GS[Zj) for / implies this for
/[—p] for any p € R^. Since B = B fl Zv by the convex extensibility of / (see
Proposition 3.18), this is further reduced to showing that every edge of polyhedron
B is parallel to \u — \v for some u, v 6 V (see (4.43)). Let E be an edge of
B. By B = argmin/ we have E fl Zv = argmin/[p] for some p € R^. For
two distinct integer points x, y on E, neither supp + (x — y) nor supp~(x — y) is
empty by B C {z e Zv z(V) = r}. By (6.60) with u e supp+(x - y) and
sufficiently small a > 0 there exists y € argmin/[g] such that y(v) > x(v) (Vv ^ u),
where q = p + a\u. Note that x ^ y since f[q](x) > f[q](y) > f[q](y)- Since
47
It will be shown in Theorem 6.42 that M^-convex functions are convex extensible.
6.8. Gross Substitutes Property 155
Proof. We may assume p = 0. Suppose that (i) in (M^-SWCSfZ]) fails, and let
a* be the maximum value of a such that x € argmin/[ax u ]. By the M-optimality
criterion (Theorem 6.26 (2)), x £ argmin/[o;Xu] if and only if
Let w & (V (J {0}) \ {u} be such that a* = A/(x; w, w) and put y = x — Xu + Xw-
Then /[a*xu](:r) = f[a*Xu](y) as well as x € argmin/[o:*Xw]. Hence follows (ii) in
(M"-SWGS[Z]). D
Proof. The implication => was shown in Proposition 6.35. We give a proof of
•4= by using Theorem 6.30 (2). It suffices to show that B = argmin/ is an M^-
convex set, since (M^-SWGSfZj) for / implies this for /[—p] for any p e Hv. Since
B = B n Zv by the convex extensibility of / (see Proposition 3.18), this is further
reduced to showing that every edge of polyhedron B is parallel to Xu ~ Xv °r Xu
156 Chapter 6. M-Convex Functions
we have x & arg min f\p]. This means that x £ E and that x — x is a scalar multiple
of x — y. In particular supp + (x — y) = supp+(x — x) = {u}. If supp~(x — y) = 0,
then x - y is a scalar multiple of \u. Otherwise, a similar argument shows that
supp~(x — y) = {v} for some v and there exists y E Er\Zv such that y(v) = y(v) — 1
and y(w) > y(w) (Vtu 7^ w). Since x — x is a scalar multiple of y — y, we have
x(u) = x(i>) + /? and y(u) = y(w) + 1//3 for some /3 > 0. We must have /3 — 1 since
£(i>) and y(u) are integers. Therefore, x — y is a scalar multiple of Xu — Xv- D
Proof. It suffices to prove (1) by showing that, for any 7 e R with 7 > inf /, there
exists some x* € dom/ satisfying f(x*) < 7 and (6.67). Suppose that x* 6 dom/
minimizes \\x* — xa\\i among all vectors satisfying f(x*) < 7. In the following,
we fix v G V and prove xa(v) — x*(v) < (n — l)(a — 1). (The inequality x*(v) —
xa(v) < (n — l)(a — 1) can be shown similarly.) We may assume xa(v) > x*(v);
put k — xa(v) — x*(v).
Claim 1: There exist wi,W2,... ,Wk € ^\{^} and 3/0 (= xa), j/i, • • • , J/fc 6 dom/
such that
By the choice of x* we have f(x* + \v — Xwf) > /(#*)> and hence /(j/j) = /(j/»-i —
x« + x»«i) < /(y*-i)-
Claim 2: For any to € V\{w} with yk(w] > xa(w) and /i 6 [0, yk(w)— xa(w) —
we
l]z> have
(Proof of Claim 2) We prove this by induction on fj,. For /z e [0, yk(w)— xa(w)-
l ] z , put x' = xa - n(xv — Xw) and assume x' e dom/. Let j* (1 < j* < k) be the
largest index such that Wjf = w. Then j/j. (tu) = j/fe(iw) > x'(tw) and supp~(j/j. —
x') = {t;}. (M-EXC[Zj) implies /(*') + /(%.) > /(^'-X,+X^) + /(%, +X»-X»).
By Claim 1 we have /(j/j. + Xw ~ Xtu) > /(%',)• This establishes Claim 2.
Claim 2 and (6.66) imply
for any w; with /z^, = yk(w) — xa(w) > 0. Hence yk(w) — xa(w) < a — I for all
w 6 V \ {u}. Then we obtain
Example 6.38. The M-proximity theorem is illustrated for the univariate M11-
convex function in Fig. 6.1 in section 6.4, where a = 2. Obviously, xa = 0 is
the minimizer of fa satisfying (6.68) and x* = 1 is the minimizer of /. We have
xa - x*\ = 1 = n(a — 1), in agreement with (6.69). •
The minimizer cut theorem (Theorem 6.28) can be adapted to scaling. Theo-
rem 6.28, except for (3), is a special case of the following theorem with a = 1.
158 Chapter 6. M-Convex Functions
Proof. We prove (2), while (1) can be proved similarly. Put xa = x + a(xv — Xu)-
We may assume max{x*(t;) | x* 6 argmin/} < xa(v); otherwise we are done. Let
x* be an element of argmin/ with x*(v) maximum and k = xa(v) — x*(v) (> 1).
The rest of the proof is the same as the proof of Theorem 6.37 (from Claim 1 until
the end). D
The algorithmic use of the above theorems, M-proximity and minimizer cut
with scaling, is shown in sections 10.1.2 and 10.1.4, respectively.
Note 6.40. An £i-norm version of Theorem 6.37 (1), with (6.67) replaced with
can be obtained from a slight modification of the proof; see Murota-Tamura [162].
Note 6.41. The M-proximity theorem (Theorem 6.37) is closely related to the
result of Hochbaum [90]. See also Moriguchi-Shioura [134]. •
The first theorem shows that the convex extension of an M-convex function
can be constructed locally.
Proof. It suffices to prove (1). The implication => is due to Theorem 6.42 and
Proposition 6.29. The converse •<= can be established by Theorem 6.30 applied to
the restriction of / to every bounded integer interval. D
Theorem 6.44. For two Afl-convex functions /i, /2 e M^[Z —> R] and x e R y ,
there exists A = (Xy y e N ( x ) ) such that
160 Chapter 6. M-Convex Functions
Theorem 6.45. The convex extension f of an M-convex function f & M[Z —> R]
on the integer lattice is a polyhedral M-convex function, i.e., f G A^[R —> R],
provided that f is polyhedral.
We denote by C[Z|R —> R] and C[R —> R Z] the sets of univariate polyhedral
convex functions with integrality (6.75) and dual integrality (6.76), respectively.
Polyhedral M-convex functions with integrality (6.75) are referred to as inte-
gral polyhedral M-convex functions, the set of which is denoted by ./Vf [Z|R —> R].
Polyhedral M-convex functions with dual integrality (6.76) are referred to as dual-
integral polyhedral M-convex functions, the set of which is denoted by ./Vf [R —> R|Z].
By Theorems 6.45 and 6.43, an integral polyhedral M-convex function is noth-
ing but a polyhedral M-convex function that can be obtained as the convex extension
of an M-convex function on integer points. Therefore, we have
where the second expression means that there exists an injection from .A/f [Z|R —» R]
to M.[Z —> R], representing an embedding of ./W[Z|R —> R] into M[Z —> R].
The effective domain of a polyhedral M-convex function is an M-convex poly-
hedron lying on a hyperplane {x e R^ x(V) = r} for some r £ R. Hence,
polyhedral M^ -convex functions can be defined as the projection of polyhedral M-
convex functions, just as M^-convex functions on integer points are defined from
162 Chapter 6. M-Convex Functions
M-convex functions via (6.4). We denote by .M^pR- ~^ R-] the set of polyhedral
M^-convex functions and by Ai^fZIR —> R] the set of integral polyhedral M^-convex
functions. The relationship between M and M*1 is described by M.n C M^ ~ Mn+i,
where M.n and M.^ denote, respectively, the sets of polyhedral M-convex functions
and polyhedral M^-convex functions in n variables.
The following are the R-counterparts of (M^-EXC[Z]) and (M^-EXC^Z]):
(M^-EXCfRj) For x,y 6 doniR/ and u & supp + (z — y), there exist
v G supp~ (x — y) U {0} and a positive number «o 6 R++ such that
where
defined similarly to (6.40), (6.41), and (6.42). Note in Theorem 6.49 (2) below that
a scaling factor j3 is allowed, unlike in the discrete case (cf. Note 6.18).
Note 6.54. Here are two remarks on QQ m (M-EXC[Rj). First, for an integral
polyhedral M-convex function / e A4[Z|R —> R], we can take a0 = 1. Second,
if (M-EXC[R]) is true at all, we can take ao = \(x(u) — y(u))/\supp~(x — y)\
independently of /; see Murota-Shioura [152] for the proof. •
164 Chapter 6. M-Convex Functions
Note 6.55. The proviso fu* > —oo in Theorem 6.49 (7) can be weakened to
fu*(xo) > —oo for some XQ. The same can be said for / > —oo in Theorem 6.49
(8) and fu > -oo in Theorem 6.50 (2). •
by the following proposition. We introduce yet another notation, 0M [Z —> ZJ, for
the set of integer-valued functions belonging to oM. [Z —> R].
Proposition 6.56.
(1) 0M[Z\R ->• R] = oM[R -> R].
(2) The convex extension of a function in oM[Z —» R] belongs to o.M[R —> R].
Proof. (1) Take / € oAi[R —> R]. For any p £ R y , argmin/[—p] is a cone that is
an M-convex polyhedron (or empty) by Proposition 6.53. Hence, argmin/[—p] =
B(/?) for a {0, +oo}-valued submodular set function p; see section 4.8. This shows
the integrality of argmin/[-p], and therefore / 6 o.M[Z|R —> R].
(2) Take / e 0M[Z'—> R]. Since / is integrally convex and / is positively
homogeneous, / can be represented as the maximum of a finite number of linear
functions. Hence, / is polyhedral and / € .M[R —> R] by Theorem 6.45. D
More precisely, we have the following, where T[R] and T[Z] denote respectively the
sets of real-valued and integer-valued distance functions with the triangle inequality.
Proposition 6.57.
(1) For f e 0M[n ->• R], we have jf e T[R].
(2) For f 6 0M[Z -> Z], we have 7/ 6 T[Z].
Proof. For (1) we apply (M-EXC[R]) to x = \V3 ~Xv2,y = Xv2 - X«i. and u = vi,
where we can take a = 1 by Proposition 6.56 (1) and Note 6.54. This yields the
triangle inequality (5.2). For (2) we use (M-EXC[Zj) in a similar manner. D
6.12. Positively Homogeneous M-Convex Functions 165
which is called the extension of 7. Proposition 5.1 as well as its proof shows
Proposition 6.58.
(1) For 7 e T[R], we have 7 e 0-M[R ->• R].
(2) For 7 e T[Z], we have fa € 0.M[Z -> Z].
Proof. Expression (6.82) is a special case of the M-convex function (2.37) appearing
in network flow problems (section 2.2), where T — V, A = {a = (u, v) u, v 6 V; u ^
v}, and /„„(£) = 7(u, v)£ (for £ > 0) and +00 (for £ < 0). See also Note 2.19. D
whenever Y^u,v&v ^uv(Xv - Xu) = a; and \uv > 0 (w, v e V). This implies / <
^> o $(/). In the case of 0Ai = o-M[Z —» Z] and T = T[Z], the opposite inequality
166 Chapter 6. M-Convex Functions
Theorem 6.61.
(1) For f e M[R —> R] and x e doniR/, define •jfiX(u,v) = f ' ( x ; ~xu + Xv)
(u,v € V). Then
6.13. Directional Derivatives and Subgradients 167
Proof. (1) Proposition 6.60 shows f ' ( x ; - ) £ oM[H —> R], from which follows
7/,z e ^"[R] by Proposition 6.57. By the definition of a sub differential and Theorem
6.52 (M-optimality criterion) we see
We have D(7/,x) e £0[R] by (5.18) and /'(*;•) = 7/,*(') by (3.31), (3.33), and
(6.84). If / e M[R -» R|Z], d n f ( x ) is an integral polyhedron by (6.76) and
1f,x(u, v) =sup{p(v) - p(u) p 6 ^R/(X)} e Z.
(2) Applying (M-EXC[Z]) to x + Xv3 ~Xv2,x + Xv2 - Xvi, and u = vl shows
the triangle inequality (5.2), and hence jftX & 1~\R]. The rest of the proof is similar
to (1), where we use Theorem 6.26 (1) instead of Theorem 6.52 and Theorem 5.5
instead of (5.18). D
The following fact shows the consistency of (1) and (2) in Theorem 6.61.
Proof, (a) =>• (b) is by Proposition 6.60, (a) =>• (c) by Theorem 6.61, and (a) =/•
(d) by Proposition 6.53. The rest is proved later in Note 8.7. D
whenever x,y € dom.R/ and 0 < A < 1 and semistrictly quasi convex if
whenever x, y 6 domp,/, f(x) ^ f(y), and 0 < A < 1. See Fig. 6.3 for an illustration
of a (semistrictly) quasi-convex function.
6.14. Quasi M-Convex Functions 169
Quasi convexity is ordinal convexity in the sense that the definition involves no
addition of function values, but relies only on comparisons. In this connection note
that, if f ( x ) is convex and <p is a nondecreasing function representing a nonlinear
scaling, then y>(/(x)) is quasi convex.
Quasi-convex functions enjoy the following nice properties:
Due to these properties, quasi convexity also plays an important role in continuous
optimization (see, e.g., Avriel-Diewert-Schaible—Zang [5]).
The concept of quasi M-convexity is defined for a function / : Zv —> Ru{+oo}
as follows. Recall the exchange axiom for M-convex functions:
The sign patterns of &f(x;v,u) and A/(j/;u, v) compatible with (implied by) in-
equality (6.91) are as follows:
Here Q and x denote possible and impossible cases, respectively. Relaxing con-
dition (6.91) to compatible sign patterns leads to two versions of quasi M-convex
functions. We say that a function / : Zv —> R U {+00} with dom / / 0 is quasi
M-convex if it satisfies the following:
The following weaker variants of (QM) and (SSQM) turn out to be useful for
our subsequent discussion:
(QM W ) For distinct x,y G dom/, there exist u G supp + (x — y) and
v G supp~~ (x — y) satisfying
(SSQMW) For distinct x,y G dom/, there exist u & supp + (x — y) and
v G supp~ (x — y) satisfying
The property (QM W ) can be expressed in two alternative forms below. The
first (6.93) may be regarded as a variant of (6.89) with discreteness in direction,
and the second (6.94) is similar to (6.50) in section 6.6.
Proof. Obviously, (6.94) implies (QM W ) and (6.93). We prove (QM W ) => (6.94)
and (6.93) => (6.94) by induction on \\x — y\\\. Suppose x,y G dom/ and f ( x ) >
f(y)', we may assume \\x—y\\\ > 2. If (QM W ) is true, there exist some u & supp+(x —
y) and v G supp~(x — y) such that Af(x;v,u) < 0 or Af(y;u,v) < 0; in the latter
case the induction hypothesis for x and y' = y + xu — Xv yields A/(x; v', u') < 0 for
some u' G supp+(x — y') C supp + (x — y) and v' G supp~(x — y') C supp~(a; — y).
If (6.93) is true, there exist u G supp+(a; — y) and v G supp~(x — y) such that
A/(x;f,u) < 0 or f(y + Xu ~ Xv) £ f ( x } ' , m the latter case we have /(x) > f ( y ' )
for y' = y + Xu — Xv and the induction hypothesis yields A/(o;; v', u') < 0 for some
u' G supp + (x — y') C supp + (x — y) and v' G supp~(x — y') C supp~(x — y). D
6.14. Quasi M-Convex Functions 171
Theorem 6.68. For f : Zv —> R U {+00} the following implications hold true.
Proposition 6.69. A set B C Zv satisfies (Q-EXCW) if and only if, for distinct
x,y £ B, there exist u e supp + (x — y) and v £ supp~(x - y) with x — \u + Xv 6 B.
Example 6.71. Whereas we have the obvious implications (B-EXC[Z]) => (Q-
EXC) =>• (Q-EXCW), these conditions are not equivalent. For instance,
The weaker version (QM W ) of quasi M-convexity for functions can be charac-
terized by the corresponding quasi M-convexity (Q-EXCW) of level sets. For any
a & R U {+00}, the level set is defined as
Note that L(/, +00) = dom/ and L(/, ao) = argmin/ for ao = min/.
Proof, ["only if"]: Let x and y be distinct elements of L(/, a). By (QM W ), we have
A/(x; v,u) < 0 or A/(y; u, v) < 0 for some u € supp+ (x — y) and v € supp~ (x — y ) .
Then, x - Xu + Xv e L(/, a) or y + \u - Xv & L(/, a).
["if"]: For any distinct x,y e dom/ with /(x) > /(y), we have x — Xu + Xv &
L(f,f(x)) for some u € supp + (x — y) and v € supp~(a; — y) by (Q-EXCW) and
Proposition 6.69. Hence f ( x — Xu + Xv) < /0*0- ^
Proof. In the proof of "only if" of Theorem 6.72, replace L(/, a) with dom/. D
Proof. The "only if" part follows from Theorem 6.72. To prove the "if" part,
we first observe that Theorem 6.4 can be strengthened to a statement that (M-
EXC[Z]) and (M-EXCioc[Z]) are equivalent if dom/ satisfies (Q-EXCW). (This can
be shown by modifying the proof of Claim 2 in the proof of Theorem 6.4.) Note that
(Q-EXCW) holds for dom/ by Theorem 6.72 and Proposition 6.73. To show (M-
EXCi oc [Z]),takeo;,y e dom/with ||x-y||i = 4 and put y = x-Xm-Xu^+Xvi+Xvz
with Ui,U2,Vi,V2 £ V and {MI,u^} D {vi,v%} = 0- In the following we assume
MI ^ «2 and vi ^ V2 (the other cases can be treated similarly). Consider a bipartite
graph G = (V+,V~;E) with vertex bipartition V+ — {ui,u2}, V" = {^1,^2} and
arc set E = {(ui,Vj) A/(x;t; j ,u i ) < +00 (ij = 1,2)}.
Claim 1: G has a perfect matching (of size 2).
(Proof of Claim 1) It suffices to show that every vertex has an edge incident
to it. Take p e Rv such that P(UI) + p(u2) — p(v\) — P(VZ) = f(y) — f ( x ) and
P(VJ) >P(UI) -Af(x;Vj,ui) for j = 1,2 (and p(v) = 0 for v e V \ {ui,u2, Vi,v2}).
Then f\p\(x) = f\p}(y) < f\p\(x - Xui + Xvj) for j = 1,2. By (Q-EXCW) for
L
(f\P\J\p}(x)) we nave f\p](x - Xu2 + Xv^ < f\P\(x) < +00 for some j € {1,2}.
Hence there is an edge incident to U2, and similarly for other vertices.
Claim 2: Inequality (6.11) is satisfied.
6.14. Quasi M-Convex Functions 173
The property (SSQM^) can be expressed in two alternative forms below. The
first (6.96) may be regarded as a variant of (6.90) with discreteness in direction,
and the second (6.97) is identical to (6.50) in section 6.6.
Proof. The proof is much the same as that for Theorem 6.67.
Proof. It suffices to show <=. (1) is immediate from (6.94) in Theorem 6.67 and
(2) is from (6.97) in Theorem 6.75. D
The minimizer cut theorem (Theorem 6.28) for M-convex functions can be
generalized for quasi M-convex functions.
Proof. The proof of Theorem 6.28 is valid here when (M-EXC[Zj) is replaced with
(SSQM^). D
The proximity theorem for M-convex functions (Theorem 6.37) can be gener-
alized for quasi M-convex functions.
Proof. It suffices to show that, for any 7 e R with 7 > inf/, there exists some
x* e dom/ satisfying f(x*) < 7 and (6.67). Suppose that x* € dom/ minimizes
| \x* —xa ||i among all vectors satisfying f(x*) < 7. In the following, we fix v 6 V and
prove xa(v)—x*(v) < ( n — l ) ( a —1). (The inequality x*(v) -xa(v) < ( n — l ) ( a — 1)
can be shown similarly.) We may assume xa(v) > x*(v). Put
(Proof of Claim 2) We prove the claim by induction on p,. For JJL 6 [0, [iw — l]z,
put x' — xa — n(xv — Xw), and assume x' 6 dom/. Note that x' 6 S and x'(v) >
x*(v), and hence Claim 1 implies f ( x ' ) > f ( y ) . Since supp~(y-o:') = {v}, (SSQM^)
for y, x', and w; e supp + (y-x') implies that if A/(y;i>,u>) > 0 then A/(x'; w, v) < 0.
By Claim 1 we have A/(j/; v, w) > 0, from which Claim 2 follows.
Claim 2 and (6.66) imply
Theorem 6.79 (Quasi M-minimizer cut with scaling). Let f : Zv —> R U {+00}
be a function satisfying (SSQM^) with argmin/ ^ 0, and assume a e Z ++ and
n=\V\. Then (I) and (2) in Theorem 6.39 hold true.
Proof. We prove (2), while (1) can be proved similarly. Put xa = x + a(xv ~ Xu)-
We may assume max{o;*(u) | x* € argmin/} < xa(v); otherwise we are done. Let
x* be an element of arg min / with x* (v) maximum. The rest of the proof is the
same as the proof of Theorem 6.78 (from Claim 1 until the end). D
Bibliographical Notes
The concept of M-convex functions was introduced by Murota [137] and that of
M^-convex functions by Murota-Shioura [151]; Theorems 6.2 and 6.3 are due to
[151].
The local exchange axiom (Theorem 6.4) is given in Murota [137], and the
weak exchange axiom (Theorem 6.5) is explicit in Murota [147].
Quadratic functions of the form (6.23) are treated in Camerini-Conforti-
Naddef [22], and their M^-convexity is observed in Murota-Shioura [151]. Propo-
sition 6.8 (characterization of quadratic M-convex functions) is due to Murota-
Shioura [155]. Quadratic functions defined by symmetric matrices of the form
(6.29), or (6.30), are treated in Hochbaum-Shamir-Shanthikumar [92], and their
M-convexity is noted by A. Shioura. Quasi-separable convex functions (6.32) are
considered by [22], and their M^-convexity is pointed out in [151]. The M^-convexity
of laminar convex functions (6.34) and minimum-value functions (6.36) is due to
Danilov-Koshevoy-Murota [34], [35]. The names laminar convex functions and
minimum-value functions are coined in this book.
The basic operations in section 6.4 are listed in Murota [141], [144], [147].
Theorem 6.13 (8) (infimal convolution) is due to Murota [137], whereas Theorem
6.15 (2) (projection of M^-convex functions) is due to [141]. The scaling operation
for M-convex functions is considered by Moriguchi-Murota-Shioura [133].
176 Chapter 6. M-Convex Functions
L-Convex Functions
(SBF[Zj) is submodularity on the integer lattice and (TRF[Zj) linearity in the di-
rection of 1. Note that we have r G Z if g is integer valued. Also recall the
submodularity inequality
We denote by £[Z —•> R] the set of L-convex functions and by £[Z —> Z] the set of
integer-valued L-convex functions.
Since an L-convex function g is linear in the direction of 1, we may dispense
with this direction as far as we are concerned with its nonlinear behavior. Namely,
instead of the function g in n = \V\ variables, we may consider the restriction g' to
177
178 Chapter 7. L-Convex Functions
an arbitrarily chosen coordinate plane, say, P(UQ) = 0 for some UQ G V, where the
restriction g' is a function in n — 1 variables defined by
with the notation V = V \ {UQ} and (po,p') G Z x Zv>. A function derived from
an L-convex function by such a restriction is called an L''-convex function.
