TOPICS ON PERFECT GRAPHS
annals of discrete mathematics General Editor
Peter L. HAMMER, Rutgers University, New Brunswick, NJ, U.S.A. Advisory Editors
C. BERGE, UniversitC de Paris, France M.A. HARRISON, University of California, Berkeley, CA, U.S.A. V. KLEE, University of Washington, Seattle, WA, U.S.A. J.H. VAN LINT, California Institute of Technology, Pasadena, CA, U.S.A. G.-C. ROTA, Massacbusetts Institute of Technology, Cambridge, MA, U.S.A.
NORTH-HOLLAND - AMSTERDAM
0
NEW YORK
0
OXFORD
NORTH-HOLLAND MATHEMATICS STUDIES
88
Annals of Discrete Mathematics (21) General Editor: Peter L. Hammer Rutgers University, New Brunswick, U.S.A.
Topics on Perfect Graphs Edited by
C. BERGE E. R. Combinatoire, Centre de Mathhatique Sociale, Paris, France
v.
CHVATAL
Department of Computer Science, McGill University, Montreal, Canada
1984 NORTH-HOLLAND - AMSTERDAM
0
NEW YORK
0
OXFORD
@ Elsevier Science Publishers B.V. 1984
All rights reserved. No part of this publication may be reproduced, stored in a retrieval sysiem or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without the prior permission of the copyright owner.
ISBN: 0 444 86587 X
Pitblisliers: ELSEVIER SCIENCE PUBLISHERS B.V. P. 0 . BOX 1991 1000 BZ AMSTERDAM T H E NETHERLANDS Sole distributors for the U . S .A . and Carinda: ELSEVIER SCIENCE PUBLISHING COMPANY, INC 52 VANDERBILTAVENUE NEWY0RK.N.Y. 10017 U.S.A.
Library of Cougres Cataloging in Publication Data Main entry under title: Topics on perfect graphs. (North-Holland mathematics studies; 88) (Annals of discrete mathematics; 21) 1. Perfect graphs. 1. Berge, Claude. 11. Chvatal, V. (Valclav) 111. Series. IV. Series: Annals of discrete mathematics; 21. QA166.16.T66 1983 51 1 ’ 5 8-3-2353 ISBN 0-444-86587-X (US.)
PRINTED IN THE NETHERLANDS
CONTENTS Introduction
vii
PART I. General results
1
C. BERGE,Minimax theorems for normal hypergraphs and balanced hypergraphs - A survey
3
J.-C. FOURNIER and M. LAS VERGNAS, A class of bichromatic hypergraphs
21
L. LOVASZ, Normal hypergraphs and the Weak Perfect Graph Conjecture
29
PART 11. Special classes of perfect graphs
43
C. BERGE,Diperfect graphs
45
C. BERGEand P. DUCHET,Strongly perfect graphs
57
V. CHVATAL, Perfectly ordered graphs
63
P. DUCHET,Classical perfect graphs
67
C. GRINSTEAD, The Perfect Graph Conjecture for toroidal graphs
97
W.-L. Hsu, The Perfect Graph Conjecture on special graphs - A survey
103
H. MEYNIEL, The graphs whose odd cycles have at least two chords
115
E. OLARUand H. SACHS,Contributions to a characterization of the structure of perfect graphs
121
G. RAVINDRA, Meyniel’s graphs are strongly perfect
145
A. TUCKER,The validity of the Perfect Graph Conjecture for &-free graphs
149
PART 111. Polyhedral point of view
159
R. GILES,L.E. TROTTER, Jr. and A. TUCKER, The Strong Perfect Graph 161 Theorem for a class of partitionable graphs M.W. PADBERG, A characterization of perfect matrices
169
vi
Conrents
PART IV. Which graphs are imperfect
179
R.G. BLAND,H.-C. HUANGand L.E. TROITER, JR. Graphical properties 181 related to minimal imperfection V. CHVATAL,An equivalent version of the Strong Perfect Graph 193 Conjecture
V. CHVATAL, R.L. GRAHAM,A.F. PEROLDand S.H. WHITESIDES, 197 Combinatorial designs related to the Perfect Graph Conjecture S.H. WHITESIDES,A classification of certain graphs with minimal 207 imperfection properties PART V. Which graphs are perfect
219
R.E. BIXBY,A composition for perfect graphs
22 1
M. BURLETand J. FONLUFT,Polynomial algorithm to recognize a Meyniel graph
225
M. BURLETand J.-P. UHRY,Parity graphs
253
V. CHVATAL, A semi-strong Perfect Graph Conjecture
279
S.H. WHITESIDES,A method for solving certain graph recognition and 28 1 optimization problems, with applications to perfect graphs PART VI. Optimizaiion in perfect graphs
299
M.C. GOLUMBIC, Algorithmic aspects of perfect graphs
301
M. GROTSCHEL, L. LovAsz and A. SCHWVER, Polynomial algorithms for perfect graphs
325
W.-L. Hsu and G.L. NEMHAUSER, Algorithms for maximum weight cliques, minimum weighted clique covers and minimum colorings of claw-free perfect graphs 357
INTRODUCTION Many challenging problems in graph theory involve at least one of the following four invariants: (i) the stability number a ( G ) (also called the independence number), defined as the largest number of pairwise nonadjacent vertices in G ; (ii) the clique covering number 8(G), defined as the least number of cliques which cover all the vertices of G ; (iii) the clique number w ( G ) , defined as the largest number of pairwise adjacent vertices in G ; (iv) the chromatic number y ( G ) (sometimes denoted also by x ( G ) ) ,defined as the least number of colors needed to color all the vertices in such a way that no two adjacent vertices have the same color. The inequality a ( G )S O(G)holds trivially for all graphs G : if k cliques cover all the vertices then no more than k vertices can be pairwise nonadjacent. (A similar observation shows that w ( G ) Sy ( G ) .In fact, if G denotes the complement of G then o ( G )= a (G) and y ( G )= 8(C?).) Graphs which satisfy this inequality with the equality sign played an important role in Claude Shannon’s 1956 paper concerning the ‘zero error capacity of a noisy channel’. In this paper, Shannon remarked that the smallest graph G with a ( G ) < 8(G) is G, the cycle of length five. It was Shannon’s work which motivated Claude Berge to make a conjecture (first presented at a graph theory meeting organized by Horst Sachs in Halle an der Saale in March 1960) concerning graphs with a ( G )= 8 ( G ) .This conjecture may be stated in many different ways which are easily seen to be equivalent. Three of them go as follows. (Sl) If a graph G has no induced subgraph isomorphic to either the chordless cycle C, whose length p is odd and at least five or the complement of such a cycle, then a ( G )= 8 ( G ) .
c,
(S2) The only minimal graphs G with a ( G )< 8 ( G )are the Cp’sand the cp’s with p odd and at least five. (S3) A graph G satisfies a ( G A )= f3(GA)for every set A of vertices (with GA standing for the subgraph of G induced by A ) if and only if no GA is isomorphic to a C, or a with p odd and at least five.
c,
An early effort by Alain Ghouila-Houri failed to produce a counterexample to this conjecture. Despite this encouraging sign, Berge felt that the conjecture might be too ambitious. Therefore he restricted himself to a weaker conjecture in the hope that it might be easier to settle. Again, this conjecture may be stated in many different ways which are easily seen to be equivalent; we list only three. vii
viii
Introduction
(W1) If a graph G satisfies a ( G A )= B(GA)for every set A of vertices then it satisfies y ( G ) = w(G). (W2) If a graph G satisfies y(GA) = w ( G A )for every set A of vertices then it satisfies a ( G ) = B(G). (W3) The class of graphs satisfying a(GA) = O(GA)for all A is closed under complementation. Clearly, the conjecture (W) is weaker than the conjecture (S). For this reason,
(S)and (W) became known as The Strong Perfect Graph Conjecture and The Weak Perfect Graph Conjecture, respectively. (Anticipating in this brief historical sketch, we note that the Weak Perfect Graph Conjecture was proved by Lovslsz in 1971. Nowadays, it is known as the Perfect Graph Theorem.) Yet another way of phrasing (W) is to say that ‘a graph is a-perfect if and only if it is y-perfect’, with ‘a-perfect’ and ‘y-perfect’ defined as ‘satisfying the hypotheses of (Wl) and (W2)’, respectively. (LovBsz’s proof of (W) made this terminology obsolete: since ‘a-perfect’ and ‘y-perfect’ are synonymous, both of them may be replaced by ‘perfect’.) The evolution of the theory of perfect graphs may be traced back to the first international meeting on graph theory held at Dobogoko (Hungary) in October 1959. At this meeting, A. Hajnal and J. SurBnyi presented an elegant result: Every triangulated graph G satisfies a ( G )= B(G).(An immediate corollary of this theorem states that every triangulated graph G is a-perfect.) Berge complemented this result by showing that every triangulated graph G satisfies y(G) = w(G). (Now the easy corollary states that every triangulated graph is y-perfect.) Berge also noticed that there are other interesting classes of graphs which are simultaneously a-perfect and y-perfect. For instance, the comparability graphs are a-perfect by virtue of Dilworth’s theorem and y-perfect by an easy ad hoc argument. Similarly, the line graphs of bipartite graphs are a-perfect and y-perfect by two different theorems of Konig. After the meeting at Halle an der Saale in 1960, the Strong Perfect Graph Conjecture received the enthusiastic support of G. HBjos and T. Gallai. In fact, Gallai provided further evidence in support of the conjecture by strengthening the results on triangulated graphs: he proved that a graph is a-perfect and y-perfect whenever each of its odd cycles of length at least five has at least two non-crossing chords. Nevertheless, Berge still felt that the weak conjecture was more promising. At a conference at Rand Corporation in the summer of 1961, he had fruitful discussions with Alan Hoffman, Ray Fulkerson and others. Later on, discussions between Alan Hoffman and Paul Gilmore led Gilmore to a rediscovery of the Strong Perfect Graph Conjecture and to an attempt to axiomatize the relevant
Introduction
ix
properties of cliques in perfect graphs. Ray Fulkerson attacked the weak conjecture from a linear programming point of view, which led to the development of his theory of ‘antiblocking polyhedra’. He proved that the conjecture was equivalent to another statement, which he found too strong to be true. For this reason, he concentrated his efforts on attempts to find a counterexample: even though he had reduced the conjecture to a certain ‘duplication lemma’, he missed its proof. (Later, when informed by Berge that the validity of the conjecture had just been established by Loviisz, he was able to supply the missing link independently in only a few hours.) Lovhsz’s beautiful proof of the Weak Perfect Graph Conjecture was found in 1971 independently of Fulkerson’s work. After more than twenty years, the Strong Perfect Graph Conjecture remains open. The question of its validity alone (or the problem of describing all minimal imperfect graphs) has become only secondary when compared with the important body of work stimulated by the conjecture over the years. Much of this work has an intrinsic interest independent of the Strong Perfect Graph Conjecture: it would not become obsolete even if the conjecture were proved. The purpose of this book is to present selected results on perfect graphs in a single volume. These are reprinted classical papers (sometimes with slight simplifications), survey papers written for this collection or new results. They concern different, and often overlapping, aspects of perfect graphs. We shall now comment on some of these aspects.
Part I. General Results When described by a reference to its cliques, a perfect graph becomes a normal hypergraph. LovBz’s proof of the Perfect Graph Theorem (1972) is given on pp. 29-42 of this volume in its original form, first in the context of hypergraph theory, and then with another characterization. A proof of Lovhsz’s theorem also appears in the article on pp. 3-19, together with various minimax equalities for normal hypergraphs - and, more specifically, for balanced hypergraphs, which motivated, since 1970,. the development of hypergraph theory. The elegant framework of hypergraph theory is also the setting of a work by Fournier and Las Vergnas (pp. 21-27), who proved in 1972 an interesting property of cliques in perfect graphs conjectured by Lovhsz.
Part II. Special Classes of Perfect Graphs Three classical examples (triangulated graphs, comparability graphs and line graphs of bipartite graphs) have been mentioned above. Results in this direction are surveyed in the papers by Duchet (pp. 67-96) and Golumbic (pp. 301-323).
Introduction
X
Another classical example is provided by the theorem of Gallai: A graph is perfect whenever each of its odd cycles of length at least five has two non-crossing chords. A companion theorem, in which ‘non-crossing’is replaced by ‘crossing’, was proved by Olaru and may be found in the paper by Olaru and Sachs (pp. 121-144). Both of these results were generalized by Meyniel in 1976 (see article this volume, pp. 115-119): A graph is perfect if each of its odd cycles of length at least five has two (or more than two) chords. Other special classes of perfect graphs may be obtained by forbidding, in addition to all the Cp’sand cp’s with p odd and at least five, an extra graph F. This has been done for F = K4 by Tucker (see pp. 149-157), and for F = K1,3by Parthasarathy and Ravindra (see articles on pp. 103-113 and pp. 161-167). The paper by Hsu (see pp. 103-113) generalizes various techniques used in this direction. As examples of ‘ad hod classes, we have included the paper of Grinstead (pp. 97-101), concerning toroidal graphs. In the first article in this volume, the reader will also find the main results about another class, the balanced graphs, whose perfectness follows from a theorem of Berge and Las Vergnas (1970): A graph is perfect whenever each odd cycle has an edge inducing only maximal cliques which contain three vertices of the cycle. Finally, a new class, the ‘strongly perfect’ graphs, is introduced by Berge and Duchet (pp. 57-61): In every induced subgraph of a strongly perfect graph, some stable set meets all the maximal cliques. This class includes the comparability graphs, the triangulated graphs, and the complements of triangulated graphs. Ravindra (pp. 145-148) and Chv6tal (pp. 63-65) have shown that it includes also the Meyniel graphs and the ‘perfectly orderable’ graphs, respectively.
Part III. Polyhedral Point of View With each graph G on n vertices, we may associate two polytopes P ( G ) and Q ( G ) .The first polytope P ( G ) is the convex hull of all the incidence vectors of stable sets in G (a stable set being a set of pairwise nonadjacent vertices). The second polytope is obtained by associating a variable x, with each vertex of G and then defining O(G)by the system of inequalities xu s 1 for each clique C, V E C
x, 3 0 for each vertex u.
Trivially, P ( G ) C Q ( G ) for every graph G. The unique character of perfect graphs is illuminated by the fact that P ( G )= Q ( G )if and only if G is perfect. This statement has been proved first by Fulkerson and then independently by Chvital; yet another proof may be found in Berge (pp. 3-19). Further results in
Introduction
xi
this direction have been found by Padberg (see pp. 169-178). Padberg’s results are used by Giles, Trotter and Tucker (pp. 161-167) to establish a class of graphs for which the Strong Perfect Graph Conjecture holds true.
Part IV. Which Graphs are Imperfect? If the Strong Perfect Graph Conjecture holds true, then the answer is ‘those containing a C, or a with p odd and at least five’; in any case, the question is equivalent to asking which graphs are minimal imperfect. A major breakthrough in this direction is due to Lovhsz (see article on pp. 29-42): Every minimal imperfect graph G with n uertices has n = (Y ( G ) w ( G )+ 1. This result led Bland, Huang and Trotter (pp. 181-192) to call a graph G partitionable if there are integers r, s greater than one and such that (i) G has precisely rs + 1 vertices; (ii) for each vertex u of G, the vertex-set of G - u can be partitioned into r disjoint cliques of size s and into s disjoint stable sets of size r. Bland, Huang and Trotter observed that a graph is imperfect if and only if it contains an induced partitionable subgraph. (The ‘only if’ part follows instantly from Lovisz’s theorem; to see the ‘if’ part, note that (ii) along with r,s 2 2 implies ( Y ( G ) =r, w ( G ) = s.) Jack Edmonds and Kathie Cameron (K. B, Cameron, Polyhedral and Algorithmic Ramifications of Antichains, Ph.D. Thesis, University of Waterloo, 1982) pointed out an immediate corollary of this observation: the class of imperfect graphs belongs to NP. Building up on Lovhsz’s theorem, Padberg (pp. 169-178) was able to establish additional properties of minimal imperfect graphs. Each of them (with n vertices) has precisely n stable sets S1,S2,.. . ,S, of size (Y = a ( G )and precisely n cliques C1,C,, . . . ,C, of size w = w ( G ) . Furthermore, each vertex is in precisely (Y stable sets S, and in precisely w cliques C,. Finally, S, n C, = 0 if and only if i = j (for some appropriate choice of indexing). For every choice of a and w greater than one, a graph satisfying Padberg’s conditions may be constructed by taking vertices u l , uz, . . . ,v, ( n = a w + 1) and making u, adjacent to u, if and only if 1 i - j 1 < w (with arithmetic modulo n ) . The resulting graph is denoted by CZil. A theorem found by ChvGtal (see pp. 193-195) shows that the Perfect Graph Conjecture may be restated as follows:
c,
(S4) Every minimal imperfect graph G has a spanning subgraph isomorphic to C:::l with (Y = (Y ( G ) and w = w ( G ) . Unfortunately, Padberg’s conditions may be satisfied by graphs radically different from CZ:l; Chvhtal, Graham, Perold and Whitesides (see pp. 197-206) described ways of constructing infinite families of graphs which, in spite of their
xii
Introduction
unwieldy structure, do satisfy Padberg’s conditions. Two of these graphs were found independently by Bland, Huang and Trotter (see pp. 181-192). The case of a = 4 and w = 3 is studied in detail by Whitesides (see pp. 207-218).
Part V. Which Graphs are Perfect? From the point of view of computational complexity, this question is definitely not just another way of asking which graphs are imperfect. T o this day, nobody has even guessed at a ‘certificate of perfection’ which could be attached to every perfect graph and whose validity could be checked in polynomial time. We believe that even just a correct guess at such a certificate (a companion to the Strong Perfect Graph Conjecture) would bring us a long way towards settling the Strong Perfect Graph Conjecture itself. Still, the question as to whether such a certificate exists at all is rarely asked. Analogous questions are answered satisfactorily for many special classes of perfect graphs; the details may be found in Golumbic (pp. 301-323). For instance, there is a polynomial-time algorithm which, given an arbitrary graph G, will find out whether G is a comparability graph or not. In fact, this algorithm will also furnish a certificate for its output: either a certain way of putting arrows on the edges of G, which certifies that G is a comparability graph, or a certain sequence of vertices in G, which certifies that G is not a comparability graph. A companion polynomial-time algorithm will accept any comparability graph along with its certificate (directions on edges), producing a largest clique and a minimum coloring as the output. It is tempting to speculate that the same pattern could be followed by a proof of the Strong Perfect Graph Conjecture. A polynomial-time algorithm, given any graph G, would produce either a certificate of perfection for G or one of the forbidden induced subgraphs (C, or with p odd and at least five in case the Strong Perfect Graph Conjecture is valid). Then a companion polynomial-time algorithm, given any perfect graph along with its certificate of perfection, would produce a largest clique and a minimum coloring. (Of course. the existence of the first of these two algorithms would imply that the class of perfect graphs belongs to P rather than merely to NP.) It is conceivable that every perfect graph can be built from ‘primitive’ perfect graphs by simple operations which preserve perfection. For instance, the role of the primitive perfect graphs could be played by comparability graphs or line graphs of bipartite graphs; the operations could consist of pasting two perfect graphs together along a clique or taking their ‘join’, as in the paper by Bixby (pp. 221-224). If the initial class of ‘primitive’ perfect graphs belonged to NP and if the operations could be carried out in a polynomial time, then the desired ‘certificate of perfection’ would follow immediately. This idea was put forth by Sue Whitesides in conversations with Chvgtal in the fall of 1977 (although it is so
c,
Introduction
...
Xlll
natural that we would not be surprised if others had proposed it earlier). in fact, it motivated her to design the algorithm for recognizing graphs with cliquecutsets, which is reproduced here (pp. 281-297). A systematic and steady progress towards answering the question ‘which graphs are perfect?’ is being made by the Grenoble School. First, Michel Burlet and Jean-Pierre Uhry (pp. 253-277) designed a polynomial-time algorithm to recognize the graphs whose perfection was established by Olaru (every odd cycle of length at least five has two crossing chords), then Burlet, and independently Whitesides, found an analogous result for the graphs of Gallai (every odd cycle of length at least five has two non-crossing chords) and quite recently Burlet and Fonlupt (pp. 225-252) designed a polynomial algorithm to recognize the Meyniel graphs (every odd cycle of length at least five has two chords). This last result implies that every Meyniel graph can be built from ‘primitive Meyniel graphs’ by an interesting new operation which preserves perfection. An intriguing sidelight to the question ‘which graphs are perfect?’ is provided by speculations about its possible interplay with the Perfect Graph Theorem: a certificate of perfection for G provides a certificate of perfection for the complement This point of view led Chvatal (pp. 279-290) to propose a ‘Semi-strong Perfect Graph Conjecture’ which suggests that a certificate of perfection could be formulated in terms of induced P4’s,and therefore apply automatically to a graph and its complement at the same time.
c.
Part VI. Optimization in Perfect Graphs The four invariants cu(G), 8(G), y(G) and w ( G ) are difficult to evaluate: each of the four problems of recognizing graphs G and integers k with a ( G ) ak, B(G)S k, w ( G ) > k and y ( G ) S k, respectively, is NP-complete. (In fact, the second and the fourth problems remain NP-complete even if the value of k is fixed, as long as it is at least three.) Nevertheless, polynomial-time algorithms for solving the four optimization problems (finding a largest set of pairwise nonadjacent vertices, etc.) in restricted classes of perfect graphs have been known for a long time. The classical ones, involving triangulated graphs and comparability graphs, are surveyed in Golumbic (pp. 301-323). Recent results of Burlet and Fonlupt, applying to the wide class of Meyniel graphs, are as yet unpublished and will appear elsewhere in a companion paper to that on pp. 225-252; nevertheless, the algorithms of Burlet and Uhry (see pp. 253-277) for solving the four optimization problems in the class of Olaru-Sachs graphs may be found in this volume. Hsu and Nemhauser (see pp. 357-369) treat the class of claw-free perfect graphs. A fundamental result in this direction is due to Grotschel, Lovasz and Schrijver (see pp. 325-356), who designed polynomial-time algorithms for
xiv
Introduction
solving the four optimization problems in arbitrary perfect graphs. Their algorithms are ingenious variations on the celebrated ‘ellipsoid method’ for linear programming, and therefore very much unlike the typical combinatorial optimization procedures. Can they be replaced by polynomial-time algorithms of a more transparent combinatorial nature?
PART I
GENERAL RESULTS
This Page Intentionally Left Blank
Annals of Discrete Mathematics 21 (1984) 3-19 @ Elsevier Science Publishers B.V.
MINIMAX THEOREMS FOR NORMAL HYPERGRAPHS AND BALANCED HYPERGRAPHS - A SURVEY C . BERGE C.N. R.S. Paris
1. Introduction
By definition, a minimax property can be written as follows: “the minimum of something is equal to the maximum of something else”. For the perfect graph G, the properties “ y ( G )= u(G)” and “cu(G) = 6(G)”are of that type. In fact, the perfect graphs have several other minimax properties which are usually stated in the terminology of “normal hypergraphs”. In 1969, we introduced a special class of perfect graphs defined as follows: for every odd cycle (el, e2,. . .,elk+,), and every sequence of distinct maximal cliques (C,,C,,. . . , Czk+I), where Cicontains the two end-points of e i , at least one of the C,’s contains three vertices of the cycle. This concept, also called “balanced hypergraph”, has been of some importance for the theory of linear programming in integers and generalizes the totally unimodular (0,1)-matrices. This paper is intended to survey and to complete the collection of minimax properties for normal hypergraphs and balanced hypergraphs. We shall remain in the context of Hypergraph Theory, but the reader can easily translate all the results in terms of graphs. In Section 2, we recall the definitions of Hypergraph Theory which are related to packing problems. In Section 3, we give the main minimax equalities for normal hypergraphs in only one theorem; a unified proof is given for Lovkz’s theorem [19], Chviital’s theorem [5],and other results. This section is mainly based on results of LOV~SZ, which give a more elegant presentation than the theory of antiblocking polyhedra (see [13]). In Section 4, we study the hypergraphs with the Menger property. These hypergraphs are not normal but have similar properties. In Section 5, we study the paranormal hypergraphs. In Section 6, we prove several minimax theorems concerning balanced hypergraphs. In particular we give an answer to a problem raised by Fulkerson, Hoffman and Oppenheim [14].
3
C. Berge
4
2. General definitions
A hypergraph H is a family (El,E2... . , E m )of non-empty subsets, called edges; U E , = X is the vertex-set, and H is often described by its incidence matrix, i.e., a (0, 1)-matrix A with m columns representing the edges and n rows representing the vertices. This matrix A has no 0-vector as a row or as a column. The rank of H is r ( H ) = max, I E, 1, and the anti-rank is s ( H )= min, I E, 1. The maximum degree A ( H ) is the maximum number of edges having a point in common. A partial hypergraph of H is a hypergraph H’ obtained from H by removing some of the edges (and the vertices which become isolated), or, equivalently, by removing some columns of the incidence matrix A, and the rows which become 0-vectors. The subhypergraph of H induced by a set S C X is the S, land by removing hypergraph Hs obtained by replacing each edge E, by E, f an edge E, if E, n S = 0. The dual hypergraph of H is the hypergraph H * defined by the transpose A * of the incidence matrix A. A set T C X is a transversal set of H if T meets all the edges; the family of all the minimal transversal sets is called the transversal hypergraph, and is denoted by TrH. min{ I TI T E TrH} is called the transuersal number, and is denoted by r ( H ) . A matching is a partial hypergraph H’ of maximum degree 1. The matching number v ( H ) is the maximum number of edges in a matching. If v ( H )= T ( N ) , the hypergraph H is said to have the Kiinigproperty. A transversal T can also be defined by its characteristic vector t = ( t l , tz, . . .,t ” ) , where t, = 1 if T 3 x,, and t, = O otherwise; such a vector t is a (0,l)-vector of the polytope
I
P
={t
I t E R ” , t 2 0, tA
1).
Similarly, a matching H‘ can be defined by a characteristic vector z = (z,, zl,.. ., z m ) ; such a vector z is a (0, 1)-vector of the polytope
O={z IzER”. 220, ArS1). Hence, P is called the transversal polytope and 0 is called the matchingpolytope. If p = ( p , ,p 2 , . . . ,p.) is a vector with non-negative integral coordinates, we define a p-matching as an integral vector of the polytope
0,= { z l z E R “ ,
z S O , Az S p f .
If 9 = (4,. 92,. . .,9,,,) is a vector with non-negative integral coordinates, we define a q-transversal as an integral vector of the polytope P,={tItER”, (30, tAsq}.
The maximum 9-value of a p-matching is denoted by v ( ~p.;4 ) = max{(q, z )
I z E N ” n 0,).
5
Minimax theorems
The minimum p-value of a q-transversal is denoted by
I
T ( H ;p , q ) = min{(p, t ) t E N " fl P,}. Clearly, T ( H ;1,1) = T ( H )and v(H; 1,l)= v(H). Proposition 2.1. We have
w ;P? 4 ) s
7 * ( H P, ;
4 )s
7w;P, q 1,
I
I
where 7 * ( H ; p , q )= max{(q, z ) z E Q,} = min{(p, t ) t E P,}. This follows immediately from the duality principle of linear programming.
3. Normal hypergraphs Let H be a hypergraph. Denote by & ( x ) the degree of a vertex x, put A ( H )= max dH( x ) , and denote by q ( H ) the chromatic index, that is, the least number of colors needed to color the edges so that no two intersecting edges have the same color. Clearly, q ( H )2 A ( H ) . H is called a normal hypergraph if every partial hypergraph H' of H satisfies q ( H ' )= A (H'). It is not difficult to see that H is a normal hypergraph if and only if its dual is the clique-hypergraph of a perfect graph. To study the properties of normal hypergraphs, the basic result is the following (LovBsz's) lemma: Lemma (Lovasz [19]). Let H = ( E 1 , E 2..., , E m ) be a hypergraph on X = { x l ,x 2 , . . . ,x " } , and let y = ( y l , y 2 , .. . ,y m )E N". Then the hypergraph obtained from H by multiplying each E, by y i is also normal.
Proof. Consider the hypergraph fi = (EI,E l , E2,E,, . . . , E m )where E ; = E l ; it suffices to show that q ( f i )= A (fi). Put q ( H )= A ( H ) = q. Case 1. The edge E l contains a vertex x with dH( x ) = A ( H ) .Then, A (I?) = q + 1, and
A (fi)< q (B)< q ( H )+ 1 = q
+ 1 = A (R).
So q ( f i ) = A ( f i ) and the proof is achieved. Case 2. The edge El contains no vertex x with d H ( x )= A ( H ) . Consider an optimal q-coloring of the edges of H. Let (1) be the color received by the edge El. Let H I be the family of edges of H having color (1) which are different from El.Each vertex x with d H ( x )= A ( H )belongs to an edge of HI,so A ( H - Hi) = 4 - 1. Since H is normal, q ( H - HI)= q - 1; therefore, we can color with q - 1
C. Beige
6
colors the edges in H - H I , and with one new color for H I +EI, we obtain a q-coloring of H. SO q ( H ) Sq
=A
Hence, q ( H )= A(I?).
( f i ) sq ( H ) .
0
In order to include also the results of Chvatal [ 5 ] and of Fulkerson [ 131, we rephrase the theorem of Lovasz as follows: Theorem 3.1. Let H = (El,E 2 , .. . , E m )be a hypergraph on X with incidence matrix A. The following conditions are equivalent: (1) H is normal, i.e.. q ( H ' )= A ( H ' ) for every H' C H ; ( 2 ) every uertex of matching polytope Q = {y y E R m , y 3 0, Ay S 1) has (0, 1)-coordinates ; (3) every vertex of the matching polytope Q has integral coordinates ; (4) v(H;l,q)=7*(H;l,q)forevery qEN"; ( 5 ) v ( H ; I , q ) = ~ ( H ; l , qforevery ) qEN'"; (6) every partial hypergraph H' satisfies v ( H ' ) = T ( H ' ) .
I
Proof. (1) implies (2). Let z be a vertex of the polytope Q ;since z is determined by a system of equalities with integral coefficients, its coordinates are rational, and the.re exist integers k , p l , p 2,... ,pm s 0 SO that kz = ( p l , p z , . . . ,p m ) . In the hypergraph fi obtained from H by multiplying each edge E, by p i , we have
Hence, A ( H ) S k, and, by the lemma, q ( H ) S k. Thus there exists a k-coloring of the edges of fi, and each color A defines a matching GAS Put y:
=
i
1 if one copy of E, is in H A , 0 otherwise.
The vector y A = ( y ; , y:, . . . ,y i), with (0, 1)-coordinates, belongs to Q ; furthermore
Since z is a vertex of 0,
7
Minimax theorems
This shows that z is a vector with (0, 1)-coordinates.
(2) implies (3). Obvious.
I
(3) implies (4). Since max{(q, y ) y E Q } is reached by a vertex of the polytope Q, we have
(4) implies (5). Put
O,={z l z E Q , ( z , q ) = y ; ( Y , q ) ] . Since Q1is a face of the polytope Q, there exists a row-vector matrix A such that
a'l
of the incidence
z E Ql 3 ( u j l , z ) = 1. It follows from (4) that each matching of H with maximum q-values covers the vertex xjl. Put
qi - 1 if xi, E Ei,
q; =
qi
otherwise.
Thus
v(H;l,q')= v(H;l,q)-1.
As above, there exists a vertex x, of H and a vector q 2 = (q;,q:, . . . ,qL) such that v ( H ;1, q 2 )= v ( H ;1,q l ) - 1.
We continue to define a sequence a
= (xi,, x,,
. . . ,x j k ) until
we have
v(H;l,qk)=O. The vector t = (tl, tZ,. . . , t n ) , where tj is the number of appearances of x j in the sequence a, is a q-transversal of H such that " ( t , l ) = j = l tj = k = v ( H ; l , q ) . Hence t is a q-transversal of minimum value, and
4,I ) =
T(H;
V ( H ;1,q).
C.Berge
8
(5) implies (6). Let H' be a partial hypergraph of H, and put ql
={
The vector q
1 if E, E H', 0 otherwise.
= (ql, q 2 , . . ,q m ) satisfies
U(H;l,q)= V ( H ' ) ,
T(H;l,q)=T(H').
Thus (5) implies u ( H ' )= T ( H ' ) .
(6) implies (1). It suffices to show that a hypergraph H which satisfies ( 6 ) is such that q ( H )= A ( H ) . Let be a hypergraph whose vertices are the matchings of H, and where an edge E, denotes the set of all matchings of H containing E,. Clearly, E, 17 E, = 0 if and only if E, f l E, # 0. Since H satisfies (6), it has the Helly property, hence v ( H )= A ( H ) . Also,
9(H)= ~ ( f f ) , T ( G = ) q(H), d(H) = v ( H ) . Hence l? is normal; since we have already shown that (1) implies (6), we get v(z?)= T ( H ) , or, equivalently,
q (HI = A ( H ) . This achieves the proof.
0
Let G be a graph, and let A be the incidence matrix whose columns represent the vertices of G and whose rows represent the maximal cliques of G. Let S ( G ) be the set of the characteristic vectors of all the stable sets of G ; clearly, the convex hull [ S ( G ) ]of S ( G ) is contained in the matching polytope Q = {z z E R", z 3 0. Az s 1). Theorem 3.1 shows something more:
I
Corollary (Chvatal [5]). [ S ( G ) ]= Q if and only if G is a perfect graph.
Proof. Clearly, a vector s = (sl,s2,. . .,s,,) belongs to S ( G ) if and only if s is a (0, 1)-vector of Q. Consequently, the result follows from the equivalence between (1) and (2) in Theorem 3.1. 0 Another proof based on the powerful theory of antiblocking polyhedra [13] has been discovered by Fulkerson and can be found in [34].
Minimax theorems
9
4. The Menger property
A hypergraph H on X =(xI,x2, ..., x n } has the Menger property if u(H;p, 1) = 7 ( H ; p ,1) for every n-dimensional vector p with non-negative integral coordinates. Lemma. Let H = ( E l ,E 2 , .. . , E m ) be a hypergraph of order n, and let q E N". The following conditions are equivalent: (i) 7 * ( H ; p q, ) is a n integer for every p E N" ; (ii) ~ * ( H ; p , q ) = ~ ( H ; p ,for q )everypEN". Proof. Clearly, (ii) implies (i). Now we shall show that (i) implies (ii). Let H be a hypergraph with the property (i) and with incidence matrix A = ((a:)); let X be a vertex of the polytope
P, = { x I x E R " , x ~ 0 xA , aq}. We write: X E Extr(P,). It suffices to show that all the coordinates of X are integral. Let us show, for instance, that XI is integral. Put
For j < n, put
dl
=
I
C a',+l
ifjEJ.
kEK
For every x E P,, the vector d
For x
{X
= X, this
= ( d , , d2,. . . , d , )
satisfies
yields
If x E P,, x # 2, the vector x cannot be in all of the bounding hyperplanes x, = o}, ;E J, Or {X (&, X ) = qk}, k E K, SO that
I
I
Let A be a positive integer; put d ( h ) = ( A d , + 1, Adz, Ad3,. . . ,Ad,).
C. Rerge
10
Clearly, the minimum of ( d ( A ) ,x ) for x E P, is obtained for at least one vertex of the polytope P,. say x ( A ) . Since the set Extr(P,) is finite, there exists a vertex 2 E Extr(P,) which satisfies = x ( A ) for an infinity of values of A. Since
+ A I,
( d, I )
(d,X)
+A
X,
for an infinity of A's, we get
( d , i) s ( d , x). From (2) and (3), it follows that = X. Thus the minimum of the linear function ( d ( A ) ,x ) = A(d, x ) + xI is obtained for x = X with an infinity of values of A. So for some integer h, we have
h(d, 2 ) + il= min (d(h), x ) = T * ( H ;d(h), 4 ) . X€PS
From (2) and (3), we obtain also
h(d, X) = min (Ah, x) = T * ( H ;id, 4). X€Pq So X, = T * ( H ; d ( h ) , q ) -T * ( H ; h d , q ) ,and XI is an integer.
0
Theorem 4.1 (Hoffman). Let H be a hypergraph. The following conditions are equivalenr (and characterize the Menger property): (i) v ( H ;p, 1) = T ( H ;p, 1) for euery p E N " ; (ii) v ( H ; p ,1) = T * ( H ; 1) ~ ,for every p EN".
Proof. (i) implies (ii) by Proposition 2.1, and (ii) implies (i) by the lemma. 0 To recognize the hypergraphs satisfying the Menger property consider a hypergraph H = (El,E 2 , .. . , E r n )on X = {xl,x2,. . . ,x,} and an integer p, 2 1. Define a hypergraph H ( x , , p , ) obtained from H by replacing x, by a set X = {x:, xf, . . . , x?} and by replacing each edge E, containing x, by p, new edges: (E, -{~,})U{.K;}, k
=
1,2,... ,pt.
For p, = 0, define H ( x , , p , )by removing from H all the edges containing x i . We shall refer to H ( x , , p , ) as the ''multiplication'' of the vertex x, by p , . For an integral vector p = ( p , ,p 2 , .. . ,p")3 0, let H ( p )be the hypergraph obtained from H by multiplying x1 by pl, and then x2 by p z , etc. Clearly,
T(H"')= T ( H ;p, l),
v(H"') = v ( H ; p ,1).
So, if X is a family of hypergraphs H satisfying v ( H )= T ( H ) , and if H E X
Minimax theorems
11
implies H@’EX, we know that every hypergraph in X satis,fies the Menger property. Example 1. Let G be a multigraph, and let a, b be two of its vertices. Consider the hypergraph H whose vertices are the edges of G, and whose edges are the sets of edges constituting an elementary chain connecting a and b in G. In fact, the multiplication of a vertex e, of H by pt is equivalent to considering a multigraph obtained from G by replacing the edge. e, by pr parallel edges (or removing e, if p, =O). Since the Menger theorem asserts that v ( H )= T ( H ) , the multiplication principle shows that H has the Menger property. Example 2 (Edmonds [6]). Generalizing Example 1, consider a set S of vertices in G, with I S 122, and a hypergraph H whose edges are the chains connecting in G two distinct vertices of S. A theorem of Edmonds [6] asserts that v ( H ) = T ( H ) ;therefore, H has the Menger property (by the multiplication principle). Example 3. Let G be a directed graph and let a be a vertex. Consider the hypergraph H whose vertices are the arcs of G, whose edges are spanning arborescences rooted in a. A theorem of Edmonds [7] asserts that v ( H )= 7 ( H ) (see an elegant proof by LovAsz [20]). Therefore H satisfies the Menger property (by the multiplication principle). Example 4 (Schrijver [30]). Generalizing Example 3 , consider a directed graph G = ( X , U ) ,and a directed graph K = (X, V )on X such that each pair of source and sink of K is connected by a directed path of K. The family (W,, W z ,. . .) of subsets W iof U such that K + W iis strongly connected (and which are minimal) constitute a hypergraph H on U, and Schrijver [30] has proved that v ( H ) = T ( H ) .Therefore H satisfies the Menger property. This result implies, by constructing appropriate graphs K, many known results, such as Menger’s Theorem for directed graphs, Gupta’s Theorem for bipartite graphs [15], Edmonds’ branching Theorem (Example 3), a special case of a conjecture of Edmonds and Giles [8] (also proved independently by D. Younger), and a theorem of Frank [lo]. Example 5 (Rothschild and Whinston [25]). Consider a multigraph G with only even degrees, and let a, a ‘ , b, b’ be four vertices. Consider the hypergraph H whose vertices are the edges of G, and whose edges are the chains linking a and a’, or b and b‘. Rothschild and Whinston 1251 have shown that H has the Menger property.
C. Beige
12
Example 6. Let G be a simple graph, and let a and b be two vertices. Consider the hypergraph H whose vertices are the vertices of G different from a, b, and whose edges are the chains linking a and b in G. By the theorem of Menger and by the multiplication principle, H satisfies the Menger property. An equivalent way to study the Menger property is to consider the transversal hypergraphs. For a hypergraph H on X = {xl,x2,. . . ,x n } , and for a non-negative integer-valued vector p = ( p , ,p 2 , .. . ,p"), let H P denote the hypergraph obtained from H by replacing each vertex x, by a set X , with IX, I = p,, and by replacing each edge E, by u { X , Ix, EE,}. Theorem 4.2. Let H be a hypergraph on X = { x I , x 2 ,..., x,,}, and let p ( p l , p 2 ,... ,p n )E N " . Then Tr[H'"]
= [Tr HI",
Tr[H"]
=
= [Tr HI"'.
The proof is easy. Let a ( H ) denote the maximum number of colors needed to color the vertices of H ss that every edge contains all the colors. Clearly cr(H)sminIE,I = s ( H ) . We shall say that H satisfies the Gupta property if a ( H P )= s(HP)for every p E N " . For instance, if G is a bipartite multigraph, the dual G * satisfies the Gupta property by a theorem of Gupta [ 151.
Theorem 4.3. A hypergraph satisfies the Gupta property if and only if its transversal hypergraph satisfies the Menger property. Proof. By Theorem 4.2, a ( ~= v~[ T r)( H P ) = ] V[(T~H)'~'],
s ( H ~=)T [ T ~ ( H =~ T) [] ( T ~ H ) " ' ] . Hence the result follows.
5. Paranormal hypergraphs A hypergraph H of order n is paranormal if 7 ( H ; p ,1)= 7 * ( H ; p ,1) for every n-dimensional vector p 3 0. By Proposition 2.1, every hypergraph with the
Minimax theorems
13
Menger property is paranormal. The converse is not true. For instance, the dual K : of the 4-clique K4 is paranormal, but v(K:)= 1 and T ( K $ = ) 2; hence K $ does not have the Menger property.
Proposition 5.1. Let H be a hypergraph on X = {x,,x 2 , . . . ,x,} with incidence matrix A. The following conditions are equivalent (and characterize the paranormal hypergaphs): (i) ~ ( H ; p , l ) = ~ * ( H ; pfor , l )a l l p E N " ; (ii) T * ( H ; ~1 ), is integral for all p E N " ; (iii) every vertex of the transversal polytope P = { t t E R ', t > 0, tA 3 1) has (0, 1)-coordinates.
1
Proof. This follows immediately from Theorem 4.1 and its lemma.
0
Several examples of paranormal hypergraphs arise in matroid theory (see Woodall [35]).Other examples follow from the theory of T-joins developed by Edmonds and Johnson [9] to generalize the famous Chinese Postman problem of Guan Megu. Let G = (X, E ) be a multigraph without loops, and let T be a non-empty subset of X . A T-join is a minimal subset F C E such that the partial graph G' = (X, F ) satisfies
I
{ x d&)=
1 modulo 2} = T.
A T-cut is an elementary cocycle w ( A ) , A C X, such that
/ A n TI = 1 modulo 2. It is easy to show that the transversal hypergraph of the T-join hypergraph is the T-cut hypergraph, and vice versa. Edmonds and Johnson have proved that every T-join hypergraph is paranormal (see also Seymour [31]). Example 1. Let G be a simple connected graph on X where I X I is even. An X-join of G is a spanning forest with only odd degrees. An X-cut is a cocycle w ( A ) where [ A I is odd. Example 2. Let a, b be two vertices of G. Let T = {a, b } . A T-join is an elementary chain connecting a and b. A T-cut is a cut between the vertices a and b.
I
Example 3. Let T = { x x E X, dG( x )3 1 modulo 2). A T-join is a minimal set of edges which need to be duplicated to get an eulerian graph. For A C X ,
Hence I o ( A ) Jis odd if and only if I T n A I is odd. So a T-cut is an elementary cocycle w ( A ) with an odd number of edges.
6. Balanced hypergraphs
Let H = ( E l ,E 2 , .. . ,E m )be a hypergraph with incidence matrix A. A cycle is a sequence (xI1,E,,,xQ,E,, . . .,xIk,E,,, x,J, where k 2 2, the x,,,’s are distinct vertices, the E p ’ s are distinct edges, and each edge Elpcontains the two vertices which are before and after it in the sequence. A hypergraph H is balanced if every cycle with an odd number of edges admits an edge containing at least three vertices of the sequence - or, equivalently, if the incidence matrix A does not contain an odd square submatrix of the the following type: 1 1 0 0 . .
. 0 0
0 0 0 0 . . . 1 1 1 0 0 0 . . . 0 1 We introduced this concept to extend theorems about hypergraphs with no odd cycles and totally unimodular matrices. Proposition 6.1. Let H be a balanced hypergraph ; then every partial hypergraph H’ is also balanced. Proof. See [l].
0
Proposition 6.2. Let H be a balanced hypergraph on X, and let S C X ; then the subhypergraph Hs induced by S i s also balanced. Proof. See [l]. Proposition 6.3. Let H be a balanced hypergraph, then its dual H * is also balanced. Proof. See [l].
0
Minimax theorems
Proposition 6.4. Let H be a balanced hypergraph on X is also balanced, p E N".
15 = {xl, x2,. .
. ,x"}; then H P
Proof. It suffices to prove the result for p = (0,1,1,. . . ,1) and for p = (2,1,1,. . . ,l). In the first case, H P = Ha, where A = X - {xl}, and the result follows from Proposition 6.2. In the second case, H P is obtained from H by replacing x , by a set X1= {xi, x;}. Let p be an odd cycle of H P which does not have an edge containing three vertices of the sequences. If p contains both xl and x;, then p = ( . .., z,E,xl, ...,xY) and the edge E which precedes x ; contains the vertices z, x i and x;, which is a contradiction. If p does not contain x;, say, then p induces on H an odd cycle with no edge containing three vertices, which is a contradiction. 0 Proposition 6.5. A hypergraph H is balanced if and only if every induced subhypergraph Hs is bicolorable (i.e., there exists a bipartition (S1, S2) of S such that every edge of Hs with more than one point meets both S1 and S,). Proof. See [l].
0
Proposition 6.6. Let H be a balanced hypergraph on X H @ )is also balanced, p E N " .
= { x l ,x2,. . . ,x,,};
then
Proof. It suffices to prove the result for p = (0,1,1,. . . , 1 ) and p = (2,1,1,. . . , 1). In the first case, the result follows from Proposition 6.1. In the second case, H(") is obtained from H by replacing x 1 by two additional vertices x I and xY, and each edge E containing x1 by two edges E ' = ( E -{xl})U{xl} and ( E - {XJ) u ( x 9 . Every bicoloring of H induces a bicoloring of H ( p (by ) giving the same color to x I and x:'), and every bicoloring of H @ in ) which x i and x; have the same color induces a bicoloring of H. So, if an induced subhypergraph K of H @ )has no bicoloring, then K contains both x i and x;' and K is induced by a subhypergraph Hs. Since Hs has a bicoloring, the contradiction follows. 0 El'=
Proposition 6.7. A hypergraph H is balanced if and only if q (Hk) = A (Hk) for every partial subhypergraph H;. Proof. See [l]. Proposition 6.8. A hypergraph H is balanced if and only if every induced subhypergraph Hs is normal.
C. Berge
16
Proof. This follows from Proposition 6.7 and the definition of normality.
0
Proposition 6.9. A hypergraph H is balanced if and only if every partial subhypergraph HL satisfies the Konig property
v(HL)= 7(Hk). Proof. This follows immediately from Proposition 6.8 and Theorem 3.1. 0 The first proof of this result was due to Berge and Las Vergnas [3], but a simpler proof was found later by LovBsz.
Proposition 6.10. A hypergraph H is balanced if and only if
v ( H s ; I , q ) = T ( H s ; l , q ) ( S C X , qEN"'). Proof. From Proposition 6.8 and Theorem 3.1. 0 Proposition 6.11. A hypergraph H is balanced if and only if
v ( H s ; l , q ) = ~ * ( H s ; l , q () S C X , q E N " ' ) . Proof. From Proposition 6.8 and Theorem 3.1. 0
Proposition 6.12. A hypergraph H is balanced i f and only i f every partial hypergraph H' has the Menger property. Proof. Let H be a balanced hypergraph on X = {xl,x2,. . . ,x.} and let p E N " . By Proposition 6.6, H'p' is balanced. By Proposition 6.9,
v ( H ;p. 1) = v(H"') = T ( H ( ~=)T) ( H ; 1~) ., So every partial hypergraph H' C H also satisfies v ( H ' ;p, 1 ) = T ( H ' ;p, 1). Hence H' has the Menger property. Conversely, let H be a hypergraph whose partial hypergraphs have the Menger property. Assume that H is not balanced; there exists an odd cycle, say (xl,El,xZ.E2,. . . ,X 2 k + l r EZk+l, x,), with no edge containing three vertices. Put H' = (El,E2,.. . , Ezk+Jand p, = 1 if i S 2 k + 1 , p , = + r: if 2 k + 1 < i 6 n. We have ~ ,= k + 1. v ( H ' ;p, 1) = k, T ( H ' ; 1) So H' does not have the Menger property, which is a contradiction.
0
The fact that every balanced hypergraph has the Menger property was first proved (by different methods) by Fulkerson, Hoffman and Oppenheim [14].
Minimax theorems
17
Proposition 6.13. A hypergraph H is balanced if and only if every partial hypergraph H’ is paranormal. Proof. Let H be a balanced hypergraph. By Proposition 6.12, H’ is paranormal. Now, let H be a hypergraph whose partial hypergraphs are paranormal. There exists an odd cycle (XI, E l , . . . ,EZk+1,XI) with no E, containing three x,’s. Define (as above) H’ and p. Then T ( H ‘ p, ; 1)=k
+ 1,
T * ( H ‘ p, ; 1 ) = ( 2 k + 1)/2.
Hence by Proposition 5.1, H’ is not paranormal, which is a contradiction.
0
Proposition 6.14. A hypergraph is balanced if and only if every partial hypergraph has the Gupta property. Proof. As above, it suffices to show that a balanced hypergraph H of order n satisfies the Gupta property. Since H P is also balanced for every p E N ” (Proposition 6.4), it suffices to show that a ( H )= s ( H ) . Let k = min I E, 1, and let (SI,Sz,. . . ,s k ) be a partition of the vertex-set X into k classes. Denote by k ( i ) the number of classes which meet E,. If k ( i ) = k for all i, all the classes are transversal sets of H, and the proof is achieved; otherwise, there exists an index j S m with k ( j ) < k. Since k ( j )< k c ( E ,1, there exists a class S, such that
Is, n~,I==2. Furthermore, there exists a class S, such that
1 S, f l E, 1 = 0. The subhypergraph HspuSqis balanced, and therefore has a bicoloring (Sb, S:). Put S : = S, for r# p, q. The partition (Sl, S : , . . . ,S : ) determines as above new coefficients k ’ ( i ) , and
k ’ ( i ) * k ( i ) ( i S m ) ; k ’ ( j ) = k ( j ) + 1. With this method, it is always possible to improve a partition (Sr, S:, . . . , S ; ) until we have k ’ ( i )= k for all i ; then we have a partition of X into k = s ( H ) transversal sets, and the proof is achieved. 0
Proposition 6.15. Let H be a balanced hypergraph. Then the transversal hypergraph Tr H has the Menger property. Proof. This follows from Proposition 6.14 and Theorem 4.3. 0
C. Berge
18
Note that this result is an answer to one open problem raised by Fulkerson, Hoffman and Oppenheim [14], who found a balanced hypergraph whose transversal hypergraph is not balanced, and asked if such a transversal hypergraph has similar properties. The above results can be applied to the class of graphs defined in the introduction, and to many hypergraphs arising in graph theory. In particular, every hypergraph whose incident matrix is totally unimodular is balanced (because this class is characterized by the Ghouila-Houri property: every subhypergraph admits a bicoloring which splits every edge in two equal parts). For instance, the dual of a bipartite multigraph is balanced. Also, if T is an arborescence on X , and if H is a hypergraph on X whose edges E, are subsets of X such that TE,is a path, it is clear that H is unimodular and, consequently, balanced. Other balanced hypergraphs have been considered by Frank [12]. We can also summarize the above results by considering the property v ( H ; p , q ) = T ( H ; 4). ~ , This equality holds for every p E N " and every q E N " if and only if h is a unimodular hypergraph. If H is a balanced hypergraph, Proposition 6.10 states that the above equality holds for every p E N " with (1. + 1;)-coordinates, and every q E N". Also, Proposition 6.12 states that the above equality holds for every p E N", and every q E N " with (0, 1)-coordinates. But the above equality does not hold for every p E N" and every 4 E N". For instance, let H be a hypergraph with edges E , = {1,4}, E , = {2,4}, E , = {3,4}, E , = {1,2,3,4}.Let p = (2,2,2,3), q = (1,1,1,2). It is easy to see that t = ( I~ ,I 2 I, 2 I, ~ ) belongs to the q-transversal polytope of H, and z = (f,f,f,i)belongs to the p-matching polytope. Since ( p , 1 ) = (q,z ) = 4.5, we have (by Proposition 2.1) T * ( H ; q~ ), = 4.5. Hence V(ff;p,
4 ) = 4 < 4.5 = T * ( H ; p ,q ) < 5 = T ( H ; p ,4 ) .
We do not know what are all the pairs (p,q ) such that v ( H ;p , q) = T ( H ;p , q ) for every balanced hypergraph H with n vertices and m edges, i.e., if there is a common generalization to Proposition 6.10 and Proposition 6.12. References [I] C. Berge, Balanced matrices, Math. Program. 2 (1972) 19-31. [2] C. Berge. Balanced matrices and property G , Math. Program. Study 12 (1980) 16>175. (31 C . Berge, C.C. Chein, V. Chvital and C.S. Seow, Combinatorial properties of polyominoes, Combinatorica 1 (1981) 217-224. [4] C. Berge and M. Las Vergnas, Sur un thCortme du type Konig pour hypergraphes, Ann. N.Y. Acad. Sci. 175 (1970) 32-40. [S] V. Chvatal, On certain polytopes associated with graphs, J. Comb. Theory, Ser. R, 18 (1975) 13RlS4. [ 6 ] J. Edmonds, Submodular functions, matroids and certain polyhedra, Combinatorial Structures and their Applications (Gordon and Breach, 1969) 6%87.
Minimax theorems
19
[7] J. Edmonds, Edge disjoint branchings, Combinatorial Algorithms, New York, 1973, 91-96. [8] J. Edmonds and R. Giles, A min-max relation for submodular functions on graphs, Studies in Integer Programming, Ann. Discrete Math. 1 (1977) 185-204. [9] J. Edmonds and E.L. Johnson, Matching, Euler Tours and the Chinese Postman, Math. Program. 5 (1973) 88-124. [lo] A. Frank, Kernel systems of directed graphs, Acta Sci. Math. (Szeged) 4 (1979) 63-76. [ l l ] A. Frank, Covering branchings, mimeo, Janos Bolyai Institute, 1979. [12] A. Frank, On a class of balanced hypergraphs, mimeo, Research Inst. for Telecom., Budapest, 1979. [13] D.R. Fulkerson, Blocking and anti-blocking pairs of polyhedra, Math. Program. 1 (1071) 16&194. [I41 D.R. Fulkerson, A.J. Hoffman and R. Oppenheim, On balanced matrices, Math. Program. Study 1 (North-Holland, Amsterdam, 1974) 120-132. [15] R.P. Gupta, An edge-coloration theorem for bipartite graphs, Discrete Math. 23 (1978) 229-233. [16] A.J. Hoffman, A generalization of max-flow min-cut, Math. Program. 6 (1974) 352-359. [17] A. Lehman, On the width-length inequality, mimeo, 1965. [18] L. LovLsz, 2-matchings and 2-covers of hypergraphs, Actes Math. Acad. Sci. Hung. 26 (1975) 433444. [19] L. Lovisz, Normal hypergraphs and the perfect graph conjecture, Discrete Math. 2 (1972) 253-267 (this volume, pp. 29-42). [20] L. LovLsz, On two min-max theorems in graphs, J. Comb. Theory, Ser. B, 21 (1976)96103. [21] J. F. Maurras, Polytopes i sommets dans [O, l]“, These doctorat d’etat, Paris VII, 1976. [22] M.W. Padberg, On the facial structure of set packing polyhedra, Math. Program. 5 (1973) 199-215. [23] M.W. Padberg, Perfect zero-one matrices, Math. Program. 6 (1974) 18&196. [24] M.W. Padberg, Almost integral polyhedra related to certain combinatorial optimization problems, Linear Algebra & Appl. 15 (1976) 69-88. [25] B. Rothschild and A. Whinston, On 2-commodity network flows, Oper. Res. 14 (1966) 377-387. [26] M. Sakarovitch, Quasi-balanced matrices, Math. Program. 8 (1975) 382-386. [27] M. Sakarovitch, Sur quelques problemes d’optimisation combinatoire, These doctorat d’etat. Grenoble, 1975. [28] A. Schrijver, Fractional packing and covering, Packing and Covering (Sti. Math. Centrum Amsterdam, 1978) 175-248. [29] A. Schrijver and P.D. Seymour, A proof of total dual integrability of matching polyhedra, Math. Centrum report ZN, 1977. [30] A. Schrijver, Min-max relations for directed graphs, Report AE 21/80, University of Amsterdam, 1980. [31] P.D. Seymour, On multicolorings of cubic graphs and conjectures of Fulkerson and Tutte, mimeo, Oxford, 1977. [32] P.D. Seymour, The forbidden minors of binary clutters, J. London Math. Soc. 12 (1976) 356-360. [33] P.D. Seymour, Discrete Optimization, Lectures Notes, Univ. Oxford, 1977. [34] L. E. Trotter, Solution characteristics and algorithms for vertex packing problems, Thesis, Cornell University, 1973. [35] D.R. Woodall, Menger and Konig systems, Theory and Applications of Graphs; Springer Verlag Lecture Notes in Math. 642 (1978) 620-635.
This Page Intentionally Left Blank
Annals of Discrete Mathematics 21 (1984) 21-27 @ Elsevier Science Publishers B.V.
A CLASS OF BICHROMATIC HYPERGRAPHS Jean-Claude FOURNIER Universitt Paris-Val de Maine, U.E.R. de Sciences Economiques et de Gestion, 94210 La Varenne St-Hilaire, France
Michel LAS VERGNAS* Universitt Pierre el Marie Curie, U.E.R. 48 - Mathtmatiques, 7.5005 Paris, France
We give a sufficient condition for bichromatic hypergraphs in terms of properties of cycles. Application: The set of inclusion-maximal cliques of a perfect graph can be partitioned into two classes such that both classes are represented at every vertex contained in at least two inclusion-maximal cliques.
1. Bichromatic hypergraphs and perfect graphs
Let G be a graph with vertex-set V. A clique of G is a subset of V inducing a complete subgraph of G. We denote by K ( G ) the hypergraph with vertex-set V whose edges are inclusion-maximal cliques of G. We recall that a hypergraph H is normal ([l],[ 5 ] )if v ( H ’ ) =7 ( H ‘ )for all partial hypergraphs H‘ of H, where v(H’) is the maximal number of pairwise disjoint edges of H’ and T ( H ’ )is the minimal cardinality of a set of vertices meeting all edges of H ’ . A partial hypergraph H’ of H is constituted by a subset of the edge-set of H [l]. It is easily seen that a graph G is perfect if and only if K ( G ) * , the dual hypergraph of K ( G ) ,is normal. By definition of a dual hypergraph [l], vertices of K ( G ) * are inclusion-maximal cliques of G and edges of K ( G ) * are all sets K,(G),x E V, where Kx(G) is the set of inclusion-maximal cliques of G containing the vertex x. A hypergraph H = ( X , ‘8) is bichromatic if there is a partition X = B + R , called a bicoloring, such that every edge E E 8 with lE 122 meets both B and R. In relation with his work on perfect graphs Lovasz asked: Is every normal hypergraph bichromatic? [5]. We answered positively this question in the following more general form:
Theorem 1 [2]. A hypergraph with no odd cycles of maximal degree two is bichromatic. * C.N.R.S. 21
22
3.-C.Foumier, M. Las Vergnas
In a hypergraph a cycle (with length k ) is defined by a sequence ( x l .E l ,x2. E2,. . . ,X k , E k , x l )of pairwise distinct vertices x I ,xZ,. . . ,Xk and pairn E, for i = 2,3,. . . , k and wise distinct edges El,E z , .. . ,E, such that x, € E,-l xI E Ek f l El. A cycle is of maximal degree two if any three of its edges have an empty intersection. An odd cycle is a cycle with an odd length k 2 3 . Clearly a normal hypergraph contains no odd cycles of maximal degree two. Hence by Theorem 1 every normal hypergraph is bichromatic. Equivalently we have the following corollary:
Corollary. The set of inclusion-maximal cliques of a perfect graph can be partitioned into two classes such that both classes are represented at each vertex contained in at least two inclusion -maximal cliques. Theorem 1 can be equivalently stated: A hypergraph with no odd cycles of maximal degree two such that any two non-consective edges of the cycle are disjoint is bichromatic (see [2]). A conjecture due to Sterboul [6] proposes a strengthening of this form of Theorem 1:
Conjecture [6]. A hypergraph with no odd cycles ( x , , El,x Z ,E 2 , .. . ,x k , Ek,x , ) such that any two non-consecutive edges are disjoint and I E, f l E,+l1 = 1 for i = 1 , 2 , . . . ,k - 1 is bichromatic. Clearly a cycle with length > 3 such that any two non-consecutive edges are disjoint is of maximal degree two. O n the other hand a 3-cycle ( x , , El,xZ, EZ, x , , E 3 ,x l ) such that I El r l E,I = I E2 fl Ejl = 1 is also of maximal degree two. Hence Sterboul’s conjecture implies Theorem 1. We emphasize that in Sterboul’s conjecture the condition 1 El fl Ek I = 1 is not required. Actually there are non-bichromatic hypergraphs containing no odd cycles such that any two non-consecutive edges are disjoint and any two consecutive edges have exactly one common vertex (for example, the collection of ( n + 1)-subsets of a set with 2 n + 1 elements). In other words Sterboul’s conjecture is not true if we require I El fl Ek I = 1. In the present note we prove a theorem intermediate between Sterboul’s conjecture and Theorem 1. The main results were first published in French by the same authors in [3]. All hypergraphs considered in this paper are finite and without multiple edges.
2. A sufficient condition for bichromatic hypergraphs
Theorem 2. A hypergraph with no odd cycles
(XI,
EI,X Z , E2,.. . ,x k , Ek, X I )
Of
A class of bichromatic hypergraphs
maximal degree two such that IE, n E,+l\= 1 for i bichroma tic.
23 = 1,2,. . . , k
-
1, is
The following lemma immediately implies Theorem 2 by induction on the number of vertices:
Lemma. Let H = ( X , 8 ) be a hypergraph containing an edge EoE 8, I Eo1 3 2 , such that H \Eo = ( X , % \{E,,}) is bichromatic. Suppose further that there is a vertex z E Eo such that every odd cycle (xl, E l ,xz,E2,.. . , xk, Ek, x l ) with maximal degree two satisfying 1 E, n E,+I1 = 1 for i = 1,2,. . . ,k - 1 contains Eo, suy E l = Eo, and has x i , xz # z. Then H is bichromatic. Proof. Let X = B + R (blue and red) be a bicoloring co of H\Eo. We assume notations such that EoC B (if E , meets both B and R there is nothing to prove). The central idea of the proof is to construct inductively a sequence co,cl, . . . , c, of bipartitions of X , c,: X = B, + R,, by interchanging colors blue and red on a set meeting all edges of H monochromatic in c,-,. Thus at each step the edges not bicolored by c , - ~become bicolored by c,. The hypothesis of the lemma ensures that this algorithm does not cycle: Let 8, denote the set of edges monochromatic in c,. We will prove that the 8,’s are pairwise disjoint. Hence by finiteness %, = B for some p , i.e., c, is a bicoloring of H. More precisely: At the beginning go = (Eo}.We set To= ( z } and define c1 by B1= B \ Toand R I = R To.In general suppose c , - ~has been defined, i 3 2. We have E , g 8,-1(the set of edges monochromatic in c , - ~ ) ;this fundamental property will be proven inductively. Hence B, resp. R, meets all edges in 8,-l. For i even, resp. i odd, let T,-, be a set contained in R, resp. in B, meeting all edges in and minimal with respect to inclusion with these properties. We define c, by B, = B,_,+T-,, R, = R,-,\T,-, if i is even, and B, = B,-l\T,-,, + T,-I if i is odd. R, = We prove the following properties by induction on p 2 1: ( 1 . p ) gP is disjoint from go, ,..., 8p-i; ( 2 . p ) every edge in gp meets Tv-land is blue in c, (i.e., contained in B,,) for p even, and red for p odd; (3.p) T, is disjoint from Eo, T 1 , .., . Tp-l. The case p = 1 is immediate. Suppose p 3 2 and ( l . i ) , (2.i), (3.i) hold for i = 1 , 2 ,..., p - 1 .
+
Proof of (1.p). An edge E E 8,, 1 C i G p - 1, meets T,-I by (1.i) and meets also T, by definition of T,.Now by (3.1), . . . ,( 3 . p - 1) the sets Eo, T I , .. . , T+ are pairwise disjoint, hence all T, with an even index i < p - 1 are red in c, and all T with an odd index are blue. It follows that all edges in g1,g 2 ,..., are
J.-C. Fournier, M. Lm Vergnas
24
bicolored in c,. O n the other hand, Eo is also bicolored since Eo\T , is blue in c, and T , is red. Hence gP is disjoint from 8", 8,,. . . , by (1.p). Since E is Proof of (2.p). An edge E E 8 ' , is bicolored in monochromatic in c,,, necessarily E meets Tp-Iand its color is that of Tp-lin c,, hence (2.p).
Proof of ( 3 . p ) . It is easily seen that T, is disjoint from To,T I , .. . , Tp-l. Suppose for instance p even. The edges in E, are blue in c, by (2.p) and T,, T z , ... , Tp-2 are red by (3.1),. . . ,(3.p - 1). Hence T, does not meet T,, T2,.... Tp-2.On the other hand T, is contained in B and T I ,T 3 , .. . ,T,-, in R . The case p odd is analogous. We have thus to prove that T, n (&\ To)= 0. This is clear for p odd since Eo C B and T, C R in this case. Suppose p is even. We show that E,, fl E = 0 for all edges E E gP. Let E, E 8, and suppose for a contradiction that there is xo E E, f l ( E o \ { z } ) . Let x, E Tp-lf l E,. By the minimality of T,-, there is Ep-lE such that E,-, fl T,-, = {x,}. Since is red in c,-~ by ( 2 . p - I), and €, is blue in c, by (2.p) we have n E, Tp-I,hence E,-, n E, C EP-,n Tn-,= { x p } , and thus E,-, n E, = { x , } . Repeating this procedure with instead of E,, and inductively, we get a sequence E P , ~ P , E p -,,. l r.~. ,xI P = z, Eo with E, E g,, x, E 7'-, and E, n = {x,-J of pairwise distinct edges and pairwise distinct vertices, hence an odd cycle (xu, Eo,x I , E l , .. . ,x,, E,, xo). By the hypothesis of the lemma this cycle is not of maximal degree two there are three edges E,, El, E,, 0 s i < j < k G p, such that E, f l El n E, # @. Suppose i, j , k are chosen with this property such that k - i is minimal. Let y E E, n El n €5. We first show that y # x,+,, xlrZ,. . . ,xk. Suppose k is even. The edge E, being blue in c k by ( l . k ) , y is different from xl, x s r . ..,& - I which are red in c, by (3.1), . . . ,( 3 . k - 1). Suppose y = x,, s even, i + 1 S s S k. If s < k the cycle (xs,Es,xs+l,E,,,,. . . ,X k , & , x s ) is an odd cycle with maximal degree two (by the minimality of k - i ) contradicting the hypothesis of the lemma. Hence s = k . Necessarily i and j are odd - if i is even, i7 being blue in c, cannot contain xk which is red in c,. But then the cycle ( x k , E,,x , + ~E,+I,. , . . ,x,, El, & ) is odd with maximal degree two, contradicting the hypothesis of the lemma. The case k odd is similar. The indices i and j have necessarily different parities - otherwise the cycle (y, E,, x,+,,E,,,, . . . ,x,, E l , y ) is odd with maximal degree two, contradicting the hypothesis of the lemma. Similarly j and k have different parities. A final contradiction arises as follows: suppose i even, j odd, k even (the case
c
A class of bichromatic hypergraphs
25
i odd, j even, k odd is similar). Since E, is blue in c, and El n R is red in c,, we have E, n El C B. Similarly El n EkC R. Therefore E, n El fl El, = 0. 0 Theorem 2 implies Sterboul’s conjecture for hypergraphs with 2- or 3-element edges: Corollary. Let H be a hypergraph with rank S 3 containing no odd cycle (xl, El,x2, Ez,. . . ,X k , E k ,xl) such that any two non-consecutive edges are disjoint 1 = 1 for i = 1,2,. . . ,k - 1. Then H is bichromatic. and 1 E, n E,+I
Proof. We show that H satisfies the hypothesis of Theorem 2. Suppose, for a contradiction, that there is an odd cycle (xl, E l , x2, E 2 , .. . ,x k , Ek, x l ) with maximal degree two such that I E, n E,+,I = 1 for i = 1,2,. . . , k - 1. We consider such a cycle with minimal length k. By hypothesis there are two non-consecutive edges, say El and El, 3 S j S k - 1, such that I El f El l I # 0. Let y E El n El ; y is different from xl, x 2 , . . . , X I , since the maximal degree is two. As I El 1, IE, I s 3 necessarily El = { y , xl, x2}, El = {y, x,, x,+J and El n El = {y}. Hence one of the two cycles (xl,El, y, El, x,+I, EI+2,x,+2,. . . ,X k , Ek, X I ) and (y,E1,x2,E2,x,,...,x,, E,, y ) is odd and thus contradicts the minimality of k. 0 This proof still works if H contains a unique edge with more than three vertices. However, if H contains several edges with cardinality 3 4 the above proof cannot be used, as an odd cycle (xl, E l , x2,. . . ,x k , E k ,xl) with maximal degree two such that 1 Ein Ei+,I = 1 may not contain any odd cycle with this property such that, furthermore, non-consecutive edges are disjoint. An example is the hypergraph with edges 01, 067, 1289, 23, 34, 4589, 567.
3. Minimal non-bichromatic hypergraphs A hypergraph H = ( X , 8)is minimal non-bichromatic if H is non-bichromatic and H \ E is bichromatic for every E E 8. Clearly a hypergraph H = ( X , ‘i9) is bichromatic if and only if there is no partial hypergraph of H which is minimal non-bichromatic. Theorem 1 and Theorem 2 have the following equivalent statements as properties of minimal non-bichromatic hypergraphs:
Theorem 1. In a minimal non-bichromatic hypergraph there is an odd cycle with maximal degree two such that any two non-consecutive edges are disjoint. Theorem 2. In a minimal non-bichromatic hypergraph there is an odd cycle with
26
J.-C. Foumier, M. Las Vergnas
maximal degree two such that, with at most one exception, all intersections of two consecutive edges are of cardinality one. This form of Theorem 1 has the following strengthening:
Theorem 1' ([4]). Let H = ( X , a ) be a minimal non-bichromatic hypergraph. For every E E % and z E E there is an odd cycle ( x l , E l = E x 2 = z,E2,..., xk,Ek,xl) with maximal degree two such that any two non-consecutive edges are disjoint and
E ,n E~ = { z } . The corresponding statement for Theorem 2 does not hold: The hypergraph with edges 12, 179,235, 34,356, 45,679, 78,89 is minimal non-bichromatic. The edge 12 is not contained in any odd cycle with maximal degree two. We have the following theorem:
Theorem 3. In a minimal non-bichromatic hypergraph which is not a graph ( o d d )cycle there are at least two different odd cycles with maximal degree two such that, with at most one exception, all intersections of two consecutive edges are of cardinality one.
Proof. Let H be a minimal non-bichromatic hypergraph, different from a graph (odd) cycle. By Theorem 2, H contains an odd cycle ( x l ,El,x2,E?,. . . ,x k , Ek, x,) with maximal degree two such that IE, f l E,,, I = 1 for i = 1,2,. . . , k - 1. Suppose there is no other such cycle. If IElnEk1 3 2 we set E o = E l ; let z E El n Ek\ { x l } ,I f ] E lf l Ek I = 1, since H is not a graph there is an edge E, with IE, I > 3. We set En = E, ; let z E E, \{x,, x , + ~}Note . that z is in no other edge of the cycle (otherwise we could form a second odd cycle with the required properties). In both cases, for these choices of En and z the hypergraph H satisfies the hypothesis of the lemma of Theorem 2. Hence H is bichromatic, a contradiction. 0 The number two in Theorem 3 is best possible: in the above example of a minimal non-bichromatic hypergraph there are only two odd cycles with the required properties.
Remark. The proof of Theorem 2 is constructive and provides a polynomial algorithm for constructing either a bicoloring of a hypergraph or an odd cycle with maximal degree two such that, with at most one exception, all intersections of two consecutive edges are of cardinality one. A rough estimate of the complexity of this algorithm is in O ( m z n ' ) ,where m is the number of edges of the hypergraph and n its number of vertices.
A class of bichromaric hypergraphs
27
References [ I ] C . Berge, Graphes et hypergraphes, 26me i d . (Dunod, Paris, 1973). [2] J-C. Fournier et M. Las Vergnas, Une classe d’hypergraphes bichromatiques, Discrete Math. 2 (1972) 407-410. [3] J-C. Fournier et M. Las Vergnas, Une classe d’hypergraphes bichromatiques 11, Discrete Math. 7 (1974) 99-106. [4] M. Las Vergnas, Sur les hypergraphes bichromatiques, Hypergraph Seminar, C. Berge and D.K. Ray-Chaudhuri, eds., Springer Lecture Notes in Math. No. 411 (Springer, Berlin, 1974) 102-1 10. [5] L. Lovasz, Normal hypergraphs and the perfect graph conjecture, Discrete Math. 2 (1972) 253-267 (this volume, pp. 29-42). [6] F. Sterboul, Communication at the Graph Theory Seminar, Paris, 1973.
This Page Intentionally Left Blank
Annals of Discrete Mathematics 21 (1984) 29-42 Science Publishers B.V.
0 Elsevier
NORMAL HYPERGRAPHS AND THE WEAK PERFECT GRAPH CONJECTURE* L. LOVASZ Mathematical Institute, Eotvos L. University, H. 1088 Budapest, Hungary A hypergraph is called normal if the chromatic index of any partial hypergraph H' of it coincides with the maximum valency in H'. It is proved that a hypergraph is normal iff the maximum number of disjoint hyperedges coincides with the minimum number of vertices representing the hyperedges in each partial hypergraph of it. This theorem implies the following conjecture of Berge: The complement of a perfect graph is perfect. A new proof is given for a related theorem of Berge and Las Vergnas. Finally, the results are applied o n a problem of integer valued linear programming, slightly sharpening some results of Fulkerson.
Introduction Let G be a finite graph and let x ( G ) and w ( G )denote its chromatic number and the maximum number of vertices forming a clique in G, respectively. Obviously, (11
x ( G ) s@(GIThere are several classes of graphs such that
x ( G )= w ( G ) ,
(21
e.g., bipartite graphs, their line graphs and complements, interval graphs, transitively orientable graphs, etc. Obviously, relation (2) does not say too much about the structure of G ; e.g., adding a sufficiently large clique to an arbitrary graph, the arising graph satisfies (1). Berge [l, 21 has introduced the following concept: a graph is perfect ( y perfect) if the equality holds in (2) for every induced subgraph of it. The mentioned special classes of graphs have this property, since every induced subgraph of them belongs to the same class. H e formulated the following two conjectures in connection with this notion:
Conjecture 1. A graph is perfect if and only if neither itself nor its complement contains an odd circuit without diagonals. * Reprinted from Discrete Math. 2 (1972) 253-267. 29
30
L. Looas2
Conjecture 2. Let a ( G )denote the stability number of G, let 6(G) denote the minimum number of cliques which partition the set of all the vertices. A graph G is perfect if and only if a ( G ' )= 6(G') for any induced subgraph G' of G. This conjecture is an attempt to explain some similarities between the properties of the chromatic number and the stability number; his next conjecture is proved in the present paper, formulated as follows.
Perfect Graph Theorem. The coniplemeitt of
c1
perfect graph is perfect as well.
Obviously, thc second conjecture of Berge would follow from the first one. However, due to its simpler form, it has more interesting applications and has been more investigated. Partial results are due to Berge [ 3 ] , Berge and Las Vergnas [4], Sachs and OIaru (61. Fulkerson [S] reduced the problem to the following conjecture, using the theory of anti-blocking polyhedra:
Conjecture 3. Duplicating an arbitrary vertex of a perfect graph and joining the obtained two vertices by an edge, the arising graph is perfect. In 91 we prove a theorem which contains Conjecture 3 . Berge has observed that the perfect graph conjecture has an equivalent in hypergraph theory, interesting for its own sake too. The correspondence between graphs and hypergraphs is simple and enables us to translate proofs formulated in terms of graphs into proofs with hypergraphs, and conversely. In $2 we deduce the hypergraph version of the perfect graph theorem from the above-mentioned conjecture of Fulkerson; the proof is short and docs not use the theory of anti-blocking polyhedra. It could be formulated in terms of graphs as well; however, the hypergraph version shows the idea more clearly. I t should be pointed out that thus the proof consists of two steps and the more difficult second step was first carried out by Fulkerson. In $3, wc give a new proof of a related theorem of Berge. Finally, in $4 we give some formulations of the results in terms of linear programming. Most of them have been observed to be equivalent to the perfect graph theorem already proved by Fulkerson.
1.
Let G, H be two vertex-disjoint graphs and let x be a vertex of G. By substituting H for x we mean deleting x and joining every vertex of H to those vertices of G which have been adjacent with x.
Normal hypergraphs and the Weak Perfect Graph Conjecture
31
Theorem 1. Substituting perfect graphs for some vertices of a perfect graph the obtained graph is also perfect. Proof. We may assume that only one perfect graph H is substituted for a vertex x of a perfect graph G. Let G’ be the resulting graph. It is enough to show that
since for the induced subgraphs of G’, which arise by the same construction from perfect graphs, this follows similarly. We use induction on k = w ( G ‘ ) . For k = 1 the statement is obvious. Assume k > 1. It is enough to find a stable set T of G‘ meeting all k-element cliques, since then coloring these vertices by the same color and the remaining vertices by k - 1 other colors (which can be done by the induction hypothesis), we obtain a k-coloring of G’. Put m = w ( G ) , n = w ( H ) , and let p denote the maximum cardinality of a clique of G containing x. Then, obviously,
k =max{m,n+p-1). Consider an m-coloring of G and let K be the set of vertices having the same color as x. Let, further, L be a set of independent vertices of H meeting every n-element clique of H. Then T = L U ( K \{x}) is a stable set in G‘. Moreover, T intersects every k-element clique of G‘. Really, if C is a k-element clique of G’ and it meets H then, obviously, it contains an n-element clique of H and thus a vertex of L. On the other hand, if C does not meet H, then C must be an m-element clique of G, and thus C contains a vertex of K \ { x } . 0
As has been mentioned in the introduction, in view of Fulkerson’s results, the perfect graph theorem already follows from Theorem 1. However, to make the paper self-contained, we give a proof of the perfect graph conjecture (which seems to be different from that of Fulkerson).
2.
A hypergraph is a non-empty finite collection of non-empty finite sets called edges. The elements of edges are the vertices. Multiple edges are allowed, i.e., more (distinguished) edges may have the same set of vertices. The number of edges with the same vertices is called the multiplicity of them. The number of edges containing a given vertex is the degree of it. The maximum degree of vertices of a hypergraph H will be denoted by 8 ( H ) . A partial hypergraph of H is a hypergraph consisting of certain edges of H.
32
L. Loua'sz
The subhypergraph induced by a set X of vertices means the hypergraph H
[ x= { E n x 1 E EH, E nx~ra).
A partial subhypergraph is a subhypergraph of a partial hypergraph (or, equivalently, a partial hypergraph of a subhypergraph). The chromatic number x ( H ) of a hypergraph H is the least number of colors sufficient to color the vertices (so that every edge with more than one vertex has at least two vertices with different colors). The chromatic index q ( H ) of H is the least number of colors by which the edges can be colored so that edges with the same color are disjoint. Obviously,
4 (HI 2 8 (HI.
(4)
Let a hypergraph be called normal if the equality holds in (4) for every partial hypergraph of it. A set T of vertices of H is called a transversal if it meets every edge of H ; T ( H ) is the minimum cardinality of transversals. Denoting by v ( H ) the maximum number of pairwise disjoint edges of H, we obviously have U(H) T(H).
(5)
Let a hypergraph be called 7-normal if the equality holds in (5) for every partial hypergraph of it. A hypergraph is said to have the Helly property if any collection of edges whose intersection is empty contains two disjoint edges. It is easily seen that normal and T-normal hypergraphs have the Helly property. Given a hypergraph H, we can consider its edge-graph G ( H ) defined as follows: the vertices of G ( H )are the edges of H and two edges of H are joined iff they intersect. O n the other hand, for a given graph G we can construct a hypergraph H ( G ) by considering the maximal cliques of G (in the settheoretical sense) as vertices of H and, for any vertex x of G, the set of maximal cliques containing x, as an edge of H ( G ) . It is easily shown that if G has no multiple edges (which can be assumed throughout this paper) then G ( H ( G ) ) =G.
(6)
Furthermore, H ( G) always has the Helly property. It is easily seen that -
( G ( H )is the complement of G ( H ) ) .Moreover, if H has the Helly property then
Normal hypergraphs and the Weak Perfect Graph Conjecture
33
Hence by (6),
X ( G )= s ( H ( G ) ) , w ( G )= s ( H ( G ) ) ,
x ( G )= T ( H ( G ) ) ,
4) = u(H(G))7
(9)
for any graph G. Equalities (7), ( 8 ) and (9) imply the following theorem:
Theorem 2. Let H be a hypergraph with the Helly property. H is normal iff G ( H ) is perfect; G is perfect if H ( G ) is normal. H i s r-normal iff G ( H )is perfect; is perfect i f H ( G ) is 7-normal.
As a corollary to Theorems 1 and 2 we have the following theorem: Theorem 1‘. Multiplying some edges of a normal hypergraph, the obtained hypergraph is normal. Theorem 2 implies that the perfect graph theorem is equivalent to the following:
Theorem 3. A hypergraph is r-normal iff it is normal. Proof. Parts “if” and “only if” of this theorem are equivalent (by Theorem 2). Thus it is enough to show that if H is normal then 7 ( H )= v(H)7 since for the partial hypergraphs this follows similarly. We use induction on n = r ( H ) . For n = O the statement can be considered to be true. It is enough to find a vertex x with the property that the partial hypergraph H’ consisting of the edges not containing x has v ( H ‘ )< u ( H ) ,since then H’ has an ( n - 1)-element transversal T and then T U {x} is an n-element transversal of H’, showing that
T ( H )n ~= v ( H ) . Assume indirectly that for any vertex x there is a system F, of n disjoint edges not covering x. Let
Ho=
U Fx7 x
where the edges occurring in more F,’s are taken with multiplicity. Ho arises from H by removing and multiplying edges, hence by Theorem 1’ it is also normal, i.e., q (Ho)= s (Ho).
L. Lovrisz
34
But obviously H,, has n . m edges, where m is the number of vertices of H. Since there are at most n disjoint edges in Ho, we have q (Ho)2 m.
On the other hand, a given vertex x is covered by at most one edge of F, ( y = x ) and by no edge of F,. Hence
S ( H o ) sm a contradiction.
- 1,
0
3. A subhypergraph of a normal hypergraph is not always normal as shown, e.g., by the hypergraph
{ { a ,b, d } , { b ,c, d } , I a ,c, 41; here { a , 6, c} spans a non-normal subhypergraph. Hypergraphs with the property that every subhypergraph of them is normal are described in the following theorem. A hypergraph is balanced if no odd circuit occurs among its partial hypergraphs (an odd circuit is a hypergraph isomorphic with the hypergraph ((1.2). { 2 , 3 } ,. . . ,(2n.2n + 1),11,2n + 1))).
Theorem 4. The following statements are equivalent : (i) H is balanced; (ii) every subhypergraph of H has chromatic number 2; (iii) every subhypergraph of H is normal. Obviously, Theorem 3 gives more equivalent formulations of (iiii). The theorem is actually due to Berge [3]. In what follows, we are going to give a new proof for the non-trivial parts of it.
Proof of Theorem 4. (iii) 3 (i) being trivial, it is enough to show (i) 3 (ii) and (ii) (iii). (I) Assume that H is balanced, though it has subhypergraphs which are not 2-chromatic. Let Hobe such a subhypergraph with minimum number of vertices. Consider the graph G consisting of the two-element edges of Ho: every vertex of H,, is considered to be a vertex of G. Now G is connected. Really, if V ( G) = X U Y, X f l Y = 0, X,Yf 0, and no edge of G joins a vertex of X to a vertex of Y, then considering a 2-coloration of Ho X and one of Hc, Y (by the minimality of Ho such 2-colorations exist) these
+
1
I
Normal hypergraphs and the Weak Perfect Graph Conjecture
35
form together a 2-coloration of Ho,since every edge E of Ho with 1 E I > 1 has at least two points in one of X , Y, and then even in this part of it there are two vertices with different colors. Since H is balanced, G is obviously bipartite. Let G be colored by two colors. Since Ho cannot be colored by two colors, there is an edge E, with I E I > 1, of HI having only vertices of the same color. Let x, y E E, x # y. Since G is connected, there is a path P of G connecting x and y. We may assume that no further vertex of E belongs to P. Then the subhypergraph spanned by the vertices of P contains an odd circuit, a contradiction. (11) Now let H be a hypergraph with property (ii); we show it has property (iii). Obviously it is enough to show T ( H )= V ( H ) .
Let T ( H )= t and consider a minimal partial subhypergraph Ho of H with the property 7(Ho)= t. If we show that Ho consists of independent edges, we are ready. Suppose indirectly El : E2E Ho, x E El n E2. By the minimality of Ho, there is a ( t - 1)-element transversal T, of Ho\{E,},i = 1,2. Put Q = TIfl T,, R, = T , \ Q , S = R I U R 2 U { x } . Obviously, X E T , , hence I S I = 2 1 R , ( + l . Since Ho S is 2-chromatic by (ii), there are two disjoint subsets of S both meeting every edge E of Ho S with IE I > 1. One of them, say M, has at most [$I S I] = I R 11 elements. Now M U Q is a transversal of Ho. Indeed, if an edge E is not represented by Q then it meets both R 1and R, if E Z E, and meets RIP,and { x } if E = E, ; thus, I E n S 122, whence E is represented by M. But I M U Q I s I R I I + I Q I = t - l , a contradiction. 0
I
1
We conclude this section with the remark that bipartite graphs are, obviously, balanced (and thus normal). On the other hand, Theorem 4 shows that balanced hypergraphs have chromatic number 2. Recently, Las Vergnas and Fournier sharpened this statement and showed that normal hypergraphs have chromatic number 2.
4.
