Advanced Topics in Algorithms Shamir

Advanced Topics in Graph Algorithms
Technical Report
Lecturer: Ron Shamir
Edited by: Sariel Har-Peled
Based on Notes from the Course

c Tel-Aviv University, Spring 1994.
2
Contents
1 Lecture 1 11
1.1 Introduction to Graph Theory . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.1.1 Basic Denitions and Notations . . . . . . . . . . . . . . . . . . . . . 11
1.1.2 Intersection Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.1.3 Circular-arc Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.1.4 Interval Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.1.5 Line graphs of bipartite graphs . . . . . . . . . . . . . . . . . . . . . 23
2 Lecture 2 25
2.1 Perfect Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.1.1 Denition of perfect graph . . . . . . . . . . . . . . . . . . . . . . . . 25
2.1.2 Some Denitions and Properties . . . . . . . . . . . . . . . . . . . . . 26
2.1.3 Perfect Graph Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 27
3 Lecture 3 35
3.1 Perfect Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.1.1 p-Critical Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.1.2 A Polyhedral Characterization of Perfect Graphs . . . . . . . . . . . 36
3.1.3 The Strong Perfect Graph Conjecture . . . . . . . . . . . . . . . . . . 36
3.2 Triangulated Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2.2 Characterizing Triangulated Graphs . . . . . . . . . . . . . . . . . . . 37
3.2.3 Recognizing Triangulated Graphs . . . . . . . . . . . . . . . . . . . . 40
4 Lecture 4 43
4.1 Recognizing Triangulated Graphs . . . . . . . . . . . . . . . . . . . . . . . . 43
4.1.1 Generating a PEO . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.1.2 Testing an Elimination Scheme . . . . . . . . . . . . . . . . . . . . . 46
4.2 Triangulated Graphs as Intersection Graphs . . . . . . . . . . . . . . . . . . 49
4.2.1 Evolutionary Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3
4 CONTENTS
4.2.2 Triangulated Graphs as Intersection Graphs . . . . . . . . . . . . . . 51
5 Lecture 5 57
5.1 Triangulated Graphs Are Perfect . . . . . . . . . . . . . . . . . . . . . . . . 57
5.1.1 Proving This Property . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.1.2 Other Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.2 Computing the Minimum Fill In . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.2.1 The Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.2.2 Fill In is NP-Hard . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.2.3 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
6 Lecture 6 67
6.1 Some Optimization Algorithms on Triangulated Graphs . . . . . . . . . . . 67
6.1.1 Computing the chromatic number and all maximal cliques . . . . . . 67
6.1.2 Finding and k . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
6.2 Comparability Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
6.2.1 Some Denitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
6.2.2 Implication Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
6.2.3 The Triangle Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.2.4 Decomposition of Graphs . . . . . . . . . . . . . . . . . . . . . . . . . 74
7 Lecture 7 77
7.1 Uniquely Partially Orderable Graphs . . . . . . . . . . . . . . . . . . . . . . 77
7.2 Characterizations and Recognition Algorithms . . . . . . . . . . . . . . . . . 84
8 Lecture 8 93
8.1 Some interesting graph families characterized by intersection . . . . . . . . . 93
8.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
8.1.2 Permutation graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
8.1.3 F -Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
8.1.4 Tolerance graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
8.1.5 Bounded and unbounded tolerance graphs. . . . . . . . . . . . . . . . 101
9 Lecture 9 103
9.1 Comparability Graph Recognition . . . . . . . . . . . . . . . . . . . . . . . . 103
9.2 The Complexity of Comparability Graph Recognition . . . . . . . . . . . . . 105
9.2.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
9.2.2 Complexity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
9.3 Coloring and Other Problems on Comparability Graphs . . . . . . . . . . . . 109
9.3.1 Comparability Graphs Are Perfect . . . . . . . . . . . . . . . . . . . 109
CONTENTS 5
9.3.2 Maximum Weighted Clique . . . . . . . . . . . . . . . . . . . . . . . 111
9.3.3 Calculating (G) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
9.4 The Dimension of Partial Orders . . . . . . . . . . . . . . . . . . . . . . . . 113
10 Lecture 10 119
10.1 Comparability invariants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
10.2 Interval graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
11 Lecture 11 127
11.1 Preference and Indierence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
11.2 Recognizing interval graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
11.2.1 A quadratic algorithm 1 . . . . . . . . . . . . . . . . . . . . . . . . . 132
11.2.2 A Linear Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
11.3 More about the consecutive 1's property . . . . . . . . . . . . . . . . . . . . 137
12 Lecture 12 139
12.1 Temporal Reasoning and Interval algebras . . . . . . . . . . . . . . . . . . . 139
12.2 Interval satisability problems . . . . . . . . . . . . . . . . . . . . . . . . . . 141
12.3 Interval sandwich problems and ISAT . . . . . . . . . . . . . . . . . . . . . . 143
12.4 A linear time algorithm for ISAT(1) . . . . . . . . . . . . . . . . . . . . . 145
12.5 Bandwidth, Pathwidth and Proper Pathwidth . . . . . . . . . . . . . . . . . 148
6 CONTENTS
List of Figures
1.1 Three pictures of the same graph. . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2 Some orientations of a triangle. . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3 (a) The graph K3.(b) is the complement of (c) and vice versa. . . . . . . . . 13
1.4 Examples of subgraphs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.5 (a) A graph G.(b) A maximal clique in G.(c) A maximum clique of G.(d) A
1-clique. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.6 Graph examples. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.7 (a) A graph G.(b) An interval representation of G.(c) A proper interval rep-
resentation for P3 .(d) A unit interval representation for P3 . . . . . . . . . . . 16
1.8 (a) A graph G.(b) A circular-arc representation of G. . . . . . . . . . . . . . 17
1.9 (a) The trac ow. (b) The intersection graph of the arcs. (c) The clock cycle. 18
1.10 The matching diagram of 4,3,6,1,5,2], and the corresponding permutation
graph. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.11 A triangulated graph which is not an interval graph. . . . . . . . . . . . . . . 20
1.12 Transitive orientations of Three comparability graphs. . . . . . . . . . . . . . 21
1.13 Two graphs which are not transitively orientable. . . . . . . . . . . . . . . . 21
1.14 (a) A graph G.(b) The line graph of G. . . . . . . . . . . . . . . . . . . . . . 23
2.1 The matrix M: each row correspond to a vertex in V , and each column to an
! ; clique K (S ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1 A graph and its clique matrix . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2 Two graphs, the left one is triangulated, and the right one is not. . . . . . . 37
3.3 Code for MCS algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.4 Two dierent possible running of the MCS algorithms. . . . . . . . . . . . . 40
4.1 The path . Vertices appear ordered from left to right according to the MCS
algorithm output. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.2 The vertices x and vj . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.3 The path . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
0
45
4.4 The path . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
00
45
7
8 LIST OF FIGURES
4.5 A forbidden structure in a chordal graph . . . . . . . . . . . . . . . . . . . . 45
4.6 Code for PEO testing algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.7 An input chordal graph and elimination order . . . . . . . . . . . . . . . . . 47
4.8 Xv n fug is a clique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.9 The intersection graph of a set of subtrees of a host tree is triangulated . . . 49
4.10 An evolutionary tree describing the properties A B C D 1 2 3, and the cor-
responding triangulated intersection graph . . . . . . . . . . . . . . . . . . . 50
4.11 The left family has the Helly property, while the right family does not . . . . 51
4.12 The paths P1, P2, and P3 must intersect . . . . . . . . . . . . . . . . . . . . 51
4.13 The tree T = (K E ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.14 The paths P0 : : : Pk 1 form a cycle . . . . . . . . . . . . . . . . . . . . . . . 54
The case B 6= Y . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
;
4.15 54
4.16 The case B = Y . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.1 The Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.2 Some examples of Meyniel graphs (which are not chordal!) . . . . . . . . . . 59
5.3 Example-A Chain Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.4 Ordering Q, in the same chain graph . . . . . . . . . . . . . . . . . . . . . . 61
5.5 Independent edges (broken lines denote non-edges) . . . . . . . . . . . . . . 61
5.6 Independent edges that form an induced C4 . . . . . . . . . . . . . . . . . . 62
5.7 C (G) for the graph G of Figure 5.3 . . . . . . . . . . . . . . . . . . . . . . . 62
5.8 a graph (left) with two possible layouts of cost 8 (center) and 6 (right) . . . 63
5.9 the graph G (left) and G' (right) . . . . . . . . . . . . . . . . . . . . . . . . 64
5.10 Example-Edges we have to add . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.11 x and v . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.1 An example for maximal cliques in a graph . . . . . . . . . . . . . . . . . . . 67
6.2 Chromatic Number Calculation . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.3 An example for nding maximal stable set . . . . . . . . . . . . . . . . . . . 70
6.4 some comparability graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
6.5 An example for implication classes . . . . . . . . . . . . . . . . . . . . . . . 71
6.6 The triangle lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
7.1 The composition graph of P3K3 K1 C4] . . . . . . . . . . . . . . . . . . . . 78
7.2 The composition graph of F1F2 F3 F4] . . . . . . . . . . . . . . . . . . . . . 78
7.3 The decomposition of a graph . . . . . . . . . . . . . . . . . . . . . . . . . . 80
7.4 Module in a graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
7.5 Uniquely partially orderable graph . . . . . . . . . . . . . . . . . . . . . . . 82
7.6 A graph with a color class that induce all vertices . . . . . . . . . . . . . . . 83
7.7 A graph G-decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
LIST OF FIGURES 9
7.8 Algorithm to construct G-decomposition . . . . . . . . . . . . . . . . . . . . 85
7.9 A triangle forcing the merge of two implication classes . . . . . . . . . . . . 86
7.10 The classes Fb Cb must be equal. . . . . . . . . . . . . . . . . . . . . . . . . . 87
7.11 Ab + Ac1 is an implication class in E n D
c. . . . . . . . . . . . . . . . . . . . . 87
7.12 A graph G and its quasi-transitive graph G (in G , edges (x y) $ (y x) are
0 0
not drawn) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
8.1 A permutation graph for the permutation 4,2,6,1,3,5] . . . . . . . . . . . . . 94
8.2 An orientation on a permutation graph . . . . . . . . . . . . . . . . . . . . . 95
8.3 An F-graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
8.4 Not an F-graph,point A have two values of the function. . . . . . . . . . . . 97
8.5 The air controllers strike problem. . . . . . . . . . . . . . . . . . . . . . . . 98
8.6 The composition of permutation diagrams. . . . . . . . . . . . . . . . . . . 100
8.7 A tolerance graph. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
8.8 An unbounded tolerance graph. . . . . . . . . . . . . . . . . . . . . . . . . . 102
9.1 Code of The TRO Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 103
9.2 Example 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
9.3 Example 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
9.4 Example 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
9.5 Procedure EXPLORE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
9.6 The main procedure of the algorithm . . . . . . . . . . . . . . . . . . . . . . 108
9.7 Two elements for each Edge- . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
9.8 Example 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
9.9 Example 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
9.10 Example 6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
9.11 Procedure EXPLORE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
9.12 The main procedure of the algorithm . . . . . . . . . . . . . . . . . . . . . . 112
9.13 Example 7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
9.14 Example 8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
9.15 Example 9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
9.16 Example 10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
9.17 Example 11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
9.18 Example 12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
9.19 Example 13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
9.20 Example 14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
10.1 a) not SP poset b) - e) SP posets . . . . . . . . . . . . . . . . . . . . . . . 121
10.2 a) interval graph b) it's interval representation c) chordal, but not interval
graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
10 LIST OF FIGURES
10.3 example for lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
10.4 example for theorem of Gilmore-Homan . . . . . . . . . . . . . . . . . . . . 123
10.5 a) interval graph b) it's clique matrix. . . . . . . . . . . . . . . . . . . . . . 124
10.6 Example for Lekkerkerker - Boland theorem . . . . . . . . . . . . . . . . . . 125
11.1 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
11.2 Wegner 1967 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
11.3 Example of function l h . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
11.4 a PQ-tree. The FRONTIER here is F J I G H A B E D C] . . . . . . . . . 134
11.5 Example of execution of CONSECUTIVE . . . . . . . . . . . . . . . . . . . 136
12.1 A consistent scenario . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
12.2 The 13-valued interval algebra A13. Single line: x interval. Double line: y interval. 140
12.3 Two interval realizations for Example 12.2 . . . . . . . . . . . . . . . . . . . 142
12.4 (a) The 5-chain (vi1 x1i vi2 x2i vi3 ). (b) An interval realization of the 5-chain. 144
12.5 Algorithm to construct Delta1-CONSISTENCY . . . . . . . . . . . . . . . . 147
12.6 Complexity of ISAT on all fragments of A3. T : satised by setting all intervals
disjoint or identical. The four asterisk cases were shown to be NPC by Webber
(1994). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
Chapter 1
Lecture 1
c Tel-Aviv University
Spring Semester, 1994
Lecturer : Ron Shamir
Scribe : Kahn Shaul
Lecture 1 : April 21
1.1 Introduction to Graph Theory