More formally, an L''-convex function is defined as follows. Let 0 denote a new
element not in V, and put V — {0} U V. A function g : Zv —> R U {+00} is called
is translated to a condition
Proof. This follows from Theorem 7.1 and the obvious implications (7.3) =>•
(SBF»[Z]) =>(SBF[Z]). D
where Cn and &n denote, respectively, the sets of L-convex functions and L^-convex
functions in n variables, and the expression £^ ~ Cn+i means a correspondence of
their elements (functions) up to the constant r in (TRF[Zj), where (7.2) gives the
correspondence under the normalization of r = 0.
By the equivalence between L-convex functions and L''-convex functions all
theorems stated for L-convex functions can be rephrased for L''-convex functions,
and vice versa. In this book we primarily work with L-convex functions, making
explicit statements for L''-convex functions when appropriate.
A set function p : 2V —>• RU {+00} can be identified with a function g : Zv —>
RU {+00} with domg C {0,1}V through
The following states that the submodularity of p is the same as the L^-convexity of g.
Proposition 7.4. Let p : 2V —> RU {+00} be a set function with dom/9 ^ 0 and
g : Zv —>• RU {+00} be the associated function in (7.5). Then we have
The proposition above shows that L^-convex functions effectively contain sub-
modular set functions as a subclass; i.e.,
where
This says that the sum of the function values at a pair of points (p, q) does not
increase when the pair is replaced with another pair (p — xx-,1 + Xx) of closer
points.
We have \p'(v) — q'(v)\ <1 (v 6 V), supp+(p' — q') C supp+(p — q), and supp (p'—
</') £ supp~(p - g). Starting with (p, Q) and applying (LI'-APR[Z]) repeatedly,
we obtain g(p) + g(q) > g(p') + g(q'). Applying (L"-APR[Z]) to (p',q') yields
9(P') + 9(l'} > 9(P") + 9(q"}- Hence follows g(p) + g(q) > g(p") + g(q").
[(7.7)=^>(SBFl'[Z])]: (SBF^Z]) for g is equivalent to the submodularity of g in
(7.2) (cf. proof of Theorem 7.1). Since domg is an L^-convex set by (7.7) and (5.15),
dome? is an L-convex set. By Proposition 7.5, the submodularity of g is equivalent
to the local submodularity of g and the latter holds if and only if
7.3 Examples
We have already seen L-convexity in network flows and in matroids (section 2.2,
section 2.4). In this section we see some other examples of L-convex functions, such
as linear functions, quadratic functions, and separable convex functions.
First we note the following facts.
Proposition 7.8.
(1) The effective domain of an L-convex function is an L-convex set.
(2) The effective domain of an L^-convex function is an L^-convex set.
Proof. (1) (SBF[Z]) and (TRF[Z]) for g imply (SBS[Zj) and (TRS[Z]) for D =
dom^. (2) Similarly, (SBF"[Z]) for g implies (SBS^Z]) for D = domg. D
182 Chapter 7. L-Convex Functions
In Example 2.1 (Poisson equation) and Example 2.2 (electrical network), we have
seen the matrices
48
In this section, V = {1,... ,n}, g denotes a real-valued function in integer variables, i.e., g :
n
Z —> R U {+00}, and p(i) is the ith component of an integer vector p = (p(i) i = 1 , . . . , n) 6 Z".
49
It is easy to verify discrete midpoint convexity (7.7) for g.
7.4. Basic Operations 183
is L-convex if dom g is an L-convex set. As special cases of (7.12) and (7.14) we see
the following.
Proposition 7.9. Let tfj € C[Z —> R] be a univariate discrete convex function.
(1) ip is L^-convex.
(2) TTie function g : Z2 -> R U {+00} de/med fry 0(p) = VOK 1 ) ~ P( 2 )) ««
L-confeo;.
Note in Theorem 7.10 (2) that we have a scaling factor (3, in contrast to the
similar statement (Theorem 6.13 (2)) for M-convex functions. Also note that g in
Theorem 7.10 (5) is the infimal convolution of g with a separable convex function.
Operations in Theorem 7.10 are also valid for L^-convex functions. In addi-
tion, restrictions are allowed for L''-convex functions. See (3.55) and (6.40) for the
definitions of the restrictions g[a,b] and gu-
Proof. (2), (3) It is easy to verify (SBF^Zj) for g[a<b] and gv. D
Note 7.12. The infimal convolution of two L-convex functions is not necessarily
L-convex, and similarly for the infimal convolution of two L^-convex functions. Such
functions are studied in section 8.3 under the names L2-convex functions and L2-
convex functions, respectively. •
7.5. Minimizers 185
Note 7.13. The proviso gu > —oo in Theorem 7.10 (4) can be weakened to
gu (po) > — oo for some po, and similarly for g > —oo in Theorem 7.10 (5). •
7.5 Minimizers
Global optimality for an L-convex function is characterized by local optimality.
Proof. It suffices to prove 4= in (1) and (2). We first consider (2). For any disjoint
Y, Z C V, condition (7.20) together with submodularity yields
which implies the optimality criterion (3.65) for integrally convex functions. Since
an L^-convex function is integrally convex (to be shown in Theorem 7.20), Theorem
3.21 establishes 4= in (2). Next, (1) follows from (2), since an L-convex function
is L''-convex (Theorem 7.3) and the right-hand side of (7.19) implies g(p - \Y) =
g(p + xv\Y)>g(p)- n
The well-known optimality criterion for a submodular set function is an im-
mediate corollary of (2) above.
Proof. Let g be the L''-convex function associated with p (see (7.5) and Proposition
7.4), and apply Theorem 7.14 (2) with p = xx- D
Although Theorem 7.14 affords a local criterion for global optimality of a point
p, a straightforward verification of (7.19) requires O(2 n ) function evaluations. The
verification can be done in polynomial time as follows. We consider a submodu-
lar set function pp defined by pp(Y) = g(p + XY) — g(p) and note that (7.19) is
equivalent to saying that pp achieves its minimum at Y = 0. This condition can be
verified in polynomial time by the submodular function minimization algorithms in
section 10.2.
The minimizers of an L-convex function form an L-convex set, a property that
is essential for a function to be L-convex.
186 Chapter 7. L-Convex Functions
Proof. Suppose D = arg min <? is nonempty. We have r = 0 in (TRF[Zj) and hence
D satisfies (TRS[Z]). For p,q 6 D, we have p V q,p A q e D by (SBF[Z]). This
shows (SBS[Z]) for D. D
Proof. It suffices to prove (1). The implication =>• is immediate from Theorem 7.10
(3) and Proposition 7.16. The converse is shown later in Note 7.47. D
then arg min 5^0 and there exists p* e arg min g with
then arg min <? ^ 0 and there exists p* e arg min g with
7.7. Convex Extension 187
Proof. (1) It suffices to show that, for any ft > inf g, there exists p* that satisfies
g(p*) < /3 and (7.22); note that there exist only a finite number of p* satisfying
(7.22). We may assume pa = 0. By (TRF[Z]) with r = 0 there exists p* such that
g(p*) < ft and p* > 0; let p* be minimal (with respect to order >) among such
vectors. We have p* (v) = 0 for some v G V and
(Proof of Claim 2) Put p — Y^i=i P-iXXi and q = ^xXj and suppose q 6 dom g.
Since V" \ Xj = argmax^y{q(v) - p ( v ) } and g(p + Xv\Xj)=g(p- XXj} > g(p) by
Claim 1, (Ll-APR[Z]) implies g(q) > g(q - \v\Xj) = g(q + Xxt).
It follows from Claim 2 and (7.21) that ^i < a for i = 1 , . . . , k, and hence
The algorithmic use of the above theorem is shown in sections 10.3.2 and 10.4.5.
If g is L-convex, the associated set function pg,p belongs to <S[R] (i.e., p9tp is sub-
modular, /?g,p(0) = 0, and pg,p(V) < +00).
188 Chapter 7. L-Convex Functions
Theorem 7.19. Let g € £[Z —> R] be an L-convex function and ~g be its convex
closure.
(1) For p 6 domg and q £ [0, I]R, we have
where p3tp denotes the Lovasz extension (4.6) of the associated set function pg>p in
(7.26), q\ > 92 > • • • > qm are the distinct values of the components of q, and
Proof. (2) For each p € Zv, let hp(q) denote the function in q e [0, I]R defined
by the right-hand side of (7.29). If p € domg, p9jp belongs to <S[R], and hence
hp is a polyhedral convex function by Theorem 4.16. With the representation
of a real vector s = p + q with p = [s\ and q = s — [s\ we define a function
h : Hv -* R U {+00} by h(s) = hp(q). We have
The above theorem implies the integral convexity of an L-convex function and
hence that of an L''-convex function.
Proof. The implication =>• follows from (SBF[Zj) and Theorem 7.20. The converse
can be shown as follows. By the integral convexity of g, the convex closure g
coincides with the local convex extension, and, by the submodularity of g, the latter
is obtained as the Lovasz extension (7.27) of pg<p in (7.26). Therefore, we have
from convex extensibility and midpoint convexity (1.3). From these follows the
discrete midpoint convexity (7.7) of g, which means L^-convexity by Theorem 7.7.
D
Note 7.22. Submodularity (SBF[Zj) alone does not guarantee convex extensibility.
For example, any function g : Z2 —> Ru{+oo} with doing = {p € Z2 | p ( l ) = p(2)}
is submodular. •
we define here the concept of L-convexity for polyhedral convex functions in general
and show that major properties of L-convex functions survive in this generalization.
A polyhedral convex function g : Rv —>.R U {+00} with doniRg ^ 0 is said
to be L-convex if it satisfies
If i =£ j, we have
and put <?fc = pk. We have pk < +00 and pk G <5[R] and therefore (SBF[Rj) for
5fe. Since p(X) = lim/c^oo pk(X) (VX C V), we have #(p) = lim^oo gk(p) for each
p G Hv. Hence follows (SBF[R]) for g. D
where the second expression means that there exists an injection from £[Z|R —> R]
to £[Z -> R], representing an embedding of £[Z|R -> R] into £[Z -» R].
The effective domain of a polyhedral L-convex function is an L-convex polyhe-
dron, which is homogeneous in direction 1. Hence, polyhedral L^-convex functions
can be defined as the restriction of polyhedral L-convex functions, just as L''-convex
functions on integer points are defined from L-convex functions via (7.2). We denote
by £^[R —> R] the set of polyhedral L''-convex functions and by £ ll [Z|R —> R] the
set of integral polyhedral L^-convex functions. The relationship between L and Lfi
is described by Cn C £„ ~ £ r a +ij where £ra and C\ denote, respectively, the sets of
polyhedral L-convex functions and polyhedral L''-convex functions in n variables.
The R-counterpart of (SBF^Z]) is the following:
The condition (7.32) below is stronger than (SBF^Rj) in that it requires the
inequality not only for nonnegative a but also for negative a.
Proposition 7.34. Let g € £[R —> R] be a polyhedral L-convex function. For any
x € R^; argming[—x] is an L-convex polyhedron if it is not empty.
Note 7.35. The proviso gu > -oo in Theorem 7.31 (4) can be weakened to
9U(po) > -°° for some po, and similarly for g > -oo in Theorem 7.31 (5). •
Note 7.36. For L^-convex functions on integer points we have seen characteriza-
tions in terms of discrete midpoint convexity and submodular integral convexity
(Theorems 7.7 and 7.21). These characterizations, however, do not carry over to
polyhedral L^-convex functions. In contrast, translation submodularity (SBF^Z])
is generalized to (SBF^R]), as stated in Theorem 7.28. •
by the following proposition. We introduce yet another notation, oC[L —» Z], for
the set of integer-valued functions belonging to o£[Z —> R].
Proposition 7.37.
(1) 0 £[Z|R->R] = 0 £[R->R].
(2) The convex extension of a function in o£[Z —> R] belongs to o£[R —» R].
Proof. (1) Take g e o£[R -* R]- For any x e Ry, argming[-z] is a cone that is an
L-convex polyhedron (or empty) by Proposition 7.34. Hence, argmingf—x] = D(7)
for a {0, +oo}-valued distance function 7; see section 5.6. This shows the integrality
of argmin<?[—x], and therefore g e 0 £[Z|R —> R].
(2) Take g e o£[Z —> R]. Since g is integrally convex and g is positively
homogeneous, g can be represented as the maximum of a finite number of linear
functions. Hence ~g is polyhedral, and <; e £[R —> R] by Theorem 7.26. D
More precisely, we have the following statements, where <S[R] and <S[Z] denote
respectively the sets of real-valued and integer-valued submodular set functions
defined in (4.10) and (4.11).
Proposition 7.38.
(1) For g e 0£[R -» R], we have pg <E <S[R].
(2) For 5 e 0£[Z -> Z], we ftave ps € <S[Z].
Proof. The submodularity of /9S is immediate from that of 3. Note also that
Pg(9) = 5(0) = 0 and pg(V) = g(l) < +00. D
where pi > £>2 > • • • > pm denote the distinct values of the components of p, and
Ui = Ui(p) = {v e V I p(v) > pi} for i = 1 , . . . , m. Denote by pz the restriction of
p to Zv, and note that we have pz : Zy —> Z U {+00} for integer-valued p.
7.9. Positively Homogeneous L-Convex Functions 195
Proposition 7.39.
(1) For p e 5[R], we have p &0£[R -> R]-
(2) For p £ S[Z], we have pz € 0C\Z -> Z].
Proof. (1) We have p € £[R —> R] by Proposition 7.25, whereas the positive
homogeneity of p is obvious. (2) follows easily from (1). D
for q G NO, which remains valid for all q & Ry since g is positively homogeneous
and the origin 0 is an interior point of -/Vo- Hence g — \P o 3>(g). D
Proof, (a) =J> (b) is immediate from Proposition 7.34. For (b) =>• (a), define p by
p(X) = g(xx) (X C V) and denote by p its Lovasz extension (7.36).
Claim 1: g(p) = p(p) for all p 6 Ry.
(Proof of Claim 1) The positively homogeneous convexity of g as well as (7.36)
yields p(p) > g(p). We may assume p e doniR.<7, since otherwise /3(p) — g(p) = +00.
Take x such that p € argming[—x , put Z) = argmin <?[-£], and let SD be its
indicator function. Since D is an L-convex cone, we have 60 € o£[R —> R]- A set
function p. defined by fi(X} — SD(XX) (X C V) belongs to <S[R] and its Lovasz
extension coincides with SD by Theorem 7.40. From this and (4.8) we see
196 Chapter 7. L-Convex Functions
In view of the expression (4.6) and the linearity of g on D we obtain g(p) = p(p).
By Claim 1 and the convexity of g, p is submodular by Theorem 4.16. In
addition we have p(0) = g(0) = 0 and p(V) = p(l) = <?(!) < +00. Therefore,
p G <S[R]. This implies p G 0£[R -> R] by Proposition 7.39 (1). D
Proposition 7.42. If g & £[R —> R] and p G domRg, then g'(p', •) G o£[R —> R].
Theorem 7.43.
(1) For g e £[R -» R] and p G dom Rff , define pg,p(X) = g'(p; Xx) (XCV).
Then
(2) For g G £[Z -> R] and p G domzg, define pg,p(X) = g(p + Xx) ~ g(p)
(XCV). Then
Proof. (1) Proposition 7.42 shows g'(p;-) e o£[R —> R], from which follows
Pgtp € <S[R], by Proposition 7.38. By Theorem 7.33 (1) (L-optimality criterion),
We have B(p9,p) € M0[H} by (4.39) and g'(p; •) = p 9 , p (-) by (3.31), (3.33), and
(4.14). If g e £[R -»• R|Z], dn,g(p) is an integral polyhedron by (6.76) and
pg,p(X) = sup{x(X) | x e dRg(p)} € Z.
(2) It is easy to see p9tp e <S[R]. The rest of the proof is similar to (1), where
we use Theorem 7.14 (1) instead of Theorem 7.33 (1) and Theorem 4.15 instead of
(4.39). D
Proof, (a) =^> (b) is by Proposition 7.42, (a) => (c) by Theorem 7.43, and (a) =>
(d) by Proposition 7.34.
(b) =>• (a): By (7.37), g is submodular and linear in the direction of 1 in a
neighborhood of each p e Hv. This implies (SBF[R]) and (TRF[R]) (see Proposi-
tion 7.23).
(b) •& (c): This follows from the relation (<5aRfl(p))" = g'(p; •) in (3.33) and the
one-to-one correspondence between .Mo[R] and o£[R —> R], which is a consequence
of (4.39) and Theorem 7.40.
(d) =>• (b): To use Proposition 7.41 let x € Rv be such that inf g'(p; -)[-x] >
—oo. Then inf g[—x] > —oo and argmingi[—x] E £o[R] by (d). By (5.18) we have
with Ap = {(u,v) | p(v) — p(u) = 7(u,u)}. This shows that argmin((/(p; •)[~:r]) is
an L-convex cone. Then (b) follows from Proposition 7.41. D
Note 7.47. We prove <£= of Theorem 7.17 (1). We have argmin#[-x] e £o[Z]
for every x € R^ by the assumption. Since an L-convex set is integrally con-
vex (Theorem 5.10), g is an integrally convex function by Theorem 3.29. By the
boundedness of dome/, the convex closure ~g of g is a polyhedral convex function
and argming[—x] = argmin<?[—x] 6 £o[Z|R]. This implies g e £[Z|R —> R] by
Theorem 7.46 and therefore g e £[Z -> R]. •
to sign patterns of g(p A q) — g(p) and g(p V q) — g(q) compatible with (implied by)
this inequality, which are given as follows:
(SSQSB) For any p.q e Zv, both (i) and (ii) hold:
Theorem 7.49. For g : "Zv —> RU {+oc}, the following implications hold true.
As is easily seen from the definitions of quasi L-convexity, most of the prop-
erties of quasi-submodular functions can be restated naturally in terms of quasi
L-convex functions, and vice versa. We will work mainly with quasi-submodular
functions.
The following are quasi versions of Theorem 7.2 for L-convex functions.
200 Chapter 7. L-Convex Functions
Proposition 7.50. Assume that g : Zv —> RU {+00} satisfies g(p) = g(p + 1) for
allpeZv.
(1) For g satisfying (QSBW) and for p, q € Zv and a & Z, we have
(2) For g satisfying (SSQSBW) andforp,q 6 Z17 with g(p) ^ (/(g) and a € Z,
we /ia«e inequality (7.39) wzi/i sinci inequality. In particular, for p, q G domg with
g(p) 7^ <?(<?) ana7 a e [0,ai - a 2 ]z, we have (7.40) wz£/i sihci inequality.
(3) For 5 satisfying (SSQSB) andforp,q e Zv and a e Z, we ftawe
in which g(p/\(q — al)) = g((p/\(q — al)) + al) = ^((p + al) Ag). Inequality (7.40)
is obvious from (7.39) since pV{o— (c<i — ce)l} = p + axx and (p+ (ai — a)l) Agr =
9 — c^Xx for a e [0,a\ — a2\z- The proofs of (2) and (3) are similar.
Proof. For the "if" part, take p,q & doing and put a = max{g(p),g (</)}. Since
p,q G L(g,a), we have p A q & L(g,a) or p V q G L(g, a); i.e., max{g(p),g(<?)} >
min{g(p A q),g(p V 5)}. The "only if" part is even easier. D
Proof. The "only if" part follows from Theorem 7.51 and the submodularity of g[x\.
For the proof of the "if" part, takep,q G domg. By (QDL) for L(g, max{g(p),g (<?)})
we have p A g G dom g or p V q G dom g. We consider the former case, where we
may assume p A q ^ p,q. For any e > 0, we can choose some x G Ry such that
g[x\(p) = g[x](q) = g[x}(p A q) - e. (QDL) for L(g[x],o) with a = g(x\(p) shows
pVq G L(g[x],a), which implies g[x](p) + g[x](q) = 1a> g[x](p/\q)+g[x](pVq)-e.
Since e > 0 is arbitrary, this means (SBF[Zj). D
Proof. We prove •<= of (1) by contradiction. Suppose that g(q) < g(p) for some
q G domg such that q — p is not a multiple of 1. We may assume that q > p
by (TRF[Zj) with r = 0 and that q minimizes maxt,ey{q'(t;) — p(v)} among such
vectors. Put X = argmax.vev{q(v) - p ( v ) } , where X ^ V. By Proposition 7.50
(1), we obtain
whereas g(p) < g(q — Xx) by the choice of q. Hence follows g(p) > g(p + xx), a
contradiction to the strict local minimality of p. The other direction => of (1) is
obvious, and (2) can be shown similarly by Proposition 7.50 (2). D
The proximity theorem for L-convex functions (Theorem 7.18) can be gener-
alized for quasi L-convex functions.
Proof. The proof of Theorem 7.18 works with (7.43) and (7.44) in place of
(L*-APR[Z]).
Bibliographical Notes
The concept of L-convex functions was introduced by Murota [140]. L^-convex
functions are defined by Pujishige-Murota [68] as a variant of L-convex functions,
together with the observation that they coincide with the submodular integrally
convex functions considered earlier by Favati-Tardella [49]. Theorems 7.1 and 7.3
are due to [68], and Theorem 7.2 is stated in Murota [147].
Discrete midpoint convexity is considered by Favati-Tardella [49] with an ob-
servation of its equivalence to submodular integral convexity. The equivalence of
discrete midpoint convexity to translation submodularity (SBF^Zj) in Theorem 7.7
is by Fujishige-Murota [68], whereas that to (L^-APRfZ]) is noted in Murota [147].
Condition (7.11) for quadratic L-convex functions is given in Murota [141].
Separable convex functions with chain conditions (7.13) are considered in Best-
Chakravarti-Ubhaya [11]. Multimodular functions are treated in Hajek [85].
The basic operations in section 7.4 are listed in Murota [141], [144], [147].
The theorems on minimizers of L-convex functions are of fundamental impor-
tance. Theorem 7.14 (L-optimality criterion) is stated in Murota [145]. Theorem
7.15 (optimality for submodular set functions) can be found as Theorem 7.2 in Fu-
jishige [65]. Theorem 7.17 (characterization by minimizers) is a corollary of Theorem
7.45 due to Murota-Shioura [152]. A thorough study of minimizers of submodular
functions is made in Topkis [202].
The L-proximity theorem (Theorem 7.18) is due to Iwata-Shigeno [105]. The
present proof based on (L^-APRfZ]) is by Murota-Shioura [154].
The construction of the convex extension of an L-convex function by means
of the Lovasz extensions (Theorem 7.19) is due to Murota [140]. The same idea,
however, was used earlier by Favati-Tardella [49] for submodular integrally con-
vex functions. The equivalence of L''-convexity to submodular integral convexity
(Theorem 7.21) is due to Fujishige-Murota [68].
Polyhedral L-convex functions are investigated by Murota-Shioura [152], to
which all the theorems in section 7.8 (Theorems 7.26, 7.28, 7.29, 7.30, 7.31, 7.32,
and 7.33) as well as Proposition 7.34 are ascribed. L-convexity for nonpolyhedral
convex functions is considered in Murota-Shioura [156], [157].
The correspondence between positively homogeneous L-convex functions and
submodular set functions (Theorem 7.40) is established for the case of Z in Murota
[140] and generalized to the case of R in Murota-Shioura [152]. Proposition 7.37 is
stated in Murota [147].
Theorem 7.43 for directional derivatives and subgradients is shown for the case
of Z in Murota [140], [141], and generalized to the case of R in Murota-Shioura [152].
Theorem 7.45 (characterizations in terms of directional derivatives, subdifferentials,
7.11. Quasi L-Convex Functions 203
and minimizers) is by [152], whereas its ramification with integrality (Theorem 7.46)
is stated in Murota [147].
The concept of quasi L-convex functions was introduced by Murota-Shioura
[154] on the basis of the idea of Milgrom-Shannon [129]. Theorem 7.52 is due to
[129], and the other theorems in section 7.11 (Theorems 7.49, 7.51, 7.53, and 7.54)
are in [154].
This page intentionally left blank
Chapter 8
By addressing the issues of conjugacy and duality, this chapter provides the the-
oretical climax of discrete convex analysis. Whereas conjugacy in convex analysis
gives a symmetric one-to-one correspondence within a single class of closed convex
functions, conjugacy in discrete convex analysis establishes a one-to-one correspon-
dence between two different classes of discrete functions with different combinatorial
properties distinguished by "L" and "M." The conjugacy between L-convexity and
M-convexity is thus one of the most remarkable features of discrete convex analysis.