Let
be a (0, 1)-matrix, no row or column of which is the 0 vector, and consider the optimization programs
L. Lovdsz
36
yA 3 w v20 min y . 1 Ax 6 1 x 20
maxw.x where 1 denotes the vector
It is well-known that if x, y run through non-negative real vectors, (10) and (11) have a common optimum. But now we are interested in integer vector solutions. Let B be a (0, 1)-matrix such that (i) any column u of B satisfies Au C 1, (ii) every maximal (0, 1)-vector with this property is a column of B. Consider two further programs:
yB 3 w y 30 min y * 1
Bx s 1 x 20 rnaxw-x Theorem 5. Assume that the optimum of (10) ( = the optimum of (11)) is an integer for any (0, 1)-vector w. Then, for any non-negative integer vector w, each of (10)-(13) has an integer optimum and an integer solution vector. Remark. The greatest part of this theorem is formulated in Fulkerson [ 5 ] as a
consequence of the perfect graph conjecture and the theory of anti-blocking polyhedra. Proof of Theorem 5. (1) First we show that (11) has a solution vector with integral entries for any (0, 1)-vector wo. For let xo be a solution of it with the greatest possible number of 0’s. Put wn= ( w , , . . . , W k ) ,
xu = (XI) xk
Obviously, x:S
WO. We
show that xo is an integer vector.
Normal hypergraphs and the Weak Perfect Graph Conjecture
37
Assume indirectly 0 < x 1 < 1, say; then w 1 = 1. Put
w: =
{
1 if x,#Oand i > l , 0 otherwise,
and
w ’ = ( w I, . . ., w ;). Let x’ be a solution of (11) with w
=
w’, then
w ’ x ’ s w o x ’ s woxo and
w ’x‘ 3 w ’xo > W”X0 - 1. Hence, both w’x’ and woxo being integers,
w‘x‘ = W(,X’ = woxo, i.e., x’ is a solution of (11) with w = wo too, and has, obviously, more 0’s than xo has, a contradiction. ( 2 ) Now we prove that also (10) has an integer solution vector for any (0, 1)-vector w. Assume indirectly that there are (0, 1)-vectors w failing to have this property and let wo be one with minimum number of 1’s. Let y o be a solution of (10) with w = wo. Obviously, we may assume that y l s 1. Put WO
= ( w l , ...9
Wk)r
Y O
= ( y l , . . . ? yk),
yl
# 0,
say, and define a (0, 1)-vector w’ = ( w I,. . . ,w ; ) by
w, if a l , = 0 , w: =
0
otherwise.
We show first that yo is not a solution of (lo), with w = w‘. For let x’ be a solution of (11) with w = w’; we may assume x I T S w’. Then
or yo(l - A X ’ = ) 0, but this is impossible since both yo and 1 - A x ’ are non-negative and their first entries are y l and 3 1 a,,w: = 1 , i.e., the inner product is non-zero. Thus, considering a solution y’ = (y I,. . . ,y:) of (10) with w = w‘ we have
-x:=l
38
L. Lovdsz
y'.l
and these being integers, y" 1 c y * 1-1.
This implies w' # wo, i.e., by the minimality property of wo, y ' can be chosen to be an integer vector. Let y " = (1, y ; , . . ., Y:)?
then y"A
3
w
since
Since
+
y " - 1s 1 y ' . 1
yo' 1,
y " is an integer vector solution of (10). (3) Put
B = ( bii; *
* *
* * *
bik : ) brk
Let H be a hypergraph on vertices 1 , . . .,s; for any 1 S i s k it has an edge
E, = { j ; b , , = 1). Now H is normal. For consider a partial hypergraph H' of it; let
1 if Ei E H ' , 0 otherwise, wO=(wl, ..., wk).
Let xo, y o be integer solution vectors of (11) and (lo), respectively. Since AX,) 1,
there is a column u of B with x o s u by property (ii) of it. Then the vertex corresponding to u has degree wou 3 woxo in H ' , i.e.,
Normal hypergraphs and the Weak Perfect Graph Conjecture
39
S(H’)2 wuxn. On the other hand, associate a color with every 1 entry of yo. For a given edge E,, consider a 1 S j S r with yla,, S 0 and give the color associated with y, to E,. If El and E, have the same color, then there is a j with a,, = a, = 1, i.e., no column of B can have 1’s o n both the i t h and tth place by (i). Hence E,, E, are disjoint, i.e., the coloring defined above is a good one, showing that p ( H ’ ) yo. 1 = ~ 0 x 0 ,
whence p ( H ’ ) = s(H’). (4) Let now w o = ( w l , . . . , wk) be a (0, 1)-vector. Consider the partial hypergraph H’ consisting of those E,’s for which w, = 1. By Theorem 3 , T ( H ’ )=
V ( H ’ )= V ,
i.e., there are v columns u,,, . . . ,ulv of B such that every row corresponding to an edge of H’ has a 1 in at least one of them. Let
1 i f j = j , ,..., j ” , Y t = [
0 otherwise, yo = 0%.. ., ys). Then
yoB z= Wo,
yo 3 0, y o . 1 = v.
On the other hand, there are u rows b,,, . . . ,b,” of B such that they correspond to edges of H’ and every column has at most one 1 in them. Putting
x, =
[
1 i f i = i , ,..., i,,
0 otherwise,
xo = (XI,.. . ,X k )
we have
Bxo
1, xo 3 0,
woxo = V ,
showing that xo, yo are solution vectors of (12) and (13). ( 5 ) Finally, let w n = (w,, . . . , w k )be an arbitrary non-negative integer vector. We show that (10)-(13) have integer solution vectors. It is enough to deal with (10) and (11). Let us multiply the edge E, of H by w,, i = 1,. . . ,k ; let H‘ denote the arising hypergraph. Then 6 ( H ’ )= p (H’l
L. Lovasz
40
since by Theorem 1’ H’ is normal. Let j be a vertex with maximum valency in H‘ and u, the corresponding column of 23. Then Au, S 1,
u, 2 0 ,
and wou, = 8(H’).
(14)
On the other hand, let the edges of H‘ be colored by p = p ( H ’ ) colors. This means that there are p (0,1)-vectors u l , . . . ,up such that a l + . . . + up = w0 and Ax s 1, x 3 0 implies arx S 1 for any 1 S 1 S p. Hence there is a (0, 1)-vector yl by part (2) of the proof such that
yrA 3 a r , y l 3 0 ,
yl . 1 = 1.
Putting
y =y1+...+y,, this vector satisfies yA 2 w,,, y 2 0, y . 1
= p,
i.e., by (14) the theorem is proved.
0
Appendix. A characterization of perfect graphs Let
be the complement of G. We prove the following theorem:
Theorem. A graph G is perfect if and only if w ( G ’ ) w ( G ’ ) SIG’J
for every induced subgraph G’ of G.
Proof. Part “only if” is trivial. To prove part “if” we use induction on I G I. Thus we may assume that any proper induced subgraph of G, as well as its complement, is perfect. Let multiplication of a vertex x by h ( h S O ) mean substituting for it h independent vertices, joined to the same set of vertices as x. This notion is closely related to the notion of pluperfection, introduced by D. R. Fulkerson. (I) As a first step of the proof we show that if Go arises from G by multiplication of its vertices then Go satisfies (G& (Go)3 I coI .
Normal hypergraphs and the Weak Perfect Graph Conjecture
41
Assume this is not the case and consider a Go failing to have this property and with minimum number of vertices. Obviously, there is a vertex y of G which is multiplied by h 2 2; let yl, . . . ,yh be the corresponding vertices of Go. Then w ( ~-oyl)w(Go- YJ 2
1
-
1
by the minimality of Go; hence w ( G o ) = w (Go - yl) = p ,
w (Go) = w (Go- yl) = r
and
( Go(= pr + 1. Put G, = Go- {y,, . . . ,yh}. Then GI arises from G - y by multiplication of its vertices, hence by [1, Theorem 11, GI is perfect. Thus, GI can be covered by ~ ( ( 2 w, )( Gso )= r disjoint cliques of GI; let C , ,. . . ,C, be these cliques, I C I I 3I CZI 2 . . .2= 1 c,I. Obviously, k s r. Since 1 GII = 1 Go1 - h = p r + 1- h,
1 c11= . . . = 1 C,-h 1 = p. +I
Let G2 be the subgraph of Go induced by C1U
IG2(= ( r - h
* *
U C,-I,+~ U {y,}, then
+ 1)p + 1< I G I ;
thus, by the minimality of GO, w (G2)w (G2) 2
I GzI.
Since w GZ)s w (Go)= p , this implies w ( G z ) 2r - h + 2 .
Let F be a stable set of r - h + 2 vertices of G,; then l F f l C I S 1 (1s i s r - h + l), hence y l E F. This implies that F U { y 2 , .. . , y h } is stable in Go. On the other hand I F U { y ~ , . . . , y h } =l r + l > w ( G O ) ,
a contradiction. (11) We show that x ( G )= w ( G ) . It is enough to find a stable set F such that w (G - F ) < w (G) since then, by the induction hypothesis, G - F can be colored by w (G) - 1 colors and, adding F as a further one, we obtain a p (G)-coloring of G. Assume indirectly that G - F contains a o(G)-clique CF for any stable set F in G. Let, for x E G, h ( x ) denote the number of CF’scontaining x. Let GOarise from G by multiplying each x by h ( x ) . Then, by Part I above,
42
L. Loua'sz
I Go\.
w(Go)w(Go)a
On the other hand, obviously
where f denotes the number of all stable sets in Go, and
a contradiction.
0
References [ 1 ] C. Berge, Farbung von Graphen deren samtliche bzw, ungerade Kreise starr sind (Zusarnmenfassung), Wiss, 2. Martin Luther Univ. Halle Wittenberg, Math. Nat. Reihe (1961) 114. [2] C. Berge, Sur un conjecture relative au probleme des codes optimaux, Commun. 13ime Assemblte Gtn. URSI, Tokyo, 1962. * [3] C. Berge, Graphes et hypergraphes (Dunod, Paris, 1970). [4]C. Berge and M. Las Vergnas, Sur un theoreme du type Konig pour hypergraphes, Ann N.Y. Acad. Sci. 175 (1970) 32-40. [5] D.R. Fulkerson, Blocking and anti-blocking pairs of polyhedra, Math. Program 1(1971) 168-194. [6] H. Sachs, On the Berge conjecture concerning perfect graphs, Combinatorial Structures and their Applications (Gordon and Breach, New York, 1969) 377-384.
PART I1
SPECIAL CLASSES OF PERFECT GRAPHS
This Page Intentionally Left Blank
Annals of Discrete Mathematics 21 (1984) 45-56 @ Elsevier Science Publishers B.V.
DIPERFECT GRAPHS Claude BERGE C.N.R.S. Paris Dedicated to Tibor Gallai for his 70th Birthday Gallai and Milgram have shown that the vertices of a directed graph with stability number a ( G ) can be covered by exactly a ( G ) disjoint paths. However, the various proofs of this result do not imply the existence of a maximum stable set S and of a partition of the vertex set into paths p i , F ~. .,. , p, such that 1 p, fl S 1 = 1 for all i. Later, Gallai proved that in a directed graph, the maximum number of vertices in a path is at least equal to the chromatic number; here again, we do not know if there exists an optimal coloring ( S , , S,, . . . ,S , ) and a path p such that l p fl S, 1 = 1 for all i. In this paper we show that many directed graphs, like the perfect graphs, have stronger properties, such as (i) for every maximum stable set S there exists a partition of the vertex set into paths which meet the stable set in only one point, and (ii) for every optimal coloring there exists a path which meets each color class in only one point. This suggests several conjectures similar to the perfect graph conjecture.
1. Path partitions
Let G be a (directed) graph defined by a set X of vertices and a set U C X x X of arcs (directed edges). A path p will be defined as a sequence ( x l , x 2 , . . . ,x k ) of distinct vertices, such that ( x l , x 2 ) E U, . . . , ( X k - 1 , x k ) E u. Every path p = ( x 1 , x 2 , . .,x k ) defines a set that we denote by p = { x , , x 2 , . . . , x k } . A family M = {pl,p2,.. . } is a path partition of G if the pi are pairwise pi = X . vertex-disjoint, and Denote by a ( G )the stability number of G, that is, the maximum number of independent vertices in G. The purpose of this paper is to enlarge the field of applications of the following classical result:
u
Theorem of Gallai-Milgram. Every directed graph G satisfies min I M 1 < cr (G), where the minimum is taken over all path partitions M. In other words, one can always cover the vertex set with exactly a ( G )paths which are pairwise disjoint. The above theorem has many applications. Let us mention the following ones:
Application 1 (Theorem of RCdei). Every tournament (complete anti-symmetric graph) contains a path which meets each vertex exactly once. 45
C. Berge
46
This follows immediately, since a tournament G satisfies a (G) = 1, and therefore min I M I = 1. 0
Application 2 (Theorem of Dilworth). Every transitive graph G satisfies minIMI = a ( G ) . A graph G = (X, U ) is transitive if ( x , y ) E U, ( y , 2 ) E U implies (x, z ) E U. Let 8 ( G ) denote the least number of cliques needed to cover X. Then, by RCdei's Theorem, the cliques of G and the paths of G define the same sets, so a (G) s e(G) = min I M I s a (G).
The required equality follows. 0
Application 3 (Theorem of Erdos-Szekeres). Let n, p, q be positive integers with n > pq, and let u = (a,, az,. . .,a.) be a sequence of n distinct integers. Then there exists either a decreasing subsequence with more than p integers or an increasing subsequence with more than q integers. Let G = ( X , U ) b e a g r a p h w i t h X = { a l , a 2,..., a , } , a n d ( a , , a , ) E U i f i < j and a, < a,. If { p lp,2 , .. . } is a partition of X into a ( G ) paths,
So we have either a ( G )> p (and there exists a decreasing subsequence with more than p integers), or max I p I > q (and there exists an increasing subsequence with more than 4 integers). [7 Many other applications are possible, and show the usefulness of the Gallai-Milgram theorem. In this section we shall give a stronger statement which combines an idea of Las Vergnas [8] and an idea of Linial [lo]. The proof is essentially the same as in the Gallai-Milgram theorem, but we obtain as corollaries results which could not be obtained directly. For an arborescence H , the root is the vertex x with dH,(x)= 0, and a sink is a vertex y with d & ( y ) = 0. If y is a sink, the maximal path of HI leading to y which does not contain a vertex z with d L , ( z )2 2 is called a terminal branch of HI. Let H be an arborescence forest of G, that is, a partial graph of G whose connected components are arborescences HI,H,, . . . . We denote by R (H) the set of the roots of the arborescences Hi, and by S ( H ) the set of the sinks of these arborescences. So a vertex in R ( H )n S ( H ) is an isolated vertex of H. Theorem 1. Let Ho be an arborescence forest of G with R ( H o )= R,, and
Diperfecr graphs
47
S(HII) = So. For every arborescence forest H with R ( H )C Rtr, S ( H )C So and 1 S ( H ) (minimum, there exists a stable set which meets every terminal branch of H.
Proof. Assume that the result holds for all graphs of order less than n, and consider a graph G of order n. Let H be an arborescence forest of G with R ( H ) C RO, S ( H ) C So, IS(H)I minimum. If S ( H ) is stable, the theorem is proved. If S ( H ) is not stable, there exists in G an arc ( b , a ) connecting two vertices a and b in S ( H ) . We have a C R ( H ) ,because otherwise, H‘ = H + (b, a ) satisfies R ( H ’ )C R o , S ( H ’ ) = S ( H ) - { b } C SO, I S ( H ’ ) I < I S ( H ) I , a contradiction. So H has an arc incident to a, say ( a l , a ) . Furthermore, d L ( a l )= 1 , because otherwise a l has at least two descendents in H which belong to S ( H ) , and H‘ = H - (a,, a ) + (b, a ) satisfies R ( H ’ )C Rll, S ( H ’ ) cSo, IS(H’)I< IS(H)J,a contradiction. = X - { a } admits fi = HX as an The subgraph C? of G induced by arborescence forest with R ( f i )C Ro, S ( f i ) C( S o - { a } ) U {al}. Now we shall show that ( S ( f i ) I is minimum in G. Otherwise, there exists in G an arborescence forest H’with R ( H ’ ) C Ro, S ( H ‘ ) C ( S I I - { U } ) U ( UIIS} (, H ’ ) I S I S(fi)I- 1. The following cases can happen: Case 1. a l E S ( H ’ ) Then H“ = H’+ (al,a ) is an arborescence of G which satisfies R ( H “ )C R,,, S ( H ” ) C So, I S(H”)I = IS(H‘)I S I S ( f i ) J- 1 = IS(H)I - 1, a contradiction. Case 2 . a l E S ( H ’ ) ,b E S ( H ’ ) Then H ” = H ’ + ( b , a )satisfies S ( H ” ) CSO, R ( H “ ) CRo, IS(H”)l = I S ( H ’ ) l S 1 S(H)I - 1, a contradiction. Case 3. a , E S ( H ‘ ) ,b E S ( H ’ ) Then IS(H’)IS I S ( f i ) l - 2 , and H ” = H ’ + ( a l , a ) satisfies R ( H ” ) CR,,, S ( H ” )c SO, I S(H”)I s I S(H)I - 1, a contradiction. Thus we have proved the minimality of (S(fi)I.So, by the induction hypothesis, there exists in G a stable set S which meets every terminal branch of H. Clearly, S meets also every terminal branch of H = + (al,a ) . 0
x
Corollary 1 (Las Vergnas [8]). Every quasi-strongly connected graph G contains a spanning arborescence with at most a ( G ) sinks.
A graph is quasi-strongly connected if for every x, y EX, there exists an ancestor common to x and y. The result follows immediately. Corollary 2 (Linial [lo]). Let M = { p l ,p 2 , .. . ,pk}be a path partition of G ; either there exists a stable set S which meets each of the pi’s,or there exists a path partition M’ with S ( M ’ ) C S ( M ) , S ( M ’ ) # S ( M ) .
38
C. Berge
Corollary 3 (Camion [5]). Let G be a strongly connected graph such that every pair of vertices is linked with at least one arc (“strong tournament”). Then there exists a hamilton circuit.
Roof. Let p be the largest circuit in G. If p does not cover the vertex set, every arc going out of p is the initial arc of a path p ’ which comes back into p (since the contraction of p gives a graph which is also strongly connected). Let a E p n p ’ be the terminal vertex of p ’ ; the graph obtained from p + p’ by removing the arc ( y , a ) E p and the arc ( y ’ , a ) E p ’ is an arborescence with root a and with sinks y and y’. By Theorem 1, G,,,. can be covered by an arborescence of root a with only one sink, either y or y ’ . By adding the arc of G which goes from that sink to a, we form a circuit larger than p, whjch is a contradiction. 0 Corollary 4. If a graph G has a basis B with 1 B I = IY (G), the vertex set can be covered by IY (G ) disjoint paths all starting from B.
Proof. A basis of G = (X, U ) is a set B C X such that each vertex is the terminal end of a path starting from B, and no two distinct vertices in B are connected by a path. A basis always exists, by a theorem of Konig. Consider the sets Bo= B , B I , B 2 ... . where B, is the set of all vertices which can be reached by a path of length i from B and by no path of length smaller than i. Since B is a basis, U B , = X. Consider a graph Ho with vertex set X,obtained by taking for each x E B,, i 3 1 , one of the arcs of G going from B,-l to x. Clearly Ho is an arborescence forest with R ( H , )= B. By Theorem 1 there exists an arborescence forest H with R ( H ) C B , I S ( H ) I s a ( G ) .However B is a basis, so R ( H ) = B . 0 When G is a strongly connected graph, one can expect to prove stronger results, that is, the existence of better path partitions. Let us mention the following conjectures: Conjecture 1 (Las Vergnas). Every strongly connected graph G with has a spanning arborescence H with 1 S(H)I =sQ ( G )- 1.
(Y
(G) 3 2
Conjecture 2 (Bermond [2]). Every strongly connected graph G can be covered with a ( G ) circuits. One can cover these two conjectures by another one: Conjecture 3. Every strongly connected graph can be covered by a circuit C and a (G) disjoint paths having at most in common with C this initial end-point.
Diperfect graphs
49
This conjecture has been proved recently for a (G) = 2 by C.C. Chen and P. Manalastas [18]. In fact we can apply Theorem 1 to prove the first conjecture for a graph G having the following property:
(P): G has a circuit which meets every maximum stable set. In fact, many graphs satisfy Property (P). Example 1 (P6sa [12]). A symmetric graph satisfies Property (P). Let ( y o , y , , y,, . . . ,y k ) be the longest path issuing from y o ; all the neighbours of y t are of the type yi with 0 S i < k ; let io be the smallest index i 2 0. Then the circuit p = (y,, y , + l , . . . ,y k , y,) contains y k and all its neighbours; therefore every maximal stable set meets the circuit p. Example 2 (Meyniel [ll]). Not every strongly connected graph satisfies Property (PI.
Consider a graph G, with vertices 1,2,. . . ,11, whose arcs are (1,2), (2,3), (1, 8), (7,8), (7,3), (9, lo), (10, l l ) , (4, l l ) , (4,5), (9,5); join all the vertices in A = {9,10,11,4,5} to 6 by arcs directed towards 6. Join 6 to all the vertices in B = { 1,2,3,7,8}by arcs directed out of 6, also add all possible arcs from A to B, but remove the arc (3,9). This graph G has stability number a ( G )= 2, but no circuit meets all the maximum stable sets. Theorem 2. Let G = ( X , U ) be a strongly connected graph with a (G) > 1 that satis,fies Property P; then there exists a spanning arborescence H with S ( H )I S a ( G ) - 1.
I
Proof. Let p be a circuit which meets every maximum stable set. Let G be the graph obtained from G by contracting p into a single vertex c. The graph G is strongly connected; so, by Theorem 1, it admits a spanning arborescence I? with root c, and G has a stable set S with IS 1 = S ( 8 ) I which meets every terminal branch of H. Clearly, c$Z S ; so, from the definition of p, we see that S is not a maximum stable set of G. Hence IS 1 S a(G)- 1. We can construct in G a spanning arborescence H with arcs of p and the images of the arcs of fi so that I S ( H ) (= IS ( f i ) (S a ( G )- 1. This arborescence H fulfills the conditions of the theorem. I7
I
Remark. This result shows that the Las Vergnas conjecture is true for symmetric
50
C. Berge
graphs; the same argument shows also that Bermond’s conjecture is true for symmetric graphs. However, we can expect better results and, in fact, Amar, Fournier and Germa [ l ] have conjectured that a symmetric graph G can be covered by [;a( G ) ] cycles. By using Theorem 1, we can also extend well known properties of tournaments to “join” of graphs. The following result has been proved by Las Vergnas [9], and later by Linial [lo]:
Proposition 1. Let G = ( X , U ) be a graph, let (A,B,. . . ,K ) be a partition of X such that GA, G B , . . , have hamilton paths (a,, a2,. . . ,a‘), (h,6 2 , . . . ,b’), . . . , ( k , , k2,.. . ,k’), respectively. If every pair of vertices in different classes is joined by at least one arc, then G contains a hamilton path starting in {a,, bl, . . . ,kl} and ending in {a’,b’,. . . , k ’ } .
Proof. Clearly it suffices to show the result for a partition (A, B) of X in two classes. In this case, the two hamilton paths of GA and GB constitute an arborescence forest Ho with R ( H o ) C{ a , ,bl}, S(Ho)C { u ’ , b’}. From Theorem 1 , this arborescence forest is not minimum, and there exists a unique path H with R ( H )C {at, btl, S ( H )C {a’,b‘l. 0
For strongly connected graphs, the same argument shows the existence of hamilton circuits. More precisely, we get the following proposition: Proposition 2 (Las Vergnas [9]). Let G = ( X , U ) be u strongly connected graph; let ( A ,B, . . . ,K ) be a partition of X such that GA, G B , . .. have hamilton circuits (or are reduced to a singleton). If every pair of vertices in different classes is joined by at least one arc, then for every integer 1, 3 S 1 C IX 1, the graph G contains a circuit of length 1.
For a tournament G this result was found by Moon in 1969. For a complete proof of Proposition 2, see [9].
2. a-diperfect graphs
A directed graph G is a -&perfect if for every optimal stable set S, there exists a partition of the vertex set into paths p l ,p 2 , .. . such that I S fl p, I = 1 for all i (and if every induced subgraph of G has the same property). Many important classes of graphs are a-diperfect.
Diperfect graphs
51
Theorem 3. Every perfect graph is a -diperfect. Proof. In a perfect graph G, there exist k = a ( G )cliques C,, Cz,. . . ,c k which partition the vertex set, and by RCdei’s Theorem, each C, is spanned by a path p,. So every optimal stable set S satisfies lSnp,l=l
( i = 1 , 2 , . . . ). 0
Theorem 4. Every symmetric graph is a -diperfect. Proof. Let G
= (X,U
) be a symmetric graph:
For an optimal stable set S of G consider a graph G ’ obtained from G by removing the arcs going into S. By the Theorem of Gallai-Milgram, G’ has a partition into k = a ( G )= a ( G ’ )paths P I ,p 2 , . . ,pk. Clearly, the p,’s are paths of G and satisfy IS n p, I = 1. 0
Proposition 3. A graph G which is a n anti-directed cycle is not a-diperfect. Proof. An odd cycle of length 2 k + 1 is given by a set of 2 k + 1 arcs u I , u z , . . . , u 2 k + l and determines a sequence of vertices (xl,x 2 , .. . ,XZktl,xI); a chord can be either an arc of G joining two non-consecutive x,’s, or an arc of G parallel to one of the u,’s. A cycle ( x l , x z , . . . ,X Z k + l , x l ) is anti-directed if (1) its length is odd and 2 5 ; ( 2 ) it has no chords; (3) each of the vertices x l , x2,x3,x4,x6,xs, . . . ,X Z k is either a source or a sink. There are two anti-directed cycles of length 9, shown in Fig. 1, but there is only one anti-directed cycle of length 5 or 7 . If G is an anti-directed odd cycle of length 2 k + 1, the set S = { x l ,x4, x 6 , . . . ,x z k } is a maximum stable set. If M is a partition into paths which meet S in only one point, then the path of M which contains xz is necessarily (xz,x,); so the path of M which contains x3 is necessarily (x4x3);so the path of M which contains xs is necessarily (xs,xh),etc. Thus none of the paths of M can be of length 2, which is a contradiction. 0 The graph in Fig. 2 is not a-diperfect (because the set { a , b } is an optimal stable set which does not have the required property), but it contains antidirected cycles of length 5. This suggests the following conjectures:
Conjecture 1. A graph G is a-diperfect if and only if G does not contain an anti-directed cycle as an induced subgraph.
C.Eerge
52
Fig. 1.
Fig. 2.
Conjecture 2. If every odd cycle (of lengths5) has a chord, then G is a-diperfect. Conjecture 3. If every odd cycle (of length is (Y -diperfect.
3 5) has
at least two chords, then G
Each of these conjectures is stronger than the following one.
3. y-diperfect graph
Gallai [6] and Roy [13] have proved independently that for a directed graph G, max 1 p I 2 (Y (G). The following stronger statement can be proved by the same argument:
Dipetfect graphs
53
Theorem 5. Let k be the maximum number of vertices in a path of G. Then for every path p with k vertices, there exists a k-coloring (S,,S 2 , . . . ,S,) such that 1 Sin p I = 1 for i = 1,2,. . . ,k. Furthermore, this k-coloring has the following property: for each x E SA,there exists an arc going from x to Ll. Proof. Let p be a path with 1 p 1 = k. Let H be a partial graph G obtained from p by adding to p as many arcs of G as possible without creating a circuit. Put A ( x )= max{ I p I p is a path of H issuing from x}; if (x, y ) is an arc of H, then A (x) > A ( y ) (because H is acyclic). If (x, y ) is an arc of G - H, then A ( y ) > A ( x )(because H + (x, y ) has a circuit). So for every arc (x, y ) of G, we have A (x) # A ( y ) , and therefore A ( x ) is a coloring function with max, A ( x )= k colors. Thus a k-coloring with the required property is defined. 0
I
Corollary. Let k = max I p 1, and let S be a set of vertices contained in a path p with ( p I = k . Then t h e s e t A = { x ( x E X - p , T ( x ) C S } U ( p - S ) induces a subgraph with chromatic number y ( G A )s k - 1 S I. Remark. A generalisation of [6] has been obtained by Bondy [3]: If a strongly connected graph G has at least 2 vertices, the longest circuit (“directed cycle”) has length S y ( G ) . The Gallai-Roy theorem for a graph H can be proved by using the Bondy theorem for the graph obtained from G by adding a new vertex xo that we join in both directions with every vertex of G. Furthermore, it contains also a theorem of Camion [5]which states that a strongly connected tournament has a circuit which meets each vertex exactly once. Note that the proof given above for Theorem 5 does not imply that every graph G has an optimal coloring and a path p which meets each color exactly once. However, we have the following proposition: Proposition 4. A directed graph G with y ( G ) = 3 has a 3-coloring ( S , , Sz, S,) and a path p = (a, b, c ) which meets each color class Siexactly once. Proof. Suppose that G is a graph for which that proposition does not hold true. Let be a 3-coloring of G, and let p = (a, b, c ) be a path of 3 vertices (which exists by the Gallai-Roy Theorem); since p does not meet the 3-color classes, we have, say, a, c E b E S,. Let S3 be the maximal stable set which contains and put S1 = 3, - Ss, S2= - S3. Thus, we have
(sI,s,,s,)
s3,
uESI, cESI,
sI,
s2
bES3.
A vertex x of S1 U S2 is necessarily in GS,,,% an isolated vertex, or a source, or a
C. Berge
54
sink. Otherwise, there is a path (y, x, z ) with x E S , , y , z E S2;since x is adjacent to a vertex u E S3, there is either a 3-colored path (y, x, u ) , or a 3-colored path (0,x, z), which is a contradiction. Let TIbe the sources, T2the sinks (which are not isolated vertices), and Tothe isolated vertices in GslUsl.Thus (TI,T2U To,S3)is a 3-coloring of G with no 3-colored path and we may assume a, c E T2 U To, 6 E S3.
Every arc between T2 and S3 is directed from S , (otherwise there would be a 3-colored path). So a E To,and p = ( a ,6, c ) meets the 3 colors of the coloring (TI~{aT } ,~ u T o - { a ) , S3).
The contradiction follows. 0 Remark. Among the applications of the Gallai-Roy Theorem, let us mention the following ones: (1) Rkdei’s Theorem. In every tournament, there exists a path which meets each vertex exactly once. ( 2 ) The Chucitaf-Komlbs Theorem [16]. Let G = (X, U ) be a graph, let U = U , + U 2 + .. . + U, be a partition of the arc set into q classes, let pl,p?,. . . ,p, be integers such that pIp2 * pq < y ( G ) . Then for some i, the partial graph G, = (X, U , ) contains a path with more than p, vertices. (3) The generalized Erdos-Szekeres Theorem [17]. Let cr = (a,, a 2 , .. .) be a sequence of plpz* . . pq + 1 distinct integers; let p l , p 2 , . .. ,pq be binary relations satisfying: for every i < j, there exists a relation p k with a,pka,.Then for some k, there exists a subsequence cr’ = (a,,,a,,, . . . ) of cr of length > p k , such that
-
a,,pkah,ahpka,,,. . . .
The proofs are easy and are left to the reader. A directed graph G is y-diperfecr if for every optimal coloring (S,, Sz,...,S k ) there exists a path p such that I p fl Si I = 1 for all i (and if every induced subgraph of G has the same property).
Theorem 6 . Euery perfect graph is y-diperfect.
Proof. In a perfect graph G there exists a clique C with I C I = y ( G ) , and by Rtdei’s theorem, C is spanned by a path p. So every optimal coloring (S1, S 2 , .. . , s k ) satisfies ( S i n F I = l ( i = 1 , 2 , ...).
55
Diperfect graphs
Theorem 7. Every symmetric graph is y-diperfect. Proof. Let (S1, S2,. . . ,S , ) be an optimal coloring of a symmetric graph G, and let G' be the graph obtained from G by removing the arcs going from S, to S, if j > i. By the theorem of Gallai-Roy, G' has a path p with k = y ( G ' ) = y ( G ) vertices, therefore it meets each of the Si'sexactly once. 0
c,
The graph in Fig. 3, which is isomorphic to the complement of a cycle of length 7, is not y-diperfect, because y ( c , ) = 4 and an optimal 4-coloring is represented by the numbers in parentheses. One can see that the only paths of length 3 3 which contain the color ( 3 ) are agb, egb, agf, egf, and none of them meets the four colors. Also, it is easy to show that the graphs in Figs. 1 and 2 are not y-diperfect, and that the perfect graph conjecture is equivalent to the statement: A simple graph is y-diperfect for each orientation of its edges if and only if G is perfect.
Theorem 8. Let G be a a-diperfect graph on X. Then there exists an optimal coloring c : X+{1,2,-. . , y ( G ) } such that each vertex x o belongs to a path of cardinality c(xo) that meets each of the colors 1,2,. . . ,c(xo) exactly once. Proof. Assume that the result is true for all graphs having chromatic number < k ; we shall show that it is also true for a graph G with y ( G )= k. Let (S1, S 2 , . . . ,S , ) be an optimal coloring of G with SIU Sz U . U S k - , = A maximal, and let xo E Sk. So y ( G A ) = k - l ; y(GAU%)=k. By the induction hypothesis, each vertex x in A belongs to a path of GA with the required properties for some coloring ( S ; ,S:, . . . ,SL-l) of GA.Since the graph
Fig. 3.
C. Berge
56
GAuIdis y-diperfect there exists a path containing xo that meets every color-class of the optimal coloring (Si, S:, . . . ,SL,, {xo}). Therefore each vertex of G belongs to a path with the required property for the optimal coloring (Sl, s;,.. . ,s;-*, Sr). 0 This result is similar to a theorem of de Werra about perfect graphs [15].
References [l] D. Amar, I. Fournier and A. Germa, Private communication, May 1981. [2] J.C. Bermond, Private communication, February 1978. [3] J.A. Bondy, Disconnected orientations and a conjecture of Las Vergnas, J. London Math. Soc. 2 (1976) 277-282. [4] J.A. Bondy and U.S.R. Murty, Graph Theory with Applications (MacMillan, London, 1972). [5] P. Camion, Chemins et circuits des graphes complets, C.R. Acad. Sci. Paris 249 (1959) 2151-2152. [6] T. Gallai, On directed paths and circuits, Theory of Graphs, Erdos and Katona, eds. (Academic Press, New York, 1968) 115-11. (71 T. Gallai and A.N. Milgram, Verallgemeinerung eines Graphentheoretischen Satzes von Rtdei, Acta Sci. Math. (Szeged) 21 (1960) 181-186. [8) M. Las Vergnas, Sur les arborescences dans un graphe orientt, Discrete Math. 1.5 (1976)27-29. [9] M. Las Vergnas, Sur les circuits dans les sommes compltties de graphes orientts, Institut de Hautes Etudes de Belgique, Colloque sur la Thiorie des Graphes, 1973, 231-244. (101 N. Linial, Covering digraphs by paths, Discrete Math. 23 (1978) 257-272. [I I] H. Meyniel, Private communication, May 1981. (121 L. P&a, On the circuits of finite graphs, Publ. Math. Inst. Hung. Acad. Sci. 8 (1963) 355-3. (131 B. Roy, Nombre chromatique et plus longs chemins, Rev. Fr. Automat. Informat. 1 (1967) 127-132. [I41 A. Sache, La Thtorie des Graphes (Presses Universitaires de France 1554, Paris, 1974). [15) D. de Werra, On the existence of generalized good and equitable edge colorings, J. Graph Theory 5, No. 1 (to appear). (16) V. Chvatal and J. Kombs, Some combinatorial theorems on monotonicity, Canad. Math. Bull. 14 (1971). [I71 P. Erdiis and G. Szekeres. A combinatorial problem in geometry, Compositio Math. 2 (1935) 46-70. [ 181 C.C. Chen and P. Manalastas, Every strongly connected digraph of stability 2 has a hamilton path, Discrete Math. 44 (1983) 243-250.
Annals of Discrete Mathematics 21 (1984) 57-61 @ Elsevier Science Publishers B.V.
STRONGLY PERFECT GRAPHS C. BERGE and P. DUCHET C.N.R.S. Paris In this paper, we investigate the class of graphs containing a set of vertices which niwts exactly once every maximal clique.
1. Examples of strongly perfect graphs Let G = ( X , E ) be a simple graph, and let cliques. A set S such that
(e
be the family of all maximal
/sncl=i ( c E ( ~ ) is called a stable transversal of G. If G and all its induced subgraphs have a stable transversal, we shall say that G is strongly perfect. Clearly, every bipartite graph satisfies such a property. Many other classes of perfect graphs are strongly perfect, as can be shown by the following theorems: Theorem 1. Let G be a graph with no induced Pq (elementary chain of four vertices without chords). Then every maximal stable set meets all the maximal cliques. Consequently, G is strongly perfect.
By a theorem of Seinsche [4]every graph with no induced P, is perfect; by the same type of argument, we shall show that every maximal clique meets all the maximal stable sets. Lemma (Seinsche). The following two conditions are equivalent: (i) G = ( X ,E ) has no induced P4; (ii) for every A C X , either the subgraph GA or its complement disconnected.
GA are
If G has an induced P4.then the set A of the four vertices in a P., is such that GA and GA are connected, hence (ii) is false. Conversely, if (ii) is false, let A be a set of vertices with minimum cardinality such that both GA and G,., are connected. Let a be a vertex of A such that GA - { a } contains two distinct connected components, say C and C'; such a 57
C. Berge, P. Duchet
58
vertex exists by the minimality of A. Since G, is connected, there exists a point b in A such that a b E E ; we may assume b E C. Then, by an appropriate choice of vertices b, c, c ’ , we may have C‘EC’, c EC,
ab, bc’, cc’ $Z E, ac‘, ac, bc E E. For B
= { a , c, c ’ , b } , we
have Ge= Pa, which contradicts (i).
Proof of Theorem 1. Assume that every graph with no induced P4 and of order less than n has the following property: “every maximal stable set meets every maximal clique”. We shall show that a graph G of order n with no induced P4 also has the same property. By the lemma, either G or G are disconnected; we may assume that G is disconnected and has two connected components GI and Gz. By the induction hypothesis, a maximal clique C1of GI, a maximal stable set S1of G,, a maximal clique C2of G, and a maximal stable set S , of Gz satisfy
Therefore, every maximal stable set of G, which is of the type S = SI U Sz, meets every maximal clique of G, which is of the type C1 or C,. 0
Theorem 2. Every comparability graph is strongly perfect. Proof. Let G be the comparability graph of an ordered set (X, S ); the strongly connected components of C are disjoint classes, and some of these are ‘terminal’ components (with no edge going out). Choose one vertex in each of the terminal components, to define a set S. Clearly, S is stable. By a theorem of Rtdei, a maximal clique C is spanned by a directed path, and by the maximality of C this path ends in a terminal component; hence C meets S in exactly one point. 0 Note that the complement of a comparability graph is not necessarily strongly perfect, as we can see with (see Fig. 1).
c6
Theorem 3. Every graph which is the complement of a triangulated graph is strongly perfect. Proof. It is known that a graph G = (X, E ) is triangulated if and only if, for every A X,the subgraph GA has a simplicia1 vertex (i.e., a vertex which belongs to only one maximal clique); see, for instance, [l].
Strongly perfect graphs
59
Let G be a triangulated graph, let x be a simplicia1 vertex, let C be the maximal clique containing x. Every maximal stable set S of G satisfies S n C# 0;otherwise, S U {x} would be also a stable set, which contradicts the meets every maximal maximality of S. Hence, in G, the maximal stable set clique 5.
c
Theorem 4. Every triangulated graph is strongly perfect. Proof. Let G = (X, E ) be a triangulated graph. By a theorem of Hajnal and Suranqi 121, it is known that G is perfect. We shall index the vertices x I ,x2,. . . so that for every k, the vertex Xk is a simplicial vertex of the subgraph induced by { x k ,& + I , . . .,xn}. Put
Tx, = {x, / I > i, [x,, x,] E E l The directed graph ( X , r ) has no (directed) circuits; furthermore, for each maximal clique C there exists a vertex x, such that { x l }U Tx, = C. Since the graph ( X , r ) has no circuit, it has a kernel S ; the maximal clique C = {x,} U Tx, meets S, by the definition of a kernel. Hence there exists a stable set S which meets all the maximal cliques. 0
2. General results
In this section, we give a characterization of strongly perfect graphs. Theorem 5. Every strongly perfect graph is perfect. Proof. Since the induced subgraphs of a strongly perfect graph are strongly perfect, it suffices to show that a strongly perfect graph G has its chromatic number -y(G) equal to w ( G ) , the maximum size of a clique. Assume that w ( G ) = k, and that the theorem is true for all the graphs G’ with w ( G ’ )< k ;let S be a stable set of G which meets all the maximal cliques. Then k - 1= w ( G ~ = -~ Y )( G ~ -(by ~ )the induction hypothesis). Hence
+
y ( G )S ( k - 1) 1 = k
= w (G) C y(G).
Hence -y(G)=w ( G ) . 0 Theorem 6. A perfect graph G = ( X ,E ) has a stable set S which meets all the maximal cliques iff no two families % =(CI,C Z , ..,ck) . and 9 = (Dl,D 2 , .. . ,Dk,) of maximal cliques (with possible repeated cliques) satisfy
C.Berge, P. Ducher
60
1%I=1 9I
and
I % I>1% I
for all x E X .
Ce, is the subfamily of the cliques of % which contain x, with possibly repeated
cliques. The two families can be described by an assignment of an integer z (C) E 2 to each clique C of G so that XcEoz(C)= 0 and Xceo, z ( C ) > 0.
Proof of Theorem 6. Let H = (X, 8)be a hypergraph on X whose edges are the maximal cliques of G. It is known that G is perfect iff the dual H * is a normal hypergraph, that is, 4 ( H ’ )= A ( H ’ ) for all H ’ c H*. Clearly, the following conditions are equivalent: (1) G has a stable set S which meets all the maximal cliques; (2) H has a stable transversal S; (3) H * has a ‘perfect’ matching (a matching which covers the vertex set). We say that a hypergraph H * is quasi-regularizable if there exists a positive integer k and a partial hypergraph H’ of kH* which is ‘regular’ (every vertex of H * has the same degree in H ‘ ) . If H * has a perfect matching, then H * is quasi-regularizable (since a perfect matching H‘ is a regular partial hypergraph of H * ) . For a normal hypergraph, the converse is also true, because if H * is normal and has a partial hypergraph H’ (with repeated edge) which is regular, then H’ is also normal, so q ( H ’ )= A ( H ’ ) , and each color of an optimal edge-coloring of H‘ defines a perfect matching of H*. So (3) is equivalent to the following condition: (4)H * is quasi-regularizable. Also, the Farkas lemma shows that (4)is equivalent to the next condition: ( 5 ) there exists no integral valuation z, E 2 (for each vertex i of H*)such that
& zi> 0
for every edge E of H *
iE
Define a family % by taking zi times the clique C, if zi > 0; define the family 9 by taking I zj I times the clique C, if zj < 0. So the two families % and 9satisfy
Hence the condition of the theorem is equivalent to (5). 0 Some examples of perfect graphs are presented in Figs. 1, 2 and 3; the numbers associated with the maximal cliques show that none of these graphs is strongly perfect.
Strongly perfect graphs
61
c6
Fig. 2.
Fig.1.
-3
-3
-3
-3 +4
Fig. 3.
References [l] P. Duchet, Classical perfect graphs (this volume, pp. 67-96). [2] A. Hajnal and J. Suranj4, h e r die Auflosung von Graphen, Ann. Univ. Sci. Budapestinensis 1 (1958) 113-120. [3] H. Meyniel, The graphs whose odd cycles have at least two chords (this volume, pp. 115-119). [4] D. Seinsche, On a property of the class of n-colorable graphs, J. Comb. Theory, Ser. B 16 (1974) 191-193.
This Page Intentionally Left Blank
Annals of Discrete Mathematics 21 (1984) 63-65 @ Elsevier Science Publishers B.V.
PERFECTLY ORDERED GRAPHS
v. CHVATAL School of Computer Science, McGill University, Montreal, Canada This note presents a good characterization of a class of strongly perfect graphs which includes all comparability graphs, all triangulated graphs and all complements of triangulated graphs.
This note concerns colouring the vertices of a graph in the usual way, with adjacent vertices receiving distinct colours. When the vertices of a graph are ordered in a sequence ul, uz, . . . , u,, a natural way of colouring them by positive integers consists of scanning the sequence from u1 to u, and assigning to each u, the smallest positive integer f ( v , ) assigned to none of its neighbours u, ( i < j ) . We shall refer to the graph with the linear order on the set of its vertices as an ordered graph and to the largest integer appearing as some f ( u , ) as the Grundy number of this ordered graph. (The Grundy function [3] of a directed acyclic graph is the unique function assigning to each vertex u the smallest nonnegative integer g ( u ) assigned to no vertex u such that there is a directed edge from u to u. If, in a graph with a linear order < on the set of its vertices, each edge uu with u < u is directed from u to u, then the Grundy function g of the resulting directed acyclic graph and the function f defined above satisfy f ( u ) = g(v) + 1 for all vertices u.) Trivially, the Grundy number of an ordered graph is at least its chromatic number; to see that the inequality may be strict, consider the graph with vertices a, b, c, d, edges ab, bc, cd and a linear order < such that a < 6, d < c. We shall refer to any of these (three) ordered graphs as an obstruction. A linear order on the set of vertices of a graph will be called (i) admissible if it creates no obstruction and (ii) perfect if, for each induced ordered subgraph H, the Grundy number of H equals the chromatic number of H. Thus, every perfect order is admissible; to prove the converse, we shall rely on the following fact.
Lemma. Let G be a graph and let C be a set of pairwise adjacent vertices in G such that each w E C has a neighbour p(w)E! C ; let the vertices p ( w ) be pairwise nonadjacent. If there is an admissible order < such that p ( w ) < w for all w E C then some p ( w ) is adjacent to all the vertices in C. Proof. Let us proceed by induction on the number of vertices in C. For each 63
64
v.ChUdtd
w E C,the induction hypothesis guarantees the existence of a vertex w * E C such that p ( w *) is adjacent to all the vertices in C except possibly w. In fact, we may assume that p ( w *) is not adjacent to w, for otherwise we are done. Now it follows that the mapping which assigns w * to w is one-to-one, and therefore it is onto. In particular, with u standing for that vertex in C which comes first in the admissible order, there are vertices b, c E C such that b * = u and c* = b. But then there is a contradiction: the vertices a, b, c, d with a = p ( b ) and d = p ( v ) constitute an obstruction. 0
Theorem 1. A linear order of the set of vertices of a graph is perfect if and only if it is admissible.
Proof. The ‘only if’ part is trivial; the ‘if’ part will be proved by induction on the number of vertices. Let G be a graph with an admissible order < of the set of its vertices and let k stand for the Grundy number of this ordered graph. By virtue of the induction hypothesis, it will suffice to show that the chromatic number of G is (at least) k. Thus it will suffice to find k pairwise adjacent vertices in G. For this purpose, consider the smallest i such that there are pairwise adjacent vertices w , + ~w, , + ~. ., . ,wk with f ( w , ) = j for all j. If i = 0 then we are done; otherwise each w, has a neighbour p ( w , ) such that p ( w , ) < w, and f ( p ( w , ) )= i. But then the lemma implies the existence of a vertex u with f ( v ) = i, adja-bent to all the vertices w,, which contradicts the minimality of i. 0 A graph will be called perfecfly orderable if it admits a perfect order. The proof of Theorem 1 shows that every perfectly orderable graph is perfect: the chromatic number of each of its induced subgraphs H equals the number of vertices in the largest clique in H. A related property has been studied by Berge and Duchet [4]: a graph is called strongly perfect if each of its induced subgraphs H contains a stable set meeting all the maximal cliques in H . (Here, as usual, “maximal” is meant with respect to set-inclusion.) It is easy to show that every strongly perfect graph is perfect (see [4]).
Theorem 2. Every perfectly orderable graph is strongly perfect.
Proof. It will suffice to find, in an arbitrary graph G with a perfect order < , a stable set S meeting all the maximal cliques in G. We claim that S can be found by the following greedy algorithm: scan the perfect ordering v I ,v2,. . . , on from u , to v. and place each uj in S if and only if none of its neighbours ui (i < j ) has been placed in S. Indeed, if the resulting stable set S is disjoint from some clique C then each w E C has a neighbour p ( w ) E S with p ( w ) < w. But then the lemma implies the existence of a vertex u E S adjacent to all the vertices in C. Thus C is not maximal. 0
Perfectly ordered graphs
65
(As Claude Berge and Pierre Duchet pointed out to me, this argument may be interpreted as a common generalization of the arguments used in their pioneering study [4] to establish strong perfection of various classes of graphs.) The converse of Theorem 2 is false: the strongly perfect graph shown in Fig. 1 is not perfectly orderable. (The decision to set 1 < 2, which can be made without loss of generality, sets off a chain reaction forcing 3 < 4 , 5 < 6 , 7<8, 3 < 2 , 9<10, 11<12, 13<14, 15<10, 9 < 1 , 16<17, 18<19, 20<21 and 16<1, whereupon the vertices 16, 1, 2, 3 constitute an obstruction.) On the other hand, it is an easy exercise to show that comparability graphs, triangulated graphs and complements of triangulated graphs are perfectly orderable. (More precisely, a graph is a comparability graph if and only if its vertices can be linearly ordered in such a way that no induced subgraph with vertices x, y, z and edges xy, yz has x < y < z . A graph is triangulated if and only if its vertices can be linearly ordered in such a way that no induced subgraph with vertices x, y, z and edges xy, yz has x < y and z < y. A graph is the complement of a triangulated graph if and only if its vertices can be linearly ordered in such a way that no induced subgraph with vertices x, y, z and the single edge xz has y < x and y < z.) We close with the obvious question: how difficult is it to recognize perfectly orderable graphs? 13
50
5
Fig. 1.
References [l] C. Berge, Farbung von Graphen, deren samtliche bzw. deren ungerade Kreise starr sind, Wiss. Z. Martin-Luther-Univ. Halle-Wittenberg 114 (1961). [2] C. Berge, Sur une conjecture relative au p r o b l h e des codes optimaux, comm. 132me assemblee gtnCrale de I'URSI, Tokyo, 1962. [3] C. Berge, Graphs and Hypergraphs (North-Holland, Amsterdam, 1973). [4] C. Berge and P. Duchet, Strongly perfect graphs (this volume, pp. 57-61).
This Page Intentionally Left Blank
Annals of Discrete Mathematics 21 (1984) 67-96 @ Elsevier Science Publishers B.V.
CLASSICAL PERFECT GRAPHS An introduction with emphasis on triangulated and interval graphs Pierre DUCHET C.N.R.S. Paris Essential results and some new ones are presented on Comparability graphs, triangulated graphs, interval graphs, and related classes of perfect graphs.
1. Introduction
It is well known that the birth of perfect graphs, in 1960, was closely related to the discovery of perfection for comparability graphs [5],for triangulated graphs [5], [48], and for interval graphs [5]. These famous classes and subclasses or related classes, studied by various authors since the 1960’s, are called classical perfect graphs (see Fig. 1). In this survey, we restrict our attention to those classes that are ‘good classes’ in the following sense: every induced subgraph of a graph in the class is also in the class. So these classes may be characterized by forbidden configurations. Fortunately, these classes also have a ‘good characterization’ in the sense of Edmonds [19]. They are also recognizable in polynomial time. Moreover, these graphs are strongly perfect (see articles in this volume, pp. 57-61 and 63-65). Hence their perfectness will not be proved here. The largest classes of classical perfect graphs are comparability graphs, triangulated graphs, and their complements (called cocomparability and cotrian gulated graphs). It would be interesting to include all of them in a wider class.
Problem 1.1. Find a well characterized class of perfect graphs that contains comparability graphs, triangulated graphs, and their complements. We summarize in Section 3 the most famous results about the characterization problem for classical perfect graphs. Sections 4, 6 and 8 deal with the most important classes. We shall try to demonstrate their usefulness from the standpoint of representative graphs (see Section 2 for definitions). In particular, simple proofs are proposed for triangulated graphs and interval graphs via, respectively, subtree hypergraphs and interval hypergraphs (Sections 5 and 7). 67
P. Duchet
68
The reader who is interested in the algorithmic point of view is referred to 1371With slight differences our terminology is taken from [7]. Our graphs are undirected simple graphs, finite without loops or multiple edges. Hypergraphs will be denoted by the form H = (X, 8'), where X is the vertex set and 8 is any family of subsets of X. ' E occurs in 8' will often be denoted by E E 8 or by E E H.
Clique : w(G): ll(G):
Walk :
Path :
set of vertices of a complete subgraph. maximum number of vertices of a clique of G. stability number. sequence xI, . . . ,x,, of vertices such that x ~ and + ~x, are adjacent for all 1 S i S n. In a closed walk, the first and last vertices coincide. (for graphs and hypergraphs) alternated sequence xlE1.. . X.E.X,,+~ of edges and vertices that satisfies x, E E,
and x , + ~E E, for 1 c i s n,
E,# E, for 1 6 i < j x,#x,
S
n,
for 1 ~i < j s n ,
X"+~#X,f o r 2 s i i n n ;
the path becomes a cycle when xn+*= x,. f ( H ) of a hypergraph H = (X, 8'): hypergraph on f ( X ) whose edges are f ( E ) for E E 8'; here f is a mapping X + Y. Contraction of H = (X, ZP) by E C X: mapping c : X-+X \ E U { e } where e is a new element; c is defined by
Image :
x
if X ~ E ,
e
if x E E.
c(x) =
The image c ( H ) is also called the contracted hypergraph from H by E or the contraction of H by E and is denoted by HIE. HIEl, ..., Em means H/EI/ . ' . /Em.These hypergraphs are determined up to isomorphism.
Graphs modulo subset: If V , , .. ., V, are subsets of vertices in a graph G = (V, E), they constitute the vertices of a new graph, denoted by G(VI,. . . , V,), in which V and V, are joined when they are adjacent in G (i.e.
rG(v,)nv,z0).
Classical perfect graphs
69
2. Representative graphs
The representative graph or intersection graph of a hypergraph H = (xl, . . . ,x, ; E l , . . . , E m is ) the graph denoted by L ( H )whose vertices are the E,, two vertices E, and E, being adjacent if and only if E, and E, have a common point. When we are interested in a class %' of hypergraphs, any hypergraph isomorphic to a member of %' will be called a (e-hypergraph. The representative graphs of %'-hypergraphs will be the %'-graphs. Helly hypergraphs, in which every (finite) family of edges with no common point contains two disjoint edges, play an important role in the domain of representative graphs. Every graph G is the representative graph of a Helly hypergraph, namely, the hypergraph C*(G ) defined as follows: the vertices of C*(G) are the maximal cliques of G ; the edges of C*(G) are the sets of maximal cliques of G that contain a given vertex of G. We define a refraction of a hypergraph H = ( X , a ) as a map r : X + X satisfying the following property: r ( E ) C E for every edge E of H. The range r ( H )= ( X ;( r ( E ) ) ,E in an edge of H ) is also called the retraction of H by r. An R -class of hypergraphs is a class (e satisfying the following two axioms: (2.1) Every Ce-hypergraph is a Helly hypergraph. (2.2) Every retraction of a Ce-hypergraph is a Ce-hypergraph.
Examples of R-classes are normal hypergraphs, balanced hypergraphs, unimodular hypergraphs, subtree hypergraphs (Section 5), and interval hypergraphs (Section 7). The above definitions are motivated by the following fundamental lemma: 2.3. Representation lemma (cf. Duchet [15], [16]). Let %' be an R-class. A graph G is a %'-graph if and only if C*(G) is a %'-hypergraph. Proof. The 'if' part is nothing but the fact that G is isomorphic to L ( C * ( G ) ) . Conversely, if H = (x,,. . . , x,; E l , .. . , E m )is a (e-hypergraph with G = L ( H ) , we may associate to each maximal clique C of G the intersection subset E, = n,,,,E,. By axiom (2.1), E, is not empty. Distinct Ec's are disjoint; the contraction by a given Ec is clearly a combination of retractions of H. Moreover, it is not difficult to see that for every vertex not in any Ec( = isolated vertex in H or vertex of degree 1) there is a retraction of H whose range is X \ x. So, by axiom (2.2), the hypergraph constructed from H via the contractions by all the Er
P. Ducher
70
successively and via the elimination of vertices not in any Ec is a %-hypergraph and is isomorphic to C * ( L ( H ) ) =C * ( G ) . 0 3. Classical perfect graphs
The most important special classes of classical perfect graphs are classified in Fig. 1. In this diagram, all inclusions between classes are indicated by arrows. They are strict inclusions. What follows is relative to a graph G = (X, E) which is finite, without loops or multiple edges. 3, I . Comparability graphs Definition. There is a partial order S on X such that x and y are adjacent if and only if x s y or y s x. See Section 4. 2.2. Cocomparability graphs Definition. A cocomparability graph is the complement of a comparability graph.
PERMUTATION
INTERVAL
THRESHOLD
Fig. 1. Classical perfect graphs.
Classical perfect graphs
71
3.3. Permutation graph Definition. There exists a permutation 7~ on X and a labelling 1,. . . ,n of X, such that i and j are adjacent if and only if . r ( i ) - TO’)< 0.
i-j
Characterization (Dushnik and Miller [ 181). Each of the following conditions is necessary and sufficient for G to be a permutation graph: (i) G and G are both comparability graphs, (ii) G is the comparability graph of an order of dimension s 2 , (iii) G is the comparability graph of the inclusion relation between a family of intervals of N, (iv) G has no induced subgraph isomorphic to one of the graphs listed in Fig. 8, or to their complements (see Section 4). Dushnik and Miller’s result has been rediscovered with an algorithmic characterization by Pnueli, Lempel and Even [65].
3.4. Circular permutation graphs Definition (Rotem and Urrutia [71]). A circular permutation diagram D, consists of: (i) two concentric circles C1 and C2 in the plane, (ii) n points l’, . .. ,n‘ on C,in the clockwise direction, (iii) n points ~ ( l .).,. ,r ( n ) on C2 in the clockwise direction, (iv) n paths 1,.. . ,fi respectively from i’ on C1 to T ( i ) on C2. Two distinct paths do not intersect in more than one point.
G is a circular permutation graph when there is a labelling 1,. . . , n of X and a permutation diagram D, satisfying the property that i and j are adjacent if and only if T and intersect.
7
Characterization (Rotem and Urrutia [71]). For a vertex u of G, define the graph G,u as follows (where V = r G ( u ) ) :
G,v and G have the same vertices, ( G , u )=~G, (G,u)x,v = Gxiv. For x E V and y E X \ V, x y is an edge of G * v if and only if x y is not an edge of G.
P. Duchet
72
Theorem 3.4. G is a circular permutation graph if and only if it satkfies both the following conditions : (i) G is a comparability graph, (ii) G,v is a permutation graph for some vertex v.
3.5. P4-free graphs Definition. P4,the path with four vertices, is not an induced subgraph of G. Characterization (see article in this volume, pp. 57-61). Every induced subgraph G’ or its complementary graph is disconnected. 3.6. Arborescence compurability graph Definition. The comparability graph of an ‘arborescent order’ means an order < X such that, for every x E X , { y , y .< x ) is a linearly ordered subset.
on
Characteriztion (Wolk [84],[85]). G does not contain P4 or C,as an induced subgraph (see Fig. 2). Another necessary and sufficient condition is the following: For every edge x y of G, we have V ( X ) C V ( Y ) or
V(Y)CW).
(Here V(s) = Tc(s) U {s} for s E X.) 3.7. Threshold graphs
Definition (Chvatal and Hammer [12]). Put X = 11,. . . ,n } . There exists a linear inequality
2 a,x, s a
,=I
with a; E aB and n = I X I such that the following holds: S C X is a stable set of G if and only if (1) is satisfied by the characteristic vector xs = ( x r ,x 2 , . . . ,x,) of S
Fig. 2.
73
Classical perfect graphs
where for all i x,
1 if i E S, if i P S.
=( 0
Characterization (Chvital and Hammer [12]). A necessary and sufficient condition for G to be threshold is that G does not have 2Kr, P, or C, as an induced subgraph (see Fig. 3).
Thus (a) The complement of a threshold graph is threshold. (b) A threshold graph is an arborescence comparability graph and a fortiori a permutation graph. Those permutations that produce a threshold graph were characterized by Golumbic [36]. Let (T and T be two sequences over some alphabet. The shuffle product is defined as follows: (T
w
T ={(TITI,.
. . ,(TkTk
I CT =
@I,.
. . ,U k
and
T = TI,.. . , T k } .
Here the a: and are subsequences, k ranges over all integers, and juxtaposition means concatenation. Theorem 3.7.1 (Golumbic [36]). The threshold graphs are precisely those permu tation graphs corresponding to sequences contained in
[ 1,2, . . . ,p ] ul [ n, n
-
1, . . . ,p
+ 11
where p and n are positive integers and w denotes shufle product. As for split graphs, thresholdness is characterized by the degree sequence of the graph. Furthermore, the entire graph is completely determined. Theorem 3.7.2 (ChvBtal and Hammer [12], [24]). Let G = ( X , E ) be a graph. The following statements are equivalent: ( i ) G is a threshold graph. (ii) (Henderson and Zalcstein [51]). There exist an integer-valued function U : X + N and a n integer t such that for distinct vertices x and y x and y are adjacent
FX 3%
V ( x ) +U ( y ) > t.
%% p4
Fig. 3.
c4
P. Duck1
74
(iii) There exist a clique C and a stable set S partitioning X and there are partitions
c = c,u
* * .
u cpucp+,.
s=s"us,u~~'usp, with p
20
and Cp+,, So possibly empty, such that for all 0 S i, j
x E C, and y E Sj are adjacent
S
p,
i s j.
3.8. Triangulated graphs
Definition. Every cycle of length 3 4 in G has a chord, that is, an edge joining two non-consecutive vertices of the cycle. In the literature, triangulated graphs are also called chordal, rigid -circuit, monotone transitive, and perfect elimination graphs. See Sections 5 and 6. 3.9. Cotriangulated graphs
Definition. A cotriangulated graph is the complement of a triangulated graph. 3.10. Split graphs Definition. There is a partition X = CU S of the vertex set of G into a clique C and a stable set S. Characterization (Foldes and Hammer [24]). Each of the following conditions is necessary and sufficient for G to be a split graph: are triangulated graphs, (i) Both G and its complement (ii) G does not have 2K2, C, or C, as induced subgraphs (see Fig. 4). A nice property of these graphs is that the splitness is completely determined by the degree sequence. More precisely, we have the following theorem:
Fig. 4.
Classical perfect graphs
75
Theorem 3.10 (Hammer and Simeone [50]). Let G = ( X , E ) be a graph with degree sequence d l 5 d Z3 * 5 d,, and put m = max{i d, 5 i - 1). Then G is a split graph if and only if
--
1
m
x d i =m(m-1)+
r=l
i=m+l
d,.
Furthermore, if this is the case, then
w ( G )= m.
The class of split graphs has been generalized by Peled [63] (matroidal graphs, which are possibly imperfect) and Foldes and Hammer [24] (matrogenic graphs, which coincide with matroidal perfect graphs).
3.11. Split comparability graphs The class of graphs that are both split graphs and comparability graphs was identified by Foldes and Hammer [24] as those graphs that do not contain 2 K z , C,, C,, D2, F2 or H (see Fig. 5 ) as an induced subgraph.
3.12. Interval graphs Definition. There is a one-to-one correspondence I between vertices of G and intervals of a totally ordered set such that X X Y E E
e
I(x)nr(y)#B.
See Sections 7 and 8.
3.13. Indifference graphs See Section 8.
3.14. Split cocomparability graphs Foldes and Hammer [24a] have shown that a split graph G is an interval graph
P
Fig. 5 .
P. Duchel
76
if and only if its complement G is a comparability graph. These graphs are exactly the graphs having n o induced subgraph isomorphic to any of the graphs of Fig. 6 ([24a]).
4. Comparability graphs A graph is a comparability graph when it admits a transitive orientation of its edges. A weaker condition was found by Ghouila-Houri: Theorem 4.1 (Ghouila-Houri [34]). A relation > is said to be a pseudo-order on a set X if we have:
a > 6, b > c 3 a > c o r c > a (pseudo-transitivity), a>b 3 notb>a
(antisymmetry).
A graph G is a comparability graph if and only if it admits an orientation of its edges that represents a pseudo -order relation. This result yields a characterization in terms of odd closed walks which had been conjectured by A. Hoffmann:
Theorem 4.2 (Ghouila-Houri [34], Gilmore and Hoffmann [35]). A graph G is a comparability graph if and only if for every closed walk with odd length ~ X , exists an edge of the form x,x,+~(where i is taken modulo x I .* - X ~ ~ + there 2p + 1). Gallai strengthened this result, considering the following relation between adjacent edges : a b ac i f and only if b and c are in the same connected component of the subgraph GrGc,, induced by the neighbours of a (in G ) in the complement of the graph G. A wreath of G is defined to be a cycle x l . . x&l that satisfies the following two conditions: (i) all x, are different,
-
2K2
c.4
c5
D2
Fig. 6.
El
Classical perfect graphs
77
-
(ii) xixi-] xixi+lfor all i (modulo p ) , where p is the length of the wreath (see Fig. 7).
Theorem 4.3 (Gallai [32]). A graph G is a comparability graph if and only if it does not have a n odd wreath. We shall say nothing of the deep structure of comparability graphs for we would plagiarize Gallai’s remarkable work [32], the conclusion of which is the following characterization b y excluded configurations:
Theorem 4.4 (Gallai [32]). G is a comparability graph if and only i f it does not contain an induced subgraph isomorphic to one of the graphs C (1 d i d 4) shown in Fig. 8 ( a ) or of the complements of the graphs K (5 S i S 19) shown in Fig. 8(b). Recently, Comparability graphs were characterized as the complement of some intersection graphs:
Theorem 4.5 (Rotem and Urrutia [72]). The complement of a graph G with n vertices is a comparability graph if and only if G is the intersection graph of the graphs f , , . . . ,f n of n continuous functions E : (0,l)- R. In fact, Rotem and Urrutia proved that the E’s may be chosen linear by intervals and intersecting in a finite number of points. Arditti et al. [2] studied the comparability graphs that completely determine (up to duality) a partial order. These graphs are named uniquely partially orderable graphs (U.P.O.). Such graphs have been characterized by Shevrin and Filippov [74] and Trotter, Moore and Sumner [78]. Theorem 4.6 (Arditti [2]). The comparability graph of an irreducible partial order is U.P.O.
Fig. 7. Wreaths.
P. Duck1
78
2n
2n 0 2 2ntl 1 2n+ &$2 1
1
"L/ n
1
D,
C, n > 6
n>2
En n > 1
F, n > 1
,/A A1
A3
A4
A5
dz A7
A0
A9
A10
Fig. 8(b). Fig. 8. Critical non-comparability graphs @(a)) or their complements (U(b)).
More generally, the comparability graph of a partially ordered set suffices to give much interesting information about this order. Theorem 4.7 (Trotter, Moore and Sumner [78]). The dimension of a partial order ( = the minimum number of linear orders whose intersection is the given order) is determined by its comparability graph.
Theorem 4.7 also appears in [47]. A nice representation of a comparability graph, generalizing the result on permutation graphs by Dushnik and Miller (see Theorem 3.3), is given by the following theorem:
Clnssicnl perfect graphs
79
Theorem 4.8 (Leclerc [57]). G is the comparability graph of an order of dimension < d if and only if it represents the inclusion relation between subtrees of some tree that has at most d pendant vertices. (In this theorem, a ‘subtree’ means a subset of vertices that induces a subtree; see Section 5.) Purely graphical strong properties, generalizing perfectness, were obtained by Greene and Kleitman [38], [40]. See [39] for a survey of these properties involving cliques and stable sets (or, equivalently, chains and antichains for partial orders); these results generalize to a great extent the famous Dilworth theorem (see [13], [28], [64], [65], [76], [30],[41], [SO]). Using the minimal cost flow algorithm of Ford and Fulkerson [25],Frank [28] gives a nice formulation and a generalization of Greene and Kleitman’s min-max relations. Following Frank, let us say that two families % and Y constituted by subsets of a given set X are orthogdnal when they satisfy the following two conditions:
C nS#
0
for every C E % and every S E 9,
(4.10)
Theorem 4.11 (Frank [28]). There exist families V1,. . . , Va and Y 1 ,... ,Yo and there exist integers a = j , > * * > js = 1 and 1 = i, < * < is = w with the following properties : i s a family of j disjoint cliques of G (for 1 S j d a ) . (i) (ii) Y, is a family of i disjoint stable sets of G (for 1 S i S w ) . and Y,, are orthogonal (1 k S s). (iii) For every j, jk 2 j 2 jk+,, (iv) For every i, ik s i s i k + l , Y, and V1,are orthogonal (1 G k s s). Corollary 4.12 (Greene and Kleitman [40]). Denote
where the maximum is taken over all families Y, consisting of i stable sets of G. Then
where the ,first minimum runs over all partitions of X into cliques C1,.. . , C,. Corollary 4.13 (Greene [38]). Denote
P. Duchet
80 w,
=max el
1 uc/, C€’%,
where the maximum is taken over all families %, consisting of j cliques of G. Then w,
=min Cmin(lS,I,j) t=1
where the first minimum runs over all partitions of X into stable sets S,, . . . ,S,.
5. Representation of triangulated graphs A subtree means here a subset of vertices that induces a subtree of a given graph. A subtree hypergraph is a finite hypergraph H = (X, %) whose edges are subtrees of some fixed tree having X as set of vertices.
Theorem 5.1 (Duchet [ 161, Flament [23]). A hypergraph H = ( X ,%) is a subtree hypergraph if and only if it has both of the following properties: (S.l.1) H is u Helly hypergraph. (5.1.2) Every cycle of H contains three edges with a non-empty intersection. It should be noted that a subtree hypergraph is not generally a balanced hypergraph.
Proof of Theorem 5.1. The proof of this theorem uses the following simple remark whose verification is left to the reader:
Lemma 5.2. Let E and F be two intersecting edges of a hypergraph that satisfy (5.1.1) and (5.1.2). Then, if we add E U F o r E f l F as a new edge, the resulting hypergraph also satisfies (5.1.1) and (5.1.2). Now, among all hypergraphs containing H as a partial hypergraph and satisfying conditions (5.1.1) and (S.1.2), we choose one, say H‘, with a maximal number of edges. Let T be the partial hypergraph of H’ constituted by the minimal edges that contain at least two vertices (‘minimal’ is relative to the inclusion relation). We are going to show that T is the required tree; H’ is a family of subtrees of T. Let E be an edge of H’, IE 122, and a, b two vertices of E. Suppose { u. b ) E H ‘ . Then, by maximality of H’, if we add { a ,b } as a new edge of H’, the resulting hypergraph contradicts (S.l.1) or (S.1.2). In both cases, noting that if
Classical perfect graphs
81
a E l x l* . . x k +Ekb , is a path in H‘ from a to b, then E , U * * * U E, is an edge of H’ by Lemma 5.2, there exist two edges of H’, say A and B,with the properties a€A, bEB
and
AflBn{a,b}=0.
By (5.1.1) we have A n B n E # 0, hence A n E and B f l E are edges of H’ with at least two vertices each. We may conclude as follows: ( I ) If E is in T, then E = { a , b } . Hence, T is a graph and, by (5.1.2), has no cycle. (2) If E’ is every edge in H ’ , every two vertices a and b of E‘ are connected by a path in H’\E‘ whose edges are included in E ’ . The conclusion easily follows, by induction on E’l . 0
I
A complete description of all trees on which a subtree-hypergraph is representable is possible (see [16], [17]). The above theorem provides a nice characterization of triangulated graphs, many times rediscovered: Theorem 5.3 (Buneman [ 101, Gavril [33]). A graph G is a triangulated graph if and only if G is a subtree graph, i.e., the intersection graph of a family of subtrees of a tree. Moreover (see [lo]), a triangulated graph is the intersection graph of subtrees of a tree whose vertices are the maximal cliques of G, the subtrees being constituted by the set of those maximal cliques that contains a given vertex of G. The proof is a simple application of Lemma 2.3 and Theorem 5.2, noting that G is a triangulated graph if and only if C*(G)satisfies (5.1.1) and (5.1.2). An example of representation by subtrees is shown in Fig. 9. Walter [81], [82] studied some graphs that are intersection graphs of subtrees of a prescribed tree.
6 Fig. 9.
4
5
a2
P. Duchet
Gavril[33] remarked that every triangulated graph is the intersection graph of a 'Sperner' family of subtrees of a tree ( = no subtree is included in another).
Corollary 5.4. A hypergraph H is the set of all maximal cliques of a triangulated graph if and only if it satisfies one of the following equivalent properties: (5.4.1) No edge of H contains another edge of H and in every cycle of length 3 3 of H there is an edge that contains three vertices of the cycle (Berge, cited in I1 1). (5.4.2) No edge of H contains another edge of H and the dual hypergraph H * is a subtree hypergraph. Analysing a cyclomatic number for hypergraphs, Acharya and Las Vergnas [ 11 have shown another interesting characterization of these hypergraphs. More precisely, let H be a hypergraph: L , ( H ) is the graph L ( H ) with weighted edges (the weight of (E, F) being I E n F I ) ; w ( H ) denotes the maximal weight of a forest of L , ( H ) ; and p ( H ) , the cyclomatic number, is defined by
(5.5) Theorem 5.6 (Acharya and Las Vergnas [I]). The cyclomatic number of a hypergraph H equals zero if and only if its edges, maximal with respect to inclusion. are the maximal cliques of a triangulated graph. Theorem 5.7 (Duchet [16]). Euery triangulated graph is representative of a family of convex polygons in the plane. Proof. Choose a tree T with vertex set X. T admits a planar embedding such that the points corresponding to vertices of T are exactly the vertices of a convex polygon. Denote by f : X + R Z t h e vertex embedding. The reader can easily verify that, for every two subtrees A and B of T, we have A f l B # 0 if and only if Conv ( f ( A)) f l Conv ( f (B)) # 0,where Conv denotes the ordinary convex hull. The conclusion follows by Theorem 5.3. 0 An interesting problem o n representative graphs of subtrees has not yet been completely solved:
Theorem 5.8 (Renz [67]). G is the intersection graph of a family of paths in a tree if and only if G is triangulated and is the intersection graph of a family F of paths of a graph such that F satisfies the Helly property.
Classical perfect graphs
83
6. Graphical properties of triangulated graphs
Various properties of triangulated graphs have been proposed. The most important one was discovered by Dirac [14] who first proved the existence of a simplicia1 vertex. This fact is at the origin of an elimination scheme in solving linear systems (see Rose [70] and Golumbic [36]). Let us give a few basic definitions: Definitions. Asimplicial vertex of a graph G is a vertex all of whose neighbours are adjacent. A minimal relative curser (relative to x and y ) is a subset of vertices whose suppression disconnects two fixed vertices of the graph ( x and y ) and which is minimal for the inclusion order with this property. In Fig. 10, { a , b } is a minimal relative cutset. A T-orientation (called ‘monotone transitive orientation’ in Rose [70]) of a graph G is an orientation of the edges of G such that the following two conditions are fulfilled: (1) The orientation has no circuit. implies
b-c
b +.a.
A path a = x , . . . x,, = b from a to b is named a minimal path when it has no chords; no edge in the graph is of the form xixj with i + 1 # j. Three technical, but simple lemmas will be useful. In what follows, G denotes a triangulated graph. Lemma 6.1. Every closed walk x 1 * * . x, . . . x,xI with n x , x , + ~or a repetition x , = x , + ~ .
Fig. 10.
32
= ( X ,E
)
has a triangular chord
84
P. Ducher
Lemma 6.2. Let a and b be two neighbours of a vertex x in G. lf x i = a . . . x, . . . x,, = b is a minimal path not passing through x, then all x, are neighbours of x.
Lemma 6.3. Suppose V and C are disjoint subsets of vertices in G. Suppose V is connected and C is a maximal clique. Then C contains a vertex not adjacent to V. Proofs. Lemma 6.1 and 6.2 are easy to verify and left to the reader. To prove Lemma 6.3, choose in V a subset W that is maximal with the properties that W is connected. and C contains a vertex s not adjacent to W. If W # V, consider in V \ W a vertex u having a neighbour w in W. C contains a vertex t not adjacent to t‘. Clearly, t is distinct from s and has a neighbour w’ in V. Moreover, s is adjacent to u. A minimal path uw . * w , . * . w‘t, where the w,’s are in W , contradicts Lemma 6.2. 0
Theorem 6.4. Each of the following assertions is a necessary and sufficient condition for a graph G = ( X , E ) to be triangulated: (1) (Dirac [14]) Every minimal relative cutset of G is a clique. (2) (Frank-Kas [29]) Every induced subgraph of G is either a clique or contains two non -adjacent simplicia1 vertices. ( 3 ) (Dirac [ 141, Lekkerkerker and Boland [%I) Every induced subgraph has a simplicia1 vertex. (4) (Rose [70]) There exists a n order x , , . . . ,x, of vertices of G such that every x, is a simplicia1 vertex of the induced subgraph G, , x m , for i = 1 , . . .,n. ( 5 ) (Rose [70]) G admits a T-orientation of its edges. (6) (Duchet [ 151) Every connected subgraph with p 2 2 vertices contains at most p - 1 cliques. The formulation of (1) makes precise a result of Hajnal and Surangi [48], who have shown a weaker property of triangulated graphs (sufficient to prove the perfectness): every minimal cutset is a clique.
Proof of Theorem 6.4. The proofs of (1) and ( I ) 3 (2) are simplified versions of cited proofs. The implication (5) j (6) is new. Let x, y be two non-adjacent vertices of a triangulated graph G, S a minimal cutset relative to x and y. and a, b two non-adjacent vertices of S. If C, and C, denote the connected components of Gx-s containing, respectively, x and y. there exists, by minimality of S, a minimal path from a to b in C, U { a ,b } and a minimal path from a to b in C, U { a , b } . These paths form a cycle of length 2 4 with no chord. This contradiction proves (1).
Classical perfect graphs
85
(1) j(2) is shown by induction on the number of vertices. If x and y are non-adjacent vertices of G, if S is a minimal cutset relative to x and y , and if C, and C, are the associated connected components, then the induced subgraph GcXushas two simplicial vertices that are not adjacent and one of them is a simplicial vertex of G. The same argument holds for Gcyus. (2) j(3) 3 (4) j ( 5 ) are trivial. (5)j (6) does not present any complication. Let G denote a T-orientation of G. The sets C, = { x } U TC(x)are cliques of G. Conversely, a clique of G induces a transitive tournament in G and so is contained in a set of the form C,. Thus, all the maximal cliques in G are of the form C,. A source in G does not produce a maximal clique. The conclusion follows. (6) implies trivially that every cycle of G, with length 3 4, has a chord. 0
Remark. The ratio k ( G ) / a ( G )where , k ( G )is the number of maximal cliques and a ( G ) is the stability number, is not bounded for triangulated graphs, as shown by the graph Gn,"having 2n vertices a l,..., a,,, b , , . .. , b,, and the following edges: a,a, for all i, j , a,b, for i < j , 6,b, for all i, j. For this graph, k(G,.,)= n and a ( G )= 2. In the following corollaries, G
= ( X ,E
) is a triangulated graph.
Corollary 6.5 (Lekkerkerker and Boland [58]). For every subset A of X , the subset X \ A \ r G ( A ) is either empty or contains a simplicial vertex of G. Proof. The proof is easy by 6.4(1) and 6.4(2). 0 Corollary 6.6 (Laskar and Shier [56]). For every vertex x of G there exists a simplicial vertex y at maximum distance :
d ( x , y ) = max d ( x , z ) . 1€X
Proof. The proof is by induction on the maximum distance, using the above corollary. 0
Various properties involving distances and centers of triangulated graphs have
P. Ducher
86
been investigated by Laskar and Shier [56]. They conjectured that every odd power of a triangulated graph is triangulated. In fact, we prove a slightly stronger result. In the following theorem, G k denotes the k-th power of G, that is, the graph with the same vertices as G, two vertices being adjacent in G' when their distance in G is at most k. Theorem 6.7 (Duchet). Let G be a graph. If G' is triangulated, so is G k + 2 .
Lemma 6.8. If G = ( V , E ) is a triangulated graph, and if VI,. . ., V, are connected subsets of V, G ( VI,. . . , Vm)is also a triangulated graph. Proof of Lemma 6.8. Let c = Cl . . . ckcl be a cycle with length 3 4 in G ( V,, .- ., V , ) . A closed walk uI * * up of G is called a C-walk if and only if uI"'up=
WI &
wz &
... &
Wk
& WI
(& denotes the concatenation of sequences) where W, is a non-empty walk in the subgraph induced by G over C,. W, is called the i-th component of the C-decomposition W , ,. . . , W,. Consider a C-walk W with minimal length p . Denote by u I , . . . ,v,, the vertices of W and W , ,. . . , wk the C-decomposition of W. By Lemma 6.1, W must contain two vertices u,, and uqr2 such that
u ~ ,= uqA2 or
u, is adjacent to uq+?.
The minimality of W implies that u, and uq+2are in different components W, and W, with
Ia-PI#l,
(wP)#(k,l)
and
(a,P)#(l,k).
Therefore W., and W, are linked in G ( V l , .. ., V , ) and the lemma is proved. 0 Proof of Theorem 6.7. f U I . . . . ,u, are vertices of G, Gk" is clearly isomorphic to G k ( V,..., l V m )where Vi = r G ( v i ) U { v i } So, . we are done. 0 Definition. A block graph is a graph whose blocks are cliques.
Corollary 6.9 (Jamison [53]). If B is a block graph, B k is triangulated for all k 3 1. Corollary 6.10. If G is a triangulated graph, so is G"" for euery integer k. The property fails for even powers (see [56]).
Classical perfect graphs
87
7. Representations by intervals
An interval hypergraph is a hypergraph isomorphic to a (finite) family of intervals of N. In other words, there exists a linear order o n the vertices of H such that every edge of H is an interval for this order. Algorithmic characterizations have been proposed by Fulkerson and Gross [31], Eswaran [20], and others. Theoretical characterizations are due to Tucker [79], Duchet [15], [ 161, Fournier [27], and Nebesky 1621. The following form seems to be most pleasant: In a hypergraph, a vertex a is said to lie between vertices b and c when every path joining b to c contains an edge containing a. Theorem 7.1 (Duchet [15]). A hypergraph H is an interval hypergraph i f and only if the following condition is fulfilled:
For every three vertices H, one of them lies between the other two.
(1)
Proof. The proof is similar to that of Theorem 5.1 and is much simpler than in 1271, 1151, [621. The following lemma is useful. Lemma 7.2. Let E, F be two intersecting edges of a hypergraph H that satis,fies ( I ) . Then, i f we add E U For E f l F as a new edge of H, the resulting hypergraph also satis,fies ( I ) . If E,C F and F Z E, a similar property holds with E \ F. The verification of this lemma is straightforward and left to the reader. Proof of Theorem 7.1 (continued). Now let H satisfy (I). Among all hypergraphs that satisfy (I) and that contain H as a partial hypergraph, we choose one, say H’, with a maximal number of edges. Let T be the hypergraph whose edges are the minimal edges of H’ with at least two vertices. We are going to show that T is a path and that H is a family of subintervals of this path. Let E be an edge of H’, I E 122, and let a, b be two vertices of E. If E # { a , b } , the maximality of H ‘ implies the existence of three vertices x, y, z and three paths xEl * . . E,y, yFI . * Fpz, zG1* . . G,x satisfying
ZP
Ifi E,,
,=I
XE
,=I
F,, z E
u G,.
k=i
{ a , b } is one of these edges. Moreover, we may suppose the following:
P. Duchet
88
(*) Property (I) holds when any occurrence of { a , b } in the sequence E, . . . F, . . Gk is replaced by every edge of H' that contains both a and b. Thus, changing { a , b ) to E, we have, for instance, that z lies between x and y . Hence a, h. z are different, { a , b } is some F, and z E E. By Lemma 7.2, there are edges A and B of H' such that
a E A, z P A, and A contains, for instance, x. b E B,
Z PB,
and B contains
Y-
Moreover, b e A, otherwise A may replace E, = { a , 6) in t h e path joining x and y, in contradiction to our assumption (*). Similarly a @ B. Thus E \ B and E \ A are edges of H' connecting a and b. We may therefore conclude with the following statements: (1) If E is an edge of T, E = { a , b } . Hence, T is a graph. (2) Every pair of vertices of an edge E in H' are linked by a path in H' whose cdges are included in E. X being an edge of H'. T is connected and, by (I) applied to T, is a path. By induction on I E I , every edge of H' is a connected subset of T. 0
Theorem 7.3 (Tucker [79]). His an interval hypergraph if and only if it does not contain as induced subhypergraph one of the list in Fig. 11. Tucker's proof is long. Using the characterization of interval graphs by Lekkerkerker and Boland [%I, whose proof is also long, Trotter and Moore [77] gave a short proof. We give here a direct short proof as a consequence of the above characterization.
c,
n r 3
01
N,
M, n > 1
n> 1
Fig. 11.
Chsical perfect graphs
89
Proof of Theorem 7.3. The hypergraphs of Fig. 10 are clearly not interval hypergraphs. Conversely, suppose H = ( X , ‘8‘) is not an interval hypergraph, every proper subhypergraph of H being an interval hypergraph. Let x be a vertex of H. Since H does not satisfy (I), there exist two vertices y and z in X \ x and three paths xEl . . . E,y, xF, . . . F,z, y G 1 .. . G,z with the properties
XP u k=l
G,,
ye 6 F,, /=I
ZE
6 E,.
,=I
We choose p and q as short as possible over y, z and all such systems of paths (x remaining fixed). If El n Fl = (?,H contains some cycle C,,. If El C F, (the case F, C El follows by symmetry), we easily find 01,O2or N ,
as induced subhypergraph. If El f l F1# 13, FIf El and E l ( F1# (?, the minimalities of H, p and q imply y E E, \ FI and z E FI\ El. The subhypergraph induced by H over the set Gk U (El f l F , ) contains C, or M,, as partial subhypergraph. 0
uL=,
An interesting property of families of intervals was discovered by Fulkerson and Gross [31], namely, that an interval hypergraph is completely determined by its ‘intersection pattern’. Theorem 7.4 (Fulkerson and Gross [31]). Let (E,),E,be a ,finite family of intervals of N, and let (E),El be a family of sets. Suppose J E ,n El 1 = ( E f l F, 1 forall i, j E I. Then the families E, and E are isomorphic.
(7.5)
This theorem was considerably strengthened to more general hypergraphs by Fournier [26]. Nevertheless, the problem of characterizing the matrices 1 E, f l El I of families of intervals is still open. Interval hypergraphs are unimodular (Berge [6]). A nice min-max relation for intervals was recently settled by Gyori [46]. Other kinds of representation via intervals have been proposed, such as ‘D-interval hypergraphs’ by Moore [61], ‘dotted-intervals’ by Duchet [ 151, and ‘interval representability’ by Trotter and Moore [77]. See also Gyarfis and Lehel [451. 8. Interval and indifference graphs
Hajos [49] was the first to note the deep interest of ‘intersection graphs of families of intervals of the real line’. Hajos’ graphs became interval graphs and
!MI
P. Ducher
play a part in various applied fields like psychology (distinguishability in hierarchies), zoology (place of species in animal evolution), genetics (gene as segment of a chromosome), archeology (seriations), geology (classifications), criminology (see [9]), etc., in fact, wherever a chronological ordering is deduced from data which only indicates contemporaries. See [3], [42], [43], [44]. [60], [4], [54], [69], [7S]. Interval graphs have many interesting properties. They are unirnodular graphs (see [S]), triangulated graphs (Hajos [49]), and their complements are comparability graphs [49]. A first characterization may be formulated in terms of betweenness, as follows. In a graph G, a vertex x is said to lie between the vertices y and z when every path joining y to z must contain x or a neighbour of x.