1.1.1 Basic Denitions and Notations
A directed graph G consists of a nite set V and an irreexive binary relation on V . We call
the elements of V , vertices. The binary relation may be represented either as a collection E
of ordered pairs, or as a function from V to its power set: Adj : V ! P (V ).
We call Adj (v) the adjacency set of vertex v, and we call the ordered pair (v w) 2 E
an edge. Clearly : (v w) 2 E $ w 2 Adj (v). In this case we say that w is adjacent to v,
and v and w are endpoints of the edge (v w). The assumption of irreexivity implies that
(v v) 2= E (v 2 V ), or equivalently, v 2= Adj (v) (v 2 V ). Adj (v) is sometimes denoted by
N (v) and is called the open neighborhood of v. The closed neighborhood of v is dened :
N v] = v + Adj (v)
We will usually drop the parentheses and the comma when denoting an edge. Thus xy 2 E
and (x y) 2 E will have the same meaning.
In an undirected graph every edge appears in both directions.
Two graphs G = (V E ) and G = (V E ) are called isomorphic, denoted G
0 0 0
= G if there
0
is a bijection f : V ! V satisfying for all x y 2 V ,

0
(x y) 2 E $ (f (x) f (y)) 2 E

0
11
12 CHAPTER 1. LECTURE 1
a b
a b d a
c d b c
c d
Figure 1.1: Three pictures of the same graph.
Let G = (V E ) be a graph with vertex set V and edge set E . The graph G 1 = (V E 1) is
; ;
said to be the reversal of G, where E 1 = f(x y) j (y x) 2 E g, that is,

;
xy 2 E 1 $ yx 2 E (x y 2 V ):
;
We dene symmetric closure of G to be the graph G^ = (V E^ ), where E^ = E E 1. ;
A graph G = (V E ) is called undirected if its adjacency relation is symmetric i.e., if

E = E 1 or E = E^ .
;
A graph is called an oriented graph if E \ E 1 = .

;
Orientation of an undirected graph is a choice of a direction on every edge in it.