Discrete duality is another distinguishing feature. It is expressed in a number of the-
orems, such as the separation theorems for M-convex/M-concave functions and for
L-convex/L-concave functions (M- and L-separation theorems) and the Fenchel-type
duality theorem. Besides formal parallelism with convex analysis, these discrete du-
ality theorems carry deep combinatorial facts, implying, for example, Edmonds's
intersection theorem and Prank's discrete separation theorem for submodular set
functions as special cases.
8.1 Conjugacy
M-convex functions and L-convex functions form two distinct classes of discrete
functions that are conjugate to each other under the Legendre-Fenchel transfor-
mation. This stands in sharp contrast with conjugacy in convex analysis, which
is a symmetric one-to-one correspondence within a single class of closed convex
functions. The conjugacy correspondence between M-convexity and L-convexity is
in fact a translation of two different combinatorial properties, exchangeability and
submodularity, on top of convexity. The relationship between submodularity and
supermodularity with respect to conjugacy is discussed first in section 8.1.1. Con-
jugacy for polyhedral M-/L-convex functions is established in section 8.1.2 and that
for integer-valued M-/L-convex functions on integer points in section 8.1.3.
205
206 Chapter 8. Conjugacy and Duality
and supermodular if
Also recall from (3.26) that the Legendre-Fenchel transform /* : Ry -> RU {+00}
is denned by
From this inequality, Submodularity (8.1), and the definition (8.3), we see that
for p = (p(l),p( 2 )) e R2 and q = (<?(!), <?(2)) 6 R2 with p(l) > q(l) and p(2) >
q(2). We claim that
and, therefore,
and, therefore,
To prove this by contradiction, suppose that for every v £ supp (x — y) there exist
pv £ Dx and qv £ Dy with pv(v) > qv(v). Then, for
we have p* £ Dx, q* £ Dy, and p*(v) > q*(v) (Vu e supp (x - y ) ) . By denning
from Theorem 7.29. By supp (x — y) C supp+(p* — ?»), on the other hand, we see
where the last equality is due to x(V) = y(V) — r (the constant in (TRF[R])).
Combining (8.7) and (8.8) results in
8.1. Conjugacy 209
Proof. (1) and (2) are equivalent, so we prove (1). We first note that /** = / for
any polyhedral convex function /.
For / G M we have $R/*(P) = argmin/[—p] G .Mo[R] (Vp £ dom/*) by
Proposition 6.53. Then /* G £ by (c)=^(a) in Theorem 7.45. (An alternative proof
is described in the proof of Theorem 8.6.)
Conversely, take g € C, and x,y G domg*. Since inf<7[—x] > —oo and
inf g[—y] > —oo, Proposition 8.3 shows that for every u € supp+(x — y) there
exists v G supp-(z - y) satisfying (8.6). Noting that (£ a r g min g [-x])* = (g')'(x', •)>
which follows from (3.30) and (3.33), we obtain
Theorem 8.4 (2) states that L^-convex functions and M^-convex functions are
transformed to each other, where L^-convex functions are submodular (Theorem
7.28) and M^-convex functions are supermodular (Theorem 6.51). It is noted that
Theorem 8.4 (2) does not imply, nor is it implied by, Theorem 8.1, which shows the
supermodularity of the conjugate of a submodular function.
Recalling the basic fact that the conjugate of the indicator function of a con-
vex set is a positively homogeneous convex function, and vice versa, we see from
Theorem 8.4 above that M-convex polyhedra ./Vfo[R] and positively homogeneous
L-convex functions o£[R —»• R] are conjugate to each other and also that L-convex
polyhedra £o[R] and positively homogeneous M-convex functions o.M[R —> R] are
conjugate to each other. On the other hand, we can identify positively homogeneous
L-convex functions o£[R —* R] with submodular set functions <S[R] (Theorem 7.40)
and positively homogeneous M-convex functions oA^[R —> R] with distance func-
tions with the triangle inequality T[R] (Theorem 6.59). We can summarize these
210 Chapter 8. Conjugacy and Duality
In addition, the polarity between M-convex cones and L-convex cones follows
from (8.9). This is because two convex cones are polar to each other if and only if
their indicator functions are conjugate to each other. Thus we obtain the following
theorem.
Theorem 8.5. A polyhedral cone is M-convex if and only if its polar cone is L-
convex. Hence, the classes of M-convex cones and L-convex cones are in one-to-one
correspondence under polarity (3.34).
where ,M[Z|R —> R] and £[Z|R —> R] denote the sets of integral polyhedral M-
convex and L-convex functions, respectively; .M[R —> R|Z] and £[R —> R|Z] denote
the sets of dual-integral polyhedral M-convex and L-convex functions, respectively;
and o.M[R —> R|Z] and o£[R —> R|Z] are their subclasses with positive homogene-
ity.52
It is known that the conjugacy relationship between M-convexity and L-
convexity holds more generally for closed proper convex functions. Recall that
the Legendre-Fenchel transformation gives a symmetric one-to-one correspondence
in the class of all closed proper convex functions (Theorem 3.2).
Proof. The proof of the "if" part is essentially the same as the latter half of
the proof of Theorem 8.4. The "only if" part needs a new approach, since the
implication (c)=^(a) in Theorem 7.45 does not carry over to nonpolyhedral convex
functions. Suppose that / satisfies (M-EXC[R]) and put g = /*. It is easy to show
(TRF[Rj) for g. The proof of the submodularity (SBF[Rj) for g consists of the
following steps.
52
The notation for dual integrality extends naturally to other classes of functions. For example,
.M^R -+ R|Z] and /^[R —> R|Z] denote the sets of dual-integral polyhedral M^-convex and
L^-convex functions, respectively.
8.1. Conjugacy 211
Note 8.7. With Theorem 8.4 we complete the proof of Theorem 6.63 (characteriza-
tions of polyhedral M-convex functions), (b) ^ (c) follows from <59R/(X)* = f'(x; •)
in (3.33) and the correspondence between £o[R] and o-M[R —> R], which is a
special case of Theorem 8.4. To show (a) •&• (c) •& (d), put g = /* and note
that d R f ( x ) = argmin<?[—x] and argmin/[—p] = dRg(p). By Theorem 8.4 and
Theorem 7.45 (characterizations of polyhedral L-convex functions), we see that
/ e M[n -*• R] & g 6 £[R -> R] <^ argming[-z] 6 C0[R] & ^(p) & M0[R}.
Note 8.8. We complete the proof of Theorem 6.45 using Theorem 6.63 ((d)
(a)) established in Note 8.7. Let / be the convex extension of / 6 M[Z —> R]. For
any p € Ry, argmin/[—p] is an M-convex polyhedron if it is not empty (Theorem
6.43). Since / is polyhedral by the assumption, Theorem 6.63 ((d) =>• (a)) sho
/e.M[R^R].
Note 8.9. Using Theorem 8.5, we complete the proof of (4.42), the representation
of an M-convex cone in terms of vectors Xu ~ Xv (u>v £ V). Let B be an M-
convex cone and D be the polar of B. By Theorem 8.5, D is an L-convex cone,
and by (5.18) it can be represented as D — {p G R^ (p,a») < 0 (i = 1,... ,m)}
for some a^ = \Ui — \Vi (i = 1,... ,m). Since B is the polar of D, this implies
B = {x 6 Rv | x is a nonnegative combination of a* (i = 1 , . . . , m)} by (3.36).
Conversely, a convex cone of this form is M-convex, as can be shown by reversing
the above argument.
Note 8.10. Using Theorem 8.5, we complete the proof of (5.21), the representation
of an L-convex cone in terms of a ring family T> C 2V. Let D be an L-convex cone
and B be the polar of D. By Theorem 8.5, B is an M-convex cone, and, by (4.39),
it can be represented as B = B(p) using p € <S[R] with p : 2V —> {0, +00}; i.e.,
with T> = dom/5, which is a ring family. Since D is the polar of B, this implies
212 Chapter 8. Conjugacy and Duality
We call (8.11) and (8.12), respectively, convex and concave discrete Legendre-
Fenchel transformations. The functions /* : R^ —* R U {±00} and h° : R^ —>
R U {±00} are called the convex conjugate of / and the concave conjugate of h,
respectively. Note that h°(p) = — (—h)'(—p).
For an integer-valued function /, f'(p) is integral for an integer vector p.
Hence, (8.11) with p 6 Zv defines a transformation of / : Zv —> Z U {+00} to
/* : Zv -> Z U {±00}; we refer to (8.11) with p e Zv as (8.11)z. We call (/*)*
using (8.11)z the integer biconjugate of / and denote it by /**. Similarly, (8.12)
with p e Zv is designated by (8.12)z and we define h°° = (h°)°.
The following fact is fundamental for the conjugacy of discrete functions.
The conjugacy theorem for discrete M-convex and L-convex functions reads
as follows.
Proof. The basic idea of the proof is to apply Theorem 8.4 to the convex exten-
sions of / and g with additional arguments for discreteness. Since (1) and (2) are
equivalent, we deal with (2).
8.1. Conjugacy 213
(i) Take / e M.^[Z —> Z]. Let / be the convex extension of / and / be the
conjugate of / in the sense of (8.3). We have f ' ( p ) = f (p) for p e Zv.
If doniz/ is bounded, / is polyhedral convex, and therefore / e .M^R —> R]
by Theorem 6.45. Then Theorem 8.4 shows /* € &[H -> R]. Since /*(p) = /*(p)
for p € Zv, (SBF^R]) for /* implies (SBF^Z]) for /*. Hence /• e C\Z -> Z].
If domz/ is unbounded, we consider the restriction /& of / to integer interval
[—fcl,fcl]z for A; E Z large enough to ensure domz/ fl [—fcl,fcl]z ^ 0- Then we
have fk & M^Z -> Z} and /£ e -C^Z -» Z] by the argument above. For each
p € domz/*, there exists kp such that /*(p) — /*(p) for all k > kp (cf. Theorem
6.42 and Proposition 3.30). Therefore, (SBF^Z]) for ft implies (SBF^Z]) for /*.
Hence/* € &[Z -> Z].
(ii) Take # € /^ [Z —> Z]. Let ^ be the convex extension of g and g* be the
conjugate of ~g in the sense of (8.3). We have
Note that the right-hand side is an integer and the coefficient of a on the left is
either ±1 or 0. By virtue of this integrality, the above inequality is satisfied by all
a e [0, I]R, if it is satisfied by some a > 0. Therefore, (8.14) holds for all a 6 [0, I]R.
Similarly, there exists qo G Zv such that
This follows from the discrete conjugacy theorem (Theorem 8.12) in combination
with Theorems 7.40, 6.59, 4.15, and 5.5. In addition, we can obtain the M^-fLfi-
version of (8.17).
The conjugacy relationship among discrete convex functions is schematized in
Fig. 8.1, where M^-convex and L^-convex functions are defined in section 8.3. This is
the ultimate picture for the discrete conjugacy relationship, which originated in the
equivalence between the base family and the rank function of a matroid (section 2.4).
In other words, the exchange property and submodularity are conjugate to each
other at various levels.
Examples of mutually conjugate M-convex and L-convex functions are demon-
strated below for integer-valued functions defined on integer points.
In the network flow problem in section 2.2, if /„ e C[Z —> Z] and ga e C[Z —>
Z] are conjugate for each arc a € A, then
in (2.42) and (2.43) are conjugate to each other. We will dwell on the conju-
gacy in network flow in section 9.6.
In a valuated matroid (V,B,w), which arises, e.g., from a polynomial matrix
(section 2.4.2),
53
An alternative proof is possible on the basis of (i)—(iii) below if (iii) is accepted as a known
fact: (i) (8.15) is equivalent to x — a(xu — Xv) 6 dn9(po), (ii) dRg(po) is an integral M^-convex
polyhedron (Theorem 7.43 (2)), and (iii) for an integral M^-convex polyhedron Q and x, y 6 Qr\Zv,
a0 = 1 is valid in (B^EXCp,]).
8.1. Conjugacy 215
Note 8.13. In section 2.1 we saw the conjugacy between M^-convex and L''-convex
quadratic functions in real variables (Theorems 2.11 and 2.16). This conjugacy
relationship does not have a discrete counterpart. Let / : Zn —> Z be a quadratic
function represented as f ( x ) = XT Ax, with a positive-definite symmetric matrix A
with integer entries, and /* : Z" —> Z be the discrete Legendre-Fenchel transform
(8.11) of /. If A satisfies (6.26) through (6.28), then / is M^-convex and /* is
L^-convex, but /* is not necessarily a quadratic function. Likewise, if A satisfies
(7.10), then / is L^-convex and /* is M^-convex, but /* is not necessarily a quadratic
function. For instance, f ( x ) = x2, where x G Z, is an M^-convex function with
8.2 Duality
Discrete duality theorems lie at the heart of discrete convex analysis. Major theo-
rems presented in this section are the separation theorem for M-convex/M-concave
functions (M-separation theorem), the separation theorem for L-convex/L-concave
functions (L-separation theorem), and the Fenchel-type duality theorem. These
theorems look quite similar to the corresponding theorems in convex analysis, but
they express, in fact, some deep facts of a combinatorial nature. Almost all dual-
ity results in optimization on matroids and submodular functions are corollaries of
these theorems.
Proposition 8.14.
(1) 7//,-/ie.Ml[Z->R], then
Proof. (1) Theorem 6.44 with /i = / and /2 = —h shows that fi + f% > 0 implies
A+]2>o.
(2) It suffices to prove the claim when g, — k e £[Z —> R]. Theorem 7.19
applied to g and — k shows this. D
(by the integrality of the function value). By (6.87) and Theorem 6.61 (2) we have
Since domz/ and domzft are disjoint M-convex sets, the separation theorem for M-
convex sets (Theorem 4.21) gives p* G Hv such that the right-hand side of (8.21)
with p = p* is nonnegative. With this p* and a* G R such that f ' ( p * ) < -a* <
h°(p*), the inequality (8.20) is satisfied.
For integer-valued / and ft, we have /•,—ft 0 G £[Z|R —> R] and, hence,
dom R /*,dom R ft° G £0[Z|R]. We may assume p0 G Zv by (5.9) and p* G Zv by
Theorem 4.21. Then f*(p*) and h°(p*) are integers, and therefore we can take an
integer a* e Z. D
Moreover, if g and k are integer valued, there exist integer-valued j3* G Z and
x* e zv.
Proof. We may assume g, —k & £[Z —> R].
(i) Suppose that domzg n domzfc j^ 0. For the convex closure ~g of g and the
concave closure A; of k, we have (?(p) > fc(p) (Vp e R17) by Proposition 8.14 (2).
Since domR<7 n domRfc ^ 0, the separation theorem in convex analysis (Theorem
3.5) gives /?* e R and x* e Ry such that g(p) > /3_* + (p,x*) > Jfc(p) (Vp e Ry)
(see Note 8.19). This implies (8.22) since g = g and k = k on Zv by Theorem 7.20.
8.2. Duality 219
The integrality assertion is proved from the facts that the integer subdif-
ferential of an integer-valued L-convex function is an M-convex set and that M-
convex sets have the property of convexity in intersection. We may assume that
inf{0(p) - k(p) | p 6 Zv} = 0. Then there exists p0 e Zv with g(pQ) - k(p0] = 0
(by the integrality of the function value). By (6.87) and Theorem 7.43 (2), we have
Since domz<? and dom^A; are disjoint L-convex sets, the separation theorem for L-
convex sets (Theorem 5.9) gives x* G R^ such that the right-hand side of (8.23) with
x = x* is nonnegative. With this x* and /?* e R such that g*(x*} < —/3* < k°(x*),
the inequality (8.22) is satisfied.
For integer-valued g and k we have g°,—k° e ,M[Z|R —» R] and, hence,
domRfi(*,domRA;0 e Ai0[Z|R]. We may assume x0 e Zv by (4.34) and x* & Zv by
Theorem 5.9. Then g*(x*) and k°(x*) are integers, and therefore we can take an
integer j3* e Z. D
Moreover, if fi and fa are integer valued, i.e., /i,/2 G M^[Z —> Z], we can choose
integer-valued p* G Zv.
Proof. The sufficiency of (8.25) and (8.26) is obvious. Conversely, suppose that
(8.24) is true and apply the M-separation theorem (Theorem 8.15) to f ( x ) = f \ ( x )
and h(x) = f i ( x * ) + h(x*) — fa(x) to obtain a* and p* satisfying (8.20). We have
a* = fi[—p*](x*) from (8.20) with x = x* and hence
This implies (8.25) and (8.26), which are equivalent to (8.27) and (8.28), respec-
tively, by the M-optimality criterion (Theorem 6.26 (2)). To prove (8.29) take
x G argmin(/i + / 2 ). Then
which, along with (8.25) and (8.26), implies x G argmin/i[— p*] n argmin/2[+p*].
Hence follows C in (8.29), whereas D is obvious. Finally, the integrality of p* is due
to the integrality assertion in Theorem 8.15. D
Note 8.18. The assumptions on the effective domains are necessary in the sepa-
ration theorems (Theorems 8.15 and 8.16). For instance, for an M-convex function
/ : Z2 —> Z U {+00} and an M-concave function h : Z2 —> Z U {—00} denned by
we have
Note 8.19. This is a technical supplement to the proof of the M-separation the-
orem (Theorem 8.15). (A similar remark applies to the proof of the L-separation
theorem.) We applied the separation theorem to / and ft, the convex and concave
8.2. Duality 221
Note 8.20. The original proof of the M-separation theorem is based on an algo-
rithmic argument for a generalization of the submodular flow problem involving an
M-convex cost function (Murota [142]). In particular, the argument is purely dis-
crete, not relying on the separation theorem in convex analysis. See sections 9.1.4
and 9.5. •
from the definitions (8.11) and (8.12) of conjugate functions, where / and h are the
convex and concave closures of / and h, respectively. We observe the following:
2. The first inequality can be strict even when / is convex extensible and h is
concave extensible, as is demonstrated by Example 1.6. A similar statement
applies to the third inequality.
The following theorem asserts that the first and second inequalities in (8.30) turn
into equalities for Mb-convex/M^-concave functions and L^-convex/L^-concave func-
tions and that all three inequalities are equalities for such integer-valued functions.
222 Chapter 8. Conjugacy and Duality
Proof. (1) Suppose that domz/ H dornzft. ^ 0. By (8.30) we may assume that
A = mf{f(x) — h(x) x e Zv} is finite. By the M-separation theorem (Theorem
8.15) for (/ - A, h), there exist a* 6 R and p* & Hv such that
for all x e Zv, which implies ft°(p*) - f ' ( p " ) > A. Combining this with (8.30)
shows (8.31) as well as the attainment of the supremum by p*. Next suppose that
domz/ n domz/i = 0 and doniR/* n doniR/i0 ^ 0. The separation theorem for
M-convex sets (Theorem 4.21) applied to BI = domz/i and B2 = domz/ gives
p* e {0, ±1}V satisfying (4.33). Putting p = p0 + cp* in (8.21) (within the proof
of the M-separation theorem) and letting c —> +00, we obtain sup — +00 in (8.31),
whereas inf — +00 by domz/ n domz/i = 0.
8.2. Duality ; 223
(2) (The proof goes in parallel with (1).) Suppose that domzg fl domzA; ^ 0.
By (8.30) we may assume that A = inf{<?(p) — k(p) \ p e Zv} is finite. By the L-
separation theorem (Theorem 8.16) for (g — A, k), there exist /?* e R and x* e Rv
such that
for all p & Zv, which impliesfc°(a;*)- ff*(x*) > A. Combining this with (8.30)
shows (8.32) as well as the attainment of the supremum by x*. Next suppose
that domzg n domzfc = 0 and domR,<?* PI domR,fc0 ^ 0. The separation theorem
for L-convex sets (Theorem 5.9) applied to D\ = domz^ and D-2 = domzff gives
x* € {0, ±1}V satisfying (5.10). Putting x = x0 + ex* in (8.23) (within the proof
of the L-separation theorem) and letting c —> +00, we obtain sup — +00 in (8.32),
whereas inf = +00 by domzg fl donizfc = 0.
(3) In the proof of (1) we can take a* € Z, p* e Zv, and c € Z. The supremum
and infimum for finite (8.33) are attained since the functions are integer valued.
(4) In the proof of (2) we can take j3* e Z, x* £ Zv, and c e Z. The
supremum and infimum for finite (8.34) are attained since the functions are integer
valued. 0
Note 8.22. In Theorem 8.21 (1) the infimum is not necessarily attained by any
x G Zv (and similarly for (2)). For example, consider / : Z —» R U {+00} and
h : Z -^ R U {-00} defined by
Note 8.23. The assumptions on the effective domains are necessary in Theorem
8.21. For the M-convex and M-concave functions / and h in Note 8.18, we have
domz/ n domz/i = 0 and domz/* n domz/i0 = 0. The identity (8.33) fails with
infimum = +00 and supremum = — oo. •
224 Chapter 8. Conjugacy and Duality
8.2.3 Implications
In spite of the apparent similarity to the corresponding theorems in convex analysis,
the discrete duality theorems established above convey deep combinatorial proper-
ties of M-convex and L-convex functions. We now demonstrate this by deriving
major duality results in optimization on matroids and submodular functions as im-
mediate corollaries of these theorems (see also Fig. 8.2). The connection to the
duality in network flow problems is discussed in Chapter 9.
Example 8.26. The Fenchel-type duality theorem for submodular set functions is a
special case of the Fenchel-type duality theorem for L''-convex functions (Theorem
8.21 (2), (4)). The conjugate functions of a submodular set function p : 2V —>
RU{+oc} and a supermodular set function \JL : 2V —» R(j{—00} (i.e., p, —/it G <S[R])
8.2. Duality 225
are defined by
with an additional integrality assertion that, for integer-valued p and n, the maxi-
mum on the right-hand side of (8.35) can be attained by an integer vector x 6 Zv.
As in Example 8.24, we consider an L''-convex function g and an L''-concave function
k associated with p and p.. We have g* = p°, k° = /it 0 , and domz<? PI donizfc ^ 0,
and, therefore, (8.35) is obtained as a special case of (8.32) and (8.34). •
Example 8.27. Prank's weight-splitting theorem for the weighted matroid in-
tersection problem is a special case of the optimality criterion for the M-convex
intersection problem (Theorem 8.17). Given two matroids (V, BI) and (V,BZ) on
a common ground set V with base families BI and $2, as well as a weight vec-
tor w : V —» R, the optimal common base problem is to find B £ B\ n B^ that
minimizes the weight w(B) — ^V€BU)(V). Frank's weight-splitting theorem says
that a common base B* e BI n B% is optimal if and only if there exist real vectors
w*, w% : V —> R such that
(i) w = w* + toj-I,
(ii) B* is a minimum-weight base of (V, B\) with respect to w"(, and
(iii) B* is a minimum-weight base of (V, BZ) with respect to w%-
In addition, the theorem states that, if w is integer valued, the vectors w^ and w\
can be chosen to be integer valued. The combinatorial content of this theorem lies
in the assertion about the existence of an integer weight splitting in the case of
integer-valued weight. Applying Theorem 8.17 to a pair of M-convex functions
yields p* satisfying (8.25) and (8.26) with additional integrality in the case of integer-
valued w. A weight splitting constructed by
has the properties (ii) and (iii) because of (8.25) and (8.26). In Example 1.21 we
derived the weight-splitting theorem (integer-weight case) from the M-separation
theorem (Theorem 8.15). •
Example 8.28. Suppose we are given two valuated matroids (V, a>i) and (V,^)
as well as a weight vector w : V —> R. The valuated matroid intersection problem
is to find B C V that maximizes w(B) + u\(B) + w 2 (B). The weight-splitting
theorem for valuated matroid intersection says that a common base B* maximizes
226 Chapter 8. Conjugacy and Duality
w(B) + LOi(B) + W2(S) if and only if there exist real vectors w*,w^ '• V —> R
such that
(i) w = wl + w^,
(ii) B* maximizes u>i[w;*], and
(iii) B* maximizes u^fu^L
where wiftu*] and u^^] are defined by (2.76). In addition, the theorem states that,
if wi, uj2, and w are all integer valued, the vectors w* and w^ can be chosen to be
integer valued. Let /i and /2 be the M-convex functions associated, respectively,
with LUi and o>2 by (2.77). Maximizing w(B) + uJi(B) + ^(B) is equivalent to
minimizing f \ ( x ) + / 2 [ — w ] ( x ) , and a desired weight splitting can be obtained from
the M-convex intersection theorem (Theorem 8.17) as in Example 8.27. •
Note that a set is M2-convex (resp., M2-convex) if and only if its indicator function
is M2-convex (resp., M^-convex).