Theorem 8.1 (Lekkerkerker and Boland [%I). A graph G is an interval graph if and only if it satisfies the following two conditions: (i) G is n triangulated graph. (ii) Among every three vertices of G, at least one is situated between the others. Moreover, condition (ii) may be replaced by the following weaker form: (ii') Among every three simplicia1 vertices of G at least one is situated between the others. Theorem 8.2 (Lekkerkerker and Boland [%I). A necessary and sufficient condition for a graph G to be a n interval graph is: (iii) G does not have as induced subgraph one of the graphs listed in Fig. 12. Proofs of Theorems 8.1 and 8.2. The class of interval hypergraphs is clearly an R-class (see Section 2) and Lemma 2.3 can be applied. Conditions (i) and (ii) are clearly necessary for a graph to be an interval graph. Conditions (i) and (ii') together obviously imply (iii). For proving the sufficiency of (iii), suppose G is a triangulated graph but not an interval graph. Then, by Theorem 7.3, the hypergraph C*(G ) possesses as induced subhypergraph one of the hypergraphs O , , O?, M,, N,, C,, (see Fig. 11).
Fig. 12.
91
Classical perfect graphs
Thus, applying Lemma 6.3, we have the following: If C*(G) contains 01,G contains A 2 or D2. If C*G contains 02,G contains A 1 or Dz or D3. If C*G contains M,, G contains some D,. If C*G contains N., G contains some E,. If C*G contains C,,, G contains E , or Ez. 0 Corollary 8.3 (Gilmore and Hoffman [35]). G is an interval graph if and only if G is a triangulated graph and the complementary graph is a comparability graph. Proof. None of the graphs A l , A2, 0, and Ep is a comparability graph. This proves the ‘if’ part. It is not difficult to verify that the conditions are necessary. If G = I!,(%‘), where 8 is a family of intervals in N, put I < J for two members I and J of 8 when
InJZ0, (VX E I ) ( V y E J ) x < y. The comparability graph associated with the partial order < is
G. 0
Such orders are termed interval orders and have been characterized by Fishburn [22] as partial orders < that satisfy
(8.4) ‘‘x < y and z < t” implies ‘‘x d t or z < y for every element x, y , z, t. ”
Corollary 8.5 (Duchet [15]). G is a n interval graph if and only if it has no induced C, and if for every closed walk x I * x,xl with n 3 5 there exist partitions A = {xPcl. . . x q } , B = {x,+, * * x,} (the indices are modulo n ) that satisfy both following conditions : (1) Every vertex of A has a neighbour in €3. (2) Every uertex of B has a neighbour in A. Corollary 8.6 (Kotzig [55]). The connected bipartite interval graphs are exactly the caterpillars ( = trees that become paths by removing the pendant vertices).
The proofs are left as exercises. We cannot conclude this section without mentioning an important family of interval graphs, the indifference graphs. Definition. Let 9 be any family of intervals of the real line and X be a finite
P. Duchet
92
subset o f points of the line. Link x and y by an edge if some member of 5!? contains both x and y ("indistingiiishability"). This forms a graph on %. The indifference graphs are all the graphs that can be obtained in this way. An equivalent definition (using the tcrminology of Berge [7]) is the 2-section of an interval hypergraph. In the context of Decision Science, Luce [60] introduced the notion of a semi-order.
Definition. A partial order < on a set X is called a semi-order if there exist a real number d > 0 and a real-valued function n : X - R with the following property: x < y e Lf(X)B u(y)+d. A necessary and sufficient condition for < to be a semi-order is (Scott and Suppes 1731) that < satisfies both condition (8.4) and the following condition: (8.7) "x
i y
and y < z"
+ "x < t or t <
2''
for every element x, y , z,
1.
This was rediscovered by Roberts in the following form:
Theorem 8.8 (Roberts [6X]). Each of fhe following conditions is necessary and sufficient for a graph G to be a n indifference graph: ( 1 ) G is an interval graph and does not contain K , . 3 (Fig. 13) as induced subgraph. ( 2 ) G is the representative graph of a family of inlervals pairwise uncomparahle by inclusion. (3) G is rhe representative graph of a family of intervals of equal length. (4) G does not have K , , ?or Dr or E l as induced subgraph (see Fig. 13) (Wegner ~31). ( 5 ) G is the comparability graph associated with a serni-order. Two others characterizations are interesting:
K1,3
E2
Fig. 13.
Classical perfect graphs
93
Corollary 8.9 (Duchet [ 151). G = ( x , E ) is an indifference graph if and only if the family of neighbours in G, that is, the family ( V ( x )= 1, ( x ) u { X ) ) , , X ,
constitutes an interval hypergraph. Corollary 8.10 (Duchet [15]). G = ( X , E ) is an indifference graph if and only i f G admits an orientation G that satis,’ies both the following conditions: ( 1 ) G has no circuit, (2) G ( x ) and &(x) are complete graphs.
An orientation satisfying 8.10(1) and 8.10(2) is a T-orientation (see Section 6) whose reverse orientation is also a T-orientation. Proofs of Corollaries 8.9 and 8-10. The ‘only if’ part of Corollary 8.9 results from our definition of indifference graphs. Conversely, if V ( X ) , =is~a family of intervals of X ordered by a total order c ,put V + ( x )= { y E V ( x ) ;x S y }
and
V - ( x )= { y E V ( x ) ;y c x}.
x and y are adjacent in G if and only if
v+(x)
n V + ( Y#) 0
or, equivalently, if
v - ( xn) v-(y) z 0. This shows that G is an interval graph. Obviously, K,,?is not a subgraph of G. This proves Corollary 8.9, and the ‘only if’ part of Corollary 8.10. The remainder of the proof is left as an exercise. 0 Fine and Harrop [21] have given a necessary and sufficient condition for an indifference graph to represent the ‘indistinguishability’ by intervals of the same length.
References [ l ] D. Acharya and M. Las Vergnas, Hypergraphs with cyclomatic number zero, triangulated
graphs, and an inequality, J. Comb. Theory, Ser. B (to appear). [ 2 ] J.-C. Arditti, Partially ordered sets and their comparability graphs, their dimension and their adjacency, Proc. Coil. Int. CNRS no 260 (Orsay 1976), Problkmes Combinatoires et ThCorie des Grdphes, Paris (lY78). (31 S. Benzer, On the topology of the genetic fine structure, Proc. Nat. Acad. Sci. U.S.A. 45 (1959) 1607- 1620. [4] S. Benzer, The fine structure of the gene, Sci. Amer. 206 (1962) 7&84.
94
P. Duchet
[5] C. Berge. Les problemes de coloration en thtorie des graphes, Publ. Inst. Statist. Univ. Paris 9 (1960) 123-160. [6] C. Berge, Some classes of perfect graphs, Graph Theory and Theoretical Physics (Academic Prcss, London. 1967) 155-165, MR 38 # 1017. [7] C. Berge, Graphs and Hypergraphs (North-Holland, Amsterdam, 1973). [8] C. Berge. Perfect graphs, in: D.R. Fulkerson, ed., Studies in Graph Theory, Part I, MAA Studies in Mathematics 11 (1973) 1-22. [91 C. Berge, Les graphes d'intervalles, in: P. Hansen and D. de Werra, eds. Proc. Cerisy Colloquium, 1980 (Presses Polytechn. Romandes, Lausanne, 1980). [lo] P. Buneman, A characterization of rigid circuit graphs, Discrete Math. 9 (1974) 205-212. ( 111 V. Chvatal, Edmonds polytopes and a hierarchy of combinatorial problems, Discrete Math. 4 (1973) 305-337. ( 121 V. Chvatal and P. Hammer, Aggregation of inequalities in integer programming, Ann. Discrete Math. 1 (1973) 145-162. [13] R.P. Dilworth. A decomposition theorem for partially ordered sets, Ann. Math., Ser. 2 51 (1950) 161-166. [ 141 G.A. Dirac, On rigid circuit graphs, Abh. Math. Sem. Univ. Hamburg 25 (1961) 71-76. [IS] P. Duchet, Representations, noyaux en thtorie des graphes et hypergraphes, Doct. Diss. Univ. Paris 6 (1979). [ 161 P. Duchet, ProprietC de Helly et problemes de representation, Problemes combinatoires et theorie des graphes, Proc. lnt. Vol. CNRS 260 (Orsay 1976), Paris (1978) 117-1 18. [I71 P. Duchet, Combinatorial properties of subtrees of a tree (in preparation). [I81 B. Dushnik and E.W. Miller, Partially ordered sets, Amer. J. Math. 63 (1941) 600-610. I191 J . Edmonds, Path trees and flowers, Canad. J. Math. 17 (1965) 44-67, [20] K.P. Eswaran, Faithful representation of a family of sets by a set of intervals, SIAM J. Comput. 4 (1975) 56-68, MR51 #14677. [21] N. Fine and R. Harrop, Uniformization of linear arrays, J. Symbolic Logic22 (1957) 130-139. [22] P.C. Fishburn, Betweenness, orders and interval graphs, J. Pure Appl. Alg. 1 (1971) 159-178. [23] C. Flarnent, Hypergraphes arbor&, Discrete Math. 21 (1978) 223-226. (241 S. Foldes and P.L. Hammer, Split graphs, Proc. 8th South-Eastern Conference on Combinatorics, Graph Theory and Computing (1977) 31 1-315. [24a]S. Foldes and P.L. Hammer, Split graphs having Dilworth number two, Canad. J. Math. 39 (1977) 6-72. [2S] L.R. Ford and D.R. Fulkerson, Flows in Networks (Princeton Univ. Press, Princeton, N.J., 1962). (261 J.C. Fournier. lsomorphisme d'hypergraphes par intersections equicardinales d'arites et configurations exclues, J. Comb. Theory, Ser. B 29 (1980) 321-327. [27j J.C. Fournier, Hypergraphes de chaines d'aretes d'un arbre, Discrete Math. 43 (1983) 29-36. [28] A. Frank, On chain and antichain families of a partially ordered set, J. Comb. Theory, Ser. B 29 (1980) 17&184. 1291 A. Frank, Some polynomial algorithms for certain graphs and hypergraphs, Proc. 5th British Comb. Conf.. Aberdeen (1975) 211-226. (301 D.R. Fulkerson, Note on Dilworth's decomposition theorem for partially ordered sets, Proc. Amer. Math. SOC.7 (1956) 701-702. [31]D.R. Fulkerson and O.A. Gross, Incidence matrices and interval graphs, Pacific J. Math. 15 (1965) 835-855. (321 T. Gallai. Transitive orienterbare Graphen, Acta Math. Sci. Hung. 18 (1967) 25-66. [33] F. Gavril, The intersection graphs of subtrees in trees are exactly the chordal-graphs, J. Comb. Theory. Ser. B 16 (1979) 47-56. 1341 A. Ghouila-Houri, Caracterisation des graphes non orientes dont on peut orienter les arites de maniere a obtenir le graphe d'une relation d'ordre, C.R. Acad. Sci. Paris 254 (1962) 1370-1371. 1351 P.C. Gilmore and A.J. Hoffman, A characterization of comparability graphs and of interval graphs, Canad. J. Math. 16 (1979) 47-56.
Classical perfect graphs
95
[36] M.C. Golumbic, Threshold graphs and synchronizing parallel processes, in: A. Hajnal and V.T. Sos, eds., Combinatorics, Coll. Math. Soc. J. Bolyai 18 (1978) 419-428. [37] M.C. Golumbic, Algorithmic Graph Theory and Perfect Graphs (Academic Press, New York. 1980). [38] C. Greene, Some partitions associated with a partially ordered set, J. Comb. Theory, Ser. A 20 (1976) 69-79. [39] C. Greene, Sperner-families and partitions of a partially ordered set, in: M. Hall and J.H. van Lint, eds., Adv. Study Inst. Series C, Math and Phys. Sci., Vol. 16, Brenkelen (1974). [40] C. Greene and D.J. Kleitman, The structure of Sperner k-families, J. Comb. Theory, Ser. A 20 (1976) 41-68. [41] J.R. Griggs, On chains and Sperner k-families in ranked posers, J. Comb. Theory (to appear). [42] J. Guex, Une nouvelle mtthode d’analyse biochronologique, Note prtliminaire, Bul. SOC. Vaud. Sci. Nat. no 351, Vol. 73, Lausanne (1977) 309-321. [43] J. Guex, Terminologie et mtthodes de la biostratigraphie moderne: commentaires critiques et propositions, Bul. SOC.Vaud. Sci. Nat., no 355, Vol. 74, Lausanne (1979) 169-216. [44] J. Guex, Datations paltontologiques et graphes d’intervalle, in: P. Hansen and D. de Werra, eds., Regards sur la thtorie des graphes, Col. de Cerisy 1980 (Presses Polytechn. Romandes, Lausanne, 1980) 243-248. [45] A. Gyarfas and J. Lehel, A Helly type problem in trees, Combinatorial Theory and its Applications, Coll. Math. SOC.J. Bolyai 4 (1970) 571-584. [46] E. Gyori, A minimax theorem for intervals, J. Comb. Theory, Ser. B (submitted). [47] R. Gysin, Dimension transitiv orienteerbaren graphen, Acta Math. Sci. Hungar. 29 (1977) 313-316. [48] A. Hajnal and J. Suranfi, Uber die Auflosung von Graphen in vollstandige Teilgraphen, Ann. Univ. Sci. Budapest Eotvos. Sect. Math. 1 (1958) 113-121. [49] G. Hajos, Uber eine Art von Graphen, Int. Math. Nachr. 11 (1957) problem 65. [SO] P.L. Hammer and B. Simeone, The splittance of a graph, Combinatorica 1 (1981) 375-384. [51] P.B. Henderson and Y.Zalcstein, A graph theoretic characterization of the PV chunk class of synchronizing primatives, SIAM J. Comp. 6 (1977) 88-108. [52] A.J. Hoffman and D.E. Schwartz, On partitions of a partially ordered set, J. Comb. Theory, Ser. B 23 (1977) 3-13. [53] R. Jamison, Powers of block graphs are chordal (submitted). [54] D.G. Kendall, Incidence matrices, interval graphs and seriation in archeology, Pacific J. Math. 28 (1969) 565-570. [55] A. Kotzig, Paare Hajos the Graphen, Casop. Pest. Mat. 88 (1963) 23G241. [56] R. Laskar and D. Shier, On powers and centers of chordal graphs, Rep. Clemson Univ. (submitted). [57] B. Leclerc, Arbres et dimension des ordres, Discrete Math. 14 (1976) 69-76. [58] C . Lekkerkerker and J. Boland, Representation of a finite graph by a set of intervals on the real line, Fund. Math. 51 (1962) 45-64. [59] L. Lovisz, Normal hypergraphs and the perfect graph conjecture, Discrete Math. 2 (1972) 255267 (this volume, pp. 29-42). [60] R.D. Luce, Semiorders and a theory of utilitydiscrimination, Econometrica 24 (1956) 178-191. [61] J.I. Moore, Graphs and partially ordered sets, Ph.D. Thesis, Univ. of South California, Columbia (1975). [62] L. Nebesky, On a certain numbering of the vertices of a hypergraph 11, Czechoslovak Math. J. (1981). [63] U.N. Peled, Matroidal graphs, Discrete Math. 20 (1977) 263-286. [64] M.A. Perles, On Dilworth’s theorem in the infinite case, Israel J. Math. 1 (1963) 108-109. [65] A. Pnueli, A. Lempel and S. Even, Transitive orientation of graphs and identification of permutation graphs, Canad. Math. J. 23 (1971) 160-175. [66] 0. Pretzel, On the dimension of partially ordered sets, J. Comb. Theory, Ser. A 22 (1977) 146-152.
96
P. Duchef
[67] P.L. Renz, Intersection representationsof graph by arcs, Pacific J. Math. 34 (1970) 501-510. [a] F.S. Roberts, Indifference graphs, in: F. Harary, ed., Proof Techniques in Graph Theory (Academic Press, New York. 1969) 139-146. [h9] F.S. Roberts, Measurement Theory, with Applications to Decision-Making, Utility and the Social Sciences (Addison-Wesley, Reading, Massachusetts, 1979). [70] D.J. Rose, Triangulated graphs and the elimination process, J. Math. Anal. Appl. 32 (1970) 597-609. [71] D. Rotem and J. Urrutia, Circular permutation graphs, Research Rep. Univ. of Waterloo (submitted). [72] D. Rotem and J. Urrutia, Comparability graphs and intersection graphs, Discrete Math. 43 (1983) 37-46. 1731 D.S. Scott and P. Suppes, Foundation aspects of theories of measurement, J. Symbolic Logic 23 (19%) 113-128. [74] L.N. Shevrin and N.D. Filippov, Partially ordered sets and their comparability graphs, Siberian Math. J. 11 (1970) 497-509. 1751 F.W.Stahl, Circular genetic maps, J. Cell. Physiol., Suppl. 70 (1967) 1-12. [76) W.T. Trotter. A note on Dilworth’s embedding theorem, Proc. Amer. Math. Soc. 52 (1975) 3s39. [77] W.T. Trotter and J.I. Moore, Characterization problems for graph partially ordered sets, lattices and families of sets, Discrete Math. 16 (1976) 361-381. [78] W.T. Trotter, J.l. Moore and D.P. Sumner, The dimension of a comparability graph, Proc. Amer. Math. SOC.60 (1976) 35-38. [79] A. Tucker. A structure theorem for the consecutive 1’s property, J. Comb. Theory 12 (1972) 15S162. [so] H. Tverberg, On Dilworth’s decomposition theorem for partially ordered sets, J. Comb. Theory 3 (1967) 305-306. [Sl] J.R. Walter. Representations of rigid cycle graphs, Ph.D. Thesis, Wayne State Univ. (1972). (821 J.R. Walter, Representation of chordal graphs as subtrees of a tree, J. Graph Theory 2 (1978) 265-267. [83] G. Wegner, Eigenschaten der Nerven Homologische-einfacher Familien in R“. Ph.D. Thesis, Gottingen (1967). [& E.S. I] Wolk, The comparability graph of a tree, Proc. Amer. Math. SOC.3 (1062) 789-795. [85]E.S. Wolk, A note on the comparability graph of a tree, Proc. Amer. Math. SOC.16 (1965) 17-20.
Annals of Discrete Mathematics 21 (1984) 97-101 Elsevier Science Publishers B.V.
THE PERFECT GRAPH CONJECTURE FOR TOROIDAL GRAPHS* Charles GRINSTEAD Department of Maihematics, Swarthmore College, Swarthmore, PA 19081, USA
1. Introduction
In what follows, we assume that our graphs are finite without loops or multiple edges. We define w ( G ) to be the size of the largest complete subgraph of G, and r ( G ) to be the vertex coloring number of G. We say that G has property P if neither G nor G (the complement of G ) contains an odd chordless cycle of length at least five as an induced subgraph. Claude Berge defined a graph G to be perfect if w ( H )= y ( H ) for all induced subgraphs H of G. He then made the following conjecture, which has never been proved. The Perfect Graph Conjecture. A graph G is perfect if and only if G has property P. Define a graph G to be critical if G is not perfect but all proper induced subgraphs of G are perfect. In 1972, Lovasz [3] proved that G is perfect if and only if G is perfect (the Weak Perfect Graph Conjecture of Berge). He did this by showing that G is perfect if and only if ( H I C w ( H ) w ( H ) for all induced subgraphs H of G. From this it follows that if G is critical, then 1 G 1 = w ( G ) w ( G ) + 1. In 1974, Padberg proved the following theorem (see [4]). Theorem 1 (Padberg). A critical graph G with n vertices has exactly n cliques of size w ( G ) with each vertex in w ( G ) maximal cliques and has exactly n independent sets of size w ( G )with each vertex in w ( G ) maximal independent sets of G. Each maximal clique of G intersects all but one maximal independent set of G, and vice versa. Alan Tucker has proved the following two results (see [5],[6]). * Reprinted from J. Comb. Theory, Ser. B 30 (1981) 7C74, @ 1981, Academic Press. 97
C. Grinstead
98
Theorem 2 (Tucker). If G is planar and G has property P, then G is perfect.
Theorem 3 (Tucker). If o(G) = 3 and G has property P, then G is perfect. We remark that if w ( G ) = 2 and G has property P, then G is bipartite, hence perfect. In this paper we shall prove that if G is toroidal (i.e., embeddable in the torus) and G has property P, then G is perfect.
2. Proof of the result
In characterizing critical toroidal graphs G, we can assume that o(G)3 4, by Theorem 3. The following two lemmas show that in fact we can assume that w ( G ) = 4.
Lemma 1. In a critical graph G, if there exists a vertex of degree at most six, then w(G)<4.
Sketch of proof. The set of neighbors of a vertex in a critical graph generates a subgraph which contains w ( G ) cliques of size w ( G ) - 1, but no cliques of size w ( G ) (see Theorem 1). No graph with fewer than seven vertices contains five cliques of size four, and no cliques of size five. Hence, w ( G ) s 4 . 0 Lemma 2. In a critical toroidal graph G, either w ( G) < 4, or G is regular of degree
six and triangulates the torus.
Proof. Let bi be the number of vertices of degree i in G, and let ri be the number of regions with i edges. By a standard counting argument using Euler’s formula, we get x
I
4b2 + 3 b , + 2 b 4 + bs =
,=7
( i -6)b, t
i =4
( 2 j -6)r,.
If there exists a vertex of degree less than six, then by a similar argument to that used in Lemma 1 , we must have w ( G ) < 4 . If every vertex has degree at least six, then the left side of (*) is zero, so b, = 0 for i 3 7, and r, = 0 for j 3 4. This means that G is regular of degree six and triangulates the torus. 0 We now state a theorem due to Altshuler which characterizes triangulations of the torus which are regular of degree six (see [l]).
The Perfect Graph Conjecture for toroidal graphs
II
al,l-t
I1
II
II
a
a1,2-t a1,3-t
1,s.t-1
11
a
1,s-t
99
II
a
l,s+l-t
Fig. 1.
Theorem 4 (Altshuler). Let G be a triangulation of the torus which is regular of degree six. Then G can be represented as in Fig. 1, where a,,,+lis identified with al.l for i = 1,2,. . . ,r, and where a,+],,is identical with aid-,, where the second subscript is taken (mods), and t is .fixed. This triangulation is denoted by T(r,s, t ) . Furthermore, s 2 3, r 2 1, and if r S 2, then t f 0, s - 1 (mod s). Finally,
I G 1 = rs.
We can now prove the following theorem. Theorem 5. If G is toroidal and has property P, then G is perfect. Proof. Assume that G is a toroidal critical graph. From Theorem 3, we may assume that w ( G ) 3 4, so from Lemma 2, we may assume that G is regular of degree six and triangulates the torus. From Lemma 1 we get that o(G)= 4; so from Theorem 1 we get that every vertex is in four cliques of size four. Now assume that r > 3. Then the neighbors of u2,2generate the graph H shown in Fig. 2, where not all adjacencies are shown at this point. Since r > 3,
Fig. 2.
100
C. Grinstead
neither a a , nor a3,zis adjacent to either u12or a1.3. Also, a2,1is not adjacent to a3.2,and al.2is not adjacent to a23. Since a2.2must be in four cliques of size four and no cliques of size five, H must contain four triangles and no cliques of size four. It is easy to check that this implies that every vertex in H must be in at least one triangle. This implies that a2I is adjacent to a1.3, for otherwise a2.1would be in no triangle in H. This implies that s = 2, which contradicts Theorem 4. If r = 3, then rs = I G I = 4 w ( G ) + 1; so s must be odd, and since we may assume that w ( G )2 4, from Theorem 3, then we may also assume that s 3 7. But then the vertices al.,,for 1S i S s, form an odd chordless cycle, which contradicts the fact that G is critical. Since I G I is odd, r # 2, so we may assume that r = 1. We may now drop the first subscript on all of the vertices. We have vertices al, az,. . . , us,where a, is adjacent to a, if i - j E D = { k 1, t, 2 ( t + l)},where all arithmetic is modulo s. From Theorem 4, we know that f f 0, - 1 (mod s), and since G is regular of degree six, t f 1 (mod s). We may also assume without loss that 1 < t < s2- 1. If f > 2, then a, is not adjacent to a, or as-l.Since a lmust be in a clique of size four with a,, either al is adjacent to us-,-l or us-, is adjacent to a,+l.Figure 3 shows the neighbors of a,. So, either s - t - 2 or s - 2t - 1 is in D. If s - f - 2 is in D,then it equals either t or t + 1. However, since s is odd, we must have s - t - 2 = t + 1, or s = 2t + 3. In this case, we can multiply all subscripts by 2, since ( 2 , s ) = 1, and get that a, and a, are adjacent if and only if i - j E D‘ = { & 1, k 2, ? 3). We will deal with this case at the end of this proof. If s - 2t - 1 is in D, then either s - 2t - 1 equals 1, t, or t + 1.It cannot equal 1, for s is odd. If it equals t, then s = 3t + 1. In this case, we can multiply all subscripts by 3, since (3, s ) = 1, and get that a, and a, are adjacent if and only if i - 1 E D’. This is the same conclusion as before. If s - 2t - 1= t + 1, then s = 3 t + 2 , so we can multiply all subscripts by 3, and arrive at the same conclusion as before. If t = 2, we get the same conclusion again. So, we are left with the case that G has s vertices and that a,and a, are adjacent if and only if i - j ED’. We may
*
Fig. 3.
The Perfect Graph Conjecture for toroidal graphs
101
assume that s 17, since o(G) 4, o(c)2 4. In this case, it is easy to find an odd chordless cycle. If s = 4m + 1, then the cycle is given by the vertices a,, a3, a+ a,, as, , . . , a4m-,a4m-2,a,. This contradicts the fact that G is critical and completes the proof. 0
References [l] A. Altshuler, Construction and enumeration of regular maps on the torus, Discrete Math. 4 (1973) 201-217. (21 C. Grinstead, The strong perfect graph conjecture for a class of graphs, Ph.D. Thesis, UCLA, Los Angeles (1978). [3] L. LovSsz, A characterization of perfect graphs, J. Comb. Theory, Ser. B 13 (1972) 95-98. [4] M. Padberg, Perfect zero-one matrices, Math. Program. 6 (1974) 180-196. [5] A. Tucker, The strong perfect graph conjecture for planar graphs, Canad. J. Math. 25 (1973) 103-114. [6] A. Tucker, Critical perfect graphs and perfect 3-chromatic graphs, J. Comb. Theory, Ser. B 23 (1977) 143-149.
This Page Intentionally Left Blank
Annals of Discrete Mathematics 21 (1984) 103-113 @ Elsevier Science Publishers B.V
THE PERFECT GRAPH CONJECTURE ON SPECIAL GRAPHS - A SURVEY* Wen-Lian HSU Department of Industrial Engineering and Management Sciences, Northwestern University, Evanston, Illinois, USA We discuss several basic techniques for proving the strong perfect graph conjecture o n special classes of graphs. Our discussion is primarily based on the neighborhood structure of these graphs. By combining these techniques we are able to prove the conjecture for more general graphs. The main classes of graphs we consider are claw-free graphs, 3-chromatic graphs, and (K,-e)-free graphs.
1. Introduction
In this paper all graphs considered are simple, i.e., finite, undirected, without loops or multiple edges. Let graph G = (V, E ) have respective vertex and edge sets V and E. Let a (G) denote the maximum size of a stable set of G and w ( G ) denote the maximum size of a clique of G. Let 8(G) denote the minimum number of cliques which cover G, and y ( G ) denote the minimum number of stable sets which cover G ( y ( G )is also called the chromatic number of G ) .Since two vertices of a clique cannot be in the same stable set, we have w ( G ) s y ( G ) . Similarly we have (Y (G) < 8 (G). A graph G is perfect if w ( H )= 8 ( H )for every induced subgraph H of G. An odd hole is a cordless odd cycle, and an odd anti-hole is a complementary graph of an odd hole. Berge’s (see [l])strong perfect graph conjecture (SPGC) asserts that a graph G is perfect if and only if G contains no odd holes or odd anti-holes. Its resolution has eluded researchers for two decades. One line of attack on this conjecture has been to look at general properties of critically imperfect graphs (p-critical graphs) - graphs that are not perfect but all of whose vertex-induced subgraphs are perfect. The major results here are as follows:
Theorem 1 [9]. A graph G with n vertices is perfect if and only if a(H)w(H)s 1 H 1 for all subgraphs H of G, where 1 H I is the number of vertices in the graph H. Thus a critically imperfect graph has n = CY ( G ) w ( G )+ 1 vertices. * This research was supported by National Science Foundation Grant ECS-8105989 to Northwestern University. in3
104
W.-L.HSU
Theorem 2 [lo]. (Let a = a ( G ) ,w = w ( G ) . ) A p-critical graph G with n vertices has exactly n cliques of size w with each vertex in o maximum cliques and has exactly n stable sets of size a with each vertex in a maximum stable sets. Each maximum clique intersects all but one maximum stable sets, and vice versa. Theorem 3 [2]. For each vertex u of a p-critical graph G, there is a unique partition of G - v into a-stable sets. In particular, the induced subgraph on any two a-stable sets in G - v must be connected. An example discovered independently by Huang [8] and by Chvital et al. [3] shows that the above three theorems alone are not sufficient to prove the SPGC (note that the addition of Proposition 1 to the properties of p-critical graphs will eliminate the counterexample). Therefore further characterization of p-critical graphs are desired. Another approach has been to look for special graphs which warrant a proof of the conjecture. In this paper we will combine several proof-techniques for special graphs and apply them to more general graphs with ‘good’ neighborhood structure. We will also discuss some relationships among various special graphs.
2. Some interesting neighborhood structures
Consider a graph G and a vertex u in G. Let N ( u ) be the set of vertices that are adjacent to u. For convenience, - we also regard N ( u ) as the induced subgraph on the neighbors of u and N ( u ) , its complement. Consider the following four neighborhood structures: (i) N ( u ) is bipartite, in which case u is called a b-vertex. (ii) N(u) is bipartite, in which case u is called a i-vertex. (iii) __ N ( u ) is complete multipartite, in which case u is called a m-vertex. (iv) N ( u ) is complete multipartite, in which case u is called a m-vertex. It is clear that these structures are inherent upon induced subgraphs. Corollary 1 implies that if every vertex of G is one of these four types, then the SPCC is true for G. We will prove several propositions and lemmas which together give Corollary 1. The following lemma is needed in our later proofs.
Lemma 1. A n odd cycle without triangle must contain an odd hole. Proposition 1. I f a p-critical graph G contains a b-vertex u, then G is an odd hole or an odd anti-hole.
The Perfect Graph Conjecture on special graphs
105
Proof. Since N ( u ) is bipartite, a maximum clique containing u has size 9 3 . Hence w ( G ) S 3 and the result follows from Tucker [15]. 0 Proposition 2. If a p-critical graph G contains a m -vertex u, then G is an odd hole or an odd anti-hole. Proof. Suppose N ( u ) is complete K-partite, then a maximum clique containing u has size K + 1. Hence o ( G ) = K + 1 and y ( N ( u ) )= K + 1. There must exist a part of N ( u ) which uses two or more colors. Consider two colors PI, p2 that are used in the same part C,. Since N ( u )is complete multipartite, no other part can use p , or p2. Let GPla be the subgraph of G \ u induced on vertices colored p i and p2.Since G is p-critical by Theorem 3, Galamust be connected. Let P be a shortest (PI,&)-path between two vertices x, y in N ( u ) with colors of x and y being p1and p2 respectively. Since x and y are in C,,they are not adjacent. Since P is the shortest possible such path, no other vertices in P are in N ( u ) . Therefore we have an odd hole uPu in G. 0 A claw is a bipartite graph Kl,3with four vertices vo, v I , u2, v 3 (Fig. 1)such that
the only edges that exist among them are (vO,u,), (vu,vz), ( u o , ~ s ) We . use {v0, ( v l , uz, v3)} to denote a claw rooted at uO.
Fig. 1. A claw
Proposition 3. Consider a p-critical graph G. If every vertex of G is a 6-vertex, then G is an odd hole or an odd anti-hole. ~
~
Proof. For each vertex u in G, N ( u ) is bipartite. Hence N ( u ) does not contain triangles. Thus in G, there can be no claw (induced subgraph) rooted at u. Therefore G is claw-free and the result follows from [12]. 0
Before stating a proposition about the m-vertices, we define the following two types of vertices. Type I. A vertex vo is a Type I vertex if there exist three other vertices 01, u2, ( u t , %), (u3, u*) E E and ( u l , u2)$Z . E (as shown in Fig. 2). It is easy to check that such a vertex is a m-vertex. u3 with (uo, uJ, ( U O , U Z ) , (uo, v3),
106
W.L. HSU
Fig. 2. Forbidden subgraph containing uw
Type 11. A vertex vOis a Type I1 vertex if there does not exist three other vertices u I , v2, v3 with ( V O , v,), ( V O , V Z ) , (v3, vl), (v3, V Z ) , (vt, ~ 2 E) E and (oo, v 3 ) eE (as shown in Fig. 3). We call u a m’-vertex.
Fig. 3. Forbidden subgraph containing uo.
It is clear that if every vertex of G is either a m”-vertexor a fi-vertex, then G is K4\e (K4 taking out one edge) free. Hence we can derive the following proposition from [ 1 I].
Proposition 4. Let G be a p-critical graph. If every vertex of G is either a m-vertex or a mi’-vertex, then G is an odd hole or an odd anti-hole. The distinction between m ‘-vertices and *-vertices will be important when we consider more general graphs in the next section. Previous SPGC proofs on special graphs rely heavily on the uniformity of those special properties (which often follow ChvBtal’s approach [4]) and, hence, are difficult to be used for a general proof of the SPGC. Our method is based primarily on a simplified version of an alternative proof (see [7]) that the SPGC is true for claw-free graphs, which uses the neighborhood approach.
3. Special graphs with mixed neighborhood properties
Theorem 4. Let G be a p-critical graph. If every vertex.of G is either a 6-vertex, a m-vertex or a m’-vertex, then G is an odd hole or an odd anti-hole.
The Perfect Graph Conjecture on special graphs
107
We will prove this theorem by proving several lemmas. Throughout this section we assume the graph G satisfies the assumptions in Theorem 4. Furthermore, we only have to consider those G which contain at least one 6-vertex; otherwise Proposition 4 would apply. Also we can assume w (G) 2 4 by Proposition 1. Let u be a 6-vertex in G. Consider the unique minimum coloring of G \ u (this coloring partitions G \ u into w a-stable sets). Since N ( u ) is bipartite, at most two vertices in N ( u )can have the same color. Let M be the set of edges in N ( u ) between two vertices with the same color. Then M is an edge matching in N ( u ) . Furthermore we have the following lemma. ~
~
Lemma 2. M i s not a maximum matching. __
Proof. Suppose M is a maximum matching. Since vertices in N ( u ) which are not adjacent to a matching edge are colored singly in N ( u ) , we have, by Konig's Theorem,
-
1 maximum stable set in N ( u ) 1 = I N ( u )1 - 1 M 1 = y ( G \ u ) =w
( G\ u)= w(G).
-
Let S be a maximum stable set in N ( u ) of size w ( G ) . Thus S U {u} is a clique of size w ( G ) f 1 in G, a contradiction. 0 Given that M is not a maximum matching, there must exist an augmenting path in N ( u )with respect to M. Let P be a shortest such path. Let the endpoints of P be xo and yo with colors j and k respectively. Then colors j and k are used exactly once in N ( u ) . Let the length __ of P be 2a f 1 (a S O ) . Denote P by x o y 1 x I y 2 x 2 x r y l + l x r + l . y,x,yo in N ( u ) with colors f ( x o ) = j, f ( y o ) = k and f ( y r )= f ( x , ) = i,, 1 G 1 s a. Since P is a shortest augmenting path no xi can be adjacent to y m VO G 1 S a, m > I 1. We will show that an odd hole in G can be constructed using the path P. ~
+
Lemma 3. a 2 2 , i.e., the length of P is greater than or equal to 5. If a =0, i.e., P consists of exactly one non-matching ( x o , y o ) edge in N ( u ) , then there can be no &&)-path connecting xo and y o in G \ u (the existence of such a path would imply an odd hole containing u in G), contradicting Theorem 3. Next we consider the case where a = l . The path P is now x o y l x l y o with f ( x o ) = j , f ( y o ) = k and f(xl) = f ( y l ) = il. Consider the induced subgraph H of G on vertices colored j , k or i, and the vertex u. Clearly w ( H ) = 3. Since we assume
Proof. -
108
W.-L.Hsu
w ( G ) z = 4 , H is a proper subgraph of G. Therefore, H must be perfect and, hence, 3-colorable. This implies that there exists a 3-coloring f’ of H in which f ’ ( x l ) = f ‘ ( y o ) and f ’ ( X o ) = f ’ ( y , ) , contradicting the fact that the coloring in G \ u must be unique (by Theorem 3). 0 ___
By Lemma
a 2 2 . Hence we x o y l x I y 2 x 2 y 3.~. y.xayo ( a 3 2). 3,
Lemma 4. Each vertex in the path P
is
a
can
denote
P
(in
N ( u ) ) by
b-vertex.
Proof. We will prove the lemma by proving several claims. Claim I. xi. 0 S i s a - 1 and (let
= yo) y,, 2 G j
s a + 1 are 6-vertices
Proof. It is sufficient to prove that they cannot be f i or f i ’ vertices. (i) They are not m-vertices. A complementary graph of the one shown in Fig. 2 is shown in Fig. 4.
Fig. 4. Complement of Fig. 2.
To show that a vertex uo in G is not a fi-vertex we only have to show that in there exist vertices u l , u2, v 3 such that, together with uo, they form an induced subgraph. as shown in Fig. 4. Now, for each x,, 0 S i S a - 1 there exist x , + ~ Y, , + ~ and u such that the only edge that exists among x,, x , + ~ ,yli2 and u in G is y,+>). Hence these x,’s are not rfi-vertices. Similarly for each y,, 2 S j s a + 1 there exists Y,-~, x , - ~and u such that the only edge that exists among y,, Y , - ~ ,x , - ~ and u in G is (y,-,,x,-J. (ii) They are not m‘-vertices. Fig. 5 shows the complementary graph of the one shown in Fin. 3.
Fig. 5. Complement of Fig. 3.
The Perfect Graph Conjecture on special graphs
109
Now, for each x,, 0 < i < a - 1 there exist x, Y , + Y~ ,,, ~ and u such that the only is (x,, Y!+~). Hence these x,’s are edge that exists among x,, Y,,~, and u in not fi’-vertices. Similar argument shows that y,, 2 < j < a + 1 are not m’vertices. 0
Proof. It is equivalent to show that they are not in G. Suppose (yl,xz)kZ E. Then we have an induced subgraph in G, as shown in Fig. 6. Now in G, xl, x2, yl, y2
it
12
Fig. 6.
must be connected by some ( i l , &)-path. Let xiQxyZ be a shortest ( i l , i2)-path connecting xI to yz. Q must contain either x2 or yl, for otherwise uxIQy2uwould be an odd hole in G. However, Q cannot contain both x2 and yl, because then ux2(along Q)ylu would be an odd hole in G. Hence assume Q contains only x2 (the case of y l is similar). Now y,xlx2(along Q)y2y3is an odd cycle in G. Since y3 is a 6-vertex, it cannot be adjacent to vertices coloied il other than xI, y l (otherwise this would give a claw). Furthermore if y3 is adjacent to any iz vertex y‘ on Q between x2 and y2 then {y3,( x l , y2, y’)} would be a claw, contradicting that y3 is a 6-vertex. Hence y3xlxz(alongQ)y2y3is an odd hole. This contradiction shows that (yl, x2)E E. By symmetry, we have (xa, ya-l)E E. 0
Claim 3. yl and x, are i-vertices. Proof. In G there exist induced subgraphs containing y,, as shown in Fig. 7. Hence y l is not a m- or fi’-vertex. Similarly we can argue that x, is not a m - or m’-vertex. 0
@-@ Fig. I .
110
w.-L.Hsu
We have shown that every vertex on the path P is a b-vertex. Now let Q 1be a shortest (i,,j)-path connecting x1 to yl in G. QI cannot contain xo, for otherwise uxo(along QI)yIuwould be an odd hole. Let yl’ be the neighbor of yl on Q1.yl’ is colored j.
Lemma 5. (yl’,yz) E E and (yl’, xl) E E. Proof. Suppose (y1’,y2) 6Z E. Then consider the induced subgraph H on vertices colored il, i2 o r j and the vertex u. Being a proper induced subgraph of G, H is perfect. Furthermore i t is easy to check w ( H ) = 3. Hence we can color H using three colors, contradicting that the coloring in G \ u is unique (by Theorem 3). Now suppose (y i, xl) 6Z E. Then in G we have a path uy;xoyIx1y2 whose edge connections are shown in Fig. 8, where u and y2 are b-vertices. Compare Fig. 8 with Fig. 6. We can apply the same argument in the proof for Claim 2 of Lemma 4 to show that Fig. 8 is impossible. Hence, ( y l , x l ) must be in E. 0
Fig. 8.
Similarly we can argue that there exists a vertex x: colored i, such that its connection in G is shown in Fig. 9. Now we are ready for the final conclusion.
Fig. 9.
Proof of Theorem 4. Following the previous discussion in this section, let us consider the odd cycle C* in G,
uy lxoylxly2(along P ) ~ , - ~ y , x . y :u. ~x
To avoid an odd hole in G, C* must contain a triangle by Lemma 1. Since N(u) is bipartite this triangle must contain either y I o r x:. Suppose there is a triangle T containing y I (the case of x: is similar); T cannot contain xl, yl or y2. Since x1 is
The Perfect Graph Conjecture on special graphs
111
only adjacent to y l and y2 in C* the existence of such a T would imply a claw (T)} rooted at x1 in G. Hence C*contains an odd hole. By the p-criticality of G, G must be an odd anti-hole.
{x,,
It is easy to check that if we assume G is claw-free, then every vertex of G must be a 6-vertex. Hence we can obtain a proof that the SPGC is true for claw-free graphs by eliminating Lemma 4 and following the rest of the discussion in this section. This proof uses the same approach as theone in [ 7 ] . However, it is much simplified. We summarize the result in Section 2 and Section 3 as follows. Corollary 1. The SPGC is true for graphs in which each vertex is one of the following types : (i) a b-vertex, (ii) a 6-vertex, (iii) a rn-vertex, (iv) a 6 -vertex (v) a m’-vertex.
4. The SPGC on some other special graphs
We have illustrated, in previous sections, how to apply known techniques to more general graphs. In this section we show that a similar approach can produce alternative proofs that the SPGC is true for other special graphs. Those proofs turn out to be much simpler than the original ones. Corollary 2. If a p-critical graph G is planar, then G is an odd hole or an odd anti-hole. Proof. By Proposition 1, we can assume w ( G ) = 4. It is well known that a planar graph has a vertex of degree 5 or less. Let u be such a vertex in G. By Theorem 2, u is contained in 4 maximum cliques. This would imply u has degree 5 and there exist four 3-cliques in N ( u ) , which can easily be shown to be impossible. 0 A circular-arc graph is defined to be the intersection graph obtained from a collection of arcs on a circle. Characterizations of circular-arc graphs are given by Tucker [16]. The following corollary is implied by the proof of Tucker [14]. However, we will indicate the implication more explicitly here.
Corollary 3. If a p-critical graph G is circular-arc, then G is an odd hole or an odd anti-hole.
112
W.-L.HSU
Proof. We first show that there do not exist two vertices x, y in G such that (x, y ) E E and (x, z ) E E .$ (y, z ) E E. Assume such a (x, y ) pair exists. Since G is p-critical, the size of a maximum clique containing x in G \ y should be w ( G ) . But then any such maximum clique together with y would give us a clique of size w ( G ) + 1 in G, a contradiction. This implies that in a corresponding circular-arc representation of G no arc is properly contained in another arc. Thus G is, in fact, claw-free and the result follows from [12].
A toroidal graph is a graph which can be drawn on a torus so that no two edges intersect. It has been shown that the SPGC holds for toroidal graphs (see [6]). We will establish the same result using Proposition 1 and the following results of [S] and [6]. Lemma 6 [5]. Let G be a p-critical graph. If for each vertex u of G, the partition of G \ u into (Y -stable sets has at least two members containing one neighbor of u, then G is an odd hole or an odd anti-hole. Lemma 7 [6]. In a critical toroidal graph G, either w ( G )< 4, or G is regular of degree six and triangulates the torus. Corollary 4. If a p-critical graph G is toroidal, then G is an odd hole or an odd anti-hole. Proof. By Proposition 1 we can assume w(G)24. Let u be any vertex of G. Since y(G \ u ) = w(G)a4, we can color G \ u using w(G) colors, Since G is p-critical, N ( u ) must use w(G)24 colors. By Lemma 7, I N ( u ) l = 6 . Hence there must be a t least two vertices which are colored singly in N ( u ) .This is true for every vertex of G. Hence the result follows from Lemma 6. 0
References [ I ] C . Berge, Graphes et Hypergraphes (Dunod, Paris, 1970). [Z] R.G. Bland. H.-C. Huang and L.E. Trotter, Jr., Graphical properties related to minimal imperfection, Discrete Math. 27 (1979) t 1-22 (this volume, pp. 181-192). 131 V. Chvatal. R.L. Graham, A.F. Perold and S.H. Whitesides, Combinatorial designs related to the strong perfect graph conjecture, Discrete Math. 26 (1979) 83-92 (this volume, pp. 197-206). [IV. ]Chvatal, On the strong perfect graph conjecture, J. Comb. Theory, Ser. B 20 (1976) 139-141. 1.51 R. Giles and L.E. Trotter, Jr., On stable set polyhedra for K,,,-free graphs, J. Comb. Theory 31 (1981) 313326. [ h ] C.M. Grinstead. Toroidal graphs and the strong perfect graph conjecture, Ph.D. thesis, UCLA. [7] W.-L. Hsu, How to color claw-free perfect graphs, Ann. Discrete Math. 11 (1981) 18Y-197.
The Perfect Graph Conjecture on special graphs
113
[8] H.-C. Huang, Investigations on combinatorial optimization, Ph.D. thesis, Cornell University. [9] L. LovAsz, A characterization of perfect graphs, J. Comb. Theory, Ser. B 13 (1972) 95-98. [lo] M. Padberg, Perfect zero-one matrices, Math. Program. 6 (1974) 18C-196. [l 11 K.R. Parthasarathy and G. Ravindra, The validity of the strong perfect graph conjecture for (I&\ e)-free graphs, J. Comb. Theory, Ser. B 26 (1979) 98-10, [12] K.R. Parthasarathy and G. Ravindra, The strong perfect graph conjecture is true for K,,3-free graphs, J. Comb. Theory, Ser. B 21 (1976) 212-223. [ 131 A.C. Tucker, Critical perfect graphs and perfect 3-chromatic graphs, J. Comb. Theory, Ser. B 23 (1977) 143-149. [14] A.C. Tucker, Coloring a family of circular-arc graphs, SIAM J. Appl. Math. 29 (1975)493-502. [15] A.C. Tucker, The strong perfect graph conjecture for planar graphs, Canad. J. Math. 25 (1973) 103-114.
This Page Intentionally Left Blank
Annals of Discrete Mathematics 21 (1984) 115-119 @ Elsevicr Science Publishers B.V.
THE GRAPHS WHOSE ODD CYCLES HAVE AT LEAST TWO CHORDS H. MEYNIEL E R 175 Combinatoire, Universifi Pierre et Marie Curie, UER 48, 7.5230 Paris, France
We prove the following theorem: If every odd cycle of length 2 5 has at least two chords, then the gruph is perfect.* This generalizes a result of Gallai and Suriinyi and also a result of Olaru and Sachs.
A k-coloring o f the vertices of G is a mapping f of the vertex set X of G into a set C of k colors such that no two adjacent vertices have the same color. A switching (relative to C ) is an operation on f defined as follows: take a connected component H of the subgraph G,, induced by the vertices of colors a, p E C. If H has at least one vertex of each color a, p, we interchange the colors (Y and p for all the vertices in H, the colors of the other vertices remaining unchanged. If H has only one vertex, of color a, say, we color this vertex with p and we color with p all the connected components of Gm,consisting of a single vertex with color a. Let 3 be the class of all graphs such that every odd cycle of length 2 5 has at least two chords, and let G E '9. For k > y ( G ) , let f be a k-coloring of G. We shall first show that it is possible to obtain from f a coloring of G with y ( G ) colors by a sequence of switchings.
Lemma 1. Let G E 9, let x be a vertex of G, and let G r ( x ) be the set of its neighbors. Consider a k-coloring of X - {x}. If y l and y 2 are two vertices of rG( x ) with distinct colors a and p, respectively, belonging to a component H of Gap,then y , and y2 are linked by an (a,P)-chain contained in rG(x). Proof. Let p [ y I y2] , = (xl = y l , xz,.. . ,x Z p = y2) be a chain of Gap of minimum length between y l and y2. The cycle p = p [ y l ,y 2 ] + [x,y l ] + [x, y z ] is odd and has only chords issuing from the vertex x. Assume that there exists a vertex a of p [ y l ,yz] which is not in rG (x); we shall show that this leads to a contradiction. * Edifors' Note. The same result was conjectured by E. Olaru (Elektron. Informationsverarb. Kybern. (EIK) 8 (1972) 147-179). Also, we have just learned from E. Olaru that S.E. Markosian and I.A. Karapotian have found independently another proof of Theorem 2.
115
H. Meyniel
116
Let x, be the last vertex in rG ( x ) which is before a on p [ y I y, 2 ] .and let x, be the first vertex in rG ( x ) which is after a on p [ y I y, 2 ] .Clearly, p [ x , ,x,] is of even length because otherwise the cycle ( x , x,, x # + ~. ., . ,x,, x ) would be an odd cycle of length 2 5 without chords. Let x k be the vertex of r G ( x ) after x, on p [ y l r y 2 ]We . have k > j + 1 , otherwise (x, x,, x , + ~. ,. . ,x,, x , + ~x, ) would be an odd cycle of length 2 5 with only one chord. As above, p [ x , , x k ]is of even length. Thus, all the portions of p [ y , ,y2] between two vertices of I‘G ( x ) are even, and therefore p [ y , .y z ] is even, a contradiction. 0
Lemma 2. Let G E %, let x be a vertex of G, and let f : X - { x } + C be a k-coloring of the subgraph G - x. Let f be the restriction o f f to the subgraph G induced by TG( x ). For every k-coloring f l of G which is obtained from f by a sequence of switchings relatively to C there exists a k -coloring g of G - x, obtained from f by a sequence of switchings relatively to C, such that g = fl. Proof. Each component H of G=@intersects G either in a single component or in several components. In the latter case, Lemma 1 guarantees that each of the several components is a single vertex. Hence, each switching on G extends into a switching or a sequence of switchings of G. 0 We remark that if GfZ 3, then the property of Lemma 2 is not always true. Consider the graph G with vertices a, b, c, d, x and edges ax, bx, dx, ab, bc, cd. The bicoloring
of G - x has a restriction
A switching on
7 gives
which is not the restriction to G of a bicoloring f l of G switchings.
-
x obtained from f by
Theorem 1. Lei G E 3, and lei f : X --+ C be a k -coloring of G. Then there exists a coloring of G with y ( G ) colors which is obtained from f by a sequence of switchings relative to C.
The graphs whose odd cycles have at least two chords
117
Proof. If the order of G is 1, then the theorem is obviously true. Let n > 1. Assume that the result is true for all graphs with order < n ; we shall show that it is true for every graph G of order n. Let C = { a I a , 2 , .. . , a k } , with k > y ( G ) = 9, and consider a k-coloring f : X + C. There exists in G a vertex x with color a # a l ,a2,. . . , aq (otherwise there is nothing to prove). Let H be the subgraph of G induced by x and the vertices of color a I ,a2,.. . ,aq.Since y ( G ) = q, the subgraph I? of H induced by rH(x) satisfies y ( H )s q - 1. By the induction hypothesis, I?, which is of order less than n and belongs to 93, has a coloring f, with y ( I ? ) colors obtained from the restriction f of f to H by a , 2 , . ., a,}. sequence of switchings relative to { a I a By Lemma 2, f l can be obtained from f by a sequence of switchings relative to { a 1a2,. , . . ,aq}.After this sequence of switchings, we can recolor x with one of the colors a l ,a 2 , . . ,aq; in other words, we can enlarge the set of vertices colored with a l ,a z ,. .. , aq.This process can be repeated until all the vertices of G are colored with a I , a z , .. . , aq.Thus we obtain a coloring of G with y ( G ) colors. 0 Theorem 2. Let G be a graph such that every odd cycle of length more than 3 has at least two chords. Then G is perfect. Proof. Consider a graph G E 3 which is not perfect and of minimal order. So, if q = y ( G ) , we have w ( G ) < y ( G ) = 9 ; furthermore, for every set A C X , A # X ,
7
Let x be a vertex of G ; so the subgraph G - x has a ( q - 1)-coloring f . Let be the restriction of f to the subgraph G induced by r G ( x ) . We have y ( G ) 3 9 - 1: otherwise, by Theorem 1, G has a (q -2)-coloring f l , which is obtained from f by switchings; so, by Lemma 2, there exists a (9 - 1)-coloring g of G - x with g = Hence y ( G ) C 9 - 1, a contradiction. Let A = { x } U rG (x), then y ( G A )3 q ; so, by (l),A = X , and x is adjacent to every other vertex. Since x was chosen arbitrarily, this shows that G is a clique. Hence y ( G ) = w ( G ) , and the contradiction follows. 0
TI.
Corollary 1 (Gallai [3]). Let G be a graph such that every odd cycle of length > 3 has at least two non-crossing chords; then G is perfect. Corollary 2 (Olaru, Sachs [6]). Let G be a graph such that every odd cycle of length > 3 has at least two crossing chords ; then G is perfect.
I in
H. Meyniel
Corollary 3 (Hajnal and Suranyi [4]). Let G be a triangulated graph (that is, every cycle of length > 3 has a chord). Then G is perfect. Proof. I n this case every odd cycle of length theorem follows. 0
3 5 has
at least two chords and the
Corollary 4 (Seinsche [7]). If a graph G has no induced P4(elementary chain with -1 verlices and no chords), then G is perfect. Proof. Every odd cycle of length at least 5 and with at most one chord contains an induced P,. 0 Corollary 5. I f a graph G has no induced subgraph isomorphic to Go= k > 1. (nb, ac, bc, c d } , then G is perfect if and only if it has no induced Proof. A graph G with no CZk+, and no Go has no odd cycle with only one chord: the unique chord would be triangular, which would yield an induced Go. 0 Remark. The method used in Theorem 2 cannot yield a proof of the perfect graph conjecture. For example, consider the graph G with vertex set {x, : 1 s i G 3 , 1 S j S 4}, where {x,,x,} is an edge iff i f p and j f q. The graph G is perfect (since its complement has a clique-hypergraph which is balanced, and, by a theorem of Berge [2, Chapter 161, this implies that this complement is perfect). Clearly. G E 3, and from the 4-coloring with classes S, = {x,] j = 1,2,3) we cannot get a 3-coloring by a sequence of switchings.
I
Problem. We denote by C , ( G ) the set of colorings of G with q colors, 1,2,. . . ,q. We say that two colorings f , g E C,(G) are q-Kempe-equiualent if there exists a sequence of colorings all in C, (G), f n = f , f l , . . . ,f k = g, such that for i = 1,2,. . . , k, the coloring f, is obtained from f,-, by a switching. The Kempe-equivalence being an equivalence relation, the classes of this relation will be called the q-classes of G. If G is the graph with vertex set {1,2,3,4} x {1,2,3,4,5} x{1,2,3,4.5}, where the vertices ( i , j, k ) and ( i ' , j', k ' ) are adjacent iff i # i f ,j # j ' and kf k', we make the following observations: (1) The projection p l ( i , j , k ) = i is a 4-coloring of G ; the projections pz(i, j, k ) = j and ps(i,j, k ) = k are 5-colorings. (2) For q 3 4, every q-coloring of G is q-Kernpe-equivalent to pl, or to p2,or to
p3.
(3) The colorings pl, p 2 and p3 belong to three different 6-classes.
The graphs whose odd cycles have at least two chords
119
(4) The coloring p 2 is q-Kempe-equivalent to pl for q 7. Thus, the number of q-classes is 1 for q = 4 ; it is 3 for q = 5 or 6; it is 1 for q 3 7. This suggests the following open problem: If in G every odd cycle has at least two chords, is the q-class unique for every q y ( G ) ?
References [I] C. Berge, Farbung von Graphen, deren samtliche bzw. deren ungerade Kreise starr sind, Wiss. 2. Matin-Luther-Univ. Halle-Wittenberg (1961) 114. [2] C. Berge, Graphes et hypergraphes (Dunod, Paris, 1970); English translation: Graphs and Hypergraphs (North-Holland, Amsterdam, 1973). [3] T. Gallai, Graphen mit triangulierbaren ungeraden Vielecken, Magyar Tud. Akad. Mat. Kutato Int. Kozl. 7 (1962) A 3-36. [4] A. Hajnal and J. Surinyi, Uber die Auflosung von Graphen in Vollstandige Teilgraphen, Ann. Univ. Sci. Budapestinensis 1 (1958) 113. [5] M. Las Vergnas and H. Meyniel, Kempe classes and the Hadwiger conjecture, J . Comb. Theory. Ser. B 31 (1) (1981) 95-104. [6] H. Sachs, On the Berge conjecture concerning perfect graphs, Combinatorial Structures and their Applications (Gordon and Breach, New York, 1970). [7] D. Seinsche, On a property of the class of n-colorable graphs, J. Comb. Theory, Ser. B 16 (1974) 191-193. [8] L. Surinyi, The covering of graphs by cliques, Studia Sci. Math. Hungar. 3 (1968) 345-349.
This Page Intentionally Left Blank
Annals of Discrete Mathematics 21 (1984) 121-144 @ Elsevier Science Publishers B.V.
CONTRIBUTIONS TO A CHARACTERIZATION OF THE STRUCTURE OF PERFECT GRAPHS Elefterie OLARU and Horst SACHS Technische Hochschule Ilmenau, DDR -6300 Ilmenau, German Dem. Rep.
Characterizing the structure of perfect graphs is one of the important, most actual, and most challenging unsolved problems of graph theory. The strong version of Claude Berge’s Perfect Graph Conjecture has given rise to numerous investigations of this problem. In this paper, results found by E. Olaru since 1969 are summarized.
1. Introduction
The graphs G = ( X , U ) considered in this paper are finite, undirected, and have neither loops nor multiple edges; X and U denote the sets of vertices and of edges, respectively. We shall use the following general terminology and notation. S ’ C S means: S’ is a subset of S ; S ’ C S means: S‘ is a proper subset of S. Let X ’ C X , U’ 5 U ; G = ( X , U ) , G’ = ( X ‘ ,U’). The graph G ’ is called a partial subgraph of G ; if any two vertices of X ’ are adjacent in G ’ if and only if they are adjacent in G, then G‘ is called a subgraph of G or, more precisely, the subgraph of G spanned by X ‘ , and denoted by [ X ’ ] ,;if X ’ = X then G ‘ is called a partial graph of G. The number of vertices of G , i.e., the cardinality I X I of X , is also denoted by n(G). A complete subgraph of G is called a clique of (or in) G . The maximum number of vertices contained in a clique of G, i.e., the ‘density’ of G, is denoted by w ( G ) . A subset S C X with [SIC= (S,@ is called a set of independent vertices or, briefly, an independent set of G . The maximum number of independent vertices of G is called the stability number of G and denoted by a ( G ) .Here, a stability system of G is an independent set S of G with 1 S I = a ( G ) . The complement of G = ( X , U )is the graph G = ( X , 0) with the property that any two vertices of X are adjacent in G if and only if they are non-adjacent in G . Clearly,
a ( C )= w ( G ) ,
w ( G )= a(G). 121
(1)
E. Olanc, H. S a c k
122
For X' C X and U ' C U, we define G-X'=[X-X']c,
G-U'=(X,U-U');
f o r x E X , u E U , w e w r i t e G - x a n d G - u insteadof G - { x } o r G - { u } , respectively. Let G ' be a subgraph of G ; an edge u of G is called a-critical with respect to G ' if
a(G'-u)=a(G')+l.
(2)
The set of all edges of G which are a-critical with respect to G itself is denoted by U', and the partial graph G' = (X, U c )of G is called the &-critical skeleron of G. If G = G' then G is called a-critical.
2. a -partitionable graphs
Let G = (X, V), let Z = {XI,X,,. . .,X m }( m a 1) denote an arbitrary partition of X (i.e., X = XI u X1u . . u Xm,X, # 0, Xi n X, = 0 ( i # j ;i, j = 1,2, . . . ,m)), and put G, = [X,Ic ( i = 1,2,. . . ,m ) . Then, evidently,
-
m
a ( G ) sT a ( G t ) , I =
(3)
m
w(Gi).
w(G)s I=
Definition 1. The graph G = (X, V ) is called a -partitionable if a ( G )= 1 or if there is a partition Z = {XI,X2,.. . ,X,} of X with m > 1 such that m
Z is then called an a-partition of G, and the graphs Gi are called the corresponding a -components of G. G is called w-partitionable if G is a-partitionable. G is called parfirionable if it is both a-partitionable and w-partitionable.
Of particular interest in graph theory are those a -and w-partitions for which all G,have a certain prescribed property. An an example, consider the following unsolved problem: Characterize all graphs G having one of the following two properties: (C) G has an a -partition with all a -components being cliques; (F) G has an w-partition with no w-component having an edge.
Structure of perfect graphs
123
Remark. If the graph G = ( X , U )has an a -partition Z such that the corresponding a-components are cliques then, obviously, the number of these components equals a (G), i.e., G can be covered by a ( G )cliques. We shall then say: G has a perfect clique -covering or G is perfectly a -partitionable.
z
If G has an o-partition 2 satisfying (F) then generates a colouring of X with exactly w ( G ) colours. We shall then say: G has a perfect colouring or G is perfectly w -partitionable. Definition 2. A graph G is called a-perfect (or w-perfect) if each subgraph of G (including G ) has a perfect clique covering (or a perfect colouring, respectively). G is called perfect if it is both a-perfect and o-perfect.
Before dealing with perfect graphs, we shall prove some general statements concerning a -partitionable graphs. Remark. If G is disconnected G is a-partitionable, for, if G has C1, C2, . . . ,C,,, ( m > 1) as its components of connectivity, then, clearly, rn
)). a ( G )= i = l a (C;
Next we prove Theorem 1. Let G = ( X , U ) be a connected graph and assume that there is a set Q C X such that (i) G - Q is disconnected, (ii) [a], is a clique, then G is a-partitionable. Proof. If
u(G - Q ) = a(G)- 1
(5)
then G is a-partitionable since (5) implies a ( G )= a ( G - Q) + a ([ a],). Therefore we assume a(G - Q)= CY(G).
(6)
Because of (i), G - Q is the union of two (non-empty) separated parts GI= (XI,Ul)and Gz = (X2,U2).The set of the stability systems of G - Q is obtained by combining in every possible way a stability system of GI with a stability system of G,; thus, because of (6), we have
(GI)+ (G2)= ( G ) .
(7)
E. Olaru, H.Sachs
124
We shall now show that Q can be partitioned into two parts a way that
a ( [ X ,U Q,IC)= a ( [ X , I c ) =a ( G ) ( i where Q1or
Q2
Ql
= 1,2)
and Q2in such (8)
may be empty. From (8) and (7) we deduce
..([Xi U Q,]C)+(Y([X~U Q,]~)=~~(G~)+(Y(G~)=(Y(G), i.e., G is a-partitionable. Let x be an arbitrary vertex of Q. We shall show that either in every stability system of GI or in every stability system of G2there is a vertex which is adjacent to x. For suppose that S1 is a stability system of GI,say, with none of its vertices adjacent to x : then, if S2 is an arbitrary stability system of G1, x is adjacent to at least one vertex of S2, for otherwise there would exist a stability system S = S , U Sz of G with none of its vertices adjacent to x - contradicting the definition of a stability system. Now let QIcQ be the set of all those vertices of Q which are adjacent to at least one vertex of every stability system of GI:then every vertex of Qz= Q - QI is adjacent to at least one vertex of every stability system of Gz.Clearly, Ql and Q2 satisfy (8), which proves the theorem. 0
A
For an arbitrary graph G = (X, V )and sets A C X, B C X, where A # 0,B # 0, n B = 0, we define
[A,Blc = { u ( u E U ; u = ( a , b ) , a € A , b EB}. Theorem 2. Let G be an a -partitionable graph and let {XI,Xz, . . . ,X,,, } be an a-partition of G. Then none of the sets [X,,XjlC (i# j; i, j = 1 , 2 , . .. ,m ) contains an a-critical edge of G.
Proof. Assume that u = (a, b) is an a-critical edge of G. Then, by definition, G - u has a stability system S with IS I = a ( G ) +1. From
/ s / = a ( G ) + i= l S n x l =
i=l
m
= C Isnx,~ i=l
1
we deduce that there is an io such that S f l X,I
a([X,]c-.)=ff(G,)+
> a(G,). This implies that
1,
which is impossible unless a, b EX,.
0
Corollary 1. lf G is (Y -partitionable and
(Y
( G )> 1 then the a-critical skeleton G'
Structure of perfect graphs
125
of G is disconnected and each connected component of G' is contained in an a -component of G. Corollary 2. Each a-critical edge of an a-partitionable graph G belongs to an a-component Gb of G and is a-critical with respect to G,. Corollary 3. A n a -critical graph G with a (G) > 1 is a -partitionable if and only if G is not connected. Note that there are a -non-partitionable graphs whose a -critical skeletons are disconnected. As an example consider G = ( X , U ) where X = {xl,x2, x3, x4,XS, x ) and
U = { X I , XZ),
(x2, x3), (x3, x4), (x4,
u {(x, x i ) 1 i = 1,2,3,4,5}
xs), (xs, XI))
(the "5-wheel"):
here G' = (X,V c )with U' = {(XI,XZ), (XZ, x3), (XS, x4), (x4, XS), (XS, XI)), thus G' is disconnected. Clearly, any graph G = ( X , U ) contains a partial graph G' = ( X , U ' )(U'C U ) with the following properties: (a) a ( G ? = a (GI; (b) every edge of G' is a-critical with respect to G'. G ' is called an a-critical partial graph of G. Theorem 3. A graph G with a ( G )> 1 is a-partitionable if and only if it contains at least one a-critical partial graph which is disconnected (i.e., by virtue of Corollary 3 of Theorem 2, which is a -partitionable). Proof. First suppose that G is a-partitionable; let Z = {Xl,Xz,.. .,X m }(rn > 1) be an a-partition of G. If there is, for some i €{2,3,. . ., m } , an edge u1€[XI,X,], such that a ( G - u ; ) = a ( G ) , then u I is omitted and G - u1 is denoted by GI. Clearly, Z is an a-partition of GI. If there is an edge u2E (XI,Xi),, such that a ( G 1- uz)= a (GI)= a ( G ) , then uz is omitted and GI - u2 is denoted by Gz, and so on. After a finite number of steps we arrive at a partial graph G, of G having the following properties: (i) a ( G p) = a ( G ) ; (ii) Z is an a-partition of G,; (iii) every edge contained in lXl,X,]Gp is a-critical with respect to G,. Corollary 2 of Theorem 2 now yields that [XI,X , ] , is empty, which implies that not every a-critical partial graph of Gp can be connected. If G contains a disconnected a-critical partial graph G' then, according to Corollary 3 of
E. Olaru, H.Sachs
126
Theorem 2, G' is a-partitionable. An arbitrary a-partition Z = {XI,X2,. . . ,X k } of G' is an a-partition of G, too, since
Theorem 3 is now proved.
0
Definition 3. A graph G is called strongly stable if for each clique C = [a], of G, a ( C - Q ) = a ( G ) . Remark 3. If G is a-non-partitionable then G is either strongly stable or complete. The converse of this statement, however, is not true: the graph consisting of two separated 5-circuits is a -partitionable though it is strongly stable. But the following theorem is valid: Theorem 4. If a graph C is strongly stable and a-partitionable, then all a components of any a-partition of G are strongly stable. Proof. Let G = (X, V )be a strongly stable graph and let Z = {XI,X,, . . . ,X,,} be an arbitrary partitlon of X such that
Assume that one of the a-components, say GI,is not strongly stable. Then there is at least one clique C = [QlC, of GI satisfying a(G1-
a)= a(C1)- 1.
(10)
Since {XI - Q, X 2 , .. . ,X.} is a partition of X - Q, (9) and (10) imply
( G - Q) C (GI- Q) + (Gz) +
*
+
( G , ) = ( G )- 1.
But this inequality contradicts the hypothesis that G is strongly stable.
0
As a simple consequence of Theorem 4 we have
Corollary 1. Let G = (X, U )be a strongly stable graph. Then there is a partition {XI,X z , . . . ,X,} (p 2 1) of X such that (a) all G, = [XI,(i = 1,2,. .. ,p ) are strongly stable and a -non-partitionable, (b) a ( G , )= cu(G).
xr=l
Remark. We shall see that, in the investigation of perfect graphs, the strongly stable graphs are of a particular significance.
Structure of perfect graphs
127
3. Perfect graphs Let x ( G ) denote the chromatic number of the graph G. Obviously,
X ( G ) Zw ( G ) and, analogously, x(G)Zw(G)=a(G). For many purposes the following reformulation of the definition of perfect graphs (see Definition 2) is more appropriate.
Definition 2’. The graph G is called (a) a-perfect if = a ( G ‘ )for every subgraph G’ of G ( G ’ = G included); (b) w-perfect if x ( G ’ )= w ( G ’ )for every subgraph G‘ of G (G’ = G included); (c) perfect if G is both a-perfect and o-perfect.
~(c’)
Note that, in fact, properties (a), (b), and (c) are equivalent: this is the content of the weak version of Berge’s Perfect Graph Conjecture and follows easily from an important result of L. L O V ~ S[lo], Z [ll]:
Theorem 5 (Lovisz [lo]). A graph G is perfect if and only iffor every subgraph G ‘ = [X’], (X’CX)
n ( G ’ ) = I X ’ l s a ( G ’ )w . (G’). Remark. x ( G )equals the minimum number of cliques covering the vertex set of G : therefore, x ( G ) is also called the clique-covering number of G denoted by O(G). We need some notions and results. 3.1. O d d holes and
(Y
-critical Hamiltonian circuits
The graph G = (X, U )is called a k-circuit (of length k 3 3) and denoted by L if X can be so arranged as a sequence xl, x 2 , . .. ,X k that
u = {(Xi,Xi+])
I i = 1,2,. . . ,k ;
Xk+1
k
= XI}.
We shall briefly write L k
=( x l ,
x2,. ,
.
9
X k , XI)
and say that xi and xi+]( i = 1,2,. . . , k ; X k c l = xl) are neighbouring on L k . If the circuit L is contairred in a graph G (as a partial subgraph) we shall say that L is a
E. Olaru. H. Sachs
128
circuit of (or i n ) G. An edge of G connecting two vertices of a circuit L of G which are non-neighbouring on L is called a chord of L. Let s = (x, y ) and s‘ = (x’, y’) be two chords of a circuit L in G where the vertices x, y, x’, y’ are pairwise distinct. If, on L, the pair x’, y ‘ separates (and is separated by) the pair x, y we shall call s, s f a pair of crossing chords of L.
Definition 4. A circuit of a graph G of length greater than three which has no chord in G is called a hole of G. Towards a characterization of the structure of &-perfect graphs the following conjecture has already become famous.
Conjecture 1 (C. Berge and P.C. Gilmore) (strong version). A graph G is a-perfect if and only if neither G nor its complement G contains an odd hole. For the sake of brevity we introduce the following notation: Let G = ( X , U ) be an arbitrary graph, let X ‘ , X ” C X with X ’ n X ” = 0;we write X’pcX” if in G every vertex of X ’ is adjacent to every vertex of X ” , and we write X’pCX“ if there is in G no edge connecting a vertex x f E X ’ with a vertex x “ E X ” ; instead of { x } p c Y we briefly write xpcY, etc. The next theorem yields a necessary and sufficient condition for the existence of an odd hole. Theorem 6. A graph H = ( X , U ) contains an odd hole i f and only if there are in H two distinct edges ( x ’ , x ) and (x,x’) such that (i) ( x I , x ) and (x,x2) are a-critical with respect to some subgraph G of H ; (ii) ( x I r x Z ) U. ~ If H satisfies the above condition then there is in H an odd hole L which contains both edges (x’, x ) and ( x . x’). Proof. 1 ( “ i f ” ) : Because of
a ( G - ( x ’ , x ) ) = a ( G ) + 1 ( i = 1,2), there are two stability systems S , and S2 of G such that x ’ E s,,XP xpCS,
-{XI}
s,, ( i = 1,2).
Now consider the subgraph G‘ of G spanned by SIU Sz U {x}. We have a ( G ’ ) =a ( G ) ,
(11)
Structure of perfect graphs
129
a ( G ’ )< x(C’),
(12)
o(G’)=2.
(13)
The first relation is clear since, by definition, I S, I = a ( G ) . Expression (11) implies that (x’, x ) and ( x , x2) are a-critical also with respect to G ‘ . If there is an a-partition of G ’ with all its a-components being cliques then, by Corollary 1 of Theorem 2, the a-critical edges (x’, x ) and (x, x2) are contained in a clique of G ‘ , i.e., ( X I , x2) is an edge of H, contradicting (ii); thus (12) follows. The subgraph spanned by S, U Sz is bipartite, thus using (11) and (ii) we conclude that G ‘ does not contain a 3-circuit; therefore, w(G’)= 2 , i.e., (13) is valid. Since a bipartite graph is a-perfect, (12) and (13) imply that G ’ contains an odd hole, say L * . Necessarily, ( x l , x ) and (x,x2) belong to L* for G I - ( x , x ’ ) ( i = 1,2) are bipartite graphs. 2 (“only if”): If H contains an odd hole L then any two adjacent edges ( x ’ , x ) and (x,x*) of L are a-critical with respect to L and, clearly, satisfy (ii). Theorem 6 is now proved. Corollary. A n a-critical graph is a-perfect if and only if all of its connected components are cliques. Definition 5. A Hamiltonian circuit of a graph H which has the property that all of its edges are a-critical with respect to H is called an a-critical Hamiltonzan circuit of H. From Corollary 1 of Theorem 2 we deduce the following statement: A graph H which contains an a-critical Hamiltonian circuit and which is perfectly a -partitionable is a complete graph.
This implies the subsequent statement:
A graph H with a(H)2 2 which contains an a-critical Hamiltonian circuit cannot be perfectly a -partitionable. Definition 5a. A subgraph G of an arbitrary graph H is called an a-critical hole of H if (i) a ( G )3 2, (ii) G contains an a-critical Hamiltonian circuit. Remark. A graph H which contains an a-critical hole cannot be a-perfect.
E. Ohm, H. Sachs
130
Conjecture 2 (see [17]). A graph H is perfect if and only if it has no a-critical hole. Remark. An odd circuit holes.
L2kCl
(k 3 2) and its complement
z2k+l
are a-critical
3.2. Critically a -imperfect graphs
For the investigation of perfect graphs we have introduced the notion of a critically a-imperfect graph and we have shown (see [13]) that, using this notion, many propositions concerning perfect graphs - new ones and also well-known ones - can be proved in a unified and relatively simple manner. We hold that the notion of a critically a-imperfect graph plays a particular role within the theory of perfect graphs and, therefore, believe that it is useful to continue the investigation of critically a -imperfect graphs until a reasonable structural characterization of perfect graphs is arrived at. (See also the Concluding Remark of this paper.)
Definition 6. A graph G is called critically a-imperfect if (A) 8 ( G ) > a(C) (recall: 8 ( G ) = x ( G ) ) ; (B) e(Cr)= a ( G r )for every proper subgraph G' of G.
Lemma 1. Every a -imperfect graph contains a critically a-imperfect subgraph.
Proof. (Clear.) Lemma 2. A graph H is a-perfect if and only if it contains no strongly stable subgraph.
Proof. If H contains no strongly stable subgraph then every subgraph H' of H (including H)has a clique-covering of exactly a (H')cliques, i.e., H is perfect. To finish the proof just note that a strongly stable graph G cannot be partitioned into a ( G ) cliques. 0 There is a tight relation between strongly stable and critically a -imperfect graphs, to be expressed in the next theorem. First we give the following definition: Definition 7. A strongly stable graph which contains no strongly stable proper subgraph is called a minimal strongly stable graph. Now we prove the next theorem.
Structure of perfect graphs
131
Theorem 7 . A graph G = ( X , U ) is critically a-imperfect if and only if it is a minimal strongly stable graph. Proof. If G is critically a-imperfect then it is a-non-partitionable (and, consequently, strongly stable) for otherwise there would exist an a-partition, say { X , ,X z } , with a ( G J a ( G Z = ) a ( G ) . The graphs G I and Gz being proper subgraphs of G, G Ican be covered with a ( G I )and G2can be covered with a ( G 2 ) cliques, thus G can be covered with a ( G )cliques -contradicting the hypothesis that G is critically a-imperfect. From Lemma 2 it follows that G (since it is critically a -imperfect) contains no strongly stable proper subgraph: this proves that G is a minimal strongly stable graph. Conversely, let G be a minimal strongly stable graph. Then, by Lemma 2, G is a-imperfect. Now, since G contains no strongly stable proper subgraph, again by Lemma 2, every proper subgraph of G is a-perfect. Theorem 7 is now proved. 0
+
Before L O V ~ Spublished Z his important result (Theorem 5), we established a couple of properties of critically a-imperfect graphs without using Theorem 5 (see [13]-[18]). In making use of Theorem 5 we shall now simplify the original proofs.
Theorem 8. A graph G = ( X , U ) is critically a-imperfect if and only if (i) I X I = a ( G ) * w ( G ) + l , (ii) each proper subgraph G’ = [ X ’ ] , ( X ’ C X ) of G satisfies
I X ’ 1G a (G‘) w (G’). *
Proof. Clearly, (i) and (ii) imply that G is critically a-imperfect (see Theorem 5). Conversely, if G is critically a-imperfect then Theorem 5 implies that (ii) is valid and that
1 X I S a ( G ) *w ( G ) + 1. Now, let x be an arbitrary vertex of G. Then, by virtue of Theorem 7 and property (B) of G (see Definition 6), e ( G - X) = a ( G )and, since a clique cannot have more than o ( G ) vertices,
n ( G - x ) = I X I - 1 G a ( G )* w ( G ) . Thus (i), too, is valid.
17
Corollary. A graph G is critically a-imperfect if and only if its complement G is critically a -imperfect.
E. Olary H. Sachs
132
Let C = (X,U ) be critically a -imperfect and let x be an arbitrary vertex of G. Because of property (€3) of C (see Definition 6), G - x is a -partitionable and Theorem 7 implies a ( G - x ) = a ( G ) . Let C,, C,, . . .,C, be an arbitrary system of cliques covering G - x and satisfying
c, = [X,],x, , cx - { x } = X,nxl=O
X,#O,
n
u x,,
,=I
( i # j , ; i , j = l , 2 ,..., a ) .
G X , ( i = 1,2,. . . ,a)denote the set of all those vertices of X, Further, let X : which, in C, are non-adjacent to x, i.e., such that x&X!
and
x p c ( X , - Xf).
Then we have the following lemma:
Lemma 3. For every x E X and for every perfect a-partition of G - x, the following propositions hold : (1) for i = 1,2,. . . , a the set X fis non-empty, ( 2 ) for eoery ser of subscripts J C{l, 2 , . . . ,a } = I, the equality
holds.
Proof. (1) If there is an i with X : = 0 then C : = EX,U {x}IC is a clique and the cliques C , ,. . .,C,-,, C:, C,,,,. . . , C, generate a perfect a-partition of G, contradicting property (A) of critically a-imperfect graphs (see Definition 6). ( 2 ) The graphs [Xi],being cliques,
immediately follows. Let us assume that there is a set of subscripts J C I (i-e.,I J 1 < (Y (C)) such that a([
u x:] )SIJl-l.
IEJ
C
This assumption implies
ujGJX,
U { x } l C either contains x - then, by our since a stability system of [ assumption, it cannot have more that I J I vertices - or it does not contain x -
Structure of perfect graphs
133
then, again, the system cannot have more than ( J i vertices since the graph [ j E J X , ] G is partitioned into exactly I J I cliques having pairwise disjoint vertex sets. The graph [ U,,,X], is covered by a ( G ) - ( J Jcliques, therefore
u
a(
[ u XI iGJ
G
)Ga(G)-IJI.
ujEIX,
u,,JX,]G
The relation I J I < a ( G ) implies that [ U {x}], and [ are proper subgraphs of G , and using property (B) (see Definition 6) we conclude that
B(G)= B(
[ u X, U { x } U u X i ] ) j€J
is1
G
contradicting (A) of Definition 6. Lemma 3 is now proved. 0 As an easy consequence of Lemma 3, we obtain the following corollary:
Corollary. (1) Every vertex of a critically a -imperfect graph G is contained in at least a(G)(pairwise distinct) stability systems of G . (2) A critically a-imperfect graph G has at least a(G)*w ( G ) + 1 (pairwise distinct) stability systems. For every vertex x EX, we define the following sets:
I xz= {xz I x z E ( X - {x}) and x 2 p c x } , X ' = { x i x 1 E ( X - {x}) and x ' p c x } ,
and the corresponding graphs:
G t x ) =[XI],, G : x ) = [ X 2 ] ~ . Lemma 4. The graph Gtx1is perfectly ( a ( G ) - 1)-partitionable, and each clique of any perfect ( a ( G ) - 1)-partition of G f , ) contains at least two vertices. Proof. Lemma 3 , part ( 1 ) yields
Lt
x i =U x:,xt#0, x:nx:=0, i-I
and the graphs [ X l ] , are cliques ( i # j ; i,j = 1,2,. . . , a ) . From Lemma 3, part ( 2 ) we obtain a ( C i X ) ) ab ( G ) - 1. If there is a stability system S ' of Gll, with 1 S' I > a ( G ) - 1 then, because of x&X', the set S' U { x }is a stability system of G with S'U { X I 1 > a ( G ) ,contradicting the definition of a stability system. Consequently, a(Gi,))= a ( G ) - 1; the graph Glx, being a proper subgraph of G, property (B) (see Definition 6) implies
I
e(G{x))= a(Cf,)). In order to establish the second part of the assertion, we first prove
(**I
a ([XI - { x ' } ] ~=)(Y ( G )- 1 for every x I E XI. By virtue of (*), X I belongs to precisely one of the sets X l , say part ( 2 ) of Lemma 3 we obtain
X I
E X : i ;from
( ~ ( [-xXii],.) ' = ( Y ( G ) -1. Now,
a ( [ X '-xli]c)s Cr([X'-{x'}]c) (Y([X']G) = a(C[,,)= a ( G ) - 1,
hence (**) is valid. Let us assume that there is a perfect ( a ( C )- lkpartition o f G:x)such that one of the cliques of this partition consists of a single vertex, say X I . Then the graph [ X ' - { X ' } ] is ~ covered by a ( G ) - 2 cliques,.thus ( Y ( [ X ' - { X ' ) ] ~ ) ~ ( Y ( G ) - ~ , contradicting (**). Lemma 4 is now proved. 0
Corollary 1. Let G be a critically a-imperfect graph, let x be an arbitrary vertex of G, and denote the valency of x with respect to G.by r ( x ; G ) . Then r ( x ; G ) sn ( G ) - 2 a ( G ) + 1. Proof. By Lemma 4, I X1I k 2a ( G )- 2, thus
r ( x ;C ) = ) X z = ( n ( G ) - I X ' I - 1 S n ( G ) - 2 a ( C ) + 1. 0 Corollary 2. For any critically a-imperfect graph G,
+
n ( G )2 2 a ( G ) 20(G)-3.
(14)
Structure of perfect graphs
135
Proof. Together with G, its complement G is also critically a-imperfect (see Theorem 8), hence, by Lemma 4,
J X'
(***I
13 2 a ( G ) - 2 = 2 w ( G )- 2;
thus n ( G ) = 1 X'
1 + I X 2I + 1 3 2a (G)+ 2 w ( G ) - 3. 0
Remark. We shall show (Theorem 11) that equality holds in (15) if and only if the critically a-imperfect graph G is an odd circuit LZq+'or the complement LZq+' of an odd circuit (q 3 2). To prove this, we need some more results. Lemma 5. Every vertex x of a critically a-imperfect graph G is contained in an odd k-circuit ( k 3 5 ) L of G such that all chords of L (if there are any) are issuing from x. Proof. Again we consider an arbitrary vertex x of G and the corresponding subgraphs G i I )= [X'],and G:=)= [X'],.By Lemma 4, a ( G t X )=) a(G)- 1. Let X 2be a subset of X 2satisfying
a ( [ X ' U PI,)= a ( G )- 1 and having, under this condition, the maximum number of vertices:
a([xlu X' u { y ) ] , ) = .(G) for every y E (x' - P). (16) [X' U PI, is a proper subgraph of G (since x e X' U X2), therefore, because of property (B) (see Definition 6),
The set X 2- X 2does not span a clique in G for, otherwise, [(X'- X') U {x}], would be a clique and using (17) we should obtain O(G)= a ( G ) , contradicting property (A) (see Definition 6) of critically a-imperfect graphs. From this and from (16) we conclude that there exist two vertices XI,x' E ( X 2- X') such that xpc{x1,xZ1, x ' p c x 2 ,
(18)
and
a ([XI U X2U {x '}Ic)
= a ( G)
(j= 1,2).
(19)
and denote the Now we omit all edges of G which connect x with vertices of remaining graph by H. We have a ( H )= a ( G )since a stability system of H that does not contain x is a stability system of G as well, and a stability system of H that contains x has,
136
E. Olaru, H. Sachs
because of (17) and because of x p H ( X z- X') and x p H ( X 'U X'), exactly a ( G ) vertices. From (18) and (19) we obtain just those conditions formulated in Theorem 6 (with respect to H), thus H contains an odd hole L passing through x. To such a hole of H there corresponds in G an odd k-circuit ( k 2 5) passing through x and having the property that all its chords (if it has any) issue from x (namely, these chords are precisely those edges of G which connect x with those vertices of X2 that lie on L). x, having been chosen arbitrarily, Lemma 5 is now proved.