For example : see Figure 1.2.
a c a c a c
b b b
Figure 1.2: Some orientations of a triangle.
Let G = (V E ) be an undirected graph. We dene the complement of G to be the graph
G = (V E ), where
E = f(x y) 2 V
V j x 6= y and (x y) 2= E g
A graph is complete if every two distinct vertices are adjacent. The complete graph on n
vertices is usually denoted by Kn , and is called a clique of size n.
A (partial) subgraph of a graph G = (V E ) is dened to be any graph H = (V E ) 0 0
satisfying V V and E E .
0 0
Given a subset A V of the vertices, we dene the subgraph induced by A to be GA =

(A EA), where EA = fxy 2 E j x 2 A and y 2 Ag.
1.1. INTRODUCTION TO GRAPH THEORY 13
(a) (b) (c)

Figure 1.3: (a) The graph K3 .(b) is the complement of (c) and vice versa.
a e a e
c d c d
b f b
a graph G a (partial) subgraph of G
a
c d
f
the induced subgraph
G {a,c,d,f}
Figure 1.4: Examples of subgraphs.
Let G = (V E ) be an undirected graph. Consider the following denitions :
A subset A V of r vertices is an r-clique@r-clique if it induces a complete subgraph,
i.e., if GA = K r . A single vertex is a 1-clique.
A clique K in G is maximal if there is no clique of G which properly contains K as a
subset.
A clique is maximum if there is no clique of G of larger cardinality.
!(G) is the number of vertices in a maximum clique of G, and is called the clique number
of G.
A clique cover of size k is a partition of the vertices V = A1 + A2 + ::: + Ak such that
each Ai is a clique.
k(G) is the size of a smallest possible clique cover of G. It is called the clique cover
number of G.
A stable set or independent set is a subset X of vertices no two of which are adjacent.
(G) is the number of vertices in an independent set of maximum cardinality. It is called
the stability number of G.
b c b b c a
a d a d
(a) (b) (c) (d)
Figure 1.5: (a) A graph G.(b) A maximal clique in G.(c) A maximum clique of G.(d) A
1-clique.
A proper ccoloring@c-coloring is a partition of the vertices V = X1 + X2 + ::: + Xc such
that each Xi is an independent set. In such a case, the vertices of Xi are "painted" with the
color i and adjacent vertices will receive dierent colors. If G has a proper c-coloring, we
say that G is ccolorable@c-colorable.
(G) is the smallest possible c for which there exist a proper c-coloring of G, it is called
the chromatic number of G.
In the graph G of gure 1.5 (a) : fb c dg and fag are a clique cover of size 2. fa cg is
an independent set in G.
Hence, !(G) = 3 k(G) = 2 (G) = 2 (G) = 3
It is always true that:
!(G) (G) and (G) k(G)
Since every vertex of a maximum clique must be contained in a dierent color class in
any minimum proper coloring, and every vertex of a maximum stable set must be contained
in a dierent clique in any minimum clique cover.
The following relations are sometimes called duality:
!(G) = (G) and (G) = k(G):
Let G = (V E ) be an arbitrary graph, consider the following denitions :
Chain: A sequence of vertices v0 v1 v2 :::vl] is a chain of length l in G if
vi 1vi 2 E or vivi 1 2 E for i = 1:::l:
; ;
Path: A sequence of vertices v0 v1 v2 :::vl] is a path from v0 to vl in G provided that
vi 1vi 2 E for i = 1::l. A path or a chain in G is called simple if no vertex occurs more than
;
once in the sequence. It is called trivial if l = 0. It is called chordless if the graph induced
by fv0 :::vlgis a path of length l. We denote the chordless path on n vertices by Pn .
Connected graph: A graph G is connected if between any two vertices there exists a chain
in G joining them.
Strongly connected graph : A graph G is strongly connected if for any two vertices x and y
there exists a path in G from x to y.
Cycle : A sequence of vertices v0 v1 v2 :::vl v0] is called a cycle of length l +1 ( or closed
path ) if vi 1vi 2 E for i = 1:::l and vlv0 2 E .
Simple cycle: A cycle v0 v1 v2 :::vl v0 ] is a simple cycle if vi 6= vj for i 6= j .
;
Chordless cycle: A simple cycle v0 v1 v2 :::vl v0] is chordless if vivj 2= E
for every i j such that f(i ; j ) > 1(mod l)g. A chordless cycle on n vertices is denoted
by Cn .
Bipartite graph: An undirected graph G = (V E ) is bipartite if its vertices can be parti-
tioned into two disjoint independent sets V = S1 + S2 i.e., every edge has one endpoint in
S1 and one in S2. We usually denote such graphs by (S1 S2 E ) to distinguish the two parts.
Complete bipartite graph: A bipartite graph G = (S1 S2 E ) is complete if for every
x 2 S1 and y 2 S2 we have xy 2 E . We denote by Km n the complete bipartite graph
(S1 S2 E ) with j S1 j= m and j S2 j= n.
Examples for using the above notation : K3 = C3 is called a triangle C 4 = 2K2 Kn n =
2Kn C5 = C 5.
1.1.2 Intersection Graphs

Let F be a family of nonempty sets. The intersection graph of F is obtained by representing
each set in F by a vertex and connecting two vertices by an edge i their corresponding sets
intersect. Examples :
1. Interval graph : The intersection graph of a family of intervals on a linearly ordered
set (like the real line) is called an interval graph. If these intervals have a unit length,
then we call the graph a unit interval graph.
P5
K
4,3
C6 K 2K 2
1,6
Figure 1.6: Graph examples.
2. Proper interval graph : A proper interval graph is constructed from a family of intervals
on a line such that no interval properly contains another.
In 1969 Roberts showed that a graph is a unit interval graph i it is a proper interval
graph.
5 2
3 2
4 5 1
6 1 6
3
(b)
(a) (c) (d)

Figure 1.7: (a) A graph G.(b) An interval representation of G.(c) A proper interval repre-
sentation for P3.(d) A unit interval representation for P3.
1.1.3 Circular-arc Graphs

Consider the following generalization of the notion of intervals on a line. If we join the two
ends of our line, thus forming a circle, the intervals will become arcs on the circle. Allowing
arcs to slip over and include the point of connection, we obtain a class of intersection graphs
called the circular-arc graphs, which properly contains the interval graphs. In other words :
a graph is a circulation arc-graph if it is the intersection graph of a family of arcs on a circle.
5
3 1
2 2
1
4 5 4
(a)
3
(b)
Figure 1.8: (a) A graph G.(b) A circular-arc representation of G.
Example 1.1 The trac ow at the corner of Holly, Wood and Vine is pictured in the next
page in gure 1.9 (a). Two lanes are called compatible if trac can ow on both of them
simultaneously without collisions. Certain lanes are compatible with one another, such as c
and j or d and k, while others are incompatible, such as b and f. In order to avoid collisions,
we wish to install a trac light system to control the ow of vehicles. Each lane will be
assigned an arc on a circle representing the time interval during which it has a green light.
Incompatible lanes must be assigned disjoint arcs. The circle may be regarded as a clock
representing an entire circle which will be continually repeated.
An arc assignment for our example is given in gure 1.9 (c). In general if G is the
intersection graph of the arcs of such an assignment (see the gure 1.9 (b) ), and if H is the
compatibility relation dened on the pairs of lanes, then clearly G is a (partial) subgraph of
H . In our example, the compatible pairs (d k) (h j ) and (i j ) are in H but are not in G.
A proper circular arc graph is the intersection graph of a family of arcs none of which
properly contains another. In a dierent generalization of interval graphs Renz 1970] char-
acterized the intersection graphs of paths in a tree. Walter 1972], Buneman 1974], and
Gavril 1974] carried this idea further and showed that the intersection graphs of subtrees of
a tree are exactly the triangulated graphs (the graphs with no induced Ck for k 4).
A permutation diagram consists on n points on each of two parallel lines, and n straight
line segments matching the points. The intersection graph of the line segments is called a
permutation graph.
If the 2n points are located randomly around a circle, then the matching segments will
be chords of the circle, and the resulting class of intersection graphs properly contains the
permutation graphs. This class is called the class of circle graphs.
A simple argument shows that every proper circular-arc graph is also a circle graph. We
may assume that no pair of arcs together covers the entire circle. For each arc on the circle
Wood Street
g e
f
j
k c da d
b i a b c e
l
(c) f
----------- ---------------
g
l
----------- ---------------
k
g Holly Street
k
f
l e j i h
h a
d Vine Avenue
i j
(a)
b c G , the circular-arc graph.