The effective domain and the set of minimizers of an M2-convex function are
M2-convex sets; the latter is a consequence of the M-convex intersection theorem
(Theorem 8.17).
8.3. M2-Convex Functions and L2-Convex Functions 227
Proposition 8.29.
(1) For an M%-convex function f, dom/ is M^-convex.
(2) For an M%-convex function f, dom/ is M^-convex.
Proof. This follows from the relation dom (/i + /2) = dom /i n dom /2 and the M-
or M^-convexity of dom/, (i = 1, 2) given in Proposition 6.7. D
Proposition 8.30.
(1) For an M%-convex function f, argmin/ is M^-convex if it is not empty.
(2) For an M^ -convex function f, argmin/ is M^-convex if it is not empty.
where /, /i^jind /2_are the local convex extensions (3.61) of /, /i, and /^respec-
tively, and /i and /2 are the convex closures (3.56) of /i and /2. Since /i + /2 is
convex, so is /. D
Theorem 8.32 (M2-optimality criterion). For an M2- convex function f 6 -M2[Z —>
R] and x €. dom /, we have
Proof. By Theorem 8.31 the optimality criterion for an integrally convex function
(Theorem 3.21) applies. We may impose the condition |V| = \Z\ because x(V) is
constant for any x e dom/. D
Theorem 8.33 (M2-optimality criterion). For M-convex functions fi, f? € M.\L —>
R] and a point x £ dom /j n dom /2, we have
if and only if
Theorem 8.35.
8.3. M2-Convex Functions and L2-Convex Functions 229
Proof. Using the M-convex intersection theorem (Theorem 8.17), we see that
For (2) and (3), note that dnfi(x) e C\[Z R] from Theorem 6.61 (2). D
where DZ denotes the integer infimal convolution (6.43) and * is the discrete
Legendre-Fenchel transformation (8.11)z- A relation conjugate to this also holds
for M-convex functions as follows.
Theorem 8.36. For integer-valued Afi-convex functions /i,/2 € M^[Z —> Z] with
dom/i n dom/2 ^ 0, we have (/i + /2)* = /i*Qz h" and (fi + /2)** = /i + h-
by the M-convex intersection theorem (Theorem 8.17). This shows that (/i +/ 2 )* >
/i* n z /2*, whereas < is obvious from (p,o;) - /i(x) - / 2 (x) < /*(g) + /*(p - <?).
The second identity follows from the first because of (8.38) and Theorem 8.12. D
Note that a set is L2-convex (resp., L2-convex) if and only if its indicator function
is L2-convex (resp., L^-convex).
230 Chapter 8. Conjugacy and Duality
for some <?i,<?2 G £[Z —»• R]- We may assume that the infimum is attained for
each p 6 domg. Namely, it is known that for an L2-convex function g there exist
L-convex functions gi, g% 6 £[Z —> R] such that
The infimal convolution g — gi^z92 is identically zero, with the infimum in (8.40)
unattained. An obvious valid choice for (8.41) is 171 = g% = 0 (identically). •
Note 8.38. For #1,52 G £[Z —> R], it can be shown that (51 DZfl^Xpo) = —oo
(3p 0 ) =*> (Si n z02)(p) = -oo ( V p e domsri+dom^)- •
The effective domain and the set of minimizers of an L2-convex function are
L2-convex sets.
Proposition 8.39.
(1) For an L-2-convex function g, doing is L^-convex.
(2) For an L\-convex function g, domg is L^-convex.
Proof. This follows from the relation dom (giOz g2) = domgi +dom<72 and the Iv-
or L^-convexity of dom <jij (i — 1, 2) given in Proposition 7.8.
Proposition 8.40.
(1) For an L%-convex function g, argming is L^-convex if it is not empty.
(2) For an I\-convex function g, argming is L\-convex if it is not empty.
Proof. By (8.41) this follows from a general fact in Proposition 8.41 below and the
L- or L^-convexity of argming, (i = 1,2) given in Proposition 7.16. D
then we have
8.3. M2-Convex Functions and l-2-Convex Functions 231
since the converse inclusion is always true, independently of (8.42). Take p* &
argmin(<7iD2 <?2)[—a:]. By (8.42) there exist p\ and p\ such that p* = p* + p2
and (giO2g2)(p*) = gi(p*) + g2(p2)- If 5i[-^](Pi) > 9i[-x](pi) for some pi, we
would have
The assumption (8.42) above is necessary for the identity (8.43) to hold. For
the functions gi and 52 in Note 8.37, for instance, we have argmin((?in z g 2 ) = Z2
but arg min gi = arg min 52 = 0- For integer-valued functions, however, (8.42) is
always satisfied.
L2-convexity implies integral convexity.
and, hence,
we have
and divide into two cases: (i) v G W and (ii) v €E V" \ W. In case (i), let i be such
that v € t/ij. Then v e C/2i follows from
Therefore, -1 < Xc/ii(f) - Xu2i(v) < 0, which implies (8.45). In case (ii), let i be
such that v G U^i- Then TJ € C/H follows from
Proof. By Theorem 8.42 this is obtained as a special case of the optimality criterion
for an integrally convex function (Proposition 3.22).
Proof. Let g be represented as (8.41) with L-convex functions g\ and 02, where
gi(p + 1) = 0j(p) (Vp) for i = 1, 2 as a consequence of g(p) = g(p + 1) (Vp). There
exist p?,p2* e Zy such that 0(pa) = 0i(p") + ^2(^2) and Pa = Pi + Pa • For anY
V C I/ we have
By the L-proximity theorem (Theorem 7.18) there exist p* & argmingi and p^ &
argmin<72 such that
The following two theorems are the counterparts of Theorems 8.35 and 8.36.
Theorem 8.45.
(1) .For0i,02 e /^[Z —*• R] withgiOzg2 > -oo and(8A2) andp e dom(0in z 0 2 ),
there exist pt e dom0j (i = 1,2) such that p = p\ +P2 and
(2) For 01,02 6 ^[Z —> Z] with giOzg2 > -oo andp e dom(0inz52), i/iere
exzsi ^ 6 dom 0j (i = 1,2) SMC/I tftai p = Pi + P2 «wrf
(3) -For 0 € >C^[Z —> Z] and p € dom0, <9z0(p) is an A^ -convex set. For
g e £2[Z —> Z] anrfp e dom0, dzg(p) is an M^-convex set.
8.3.3 Relationship
The relationship between M2- and L2-convex functions is discussed here. The first
theorem shows the conjugacy relationship between M2- and La-convex functions.
Theorem 8.48. The two classes of functions M.^\L —> Z] and H^\L —> Z] are
in one-to-one correspondence under the discrete Legendre-Fenchel transformation
(8.11)z, and similarly for M\[Z -> Z] and £ 2 [Z -> Z].
Proof. This is due to (8.38) and Theorems 8.36, 8.46, and 8.12. D
where c : Zv —> Z U {+00} and 0 ^ B C Zv. The canonical "convex" case consists
of problems in which
8.4. Lagrange Duality for Optimization 235
with
where SB '• Zv —> {0, +00} is the indicator function of B. We say that the problem
P is feasible if f ( x ) < +00 for some x G Zv.
Next we embed the optimization problem P in a family of perturbed problems.
As the perturbation of / we consider F : Zv x Zu —> Z U {+00}, with U being a
finite set, such that
236 Chapter 8. Conjugacy and Duality
Here the second condition (8.55) means that the integer biconjugate of F(x,u) as
a function in u for each fixed x coincides with F(x, u) itself.
Note 8.50. By Proposition 8.11, the condition (8.55) is satisfied if, for each x,
either F(x, •) = +00 or the integer subdifferential of F(x,u) with respect to u is
nonempty for each u € domF(:r, •). Recall that an integer-valued M^- or Lj-convex
function has a nonempty integer subdifferential (Theorems 8.35 and 8.45). •
For each x 6 Zv, the function K(x,-) : y >-* K(x,y) is the concave discrete
Legendre-Fenchel transform of the function —F(x, •) : u H-> —F(x,u).
Our assumptions (8.54) and (8.55) on F(x,u) guarantee the following.
Proposition 8.51.
(1) F(x,u) = sup{K(x,y)-(u,y)\y£Zu}
( 2 ) f ( x ) = sup{K(x,y) yeZu}(xeZv).
Proof. (1) Abbreviate F(x,u) and K(x,y) to F(u) and K(y), respectively. We
have F'(y) = -K(-y) by (8.58), while F(u) = F"(u} by (8.55). Therefore,
We say that the problem D is feasible if g(y) > — oo for some y 6 Zu. If the problem
P is feasible, we have g(y) < +00 for all y € Zu, since g(y) < K(x,y) < f(x] for
all x € Zv and y e Zu.
8.4. Lagrange Duality for Optimization 237
We write min(P) instead of inf (P) if the problem P is feasible and the infimum is
finite, in which case the infimum is attained (i.e., opt(P) ^ 0), and similarly for
max(D).
Proof. We have
Our main interest lies in the strong duality, namely, in the case where the
inequality in the weak duality turns to an equality with a finite common value.
Theorem 8.53.
(1) g(y) = -v*(-y).
(2)sup(D) = p"(0).
(3) inf(P) =¥>(()).
(4) inf(P) = sup(D) <=> ip(0) = <p"(0).
(5) Suppose inf(P) is finite. Then min(P) = max(D) <S=> dz<p(Qi) ^ 0.
(6) //min(P) = max(D), then opt(D) = -dz<p(0).
mfu((p(u) + (u, y)} = g(y) is shown in the proof of (1). This implies the claims when
combined with the weak duality (Theorem 8.52).
Theorem 8.54 (Saddle-point theorem). Both inf(P) and sup(D) are finite and
min(P) = max(D) if and only if there exist x £ Zv and y 6 Zu such that K(x,y)
is finite and
Proof. By Proposition 8.51 (2) we have f(x) = supyK(x,y) for any x € Zv,
whereas g(y) = inf x K(x,y) for any y G Zu by the definition (8.60). In view of the
weak duality (Theorem 8.52) and the relation
we see that
We single out the case of r = 0 because the technical development in this special
case can be made within the framework of M-/L-convex functions, whereas the
general case involves M2-/L2-convex functions. Throughout this section we assume
(REG), i.e., that B is an M-convex set.
We use the subscript r to denote the quantities derived from Fr; namely,
8.4. Lagrange Duality for Optimization 239
Proposition 8.56.
Proof. It suffices to prove (2), since (1) is its special case with r = 0. Assume
x e dome. Substituting (8.61) into (8.64) we obtain
Proof. (1), (3) The expression of Ko(x,y) in Proposition 8.56 (!) shows these.
(2) In the expression of Kr(x,y) in Proposition 8.56 (2) we have SB-X + f G
M2[L —> Z] or = +00 (see Note 6.17). Then the conjugacy in Theorem 8.48 implies
this.
(4) In the expression of Kr(x,y) in Proposition 8.56 (2) the second term
(S-Bazf[y])(—x) is M-convex or e {+00, — 00} since it is the integer infimal con-
volution of two M-convex functions (Theorem 6.13 (8)). D
In the case of M-convex programs the dual objective function gr and the
optimal value function ipr are well behaved, as follows.
This shows ^o is L-concave or = — oo, since the sum of two L-concave functions is
again L-concave provided the effective domains of the summands are not disjoint.
(3) We have (po(u) = inf z (c(x) + SB(X + u)) = (cDz <5_ B )(—u). The assertion
follows from Theorem 6.13 (8).
(4) It follows from Fr(x,u) = FQ(X,U) + r(u) that <pr(u) = +00 unless u e
domr and that (pr(u) = <fo(u) + r(u) if u € domr. If (p0 e M[Z —> Z], then
ipr G M-2\L —> Z] or ipr = +00. If <fo(u) e {+00, —oo}, then <pr(u) e {+00, —oo}.
(2) First recall the relation gr(y) = —(pr*(-y) (Theorem 8.53 (1)). If (pr e
Mz[Z —> Z], the conjugacy in Theorem 8.48 implies the L2-concavity of gr. If
(pr = +00, then gr = +00. If (pr(u) = —oo for some u, then gr = — oo. D
Strong duality holds true for M-convex programs with the Lagrangian function
Kr(x,y).
Theorem 8.59 (Strong duality). Assume (REG), (OBJ), and that the problem P
8.4. Lagrange Duality for Optimization 241
Proof. Since </v(0) is finite by the assumption, <pr is M2-convex by Theorem 8.58
(4) and dz<fr(0) 7^ 0 by Theorem 8.35. Then the assertions follow from Theorem
8.53. D
Example 8.60. Let us consider the case where c(x) is a linear function on another
M-convex set B' C Zv. The primal problem with c(x) = (x,w) + SB'(X) (where
w € Zv denotes a weight vector) reads as follows:
This is the polymatroidal version of the optimal common base problem explained
in Example 8.27. The optimal solution y = y* to D gives the weight splitting
w* = w — y* and w\ = y*.
For a concrete instance, take V = {1,2},
We have seen that g is L2-concave, i.e., —g € £a[Z —> Z], in the "convex" case
where (REG) and (OBJ) are satisfied.
We are now interested in the reverse process, i.e., how to restore the primal
problem P from the dual D in a way consistent with the general duality framework of
section 8.4.1. We embed the dual problem D in a family of maximization problems
defined in terms of another perturbation function G(y, v) such that G(y, 0) = g(y)
and -G(y, •) e £2[Z -> Z]. Namely,
As can be imagined from the corresponding constructions in convex analysis (cf. sec-
tion 4 of Rockafellar [177]), K(x,y) and f ( x ) thus constructed do not necessarily
coincide with the original K(x,y) and f ( x ) . We show, however, that the dual of
the dual comes back to the primal in the canonical case with a bounded M-convex
set B using the Lagrangian function Kr.
Proof. The definitions (8.67) and (8.69) show Gr(y, •) = ~(Kr(-, y))* and Kr(; y) =
(-Gr(y,-))' for each y. Since Kr(-,y) 6 M2[Z -> Z] or = +00 by Theorem 8.57
(4) when B is bounded, we have KT(-,y) — (Kr(-,y))" for each y. Hence follows
Kr = Kr. Then Proposition 8.51 (2) and (8.70) imply / = /. D
in which c[—v] is M-convex. On the other hand, Theorem 8.58 (2) shows that
is L2-concave, gr(y) = —oo for all y, or gr(y) = +00 for all y. By replacing c with
c[—v] we obtain the claim. D
The optimal value function jr, defined by (8.68) with reference to Gr, enjoys
the following properties.
Proof. (1) Using Proposition 8.51 (2), (8.69), and Proposition 8.62 we obtain
244 Chapter 8. Conjugacy and Duality
(2) Since / 6 M.i\L —> Z] or = +00, the assertion follows from (1) and the
conjugacy between £2[Z —> Z] and M.i\L —> Z]. D
Theorem 8.65. Assume (REG), (OBJ), and iftai ifte problem P is feasible and B
is bounded.
(1) min(P) = 7r(0) = 7r°°(0) = max(D r ).
(2)opt(P)=d z (-7,)(0)^0.
Proof. The proof is essentially the same as that of Theorem 8.59. To be specific,
we have the following chain of equivalence: x & dz(—7r)(0) •<=> Tr(^) ~ 7r(0) <
(v, -x) (Vv e Zv) &-mfv((v, -z)-7 r (t>)) = 7r(0) <=> f ( x ) = jr(0). This implies
the claim when combined with the weak duality (Theorem 8.52). D
Bibliographical Notes
The conjugacy relationship between M-convexity and L-convexity was established
first for integer-valued functions (Theorem 8.12) by Murota [140], whereas the
present proof is based on Murota [147]. The conjugacy theorem for polyhedral
M-/L-convex functions (Theorem 8.4) is due to Murota-Shioura [152]. The polar-
ity between M-/L-convex cones in Theorem 8.5 is stated in [147] and Proposition
8.11 for the integer biconjugate is in [140]. Theorem 8.1 is a special case of a theorem
.of Topkis [202], stated explicitly as Corollary 2.7.3 in Topkis [203].
The M-separation theorem (Theorem 8.15) is given in Murota [137], [140],
[142] and the L-separation theorem (Theorem 8.16) in [140]. The Fenchel-type
duality theorem for M-convex functions originated in [137] (see also [140]); the
present form (Theorem 8.21) is in Murota [147]. The M-convex intersection theorem
(Theorem 8.17) is in [137], [142]. The Fenchel-type duality theorem for submodular
set functions described in Example 8.26 is due to Fujishige [62]. The weight-splitting
theorem for weighted matroid intersection in Example 8.27 is due to Frank [54], and
that for valuated matroid intersection in Example 8.28 is due to Murota [135]; see
also Theorem 5.2.40 of Murota [146].
M2-convex and L2-convex functions were introduced by Murota [140], to which
Theorems 8.35, 8.36, 8.45, and 8.46 and the conjugacy theorem (Theorem 8.48) ar
ascribed. Theorems 8.31 and 8.42 (integral convexity) as well as Theorem 8.49 are
due to Murota-Shioura [153]. Theorems 8.32 and 8.43 (M2-/L2-optimality criteria
and Theorems 8.34 and 8.44 (M2-/L2-proximity theorems) are given by Murota-
Tamura [162]. See Tamura [198] for Notes 8.37 and 8.38.
The Lagrange duality of section 8.4 is developed in Murota [140]. See Nemhauser-
Rinnooy Kan-Todd [166] and Nemhauser-Wolsey [167] for the subadditive duality.
Chapter 9
Network Flows
245
246 Chapter 9. Network Flows
should be equal to the given x. The problem is described by a graph G = (V, A),
an upper capacity c : A —> R U {+00}, a lower capacity c : A —> R U {—oo}, a
cost vector 7 : A —> R, and a supply vector x : V —> R, where it is assumed that
c(a) > c(a) for each a G A The variable to be optimized is the flow £ : A —> R.
Minimum cost flow problem MCFP0 (linear arc cost)55
is additionally imposed on the flow £. We refer to the problem with (9.6) in place
of (9.5) as the minimum cost integer-flow problem.
To discuss the relationship to convex analysis it is convenient to consider a
more general form of the minimum cost flow problem. The generalization is twofold.
First, the linear arc cost /^ae.4 7( a )£( a ) is replaced with a nonlinear cost represented
by a separable convex function ^2aeA / 0 (£(a)) with a family of univariate polyhedral
convex functions fa S C[R —> R] indexed by a G A. Second, with a polyhedral
convex function / : Rv —> R U {+00}, an additional term /(d£) for the flow
boundary <9£ is introduced in the cost function as a generalization of the supply
specification <9£ = x.
Minimum cost flow problem MCFPs (nonlinear cost)56
55
MCFP stands for minimum cost flow problem.
56
We have MCFP; for i = 0, 3 and not for i = 1, 2. This is for consistency with section 9.2.
9.1. Minimum Cost Flow and Fenchel Duality 247
for a € A and / is the indicator function 6{xj of the singleton set {x}.
Among the four nice properties of MCFP0 listed above, the optimality crite-
rion by potentials is generalized to MCFPa, as we will see in section 9.1.3, whereas
the other three fail to survive for a general /. In considering the integer-flow version
of the problem it is natural to assume /„ e C[Z —> R] (or /„ e C[Z|R —> R]) for
each a e A, but it is not clear what combinatorial property to impose on / to ensure
the integrality of optimal solutions. M-convexity gives an answer to this, as we will
see in section 9.4.
9.1.2 Feasibility
For the minimum cost flow problem MCFPo, a feasible flow means a function £ :
A —» R that satisfies
where the summation is taken over all arcs a connecting X \ Y and Y\X. D
If a flow £ meets the capacity constraint (9.12), its boundary x = <9£ satisfies
for all X C V and also x(V) = 0 = «(V). This means x e B(re), where B(K) is the
base polyhedron (4.13) associated with n.
The above argument shows that the condition x € B(K) is necessary for
MCFPo to be feasible. It is also sufficient, as stated in the following theorem.
That is,
Proof. This follows from the max-flow min-cut theorem or a variant thereof, called
Hoffman's circulation theorem (see, e.g., (2.65) of Fujishige [65] or Theorem 3.18 of
Cook-Cunningham-Pulleyblank-Schrijver [26]). D
where inf /[—p] and mi fa[6p(a)] with a € A mean the infima of the reduced cost
functions. The inequality (9.25) gives a lower bound for the minimum of F3. In
particular, if
for some p, then £ is an optimal flow satisfying (9.25) with equality. This statement
is true for any functions / and fa (a & A).
The converse is also true under a fairly general assumption that / and fa
(a €. A) are convex.
Theorem 9.4 (Potential criterion). In the minimum cost flow problem MCFPs
with polyhedral convex f and fa (a & A), we have the following:
(1) For a feasible flow £ : A —> R, the two conditions (OPT) and (POT) below
are equivalent.
(OPT) £ is an optimal flow.
(POT) There exists a potential p : V —» R such that
(i) £(o) € &Tgmmfa[Sp(a)] for every a & A, and
(ii) d£ € argmin/[-p].
(2) Suppose that a potential p : V —> R satisfies (i) and (ii) above for an
optimal flow £. A feasible flow £' is optimal if and only if
250 Chapter 9. Network Flows
Proof. (1) (POT) =$> (OPT) is already shown. To prove (OPT) => (POT), suppose
that £ is an optimal flow. Putting
we see
where inf FS is finite and x = <9£ attains the infimum of the last expression. Noting
that /A is a polyhedral convex function (see Note 2.17) and dom/^ n dom/ ^ 0,
we apply the Fenchel duality theorem in convex analysis (Theorem 3.6 and (3.42))
to obtain p : V —> R such that
The second equation shows (ii) in (POT). We will show that the first equation above
implies (i) in (POT). It follows from (9.26) and (9.21) that
On the other hand, the optimality of £ implies /yi(d£) = ^2a€A /a(£( a ))> which, in
combination with (9.21), yields
Substituting (9.28) and (9.29) into the first equation in (9.27) shows
9.1. Minimum Cost Flow and Fenchel Duality 251
by (2.34) and (2.35), condition (i) in (POT) says that flow £(a) and tension r?(a) =
—6p(a) should satisfy the constitutive equation in every arc a G A. In the case of lin-
ear arc cost, the characteristic curve Fa takes the form of Fig. 9.1, and, accordingly,
condition (i) in (POT) is expressed as
In the conventional case MCFPo with linear arc cost, the optimality criterion
can be reformulated in terms of negative cycles in an auxiliary network. For a
feasible flow £ : A —> R, let G^ = (V, A^) be a directed graph with vertex set V and
arc set A^ = A^ U jB| consisting of two disjoint parts:
252 Chapter 9. Network Flows
Theorem 9.5 (Negative-cycle criterion). For a feasible flow £, : A —> R to the mini-
mum cost flow problem MCFPo, conditions (OPT) and (NNC) below are equivalent.
(OPT) £ is an optimal flow.
(NNC) There exists no negative cycle in (G^,l^) with t% o/(9.34).
Proof. By (9.31), (9.32), and the definition (9.34) of £?, condition (i) of (POT) in
Theorem 9.4 is equivalent to
whereas condition (ii) of (POT) is void for MCFPo. On the other hand, the existence
of a potential p : V —*• R satisfying (9.35) is equivalent to (NNC), as is well known
in network flow theory. Hence follows the equivalence of (NNC) and (OPT) by
Theorem 9.4. D
The minimum cost flow problem MCFP0 is endowed with remarkable inte-
grality properties:
1. An integer-valued optimal flow exists if the upper and lower capacities and
the supply vector are integer valued (primal integrality).