Remark. The following proposition (see [13], Theorem 10) is also valid: Proposition. Every edge (x,y ) of a critically a-imperfect graph is contained in an odd k-circuit ( k 2 5 ) all of whose chords (if there are any) are issuing from x, and, simultaneously, in an odd k'-circuit ( k ' 3 5 ) all of whose chords (if there are any) are issuing from y . Corollary 1. If the critically a-imperfect graph G is not an odd circuit Lk ( k 3 5 ) then every vertex x of G is contained in a 3-circuit. 0 Proof. According to Lemma 5, every vertex x of the critically a-imperfect graph G lies on an odd circuit L of length 2 5 which, except for those (possibly existing) chords issuing from x, has no chords. The critically a-imperfect graph G is not an odd k-circuit with k 3 5, hence the vertex set of L spans a subgraph of G which contains at least one 3-circuit (since, otherwise, G would contain an odd hole as a proper subgraph). Evidently, x belongs to this 3-circuit.
Corollary 2. If every odd k-circuit ( k 3 5 ) of a graph H has a pair of crossing chords then H is a-perfect. Proof. Suppose H were not a-perfect. Then, by Lemma 1, H contains a critically a-imperfect subgraph G. Lemma 5 implies that every vertex of G lies on a circuit of length 2 5 without crossing chords. This contradicts the hypothesis of the corollary. 0
Remark. In 1968 (see [16]), we conjectured that the existence of two chords in every odd k-circuit ( k 2.5)of H implies the perfectness of H. In 1976, Olaru [17] proved the following statement:
If a graph H satisfies the following conditions:
Structure of perfect graphs
137
(i) every odd k-circuit ( k 3 5 ) of H has at least two chords, (ii) no 5-circuit of H has exactly one pair of crossing chords, then H is a-perfect. Only recently, Markosjan and Karapetjan (see also [12]) proved the abovementioned conjecture:
Theorem 9 (Olaru, Markosjan and Karapetjan). If every odd k-circuit ( k 2 5 ) of a graph H has at least rwo chords then H is perfect. Theorem 10. A critically a-imperfect graph G is (1) an odd circuit L 2 k + l ( k 3 2) if and only if o(G)= 2; (2) the complement of an odd circuit L Z k + l ( k 3 2) if and only if a ( G )= 2.
Proof. (l)(a) First we show: An odd circuit L 2 k + ) ( k 2 2 ) is critically a imperfect. For every vertex x of L Z k + l , the graph L2k+l - x is bipartite and, therefore, a-perfect. L 2 k + l itself, however, is not a-perfect for n ( L Z k +=l )2k + 1, a ( ( L Z k + l ) = k, and @ ( L Z k + l ) = 2 imply 8(L2k+l)> a (L2k+l); thus L 2 k + l is critically a -imperfect. (b) Now we assume that G is critically a-imperfect with o(G)= 2. If G is not an odd circuit then each vertex of G lies on a 3-circuit (see Corollary 1 of Lemma 5), contradicting w ( G ) = 2; thus G must be an odd circuit. (2)(a) First we show: The complement L Z k + l of an odd circuit L Z k + l (k 2 2) is critically a -imperfect. For every vertex x of L Z k + l , the graph i 2 k + l - x is the complement of a bipartite graph; therefore, each proper subgraph of L 2 k + I is perfectly a partitionable (and hence a -perfect). & k + I itself, however, is nor a-perfect for n(LZk+1)=2k 1, ( Y ( L 2 k + 1 ) = 2 , and @ ( L 2 k + I ) = ( Y ( L Z k + l ) = k imply that 8 ( L Z k + I ) > a ( L 2 k + I ) ; thus E 2 k + l is critically a-imperfect. (b) Now we assume that G is critically a-imperfect with a ( G )= 2. Suppose of G does not contain an odd hole. that the complement If G contains an odd circuit, but no odd hole, then it must contain a 3-circuit; but that contradicts the hypothesis o(c)= a ( G ) = 2. So G cannot contain an odd circuit, i.e., is a bipartite graph. But that is impossible since it would imply that G is a-perfect. Consequently, G must contain an odd hole; thus G contains the complement i 2 k + l ( k 2 2) of an odd circuit L2k+l as a subgraph. Since iZn+, is not perfectly a-partitionable this is possible only if G = E Z k + l . Theorem 10 is now proved. 0
c
E. Olaru, H. S a c k
138
Corollary. A graph H with w ( H )= 2 or neither H nor H contains an odd hole.
a (H) =2
is a-perfect if and only if
Proof. The assertion follows from Lemma 1 and Theorem 10. 0 Now we are able to prove the following theorem:
Theorem 11. A critically a-imperfect graph G is an odd circuit LZk+, ( k 2 2), or the complement of such a circuit, if and only if n ( G )= 2 a ( G ) + 2 w ( G ) - 3.
(20)
(See also the Remark following Corollary 2 of Lemma 4.)
Proof. (a) For an odd circuit
LZk+l
(k 3 2), clearly,
n(Lzk+i)=2k+ 1, (Y(Lzk+i)=k
and O(Lzk+i)=2,
so
L z k + l satisfies (20). If a graph G satisfies (20) then its complement too, satisfies (20), for a ( G ) =w ( G ) and w ( G ) = a ( C ) . (b) If G is a critically a-imperfect graph satisfying (20) then, by virtue of Theorem 8, n ( G ) = a ( G ) * w ( C ) + 1, therefore a (G) * o ( G ) + 1 = 2 a ( C ) 2 w ( G ) - 3 , implying w ( C ) = 2 or a ( G ) = 2 . The assertion now follows from Theorem 10. 0
c,
+
Let G = (X, U ) , Y C X;we define
r , ( Y ) = { x I x EX, x$Z Y, and there is a y E Y with ( x , y ) E U } ; we write Tc(y)instead of T C ( { y } ) .
Theorem 12. A critically a -imperfect graph G = ( X , U )or its complement G is an odd circuit L Z k + l ( k 3 2 ) if and only if for every maximum clique C of G (i.e., for every clique C = [Q], with I Q I = w ( C ) )
I rc(Q)ls w ( G ) + 1.
(21)
Proof. If the critically a-imperfect graph G or its complement circuit L 2 k + l (k 2 ) then (21) clearly holds. We shall show that (21) implies
n ( C )= 2 a ( G ) + 2 w ( G ) - 3 ; the assertion then follows from Theorem 11.
0
is an odd
Structure of perfect graphs
139
From Lemma 4 and the Corollary of Theorem 8 we derive the following statement:
For every vertex x EX, (Y([X2]c)= a(G)-l,
i.e., w ( [ X ' ] , ) = w(G)- 1, where X 2= Tc(x). From this it follows that every vertex x of G is contained in a maximum clique, say C = C(x) = We consider two cases. (I) There is a vertex y E Q, y # x such that
[a],.
([XI U {y)]c)
= a (G) = 1,
where XI = I'c(x). [X'U{y}],, being a proper subgraph of G, property (B) of critically a imperfect graphs (see Definition 6) yields
O([X' U{y}]c)= a ( G ) - l = e ( [ X ' ] , ) . Using Lemma 4 we conclude from this relation that there exist in X' at least two vertices which in G are adjacent to y. By hypothesis,
I rc(Q)Is w(G)+
1,
implying
I rc (x)-
(Q -{x))I
I Tc (Q) I - 2
w(G) - 1,
thus
I X 2I = I T C ( XI =) I T c ( x ) - ( Q -{XI) I + I Q - { x } 1
2w(G)-2,
and because of Corollary 2 of Lemma 4 (see (***)),
I X21 = 2 w ( G ) - 2 . (11) For every vertex y E Q, we have a ([XI U { y ) ] ~ = ) a (G). Let y E Q, y # x. By Lemma 4 (see (**)), (Y
([x2- { x '}I c ) = (Y (G ) - 1
or, equivalently, ~ ( [ X ' - { X ~ } ]w(G)-1 ~)=
for every x 2 E X 2 .
This implies the existence of a maximum clique C' = [ Q']c with x E Q', y $Z Q '.
E. Olaru, H. S a c k
140
For Q’ either case (I) holds (then 1 X z I = 2 w ( G ) - 2) or case (11) holds, i.e., &([XIU {y‘}],) = a(G) for every y ’ E 0‘. Now, in Q’ - { x } there is a vertex z such that ( y , z ) fZ U for, otherwise, Q ’ U { y } would span a clique in G with w ( G ) + 1 vertices, contradicting the definition of w ( G ) . The edges ( x ,y ) and ( x , z ) with ( y , z ) U are a-critical with respect to G ; consequently, because of Theorem 6, G is an odd circuit L 2 k + , (k 3 2) and thus, clearly,
1 X 2 1 = 2 w ( G )- 2 . So, in every case, we have obtained IX21=2w(G)-2 and, analogously,
1 x’I = 2 w ( G ) - 2
=2a(G)-2.
Hence,
n ( C )= I X iI + 1 X21+ 1 = 2 w ( G ) + 2 a ( G ) - 3 , proving Theorem 12. 0
Remark. Clearly, a critically a -imperfect graph G is a-non-partitionable (see proof of Theorem 7); this, in connection with Theorem 1, implies: (i) G is connected, (ii) no separating vertex set of G can span a clique in G . Theorem 13. Let T be a separating vertex set of the critically a-imperfect graph C = ( X , U ) such that a ( C - T )= a ( C ) . Then T spans a connected subgraph in the complement G of G . Proof. Assume that [TIt: is disconnected. Then T can be partitioned into two parts A and B such that A#0,
B#O,
A n B = 0 , A U B = T and ApcB.
Since, by hypothesis, C - T is disconnected, it can be decomposed into two subgraphs GI = [ X , ] , and G2= [X2IGsuch that
X , # 0, X 2# 0, X , f l X 2 = 0, X , U X , = X - T and
X,P,X,.
Structure of perfect graphs
141
B # 0 implies that [XIU X2 U AIc is a proper subgraph of G ;therefore, because of property (B) of G (see Definition 6), 8 ([Xi U Xz U A ]G) =
([Xi U X2 U A ] G ) = ( G ) .
(23)
Let denote a perfect a-partition of [Xi U X , U A]G and let, for i = 1,2, A, be the set of all those vertices of A contained in parts (ie., cliques) of the partition Z which also contain at least one vertex of X,. Then A , n A, = 0 and A 1U A, = A since two vertices x1 E XI and x2 E Xzcannot be contained in the same clique, and in Z there is no clique all of whose vertices belong to A (because of (22) and (23)). It is easily seen that
a ( [ X UA,],)=a(G,)
(i =1,2).
(24)
By means of a perfect a -partition of [XIU XzU BIG we define in an analogous manner the sets B1 and B2 satisfying:
B~n B , = 0,
B, u
= B,
a ( [ X UB,].)=a(G,) (i=l,2).
(25)
Further, we prove
a([Xi U A i U B , ] G ) = a ( G i ) (i =1,2). Let S be a stability system of [XlU A, U B1IG.If S C X , then, clearly, (26) is true (for i = 1); if S n A , # 0, then S n B , = 0 since AlpcBl (because of ApGB); analogously, if S n B , # 0, then S n A , = 0. The subscript 1 can be replaced by 2, and we conclude from (24) and (25) that (26) is valid in every case. Because of property (B) of G (see Definition 6), we obtain from (26) and (22): 8 ( G )= O([XiU X2U A U B I G )
e([x,u A, u B , I ~ ) +e([x2u A, u ~ = a([Xi U A I U Bi]c)+
(~([xz U A2U
~
1
~
)
&]G)
+
= a ( G , ) a(G2) = ( Y ( G ) ,
contradicting property (A) (see Definition 6) of critically Theorem 13 is now proved. 17
(Y
-imperfect graphs.
Theorem 13 has some important consequences. As a first conclusion, we obtain the following corollary: Corollary 1. Let G be a critically a-imperfect graph and let [Q], be an arbitrary clique of G. Then G - Q is connected.
142
E. Olaru, H.Sack
Proof. We may assume the clique [QIG to be maximal, so I Q 13 2. We have a(G - Q ) = a ( G ) .The graph [ Q ] B consists of isolated vertices only, hence, by Theorem 13, Q cannot separate G. 0
Remark. Corollary 1 of Theorem 13 can be used to immediately prove the following well-known theorem (see also [13], Theorem 1 (p. 150) and Folgerung aus Satz 4 (p. 153)): Theorem (Hajnal and Suriinyi [9], Berge [4]). Every triangulated graph is perfect. Let X' denote a separating vertex set of G = ( X , U ) and let C = [ YIGbe a connected component of [ X - XrlG. If in G every vertex of X' is adjacent to at least one vertex of Y then C is called a normal component with respect to X ' ; if [ X - X'Ic has at least two components which are normal with respect to X'then X' is called a normal separating vertex set of G. Corollary 2. Let G = ( X , U ) be a critically a-imperfect graph and let x be an arbitrary vertex of G. Then (a) the graph Glxl= [ X ' ] , is connected, and (b) XI is a normal separating vertex set of G.
Proof. (a) Clearly, for every x E X the vertex set X ' = r c ( x ) is a separating vertex set of the critically a-imperfect graph G. Further, a(G - X ' ) = a ( [ X 2U { x } ] ~ = ) a ( G ) since every vertex of the critically a-imperfect graph is contained in at least a ( G )stability systeps of (see Corollary to Lemma 3). The assertion now follows from Theorem 13. (b) Proposition (a) applied to G says that [ X ' ] , is connected: thus X' separates G, and G -X'has precisely two connected components, namely, [X2]a and the isolated vertex x. Since, by definition, xpc;-X'all we have to show is that in C every vertex of XI is adjacent to some vertex of X 2 . Assume that there is a vertex x ' E X ' which, in G, is not adjacent to any vertex of X2.Then, in G, x is adjacent to every neighbour of X I , i.e., the graph GixlJ= [Tc(x')lCis disconnected, contradicting (a). 0
c
It is convenient to restate Corollary 2 of Theorem 13 in the following equivalent form.
Corollary 2'. For every vertex x of a critically a-imperfect graph G the following statements hold : (a') the graph Gixl is connected, (b') the set X z of the neighbours of x is a normal separating vertex set of G.
Structure of perfect graphs
143
Remark 1. Corollaries 1 and 2 of Theorem 13 can be proved without using LovBsz’ theorem (see Theorem 5). Remark 2. Proposition (a) of Corollary 2 of Theorem 13 can be used to prove Dilworth’s theorem (see [13], Th. 3 (p. 150 and p. 158)): Theorem (Dilworth [S]). Every transitively orientable graph (‘comparability graph’) is a -perfect. Proposition (b) of the same corollary yields a simple proof of Gallai’s theorem (see [13], Th. 2 (p. 150 and pp. 164-165)):
Theorem (Gallai [7]). Every graph G having the property that each of its odd circuits (of length 2 5 ) is triangulated by some of its chords is a-perfect. Concluding Remark. The strong version of the Perfect Graph Conjecture (see Conjecture 1) is equivalent to the following: Conjecture 1’. The only critically a-imperfect graphs are the odd circuits of length 3 5 and their complements. By virtue of Theorem 10, Conjecture 1’ can also be given the following equivalent formulation:
Conjecture 1”. The only critically a-imperfect graphs G with a(G)> 3 are the odd circuits of length 2 7 . Therefore, it remains a central task to investigate the structure of critically a -imperfect graphs. In this direction, a somewhat weaker result obtained by Olaru in 1969 (see [13], Satz 14 and Folgerung (p. 169) or [17], Satz 3.1 (p. 99)) should be mentioned:
An a-imperfect graph G = ( X , V ) which has the property that G - u is a-perfect for every edge u E U is called edge-critically a-imperfect. Clearly, an edge-critically a -imperfect graph without isolated vertices is (vertex-) critically a -imperfect. Theorem 14. Every edge-critically a-imperfect graph consists of an odd circuit of length 2 5 and, possibly, some additional isolated vertices.
E. Oluru. H. Suchs
144
References [ I ] C. Bergc, Graphes et Hypergraphcs (Dunod, Paris, 1970). 121 C. Berge, Firbung von Graphen deren s8mtliche bzw. deren ungerade Kreise starr sind (Zusamrnenfassung). Wiss. Z . Martin-Luther-Univ. Halle 10 (1961) 114-1 15. [3] C. Berge, Une application de la theorie des graphes a un problkme de codage, in: Caianello. ed.. Automata Theory (Academic Press. New York-London. 1966) 25-34, [4] C. Berge, Some classes of perfect graphs, in: Graph Theory and Theoretical Physics, Chapter 5 (Academic Press, London-New York, 1967) 355-166. [ S ] R.P. Dilworth, A decomposition theorem for partially ordered sets, Ann. Math. 51 (1950) 161. [h] D.R. Fulkerson. Anti-blocking polyhedra, J. Comb. Theory 12 (1972) 50-71. 171 T. Gallai. Graphen init triangulicrbaren ungeraden Vielecken, Magyar Tud. Akad. Mat. Kutat6 Int. K(izl. 7A (1962) 3-37. [ X I P.C. Gilmore and A.J. Hoffman, A characterization o f comparability graphs and of interval graphs, Canad. J . Math. I6 (I‘J64) 53Y. [ 9 ] A. Hajnal and J. Suranyi. IJber die Aufltisung von Graphen in vollstandige Teilgraphen, Ann. Univ. Sci. Budapest I (19%) 1 1 3 . [lo] L. Lovasz. A characterization of perfect graphs, J. Comb. Theory 13 (1972) 95-98, [ I ] ] L. Lovasz, Normal hypergraphs and the perfect graph conjecture, Discrete Math. 2 (1972) 253-267 (this volume, pp. 29-42). (121 S.E. Markosjan and I.A. Karapetjan, Perfect graphs (in Russian), Akad. Nauk Armjan. SSR Dokl. 63 (1976) 292-296. [ 131 E. Olaru, Beitrage m r Thcorie der perfekten Graphen, Elektronische Informationsverarbeitung und Kybernetik (EIK) S (1972) 147-172. [14] E. Olaru. Zur Charakterisierung perfekter Graphen, EIK 9 (1973) 543-548. [IS] E . Olaru, Uber perfekte und kritisch imperfekte Graphen, Ann. St. Univ. Iagi 9 (1973) 477-486. [I61 E. Olaru. Uber die Uberdeckung von Graphen init Cliquen, Wiss. Z. TH Ilmenau 15 (1969) 115-120. 117) E. Olaru. Zur Theorie dcr perfekten Graphen, J . Comb. Theory, Ser. B 23 (1977) 94-105. [IS] H. Sachs, On the Berge conjecture concerning perfect graphs, in: Cornbinatorial Structures and Their Applications (Gordon and Breach. New York. 1970) 377-384. [ 191 W. Wessel, Some color-critical equivalents of the strong perfect graph conjecture, in: Beitrage
zur Graphentheorie und deren Anwendungen (Proc. Int. Koll. “Graphen theorie und deren Anwendungen”. Oberhof (DDR), I ( k l 6 April 1Y77) (Math. Gcs. D D R , TH Ilmenau, 1977)
30(k309.
Annals of Discrete Mathematics 21 (1984) 145-148 @ Elsevier Science Publishers B.V.
MEYNIEL’S GRAPHS ARE STRONGLY PERFECT G. RAVINDRA* ER 175 Combinatoire, Uniuersite‘ Pierre el Marie Curie, UER 48, 75230 Paris, France Meyniel (see article this volume, pp. 115-119) proved that a graph is perfect whenever every odd cycle of length at least five has at least two chords. This paper strengthens this result by proving that every graph satisfying Meyniel’s condition is strongly perfect (i.e., each of its induced subgraphs H contains a stable set which meets all the maximal cliques in H ) .
The graph-theoretic notions used here are those of [3]. A graph is called perfect if the chromatic number of each of its induced subgraphs H equals the number of vertices in the largest clique of H ; it is called strongly perfect [4] if each of its induced subgraphs H contains a stable set which meets all the maximal cliques in H. (Here, as usual, ‘maximal’ is meant with respect to set-inclusion.) It is easy to show that every strongly perfect graph is perfect. Meyniel [5] proved that a graph is perfect whenever each of its odd cycles of length at least five has at least two chords. The purpose of this paper is to strengthen Meyniel’s result as follows: Theorem. If every odd cycle of length at least ,five in a graph G has at least two chords then G is strongly perfect.
The following useful observation has been made by Meyniel. Lemma 1. If a graph G = ( V ,E ) contains an odd cycle [xo, xl,. . . , xzr,xd] such that the path [x,, xz, . . .,x Z f ]is chordless and xo is nonadjacent to at least one x k , then G contains an odd cycle of length at least five with at most one chord.
Proof. If x o x z E E then consider the largest i such that xo is adjacent to x1x2.- ax, and the smallest j such that j > i and xoxJ E E : the cycle [xo, x,, . . . ,x,, xo] has no chords, the cycle [xo,x , - ~ ,x,, . . . ,x,] has precisely one chord, and one of these two cycles is odd. If xoxz! i ! E then consider the smallest even j such that xoxJ E E and the largest i such that i < j - 2 and xox, E E : the odd cycle [xo,x,, ..., x,,xo] has at most one chord. 0 * Current address: Regional College of Education, Ajmer 305 004, India. 145
G. Ravindra
146
By a starter in a graph G, we shall mean a cycle wu0ul * . . U k W such that (i) uo is adjacent to none of the vertices u2, u 3 , . . . ,uk, (ii) w is not adjacent to uI, (iii) some stable set S, containing uI and u k , meets all the maximal cliques in
G - uO.
Lemma 2. lf n graph contains a starter then it contains an odd cycle of length at least five with at most one chord.
Proof. We shall present an informal description of an efficient algorithm which, given any starter in G = (V, E), finds the desired cycle. No generality is lost by assuming that (iv) [ ul, u z , . . . ,u k , w ] is the shortest path from u1 to w with the next-to-last vertex in S and all the vertices except the first and the last nonadjacent to uo. In particular, (iv) implies that (v) the path [ul, u 2 , .. .,uk] has no chords. Next, we may assume that (vi) every u, adjacent to w has an even subscript r ; otherwise the desired cycle can be found at once by applying Lemma 1 to the odd cycle W U o * . U,W. Write y E S" if y E S and y is adjacent to two consecutive vertices u,, u , + ~on the path [uO,u I , .. . , u k ] . We may assume that (vii) no y E S" is adjacent to uo; otherwise the desired cycle can be found at once by applying Lemma 1 to any odd cycle [ y u o - . u,y]. Now it follows that (viii) no y E S" is adjacent to w ; otherwise (iv) would be contradicted by uoul * * . u,yw such that i is the smallest subscript with yo, E E. Next, we may assume that (ix) each y E S * is adjacent t o at least three vertices on the path u0ul Uk ; otherwise the desired cycle is [ w , u,, . .. ,u,, y , u , + ~.,. . , u,, w ] with r standing for the largest subscript such that r < j, wu, E E and s standing for the smallest subscript such that s 3 j + 1, wu, E E. Now it follows that u,, u,+] on the path (x) each y E S " is adjacent to precisely three vertices [ U h 01,. . . , u t ] ; otherwise (iv) would be contradicted by uI . u,yu, - Dk such that r is the smallest subscript with yu, E E and s is the largest subscript with yu, E E. u, and u , + ~the , substitution Now observe that, for any y E S* adjacent to of y for u, in the original starter yields a new starter with a smaller S * .Repeating this operation as many times as possible, we eventually obtain a starter wVo& * Vkw satisfying (iv) and having an empty S*. Since k is even by (vi) and since V,, 6k E S, there is an edge fi,Gf+l with j > 0 and neither endpoint in S (it
-
4
4
Meyniel's graphs are strongly perfect
147
suffices to set j = i - 2 with i being the smallest even subscript such that 17, E S ) . Let C be any maximal clique extending 6,17,+~in G - uo. Since S * is empty, C is disjoint from S, a contradiction. 0
Proof of the Theorem. We shall present an informal description of an efficient algorithm which, given any graph G, finds either a starter in some induced subgraph of G or else a stable set meetirlg all the maximal cliques in G. The set of neighbours of a vertex u in G will be denoted by N ( u ) . First, choose a vertex t and a component H of G - t - N ( t ) so that the number of vertices in H is minimized (over all choices of t and H ) . If H = 0 then set S = { t } ;otherwise choose a vertex uo in H and denote by F that component of G - uo - N(uo) which contains t. (For future reference, note that each vertex x in H - uo - N ( u o ) belongs to F : otherwise the component of G - uo - N(uo)containing x would be fully contained in H - uo, and so it would have fewer vertices than H.) Apply the algorithm recursively to G - uo. When a stable set S meeting all the maximal cliques in G - vo is returned, search for vertices u l , u2 in G - uo such that uI E N(u o )fl S
and
v2 E N(vl)fl F.
If such vertices cannot be found then the stable set ( S - N(uo))U {uo}meets all the maximal cliques in G. (To verify this claim, assume that some maximal clique Q in G is disjoint from ( S - N(uo))U {uo}. Since uoE? Q, there is a vertex x E Q - N(uo) and there is a vertex uI E Q fl S. Of course, u1 E N(uo). If u1 E N ( t ) then we may set u2 = t ; otherwise u1E H and so x E H U N ( t ) , in which case we may set u2 = x.) If uI and u2 are present, search for vertices w and z , distinct from uo, u l , u2 and each other, such that w E N(u o )- N ( u l ) and
z E N ( w )fl F f S. l
If such vertices cannot be found then S meets all the maximal cliques in G. (To verify this claim, assume that some maximal clique Q in G is disjoint from S. Since u1E S, at least one vertex w E Q is nonadjacent to u I . Note that uo E Q, and so either (Q - vo) fl H # P, or else Q - v o c N ( t ) . Let Q " be any maximal clique in G - uo which extends Q - uo in the first case and (Q - uo)U { t } in the second case. In either case, we have Q * H U N ( t ) U { t } , and so Q * - N(uo) F. The vertex z common to Q * and S must be outside N(uo):otherwise Q could have been extended to Q U {z}.) If w and z are present then any path from uz to z in F yields a starter with v k = z. 0 Acknowledgements
The author thanks Prof. V. Chvatal for simplifying the proof of the main theorem and rendering the presentation more lucid. The author is also grateful to Prof. Claude Berge for fruitful discussions.
148
G. Ravindra
References [ 11 C. Berge, Farbung von Graphen, deren samtliche bzw. deren ungerede Kreise starr sind, Wiss. 2. Martin-Luther-Univ. Halle-Wittenberg 114 (1961). [2] C. Berge, Sur une conjecture relative au probleme des codes optimaux. comm. t 3 h e assemblee gtnerale de I’URSI, Tokyo (1962). [3] C. Eierge, Graphs and Hypergaphs (North-Holland, Amsterdam, 1973). (41 C. Berge and P. Duchet, Strongly perfect graphs (this volume, pp. 57-61). [ 5 ] H. Meyniel, On the Perfect Graph Conjecture, Discrete Math. 16 (1976) 339-342.
Annals of Discrete Mathematics 21 (1984) 140-157 @ Elsevier Science Publishers B.V.
THE VALIDITY OF THE PERFECT GRAPH CONJECTURE FOR K,-FREE GRAPHS Alan TUCKER Department of Applied Mathematics, State University of New York. Stony Brook, N Y 11 794, USA
1. Introduction
This paper builds on results based on D.R. Fulkerson's antiblocking polyhedra approach to perfect graphs to obtain information about critical perfect graphs and related clique-generated graphs. Then we prove that the Perfect Graph Conjecture is valid for 3-chromatic graphs. Fulkerson felt that a proof of the Perfect Graph Theorem would involve exactly the kind of duality that existed in his theory of blocking and antiblocking polyhedra [S], [6]. He proved what he called the Pluperfect Graph Theorem [7]. Lov6sz [8]independently developed some of the antiblocking theory, stated in hypergraph terms, and showed that with a fairly simple lemma, the Pluperfect Graph Theorem implied the Perfect Graph Theorem. Lovkz [Y] obtained a related result (Theorem 1 below) with a valuable implication about critical perfect graphs. A critical perfect graph, p-critical for short, is an imperfect graph all of whose proper induced subgraphs are perfect. The major results about p-critical graphs, especially those of Padberg [lo], are based o n an antiblocking polyhedra approach (also see [ 3 ] ) .
Theorem 1 [Y]. A graph G with n vertices is perfect if and only if a ( C ) w ( G ) for all induced subgraphs G' of G. Thus a p-critical graph G has ti a ( G ) w ( G ) + 1 vertices.
11.
=
Theorem 2 [lo]. A p-critical graph with n uertices has exactly 11 cliques of size w ( G ) with each vertex in w ( G ) maximal cliques and has exactly ri stable sets of size a ( G ) with each vertex in a ( G ) maximal stable sets. Each maximal clique intersects all but one maximal stable sets, and vice versa. We define a graph G to be pseudo-p-critical if (a) G has n = a ( G ) w ( G ) +1 vertices; 149
150
A. Tucker
(b) for all x in G, a (G - x ) = B(G - x ) and w ( G - x ) = y(G - x ) , (c) G has the properties listed in Theorem 2. Clearly a p-critical graph is pseudo-p-critical.
2. Unique covers and connectivity in G - x
Throughout this section, we assume that G = (V, A ) is a pseudo-p-critical graph with a = a ( G ) , w = w ( G ) , and n = I V ( = a w + 1. Then for any x E V, G - x is covered by a cliques of size o and also by w stable sets of size a (by property (b)). If C is one of the a cliques in the cover of G - x , then a ( G - x - C )= a - 1 and so if Z is the maximal stable set such that I f l C = 0, then x E Z. On the other hand, for each x E Z, one clique of the a clique cover of G - x will not intersect I; so C must be in all those clique covers. Further, C cannot be in any other clique cover of G - y for Z, or else C would have to intersect f just as it intersects all maximal independent sets in G - y. A similar argument applies to Z's appearances in stable set covers of G - x. So we have proved the following theorem:
ye
Theorem 3. I f G is pseudo-p-critical and I and C are an a-stable set and an w-clique, respectively, of G with Z f l C = 0, then C is in an a clique cover of G - x if and only if x E Z, and Z is in an w stable set cover of G - x if and only if x E C. CoroUary 3.1. Let G be pseudo-p-critical and x be any vertex in G. Then the w stable set cover (w-coloring) and the a clique cover of G - x are unique. Corollary 3.2. Let G be pseudo-p-critical, x be any vertex in G,and I,,Zz be two color classes in a w-coloring of G - x. Then the subgraph induced by Z, f l I2 is connected.
Proof. If not connected, the color classes could be interchanged in one component to get a different coloring, violating Corollary 3.1. 0 Corollary 3.3. Let G be pseudo-p-critical and C be a maximal clique of G. Then as y ranges over the w different vertices in C,the clique covers of the G - y contain all the other n - 1 ( = a w ) maximal cliques.
Proof. If the same clique was in covers for different y's, then by Theorem 3 the y's must be in a (maximal) stable set. 17 Let N ( x ) denote the set of x and the vertices adjacent to x.
The validity of the Perfect Graph Conjecture for K4-free graphs
151
Theorem 4. Let G be pseudo-p-critical and x be any vertex in G. Then G - N ( x ) is connected.
Proof. Suppose G - N ( x ) is not connected. Let GI be one component of G - N ( x ) and let Gz be the rest of G - N ( x ) . Let C1be a clique of the a clique cover of G - x and which is contained in GI U N ( x ) . Let Czbe defined similarly for GzU N ( x ) . Let a, = a ( G i ) , for i = 1,2. Clearly, a = (G - N ( x ) ) + 1 = a l + a Z + 1 ,and a ( G - C , ) = a , for i = 1 , 2 . This implies a ( G , - C , ) = a , , for i = 1,2. But then a -1 = a ( G - C1- Cz)= a ( ( G- N ( x ) ) - C1- G ) + 1 = a (GI- C , )+ a (GZ- Cz)+ 1 = a 1+ a z + 1 = a, a contradiction. 0
3. Maximal clique graphs Let us define M ( G ) , the maximal clique graph of the graph G, to have one vertex for each maximal clique of G (of size w (G)) and an edge between vertices which correspond to intersecting cliques. First we prove a maximal clique duality for pseudo-p-critical graphs. Then we prove that M ( G ) is pseudo-p-critical if G is pseudo-p-critical. Theorem 5. If G is a pseudo-p-critical graph, each maximal clique in M ( G ) corresponds to a vertex of G.
Proof. Since by Theorem 2, each vertex x of G is in w maximal cliques, then the corresponding w vertices in M ( G ) form a clique of size w. Now suppose that there exists a set of m, m 2 w, pairwise intersecting maximal cliques in G with no common vertex. Let C,, C,, . . . , C,,, be this set of cliques and I,,l 2 ,... , I,,, be the associated maximal stable sets with Cin Ii = 0. For any two ck,C, there is a vertex y E c k n C, and then by Theorem 3 the associated I,, I, are both in the stable set cover of G - y and so Ik f l I, = 0. Since this is true for any two of the I’s, the rn I’s are mutually disjoint. Since m 2 w, then it must be that m = w. Further, these 1 ’ s must contain all but one vertex of G, call it z , and so form a stable set cover of G - z. Then by Theorem 3, each Ck contains z. 0 Theorem 6. If G is a pseudo-p-critical graph with n = a w + 1 vertices, then M ( G ) is also pseudo-p-critical with n vertices and w ( M ( G ) )= w and a ( M (G)) = a.
Proof. By Theorems 2 and 5, we know that M ( G ) has n vertices, that w ( M ( G ) )= w, and that M ( G ) has exactly n cliques of size w. Since for any x, G - x has a cover of a (disjoint) maximal cliques, it follows that a ( M ( G ) )= a.
152
A. Tucker
Conversely, any stable set of a disjoint maximal cliques is a cover of G - z, for some z, and so by Corollary 3.1 M ( G ) has exactly n maximal stable sets. Clearly each maximal clique contains w vertices and by Theorem 3, is in a clique covers and so each vertex of M ( G ) is in w maximal cliques and a maximal stable sets. Also the maximal clique of M ( G )corresponding to the set of maximal cliques of G containing the vertex x is uniquely disjoint from the maximal stable set of M ( G ) corresponding to the a cliques covering G - x. Thus M ( G ) satisfies the properties of Theorem 2. If C is any maximal clique of G and I is the associated disjoint maximal stable set, then every other maximal clique of G contains one of the vertices in I and so the a cliques in M ( G ) corresponding to the vertices in I constitute a clique cover of M ( G ) - C of size a. Thus B ( M ( G ) - C ) = a. O n the other hand, by Corollary 3.3, the w stable sets of M ( G ) , corresponding to the clique covers of G - x for the w different x’s in the maximal clique C, cover M ( G ) - C. So y ( M ( G ) )- C ) = w. Corollary 6.1. If M ( G ) is the maximal clique graph of a pseudo-p-critical graph G,then for a vertex C in M ( G ) , M ( G ) - N ( C ) is connected.
Proof. This is an immediate consequence of Theorems 4 and 6. 0 For a given graph G, defined the skeleton S ( G )to be the subgraph of G with the same vertex set as G and only those edges of G that are part of a maximal clique of G.The following theorem follows from Theorems 5 and 6. Theorem 7. If G is pseudo-p-critical, then M ( M ( G ) )= S ( G ) and so S ( G ) is also pseudo -p-critical.
4. Perfect 3-chromatic graphs In this section, we prove that Berge’s Perfect Graph Conjecture is true for 3-chromatic graphs. First we prove a lemma that is of some interest in its own right. Lemma 8. Let G be pseudo-critical and let x be a vertex of G. Then, in the w -coloring of G - x , each i - j edge is on an i - j path whose terminal vertices are in N ( x ) . Proof. Let F stand for the subgraph of G induced by x along with all the i-vertices and all the j-vertices. If the lemma is false then, by Menger’s theorem,
The validity ofthe Perfect Graph Conjecture for K,-free graphs
153
there is a vertex s such that x and some i - j edge uv belong to distinct components of F - s. Let H stand for that component of F - s which contains the edge uu. We may assume that s is an i-vertex. Now consider the cover of G - s by CK cliques. Each of these cliques except one includes one vertex of each color; the exceptional clique includes x and one vertex of each color distinct from i. Thus for each i-vertex w other than s there is a unique j-vertex f ( w ) such that w and f ( w ) belong to the same clique in the cover. Since f is one-to-one and since f ( w ) E H if and only if w E H, it follows that the number of i-vertices in H equals the number of j-vertices in H. Choose a j-vertex t in H and consider the cover of G - t by CK cliques. Each of these cliques except one includes one vertex of each color; the exceptional clique includes x and one vertex of each color distinct from j. Thus for each i-vertex w in H there is a unique j-vertex g ( w ) in H such that w and g ( w ) belong to the same clique in the cover. Since g is one-to-one, it follows that the number of i-vertices in H is at most the number of j-vertices in H - r, a contradiction. 0
Theorem 9. Every graph containing no holes, no anthiholes, and no K4 is perfect. Proof. Assume the theorem false and let G be a counterexample with n vertices and rn edges such that no counterexample has fewer than n vertices and no counterexample with n vertices has fewer than rn edges. Since n is minimal, G is p-critical; since rn is minimal, every edge of G belongs to a triangle. (To clarify the second point, assume that an edge e of G belongs to no triangle. Since G - e retains the n triangles covering each vertex three times, we have CK (G - e ) s n / 3 = a ( G ) + f, and so G - e is imperfect. Furthermore, G - e has no holes, for otherwise G would have a hole. Finally, G - e has no antiholes of length at least seven, for otherwise G - e would have a K4. Thus G - e is another counterexample, a contradiction.) By Theorem 2, each N ( x ) in G induces three triangles. If these three triangles are arranged as in Fig. 1 for every x, then an easy argument (whose details we omit) shows that G is completely determined: its vertices can be enumerated as u , , v z , . . . ,u. ( n = 3 a ( G ) 1 ) in such a way that, with subscript arithmetic modulo n, vertices v, and v, are adj icent if and only if I i - j 1 s 2. But then G is either an antihole (if a ( G ) = 2) or tqsily seen to contain a hole (if a ( G ) 2 3). (Generalizations of the latter case are given in Tucker [13] and Chv5tal [4].) Now choose a vertex x such that the three triangles xcr', xdr and xab are not arranged as in Fig. 1 and consider the unique 3-coloring f of G - x. By the second part of Theorem 3, we may assume that
+
f ( c ) = 2, f ( r ' ) = 3 and f ( d ) = 1 , f ( r ) = 3 and f ( a ) = 1 , f ( b ) = 2 .
A. Tucker
154
X
Fig. 1.
Thus w e may assume that N ( x ) is arranged as in Fig. 2, possibly with r = r ' . If e is an edge on a 1-2 path w I ,w 2 , . .., wk with w I , wk E N ( x ) and w 2 ,w,, . . . ,Wk-I fZ N ( x ) then we shall write if w I = a and
Wk
= d,
e E S 2 if w I = b and
wk
= c,
e E S 3 if w I = a and
wk
=
e E SI
b
b2 2c r'3
3s & r2
Fig. 2.
(see Fig. 2). We claim that every 1-2 edge in G - x belongs to precisely one of the three sets S. To justify this claim, we first observe that every 1-2 path W I , W 2 , . . . , Wk with W l , Wk E N(X) and W 2 , W 3 , . . . , Wk-1 p N(X) must have { w l ,wk} = { a , b } , { b , c } or {a,b ) : otherwise w I , wk would be nonadjacent and have distinct colors, in which case the shortest path from w I to wk through w2, w 3 , ..., wk-1 would, together with x, yield a hole. This observation alone implies that SI,S2, S3 are pairwise disjoint; combined with Lemma 8, it shows that every 1-2 edge belongs to exactly one S,. In addition, note that Sl and Sz are nonempty since (by Corollary 3.2) the 1-2 subgraph of G - x is connected. Corollary 6.1 with the triangle abx in place of C guarantees that the hypergraph consisting of all the triangles disjoint from abx is connected. In particular, some triangle disjoint from abx and with a 1-2 edge in S1 shares its 3-vertex z with a triangle disjoint from abx whose 1-2 edge does not belong to S , . For simplicity of exposition, let us first assume that z#r
and
zZr'.
Let the three triangles which contain z be named C1,C2,C3 in such a way that one of the following five statements is true.
The validity of the Perfect Graph Conjecture for &free graphs
155
(i) each C, has its 1-2 edge in a different S,. (ii) C1has its 1-2 edge in SI and Cz, C3 have their 1-2 edges in S3, (iii) C1has its 1-2 edge in S1 and Cz, C3 have their 1-2 edges in Sz, (iv) C1has its 1-2 edge in S1 and Cz, C3 have their 1-2 edges in S1, (v) C1has its 1-2 edge in S3 and Cz, C3 have their 1-2 edges in S1. In case (i) or (ii), let vI, vz,. . . ,us be a shortest path with v I = z , v z = a and v2v3, v3v4,. . . ,vo-lvs E S3, and v,# b for all i. Let w l , w z , .. ., w, be a shortest path with w, = a, wl = d and w 1 w 2 w 3 ., ..,wl-lwl E S1 and wlwl+lE C1for some k. Now consider the closed walks V I V Z " ' V ~ W ~ " 'W k V i
and
V r V 2 " ' V S X W ~ W ~ - ~ " 'W k + I V I
of lengths s + k - 1 and s + 1 + t - k, respectively. Since t is odd (each w, is colored 1 if j is odd and 2 if j is even), precisely one of the two walks is odd. We claim that this odd walk L is a hole. To justify this claim, consider first the case when the odd walk is
L
=
vIv2"'
Vsw2. ' . Wkvl.
Since s + k - 1 is odd and s 2 2, k 2 1, we have s + k - 1 2 3. In fact, we cannot have s + k - 1 = 3, for then either s = 2, k = 2 and z is in two triangles with 1-2 edges in Sl, or else s = 3, k = 1 and s is not minimal. Furthermore, L has no chords v,v, (by minimality of s), no chords w,w, (by minimality of t ) , no chords v,w, (since z is in only one triangle whose 1-2 edge belongs to S , ) and no chords v,v,, 1 < i < s (for such chords would belong to both Sl and S3).Thus L is a hole. Next consider the case when the odd walk is
L
=vIvZ"'~sxwiwi-I"'Wk+lvI.
Since s + 1 + t - k is odd and s 2 2 , t - k + 1 3 3 , we have s + 1 + t - k 3 5 . As before L has no chords QV,, no chords w,w, and no chords v,w,, i < s. There are no chords beginning at x since N ( x ) = {a, b, c, d, r, r'}. To prove that there are no chords v,w, (i.e., aw,), recall that no vertex of G has four pairwise nonadjacent neighbors. Since x , us-, and wz are pairwise nonadjacent neighbors of v,, then or wz. The first endpoint w, of chord v,w, would have to be adjacent to x, two options are easily eliminated; the third is unavailable since the 1- or 2-vertex w, cannot be adjacent to both the 1-vertex v, and the 2-vertex w2. Thus L is a hole. In case (iii), let vl, vz,. . . , v, be a shortest path with vr = z and v, = x , and V Z v 3 , v3v4,.. . ,U , - ~ U , - ~E S2 and v,# b for all i. Let w,, w z , .. . , wl be as before. By the same reasoning used above, one can verify that one of the closed walks ViVz'
is a hole.
' *
V,Wi
* *
WkVi
and
ViVz'
.
'
V,WrWl-r
* * .
Wk+IVI
A. Tucker
156
Case (iv) reduces to (iii) by symmetry. In case (v), let ul, u 2 , . . . , us be a shortest path with u I = z, us = a and u203,. . . , E SI. Let wI,w 2 , .. . ,wI be a shortest path with w1 = a, w, = b and w ,w 2 , .. . , wl-l wIE S,. Note that t is even since a, b have distinct colors and that 1 < k < t - 1 since C1is disjoint from abx. Now one of the two closed walks UlU2"'
U,WI"'
WklJl
and
WIU~"'U,WIWI-I"'
Wk+iUi
is odd. Again this odd walk will be a hole. Finally, if z = r or z = r', then one of the z's triangles contains x, another, which we will call CI,has its 1-2 edge in SIand the third C2has its 1-2 edge in S2 or S,. If C2has its 1-2 edge in S3 then we may proceed as in case (v). If Czhas its 1-2 edge in S2 then we may assume z = r' (the case z = r reduces to z = r' by symmetry; it is also possible that z = r = r') and proceed as in case (iii) with r'x in place of u 1 u 2 .. . us. 0 Corollary 9.1. Euery planar graph containing no holes is perfect.
Proof. It will suffice to prove that every planar graph G containing no holes is w(G)-colorable. We shall prove this by induction on the number of vertices in G. If o(G)s 3 then the conclusion is immediate: since no planar graph contains an antihole of length more than five, G satisfies the hypothesis of Theorem 9. If w ( G ) = 4 then G is easily seen to be the union of graphs G,and GZ, each of them having fewer vertices than G, such that GI n G2 is a triangle. By the induction hypothesis, GI and G2 are four-colorable; now G is easily seen to be four-colorable.
0
References [ I ] C. Berge, Farbung von Graphen, deren samtliche bnv. deren ungerade Kreise starr sind, Wiss. 2. Martin-Luther-Univ. Halle-Wittenberg Math.-Natut. Reihe 114 (1961). 121 C. Berge, Introduction a la theorie des hypergraphes, Lecture Notes, UniversitC de Montreal (Summer 1971). [ 3 ] V. Chvatal, On certain polytopes associated with graphs, J. Comb. Theory, Ser. B 18 (1975) 138- 154. [4] V. Chvatal. On the strong perfect graph conjecture, J. Comb. Theory, Ser. B 20 (1976) 139-141. [S] B.R. Fuikerson, Blocking and anti-blocking pairs of polyhedra, Math. Program. 1 (1971) 168- 194. [ 6 ] D.R. Fulkerson. Anti-blocking polyhedra, J. Comb. Theory 12 (1972) 50-71. 171 D.R. Fulkerson, On the perfect graph theorem, in: T.C. Hu and S. Robinson, eds. Mathematical Programming (Academic Press, New York, 1973) 6%76. 181 L. Lovasz, Normal hypergraphs and the perfect graph conjecture, Discrete Math. 2 (1972) 25F267 (this volume, pp. 29-42). "41 L. Lovasii, A characterization of perfect graphs, J. Comb. Theory, Ser. B 13 (1972) 95-98.
The validity of the Perfect Graph Conjecture for K4-free graphs
151
[lo] M. Padberg, Perfect zero-one matrices, Math. Program. 6 (1974) 18CL196. [ll] K.R. Parthasarathy and G . Ravindra, The strong perfect graph conjecture is true for K,,-free graphs, J. Comb. Theory, Ser. B 22 (1976) 212-223. [12] A . Tucker, The strong perfect graph conjecture for planar graphs, Canad. J. Math. 25 (1973) 103-114. [13] A . Tucker, Coloring a family of circular arcs, SIAM J. Appl. Math. 29 (1975) 493-502.
This Page Intentionally Left Blank
PART 111
POLYHEDRAL POINT OF VIEW
This Page Intentionally Left Blank
Annals of Discrete Mathematics 21 (1984) 161-167 @ Elsevier Science Publishers B.V.
THE STRONG PERFECT GRAPH THEOREM FOR A CLASS OF PARTITIONABLE GRAPHS Rick GILES* School of Computer Science, Acadia University, Wolfville, NS, Canada
L.E. TROTTER, Jr.** School of Operations Research and Industrial Engineering, Cornell University, Ithaca, NY 14853, USA
Alan TUCKER*** Department of Applied Mathematics, State University of New York, Stony Brook, NY 1 1 794, USA A simple adjacency criterion is presented which, when satisfied, implies that a minimal imperfect graph is an odd hole or an odd antihole. For certain classes of graphs, including K,,,-free graphs, it is straightforward to validate this criterion and thus establish the Strong Perfect Graph Theorem for such graphs.
If graph G is imperfect but contains no imperfect proper induced subgraph, then G is minimal imperfect. Examples of such graphs are chordless odd cycles of length at least five and their complements, respectively termed odd holes and odd antiholes; Berge has conjectured that these are the only examples of minimal imperfect graphs. Conjecture 1 (Strong Perfect Graph Conjecture [l]). G is a minimal imperfect graph if and only if G is an odd hole or an odd antihole.
Whereas the reverse implication (‘if’) here is clearly valid, the forward implication (‘only if’) has been verified only for special classes of graphs [2,8, 12, 13, 15, 16, 17, 18, 191 and its general form has remained unsettled for some 20 years. We shall assume henceforth that G is a minimal imperfect graph. A direct approach to establishing Conjecture 1 leads naturally to the consideration of two types of information: properties of G which are a direct consequence of minimal imperfection and sufficient conditions for a minimal imperfect graph to be an * Partially supported by NSF Grant MCS8002987. ** Partially supported by NSF Grant ENG7809882. * * * Partially supported by NSF Grant MCS7904489. 161
R. Giles el al.
162
odd hole or an odd antihole. Considering first the former, let a and w denote the respective sizes of a largest stable set and a largest clique of G ;also let n = I V 1, where V is the vertex set of G. Certain direct consequences of minimal imperfection are immediate, e.g., that G is connected and that a 2 and w 3 2. From Lovasz’ characterization of perfect graphs [7] it follows easily that n=aw+1
(1)
for each u E V, V - u (denoting the set V - { u } ) may be partitioned into cliques of size w and into stable sets of size a.
(2)
and
A clique of size w (stable set of size a )is termed an w-clique (a-stable set). Padberg [9] has established:
G contains exactly n w-cliques and exactly n a-stable sets.
(3)
To each w-clique C in G there corresponds a unique a -stable set S in G such that C n S = 0.
(4)
The n x n incidence matrix of w-cliques with vertices of G is nonsingular with all row and column sums equal to w.
(5)
Straightforward consequences of (3)-(5) are the further properties (see [3,18]): The partitions described in (2) are unique.
(6)
For each u E V, the w-cliques containing u correspond (in the sense of (4))to the a-stable sets which partion V - u.
(7)
Additional properties of this type are described in [3, 11, 181. Next consider aspects of odd holes and odd antiholes which characterize them among minimal imperfect graphs. For example, recalling that G is minimal imperfect, clearly w = 2 implies
G is an odd hole and (8)
a = 2 implies G is an odd antihole.
Following [4] we denote by C : the graph with vertices uo, u l , . . . ,u,-, and edges {u,, a+,}, where 0 5 i 5 n - 1, 1 S j 5 k, and the index i + j is to be computed here (and below) modulo n. For n = 5,7,. . . , odd holes are of the form Ci with k = 1 and odd antiholes are given by k = (n - 3)/2. The following result (see [lo, 14, 171) shows that when the cliques of a minimal imperfect graph are ‘circular’ as in C:,then the graph must be an odd hole or an odd antihole. If G is isomorphic to C:for some k or an odd antihole.
3
1, then Gsis an odd hole
(9
The Strong Perfect Graph Theorem
163
This result has been strengthened by Chvdtal [4]:
If G contains a (spanning) subgraph isomorphic to Cf:for some k 2 1, then G is an odd hole or an odd antihole.
(10)
It is evident that the symmetry attributed minimal imperfect graphs by ( 5 ) is suggestive of the symmetry of graphs of the form Ci. Property (3,however, as well as properties (3), (4), (6) and (7), also holds (see [3, 181) for any graph which is parfitionable in the sense of (1) and (2). Furthermore, imperfect partitionable graphs which are not minimal imperfect have been given in [3,51. Thus a proof of Conjecture 1 based on (lo), i.e., showing that a minimal imperfect graph must contain a spanning subgraph isomorphic to Ci, should require the use of deeper information about minimal imperfection than that embodied in (5). In this paper we present the following strengthening of (10) which is directly linked to the partitioning information of (2). Theorem 1. If, for each v E V, the partition of V - v into a-stable sets has (at least) two members containing a single neighbor of v, then G is an odd hole or an odd antihole. Note that if for any v E V, the partition of V - v into a -stable sets has two members, S1 and S2, each containing a single neighbor of v, say v1 and u2 respectively, and v1 and v 2 are not adjacent, then G must be an odd hole. This follows from the fact that (6) implies that the subgraph H generated by S , U Sz U {v} is connected; hence a shortest cycle in H containing u, v1 and u2 is an odd hole. For certain classes of graphs the hypothesis of Theorem 1 is easily verified. A graph is K1,,-free if it does not contain
as an induced subgraph. Parthasarathy and Ravindra [12] have validated Conjecture 1 for K,,-free graphs. If the minimal imperfect graph G is KI.3-free, then the neighbor set of each vertex may be partitioned into two cliques, implying that each vertex has at most 2 0 - 2 neighbors. Since for each u E V, V - v partitions into w a-stable sets by (2), at least two of these a-stable sets contain only one of the 2w - 2 neighbors, and so validity of the hypothesis of Theorem 1 follows immediately. This method of obtaining the result of Parthasarathy and Ravindra was developed independently in [6, 191. Corollary 1 [12]. The Strong Perfect Graph Conjecture is valid for K1,3-freegraphs.
164
R. Giles et al.
Proof of Theorem 1. By (lo), it suffices to demonstrate that the vertices of G can so that for each 1 = 0,1,. . . ,n - 1, the vertex set be ordered uo, u,, . . . , {ul+,:0 =sj s w - 1) defines an w-clique of G; i.e., we will show that G contains the spanning subgraph C:i:l. Arbitrarily select uo E V and let {S,: 1 S j S w } denote the a-stable set partition of V - uo described in (2) (see Fig. 1). By hypothesis we may assume that uo is adjacent to only one vertex of SI, say u , E S , . Thus u 1 is adjacent only to uo among the vertices of the a-stable set Sl+m = SI- uI + uo (i.e., where we use the notation S + u for the set S U { u } ) and we have that {Sl+,: 1 S j 6 w } is the a-stable set partition of V - u1 (see Fig. 2). Continuing now inductively, suppose that k is fixed, 1 S k < n - 1, and that (i)-(iii) hold for 1 C 1 k. (i) uo, u , , . . . , uI are distinct vertices of G. is adjacent to u l . (ii) In S l + , = Sl - uI + ul-l only (iii) { S , + , :1 s j S w } is the a-stable set partition of V - uI. To complete the induction we verify (i)-(iii) for I = k + 1. Combining (ii) and (iii) for 1 = k with the hypothesis of Theorem 1 shows that some a-stable set St+,, 1 S j S w - 1, contains a single neighbor & + I of V k . Consider the case k < o depicted in Fig. 3 and first suppose u k + l E Sl+w. Since is in the a-stable set
00. . .
s,
...o
Fig. 1. a-stable set partition for V - u,.
Fig. 2. a-stable set partition for V - u , .
165
The Strong Perfect Graph Theorem
. . . s,,
...
..
s,,
. ..
...
s,,
..
...
s,
,
(00
6..
*
s,
*
..
*..
0) 0)
Fig. 3. a-stable set partition for V - uk,with k < w.
partitions for both V - uo and V - u k , it follows from (7) that the o-clique corresponding to & + I as in (4) contains both uo and u k . Thus uo and v k are adjacent and uo E Sl+w implies u k + l = uo. Substituting u k for uo in Sl+, yields the a-stable set partition of V - uo given by . . , S,, S1+, - v o + U k , SZ+-,. . .,s k + o } which is distinct from that specified by (iii) for I = 0. As this is in contradiction to (6), we conclude u k + l k ? S S l + w . A similar argument shows that u k + l $Z Si+-, for 1 S i S k - 1. Thus u k + l is a member of one of the a-stable sets &+I,. . . ,S,, and no generality is lost by assuming u k + 1 E S k + l . Note also that in this case (i) clearly holds for I = k 1. For the case k 2 w, we obtain, arguing as above, that if u k + l E & + j , 2 s j s w - 1, then u k + l = u k - o - l + j , again leading to a contradiction of (6) for V - u k - - - l + j . Thus in this case also u k + l E s k + l and it is again clear that (i) holds for 1 = k + 1 (see Fig. 4). Defining Sk+I+, = - u k + I + O k , we see immediately that (ii) holds for 1 = k + 1; (iii) follows by exchanging u k for u k + l in s k + 1 . An application of (iii) now shows that for 0s I s n - 1, the a-stable set S,+, is a member of the a-stable set partitions for
+
.
Fig. 4. a-stable set partition for V - u,, with k
3 o.
The Strong Perfect Graph Theorem
167
V - u l , .. . , V Thus it follows from (7) that for any choice of I, 0 s 1 s n - 1, the vertices uI,. . . ,ul+,-l constitute an w-clique of G, which completes the proof.
References [l] C. Berge, Farbung von Graphen, deren samtliche bzw. deren ungerade Kreise starr sind, Wiss. Z. Martin-Luther Univ. Halle-Wittenberg Math.-Natur. Reihe 114 (1961). [2] C. Berge, Graphs and Hypergraphs (North-Holland, Amsterdam and American Elsevier, New York, 1973). [3] R.G. Bland, H.-C. Huang and L.E. Trotter, Jr., Graphical properties related to minimal imperfection, Discrete Math. 27 (1979) 11-22 (this volume, pp. 181-192). [4] V. Chvatal, On the strong perfect graph conjecture, J. Comb. Theory, Ser. B 20 (1976) 139-141. [5] V. ChvBtal, R.L. Graham, A.F. Perold and S. Whitesides, Combinatorial designs related to the perfect graph conjecture, Discrete Math. 26 (1979) 83-92 (this volume, pp. 197-206). [6] F.R. Giles and L. E. Trotter, Jr., On stable set polyhedra for K,,3-free graphs, J. Comb. Theory, Ser. B 31 (1981) 313-326. [7] L. Lovasz, A characterization of perfect graphs, J. Comb. Theory, Ser. B 13 (1972) 95-98. (81 H. Meyniel, On the perfect graph conjecture, Discrete Math. 16 (1976) 339-342. [9] M.W. Padberg, Perfect zero-one matrices, Math. Program. 6 (1974) 180-196. [lo] M.W. Padberg, Characterizations of totally unimodular, balanced and perfect matrices, in: B. Roy, ed., Combinatorial Programming: Methods and Applications (Reidel, Boston, MA, 1975) 275-284. [ l l ] M.W. Padberg, Almost integral polyhedra related to certain combinatorial optimization problems, Linear Alg. Appl. 15 (1976) 69-88. [I21 K.R. Parthasarathy and G. Ravindra, The strong perfect-graph conjecture is true for K,,,-free graphs, J. Comb. Theory, Ser. B 21 (1976) 212-223. [13] K.R. Parthasarathy and G. Ravindra, The validity of the strong perfect graph conjecture for ( K , - e)-free graphs, J. Comb. Theory, Ser. B 26 (1979) 98-100. [14] L.E. Trotter, Jr., A class of facet producing graphs for vertex packing polyhedra, Discrete Math. 12 (1975) 373-388. [15] L.E. Trotter, Jr., Line perfect graphs, Math. Program. 12 (1977) 255-259. [16] A.C. Tucker, The strong perfect graph conjecture for planar graphs, Canad. J. Math. 25 (1973) 103-114. [17] A.C. Tucker, Coloring a family of circular arcs, SIAM J. Appl. Math. 29 (1975) 493-500. [18] A.C. Tucker, Critical perfect graphs and perfect 3-chromatic graphs, J. Comb. Theory, Ser. B 23 (1977) 143-149. [19] A.C. Tucker, Berge’s strong perfect graph conjecture, Ann. New York Acad. Sci. 319 (1979)
53s.535.
This Page Intentionally Left Blank
Annals of Discrete Mathematics 21 (1984) 169-178 @ Elsevier Science Publishers B.V.
A CHARACTERIZATION OF PERFECT MATRICES Manfred W. PADBERG New York University, New York, N Y 1ooo6, USA A zero-one matrix is called perfect if the polytope of the associated set packing problem has integral vertices only. By this definition, all totally unimodular zero-one matrices are perfect. In this paper we give a characterization of perfect zero-one matrices in terms of forbidden submatrices. The notion of a perfect zero-one matrix is closely related to that of a perfect graph as well as that of a balanced matrix as introduced by Berge. Furthermore, the results obtained here bear on an unsolved problem in graph theory, the Perfect Graph Conjecture.
1. Introduction
In this paper, we consider the polytope defined by the constraints of the following set packing problem: maxc x Axse xi = O o r 1 V j i E N = { l ,
..., n } ,
where A is an m X n matrix of zeroes and ones having no zero columns, eT = (1,. . . ,1) is the vector having all m components equal to one, and c is an arbitrary vector of reals. This problem has recently obtained much attention, see e.g. [l], [2], [6], [14], [18]. By (LP) we denote the linear programming problem obtained from (P) by dropping the integrality requirement on x. If the matrix A involved in problem (P) is totally unimodular [lo], then all basic feasible solutions to (LP) are integral, i.e., for any vector c the integer programming problem (P) can be solved as an ordinary linear programming problem. Generally, the matrix A encountered in (P) is not totally unimodular. Nevertheless, for certain matrices A the property that all basic feasible solutions to (LP) are integral, remains true (see Section 2 for relevant examples, also [5]). We call such matrices perfect zero-one matrices. Using some results from graph theory, we give a complete characterization of perfect matrices A in terms of forbidden submatrices. We give examples that show that - as in the case of balanced matrices, see [5] - it is not possible to characterize perfect zero-one matrices by means of forbidden determinantal values (as is the case for totally unimodular matrices). Indeed, given any natural number k, there exists a perfect zero-one matrix A such that it has a minor with determinant k. 169
M. W.Padberg
170
This paper is an abbreviated version of the original paper [15]; see also [16] where the main result of this paper is derived using a different proof technique.
2. Perfect Zero-One matrices
Let A be any m x n matrix of zeroes and ones having no zero column, and define the polytopes P and PI as follows:
P = { X E W I A X S ~ , X , ~ O ,j = 1,..., nl, P, = conv{x EW"1 AX c e, x, = o or 1, j = I , . . , n),
(2.1)
where e T= (1,. . . , 1 ) has m components all equal to one. The matrix A is called perfect if P = P I , i.e., if the polytope P defined in (2.1) has only integral vertices. Note that dim P = dim PI = n holds. Denote G the (intersection) graph associated with the matrix A, i.e., the nodes of G correspond to the columns of A and two nodes of G are linked b y an edge if the associated columns a' and a' of A have at least one + 1 entry in common, see e.g. [14]. A clique in G is a maximal complete subgraph of G. Let C denote the node set of any clique in G. Then by Theorem 2.1 of [14], the inequality
2 afx,
GI,
a; =
,=I
r
1 ifjEC (2.2)
0 if not
yields a facet of P I , i.e., a face of dimension n - 1 of P,. Clearly, every facet of the polytope PI is essential in defining P I .Hence it is a necessary condition for A to be perfect that A contain the incidence (row-) vectors of all cliques of the associated graph G. In order to characterize perfect matrices we can thus restrict ourselves to considering 'clique'-matrices, i.e., matrices A which contain the incidence vectors of all cliques of the associated graph G. Let A be any clique-matrix of size rn x n and let G be the associated graph. Let G denote the complement of G and denote B a clique matrix of G. Similarly to (2.1) define the polytopes Q and Q,, respectively, as follows: Q = { x E R " ( B x < ~ , x , z - o , j = 1 , ..., n),
I
0, = conv1x E R" BX s e, x, = o or I, j
= I,. .
.,n ) ,
(2.3)
where i?'=(l,. . . , l ) has all components equal to one and is dimensioned compatibly with B. Note that B has no zero column and that dim Q = dim QI = n holds. The vertices of Q, correspond to complere subgraphs of G, and vice versa. Furthermore, every maximal independent node set in G defines a clique of d (and vice versa). Consequently, there exists a (incomplete) 'duality' relation
171
A characterizationof perfect matrices
between the vertices of PI and the facets of Q (and, hence, between the vertices of Q l and the facets of P), see e.g. [ll].In the terminology of Fulkerson [7], Q (P, resp.) is the anti-blocker of Pl ( Q I , resp.). Let A be any clique-matrix of size m X n. By Theorem 1 of [ 6 ] ,A is perfect if and only if the associated graph G is perfect. Consequently, A is imperfect if and only if the graph G contains an induced subgraph of G’ which is almost perfect, i.e., G’ is imperfect, but every proper induced subgraph G’ is perfect, see [13], [15], [17]. Since induced subgraphs of G having k nodes correspond uniquely to m X k submatrices of A (and vice versa) we can make, without loss of generality, the assumption that G is an almost perfect graph. Denote a ( G ) the maximum cardinality of a stable (independent) node set in G and define C Y ( ~likewise ) for G. Lemma 1. Let A be a clique-matrix of an almost perfect graph G, LY = a ( G )and o = (Y ( G ) .Then xi < o provides a facet of QI and x = ( l / w ) e is a fractional vertex of P.
cY=,
+
Proof. Denote by e’ the row vector with n - 1 components equal to 1 and having a zero in the j-th component. Since G is almost perfect it follows from Theorem 1 of [13] that
e’x = a! for j = 1 , . . . , n. max XEP Define Hito be the halfspace given by
H, = { X E W ”I e ’ x s a } f o r j = 1 , ..., n. Hi for j = 1 , . . . , n or, equivalently,
Consequently P
n H’.
PCH=
j=l
By definition of e’, j apex at
x-
f f =-
n-1
=1,.
. . ,n, H is a pointed polyhedral convex cone with its
e.
By Theorem 1 of [13], we have that a ( G ) oa! (G) = n - 1 holds, since G is almost perfect. Hence it follows that Ai<-
wff
e = e. n-1 Consequently, since H is pointed, f f 1 x- =e=-e w n-1
M.W. Padberg
172
is a vertex of P, satisfying 0 < f < e. Consequently, the inequality provides a facet of Q,. 0
El=,xi s o
Remark 1. By Theorem 1 of [12]it follows that G is almost perfect if and only if G is almost perfect. Consequently, we have that = (l/cr)e is a fractional vertex of 0 and x, 6 cr provides a facet for P I .
x;=,
Remark 2. Since ff = ( l / w ) e is a vertex of P, we can reorder the rows of the matrix A as follows: A =
(2;)
where the row sums of A , all equal w and the row sums of A, are all (strictly) less than w. Furthermore, A, is of size m x n with m 2 riZ 2 n and A , contains a (at least one) nonsingular submatrix of size n X n. By Remark 1, we have a similar partitioning of B into B1and B2,where B , has m rows, m 2 n, having row sums equal to cr and B1contains a n X n nonsingular submatrix.
Remark 3. In graphical notation, Remark 2 implies that every almost perfect graph G contains at least I G I maximum cliques of the cardinality w = cr (G). This has been observed earlier by Sachs and can be found, though without proof, in [17]. F o r j = l ...., n define P,
=P
n {X E wn 1 X,
= 0)
(2.4)
and let 0, be defined analogously with P replaced by Q. Intersecting P with x, = 0 corresponds to the operation of deleting from the graph G node j and all edges incident to it. Hence we have
Remark 4. If A is a clique-matrix of an almost perfect graph G, then Pi PI (0,C Q,, resp.) holds for j = 1,. . . ,n. Lemma 2. Let A be an m X n clique-matrix of an almost perfect graph G, then A contains an n x n nonsingular submatrix A , whose column and row sums are all equal to w = a ( G ) . Furthermore, any row of A which is not in A l is either componentwise identical to some row of A , , or has a row sum strictly less than o.
Proof. By Remark 1, the point
3 = ( l / a ) e is a fractional vertex of
Q. The point
x’ = ( l / a)el satisfies x’ E Qi for j = 1,. . . ,n where e’ is the vector defined in the proof of Lemma 1. Since QiC Q, holds for j = 1,. . . ,n it follows by
convexity that X E 0,where
A characterization of perfect matrices
173
n
2
C
= ( l / n ), = I x 1 = ( w / n ) e .
(2.5)
c,"=,
Since i, = w holds it follows from Carathiodory's theorem [11] that there exists a n x n nonsingular submatrix A , of the matrix A , defined in Remark 2 such that n
i T = y T A , , y,>O
f o r i = l , ..., n
and
C y = 1
(2.6)
,=I
holds. Let Bl be any n X n nonsingular submatrix of the matrix Remark 2. Then we obtain from (2.6)
B, defined
in
Define D = All?:. Then D is an n X n nonsingular matrix of zeroes and ones since the columns of B: are (a subset of) vertices of PI (satisfying eTx < a with equality). Furthermore, D cannot contain a row consisting of + 1 entries only. For, if there were such a row, then necessarily B , would have to be singular since the row sums of B , all equal a. Then we have that
n-1 yTD= eT n where yi 2 0 for i = 1 , . . . , n and
x:=lyi = 1. From (2.7) we obtain
yTDe= n - 1. Consequently, 'yi > 0 implies that the row sum of row i of D equals n - 1, since y, = 1. Suppose now that -yl S * . . < yk with k < n satisfies yl > 0, yk+,= * = yn = 0. Since D is nonsingular we can rearrange the rows and columns of D such that D has the form
c:=, -
with 0 1
DI=(
. . .
. .
:)
, D2=(I
'
.
'
'
:)
. . . .
. 1 0
where Dl is of size k X k and has zeroes only in the main diagonal, D, is of size k x ( n - k ) and consists entirely of ones. If k < n, obviously, (2.7) cannot have a solution satisfying yi 3 0 and y, = 1. On the other hand y, = 1
z:=l
c:=,
M.W. Padberg
174
implies k 3 1. Consequently, k = n and D has the general form E - R, where E is the n X n matrix consisting entirely of ones and R is an n X n permutation matrix. Consequently, from (2.7) we have that
To complete the proof of Lemma 2, we note that the relation A,BT = E - R implies that A ; ' = - a E - B:RT. n-1
Suppose now that the matrix A contains a row aT satisfying are = w which is no1 contained in the submatrix A , . Then we have
0 S aTA = eT- aTB:RT= Z
eT.
(2.8)
From Ze = 1 and the integrality of Z it follows that such a row aT is componentwise identical to some row in A , and thus Lemma 2 follows. Remark 5. In graphical notation, Lemma 2 implies that every almost perfect graph G has exactly I G I maximum cliques of cardinality w = a(G) and 1 G I maximal independent node sets of cardinality a = a ( G ) . Lemma 2 suggests the following definition:
Definition. Let A be a zero-one matrix of size m X n, m 2 n. A is said to have property 7rB.,, if the following conditions are met: (i) A contains an n X n nonsingular submatrix A , whose row and column sums are all equal to p. (ii) Each row of A which is not a row of A l either is componentwise equal to some row of A , or has a row sum strictly less than p. Remark 6. Let A be an m X n matrix of zeroes and ones and G its associated intersection graph. A is a clique-matrix, i.e., A contains as row vectors the incidence vectors of all cliques in G, if and only if A does not contain any m x k submatrix A ' having the property 7rp.k with p = k - 1 and p 2 2. We will sketch the proof of Remark 6 only, since the assertion of Remark 6 is probably a known result. The necessity of the condition is obvious. To prove sufficiency, let C be the node set of a clique in G such that the associated incidence vector a.1 with aj = 1 if j E C,a, = 0 otherwise, is not contained in A. Define P ' = P f l { x E B P " I x j = O , V j E N - C } and P ~ = P ' n { x E W " ) x j = O } for all j E C. Obviously, P' must have at least one fractional vertex. We now
A characterization of perfect matrices
175
distinguish two cases: (i) Pi has integral vertices only for all j E C or (ii) there exists a j E C such that Pi has a fractional vertex. In case (i), we can show by an argument completely analogous to the one used in the proof of Lemma 1 that A contains an m X k submatrix A’ having property m6.k with p = k - 1, p 3 2 and k = 1 C I. In case (ii), we can restrict attention to any Pi having a fractional vertex for j E C. Abusing slightly the notation, we redefine C to be C - (i},redefine the polytopes P’, Pi with respect to the new set C and find again the two cases mentioned above. (Note that the new set C defines a complete subgraph in G, which is no longer a clique in G.This, however, does not affect the argument.) Clearly, case (ii) can happen only finitely many times and, finally, we obtain a set C having at leust three elements and which is such that case (i) prevails. This completes the outline of the proof of Remark 6. The following theorem states a necessary and sufficient condition for an arbitrary zero-one matrix to be perfect: Theorem 1. Let A be any zero-one matrix of size m X n. The following two conditions are equivalent: (i) A is perfect. (ii) For p 5 2 and 3 G k C n, A does not contain any m x k submatrix A‘ having the property T @ . k . Proof. Suppose that A is perfect and that (ii) is violated. Then there exists an m X k submatrix A ’ of A having property r 6 . k for some p 3 2 and k 5 3. Suppose the columns of A have been ordered such that A ’ coincides with the k first columns. Then 2 defined by Zj = 1/p for j = 1,.. . ,k, Zj = 0 for j = k + 1 , . . . , n, is a fractional vertex of the polytope P defined in (2.1). Hence, by definition, A cannot be perfect. O n the other hand, suppose that A is such that (ii) holds. Then A must be perfect. For if not, then by Remark 6 and Theorem 1 of [6] the intersection graph G associated with A must be imperfect. Consequently G contains an induced subgraph G’ that is almost perfect. Again by Remark 6, the clique-matrix of G’ is an m x k submatrix A ‘ of A where k = 1 G’l. Let A ” denote the m ’ x k submatrix of A ’ whose rows correspond to the cliques of G’. By Lemma 2, A ” has the property T 6 . k with p = cr 3 2. Since the m - m’ truncated rows of A not contained in A ” are dominated by some row in A ” ,the m x k submatrix A ’ of A must also have property T 6 . k with p = cr(G’).Thus (ii) cannot be satisfied by A . 0
(c’)
Corollary 1. L e t A be a zero-one matrix of size m X n. A is almostperfecf, i.e., A is the clique-matrix of an almost perfect graph G, if and only if the following two conditions are met:
176
M.W.Padberg
(i) A has the property r 6 . k for some p satisfying 2 S p sz [ n / 2 ] and k = n. (ii) A does not contain any m X k submatrix having property q , k for p 2 2 and 3 G k S n - 1. Remark 7. The strong perfect graph conjecture [ 2 ] ,[ 3 ] ,if true, now is reduced to proving that the only zero-one matrix A of size m X n satisfying the conditions (i) and (ii) of Corollary 1 and 2 G p < [ n / 2 ]is the circulant of odd size having exactly two positive entries in every row and column, i.e., A is the clique-matrix of an odd cycle without chords. A characterization of almost perfect matrices in graphical terms appears advantageous if one wants to check the perfection of a zero-one matrix. For, similarly to the criterion for the total unimodularity of a matrix A, a direct check of the perfection of a zero-one matrix via the necessary and sufficient criterion of Theorem 1 is - computationally - an impossible task, whereas graphical criteria - at least i n the context of total unimodularity - are relatively easily verified or known to be satisfied by the physical conditions of a problem under consideration. Example 1. Since a zero-one matrix A is perfect if it is the clique-matrix of a perfect graph (and vice versa), the clique-matrices of perfect graphs furnish examples of zero-one matrices satisfying the condition (ii) of Theorem 1. Among the graphs known to be perfect are the rigid circuit graphs [7] (or ‘triangulated’ graphs [2]),the comparability graphs and the ‘i-triangulated’ graphs, see e.g. (21. The example of rigid circuit graphs provides examples of zero-one matrices which are perfect, but nor totally unimodular. (I am indebted to D.R. Fulkerson for this example.)
Example 2. Consider the graph G of Fig. 1 and its associated clique-matrix A in Fig. 2. The submatrix A ’, made up of columns 1 , 2 , 3 and 4 and rows 1,6, 11 and 16, has a determinant of 3. Consequently, A is not totally unimodular. Due to the simple structure of A, we find by inspection that A is perfect. (Checking condition (ii) of Theorem 1 amounts to proving that G does not contain an odd cycle without chords.) Furthermore, the matrix A provides us with an example where AT, the transpose of a perfect matrix, is not perfect. This is remarkable since it is different from total unimodularity. The above example can be generalized to prove that given any natural number k there exists a perfect matrix A having a minor whose determinant in absolute value equals k. (Replace K4by K k + ,and the G,, i = 1,. . . , 4 , by k + 1 copies of GI,say, each of which is connected to & + I in a similar fashion as done above.) This indicates why a characterization of perfect matrices in terms of forbidden matrices is appropriate rather than a characterization in terms of forbidden subdeterminants (which is not possible here, but possible for totally unimodular matrices).
A characterization of perfect matrices
18
17\
Fig. 1.
I
2 3
4 S 6 7 X 9
10 11 12 13 14
0 I 0 0 0 0 1 1 0 0 0 0 1 1 0 0 0 0 1 1
0 0 0 0 0 0 0 0 0 I 0 0 I 0 0 0 1 1 1 0 0 0 0 1 0 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 0 0 1 1 0 0 0 0
A=
1 0 0 0 0
1 0 0 0 0
0 0 0 0 0 0 0 0 0 0 1 0 1 1 0 0 0 0 0 o 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 0
0 0 0 0 0 0 0 1 0 0 0 0 0 0 I 1 0 0 0 I 0 0 1 0 0 0 0 0 0 0 0 o o o o 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 I) 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 Fig. 2.
0 0 0 0 0 0 0
IS 16 17 I X
19 20
0 0 0 0 0 0 0 0 0 0
0 0 0 0 0
0
0 0
0
0
1 0
I 0 0 1 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
I
o o o o o o
1
1 0 0 0 0 0 1 I 0 0 0 0 0 1 0 0 0 0 0 0 1 0 0 0 0 0 I 1 0 0
0 0 0 0 0 0
Ll
0 0 0
I
I
0
0 0 0 0 1 I 0 0 0 I 0 0 I 0 0 0 0 0 0 0
178
M.W. Padberg
References [I] E. Balas and M.W. Padberg, On the set covering problem, Oper. Res. 20 (1972) 1152-1161. [2] E. Balas and M.W. Padberg, On the set covering problem 11: an algorithm, Oper. Res. 23 (1975) 74-90. (31 C. Berge, Graphes et hypergraphes (Dunod, Paris, 1970). [4] C. Berge, Introduction B la thtorie des hypergraphes, Lecture Notes, Universit6 de Montrtal (Summer 1971). [S] C. Berge, Balanced matrices, Math. Program. 2 (1972) 19-31. [6] V. Chvhtal, On certain polytopes associated with graphs, J. Comb. Theory, Ser. B 18 (1975) 138-154. 171 D.R. Fulkerson, Blocking and antiblocking pairs of polyhedra, Math. Program. 1 (1971) 168-194. [8] D.R. Fulkerson, On the Perfect Graph Theorem, in: T.C. Hu and S.M. Robinson, eds., Mathematical Programming (Academic Press, New York, 1973) 69-77. [9] F.R. Gantmacher, Matrix Theory, Vol. 2 (Chelsea Publishing Company, New York, 1964) (translated by K.A. Hirsch). 1101 R. Garfinkel and G . Nemhauser, Integer Programming (Wiley, New York, 1972). (111 B. Griinbaum, Convex Polytopes (Wiley, New York, 1966). [12] L. Lovhsz, Normal hypergraphs and the perfect graph conjecture, Discrete Math. 2 (1972) 253-268 (this volume, pp. 29-42). [13] L. Lovasz, A characterization of perfect graphs, J. Comb. Theory, Ser. B 13 (1972) 95-98. [14] M.W. Padberg, On the facial structure of the set covering problem, Math. Program. 5 (1973) 199-215. [15] M.W. Padberg, Perfect zero-one matrices, Math. Program. 6 (1974) 18CL196. [16] M.W. Padberg, Almost integral polyhedra related to certain combinatorial optimization problems, Linear Algebra & Appl. 15 (1976) 69-88. [17] H. Sachs, On the Berge conjecture concerning perfect graphs, Combinatorial Structures and their Applications (Gordon and Breach, New York, 1970) 377-387. [IS] L. Trotter, Solution characteristics and algorithms for the vertex packing problem, Technical Report no 168, Ph.D. Thesis, Cornell University (January 1973).
PART IV
WHICH GRAPHS ARE IMPERFECT
This Page Intentionally Left Blank
Annals of Discrete Mathematics 21 (1984) 181-192 @ Elsevier Science Publishers B.V.
GRAPHICAL PROPERTIES RELATED TO MINIMAL IMPERFECTION* R.G. BLAND School of Operations Research and Industrial Engineering, Cornell University, Ithaca, NY 14853, USA
H.-C. HUANG** Department of Economics and Statistics, Nanyang University, Singapore
L.E. TROTTER, Jr.*** School of Operations Research and Industrial Engineering, Cornel Uniuersity, Ithaca, NY 14853, USA Say that the graph G is partitionable if there exist integers a 2 2 , w 2 2 , such that V ( G ) \u into stable sets of size a and into cliques of size w . An immediate consequence of Lovasz’ characterization of perfect graphs is that every minimal imperfect graph G is partitionable with a = a ( G )and w=w(G). Padberg has shown that in every minimal imperfect graph G the cliques and stable sets of maximum size satisfy a series of conditions that reflect extraordinary symmetry in G. Among these conditions are: the number of cliques of size w ( G ) is exactly I V ( G ) I ;the number of stable sets of size a ( G ) is exactly I V ( G ) / every ; vertex of G is contained in exactly o ( G ) cliques of size w ( G ) and a ( G ) stable sets of size a ( G ) ; for every clique Q (respectively, stable set S) of maximum size there is a unique stable set S (clique Q ) of maximum size such that Q n S = 0. Let C: denote the graph whose vertices can be enumerated as u I , . . . , u, in such a way that u, and u, are adjacent in G if and only if i and j differ by at most k, modulo n. Chvital has shown that the Perfect Graph Conjecture is equivalent to the conjecture that if G is minimal imperfect with a ( G ) = a and w ( G ) = w , then G has a spanning subgraph isomorphic to
I V ( G )I = a m + 1 and for every u E V ( G )there exist partitions of
c:::
I.
Padberg’s conditions are sufficiently restrictive to suggest the possibility of establishing the Perfect Graph Conjecture by proving that any graph G satisfying these conditions must contain a spanning subgraph isomorphic to C:::,, where a ( G )= a and w ( G )= w. It is shown here, using only elementary linear algebra, that all partitionable graphs satisfy Padberg’s conditions, as well as additional properties of the same spirit. Then examples are provided of partitionable graphs which contain no spanning subgraph isomorphic to C:::,, where a ( G )= a and w ( G )= w.
* Reprinted from Discrete Math. 27 (1979) 11-22. ** Research partially supported by NSF Grant ENG 76-09936.
*** Research partially supported by NSF Grant ENG 76-09936 and Sonderforschungsbereich 21 (DFG). 181
R.G. Bland ef al.
182
1. Introduction
A graph G is called perfect if every induced subgraph H of G has chromatic number y ( H ) equal to w ( H ) , the clique number of H (the size of a largest clique in H). The notion of perfection is due to Berge (see [l],[ 2 ] , [3]) who offered two conjectures concerning perfect graphs. Lovisz [ll] proved the weaker of the two conjectures:
Theorem 1 (The Perfect Graph Theorem [ll]). A graph G is perfect if and only if the complement graph d is perfect. Thus G is perfect if and only if for every induced subgraph H of G the srability number a ( H ) , the size of a largest stable set (or anticlique), is equal to 6 ( H ) , the minimum number of cliques whose union covers H. Let n ( G ) denote the number of vertices in graph G. One of Lovisz’ 1121 proofs of Theorem 1 established the following stronger result.
Theorem 2 [12]. G is perfecr if and only if a ( H )- w ( H )2 n ( H ) for all induced subgraphs H. However, the stronger of Berge’s conjectures remains unsettled.
Conjecture 1 (The Perfect Graph Conjecture [l]). G is perfect if and only if in G and
every odd cycle of length at least five has a chord.
Chordless odd cycles of length at least five have been termed odd holes and their complements odd antiholes. Obviously, any graph that has an odd hole or an odd antihole is not perfect. The converse has been established only for special classes of graphs [2], [13], [16], [17], [IS], [19], [21], [22], 1231. An imperfect graph G is called minimal imperfect if it has no imperfect proper induced subgraph. It is evident that every minimal imperfect graph G has the following properties.
G is connected; a ( G ) 2 2 and
(1.la)
o(G)a2;
(l.lb)
w (G) = 2 if and only if G is an odd hole and a (G ) = 2 if and only if G is an odd antihole;
(Llc)
G is minimal imperfect.
(l.ld)
It is easy to deduce from LovBsz’ characterization of perfect graphs (Theorem 2 above) that every minimal imperfect G must have the additional properties
183
Graphical properties related to minimal imperfection
n =aw+1
(1.2a)
and for every vertex u of G there is a partition of V \ v into cliques of size w and a partition of V \ v into stable sets of size a, (1.2b) where a = a ( G ) , w = w(G), n = n(G), and V'= V(G), the vertex set of G. Further properties of minimal imperfect graphs have been determined by Padberg [14]. Say that a clique (respectively, a stable set) of size k is a k-clique (k-stable set). Padberg has shown that if G is minimal imperfect, then
G has exactly n w-cliques, every vertex of G is in exactly w-cliques, and the n x n incidence matrix of o-cliques with vertices of G is nonsingular.
w
(1.3)
By (1.ld) the analogous property for stable sets of a minimal imperfect graph also holds. Say that a graph G with n vertices is partitionable if there exist integers a 3 2 and w 3 2 such that G satisfies (1.2). Clearly, G is partitionable if and only if is partitionable. The comments above indicate that every minimal imperfect graph is partitionable. In fact, it is easy to see that a graph is minimal partitionable if and only if it is minimal imperfect. In the next section we show that every partitionable graph satisfies (1.3). The proof uses only elementary linear algebra. Related results of Padberg [15] and some new results of the same spirit are also derived from (1.2); these results further emphasize the rigid symmetries of partitionable graphs and, therefore, of minimal imperfect graphs. Among the new results is one related to adjacency in polyhedra associated with minimal imperfect graphs. The algebraic nature of the derivations is in the spirit of the previous work of Fulkerson [7], [8],[9] and Padberg [14], [15] on perfect graphs. The simplicity of the derivation of (1.3) from (1.2) and the additional generality that it affords may further elucidate Padberg's interesting results. Chvhtal [5] denotes by Ct the graph whose vertices can be enumerated as uI,. . . , un so that u, is adjacent to uj if and only if the indices i and j differ by at most k modulo n. (Properties of the complements of graphs of the form Ci are studied in [20].) Clearly, for 1 2 2 C:,,,is an odd hole and Ci;:, is an odd antihole. Chvfital [5] proves that Conjecture 1 is equivalent to the following: Conjecture 2 [5]. If G is minimal imperfect with a ( G )= a and w ( G ) = o,then G has a spanning subgraph isomorphic to CZ:l.
In light of the apparently stringent symmetries embodied in (1.3), one might hope to prove Conjecture 2 and, therefore, Conjecture 1, by showing that any
184
R.G. Bland et a1
graph G on ao + 1 vertices, where a = a ( G ) , w = o(G), such that G and G satisfy (1.3), has a spanning subgraph isomorphic to CZ:,.In Section 3 we give two counterexamples, graphs which are partitionable (and hence satisfy 1.3), but which have no spanning subgraph isomorphic to CZ:l, where a = a ( G ) , w =w(G).
2. Properties of partitionable graphs
We will now proceed to derive Padberg’s property (1.3) from (1.2). The derivation may lend some insight into what additional structure associated with minimal imperfection must be invoked along with (1.3) in order to establish Conjecture 1. For the remainder of this section it is assumed that G is a partitionable graph with V ( G ) = V and that a 2 2 and w 2 2 are integers such that { G , a , w } satisfies (1.2).
Claim 1. a ( G ) = a and w ( G ) = w .