(b)
Figure 1.9: (a) The trac ow. (b) The intersection graph of the arcs. (c) The clock cycle.
draw the chord connecting its two endpoints. Clearly two arcs overlap i their corresponding
chords intersect.
1 2 3 4 5 6 1 6 5
2
4
4 3 6 1 5 2 3
Figure 1.10: The matching diagram of 4,3,6,1,5,2], and the corresponding permutation
graph.
1.1.4 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 f of a linearly ordered set (like the real line) such that
two vertices are connected by an edge of G i their corresponding intervals have nonempty
intersection. We call f an interval representation for G.
Application to scheduling :
Consider a collection c = ci of courses being oered by a major university. Let Ti be
the time interval during which course ci 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 :
cicj 2 E $ Ti \ Tj 6= :
Each color corresponds to a dierent classroom. The graph G is obviously an interval graph,
since it is represented by time intervals. This example is especially interesting because e-
cient 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).
An application of interval Graph
Suppose c1 c2 ::: cn are chemical compounds which must be refrigerated under closely
monitored conditions. If compound ci must be kept at a constant temperature between ti
and ti degrees, how many refrigerators will be needed to store all the compounds ?
0
Let G be the interval graph with vertices c1 c2 ::: cn and connect two vertices whenever
the temperature intervals of their corresponding compounds intersect. Intervals on a line
satisfy the Helly Tproperty : if a set S = fTi:::Tkg of intervals such that, Ti \ Tj 6= for
every i 6= j then ti S Ti 6= . Hence, if fci1 ci2 ::: cikg is a clique of G, then the intervals
2
ftij tij ] j j = 1 2 ::: kg will have a common point of intersection, say t. A refrigerator set
0
at a temperature of t will be suitable for storing all of them. Thus, a solution to the problem
will be obtained by nding a minimum clique cover of G. This is a polynomial problem on
interval graph, although it is NP-hard for arbitrary graphs.
Interval graphs satisfy many interesting properties, one of them is that interval graph is
an hereditary property: An induced subgraph of an interval graph is an interval graph.
x
d
b c
y z
Figure 1.11: A triangulated graph which is not an interval graph.
Hereditary properties abound in graph theory, such as planarity and bipartiteness. The
next property of interval graphs is also an hereditary property.
Triangulated graph property :
Every simple cycle of length strictly greater than 3 possesses a chord. Graphs which
satisfy this property are called triangulated graphs.
Theorem 1.2 (Hajos1958]) An interval graph satises the triangulated graph property.
Proof: Suppose the interval graph G contains a chordless cycle v0 v1 v2 :::vl 1 v0] with
l > 3. Let Ik denote the interval corresponding to Vk . For i = 1 2 ::: l ; 1 choose a point
;
pi 2 (Ii 1 \ Ii). Since Ii 1 and Ii+1 do not overlap, the Pi constitute a strictly increasing
; ;
or strictly decreasing sequence. Therefore, it is impossible for I0 and Il 1 to intersect,

;
contradicting the criterion that V0 Vl 1 is an edge of G.

;
Not every triangulated graph is an interval graph. Consider the tree T given in gure
1.11, which certainly has no chordless cycles. The intervals Ia Ib and Ic of a representation
for T would have to be disjoint, and Id would properly include the middle interval, say Ib.
Where, then could we put Iy so that it intersects Ib but not Id ? Clearly we would be stuck.
Transitive orientation property.
Each edge can be assigned a one-way direction, in such a way that the resulting oriented
graph G = (V E ) satises the following condition :
( ab 2 E and bc 2 E )imply ac 2 E (8a b c 2 V ) ?
The A graph The suspension bridge graph The house graph
Figure 1.12: Transitive orientations of Three comparability graphs.
An undirected graph which is transitively orientable is sometimes called a comparability

graph. Figure 1.12 shows a transitive orientation of the A graph and of the suspension bridge
graph. The odd length chordless cycles C5 C7 C9::: and the bull's head graph ( see Figure
1.13 ) cannot be transitive oriented.
C The bull’s head graph

5
Figure 1.13: Two graphs which are not transitively orientable.
Theorem 1.3 (Ghouila - Houri 1962]). The complement of an interval graph satises the
transitive orientation property.
Proof: Let fIv j v 2 V g be an interval representation for G = (V E ). Dene an
orientation F of the complement G = (V E ) as follows :
xy 2 F $ Ix < Iy (8xy 2 E ):
Here Ix < Iy means that the interval Ix lies entirely to the left of the interval Iy . (Remember,
they are disjoint). Clearly (?) is satised, since Ix < Iy < Iz implies Ix < Iz . Thus, F is a
transitive orientation of G.
Theorem 1.4 (Gilmore and Ho man 1964) An undirected graph G is an interval graph
i G is a triangulated graph and its complement G is a comparability graph.
We will see the proof in a later lecture.
We have seen already that every interval graph satises the two conditions. Looking
back each of the graph in Figure 1.12 can be properly colored using three colors and each
contains a triangle. Therefore for these graphs, their chromatic number equals their clique
number. We will see later in the course that any triangulated graph and any comparability
graph also satises the following properties :
;Perfect property
For each induced subgraph GA of G, (GA ) = !(GA ).
The chordless cycles C5 C7 C9 are not - perfect. A dual notion of - perfection is the
following:
alphaperfect@ ;Perfect property
For each induced subgraph GA of G, (GA ) = k(GA ).
Theorem 1.5 The perfect graph theorem (Lova!sz 1972]).
A graph is ;perfect i it is ;perfect.
We will prove this theorem in lecture 2.
1.1.5 Line graphs of bipartite graphs

Let G = (V E ). The line graph or the edge graph of G denoted L(G) is the graph with
V (L(G)) = E (G) and
(e1 e2) 2 E (L(G)) $ e1 and e2 have a common endpoint in G:
1 2 12 14
3
43 25
5 4
G 45 L(G)
Figure 1.14: (a) A graph G.(b) The line graph of G.

Being a line graph of a bipartite graphs is an hereditary property. (Removing a vertex
in L(G) is the same as removing the corresponding edge in G, which remains bipartite).
An Independent set in L(G) corresponds to a matching ( a set of independent edges) in
G. A Clique in L(G) corresponds to a set of pairwise adjacent edges in G. If G is bipartite
it contains no triangles, so every set of pairwise adjacent edges in G) is a star. Therefore, if
G is bipartite, a clique cover of L(G) corresponds to a vertex cover of G.
Let us call a graph LGBG if it is the line graph of a bipartite graph. Konig (1931)
proved that the edge chromatic number of a bipartite graph equals its maximum degree.
By the above, this is equivalent to saying that if G is LGBG then (G) = !(G). The
Konig-Hall theorem (1931) states that the size of a maximal matching in a bipartite graph
equals its minimum vertex cover. This is equivalent to saying that (G) = K (G) if G is
LGBG. Together with the fact that being LGBG is hereditary, any of these two statements
imply that LGBG S are perfect. From the perfect graph theorem the two statements are
0
equivalent.
Chapter 2
Lecture 2
Scribe : Erez Hartuv
Lecture 2 : April 28
2.1 Perfect Graphs

In this lecture the graphs we consider are undirected. When we refer to a graph G we mean
the Graph G(V E ), with vertex set V and edge set E .
2.1.1 Denition of perfect graph

The following three conditions dene perfection properties of a graph G(V E ).
(P1) !(GA ) = (GA) 8A V
(P2) (GA ) = k(GA ) 8A V
(P3) !(GA ) (GA ) jAj 8A V
By duality, a graph G satises (P1) i its complement G satises (P2).
In this lecture we'll prove the perfect graph theorem which says that (P1),(P2) and (P3)
are equivalent. A graph which satises one of the perfection properties satises all (by the
theorem) and is called perfect graph.
Remark: The equivalence holds for nite graphs only.
25
Lemma 2.1 Let the graph G(V E ) satisfy: (G) = k(G). If K is a minimum clique cover,
and S is a maximum independent set, then each clique in K has exactly one vertex in common
with S .
Proof: By denition, no clique can intersect an independent set in more than one ele-
ment. Therefore the intersection of S and each clique in K may be empty, or one element.
Remember that each vertex is contained in some clique. If there is a clique whose intersection
with S is empty then, jK j > jS j, contradiction to the assumption of equality.
Lemma 2.2 Let G be an induced subgraph of H then:
1. ! (H ) ! (G)
2. (H ) (G)
3. (H ) (G)
4. k (H ) k(G)
Lemma 2.3 If D V intersects every maximum independent set of G in exactly one vertex
then: (GV D ) = (G) ; 1
;
Proof: Denote = (G). Let U be a maximum independent set of G. By the assumption