2. An integer-valued optimal potential exists if the cost vector is integer valued
(dual integrality).
Theorem 9.6 (Integrality). Suppose that the minimum cost flow problem MCFPo
has an optimal solution.
(1) [Primal integrality] // c : A —> Z U {+00}, c : A —> Z U {-oo}, and
x : V —> Z, then there exists an integer-valued optimal flow £ : A —> Z.
(2) [Dual integrality] The set of optimal potentials
II* = {p p : optimal potential}
is an L-convex polyhedron. If^:A^Z, then II* is an integral L-convex polyhedron
and there exists an integer-valued optimal potential p : V —> Z.
Proof. (I) Let p be an optimal potential. By (9.31) and (9.32), a flow £ is optimal if
and only if it is a feasible flow with respect to a more restrictive capacity constraint
c*(a) <£,(a) <c*(a) with
9.1. Minimum Cost Flow and Fenchel Duality 253
for each a & A. Since c*(a) and c*(a) are integers for every a € A, the claim follows
from (9.19) in Theorem 9.3.
(2) Since condition (i) of (POT) in Theorem 9.4 is equivalent to (9.35) in the
proof of Theorem 9.5, II* coincides with the polyhedron described by (9.35) with
an optimal £. This implies the L-convexity of II* (see section 5.6). The integrality
assertion follows from Proposition 5.1 (4). D
The nice features of the minimum cost flow problem discussed so far (Theorems
9.4, 9.5, and 9.6) are derived mainly from the combinatorial structure inherent
in the underlying graph, as well as the convexity of the cost functions. Further
combinatorial properties stemming from the M-convexity of the cost functions will
be investigated in section 9.4 and section 9.5.
Note 9.7. Here is a comment on the definition of the coboundary. In this book
we follow the convention of defining Sp(a) by
The boundary d£(v) is defined to be the amount of flow leaving v and the tension 77
is defined as 77 = -8p (see (9.1), (9.20), and (9.22)). Then follows the fundamental
identity
The notations div and A in Rockafellar [178] are related to ours as div = d and
A = -5. m
and the evaluation of JA amounts to solving a minimum cost flow problem with
nonlinear arc cost /„ but without boundary cost /. Thus, the minimum cost flow
problem MGFPs with boundary cost can be understood as a composition of the
254 Chapter 9. Network Flows
where
arc set is A = {(^1,^2) | v € V}, with v\ € V\ and v-z € V-z denoting the copies of
v € V. We define the boundary cost function / : R/1 x R72 -» R U {+00} by
and assume that the arc cost functions fa (a e A) are identically zero without
capacity constraints. Note that x\ = —x2 if (xi,x2) = d£ for a flow £ in this
network. Assuming inf (/i — /la) > — oo, let £ be an optimal flow, which exists since
/ is a polyhedral convex function. Let (piiPi} 6 R^1 x R^2 be an optimal potential
satisfying (POT) in Theorem 9.4. Condition (i) of (POT) implies pi = p2- Since
for x = d£|vi and p = p\. This implies the Fenchel duality (3.41) for f\ and h^\ see
also (3.30) and (3.42).
The nice properties described in section 9.1 are maintained if -B is a base polyhedron
represented as B = B(p) with a submodular set function p : 2V —> RU{+oo}. Such
a problem described by some p G <S[R] is called the submodular flow problem.
Submodular flow problem MSFPi (linear arc cost)57
57
MSFP stands for M-convex submodular flow problem. We use the notation MSFPj with
i = 1, 2, 3 to indicate the hierarchy of generality in the problems.
256 Chapter 9. Network Flows
Note that the M-convex submodular flow problem with a {0, +00}-valued / reduces
to the submodular flow problem MSFPi. In the integer-flow version of the problem
we assume c : A -> Z U {+00}, c : A -> Z U {-oo}, and / 6 M[Z -> R] (or
/eM[Z|R->R]).
A still further generalization is possible by replacing the linear arc cost in T%
with a separable convex function. Namely, using univariate polyhedral convex func-
tions fa e C[R -» R] (a € A), we consider Y^aeA /a(£(°0) instead of ^2aeA 7(a)£(a)
to obtain MSFPa below, a special case of MCFPs with / being M-convex.
M-convex submodular flow problem MSFPs (nonlinear arc cost)
The converse is also true; i.e., MSF?3 can be put into a problem of the form of
MSFP2, as is explained in Note 9.8.
Throughout this chapter we assume
9.2. M-Convex Submodular Flow Problem 257
since d£(V) = 0 for any flow £ and d£ & dom / = B — B(/j) is imposed.
In subsequent sections we will see that the optimality criteria in terms of
potentials and negative cycles, as well as efficient algorithms for the conventional
minimum cost flow problem MCFPo, can be generalized for the M-convex submod-
ular flow problem.
Note 9.8. The problem MSFP3 on G = (V,A) can be written in the form of
MSFP2 on a larger graph G = (V, A). We replace each arc a = (u, v) G A with a
pair of arcs, a+ = (u, i>~) and a~ = (v£,v), where v+ and v~ are newly introduced
vertices. Accordingly, we have A = {a + ,a~ | a G A} and V = V(J{v^,v~ \ a G A}.
For each a G A we consider a function /0 : R2 —> R U {+00} given by
Note 9.9. The cost function FS of MSFPs consists of two terms, the separable arc
cost X^aeA /o(£(a)) and the M-convex boundary cost /(<9£). Noting that the former
is M^-convex, one might be tempted to consider a (nonseparable) M^-convex cost
function defined on the arc set. The integer-flow version of such a problem, however,
contains the Hamiltonian path problem, a well-known NP-complete problem, as a
special case.
Suppose that we want to check for the existence of an (s, i)-Hamiltonian path
in a directed graph G = (V, A), where we may assume s ^ t G V and 5~s = 6+t = 0.
We construct another directed graph G = (V, A) by replacing each arc a = (u, v) G
A with three arcs connected in series:
M is another partition matroid defined similarly (with + replaced with —), and M°
is the graphic matroid in which B° C. A° is a base if and only if {a G A \ a° G B°} is
258 Chapter 9. Network Flows
a tree of the original graph G. Let Q be the set of characteristic vectors of a subset
B of A such that B H A+ is a base of M+, B n A~ is a base of M~, and B r\ A° is
an independent set of M°. Then a {0, l}-flow £ in G with £ e Q and <9£ = Xs —
corresponds to an (s, i)-Hamiltonian path in G. Since Q is an M^-convex set, the
constraint | e Q can be represented by a {0, +oo}-valued M^-convex cost fun
on the arc set A. •
Moreover, if c, c, and p are integer valued and the problem is feasible, there exists
an integer-valued feasible flow £ : A —> Z.
Proof. Let K be the cut capacity function denned by (9.16). By Theorem 9.3 a
feasible flow exists if and only if B(/c) n B(p) 7^ 0. The latter condition is equiva-
lent to
In a feasible problem with integer-valued c, c, and p, both B(K) and B(p) are
integral base polyhedra (integral M-convex polyhedra), and B(K) n B(p) n Zv is
nonempty by (4.32). Then (9.19) in Theorem 9.3 guarantees the existence of an
integer flow £ : A -+ Z with d£ G B(K) n B(p) n Zv. D
Note 9.11. The necessity of (9.54) is easy to see. For any X C V, the net amount
of flow entering X is equal to zero:
9.3. Feasibility of Submodular Flow Problem 259
Combining these two yields (9.54). Theorem 9.10 claims that this "obvious" neces-
sary condition is in fact sufficient. •
Note 9.12. In a feasible submodular flow problem, the set of boundaries of feasible
flows, dE = {d£ £ : feasible flow}, is an M2-convex polyhedron, and it is an integral
M2-convex polyhedron if c, c, and p are integer valued. This can be seen from the
proof of Theorem 9.10. •
The maximum submodular flow problem is to find a feasible flow £ that max-
imizes £(ao) for a specified arc ao € A.
Maximum submodular flow problem maxSFP
A max-flow min-cut theorem holds for this problem. Note that for any X C V
with OQ G /\+X we have an "obvious" inequality:
where this common value can be +00. If c, c, and p are integer valued and (9.61)
is finite, there exists an integer-valued maximum flow £ : A —> Z.
Proof. Divide the arc ao = (u, v) into two arcs in series, say, ao = (u, w) and
a0 = (w,v), and denote by G = (V,A) the resulting graph, where V = V U {w}
and A = A U {o0}. Define the capacities of a0 by C(OQ) = t and c(a0) = +00 with
a parameter i, and let p be defined for all subsets of V by p(X U {w}) = p(X) for
X C V. The maximum in (9.61) is equal to the maximum (or supremum) of t such
260 Chapter 9. Network Flows
that the submodular flow problem on G = (V, A) is feasible. With this relationship,
Theorem 9.10 implies (9.61) as well as the integrality assertion. D
If £ and X attain the maximum and the minimum in (9.61), respectively, and
if £(ao) < c(a0), then we have
by the M-optimality criterion in Theorem 6.52 (1). These expressions show how the
conditions in (POT) can be verified efficiently for a given p. It is also mentioned that
these expressions lead to another optimality criterion in terms of negative cycles, to
be established in section 9.5, and furthermore to the cycle-canceling algorithm for
the M-convex submodular flow problem, to be explained in section 10.4.3.
9.4. Optimality Criterion by Potentials 261
This shows that argmin/[—p] coincides with the base polyhedron B(gp) associated
with the set function gp denned by
This expression is used in the primal-dual algorithm for the M-convex submodular
flow problem, to be explained in section 10.4.4.
We go on to discuss integrality properties of the M-convex submodular flow
problem. This generalizes the well-known facts (Theorem 9.6) for the minimum cost
flow problem MCFP0. Recall the notation M[Z\R -> R] and M[R -»• R|Z] for the
sets of integral and dual-integral polyhedral M-convex functions, respectively.
Theorem 9.15. Suppose that an optimal solution exists in the M-convex submod-
ular flow problem MSFP3 with fa <E C[R -> R] (a e A) and f e X[R -> R].
(1) The set of the boundaries of optimal flows,
is an L-convex polyhedron.
(2) [Primal integrality] If fa £ C[Z|R -> R] (a e A) and f e X[Z|R ->• R],
then 95* is an integral M^-convex polyhedron, and there exists an integer-valued
optimal flow £ : A —> Z.
(3) [Dual integrality] ///„ e C[R -> R|Z] (a G A) and f 6 M[R ->• R Z], then
II* is an integral L-convex polyhedron, and there exists an integer-valued optimal
potential p : V —> Z.
Proof. (1) Let p be an optimal potential. Since argmin/ a [5p(a)] forms an interval,
say, [c*(a),c*(a)]R, condition (i) in (POT) of Theorem 9.14 can be expressed as
c*(a) < £(a) < c*(o) (a 6 A). Just as in (9.16) and (9.18), the set of <9£ for such £
coincides with the base polyhedron B(K*) for K* defined by
262 Chapter 9. Network Flows
Combining this with (9.65) we obtain 9H* = B(K*) nB(<7 p ), which is an M2-convex
polyhedron. Now let £ be an optimal flow. Potentials p satisfying (i) in (POT) form
an L-convex polyhedron, say, DI, by (9.63), whereas those satisfying (ii) in (POT)
form another L-convex polyhedron D% by (9.64). Therefore, II* = DI n D% is an
L-convex polyhedron.
(2) Both B(K*) and B(0P) are integral M-convex polyhedra. The integrality
of dE* = B(/c*) n B(gp) follows from (4.32).
(3) Both DI and D% are integral L-convex polyhedra by Theorem 6.61 (1).
The integrality of II* = DI n D2 follows from Theorem 5.7. D
For linear arc cost, with fa given by (9.50), the integrality conditions are
simplified as follows:
Finally, we state the optimality criterion for the integer-flow version of the
M-convex submodular flow problem MSFPa. This is a corollary of Theorems 9.14
and 9.15.
is an M-2-convex set.
(4) If the cost functions are integer valued, i.e., if fa e C[Z —> Z] (a e A) and
f £ M[Z —> Z], then there exists an integer-valued potential p :V—>Zin (POT).
Moreover, the set of integer-valued optimal potentials,
is an L-convex set.
In connection to (i) and (ii) in (POT) in Theorem 9.16, note the equivalences
9.5. Optimality Criterion by Negative Cycles 263
Note 9.17. The Fenchel-type duality theorem for M-convex functions (Theorem
8.21) is essentially equivalent to the optimality criterion for the M-convex submod-
ular integer-flow problem (Theorem 9.16). See section 9.1.4 and note that, for an
M-convex function /i and an M-concave function /i 2 > f ( x i , X 2 ) = /i(a^i) — /i2(—#2)
is an M-convex function.
Proof. As is well known in network flow theory, (NNC) is equivalent to the existence
of a potential p : V —> R such that
By (9.31), (9.32), (9.64), and the definition (9.71) of i^ this condition is equivalent
to conditions (i) and (ii) of (POT) in Theorem 9.14. Hence follows the equivalence
of (NNC) and (OPT) by Theorem 9.14. D
Note 9.19. In a problem with dual integrality the arc length ^ is integer valued.
The integrality of ^(a) for a e A* U B* is due to (9.67) and that for a e Q is
by Theorem 6.61 (1). For integer-valued £^, we can take an integer-valued p in the
proof of Theorem 9.18. •
Proof. This is similar to the proof of Theorem 9.18. Note, however, that Theorem
9.16 is used here in place of Theorem 9.14. D
Note 9.21. The M-convex intersection problem introduced in section 8.2.1 can
be formulated as an M-convex submodular flow problem. Given two M-convex
functions /i,/2 : Zv —> R U {+00}, we consider an M-convex submodular flow
problem on the bipartite graph G = (Vi U V2,A) in Fig. 9.3, where V\ and Vi are
copies of V and A = {(v\, v%) v e V} with vi £ V\ and v% € Vi denoting the copies
of v e V. The boundary cost function / : ZVl x ZV2 -> R U {+00} is defined by
/(zi,X2) = f i ( x i ) + f2(^X2) for Xi G ZVl and X2 6 Z^ 2 , whereas the arc costs are
9.5. Optimality Criterion by Negative Cycles 265
The following theorem shows that £ is a feasible flow with an improvement in the
objective function
This gives an alternative proof for "(OPT) =>• (NNC)," which has already been
established in Theorem 9.20 with the aid of Theorem 9.16.
Theorem 9.22. For a feasible integer flow £ to the M-convex submodular integer-
flow problem MSFP2 with (9.72), let Q be a negative cycle with the smallest number
of arcs in (Gj,^). Then £ in (9.75) is a feasible integer flow and
266 Chapter 9. Network Flows
The rest of this section is devoted to the proof of Theorem 9.22. The key
ingredient of the proof is the unique-rain condition, denned as follows.
For a pair ( x , y ) of integer vectors satisfying x G doniz/ and ||x — y\\<x> = 1, we
consider a bipartite graph G(x, y) = (V+, V~\E) with vertex sets V+ = supp+(:r -
y) and V~ = supp^ (x - y) and arc set
The following proposition gives a necessary and sufficient condition for a bi-
partite graph to have a unique minimum-weight perfect matching. It also shows
that the unique-min condition for a pair of integer vectors can be checked by an
efficient algorithm.
Proof. This follows from the complementarity (Theorem 3.10 (3)) in the linear
program formulation in the proof of Proposition 3.14. D
The following is the key fact. Note that <9£ G domz/ and \\d£ — <9£||oo — 1-
9.5. Optimality Criterion by Negative Cycles 267
We have \V+\ = \V \ — m for m = ||<9£ — d£||i/2 and the weight of arc (u,v)
equal to A/(9£; v, u). We may think of G(d£, d£) as a subgraph of the graph Gj of
section 9.5.1 by regarding E as a subset of C^ in (9.73). Then Q n C% determines a
perfect matching in G(d£,d£).
Let M — {(ui,Vi) | i = 1,...,TO}be a minimum-weight perfect matching in
G(d£,d£) and p be an optimal potential in Proposition 3.14. Note that M is a
subset of
Since Q' is a union of disjoint cycles with \Q'\ = \Q\ and Q is a negative cycle with
the smallest number of arcs, (9.78) implies that Q' is also a negative cycle with the
smallest number of arcs.
To prove by contradiction, suppose that (d£, <9£) does not satisfy the unique-
min condition. Since (KJ, vt) e (7| for i = 1,..., m, it follows from Proposition 9.24
that there exist distinct indices i*. (fc = 1,..., q; q > 2) such that (uik, vik+1) e C£
for fc = 1 , . . . , g , where i g+ i = i\. That is,
i.e.,
268 Chapter 9. Network Flows
For k — ! , . . . , < / , let P'(vik+1,Uik) denote the path on Q' from Vik+1 to Uik,
and let Q'k be the directed cycle consisting of arc (u$ fc , Vik+1) and path P'(vik+1, itj fc ).
Obviously,
where the union here (and also below) means the multiset union, counting the
number of occurrences of elements. A simple but crucial observation is that
where (9.79) and (9.78) are used. This implies that ^(<2'fc) < 0 for some k, which,
however, is a contradiction, since Q'k has a smaller number of arcs than Q'. This
completes the proof. D
Proof of Theorem 9.22: It follows from Propositions 9.25 and 9.23 as well as
the definition of / that
whereas
We may think of f(y) as the minimum cost to meet a demand specification y at the
exit, where the cost consists of two parts, the cost f(x) of supply or production of
x at the entrance and the cost 5^aeA/a(^( a )) of transportation through arcs; the
sum of these is to be minimized over varying supply x and flow £ subject to the
flow conservation constraint <9£ = (x, — y, 0). A similar interpretation is possible
for g(q). We regard / and g as the results of transformations of / and g by the
network; (9.81) and (9.82) are called transformations of flow type and of potential
type, respectively.
270 Chapter 9. Network Flows
Theorem 9.26. Assume fa,ga <E C[Z -> Z] for each a <E A. For f,g : Zs
Z U {+00}, let /, g : ZT -> Z U {±00} be the functions induced by G = (V, A- S, T)
according to (9.81) and (9.82), where it is assumed that f > —oo, f ^ +00, g > —oo,
and g ^ +00.
where S' = {v1 v& S} and T" = {v" v e T}. Let / : Zs/ -> Z U {+00} be the
indicator function of the set of the characteristic vectors of bases, i.e., / = SB for
B = {xxr | X e B}, and, for each arc a 6 A, define fa to be the indicator function
of {0,1}. Then the induced function / : ZT —> Z U {+00} represents the family B
in the sense that / = 6^ for B = {XY" \Y&B}. Then the M-convexity of / stated
in Theorem 9.26 (1) shows that (T,B) is a matroid.
The second case is where (5, T) = (V, 0). Then the induced functions / and
g are constants, having no arguments, and the conjugacy asserted in Theorem 9.26
(3) amounts to a min-max relation
for
9.6. Network Duality 271
This is the discrete counterpart of (9.36), showing the duality nature of the assertion
of Theorem 9.26 (3).
The third case is where 5 = 0. Then the induced functions are
which are identical to (2.42) and (2.43), respectively, and the claims in Theorem
9.26 reduce to the facts observed in section 2.2.2.
Whereas Theorem 9.26 deals with integer-valued functions defined on integer
points, similar statements are true for functions of type Z —> R and R —> R. Note
that the conjugacy assertion is missing in the case of Z —> R.
Theorem 9.27. Assume fa,ga 6 C[Z -> R] for each a e A. For f,g : Zs ->
R U {+00}, let /, g : ZT -> R U {±00} be the functions induced by G = (V, A; 5, T)
according to (9.81) and (9.82), where it is assumed that f > — oo, / ^ +00, g > —oo,
and g ^ +00.
Theorem 9.28. Assume fa,ga £ C[R —> R] for each a e A. For f,g : Rs —>
RU{+oo}, let f,g : RT -» RU{±oo} be the functions induced by G = (V,A;S,T)
according to
where it is assumed that f > -oo, / ^ +00, g > -oo, and g ^ +00.
272 Chapter 9. Network Flows
by a refinement of the augmenting path argument used in Note 2.19 for the special
case of S = 0. We take & e ZA and z; e Zs for i = 1,2 such that
We search for a kind of augmenting path with respect to the pair (£1, £2) that yields
the desired inequality (9.84). If we are not successful in finding such a path, we
modify the flow pair to a new pair with smaller ^-distance ||£i — &\\i, so that we
can eventually find an appropriate augmenting path.
58
For w G T, xfu is the characteristic vector of {w} in ZT.
274 Chapter 9. Network Flows
Before giving a formal proof we explain the idea of the proof in a typical
situation. Consider the difference in the flows, £2 — Ci 6 Z'4, for which we have
9(& ~ £1) = (xy ~ xi,yi - 2/2,0). Since u € supp+(<9(£2 - £1)), there exists a
simple path, say, PI, compatible with £2 ~~ £1 that connects u to some vertex v\
in supp~(i9(^2 — £1)) = supp + (xi - £2) U supp~(j/i - yi). This is an augmenting
path with respect to the pair of flows £1 and £2- Suppose that we have the case of
vi E supp + (xi — £ 2 ). By (M-EXC[Zj) for / we obtain HI <E supp"(xi — £2) such
that59
Since u\ E supp+(9(^2 — £i))> there exists a simple path, say, P2, compatible with
£2 ~ £1 that connects HI to some v% € supp~(<9(£2 ~ Ci)) = supp + (xi — ^2) U
supp~(yi — 7/2)- Suppose further that v? E supp~(yi — 2/2) and P% is vertex disjoint
from PI. Putting v — V2 we represent the path PI U P% by TT : A —> {0, ±1}
such that supp+(7r) C supp+(^2 — 6): supp'(?r) C supp^(^ 2 - Ci)> and dir =
(Xui - X?i > Xu - X« > 0). For the augmented flows £{ = ^ + TT and ^2 = 6 - TT a
the new bases x{ = £1 - x^ + xf j and x'2 = x2 + X^ - xf1, we have
and
since
which shows the inequality (9.84). Having presented the rough idea, we are now in
a position to start the proof that works in general.
We shall construct a pair of flows £,( and £2 that satisfy (9.87) for some v E
supp~(yi — 2/2) and £^,£2 6 Zs, and also
where
59
For w 6 S, Xw is the characteristic vector of {w} in Zs.
9.6. Network Duality 275
The frontier vertex is updated to this w' and the boundaries to b( =b\~Xw+ Xw1,
62 = 62 + Xw - Xw'- The flows remain the same: ip{ = ipi and ip'2 = y? 2 -
A crossover updates ((/PI, tp2,61, 62, w) with reference to another flow-boundary
tuple (</^,</?2,6i,62, U ) °) with the same frontier vertex w° = w. The flow-boundary
tuple is updated to
according to whether
we have
which shows (TRF[Zj) for g. Suppose that g(q) is finite for q = qi,q2- There exist
(77i,Pi,7*i) and (772,252,7*2) such that
Here we have
with
Therefore, we have
Fix y with f ( y ) finite. By Theorem 9.16 (4) there exists (p*,q*,r*) such that
Bibliographical Notes
The minimum cost flow problem treated in section 9.1 is one of the most fundamen-
tal problems in combinatorial optimization. For network flows, Ford-Fulkerson [53]
is the classic, whereas Ahuja-Magnanti-Orlin [1] describes recent algorithmic devel-
opments; see also Cook-Cunningham-Pulleyblank-Schrijver [26], Du-Pardalos [43],
Korte-Vygen [115], Lawler [119], and Nemhauser-Wolsey [167]. Thorough treat-
ments of the network flow problem on the basis of convex analysis can be found in
Iri [94] and Rockafellar [178].
The submodular flow problem was introduced by Edmonds-Giles [46] using
crossing-submodular functions. The present form avoids crossing-submodular func-
tions on the basis of the fact, due to Fujishige [61], that the base polyhedron defined
by a crossing-submodular function can also be described by a submodular function.