Proof. From (1.2) it is clear that a ( G )b a and o(G)2 w. Suppose that G has a
stable set S with 1 S I = a + 1. Since a 5 2, w 3 2 property (1.2a) implies that n a 2a + 1 > a + 1, so there exists u E V \ S. By (1.2b) there is a partition of V \ u into a w-cliques of G, which implies that two distinct vertices of the stable sets S are in a common clique, a contradiction. Thus I S 1 s a for all stable sets S in G. Similarly 1 Q 1 s w for all cliques Q in G.
Claim 2. For every clique Q in G there exists an a-stable set S in G such that QnS=@. Roof. For any o E Q there is a partition of V \ u into stable sets SL, S 2 , .. . ,S,. By Claim 1 10 \ u I < w - 1. Since each vertex in Q is in at most one of S , , S I , .. . ,S,, we have Q f l Si = 0 for some 1 S i 6 w.
In the case where the partitionable graph G is minimal imperfect, the following Claims 3-5 culminating in Proposition 6 are either due to Padberg [14] or follow easily from his results. In what follows 0, denotes the k-vector (0,. . . , O ) , l k denotes the k-vector (1,. . . , l), and the superscript t is the transpose operator. Claim 3. There are n cliques in G whose n and has all row and column sums equal to
X
n incidence matrix is nonsingular
w.
Graphical properties related to minimal imperfection
185
Proof. Let Q = { v l , . . . ,uw}be any o-clique of G. Let c, be the incidence vector of Q and for i = 1,.. . ,w, let Mi be the a X n clique-vertex incidence matrix of some partition of V \ ui into w-cliques. We then define C to be the matrix
Since each M , is an a X n matrix and n = a w + 1, C is an n X n matrix. Also note that the row and column sums of C are all w. Therefore y = (l/w)l,, is a solution of the system y C = 1,. We will prove that C is nonsingular by showing that y is the unique solution. Suppose that for some y # y we have yC = 1,. Since j j > 0, we lose no generality in assuming that y 2 0. For i = 1,. . . ,n let e, denote the ith unit vector, and let c, denote the ith row of C. By Claim 2 there corresponds to each c, an incidence vector a, of an a-stable set such that a, . c, = 0. Hence for any i = 1,. . . ,n we have a = (1, - c l ) .a, = [(y - e , ) C ] .a, = (y - er)(Ca:) =
c
,=1
y,(c, * a z ) + ( y , - l)(c, . at).
f#I
But a = ( n - l)/w, c, a, = 0 or 1 for i# j , and c, - a, = 0. Therefore, since y we have
2 y j 3 -n. - 1 i=l
w
I#i
On the other hand, 11 =
1, * 1, = (yC). 1" = y(C1',) = y . ( w l . )
Hence we have that
From (2.1) and (2.2) we see that y
= (l/w)ln.
=w(y
*
1.1.
20
R.G. Bland et al.
186
It is important to note that the incidence vectors ai in the proof of Claim 3 must in fact satisfy ai cj = 1 for all i f j . We let Sidenote the a -stable set with incidence vector ai and denote by A the n X n matrix whose ith row is ai, i = l , ..., n. Claim 4. Let E and I denote the n X n matrix of ones and the n X n identity matrix, respectively. Then CA' = E - I, so C-' = ( l / w ) E - A', i.e., C-' has as the entry in its jth row and ith column w-1
, if uj E Si, otherwise.
-
Proof. This follows immediately from the fact that ai ci CE = WE,as implied by Claim 3.
= 1 for
i# j and from
Claim 5. G has exactly n w-cliques.
Proof. Let c be the incidence vector of any maximum clique in G. Then y = cC-' is the unique solution of the system yC = c. From Claim 4 we see that Yi
0 if ai * c = 1, 1 if ai - c =O.
=
Also w
c
i=l
yi
=y
(a:=)(yC). 1" = c . 1,
cT=,
= w,
implying that y, = 1. Thus y = ( y , , . . . ,yn) has y, for all j# i, so c = c,.
=
1 for some i and y, = 0
It follows from Claim 5 and the proof of Claim 3 that the a-stable set S of Claim 2 having S n Q = 0 for some w-clique Q is unique. The results above are summarized in the following proposition.
Proposition 6. Let a 3 2 be integers and let G be a graph that satisfies (1.2). Then G has exactly n = (YW+ 1 (maximum) w-cliques. There is a 1-1 correspondence between w-cliques and a-stable sets, pairing each o-clique Q with the unique a -stable set S having S n Q = 0. The incidence martix of @-cliques with vertices is nonsingular with all row and column sums equal to w. Furthermore G, the complement of G, and 6 = w. W = (Y also satisfy (1.2).
Graphical properties related to minimal imperfection
187
Since minimal imperfect graphs are partitionable, Proposition 6 includes Padberg's property (1.3) for minimal imperfect graphs. Further properties follow from those outlined above. These will be stated in terms of cliques; analogous properties obviously hold for stable sets. Claim 7. The partition of the vertices of G\vi into w-cliques is unique. Proof. A partition of the n - 1 = a w vertices of G \ v, into w -cliques represents a solution of the system yC = (1, - e!). Since C is nonsingular, y is unique. Claim 8. For v E V, the a a-stable sets that contain v correspond (in the sense of Proposition 6) to the a w-cliques in the unique partition of the vertices of G \ u into w-cliques. Proof. Let v E V and let {Ql, Q 2 , . .. , Q,} be the unique partition of V \ u into w-cliques. Recall that for 1 i G a, Si is the unique a-stable set having S, fl Q, = 0. Since IS, n 0,I = 1 for all j # i, we have 1 S, n (V\{v})i = a - 1. But 1 S, I = a, so it must be that v E Si. Let B be the p x n incidence matrix of all maximal cliques of G ;p 2 n since G may have maximal cliques Q with 1 Q I < w. Let R : denote the nonnegative orthant of R " and let BG and PG be the polytopes
PG = { x E R : : Bx'G l;} and
BG
= convex
hull
{X
E {0,1}" : B x ' S 1;).
Clearly BG C PG and the set of integer extreme points of PG is the same as the set of extreme points of BG;it is precisely the set of incidence vectors of stable sets of G. Furthermore, since C is a nonsingular submatrix of B and Cx' = 1; for x = (l/w)l,,, it follows that (l/w)ln is an extreme point of PG. Padberg proved the next claim. Claim 9 [15]. If G is minimal imperfect, then the point (l/w)ln is the unique fractional extreme point of PG. The following lemma, which can be easily proved, will be useful in the proof of Claim 9. Lemma 9.1. Let 61,.. . ,S. l/S1+. * + 1/S, 3 n / s .
-
be
positive
with
+
S , +. . . 6, s nE.
Then
R. G. Bland et al.
188
Proof of Claim 9. Note that every x E Pc must satisfy Cx' s lt,, or equivalently, s' = 1: - Cx' 3 From Claim 4 C ' = ( ( l / w ) E - A '), and as already observed, C-'l: = (l/w)lk. Let LT = 1, . s. Then 1
c)w
lL+ A'S'.
(2.3)
Observe that I1
I;x
= (l -V -+C Ta , w
and so 1,
' X
z a
if and only if
CT 4
1.
Since s 3 0" wc have u 2 0, and from (2.3) and (2.4) we see that 1, . x 3 a if and only if I is a convex combination of the extreme point ( l / w ) l n of Pc and the extreme points of PG corresponding to a-stable sets in G. Suppose that x = (x,, . ..,x,) is a fractional extreme point of PG. Then clearly 1 > x, > 0 for j = 1.. . . , n, since G is minimal imperfect. Hence there exist n cliques of G yhose n X n incidence matrix D is nonsingular and has Dx' = 1:. Let 1,D = 6 = (S,,. . . , a n ) and for j = 1,. . . ,n let 8' = (a;, . . . ,a',,) have 6: = 0 and 6: = 6, for all k # j . Since G is minimal imperfect, the linear form 6 ' . y achieves a maximum over PG at an extreme point y = y' that is the incidence vector of a stable set in G. Since x # y' E PG and D is nonsingular, we have that n = 6 . x > 6 . v'. and since 6 . y' is an integer, n - 1 2 6 . y' 3 6' . y'. Furthermore, because y = y' maximizes 6' . y over y € PG, we see that n
-
1 3 6' . x = 6 . x - 6'X' = n - S'X,.
Thus 6,x, 3 I , and since 6 . x = n we obtain 6,x, = 1. or x, = l / & for j = 1,. . . ,n. Also note that 6, + . . . + 6, d nw, since w ( G ) = w and each row of D is the incidence vector of a clique in G. I t now follows from Lemma 9.1 that
I,, x
=
1/al
+ . . . + 1/6,
2 n/w
> a.
Thus (2.4) implies that x = (l/w)l,. I t is easy to derive from Claim 4 another result of Padberg [15]: the fractional extreme point ( l / w ) l n is adjacent to each of a l , . . . ,a, in Pc. While the proof of Claim 9 uses minimal imperfection (the result obviously fails if G is not minimal imperfect), the following related result is true for all partitionable graphs G.
Claim 10. The subgraph of G induced by the symmetric difference of any two a -stable sets is connected.
Graphical properfies related to minimal imperfection
189
Proof. Suppose that S, and S,, i f j, are two maximum stable sets; let a, and a, be the associated incidence vectors. Let S,,u S,, = S, \ S, with S,l n S,, = 0 and let
S,, u S,, = S, \S, with S,, n SJ2= 0, where S,, U S,, is a component of the subgraph induced by ( S , \ S , ) U (S, \ S,). Thus S,lU S,, and S,, U S,, are stable sets, and consequently S=S,,US,,U(S,nS,)
and
S=S,,uS,,U(S,nS,)
s,
are a-stable sets. Let 6 and d be the incidence vectors of and respectively. Now, 6 + Ci = u, + a,, so these four vectors are linearly dependent. Hence, by the linear independence of the y1 incidence vectors of a-stable sets, it must be that either a, = 6 and a, = Ci or a, = 6 and a, = 6. Therefore the subgraph induced by (S, \S,) U (S, \ S , ) is connected. Chvatal [4]has shown that if B is a polytope that is the convex hull of the incidence vectors of the stable sets in a graph G, then two extreme points of B are adjacent if and only if the symmetric difference of the corresponding stable sets induces a connected subgraph of G. Thus, Claim 10 together with Chvatal's result implies that each pair of extreme points of BG corresponding to a -stable sets is adjacent in BG.Claim 9 can then be used to prove that the extreme points corresponding to the n incidence vectors of a-stable sets in any minimal imperfect graph G and the unique fractional extreme point ( l / w ) l n arc mutually adjacent extreme points of PG;i.e., they form an n-simplex in the skeleton of PG.
3. Examples
In light of the stringent symmetries embodied in (1.3), one might hope to prove Conjecture 2 by demonstrating that any imperfect graph with property (1.3) has a spanning subgraph isomorphic to C:l,' where a ( G )= a, w ( G ) = w . We will now describe two counterexamples, partitionable graphs that have no spanning subgraphs of the form CZ;,. Consider the graph G I of Fig. I , having n ( G , )= 10 and a ( G , )= w ( G , )= 3 . Table 1 describes for each u E V(GI) partitions of V ( G I ) \u into 3-cliques and 3-stable sets. Thus G I is a partitionable graph and, therefore, satisfies (1.3). Yet GI has no spanning subgraph isomorphic to C;", as is evident from the fact that there is only one 3-clique Q' in G I ,that has I Q n Q' 1 = 2 , where Q = {1,2, A } ,a 3-clique of GI. A 5-hole is induced by the subset { I , 3,5,7,9} of V(G,).
R.G. Blander al.
1
6 Fig. I . The partitionable graph G , .
Table 1. Partitions of G , \ u Vertex deleted
1
7
3 4
S 6 7 8 9 A
Clique partition 234, 678, 59A 345. 89A, 167 12A, 456, 789 123. 678, 59A 234, 89A. 167 12A,345, 789 123. 456, 89A 234, 167, 59A 12A,345. 678 123. 456. 789
Stable set partition 369, 41AA,258 369, 47A, 158 47A, 158, 269 158, 269, 37A 269, 37A. 148 37A, 149. 258 149, 36A. 258 149, 36A, 257 36A, 257, 148 369, 257, 148
Consider the graph G2 of Fig. 2, which has n(G2)= 13, w ( G 2 )= 3, and (Y (G2)= 4. Table 2 lists partitions for each u E V(G2)of V(G2)\u into 3-cliques and into 4-stable sets. The 3-clique 0 = {1,5,9} in G2 has I 0 r l Q ' I # 2 for all 3-cliques Q' in G2, so the partitionable graph G2 has no spanning subgraph isomorphic to C%. G2 contains a 5-hole induced by {1,2,5,6,8} and a 7-hole induced by {2,3,4,7,9, B, D}. Note that the pair {3,8} is in neither a maximum clique nor a maximum stable set in G I .The pairs (6, A } and {S, C} have the same open status in G2.Since (1.2) depends only on the cliques and anticliques of maximum size, any such open pair can be added as an edge without disturbing property (1.2). Among the partitionable graphs of the form C:,only odd holes and odd antiholes have no such open pairs.
Graphical properties related to minimal imperfection
191
1
Fig. 2. The partitionable graph G,.
Table 2. Partitions of G,\ u Vertex deleted 1
2 3 4
5 6 7 8 9 A B C D
Clique partition
Stable set partition
23A, 456, 789, BCD lCD,345, 678, 9AB 12D, 456, 789, ABC 159, 23A, 678, BCD lCD,23X, 467, 9AB 12D, 345, 789, ABC lCD,238. 456, 9AB 159, 23A,467, BCD 12D, 345, 678, ABC 159, 238, 467, BCD ICD, 23A, 456, 789 12D, 345, 678, 9AB 159, 238, 467, ABC
257B, 369C, 48AD 148B. 369C, 57AD 148B, 26YC, S7AD 137B. 26YC, SXAD 137B, 269C, 48AD 137B, 249C, 58AD 136B, 249C. S8AD 136B, 249C, 57AD 136B. 2S7C, 48AD 148B, 257C. 369D 148A, 257C, 369D 148A, 257B, 36YD 148A, 257B, 369C
Much of our work here, including the examples G I and Gz, was first announced in [lo]. The authors have recently learned of the related work in [6], where explicit procedures for constructing partitionable graphs are given. G I and Gz are among the partitionable graphs discussed in [6].
R.G. Bland et al.
192
References [ I ] C. Bcrgr. Firbung von Graphen, deren samtliche hzw. deren ungerade Krcise \tarr sind. Wiss. Z. Martin-Luther-Univ. Halle-Wittenberg Math.-Natur. Reihe I 14 (1961). 121 C. Berge. Graphs and Hypergraphs (North-Holland, Amsterdam. 1973). 131 C. Berge. The history of the perfect graph conjecture (to appear). 141 V. Chvital. O n certain polytopes associated with graphs. J . Comb. Theory, Scr. B 18 (1975) 138- 154. 151 V. Chvatal. O n the strong perfect graph conjecture. J. Comb. Theory, Ser B 20 (1Y76) 139-141. 161 V. Chviral, R.L. Graham. A.F. Perold and S. Whitcsidrs, Combinatorial designs relatcd t o the strong perfect graph conjccture. Centre d e Recherches Mathematiques. Univcrsiti. de Montreal. Publication 278. 171 D.R. Fulkerson. Blocking and anti-blocking pairs o f polyhedra. Math. Program. 1 (1971) 168- 194. [XI D.R. Fulkerson. Anti-blocking polyhedra. J. Comb. Theory I2 (1972) S(k71. 191 D.R. Fulkcrson. On the perfect graph theorem. in: T.C. H u and S.M. Robinson. cds.. Mathematical Programming (Academic Press, New York, 1973) 69-76. 101 H.-C. Huang. investigations on combinatorial optimization, Ph.D. Thesis. School of Organization and Management, Yale University (School of O R / I E , Cornell University, Technical Report No. 308 (1976)). [ 111 1.. Lovrisz. Normal hypergraphs and the perfect graph conjecture. Discrete Math. 2 (1972) 253-267 (this volume, pp. 29-42). [ 121 L. Lovasz. A characterization of perfect graphs, J . Comb. Theory. Scr. B 13 (1972) 95-98. [ 131 H. Meyniel. O n thc perfect graph conjecture. Discrete Math. I6 (1976) 339-342. [ 141 M. Padberg. Pcrfcct zero-one matrices, Math. Program. 6 (1974) IX(k196. 1 IS] M. Padberg. Almost integral polyhedra rclated to certain combinatorial optimization problcms. Linear Algebra & Appl. IS (1976) 69-88. [ 161 K.R. Parlhasarathay and G. Ravindra, The strong perfect graph conjecture is true for K,,,-free graphs. J. Comb. Theory, Ser. B 21 (1976) 212-223. [ 171 K.R. Parthasarathay and G. Ravindra, The validity of the strong perfect-graph conjccture for (K4- *)-free graphs, J. Comb. Theory, Ser. B 26 (1079) 98-100. [ 181 H. Sachs. O n the Berge conjecture concerningperfect graphs. in: Combinatorial Structures and Thcir Applications (Proceedings of the Calgary International Conference, 1969) (Gordon and Breach, New York, 1970) 377-384. 1191 L.E. Trotter, Jr.. Line perfect graphs. Math. Program. I 2 (1977) 255-259. [ZO] L.E. Trotter. Jr.. A class of facet producing graphs for vertex packing polyhedra. Discrete Math. 12 (1975) 373-388. [21] A.C. Tucker. The strong perfect graph conjecture for planar graphs, Canad. J. Math. 25 (1973) 103-114.
(221 A.C. Tucker, Coloring a family of circular arcs, S l A M J. Appl. Math. 29 (1975) 493-500. [23] A.C. Tucker. Critical perfect graphs and perfect 3-chromatic graphs. J. Comb. Theory, Ser. B 23 (1077).
Annals of Discrete Mathematics 21 (1984) 193-195 0 Elwvicr Science Publishers B.V.
AN EQUIVALENT VERSION OF THE STRONG PERFECT GRAPH CONJECTURE
v. CHVATAL Sciiool
01
Computer Scieiice. McGill Uiiiuenity, Monrrenl.
C~iiiiiih
When G is a graph, a ( G )denotes the largest size of a stable (independent) set of vertices in G and w ( G ) denotes the largest size of a clique in G. A graph is called perfect if each of its induced subgraphs H is w(H)-colorable. This notion comes from Claude Berge [I], [ 2 ] , who conjectured that every minimal imperfect graph is isomorphic to a cycle whose length is odd and at least five or to the complement of such a cycle. This conjecture is known as the Strong Perfect Graph Conjecture. An equivalent version of the Strong Perfect Graph Conjecture was proposed in [ I ] ; the purpose of this note is to present the argument from [ 11 in a slightly different form. We shall denote by C: the graph with vertices u I , u2,. . . , u,, such that u, is adjacent to u, if and only if 1 i - j I d k, with subscript arithmetic modulo n. Theorem. Let G be a graph which contains a spanning subgraph isomorphic to Czi:l with a = a ( G ) and w = w ( G ) . If a 3 3 and w a 3 then G is not minirnal imperfect.
Proof. Writing n = a w + 1, we may enumerate the vertices of G = (V, E ) as uI, u?, . . . , u,, in such a way that (with subscript arithmetic modulo n ) u,u,+, E E
for all i and for all j
=
1,2,. . . , w - 1.
=
1,2,. . . , a
(1)
Now we may assume that
U , ~ , + ,E< ~ Efor all i and for all j
-
1;
otherwise the subgraph induced by u,, v , + ~. ., . , u,+,, is not w-colorable, and so G is not minimal imperfect. Next, writing w, = u,, for all i = 1 , 2 , . . . , n, we record ( 2 ) as
w , w , + , E E for all i and for all j 193
= 1,2,..
., a
-
1.
(3)
V.Churital
194
From (3), we conclude easily that every clique of size w consists of vertices I)u for some i. To put if differently,
w , , w , ~ .~. ,, w,,,,, .
every clique of size w consists of consecutive u,'s.
(4)
Similarly, (1) implies that every stable set of size a consists of consecutive wi's. If P stands for the set of vertices u,, u2, uO+1, uw+3 and 2,3,. . . ,a - 1 then (4)implies easily that
(5) uku+2
every clique of size w includes a vertex from F.
with k
=
(6)
Similarly, if Q stands for the set of vertices W ~ + I - ~ ~ w, ~ -W ~~ + Z~ - ~. w~ ~, ~ + I and -~ wLu with k = 0, 1,. . . , w - 3 then ( 5 ) implies easily that every stable set of size a includes a vertex from Q. Note that uz = Since a 3 3 and w
(7)
uu+3 = W n + l - 3 o , uzw+2= w ~ + and ~ - u,~ =~wo. five distinct vertices belong to both P and Q. Hence
LL,+~ = w.+~-,,
3 3, these
J Pu 01s a + w
-
1.
(8)
The rest is easy: it sutiices to consider the subgraph H of G induced by all the vertices outside F and 0. By (6), (7) and (S), we have w(H)sw-l,
a ( H ) s ( ~ - l and
JHl>(a-l)(u-l),
respectively. It follows that H is imperfect, and so G is not minimal imperfect. 0
Corollary. The Strong Perfect Graph Conject~treis true if and only if euery minimal imperfect graph G contains a spanning subgraph isomorphic to Czi;, wirh a = a ( G ) and w = w ( G ) . Proof. To establish the 'if' part, consider a minimal imperfect graph G. By the assumption, G contains a spanning subgraph isomorphic to Czil, with LY = a ( G ) and w = w ( G ) ; by the above theorem, a s 2 or w 5 2 . It is an easy exercise to show that G is isomorphic to Czo+l in case w S 2 and that G is isomorphic to the complement of C2w+l in case a s 2 . To establish the 'only if' part, it suffices to observe that the cycle of length 2 k + 1 with k 2 2 is isomorphic to C:ull. with a = k, w = 2 and that the complement of this cycle is isomorphic to Czi;, with a = 2, w = k. 0
An equivalent version of the Srrong Perfect Graph Conjecture
195
References [l] C. Berge, Farbung von Graphen, deren samtliche bzw. deren ungerade Kreise starr sind, Wiss. 2. Martin-Luther Univ. Halle-Wittenberg 114 (1961). [2] C. Berge, Sur une conjecture relative au probltme des codes optimaux, Cornrn. 136me assemblee genCrale de I’URSI, Tokyo (1962). [3] V. Chvatal, On the Strong Perfect Graph Conjecture, J. Comb. Theory 20 (1976) 139-141.
This Page Intentionally Left Blank
Annals of Discrete Mathematics 21 (1984) 197-206 @ Elsevier Science Publishers B.V.
COMBINATORIAL DESIGNS RELATED TO THE PERFECT GRAPH CONJECTURE*
v. CHVATAL School of Computer Science, McGill University, Montreal, Canada
R.L. GRAHAM Bell Laboratories, Murray Hill, New Jersey, U.S.A.
A.F. PEROLD Harvard Business School, Boston, Massachusetts, USA
S.H. WHITESIDES School of Computer Science, McGill Uniuersity, Montreal, Canada
Introduction
Our graphs are ‘Michigan’ except that they have vertices and edges rather than points and lines. If G is a graph then n = n ( G )denotes the number of its vertices, a = a ( G ) denotes the size of its largest stable (independent) set of vertices and w = w ( G ) denotes the size of its largest clique. The graphs that we are interested in have the following three properties: (i) n = a w + l , (ii) every vertex is in precisely a stable sets of size a and in precisely w cliques of size w , (iii) the n stable sets of size a may be enumerated as S , , Sz,. . . , S, and the n cliques of size w may be enumerated as C1,Cz,. . . , C, in such a way that S, n C, = 0 for all i but S, n C,#B whenever i # j . We shall call them (a,w)-graphs. This concept, contrived as it may seem at first, arises quite naturally in the investigations of imperfect graphs; we are about to explain how. In the early 1960’s, Claude Berge [I], [ 2 ] introduced the concept of a perfect graph : a graph is called perfect if and only if, for all of its induced subgraphs H, the chromatic number of H equals w ( H ) . Berge formulated two conjectures concerning these graphs: * Reprinted from Discrete Math. 26 (1979) 83-92. 197
(I) a graph is perfect if and only if its complement is perfect; (11) a graph is perfect if and only if it contains no induced subgraph
isomorphic either to a cycle whose length is odd and at least five or to the complement of such a cycle. The concept of a perfect graph turned out to be one of the most stimulating and fruitful concepts in modern graph theory. The weaker conjecture (I), proved in 1971 by Lovasz [lo], became known as the Perfect Graph Theorem. The stronger conjecture (II), still unsettled, is known as the Perfect Graph Conjecture. A graph is called minimal imperfect if it is not perfect itself but all of its proper induced subgraphs are perfect. Clearly, every cycle whose length is odd and at least five is minimal imperfect, and so is its complement. The Perfect Graph Conjecture asserts that there are no other minimal imperfect graphs. The first step towards a characterization of minimal imperfect graphs was made again by Lovasz [ I l l : every minimal imperfect graph satisfies n = a w + 1. It follows from this that, in a minimal imperfect graph G, for every vertex u E G, the vertex set of G - u can be partitioned into a cliques of size w, and into w stable sets of size a. Further refinements along this line are due to Padberg [12]: every minimal imperfect graph is an (a,w)-graph. (Bland et al. [3] strengthened Padberg's result by proving that every graph satisfying (1) in an (a,w)-graph.) Hence characterizing (a,w )-graphs might help in characterizing minimal imperfect graphs. It is easy to construct (a,w)-graphs for every choice of a and w such that a 2 2 and o 3 2: begin with vertices uI, u 2 , .. . , and join 0, and u, by an edge if and only if I i - j I s w - 1, with subscript arithmetic modulo aw + I . The resulting graph, denoted by CZil, is an (a,w)-graph. If w = 2 then CEiI is simply the odd cycle Czo+,; if (Y = 2 then C2i:l is the complement of the odd cycle C,,,,. If a 2 3 and w 3 3 then C Z : , contains several pairs of nonadjacent vertices u, w such that joining u to w by an edge destroys no stable set of size a and creates no new clique of size w. Hence the graph obtained by joining u to w is again an (a,w)-graph. However, calling this graph new smacks of cheating: the structure of the largest stable sets and of the largest cliques has remained unchanged. To avoid such quibbling, we shall consider normalized ( a ,w)-graphs in which every edge belongs to some clique of size w. (As we shall see in a moment, every (a,w)-graph contains a unique normalized (a,[.')-graph.) The purpose of this note is to present two different methods for constructing normalized (a, w)-graphs other than C ~ , The ~ ~smallest . of these graphs is the (3,3)-graph shown in Fig. 1. (This graph and the (4,3)-graph of Fig. 4 were
Combinatorialdesigns related to the Perfect Graph Con]ecture
199
Fig. 1.
independently presented in [3] as examples of (a,o)-graphs different from CZll; see also [9].) The problem of characterizing all the normalized (a,w)-graphs can be given at least two additional interpretations. First, with each (a,o)-graph we may associate two zero-one matrices X , Y of dimensions n x n such that the rows of X are the incidence vectors of the stable sets S , , Sz, . . . ,S, and the columns of Y are the incidence vectors of the cliques C1,C,, . . . ,C.. If I denotes the n X n identity matrix and if J denotes the n X n matrix filled with ones then clearly
JX = XJ =aJ, JY = YJ = wJ, X Y
=J
- I.
(2)
In the terminology of Bridges and Ryser [4],the matrices X and Y form an ‘(n,O,l)-system on a, o’.Conversely, with each pair of zero-one matrices X, Y satisfying (2), we may associate a graph G with vertices u l , uz, . . . , u, such that u, is adjacent to u, if and only if Yri= Ysi= 1 for some j . Let us show that G is a normalized (a,o)-graph. To begin with, each column of Y generates a clique of size w in G and, since X Y is a zero-one matrix, each row of X generates a stable set of size a in G. To show that G has no other cliques of size w, consider an arbitrary clique of size w and denote its incidence vector by d. Clearly, Xd is a zero-one vector. In fact, since J ( X d ) = (JX)d = aJd, the vector Xd has a w = n - 1 ones and one zero. Hence Xd is one of the columns of J - I = X Y . Finally, since X is nonsingular, d must be a column of Y.A similar argument shows that every stable set of size a in G arises from some row of X . Hence G is an (a,w)-graph; since each edge of G belongs to some clique of size w, G is also normalized.
V. Chvcital er al.
200
The matrix interpretation makes it easier t o clarify the role of normalized (a,w)-graphs. Consider an arbitrary (a,w)-graph G and delete all those edges which belong to no clique of size o.To show that the resulting graph H is an (a,w)-graph, it will suffice to show that every stable set of size a in H was also stable in G. Beginning with G, define X and Y as above; in addition, let d denote the incidence vector of an arbitrary stable set of size a in H. Since the cliques of size o are the same in G and H, the vector d Y is zero-one. Since (dY)J = d ( YJ) = wdJ, the vector dY is one of the rows of X Y . Since Y is nonsingular, d is one of the rows of X , which is the desired conclusion. Hence H is the unique normalized (a,w)-graph contained in G. I n the next section, we shall make use of the fact that the equations (2) imply
YX
= X - ' X Y X = X-'(J
- I ) X = X-IJX
- I = J - I.
(The above observations are due to Padberg [12].) Before proceeding, let us point out a simple fact which may be useful in constructing (a,w>graphs. For the moment, we shall refer to each pair of matrices ( X , Y ) satisfying (1) as a solution. Now, let r and s be positive integers such that r + s = n. Let A, A * be n X r matrices, let B, B * be n X s matrices, let C, C* be r x n matrices and let D , D * be s x n matrices. Finally, let us write
XI = ( A ,B * ) , Xz= ( A *, B ) , X 3 = (A,B ) , X4 = (A *, B * ) and
We claim the following: if ( X I ,Y l ) ,( X 2 ,Yz),(X,, Y 3 )are solutions then ( X + Y4) is a solution. The proof is straightforward: since
+ B * C* = J - I, A * C* + B D = J - I,
XI Y l = A C
Xz Yz
X,Y,=AC+BD=J-I, we have AC
= A * C * , BD = B*D*
X4Y4 = A * C * + B * D * = J
and so
- I.
Similarly, the equations JX4= X4J = (YJ and JY, = Y,J routinely. It may be also interesting to note that:
= wJ
follow quite
if ( Y , ,X I ) ,( Y2,X z ) ,( Y3,X,)are solutions then ( Y 4 ,X4)is a solution.
Combinatorial designs related to the Perfect Graph Conjecture
20 1
The point is that the equations
imply
X k y k
=J
- I for each k = 1,2,3. Now X4Y4= J
-I
as above, and so
Y4X4= J - I.
An alternative interpretation of ( a ,w )-graphs concerns a packing problem. With a slight abuse of the standard notation, let K , denote the directed graph on n vertices such that, for every ordered pair of vertices u and w, there is a (unique) directed edge from u to w . Similarly, let Ka,,denote the complete bipartite graph in which each edge is directed from the a-set. As above, let n stand for aw + 1. We claim that normalized ( a ,@)-graphs correspond to partitions of the edge-set of K,, into n disjoint copies of Ka.u.With every such partition, one may associate n X n matrices X , Y such that the j-th column of X is the incidence vector of the a-set of the j-th copy and such that the i-th row of Y is the incidence vector of the o-set of the i-th copy. It is not difficdt to verify that these matrices satisfy (2), Conversely, with every pair of zero-one matrices satisfying (2). one may associate a partition of Kn into n disjoint copies of Km,w by making the directed edge uiui belong to the k-th copy if and only if X t k = y k j = 1. Incidentally, if the directions of the edges are ignored then these partitions become covers of the undirected K, by n copies of undirected Kn.w such that each edge is covered precisely twice. Designs of this kind have been studied by C . Huang and Rosa [6], [7], [8]. Finally, let us return to the link between the problem of characterizing (a,&)-graphs and the Perfect Graph Conjecture: it is not clear that a solution to the former would indeed help to settle the latter. In fact, Tucker [13] succeeded in proving the Perfect Graph Conjecture for all graphs G with w ( G ) = 3 without characterizing ( a ,3)-graphs. By virture of Padberg’s theorem the Perfect Graph Conjecture may be stated as follows: every ( a ,w)-graph G with a 3 3 and w 2 3 contains a smaller induced imperfect graph. We shall say that an (a,w)-graph G is of type I if it contains a set W of a + w - 1 vertices such that W n S # P, for all stable sets of size a and W f l C# 0 for all cliques of size w . Otherwise we shall say that G is of type 11. It is easy to see that every ( a ,@)-graph of type I contains a smaller induced imperfect graph (namely, the graph G - W with ( a - l ) ( w -1)+1 vertices and a(G - W ) S a - 1, w ( G - W ) S w - 1). Hence the Strong Perfect Graph Conjecture would follow if every (a,w)-graph with a 3 3 and w 3 3 were of type I. Unfortunately, this is not the case: the (4,4)-graph constructed in Section 2 of this paper is of type 11. (In [ 5 ] , it has been shown that every CZ:, with a 2 3 and w a 3 is of type 1.1
V.Chvatal er al.
202
1. The 6rst method Each graph C&'lb+lcan be seen as arising from Ctill by a simple construction which, vaguely speaking, leaves most of the graph unchanged and increases the total number of vertices by w. We are about to show that the same construction applies in a more general setting: if some set of 2 0 - 2 vertices of an (a,w)-graph G induces a subgraph resembling a piece of Cz-'then a simple local change in G creates an (a + 1, w)-graph H. More specifically, the properties required of the 2w - 2 vertices uI, u 2 , .. . ,u2w-2in G are that each Of the Sets c k = { u k + l , u k + 2 , . . . , with k = 0,1,. . .,w - 2 is a clique, and that for each k = 2 , 3 , . . . ,o - 1, either c k - 1 is one of the a cliques partitioning G - u k - I or else Ck-2is one of the a cliques partitioning G - u o + k - l . The graph H has w new vertices ul,u 2 , .. . ,u, in addition to the old aw + 1 vertices of G. The adjacencies in H are best described in terms of its cliques of size W .First of all, we delete edges which belong to the w - 1 cliques C,specified above and no others. Each c k is replaced by two cliques,
c;=
{0k+lr uk+2,.
..
9
&-I,
a2,.
..
3
ak+I},
c ~ = { u k + 2 , u k + 3 , . . . . , a ~ , v o , . . . , v w + k } .
Finally, we introduce the clique C*= {al,u 2 , .. . ,u w } .In case w = 3, the passage from G to H is schematically illustrated in Fig. 2 . Before proving that H is indeed an (a +l,w)-graph, let us consider a few examples. To begin with, take G = C: and consider four consecutive vertices in the natural cyclic order. If these four vertices are labeled as uI, u2, u3, v4 then H = C&;however, if they are labeled as ol, v3, u2, u4 then H is the graph of Fig.
H a2 Fig. 2.
a3
v3
v4
Combinatorial designs related to the Perfect Graph Conjecture
203
1. Next, let G be the graph of Fig. 1. The three choices (Ul, uz,
0 3 , u4) =
@,I,2,3),
( V l , UP, u3,u4)
= (2,0,1,9),
(01, u 2 , u 3 , u 4 )
= (3,1,2,0),
lead to the (4,3>graphs shown in Figs. 3 , 4 and 5. These three graphs together
Fig. 3.
Fig. 4.
204
V. Chva'tal et al.
with C:, and the graph shown in Fig. 6 are in fact the only normalized (4,3)-graphs. Now, let us establish that: for every vertex u E H, the vertex set of H - u can be partitioned into a + 1 cliques of size w.
(3)
First, we consider the case u E G. By (2), the vertex set of G - u can be partitioned into a cliques of size w. If one of these cliques is some c k then replace this C, by C ; and C ; ; otherwise simply add C* to the a cliques. Second, we consider the case u6Z G. Now u = ak for some k. If 1 < k < w then, by the assumption, either C = Ck-lbelongs to the partition of G - uk-I or else C = ck-* belongs to the partition of G - t ) u + k - l . In either case, replacement of C by CL--2
Fig. 5 .
Fig. 6.
Combinatorial designs related to the Perfect Graph Conjecture
205
and C[-l yields the desired partition of G - ak. Finally, if k = 1 then add C:I to the partition of G - u, ; if k = w then add C L to the partition of G - u,-,. With the help of (3), proving that H is an ( a + 1, w)-graph becomes a routine matter. Let n stand for ( a + 1)w + 1 and let Y denote the n x n zero-one matrix whose columns are the incidence vectors of the ( a - 1)w 2 cliques of size w inherited by H from G and of the 2w - 1 new cliques C*, C ; , C ; , 0 s k s w -2. By this definition and by construction of H,we have JY = YJ = wJ. By (3), there is an n X n zero-one matrix X such that YX = J - I and JX = (a 1)J. As we have seen in the preceding section, these equations imply XY = J - I. In addition,
+
+
,
XJ
=
0
I
X(YJ)=
w
(J - I)J
n-1
7 -
w
J
= (a
+ 1)J.
Since each edge of H belongs to some clique of size w, the rows of X are the incidence vectors of stable sets. As in the preceding section, H had no other stable sets of size a 1. Hence H is an ( a + 1, w)-graph.
+
2. The second method
It seems that characterizing all the (a,w)-graphs may be a rather difficult problem. At the moment, we can’t even characterize those (a,w)-graphs which have circular symmetries. For these graphs, the associated matrices X , Y assume the form
x = c z’,Y = C Z’ j€B
j€A
where Z is the permutation matrix of a cycle and
(4)
IAI=a, I B ( = w . The condition XY
=J -I
reduces to
A + B ={1,2, ..., a w }
(5)
with addition modulo n = a w + 1. The graphs CZil correspond to, say, A ={1,2 ,..., w } and B ={O,w,2w ,..., ( a -l)w}. We are going to describe a more general class of solutions A, B to (4) and (5). Consequently, we shall obtain new ( a ,w )-graphs with circular symmetries. When n - 1 = m l m 2 .* mk for some integers mi greater than one, then we can consider the sets M I ,M2,. . . ,Mk defined by
-
Z=l
M.= (0, 1,. . . ,n -21. Clearly, {1,2,. . . , k } then
NOW, if niGsmi = a for some
s
V.Chvatal et al.
2ob
A
=cM,,
B=l+ZsMi
iES
satisfy (4) and (5). For example, if a
= w =4
then n - 1 = 24 and so we consider
{0,1}+{0,2}+{0,4}+{0,8}={0,1, ..., 15). Now we might choose A = {0,1) + {0,2)
B
= {0,1,2,3],
+{0,4}+{0,8}={1,5,9,13}, but instead we shall choose =1
A = {O, 1) + (0,4} = {0,1,4,5}, B = 1 + {0,2}+{0,8} = {1,3,9,11}. The latter choice yields X=Z"+Z'+Z4+Z5, Y
=z1+23+zY+z1'.
The corresponding (4,4)-graph G has vertices o0, u l , . . . ,u16 such that ui and ui are adjacent if and only if j
-
i €{2,6,7,8,9,10,11,15}
with arithmetic modulo 17. Clearly, this graph cannot be obtained by the method of the preceding section. References [I] C. Berge, Farbung von Graphen deren samtliche bnv. ungerade Kreise starr sind (Zusammenfassung), Wiss. 2. Martin Luther Univ. Halle-Wittenberg, Math. Nat. Reihe (1961) 114. 121 C. Berge, Sur une conjecture relative au probleme des codes optimaux, Commun. 13itme AssemblCe GCn. URSI, Tokyo (1962). 131 R.G. Bland, H.-C. Huang and L.E. Trotter, Jr., Graphical properties related to minimal imperfection, Discrete Math. 27 (1979) 11-22 (this volume, pp. 181-192). [4] W.G. Bridges, Jr. and H.J. Ryser, Combinatorial designs and related systems, J. Algebra 13 (1969) 432-446. [5] V. ChvBtal, On the strong perfect graph conjecture, J. Comb. Theory, Ser. B 20 (1976) 139-141. [6] C . Huang and A. Rosa, On the existence of balanced bipartite designs, Utilitas Math. 4 (1973) 55-75. [7] C . Huang, On the existence of balanced bipartite designs 11, Discrete Math. 9 (1974) 147-159. [8] C. Huang, Resolvable balanced bipartite designs, Discrete Math. 14 (1976) 319-335. 191 H.-C. Huang, Investigations on combinatorial optimization, Cornell University, O.R. Dept., Tech.Report No. 308 (August 1976). 1101 L. Lovasz, Normal hypergraphs and the perfect graph conjecture, Discrete Math. 2 (1972) 253-267 (this volume, pp. 29-42). [I 11 L. Lovisz, A characterization of perfect graphs, J. Comb. Theory, Ser. B 13 (1972) 95-98. [12] M.W.Padberg. Perfect zero-one matrices, Math. Program. 6 (1974) 180-196. [I31 A. Tucker, Critical perfect graphs and perfect 3-chromatic graphs, J. Comb. Theory, Ser. B 23 (1977) 143-149.
Annals of Discrete Mathematics 21 (1984) 207-218 @ Elsevier Science Publishers B.V.
A CLASSIFICATION OF CERTAIN GRAPHS WITH MINIMAL IMPERFECTION PROPERTIES S.H. WHITESIDES School of Computer Science, McGill University, Montreal, Canada The family of (a, w)-graphs are of interest for several reasons. For example, any minimal counter-example to the Perfect Graph Conjecture belongs to this family. This paper accounts for all (4,3)-graphs. One of these is not obtainable by existing techniques for generating (a + 1, w)-graphs from (a, o)-graphs.
1. Introduction
A graph G is said to be perfect if for each induced subgraph G ' of G the size of the largest clique of G' is equal to the chromatic number of G'. The Perfect Graph Conjecture of Berge asserts that a graph is perfect if and only if it contains no induced subgraphs which are holes or antiholes, where a hole is a chordless cycle of odd length at least 5 , and an antihole is the complement of a hole. Lovasz [12], [13] proved a weaker conjecture of Berge: a graph is perfect if and only if its complement is perfect. The Perfect Graph Conjecture has been established for several classes of graphs, including planar graphs [16], circular arc graphs [17], KIJ-free graphs [15], [19], 3-chromatic graphs [18], and graphs with maximum degree at most 6 [71. Another way to state the Conjecture is to,say that any imperfect graph whose proper induced subgraphs are all perfect must be either a hole or an antihole. Padberg [14] showed that such a minimally imperfect, or critical, graph must be an ( a ,w)-graph, defined below. Definition. G is an (a,w)-graph if and only if (i) its vertex set has size a w + l ; (ii) its largest stable set has size a , and its largest clique has size w ; (iii) each vertex is in precisely a stable sets of size a and w cliques of size w ; (iv) each clique of size w is disjoint from precisely one stable set of a, and each stable set of size a is disjoint from precisely one clique of size w.
From now on, we denote the number aw + 1 by n. An ( a ,w)-graph is said to be normalized if each of its edges belongs to at least 207
S.H. Whitesides
208
one clique of size w. Each (a,w)-graph contains a unique normalized ( a , w ) subgraph, because removing edges which belong to no cliques of size w does not create any new stable sets of size a ((61, [ 181). The paper of Chvatal, Graham, Perold, and Whitesides [6] establishes two additional contexts in which (a,w)-graphs arise. First of all, there is a correspondence between normalized (a,w)-graphs and solutions to the system of equations
JX=XJ=aJ,
J Y = YJ=wJ, XY = J - I ,
(1)
where X and Y are matrices of 0’s and l’s, J has all entries 1, I is the identity matrix, and all these matrices are n x n. Bridges and Ryser [4]call the above matrices X and Y an ( n ,0,1) system on a, w . Second, there is a correspondence between normalized ( a ,w)-graphs and packings of the complete graph K,, by complete bipartite graphs Ka.w with each edge of K , covered exactly twice. C. Huang [9], [ 101 and C . Huang and Rosa [8] have studied such packings. The graphs denoted by CZ-’are ( a ,w)-graphs; they have vertices vo, . . . , v),,, with u, adjacent to u, whenever there is a d such that O < d < w and d = i - j or j - i (mod n ) . Holes and antiholes are of this type. In [6], methods are given for constructing (a,w)-graphs which are not of this type. The purpose of this paper is to describe all normalized (4,3)-graphs. One of these is a graph which is neither C:, nor a graph obtainable by the methods of [6]. Of course, none of these graphs is a counterexample to the Perfect Graph Conjecture. as Tucker [18] has shown the conjecture holds for graphs with w C 3. 2. Properties of ( a ,w)-graphs
We now list several well known properties of ( a ,o)-graphs which we will use frequently throughout this paper. For convenience, we will use the word clique (stable set) to refer to a clique (stable set) of maximum size only. Recall that in an (a,wkgraph, the property of being disjoint pairs off the cliques with the stable sets. We will denote the stable sets by S , , . . . ,S, and the cliques by T I , .. . ,T,, where S, n T, = 0.
Lemma 1. Zf G is an ( a ,w )-graph, then G contains exactly n = aw (of size w ) and exactly a w + 1 stable sets (of size a ) .
+ 1 cliques
Proof. This follows easily from the definition of an (a,w)-graph.
0
Lemma 2. For each vertex v in an (a,w)-graph G, the vertices of G - v are partitioned by the stable sets (cliques) which are paired with the cliques (stable sets) containing v.
Certain graphs with minimal imperfection properties
209
Proof. Let X be the matrix whose rows are the incidence vectors of the stable sets S1,.. . ,S,, and let Y be the matrix whose columns are the incidence vectors of the cliques TI,..., T,. Then we know by [6] that
JX
= XJ = aJ,
JY
=
YJ
= oJ,
and
XY
= J - I,
where J is the n x n matrix of l's, Also,
YX =X-'XYX=X-'(J-I)X =X-'JX-I=J-I. The lemma now follows from the fact that Y X = J - 1.
0
Remark 1. A consequence of Lemma 2 is that the pseudo p-critical graphs of Tucker [lS] are precisely the ( a , w)-graphs. Lemma 3. Let G be a graph, and de,finea graph M ( G ) by making the vertices of M ( G ) correspond to the cliques of G and making vertices in M ( G ) adjacent whenever the corresponding cliques intersect. If G is an (a,@)-graph, then so is M(G).
Proof. See [MI.
3. Generation of (4,3)-graphs from (3,3)-graphs We assume throughout the rest of this paper that G is a normalized (4,3)-graph whose cliques are 'triangles' T I , .. . , TI, and whose stable sets are S1,. . . , Sls,where S, n T, = P, if and only if i = j. By an i,, . . . ,ik-vertex,we mean a vertex which belongs to the stable sets S,,, . . . ,S k .By Axyr, we mean a triangle whose vertices are x, y, and z. Remark 2. We emphasize that Lemma 2 says the following: if a vertex u belongs to distinct triangles T,,, T,, and T,, then each other vertex of G is exclusively an il-vertex, an &vertex, or an &vertex. Also, it says that if distinct and Tk intersect, then there are no j,k-vertices. triangles Define a graph K ( G )from G as follows. Make the vertices of K ( G ) ,like the vertices of M ( G ) , correspond to the triangles of G. This time, however, make vertices adjacent whenever the triangles to which they correspond intersect in an edge. We first show that the maximum degree of K ( G )is at most 2. Then we use this fact to prove that K ( G ) contains a path of length at least 5 if and only if G can be generated from a (3,3)-graph by the first construction method of ChvAtal [6].
S. H.Whitesides
210
Lemma 4. The graph K ( G ) has maximum degree at most 2. Proof. If a triangle of G met three other triangles in edges, then the four triangles together would give rise to a clique of size 4 in M ( G ) . However, M ( G ) is a (4,3)-graph according to Lemma 3. 0 Lemma 5. No edge of G is in three triangles. Proof. Suppose triangle T, with vertices a, b, and c shared an edge ab with triangles T, and T,. Then since T , would intersect S, and Sk, c would be a j , k-vertex, which is ruled out b y Lemma 2. 0
Lemma 6. Suppose that H is a (3,3)-graph containing the configuration of four vertices shown in Fig. l(a) and that either {UI, 0 2 , v3} is one of three triangles partitioning H - v4 or {u2,v3,u4} is one of three triangles partitioning H - ul. Then replacing the con,figurationshown in Fig.'l(a) by the configuration shown in Fig. 1 ( b ) generates a (4,3)-graph. Proof. This is the construction of [6] specialized to the generation of (4,3)-graphs from (3,3)-graphs. 0
Lemma 7. I f K ( G ) contains a path of length at least 5, then G contains the configuration of triangles shown in Fig. 2, and (i) d is the only common neighbor of b and f ; (ii) vertices b and g can have common neighbors only if b and g are adjacent (similarly for a and f ) .
-_-._-- .-. ,.-____-------____ -... '.*,. -----r-r=<--------
-*-- --__/-
,--,,
,,
L v1
=
s v2
b v3
v4
v1
v2
v3
v4
(a) (b) Fig. 1. Configurationreplacement generating a (4,3)-graph. An edge shown dashed is to be included provided that, in H, it belongs to a triangle whose third vertex is in H - {u,, u2, u,, u4}.
Fig. 2. Configuration contained in G, where K(G)contains a path of length 5. Some adjacencies may not be shown.
Certain graphs with minimal imperfectionproperties
211
Proof. If K ( G )contains a path of length at least 5, then Lemma 5 implies that G contains the configuration in Fig. 2. (i) Suppose that T6 is the third triangle containing b and that T7 is the third triangle containing f. Then each vertex u distinct from b and f must be an i, j-vertex for some i in {l,2,6} and j in {4,5,7}. In particular, if u is a 3-vertex, then by Remark 2 it must be a 3, 6 , 7-vertex, as T3 intersects TI, Tz, T4, and Ts. Hence T6 and T7 do not intersect, again by Remark 2. Therefore, b and f have no common neighbors other than d. (ii) Suppose, by way of contradiction, that b and g are not adjacent but have a common neighbor h. The intersections of S,, ..., Ss with a , . . . , g can be determined by repeated application of Remark 2: a E S2f l Ss, b E S3, c E S4, d E SIfl Ss, e E Sz, f E S3, and g E S , f l S4. Consequently, h must be a 2,5vertex. Hence h is not adjacent to a, as a is also a 2,5-vertex. Suppose h = a. Let TG be a triangle containing edge ag. By part (i), the third vertex of TGis neither b nor f. The five vertices of G outside T Ithrough T6 must include one vertex from S3 and two vertices from each of S1, Sz, S4, Ss, and Sg. Remark 2 provides a contradiction: the 4-vertices must be I-vertices, and the 5-vertices must be 2-vertices, so the 6-vertices must coincide with the 3-vertex. Hence h # a. By part (i), the 2,5-vertex h is not adjacent to f. Therefore, since G is assumed normalized, edge bh belongs to some triangle T6with a 1,4-vertex i distinct from a , . . . ,h. Edge gh belongs to some triangle T7 with a 3,6-vertex j distinct from a , . . . , i. Vertices c and f are 6-vertices. If i were a 7-vertex, then either a or c would be a 7-vertex. Since c is a 6-vertex, it cannot be a 7-vertex, so a would be a 7-vertex. Hence d and g would also be 7-vertices. However, g being a 7-vertex would contradict the fact that g belongs to T7.Therefore, b must be the 7-vertex of Th.Then e is also a 7-vertex. Each of S,, . . . ,Sg contains exactly one of the unlabelled vertices of G. It follows that there is a 1,4-vertex, a 2,5-vertex, and a 3,6-vertex. However, two of these must be 7-vertices, which is impossible. 0
Lemma 8. If K ( G )contains a path of length at least 5 , then G can be obtained by the construction in Lemma 6. Proof. Assume G contains the configuration shown in Fig. 2. Let H be the graph obtained from G by removing c, d, and e, making b adjacent to f and g, and making a adjacent to f. Triangles TI,. . . , Ts are destroyed by the removal of c, d, and e. We need to know that a, b, f, and g do not form a clique in H. If ag were an edge of G, then by Lemma 7, af and bg would also be edges. Then a , . . . ,g would constitute a component of G, as each of these vertices is on three triangles from the set. Also, the remaining six vertices could not possibly have each vertex
S.H.Whitesides
212
in exactly three triangles. Hence, ag is not an edge of G, and vertices a, b, f, and g do not form a clique in H. It is easy to check, using Lemma 7, that Aabf and Abfg are the only triangles which appear in H but not in G. Consequently, H has clique size 3 and contains ten triangles. Also, each vertex of H belongs to three triangles. Consider how the stable sets of G intersect {a,. . . ,g } . With the exception of S : , any stable set containing (I contains d : if a is an i-vertex, b is a j-vertex, and e is a k-vertex, then d must be an i-vertex again, as long as d is not in T,. Hence, three stable sets contain both a and d. Similarly, three stable sets contain both d and g. Since vertex d belongs to S, and S s , it follows that the two remaining stable sets to which d belongs both contain a and g. Further consideration along these lines shows that G has, in addition to S , , . . . ,S5 and the two stable sets containing a, d, and g, three stable sets containing b and e and three stable sets containing E and f. This accounts for all thirteen of the stable sets S, of G. This information. along with other information we are about to produce, is represented in Fig. 3.
Stable sets of G
Stable sets of H
1 cf-- I 0 1 cf-- I Fig. 3. Stable sets of G and H,where K ( G ) has a path of length at least 5.
Certain graphs with minimal imperfection properties
213
With the exception of S3, each S, loses a vertex of T3 in the formation of H from G. The description of the stable sets of G shows that the remaining vertices of S , , i# 3, continue to be mutually non-adjacent in H, as no stable set of G contains both b and g or both a and f. We will call the set of remaining vertices of S,, i# 3, the image of S,. Since S3 is the only stable set of G containing both b and f, the joining of b to f produces in H two stable sets Sb and Sf of size 3 containing b and f, respectively. For future reference, let S3= {b,f,x,y}, and note that S6 = (6,x, y } and S, = {f,x, y}. Also, note that {b, e, x, y } is a stable set (one of the s6, . . .,S,,) whose image is Sb and that {e, f, x, y} is a stable set (one of the S6,. ..,S,,) whose image is Sf. We now check that the stable sets of H have the properties of stable sets in (3,3)-graphs. Clearly, the stability number of H is 3. The vertices of each stable set of H together with one of c, d, e form a stable set of G. By Remark 2, each vertex of G except c is an i-vertex for exactly one i in {1,2,3). Each vertex except d is a j-vertex for exactly one j in {2,3,4}. Consequently, each 1-vertex except d is also a 4-vertex. It follows that S1 and S4 have the same image. Similarly, S2 and Ss have the same image. As previously stated, {b, e, x, y } has image Sb and {c,f, x , y } has image S,. By referring to the list of stable sets of G given in Fig. 3, we can now deduce that there are n o additional agreements among the images of these stable sets. This fact is shown in Fig. 3. Hence H contains exactly ten stable sets. Each vertex of H is a 1,4-vertex, a 2,5-vertex, or a 3-vertex of G. Fig. 3 shows that each of these vertices lies in a pair of stable sets of G which have a common image in H. Hence, each vertex of H lies on at most three stable sets of H. Then, counting vertex-stable set incidences in H shows that each vertex lies in exactly three stable sets. Next, w e check that each of the ten triangles of H is disjoint from exactly one stable set of H, and vice versa. The common image of SI and S4 is the only stable set of H disjoint from Aabf, and the common image of Sz and Ss is the only stable set disjoint from Abfg. For 6 i 13, T, belongs to H, and T, is disjoint from a stable set of H if and only if that stable set is the image of the stable set S, disjoint from T, in G. Thus, each triangle of H is disjoint from a unique stable set. Since the images of s6,.. . ,SI3together with the images of S , and S 2 account for all the stable sets of H, each stable set of H is disjoint from at least one triangle of H. It follows that the property of being disjoint pairs the triangles and stable sets of H. We have now shown that H is a (3,3)-graph. By Lemma 2, H - a is partitioned by the triangles of H disjoint from the stable sets of H containing a. Since the common image of S2 and S s contains a and is disjoint from Abfg, Abfg is in this partition of H - a by cliques. It fdlows that we can generate G from H by the construction of Lemma 6 by choosing v,, v2, u3, and u4 to be a, b, f, and g, respectively. 0
214
S. H. Whitesides
4. Short chains of triangles
In this section, we show that if G is a (4,3)-graph, then K ( G ) must contain a path of length at least 4. Lemma 9. K ( G ) cannot have maximum path length 1 Proof. If K ( G ) had no paths of length greater than 1, then every vertex of G would have degree 6. The union of the vertices of any triangle and their neighbors would contain fifteen vertices, which is impossible. 0
Lemma 10. K ( G ) cannot have maximum path length 2. Proof. Suppose K ( G ) had maximum path length 2. Then we could assume G contains the configuration shown in Fig. 4. Notice that vertices a , . . . ,h are distinct. By Remark 2, vertex a, a vertex f of T3, and a vertex h of T4 are 2-vertices. Vertices e, d, and g are 1-vertices, b is a 4-vertex, and c is a 3-vertex. The remaining five vertices of G must consist of a 1-vertex i, a 2-vertex j , and three 3,4-vertices k, I, and m. Vertex a lies on T Iand two additional triangles TS and T6- The 1-vertex of T5 and T6 must be i. Vertex d lies on T2 and two additional triangles T7and T8.The 2-vertex of T, and TS must be j . Some 3,4-vertex k must lie on two of Ts, T6, T7, and T8.Since T5and T6 meet at a and i, and T7and T, meet at d and j , we may assume that k belongs to T5 and T7. Vertex i is the 7-vertex of T5.The 6-vertex of T5 must be k, and the %vertex of T7must be k. Thus T6 and TSare disjoint, by Remark 2. Thus the 3-vertex of T, is a 3,4-vertex not in TS,and the 3-vertex of T8 must be m. Note that m must then be the 7-vertex of T,. Vertex i is on Ts, T,, and a third triangle whose 3-vertex cannot be c, k, or 1, as K(G) has maximum path length 2. Hence this third triangle contains the
Fig. 4. Configuration contained in G, where K(G)has maximum path length 2. Some adjacencies may not be shown.
Certain graphs with minimal imperfection properties
215
3,4,7-vertex m. But vertices i and m cannot be adjacent as they are both 7-vertices. Thus no G exists whose K ( G ) has maximum path length 2. 0
Lemma 11. K ( G ) cannot have maximum path length 3. Proof. Suppose, by way of contradiction, that K ( G ) has maximum path length 3. Then we may assume that G contains the configuration shown in Fig. 5. It follows from Lemma 4 that vertices a , . . . , i are distinct. By Remark 2, a, e, a vertex f of T,, and a vertex i of T, are 2-vertices. Furthermore, b is a 3-vertex, d is a 1-vertex, c is a 4,5-vertex7 h is a 1,4-vertex, and g is a 3,5-vertex. The remaining vertices of G consist of two 1,4-vertices j and k, and two 3,5-vertices I and m. By Lemma 2, G - h is partitioned by triangles TI, T4, and two additional triangles T, and T7.We may assume T, = {e, j , 1 ) and T7= {i, k, m } . Since h is a 6-vertex, i cannot be a 6-vertex. Vertex m cannot be a 6-vertex because m is a 3-vertex, and T, meets T6. Thus the 6-vertex of T7 must be k. Since k is a 1,4,6-vertex, G - k is partitioned by TI, T,, T, and a fourth triangle T8 which must be {h, i, m } . The 8-vertex of Ts must be b, and consequently, e must be the 8-vertex of T,. Since g is a 5-vertex, and Txmeets Ts,f must be the 8-vertex of T,. If c were a 4,5,6-vertex, then T4, Ts, T6, and {a, k, m } would partition G - c. However, {a, k, m } cannot be a triangle because it and triangles T,, Tx, and Ts would correspond to a path of length 4 in K ( G ) . Hence the 6-vertex of T i is not c, but a, as 3-vertex b is not a possibility since T3 meets T,. Consideration of G - f shows that {a, e, g, j , k, 1 ) must be partitioned by two triangles. However, 1,4,6,8-vertex k cannot be adjacent to a 2-vertex in this set, as a is also a 6-vertex, and e is also an 8-vertex. Thus no G exists whose K ( G ) has maximum path length 3. 0
I
f
Fig. 5. Configuration contained in G, where K ( G ) has maximum path length 3. Some adjacencies may not be shown.
S.H. Whitesides
216
5. The (4,3)-graphs
The (4,3)-graphs which are obtainable from (3.3)-graphs by the method of Lemma 6 are listed in [6]. One of these graphs was discovered independently by H.-C. Huang [3], [ 111. The following theorem completes the description of all normalized (4,3)-graphs.
Theorem. The normalized (4,3)-graphs consist of the graph shown in Fig. 6 together with those graphs which can be Constructed from (3,3)-graphs by the method of Lemma 6. Proof. Suppose G is a normalized (4,3)-graph which cannot be constructed from a (3,3)-graph by t h e method of Lemma 6. Then by Lemmas 8, 9, 10, and 11, K ( G ) has maximum path length 4. We may assume the configuration shown in Fig. 7 appears in G, where a , . . . ,h are distinct vertices.
e
C
a
i
m
k
Fig. 6 . The normalized (4,3)-graph not constructible by the method of [6].
Fig. 7. Configuration contained in G, where K ( G ) has maximum path length 4. Some adjacencies may not be shown.
Certain graphs with minimal imperfection properties
217
Vertex e lies on T3, T4,and an additional triangle. Since the maximum path length of K ( G ) is 4, e has degree 5. If e were adjacent to two vertices not in { a , . . . , h } , then exactly three vertices of G would not be adjacent to either of b or e. Since every stable set which contains b also contains e with the exception of S3, there are three stable sets containing both b and e. Thus the three vertices adjacent to neither b nor e would be mutually non-adjacent and would form with b and e a stable set of size 5. Hence e is adjacent to a fourth vertex in { a , . . . ,h } . The 1,bvertex h of Ts is the only possibility. Since G is normalized, edge eh belongs to some triangle T6whose remaining vertex i is a 3,s-vertex. Consider the remaining vertices. There is a 2,6-vertex j , a 3,5-vertex k, and two 1,4-vertices 1 and m. These vertices are precisely the ones adjacent to neither b nor e, so the subgraph they induce contains exactly three non-edges but no stable set of size 3. Since each triangle containing j has a 1-vertex, j must be adjacent to at least two 1-vertices by Lemma 5. Both these vertices must in fact be 1,4-vertices, as j is not adjacent to d. Similarly, k must be adjacent to two 1,4-vertices. If neither j nor k were adjacent to h, then both would be adjacent to 1 and to m. Then the subgraph induced by { j , k, 1, m } would contain at least four edges, which is impossible. Therefore, we may assume that j is adjacent to h (by symmetry). The 6-vertex of Ts must be g. Consider the two stable sets containing g other than S2or Sb.Each one either contains c, and hence f and i, or contains d, and hence a and i. Hence { g , c , f , if and { g , d, a, i) are both stable sets. Consequently, g is not adjacent either to i or to f. Also, i and j must be adjacent, or { g , i, a, d, j } would be a stable set: j is not adjacent to d, and since j is a 2-vertex, it is not adjacent to 2-vertices a and g. Note that at this point, we know all three triangles containing h, namely T s , T6, and a triangle consisting of h, i, and j . By Remark 2, G without 1,4-vertex h is partitioned by TI, T,, and two additional triangles. The triangle containing g must have k as its 3,S-vertex and must contain a 1,4-vertex 1. The vertices i, j , and rn form the remaining triangle. At the moment, we know only two of the triangles containing g : TSand Agkl. N o additional adjacencies among b, h, 1, and k can be made, so g must be adjacent to a fifth vertex. The only possibility is m, and m must be joined to k to form the missing triangle. We have now found all three of the edges in {j , k,1, m } , namely, km, kl, and jm. Vertices j and k must each have at least one more neighbor. The only possibility is that j is adjacent to f, and k is adjacent to a. Vertex m must be the 1-vertex of the triangle containing edge j f . Vertex i is in T6,Ajmi, and Aijh. Hence l cannot be adjacent to i. Vertex 1 must then be adjacent to a and f, and a must be the common neighbor of f and 1. Hence G is the (4,3)-graph shown inFig. 6. 0
218
S.H.Whitesides
Acknowledgement The author wishes to thank V. Chvital for the representation of the graph shown in Fig. 6.
References [I] C. Berge, Farbung von Graphen deren siimtliche bzw, ungerade Kreise starr sind (Zusammenfassung), Wiss. Z. Martin Luther Univ. Halle Wittenberg, Math. Nat. Reihe (1961) 114. [2] C. Berge, Sur une conjecture relative au probleme des codes optimaux, Commun. 13ibme Assemblte Gn. URSl (Tokyo, 1962). 131 R.G. Bland, H.-C. Huang and L.E. Trotter, Jr., Graphical properties related to minimal imperfection, Discrete Math. 27 (1979) 11-22 (this volume, pp. 181-192). [4] W.G. Bridges, Jr. and H.J. Ryser, Combinatorial designs and related systems, J. Algebra 13 (1969) 432-446. [5] V. Chvatal, On the strong perfect graph conjecture, J. Comb. Theory, Ser. B 20 (1976) 139-141. [6] V. Chvatal, R.L. Graham, A.F. Peroid and S.H. Whitesides, Combinatorial designs related to the strong perfect graph conjecture, Discrete Math. 26 (1979) 83-92 (this volume, pp. 197-206). [7] C. Grinstead, The strong perfect graph conjecture for graphs with maximum degree six (to appear). [S] C. Huang and A. Rosa, On the existence of balanced bipartite designs, Utilitas Math. 4 (1973) 55-75. [9] C. Huang, On the existence of balanced bipartite designs 11, Discrete Math. 9 (1974) 147-159. (lo] C. Huang, Resolvable balanced bipartite designs, Discrete Math. 14 (1976) 319-335. [ 1I] H.-C. Huang, Investigations on combinatorial optimization, Cornell University O.R. Dept. Tech. Report #308 (August 1976). 112) L. LOV~SZ, Normal hypergraphs and the perfect graph conjecture, Discrete Math. 2 (1972) 253-267 (this volume, pp. 29-42). [13] L. Lovasz, A characterization of perfect graphs, J. Comb. Theory, Ser. B 13 (1972) 95-98. [14] M.W. Padberg, Perfect zero-one matrices, Math. Program. 6 (1974) 180-196. [15] K. Parthasarathy and G. Ravindra, The strong perfect graph conjecture is true for K,,,-free graphs, J. Comb. Theory, Ser. B 21 (1976) 213-223. [16] A. Tucker, The strong perfect graph conjecture for planar graphs, Canad. J. Math. 25 (1973) 103-114. [17] A. Tucker, Coloring a family of circular arcs, SIAM J. Appl. Math. 29 (1975) 493-502. [18] A. Tucker, Critical perfect graphs and perfect 3-chromatic graphs, J. Comb. Theory, Ser. B 23 (1977) 143-149. [19] A. Tucker, Berge’s strong perfect graph conjecture, Ann. N.Y. Acad. Sci. (1979) 530-535.
PART V
WHICH GRAPHS ARE PERFECT
This Page Intentionally Left Blank
Annals of Discrete Mathematics 21 (1984) 221-224 @ Elsevier Science Publishers B.V.
A COMPOSITION FOR PERFECT GRAPHS Robert E. BIXBY* Department of Mathematical Sciences, Rice University, Houston, TX 77251-1892, USA
We describe a method, introduced by Cunningham and Edmonds, for composing two graphs, and show that this composition preserves perfection. The proof is based on a method for converting efficient covering afgorithms on the component graphs into an efficient covering algorithm for the composed graph.
1. Introduction Let H and K be disjoint (simple) graphs with distinguished vertices u and U, respectively. (For an introduction to our graph-theoretic notation, see [ 11.) Define a graph H o K, called the composition of H and K, by
V ( H 0 K ) = ( V ( H )u V(K))\{U,81,
E ( H 0 K )= E ( H - U ) U E ( K - U ) U { X Y : X U E E ( H ) ,V Y E E ( K ) } . Thus, H o K is constructed from H and K by deleting u and U, and forming a complete bipartite subgraph on the neighbor sets of u and u. (This composition is one instance of a more general construction studied in great detail in [3].) Our main result for H K is the following: 0
Composition Theorem. If H and K are perfect, then H 0 K is perfect.
For an interesting polyhedral generalization of the above theorem see [ 2 ] . Let G be a graph, and let a ( G ) denote the maximum cardinality of an independent set of vertices in G. If G has a distinguished vertex x , let G‘ denote G - x , and let G” denote G with x and all its neighbors deleted. Clearly,
+
a ( H o K ) = max{a (H’) a(K’’),a ( H ” )+ a ( K ’ ) } . Our proof of the Composition Theorem is based on constructing a size a ( H K ) clique cover for H K . We begin, in the next section, with a treatment of vertex 0
0
* Research partially supported by Air Force Office of Scientific Research Grant AFOSR-82-0004, at Northwestern University. 221
R.E. Bixby
222
replication. This material is not strictly necessary for the proof of the Composition Theorem (in particular, vertex replication is well-known to preserve perfection [4]); rather, we present it for algorithmic reasons. In Section 3, the Composition Theorem is proved. In Section 4, we examine weighted covering of HoK.
2. Vertex replication
For a graph H with a distinguished vertex u, and for an integer k 3 1, H k denotes the graph obtained by replicating u, k - 1 times. Thus, H k has k copies of u (denoted, below, as ui, . . . ,u k ) , no two of which are joined but each with the same neighbor set as u. Note that Hi = H. Note, also, that the construction of H k may be viewed as a special case of the composition H o K - letting K be the star graph with k + 1 vertices and k edges, and such that u has degree k , yields H k= H o K . Assuming H is perfect and that an efficient algorithm is available for constructing minimum clique covers in H (and all its induced subgraphs), we show how to efficiently construct a clique cover for any replicate H k . In particular, we show how to construct a cover for Hk+'from one for H k ,k 3 1. Let X , , . . . , X , be a minimum clique cover for H k , and assume u, EX, ( I s i s k ) . Let
G
= H - ((Xi \ { Ui}) U *
'
. U (xk \ { UI, 1)).
Let Y 1 , ... , Y, be a minimum clique cover for G. There are, then, two cases to consider: Case 1. s ~ - k. n Then, Y1,..., Y,, X , ,...,X k , ( X l \ { u l } ) u { ~ ~is+ la} minimum cover of H k " .(Thus, s = n - k.) Case 2 . s 2 n - k + 1. Then, {u~,~}, X I , .. . ,X.is a minimum cover of Hktl, . . ,X , is a minimum cover of G. (Thus, s = n - k + 1.) since { u } , Remarks. (a) Case 2 requires some justification. Since H is perfect and G is an induced subgraph of H, G is perfect. Hence, the existence of a minimum cover { u } ,& + I , . . . ,X , for G implies the existence of an independent set I of G, such that u E I and I I I = n - k + 1. It follows that
I'
=( I \{u})
u {Ul, - .. , U k + l l
is independent in H k + ' ;but 11'1 = n + 1, and so {&+I}, X i , .. . ,X , is a minimum cover. (b) For a particular r 2 2, it will not generally be necessary to carry out the above computation for all H k (2 S k S r) in order to obtain a cover for H', even
A composition forperfeci graphs
223
if r < deg(u). In particular, if Case 2 occurs for some k, then X1,.. .,A,, {uk+l},.. . , { u s } is evidently a minimum cover for H' ( s 2 k + 1). (c) Note that the graph G changes in a simple way from one value of k to the next. The only new Xi \{ui}is the one for i = 1. Hence, to compute G we need simply find the Y, (in Case 1) that contains ul, and then delete Y , \{ul} from the old G.
3. Clique covers for H o K
Assume H and K are perfect, k = (Y ( H ' )- (Y (H") 3 a ( K ' )- a(K"), and k > 0. (If k = 0, a minimum cover is trivial to construct.) Let X 1 , .. . ,X,, be a minimum clique cover for H k ,where u, EX,(1 i s k ) , and let Y I ., . . , Y,,,be a minimum clique cover for K k , where u, E Y , ( 1 s i s k ) . Let Z, = ( X , \ { u , } ) U ( Y , \ { u , } ) for l s i s k , and let {zk+l,
- .., Zm+n-k}= {xk+l,. . ., x n ,
yk+I,.
.. Y m } . 7
Now, n = (Y ( H ' ) and rn = (Y ( K " )+ k, by perfection and the choice of k. Hence, rn n - k = (Y ( H ' ) a (K"),and therefore Z1,.. . ,Z,,,+"-kprovides a minimum covering for H 0 K. This proves the Composition Theorem.
+
+
4. Weighted covers
Assume H and K are perfect, and let w be a real-valued weight vector defined on V ( H 0 K ) . We sketch a method for efficiently constructing a minimum-weight weighted clique cover for H 0 K, given efficient algorithms for the same problem on H and K, respectively. The procedure is virtually identical to that described in Section 3, but is more difficult to write down. We begin with some notation. Let G be a graph, and let c be a real-valued vector defined on the vertices of G. Let '23 be a family of cliques in G, and let 8 be a real-valued vector indexed on the members of '23. For any vertex x of G, let Bx= { B E '23 : x E B } . Finally, for B ' c '23, we write O('23') = (0, : B E B'). Then, the pair (O,'23) is called a c-couer of G if
c
6(Bt)3 C,
( i E V(G)).
We also use the notation c, : I is independent in G
R.E. Bixby
224
Now, assume fi = a (w, H')- a (w, H") P a ( w , K') - a (w, K").Let w. = w, = p, and let (T,B) and (p,'3) be w-covers of H and K that minimize T ( & ) and p('3), respectively. We construct a w-cover (&a)for H o K that minimizes e(B).The construction begins with the application of the following procedure: Step 1. If 7rc = 0 for each C E B,, STOP. Otherwise, let C E &,, and D E Bu be such that TC, pD >O. Step 2. Set B +(C\{u})U(D\{o}), B' c B' U { B } (initially 8b+min{.rrc,pD}, 7rc + m - 0; and p D+ p D - 0;. Go to Step 1. Now, define B = 8'U @\&) U (6\ 6"), and for B E B define
eB =
{
e:, T,,
p8
B' = 0),
if B EB', if BE&\&., if B E D \ % , ;
(&B)can then be proved to be a w-cover of H o K that minimizes
O@).
References [I] J.A. Bondy and U.S.R.Murty, Graph Theory with Applications (North-Holland, Amsterdam, 1976). [2] W.H. Cunningham, Polyhedra for composed independence systems, in: A. Bachem, M. Grotschel and B. Korte, eds., Bonn Workshop on Combinatorial Optimization, Annals of Discrete Math. 16 (North-Holland, Amsterdam, 1982) 57-67. [3] W.H. Cunningham and J. Edmonds, A combinatorial decomposition theory, Canad. J. Math. 22 (1980) 734-765. [4] L. Lovisz, Normal hypergraphs and the weak perfect graph conjecture, Discrete Math. 2 (1972) 252-267 (this volume, pp. 2942).
Annals of Discrete Mathematics 21 (1984) 225-252 0 Elsevier Science Publishers B.V.
POLYNOMIAL ALGORITHM TO RECOGNIZE A MEYNIEL GRAPH M. BURLET U.S.M.G.-IMAG-BP53 X-38041 Grenoble Cedex, France
J. FONLUPT IMAG-BP 53 X-38041 Grenoble Cedex, France We show that the recognition problem of a Meyniel graph is polynomial by presenting a polynomial recognition algorithm. The Meyniel graphs are the graphs in which every odd elementary cycle ( 3 5) has at least two chords. These graphs are known to be perfect. As a by-product we obtain new results concerning some classes contained in Meyniel graphs: triangulated graphs, o-triangulated graphs and parity graphs.
1. Introduction
For a graph G = (V, E) we denote by a (G) ( w ( G ) , respectively) the maximum cardinality of a stable set of G (of a clique of G, respectively) and by B(G)( y ( G ) ,respectively) the minimum cardinality of a covering of G by cliques (by stable sets, respectively). Perfect graphs, introduced by Berge [l], are those graphs G for which a(G’)=d(G’) for all induced subgraphs G‘ of G. As a result of the perfect graph theorem (Lovhsz [lo]) perfect graphs may also be characterized by the following relation: w ( G ’ ) = y ( G ’ ) for all induced subgraphs G’ of G. An interesting aspect of perfect graphs is their algorithmic study, i.e., to solve the following problems: maximum weighted stable set of G, minimum weighted covering of vertices of G by cliques, maximum weighted clique of G, minimum weighted covering of vertices of G by stable sets. All this means that we want to find algorithms of polynomial order in I V [to solve these four problems. The possibility of solving these four optimisation problems for perfect graphs is a consequence of the definition of perfect graphs: the relation a(G’)= e(G’) for any induced subgraph G’ of G is nothing else than a ‘strong duality relation’. The general papers of Grotschel, Lovasz and Schrijver [7,8] give polynomial algorithms for the four above-mentioned problems. Their algorithms, which are based on the ellipsoid method, unfortunately give no idea of the structure of perfect graphs, and, at the present time, appear to have no great combinatorial interest. Furthermore, these algorithms may not work if G is not perfect. This 225
226
M. Burlet, J. Fonlupt
brings us to another equally important basic question - the recognition problem, i.e., to recognize in polynomial time whether a graph G is perfect. This problem is very difficult (its solution would likely give a positive or negative answer to the strong perfect graph conjecture) but has been solved for some important classes of perfect graphs: bipartite graphs and their line graphs, triangulated graphs, comparability graphs, parity graphs, and their complements. The recognition problem has also been solved for a number of subclasses of these latter graphs (see Golumbic [6]). In this paper we want to give a recognition algorithm which runs in polynomial time in I V 1, for an important class of perfect graphs: Meyniel graphs. This will imply that we have a good characterization, in the sense of J. Edmonds, of Meyniel graphs. Definition 1. A Meyniel graph is a graph in which every odd elementary cycle of length at least five contains at least two chords. Meyniel [ 1I] has shown that these graphs are perfect. Later, Ravindra [I21 proved that they are strongly perfect in the sense of Berge and Duchet. Note that this definition is not a good characterization of Meyniel graphs, since we need to enumerate all the odd elementary cycles to know whether or not a graph is a Meyniel graph. A by-product of our algorithm is a good insight into the structure of Meyniel graphs. The results of this paper will be used in a second paper to design good algorithms for the four optimization problems. In Section 2 we introduce some definitions and notation. I n Section 3 we define an operation (amalgamation) which constructs a Meyniel graph from two Meyniel graphs. We also state the main theorem of this paper. In Section 4,we give a constructive proof of this theorem. In the last section, we show how the main theorem may be used to give not only a good characterization of, but also a polynomial recognition algorithm for Meyniel graphs. Furthermore, we will show briefly that this characterization and this recognition algorithm can be specified for certain classes of perfect graphs found in the class of Meyniel graphs.
2. Notation and definition
Notation. Let G = ( V ( G ) , E(G)) be a graph with vertex set V(G) and edge set E ( G ) .When no confusion is possible we will write V and E for V(G) and E ( G ) . The complementary graph of a graph G will be denoted by (?. For a graph G= (V, E ) and S C V we shall denote by G ( S )the subgraph induced by S, and by G ( S ) its complementary graph.
Polynomial algorithm to recognize a Meyniel graph
227
We denote by T ( x ) the set of vertices adjacent to x , i.e.,
W )= {Y
E
v I (x, Y 1E El.
For A C V the intersection T ( x )n A will be denoted by T(x,A ) . Occasionally, when H = G ( A ) is a subgraph of G we shall write T(x,H ) for T(x,A). Terminology and definition. A minimal chain is an elementary chain which is an induced subgraph. The parity of a minimal chain is the parity of the number of its edges. A hole is an elementary cycle of length at least four without chords. The parity of a hole is the parity of the number of its edges. We say that two chords ( x , y ) and (2,t ) of an elementary cycle cross, if the vertices x, z, y , 1 are different and in this order on the cycle. A short chord of an elementary cycle is a chord which subtends a triangle. For a graph G = (V, E ) , a universal vertex for a subgraph G(A ) is a vertex x of V \ A such that T(x,A ) = A. We can also say that x is universal for A. A universal vertex of a graph G is a vertex x adjacent to all the vertices in G (except itself). A vertex partially adjacent to a subgraph G ( A ) is a vertex x of V \ A which is adjacent to some but not all vertices in A. A simplicia1 vertex of a graph G is a vertex whose neighbours form a clique. Let us define two families of graphs belonging to the class of Meyniel graphs: Complete multipartite graph. A complete multipartite graph is a graph in which the vertices can be partitioned into stable sets, where two vertices are adjacent iff they belong to different classes. Furthermore we assume there exist at least three classes with cardinality greater than two. Basic Meyniel graph (cf. Fig. 1). A basic Meyniel graph G = (V, E ) is a connected graph, where V can be partitioned into A, K, S such that: G ( A ) is a two-connected bipartite graph (which is not an edge); G ( K )is a clique; x E A, y E K J ( x , y ) E E ; S isastablesetof G;VxES, lr(x,A)lSl. In the class of Meyniel graphs, the notion of simplicia1 disconnecting clique plays a very important role.
Definition 2. In a graph G = (V, E ) , a simplicia/ disconnecting clique is a separating set K of G such that (i) G ( K ) is a clique, (ii) G( V \ K ) has at least two connected components containing more than one vertex and a universal vertex for G ( K ) .
M.Burlet. J. Fonlupt
228
I I
I
I
I
I I
I I I
I I I I
I I I
I
"3
I
Basic Meyniel graph
Expanded cycle
Fig. 1.
Definition 3. An expanded cycle (cf. Fig. 1) is a graph spanned by an odd elementary cycle. The only other edges outside this cycle are two short crossing chords. In other words, it can be labelled U ~ , U ~ , . . . , U ~ ~ + I such that ( u l , u 2 . . . . , u Z k )is an even hole and 1 ) 2 k + l is adjacent to only ol, u2 and U Z k . The cardinality of an expanded cycle is its number of vertices.
3. Amalgamating Meyniel graphs
Let us define a particular operation which will allow the construction of a graph G from two graphs, GIand G2.This operation is essential in this paper.
3.a. Description of the amalgamation of hvo graphs Let xi be a vertex of Gj and K, C T(xi) (i = 1,2) such that (a) KI and K , are cliques with the same cardinality, (b) z E K,,t E r(xi)\Ki j (z, t)€E(Gi) (i = 1,2), (c) T(Xl)\KI e T(xz)\K2=0. G is the amalgam of GI and G2(d. Fig. 2) if G is obtained from GIand G2by (i) one-to-one identification of the vertices of KI with the vertices of K2, (ii) creation of an edge between every vertex of r ( x l ) \ K Iand every vertex of
=o
r(xz)\ Kz,
(iii) deletion of the vertices xi (i = 1,2).
Fig. 2.
We shall denote the resulting graph by
G
= (GI,X I , Ki)@(G2,~
2
Kz). ,
Conversely, if for a graph G = (V, E), we can find two triples (Gi,x,, K , ) (i = 1,2) which satisfy the conditions (a), (b), (c) such that G = (G1,xl,KI)@(Gz,x2,Kz) we will say that G has then been decomposed by the amalgam operation (amalgam decomposed, @-decomposed) into two graphs GI and G2.
Remark 1. We warn the reader that the special case of amalgamation with KI = T(xl) and K2 = r(x2) is different from the usual operation of identifying a clique Kl in GI with a clique K2in G2. It may be interesting to note that amalgamation preserves perfectness. This result is very important in itself. However its proof will be given in a forthcoming paper in a more general set up. Here we shall only be concerned by the importance of the amalgamation in connection with Meyniel graphs. Definition 4. A proper decomposition of G is an amalgam decomposition of G into GI and G2 such that
IV(Gi)I
M.Burlet, J. Fonlupt
230
G
= (G,, x,,
K,)@(G,, x,.K,)
A @-proper decomposition of G
Fig. 3.
3.6. Decomposition theorem
Theorem 1. Zf GI and G, are Meyniel graphs, then G = (GI,x i , KI)@(G2,x2, K , ) is a Meyniel graph, i.e., the amalgamation of two Meyniel graphs i s a Meyniel graph. Proof. We have to prove that every odd cycle ( 2 5 ) of G has two chords. First of all we remark that any subgraph of G is either isomorphic to a subgraph of Gi ( i = 1,2) or the amalgamation of two subgraphs of Gi (i = 1,2), one of GI, the other of G2.So if there exists a counter-example G to this theorem, we can take one with the smallest number of vertices. So we shall assume that G = (GI,xl, K , ) @ ( G 2x2, , K z ) is either an odd hole or an odd cycle with one short chord. Let us call K the clique of G obtained by identification of K , and K z , and let N, be the vertices of G which are the exact copy of r ( x i ) \ K i for i = 1,2. By inspection of the subgraph of G induced by K U NIU N2 we can eliminate the following cases: 1
~
~
3
I N I ) > l and
, IN21>1,
I N I ) = l , IN2(>1 and
IK(a1,
) N , I = l , ) N I I > l and
IKI31,
) N I l = 1 , I N z l = l and
IKI>l.
So we have to study only the following four cases:
Polynomial algorithm to recognize a Meyniel graph
23 1
= 1 and IN2/= 1, NII=O and lN21=0, = 1 and IN21G3, = 1 and IN21>3.
(a) Let us call y the vertex of K obtained by identification of y l E T ( x l )and yz E T(xz)(cf. Fig. 4). Let z, be the unique vertex of T ( x i ) \ { y ,for } i = 1,2 and let 2 ’ and z ” be respectively the exact copy of z I and z 2 in G. Suppose that there exists a vertex of G, nonadjacent to x I and a vertex of G2 nonadjacent to x 2 . Then (y, z ’ ) is a separating set for these two vertices, similarly for ( y , 2”). This is impossible in our situation for G. So we may assume that G I is a clique defined on the three vertices ( x l ,y l , zl). Therefore the amalgamation G, of G t and Gz,is isomorphic to G2,which is impossible since Gz is supposed to be a Meyniel graph. (Fig. 4 shows that the amalgamation (GI,x I ,K , ) @ ( G 2x, z . K z ),where G I and G2(x2, yz, z 2 ) are triangles, gives a graph G isomorphic to G2.) Cases (b), (c) and (d) are treated in a similar way. 0 We now state the main theorem of this paper.
Theorem 2 (Decomposition Theorem). A connected graph G = (V, E ) , which is not a basic Meyniel graph, is a Meyniel graph if and only if G can be @-properly decomposed into two Meyniel graphs GI and G2. The preceding theorem has shown that the condition is sufficient. The proof that the condition is necessary will be quite long and will constitute all of the next section. Note that in the proof of this theorem, we will use the fact that when G is decomposed into G Iand G2,GI and G2are isomorphic to certain subgraphs of G. If G is Meyniel, then GI and Gz are also Meyniel.
Fig. 4.
M. Burlet, J. Fonlupt
232
4. Proof of the main theorem
To prove Theorem 2, we present a sequence of polynomial time procedures. Combining appropriately all these procedures, we get a constructive proof of Theorem 2 by an algorithm polynomial in 1 VI. For each procedure, when necessary, some remarks or special results are established before the full description of this procedure. The outputs of each procedure are followed by ‘stop’ or ‘execute procedure P,’. Instruction ‘stop’ is always accompanied either by a proof that G is not a Meyniel graph or by a proof that G satisfies Theorem 2. lnstruction ‘execute procedure P,’ means that we need to enter into procedure P, and the results obtained up to this point are the inputs of the procedure Pi. The algorithm to prove Theorem 2, which is proposed in this section, begins by verifying whether the graph G is a basic Meyniel graph. If it is not one, it tries to properly decompose G, by a special case of @ (G = (Gl,xl,K1)@(G2,xZ,K2), with K , = T ( x I )and K , = r(x2)). If such a decomposition exists, Theorem 2 is proven; if not, then there exists a hole. The algorithm detects such a hole, verifies whether this hole is even, and then examines the ‘neighbourhood’ of this hole. It may detect either of the two following configurations: a certain ‘multipartite structure’ or a certain ‘expanded basic Meyniel subgraph’ (cf. Definitions 6 and 8); if not then G is not Meyniel. The configuration detected permits the algorithm to properly decompose this graph by the operation @. For the @-decomposition found in the algorithm, i.e., G = G1@G2,GI and G, are isomorphic to certain subgraphs of G.