U \ D = fig. Therefore U ;fig is an independent set of GV D . Therefore (GV D ) ; 1.
If (GV D ) > ; 1 then (GV D ) = (because of the monotonicity shown in Lemma 2.2).
; ;
; ;
It follows that GV D contains a maximum independent set of size , therefore G contains

;
this maximum independent set of size , which does not intersect D, contradiction.
2.1.2 Some Denitions and Properties

Given two graphs G1 = (V1 E1) G2 = (V2 E2) , we dene their union and join.
Denition Union: G1 G2 = (V1 V2 E1 E2)
Denition Join: G1 + G2 = (V1 V2 E = E1 E2 fuvju 2 V1 v 2 V2g)
Lemma 2.4 if G1 G2 are perfect, then their union and join are perfect.
Proof: Denote the parameters of each graph Gi as: i !i i ki i = 1 2
The parameters value for G1 G2 are:
! = max(!1 !2) = max(1 2)
2.1. PERFECT GRAPHS 27
k = k1 + k2 = 1+ 2
The parameters value for G1 + G2 are:
! = !1 + !2 = 1 + 2
k = max(k1 k2) = max( 1 2)
These relations together with the perfection of G1, G2 imply the claim.
Denition G x: Let G(V E ) be an undirected graph with vertex x 2 V . The graph
G x is obtained from G by adding a new vertex x which is connected to all the neighbors
0
of x: G x = (V fx g E = E fx uju 2 Adj (x)g).

0 0 0
Lemma 2.5 8x 6= y (G x) ; y = (G ; y) x
Proof: Let x denote the new vertex (the copy of x). Both graphs has the same vertex
0
set: V + x ; y. In both graphs, uv is an edge i one of the following exist:

0
1. uv 2 E u 6= y v 6= y
2. v = x ux 2 E u 6= y
0
Denition More generally, if x1 : : : xn are the vertices of G and h = (h1 : : : hn ) is a

vector of non-negative integers, then H = G h is constructed by substituting for each xi
h
a stable set of hi vertices x1i : : : xi i and joining xsi with xtj i xi and xj are adjacent in G.
We say that H is obtained from G by multiplication of vertices.
Remark. The denition allows hi = 0, in which case H includes no copy of xi. Thus, every
induced subgraph of G can be obtained by multiplication of G by appropriate (0,1)-valued
vector. For example G = G (1 : : : 1).
2.1.3 Perfect Graph Theorem

We now reach the main part of the lecture, in which we prove two lemmas that will help to
prove the perfect graph theorem.
Lemma 2.6 (Berge1961]). Let H be obtained from G by multiplication of vertices.
1. If G satises (P1), then H satises (P1).
2. If G satises (P2), then H satises (P2).
Proof: The proof is by induction on the number of vertices in G. The lemma is true if
G has only one vertex. We shall assume that (1) and (2) are true for all graphs with fewer
vertices than G. Let H = G h. If one of the coordinates of h equals zero, say hi = 0, then
H can be obtained from G ; xi by multiplication of vertices.
But, if G satises (P1) respectively (P2)], then G ; xi also satises (P1) respectively
(P2)] (G ; xi is an induced subgraph of G, therefore (P1) on G implies (P1) on all induced
subgraphs of G). Then by the induction hypothesis H satises (P1) respectively (P2)].
Thus, we may assume that each coordinate hi 1, and since H can be built up from a
sequence of single-copy multiplications, it is sucient to prove the result for H = G x. Let
x denote the added "copy" of x.
0
Assume that G satises (P1). Since x and x are nonadjacent, !(G x) = !(G). Let G
0
be colored using !(G) colors. Color x the same color as x. This will be a coloring of G x
0
in !(G x) colors. Hence, G x satises (1).

Next assume that G satises (P2). We must show that (G x) = k(G x). Let K be
a (minimum) clique cover of G with jK j = k(G) = (G), and let Kx be the clique in K
containing x. There are two cases.
Case 1: x is contained in a maximum stable set S of G, i.e., jS j = (G). In this case
S fx g is a stable set of G x, so
0
(G x) = (G) + 1: (2.1)
Since K ffx gg covers G x, we have that
0
K x covers G x always
k(G x) k(G) + 1 (P=2) (G) + 1 equation
= 2:1 (G x) k(G x):
0
ff gg
Thus, (G x) = k(G x).

Cases 2: No maximum stable set of G contains x. In this case,
(G x) = (G): (2.2)
(G is an induced subgraph of G x, therefore (G x) (G). If (G x) > (G) it
means that we switched one vertex from a maximum stable set in G with x x , to obtain a
0
maximum stable set in G x with cardinality greater by one. But that means x must be an
element in a maximum stable set in G, contradicting case 2 assumption).
Since each clique of K intersects a maximum stable set exactly once (Lemma 2.1), this
is true in particular for Kx. But x is not a member of any maximum stable set. Therefore,
D = Kx ; fxg intersects each maximum stable set of G exactly once, so by Lemma 2.3
(GV D ) = (G) ; 1:
; (2.3)
This implies that
= (GV D ) equation
k(GV D ) induction
; ; = 2:3 (G) ; 1 equation
= 2:2 (G x) ; 1.
Taking a clique cover of GV D of cardinality (G x) ; 1 along with the extra clique
D fx g, we obtain a cover of G x of cardinality (G x). Therefore we have a clique cover
;
0
with the lowest possible number of cliques, because always k , so

k(G x) = (G x).
Lemma 2.7 (Fulkerson1971], Lovasz1972]). Let G be an undirected graph each of whose
proper induced subgraphs satises (P2), and let H be obtained from G by multiplication of
vertices. If G satises (P3), then H satises (P3).
Proof: Let G satisfy (P3) and choose H to be a graph having the smallest possible
number of vertices which can be obtained from G by multiplication of vertices but which
fails to satisfy (P3) itself. Thus,
!(H ) (H ) < jX j (2.4)
where X denotes the vertex set of H , yet (P3) does hold for each proper induced subgraph
of H.
As in the proof of the preceding lemma, we may assume that each vertex of G was
multiplied by at least 1 and that some vertex u was multiplied by h 2. Let U = fu1 : : :uh g
be the vertices of H corresponding to u. The vertex u1 plays a distinguished role in the proof.
By the minimality of H , (P3) is satised by HX u1 , which gives
;
(P 3)
jX j ; 1 = jX ; u1j !(HX u1 ) (HX u1 )
; ;
!(H ) (H )
equation 2:4
jX j ; 1
Thus, equality holds throughout, and we can dene
p = !(HX u1 ) = !(H )
;
q = (HX u1 ) = (H )
;
and
pq = jX j ; 1: (2.5)
Since HX U is obtained from G;u by multiplication of vertices, and using the assumption
;
that every proper induced subgraph of G satises (P2), Lemma 2.6 (Berge) implies that
HX U satises (P2). Thus, k(HX U ) = (HX U ) (HX u1 ) = q. Therefore, HX U can
; ; ; ; ;
be covered by a set of q complete subgraphs of H , say K1 : : : Kq (this is not necessarily a
minimum cover). Without lost of generality we may assume that these q cliques are pairwise
disjoint and that jK1j jK2j : : : jKq j. Obviously,
q
X
jKi j = jX ; U j = jX j ; h equation
= 2:5
pq ; (h ; 1)
i=1
Since jKi j p, at most h ; 1 of the cliques fail to contribute p to the sum. Hence,
jK1j = jK2j = : : : = jKq h+1 j = p: ;
Let H be the subgraph of H induced by X = K1 : : : Kq h+1 fu1g. Thus