See Fujishige [65] for other equivalent neoflow problems, such as the independent
flow (Fujishige [59]) and the polymatroidal flow (Hassin [87], Lawler-Martel [120],
[121]). The M-convex submodular flow problem was introduced by Murota [142].
Section 9.3 is a collection of standard results on the submodular flow problem.
Theorems 9.10 and 9.13 are taken from Fujishige [65] (Theorems 5.1 and 5.11,
respectively), where the former is ascribed to Frank [56].
The optimality criterion by potentials for the M-convex submodular flow prob-
lem was established by Murota [142] for the integer-flow version (Theorem 9.16) and
adapted to the real-valued case (Theorem 9.14) with the integrality assertion (The-
orem 9.15) in Iwata-Shigeno [105] and Murota [147].
The optimality criterion by negative cycles (Theorem 9.20) was established
by Murota [140], [142] for the integer-flow version and adapted to the real-valued
case (Theorem 9.18) in Murota-Shioura [152]. Theorem 9.22 is in [140], [142].
Proposition 9.23 is a reformulation of the unique-max lemma due to Murota [135].
Proposition 9.25 is also from [135]; the proof technique using (9.80) originated in
Fujishige [59].
Transformation by networks was found first for M-convex functions / s M [Z —>
R] by Murota [137]; the proof given in section 9.6.2 is due to Shioura [188], [189].
Transformation of L-convex functions is stated explicitly in Murota [145]. The ex-
tension to polyhedral M-convex and L-convex functions is made in Murota-Shioura
9.6. Network Duality 279
[152]. Theorems 9.26, 9.27, and 9.28 are explicit in Murota [147]. Induction of ma-
troids through graphs is due to Perfect [171] (for bipartite graphs) and Brualdi [21];
see also Schrijver [183], Welsh [211], and White [213]. The alternative proof of the
M^-convexity of a laminar convex function described in Note 9.31 is communicated
by A. Shioura.
This page intentionally left blank
Chapter 10
Algorithms
Algorithmic aspects of M-convex and L-convex functions are discussed in this chap-
ter. Three fundamental optimization problems tractable by efficient algorithms are
(i) M-convex function minimization, which is a nonlinear extension of the minimum-
weight base problem for matroids; (ii) L-convex function minimization, which in-
cludes submodular set function minimization as a special case; and (iii) minimiza-
tion/maximization in the Fenchel-type min-max duality, which is equivalent to the
M-convex submodular flow problem.
281
282 Chapter 10. Algorithms
property alone does not ensure finite termination in general, although it does if /
is integer valued and bounded from below.
Let us derive an upper bound on the number of iterations by considering the
distance to the optimal solution rather than the function value.
Proposition 10.1. If f has a unique minimizer, say, x*, the number of iterations
in the steepest descent algorithm is bounded by \\x° — x*\\\/1, where x° denotes the
initial vector found in step SO.
where
Note 10.3. For integrally convex functions we have the local optimality criterion
for global optimality (Theorem 3.21). This naturally suggests the following.
10.1. Minimization of M-Convex Functions 283
The steepest descent algorithm for M-convex functions is a special case of this. It is
emphasized that no efficient algorithm for step SI is available for general integrally
convex functions. •
By the M-proximity theorem (Theorem 6.37 (1)), the set B always contains a global
minimizer of / and, at the termination of the algorithm in step S2, x is a global
minimizer by the M-optimality criterion (Theorem 6.26 (1)). The number of itera-
tions is bounded by [Iog2(-ft'oo/4n.)] • Some remarks are in order concerning step SI.
If the function / is such that the scaled function / remains M-convex, step SI can
be done in O(F • n4) time by the steepest descent algorithm with tie-breaking rule
(10.2). This time bound follows from Proposition 10.2 with K\ < 4ra2. For a general
/, however, / is not necessarily M-convex (see Note 6.18) and no polynomial bound
for step SI is guaranteed, although we have an obvious exponential time bound
0(F-(4n)"n 2 ).
On the basis of Theorem 6.76 (quasi M-optimality criterion) and Theorem
6.78 (quasi M-proximity theorem), the steepest descent scaling algorithm can be
adapted to the minimization of quasi M-convex functions satisfying (SSQM^).
284 Chapter 10. Algorithms
The set B° is intended to represent the central part of B, i.e., the set of vectors of
B lying away from the boundary. The set B° is nonempty if B is M-convex; see
Proposition 10.6 below.
Domain reduction algorithm for an M-convex function / S Ai[Z —*• R]
SO: Set B :=dom/.
SI: Find a vector x <E B°.
S2: Find u, v € V (u ^ v) that minimize f(x — Xu + Xv)-
S3: If f ( x ) < f(x — Xu + Xv), then stop (x is a minimizer of /).
S4: Set B := B D {y € Zv \ y(u) < x(u) - 1, y ( v ) > x(v) + 1} and go to SI.
The vector x G B° in step SI can be found with O(n2log2K00) evaluations of /
by the procedure to be described below. The set B forms a decreasing sequence
of M-convex sets, which contain a minimizer of / because of the M-minimizer cut
theorem (Theorem 6.28). Since x is taken from the central part of B, UB(W) — IB(W)
for w e {u, v} decreases with a factor of (1 — ~) and hence the number of iterations
is bounded by O(n2 Iog2 -ftToo)- The above algorithm, therefore, finds a minimizer of
/ with O(n 4 (log 2 Koo)2) evaluations of / provided that it is given a vector in dom /.
(see Theorem 3.8 in Pujishige [65]). We prove (i) here; a similar argument works for
(ii). Fix X C V and put p = xx, Pv = 1 — Xv (v € X), and k — \X\. It follows from
At the end of the algorithm we have a = 1 and hence x = i by (10.6). This means
S(l) n dom/ = {x}, since x(V) = y(V) for any y G dom/ by Proposition 6.1 and
since x(V) < y(V) for any y e S(£) distinct from x. Furthermore, we see from the
second condition in (10.5) that a; is a minimizer of /, since {x} — S(K) n dom/ D
S(f)nargmin/^0.
The outline of the algorithm reads as follows, where step SI for the a-scaling
phase is described later and K^ is defined in (10.3).
10.1. Minimization of M-Convex Functionsiions 287
Proof. At the beginning of the a-scaling phase we have \\x°— (°\\\ < n(n— l)(2a—1)
by (10.6). Since step S3 in the a-scaling phase can be done with n evaluations of
/, the a-scaling phase terminates in O(F • n3) time. The number of scaling phases
is equal to \\og2(Koc/2n)~\. D
On the basis of Theorem 6.79 (quasi M-minimizer cut with scaling), the do-
main reduction scaling algorithm can be adapted to the minimization of quasi
M-convex functions satisfying (SSQM^) provided that the effective domain is a
bounded M-convex set.
288 Chapter 10. Algorithms
be the base polyhedron associated with p. Recall that a point in B(/o) is called a
base and an extreme point of B(p) is an extreme base. For any base x and any
subset X we obviously have
maximizer on the left-hand side. For any u € supp~(z) and v €. supp + (x), there
exists a subset Xuv such that u 6 Xuv C V \ {v} and x(Xuv) — p(Xuv). Put
X = U u6 sup P -(x) a6,Upp+(a,) Xuv. We have x(X) = p(X) by (4.23) and x'(V) =
x(X) since supp~(x) C X and supp+(a;) C V \ X.. The integrality assertion can be
established by the same argument starting with an integral base x that maximizes
x~(V) over all integral bases. D
The min-max relation (10.11) shows that we can demonstrate the optimality
of a subset X by finding a base x with x~(V) — p(X). But how can we verify that
a vector x belongs to B(p)? By definition, x e B(p) if and only if mmx(p(X) —
x(X)) = p(V) - x(V) = 0. Thus, testing for membership in B(p) for an arbitrary
x seems to need a submodular function minimization procedure.
To circumvent this difficulty, we recall from Note 4.10 that we can generate an
extreme base by the greedy algorithm. Let L = (vi,v<2, • • •, vn) be a linear ordering
of V and define L(VJ) = {i>i, t>2, • • •, vj} for j — 1,...,n. Then the extreme base y
associated with L is given by
where we may assume |J| < n by the Caratheodory theorem. Combining (10.12)
and (10.13) shows that any base can be represented by a list of linear orderings
{Li | i € /} (that generate {j/j | i 6 /}) and coefficients of the convex combination
{A; | i e /}. With this representation of x we can be sure that x is a member of
B(p). For u, v € V and i e I we use the notation
Proof. (1) is obvious. By (10.12) the condition in (2) implies p(W) = yi(W) for
every i € /. Then x(W) = Eie/^W = P(W)- (3) follows from (1), (2), and
(10.10). D
290 Chapter 10. Algorithms
Note 10.11. Because the minimizers form a ring family (see Note 4.8), there exists
a unique minimal minimizer as well as a unique maximal minimizer of a submodular
set function p : 2V —> R. Given an optimal base x with the representation x =
Sie/^il/i m (10.13), we can compute the minimal and maximal minimizers in
strongly polynomial time as follows.
With the notation T>(x) for the family of tight sets at x, introduced in (4.22)
of Note 4.9, we have a representation
for the family of the minimizers of p. Noting that T>(x) is a ring family, let Gx =
(V,AX) be the directed graph associated with T>(x), as defined by (4.20) in Note
10.2. Minimization of Submodular Set Functions 291
4.7. This is equivalent to saying that (u, v) € Ax if and only if v G dep(x, u), where
dep(x,w) means the smallest tight set at x that contains u. Then the minimal
minimizer can be identified as the set of vertices reachable from supp~ (x) in Gx
and the maximal minimizer as the complement of the set of vertices reachable to
supp+ (x) in Gx. Moreover, the graph Gx enables us to enumerate all the minimizers.
By (10.13) we have
Since each yi is an extreme base, we can easily compute dep(yi,u). For each u e V,
we start with D := V and update D to D \ {v} as long as D \ {v} € T^(yi)
for some v £ D \ {u}; we obtain D = dep(yi,u) at the termination. We can thus
compute dep(yi, u) with O(n 2 ) evaluations of p. The number of function evaluations
can be reduced to O(n) if a linear ordering Lj = (vi,vz,... ,vn) generating yi
is also available; assuming u = Vk, start with D = {vi,f2> • • • ,Vk-i} and, for
j = k - 1, k - 2 , . . . , 1, update D to D \ {vj} if D \ {vj} e £>(&). Therefore, the
graph Gx can be constructed with O(n 3 |/|) or 0(n 2 |/|) evaluations of function p,
where it is reasonable to assume |/| < n. •
Note 10.14. It has been assumed that the submodular function p is finite valued for
all subsets. This assumption is not restrictive but for convenience of description. Let
p 6 <S[R] be a submodular set function on V taking values in RU{+oo}; T> = domp
is a ring family with {0, V} C T>. For v € V we denote by Mv the smallest member
of T> containing v and by Nv the largest member of T> not containing v. For X C V
we denote by X the smallest member of T> including X. We assume that we can
compute Mv for each v & V efficiently, say, in time polynomial in n. Then we can
compute X and Nv by
292 Chapter 10. Algorithms
Let us assume, without loss of generality, that the length of a maximal chain of V
is equal to n = \V\. Consider now a set function ~p : 2V —> R denned by
with c e R^ given by
As is shown below, (i) p is a finite-valued submodular function and (ii) X £ arg min p
implies X e arg min p. Thus we can minimize p via the minimization of p.
Proof of (i): p is obviously finite valued. If X e V, v £ X, and Y = X U {v} e
T>, we have YL)NV = NVU {v} and Y D Nv = X and, therefore,
This means that n(X) = p(X}-\-c(X) is nondecreasing on T>. Putting Ji(X) — n(X)
for X C V and noting X U F = XUY and X n F D X n F, we see
Note 10.15. The method in Note 10.14 can be used to minimize a submodular
function defined on a (finite) distributive lattice. By a lattice we mean a tuple
(S, V, A) of a nonempty set S and two binary operations V and A on S1 such that
holds true. A ring family T> is a typical distributive lattice, where (S, V, A) =
(Z>, U , n ) . The converse is essentially true: any distributive lattice (S, V , A ) can be
represented in the form of a ring family (Birkhoff's representation theorem). The
size of the underlying set of the ring family is equal to the length of a maximal
chain of S. A function p : S —> R defined on a distributive lattice (S, V, A) is said
to be submodular if it satisfies
10.2. Minimization of Submodular Set Functions 293
(see (10.14) for the notation -<i) and searches for a directed path from P — supp + (x)
to N = supp~(a;). If there is no such path, the algorithm terminates by setting W
to be the set of vertices reachable to N. Then W and x satisfy all the conditions in
Proposition 10.9, and hence W is a minimizer of p and a; is a maximizer of x~(V).
Otherwise, it modifies {(Li, Aj) | i 6 /} with reference to a path from P to N.
Schrijver's algorithm for submodular function minimization
SO: Take any linear ordering LI and set / := {!}, AI := 1.
SI: Construct the graph G = (V,A) for {(Li, A,) | i e /} by (10.17).
S2: Set P := supp+(x), TV := supp~(x) for base x = £)ie/ A^ in (10.13).
S3: If there exists no directed path from P to AT in G, let W be the set
of vertices reachable to N and stop (W is a minimizer of p).
S4: Update {(I/j,Aj) i 6 /} and go to SI.
We now describe the concrete procedure for step S4, where a directed path exists
from P to ./V in G. Let d(v) denote the distance (= minimum number of arcs in a
directed path) in G from P to v.
We choose s, t e V as follows (a lexicographic rule). We fix a linear ordering -<Q
of elements of V; this is independent of the linear orderings Lj. Let t be the element
in N reachable from P with d(v) maximum; in the case of multiple candidates,
choose the largest with respect to -<Q. Let s be the element with (s,t) € A, d(s) =
d(t) — l; in the case of multiple candidates, choose the largest with respect to -<o- Let
a be the maximum of |(s,i]^J over i e I and let k e I be such that |(s, t]_<J = <*.
294 Chapter 10. Algorithms
which is obtained from Lk by moving vp+j to the position just before vp = s, and
let Zj be the extreme base associated with L*.
Proposition 10.17. For some S > 0, j/fc + 6(xt ~ Xs) can be represented as a
convex combination of {zj \ j = 1 , . . . , a}.
and, therefore,
where the inequalities follow from submodularity; for instance, for h = p+j, we have
in which (Vp-i U {vp+j}) U Vp+j-i = Vp+j and (Vp-i U {vp+j}) n Vp+j-! = Vp-!.
The sign pattern of (10.18), as well as that of (xt~Xs)(vh), for h with p < h < p+j
looks like:
Note also that each row sum is equal to zero since Zj(V) = yk(V). If all the diagonal
entries marked by @ are strictly positive, we can represent xt ~ Xs as a nonnegative
combination of {zj —y^ \ j = 1 , . . . , a} with a positive coefficient for j — a; namely,
10.2. Minimization of Submodular Set Functions 295
X t ~ X s = E°=i »j(zi ~ yk) with //j > 0 (j = 1,..., a - 1) and p,a > 0. Then the
claim is true for 5 = 1/($3?=1 Mj)- ^ a diagonal entry, say, in the j'oth row, vanishes,
then Zj0 = yk and the claim is true for 6 = 0. D
Note that x' can be represented as a convex combination of Y U {yk} and, moreover,
{yk} can be dispensed with if x'(t) < 0. By a variant of Gaussian elimination we
can obtain a convex combination representation of x' using at most n vectors from
Y U {yk}- We update {(Li, A,) | i e /} according to this representation. Since
\Y\ + 1 < 2n, step S4 can be done with O(n3) arithmetic operations.
j — 1,..., a, (s, £]_<• is a proper subset of (s, t]->k. This implies a' < a. If a' = a,
then /3' < (3, since x'(t) = x'(t') < 0 and L^ disappears in the update. Hence (c).
It follows from (c) that d(v) increases for some v € V in O(n 4 ) iterations
because each of d(t), t, s, a, (3 is bounded by n and there are at most n pairs
(d(t),t) if d does not change. For each v € V, d(v) can increase at most n times.
Therefore, the total number of iterations is bounded by O(n 6 ). D
Note 10.19. A more detailed analysis of Vygen [208] yields an improved bound of
O(n 5 ) on the number of iterations in Schrijver's algorithm. •
We start with a scaling algorithm for minimizing a submodular set function p : 2V —>
R. The algorithm is weakly polynomial for integer-valued p. It is emphasized that
the value of M in (10.8) need not be computed.
Recall from Proposition 10.8 that the problem dual to minimizing p is to
maximize x~(V) over x e B(/o). To add flexibility in solving this maximization
problem we introduce a scaling parameter 6 > 0 to relax (or enlarge) the feasible
region B(p) to B(p + Kg), where
The function Kg is submodular and therefore B(p + K,j) — B(p) + B(K^) by Theorem
4.23 (1). For a concrete representation of B(K^) we observe that Kg is the cut
capacity function associated with a complete directed graph G = (V, A), where
and the arc capacities are all equal to 6; indeed, Kg coincides with K in (9.16) with
c(a) = 8 and c(a) = 0 for every arc a € A. By a S-feasible flow we mean a function
if : A —> R such that for each (u,v) € A we have (i) 0 < <p(u,v) < 5 and (ii)
either <p(u,v) = 0 or <p(v,u) — 0. Then B(KS) = {d(p \ ip is 5-feasible} by (9.18) in
Theorem 9.3. Thus our relaxation problem with parameter 6 reads as follows:
Maximize z~(V) over z = x + dtp with x e B(p) and ^-feasible ip. (10.21)
The algorithm consists of scaling phases, each of which corresponds to a fixed
parameter value 5. We start with an arbitrary linear ordering, the extreme base x
associated with it, zero flow (p = 0, and
10.2. Minimization of Submodular Set Functions 297
Proof. Since S C W C V \ T, we have z(v) < 6 for every v £ W and z(v) > —5
for every v € V \ W. Therefore,
Since x(W) = p(W] by the nonexistence of active triples and Proposition 10.9 (2)
and d<p(W) > 0 by the nonexistence of arcs leaving W in Gv, we have z~(V) >
p(W) — n5. Since dtp(v) < (n — 1)5 for every v 6 V, we have
298 Chapter 10. Algorithms
With 5 < A/n 2 we have x~(V) > p(W) - n25 > p(W) - A, whereas x~(V) < p(Y)
for all Y C V by (10.10). Hence W is a minimizer of p. D
On the basis of Proposition 10.20 above we terminate the scaling phase if nei-
ther augmenting path nor active triple exists. Otherwise, while keeping z invariant,
we aim at "improving the situation" by either
1. eliminating an active triple or
2. enlarging the reachable set W.
With a view to eliminating an active triple (i, u, v) we modify Li by swapping u and
v in Li; denote the old pair (L/i,yi) by (Lk,yk) with a new index fc. The extreme
base associated with the updated Li is given by yi — yk + /3(Xu — Xv) with
Procedure Double-Exchange(i,«, v)
SI: Set /3 := p(Li(u) \ {v}} - p(Lt(u)) + yi(v), a := mm(^(u, v), A;/3).
S2: If a < \i/3, then let k be a new index and set
/ := 7 U {k}, Afc := A* - a//?, A* := a/j3, yk := j/», Lk := Lt.
S3: Set ^ := yi + j3(\u - Xv), x := x + a(xu ~ Xv), <p(u, v) := (p(u, v) - a.
Update Li by swapping u and v.
The overall structure of the algorithm, which we name the IFF scaling algo-
rithm, is described below. Reduce(o;,/) is a procedure that computes an expression
of x as a convex combination of at most n afflnely independent extreme bases cho-
sen from the current extreme bases indexed by I; this can be done by a variant
of Gaussian elimination. Parameter A for the stopping criterion in step S4 should
satisfy (10.24); we take A = 1 for integer-valued p.
10.2. Minimization of Submodular Set Functions 299
Proof. This can be derived from the properties listed in Proposition 10.22. D
Proposition 10.22.
(1) The number of scaling phases is O(log 2 (M/A)).
(2) The first scaling phase calls Augment O(n 2 ) times.
(3) A subsequent scaling phase calls Augment O(n 2 ) times.
(4) Between calls to Augment, there are at most n — 1 calls to nonsaturating
Double-Exchange.
(5) Between calls to Augment, there are at most 2n3 calls to saturating Double-
Exchange.
(6) We always have \I\ < In.
(7) Reduce(a;, /) with \I\ < 2n can be done in O(n 3 ) arithmetic operations.
Proof. (1) We have 6 < M/n2 in step SO, since x+(V) = x(X) < p(X) < M for
X = supp+(:c). The number of scaling phases is bounded by log 2 ((M/n 2 )/(A/n 2 )) =
log2(M/A).
(2) Let x denote the initial base in step SO. Then z (V) = x (V) at the
beginning of the scaling phase. Throughout the scaling phase we have z~(V) <
z(V) = x(V) as well as z~(V) < 0. Since Augment increases z~(V) by <5, the
number of calls to Augment is bounded by
300 Chapter 10. Algorithms
(3) At the beginning of a subsequent scaling phase, we have z~(V) > p(W) —nS
by Proposition 10.20, where z = x + dtp with (p = 2ip and 6 — 26. Since
this implies that z~(V) > p(W) — 2n8 — «2<5/4 at the beginning of the scaling phase.
Throughout the scaling phase we have
The following is a key property of the scaling algorithm that we make use of in
designing a strongly polynomial algorithm. Recall that a scaling phase ends when
the algorithm reaches step S4.
Proposition 10.23. At the end of a scaling phase with parameter S, the following
hold true.
(1) If x(w) < —n28, then w is contained in every minimizer of p.
(2) // x(w) > n2S, then w is not contained in any minimizer of p.
Proof. We have x~(V) > p(W) — n2S by Proposition 10.20, whereas, for any
minimizer X of p, we have p(W) > p(X) > x(X) > x ~ ( X ) . Hence, x~(V) >
x ~ ( X ) - n25. Therefore, if x(w) < —n2S, then w e X. On the other hand,
Using the scaling algorithm as a subroutine, we can devise a strongly polynomial al-
gorithm for submodular function minimization. The strongly polynomial algorithm,
which we call the IFF fixing algorithm, exploits two fundamental facts:
• The minimizers of p form a ring family (see Note 4.8).
10.2. Minimization of Submodular Set Functions 301
• A ring family can be represented as the set of ideals of a directed graph (see
Note 4.7).
Let D° be the directed graph representing the family T>° of minimizers of p. This
means that W C V is a minimizer of p if and only if it is an ideal of D° such
that min£>° C W C max 2}°. The algorithm aims at constructing the graph by
identifying arcs of D° or elements of (mini?0) U (V \ max£>°) one by one with the
aid of the scaling algorithm.
To be more specific, we maintain an acyclic graph D = (U, F) and two disjoint
subsets Z and H of V such that
Using the notation T(Y) = Uusr ^(u) f°r ^ — U, we define a function p : 2U —»•
Rby
Suppose that 77 > 0 and let u G U be the vertex that attains the maximum in
(10.26). Then
and we have at least one of the following three cases: (i) p(U) > 77/8, (ii) p(R(u) \
{u}) < -77/3, or (iii) p(R(u)) - p(U) > 77/8.
Case (i): If p(U) > 77/8, we invoke a procedure Fix + (p, D, 77), described below,
to find an element w 6 U that is not contained in any minimizer of p. Since T(w)
cannot be included in any minimizer of p, we add T(w) to H and delete w from D.
Case (ii): If p(R(u) \ {u}) < — 77/8, we invoke another procedure Fix~(/5, D, 77),
described below, to find an element w € U that is contained in every minimizer of
p. Since T(w) must be included in every minimizer of p, we add T(w) to Z and
delete w from D.
Case (iii): If p(R(u)) — p(U) > 77/8, we consider the contraction of p by R(u),
which is a submodular function p* on U \ R(u) defined by
The procedures F\x.~(p,D,r]) and Fix + (/5, D,rj) are the same as the IFF scaling
algorithm except that they start with 6 = 77 and a linear extension of the partial
order represented by D and return w in step S4. We put h = \U\.
10.2. Minimization of Submodular Set Functions 303
Proposition 10.25. The following statements hold true for Fix ± (p, D,rj).