4.a. Is G a basic Meyniel graph?
Procedure PI. In this procedure we verify if G is a basic Meyniel graph. input : G =(V,E). (i) G is a basic Meyniel graph; stop. output : (ii) G is not a such graph; execute procedure P2. complexity : O ( I V 1).’ description: Delete S, the set ,of all the simplicia1 vertices. Detect then K, the set of all the universal vertices. Verify if G ( V \ ( K U S)) is two-connected bipartite. Verify that S is a stable set of G and x E S 3 IT(x,A)I 1. 4.b. 1s G properly decomposable by a special case of @ ?
In this subsection, we want to detect if G can be @-properly decomposed, G = ( G l , x , , K I ) @ ( G 2 , x 2 , K where 2), @ is restricted to the case where K , = I’(x,) and K 2 = T(x,). Proposition 3. Let G
= ( V ,E
) be a Meyniel graph and a disconnecting clique K
Polynomial algorithm to recognize a Meyniel graph
233
of G such that: there exists a universal vertex x for G(K) and G ( V \ K) has at least two connected components with more than one vertex, then there exists a simplicial disconnecting clique K1 included in K. Proof. Let K1 be a clique included in K which is minimal for this property: G( V \ K1) has at least two connected components with more than one vertex. Let GI = (Vl,E1)be the connected component of G ( V \ K1) containing x. Let G2= ( V2,E2) be another component of G ( V \ K , ) containing more than one vertex. Let y be a vertex of GZsuch that T(y, Kl) is maximal by inclusion among the sets T ( z , K 1 )where z ranges over the vertices of GZ.We claim that r(y,K ) = KI. Suppose not, then there exists z E Vzsuch that T ( z ,K1)Z T(y,K1),and also a minimal chain C(y, z ) joining y to z with all its vertices in G2.We can now find a chain (ul = y, u2, . . . , u,) whose vertices are in C(y, z ) such that:
T(u,,K1)C T(y,KI) V i Z t
and T(u,,K l ) f T(y, KI).
Let u E T(ut,K1)\T(y,Kl) and u E T(y, K,)\T(u,,Kl). Let I = max{i u E T(ui,K1), i = 1 , . . . , t } . Note that 1 < t. The induced subgraph by (ur,u l t l , . . . ,u,, u, u ) is a hole. If G is a Meyniel graph this hole is even, but adding x to this hole we obtain an odd cycle with only one chord. So T(y,K1)=K,. This proves the proposition for G2, and similarly for all the connected components of G ( V \ K,) with more than one vertex. 0
I
Procedure PZ. In this procedure we check if G contains a simplicial disconnecting clique K. A non-basic Meyniel graph G = (V, E). input : output : (i) A simplicia1 disconnecting clique K, G is properly 0decomposable, G = G1@GZ(cf. Remark 2); stop. (ii) The non-existence of such a clique K ; execute procedure P,. complexity : O ( I v I"). description: For each pair of nonadjacent vertices yl and y,, find out if K = T(yJ n T(yz)is a disconnecting clique such that yl, y2 are in different components HI,Hz of G(V \ K ) with I V(H1)I> 1 and 1 V(H*)I> 1. Remark 2. If Pz succeeds in finding a disconnecting clique K then G decomposes. More precisely, consider yi and Hi as defined in procedure Pz( i = 1,2). Let G1be an exact copy of G({yl}U K U V ( H 2 )where ) x1 corresponds to yl. Let G2be an exact copy of G({yz} U K U ( V \ ( V ( H z )U K))) where xz corresponds to y2. We have G = (GI,x l , T(xl))@(Gz,X Z , ~ ( x z ) ) .
M.Burlet, J. Fonlupt
234
4.c. Detection of a hole The following proposition shows that a graph G = (V, E), which is neither a basic Meyniel graph nor decomposable by Pz, is non-triangulated, and thus contains a hole. Recall that a triangulated graph is a graph in which every cycle of length at least four has at least a chord. A triangulated graph is a Meyniel graph. Definition 5. A basic triangulated graph G = ( V ,E) is a graph wherein there exists a partition of its vertex set into a stable set S and a clique K (cf. Fig. 5). These graphs are also called split graphs. They are a particular class of basic Meyniel graphs.
Basic triangulated graph
Fig. 5.
Proposition 4. A triangulated graph G = ( V , E ) , with no simplicial disconnecting clique K , is a basic triangulated graph.
Proof. Induction on I V I . The proposition is true for I V 1 = 1,2,3. Assume that the proposition is true for I V J = n - 1. We want to prove it true for I VI = n. The graph G contains a simplicia1 vertex u. Thus G(V\{u))satisfies the hypothesis of the proposition. By the induction hypothesis, the set V \ { u } can be partitioned into a stable set S , and a clique K , . If T ( u ) CK , , the proposition is true, if not, let u be the only vertex of T ( u ) not found in K , ; the subgraph induced by ( { u } U T ( o ) \ { u } )is a clique denoted K2. We have T ( u ) C ({u1
u n u 1)-
If the vertices of V \ K 2 form a stable set, the proposition is true; otherwise, there exists a component of G ( V \ K 2 ) containing at least two vertices, thus K2\{u} is a separating set of G and G ( { u } , { u } is ) another connected component with more than one vertex of G(V\(K2\{u))).Moreover u is a universal vertex €or K2\{u}, we can apply Proposition 3. We have a contradiction. Cl
Polynomial algorithm to recognize a Meyniel graph
235
Procedure P3. In this procedure we detect a hole in G. G = (V, E ) after an unsuccessful application of P2. input : (i) There is an odd hole, G is not a Meyniel graph; stop. output : (ii) There is an even hole C = ( u , , u2, . . . ,u 2 k ) ; execute procedure P4. complexity : O ( I V I"). description : Choose an edge (x, y ) of G. Delete from G the edge (x, y ) and the vertices T ( x ) nT ( y ) . If x and y are linked by a chain in the resulting graph, take a minimal chain in that graph from x to y . This chain with the edge (x, y ) forms a hole (even or odd). Repeat the procedure, until we have found a hole. 4.d. Detection of an expanded cycle Proposition 5. Let C = ( u , , u 2 , . . . , u 2 k ) be an even hole in a Meyniel graph G = (V, E ) and let x E V \ V(C). One and only one of the following is true (see Fig. 6): (i) G ( V(C) U {x}) is a bipartite graph, (ii) x is a universal vertex for C, (iii) G ( V(C) U {x}) is an expanded cycle. Proof. The first two cases are mutually exclusive. Suppose that G(V(C)U{x}) is not a bipartite graph, and that x is not universal for C. Then since G is a Meyniel graph, x is adjacent to two adjacent vertices of C and is not adjacent to at least a vertex of C. By a suitable relabelling of vertices of C, we can assume that x is adjacent to u , and u2 and not adjacent to u3. Let U, be the vertex of C with the least index ( t > 3 ) adjacent to x. If t < 2k, G(x, uz, u 3 , . . . ,u , ) is a hole of G ; if this hole is odd, G is not a Meyniel graph. If this hole is even (x, u,, uz, . . . ,u,) is an odd cycle with exactly one chord. Therefore t = 2k, which means that G ( V(C) U {x}) is an expanded cycle. 0
Fig. 6.
236
M. Burlet, J. Fonlupi
Proposition 6. Let G I = ( V,, E l ) be a two-connected bipartite induced subgraph of a Meyniel graph G = ( V, E ) , and x E V \ V,. One and only one of the following is true : (i) G(V, U { x } ) is a bipartite graph, (ii) x i s a uniuersal verrex for G ( V , ) , (iii) G ( V, U {x}) contains an expanded cycle.
Proof. Suppse that the first two assertions are false. An edge of El will be called a red edge, if x is adjacent to the two ends of this edge, otherwise it will be called a black edge. There exists a red edge (u, u , ) adjacent to a black edge (u, uz), since G ( V,) is a connected graph. Since G ( VI \ { u } ) is a connected graph, there exists a minimal chain in G ( V, \ { u } ) linking u , and u2. Let C, be a subchain of this one linking a vertex u ‘ to a vertex u ” such that (u, u ’ ) is red, (u, u ” ) is black, and u is not adjacent to any other vertex of C,. Then G(V(C,) U { u } ) is an even hole of (3. By Ihc preceding proposition G ( { x } U { u } U V(C,)) is an expanded cycle of G. 0 We can remark that this proof is algorithmic (in O( I V 12) . To color the edges of G, takes at most O ( I V 1)’ steps. To find the chain joining u , and u2 requires at most O ( I V )1: elementary operations. Now, to find o f and u ” requires O ( I Vl) operations. Then we have to check that G ( { x }U { u } U V(C,)) is an expanded cycle (if not, G is not a Meyniel graph). We can now describe procedure P.,.
Procedure P4. G = ( V , E ) and an even hole of G, C = ( u l r u z , . . ., u z k ) . input: One and only one of the following: output : (a) A complete multipartite induced subgraph; execute Ps. (b) A subgraph which is an expanded cycle of cardinality 5 ; execute P5. (c) A subgraph whose set of vertices is partitioned into C and K, I C I > 5 and G ( C )an expanded cycle. K is a clique containing all the universal vertices of G for G(C); execute P,. (d) G is not a Meyniel graph (a forbidden subgraph found); stop. complexity : o ( I v 13). description : V’ := V(C). Step 1. If there exists a non-universal vertex x for G ( V ’ ) ,adjacent to the vertices of an edge of G(V’), go to step 2; else, if there exists a non-universal vertex x for G ( V ’ ) ,such that Ir(x,V’)I 2 2, go to step 3; else, go to step 4. Step 2. We can apply here Proposition 6. After at most O ( 1 VI2) operations, we find either an odd hole, we have output (d), or an expanded cycle whose ) vertices ( u , ,i = 1,2,. . . ,2k + 1) are contained in V’ U { x } , ( u , , .. . , u ~an~even hole, and k k + l adjacent to u , , u2, u2k.
Polynomial algorithm to recognize a Meyniel graph
237
If k = 2 , we have output (b). If k > 2 and if all the universal vertices of G for the given expanded cycle induce a clique, we have output (c). If k > 2 and there exist two non-adjacent universal vertices w and w ’ , then ( u , , us,u4, w, w‘) is an expanded cycle of cardinality 5 , we have output (b). Step 3. If G ( V‘U {x}) is bipartite, V’:= V’U {x} go to step 1; else, G is not a Meyniel graph, we have output (d). Step 4. Let V, be the set of all universal vertices for G(V’). Find a minimal chain ( u , , uz,. . . ,u , ) whose vertices are in V \ ( V ’ U V,) and, among these vertices, only u , and u, are adjacent to some vertices in V’. If such a chain does not exist go to step 5; else, V‘:= V’ U ( u , , u2,. . . , u,). If G ( V’) is bipartite, go to step 1; else, G is not Meyniel, we have output (d). Step 5. At this point, as will be shown in the following remark, the universal vertices for G ( V’) do not induce a clique, or else, G is not a Meyniel graph, we have output (d). Let ( u l , u2 r . . . , u z k ) be a hole of G(V’), w and w ’ , two non-adjacent universal vertices for G ( V’). If k = 2, then G(ul,uz,u3,u4, w, w ‘ ) is a complete multipartite induced subgraph, we have output (a). If k > 2, then G ( u l , u3,u4, w, w’)is an expanded cycle, we have output (b). Complexity. Each step of this algorithm requires at most O ( I VI‘) elementary operations. We can enter at most I VI times into step 1. This procedure requires at most O ( I V )1’ operations. Remark 3. Let V, be the universal vertices for G(V’) when we enter step 5 of this algorithm. A t this point, G ( V‘) is a two-connected (bipartite) component of G ( V \ V,). Therefore, if G(V2) is a connected component of G ( V \ ( V , U V’)) there exists at most one vertex x of V’ such that V, U { x } disconnects V2 from V’\{x}. If such a vertex exists, let y be a vertex of V’ adjacent to x. Otherwise let y be any vertex of V’. Note that 1 V’I > 1, and y is universal for Vl or V, U { x } . Thus if VI is a clique and G is a Meyniel graph, Proposition 3 may be applied; this implies 1 V21= 1 otherwise procedure Pz would have found a proper decomposition of G by operation @. On the other hand, if all the connected components of G ( V \ ( V , U V’)) have exactly one vertex, G is a basic Meyniel graph which is impossible on entering procedure P4.This implies that if we enter step 5 of the preceding procedure, and if moreover V, induces a clique, G is not a Meyniel graph. 4.e. Multipartite structure Definition 6. An induced subgraph G ( V,) of a graph G = (V, E ) will be called a multipartite structure (m-structure) if V1 can be partitioned into k subclasses ( k a 2) W1, W 2 , .. . , W ksuch that G( Wi) is a connected component of G( V,),
238
M.Burlet, J. Fonlupt
1 s i s k, and at least two subclasses of the partition contain at least two vertices each (clearly this partition is unique). The subclasses with more than one vertex will be called proper subclasses. By a suitable indexing of the subclasses we shall always assume that W,, W2,. . . , W, are proper subclasses of the m-structure. We shall denote the partition of an m-structure ( W , ,W,, . . ., W,, K) where K is the union of all the non-proper subclasses.
Lemma 7. A n m-structure G(V,) of G = ( V , E ) is maximal (with respect to uertex inclusion) if and only ifthere does not exist a uertex x E V I‘ V, such that x is universal for a proper subclass of G ( V,).
Proof. Let x E V \ V,, not universal for any proper subclass of G(V,). Suppose that G(VI U { x } )is an m-structure, so there exists an edge in G ( V, U { x } )linking x to at least one vertex of any proper subclass of the m-structure G(VI). This implies that there exists exactly one connected component of G( V, U { x } ) with more than one vertex, so that G(VIU {n}) is not an m-structure. On the other hand, if x is universal, say for WI, a proper subclass of G(V,), G ( W,), will be a connected component of G ( V, U { x ) )and G((V, U {x})\ W , ) contains at least a connected component of more than one vertex. 0 We now give some results concerning maximal m-structures in Meyniel graphs.
Proposition 8. If G ( V,) is a maximal m -structure of a Meyniel graph G = ( V, E) and x E V \ V,, then x cannot be adjacent to two proper subclasses.
Proof. Let W, and W, be two proper subclasses of G ( V,). Suppose T(x, W , )# 0 and T ( x , W 2 )# 0. Since x is not universal for Wl and G (W , )is connected, there exist two vertices u,, u2 in W , such that ( u , , u2) E E, ( u , , x ) E E and (u2,x ) E E. Similarly there exist two vertices w , , w2 in W2 such that (w,, w2)6Z E, ( w l , x ) E E and ( w 2 ,x ) P E. Then (x, u,, w2,u2, w l ) is an odd cycle with only one chord ( u , , w,), which gives a contradiction. 0 Proposition 9. If G(V,) is a maximal m-structure of a Meyniel graph G = ( V ,E),whose partition is ( W , , W2, . .., W,, K ) with r > 2, there does nor exist a chain in V \ V ,joining two distinct vertices respectively adjacent to two distinct proper subclasses. Proof. If this proposition is false, let us consider the chain with the least number t of vertices contradicting this theorem. Let (u,, . . . ,u,) be such a chain. By the preceding proposition t > 1. We can suppose that
Polynomial algorithm to recognize a Meyniel graph
239
0, r(u’,W2)Z 0, U U l ,
WJZ
I‘(uj,Wi)=fl V l < j < t and l S i S r . Choose uoE T(u,,W,) and uc+lE T(u,,W2) and x E W,. Then (x, xo, u l , . . . , u,, ur+])is an odd cycle with exactly one chord (uo, urcl). This is impossible.
Remark 4. Let G(VI) be a subgraph of G = ( V , E ) . If G(V,) is a complete multipartite graph it is obvious that G(VI) is an m-structure. If G(V,) is an expanded cycle of cardinality 5 , G(V,) is also an m-structure. Moreover a maximal m-structure containing a complete multipartite graph or an expanded cycle with 5 vertices cannot be a graph composed of a complete bipartite graph and a clique K universal for this complete bipartite graph. Procedure Ps. A graph G = (V,E ) with the initial m-structure M given by input: procedure P4 (a complete multipartite graph or an expanded cycle with 5 vertices). (i) G is not a Meyniel graph; stop. output : (ii) A maximal m-structure with at least 3 proper subclasses. If G is @-properly decomposable then G = G1@G2;else G is not a Meyniel graph (cf. Remark 5 ) ; stop. (iii) A maximal m-structure with exactly 2 proper subclasses W , , W, such that W1 and W 2 are not both stable sets; execute Procedure P6. complexity : O( 1 V 1.)’ description: While there exists in V \ V ( M )a universal vertex x for a proper subclass of M do M:= G ( V ( M ) U {x}). If there exists in V \ V ( M )a vertex u adjacent to two proper subclasses of M, G is not a Meyniel graph, we have output (i); else we have output (ii) or (iii). Remark 5. If G(V,) is a maximal m-structure of a Meyniel graph with at least three proper subclasses, then the main theorem is true. Let ( W , , W2,.. . , W,, K) be the partition of the maximal In-structure. Delete from G the clique K and all the edges joining any vertex of Wl to any vertex of (W,, . . . , W,). By the preceding propositions W, and W , do not belong to the same connected component of this new graph. Let V’ be the connected component containing W, and V” the vertices of the other connected components. Consider the graph G I obtained by adding to an exact copy of G ( V’ U K) a new vertex xl, such that r(xl)is the entire image of WI U K. Similarly let G2be the graph obtained by
u:=,
M.Burlet, J. Fonlupt
240
adding a vertex x2 to an exact copy of G ( V ” U K), where image of ((U W, u K ). If G is a Meyniel graph we have
:=,
G
= (GI, xi, Ki)@(Gz,~
2
r(x,)is the entire
K2), ,
where we have chosen for K, the image of K in G, (i = 1,2). If we want to verify that G is Meyniel it is necessary to verify Propositions 8 and 9 (in O(l VI2)), which then permits the @-decomposition. 4.f. Expanded basic Meyniel graph
Definition 7. Let G(V,) be an induced subgraph of G = ( V , E ) . A subgraph G(V2) will be called a partitive subgraph if (i) Vz is strictly included in V, and V,l> 1, (ii) G(V2) is connected, (iii) u E V, \ V, j I‘( u. V2) = 0 or T(u, V2)= V2.
I
Definition 8 (cf. Fig. 7). An induced subgraph G ( V’) of a graph G = (V, E) will be called an expanded basic Meyniel subgraph if V’ can be partitioned into K , U , , .. . , U, ( f 3 2 ) such that: (a) u E K 3 V ‘ C { u } U T ( u ) (note that this implies that K is a clique), (b) if U, 1 2 2 , 1 s i zz f then G ( U ) is a partitive subgraph of G(V’), (c) there exists at least a partitive subgraph G ( U , )of G(V‘), 1 S i S t, (d) if we choose a vertex w, in each U,,the subgraph induced by u , , u2,.. . ,U, is a bipartite connected graph called the skeleton of G(V’), (e) for all LIE! V’, u is not adjacent, at the same time, to some vertex of U, and to some vertex of U,, where Ui and V, are represented in the skeleton by two adjacent vertices.
I
m-structure W v3
Expanded basic Meyniel graph Fig. 7.
Polynomial algorithm to recognize a Meyniel graph
231
The partition of G ( V’), mentioned above, will be denoted by ( K , U1,. . . . , U , ) and we will assume that U1is a partitive subgraph of G ( V’). It is important to note that an expanded basic Meyniel subgraph is not necessarily a Meyniel graph. We present here two procedures permitting us to detect, from certain P4 and P5 output configurations, an expanded basic Meyniel subgraph in a Meyniel graph.
Procedure Ps. A graph G = (V, E ) and a maximal m-structure with exactly two input : proper subclasses ( W1, W2,K ) given by P,. An expanded basic Meyniel subgraph G ( V’). If G is @-properly output : decomposable then G = GI@Gz;else G is not a Meyniel graph. (Refer to Remark 7;); stop. complexity : 0 ( 1 V 1.)’ G ( K U W1U W2) is a disguised expanded basic Meyniel subdescription: graph (cf. Remark 6). Detect the connected components of G(W,) and of G(Wz) which together with K constitute the partition of Definition 8. Remark 6. The only nontrivial point in order to prove that G ( V’) given by P, is effectively an expanded basic Mayniel subgraph of G is the condition (e) of Definition 8. But Ps assures us that this condition is satisfied when we enter P,. Procedure P7. A graph G = (V, E ) and a subgraph G(ul, u 2 , . . . , V 2 k + l ) ( k > 2 ) input : composed of an even hole ( u 1 , u 2 ,..., u z n ) and a vertex U 2 k i l adjacent to u l , uz, U Z k . Furthermore, all the universal vertices of G for this subgraph induce a clique K . (i) G is not a Meyniel graph; stop. ouput: (ii) An expanded basic Meyniel subgraph G(V‘). If G is @properly decomposable then G = G,@G,; else G is not a Meyniel graph; stop. (Refer to Remark 7.) complexity : o(i ~ 1 3 1 . description : For a vertex ui, 1 s i S 2 k , consider the subset of vertices adjacent to ui-l and ui+, (with the subscript arithmetic taken modulo 2 k ) ; consider the connected component U , containing u, induced by this subset. I f G ( K U U , ) is not an expanded basic Meyniel subgraph we have output (i); else we have output (ii).
u:El
M.Burlet, J. Fonlupt
242
Proposition 10. In a Meyniei graph G = (V, E ) , the subgraph G ( K U detected by Procedure P, is an expanded basic Meyniel subgraph.
u?:lU , )
Proof. Note that uZk+,E U1. Let u E K and u, E U,. Since ( v , , . . ., u,+ u,, a+,,.. . , v Z k ) is an even hole and u is adjacent to each v,, u is also adjacent to u,. This proves the first condition of the Definition 8. Let us show that u, E U, and u,+,E U,+, implies (u,, u , + , ) EE. Suppose it is not true: we can find u:+, and uY+, E U,+,such that (u,'+,,u,'+,)EE, (u,,u,',,)E E and (u,, u,'+,)gE. But then the cycle (ul, u 2 , . . . , or-,, u,, u,'+,,u,':~,u , , ~., . . ,u Z k )is an odd cycle containing exactly one chord. This is a contradiction, thus we have proven the second condition of the definition. Suppose there exists a vertex z P K U U, and z is adjacent to u, E U, and u , - , E L. Consider the hole C = (q,. . . ,u,-~, Y-,,u,, u , + ~. ., . ,UN). The vertex z cannot be universal for C otherwise z would be universal for the given hole ( u , , u z , ..., u Z k )and thus z would beiong to K. Proposition 5 proves that the vertex z is adjacent to three consecutive vertices of C, say u,-,, u,, v l + , . Hence (2,u , - , ) E E, otherwise the subgraph induced by ( u , , . . . ,u,-,,u,, u,+,, . . . ,u2&,z ) is an odd cycle with only one chord. Thus z E U, and we have a contradiction. 0
u?:,
Lemma 11. Let G( V , ) be a connected subgraph of a Meyniel graph G = (V, E ) with V, 1 > 1, and let z be a universal vertex for G (V,). If there exists a chain C ( x ,z ) in G(V \ V,) joining a vertex x to z where x is partially adjacent to V,, then there exists a minimal chain C,(y,z ) joining y to z where y is partially adjacent to V, and all the other vertices on C , ( y , z )are universal for VI.
I
Proof. By deleting eventually some vertices of the initial chain C(x,z), we obtain a minimal chain cl= (0,= y , uz, . . . ,u k = 2 ) whose vertices are in c(x,2 ) such that y is partially adjacent to V1and f ( o , , V,) = 0 or r ( v , , V,) = V,, V i # 1. We prove now that this chain C, is the desired chain of the lemma. If k = 2 , the lemma is proven. Suppose k > 2 and that u2 is not universal for G(VI). Let i o = min{i I'(u,, V,) = V1,uI E C , } .Consider two adjacent vertices ( w , w ' ) in V1 such that ( y , w') $ZE and ( y , w ) E E. The subgraph induced by ( w , w ' , u I , u 2 , ..,. u,) contains an odd cycle ( 35 ) with at most one chord. This is a contradiction. Suppose now that there exists a vertex of CI which is not adjacent to any vertex of G ( V,). We then can find a chain (v,, .. .,v,,,+,) Go, t 3 2 ) whose vertices are in C , ,such that only v, and ub+, are universal for G( V,). Furthermore u,-] is adjacent to at least one vertex w of V,. The subgraph induced by (uX,-,, u,, . . . ,u,,,, w ) contains an odd cycle ( 35) with at most one chord. This completes the proof. 17
1
Polynomial algorithm to recognize a Meyniel graph
243
uf=,
Theorem 12. Let G(V’) ( V =K U V,) be an expanded basic Meyniel subgraph of a Meyniel graph G = (V, E ) . Then there exists no chain in G ( V \ V‘) joining a vertex x to a uertex y where x is partially adjacent to U,, and y is adjacent to a vertex of U u,.
:=,
Proof. Let C(x,y ) be a chain which contradicts the theorem. The skeleton of the expanded basic Meyniel subgraph is connected, so we can concatenate the chain C(x,y ) to a suitable chain C(y,z ) to obtain a chain (vl = x , v 2 , .. . ,u, = z ) whose V,), is universal vertices belong to V \( K U U,) and where z , a vertex of for G ( U,).Moreover, by the preceding lemma we can suppose all vertices of this chain are universal for G(Ul) except u l . Let ui be the vertex of this chain with the smallest index belonging to U,). We have I < i o S r . But v+ is adjacent to ui, and also adjacent to some vertex of U1,which is impossible by definition of an expanded basic Meyniel subgraph. 0
(uf=,
(ul=,
Notation (cf. Fig. 8). Let G( V’) be an expanded basic Meyniel subgraph. Recall that we have V’ = (K U U,)) where K and U, satisfy Definition 8. For the remainder of this section, let us define: (i) is the union of U , and the set of vertices belonging to some connected component of G ( V \ ( K U U,)) such that this component contains at least a subgraph G(U,) i # 1, (ii) is the set of elements of V \ contained in some connected component of G(V\(K U U , ) )such that this component is adjacent to U1, (iii) K = { u E K ~ ( u P2) , z 01.
Cut=,
e,
vl
v2
1
M. Burlet, J. Fonlupt
244
Theorem 13. Let G ( V‘) be an expanded basic Meyniel subgraph, and V,, Vz,K as defined aboue. Then: (a) U ,is a partitive subgraph of G ( (b) each uertex of K is adjacent to all the vertices of universal for U , .
el),
v,
Proof. (a) This part is a reformulation of the preceding theorem. (b) Let z I be a vertex of linking z, and zz such that
K. Let z z E UI be such that there
exists a chain
e (*1
0 # ( V ( C ) \ { Z l , 2 2 ~ ) CQz.
Let C be the shortest among those chain satisfying (*). Let z be a universal vertex of G ( V’\K) for UI.Note that z is adjacent to 2 , and zz and to no other vertex of C. Therefore if I V(C) fl > 1 the subgraph induced by (V(C) U { z } ) contains either an odd hole 5 5 or an odd cycle 3 5 with only one chord. So we can assume that V(C) n ={z3}. Suppose that part (b) of the theorem is false. Consider a shortest chain C ( y l ,y z ) joining y , E to some y2 E V‘\(K U U , )where y l is universal for U , and not adjacent to a vertex z, E I% Let zz and z3 be defined relatively to z1 as stated above. Since the skeleton of G(V’) is connected, consider a chain C ( y Z r y 3joining ) y 2 to y?, whose vertices are in V‘\(K U U , ) and where y l is universal for U,. Concatenating C ( y l , y 2 )and C ( y z , y 3 )we obtain a chain CI= ( u , , . . . ,u,, . . . ,u I ) with u , = y , , u, = y z and of = y 3 . We can suppose that C(y,,y 3 ) is a minimal chain, which obviously implies that CI itself is a minimal chain. If r = 2 , u, cannot be adjacent to Ul (cf. Definition 8). Therefore G(zl,z2,z3,ul,u z ) is an odd cycle with only one chord; this is impossible. Suppose r > 2 and uz not universal for UI. Let u, be the vertex of Cl with the smallest index such that (ub,z r ) EE; note josr. The vertices u, with 1 < i S j,, cannot be adjacent to UI by the way we chose the chain C ( y l , y 2 ) . If u,, is universal for U , , the subgraph induced by (zl,1 2 , u I ,u z , ... , u,) is not Meyniel. If u, is not universal for U , ,the subgraph induced by (z3, zzrz,, u I , uz, . . . , uk,)is not Meyniel. Thus uz has to be universal for U , .This enables us to show that all the vertices of the chain ( u , , . . . ,u , ) are universal for U , . If this were not the case, we could find two vertices of CI,ub and u, UO> io + 1 b 3) such that u,, D,, u2,. ..,u,, are universal for U , but not u,,,, . . . ,u,-,. The subgraph induced by the vertices (zz, u,-,,. . . ,u,) is not Meyniel, which leads to a contradiction. Therefore u,-, (which V’) is universal for U , and adjacent to U, (which P V’ and is universal for U , )and this is impossible by definition of an expanded basic Meyniel graph. This completes the proof.
v21
vz
v,
Remark 7 (see Fig. 9). Take an exact copy of G ( V \ e2)and let GI be this copy where the set of vertices corresponding to those of U , reduced to a single vertex,
Polynomial algorithm to recognize a Meyniel graph
245
which we will call x,. Let K1 be the clique of GI which is the exact copy of K. Take an exact copy of G ( U UIU I?) and let Gz be this copy to which we add a new vertex xz, adjacent only to all the vertices corresponding to those of K U U I . Let K 2 be the clique of Gz which is the exact copy of I?. If G is Meyniel, G = ( G I , x , , K l ) ~ ( G 2 , x Z , KIfZ )we . want to verify that G is Meyniel, it is necessary to verify that UIis a partitive subgraph of G( and that every vertex of I? is adjacent to every universal vertex of for G ( V , )(in O ( 1 V IZ) , which would then permit the @-proper decomposition.
v2
el
vl)
5. Polynomial recognition algorithm of a Meyniel graph and applications 5.a. Polynomial recognition algorithm
We are going to give the description of a polynomial recognition algorithm of a Meyniel graph. Let G o = ( V " , E " ) be the graph to be recognized. The preceding algorithm enables us in polynomial time to detect for G o one and only one of the following situations: (i) G o is not a Meyniel graph, (ii) G" is a basic Meyniel graph, (iii) Go is not a basic Meyniel graph, but there exists a @-proper decomposition into two graphs G ' = (VI, E l ) and GZ= (V', EZ)and, moreover, G o is a Meyniel graph if and only if G ' and GZare Meyniel graphs. The idea of the recognition algorithm presented below will be to apply the algorithm of the preceding section to the graph G o and, inductively, to each new graph possibly obtained by this algorithm.
M.Burlei. J. Fonlupi
216
Definition 9. Let Ce be the family of graphs obtained by the recognition algorithm applied to G".A t the start of the algorithm, the family Ce contains only one element, G". When the algorithm stops the graphs of Ce will have been denoted G' ( i = O , l , . . . , ICeI-l), where G ' = ( V ' , E ' ) . The recognition algorithm Begin
%:=(GOj. k:=O. while some graphs in % have not yet been examined by the decomposition algorithm Begin Apply the algorithm of Section 4 on this graph. If this graph is not Meyniel, GO is not Meyniel; stop. If this graph is a basic Meyniel graph go to continue. If this graph is decomposed into two new graphs, call them G k + 'G , ktZ. %:= 92 U { G k + U ' } {Gk+'}. k:= k +2. continue: end
end
Since the algorithm permits only proper decompositions, it is obvious that it will stop. Definition 10. Let k(G")be the cardinality of the largest family % which can be obtained by applying the recognition algorithm on Go.The graphs contained in this family are indexed from 0 to k(Go)- 1. If a graph G' of a family 9 is decomposed by a @-decomposition into G' and G', G' and G h are called successors of G'. Let K' be a clique of maximum cardinality of G', and let r, = 1 V ' l - / K ' l . To prove that the recognition algorithm is polynomial we only have to show that k(G")is of polynomial order in I V o J . Theorem 14. If G' and G kare successors of G ' then (a) I V'I < I V'I and I V kI < 1 V ' 1, 1 K J I K' I and 1 K kI 1 K ' 1, r, c r, and rk d r,. (b) r, + rk S r, + 1.
I
Proof. (a) This part is a direct consequence of the definition of a @-proper decomposition.
Polynomial algorithm to recognize a Meyniel graph
247
(b) We shall prove part (b) for Go, G ' and G2. Let Go= (GI, xl, K1)@(G', x2, K2).Let K be the clique of G" obtained by the identification of K , and K2of the operation Qi. Let V? (resp. V:) be the subset of vertices of G o corresponding to the vertices V1\({xl}U Kl) (resp. V'\({x,} U K2))of G ' (resp. G'). Let N': (resp. N:) be the subset of vertices of Gocorresponding to the vertices T(xl)\K1(resp. r ( x 2 ) \ K 2 )of GI (resp. G2). Note that ro= I V?l+ I GI+IKI - IK'I. First case : V? n K" # 0 and @ n K" # 0. n KO lie respectively By definition of the amalgam operation, V? n KO and in NY and N:. Again, by definition of @, KO = (KO n N?)U (KO n Nz) U K. The subgraph of G ' induced by {xl}, K I and the subset of vertices of G ' corresponding to NY n KO is a clique of cardinality: I K 1 + 1+ 1 N': n KO 1 . Therefore r, s I V ( G ' ) l - ( l K l + 1+ 1N':n K o l ) = I V?l- I N? n K"1. Similarly r z S I V':l- I Ny n K" 1. As ro = 1 fll+ 1 El- IN': n KO1 - IN; n K"I, we get
rl + r2 s ro. Second case: V; r) K" = 0. (The case V? n KO = 0 is similar). There exists in GI a clique of cardinality I K'I. As I V(G')I = I K we have rl
S
I + I V?l+ 1,
1 V':l+ IKI + 1 - IK"I.
Moreover ({x2} U K 2 ) induces a clique of cardinality IK r z d ) V ( G 2 ) ) - ) K J - 1 = I V ~ Iand , we get
rl+r2
1 + 1 in
G'. Therefore
0
Theorem 15. We have k (G") d I V" I (rg + 1). Proof. Let %* be a family of cardinality k(G") obtained by the recognition algorithm applied on Go.The cardinality of the family of 'descendants' of G ' (a graph of %*) is equal to k(G'). This theorem is obviously true if 1 Vo1 = 1 or 1 V") = 2. So we shall suppose I V"I 3 3 and we shall prove this theorem by induction on I Vo1. If G o is a basic Meyniel graph, k(G") = 1 and the theorem is true. If Go is not a basic Meyniel graph, but its two successors in %* are basic Meyniel graphs, k ( G o )= 3 and the theorem is true. If Go is not a basic Meyniel graph and exactly one of the two successors, for instance G2, a basic Meyniel graph in %*,
k(G") = 2 + k(G') < 2 + 1 V' I ( r : + 1).
M.Burlet, J. Fonlupt
248
Since I V 11 < I V"Jand r1S
Tor
k(Go)S 1 Vol(r:+1) (since G ' is not a basic Meyniel graph, rl 3 1).
If G ' and G' are not basic Meyniel graphs,
k(Go)= 1 + k ( G 1 ) +k(G'),
k( G ' )s 1 V'I ( r ? + 1) < I v"J( r : + l), k ( G 2 ) s V'l(rg+ [ 1 ) < [ V"l(r:+l),
k (GO)s I V"1 ( rf + r: + 2), but we already know 2 S rl S To, 2 s r, S ro and rl + r2 S ro + 1. Therefore rf + r: + 2 S r i + 1 and the theorem is proven. 0
Theorem 16. The recognition algorithm proposed in this section is polynomial. Proof. The cardinality of 9 is bounded by 1 V"I'. Note that the bound given by the last theorem is very rough, but it already suffices to prove that the algorithm is polynomial. 0 5.b. Applications
Theorem 17. Characterization of Meyniel graphs. A graph is a connected Meyniel graph if and only if it can be built from basic Meyniel graphs by repeated applications of the amalgam operation. Proof. The sufficient condition is true by Theorem 1, because a basic Meyniel graph is a Meyniel graph, The necessary condition is given by Theorem 2. 0
Remark 8. The recognition algorithm given in this section allows us to detect, in a non-Meyniel graph, an odd cycle ( 55 ) not containing two chords. Let us define two particular cases of amalgamation which are already known.
Debition 11. If in the operation @, where G = ( G l , x l , K I ) ~ ( G 2 , ~ z , K wez ) , restrict K, = 0 (i = 1,2) we obtain an operation sometimes called composition of GI and G2. For this operation, we will note simply G = ( G I , ~ ~ ) @ l ( G 2(cf. ,~2) Fig. 10). If we restrict K, = T(x,),i = (1,2), we obtain this time a particular case of the usual operation of identifying a clique Kl in GIwith a clique K 2in G2.For ( G ~Such , a relation will this operation we will note simply G = (GI,x ~ ) @ ~ xz). presume, from now on, that T(xl) and r ( x z ) are cliques of the same cardinality.
Polynomial algorithm to recognize a Meyniel graph
249
\
I Bipartite
G*
I Fig. 10.
It is well known that the operation cD2 preserves perfectness. That is, if G , and G2are perfect graphs, G is also a perfect graph. This result is also true for @,, as given to us by Chvital [4] (cf. Remark 1). The results and algorithms obtained for the class of Meyniel graphs generate some properties for other classes of perfect graphs (which are also Meyniel) as we now outline. One of the earliest known classes of perfect graphs was the class of triangulated graphs. We know (Hajnal and Surinyi [ 9 ] ) that every minimal separating set of a triangulated graph is a clique. Dirac and others characterized these graphs in terms of simplicial vertices: a graph G is triangulated if and only if G and all its induced subgraphs contain a simplicia1 vertex. This last result yields a procedure for constructing all triangulated graphs: we can obtain all of them from a single vertex by iterated additions of simplicial vertices. We now give a new characterization theorem of this class.
Theorem 18 (Characterization of triangulated graphs). A graph is a connected triangulated graph, if and only if it can be obtained from basic triangulated graphs by repeated applications of operation dj2. Proof. Remark 2 and Proposition 4 show that a non-basic triangulated graph G is @2-decomposableinto GI and G2,both triangulated, where I V(G,)J< I V(G)I and I V(G2))< I V(G)I. Conversely, we know that Q2 applied on two triangulated graphs produces another triangulated graph. 0
250
M. Burlet, J. Fonlupt
Another class found within the class of Meyniel graphs which contains the triangulated graphs is the class of o-triangulated graphs, a class extensively studied by Gallai [5]: An o-triangulated graph is a graph wherein every odd elementary cycle contains at least two uncrossing chords. Gallai [ S ] , and later Surgnyi [14], showed that these graphs are perfect. Their proofs, however, do not give us polynomial recognition algorithms. Burlet and Fonlupt (see Burlet [2]) provided a constructive characterization of these graphs, which led to a polynomial recognition algorithm. We give here a characterization theorem for this class of graphs, whose proof (quite similar to that of Theorem 17) gives a polynomial recognition algorithm of these graphs.
Definition 12. A basic o -triangulated graph is (1) a basic Meyniel graph, or (2) a connected graph G = (V, E) where V can be partitioned into A, K , S such that: (i) A can be partitioned into at least three stable sets, each containing at least two vertices, and two vertices of A are adjacent if they belong to different stable sets, (ii) G ( K )is a clique, (iii) x E A, y E K 3 (x, y ) E E, (iv) S is a stable set of G and x € S j / r ( x , A ) l s l .
Theorem 19 (Characterization of o-triangulated graphs). A graph is a connected o-triangulated graph if and only if it can be built from basic o-triangulated graphs by repeated applications of operation G2. Proof. We use a similar proof to that of Theorem 17. Necessary condition. We have to show that if G, ( i = 1,2) is o-triangulated and G = G1!D2G2,then G is o-triangulated. This part of the proof is similar to that of Theorem 1. Suficient condition. This part consists in showing that an o-triangulated graph G, which is not a basic o-triangulated graph, can be dj2-properly decomposed into two o-triangulated graphs, G = G1dj2G2.To prove this let us follow the steps in the algorithm of Section 4. The procedure PI detects whether G is a basic Meyniel graph; a basic Meyniel graph is o-triangulated. Procedure P2detects the existence of a simplicia1 disconnecting clique leading to a G2-proper decomposition of G, G = G1cD2G2. The graphs GI and G2 are isomorphic to some subgraphs of G ; therefore, they are o-triangulated. Procedure P4applied to an o-triangulated graph G can only give, for a result, G(Vl), a complete multipartite subgraph of G (output (a)). Procedure Ps, using G(VI), can only give a maximal m-structure G(V') (V, V') which is a complete multipartite graph. This can be seen by showing that a proper subclass of this m-structure cannot contain an edge ( u , , u2). For
25 1
Polynomial algorithm to recognize a Meyniel graph
then let u3 be a vertex of this subclass such that ul E r(u3) and v , E T(u3)( u 3 always exists); let u4 and us be two non-adjacent vertices of another proper subclass of G ( V’). So the subgraph G(u,,u,, us, u4, us) is an expanded cycle, which is not possible. The m-structure G ( V’) is then a complete multipartite graph, containing at least three proper subclasses (cf. Remark 4). Let ( W,, Wz,. . . , W,, K ) be the partition of V‘ satisfying Definition 6 . Let us show that there does not exist a chain C(w :, w ’:) linking two vertices of W,, w : and w‘:,and whose other vertices do not belong to V ’ .Let Cl be such a minimal chain, let w, E W, (j# i ) and wk E Wk ( k # i, k # j ) . The subgraph G ( V ( C , )U { w,}U { wk}) is a cycle containing only two chords, which are two short crossing chords, and this is impossible. Thus, for all w E W,, w may be a cutnode of G ( V \ K ) . In this case K U { w } is a disconnecting clique satisfying Proposition 3, and the connected components of G(V \ ( K U { w}))not containing any vertices of V’, contain at most one vertex. This latter result follows because procedure P2 has not @,-properly decomposed G. The graph G is clearly a basic o-triangulated graph. This completes the proof. 0
u;=,
Remark 9. The polynomial algorithm briefly described in the preceding proof enables us to detect an odd cycle ( 3 5 ) not containing two non-crossing chords in a non-o-triangulated graph. There exists also a third important class of graphs related to the Meyniel graphs: parity graphs. These are the graphs in which every odd cycle of length at least five contains two crossing chords. Olaru-Sachs [13] have shown that these graphs are perfect. Burlet and Uhry [3] provided a constructive characterization (and a polynomial recognition algorithm of these graphs), as follows:
Theorem 20 (Characterization of parity graphs). A graph is a connected parity graph if and only if it can be built from bipartite graphs and cliques by repeated applications of operation GI. Acknowledgement The authors wish to acknowledge Professor V. Chvtital and an unknown referee for their helpful comments regarding the presentation of this paper.
References [l] C. Berge, Farbung von Graphen deren samtliche bzw. ungerade kreise starr sind (Zusammenfassung), Wiss. 2. Martin-Luther-Univ. Halle-Wittenberg Math.-Natur. Reihe (1961) 114-1 15.
252
M.Burlet, J. Fonlupr
[21 M.Burlet, Etude algorithmique de certaines classes de graphes parfaits, These troisitme cycle, Grenoble (septembre 1981). [3] M. Burlet and J.P. Uhry, Panty graphs (this volume, pp. 253-277). [4] V. Chvatal, private communication. 151 T. Gallai, Graphen mit triangulierbaren ungeraden Vielecken, Magyar Tud. Akad. Mat. Kutato In!. Kii71. 7 (1962) A .%Xi. [6] M.C. Golumbic. Algorithmic Graph Theory and Perfect Graphs (Academic Press, New York, 1980). [7] M. Grotschel, L. Lovisz and A. Schrijver, The ellipsoid method and its consequences in combinatorial optimization, Cornbinatorica 1 (1981) 169-197. [S] M. Grotschel, L. Lovisz and A. Schrijver, Polynomial algorithms for perfect graphs (this volume, pp. 325-356). [9] A. Hajnal and J. Suranyi, Uber die Auflosung von Graphen in vollstandige Teilgraphen, Ann. Univ. Sc. Budapestinensis 1 (1958) 113. [lo] L. Lovisz, Normal hypergraphs and the perfect graph conjecture, Discrete Math. 2 (1972) 253-267 (this volume, pp. 29-42). [ll] H.Meyniel, On the perfect graph conjecture, Discrete Math. 16 (1976) 339-342. 1121 G. Ravindra, Meyniel's graphs are strongly perfect (this volume, pp. 145-148). [13] H. Sachs, On the Berge conjecture concerning perfect graphs, Combinatorial Structures and their Applications (Gordon and Breach, New York, 1970) 377-384. [I41 L. Surinyi, The covering of graphs by cliques, Studia Sci. Math. Hungar. 3 (1968) 345-349.
Annals of Discrete Mathematics 21 (1984) 253-277 0 Elsevier Science Publishers B.V.
PARITY GRAPHS M. BURLET U.S .M . G.-IMA G -BP 53X-38041 Grenoble Cedex. France
J.P. UHRY C.N.R.S.-IMAG-BP53X-38041 Grenoble Cedex. France A graph G = (V, E ) is a parity graph if and only if for every pair of vertices (x, y ) of G all the minimal chains joining x and y have the same parity. A characterization of these graphs is given by a condition on odd cycles. This characterization shows that these graphs are perfect. These graphs are then studied from the algorithmic viewpoint. Polynomial algorithms are defined t o recognize them. and for the following problems: maximum stable set, minimum coloring, minimum covering by cliques, maximum clique.
1. Introduction
If some classes of perfect graphs have been characterized, it is rather strange that the algorithmic aspect is seldom studied. In particular, there are the classical problems of perfect graphs (maximum stable set, minimum coloring), but also the main problem of the recognition of these classes of graphs in polynomial time. These problems remain unsolved for many classes of perfect graphs (Meyniel’s graphs [9], claw-free perfect graphs [7], perfect 3-chromatic graphs [13], perfect planar graphs [12]) with the exception of the general paper of Grotschell, Lovasz and Schrijvers [6], which gives a polynomial algorithm of maximum weighted independent set for all perfect graphs. This algorithm based on the ellipsoid method unfortunately gives no idea of the structure of perfect graphs, and at the present time appears to be of no great combinatorial interest. In this study, we have dealt with parity graphs, a particular class of perfect graphs which is a fairly natural extension of bipartite graphs. Definition 1. A minimal chain is an elementary chain which is a subgraph.
In the following, we will simply write minimal chain so that there is no ambiguity. In the graph of Fig. 1, only chains (x, z , t, u, y ) and (x, z , u,u, y ) are minimal. 253
254
M.Bwkt, J.-P. Uhry
Fig. 1.
DeBhition 2. We call parity of a minimal chain the parity of its number of edges. In particular, if two vertices x and y are adjacent and the only minimal chain joining them is the chain reduced to the edge (x, y ) , this chain is odd.
Definition 3. A graph G = (V,E),simple, undirected is called a parity graph if and only if for every pair of vertices x and y of V, all the minimal chains joining x and y have the same parity. We notice immediately that this notion of parity graph is a generalization of bipartite graph. The graph in Fig. 2 is an example of a non-bipartite parity graph.
Fig. 2.
In Section 2, we give a characterization of these graphs, specifying the structure of the odd cycles they contain. A theorem of Sachs [lo] enables us to confirm that these graphs are perfect. In Section 3, we prove some properties of these graphs, and we specify their minimal separating sets. These results can be compared to those of Gallai [5] for o-triangulated graphs. In Section 4, a polynomial algorithm for parity graph recognition is given. It is also shown that these graphs are in fact built with two classes of perfect graphs: bipartite graphs, and cographs studied by Corneil, Lerchs and Stewart [3], here called 2-parity graphs. In Section 5, polynomial algorithms are defined for the four above-
Parity graphs
255
mentioned problems (in cardinality and in weight). This is another proof of the fact that these graphs are perfect.
2. Characterization
Definition 4. We say that two chords (x, y ) and ( z ,t ) of an elementary cycle cross if the vertices x, 2, y , t are in this order on the cycle. Theorem 1. A condition which is necessary and sufficient for a graph G = ( V, E ) to be a parity graph, is that every odd elementary cycle have two crossing chords. Proof. Necessary condition. The condition is necessary for a cycle of 5. Suppose we admit the property on an odd cycle of cardinality k ( k > 5 ) and prove it is still true for a cycle of cardinality k + 2 . It is easy to check that such an odd cycle contains at least two chords, and two chords which do not cross make at least one odd cycle, whose cardinality is lower or equal to k, and in which the induction gives the property. SufJicient condition. Take a graph G which verifies the condition and which is not a parity graph. As the structure we want for G is hereditary, we shall choose a counter-example which is minimal with respect to the vertices. This counterexample has two vertices x and y joined by an even minimal chain C,(x,y ) and an odd minimal chain C2(x, y ) with separate vertices (cardinality of C2(x,y ) 3 3 ) . As the counter-example is minimal, all the vertices of G are on the elementary cycle A made up for C,(x, y ) and C2(x, y ) . The proof refers to Fig. 3. V
Fig. 3.
I
256
M.Burlet, J.-P. Uhry
The chords of cycle A join a vertex of C,(x, y ) to a vertex of C,(x, y ) . Let us consider an edge joining 1 to the vertex of C,(x,y ) which is the nearest to y. This vertex cannot be z ' because the chains (1,2,3,. .. ,z, y ) and (1, z', y ) would have different parities, which contradicts the fact that G is minimal. This vertex cannot be t' for the same reason, with (x, l', 2', . . . ,t') and (x, 1, t'); and so on. We realize the only possible chord coming from 1 ( 2 respectively) is the chord (1, 1') ((z, z ' ) respectively). Now let us look at the chords coming from a vertex, different from 1, and as near as possible to x, on chain C,(x, y). Let u be this vertex, and let us consider the chord joining v to u' where u' is the nearest to y . This chord cannot end on a vertex at an odd distance from y (as in Fig. 3); in this case, the two minimal chains (x, 1 , . . . , u, u ' ) and (x, 1',2', . . . ,u ' ) would have different parities and this contradicts the minimality of the counter-example. This chord cannot end on a vertex which is at an even distance to y for the same reason. We should consider ( u , . . . ,t, z , y ) and ( v , u', . . . , z ' , y). It has therefore been shown that a vertex different from 1 and t could not be the initial end of a chord. The cycle A cannot contain more than two chords (1,l') and (2,2'). So cycle A does not contain 2 crossing chords, leading to the contradiction. 0
Theorem 2 (Sachs [lo]). Parity graphs are perfect. Without proof. we will quote an obvious corollary.
Corollary 2.1. A graph G = ( V ,E ) is a parity graph if and only if it does not contain any of the following con,figurations: (1) A,, . I : infinite family of odd cycles, without chord, on 2 k + 1 vertices, (2) A z k + , :in,finitefamily of odd cycles, with only one short chord, (3) As: cycle on 5 vertices with two adjacent chords at the same vertex. Here, it would seem interesting to recall two connected results: (a) o-triangulated graphs. A graph G in which every odd elementary cycle contains at least two short uncrossing chords is a perfect graph (see Gallai [ 5 ] and Suranyi [ll]). (b) A more general result ofMeynie1 [9]. A graph G, which has two chords in every elementary cycle, is a perfect graph. In Fig. 4. we give the example of a graph which contains two chords in each of its odd cycles but which is, however, neither an o-triangulated graph nor a parity graph.
Parity graphs
251
Fig. 4.
3. Description and properties
Notation. We shall denote, for a graph G = ( V , E ) and S C V, G ( S ) , the subgraph induced by S. We shall denote by T ( x ) the vertices adjacent to x, T'(x) ={y E v
I
(X,Y)E
El.
For A C V, T ( x )n A will be denoted by C (A). Occasionally, when H = G ( A ) is a subgraph of G, we will write
E ( H ) for
rX(A).
Definition 5. We call true twins two vertices x and y joined by an edge and having the same adjacents, i.e., T ( x )= T ( y ) . We call false twins two vertices x and y which are not joined but have the same adjacents. We know from Lovasz [8] that the operation which consists in adding one twin, true or false, of a vertex to a perfect graph is an operation which builds a new perfect graph. In addition, this operation applied to a parity graph retains this property. This is wrong, however, for an o-triangulated graph (Fig. 5), but true for a Meyniel graph [91.
Definition 6. A graph without twins will be called primal. In Fig. 6, we give an example of the reduction of a parity graph, and a primal parity graph.
Definition 7. In a parity graph G = ( V, E ) , we call partition induced by a vertex x, the ordered partition of the vertices of G, denoted ( P x , L ) ,where P, ( I x
M.Burlet, J.-P. Uhry
258
respectively) is the set of vertices of V joined to x by an even minimal chain (odd minimal chain respectively). We consider that x is joined to x by an even minimal chain. Notation. We may consider only the restriction of the bipartition, induced by x, to a subset A of V. We put Px(A)=P,nA,
Z,(A)=Z,nA.
Parity graphs
259
Lemma 1. A n y minimal separating set A of a parity graph G = (V, E) can be partitioned in two parts denoted R and B which have the following property. The vertices of R induce the same partition of vertices V - A, and the vertices of B the opposite partition of V - A. Vrl, V r 2 € R P,l(V-A)=P,2(V-A)
and then
LI(V - A ) = Z,( V - A ) ; Vrl E R, Vbl E B
P,l(V - A ) = Ibl( V - A )
and then Itl(V - A ) = p b l ( V - A ) . Proof. We suppose IA 1 > 1 (otherwise, it would be obvious). Let CXI and CX2 be two different connected components of the subgraph built on V - A (see Fig. 7). CX1
Let xIbelong to CXI,x2belong to CX2and two different vertices z and t to A. As A is minimal, there exists a minimal chain C1(xI,x 2 ) joining xIand x2 only containing the vertex z in A and a minimal chain C2(xI,xz) only containing t in A. It is easy to conclude, because chains CI(xI,x z ) and C2(xl,x2) have the same parity, that z and t have the same or a different parity for x1 and x2. By changing first xI then x2, the property easily follows. 0
Lemma 2. Let A be a minimal separating set partitioned in R and B as in Lemma 1. Then V n E R , V r 2 € R , ( r l , r 2 ) E E3 T,,(V-A)=T,*(V-A).
M.Burlet, 3.-P. Uhry
260
Proof. Suppose the property is false. Let x E r,,(V - A ) and x $Z r,2( V - A). From the definition, a minimal odd chain which does not contain r , , joins r2 to X.
We have therefore found the odd cycle ( T I , r2, u , . . . ,x ) (see Fig. 8). The only possible chords on this cycle are chords issuing from r l . Thus this graph is not a parity graph (see Theorem 1). 0
Fig. 8 .
Lemma 3. With the same hypotheses as those of the preceding lemma rl E R ,
rz E R, ( r l ,r2)EE
3 T J B ) = T,(B).
Proof. If ( r 2 ,b l ) does not exist, since ( r l . x ) and ( r z , x ) do not exist, there is one minimal chain from r l to x whose length is two and another from rz to x whose length is three (see Fig. 9). This is a contradiction to Lemma 1. El
Fig. 9.
With the same hypothesis us before, u connected component of the Lemma subgraph induced by the vertices of R has no minimal chain of length three. Proof. Suppose there is such a chain ( r , , r2,r3, r4). Let us consider a vertex x of V - A which is adjacent to r l . From Lemma 2, we know that x is also adjacent to r2, r 7 , and r4. The subgraph induced by these vertices is a 5-cycle, whose chords may only be those issuing from x. Here, there is a contradiction. 0
Remark 1. This property remains true if for the set R we take the adjacents of any one of the vertices of a parity graph. (They form a separating set which need not be minimal.)
Parity graphs
26 1
Definition 8. We shall call a 2-parity graph a graph of which the length of all minimal chains is less than or equal to two. As an example, a clique is a 2-parity graph. A complete multipartite graph is a 2-parity graph. (We call a complete multipartite graph a graph in which we can separate the vertices into stable sets, and where two vertices are joined if they belong to different stable sets. Such a graph is an o-triangulated graph Gallai' [5].) We give below a property characteristic of a 2-parity graph (for other properties, see Corneil, Lerchs and Stewart [ 3 ] ) . Lemma 5 . In a 2-parity graph G true or false.
= (V,
E ) , there are always at least two twins,
Proof. For a clique, it is obviously true. Otherwise, let x and y be two vertices which are not joined. There exists a minimal separating set A which separates x from y . Let CX, and CX, be their respective connected components, in the subgraph induced on V - A. For each vertex z of CX,, z is adjacent to all the vertices of A, otherwise there would be a minimal chain of length three. When CX, = { x } and CX, = { y } then x and y are false twins, else at least one has cardinality greater than one. Suppose, for example, I CX, I > 1. The proof continues by induction in CX,. Twins in subgraph CX, will effectively be twins in the initial graph, because they have the same adjacents in A. 0 From Lemma 5 , we immediately obtain an algorithm for building all 2-parity graphs (see Fig. lo). ?
0 -
I
- I -h Fig. Iff.
M.Burlet,J.-P. Uhry
262
Theorem 3. In a primal parity graph, every minimal separating set is a bipartite graph (one of the two parts can be empty).
Proof. This theorem directly follows the preceding lemmas, where we have shown that a connected component of the graph G ( R )(or G ( B ) )was a 2-parity graph and that for every rl and rz, two vertices of such a component,
T.,(B)= TJB) and T J V - A ) = C2(V- A). True or false twins, relating to this component, are therefore twins in graph G. Thus, by removing true or false twins, we can reduce this component to only one vertex. D Two properties below are given without proof. Theorem 4. A necessary and sufficient condition for a graph G = ( V ,E ) to be a parity graph is that every relatively minimal separating set A can be divided into two parts denoted R and B (one of them can be empty). In this partition, the vertices of R ( B respectively) must induce the same ordered partition (I, P ) of V - A ((P,I ) respectively). Lemma 6. In a parity graph, i f two adjacent vertices x and y have distinct neighbourhoods, then I, = Py.
4. A constructive polynomial algorithm for recognition
To recognize a parity graph, we shall ‘hang’ it by a vertex, and study the structure of the different levels (defined below). It will not only lead us to the polynomial algorithm, but also to a theorem which gives a very easy construction characterization of parity graphs. This is at the basis of the optimization algorithm described in Section 5. We shall study here the partition (Pa,I.) induced by a well-defined vertex a of a parity graph G = (V, E ) . Definition 9. We shall call C ( x ,y ) a shorter chain (in number of edges) joining x to
y.
Definition 10. We shall call level i (denoted N , )of G, the set of vertices of V at a distance i from a :
N,
= {x E
v 1 I c ( a , x ) l =i).
Parity graphs
263
The vertices of N , are linked to a by minimal chains of the parity of i. No only contains vertex a. N1is the set of neighbourhoods of a. rn will be the cardinality of the longest minimal chain issuing from a.
Remark 2. Dijkstra's shortest path algorithm enables the partition of V into levels to be determined in polynomial time. Remark 3. An edge (x, y ) E E has its extremities either on the same level, or on two successive levels. In Fig. 11 we give two possible 'hangings' of the graph of Fig. 6. 11 a
10
Fig. 11.
Lemma 7. All vertices of a connected component of the subgraph G ( N i )have the same adjacents at the level Ni-,. Proof (see Fig. 12). Let x E N , and y E Ni and (x, y ) E E. Let x ' E I', ( N i - , )and y ' E r,(N$-,).Suppose that (y, x') E E. As the edge (x', y ) does not exist, by concatenating a chain C ( a , x') to the chain (x', x, y), it is possible to find a chain joining a to y which has the same parity as i + 1. As there exists, by definition, one chain from a to y having the same parity as i, we have a contradiction. 0
M.Burlet, J.-P. Uhry
264
a
NO
Y'
Y
X
Ni-1
Ni
Fig. 12.
This lemma shows that the subgraph induced by a level is a 2-parity graph (see Lemmas 4 and 5). We remove the level N,,and we denote by G! , G : , Gf the different connected components obtained in the subgraph induced by the lower le6els: N+,u N,+*u N,. We denote also by X!'the vertices of N, adjacent to component G,k (see the example of Fig. 11).
Remark 4. The sets X: are not necessarily disjointed and do not necessarily cover all the vertices of Ni. Lemma 8. For every x E Xfand for every y E X:, we have
(x,y)EE
+ K(G:)=f,(G:).
Proof. If the property is false, there exists a configuration, illustrated by Fig. 13, where (x, u ) $Z E. There exists a chain C(a, u ) of the parity of i + 1. (u,y , x ) is a minimal chain of length two; this means that we can find a minimal even chain (u, . . . , v, x ) having all its vertices in G : . Linking this chain to a chain C ( a , x ) , we obtain a chain of parity of i which connects u to a. This is a contradiction.
0
Lemma 9. For every x E X , " and every y E X : , we have
r, (N,- x:) = r,(N,- x;). Proof. First case: x and y non-adjacent (see Fig. 14). There exists an even minimal chain joining x and y and having its other vertices in G : (otherwise, by
Parity graphs
265
Fig. 13.
Fig. 14.
linking this chain with a chain C ( a ,x ) we would obtain a minimal chain joining a and y whose parity would be different from the parity of i). If this chain is lengthened by the edge (y, z ) we obtain an odd minimal chain joining x and z. Consider the level N, ( j < i ) where j is chosen as large as possible, in such a way that there exists a vertex 6 EN, connected to x (to y respectively) by a chain C(b,x ) (C(b,y ) respectively).
M.Burkt, J.-P. Uhry
266
By linking these two chains we obtain a minimal chain. Indeed this last chain can have only horizontal chords by selection of j . Moreover, if this last chain contained a horizontal chord (u, u ) then u and u would have the same adjacents below (see lemma S), contradicting the choice of j . This chain is even; thus there is a contradiction. Second case: x and y are adjacent (see Fig. 15). There exists u E N,,, such that u is adjacent to x and to y (see Lemma 8). The vertices x, y, and z have a common adjacent u at level N,-, (see Lemma 7). (x, u, z, y, u ) is an odd cycle. It must contain two crossing chords (Theorem 1). This is impossible. Hence there is a contradiction. 5
Fig. 15. (x. y ) E E, (x, z ) E E
Definition 11. Associate at a given level N, a graph G' = (V', E ' ) . The vertices of V' are in one-to-one correspondence with the X f . Two vertices of V' are connected by an edge of E' if the corresponding Xf and X f ' have a non-empty intersection, and are not included one in the other:
(xfn x:) # 0, xf- (x:n x:) # 0, x:'-x:n x: # 0.
u,
We shall denote Vl = (Xf) where k covers the set of vertices of a connected component. It will be assumed that G ' has p connected components. It I S easy to check that the family %IL of the sets V!, where 1 covers (1,2,.. . ,p), is a nested family, partially ordered by inclusion. Lemma 10. (a) Let a given CJf be minimal by inclusion. Define K = {k E Vf}. An edge ( z , w ) in E cannot exist where z E Xf - X ! ' and w E Xf'for 1 and 1' belonging to K .
/x:
Parity graphs
267
By removing true or false twins, we can then reduce into a stable level the subgraph G (Uf). (b) If, in addition, there exists an edge (x, y ) in E on the level N, with y $Z U : and x E U : , then all vertices of U ! are false twins of the subgraph G induced by the level indicated by 0,1,2,. . . ,i - 1, i (after the reduction described in (a)). Proof. For a U : only composed of one X f , , the proof is obvious. The proof is given only in the case where U, is composed of two X : , namely X,' and X f . There is no difficulty in generalizing the proof. Suppose the edge (x, y ) exists. The same reasoning applies to the edge ( 2 , y ) . The existence of (x, y ) involves that of ( 2 , y ) . On the other hand, there exist u and v, two vertices of N,,, with (u, z ) E E, ( u ,x ) E E, (v, x ) E E and ( v , y ) E E (see Lemma 8). A five cycle has been noted without any crossing chords ( z , u, x, v, y ). Hence (see Fig. 16), a connected component of the induced subgraph G ( U f )has the same set of successors in N,,, (see Lemma 8); remember that U : is minimal in the family %.
Fig. 16
Such a connected component has the same set of predecessors at level N , - , (Lemma 7). This connected component is thus a 2-parity graph. In addition, as soon as there exists a vertex y E N, - U,' and an edge ( y , x ) E E with x E U,',all vertices of U : have y in their adjacents (Lemma 9). We can therefore reduce every one of the connected components of the subgraph G ( U : )into single vertices, by successive removal of true and false twins. If there is no edge ( y , x ) as above, the proof is finished. If such an edge exists, then the vertices just created have the same adjacents in N,. They belong therefore to the same connected component of G ( N , ) .They have the same adjacents in the above level (Lemma 7). They are thus false twins of the subgraph G ( Ul=(,N,). This completes the proof. 0 The previous lemma is the fundamental argument of the recognition algorithm given later on, which will allow us to demonstrate Theorem 5 below.
M.Burlet, .I.-P. Uhry
268
Definition 12. We define extension of a graph G by a bipartite graph B = ( X , U X z , A ), the operation of which consists in generating a new graph obtained by identification of certain vertices of XI with a set of false twins of G (possibly formed of a single vertex). Theorem 5. Every parity connected graph G = (V, E ) is obtained from a single tiertex by the following operations :
4, creation of a false twin,
42 creation
of a true twin,
d3 extension by
Q
bipartite graph,
applied successively and in any order. It is obvious that a graph obtained by the three operations +,, #2, (p3 indicated in Theorem 5 is a parity graph. The following algorithm will in fact check that a graph which satisfies Lemmas 7 , 8 , 10(a) and 10(b) can be obtained by these three operations beginning with a single vertex (see Fig. 17).
.
Parity graph
I
false twins
rartite graph
I Fig. 17.
The algorithm consists in looking for bipartites which eventually extend (as in (p3) a graph hung by a vertex a .
We carry out this research by beginning from the lower level N,,, and then climbing from level to level. As soon as we detect that one of these bipartites extends the remainder of the graph, we suppress this bipartite. It is the inverse operation of 4 3 . More precisely, beginning at the level N,, we are going to reduce into a stable
Parity graphs
269
level, in every iteration, one of the levels successively met, and this by the inverse of &, +2 and 43. It is clear that the level N, can be reduced to a stable set by the sole operations of reduction of a 2-parity graph (4, and &).
Lemma 11. If every level bipartite.
N,
(j> i ) is a stable set, then all the graphs G f are
Proof. Obvious. 0 Let us imagine that the levels N,, j > i, are reduced into independent sets. We describe the iteration which consists in reducing N, into a stable set. Consider (Xt,Xf,. . . ,Xf), where we suppose that we create q connected components G!, G f , .. . ,GP, by deleting N,.Let the family "11 = { U!, 1 E (1,2,. . . , p ) } be defined as in Definitions 11. On the level N,, vertices can be found which do not belong to one of X:,k E (1,2,. . . ,q). We regroup these vertices in a set denoted by UP". One iteration of the algorithm consists in checking that these U ! satisfy the properties of Lemma 10, and in this case reducing them to stable sets. To be able to reduce them to stable sets, it is necessary that any connected component of G ( U ! )has the same adjacents at the levels N-,, N,and N,,,. It is the level N,,, which forces us to handle the U ! in the order of inclusion (see Fig. 11). If once reduced to a stable set a U ! confirms Lemma 10(b),then the vertices of this U : are false twins for the subgraph G(u;-olV,). We must then delete the bipartites which correspond to them. The iteration on the level i is then continued until the level N, is itself reduced to a stable set.
Recognition algorithm for a parity graph
0 - Do i : = m go to 2. 1 - Take one minimal Ui' in the family %. Let K = {k 1 X,"E U!}. - Check that the family X:,k E K (which can be reduced to a single X:, containing only one vertex) satisfies Lemmas lO(a), 7 and 8.
- Reduce this Ut to a stable set by 4;' and 4;'. - If there exists an edge as in Lemma 10(b), check Lemma 9, and by 4;' delete the bipartite G f where k E K . - Delete this U ffrom the family %.* * Check in this step that this CJ: is a maximal element of the family 2.This should allow us to discover more easily the Xf-,.
M. Burlet, J.-P. Uhry
270
- If the family &! is not empty, go to 1. 2 - The only vertices from level N, belonging to a connected component of G ( N , )are vertices without successors at the level Nil. (They can contain, among other vertices, vertices of U ? + ' . ) - Check that the connected components of level N, on these last vertices satisfy Lemma 7. Then reduce them to stable sets by the inverse of 4, and (62.
If if 0 then do i:= i - 1 go to 1. If i = 0 End. It is clear that this algorithm stops at a finite number of iterations, and it reduces any parity graph to a single vertex. We have just established that a parity graph can be constructed following the operations indicated in Theorem 5. The main operation of this algorithm is to recognize that a graph is a 2-parity graph. This operation, itself polynomial, is repeated a polynomial number of times (in [ 3 ] an algorithm is indicated to recognize a 2-parity graph whose running time is H').
Remark 5. We have just demonstrated that a graph is a parity graph, if and only if it verifies Lemmas 7, 8 . 9 and 10. If the graph is not a parity graph, the algorithm finds the lemma in fault. The proof of the lemma in question then allows us to detect an odd cycle which does not have two crossing chords. We may use this algorithm to exhibit in any non-parity graph an odd cycle which does not verify the hypothesis of Theorem 1 .
Remark 6. A new problem, apparently similar to that of recognizing a parity graph, is as follows: Let G be a graph, and (x, y ) a pair of vertices of G. 'Are all minimal chains joining x to y even?'. A polynomial time algorithm for this problem would result in a polynomial time algorithm for determining whether a graph G has no chordless odd cycle. In fact, enumerate all the minimal chains of cardinality 2. This is polynomial. Let (x, 2, y ) be such a chain. Remove the vertex z , and the vertices r, - { x } -{y}. If all the minimal chains now joining x to y are even, then x and y do not belong to a chordless cycle, otherwise x and y do belong to an odd chordless cycle. The same algorithm, in the complementary of G, would allow us to check the hypothesis of the strong perfect graph conjecture. This problem appears much more difficult.
Parity graphs
271
5. Polynomial algorithms of optimization Interest in the characterization in the previous section aims at yielding the construction of polynomial algorithms for the four problems of perfect graphs. These algorithms are not related to the more general algorithms, such as that proposed in [6], since they are based on the algorithm of the maximum weighted independent set in a bipartite graph (Ford and Fulkerson [4]). This is also the case, for example, for the algorithms proposed in this book [4] for comparability graphs. Here, we try to solve the problems with weight function: c non-negative, integral, defined on V. We might think, in fact, that it is sufficient to solve the problem of maximum cardinality, as the transformation of a vertex into false or true twins enables a problem with an integer-valued cost function to be transformed into a cardinality problem. This transformation constructs a parity graph from a parity graph. Unfortunately, this transformation is not polynomial and, consequently, we shall study general problems of optimization. Let K be the matrix of all maximal cliques of G = (V, E ) where 1 V I = n. Each line of K is a characteristic vector of a maximal clique of G in {0,l}".Also, S is the matrix of all the maximal stable sets of G. The four problems paired by duality are as follows:
I
x, 3 0 ,
Kx
G
1,
c,x, = z (Max);
x,
I12
yt 2 0 , yK 3 c,
yi = W (Min);
0,
We solve each pair simultaneously: Maximum weighted stable,
Minimum covering by cliques,
Maximum weighted clique,
Minimum covering by stable sets.
The method used is classical since the works of Edmonds on matching, for example [ 2 ] ,and consists in finding for any non-negative integral weight function c, defined on V, a pair of dual integral solutions, x and y, which satisfy the complementary slackness theorem. For the three operations which transform G * into graph G, we are content
M.Burlei, J.-P. Uhry
272
here to give the transformation on the weight. Starting from a dual optimal integral pair defined on G*, x * and y * , we shall also give the indications which enable a dual optimal integral pair, x and y defined on G, to be obtained.
Notations. We shall denote by (Y the vertex of G * with respect to operations bl, &, 43,and by c, its weight; KP, KP, . . . , KP are all the maximal cliques of G * containing the vertex a, and when associated with the components of y * strictly positive, they are denoted respectively by y *(KP),y * ( K ; ) ,. . . , y *(KP). These I cliques are the restrictions of 21 maximal cliques of graph G in the case of false twins, denoted respectively KY, K4,. ..,KT, K t , K ; , . . . ,K f . They are the restrictions of 1 maximal cliques in the case of true twins, denoted respectively KY",K f h , .. . , Kpb. 5.a. Maximum weighted independent set. Maximum covering by maximal clique
5.a.l. False twins a and b. This is a particular case of 5.a.3, where
B = ( X , u X z , A ) , X,=({a}.{b)), X2=+, A =4. s.a.2. True fwins a and b (see Fig. 18). We define c,
= max(c,,
cb) = c,.
Fig. 18.
Transformation on the pair of optimal solutions. x * ( ( Y ) = 1, x ( a ) = 1, x ( b ) = O ,
(i) Case where
y ( K f b )= y *(KP),. . .,y(KY4 = y *(KP) (ii) Case where x * ( a ) = O , x ( a ) = O , x ( b ) = O ,
y*(KY). Y ( K ' ; ~ ) = ~ * .( .K. , ~y ()KY>= , The components of x and y which are not defined are unchanged. This pair x and y fulfills the complementary slackness conditions.
S.a.3. Extension by a bipartite. We suppose the bipartite B only fixed by two false twins, a arld b.
= (XIU X,, A
) is
Panty graphs
213
The reader will be released from this constraint without any difficulty. We define c, = s s , where s:ax stands for the value of a maximum weighted stable in B, denoted SL,, and s f b stands for the value of a maximum weight stable in B, where vertex a and vertex b have been eliminated. (In general, we shall denote the sets in capital letters, and the value in lower case letters.) The idea of this transformation is quite simple (see Fig. 19). As soon as a vertex in the neighbourhood r,(G*) of a or b is in the maximum weight independent set of G,it contains Sfb. In the other case, it contains S L , . We translate therefore by c, the regret which results from putting a in the maximum weighted stable of G*.
a
Fig. 19.
Transformation on the pair of optimal solutions. (i) Case where x * ( a )= 1. We obtain x by withdrawing from x * its component x * ( a ) ,and by extending this vector by a characteristic vector of S:ax. (ii) Case where x * ( a ) = 0. We complete in this case by a representative vector of s,",. To obtain y, the process is slightly more elaborate. It is necessary to find an integer solution to the linear system
(iii) Case where x*((u)= 1. We know that y*(K:)+y*(K;)+.-.+y*(KY)= c,. We shall take ta+tb= c,.
M.B ~ l e tJ.-P. , Uhv
274
This linear system is of the transportation type. If its right-hand side is integral, an integer solution to the linear system may be found easily. It remains to determine 5. and tbin order that the solutions x and y found satisfy the Complementary slackness conditions. U X z , A ) a new bipartite graph B ' = Let us construct from B =(XI (X;U Xi,A'),
Xi = X2U { P I , Xi= XI,A ' = A
U ((p, a ) , ( P , b ) ) .
We define cg = c,. Solve the maximum weighted stable set in B' and let x' and y' be the integer optimal pair of dual solutions obtained. There exists in B' a stable set of maximum weight containing f?:
and such that €3'
Smax
=S
L .
The complementary slackness theorem applied to B' allows us to affirm (a) Y '((P,a 1) + Y '((P,b 1) = c, (b) only the edges of bipartite B', with one of their extremities always in a maximum weighted stable set, have a possible positive component. We shall define ?
6.
= Y '((BYa 11,
(b
= Y '(@,
1)-
It is easily proved that the vector x previously described and the vector y obtained by solving the system (1) and restricting y' on the edges of B, fulfills the complementary slackness conditions. (iv) Case where x *(a)= 0. This case is identical to the preceding case in + . + y *(KP), which can everything, except for the quantity y *(KP)+ y *(G) be greater than c, and no longer equal. We shall substitute in this case for la and &, any integer solution of the following linear system of inequalities: [a
a c a ,
(a
Cbr
6 + (b
= y *(KP) y * ( K ; )
* ' *
+ y *(KP).
In both cases, we were able to obtain a pair of dual integer solutions, x and y, satisfying the complementary slackness conditions.
5.6. Maximum weighted clique. Minimum covering by maxim.al cliques The notations used here are very similar to those in the preceding section.
215
Parity graphs
5.b.l. False twins a and b. Identical with 5.b.3. 5.b.2. True twins a and b. We define c, = c, + c b . (i) Case where x * ( a ) = 1, then x ( a ) = 1 and x ( b ) = 1. (ii) Case where x * ( a ) = O , then x ( a ) = O and x ( b ) = O . By a similar procedure to 5.a.1, we obtain in both cases an integer vector y which satisfies with x the conditions of the complementary slackness theorem. 5.b.3. Extension by a bipartite. We shall mention only the case where the maximum weighted clique is in G*. We define c, = max(c,, c b ) = c,. We find the maximum weighted clique in G*, and in B. We take for the primal solution x the characteristic vector on G = (V, E ) of the clique which has the heavier weight. In order to find a dual solution y defined on G, we shall proceed as in 5.a.3 by starting from a dual solution for the maximum weighted clique problem on graph B'. ( B ' is defined as before; the vertex p of B' supports the weight of the maximum clique defined by x and reduced by c,.) The y ' dual solution on B' has strictly positive components on stables which are the restrictions of stables in B. The latter ones can be partitioned, due to the fact that they contain, or d o not contain, the vertices a and b in three families, denoted by Y ; b , Y%, and Y:b. The dual solution y * on graph G * has strictly positive components on stable sets which can be partitioned into two families, denoted by Y:* and 9';'. A value is assigned to each of the stable sets of these families. This value is initially the (integer) component of vector y ' or vector y * corresponding to the stable set. We shall now juxtapose two stable sets, one from a family defined in B, the other from a family defined in G*, in such a way as to obtain a stable set defined in G. Each one of the stable sets thus obtained will correspond to a component of the vector y which is sought. This component is defined by the minimum of the values associated with the two stable sets forming it. We now deduct the value of this minimum from the quantities corresponding to the two stable sets. The stable sets associated with a zero value are then removed. We continue to define y in the same way, for the remainder of the families. It is, however, necessary to proceed in a fixed order. We first juxtapose the family Y : b with the family 92' ( a is replaced by a and b ) , then the family with 9 ' : ' ( a is now replaced by a ) . At this stage, the dual constraints corresponding to vertices a and b are satisfied. ' : and finally juxtapose the We can then juxtapose Yp,"'( a is removed) with 9
9'2
276
M. Burlet, J.-P. Uhry
rest of the latter with Y$*.This procedure allows us to find an integer dual solution y which satisfies with x the complementary slackness theorem.
Remark 7. Subsection 5.b.3 supposes an algorithm is known for covering the vertices of a bipartite graph by stable sets (dual problem of maximum weighted clique). Such an algorithm, apparently never described, is not difficult to imagine. The final section S.b, very short and very technical, obscures the fact that the two problems, maximum cardinality clique and minimum coloring, have an almost greedy solution. Yet, here again, we find that a parity graph is very much like a bipartite graph: the maximum cardinality clique problem is easier than the maximurn cardinality independent set problem. A general theorem follows from subsections 5.a.3 and S.b.3. Theorem 6 . if we extend a perfect graph by a bipartite graph, we obtain a new perfect graph.
Proof. Subsections S.a.3 and 5.b.3 are two proofs of this theorem. Indeed, they do not take into account the fact that G * is a parity graph, only that G * is a perfect graph. This theorem is even easier to deduce. It suffices in fact to prove that the cardinality of a maximum clique is equal to the chromatic number in all perfect graphs extended by a bipartite. It gives a direct demonstration of Theorem 2 (when we know Theorem 5).
References [I ] C. Berge. Graphes et hypergraphes (Dunod, Paris, 1970); English translation: Graphs and Hypergraphs (North-Holland, Amsterdam. 1973). [2) M. Boulala and J.P. Uhry, Polytopes des indtpendants d’un graphe serie-parallele, Discrete Math. 27 (1979) 225-243. 131 D.G. Corneil. H. Lerchs and L. Stewart. Cographs: a new class of perfect graphs, Discrete Appl. Math. (to appear). [4] L.R. Ford and D. R. Fulkerson, Flows in Networks (Princeton University Press, Princeton, NJ, 1962). [S] T. Gallai. Graphen mit triangulerbaren ungeraden vielecken, Magyar Tud. Akad. Mat. Kutato Int. Kozl. 7 (1962) A.3-36. [6] M. Grotschell, L. Lovisz and A. Schrijver, Polynomial algorithms for perfect graphs (this volume. pp. 325-356). [7] Wen Lian Hsu. How to color claw free perfect graphs, Ann. Discrete Math. 11 (1981) 189-197. [XI L. Lovisz. Normal hypergraphs and the perfect graph conjecture. Discrete Math. 2 (1972) 253-267 (this volume, pp. 29-42). (91 H. Meyniel. On the perfect graph conjecture, Discrete Math. 16 (1976) 33%342.
Parity graphs
277
[lo] H. Sachs, On the Berge conjecture concerning perfect graphs, Combinatorial Structures and their Applications (Gordon and Breach, New York, 1970) 377-384. [ l l ] L. Surinyi, The covering of graphs by cliques, Studia Sci. Math. Hungar. 3 (1968) 345-349. [12] A.C. Tucker, The strong perfect graph conjecture for planar graphs, Canad. J. Math. 25 (1973) 103-114. [13] A.C. Tucker, Critical perfect graphs and perfect 3-chromaticgraphs, J. Comb. Theory, Ser. B 23 (1977) 143-149.
This Page Intentionally Left Blank
Annals of Discrete Mathematics 21 (1984)279-280 @ Elsevier Science Publishers B.V.
A SEMI-STRONG PERFECT GRAPH CONJECTURE
v. CHVATAL School of Computer Science, McGill University, Montreal, Canada
A graph is called perfecr if, for each of its induced subgraphs F, the chromatic number of F equals the number of vertices of the largest clique in F. This notion has been introduced by Claude Berge [l], [ 2 ] , who proposed the following two conjectures: (1) a graph is perfect if and only if its complement is perfect, (2) a graph is perfect if and only if it has no induced subgraph isomorphic to an odd cycle of length at least five or to the complement of such a cycle. The first of these conjectures has been proved by Lovasz [3]; the second, known as the Strong Perfect Graph Conjecture, remains open. The purpose of this note is to interpose a new conjecture between (1) and (2). Let us say that a graph G has the P4-structure of a graph H if there is a bijection f between the set of vertices of G and the set of vertices of H such that a set S of four vertices in G induces the path P4of length three in G if and only if f ( S ) induces a P4in H. The conjecture interposed between (1) and (2) is: (1;) if G has the P4-structure of a perfect graph then G is perfect.
Trivially, (1;) implies (1); to show that (2) implies (l;), it will suffice to establish the following fact.
Theorem. The only graphs having the P4-structure of an odd cycle of length at least ,fiue are the cycle itself and its complement. Sketch of Proof. When the cycle has length five or seven, the assertion can be verified by an exhaustive search. Now let n be an integer greater than eight and let a graph G have the P4-structure of the cycle C, of length n. The vertices of G may be labeled as u , , u z r . .. , u. in such a way that, with the subscript arithmetic taken modulo n, a set S induces a P4 in G if and only if S = {?I,+,, u , + ~u, , + ~ , for some i. Now each subgraph G, of G induced by {a+,,u , + ~. ,. . , u , + ~has } the P4-structure of the path P8 of length seven. Further exhaustive search reveals that there are precisely four such graphs: P8itself, its complement p8,the graph F shown in Fig. 1, and its complement I? Now it is easy to verify that 279
v. Chvcitaf
280
Fig. 1
G, Pa implies G,,, = Pa, 2
-
G, = p, implies G,,, = Pa, G,
G,
2
I (
F implies G,+, = F,
P implies G,+, =E
In particular, if some G,is isomorphic to Pathen every G,is isomorphic to Pa.In this case, it is easy to verify that G is isomorphic to C.. Similarly, if some G,is isomorphic to pa then G is isomorphic to The remaining two options are excluded whenever n is odd: in this case, we would have G,,, = GI whenever k is odd, which is incompatible with G,,, = GI.
cn.
References [l] C.Berge, Fiirbung von Graphen deren samtliche bzw. ungerade Kreise starr sind (Zusammenfassung), Wiss. Z. Martin-Luther-Univ. Halle-Wittenberg, Math.-Natur. Reihe (l%l), 114. [2] C . Berge, Sur me conjecture relative au problbme des codes optirnaux, Commun. 13ibme Assemblte Gtn. URSI, Tokyo (1%2). [3] L. Lovhz, Normal hypergraphs and the perfect graph conjecture, Discrete Math. 2 (1972) 253-267 (this volume, pp. 29-42).
Annals of Discrete Mathematics 21 (1984) 281-297 @ Elsevier Science Publishers B.V.
A METHOD FOR SOLVING CERTAIN GRAPH RECOGNITION AND OPTIMIZATION PROBLEMS, WITH APPLICATIONS TO PERFECT GRAPHS* Sue H. WHITESIDES School of Computer Science, McGill University, Montreal, Canada In this paper, a polynomial time algorithm for finding clique cutsets is used to provide polynomial time algorithms for testing membership in certain families of graphs and for solving certain optimization problems in them. These optimization problems are: finding a maximum weight clique or stable set, a minimum coloring, and a minimum weight fractional cover by cliques or stable sets. The method of the paper is quite general. Applications arc given for perfect graphs: The recognition and optimization problems above are solved for clique separable graphs, and a recognition algorithm for i-triangulated graphs is given. answering questions of Gavril.