0 0
;
jX j = p(q ; h + 1) + 1 h<2 pq + 1 equation

0
=

(2.6) 2:5
jX j:
Therefore H is a proper induced subgraph of H . So by the minimality of H , H satises
0 0
(P3):
!(H ) (H ) jX j: 0 0 0
(2.7)
But p = !(H ) !(H ), so 0
equation 2:7 equation 2:6

(H ) jX j=p > q ; h + 1
0 0
Let S be a stable set of H of cardinality q ; h + 2. Certainly u1 2 S , for otherwise S

0 0 0 0
would contain two vertices of a clique (by the denition of H ). Therefore, S = S U = 0 0
S + (U ; fu g) is a stable set of H with (q ; h + 2) + (h ; 1) = q + 1 vertices, contradicting

0 0
the denition of q.
Theorem 2.8 The Perfect Graph Theorem (Lovasz 1972]). For an undirected graph G =
(V E ), the following statements are equivalent:
(P1) !(GA ) = (GA) 8A V
(P2) (GA ) = k(GA ) 8A V
(P3) !(GA ) (GA ) jAj 8A V
Proof: The three statements hold for a one vertex graph. We may assume that the
theorem is true for all graphs with fewer vertices than G.
(P1) ) (P3). Suppose we can color GA in !(GA ) colors. Every optimal coloring (a
partition to stable sets) C1 : : : Cm satises:
m
X at most (GA ) vertices of a given color
jAj = jCij m max
i
jCij (GA ) (GA ) (P=1) !(GA ) (GA )
i=1
Therefore,
jAj !(GA ) (GA )
(P3) ) (P1). Let G = (V E ) satisfy (P3). Then each proper induced subgraph of G
satises (P3), and by induction satises (P1) (and also (P2)). It is sucient to show that
!(G) = (G).
If we had a stable set S of G such that !(GV S ) < !(G), we could then paint S in one
color and paint GV S in !(G) ; 1 other colors, and we would have !(G) = (G).
;
Now let's assume that such S does not exist. It means that for every stable set S of G
GV S has a clique K (S ) of size !(G). Let I be the collection of all stable sets of G. Keep
in mind that 8S 2I 9 ! ; clique K (S ) in GV S and in particular S \ K (S ) = .
;
Construct a (0,1)-matrix M with rows corresponding to vertices, and a column i corre-

sponding to each independent set S1 : : : S . Mij = 1 i vertex i belongs to K (Sj ). The
jI j
matrix is illustrated in gure 2.1.
In the column of T In each column at most

all 0. one 1. all other 0.
K(S 1) K(S 2) K(S |I| )
1 h1
.
0 0 1 .
0 0 0
T 0 0 0 .
0 1 0
n hn
Figure 2.1: The matrix M: each row correspond to a vertex in V , and each column to an
! ; clique K (S )
For each xi 2 V , let hi denote the number of cliques K (S ) which contain xi(=the number
of ones in the i'th row). Let H = (X F ) be obtained from G by multiplying each xi by hi
if h = (h1 : : : hn), then H = G h]. On the one hand, by Lemma 2.7,
!(H ) (H ) jX j: (2.8)
On the other hand,
X summing the matrix M every clique in I
over rows and columns: X
jX j = hi = jK (S )j is of size !:
= !(G)jIj: (2.9)
xi V
2 S 2I
Since every clique in G x contains at most one "copy" of vertex x 2 G, we have

!(G x) !(G). Therefore,
!(H ) !(G): (2.10)
If a maximum stable set contain some "copy" of xi then it will contain all its "copies"
(i.e. all hi "copies"). To nd (H ), we sum up the hi's of the vertices which induce a stable
set T (in G) , and do so to all the stable sets of G which are collected in I , searching the one
with a maximum sum (note that all maximum stable sets in H are obtained in this way).
Hence, X
(H ) = max
T
hi 2I
xi T2
(Note that the maximum stable set in H does not necessarily correspond to the maximum
stable set in G. The multiplicities may play a role too).
For each T , the sum above can be computed in two ways (gure 2.1):
1. Move row by row, but only for rows of vertices in T , and sum up the ones in each
row(= hi).
2. Move column by column, and sum up the the ones in each column which correspond
to vertices in T .
Hence, X X
max
T
hi = max
T
jT \ K (S )j]:
2I
xi T
2
2I
S 2I
T is a stable set and K (S ) is a clique, so

T \ K (S ) 1:
K (T ) is a clique in GV T , so
;
T \ K (T ) = 0:
In the second way of summing, at least one term will be zero, and all terms are at most one
(compare gure 2.1). So we conclude:
X
(H ) = : : : = max
T
jT \ K (S )j] jIj ; 1: (2.11)
2I
S 2I
Let us summarize what we have shown:
jX j = !(G)jIj equation 2:9
(H ) jIj ; 1: equation 2:11
!(H ) !(G): equation 2:10
Then,
(H )!(H ) !(G)jIj ; 1] = jX j jIj ; 1 < jX j
jIj
contradicting equation 2.8.
(P2) , (P3). G satises (P2) duality
, G satises (P1) proved above
, G satises (P3) duality
, G
satises (P3).
Corollary 2.9 A graph G is perfect i its complement G is perfect.
Proof: By duality, G satises (P2) , G satises (P1) . By the Perfect Graph Theorem
each of these conditions is sucient and necessary for perfection.
Corollary 2.10 A graph G is perfect i every graph H obtained from G by multiplication
of vertices is perfect.
Proof: One direction is trivial. The other direction follows immediately from Lemma 2.6
and the Perfect Graph Theorem.
The last two corollaries are useful in checking perfection: we can choose to check the
complement, and we can reduce non-adjacent vertices with identical neighborhoods and
check on the resulting simpler graph.
Chapter 3
Lecture 3
Scribe : Dekel Tsur
Lecture 3 : May 5
3.1 Perfect Graphs

3.1.1 p-Crirical Graphs
Denition An undirected graph G is called p-critical if G is not perfect, but every proper
induced subgraph of G is perfect.
Note: If G is p-critical then for all vertices x of G
(G ; x) = k(G ; x) w(G ; x) = (G ; x)
Theorem 3.1 If G is a p-critical graph on n vertices, then:
1. (G)w(G) = n ; 1
2. for all vertices x of G, (G) = k (G ; x) and w(G) = (G ; x)
Proof: By the Perfect Graph theorem, since G is p-critical we have (G)w(G) < n and
(G ; x)w(G ; x) n ; 1 for all vertices x. Since and w are monotone,
n > (G)w(G) (G ; x)w(G ; x) n ; 1
Hence, (G)w(G) = n ; 1 (G) = (G ; x) = k(G ; x) and w(G) = w(G ; x) = (G ; x)
35
3.1.2 A Polyhedral Characterization of Perfect Graphs
Let A be an m
n matrix. We dene the two polyhedra
P (A) = fxjAx 1 x 0g
PI (A) = convex ; hull(fxjx 2 P (A) x integralg)
where x is an n-vector, 1 is the m-vector of all ones, and 0 is the n-vector of all zeros. Clearly
PI (A) P (A), and for (0 1) matrix A without zero columns, P (A) and PI (A) are bounded
in the unit hypercube in Rn . The clique matrix of a graph G = (V E ) is a (0 1) matrix where
if V = f1 2 : : : ng and C1 C2 : : : Cm are all the maximal cliques of G numbered arbitrarily,
then Aij = 1 i j 2 Ci. See Figure 3.1 for an example.
1
2 3
6
6
6
0 1 1 1 0 7
7
7
2 6
6 7
7
6
6
6
1 1 0 0 0 7
7
7
3 4 6 7
4
0 0 0 1 1 5
Figure 3.1: A graph and its clique matrix