(1) The number of scaling phases is O(log2 n), where n = \U .
(2)//
Proof. (1) Assume 8 < n/(3n3). By (10.28) we have x(U) = p(U) > ry/3 > hz5
and hence x(w) > n26 for some w & U. Therefore, the number of scaling phases
in Fix+ is bounded by [Iog2(3n3)] = O(log 2 n). For Y in (10.27), we have x(Y) <
p(Y) < —rj/3 < -n38 and hence x(w) < -n28 for some w G Y. Therefore, the
number of scaling phases in Fix" is bounded by O(log2 n).
(2) Let x denote the initial base in step SO. By the proof of Proposition
10.22 (2), the number of calls to Augment is bounded by x+(U)/S, whereas x+(U) <
nrj by (10.29). Since S — 77, the number of calls to Augment is bounded by n. D
Proof. In Case (i), (10.28) is obviously satisfied. Condition (10.27) is satisfied with
Y = R(u) \ {w} in Case (ii) and Y = U \ R(u) in Case (iii). To show (10.29) for
304 Chapter 10. Algorithms
by the submodularity of p and the definition of n. This proves (10.29) for (p, D, 77).
Finally, for each u e U \ R(u), put J?*(u) = R(u) \ R(u) and observe
We are now in the position to assert the correctness and the strong polyno-
miality of the IFF fixing algorithm.
Proof. This follows from Proposition 10.25 and Proposition 10.22 (3)-(7). D
Proposition 10.28. The IFF fixing algorithm finds the maximal minimizer of p.
Thus, we can guarantee an upper bound on the number of iterations. Let p° be the
initial vector found in step SO. If g has a minimizer at all, it has a minimizer p*
satisfying p° < p* by (10.31). Let p* denote the smallest of such minimizers, which
exists since p* A q* € argming for p*, q* € argmin <?.
306 Chapter 10. Algorithms
Proposition 10.30. In step SI, p < p* implies p + xx < P*• Hence the number
of iterations is bounded by \\p° - p*\\i-
whereas g(p*) < g(p*Vp') since p* is aminimizer of g. Hence g(p') > g(p*/\p'). Here
we have p' = p + Xx and p* A p' — p + Xx\Y, whereas X is the minimal minimizer
by the tie-breaking rule (10.33). This means that X \ Y = X; i.e., X n Y = 0.
Therefore, p' = p + xx < P*- D
Proposition 10.31. For an L-convex function g with finite K\, the number of
iterations in the steepest descent algorithm with tie-breaking rule (10.33) is bounded
by KI . Hence, if a vector in dom g is given, the algorithm finds a minimizer of g
in O((a(n)F + r(n))Ki) time.
Proof. We have \\p° — p*||i < KI since p°(v) = p*(v) for some v e V. Then the
claim follows from Proposition 10.30. D
The steepest descent algorithm above applied to this L-convex function g yields the
following algorithm for the L''-convex function g.
Steepest descent algorithm for an L^-convex function g G £&[Z —>• R]
SO: Find a vector p e doing.
SI: Find e € {1, —1} and X C V that minimize g(p + £Xx)-
S2: If g(p) < g(p + exx), then stop (p is a minimizer of g).
S3: Set p := p + sxx and go to SI.
Step SI amounts to minimizing a pair of submodular set functions
10.3. Minimization of L-Convex Functions 307
This is a translation of the tie-breaking rule (10.33) for g in (10.35) through the
correspondence
where p = (0,p) & Z x Zv. Since (l,Xv\x-) cannot be minimal in the presence of
(0, Xx+); we choose (1, X+) in the case of min/9+ = min/9~. At the termination in
step S2, p is a global minimizer by Theorem 7.14 (2) (L-optimality criterion).
In view of the complexity bound given in Proposition 10.31 we will derive a
bound on the size of dom g in terms of the size of dom g. Let K\ (g) be denned by
(10.34) for g. The li-size and ^-size of doing are denoted, respectively, by
Proof. Take p = {po,p} and q = (qo,q) in dom g such that Ki(g) = \po — qo\ +
\\p - q\\i and either (i) po = qQ or (ii) p(v) = q(v) for some v € V. We may assume
Po > Qo and p > q since p V q,p A q e domg and \\(p V q) - (p A q)\\i = \\p - q\\i-
The vectors p' = p — pol and q' = q — qol belong to domg. In case (i), we have
Ki(g] = \\p-q\\\ = \\p'~<l'\\i < KI. Incase (ii), we havep 0 -<?o = q'(v)-p'(v) and
The steepest descent algorithm could be used for minimizing quasi L-convex
functions satisfying (SSQSBW) because of Theorem 7.53 (quasi L-optimality crite-
rion). Note, however, that the set function pp of (10.32), to be minimized in step
SI, is not necessarily submodular and hence no efficient procedure is available for
step SI.
308 Chapter 10. Algorithms
rithm, the primal-dual algorithm, and the conjugate scaling algorithm. Because
the optimality criterion for the M-convex submodular flow problem is essentially
equivalent to the duality theorems for M-/L-convex functions, these algorithms can
be used for finding a separating afnne function in the separation theorem and the
optimal solutions in the minimization/maximization problems in the Fenchel-type
duality.
we obtain
where / e M[Z -> Z] and fA € M(Z -> Z]. Putting g = /*, ga = /0* for
a € A, and
310 Chapter 10. Algorithms
we have gA = fA* (see (9.28)) and also g e £[Z -> Z], #0 e C[Z -> Z] for a € v4,
and gA e £[Z -> Z]. The Fenchel-type duality (Theorem 8.21 (3)) gives
The function
The feasibility of this approach is guaranteed by the following facts if the given
functions, / and fa for a € A, can be evaluated.
1. We can evaluate g by applying an M-convex function minimization algorithm
to /. Similarly, we can evaluate ga for a e A.
2. We can find a minimizer p* in step SI by applying an L-convex function
minimization algorithm to g.
3. We can find a member of B* by applying an M-convex function minimization
algorithm to /[—p*}.
4. We can find a flow £* in step S2 as a feasible flow to the submodular flow
problem defined by c*, c*, and B*. This can be done, e.g., by the successive
shortest path algorithm described in section 10.4.2.
61
We have seen (10.44) in (9.83) as a special case of Theorem 9.26 (3).
10.4. Algorithms for M-Convex Submodular Flows 311
and define
In order to reduce the discrepancy ||z - d£\\i = X^ev \x(v) ~ <9£(v)|, tne
algorithm augments a unit flow along a shortest path P from S+ to S~ (shortest
with respect to the number of arcs) and modifies £ to £ given by
Obviously, £ satisfies the capacity constraint (10.49). The algorithm also updates
the base x to
which remains a base belonging to B; see Note 10.33. For the initial vertex s G S+
of the path P, either d£(s) increases or x(s) decreases by one and hence |z(s) —
<9£(s)| reduces by one. A similar result is true for the terminal vertex of P in S~,
whereas x(v) = d£(v) is kept invariant at every inner vertex v of P. Therefore,
each augmentation along a shortest path decreases ||z — <?£||i by two. Repeating
this process until the source S+ and consequently the sink S~ become empty, the
algorithm constructs a pair (£, z) with df, — x. Then £ is a feasible flow satisfying
both (10.49) and (10.50).
312 Chapter 10. Algorithms
and go to SI.
where cg(-, •, •) means the exchange capacity defined in (10.4). It can be shown that
stays in B; see Lemma 4.5 of Fujishige [65]. The successive shortest path algorithm
can be adapted to finding a real-valued feasible flow to a nonintegral submodular
flow problem. For a polynomial complexity bound of the algorithm, it is important
to choose, in the case of multiple candidates, an appropriate shortest path with
reference to some lexicographic ordering. The successive shortest path algorithm
can be generalized for optimal flow problems involving cost functions; see [65],
Fujishige-Iwata [66], and Iwata [98] for such algorithms for MSFPi (without M-
convex cost), whereas such an algorithm for the integer-flow version of MSFP2
(with M-convex cost) is given in Moriguchi-Murota [132]. •
Note 10.35. A number of algorithms are available for finding a feasible submodular
flow; e.g., Fujishige [59], Frank [56], Tardos-Tovey-Trick [199], and Fujishige-Zhang
[70]. The reader is referred to Fujishige [65], Fujishige-Iwata [66], and Iwata [98]
for expositions. •
cycle in an auxiliary network (Theorem 9.20) and that a nonoptimal flow can be
improved by augmenting a flow along a suitably chosen negative cycle (Theorem
9.22). These two facts suggest the following cycle-canceling algorithm, which works
on the auxiliary network (G^,^) introduced in section 9.5.
Cycle-canceling algorithm for MSFP2 (integer flow)
SO: Find a feasible integer flow £.
SI: If (G^,^) has no negative cycle, then stop (£ is an optimal flow).
S2: Let Q be a negative cycle with the smallest number of arcs.
S3: Modify £ to £ of (9.75) and go to SI.
The objective function I^ decreases monotonically by Theorem 9.22. This
property alone does not ensure finite termination in general, but it does if F^ is
integer valued and bounded from below. In the special case with dom/ C {0,1}^,
which corresponds to the valuated matroid intersection problem (Example 8.28),
some variants of the cycle-canceling algorithm are known to be strongly polynomial
(see Murota [136]).
Cycle-canceling algorithms can also be designed for problems with real-valued
flow on the basis of Theorem 9.18.
Note 10.36. For MSFPi (submodular flow problem with linear arc cost and with-
out M-convex cost), a number of cycle-canceling algorithms are proposed, includ-
ing Fujishige [59], Cui-Fujishige [27], Zimmermann [222], Wallacher-Zimmermann
[210], and Iwata-McCormick-Shigeno [103]. •
with gp(V) = 0 by (9.51), and B(gp) is the base polyhedron (M-convex polyhedron)
associated with gp.
The algorithm maintains a pair (£,p) 6 R"4 x Zv of a feasible flow £ and
an integer-valued potential p that satisfies (10.59) and repeats modifying (£,p) to
increase the set of arcs satisfying (10.57) and (10.58). We say an arc is in kilter
with respect to (£,p) if it satisfies (10.57) and (10.58) and out of kilter otherwise.
Note that exactly one of (10.57) and (10.58) fails for an out-of-kilter arc.
An initial (£,p) satisfying (10.59) can be found as follows. For any feasible
flow £ we consider a graph Gf — (V, Q) with vertex set V and arc set
and define the length of arc ( u , v ) as /'(<9£; — Xu + Xv), which is an integer by the
assumed dual integrality (Note 9.19). By solving a shortest path problem we can
find an integral vector p such that
where the dependence on p is implicit in the notation. Note that D^(v) is the set
of arcs leaving v for which (10.58) fails, D7 (v) is the set of arcs entering v for which
(10.57) fails, and {D^(v) \ v £ V} gives a partition of the set of out-of-kilter arcs.
If D^(v) = 0 for all v € V, condition (i) of (POT) is satisfied and the current
(£,£>) is optimal. Otherwise, the algorithm picks62 any v* € V with nonempty
D^(v*) and tries to meet the conditions (10.57) for a 6 D7(v*) and (10.58) for
a e Dt(v*) by changing the flow to £'. The flow £' is determined by solving the
following maximum submodular flow problem on G = (V, A) with a more restrictive
capacity constraint c*(a) < £'(a) < c*(a) with
62
The original primal-dual algorithm' picks an out-of-kilter arc and tries to meet (10.57) and
(10.58) for that arc. The present strategy to pick a vertex is to improve the worst-case complexity
bound.
10.4. Algorithms for M-Convex Submodular Flows 315
for each a 6 A.
Maximum submodular flow problem maxSFP
it is intended to~ meet the conditions (10.58) and (10.57) by increasing the flow
in a G D£(V*') to the upper capacity and decreasing the flow in a G D^(v*) to
the lower capacity. The maximum submodular flow problem is a special case of
the feasibility problem for the submodular flow problem and a number of efficient
algorithms are available for it (see section 10.4.2).
Let £' be an optimal solution to maxSFP above. If D^/(v*) is empty, the
algorithm updates £ to £' without changing p. The condition (10.59) is maintained
because of (10.65). If D^(v*) is nonempty, the algorithm finds a minimum cut
W C V (explained later) and updates p to
where A+W and A W mean the sets of arcs leaving W and entering W, respec-
tively, as in (9.14) and (9.15). The following proposition states that the flow value
£'(D£(v*))-£'(Dt(v*)) is bounded by v(W] for any W C V with v* e W and that
this bound is tight for some W, which is referred to as a minimum cut in the above.
A minimum cut can be found with the aid of an appropriately defined auxiliary
network.
for such a W. Then we obtain (10.68) from (9.61) in Theorem 9.13. Note that
d(p + Xw) — g(p) = 9p(W) in v(W) corresponds to p in (9.61). Finally, (10.69),
(10.70), and (10.71) are shown in (9.62). D
Proposition 10.38. #£' G B(<jy) for an optimum flow £' mmaxSFP and potential
p' = p + xw with a minimum cut W.
Proof. It follows from (10.65), (10.69), and discrete midpoint convexity (7.7) that
The following proposition shows the key properties for the correctness and
complexity of the primal-dual algorithm.
Proposition 10.39.
(1) The set of out-of-kilter arcs is nonincreasing.
(2) For each arc a, |7P(a)| is nonincreasing while the arc stays out of kilter.
(3) The potential is changed in at most \V\ iterations and, each time the
potential is changed, the value of max.a€rj^(v*) |7p(a)| decreases at least by one.
(4) Each time (£,p) is changed, the value of
we may assume that a is in kilter with respect to (£',p) and (i) a G A+VP7 \ Dt(v*)
or (ii) a e A~VF \ D^(v*). In case (i) we have £'(a) = c*(a) from (10.70), whereas
Thus the conditions (10.57) and (10.58) are preserved in either case.
(2) By (10.73), it suffices to show that, if (i) 7p(a) > 0, a e A+VF, or
(ii) 7P(a) < 0, a e A~W, then a is in kilter with respect to (£',p')- In case 0)i we
have a e <\+W\D+(v*), from which follows £'(a) = c*(a) = c(a) by (10.70). Since
7P'(a) > 0, a is in kilter with respect to (£',p')> and similarly for case (ii).
(3) If the potential does not change, we have D^i(v*) = 0 as well as D^(v) C
D^(v) for all w 6 V^ by (1). Therefore, the potential must be updated in |V|
iterations. Suppose now that p is changed to p'. An arc a & Df(v*).\ A+W is in
kilter with respect to (£',£>'), since 7 P /(a) < 7P(a) < 0 and £'(a) = c*(a) = c(a) by
(10.71). Similarly, a e D7(v*) \A~W is in kilter with respect to (£',p') because of
(10.70). Fora e D+(w*)nA+VF, we have |7j/(a)| = |7P(a)|-l froni7 p /(a) =7 p (a) +
1 and 7P(a) < 0. Similarly, for a e D~^(v*) n A^H 7 , we have |7P'(a)| = 7 P (a)| - 1
from 7P'(a) = 7P(a) — 1 and 7P(a) > 0. Therefore, maxo6£){(«*) |7P(«)| decreases at
least by one.
(4) When the potential changes, TV decreases because of (3). When the po-
tential remains invariant, N decreases because of Z?£/(u*) =0. D
Primal-dual algorithms can also be designed for problems without dual inte-
grality by using real-valued potential functions.
Note 10.40. The framework of the primal-dual algorithm for MSFPi (submodular
flow problem without M-convex cost) was established in Frank [55] and Cunningham-
Frank [30]. See Fujishige [65], Fujishige-Iwata [66], and Iwata [98] for expositions.
•
where the supremum is taken over integer points. Replacing g(p) with ga(p) =
g(ap)/a (p € Zv) in this expression we define
which we call the conjugate scaling of / with scaling factor a £ Z++. Note that
/^ is again a dual-integral polyhedral convex function. It is easy to see that
and that doniR,/^ = doniR,/ provided /^ > —oo. Figure 10.2 illustrates the
conjugate scaling of a univariate function / with a = 2.
We are now in the position to present the conjugate scaling algorithm. Initially
the algorithm finds a feasible flow £o and an integer-valued potential po satisfying
(10.59) and applies the conjugate scaling to the objective function rewritten as
320 Chapter 10. Algorithms
where we regard g as a member of £[Z —> R]. Recall the notation n = \V\.
if the initial potential po in step SO is computed from a shortest path on the graph
Gj = (V^Cf), as explained in section 10.4.4, where in the second term on the
right-hand side we consider only those x, u, v for which f ' ( x ; —\u + Xv) is finite. If
/ 6 .M[Z|R —> R], the second term can be bounded as
which implies
Bibliographical Notes
The tie-breaking rule (10.2) for M-convex function minimization, as well as Propo-
sition 10.2, is due to Murota [148]. Variants of steepest descent algorithms are re-
ported in Moriguchi-Murota-Shioura [133] with some computational results. Scal-
ing algorithms for M-convex function minimization, including the one described in
section 10.1.2, were considered first in [133], although the proposed algorithms run
in polynomial time only for a subclass of M-convex functions that are closed un-
der scaling. A polynomial-time scaling algorithm for general M-convex functions is
given by Tamura [197]. The domain reduction algorithm in section 10.1.3 is due to
Shioura [190] and its extension to quasi M-convex functions is observed in Murota-
Shioura [154]. The domain reduction scaling algorithm in section 10.1.4, with its
extension to quasi M-convex functions, is due to Shioura [192]. Minimizing an M-
convex function on {0, l}-vectors is equivalent to maximizing a matroid valuation,
for which a greedy algorithm of Dress-Wenzel [41] works; see also section 5.2.4 of
Murota [146].
322 Chapter 10. Algorithms
Application to
Mathematical Economics
323
324 Chapter 11. Application to Mathematical Economics
while negative numbers are interpreted as outputs from them. Given a price vector
p = (p(k) : k € K) 6 R^ of commodities, each producer (assumed to be male) in-
dependently schedules a production in order to maximize his profit, each consumer
(assumed to be female) independently schedules a consumption to maximize her
utility under the budget constraint, and all agents exchange commodities by buy-
ing or selling them through money. An important feature of this model is that the
independent agents take the price as granted; i.e., they assume that their individual
behaviors do not affect the price. Such an economy is called a competitive economy.
We assume that a producer I e L is described by his cost function GI : ZK —>
RU {+00}, whose value is expressed in units of money. He wishes to maximize the
profit (p,y) — Ci(y) in determining his production y = yi e ZK. This means that
yi is chosen from the supply set
where /3h '• RK —> R, and accordingly her schedule (x, m) = (x/j, mh) should belong
to her budget set
We assume that a consumer h is associated with a utility function Uh '• "ZK x R —>
RU {—00} that is quasi linear in money; namely,
11.1. Economic Model with Indivisible Commodities 325
with a function63 [//, : Zx —> Ru{—oo}. Consumer h maximizes Uh under the bud-
get constraint; that is, (x,m) — (xh,TTih) ls a solution to the following optimization
problem:
(see Fig. 11.1). Under the assumption that domUh is bounded0 and mh is suffi-
ciently large, we can take
63
In economic terminology, U^ is called the reservation value function, although we refer to it
as the utility function in this book.
64
The boundedness of dom I//, is a natural assumption because no one can consume an infi-
nite number of indivisible commodities. This assumption is also convenient for concentrating on
discreteness issues in our discussion.
326 Chapter 11. Application to Mathematical Economics
That is, each agent achieves what he or she wishes to achieve, the balance of supply
and demand holds, and an equilibrium price vector is nonnegative. Denoting the
total initial endowment of indivisible commodities by
where the right-hand side is a Minkowski sum in ZK. It is noted that money balance
Note 11.1. A special case of our economic model with L = 0, where no producers
are involved, is called the exchange economy. A difficulty of indivisible commodities
already arises in this case, as we will see in section 11.2. •
Note 11.2. Commodities that can be represented by real-valued vectors are called
divisible commodities. A framework for the rigorous mathematical treatment of
equilibria in economies of divisible commodities was established around 1960 usin
convexity, compactness, and fixed-point theorems as major mathematical tools. See
Debreu [37], [38], Nikaido [168], Arrow-Hahn [4], and McKenzie [128].
where domC/i = domC/2 = S (see Fig. 11.2). The demand correspondences D\ and
_D2, calculated according to (11.8), are also given in Fig. 11.2. For instance, for
328 Chapter 11. Application to Mathematical Economics
p = (p(l),p(2)) with 0 < p ( l ) < 1 and 0 < p(2) < 1, we have Di(p) = {(1,1)}; for
p = (1,1), we have D^p) = {(1,1), (0,1), (1,0)} and £>2(p) = {(1,1)}.
Given a total initial endowment x°, an equilibrium is a tuple (xi,x 2 ,p) e
Z2 x Z2 x R2. such that
For x° = (1, 2), for example, the tuple of x\ = (0,1), x? = (1,1), p = (2,1) satisfies
the above conditions and hence is an equilibrium.
Another case, x° = (1,1), is problematic. As we have seen in (11.15), a
nonnegative vector p is an equilibrium price if and only if x° € Di(p) + D^p).
Superposition of the diagrams for Di(p) and £>2(p) in Fig. 11.2 yields a similar
diagram for the Minkowski sum Di(p) + Dz(p), shown in Fig. 11.3. We see from
this diagram that no p satisfies (1,1) 6 Di(p) + D^p) and hence no equilibrium
exists for x° = (1,1). The diagram consists of eight regions, corresponding to the
eight points in [0, 2]z x [0, 2]z except {(1,1)}. Hence an equilibrium exists for every
11.2. Difficulty with Indivisibility 329
x° £ ([0,2] z x [0,2] z ) \ {(1,1)} and not for x° = (1,1). Let us have a closer look at
the problematic case to better understand the discreteness inherent in the problem
and to identify the source of the difficulty.
In view of the established mathematical framework for divisible commodi-
ties, we consider an embedding of our discrete problem via a concave extension of
the utility functions. Denote by Ui and U2 the concave extensions of U\ and U%,
respectively. Obviously, doniRt/i = dom.Rt/2 = S, where S = [0, I]R x [0, I]R, and
DI(P) + D2(p). On the other hand, the original discrete problem has an equilibrium
if and only if x° e Di(p) + D2(p), as noted already. For p = (3/2,3/2), we have
I>i(p) = {(0,1), (1,0)}, D2(p) = {(0, 0), (1,1)}, and
which has a hole at (1,1) (see Example 3.15 and Fig. 3.4). This hole is the very
reason for the nonexistence of an equilibrium for indivisible commodities.
where Xi is the ith unit vector and a maximum taken over an empty set is denned
to be — co. A more compact expression of this exchange property is
• Submodularity:
This says that special ascent directions work for increasing a utility function.
• (—M^-GSfZ]): Ifx £ argmax£/[-p+p 0 l], P < <?, Po < <?o, and argmax[7[-g +
qrol] ^ 0, there exists y e argmax/7[—q + </ol] such that
(i) y ( i ) > x(i) for every i £ K with p(i) = q(i) and
(ii) y(K)<x(K) if p0 = qo-
This is a version of the gross substitutes property. Recall the brief discussion
on the gross substitutes property at the beginning of section 6.8.
• (—M^-SWCSfZ]): For x € argmint/[-p], p <E RK, and i € K, at least one of
(i) and (ii) holds true:
(i) x e argmaxC/[—p — a\i] for any a > 0,
(ii) there exist a > 0 and y e argmaxt/[—p - a\i] such that y(i) = x(i) — 1
and y(j) > x(j) for all j € K \ {i}.
This is another version of the gross substitutes property, called the stepwise
gross substitutes property.
• M^-convexity of maximizers: For every p € R^, argmax[/[—p] is an M13-
convex set.
The properties above are essential features of M^-concave functions and, in
fact, the last four characterize M^-concavity, as follows (reformulations of Theorems
6.24, 6.34, 6.36, and 6.30). Note that submodularity and concave extensibility (even
together) do not imply M^-concavity.
Let us consider the special case of set functions to discuss the relationship of
the above theorems to the results of Kelso-Crawford [111] and Gul-Stacchetti [84].