1. Introduction
In this paper, we will show how to use a polynomial time algorithm for determining whether a graph has a clique cutset (Whitesides [IS]) to obtain a fast (i.e., polynomial time) test for membership in certain families of graphs. In these families, we then provide fast solutions to the following optimization problems: finding a minimum coloring, a maximum weight clique, a maximum weight stable set, a minimum weight clique cover, and a minimum weight stable set cover. In particular, we will show that for any hereditary family F of graphs, these problems, with the exception of finding minimum weight stable set and clique covers, can be solved in polynomial time for arbitrary members of F whenever they can be solved in polynomial time for those members that have no clique cutsets. The same is true for recognizing members of a given family. If graphs in F are perfect, then the minimum weight cover problems can also be solved by our method. The motivation for seeking a fast way to find clique cutsets came from the problem of recognizing perfect graphs. It is well known that joining two perfect graphs along a clique contained in both is a perfection-preserving operation; hence, a natural first step in testing a graph for perfection would be to look for a
* This work was supported by a Dartmouth College Junior Faculty Fellowship. 281
282
S.H.Whitesides
clique cutset, and in case of success, to test the graphs joined together along the cutset for perfection. While we do not have a general test for perfection, we can nevertheless show how to recognize some subfamilies of perfect graphs using this idea. As an application of our general methods, we will answer several questions raised by Gavril [lo]. In particular, we will show how to solve the optimization problems mentioned above for ‘clique separable’ graphs. These graphs, studied by Gallai [8] and Gavril [lo], are perfect and include as a proper subfamily the ‘i-triangulated’ graphs of Gallai [8]. Gallai’s graphs are defined by the property that their odd cycles of length at least five always have a pair of noncrossing chords. Chordal (‘triangulated’) graphs, which include bipartite graphs, are examples of Gallai graphs. Grotschel, Lovlsz and Schrijver [131 have given a polynomial time algorithm, based on the ellipsoid method, for finding a maximum weight stable set and a minimum coloring for perfect graphs in general. Hence, fast ways to solve those two problems are already known for the clique separable graphs in our application. Nevertheless, one might ask for a more structure related solution. The method in this paper is quite general and does not depend on having the property of perfection, except in the case that a minimum weight cover by stable sets or cliques is desired. It is hoped that the reader will find the general approach given here to be not only conceptually simple but also structurally elucidating for particular families of graphs of interest. We will begin in Section 2 by reviewing the way in which any graph gives rise to a ‘clique cutset tree’ that represents the structure of the graph with regard to its clique cutsets. The clique cutset tree can be constructed quickly using the algorithm of Whitesides [19] for finding clique cutsets, which w e briefly describe at the end of Section 2. In Section 3, we will give an algorithm for reconstructing a graph from its clique cutset tree. At each step in the algorithm, a ‘primitive’ subgraph is ‘glued’ onto the partially reconstructed graph at a clique. In Section 4, we will use the fact that every graph can be built up in this way to show how to solve recognition and certain optimization problems for graphs when solutions to these problems are known for their clique cutset-free subgraphs. We will turn to applications in Section 5, where we review clique separable and Gallai graphs and show how to solve various optimization problems for these. In Section 6, we will give recognition algorithms for clique separable and Gallai graphs. Section 7 is devoted to concluding remarks. Throughout this paper, we will assume that our graphs are finite, undirected, and connected, and that they have no loops or multiple edges. A graph with vertex set V and edge set E will be denoted G, G (V), or G (V, E ) . The number of vertices in G will be denoted by n. A clique of G is a set of mutually adjacent vertices of G, but need not be a maximal such set. A cufset of G is a set of
Solving certain graph recognition and optimizationproblems
283
vertices whose removal disconnects G, and a clique cutset is a cutset that is also a clique. The weight w ( S ) of the graph induced by a set S of vertices is the sum of the weights of the vertices in S. A hereditary family is called an h-family. Given an h-family F of graphs, the problems for which we will seek solutions are the following: (I) membership: determine whether a given graph G is a member of the family F. (11) maximum clique : Given a graph G that belongs to F and whose vertices have been assigned nonnegative integer weights, find a clique in G of maximum weight. (111) minimum coloring: Given a graph G belonging to F, find a coloring of the vertices of G, adjacent vertices being assigned different colors, that uses the minimum possible number of colors. (IV) maximum stable set: Given a graph G that belongs to F and whose vertices have been assigned nonnegative integer weights, find a stable set of maximum possible weight. In case the graphs in the family F are all perfect, then two additional problems that will interest us are: (V) minimum couer by cliques: Given a graph G( V) that belongs to a family F of perfect graphs and that has nonnegative integer weights on its vertices, find an assignment of numbers yi ( Ci) to the maximal cliques Cisuch that the sum of the yi’s is as small as possible subject to the following condition. For each vertex 0
E
v,
(VI)minimum cover by stable sets: This problem is the same as (V) above, except that ‘clique’ is replaced by ‘stable set’.
2. Constructing clique cutset trees
Suppose that A is a cutset of G ( V), and let Vi be the vertex set of any of the connected components of G( V - A ) . Then we say that the graph G( V, U A ) is a child of G produced by A. (This terminology differs from that of Gavril [lo], where G ( V, U A ) is called a ‘leave’ of G. We have changed this terminology because we will want to use the term ‘leaves’ to mean degree 1vertices of a tree other than the root.) Following Gavril [lo], we now construct a labeled tree T ( G ) from G as follows. Let the root of T ( G )be labeled (G,A), where A is a minimal clique cutset of G. (See Fig. 1.) Then for each child of G produced by A, create in the tree T ( G )a child of the root. Once these new vertices have been added to the
S. H. Whitesides
284
G 2 15
3
Fig. I A graph G and a cutset tree T ( G )for G. Here, G = ( I , . . ., IS} and A = (4, S}; G, = {l,. . . ,S}, a chordless cycle of length 5 ; G2= { 4 , . ... IS} and A , = ( 7 } ; G , = ( 4,..., 8}, a clique; G,= {7,Y. 10.1 I ) and A, = {Y. 10); G, = {7,12,13, I4.15) and A , = {12.13); G , = (7.9, lo), G, = (9,10, ll}, G, = (7.12.13). and G , = (12,13,14.15}, where G,. . . . G, are cliques. (We have denoted induced subgraphs of G by their vertex sets.)
.
tree, look for minimal clique cutsets for the corresponding subgraphs of G so that the new vertices can be labeled with a pair consisting of the name of the subgraph and the name of the clique cutset. Then continue this process, making a vertex in the tree a leaf whenever the subgraph of G to which it corresponds contains no clique cutset. Give the label of such a leaf a dash, ‘-’,for a second component. If in using this notation it should happen that some tree vertex denofed ( H , B), say, turns out to be a leaf, interpret ‘B’as a dash. We call this tree a clique cutsef tree for G. In general, of course, a graph has several clique cutset trees. The following observation will be needed to show that the algorithms given in later sections are polynomial.
Lemma 1 (Gavril [lo]). Any clique cufsef tree T ( G )for a graph G has at most n2 vertices, where n is the number of vertices of G. Proof. Note that each nonleaf of T ( G )has at least two children, so that for each tree vertex other than the root, there is at least one other tree vertex with the same parent. Suppose that tree vertices ( H , , A , )and (H,,A,),which might be leaves, are both children of (23, A). Then there is an ordered pair of nonadjacent vertices (a, u , ) in G such that 0, E H, - A and u, E H, -A, if i. To each vertex
Solving certain graph recognition and optimization problems
285
( H , , A , )of T ( G )other than the root, associate such an ordered pair (v,,u,). No ordered pair can possibly be associated with more than one tree vertex. Hence the number of vertices in T ( G ) is at most one more than twice the number of nonedges in G, so n 2 is certainly an upper bound for the size of T ( G ) . 0 Another useful fact about clique cutset trees is given in the following Lemma. Lemma 2 (Gavril [lo]). If (H’,A ’) is a vertex in a clique cuset tree T (G ) of a graph G, then A ’ is a clique curset of G as well as of HI. Proof. Note that if (H’,A ’ ) is a child of ( H ,A ) in T ( G ) ,then any clique cutset of H‘ is a clique cutset of H in G also. 0 We now give a brief description of the algorithm of Whitesides [19] for finding whether a graph has a clique cutset in time O(n3).See [19] for further details, including a proof of correctness. Algorithm 1 (Whitesides [19]). Finding whether a graph G ( V ) has a clique cutset. Step (0). Look for a chordless cycle of length at least four. If no such cycle exists, then either it’s easy to find a clique cutset, or G must be a clique and has no clique cutset. If the desired cycle is found, denote its vertex set by S. Step (1). Find a component C of G ( V - S ) , and then find the set R consisting of each vertex in S that has at least one neighbor in C. (Since G ( V ) is connected, R is not empty.) Step ( 2 ) . Search for a pair of nonadjacent vertices u and u in R. If no such pair exists, R must be a clique cutset of G(V), as R separates C from the nonempty set S - R in G ( V ) . Step (3). Find a chordless path P whose end vertices are u and u and whose intermediate vertices all lie in C. Step (4). Adjoin the intermediate vertices of P to S. Step (5). Repeat steps (1)-(4) until either a clique cutset has been found or S has been enlarged to all of V, in which case G ( V ) has no clique cutset. Now it is easy to give the main result of this section. Theorem 1. A clique cutset tree can be constructed for a graph G ( V ) with n uertices in O ( n 5 )time. Proof. By Lemma 1, there are at most n * vertices in such a tree. By Algorithm 1, the children of each tree vertex can be found in O(nJ) time. 0
286
S. H. Whitesides
The time complexity given in Theorem 1 may be too pessimistic. Our main point, however, is that a clique cutset tree can be found in polynomial time.
3. Reconstructing graphs from primitive subgraphs
The purpose of this section is to show how to reconstruct a graph G in O ( n S ) time by a sequence of ‘gluing’ operations. Each of these operations consists of adjoining a ‘primitive’ subgraph to a partial reconstruction R of graph G. The primitive subgraph is adjoined to R at some clique of G that is contained in both. By a primitive subgraph, we mean a connected subgraph that has no clique cutset. We will later make use of this reconstruction process to find, given some h-family F of graphs, polynomial time solutions to problems (I)-(IV) of Section The basic idea of the reconstruction algorithm is simple. Begin with a primitive subgraph Gf G that is associated with a leaf of a clique cutset tree T ( G ) for G, and then one by one, adjoin the other primitive subgraphs associated with the leaves of T ( G ) . The order in which this is done is such that the partially reconstructed graph R and the primitive subgraph being adjoined always have a clique in common. At each stage, R is connected. Hence the addition of a new primitive subgraph never connects several components at once. The reconstruction algorithm will have a step in which the leaves of a clique cutset tree are numbered. These numbers will give the order in which the primitive subgraphs associated with the leaves are to be adjoined to the partial construction R of G. In this context, we will say that a maximal subtree with root r in T ( G )has been numbered if, and only if, all its leaves have been assigned numbers. Algorithm 2. Reconstruction of a graph G from primitive subgraphs. Step (0). Check whether G is primitive. If so, stop. Step (1). Construct a clique cutset tree T ( G ) for G. Step (2). Number T ( G )by the following procedure. (See Fig. 2.)
Fig. 2. A numbering for the leaves of the tree T ( G )of Fig. I. The numbers on the edges represent the order in which the Algorithm visits the children. The numbers on the leaves represent the ‘gluing’ order assigned to the leaves.
Solving certain graph recognition and optimization problems
287
(i) Initialize variable r to be the root of T ( G )and counter i to 1. Next repeat (ii) and (iii) below until the root of T ( G ) has been numbered. (ii) If all children of r = ( H ,A ) have been numbered, mark r as numbered; reset r to the parent of r (stop if r is the root of T ( G ) ) ;return to (ii). (iii) Otherwise, choose an unnumbered child c of r. If c is a leaf of T ( G ) :label c with the cutset A associated with r ; label c with the current value of i ; mark c as numbered; increment i ; return to (ii). If c = (HI,A ’ ) is not a leaf: find the (unique) leaf ( H ” ,- ) such that ( H ” ,- ) is a descendant (although not necessarily a child) of ( H ’ , A ’ )and such that the cutset A associated with r belongs to H ” ; label (H”,- ) with A and with the current value of i ; mark ( H ” ,- ) as numbered; increment i ; reset r to the parent of (H”,-); return to (ii). Step (3). Find the leaf (H, - ) numbered 1 during step (2), and set the initial partial reconstruction R of G equal to H. Then process the remaining leaves of T ( G )from the smallest numbered leaf to the largest. To process a leaf ( H , - ) labeled with cutset A during step (2), adjoin H to R by identifying the copy of A belonging to H with the copy of A belonging to R. (See Fig. 3.) In step (2), note that each time the program returns to (ii), another node of T ( G ) has been marked as numbered. Also note that a node is not marked as numbered until all of its descendants have been so marked. Consequently step (2) terminates after I T ( G ) (iterations with a numbering of the leaves of T ( G ) . It is easy to show inductively that after the first leaf has been numbered, each time the program returns to step (2ii), the cutset associated with the current value of r belongs to the union of the graphs corresponding to the leaves of T ( G ) that have already been numbered. Also note that if a leaf is numbered while r = ro = (Ho,A@), then the leaf contains the cutset A. associated with ro.
4. Solving optimization problems in graphs
In this section, we show how the optimization problems (11)-(VI) listed at the end of Section 1 can be solved in polynomial time once these problems can be solved in polynomial time for the primitive (i.e., clique cutset-free) members of an h-family F of graphs. We begin with problems (11)-(IV), €or which the property of perfection is immaterial.
Theorem 2. Let F be an h-family of graphs, and let F’ denote the set of its primitive members. Suppose there is a polynomial time algorithm for finding a maximum weight clique for graphs in F’. Then there is a polynomial time algorithm for finding a maximum weight clique for graphs in F.
288
S.H. Whitesides
Fig. 3. Reconstructing the graph G of Fig. 1 according to the gluing order shown in Fig. 2.
Proof. Let G be a graph in F, and suppose that G has n vertices. Then according to Theorem 1, a clique cutset tree T ( G ) can be constructed using Algorithm 1 in O ( n 5 )time. By Lemma 1, there are at most n z leaves in T ( G ) ,so maximum weight cliques for each of the subgraphs of G associated with a leaf of T ( G )can be found in time O ( n 2 p ( n ) ) where , p ( n ) is a polynomial bound for the time to find a maximum weight clique for an n-vertex primitive graph in F ’ . Suppose that (HIA ) is a nonleaf vertex of T ( G ) ,and denote its children, some of which may be leaves, by (Hl,Al),. . . ,(Hk, Ak). For 1 d i S k, let Ci be a maximum weight clique for Hi. Then the C, of largest weight among these cliques is a maximum weight clique for H. Consequently, finding a maximum weight clique for G is simply a matter of processing T ( G )from its lowest level
Solving certain graph recognition and optimization problems
289
up to its root, selecting at each nonleaf vertex the maximum weight clique found among the graphs associated with its children. This can be done in O ( n 2 )time, so a maximum weight clique for G can be found in polynomial time. 0 The argument just given in the proof of Theorem 2 applies to finding a minimum coloring, so we have the following result.
Theorem 3. Suppose there is a polynomial time algorithm for,finding a minimum coloring for graphs in the set F' of primitive members of some h-family Fof graphs. Then there is a polynomial time algorithm for finding a minimum coloring for all graphs in F. To find a maximum weight stable set, we use a result of Boulala and Uhry which we describe in the next Remark.
Remark 1 (Boulala and Uhry [2]). Let G( V, E ) be a graph whose vertices are weighted with nonnegative integers, and suppose that G is composed of two V2,E2),where V, U V, = V, El U Ez = E, and subgraphs GI(V1,El) and G2( V, n V2 is the vertex set of a clique A of G. Consider the stable sets of G I that contain no vertex of A, and let 3, be one of maximum weight. Also, consider for each vertex a in A the stable sets of GI that contain a, and let S, be one of maximum weight. Now change the weights on the vertices in A : Assign to each vertex a the new weight max(0, w ( S , ) - ~ ( 3 , )Let ) . ST be a maximum weight stable set for Gz once the weights in A have been changed. Finally, denote by S the following set of vertices of G: If ST n A is empty, let S = ST U 3,; otherwise, let S = S : U S,, where S2 n A =a. According to Remark 4.3 of Boulala and Uhry [2], S is a maximum weight stable set of G with respect to the original weighting of the vertices, and its weight is w (ST) w (gl).
+
Now we are ready to find maximum weight stable sets using the clique cutset tree method.
Theorem 4. Suppose there is a polynomial time algorithm for ,findinga maximum weight stable for graphs in the set F' of primitive members of some h-family F of graphs. Then there is a polynomial time algorithm for finding a maximum weight stable set for all graphs in F. Proof. Let G be a graph in F. Using Algorithm 2, reconstruct G from primitive subgraphs, gluing one primitive subgraph at a time to the most recent partial reconstruction of G. At each step after the initial one, let the graph GI of Remark 1 be the primitive subgraph just glued onto the partial reconstruction of
S.H. Whitesides
290
G, and let G2be the graph onto which GIwas glued. Thus the graph GI,in which several maximum weight stable set problems must be solved in order to compute new weights for the vertices in the clique, is primitive. Hence, and the S,’s can be found in O ( p ( n ) n ) time, where p ( n ) is a polynomial bound on the time needed to find a maximum weight stable set for a primitive graph with n vertices. Then from these stable sets, compute new weights for the vertices of A = G ,n G2,and recursively solve the problem of finding the stable set S: of Remark 1. Since there are fewer than n 2 primitive graphs involved in the reconstruction process, a maximum stable set for G can be found in polynomial time. 0
s,
Having solved optimization problems (II)-(IV) for families of graphs whose primitive members have solutions to these problems, we turn to families of graphs whose members are all perfect. A graph G is perfect if, and only if, each induced subgraph H of G has chromatic number equal to the size of the largest clique in H. See Berge [l] for a discussion of these graphs. Chvatal[6] has shown that precisely for the perfect graphs can the problem of finding a maximum weight stable set be formulated as the following linear programming problem: n
maximize
2 w,x, subject to
i=l
Kx s 1, x bO,
where K is the maximal clique versus vertex incidence matrix, and the w,’s are the weights for the vertices.
If the weights w, are nonnegative integers, solving the dual LP problem, minimize x y , subject to
yK
2 w,
y 301
amounts to finding a minimum weight clique cover. Since the complement of a perfect graph is also perfect (Lovisz [15]), the problem of finding a maximum weight clique can also be formulated as an LP problem. Solving the dual of this problem amounts to finding a minimum weight cover by stable sets. These facts, in conjunction with Theorems 2 and 4, give us a way to solve problems (V)and (VI).
Corohry to Theorem 4. Suppose the graphs in the h-family F of Theorem 4 are
Solving certain graph recognition and optimization problems
291
all perfect. Then there is a polynomial time algorithm for finding a minimum weight clique couer for graphs in F. Proof. Let G be a graph in F, and reconstruct G from primitive subgraphs, using Algorithm 2 as before. At each stage of the reconstruction, let G1 be the last primitive piece added, and let GZbe the partial reconstruction to which GI was adjoined. Let A denote the clique G1n Gz, and change the weights on the vertices of A as follows. To each a E A, assign a new weight max (0, w ( S o )w ( s ) } ,where, of all the stable sets of GI containing a, So is one of maximum weight and of all the stable sets of G1 containing no vertex of A, is one of maximum weight. Then, recursively find a minimum weight clique cover for G , once these new weights have been assigned. By LP duality, the weight of this cover will be the weight of a maximum weight stable set for G,. Next, change the weights on the vertices of A again. This time, assign to each a E A the nonnegative weight w ( a )- ( w (So)- w where the weights are the original ones. This makes all maximum weight stable sets of G1 containing a have weight equal to w ( s ) . Now find a minimum weight clique cover for G1with these new vertex weights. The weight of this cover will be equal to w (3). Finally, take the union of the clique covers for GI and Gz. This union is indeed a clique cover for G. The cover has minimum weight because its weight is equal to that of a maximum weight stable set for G. Again, the subgraph in which multiple subproblems must be solved is primitive, so the algorithm runs in polynomial time. 0
(s)),
Of course, a proof analogous to the one just given holds for finding a minimum weight cover by stable sets, so Theorem 2 has a similar corollary.
5. Applications to clique separable and Gallai graphs A chord (u, u ) of a cycle in a graph is a pair of adjacent but nonconsecutive vertices of the cycle. Two chords (u, u ) and (x, y ) are said to cross if u, x, u and y are distinct and appear in this order on the cycle. In [8], Gallai studied the family of graphs defined by the property that odd cycles of length at least five always have a pair of noncrossing chords. The chordal (sometimes called ‘triangulated’) graphs of Dirac [7], for example, satisfy this condition; these graphs, which have been studied algorithmically by Gavril [9] and Rose [17], are defined by the property that every cycle of length at least four has a chord. Gallai gave several characterizations of his graphs and showed them to be a-perfect. Suranyi [18] later gave a simpler proof of the same result. A review of Gallai’s graphs, called ‘i-triangulated’, appears in Berge [l].
S. H. Whitesides
292
In order to describe a result of Gallai’s that we use, we need to define two special types of graphs. We say that a graph G is a type 1 graph if its vertices can be partitioned into two disjoint sets V, and V2 such that G(VJ is a connected bipartite graph, G(V2) is a clique, and every vertex of Vl is adjacent in G to every vertex of V,. Here, one of V1, V2 may be empty. We say that G is a type 2 graph if it is a complete k-partite graph for some k.
Remark 2. Problems (I)-(VI) can be solved in polynomial time for graphs that are type 1 or 2; also, it is easy to see that these graphs are perfect.
To see that a graph G can be quickly tested for being type 1, find all vertices of degree n - 1, where n is the number of vertices of G, and then check that the induced subgraph on the remaining vertices is connected and bipartite. If this induced subgraph is a stable set, a vertex of degree n - 1 can be added to it to make it connected and bipartite. Suppose that G has been found to be type 1. Then talung the vertex of maximum weight on each side of the bipartite subgraph of G together with all the vertices of degree n - 1 gives a maximum weight clique for G. A maximum weight stable set can be obtained by choosing the heavier side of the bipartite subgraph of G or the vertex of degree n - 1 that has maximum weight, as appropriate. It is easy to check whether a given graph is type 2 because in type 2 graphs, nonadjacency is an equivalence relation on the vertices. Burlet and Uhry [ 5 ] give polynomial time algorithms for solving problems (II), (IV), (V), and (VI) for the family of ‘parity’ graphs; this family includes the bipartite graphs. Using this fact, it is easy to see that these problems have polynomial time solutions for the type 1 graphs. It is easy to check that the remaining problems concerning type 1 and type 2 graphs can be solved in polynomial time. A result of Gallai’s that we need is given in the next Remark.
Remark 3 (Gallai IS]). A Gallai graph that is primitive must be a type 1or type 2 graph.
Gavril [ 101 calls a graph clique separable if each of its primitive subgraphs is type 1 or type 2. The Gallai graphs are a strict subfamily of the clique separable graphs as Fig. 4 shows. The clique separable graphs form quite a large family. Note, for example, that since any chordal graph that is not a clique must have a clique cutset (Hajnal and Surinyi [14]), the chordal graphs are certainly Gallai and hence clique separable. Also, connected bipartite graphs are clique separable, as they are tvue 2.
Solving certain graph recognition and optimization problems
293
2 Fig. 4. A clique separable graph G that is not Gallai (i-triangulated). Vertices 3 , 4, 6, 8, and 7 constitute an odd cycle that has only (3,6) as a chord. Nevertheless, A = (3,5,6} is a clique cutset, where the children of G produced by A are GI = 11,. . . ,6} and G, = {3,5,6,8}. Both GI and G, are complete 3-partite graphs.
It is easy to see that clique separable graphs are perfect, as they can be reconstructed from primitive perfect pieces by gluing these onto cliques of the partial reconstructions, an operation that preserves perfection. Now w e are ready to apply the results of Section 4.
Theorem 5. Let F be the family of clique separable graphs. Then for F there are polynomial time algorithms for solving problems (I)-(VI). Proof. To test whether a graph G belongs to F, build a clique cutset tree T ( G ) , and test that each subgraph of G associated with a leaf of T ( G )is type 1 or 2 . All this can be done in polynomial time, by Lemma 1, Theorem 1, and Remark 2 . If G is known to be clique separable, then by Remark 2 and Theorems 2, 3, and 4 and their corollaries, there are polynomial time algorithms for solving problems (11)-(VI). 0
Theorem 5 satisfies Gavril’s [lo] request for polynomial time solutions to the minimum clique cover and maximum stable set problems €or clique separable graphs. He solved the recognition problem and the problem of finding a minimum coloring and a maximal solution. His algorithm depends on finding clique cutsets with special properties. Algorithm 1 makes these properties unnecessary, and thus simplifies the details of his algorithm. Gavril also asked for a fast recognition algorithm for the family of graphs whose primitive members can be type 1, type 2, or ‘transitive orientable’ (‘comparatability’) graphs. Such an algorithm is easy to provide, as Gilmore and Hoffman [ll]and Pnueli, Lempel and Even [16] have given fast recognition algorithms for these latter graphs.
6. Gallai graphs Since the family of clique separable graphs strictly contains the Gallai graphs, we already have solutions to (11)-(IV) for Gallai graphs. Gavril [lo] asked for a
294
S. H.Whitesides
polynomial time membership test for this strict subfamily, and we are about to give such a test. We need the next theorem in order to do this.
Theorem 6 . Suppose that G is a clique separable graph, and suppose that C is an elementary cycle in G of odd length at least 5. Then either C has a pair of noncrossing chords or C has exactly one chord. In the latter case, the chord is ‘short’,meaning that it joins two vertices that are two steps apart around the cycle. Proof. Suppose C fails to have a pair of noncrossing chords. Then in any clique cutset tree T ( G )for G, C is not a subgraph of a graph associated with a leaf, because type 1 and type 2 graphs are Gallai. Hence there must be a tree vertex ( H , A ) in T ( G )with children (HI,Al),. . .,(Hkr Ae),where possibly some of the (H,,A,)%are leaves, such that C is not a subgraph of H, for any i, 1 S i S k. In fact, there must be i and j , 1 S i f j ik, such that C contains a vertex of H, - A and a vertex of H, - A. Of course, C must then contain at least two vertices of A and hence a chord. Cycle C cannot contain more than two vertices of A because it fails to have a pair of noncrossing chords. Hence C is a subgraph of H, U H,. One of C fl H,, C n H, must be an induced chordless odd cycle and the other, a chordless even cycle. The chordless odd cycle must have length 3, since we already know that longer odd cycles always have chords. Hence the chord given by C n A is ‘short’. 0
Now we can give a membership test for Gallai graphs. It runs in O ( n S )time. Algorithm 3. Testing membership of a graph G in the family of Gallai graphs. Step (1). Check whether G is clique separable, using Algorithm 2 to construct a clique cutset tree T ( G )for G. Step (2). If G is clique separable, process each nonleaf vertex of T ( G ) as follows. Suppose the children of (H,A ) are ( H I ,A1),. . . , ( I l k ,A k ) , some of which may be leaves. Consider pairs of the form (e, u ) , where e is an edge of A ; is a vertex of H, - A for some i, 1 s i s k, and is adjacent to both endpoints x and y of e. For each ifi, 1 S j s k, construct a graph S, as follows. Begin with S, = Hi- A, but then remove all vertices adjacent to both x and y . Finally, add in x and y to S,, putting in all induced edges except edge e.
Next, test whether there is a path from x to y in Sfi If there is, G is not Gallai: The existence of a path from x to y in S, implies the existence of a chordless path from x to y. If this chordless path has an odd number of vertices, it forms with e a chordless path of length at least 5 in G, as x and y have no common neighbors in S,. If the chordless path has even length, then in G it forms with the vertex II of H, a cycle of odd length at least 5 that has only one chord. Repeat step 2 until all nonleaf vertices have been processed. If the processing
Soloing certain graph recognition and optimization problems
295
of these vertices did not yield a proof that G fails to be Gallai, then by Theorem 5 , G is indeed a Gallai graph. The total time to process all the nonleaf vertices of T ( G )is at most O ( n 5 ) . Burlet [3] has extensively studied Gallai graphs from an algorithmic point of view and has independently given a recognition algorithm for them as well as algorithms for solving the Optimization problems. His algorithms depend on finding clique cutsets with special properties, which are not necessary from our point of view. Burlet works with simplicia1 clique cutsets, meaning clique cutsets that have no edge on a chordless even length cycle. In the next theorem, we again illustrate the usefulness of the clique cutset idea as a tool for structure elucidation by showing in a different way from Burlet that these simplicia1 cliques exist. Also see Burlet and Fonlupt [4].
Theorem 7 . Suppose that G is a Gallai graph and that T ( G ) is a clique cutset tree for G that has been constructed subject to the following rules: (i) At each stage of the construction, choose a minimal cutset. (ii) Construct children for a tree uertex ( H , A ) if, and only if, H fails to be a 2-connected graph of type 1 or 2. Then if (H,A ) is a tree vertex whose children are all leaves, A is a simplicial clique of G. Proof. Suppose by way of contradiction that ( H , A ) is a tree vertex whose children ( H I ,-), . . . ,(Hk,- ) all are leaves, but that the clique cutset A associated with vertex ( H , A )contains an edge e whose endpoints x and y are consecutive vertices on a chordless even length cycle C. Then A n C = e. Let K = A U C, and note that K has no clique cutset: Such a cutset would have to separate a vertex u E A - e from a vertex u E C - e. But C U u is an odd cycle of length at least 5 and so must have another chord in addition to e, because G is Gallai. The existence of this additional chord prevents the separation of u from u by a clique cutset. Now imagine ‘percolating’ K through T ( G ) . That is, observe which graphs associated with tree vertices have K as a subgraph. Since K has no clique cutset, it must be a subgraph of one of the children Hi of H. In other words, the cycle C belongs to Hj. Note that in each child H, of H,j # i, any vertex u adjacent to both x and y must in fact belong to A. This means that if H,, which is type 1 or type 2, contains a chordless even cycle in which x and y appear consecutively, then H, must be a bipartite graph completely connected to the clique A - e . Consequently, there must be some Hi, j # i, in which x and y do not appear consecutively on a chordless cycle of even length. Otherwise, H itself would consist of a connected bipartite graph completely connected to the clique A - e, and H would be 2-connected. This contradicts the properties of T ( G ) .
296
S. H.Whitesides
Let H, be a child of H in which x and y do not appear as consecutive vertices on a chordless cycle of even length. This means that if H, is a type 2 graph, one of x and y must be the lone vertex in its class, and so this vertex is adjacent to all other vertices of H,. If H, is type 1, then again, one of x and y must be adjacent to all the other vertices of H,. Otherwise, one of x and y would be an articulation point for the bipartite piece of H, and would form with A - e a smaller clique cutset for H, contradicting minimality of A. Say x is the endpoint of e that is adjacent to all other vertices of H,. Then y is adjacent to no vertex of H, - A. This is because any common neighbor of x and v in H, would form with C and the chordless even cycle in H, a cycle of odd length at least 5 containing only one chord. But this means that x together with A - e is a clique cutset for H, contradicting the minimality of A. 0 7. Conclusions
We have seen that having a quick way to find clique cutsets proves polynomial time algorithms for testing membership in families of graphs whose primitive members can be recognized in polynomial time. We have also seen that certain optimization problems can be solved quickly in these families by using the clique cutset tree approach. As a application, we pointed out that for clique separable graphs there are polynomial time algorithms for all these problems. Also, we provided a recognition algorithm for Gallai graphs which is different from the one of Burlet [3] and Burlet and Fonlupt [4]. We have seen that the clique cutset tree approach can provide a powerful tool for structure elucidation. Again, we point out that this approach is in no way restricted to perfect graphs. We close with two questions, one in the realm of perfect graphs. First, can all perfect graphs be obtained from some collection of ‘primitive’ perfect graphs by a sequence of perfection preserving operations? The second question is simply a more general version of a question posed by Gavril [lo], who asked for a characterization of clique separable graphs in terms of conditions on cycles. We ask for choices of ‘primitive’ graphs giving rise to interesting families that can be defined in another way, by conditions on cycles, for example.
Acknowledgement The author wishes to thank V. Chvatal for numerous stimulating and enlightening discussions about this work. Among other things, he suggested that Theorem 4 might be true and pointed out several of the references. Also, the author wishes to thank the Computer Science Department at Cornell University for its hospitality during the academic year 1981-1982.
Soloing certain graph recognition and optimization problems
297
References [ 11 C. Berge, Graphs and Hypergraphs (North-Holland, Amsterdam, 1972). [2] M. Boulala and J.P. Uhry, Polytope des indipendants d’un graphe strie-parallkle, Discrete Math. 27 (1979) 225-243. [3] M. Burlet, Etude algorithmique de certaines classes de graphes parfaits, thesis, I’Universitk Scientifique et MCdicale de Grenoble and I’Institut National Polytechnique de Grenoble (1981). [4] M. Burlet and J. Fonlupt, Polynomial algorithm to recognize a Meyniel graph, Research Report 303, Laboratoire d’hformatique et de MathCmatiques Appliquees de Grenoble (1982). [ 5 ] M. Burlet and J. P. Uhry, Parity graphs (this volume, pp. 253-277). [6] V. Chvital, On certain polytopes associated with graphs, J. Comb. Theory, Ser. B 18 (1975) 138-154. (71 G.A. Dirac, On rigid circuit graphs, Abh. Math. Sem. Univ. Hamburg 25 (1961) 71-76. [8] T. Gallai, Graphen mit triangulierbaren ungeraden vielecken, Magyar Tud. Akad. Mat. Kutato Int. Kozl. 7 (1962) 3-36. [Y] F. Gavril, Algorithms for minimum coloring, maximum clique, minimum covering by cliques and maximum independent set of a chordaI graph, SIAM J. Comput. 1 (1972) 18&187. [lo] F. Gavril, Algorithms on clique separable graphs, Discrete Math. 19 (1977) 159-165. [ l l ] P.S. Gilmore and A.J. Hoffman, A characterization of comparability graphs and of interval graphs, Canad. J. Math. 16 (1964) 539-548. [12] M.C. Golumbic, Algorithmic Graph Theory and Perfect Graphs (Academic Press, New York, 1980). [13] M. Grotschel, L. Lovasz and A. Schrijver, The ellipsoid method and its consequences in combinatorial optimization, Combinatorica 1 (1981) 169-197. [14] A. Hajnal and J. Suranyi, Uber die auflosung von graphen in vollstandige teilgraphen, Ann. Univ. Sci. Budapestinensis 1 (1958) 113-121. [15] L. Lovasz, Normal hypergraphs and the perfect graph conjecture, Discrete Math. 2 (1972) 253-267 (this volume, pp. 2942). (161 A. Pnueli, A. Lempel and S. Even, Transitive orientation of graphs and identification of permutation graphs, Canad. J. Math. 23 (1971) 160-175. [17] D.J. Rose, Triangulated graphs and the elimination process, J. Math. Anal. Appl. 32 (1970) 597-609. [18] L. Surinyi, The covering of graphs by cliques, Studia Sci. Math. Hungar. 3 (1968) 345-349. [19] S.H. Whitesides, An algorithm for finding clique cut-sets, Inf. Process. Lett. 12 (1981) 31-32.
This Page Intentionally Left Blank
PART VI
OPTIMIZATION IN PERFECT GRAPHS
This Page Intentionally Left Blank
Annals of Discrete Mathematics 21 (1984) 301-323 @ Elsevier Science Publishers B.V.
ALGORITHMIC ASPECTS OF PERFECT GRAPHS Martin Charles GOLUMBIC* IBM Israel Scientific Center, Technion City, Haifa, Israel
1. Introduction Consider a collection C = ( c # }of courses being offered by a major university. Let T, be the time interval during which course c, is to take place. We would like to assign courses to classrooms so that no two courses meet in the same room at the same time. This problem can be solved by properly coloring the vertices of the graph G = (C,E) where cici E E e T, f l T, # 0. We may interpret each color as corresponding to a different classroom. The graph G is an interval graph, since it is represented by intersecting time intervals. This example is especially interesting because efficient, linear-time algorithms are known for coloring interval graphs with a minimum number of colors. (The minimum coloring problem is NP-complete for general graphs.) In this paper we will survey a number of topics in algorithmic graph theory which involve classes of perfect graphs. We will also discuss some recent applications of perfect graphs to computer science. The intention of this article is to provide an understanding of the main research directions which have been investigated and to suggest possible new areas of research. The sections of this paper are numbered to correspond with the chapters of the author’s book, Algorithmic Graph Theory and Perfect Graphs. The interested reader is referred to this book for further study.
2. The design of efficient algorithms
Algorithmic complexity analysis deals with the quantitative aspects of problem solving. It addresses the issue of what can be computed within a practical or reasonable amount of time and space by measuring the resource requirements exactly or by obtaining upper and lower bounds for them. Complexity is actually * This work was supported in part by the National Science Foundation under Grant No. MCS 78-83820. 301
302
M.C. Golurnbic
determined at three levels: the problem, the algorithm, and the implementation. Naturally, we want the best algorithm which solves our problem, and we want to choose the best implementation of that algorithm. Consider the problem of determining whether an undirected graph G is connected. A mathematically elegant solution is the following: G is connected if and only if I + M + M 2+ M 3+ + M"-'has no zero entries where M is the adjacency matrix of G, I is the identity matrix, and n is the number of vertices of G. However, using this theorem as an algorithm would require much more work (matrix multiplication and addition) than is actually needed to test connectivity. A better way would be to traverse the edges of the graph. The following algorithm will test connectivity and find a spanning tree efficiently.
--
Standard Spanning Tree Algorithm (SST) Step I. Start with a tree T consisting of one arbitrary vertex and no edges. Step ZZ. If T contains all the vertices of G, then STOP [ T is a spanning tree]. Otherwise, do step 111. Step 111. Add to T an edge ( x , y ) which joins a vertex y not yet in T to a vertex x already in T. If no such edge exists, then STOP [there is no spapning tree; G is not connected]. Otherwise, go to step 11. In Step 111 of our algorithm there may be several edges (x, y ) eligible to be added to T. We call such an edge a candidate edge. Various priorities can be established to guide the choice of candidates, and each priority will yield a slightly different algorithm. If candidates are stored in a queue, then SST gives a breadth-first search (BFS) of G. Storing candidates in a stack SST does a depth-first search (DFS). If the edges have costs associated with them, and if the candidate with minimum cost is always chosen, then SST produces a minimum cost spanning tree (MST).Similarly, shortest path algorithms and critical path algorithms can also be designed by adapting SST with a suitable priority for choosing candidates. The complexity of the spanning tree algorithm depends on how the graph is stored and whether anything special is done to the candidate edges. Table 1 summarizes these complexities. A graph problem is said to be linear in the size of the graph if it has an algorithm which can be implemented to run in O ( n + e ) steps on a graph with n vertices and e edges. Thus testing connectivity is a linear graph problem. This is usually the best that one could expect for any nontrivial graph problem since every vertex and every edge would probably have to be examined at least once. A problem is called polynomial if it has an algorithm which can run in O @ ( n ) ) steps where p is a polynomial function.
Algorirhmic aspects ojperfecr graphs
303
Table 1 Complexity of the standard spanning tree algorithm Candidates in a Stack (DFS) Queue (BFS) Reverse heap (MST)
Adjacency matrix stored as an array
Adjacency sets stored as lists or sequentially
O(nZ)
O ( n+ e )
O(n2)
O(n + e )
+
O(n2 e log e )
O ( n + e loge)
The algorithmic graph problems that we will examine in this survey paper include recognizing various classes of perfect graphs and finding minimum colorings, minimum clique covers, maximum cliques, and maximum stable sets. We will be particularly interested in special purpose polynomial algorithms designed to solve these problems for particular classes of perfect graphs. The reason such algorithms are important is that for arbitrary graphs these last four problems are NP-complete, that is, they are in a large class of problems which all currently require an exponential amount of running time and which are all related in such a way that if any one of them could be solved in polynomial time, then so could all problems in this class.
3. Perfect graphs An undirected graph G equivalent conditions:
= ( V , E ) is
perfect if it satisfies any of the following
w(GA)= , y ( G A )
(for all A C V ) ,
(PI 1
a(GA)= e(GA)
(for all A C V),
(P2)
o ( G A ) a ( G A ) a l A1 (for all A C V).
(P3)
The equivalence of (PI)-(P3) is known as the Perfect Graph Theorem. An open question whose solution has eluded researchers for two decades is to prove or disprove the following conjecture of Claude Berge.
Strong Perfect Graph Conjecture (SPGC). An undirected graph G is perfect if and only if in G and in G every odd cycle of length 2 5 has a chord. Although proving the SPGC seems to be a mathematical rather than an algorithmic problem, it does raise an interesting algorithmic question.
304
M.C. Golumbic
Is there a polynomial algorithm which recognizes whether or not an undirected graph G has an odd chordless cycle of length 3 5 ? We have no answer to this question. However, if there is such an algorithm and if the SPGC is true, then it would answer another open question: Is there a polynomial algorithm which recognizes whether or not an undirected graph G is perfect? In a very recent paper, Grotschel, Lovasz and Schrijver [16] have shown that the ellipsoid method of solving linear programming problems can be applied to obtain a polynomial algorithm to find maximum stable sets and minimum colorings for perfect graphs. Also, since G is perfect if and only if its complement G is perfect, this same approach can be used to find maximum cliques and minimum clique covers. The major importance of this result is that it generalizes what had been known for certain classes of perfect graphs. Although the complexity of the algorithm is polynomial, it may not be practical to implement. As the authors point out, it is not intended to compete with the special purpose algorithms designed to solve these problems for interval graphs, comparability graphs, triangulated graphs, and other classes of perfect graphs which so often arise in applications.
4. Triangulated graphs
An undirected graph G is called triangulated if every cycle of length strictly greater than 3 possesses a chord, that is, an edge joining two nonconsecutive vertices of the cycle. In the literature, triangulated graphs have also been called chordal, rigid -circuit, monotone transitive and perfect elimination graphs. A vertex x of G is called simplicial if its adjacency set Adj(x) induces a complete subgraph of G, i.e., Adj(x) is a clique (not necessarily maximal). Dirac [4], and later Lekkerkerker and Boland [18],proved that a triangulated graph always has a simplicial vertex (in fact at least two of them), and using this fact Fulkerson and Gross [7] suggested an iterative procedure to recognize triangulated graphs based on this and the hereditary property, namely, repeatedly locate a simplicial vertex and eliminate it from the graph, until either no vertices remain and the gruph is triangulated or at some stage no simplicial vertex exists and the gruph is not triangulated. The correctness of this procedure is given in Theorem 4.1. Let us state things more algebraically. Let G = (V, E) be an undirected graph and let cr = [ u , , v 2 , .. . ,urn]be an ordering of the vertices. We say that u is a perfect vertex elimination scheme (or perfect scheme) if each ui is a simplicial vertex of the induced subgraph G{”,.
305
Algorithmic aspects of perfect graphs
In other words, each set
I
Ai = {vj E Adj(vi) j > i } is complete. For example, the graph G1 in Fig. 4.1 has a perfect vertex elimination scheme u = [ a , g, b, f , c, e, d ] . It is not unique; in fact G, has 06 different perfect elimination schemes. In contrast to this, the graph G2 has no simplicial vertex, so we cannot even start constructing a perfect scheme - it has none.
r
GI
G2
Fig. 4.1. Two graphs, one triangulated and one not triangulated.
Theorem 4.1. A n undirected graph G is triangulated if and only if it has a perfect vertex elimination scheme. Moreover, any simplicial vertex can start a perfect scheme. A proof of this theorem can be found in Golumbic [ 131 and uses the following important lemma of Dirac [4].
Lemma 4.2. Every triangulated graph G = ( V ,E ) has a simplicial vertex. Moreover, if G is not a cliqae, then it has two nonadjacent simplicial vertices. From Lemma 4.2 we learn that the Fulkerson-Gross recognition procedure affords us a choice of at least two vertices for each position in constructing a perfect scheme for a triangulated graph. Therefore, we can freely choose a vertex v, to avoid during the whole process, saving it for the last position in a scheme. Similarly, we can pick any vertex vn-l adjacent to u, to save for the ( n - 1)st position. If we continued in this manner, we would be constructing a scheme backwards! This is exactly what Rose, Tarjan, and Lueker [23] have done in order to give a linear time algorithm for recognizing triangulated graphs using lexicographic breadth-first search. This method can be found in Golumbic [13]. In an unpublished work, R. E. Tarjan [26] has shown that maximum cardinality search (MCS) can also be used to recognize triangulated graphs.
306
M. C. Golumbic
Algorithm 4.1. Maximum cardinality search. Input: The adjacency sets of an undirected graph G = (V, E). Output: An ordering v of the vertices. Method: The vertices are numbered from n to 1 in the order that they are selected in line 3. This numbering fixes the positions of an elimination scheme u. For each unnumbered vertex x, the label of x will consist of the number of numbered vertices adjacent to x. The vertices can then be ordered according to their labels. Ties are broken arbitrarily. The algorithm is as follows: 1. 2. 3. select:
4. 5. update: 6.
assign the label 0 to each vertex; for i + n to 1 by - 1 do pick an unnumbered vertex u with largest label; u ( i ) t u ; [this assigns to o the number i ] for each unnumbered vertex w E A d j ( o ) do add 1 to label ( w ) ; end end
The fact that maximum cardinality search can be used to recognize triangulated graphs is demonstrated by the next theorem.
Theorem 4.3. An undirected graph G = ( V ,E ) is triangulated if and only if the ordering u produced by Algorithm 4.1 is a perfect vertex elimination scheme.
Proof. If G has only one vertex, then the proof is trivial. Assume that the theorem is true for all graphs with fewer than n vertices and let u be the ordering produced by Algorithm 4.1 when applied to a triangulated graph G. By induction, it is sufficient to show that v = a(1)is a simplicia1 vertex of G. Claim. G may not contain a chordless path satisfying the property
j~
= [u , 0 1 ,u z , . . . , ok, w ] with
k
3
1
Suppose G contains such a path p, and choose p such that u - ' ( K ) is largest possible. Since u was numbered before -uk and since U k , but not u, is adjacent to w. there must be some vertex x such that u - ' ( u ) < a - ' ( x ) which is adjacent to u but not to u k . Let j be the largest index such that x is adjacent to u, where we let uo = u. Then the path p ' = [ x , u,, . .. ,u k , w ] must be chordless, since its only possible chord x w would give a chordless cycle of length 2 4 . If a - ' ( x ) < u - ' ( w ) , then p ' would satisfy (1) and contradict the maximality of a - ' ( u ) since &'(ok)< u - ' ( u ) < u - ' ( x ) . So it must be that a - ' ( w ) < u - ' ( x ) . But this implies
Algorithmic aspects of perfect graphs
307
that p" = [ w, u k , . . . ,u,, x ] satisfies (1) and also contradicts the maximality. It follows that no such path ,u can exist in G, which proves the claim. Now let u = ( ~ - ' ( land ) suppose that u is not simplicia]. Choose u, w E Adj(u) with uw $Z. E so that ( T - ' ( u ) < a - ' ( w ) .Then the path [u, u, w ] satisfies (l),which contradicts the claim. Therefore, u is simplicia1 and, by induction, cr is a perfect elimination scheme. The converse follows from Theorem 4.1. 0 The complexity of Algorithm 4.1 is linear in the size of G. One such efficient implementation is the following. Let Sibe the set of unnumbered vertices whose label is i, and let Si be represented by a doubly linked list. For each vertex we store its label i and a pointer to its position in the set Si. When a vertex u is numbered it is removed from its set, and we move each adjacent vertex w up by one set; this can be executed in O(1 +degree(u)) steps. Thus, the entire algorithm will be O(l VI + IE I). In order to use MCS to recognize triangulated graphs, we need an efficient method to test whether or not a given ordering cr of the vertices is a perfect elimination scheme. Such an algorithm is given in Rose, Tarjan and Lueker [23] and has complexity O(l V ( + ( E I). (See also Golumbic [13], pp. 88-91.)
Fast algorithms for the coloring, clique, stable set and clique couer problems on triangulated graphs Let G = (V, E ) be a triangulated graph, and let cr be a perfect elimination for G. It was first pointed out by Fulkerson and Gross [7] that every maximal clique was of the form {u}UA, where
I
A, = {x E Adj(u) cr-'(u)< cr-'(x)}.
However, some of these sets { u } U A, will not be maximal, and we would like to filter them out. This can be accomplished in order to find the chromatic number and maximal cliques of a triangulated graph in O ( I V I + I E I) time. The problem of finding the stability number (Y (G) of a triangulated graph and a clique cover of size a ( G )is solved by Gavril [ 8 ] .A linear implementation of his algorithm can be obtained by using techniques of Rose, Tarjan and Lueker ~31. Let u be a perfect elimination scheme for G = ( V, E ) . We define inductively a sequence of vertices y l , y2,. . . ,yt in the following manner: yl = ( ~ ( 1 )y,; is the first vertex in (T which follows yz-l and which is not in A,, U A, U . . . U Ay,-,;all vertices following y, are in A,, U . . . U A,,. Hence, V = {yl, y2,. . .,y l } U A,, u . . . U A,,. The following theorem applies. Theorem 4.4 (Gavril [8]). The set {y,, y2,. . . , y l } is a maximum stable set of G,
M.C. Golumbic
308
and the collection of sets Yi= {yi}U A, ( i = 1,2,. ..,t ) comprises a minimum clque cover of G.
Proof. The set {y,, y2,. ..,y I } is stable since if yjyi E E for j < i, then yi E A , which cannot be. Thus, a ( G ) at. O n the other hand, each of the sets Y, = { y i ) U A , is a clique, and so { Y1,. . . ,Y l } is a clique cover of G. Thus, a ( G )= B(G)= t, and we have produced the desired maximum stable set and minimum clique cover. 0
5. Comparability graphs
An undirected graph G = (V, E ) is a comparability graph if there exists an orientation (V, F) of G satisfying
FflF-'=0, F+F-'=E,
F2CF,
I
where F2 = { a c ab, bc E F for some vertex b } and F-' is the reversal of F. The relation F is a strict partial ordering of V whose Comparability relation is exactly E, and F is called a transitive Orientation of G (or of E). Comparability graphs are also known as transitively orientable graphs and partially orderable graphs. Examples of some comparability graphs can be found in Fig. 5.1. Let us see what happens when we try to assign a transitive orientation to the 4-cycle (Fig. 5.2(a)). Arbitrarily choosing ab E F forces us to orient the bottom edge toward b and the top edge toward d (for otherwise transitivity would be violated). These in turn force the remaining edge to be oriented toward d.
A A The A graph
The suspension bridge graph
Fig. 5.1. Transitive orientations of two comparability graphs.
309
Algorithmic aspects of perfect graphs
"Tt
alI
b
d
e
Fig. 5.2. Examples of forcing. The arbitrary choice of ab E F forces the other indicated orientations.
Applying the same idea to the graph in Fig. 5.2(b) we find that a contradiction arises, namely, choosing ab E F forces successively the orientations cb, cd, cf, ef, bf, ba. This graph is not a comparability graph. We now make the notion of forcing more precise. Define the binary relation r on the edges of an undirected graph G = (V, E ) as follows:
ab
a'b' iff
[
either a or
= a'
and b b ' E E,
b = b' and a a ' g E.
We say that ab directly forces a'b' whenever ab r a'b'. Since E is irreflexive, ab T a b ; however, a b f ba. The reader should not continue until he is convinced of this fact. The reflexive, transitive closure f * of r is easily shown to be an equivalence relation on E and hence partitions E into what we shall call the implication classes of G. Thus edges ab and cd are in the same implication class if and only if there exists a sequence of edges
ab
= aobo
r a l b l r . . . r akbk= cd,
with k t 0.
Such a sequence is called a r-chain from ab to cd, and we say that ab (eventually) forces cd whenever ab r* cd.
Examples. The graph G of Fig. 5.3 has 8 implication classes:
A l = { a b } , A , = { c d } , A , = {ac,ad, a e } , A., = {bc,bd, b e } ,
A;'={ba}, A;'={dc}, A;'={ca,da,ea}, A;'={cb,db,eb}.
Fig. 5.3.
M.C. Golumbic
310
On the other hand, the graph in Fig. 5.2(b) has only one implication class: A
= {ab, cb,
cd, cf, ef, bf, ba, bc, dc, fc, fe, f b } .
Let A be an implication class of an undirected graph, G, and let A = A U A - I denote the symmetric closure of A. It can be shown that if G has a transitive orientation F, then either F n A = A (F completely agrees with A ) or F n A = A - I (F completely disagrees with A ) and, in either case, A f l A-' = 0. The converse of this is also valid, namely, if A f l A-' = 0 for every implication class A, then G has a transitive orientation.
Remark. Many readers may wonder whether an arbitrary union of implication A, satisfying F fl F-' = 0 and F + F-' = E is necessarily a classes F = transitive orientation of G. The answer is no. As a counterexample, consider a triangle which has 8 = 23 such orientations two of which fail to be transitive.
u,
Methods for determining the exact number of transitive orientations t ( G )of a given undirected graph G have been developed by Shevrin and Filippov [24] and Golumbic [ 111, and a characterization of uniquely partially orderable graphs (i.e., t ( G )= 2) is given in Shevrin and Filippov [24] and Trotter, Moore and Sumner [27]. These results and others are discussed in detail in Golumbic [ 131. We shall now describe an algorithm for calculating transitive orientations and for determining whether or not a graph is a comparability graph. This technique is a modification of one first presented by Pnueli, Lempel and Even [20]. A discussion of its computational complexity will follow. Let G = (V, E ) be an undirected graph. A partition of the edge set E = 8 ,+ 8, + . . . + & is called a G-decomposition of E if B,is an implication class of &, . . . + Bkfor all i = 1,2,. . .,k. A sequence of edges [xlyl,x2, y2,. ..,x t y k ) is called a decomposition scheme for G if there exists a G-decomposition E = 8 , + +. . . + Bk satisfying x,y, E B, for all i = 1,2,. . .,k. In this section the term scheme will always mean a decomposition scheme. For a given G-decomposition there will be many corresponding schemes (any set of representatives from the B i ) . However, for a given scheme there exists exactly one corresponding G-decomposition. A scheme and G-decomposition can be constructed by the following procedure:
+
Algorithm 5.1. Decomposition Algorithm. Let G = (V, E ) be an undirected graph. Initially let i = 1 and E l = E. Step 1. Arbitrarily pick an edge e, = x,y, E E,. Step II. Enumerate the implication class B, of E, containing x,y,. Step III. Define E,,, = E, - b,.
Algorithmic aspects ofperfect graphs
Step IV. If E,,' = 0, then let k go back to step I.
=i
311
and STOP; otherwise, increase i by 1 and
Clearly, the Decomposition Algorithm yields a scheme [ x I y l , .. . ,x k y k ] and corresponding G-decomposition L?, + . + & for any undirected graph G. Moreover, if y,x, had been chosen instead of x,y, for some i, then B;' would replzce B, in the G-decomposition. Applying the algorithm to the graph in Fig. 5.3, the scheme [ac, bc, d c ] gives the G-decomposition for which Bl = A3, B2 = A4+ A ? and B, = A;'. In this example notice that although ba and bc were not r-related in the original graph, once B1 is removed they become r-related in the remaining subgraph and their implication classes merge. In general, it can be shown that each implication class of E,,, will be the union of either one or two implication classes of E,. The next theorem legitimizes the use of G-decompositions as a constructive tool for deciding whether an undirected graph is a comparability graph, and, if so, producing a transitive orientation. Proofs of this theorem can be found in Golumbic [ll] or Golumbic [13].
--
Theorem 5.1 (The TRO Theorem). Let G = (V,E ) be an undirected graph with G-decomposition E = 8, + . * + Bk. The following statements are equivalent: (i) G = (V, E ) is a comparability graph ; (ii) A n A = 0 for all implication classes A of E ; (iii) Bin B;' = 0 fur i = 1,. . . ,k.
+
Furthermore, when these conditions hold, B , . + B, is a transitive orientation of E. By combining the TRO Theorem with the Decomposition Algorithm, we obtain an algorithm for recognizing comparability graphs and assigning a transitive orientation. Algorithm 5.2. TRO Algorithm. Input: An undirected graph G = ( V ,E ) . Output: A transitive orientation F of edges of G if FLAG has final value 0, or a message that G is not a comparability graph if FLAG has final value 1. Method: The entire algorithm is as follows: initialize: i t l ; E, c E ; F t g ; F L A G t O ; I: arbitrarily pick an edge x,y, E E, ; 11: enumerate the implication class B, of E, containing x,y, ; if B, r 7B;' = 0 then add B, to F ;
312
M.C. Golumbic
else
F L A G c l ; [G is not a comparability graph]; 111: define E,+,+E, - 8, ;
IV: if E + ,= 0 then k t i ; STOP [F is a transitive orientation of GI: else i t i + 1; go to I; The sequence of arbitrary choices made in line I of the algorithm determines which of the many transitive orientations of G is produced by the algorithm. A different scheme may give a different transitive orientation. But, when you try out a few different schemes you will notice a remarkable phenomenon: No matter how the arbitrary choices for G are made, the number of iterations k will always be the same. This phenomenon is actually true for any graph G. A characterization of the underlying mathematical structure which causes it is given in Golumbic [ll], 1131. A more detailed version of Algorithms 5.1 and 5.2 will suggest how we may construct a G-decomposition and test transitive orientability of an undirected graph C = ( V , E ) in O ( 6 l E l ) time and 0 ( ( V l + l E J )space where 8 is the maximum degree of a vertex. Let G = ( V ,E ) be an undirected graph with vertices uI, u 2 , . . .,u,. In the algorithm below we use the function
[4
u,u,E E, u,u, has been assigned to Bk, CLASS(i, j ) = u,u, has been asigned to B i ' , u,u, E E has not yet been assigned, and ICLASS(i, j ) l denotes the absolute value of CLASS(i, j ) . if if if undefined if
Algorithm 5.3. Decomposition Algorithm (detailed version). Input: The adjacency sets of an undirected graph G = ( V ,E ) with vertices UI, u 2 , . . . , 0,. Output: A G-decomposition of the graph given by the final value of CLASS, and a variable FLAG which is 0 if the graph is a comparability graph and 1 otherwise. If the algorithm terminates with FLAG equal to 0, then a transitive orientation of G is obtained by combining all edges having positive CLASS. Method: The algorithm proceeds until all edges have been explored. In the k th iteration an unexplored edge is placed in Bk (its CLASS is changed to k). Whenever an edge is placed into Bk it is explored using the recursive procedure of Fig. 5.4 by adding to Bk those edges r-related to it in the graph Ek.(Notice that uiu, E Ek if and only if either (CLASS(I,j ) l equals k or is undefined throughout the k th iteration.) The variable FLAG is changed from 0 to 1 the first time a B, is found such that Bk n B i ' # 8. At that point it is known that G is
Algorithmic aspecrs of perfect graphs
313
not a comparability graph (by Theorem 5.1). The algorithm is as follows.
k+O; FLAGcO; initialize: for each edge v,v, in E do if CLASS(i,j) is undefined then do
k+k+l; CLASS(( j ) + k ; CLASS(j, i) + - k ; EXPLORE(i, j ) ; end; end : loop 1 :
loop 2:
procedure EXPLORE(( j): for each m EAdj(i) such that [mEAdj(j) or JCLASS(j,m ) l < k] do if CLASS(i, m ) is undefined then do CLASS(i, m)+ k ; CLASS(m, i)+- k ; EXPLORE(i, m ) ; end else if CLASS(i, m ) = - k then do CLASS(i, rn)+k; F L A G c l ; EXPLORE(i, m ) ; end end loop I for each m EAdj(j) such that [mPAdj(i) or lCLASS(i,m)l< k] do if CLASS(m, j ) is undefined then do CLASS(m, j ) + k ; CLASS(j, m )+ - k ; EXPLORE(m, j); end else if CLASS(m, j ) = - k then do CLASS(m,j)+k; F L A G c 1 ; EXPLORE(m, j); end end loop 2 return end Fig. 5.4.
Complexity analysis. We begin by specifying an appropriate data structure. The adjacency sets are stored as linked lists sorted into increasing order. The element of the iist Adj(i) which represents edge u,u, will contain j, CLASS(i, j ) , a pointer to CLASS(j, i), and a pointer to the next element on Adj(i). The storage requirement for this data structure is O(1 V ( +lE I), and the entire initialization of the data structure can be accomplished in linear time. The crucial factor in the analysis of our algorithm is the time required to access or assign the CLASS function. Consider the first loop of EXPLORE(i, j ) . Two temporary pointers simultaneously scan Adj(i) and Adj(j) looking for values of m which satisfy the condition in the for statement. This loop can be executed in O(d, d,) steps. The second loop is done similarly, hence the time complexity of EXPLORE( i, j ) is 0 ( d , + d, ).
+
314
M.C. Golumbic
In the main program, a pointer scans each adjacency list successively in the for loop implying a time complexity of O( IE I). Finally, the algorithm calls EXPLORE once for each edge or its reversal (both if their implication classes are not disjoint). Therefore, since
it follows that the time complexity for the entire algorithm (including preprocessing the input) is at most 0 ( 6 1 E I).
Coloring and other probiems on comparability graphs Suppose that G is a comparability graph, and let F be a transitive orientation of G.A height function h can be placed on V as follows: h ( u ) = 0 if u is a sink; otherwise, h ( u ) = 1 + max{h(w)(,uw E F}.The height function can be assigned in linear time using a recursive depth-first search, and it is a proper vertex coloring of G. The number of colors used will be equal to the number of vertices in the longest path of F, and since, by transitivity, every path in F corresponds to a clique of G, the height function will yield a coloring which uses exactly w ( G ) colors which is the best possible. Therefore, from the transitive orientation F we can assign a minimum coloring to G using the height function in O ( 1 V I + 1 E I) steps, and, at the same time, calculate a maximum clique of G. We will illustrate this by solving the more general problem of finding a maximum weighted clique of a comparability graph. (If all vertices have the same weight, then the problem is reduced to the usual problem of finding a clique of maximum cardinality.) In general the maximum weighted clique problem is NP-complete, but when restricted to comparability graphs it becomes tractable. Algorithm 5.4. Minimum Coloring and Maximum Weighted Clique of a Comparability Graph. Inpur: The adjacency sets of a transitive orientation F of a comparability graph G = (V, E ) and a weight function w defined on V. Output: A minimum coloring of G and a clique K of G whose weight is maximum. Method : We use a modification of the height calculation technique employing the recursive depth-first search procedure SEARCH in Fig. 5.5. To each vertex u we associate its COLOR and its cumulative weight W ( u ) which equals the weight of the heaviest path from u to some sink. A pointer is assigned to u designating its successor on that heaviest path. Once the cumulative weights are
315
Algorithmic aspects of perfecr graphs
assigned the clique K is calculated beginning the line labeled retrace. The algorithm is given in the form of a procedure.
retrace:
procedure MAXWEIGHT CLIQUE (V, F ) : for all u E V do if u is unsearched then SEARCH (V); end select y E V such that W(y) = max{ W ( u ) u E V); K -{y}; y +-POINTER(),); while y # A do K +-K U {y}; y +POINTER(y); end return K ; end
1
procedure SEARCH( 0 ) : if Adj(u) = 0 then do W ( u ) = w ( u ) ; POINTER(u)+A; COLOR(u)+O; end else do for all x E A d j ( u ) do if x is unsearched then SEARCH(x); end select y E Adj(u) such that W(y) = max{ W ( x ) x E Adj(u)}; W(u)+ w ( u ) + W(y); POINTER(u)+y; select z E Adj(u) such that COLOR(z) = max{COLOR(z)I z E Adj(u)}; COLOR(u)- 1 +COLOR(z); end return end
I
Fig. 5.5.
We conclude with an interesting polynomial-time method for finding a ( G ) , the size of the largest stable set of a comparability graph G. We transform a transitive orientation (V, F ) of G i n t o a transportation network by adding two new vertices s and t and edges sx and y t for each source x and sink y of F. Assigning a lower capacity of 1 to each vertex, w e initialize a compatible integer-valued flow and then call a minimum-flow algorithm. The value of the minimum flow will equal the size of the smallest covering of the vertices by cliques which in turn will equal the size of the largest independent set since every comparability graph is perfect. Such a minimum flow algorithm can run in polynomial time.
M.C. Golumbic
316
6. Split graphs
An undirected graph G = ( V , E ) is a split graph if there is a partition V = S + K of its vertex set into a stable set S and a complete set K. Since a stable set of G is a complete set of the complement G, and vice versa, G is a split graph if and only if its complement G is a split graph. Fd d e s and Hammer [6] have given the following characterization of split graphs. Theorem 6.1. Let G be a n undirected graph. The following conditions are rquiualen t : (i) G is a split graph, (ii) G and G are triangulated graphs, (iii) G contains no induced subgraph isomorphic to 2K2, C., or C,. An alternate characterization of split graphs in terms of degree sequences is the following result of Hammer and Simeone [ 171. Theorem 6.2. Let G = ( V ,E ) be an undirected graph with degree sequence d , 3 d z 3 .. . 2 d,, and let m = max{i d, z= i - 1). Then, G is a split graph if’ and only if
I
2: d, = m ( m
-
I)+
2
d,.
,=rn+l
,=I
Furthermore, if this is the case, the m vertices of largest degree wit1 be a maximum complete set of G. A simple recognition algorithm for split graphs can be designed by applying Theorem 6.2. If this is done, it can easily be seen that the complexity of recognizing split graphs is O ( n logn). The same complexity applies for the clique problem and the stable set problem on split graphs. However, the Hamiltonian circuit problem on split graphs is NP-complete. 7. Permutation graphs
Let r = [ r lm,2 , . . . , r n ]be a permutation of the numbers 1,2,. . . , n. We define the undirected graph G [ T ] = (V, E ) as follows:
v
={u1.u2
,..,, u,,]
and (u,, u , ) E E
iff
( i -j)(r;I- m y ’ ) < 0.
Algorithmic aspects of perfect graphs
317
Two vertices are joined by an edge if they occur out of their proper order reading the sequence T left to right (see Fig. 7.1).
Fig. 7.1. The graph G[4,1,3,5,2].
If we reuerse the sequence T,each pair of numbers which occur in the correct order in 7i- will now be in the wrong order, and vice versa. Thus, the permutation graph we obtain will be the complement of G [ T ] .This shows that the complement of a permutation graph is also a permutation graph. Another property of the graph G [ r ]is that it is transitively orientable. If we orient each edge toward its larger endpoint, then we will obtain a transitive orientation E For, suppose ( u i , u j ) E F and ( u j , ut)E F, then i < j < k and T;'> my' > TL', which implies that ( u i , uk)E E This is only half of the story; we actually have the following result of Pneuli, Lempel and Even [20]. Theorem 7.1. A n undirected graph G is a permutation graph if and only if G and G are comparability graphs.
Theorem 7.1 suggests an algorithm for recognizing permutation graphs, namely, applying the transitive orientation algorithm to the graph and to its complement. If we succeed in finding transitive orientations, then the graph is a permutation graph. To find a suitable permutation we can follow the construction procedure in the proof of the theorem, which can be found in Golumbic [13]. The entire method requires O ( n 3 )time and O ( n Z )space. Permutation graphs are useful in a number of applications (Even, Pnueli and Lempel [ 5 ] ,Tarjan [25],Golumbic [ 131). Of particular interest in this context is the following very efficient coloring algorithm for G [ T ] . Algorithm 7.1. Coloring a Permutation Graph. Input: A permutation T = [T,, T ~ ,. ..,T"]of the numbers {1,2,. . . ,n } . Output: A coloring of the vertices G [ T ]and the chromatic number y, of G [ T ] . Method: The vertices of G [ T ]are assigned colors in the order m,7r2,. . . , r n , although the graph itself is never actually calculated. A counter k will keep track of the total number of colors used so far, and an array LAST(c) will contain the number of the vertex which was the last to receive color c. During the jth time
M.C. Golumbic
318
through the loop we color T, with the smallest color q satisfying rj2 LAST(q). The entire algorithm is as follows: procedure
k t o ; for i + 1 to n do LAST(i)+O; for j + l to n do m t rnin{q [ nj 3 LAST(q)}; COLOR(rj)+m; LAST( m )t rj; k +max{k, m}; end loop
1. initialize: 2. loop: 3. 4. 5. 6. 7.
end
X+k; end
Example. Let us illustrate Algorithm 7.1 on the permutation T = [4,1,3,5,2]. After the initializations in line 1 the fotlowing assignments will be made in the loop: j +Z rn +2
I +3
jt 4
j--S
rn +2
m+l
m -3
COLOR(l)+2 LAST(2)tl k +2
COLOR(3)+2 LAST(2)+3 k +2
COLOR(S)+1 LAST(l)+5 k6 2
COLOR(2)+3 LAST(3)+2 k +3
j-I
m t l COLOR(4)c 1 LAST(l)t4 k-1
Thus the chromatic number of G [ r ]is 3 and a 3-coloring has been assigned. The complexity of Algorithm 7.1 is O ( n l o g x ) if line 3 is implemented using binary search. A proof of the correctness of this algorithm can be found in Golumbic [13]. Algorithm 7.1 can be used to color any permutation graph G in O(rz log n ) time provided we are given the permutation r and the isomorphism G + G [ r ] .If we do not have r,then we should use Algorithm 5.4.
8. Interval graphs An undirected graph G is called an interval graph if its vertices can be put into one-to-one correspondence with a set of intervals 9 of a linearly ordered set (like the real line) such that two vertices are connected by an edge of G if and only if their corresponding intervals have nonempty intersection. We call 9 an interoal representation for G. (It is unimportant whether we use open intervals or closed intervals; the resulting class of graphs will be the same.) The following characterization of interval graphs is due to Gilrnore and Hoffman [lo].
Algorithmic aspects of perfect graphs
319
Theorem 8.1. An undirected graph G is an interval graph if and only if G is a triangulated graph and its complement G is a comparability graph.
The coloring, clique, stable set, and clique cover problems can be solved in polynomial time for interval graphs by using the algorithms of Sections 4 or 5 , and a recognition algorithm could be obtained by combining the algorithms for triangulated graphs and comparability graphs. However, the recognition algorithm presented in Booth and Lueker [3] is asymptotically more efficient. They have shown that a data structure called a PQ-tree can be used to obtain a linear algorithm. Interval graphs have become particularly useful mathematical structures for modeling real world problems. The line, on which the intervals rest, may represent anything that is normally regarded as one-dimensional. The linearity may be due to physical restriction such as blemishes on a microorganism, speed traps on a highway, or files in sequential storage in a computer. It may arise from time dependencies as in the case of the life span of persons or cars, or jobs on a fixed time schedule. A cost function may be the reason as with the approximate worth of some fine wines or the potential for growth of a portfolio of securities. The task to be performed on an interval graph will vary from problem to problem. If what is required is to find a coloring or a maximum weighted stable set or a large clique, then fast algorithms are available. If a Hamiltonian circuit must be found, then there are no known efficient algorithms (unless the graph has more structure than just being an interval graph). Also, the speed with which such a problem can be solved will depend partially on whether we are given simply the interval graph G, or, in addition, an interval representation of G. We have already seen one application of interval graphs in the opening paragraph of this article. The interested reader is referred to Roberts [21], [22] and Golumbic [13] for numerous other applications. We will discuss here a recent application of interval graphs to optimal macro substitutions suggested by Golumbic, Goss and Dewar [ 151. The compiler or interpreter for a microcomputer system may be regarded as a byte sequence which resides in main memory. Due to restrictions on the size of main memory, it is desirable to compact this byte sequence. One technique is to define a set of macro substitutions which allow occurrences of specified byte subsequences to be replaced by single bytes. The subsequences are restored dynamically at run time by use of an associated table. Fig. 8.1 shows a sequence of hexidecimal digits of length 36. Since the digits E and F do not appear, they may be used to indicate macros. Choosing E = 6A2 and F = 43B96 the original sequence may be reduced to length 20. Notice that when two macros overlap, only one can be replaced. This overlapping phenomenon, therefore, restricts how the macro table may be applied.
M.C. Golumbic
320
~
Original Sequence: 6&C43B%OD60661C78-243B96&3C&25 Macro Table: E = 6A2
F = 43B96
T \
OVERLAP
Abbreviated Sequence: ECFOD6OElC78EFA23CE5
Fig. 8.1. Macro substitution.
The problem to be solved is to choose an optimal set of macro substitutions and an order for performing the substitutions which minimizes the total length of the byte sequence and associated table. Formally we require the following. Input: Output:
A byte sequence B of length n.
A set of m macros each of length s k and an order for performing the substitutions such that the total length of the abbreviated sequence and macro table is minimized.
The reason for specifying a bound on the length of the macros is that in practice we may want them to be very short compared to the length of the original sequence. Notice that there are actually two aspects to the problem: (1) choosing a macro set, and (2) using the macro set optimally. Let B = ( b , ,bz,. . . ,b,) be a sequence of bytes and let k be a fixed constant. The length of B is denoted by I B I = n. A subsequence ( h , .. . ,b,) of B is denoted by B [ i . j ] . Clearly, l B [ i , J ] l= J - i + 1. The weighred interval graph G = (V, E, w ) that we will associate with B is defined as follows: The vertex set V consists of all intervals [ i , j ] satisfying 1 s J - i k - 1 ; two vertices u = [ i , J ] and u = [ i ' , j ' ] are connected by an edge iff they intersect, i.e., either i f d j d j' or i S J ' s j ; the weight w ( u ) of a vertex u = [ i , J ] is equal to j - i which represents the number of bytes that would be saved by replacing B [ i , j ] by a single byte. I t is easy to see that the number of vertices of G is slightly less than kn and the number of edges is less than but on the order of k'n. Furthermore, the graph does not actually have to be calculated and stored since any query about adjacency of vertices can be answered by a simple comparison of the indices of their corresponding subsequences. Let M be a subset of V and let
1
B [ M ] = { B [ i , j ] [ i . J ]E M ) . We may think of B [ M ] as the macro table generated by M. To perform the
Algorithmic aspects of perfect graphs
32 I
macro substitutions we would find all occurrences of these macros and then choose a subset of the occurrences, no two of which intersect, to be abbreviated. Such a subset corresponds precisely to a stable set of the interval graph G. (Notice that this model does not permit embedding one macro in another macro.) Moreover, t o make the abbreviated sequence as short as possible, we would like a stable set whose weight is maximum. (The weight of a subset of vertices is the sum of the weights of its members.) This method is summarized in Fig. 8.2. procedure SUBSTITUTION(A4): C ( M ) t { [ i j, ] E V B [ i ,j ] = B [i ‘ , j ’ ] for some [ i ’ , j ’ ] E M } ;
I
X ( M ) t M A X I M U M WEIGHTED STABLE SET OF THE INDUCED SUBGRAPH GC(M); SAVINGS(M)+
c
“EX(M)
w(u)-
c
w(u);
“EM
end Fig. 8.2. Finding an optimal macro substitution for a given set of macros,
The set C ( M ) consists of all intervals representing candidate subsequences which may be replaced using the macro table B [ M ] . Of these candidates only the subsequences represented by X ( M ) will be replaced. The SAVINGS is calculated by summing the savings obtained for each macro substitution and subtracting the cost of storing the macro table. Using SUBSTITUTION we obtain the following algorithm which gives an optimal solution to the general problem. Algorithm 8.1. V such that IM I = rn do loop: for all M call SUBSTITUTION(M); end loop return the M and X ( M ) whose SAVINGS(M) is maximum; T h e number of passes through the loop in Algorithm 8.1 is on the order of (?:) since G has O ( k n ) vertices. (In practice, some of the subsets M may be ruled out due to other criteria, for example, by requiring that macros begin with certain designated bytes. This would lower the number of passes.) The complexity of SUBSTITUTION depends on how efficiently we are able to find C ( M ) and X ( M ) for a given M . Using a modification of the deterministic pattern matching algorithm of Morris and Pratt [19’], C ( M ) can be calculated in O(rn(k n ) ) time. See also Aho, Hopcroft and Ullman [l], Chapter 9. Since a maximum stable set of a n interval graph G = (V, E ) may be found in time O(l VI + IE I), X ( M ) can be calculated in O ( k ’ n ) time. Hence, we conclude
+
322
M.C. Golumbic
that the worst case complexity of SUBSTITUTION is O ( m ( k + n ) + k ' n ) and the worst case complexity of Algorithm 8.1 is
which is, in terms of the length of the input sequence, a polynomial whose degree depends on the constant m. Notice that our model has not allowed the embedding of macros in other macros. A reason for this could be that it is impractical to implement the stack nccessary to allow embedding. In some applications one may choose to allow embedding. If this is the case, a similar model can be designed which uses overlap graphs rather than interval graphs. An overlap graph is the same as an interval graph in which there are no edges between pairs of vertices whose corresponding intervals have one properly contained in the other. Our Algorithm 8.1 and SUBSTITUTION will also be optimal using the overlap graph model. Their respective complexities, in this case, will each be raised by one power of kn. This follows from the fact that a maximum weighted stable set of an overlap graph G = (V, E ) can be calculated in O ( I V 1. I E I) time (see Gavril [Y] and Golumbic [13], Chapter 1 I). The problem of macro substitution was recently applied to MICRO SPITBOL for an Incoterm SPD20/40 supporting 64K of main memory. The byte sequence for MICRO SPITBOL required 23,110 bytes of storage. There were 176 unused opcodes which were designated to represent macros. That is, n = 23110 and m = 176 and we set k =20. Since the time complexity of Algorithm 8.1 would be high for this application, an effective technique for finding a near optimal solution was needed. A combination of heuristics and SUBSTITUTE reduced the size of the sequence to 17.920 bytes and produced a macro table of 962 bytes. This represents a saving of 4,228 bytes of main storage, a saving of 20%. It should be pointed our that an increased cost of obtaining a very good macro substitution may be justified by the fact that this is done only once per compiler and machine and the result presumably will be used many, many times.
References [ I I A.V. Aho, J.E. Hopcroft and J.D. Ullman, The Design and Analysis of Computer Algorithms (Addison-Wesley, Reading, MA, 1976). [2] C. Berge, Graph and Hypergraphs (North-Holland, Amsterdam, 1973). [ 3 ] K.S. Booth and G.S. Lueker, Testing for the consecutive ones property, interval graphs, and graph planarity using PQ-tree algorithms, J . Comput. Systems Sci. 13 (1976) 335-379. 141 G . A . Dirac, On rigid circuit graphs, Abh. Math. Sem. Univ. Hamburg 25 (1961) 71-76.
Algorithmic aspects of perfect graphs
323
[5] S. Even, A. Pnueli and A. Lempel, Permutation graphs and transitive graphs, J. Assoc. Comput. Mach. 19 (1972) 4 W 1 0 . [6] S. Foldes and P.L. Hammer, Split graphs, in: F. Hoffman et al., eds., Proc. 8th Southeastern Conf. on Combinatorics, Graph Theory and Computing (Congressus Numerantiurn XIX, Utilitas Math., Winnipeg, 1977) 311-315. [7] D.R. Fulkerson and O.A. Gross, Incidence matrices and interval graphs, Pacific J. Math. 15 (1965) 835-855. IS] F. Gavril, Algorithms for minimum coloring, maximum clique, minimum covering by cliques, and maximum independent set of a chordal graph, SIAM J. Comput. 1 (1972) 180-187. [9] F. Gavril, Algorithms for a maximum clique and a minimum independent set of a circle graph, Networks 3 (1973) 261-273. [The class of circle graphs is equivalent to the class of overlap graphs.] [lo] P.C. Gilmore and A.J. Hoffman, A characterization of comparability graphs and of interval graphs, Canad. J. Math. 16 (1964) 539-548. [ l l ] M.C. Golumbic, Comparability graphs and a new matroid, J. Comb. Theory, Ser. B 22 (1977) 68-90, [ 121 M.C. Golumbic, The complexity of comparability graph recognition and coloring, Computing 18 (1977) 199-208. 1131 . . M.C. Golumbic, Algorithmic Graph Theory and Perfect Graphs (Academic Press, New York, 1980). 1141 M.C. Golumbic, A remark on the NP-completeness of the threshold dimension problem, Bell Laboratories Technical Memorandum (1982). [15] M.C. Golumbic, C.F. Goss and R.B.K. Dewar, Macro substitutions in MICRO SPITBOL - A combinatorial analysis, in: F. Hoffman et al., eds., Proc. 11th Southeastern Conf. on Combinatorics, Graph Theory, and Computing, Congressus Numerantium (Utilitas Math., Winnipeg, 1980) 485-495. [16] M. Grotschel, L. Lovrisz and A. Schrijver, The ellipsoid method and its consequences in combinatorial optimization, Combinatorica 1 (1981) 169-197. [17] P.L. Hammer and B. Simeone, The splittance of a graph, Combinatorica 1 (1982) 275-284. [18] C.G. Lekkerkerker and J. Ch. Boland, Representation of a finite graph by a set of intervals on the real line, Fund. Math. 51 (1962) 45-64. [19] J.H. Morris and V. R. Pratt, A linear pattern matching algorithm (Tech. Report No. 40, Computing Center, University of California, Berkeley, Calif., 1970). [20] A. Pnueli, A. Lempel and S. Even, Transitive orientation of graphs and identification of permutation graphs, Canad. J. Math. 23 (1971) 160-175. (211 F.S. Roberts, Discrete Mathematical Models, with Applications to Social, Biological and Environmental Problems (Prentice-Hall, Englewood Cliffs, New Jersey, 1976). [22] F.S. Roberts, Graph Theory and Its Application to Problems of Society, NSF-CBMS Monograph No. 29 (SIAM Publ., Philadelphia, Pa., 1978). [23] D.J. Rose, R.E. Tarjan and G.S. Lueker, Algorithmic aspects of vertex elimination on graphs, SIAM J. Comput. 5 (1976) 266-283. [24] L.N. Shevrin and N.D. Filippov, Partially ordered sets and their comparability graphs, Siberian Math. J. 11 (1970) 497-509. [25] R.E. Tarjan, Sorting using networks of queues and stacks, J. Assoc. Comput. Mach. 19 (1972) 341-346. [26] R.E. Tarjan, Maximum cardinality search and chordal graphs (Stanford Univ. Lecture Notes CS 259, unpublished, 1976). [27] W.T. Trotter, Jr., J.I. Moore and D.P. Sumner, The dimension of a comparability graph, Proc. Amer. Math. SOC.60 (1976) 35-38.
. -
This Page Intentionally Left Blank
Annals of Discrete Mathematics 21 (1984) 325-356 @ Elsevier Science Publishers B.V.
POLYNOMIAL ALGORITHMS FOR PERFECT GRAPHS M. GROTSCHEL Institut fur Mathematik, Universifiit Augsburg, Augshurg, W. Germany
L. LOVASZ Institute of Mathematics, Eiitviis Lorand University, H - 1088 Budapest, Hungary
A. SCHRIJVER Department of Econometrics, Tilburg University, Tilburg, The Netherlands We show that the weighted versions of the stable set problem, the clique problem. the coloring problem and the clique covering problem are solvable in polynomial time for perfect graphs. Our algorithms are based on the ellipsoid method and a polynomial time separation algorithm for a certain class of positive semidefinite matrices related to Lovasz’s bound 8 ( G ) on the Shannon capacity of a graph. We show that 9 ( G ) can be computed in polynomial time for all graphs G and also give a new characterization of perfect graphs in terms of this number 9(G). In addition we prove that the problem of verifying that a graph is imperfect is in NP. Moreover, we show that the computation of the stability number and the fractional stability number of a graph are unrelated with respect to hardness (if P # N P ) .
1. Introduction and notation It is well known that the stable set problem, the clique problem, the chromatic number problem and the clique cover problem are NP-complete problems for general graphs, cf. [3]. The purpose of this paper is to show that these problems, and even their weighted versions, are solvable in polynomial time for perfect graphs. The algorithms presented here are based on the ellipsoid method (cf. [XI, [2], [4])and on a computationally tractable characterization of the number 8 ( G ) introduced by Lovasz 1101in connection with the Shannon capacity of a graph. In the remaining part of this section we shall introduce our notation and state the problems we shall investigate. The second section gives a brief review of the ellipsoid method and some properties of this method which are important for our purposes. In Section 3 we show that the stable set problem is unrelated to the fractional stable set problem with respect to hardness for general graphs. The Shannon capacity and the numbers 6 ( G ) , &,(G), which are important for the design of our algorithms, are treated in Section 4, and a polynomial separation algorithm for a certain class of positive semidefinite matrices related to 6 ( G ) is presented in Section 5. This algorithm is utilized together with the ellipsoid 325
M.Grotschel et a/.
326
method in Sections 6 and 7 to obtain polynomial time algorithms for the weighted versions of the stable set problem, clique problem, coloring problem and clique cover problem on perfect graphs. All graphs in this paper are finite and undirected. Since loops and multiple edges do not play a role for the concepts we consider, we assume that all graphs are without such edges, i.e., are simple. A graph is denoted by G = (V(G), E ( G ) )where V(G) (or just V) is the vertex set and E ( G )(or just E ) is the edge set of G. An edge connecting two vertices i and j is denoted by ij, and we say that two vertices are adjacent if they are equal or connected by an edge. The complementary graph of a graph G is defined as the graph G with V(G) = V(G) and in which two different vertices are adjacent if and only if they are nonadjacent in G. A stable set of a graph G is a set of vertices W C V(G) such that any two vertices of W are nonadjacent in G, and a clique of G is a set of vertices C c V(G) such that any two vertices of C are adjacent in G. The maximum cardinality of a stable set in G is called the stability number of G and is denoted by a ( G ) .The maximum cardinality of a clique in G is called the clique number of G and is denoted by w ( G ) . Clearly, a stable set of G is a clique of G, and vice versa, thus a ( G )= w ( G ) and o ( G )= a(G)hold. A k-coloration of G is a partition of V(G) into k stable sets of G, and the least integer k for which G admits a k-coloration is called the chromatic number of G, denoted by x ( G ) . A k-clique couer of G is a partition of V(G) into k cliques of G, and the least k for which G admits a k-clique cover is called the clique couer number of G which is denoted by p ( G ) . By definition, every k-coloration of G is a k-clique cover of and vice versa, which implies x ( G ) = p ( G ) and p ( G )= x(G). The problem of finding the stability number (clique number, chromatic number, clique cover number) of a graph is called the srable set (clique, coloring, clique couer) problem. These four problems have natural weighted versions. Given a graph G = (V, E ) and a ‘weight’ w, E E , for all u E V (H, is the set of positive integers), then the weighted srable set problem (weighted clique problem) is to find a stable set W (a clique C) of G such that the sum of the weights of the vertices in W (in C) is as large as possible. The weighted coloring problem (weighted clique problem) is the following: find stable sets W,, Wz,.. ., W, (cliques C1,Cz,.. . ,C,) and positive integers y l , y z , . . . ,y , such that for all u E V, y , 3 w, y , 2 w , ) holds, and such that y , is as small as possible. The optimum values of these four problems are denoted by a, (G), ow(G), xW(G), pw( G ) and are called the weighted stability, weighred clique, weighted chromatic, weighred clique couer number. It is obvious that for any graph G, a ( G ) S p ( G ) and w ( G ) S x ( G )hold. A graph G is called perfect if
c,
c,,,
(c,,,
c:=,
Polynomial algorithms for perfect graphs
327
a ( G [W ] )= p ( G [W ] ) for all W C V ( G ) where G [W ] denotes the subgraph of G induced by the vertex set W C V ( G ) . Lovasz [9] has shown the so-called perfect graph theorem, namely, that a graph G is perfect if and only if its complement G is perfect. So the perfect graph theorem is equivalent to the following: a graph G is perfect if and only if w ( G [ W ] ) = x ( G [ W ] )for all W C V ( G ) . Due to the perfect graph theorem it suffices to design polynomial time algorithms only for the weighted stable set and the weighted clique cover problem in order to obtain polynomial time algorithms for all the four problems described above on perfect graphs. Namely, suppose we have a polynomial time algorithm for the weighted stable set problem on perfect graphs and we want to find the maximum weighted clique in a perfect graph G. Then, obviously, the set of maximum weighted cliques of G equals the set of maximum weighted stable sets of G. Since by the perfect graph theorem G is perfect, we can apply our polynomial time algorithm to calculate a maximum weighted stable set in G and thereby obtain a maximum weighted clique in G. Similarly, if we have a polynomial time algorithm for the weighted clique cover problem in perfect graphs we can obtain a minimum weighted coloring of a perfect graph G by applying our polynomial time algorithm to the (perfect) complementary graph G. Therefore, we shall concentrate in the sequel on designing polynomial time algorithms for the weighted stable set and clique cover problem on perfect graphs, keeping in mind that these also yield polynomial time algorithms for the weighted clique and coloring problem on perfect graphs. There are various classes of graphs known for which the weighted versions of the stable set, the clique, the coloring or the clique cover problem can be solved in polynomial time. For a survey of such results see [3]. These classes of graphs include several classes of perfect graphs, e.g., bipartite, triangulated and comparability graphs as well as line graphs of bipartite graphs. Recently, Hsu [6] has shown that the coloring problem, and Hsu and Nemhauser [7] have shown that the clique and clique cover problem, are solvable in polynomial time for claw-free perfect graphs.