Theorem 3.2 (Chvatal, 75) Let A be the clique matrix of an undirected graph G. Then
G is perfect i P (A) = PI (A).
Polyhedron P is called integral if all the coordinates of all its extreme points are integers,
i.e. if P (A) = PI (A). In other words, theorem 3.2 is: G is perfect i P (A) is integral
polyhedron. In particular, this implies that optimization problems which can be posed as
integer programming problems with constraints PI (A), can be solved in polynomial time on
perfect graphs, by solving the corresponding linear program.
3.1.3 The Strong Perfect Graph Conjecture

It is easy to see that the odd cycle C2k+1 (for k 2) is a p-critical graph. By the Perfect
Graph theorem its complement C2k+1 is also p-critical. C2k+1 and C2k+1 (called odd hole and
odd anti-hole respectively) are the only known p-critical graphs. In 1960 C. Berge conjectured
that there are no other p-critical graphs. This conjecture is called the strong perfect graph
conjecture (SPGC).
The strong perfect graph conjecture can be rephrased in several equivalent forms:
3.2. TRIANGULATED GRAPHS 37
(1) G is perfect i it does not contain induced subgraph isomorphic to an odd hole or an
odd anti-hole.
(2) G is perfect i every cycle of length 5 in G or G contains a chord.
(3) The only p-critical graphs are C2k+1 and C2k+1 (for k 2).
The strong perfect graph conjecture remains an open question. However, the conjecture
was proved on several classes of graphs like planar graphs, circular arc graphs, graphs with
w 3, graphs with maximal degree 6, and also K1 3 - free, K4 - free, -free, and
-free graphs. (a graph is X-free if it does not contains induced subgraph isomorphic to X ).
3.2 Triangulated Graphs

3.2.1 Introduction
Denition An undirected graph G is called triangulated if every cycle of length > 3
contains a chord. In other words G is triangulated if it does not contain an induced subgraph
isomorphic to Cn for n > 3.
Being triangulated is a hereditary property, meaning that every induced subgraph of a
triangulated graph is also triangulated. Triangulated graphs have also been called chordal,
rigid-circuit, monotone transitive and perfect elimination graphs. For an example of trian-
gulated graph see Figure 3.2.
p
a b c d
x y q z
G1
g f e G2
r
Figure 3.2: Two graphs, the left one is triangulated, and the right one is not.
3.2.2 Characterizing Triangulated Graphs

Denition A vertex x in G is called simplicial if Adj (x) induces complete graph in G (i.e.,
Adj (x) is a clique).
We will prove that every triangulated graph has a simplicial vertex. Since being triangu-
lated is a hereditary property, it gives us an iterative procedure for identifying triangulated
graphs: Search for simplicial vertex, remove it from the graph and repeat until there is no
simplicial vertex (in that case, the graph is not triangulated) or there are no vertices left
(and then the graph is triangulated).
Denition Let G be an undirected graph and let
= v1 : : : vn] be an ordering of the
vertices.
is called perfect elimination order (PEO) if every vi is a simplicial vertex in the
induced subgraph G vi ::: vn . In other words, each set Xi = fvj 2 Adj (vi)jj > ig is complete.
f g
In Figure 3.2,
= a g b f c e d] is one of the perfect elimination orders of G1. In
contrast to this, G2 has no simplicial vertex, so obviously it has no perfect elimination order.
Denition A subset S V is called a vertex separator for nonadjacent vertices a b (or a-b
separator), if in GV S a and b are in dierent connected components. If no proper subset of
;
S is an a-b separator, S is called minimal vertex separator.

The example will be again on the graphs in Figure 3.2. In G2, the set fy zg is a minimal
vertex separator for p and q, whereas fx y zg is a minimal vertex separator for p and r. In
G1 all the minimal vertex separators have 2 vertices which are neighbors.
Theorem 3.3 (Dirac 61, Fulkerson and Gross 65)
For undirected graph G the following statements are equivalent:
(1) G is triangulated.
(2) G has a perfect elimination order. Moreover, any simplicial vertex can start a perfect
order.
(3) Every minimal vertex separator induces a complete subgraph of G.
Proof:
(3) ! (2) Let a x b y1 : : : yk a] (k 1) be a simple cycle. Any minimal a-b separator must
contain vertices x and yi for some i, but from (3) it follows that xyi 2 E , which is a
chord of the cycle.
(1) ! (3) Let S be an a-b separator. We will denote GA GB the connected components of GV S
containing a and b respectively. Since S is minimal, every vertex x 2 S is a neigh-
;
bor of a vertex in A and a vertex in B . For any x y 2 S there exist minimal paths
x a1 : : : ar y](ai 2 A) and x b1 : : : bk y](bi 2 B ). Since x a1 : : : ar y b1 : : : bk x]
is a simple cycle of length 4, it contains a chord. For every i j aibj 2= E (S is a-b
separator) and also aiaj 2= E bibj 2= E (by the minimality of the paths) Necessarily
xy 2 E is the chord.
To prove the rest of the theorem, we will rst prove the following lemma:
Lemma 3.4 (Dirac 61) Every triangulated graph G has a simplicial vertex. If G is not
complete then it has two nonadjacent simplicial vertices.
Proof: If G is complete the lemma is trivial. We will prove the lemma for noncomplete
graphs by induction on the number of vertices in G. The lemma is trivial for graphs with
1 or 2 vertices. Assume the statement is correct for graphs with n ; 1 vertices and let G
be a triangulated graph with n vertices. Let a b be two nonadjacent vertices in G, and let
S be minimal a-b separator. Denote by GA GB the connected components of GV S which ;
contain a b respectively. We shall look at GA+S :

1. If GA+S is complete, then each of its vertices is a simplicial vertex in GA+S . Particularly,
every vertex in A which is simplicial in GA+S is also simplicial in G (as there are no
edges between A and B ).
2. If GA+S is not complete, by induction it contains two nonadjacent simplicial vertices.
Since S is complete (we already proved 1 ! 3 in Theorem 3.3) at least one of the
simplicial vertices is in A and like before this vertex is simplicial in G.
We showed that A contains a simplicial vertex in G. In the same manner, B contains a
simplicial vertex in G. These two vertices are not adjacent.
Now we will continue the proof of Theorem 3.3:
(1) ! (2) By Lemma 3.4, G contains a simplicial vertex x. GV x is triangulated and contains
;f g
less vertices, so by induction it has a perfect elimination order

i1 : : :
in 1 ]. Since x
is simplicial x
i1 : : :
in 1 ] is a perfect elimination order in G.
;
(2) ! (1) Let C be a simple cycle in G. Let x be the rst vertex from C to appear in the perfect
elimination order. In the induced graph before eliminating x, the two neighbors of x
in the cycle still exist, and since x is simplicial in that graph, these two vertices form
a chord in the cycle. Hence, every cycle in G contains a chord.
1. assign label 0 to each vertex

2. for i n to 1 by ;1 do
3. select: pick an unnumbered vertex v with largest label
4.
(i) v
5. update: for each unnumbered vertex w 2 Adj (v) do
6. add 1 to label(w) end for
7. end for
Figure 3.3: Code for MCS algorithm
3.2.3 Recognizing Triangulated Graphs

If a graph G is triangulated and is not complete, it has two nonadjacent simplicial vertices.
Also every induced subgraph of G is also triangulated. Therefore it is possible to choose
any vertex v and decide it will be last in the PEO (it is always possible to choose another
simplicial vertex instead of v). Afterwards, it is possible to choose a vertex u adjacent to
v and put it in (n ; 1)st position. It is possible to continue choosing vertices backwards
to the perfect elimination order. Using the above idea will give us a linear time algorithm
for recognizing triangulated graphs. The following linear time algorithm (called Maximum
Cardinality Search) by Tarjan and Yannakakis (1985), receives a graph G given by adjacency
list for each vertex, and outputs an order
of the vertices, with the property that if G is
triangulated then
is a PEO. We will show later how to check if this vertex order is a perfect
elimination order (i.e. if the graph is triangulated) in linear time. The code is in Figure 3.3,
and a running example is given in Figure 3.4.
3 2 1 2
7 6
5 8 5 4
6 9 8 7
1 4 9 3
Figure 3.4: Two dierent possible running of the MCS algorithms.
Time Complexity
Let Si be the set of all vertices v that have not been numbered yet, and have label(v) = i.
Each Si is represented by a doubly linked list. The set of active Si-s (all those Si-s which
were non-empty at some point) will also be represented by a doubly linked list in increasing
order of i,with a pointer to the non-empty set with largest label. For each vertex we will
store its label i and a pointer to its position in Si. When v receives a number we take v
out of its list and move each neighbor of v one list upwards. This takes O(1 + degree(v)),
therefore the overall time complexity is O(jV j + jE j).
Chapter 4
Lecture 4
Scribe : Izik Pe'er
Lecture 4 : May 12
4.1 Recognizing Triangulated Graphs