The gross substitutes property was denned originally for a set function U : 2K —> R
in [111]. It reads as follows:
(GS) If X e argmaxt/"[-p] and p < q, there exists Y e argmaxC/[—q]
such that {i 6 X \ p ( i ) = q(i)} C Y,
where
Two novel properties introduce are introduced in [84]; i.e., the single improvement
property:
for a set function U. It is also observed that (NC) is implied by the strong no
complementarities property:
To derive these results from our previous theorems, we identify a set function
U : 2K -> R with a function U : ZK -> R U {-oo}, where dom U = {0,1} K , by
When translated for U, the properties (GS) and (SI) for U may be written respec-
tively as follows:
65
To be precise, this equivalence is shown in [84] for a monotone nondecreasing U, where p and
q can be restricted to nonnegative vectors.
11.3. M^-Concave Utility Functions 333
from the multiple exchange axiom for a matroid (see Theorem 4.3.1 in Kung [117])
and Theorem 11.7. Thus, the three conditions (GS), (SI), and (NC) for U are all
equivalent to the M^-concavity of U. Obviously, the special case of (SNC) with
|/| = 1 coincides with the exchange axiom (—M I '-EXC[Z]) for U and, therefore,
Note 11.9. We mention here some examples of M^-concave utility functions treated
explicitly or implicitly in the literature of indivisible commodities. See also the
examples of M^-convex functions shown in section 6.3.
(1) A separable concave function
using the notations for subdifferentials defined in (8.19) and (6.86). Hence, p G R+
is an equilibrium price if and only if it satisfies (11.22) for some x^ G domC/h,
(h G H) and yi G domC1; (I G L) meeting the supply-demand balance (11.14) for
the given total initial endowment x°. Furthermore, if ((xh h G H), (yi I G L),p)
is an equilibrium, the set P*(x°) of all equilibrium prices for x° is expressed as
11.4. Existence of Equilibria 335
This is the integer infimal convolution (6.43) of producers' cost functions and the
negatives of consumers' utility functions. Owing to our assumptions that
• dom Uh is bounded for each h £ H, and
• domC; is bounded for each / € L,
the infimum in (11.24) is attained for each z in
Proposition 11.10. For a total initial endowment x°, the following statements
hold true under the boundedness assumption on dom Uh for h e H and dom C\ for
l&L.
(1) There exists an equilibrium if and only if x° G dom* and (—dR^(x°)) D
R*^0.
(2)p'(x°) = (-dR<Sf(x°»nnx.
(3) I f ( ( x h h e H), (yi | I e L)) satisfies (11.9), (11.10), and (11.14) for some
p (not necessarily nonnegative), then
and ((XH | h e H), (yi \ I e L),p') is an equilibrium for any p' e (—0R\l/(a;°)) nR^.
which shows
Therefore, we have
336 Chapter 11. Application to Mathematical Economics
Figure 11.4. Aggregate cost function 'f and its convex closure fy for an
exchange economy with no equilibrium.
We shall present three theorems on the existence of equilibria. The first is for
an exchange economy (with L = 0).
Putting z = x° + axi and letting a —> +00, we see that p(i) > 0.
We next consider the existence of equilibria in the general case with consumers
and producers. To separate discreteness issues from topological ones, we consider
an embedding of our discrete model in a continuous space, just as we have done for
the simple exchange economy in section 11.2.
Utility functions Uh, being M^-concave, are extensible to concave functions
Uh '• Rx -> RU{-oo}. Similarly, cost functions Ci, being M^-convex, are extensible
to convex functions Ci : RK —» RU {+00}. Then demand correspondences Dh for
h £ H and supply correspondences Si for I £ L are defined, respectively, as
Note that Dh(p) and §i(p) coincide respectively with the convex hulls Dh(p) and
Si(p) of Dh(p) and Si(p). We define an equilibrium in the continuous model as a
338 Chapter 11. Application to Mathematical Economics
as well as the supply-demand balance (11.14) and the price nonnegativity (11.12).
The following theorems state that the existence of an equilibrium in the con-
tinuous model implies that of an equilibrium for indivisible commodities under
our assumption of M^-concavity/convexity. We emphasize here that this is by no
means a general phenomenon, as we have seen in section 11.2, but is peculiar to M^-
concavity/convexity. For topological issues, the reader is referred to the literature
in economics indicated in Note 11.2.
Theorem 11.14. Suppose that utility functions UH (h G H) are Afl -concave and
cost functions Ci (I € L) are Afl-convex. If the derived continuous economy has
an equilibrium satisfying (11.33), (11.34), (11.14), and (11.12) for a total initial
endowment x° e Z^, there exists an equilibrium of indivisible commodities for x°
satisfying (11.9), (11.10), (11.14), and (11.12).
Proof. By Theorem 11.7, Dh(p) and Si(p) are M^-convex sets if they are not empty.
Hence, the claim follows from Theorem 11.15 below. D
Theorem 11.15. Suppose that, for eachp € R+, demand sets DH(P) (h £ H) and
supply sets Si(p) (I € L) are Afl-convex if they are not empty. If the derived contin-
uous economy has an equilibrium satisfying (11.33), (11.34), (11.14), and (11.12)
for a total initial endowment x° e Z+, there exists an equilibrium of indivisible
commodities for x° satisfying (11.9), (11.10), (11.14), and (11.12).
On noting Dh(p) = Dh(p) and §i(p) = Si(p), as well as Theorem 4.12 and Propo-
sition 3.17 (4), we see
By Theorem 4.23 (3), on the other hand, J2h€H ^h(p)-^2ieL &i(p) *san M^-convex
set, which is hole free by Theorem 4.12. Hence follows
Note 11.17. Inspection of the proof of Theorem 11.15 shows that it relies solely
on the following two properties of M^-convex sets: (i) M^-convex sets are hole free
and (ii) Minkowski sums of M^-convex sets are M^-convex sets. This suggests a
generalization of Theorem 11.15 for a class J- of sets of integer points such that (i)
5 e T, x e ZK =>• S = 5 n ZK, x - S e T, and (ii) Si, 52 E f => Si + S2 e T
(cf. Proposition 3.16). Namely, if Dh(p) e T (h e H) and Si(p) e F (I £ L) for
each p e R^f, where we assume 0 € F, and if the derived continuous economy has
an equilibrium, then there exists an equilibrium for indivisible commodities.
Note 11.18. In an exchange economy with two agents, an equilibrium exists under
the L11-concavity of utility functions or the L1'-convexity of demand sets.
(1) If the utility functions C/i and C/2 are nondecreasing L^-concave functions,
an equilibrium exists for any x° & domC/i + dom[/2 (cf. Theorem 11.13).
(2) Suppose that the utility functions C/i and f/2 are L^-concave. If the derived
continuous economy has an equilibrium for x° 6 Z^f, there exists an equilibrium
for indivisible commodities (cf. Theorem 11.14).
(3) Suppose that, for each p e R+, the demand sets -Di(p) and D2(p) are iJ-
convex if they are not empty. If the derived continuous economy has an equilibrium
for x° & Z+, there exists an equilibrium for indivisible commodities (cf. Theorem
11.15).
Proof: (1) In the proof of Theorem 11.13, * is an L2-convex function and,
therefore, 5a*(x°) ^ 0 by Theorem 8.45.
(2) In the proof of Theorem 11.14, D\(p) and D2(p) are lAconvex or empty.
Hence the claim follows from (3) by Proposition 7.16.
(3) This follows from the proof of Theorem 11.15 by virtue of the convexity
in Minkowski sum for L^-convex sets (Theorem 5.8). •
340 Chapter 11. Application to Mathematical Economics
Obviously, we have
for the aggregate cost function ^ defined in (11.24). For the total initial endowment
z° e ZK, we put x° = (-x°(K),x°) € Z<°> UK .
The function \I>, being an integer infimal convolution of M-convex functions,
is also M-convex and can be evaluated by solving an MSFP2 (see Note 9.30). A
concrete description is given below.
The instance of MSFPa for the evaluation of ^(x°) is defined on a directed
bipartite graph G = (V+,V~,A) with vertex partition (V+,V~) and arc set A
given by
11.5. Computation of Equilibria 341
Note that V+, Vt+, and VJj" are copies of {0}(JK. Figure 11.5 illustrates the graph G
for H = {a, /?}, L = {A, B}, and K = {1,2}. For each arc a € A, we put c(a) = —oo,
c(a) = +00, and 7(0) = 0. Using the indicator function 5^°} '• %Ve —> {0, +00} of
{x°}, we define a function / : Zv+uv~ -> R U {+00} by
Proof. (1) is obvious from the definitions of Ci, Uh, and 6^°}- By the structure
of G, for each k G {0} U K, we have
66
The notation —d£\v- designates the restriction of —d£ to Vh . Hence x*h = — d(,\v- means
*J(fc) = -ae(fc)forfcev- h -.
342 Chapter 11. Application to Mathematical Economics
This means (2), since w* = x° by (1). Since 7(0) = 0, c(a) = —oo, and c(a) = +00
for any a e A, condition (i) of (POT) of Theorem 9.16 implies p(d+a) — p(d~a) = 0
for all a G A. This implies (3). It follows from (3) that
We also have
Then (4) follows from (ii) of (POT) in Theorem 9.16, (11.35), and (11.37).
Any algorithm for the MSFP2 will find a tuple ((x*h h e H), (yf I € L),p),
which gives an equilibrium for the total initial endowment x° if the optimal potential
p happens to be nonnegative. We go on to consider the set of all nonnegative optimal
potentials or, equivalently, the set of all equilibrium price vectors. We already know
from Theorem 11.16 that the set is an L^-convex polyhedron, and our objective here
is to derive a concrete description in terms of a linear inequality system. With this
knowledge, the existence of an equilibrium price vector can be checked by solving
11.5. Computation of Equilibria 343
in which we have
by (3.30). By Theorem 6.26 (2) (M-optimality criterion), we have y^ e arg min Ci [—p]
if and only if
By defining
we obtain a concrete representation of P*(x°) as follows. Note that l(j) < +00,
u(j) > — oo, and u(i,j] > —oo for any i,jEK(i^ j).
Theorem 11.21. The set P*(x°) of all equilibrium price vectors is an ifi-convex
polyhedron described as
Similarly, the smallest equilibrium price vector can be found by solving another
linear programming problem:
Both (11.44) and (11.45) can be easily reduced to the dual of a single-source shortest
path problem.
Theorem 11.22. There exists an equilibrium price vector if and only if the problem
(11.44) is feasible. The smallest and the largest equilibrium price vectors, if any,
can be found by solving the shortest path problem.
Bibliographical Notes
The unified framework for indivisible commodities by means of discrete convex
analysis is proposed in Danilov-Koshevoy-Murota [34], [35], to which Theorems
11.14 and 11.15 as well as Notes 11.9 and 11.17 are ascribed. Theorem 11.16 for
the structure of equilibrium prices and Note 11.18 are by Murota [147].
The gross substitutes property was introduced by Kelso-Crawford [111] and
investigated thoroughly by Gul-Stacchetti [84], in which the equivalence of (GS),
(SI), and (NC) is proved. The connection of these conditions to M^-concavity was
pointed out by Fujishige-Yang [69] for set functions, with subsequent generaliza-
tions by Danilov-Koshevoy-Lang [33] (Theorem 11.6) and Murota-Tamura [160]
(Theorems 11.4 and 11.5). See Roth-Sotomayor [180] for more on (GS).
11.5. Computation of Equilibria 345
Application to
Systems Analysis by
Mixed Matrices
347
348 Chapter 12. Application to Systems Analysis by Mixed Matrices
Example 12.1. Consider the electrical network in Fig. 12.1, which consists of
five elements: two resistors of resistances Ti (branch i) (i = 1,2), a voltage source
(branch 3) controlled by the voltage across branch 1, a current source (branch 4)
controlled by the current in branch 2, and an independent voltage source of voltage
e (branch 5). The element characteristics are represented as
where & and rn are the current in and the voltage across branch i (i = 1 , . . . , 5)
in the directions indicated in Fig. 12.1. We then obtain the following system of
equations:
12.1. Two Kinds of Numbers 349
The upper five equations are the structural equations (Kirchhoff's laws), while the
remaining five are the constitutive equations.
The nonzero coefficients, ±1, appearing in the structural equations represent
the incidence relation in the underlying graph and are certainly accurate in value.
The entries of —1 contained in the constitutive equations are also accurate by defi-
nition. In contrast, the values of the physical parameters ri, r2, a, and (3 are likely
to be inaccurate, being only approximately equal to their nominal values on account
of various kinds of noises and errors.
The unique solvability of this network amounts to the nonsingularity of the
coefficient matrix of (12.1). A direct calculation shows that the determinant of this
matrix is equal to r-2 + (1 — a)(l + /3)ri, which is highly probably distinct from zero
by the independence of the physical parameters {ri,r2,a,/?}. Thus, the electrical
network of this example is solvable in general or, more precisely, solvable generically
with respect to the parameter set {ri,r 2 ,a,/?}. The solvability of this system will
be treated in Example 12.11 by a systematic combinatorial method (without direct
computation of the determinant). •
Example 12.2. Consider a hypothetical system (Fig. 12.2) for the production
of ethylene dichloride (C2H4C12), which is slightly modified from an example used
in the Users' Manual of Generalized Interrelated Flow Simulation of The Service
Bureau Company.
Feeds to the system are 100 mol/h of pure chlorine (Ck) (stream 1) and 100
mol/h of pure ethylene (C2H 4 ) (stream 2). In the reactor, 90% of the input ethylene
is converted into ethylene dichloride according to the reaction formula
At the purification stage, the product ethylene dichloride is recovered and the unre-
acted chlorine and ethylene are separated for recycling. The degree of purification is
described in terms of the component recovery ratios 01, 02, and 03 of chlorine, ethy-
lene, and ethylene dichloride, respectively, which indicate the ratios of the amounts
recovered in stream 6 of the respective components over those in stream 5. We
consider the following problem:
350 Chapter 12. Application to Systems Analysis by Mixed Matrices
Given the component recovery ratios ai and a-2 of chlorine and ethylene,
determine the recovery ratio x = 03 of ethylene dichloride with which a
specified production rate y mol/h of ethylene dichloride is realized.
Let un, Ui2, and Ui3 mol/h be the component flow rates of chlorine, ethylene,
and ethylene dichloride in stream i, respectively. The system of equations to be
solved may be put in the following form, where u is an auxiliary variable in the
reactor and r (= 0.90) is the conversion ratio of ethylene:
Example 12.3. Consider the mechanical system in Fig. 12.4 consisting of two
masses m\ andTO2,two springs fci and fc2, and a damper /; u is the force exerted
12.1. Two Kinds of Numbers 351
We may regard {rni,m2, &i, &2, /} as independent system parameters and other
nonvanishing entries (i.e., ±1) of F, A, and B as fixed constants. The Laplace
transform67 of the equation (12.4) gives a frequency domain description
where
(M-Q) Q = (Qij) is a matrix over K and
(M-T) T = (Tij) is a matrix over F such that the set of its nonzero
entries is algebraically independent over K.
We usually assume
with
where
are mixed matrices with respect to (K, F). Also note that A(s) is a mixed matrix
with respect to (K(s), F ( s ) ) in spite of the occurrence of the variable s in both Q(s)
and T(s), where K(s) and F(s) denote the fields of rational functions in variable s
with coefficients from K and JF, respectively.
Mixed polynomial matrices are useful in dealing with linear time-invariant dy-
namical systems. The variable s here is primarily intended to denote the variable for
the Laplace transform for continuous-time systems, though it could be interpreted
as the variable for the z-transformes for discrete-time systems.
which the determinant does not vanish in jF. The row set and the column set of A
are denoted by R and C, respectively. For / C R and J C C, the submatrix of A
with row indices in / and column indices in J is designated by A[I, J].
We start with the nonsingularity of a mixed matrix.
with e(I, J) e {1, —1}. If A is nonsingular, we have det A^Q and hence det Q[I, J] •
det T[R \ /, C \ J} ^ 0 for some / and J. The converse is also true, since no
cancellation occurs among nonzero terms on the right-hand side by virtue of the
algebraic independence of the nonzero entries of T.
The right-hand side of the identity (12.9) is a maximization over all pairs
(/, J), the number of which is as large as 2' fl l + l c 'l, too large for an exhaustive
search for maximization. Fortunately, however, it is possible to design an efficient
algorithm to compute this maximum on the basis of the following facts:
• The function p ( I , J) = rankQ[7, J] can be evaluated easily by Gaussian elim-
ination.
• The function r(I, J) = rankTf/, J] can be evaluated easily by finding a max-
imum matching in a bipartite graph representing the nonzero pattern of T.
• The maximization can be converted, with the aid of Edmonds's intersection
theorem for matroids (a special case of Theorem 4.18), to the minimum of an
L''-convex function.
To state the main theorem of this section we need another function 7 : 2R x
2° -> Z denned by
Note that 7(7, J) represents the number of nonzero rows of the submatrix T[I, J}.
358 Chapter 12. Application to Systems Analysis by Mixed Matrices
Proof. (12.10) can be proved from (12.9) with the aid of Edmonds's intersection
theorem for matroids (a special case of Theorem 4.18) and (12.11) can be derived
from (12.10) using the formula
Theorem 12.9 (Konig-Egervary theorem for mixed matrices). For a mixed matrix
A = Q + T, there exist I C R and J C C such that
(i) |/| + |J| -rankQ[7, J] = \R\ + \C\ -rankA, and
(ii) rankT[J, J] =0.
Proof. It is easy to see that p(I(JJ) = p(R\I, J) +1/| is a submodular set function
on R U C (see (2.70)). This is equivalent to the L''-convexity of gp by Theorem 7.1,
and similarly for gT and <?7. D
We can rewrite the right-hand sides of (12.10) and (12.11) using these Lfi-
convex functions. Namely, we see
12.4. Degree of Determinant of Mixed Polynomial Matrices 359
where 1 e ZRuC. Note that both gp + gT and gp + #7 are lAconvex and therefore
(gp + gT)' and (gp + g^)9 are M^-convex. As for (12.12) we observe that
This shows that the right-hand side of (12.12) is the minimum of an lAconvex
function over an L^-convex set. The discrete convexity implicit in Theorem 12.8
is thus exposed. A concrete algorithmic procedure for computing the rank of A =
Q + T is described in section 4.2 of Murota [146].
Example 12.12. The generic solvability of the chemical process simulation prob-
lem in Example 12.2 is denied by Theorem 12.8 applied to the Jacobian matrix in
Fig. 12.3. In (12.12) the minimum is attained by
Note that 7(7, J) = 0, |J| = |J| = 10, and p ( I , J ) = 3, for which p ( I , J ) - \I\ -
\J\ + \R\ + \C\ = 3 - 10 - 10 + 16 + 16 = 15 < \R\ = \C\ = 16. This shows that t
Jacobian matrix in Fig. 12.3 is singular, and hence the simulation problem is not
solvable in general.
Theorem 12.13. For a square mixed polynomial matrix A(s) = Q(s) + T(s),
degdet^ = max{degdetQ[7, J]+degdetT[R\I,C\ J] \ \I\ = \ J\,I C R, J C C},
(12.14)
where both sides are equal to —oo if A is singular.
The inequality turns into an equality if the highest degree terms do not cancel one
another. The algebraic independence of the nonzero coefficients in T(s) ensures
this. D
The right-hand side of the identity (12.14) is a maximization over all pairs
(7, J), the number of which is as large as '2\R\+\C\, too large for an exhaustive search
for maximization. Fortunately, however, it is possible to compute this maximum
efficiently by reducing this maximization problem to the M-convex submodular flow
problem.
To see the connection to M-convexity, we define functions /Q,/T : ZRuC —>
Z U {+00} with dom/ Q ,dom/ T C {0,1} RUC by
Both fq and fa are M-convex functions. The right-hand side of (12.14) can now
be identified as the negative of an integer infimal convolution of these M-convex
functions. Namely,
12.4. Degree of Determinant of Mixed Polynomial Matrices 361
where 1 € ZRUC. This reveals the discrete convexity implicit in Theorem 12.13
and also shows an efficient way to compute the degree of determinant of a mixed
polynomial matrix A(s) = Q(s) + T(s), since the infimal convolution of M-convex
functions can be computed efficiently by solving an M-convex submodular flow
problem, as we have seen in Note 9.30. Such an algorithm is described in detail in
section 6.2 of Murota [146].
Example 12.14. The dynamical degree of the mechanical system in Example 12.3
can be computed by Theorem 12.13 applied to A(s) = Q(s) + T(s) in Example 12.5.
In (12.14) the maximum value of 4 is attained by I = {1,2,5,6} and J = {1,2,5,6}.
Hence the dynamical degree is equal to four. •
Bibliographical Notes
This chapter is largely based on Murota [146]. The observation on two kinds of num-
bers and the concept of mixed matrices are due to Murota-Iri [149], [150], in which
Theorem 12.7 is given. Theorem 12.8 is taken from [146]. The Konig-Egervary the-
orem for mixed matrices (Theorem 12.9) is due to Bapat [7] and Hartfiel-Loewy [86].
The connection between mixed polynomial matrices and M-convexity explained in
section 12.4 is due to Murota [143] and a related topic can be found in Iwata-Murota
[104].
Applications of matroid theory to electrical networks are fully expounded in
Iri [95] and Recski [175]. When gyrators are involved in electrical networks, a
generalization of mixed matrices to mixed skew-symmetric matrices is useful, as is
explained in section 7.3 of Murota [146]. See Geelen-Iwata [75] and Geelen-Iwata-
Murota [76] for recent results on mixed skew-symmetric matrices.
Matroid theory also finds applications in statics and scene analysis (Graver-
Servatius-Servatius [80], Recski [175], Sugihara [195], Whiteley [215], [216], [217]).
For planar truss structures, in particular, a necessary and sufficient condition for
generic (infinitesimal) rigidity can be expressed in terms of unions of graphic ma-
troids, where a matroid union is a special case of the Minkowski sum of two M-
convex sets. It is noted that the rigidity of a truss structure can be represented by
a rank condition on a matrix associated with the truss, but that this matrix does
not fall into the category of mixed matrices. Recent results on rigidity in nongeneric
cases are surveyed in Radics-Recski [174].
This page intentionally left blank
Bibliography
363
364 Bibliography
[216] W. Whiteley: Some matroids from discrete applied geometry, in: J. E. Bonin,
J. G. Oxley, and B. Servatius, eds., Matroid Theory, American Mathematical
Society, Providence, RI, 1996, 171-311. (Cited on p. 361)
[217] W. Whiteley: Rigidity and scene analysis, in: J. E. Goodman and J.
O'Rourke, eds., Handbook of Discrete and Computational Geometry, CRC
Press, Boca Raton, FL, 1997, 893-916. (Cited on p. 361)
[218] H. Whitney: On the abstract properties of linear dependence, American Jour-
nal of Mathematics, 57 (1935), 509-533. (Cited on pp. 5, 6, 8, 74)
[219] Z. Yang: Equilibrium in an exchange economy with multiple indivisible com-
modities and money, Journal of Mathematical Economics, 33 (2000), 353-365.
(Cited on p. 327)
[220] L. A. Zadeh and C. A. Desoer: Linear System Theory, McGraw-Hill, New
York, 1963. (Cited on pp. 351, 355)
[221] U. Zimmermann: Minimization of some nonlinear functions over polyma-
troidal network flows, Annals of Discrete Mathematics, 16 (1982), 287-309.
(Cited on p. 176)
[222] U. Zimmermann: Negative circuits for flows and submodular flows, Discrete
Applied Mathematics, 36 (1992), 179-189. (Cited on p. 313)
This page intentionally left blank
Index
379
380 Index
tension, 53
another convention, 253
terminal vertex, 52, 53
tight set, 108
transformation by network, 269
of flow type, 269
of potential type, 269
transitive, 107, 119
translation submodularity, 23, 44, 178
triangle inequality, 24, 122
two-stage algorithm, 310
unimodular
totally, 88
unique-min condition, 266
unit contraction, 45
unit demand preference, 334
univariate function, 10
discrete convex, 95
polyhedral convex, 80
utility function, 324