2. The ellipsoid method
Based on an algorithm due to Shor [13], Khachiyan [S] recently devised a method which solves linear programming problems in polynomial time; for nrnofs. see 121. This so-called ellipsoid method can be used to derive the
328
M.Grotschel er al.
polynomial solvability of a more general class of problems, in particular to obtain a powerful tool for solving combinatorial optimization problems as was described by Grotschel et al. [4].In this section we give a brief survey of this method and state those theorems of Grotschel et al. [4]which are of interest for the design of polynomial time algorithms on perfect graphs. A conuex body is a closed, bounded, fully dimensional, and convex subset of R", n 3 2. More precisely, if we speak of a convex body K we always assume that the following information is known: the integer n 3 2 with K C R", two rational numbers 0 < r S R, and a vector a. E K such that
S(a,, r ) C K 5 S(ao,R ) , where S(a,,,s ) = {x E R" IIIx - aoIIS s } (11. ( 1 is the euclidean norm), denotes ball with center ao and radius s. Therefore, we also denote a convex body by quintuple (K; n, ao,r, R ) where we assume that n 5 2, a. E Q", 0< r G R given explicitly. The following two problems are of particular interest and - as we shall later - polynomially related. Assume that a convex body ( K ; n,a,,,r, R ) is given.
the the are see
(2.1) Optimization Problem. Given a vector c E QD" and a rational number E > 0, find a vector y E 6)" such that d(y, K) =z E and cTx =z cTy + E for all x E K (i.e. y is almost in K and almost maximizes cTx on K). (2.2) Separation Problem. Given a vector y E Q" and a rational number S > 0,
conclude with one of the following: (i) asserting that d ( y , K )4 6 (i.e., y is almost in K); or (ii) finding a vector c E Q" such that IIc 11 2 1 and for every x E K, cTx 6 cTy+ S (i.e. finding an almost separating hyperplane). (Here d ( . ,.)denotes the distance function, i.e., d ( x , y) = \ \ x - y 11 and d(y, K ) = inf{d(x, y ) Jx E K ) . ) The method of Yudin and Nemirovskii [14], section 4.5, would enable us to show that the following third problem is also polynomially related to problems (2.1) and (2.2) above:
(2.2') Feasibility Problem. Given a vector y E Q" and a rational number S > 0. conclude with one of the following: (i) asserting that d ( y ,K) S 6, or (ii) asserting that d (y, R"\ K ) s 6. Clearly this problem is easier than the separation problem. However, in the
Polynomial algorithms for perfect graphs
329
applications in this paper, we shall obtain almost separating hyperplanes automatically. To speak of a polynomial time algorithm for a convex body K we have to specify how we measure the input length of K. Whenever something is encoded we assume that the (usual) binary encoding is used. A rational number is encoded by encoding the numerator and the denominator. If x E Q" (Q is the set of rational numbers) then Ilx denotes the maximum of the absolute values of the integers appearing as numerator or denominator in the coefficients of x. In other words, to encode x at least log IIx 1 1 + n places are necessary. (In this paper all logarithms have base two.) For a convex body ( K ;n, a,,, r, R ) we assume that the parameters n, allE Q", r E 4p and R E Q are coded. If X is a class of convex bodies then the input of the optimization (or separation) problem for Yt is the code of some member ( K ; n, acl,r, R ) E X, of a vector c E Q" and of a rational number F > 0 (of a vector y E q" and a rational number 6 > 0). The length or size of the input is the length of this (binary) encoding. Thus, the length of the input is at least
)I
+ 1%
It r
1 1 3
It It= + log It Y 11%
+ log R
where y = E or y = 6. An algorithm to solve the optimization (separation) problem for the class Yt is called polynomial if its running time is bounded by some polynomial of the size of the input. (2.3) The ellipsoid method. Given a convex body ( K ;n, all,r, R ) , a linear objective function cTx with IIc ( 1 3 1 and a number E > 0 (the required accuracy). We assume that there is a subroutine SEP(K, y, 6 ) which for the given convex body K, a vector y E Q" and a rational 6 > 0 either concludes that y E S ( K ,6 ) = {x E R" d ( x , K ) s 8 ) or yields a vector d E Q" such that dTx S dTy + 6 for all x E K, i.e. SEP solves the separation problem for K. We first define the following numbers:
1
, (2.3.1) N:=4n2[ l o g m ] r&
(2.3.2)
24-N a:=- R300n
12Gl
(2.3.3) p : = 5 N (log 7 and then proceed as follows:
(2.3.4) Set x , ~= a. (center of the first ellipsoid), A,i:= R'I, (I,,is the ( n , n)-identity matrix);
M.Griirschel et al.
330
(2.3.5) for k = 0 to N - 1 do; (2.3.6) Run the subroutine SEP(K, xk,6). (2.3.7) If SEP(K, & , 6 ) concludes that Xk E S ( K , a), we say that k is a feasible index and set a : = c. (2.3.8) If SEP(K, xk, 6 ) yields a vector d E R" such that (Id (1 3 1 and sup{dTx x E K } s dTxk+ 8, we call k an infeasible index and set a : = - d.
1
(2.3.9) bk:= AkaI d a T A k a , (2.3.10) x : : = x ~
+-n +1 1
bkt
Above, the sign means that the left-hand side is obtained by rounding the binary expansion of the right-hand side after p places behind the point. Since by construction x o E K , the set of feasible indices is nonempty; moreover, we can show the following theorem, cf. (41. (2.4) Theorem. Let j be a feasible index for which
< N, k feasible}.
C'X,
= max{cTxk1 0 s k
C'X,
3 S U ~ { C ~ Xx E K} - E .
Then
1
0
Cleai.j, the number N of iterations of the ellipsoid met..od is polynomial in the size of the input. One can also show that the entries of the intermediate vectors Xk and matrices A k ,0 S k S N, are polynomially bounded. Furthermore, the number 6 used to run the separation subroutine is polynomial in the input length. Thus, the ellipsoid method is a polynomial algorithm for the optimization problem for K if and only if the subroutine SEP is a polynomial algorithm for the separation problem for K. This implies, in particular, that whenever there is a polynomial separation algorithm for a class of convex bodies X there is also a polynomial optimization algorithm for .X (via the ellipsoid method). It is of particular importance that this implication also holds the other way round, namely:
Polynomial algorithms for perfect graphs
331
(2.5) Theorem. Let X be a class of convex bodies. There is a polynomial algorithm to solve the separation problem for the members of YC, if and only if there is a polynomial algorithm to solve the optimization problem for the members of 3%. 0 Note that according to our definition neither the optimization nor the separation problem are solved exactly; in both cases we allow for a small error. This is necessary because the problem classes that are covered by Theorem (2.5) may also contain instances with a unique optimal solution which has irrational coefficients. But irrational numbers cannot be represented exactly. In case our class of convex bodies X is a class of polytopes, then the optimization problem for X is nothing but a linear programming problem. If in addition all members of X have a rational defining inequality system, then both the separation and the optimization problem can be solved precisely, we shall say in the strong sense. Moreover, it is also possible to construct a dual optimal solution in polynomial time. If P C R" is a polytope with rational vertices, define T ( P )to be the maximum of the absolute values of numerators and denominators occurring in the entries of vertices of P. The pair (P, T ) is called a rational polytope if T ( P )< T.The input size of a rational polytope is at least n [log T I . It is not difficult to prove that if (P, T ) is a rational polytope, then P C S(0, n T ) and if P is fully dimensional then S(ao,(nT)-2"3)C P for some point a".
+
(2.6) Theorem. Let ?€be a class of fully dimensional rational polytopes such that the optimization (or equivalently the separation) problem for YC can be solved in polynomial time. Then the following holds : (a) There is a polynomial optimization algorithm for YC in the strong sense, i.e., which for every member P E YC and every rational vector c finds a vector y E P such that cTy = max{cTx x E P } . (b) There is a polynomial separation algorithm for Yl in the strong sense, i.e., which for every member P E X,P C R", and every rational vector y either asserts that y E P or finds a rational vector c with Ilc I( 3 1 such that c T x < c T y for all x E P. In case y E P the algorithm also yields vertices xu, xI, . . . ,x, of P and rational numbers A,,, A l , . , . ,A, 3 0 such that A, = 1 and A,x, = y. (c) There exists a polynomial algorithm which for every P EX, P C R " and c E Z"provides facets aTx G b, ( i = 1,. . . , n ) of P and rational numbers A, 2 0 ( i = 1,. . . ,n ) such that A,a, = c and A h , = max{cTx x E P } . 0
I
c:=,,
c:=,
c:'=,
x:=,, 1
Case (c) of Theorem 2.6 will play an important role in the sequel, since it will provide us with a method to construct a minimum weighted clique cover from a maximum weighted stable set. A further class of convex bodies will be of interest for our purposes. Let R': be
M. Grotschel et al.
332
the nonnegative orthant and K C R" be a convex body such that there are reals r and R, O < r S R, witl-
Rw:flS(O,r)CKcR:nS(O. R ) , 0 sx
(2.7)
+ x E K.
y EK
(2.8)
The anti-blocker A ( K ) of K is defined by
(2.9)
A(K):={yER':(yTxS1 foreveryxEK).
It is easy to see that A ( A ( K ) )= K for a K satisfying (2.7) and (2.8),and that in case K is a polytope with vertices xI,x2,. . . ,xk then A ( K ) = { y E R: y T x i s I , i = 1.. . . , k } . Moreover, if X is a class of convex bodies satisfying (2.7) and (2.8) we set A ( x )= {A ( K ) K E XI.
1
1
(2.10) Theorem. Let .X be a class of convex bodies satisfying (2.7) and (2.8).
Then the optimization problem for X can be solved in polynomial time i f and only if the optimization problem for A ( X ) can be solved in polynomial time. 0
3. The fractional stable set problem To be able to utilize the ellipsoid method for combinatorial optimization problems one has to associate a class of convex bodies with the problem class under consideration. Natural candidates are usually the convex hulls of the incidence vectors of feasible solutions. In case of the stable set problem this is done as follows. Let G = ( V , E ) be a graph with n vertices. For every W V(G) denote by x the (node-) incidence vector of W, i.e. x ,"= 1 if u E W and x,"=O if LIEW. Then
I
P ( G ) : = c o n v { x W E R " W C V(G)isastablesetof G }
(3.1)
is called t h e stable ser polyrope of G. Clearly, every weighted stable set problem on G can be solved as a linear programming problem over P ( G ) . The polytope P ( G ) is fully dimensional, has O/l-vertices, and is contained in the unit hypercube, thus P ( G ) is a rational polytope. If we were able to design a polynomial separation algorithm for P ( G ) , then by Theorem (2.5) the weighted stable set problem would be solvable in polynomial time. Since this problem is NP-complete, we cannot expect to find a polynomial separation algorithm for P ( G ) in general. A usual approach to solve difficult optimization problems is to consider tight relaxations of the problem in question which are polynomially solvable, and then proceed by branch-and-bound methods. A natural relaxation of the stable set
Polynomial algorithms for perfect graphs
333
problem is the so-called fractional stable set problem. By definition, no two vertices of a stable set are adjacent. Thus, given any clique C of a graph G, at most one vertex of a stable set can belong to C. This implies that for every clique C C V(G) and every incidence vector x w of a stable set W G V ( G ) the so-called clique inequality
is satisfied. For any graph G with n vertices we call
1
P * ( G ) : = (x E R" x, a 0 for all u E V(G) and
the fractional stable set polytope of G. P*(G) is clearly a rational polytope. Since obviouly P ( G ) C P*(G), the LP-solution over P*(G) provides an upper bound for the weight of the optimal stable set in G. For a given graph G and an objective function w : V+Z+ let us define the following parameters: a,(G):=max{wTx I x EP(G)}, a*,(G):=max{wTxI x EP*(G)}, a*(G):=max
{
x, I x E P * ( G ) } . "E"
The number a *(G) is called the fractional stability number of G, a *,(G)is called the fractional weighted stability number of G, and as mentioned earlier a, (G) is called the weighted stability number of G. By definition we have a ( G ) 6 a *(G) and a, (G) s a *,(G). At first sight the polytope P*(G)looks rather innocent. It is easy to see that its facets are the trivial inequalities xu S O for all u E V(G) and the clique inequalities & e C ~ v 6 1 for all maximal cliques C V(G) (maximal with respect to set inclusion). However, it is not known how to find all maximal cliques efficiently, even worse, there are classes of graphs (even perfect ones) such that the number of maximal cliques grows exponentially in I V(G)I. So there is no way to represent the constraint system of P * ( G ) efficiently. By Theorem (2.5) this is not necessarily crucial, since it is not the number of inequalities which matters; what matters is whether one can find a violated hyperplane in polynomial time. Since the constraint system of P * ( G )looks quite simple one might hope to find a polynomial time separation algorithm for P*(G).But this is very unlikely as the complexity of the separation problem for P*(G) is closely related to the complexity of the weighted clique problem. More precisely:
334
M.Grotschel et al.
(3.3) Proposition. Let % be a class of graphs. Then there is a polynomial algorithm to solve the weighted fractional stable set problem for every member of % if and only if there is a polynomial algorithm to solve the weighted clique problem for every member of %.
Proof. For every member G of %, the weighted clique problem can be solved in polynomial time if and only if the linear program max w'x, x E Q(G) can be solved in polynomial time, where Q(G):=conv{xC E R " ( C C V ( G ) is a clique}. By definition, the anti-blocker of O(G) is A (Q(G)) = { y E R: yTx' s 1 for all cliques C C V ( G ) }i.e. , A ( Q ( G ) )equals P * ( G ) .Thus by Theorem (2.10) the linear program max wTx, x E Q ( G )can be solved in polynomial time for every G E % if and only if the linear program max wTx, x E P * ( G ) can be polynomially solved for every G E %. 0
I
It follows from the examples in [3] that there are various classes of graphs for which the weighted clique, and hence the fractional stable set problem, are solvable in polynomial time. However, since the weighted clique problem is NP-complete for the class of all graphs, Proposition (3.3) implies that the weighted fractional stable set problem is NP-equivalent. Proposition (3.3) therefore states that considering the fractional stable set problem instead of the stable set problem does not offer considerable advantages. Moreover, the problems of computing a, (G) and a :(G) seem to be unrelated with respect to difficulty. For planar graphs ow(G) (the weighted clique number) and hence a*,(G)can be computed easily in polynomial time, while the determination of a, ( G ) for planar (even cubic planar) graphs is NP-complete; cf. [3]. So for the complementary graphs of planar graphs the determination of o, ( G ) and hence a Z(G) is NP-equivalent, while a, ( G ) can be computed in polynomial time. Although for general graphs the fractional stable set problem does not seem to be useful for computing a w ( G ) ,the situation for perfect graphs is quite particular. Namely, Fulkerson has shown the following (see also [ 11): (3.4) Theorem. Let G be a graph. Then P ( G )= P*(G)holds i f and only if G is perfect. 0
In other words, Theorem (3.4) implies that for every perfect graph G and every objective function w , a, ( G ) = a * , ( G )holds. Therefore a computationally efficient procedure determining a Z(G) would yield the desired weighted stability number. As we shall see later a , ( G ) and a*,(G) can be computed in polynomial time for perfect graphs, however, we do not make direct use of P ( G ) resp. P*(G),but rather obtain this result via a detour which will be described in the next section.
Polynomial algorithmsfor perfect graphs
335
4. The Shannon capacity, 6 ( G ) and 6,(G)
The stable set problem has found some nontrivial applications in coding theory, in particular in finding the zero error capacity of a discrete memoryless channel; cf. [12]. Let us denote by G * H the Cartesian product of the graphs G and H, i.e. V ( G .I€)= V ( G ) x V ( H ) and two vertices (u, u ) , ( u ' , u ' ) E V ( G . H ) are adjacent if and only if u is adjacent to u ' in G and u is adjacent to u' in H. G denotes the Cartesian product of k copies of G. As an interpretation, consider a graph G whose vertices are letters in an alphabet and in which two vertices are adjacent if and only if they are 'confoundable'. Then the maximum number of one-letter messages which can be sent without danger of confusion is clearly a(G),moreover, a ( G k )is the maximum number of k-letter messages such that any two of them are inconfoundable in at least one coordinate place. It is easy to see that there are at least C Y ( Ginconfoundable )~ k-letter words, but in general there may be many more such words. To measure the largest rate at which one can transmit information with an error probability exactly equal to zero, Shannon [ 121 introduced the following number:
which is now called the Shannon capacity of graph G. From the fact that a ( G k " ) 3a ( G k ) a ( G e )it directly follows that
O(G) = lim k Va(Gk)
(4.2)
-m
and, since ( Y ( G )s~ a ( G k ) ,that
Shannon f12] obtained an upper bound for O(G) by showing
O(G)Sa*(G).
(4.4)
However, both inequalities a ( G )
vj;
M.Grotschel et al.
336
better bounds for the Shannon capacity. One such parameter, called d(G), introduced by Lovasz [ 101, will play a key role in our further development. Let G be a graph and assume that its vertices are labeled 1 , 2 , . . . ,n. We say that a system ( u , ,*~, . . . , u r nof) vectors in an Euclidean vector space is an orthonormal representation of G if each vector ui has length one and if, for every pair i, j of nonadjacent vertices of G, the vectors ui and uj are orthogonal. It is obvious that every graph has an orthonormal representation, e.g., take a set of n orthonormal vectors. Let %(G) be the set of all orthonormal representations of G, and U be the set of vectors of unit length, then set
Lovasz [lo] has given various characterizations of this number which we shall list in the sequel. Using the complementary graph C?, 6(G) can be defined alternatively as follows:
This formula can be used to show that 6 ( G ) is not greater than the fractional and d stability number. Let (u,, .. . ,u.) be an orthonormal representation of be a vector of unit length such that these vectors maximize (4.6), i.e., 6 ( G ) = 6.:,,( d T ~ , ) Let 2 . C be any clique of G. Then, by definition, the vectors u , , i E C, are pairwise orthogonal, and so (dTu,)’ dTd = 1. Defining the vector x = ( x l , .. . ,x , ) ~ by x,:= (d’u,)’a 0 we obtain X I E C x l S 1 for all cliques C C V(G), and thus x E P*(G). This implies
2
( d ” ~ ,=) ~ xI
6 ( G )= ,=I
a*(G).
,=I
(4.7)
The formulas (4.5) and (4.6) do not seem to be very handy computationally, but there are other characterizations of d(G) which use representations of G by means of symmetric matrices. For any graph G with n vertices we set
d ( G ) : = { A= ( a , , ) I A is a symmetric (n,n)-matrix such that ali = 1 if i = j or if i and j are nonadjacent}; (4.8) then 6 ( G ) can be described as the following minimum:
d ( G ) = min{A(A
)IA E d ( G ) }
(4.9)
where A ( A ) denotes the largest eigenvalue of A. Equation (4.9) implies that 6 ( G ) is an upper bound on the stability number and on the Shannon capacity of G. Namely. suppose a(G)= k, then by definition (4.8) every matrix A E d ( G ) has a principal (k,k)-submatrix, say Ak, all of whose entries are one. Since
337
Polynomial algorithms for perfect graphs
A ( A ) s A(A,) and k is an eigenvalue of Ak, we have A ( A ) 2 k for all A E d ( G ) , i.e., 6 ( G ) a a ( G ) .Now Lovasz proved
-
(4.10)
6(G H )= 6 ( G ) a ( H ) for all graphs G and H, which implies that a(G*)G6(Gk)= 6(G)k
and hence
O ( G ) s6(G).
(4.1 1)
The number 6 ( G ) can also be characterized as a maximum of the sum of the entries of certain matrices representing G. Denoting the trace b,, of a matrix B by tr(B), we define
x:=,
I
B(G):={B = (b,,) B is a symmetric positive semidefinite (n, n)-matrix with tr(B) = 1 such that b,, = O whenever i,j EE(G)};
(4.12)
then Lovasz [lo] showed that
6 ( G ) = max
{ ,$,b , I B w3)). E
(4.13)
Thus, 6 ( G ) can be considered as a maximum (cf. (4.6) and (4.13)), and as a minimum (cf. (4.5) and (4.9)). Among these characterizations of 6(G), (4.13) will be the most important one in our subsequent investigations. As a side remark we want to mention that a complementary slackness relation links the two characterizations (4.9) and (4.13) of 6(G). Namely, suppose B E 93(G), A E d ( G ) and A(A) is the largest eigenvalue of A, then
B ( A ( A ) I . - A ) = 0 e B is optimal for (4.13) and A is optimal for (4.9).
(4.14)
The inequalities (4.3), (4.7) and (4.11) imply that
a(G)<@(G)~6(G)scr*(G)
(4.15)
holds for all graphs G. We remarked earlier that for the pentagon C,, a (C,) = 2 < O(C5)= V's < a *(C,) = I. Since for the pentagon O(C5) equals d ( C s ) ,the last inequality in (4.15) may also be strict. Haemers [ 5 ] showed the existence of graphs G with O(G) < 6(G). Therefore all these four numbers, a ( G ) ,6(G), O(G) and a * ( G ) ,are different in general. However, for a perfect graph G, Theorem (3.4) implies that equality holds in all inequalities (4.15). A graph G = (V, E ) is called critically imperfect if G is not perfect but if the vertex deleted subgraph G - ZI is perfect for all u E V. There are only two
M.Grotschel et al.
338
classes of critically imperfect graphs known, namely, the cycles of odd length and their complementary graphs. We shall now prove that for critically imperfect graphs, a(G)< 8 ( G )< a * ( G ) . Padberg [ l l ] showed that if G is a critically imperfect graph with n vertices then n = a (G)w(G)+ 1, and that every critically imperfect graph has exactly n stable sets of cardinality a(G)and n cliques of cardinality w(G).He also proved that every vertex of G is contained in exactly a(G)maximum stable sets and in exactly w ( G ) maximum cliques. Moreover, Padberg [l 11 showed that the so-called stable set-vertex incidence matrix of a critically imperfect graph is nonsingular, i.e., S = (s,)) is an (n,n)matrix whose rows correspond to the n maximum stable sets of size a ( G ) , whose columns correspond to the n vertices of G, and where s,, = 1 (s,, = 0) if vertex j belongs (does not belong) to the maximum stable set i. By the properties of critically imperfect graphs mentioned above, S is an (n,n)-matrix such that every row and every column contains exactly a(G)ones. Now consider for a critically imperfect graph the matrix B':=STS. The properties of S imply that B' is positive define. An entry bl; of B' counts the number of maximum stable sets to which both i and j belong. Hence bh = a (G), and b:, = 0 if ij E E. Moreover, the sum of the entries of each row (or column) of B' equals (Y (G )'. Let A I 2 A: 2, . - 2 A L be the eigenvalues of B'. Since B ' is positive definite we have A A > 0; moreover, since B' is integral, det(B') 3 1. It follows that
2 A:
i=l
= tr(8') = na(G) and
fi A:
= det(B')a 1
i=l
Using a rough estimate we obtain A L 2 (fly:,: A :)-I vations imply that the matrix
2 ( n a (G))-"+'. These
obser-
is a positive definite matrix contained in B ( G )whose smallest eigenvalue A { is not smaller than (na(G))-"and for which b: = a ( G ) .Hence, if we subtract (na(G))-"from every element of the main diagonal of B" and then multiply the - n ) we obtain a matrix B which has resulting matrix by (na(G))"/((na(G))" trace one, is positive semidefinite, and satisfies b,)= 0 if i,j E E ( G ) . Thus B E SB(G),and since a ( G ) 2 2
c:l=l
and it follows from (4.13) that for a critically imperfect graph G (4.16)
339
Polynomial algorithmsfor perfect graphs
We now prove that 6 ( G ) < a * ( G ) . Padberg [ll] has shown that for a critically imperfect graph G, a * ( G )= cu(G)+l / w ( G ) ,and that there is a unique point y E P * ( G ) ,namely,
c:=,
which satisfies y , = a * ( G ) .B y (4.6)there exists an orthonormal representation ol,. . . ,u, of G and a vector d of unit length such that 6 ( G )= (dTu,)2. As shown in the section following formula (4.6), the vector x E R " with x, = (dTo,)', i = 1 , . . . ,n is contained in the fractional stable set polyhedron P*(G).Now suppose that 6 ( G )= a * ( G ) ,then
c:=,
hence by the uniqueness of y we have , 1 -- - yi = xi = (d'u,P, i = 1,. . . ,n. 4G) Thus for the orthogonal representation u,, . . . , ZI, of G and the vector d E U we have
and therefore formula (4.5) implies that 6(C?)Sw ( G )= a ( G ) . However, the complementary graph of a critically imperfect graph is critically imperfect too, so 6 ( G ) s a ( G ) contradicts (4.16). This implies that 6 ( G ) cannot equal a * ( G ) . Summing up we have shown that for any critically imperfect graph G
1
a ( G )< 6 ( G )< LY *( G ) = ( G )+ w(G)
(4.17)
a
Every imperfect graph contains an induced subgraph which is critically imperfect, i.e., a subgraph for which (4.17) holds, while for every induced is satisfied. This subgraph G' of a perfect graph G, a(G')= 6(G') = a *(G') implies the following characterization of perfect graphs. (4.18) Theorem. A graph G is perfect if and only if the following holds:
a ( G [ W ] ) = 6 ( G [ W ] forall ) WcV(G). 0 Our discussion above yields a further characterization, namely, a graph G is perfect if and only if a * ( G W [ ] )= 6 ( G [ W ] )for all W C V ( G ) . The number 6 ( G ) is not necessarily equal to a ( G ) , even worse, Konjagin (unpublished) has constructed a sequence of graphs G, with n vertices such that
M.Grotschel el al.
340
a ( G n )= 2 and 8 ( G m ) - + mThis . implies that there is no function f at all such that
6 ( G ) S f(cr(G)) holds for all graphs G. It seems to be an interesting problem if there exists a polynomially computable function q ( G ) and a function f such that a ( G )S d G ) f(Q(G)). An efficient way to calculate 8 ( G )provides us only with a good algorithm for the unweighted stable set problem in perfect graphs. In order to cover the weighted case too we now generalize 8 ( G )t o a weighted version 8,(G). Assume that a graph G = (V, E ) and a weight function w : V - + Z, are given. Define the graph G, to be the graph arising from G by replacing each vertex u of G by w. pairwise nonadjacent new vertices and where two vertices of G, are adjacent if and only if their originals in G are different adjacent vertices. This construction implies that a, (G) = (Y (G, ) holds. Moreover, Lovdsz [9] has shown that if G is perfect, then G, is also perfect (in fact, this is the key lemma for the perfect graph theorem). Hence for perfect graphs we have a , ( G ) = 8(G,) = a * , ( G ) Therefore . we define
O , ( G ) : = S(G,).
(4.19)
Note, however, that the existence of a polynomial algorithm for calculating 9 ( G ) does not give a polynomial algorithm for 6,(G) by applying this algorithm to G,, since no algorithm making up G, from G and w is polynomial in the input length O (I E I + log 11 w ]Im). A characterization of 8,( G )using the set SB(G)defined in (4.12) and avoiding this construction is given in the following theorem. (4.20) Theorem. Let G = ( V, E ) be a graph and w : V -+ B , a weight funcfion,
then (4.21)
Proof. Let M be the maximum in (4.21). We first prove M 6 6,(G) using formula (4.13) for 8(G). Suppose B = ( 6 , ) E 9 ( G ) . Replace every entry bij by a ( w i , wj)-matrix all of whose entries are (wiwj)-i'2bij to obtain an (r, r)-matrix B', where r = wi. B' is clearly symmetric and satisfies
c:=,
tr(B') =
2 bii =
,=I
i=l
bii = 1;
moreover, the definition implies that bh = 0 if i,jE E ( G ) . It is also easy to see that B' is positive semidefinite. Furthermore, simple calculation shows " bii = G j b i j
2
#.I = 1
i.j = I
341
Polynomral algorithmsfor perfecr graphs
which implies M G 6,(G). Conversely, with each matrix B ' E B(G,) we can associate a matrix B E %(G) such that b:, s G J 6 , holds, which proves 6, (G)<M. Namely, take B' E B(G,). Replace the (w,, w,)-submatrix induced by the copies of i and j, by the sum of its entries divided by V F , , and eventually add a nonnegative number to any diagonal entry to make the trace equal to one.
c:,,=, c:,=,
5. A separation algorithm for a class of positive semidefinite matrices with trace one
In this section we shall describe a polynomial time separation algorithm for the class of positive semidefinite matrices %(G) defined in (4.12). Every set 9 ( G ) is clearly convex and bounded, but not fully dimensional. Since the ellipsoid method, as described in Section 2, can only be applied to convex bodies, we have to replace the sets B(G) for technical reasons by fully dimensional ones. For every (n, a)-matrix B = (b,,),i.e., B E R""", and every graph G we define n + (z) - I E ( G ) I - 1. the following projection operation R"""+ R" where ii:= Discard from B all elements below the main diagonal; all lE(G)I elements b,,, i < j, corresponding to different adjacent vertices i,j E V ( G ) ;and the element b,,,,. Denote the vector of R" obtained from B in this way by B, i.e., B is an ii-vector whose components are indexed by F = {ii i = 1,. . . , n - 1) U {zj i < j and ij$Z E ( G ) } . Conversely, for every vector B E R E we define an extension operation n-1 R" + R n x n by setting b,,:=6,, i = 1,. . , , y1 - 1; b,.:= 1 b,, ; b,, = bJZ= 0 for i , j E E ( G ) ; b,, = b,, = 6,for i , j E E ( G ) , i < j. By definition the (n, n)-matrix B obtained by extending B E RE is a symmetric matrix with trace one. Now set
1
1
c,=,
I
%I(G):={BE R x B E B(G)};
(5.1)
a ( G ) is convex, since it is the projection of a convex set. We now prove that g ( G ) is fully dimensional and bounded, i.e. that a ( G ) is a convex body. Denote by. B, the projection of the matrix 1 B.:=;
I,,E B(G)
(Zn is the (n,n)-identity matrix); then the following holds:
Set r = (nZ
M.Grotschel et al.
342
and B E R""" is the extension of B then IIB - B, 1) S r implies B, 11 =s Vir. Now take any B E S(B., r ) and let B be its extension. If bij, i# j , is any nonzero entry of B, then because of symmetry bii = bji# 0. Now 11 B - B. (1 G VGr implies fibijs
B EW"
llB
-
element of B, then IIB - B, (1 S V%r implies
It follows that
are Now by Gershgorin's theorem, all eigenvalues of the extension B of positive, hence B f % ( G ) ,which proves that B E $?(G). Using the fact that for positive semidefinite matrices b,,b,J z b; holds and that the trace of B E %(G)is one, it easily follows that B ( G ) CS ( B . , 1). Expression (5.2) shows that the logarithms of the radii r = l/nZ-\m and R = 1 as well as the numerators and denominators of the interior point B,, are bounded in absolute value by a polynomial in n which is fixed over all graphs G, i.e., given a graph G with n vertices, these numbers can be computed in polynomial time. So we may apply Theorem (2.5) to compute 6,(G)via Theorem (4.20) by using the projection @(G)of B(G). More precisely, define the following class of convex bodies:
C%:
={a ( G )
I G is a graph with 1 V(G) I 3 2).
(5.3)
In order to solve the optimization problem for B it is sufficient to find a polynomial separation algorithm for 5. We shall now show that the separation problem for 8 is solvable in pdynomial time, even in the strong sense. Given a graph G, then this problem is the following: (5.4) Problem. Given a vector I? E R E , conclude with one of the following: (i) asserting that B E a(G),or (ii) finding a vector D E W" such that IIfill2 1 and for every E B(G),
DTXs DTB.
x
In principle, this separation problem reduces to checking the positive semidefiniteness of a symmetric (n, n)-matrix. Thus, given B E R" we extend B to a symmetric (n,n)-matrix B = ( b , ) with trace one and biJ = bji = O if I , j E E ( G ) . To assert that B E B(G)we have to prove that B is positive semidefinite. There are various characterizations of positive semidefiniteness which can be used for the design of an efficient proof of this property. For
Polynomial algorithms for perfect graphs
343
instance, polynomial time algorithms can be obtained from Gaussian elimination and Cholesky decomposition. In Gaussian elimination we allow pivots on the main diagonal only; if the rank of the matrix is found and only positive pivots have been carried out, then the matrix is positive semidefinite. Cholesky decomposition can be used in a similar way. A further method is to compute the smallest eigenvalue A. ; if A. is nonnegative then the matrix is positive semidefinite. This algorithm may be fast in practice but is not necessarily polynomially bounded. The method we shall describe now is based on Gaussian elimination and certain determinant calculations. In the following we assume that a graph G with n vertices is given. B(G) and @(G)are the sets defined in (4.12) resp. (5.1)
(5.5) Separation algorithm for B(G), SEP(G, B). Given a vector ii=n+(;)-[E(G)I-l. (5.5.1) Extend
B
B E Q', where
to a symmetric (n,n)-matrix B with trace 1.
(5.5.2) Use Gaussian elimination to compute the rank, say k, of
B
and a principal (k, k)-submatrix of B having full rank. (Note that Gaussian elimination automatically gives a nonsingular (k, k )submatrix, say BIJ(where I is a row- and J a column-index set with I I I = I J I = k), of B.Since tr(B) = 1, k is nonzero. It is well-known that if E1J is nonsingular and has the same rank as B,then both principal (k, k>submatrices Blr and Bjj are nonsingular, in fact de t(Bl,)det( Bjj ) = det(BIJ)'. Let us denote the principal (i, i)-submatrix of B consisting of the first i rows and columns of B by B i . For ease of exposition we assume that the principal (k, k)-submatrix Bk of B has rank k and is the one obtained in step (5.5.2). One can easily show that B is positive semidefinite if and only if B, is positive definite. Moreover, positive definiteness is easy to check, namely, B.4 is positive definite if and only if det(Bi)>O for i = 1,. .. ,k. Therefore our algorithm continues as follows.) (5.5.3) Compute det(Bi) for i = 1,. . . ,k.
(5.5.4) If det(Bi) > 0 for i = 1,. . . ,k, then B is positive semidefinite and hence B E @(G) is proved. stop! (If the test in (5.5.4) is failed, then B is not positive semidefinite and we have to calculate a separating hyperplane.)
M. GrotscheI ei al.
344
(5.5.5) Let t be the smallest index such that det(B,) S 0 and define a vector d = ( d , ,. . . ,d,)T as follows: di:=O for all i > r, d,:=-l
if t = 1,
d , : =( - l)'det(B,),
i
=
1,. .. , t, if t > 1,
where Biz denotes the (r -1,r-1)-submatrix of B, obtained by removing the i-th row and r-th column from B,.
(53.6) Define the following vector
dii:= d ; - d f , i
D E Q"
= 1,.
. . ,n - 1,
dij:= -2didj, for i , j e E ( G ) and i < j , and return D (if 111)11 < 1 we have to scale such that 1 1 011 2 1). 0 The vector D E W" gives the desired separating hyperplane in case B does not belong to B ( G ) . More exactly
D x s d : s DB
for all X E L%(G).
(5.6)
To prove (5.6) define the (n,n)-matrix D = (d,,) by setting d,, = d,d,, i.e., D = dd'. It is obvious from the definitions of D and 0 that for every vector X E R" and its extension X E R""" we have n
D'X
= d,,,
-
C d,X,d, = dZ.- dTXd.
r., = I
Now if E @ ( G ) then , the extension X is positive semidefinite, i.e., dTXd 2 0, which implies D ' x S d : . If B and B is the extension of 6 then the following holds:
x=
d'Bd =
2 didjbij= det(B,)det(B,-JG t.j
=1
0,
(5.7)
where in case t = 1, det(Bo) is assumed to be one. (5.7) can be obtained by exploiting the definition of d, cf. ( 5 . 5 . 9 , and using determinant expansions. Thus, (5.7) shows that DB == d : and (5.6) is proved. Altogether we have used Gaussian elimination once in (5.5.2) and we have performed at most 2n determinant calculations in (5.5.3) and (5.5.5). Since Gaussian elimination and determinant calculation can be done in O ( n 3 )time the overall running time of our separation algorithm is at most O(n") (not considering the length of numbers). Summarizing the discussion above we get the following theorem.
Polynomial algorithms for perfect graphs
345
(5.8) Theorem. There exists an algorithm SEP( . , . ) such that for any graph G with n vertices and any vector B E Q', 6 = n + (2)- IE(G)I - 1 , SEP(G, B) asserts whether B E B(G)or produces a vector D E Q Esuch that < BB for all Ea(G). The running time of SEP(G,B) is bounded by a polynomial in n and in
Dx
x
r 1%
It B 11-1
'
To give an example for the sets 93(G), & ( G )and the separation algorithm for & ( G ) we consider the graph K 2 which has two vertices and no edges. Then
4 (K2)= {(a,b)T1 0 s a s 1, ( a - f)' + 6' 6 i}, i.e., B((K2)is the ball in R2 around the point (f,O)' with radius f. Consider the point B = (4,l)'E [wz. The extension of B is the matrix B = j,1 i If ) .
-:.
B has rank 2 and det(B1)=$, det(Bz)=det(B)= So B is not positive semidefinite and the smallest index t with det(B,) < 0 is t = 2. Using (5.5.5) and (5.5.6) we obtain d , = - 1 , det(B12)= - 1, d2 = det(BZz)= i, and hence d,, = d:-d:=-$,d,2=1,i.e.,D=(-~,1)T.Thus,by(5.6)wehave
--
--
DX= -ia+b
)!j
for all 2 E 3(K2). The set %((Kz), the hyperplane shown in Fig. 5.1.
'b 1-
- 8 -I
fix = f and the point B are
-
B-.
a
/
Fig. 5.1.
M. Gliitschel el al.
346
6. Polynomial algorithms to compute & (G ) for all graphs and to solve the weighted stable set and clique problem in perfect graphs
We shall now use the separation aigorithm SEP( . ,. ) for B ( G )described in the foregoing section as a subroutine of the ellipsoid method to compute 6, ( G ) for every graph G and to find a maximum weighted stable set (or clique) of a perfect graph. Recall that for any graph G with a weight function w : V + Z + Theorem (4.20) implies &(G)=max
{
ij=l
G j b i jIB = ( b i j ) E 9 ? ( G ) ) ,
i.e. 6,(G) is the maximum value of a programming problem with a convex feasible region and with a linear objective function (having a particular form). We saw that it is necessary to replace B(G) by the convex body B ( G ) (cf. (4.12) and (5.1)) for technical reasons. We therefore have to replace the optimization problem over B(G) by a corresponding optimization problem over B(G). So, given an objective function c,.,c,,b,,such that c,, = c,, for B(G) as above, we then define the following objective function C for a ( G ) by setting E,,:=c,, - c,,,
C,,:=2c,, if
i
=
1,.
. . , n - 1,
i,jP E ( G ) ,i < j .
Defining
it is then easy to see that
in order to approximate 6, ( G ) it is sufficient to approximate &c (G) Therefore, for c,, = V w,w,. Since the numbers d G occurring in the objective function are not necessarily rational, we have to approximate these numbers in such a way that the optimum of the problem with the perturbed objective function does not differ too much from the true optimum. Suppose we want to calculate 6, (G) up to an error E > 0, then we claim that the following approximation of the V G ' ,is sufficient. Using, e.g., the method of continued fractions, determine rational numbers u,, with
Polynomial algorithms for perfect graphs
347
where, in addition, the denominators of the u,, are at most 2 n ( n solve the program
+ l ) / e . Then we
Assume that B E B ( G )is the true optimum solution of this program, then since 1) B 1) Q 1 , no component & of B is larger than one in absolute value. This implies that the error we make with respect to the original objective function is com pone n t wise
As the number of components ti of B is at most n(n + 1)/2- 1 and the error in u,, is also at most ~ / 2 n ( + n 1 ) we have from (6.1) that
1
8 u
( G) + u n n
- 6,
( G )1 < E /2 .
In other words, if the u,, are chosen according to (6.2) then the desired number 19,( G ) is contained in the interval (8,( G )+ u,, - ~ / 2 &,, , ( G )+ u,, + ~ / 2 )This . implies that if we compute 8 , ( G ) + u , , up to an error ~ / we 2 obtain 6,(G) within an accuracy of E. Such an approximation can be achieved with the ellipsoid method. So suppose a graph G, 1 V ( G )1 = n, with weight function w : V(G)-+ Z, and a required accuracy E > 0 for 9, ( G ) are given, then the following algorithm THETA ( G ,w, E, 7 ) finds a number T with 17 - aW (G)(< E. (6.3) Algorithm. THETA(G, w, E, 7).The graph G = ( V ( G ) ,E ( G ) ) , 1 V ( G ) )= n 3 2, the natural numbers w,,i E V ( G ) ,and the rational number F > 0 are the input of the algorithm, while the number 7 is the output of the algorithm.
vGj,
(6.3.1) Approximate the numbers 1 i s j s n, by rationals u,j satisfying (6.2) and whose denominators are at most 2 n ( n + 1 ) / ~ Set . ri,, := u,, - u,,,
i
=
1 , . .., n
-
1 and ri,j: = 2u,, for i
< j, i, j E E ( G ) .
(We now approximate the optimum value of the program max{GTB B E B ( G ) }+ unn up to an error ~ / using 2 the ellipsoid method.)
1
(6.3.2) Set r = l / n z f i , R = 1, E : = ~ / 2 -~ / 2 n ( + n 1) and define the paramters N, 6, p as in the ellipsoid method (2.3).(For the choice of the radii, cf. (5.2), the accuracy E is chosen according to the previous discussion.) (6.3.3) Set Ao:=RZZ,and choose as center xo of the first ellipsoid the projection B, E B ( G )of (l /n )Z n .
M.Grotschel et al.
348
(Recall that ii = n + (;) - [ E ( G ) (- 1, and that by (5.2) B, is an interior point of $%(G).) (6.3.4) for k = 0 to N - 1 do; 1. Run the separation algorithm SEP(G,xk) defined in (5.5). 2. If x k E & ( G ) then set a:= a. 3. If X k E B ( G ) and if 0 is the vector returned by SEP(G,xk), cf. (5.5.6), then set a:= - D. 4. Make the updates of the ellipsoid method as described in (2.3.9)-(2.3.11). End; (6.3.5) Let cr be the value of the best feasible solution of rnax{z'i'B B E a ( G ) } found in (6.3.4). Set T : = m + u,, and return T. 0
I
Algorithm (6.3) describes how we can approximte 9, (G) in polynomial time. Letting w be the vector all of whose components are one, we can use algorithm THETA to compute 6 ( G ) up to any given accuracy in polynomial time for every graph G. So 9( G) is not only well-characterized by the formulas (4.3,(4.6), (4.9) and (4.13), it is also well-behaved computationally. Theorem (3.2) and Theorem (4.20) imply that for perfect graphs the numbers a,(G) and & ( G ) coincide. Moreover, since our weight function w is integer valued, we know that the value a, (G) of the optimum weighted stable set is an integer. Therefore, in order to find the optimum value of a weighted stable set problem on a perfect graph we only need to approximate 6,(G) up to an error E c f with the algorithm THETA(G, w, E , T ) and round T to the next integer to obtain a, (G). We can also use the algorithm THETA to find a maximum weighted stable set explicitly. This goes as follows. Let a graph G = (V, E) with n z= 2 vertices, weights wi E Z, for all i E V and an accuracy O < E s $be given. (6.4) Algorithm. STABLESET(G, w, E ) .
(6.4.1) Initialization and first guess for a(G): 1. Run THETA(G, w, E , T ) and round T to the next integer, say f. (Clearly, r 3 a ( G ) and if G is perfect then t = a (G).) 2. If 1 r - T 13 E, then stop and conclude that G is not perfect. 3. Call all vertices of G unlabeled. (6.4.2) Termination check If all vertices of the present graph G are labeled, then do:
349
Polynomial algorithms for perfect graphs
1. If V ( G )is not stable then stop and conclude that G is not perfect. 2. If V ( G )is stable then V(G) constitutes a maximum weighted stable set of the original graph. stop! (6.4.3) Choose an unlabeled vertex z, E V and tentatively remove u from G, i.e., set G‘ = G - u and set w L = w, for all u E V \{ u } . (6.4.4) Run THETA(G’, w ’ , E, T ) and round T to the next integer, say s. If I s - T 12 E , then stop and conclude that G is not perfect. (6.4.5) If s = t, then remove u definitely, i.e., set G = G ’ and w
= w’.
(6.4.6) If I V(G)I = 1, then label the remaining node. (6.4.7) If s < t, then label u. (6.4.8) Go to (6.4.2). 0 To prove the correctness of the algorithm suppose first that G is a perfect graph. Then for every induced subgraph G’ of G, 6,.(G’) = a,.(G’). In step (6.4.1) we approximate & ( G ) by F G $ , and, therefore, rounding 7 to the next integer gives the true value t for a ( G ) .Now we remove a vertex u from G. If the number s calculated in (6.4.4) satisfies s = f then we know that G - u contains a stable set which is a maximum weighted stable set of G, so we can remove u without distroying all optjmum solutions of the stable set problem for G, and we can continue with this procedure. If however s # t, then all optimum stable sets of G necessarily contain vertex u. Thus we label u, keep u in our vertex set and continue. This way we will finally end up with a graph G ’ whose set of vertices is labeled, i.e., none of the vertices can be removed without reducing s = a,.(G‘)= a, ( G ) = c. This means that every vertex of G’ is contained in all optimum stable sets of G’. In other words, the vertex set of G ’ is itself a stable set, and since a,.(G’) = a , (G), this vertex set is a maximum weighted stable set of G. Thus if G is perfect, then STABLESET will produce an optimum weighted stable set of G. If however G is not perfect, then STABLESET may detect the imperfectness of G but may also deliver a maximum weighted stable set (without recognizing the imperfectness of G). If in step (6.4.1) or (6.4.4) we find that ] t - T I 3 E resp. I s - T 1 2 E then the interval ( T - E , T + E ) contains no integer. Since cu,.(G’) is an integer for all induced subgraphs G‘ of G and THETA guarantees 6,.(G‘)E (T - E , T + E ) this implies a,.(G’)# IL(G‘),i.e., by Theorem (4.18) we can conclude that G is not perfect. It may however happen that in every step (6.4.1) and (6.4.4) the approximation 7 of 6,JG’) is in the &-neighborhood of an integer and we will end up in step (6.4.2) with an induced subgraph G’ of G whose vertex set V’ is labeled. If V‘ is not a stable set, then V’ is not a solution of our
M . Grotschel et nl.
350
stable set problem. Since the algorithm works for perfect graphs, we can conclude that G is not perfect. If V' is a stable set we have to show that V' is in fact a maximum weighted stable set of G. This can be seen as follows. Suppose G" is the last subgraph of G created during the algorithm such that a vertex, say u, was definitely removed from G". Then we know that all other vertices of G" will finally be labeled, so V(G")= V' U { u } . Moreover, u was removed because the number s obtained in (6.4.4) by running THETA(G"- u, w ' , E ,T ) satisfies s = t. Since V' is stable, G' = G"- u is a perfect graph, so we have s = a,.(G'), and since t is an upper bound for LZ, (G), s is the weighted stability number of G. This proves our claim. The following (imprecise) argument shows that in case of imperfect graphs all outcomes described above are possible. Consider the pentagon G, then 9(C,) = L 5 = 2.236.. . . Suppose we run STABLESET with E = 4, then the T we get in step (6.4.1) may equal 2.1 or 2.6. If T = 2.1 then t = 2 and the algorithm will continue finding a maximum stable set of C5. If T = 2.6 then t = 3, and since the removal of every vertex from Cc results in a perfect graph we shall get s = 2 every time we execute step (6.4.4). This means that finally all vertices of C, will be labeled, but these of course do not solve our problem. If however we had chosen E =0.1, then ~ € ( . \ / 5 - 0 . 1 , V ' ~ + O . I ) a n d w e o b t a i nt = 2 a n d l r - ~ I s 0.13 > E . This implies that the algorithm would stop in step (6.4.1) concluding that C, is not perfect. Thus algorithm STABLESET has two possible outcomes. Either a maximum weighted stable set of G is found or imperfectness of G is proved. Note that in the former case it does not prove perfectness. STABLESET can also be used to find a maximum weighted clique of a perfect graph G. We simply run STABLESET on the complementary graph G which by the perfect graph theorem is perfect again. Summarizing the observations of this section we obtain the following theorem.
(6.5) Theorem. (a) There exists an algorithm which for any graph G = ( V , E ) , 1 VI 2 2, any weight function w : V-, 8, and any rational E > 0 ,findsa number r such that I T - a,(G)J < F holds. The running time of this algorithm is polynomial in I V ( , [log((w 11-1 and rlog I1
f
11-1 .
(b) There exists an algorithm which for any perfect graph G = ( V , E ) , I V I 3 2 and any weight function w : V+H+ finds a maximum stable sef (resp. maximum weighfed clique) and the running time of which is polynomial in I VI and [log It w 11-1 '
Algorithm THETA and inequality (4.16) can be combined to design a polynomial time nondeterministic algorithm which checks the imperfectness of a
Polynomial algorithms for perfect graphs
35 1
given graph. Namely, suppose G' is a critically imperfect graph with n vertices, then choosing a suitable E , e.g., E = in-'", we run THETA(G', e, E , T ) where e is the vector all of whose components are one. By Theorem (2.4), the choice of F , and inequality (4.16) the number T we obtain satisfies T E(8(G') - E
, ~ ( G ' E) )+,
T
1 W(G)+€
and therefore T - F > a ( G ' ) , T + E < cu(G)+ 1. Since lo g ( ( €(Ix is polynomial in n, THETA runs in time polynomial in n. In other words, given a critically imperfect graph, we can verify its imperfectness in polynomial time. As every imperfect graph contains a critically imperfect graph, say G', we can guess this graph G ' and then apply the algorithm described above. This shows that verification of imperfectness is an NP-problem, hence verification of perfectness is a co-NP problem. Note that if the strong perfect graph conjecture is true, then this fact is trivial, since checking imperfectness would then be possible by guessing an odd hole or antihole.
7. A polynomial algorithm for the weighted clique cover and coloring problem for perfect graphs
The separation algorithm for g ( G ) presented in Section 5 provides us - as shown in Section 6 - via the ellipsoid method with a polynomial time algorithm for solving the weighted stable set problem for perfect graphs. Seen from a different point of view this means that the class of linear programming problems maxcTx, x E P ( G ) = conv{x W stable set in G}, G a perfect graph, is solvable in polynomial time. Since P ( G )is a fully dimensional rational polytope, Theorem (2.6) implies that the optimization problem as well as the separation problem for P ( G ) ,G perfect, are solvable in polynomial time, even in the strong sense. By Theorem (3.2), for a perfect graph G the stable set polytope P ( G ) equals the fractional stable set polytope P * ( G ) , thus for this class of graphs we can decide in polynomial time whether a given vector y belongs to
1
P ( G ) = P * ( c ) = ( ~ ( ~all, ~ o E ~v ~, ~
c x, s I
"EC
for all cliques
c c v(G)/ .
A different approach to solving the separation problem for P ( G ) not using
Theorems (2.5) and (2.6) is of course to apply the algorithm which finds a maximum weighted clique, where the given vector y 3 0 is used as the vector defining the objective function. If the maximum clique, say C, satisfies y r ~ ='
M. Grotschel et al.
352
cuEcyu d 1 then y E P ( G ) ,otherwise this clique inequality provides a separating hyperplane. Since the optimum clique algorithm is nothing but the optimum stable set algorithm applied to the complementary graph G (which is also perfect), we can use the algorithm STABLESET directly to solve the separation problem for P ( G ) , G perfect. Theorem (2.6) has a further important consequence. Since the optimization problem for the class of rational polytopes P ( G ) , G perfect, is solvable in polynomial time we can find facets of P ( G ) and rationals A, 2 0 satisfying the conditions of statement (c) of (2.6). Since the facets of P ( G ) are of the form - x u SO, u E V ( G ) ,and cvGcx, G 1, C c V ( G )a maximal clique, Theorem (2.6) (c) implies that for any objective function w : V(G)-+Z+ we can find in polynomial time (maximal) cliques C C V(G) and positive rational numbers A,, i = l , ..., r S l V ( G ) l such that
Suppose for a node u E V ( G )strict inequality holds in the above inequality, say u,:=C:=l,,,c,A, - w , > 0, then pick a clique, say C,, which contains u. If A, s u, then replace C, by the clique C, \ { u } , otherwise add the new clique C, \ { u } as clique C,&, to our list of cliques and define new parameters as follows: A,:= A, - u L , A,,, := u,. Then the sum of the A,'s still equals a, ( G ) and the gap u, of the inequality corresponding to u is either zero or is strictly reduced. By continuing this process we end up wirh a list of cliques C,,, . . , C, and positive rationals A , , . . . A, such that
.
2 A, = a,(G). I
=I
8
=I.
A, = w , for all u E V(G). C i E C ,
Note that in the algorithm described above only those vertices u E V ( G )were considered for which the inequality - x, d 0 had a positive multiplier A,. Since for every such vertex at most one additional clique was added, we still have r 6 1 V(G)\. By definition, for a perfect graph G the stability number a ( G ) equals the clique cover number p ( G ) . Moreover. since for a perfect graph G the graph G, (cf. Section 3) is also perfect and as a,(G)= a(G,), p w ( G ) =p ( G , ) , the weighted clique cover number pw ( G ) equals the weighted stability number a w ( G ) .This implies that for a perfect graph G, algorithm THETA (or STABLESET) also calculates pw (G), thus, by definition, there exist integers A, > 0 which satisfy (7.1). Note that the numbers constructed by the algorithm of Theorem (2.6) ( c ) (plus scaling afterwards) need not be integral in general.
Polynomial algorithms for perfect graphs
353
However, we can find such integers for perfect graphs. We first show how this can be done in the cardinality case, i.e., w = (1,. . . ,l)T. Assume that G = ( V ,E ) , I V 1 3 2, is a perfect graph and we want to find a clique cover of G. (7.2) Algorithm. Cardinality clique cover.
(7.2.1) Apply the algorithm of Theorem (2.6) (c) to find cliques CiC V and (possibly nonintegral) rationals hi > 0, i = 1,. ..,t satisfying (7.1). (We claim that every clique C, with hi > 0 intersects every stable set of G of cardinality a ( G ) . Suppose C, is such a clique and W C V is a maximum stable set with W f l Ci= 0. Then (7.1) implies
i#/
a contradiction. Therefore we continue as follows.) (7.2.2) Remove clique CI from G, i.e., set G:= G - CI. If V ( G )= 0. stop. Otherwise go to (7.2.1). (Note that the graph G' obtained from G by removing clique CI satisfies a ( G ' )= a ( G )- 1 since every maximum stable set of G will lose one vertex. So after exactly a ( G )executions of (7.2.1) and (7.2.2) we have found a ( G )cliques which cover G.These cliques are those which have been removed in (7.2.2). Note also that every vertex of V is contained in exactly one such clique and that these cliques are not necessarily maximal cliques of G.) 0 Since the algorithm of Theorem (2.6) (c) can be shown to be polynomial in I V I the overall running time of algorithm (7.2) is also polynomial in I V I. We shall now extend this algorithm to the weighted case. Assume that a perfect graph G = ( V ,E ) and a weight function w : V +H, are given. (7.3) Algorithm. Weighted clique cover.
(7.3.1) Apply the algorithm of Theorem (2.6) (c) to find cliques C C V and (possibly nonintegral) rationals hi 0, i = 1 , . . . , f, satisfying (7.1).
=-
(7.3.2) Set A ::= LA,J, i = 1,. ..,t,
and construct the graph G,.. (Since hi - A : < 1 we obtain from (7.1)
M.Grotschel el al.
so the graph G,. obtained by replacing every node of v by w l nonadjacent copies and linking two nodes in G,. by an edge if their originals in G are adjacent has less than I V )z vertices. This implies that G,, can be constructed from G in time polynomial in I V ( and [log II w 11-1
(7.3.3) Apply algorithm (7.2) to G,, to obtain cliques D : , i = 1,. . . ,a(G,.) covering each vertex of G,, exactly once. (Since every clique D ;of G,. contains at most one copy of every vertex u E V ( G ) ,every DI corresponds to a clique, say D,, of G. Note that for 0: # 0; the corresponding cliques D , , 0,of G may be identical.) (7.3.4) Construct the cliques DI,. . . ,Do(G,,) of G corresponding to the cliques DI,. . . ,DL(G,.) of G,.. Let B I , .. .,B, be the different cliques occurring in the sequence D,, i = 1 , . . .,a(G,.) and let p,, j = 1 , . . . ,r, be the number of times clique B,occurs in the sequence D,, i = 1,. . . ,a(G,.). Then proceed as follows: Set k = t and for j = 1to r do; If B, is equal to one of the cliques C,, i E 11,. . . ,I } , then set A: := A + p,. Otherwise set k : = k + l , AL:=p, and Ck:=B,. 0
We claim that the cliques C, and integers A:, i = 1,. . . ,s, defined in step (7.3.4) solve the clique cover problem considered. Obviously, in the above algorithm every vertex u E V is covered w, - w : times after the execution of step (7.3.2). By applying algorithm (7.2) to the graph G,. and making the construction described in (7.3.4) every vertex will be covered a further w L times. So the cliques C1,. . . ,C, and integers A :, .. . , A I satisfy
Similarly, note that G,. is designed in such a way that a ( G , . ) = a,(G)-C:=I(A,- [ A , ] ) holds. Since G is perfect G,, is also perfect, so p(G,.) = XI=, p, = a(G,.) which implies that the A: defined in (7.3.4) satisfy ,=1
A', = a , ( G ) = p, (G).
Thus algorithm (7.3) produces the desired solution of the clique cover problem for a perfect graph. Since the algorithm STABLESET, the algorithm of Theorem (2.6) (c) and the algorithm (7.2) run in time polynomial in 1 V(G)1 and [logl)w 11-1 the overall
Polynomial algorithms for perfect graphs
355
running time of algorithm (7.3) is polynomial in I V(G)I and [log\(w 11-1 for every perfect graph G and every objective function w : V(G)+Z+. As before it is now easy to obtain a polynomial time algorithm for the weighted coloring problem for perfect graphs. Since the weighted chromatic number x w ( G )equals the weighted clique cover number p , ( G ) of the complementary graph G we simply apply algorithm (7.3) to the perfect graph G which will yield the desired optimum weighted coloring of G.
8. Conclusions
In the previous section we have described polynomial time algorithms for various linear programming problems on perfect graphs. All these algorithms are based on the ellipsoid method and use a polynomial time separation algorithm for a convex, nonpolyhedral set. Although these algorithms are polynomial (and thus are theoretically good) we do not recommend them for practical use. Just for curiosity we have done some computational experiments with the separation algorithm for G(G) described in Section 5. As expected, the numerical problems were such that even for small problem sizes, say V(G)I equal to 10 or 20, it was almost impossible to obtain a correct answer. An alternative approach is to use (4.9) for the design of a polynomial algorithm to compute 6(G). This amounts to minimizing a convex function on an affine space. In principle, this can be done by the ellipsoid method in polynomial time. In practice, it is probably better to use some simpler descent method. The first experiments with this dual approach seem to be more promising. Our analysis of these problems should be viewed as a theoretical contribution showing that certain programming problems for perfect graphs are indeed polynomially solvable. Future research should be directed toward finding practically good algorithms for these problems. These algorithms should have a more combinatorial nature and should not suffer from the numerical instability (due to our present-day computer technology of fixed precision arithmetic) of the ellipsoid method and the separation problem for B ( G ) .
I
References [ l ] V. Chvital, O n certain polytopes associated with graphs, J. Comb. Theory, Ser. B 18 (1075) 138-154. [2] P. Gacs and L. Lovasz, Khachian’s algorithm for linear programming, Math. Program. Studies 14 (1981) 61-68.
356
M. Grotschel et al.
131 M.R. Carey and D.S. Johnson, Computers and Intractability, A Guide to the Theory of NP-Completeness (Freeman, San Francisco, 1979). [4] M. Grotschel, L. Lovasz and A. Schrijver, The ellipsoid method and its consequences in combinatorial optimization, Combinatorica 1 (1981) 169-197. [ 5 ] W. Haemers, On some problems of Lovasz concerning the Shannon capacity of a graph, IEEE Trans Inform. Theory IT-25 (1979) 231-232. [6] W.-L. Hsu, How to color claw-free perfect graphs, Ann. Discrete Math. 11 (1981) 189-197. [7] W.-L. Hsu and G.L. Nemhauser, Algorithms for minimum coverings by cliques and maximum cliques in claw-free perfect graphs, Discrete Math. 37 (1981) 181-191 (this volume, pp. 357-369). [8] L.G. Khachiyan, A polynomial algorithm in linear programming, Dokl. Akad. Nauk SSSR 244 (1979) 1093-1096 (English transl. Soviet Math. Dokl. 20 (1979) 191-194). [9] L. Lovasz, Normal hypergraphs and the perfect graph conjecture, Discrete Math. 2 (1972) 253-267 (this volume, pp. 29-42). [lo] L. Lovasz, On the Shannon capacity of a graph, IEEE Trans. Inform. Theory IT-25 (1979) 1-7. [ I I] M.W. Padberg, Almost integral polyhedra related to certain combinatorial optimization problems, Linear Algebra & Appl. 15 (1976) 6Y-88. [ 121 C . Shannon, The zero error capacity of a noisy channel, IRE Trans. Inform. Theory IT-2 (1956) 8-19. [I31 N.Z. Shor, Convergence rate of the gradient descent method with dilatation of the space, Kibernetika 2 (1970) 80-85 (English transl. Cybernetics 6 (1970) 102-108). [ I41 D.B. Yudin and A. S. Nemirovskii, Informational complexity and effective methods of solution for convex extremal problems, Ekonomika i Mat. Metody 12 (1976) 357-369 (English transl.: Matekon: Transl. of Russian and East European Math. Economics 13 (1976) 24-25).
Annals of Discrete Mathematics 21 (1984) 357-369 @ Elsevier Science Publishers B.V.
ALGORITHMS FOR MAXIMUM WEIGHT CLIQUES, MINIMUM WEIGHTED CLIQUE COVERS AND MINIMUM COLORINGS OF CLAW-FREE PERFECT GRAPHS* Wen-Lian HSU Department of Industrial Engineering and Management Science, Northwestern Universiry. Evanston, Illinois, V.S.A.
George L. NEMHAUSER School of Operations Research and Industrial Engineering, Cornell University, Ithaca, New York, U.S.A.
For the class of claw-free perfect graphs with weights on the vertices, we give efficient algorithms for finding a maximum weight clique and a minimum weighted clique cover. We also present an algorithm for minimum cardinality coloring of these graphs, but the minimum weighted coloring problem remains unsolved.
1. Introduction
One of the unsolved problems for perfect graphs is to find efficient combinatorial algorithms for the maximum weight independent set, maximum weight clique, minimum clique cover and minimum vertex coloring problems. * * Graphs that do not contain claws, odd holes and odd anti-holes are known to be perfect (see [13], [9], [3]). We consider this family of perfect graphs. Efficient combinatorial algorithms for the maximum independent set problem on claw-free graphs, including the weighted case, have been given by Minty [ 121 and Sbihi [14]. However, one cannot expect to have efficient algorithms for the maximum clique or minimum clique cover problems on claw-free graphs because these problems are NP-hard ([7]). The edge coloring problem on general graphs can be reduced to vertex coloring on claw-free graphs [7]. Holyer [15] has shown that minimum edge * This research has been supported, in part, by National Science Foundation Grant ECS-8005350 to Cornell University. The results of this paper have appeared in Hsu’s unpublished Ph.D. Dissertation 17) and also in Refs. [6], [8]. * * Grotschel, Lovasz and Schrijver [4] use Khachian’s ellipsoid method [ l l ] to obtain polynomial algorithms for a number of combinatorial optimization problems including the four problems just noted on perfect graphs. Although these results are theoretically very significant, the ellipsoid algorithm is not practically efficient. 351
358
W.-L.Hsy G.L.Nemhauscr
coloring of general graphs is NP-hard and consequentially minimum vertex coloring on claw-free graphs is NP-hard as well. However, when we assume perfection as well as claw-freeness, these problems become tractable, as we have shown in our recent papers ([6], [8]). Here we present the combinatorial algorithms of these papers for maximum weight cliques, minimum weighted clique covers and minimum cardinality vertex-colorings of claw-free perfect graphs. The coloring problem in the weighted case remains unsolved.
2. An algorithm for maximum weight cliques on claw-free perfect graphs The algorithm is a simple consequence of the following lemma.
Lemma 1. Let G be a claw-free perfect graph and let u be any vertex of G. Let N ( u ) be the neighbors of u and let Gnbe the subgraph of G induced on N ( u ) . the complement of G O , is bipartite.
a",
Proof. Since the subgraph induced by N(u) U { u } is claw-free and every vertex of G" is adjacent to u, Gncannot have triangles. Furthermore, Gnhas no odd anti-holes so that Gohas no odd holes. A graph without triangles or odd holes has no odd cycles. 0
Since G" is bipartite, we can find a maximum weight vertex packing (independent set) P, on Go efficiently by the O(n3)maximum flow algorithm of Karzanov [lo]. f, U { u } is a maximum weight clique in G that contains u. Thus by enumerating over all u E G we can find a maximum weight clique of G in O ( n 4 )time.
3. An algorithm for minimum weighted clique covers Let A be the clique-vertex incidence matrix (abbreviated as clique matrix) of the perfect graph G, i.e., each column of A corresponds to a vertex of G and each row of A corresponds to a clique of G. Chvital [l] proved that every extreme point of the polyhedron P = {x E R : Ax d 1) is a (0,1)-vector, from which it easily follows that there is a one-to-one correspondence between extreme points of P and packings of G. Hence we can formulate the maximum weight vertex packing problem on a perfect graph as the linear program*
1
' Rows of A that correspond to non-maximal cliques are obviously superfluous. However, we retain them in the problem description because our clique cover algorithm uses non-maximal cliques.
Algorithms
(WW
359
max w . x, Ax s 1, x 30,
where w is the vector whose components are the vertex weights wt(u) for all u E V. Without loss of generality, we will assume that w is strictly positive since vertices with non-positive weights can always be assigned a value of zero in some optimal solution. The dual of the above program is the linear programming formulation of the weighted clique cover problem WCC)
min 1 . y, yA 2 w, y 30.
When w is an integer vector, perfection implies that (WCC) always has an integral optimal solution (see [2]). However, our approach can solve (WCC) for any real w 3 0. Of course, if an integral y is required then w is also required to be an integer vector. Since G is claw-free, Minty’s algorithm can be used to find a maximum weight packing P of G. Our approach is to begin with P and then to find a feasible solution y * to (WWC) with the property that 1 . y * = wt(P), which implies (by weak duality) that y * is an optimal solution to the weighted clique cover problem.* We use the fact, also implied by duality, that for u E P there exists a clique C containing u and an E , 0 < E < wt(u), such that P is a maximum weight packing on graph G with weights given by wt(u)- E for u E C and wt(u) otherwise. The difficult problem is to find such a weighted clique efficiently. A recursive application of the weighted clique-finding procedure on reduced graphs will yield an optimal weighted clique cover for graph G. To make this idea precise, we first define graph reduction and then the construction of the clique cover for G from the weighted cliques. Finally, we will give the algorithm for finding the weighted cliques. ) V ( H 1 and ) E ( H 2 )C E ( H I ) .Let Let HI,Hz be two graphs such that V ( H 2 C wtl( .) and wtz(. ) be weight functions on the vertices of H1 and H z respectively. Define the reduced graph of H I by H 2 , HIIH2,to be the graph H’ with
I E ( H ’ )= { e 1 e = (x, y ) E E ( H ~ )x,, y E v(H’)I,
V ( H ’ )= { u u E V ( H I )and wtl(u) > wt2(u) if u E V ( H 2 ) } ,
* We also used this approach in the unweighted case [9]; however, the technique used here to construct the clique cover is quite different from the technique that we gave in [9].
W.-L.Hsu,G.L. Nemhauser
360
and weight function wt'( - ) defined on V ( H ' ) by wt'(u)
wt,(u)
if u E V (HI ) \V ( H 2 ) ,
wtl(u)-wt2(u)
if u E V(H,)n V ( H 2 ) .
=
Given a maximum weight packing P and u f P, the algorithm finds a clique C:, with weight yc: > 0 such that p is a maximum weight packing on the reduced graph G I C f , .The algorithm continues the clique finding and reduction procedure until the weight of u is reduced to zero. At each step the optimality of P is maintained. Let C, be the set {C:}rA;'of weighted cliques containing u that have been deleted during the process and call the reduced graph at this stage G,. By construction
and P\{w} is a maximum weight packing in G,. Then we replace G by G, and P by P \ { u } and apply the algorithm recursively. When P becomes empty, the reduced graph will also be empty. Hence a recursive application of the algorithm will give rise to I P 1 sets { C",},u, E P, of weighted cliques such that each vertex u of G is covered by at least wt(o) cliques and for each u, E P K(u
c'
yr;,= wt(uJ) in G.
,=I
Thus
so that the set u,",{Cu,} is an optimal solution to the weighted clique cover problem. During the procedure, we also have to consider a graph with weights on the edges, which is defined as follows. Let G be a weighted claw-free perfect graph with positive weight on each vertex. Let u be a vertex in a maximum weight packing P in G. Since G is claw-free, 10 n N ( u ) l S 2 for any packing Q in G. From Lemma 1, we know that Go is bipartite. For all (x, y) E E ( G " ) define edge weights by wt((x,y))=max{wt(Q)( Q is a packing in G, { x , y } E 0)-wt(P\{u}) and for all u E N ( u ) , define tower vertex weights by
1
wt(u) = max{wt(Q) Q is a packing in G, Q -wt(P\{u}).
n N ( u )= { u } }
Algorithms
361
Note that wt((x,y)) and wt(u) can be found by Minty's algorithm [4]. Construct a weighted graph H, as the graph Gowith usual vertex weights as in G and edge weights and lower vertex weights as above. It is easy to show that P is optimal if and only if (1) wt(u) s wt(u) for all u E N ( u ) ; (2) wt((x, y ) ) s wt(u) for all (x, y ) E E ( H , ) . Therefore to maintain the optimality of P for the reduced graphs, we select the cliques {Cl}and their weights {yc;,}so that in the reduced graphs (1) and (2) are satisfied. The following algorithm selects the collection of cliques C,,: begin let G" be the subgraph of G induced on N ( u ) ; construct H, from G ;
C"+0; while wt(u)>O do begin w , +maximum lower vertex weight in H , ; w z +maximum edge weight in H, ; Wmax +max{ w 1, wzj; if w,,, < wt(u) then comment this is case 1 in the proof; begin let C be the clique { u } ; yc +wt(u)- w,,,; Cu + Cu U { C } ;wt(u)+ Wmax; end else comment now w,,, = wt(u); this is case 2 in the proof; begin A , +{v u E V(H.) and wt(v) = wmax}; comment A , is the set of vertices that must be included in the next clique; A r + { x x E V ( H , )and 3 y E V ( H , ) s.t. wt((x, y)) = w,J; E~ {(x, y wt((x, Y 1) = W m a J ; comment our next clique must include one end point of each edge in E Z ; HI+ a graph with V ( H I ) =A , U A 2 and E ( H , ) = E z with the same vertex weights and edge weights as in H , if A , = 0 then begin comment now w,,, = w 2 > w I and E , # 0 ; Fix an arbitrary bipartite partition of G'' and partition HI accordingly;
1 1
+
I
W.-L.Hsu,G.L. Nemhauser
362
end: else begin comment from linear programming duality there exists a clique in N ( u ) containing A l and one end point of each edge in E,. Hence the vertex set of HI can be partitioned into Vi and V2 such that all edges of Hi have o n e end in V, and the other in V2, and A I C V,. Furthermore, Vl is an independent set in G"; the procedure for constructing this partition, called PART(A,, H I ,G"), is given in the Appendix; PART(A,, HI,G"); let the LHS (left-hand side) of HI be V , ; end ; 514--W,,,-max{wt(v)I ~ P L H Sof H , } ; s1+w,,, - max{wt((x, y ) ) (x, y ) E E(HI)but {x, y } n (LHS of HI)= 81; s * +-min{sl,sz}; let C be the clique (LHS of H I ) U { u } in G with weight s * ; C,+C, U{C}; y c + - s * H,, + H , , / C ; wt(u)+-wt(u)-s*; for each u of H , in C \ { u } do w t ( u ) + w t ( u ) - s * ; for each ( x , y ) of E ( H . ) with either x or y in C do
1
wt((x, y )) + wt((x9 y 1) - s *; end; end; comment C,, is now the output of the algorithm; end.
Theorem 1. The algorithm outputs a ,finite collection of cliques {CL}such that P\!{U} is a maximum weight packing in the graph reduced by these cliques. Furthermore. yc;, = wt( u ).
z,
Proof. By definition, w,,,+wt(P\{u}) is the weight of a maximum weight packing in G that does not contain vertex u. Since P is a maximum weight packing in G, we have wmsrs wt(u). Consider two cases: (1) w,,,,< wt(u). Let C be the clique { u } in G with weight (wt(u)- wmaX). Clearly P still is a maximum weighted packing in GIC since now wt(u) = w,,, . (2) w,,,.,~ = wt(u). The optimality conditions (1) and (2) are satisfied in the reduced graph by the construction of C with weight y , = s*.
Algorirhms
363
To show the finiteness of the algorithm, we observe that s * > 0 and at each clique addition into C., one of the following will happen: (i) The weight of u will be reduced so that A l U A2# 0 in H,, (case (1)). (ii) At least one more vertex u of H, with wt(u) > 0 will be added to A l in the next iteration (when s * = sl in case (2)). It should be noted that once a vertex u is added to A 1 it will remain there until wt(u) is reduced to zero (since wt(u) = wt(u) at that stage). (iii) At least one more edge (x, y ) of H,, with wt((x, y)) > 0 will be added to E2 in the next iteration (when s * = s2 in case (2)). Again, once an edge (x, y ) is added to E2it will remain there until wt(u) is reduced to zero (since wt((x, y)) = wt(u) at that stage). Finally, once all vertices u of H , with wt(u) > 0 are added t a A l and all edges (x, y ) of H,, with wt((x, y)) > 0 are added to E2, then s * = s1 = sz = w,,,.~= wt(u). One more deletion of a clique with weight s * will reduce wt(u) to zero. Hence the number of iterations ( = number of Cl's) is bounded by O(l V(Ht)I+ IE(H'$I). Since W(u) is reduced by s * > O at each step, we have x,yc: = wt(u). Consider the efficiency of the algorithm. Since the final clique collection is composed of a set { C. u E P}, we only have to consider the construction of each C,,. The construction of H,, requires at most O(l V(d'))I') calculations of edge weights and lower vertex weights. Each such calculation requires an application of Minty's weighted packing algorithm, which is polynomially bounded. By Theorem 1, the number of iterations in constructing C,,is bounded by O(1 V(HO.))+ IE(H;)I). The procedure PART takes no more than O(l V(H.)21) time; the calculation of sI,s2 and s * takes at most O(max(1 V(Hu)l,lE(Hy)l)) time. Therefore, the entire algorithm is polynomially bounded.
1
4. An algorithm for minimum cardinality colorings on claw-free perfect graphs* Let f be a function mapping each vertex of G to a color. f is said to be a coloring of G if for any two vertices x and y, (x, y ) E E j f ( x ) # f ( y ) . Now consider a coloring f of G. Let i, j be two colors in f and consider the subgraph Gij induced on vertices colored i or j . An (i,j>path in G is a simple path with vertices colored alternately in i and j such that no two non-consecutive vertices are adjacent. Note that such a path may be a simple cycle.
R o p i t i o n 1. Each component of Gi, is either a single vertex or an ( i , j>parh. Actually, the presentation of this section only assumes that our graphs are clawdrtc and without odd holes or odd anti-holes.The algorithm gives yet another proof that this classof graphs is perfect.
W.-L. Hsu, G.L. Nemhauser
364
Roof. It suffices to show that each vertex in a component of Gij has at most two neighbors. Suppose not, and assume without loss of generality that y has at least three neighbors and f (y ) = i. Then since f is a coloring, all the neighbors of y are colored j and form an independent set. But then y and its neighbors contain at least one claw centered at y.
By switching of colors in an ( i , j>component, we mean changing all i's to j and all j ' s to i. Note that such a switching will still result in a coloring of G. Hence if x and y do not have the same color and are in distinct (f(x),f(y))-components, then we can switch the colors in one component so that x and y will have the same color. We use the fact that the number of colors in a minimum cardinality coloring is equal to the size of a maximum clique, w ( G ) , which can be found by the algorithm of Section 2. The algorithm colors the vertices successively to maintain a coloring that uses no more than o(G)colors. Thus a vertex is assigned an arbitrary color different from those of its already colored neighbors so long as these neighbors have used fewer than w ( G ) colors. However, if these neighbors have already used w ( G ) colors then we switch colors along appropriate (i,j)-paths. These paths are identified from the solution of an edge-matching problem as described below. Suppose that in a coloring of G \{u), N ( u ) uses w ( G ) colors. As in Section 2, let Gobe the bipartite graph which is the complement of the graph G" induced on N ( u ) . Since G is claw-free, at most two vertices in N ( u ) can have the same color. Let A4 be the set of edges in between two vertices with the same color. Then M is an edge-matching in G".
c"
kmma 2. If N ( u ) uses w ( G ) colors then M is nor a maximum cardinality matching in
c".
Roof. Let K be the set of colors that color two vertices of N ( u ) . Let
I
V, = { u u E N ( u ) , f ( u ) E K } . Then IV,I=21KI and J N ( U ) \ V I I = W ( G ) - I KHence J.
I N(u ) I = 1 Vi I 1 N ( u ) \ Vi I = 2 I K I + f
=w(G)+
0
(G ) - I K
1
I K 1.
Since Guis bipartite and the size of a maximum packing S in is =sw ( G ) - 1 , by Konig's theorem, the size of a maximum matching in G" is equal to IN(u)l-
Is 123 4 G ) + IK I - ( w ( G ) = IKl+
1.
1)
Algorithms
But IM 1 = 1 K
1.
Hence M is not maximum.
365
0
By Lemma 2 we have an augmenting path P in G" with respect to the matching M. O u r theorems show how this augmenting path identifies (i,;)-paths for color switching so that the number of colors for N ( u ) is reduced by one.
Theorem 2. If P consists of a single edge ( x ,y ) where f ( x ) = i and f ( y ) = j , then by switching colors along the (i,j)-path in G containing y and assigning vertex u the color j , we obtain a coloring for G that uses w ( G ) colors. Proof. It suffices to show that the (i,j)-path in G containing y does not contain x. Since (x, y ) kZ E ( G ) , an (i,j)-path joining x to y must contain at least two intermediate vertices. Since ( x , y ) $Z M, none of the intermediate vertices is in N ( u ) . But this implies that the cycle formed by the (i,j)-path together with the two edges ( u , x ) and ( u , y ) is an odd-hole. 0 Theorem 3. Let P = ( x o y l x l y *z* . xlyl+I' . . x,yo) be an augmenting path with respect to M with a 3 1 and suppose that P cannot be shortened by the existence of an edge ( x i , y m ) , 13 1 , m > 1 + 1 . Then by switching colors along a n (i,j)-path we can obtain a different coloring of G \ { u } with a corresponding matching M ' and augmenting path P' such that P' is shorter than P.
A constructive proof of this theorem is given in [6]. Here we will just show the color switches that we require without justifying that they are valid. Three types of switches are required as shown in Figs. 1 , 2 and 3 . In each case the figure gives the relevant part of Gowith matching edges indicated by heavy lines. Fig. 1 is for a = 1. It can be shown that there does not exist an (i,j)-path joining xo and y , and an (i, k)-path joining x I and yo. It there is no (i,j)-path joining xu and y l , then by switching colors along an (i,j)-path containing y l and revising the matching accordingly, we obtain the augmenting path consisting of the single edge (yo, x l ) .
Fig. 1.
W.-L.Hsu, G.L. Nemhauser
366
Figs. 2 and 3 are for a > 1. The vertex coloring on P is given by f ( x o ) = j , f ( y o ) = k and XI) = f ( y r ) = irr 1 sz I S a. In Fig. 2 it is assumed that there exists t, 1 g t S a, such that ( y I , x l + l ) EE ( c o ) . It can be shown that the ( i l , il+l)-path containing ( y l , y I c l )contains neither xI nor x,+~.Thus by switching colors along this (it, I+l)-path and revising the matching accordingly, we obtain the shorter augmenting path through x , - ~y , ~ yI~ +2-+ ~
@
Yt+2 it+2
i t r 1 Xt+1
Yt.1
it
4 - i1t xt-1 xt
Yt
it.1
W Fig. 2.
Fig. 3.
In the final case, as shown in Fig. 3, it is assumed that Vt, 1 S 1 G a, ( y , , x ~ +E~ E ) ( G )so that the switch shown in Fig. 2 is not possible; furthermore, it is assumed that there is an (il,j)-path 0 from y r through x 1 and X o so that setting f ( y , ) = j as in Fig. 1 is also not permissible. Then it can be shown that the colors can be switched as follows: f ( X o ) = f ( y l ) = i2, f ( y 2 )= j , f ( x z ) = il, and switch colors on the (il,j)-path Q. The new matching yields an augmenting path ( x 2 y 3 . . x,yo) that has four fewer edges than P. The inductive application of these color changes ultimately yields an augmenting path of length one so that Theorem 2 can be used to color u. The running time of the algorithm is determined by (1) the number of iterations, which is O ( n ) ; (2) finding an augmenting path in the complement graph of a neighbor of a vertex, which can be done in O(n2')time, see [ 5 ] ; (3) shortening an augmenting path, which is done by testing for the existence of certain (i,j)-paths in G", which can be done in O ( n 2 )time; the number of shortening is at most O ( n ) .
Algorithms
367
At each iteration we might have to do both (2) and (3), which can be bounded by O ( n ’ ) . Hence the entire algorithm is bounded by O(n4).
Appendix
To find an appropriate partition of the vertices of HI in the case of A , # 0, our algorithm uses the procedure PART(A, H, K ) where A is an independent set of K ; H = ( VH;E H )and K = ( VK;EK)are bipartite graphs such that VH= V K , EH C EK and the vertex set of H can be partitioned into V1and V 2such that all edges of H have one end in V1 and the other in Vz, and A C Vl. Furthermore, Vl is an independent set in K. The procedure PART will make a correct bipartite partion of the vertex set of H and output V1 (the LHS of H ) . procedure PART(A, H, K ) : begin 1 V, t { u E VH there is an even length path in H leading from v to some u €A}; 2 Q + a list of components of H which have not been partitioned; 3 I=1; while I# 0 do begin 4 I=O; for each component B in Q do 5 begin if there exists an edge (v, x ) in K connecting v E V1 to x E B 6 then do begin comment the vertex x must be placed on the RHS of H ; 7 delete B from Q ; comment it is easy to determine a partition of a connected bipartite graph according to whether the length of a path between two vertices is even or odd; arrange the vertex set of B such that x is on the RHS; 8 if (LHS of B ) # 0 then do 9 begin V1+ VI U (LHS of B ) ; 10 11 f t l ; end end
1
368
W.- L.Hsu, G.L. Nemhauser
else comment restore this component to Q for future consideration; end comment ‘ I = 0 indicates that V, has not been augmented in this loop; end 12 K ’ t t h e subgraph of K induced on U Q ; 13 Make an arbitrary bipartite partition of K’; 14 Vi + Vi U (LHS of K ‘ ) ; end The correctness of the algorithm can be argued as follows. In line I we have that A C V, and VI is an independent set in K. Each time VI is augmented in line 10, the LHS of B is independent of the old V, since otherwise there cannot exist a correct partition. In line 14 no vertex of K’ is adjacent to any vertex of the old V,. Hence the final V, is an independent set in K . Let m be the initial size of Q. Then we have to execute line 6 through line 11 at most m ( m + 1)/2 times. Therefore, this algorithm is polynomially bounded.
References \ I ] V . Chvdtai. On certain polytopes associated with graphs, J. Comb. Theory, Ser. B. 18 (1972) 13S154. [2] D.R. Fulkerson. On the perfect graph theorem, in: T.C. Hu and S.P. Robinson, eds., Mathematical Programming (Academic Press, New York, 1973). [3] R. Giles and L.E. Trotter, On stable set polyhedra for K,,3-freegraphs, J. Comb. Theory, Ser. B 31 (1981) 313-326. (41 M. Grotschel, L. Lovasz and A. Schrijver, The ellipsoid method and its consequences in combinatorial optimization, Combinatorica 1 (1981) 169-197. [S]J.E. Hopcroft and R.M. Karp, An nS’* algorithm for maximum matching in bipartite graphs, SIAM J. Comput. 2 (1973) 225-231. 161 W.-L. Hsu, How to color claw-free perfect graphs, Ann. Discrete Math. 11 (1981) 189-197. 171 W.-L. Hsu. Efficient algorithms for some packing and covering problems on graphs, Ph.D. Dissertation, School of Operations Research and Industrial Engineering (Cornell University, Ithaca, New York. 1980). 181 W.-L. Hsu and G.L. Nemhauser, A polynomial algorithm for the minimum weighted clique cover problem on claw-free perfect graphs, Discrete Math. 38 (1982) 65-71. [Y] W.-L. Hsu and G.L. Nemhauser. Algorithms for minimum colorings by cliques and maximum cliques on claw-free perfect graphs, Discrete Math. 37 (1981) 181-191. 1101 A.V. Karzanov, Determining the maximal Row in a network by the method of preflow, Soviet Math. Dokl. 15 (1974) 434-437. [ 111 L.G. Khachian, A polynomial algorithm in linear programming, Soviet Math. Dokl. 20 (1979) 191-194
Algorithms
369
(I21 G.J.Minty, On maximal independent sets of vertices in claw-free graphs, J. Comb. Theory, Ser. B 28 (1980) 284-304. [13] K.R. Parthasarathy and G. Ravindra, The strong perfect graph conjecture is true for K,,,-free graphs, J. Comb. Theory, Ser. B 21 (1976) 212-223. [14] N. Sbihi, Algorithmes de recherche d'un stable de cardinalitt maximum dans un graphe sans ttoile, Discrete Math. 29 (1980) 53-76. 1151 I. Holyer, The NP-completeness of some edge-partition problems, SIAM J. Comput. 10 (1981) 7 13-7 17.
This Page Intentionally Left Blank