4.1.1 Generating a PEO
In lecture 3 we have introduced the MCS algorithm for generating a PEO for a triangulated
graph, and proved that it is a linear time algorithm. We shall prove the correctness of this
algorithm.
Theorem 4.1 (Tarjan and Yannakakis 85)
A graph G = (V E ) is triangulated if and only if the MCS algorithm produces a permutation
1 which is a PEO.
Proof:
) Trivial.
( We shall prove the claim by induction on the number of vertices n.
1We shall refer to each permutation simply as a one-to-one function from f1 : : : ng to V , so that (1)
is the rst vertex in the order, (2) is the second, etc. Surely, for each vertex v, ;1(v) is v's serial number
in the order.
43
The theorem is trivially true for n = 1. Assume correctness for every graph with less
than n vertices. Let G(V E ) be a triangulated graph with n vertices, and let
be
the permutation generated by the MCS algorithm with input G. According to the
induction hypothesis it suces to show that the vertex v =
(1) is simplicial.
Claim 4.2 G cannot contain an induced chordless path = u = v0 v1 : : : vk w] for
k 1 so that
8i 1 i k :
1(vi) <
1(u) <
1(w)
; ; ;
(4.1)
Proof: Assume, by contrary, that such a path exists. Let be a path satisfying 4.1,
such that
1(u) is maximal. u was numbered before vk , despite the fact that w is a
;
neighbor of vk and not of u. According to the maximum cardinality rule, this implies
there exists a vertex x, such that x is a neighbor of u and not of vk , and x was numbered
before u, as in gure 4.1.
v v u w
0 k
Figure 4.1: The path . Vertices appear ordered from left to right according to the MCS
algorithm output.
Let j be the maximal index such that vj is a neighbor of x. Surely j is well dened,
since u = v0 is a neighbor of x, and j < k, by the denition of x. It follows from
the maximality of j that the induced path = x vj : : : vk w] is chordless (xw 62 E ,
0
otherwise = x vj : : : vk w x] is a chordless cycle of length 4, contradicting G's

0
chordality).
If
1(x) <
1(w) (as in gure 4.3), then for each j i k,
1(vi) <
1(u) <
; ; ; ;

1(x), so the path meets the condition 4.1, and contradicts the maximality of
; 0

1(u), since
1(x) >
1(u). Hence
1(x) >
1(w) (as in gure 4.4), and the
; ; ; ; ;
path = w vk : : : vj x] satises 4.1, contradicting the maximality of

1(u).
00 ;
It follows that there is no path satisfying 4.1.

Back to the proof of Theorem 4.1:
4.1. RECOGNIZING TRIANGULATED GRAPHS 45
v v v u w
j 0 k
Figure 4.2: The vertices x and vj
vj v u x w
k
Figure 4.3: The path 0
v v u w x
j k
Figure 4.4: The path 00
v u w
Figure 4.5: A forbidden structure in a chordal graph

Denote v =
1(1), and assume, by contrary that, v is not a simplicial vertex. Then
;
there exist u w 2 Adj (v) such that uw 62 E , and

1(u) <
1(w) (as in gure 4.5.
; ;
It follows that the path u v w] satises 4.1 of Claim 4.2, yielding a contradiction.
4.1.2 Testing an Elimination Scheme

In order to recognize whether a graph is triangulated, we can use the MCS algorithm to get
an elimination scheme, but we have yet to describe how we can test whether this elimination
scheme is a PEO.
Naive Algorithm
Simulate the elimination process, and for each vertex to be eliminated, check whether its
noneliminated neighbors are a clique. This strategy results in a total time complexity of
O(jV j jE j).
Ecient Algorithm
We can improve complexity if we collect a list of desired neighbors for each vertex, and check
all the members in this list only once.
begin
1. for each vertex v do A(v)

2. for i 1 to n ; 1 do
3. v
(i)
4. X fx 2 Adj (v)j
1(v) <
1(x)g
; ;
5. 6 then
if X =
6. u
(minf
1(x)jx 2 X g)
;
7. concatenate X n fug to A(u)

end if
8. if A(v ) n Adj (v ) =
6 then return FALSE
end for
9. return TRUE
end
Figure 4.6: Code for PEO testing algorithm

4.1. RECOGNIZING TRIANGULATED GRAPHS 47
Example
8
3 4 ‘2
5 7
Figure 4.7: An input chordal graph and elimination order

When applying the algorithm in 4.6 to the input graph and elimination order in 4.7, the
algorithm does the following iterations:
Iteration 1 u=3 A(3) = f4 8g X = 3 4 8
Iteration 2 X = 4 7 u=4 A(4) = f7g A(2) Adj (2)
Iteration 3 X = 4 5 8 u=4 A(4) = f7 5 8g A(3) Adj (3)
Iteration 4 X = 5 7 8 u=5 A(5) = f7 8g A(4) Adj (4)
Iteration 5 X = 6 7 8 u=6 A(6) = f7 8g A(5) Adj (5)
Iteration 6 X = 7 8 u=7 A(7) = f8g A(6) Adj (6)
Correctness of the Algorithm

Theorem 4.3 The algorithm in gure 4.6 returns TRUE if and only if
is a PEO.
Proof:
( Suppose the algorithm returns FALSE on iteration number
1(u). This may happen
;
only if in stage 8 at that iteration there exists an unnumbered vertex w 2 A(u) n Adj (u).
The vertex w was added to A(u) at stage 7 of a prior iteration, number
1(v). It
;
follows that
1(v) <
1(u) <
1(w), for u w 2 Adj (v), but since uw 62 E (as in
; ; ;
gure 4.5),
is not a PEO.
) Suppose
is not a PEO, and the algorithm returns TRUE. Let v be a vertex with
maximal
1(v), such that the set Xv = fwjw Adj (v)
1(v) <
1(w)g does not
; ; ;
induce a clique. The vertex v is well dened, because if all of the Xv 's are cliques, then

is a PEO. Let u be the vertex dened in stage 6 of iteration
1(v). In stage 7 of this
;
iteration, Xv n fug is added to A(u). Since the algorithm returns TRUE in stage 8 of
iteration number
1(u), the condition of stage 8 (i.e. The condition A(u) Adj (u))
;
is not met, so every x 2 Xv n fug is a neighbor of u.

Since
1(v) is chosen to be the maximal vertex for which Xv is not a clique, and v is
;
prior to u, Xu is a clique, so all u's neighbors in Xv n fug are adjacent to each other
(as in gure 4.8, hence Xv is a clique, and a contradiction follows.
v u
xvu
Figure 4.8: Xv n fug is a clique
Complexity of the Algorithm

Theorem 4.4 The algorithm 4.6 runs in time O(jV j + jE j).
Proof: We shall hold
and
1 as vectors, and for each v we shall hold Adj (v ) and A(v )
;
as lists, allowing duplicates to be concatenated to A(v). Stage 8 of iteration

1(v) can ;
be implemented in O(jAdj (v)j + jA(v)j) by holding a bit vector Test of length n, and the
condition A(v) n Adj (v) = is implemented in the following way :
1. for each w 2 Adj (v), set the bit Test(w) to 1.
2. for each w 2 A(v), if Test(w) is 0, return FALSE.
3. for each w 2 Adj (v), clear the bit Test(w) to 0.
4. return TRUE
4.2. TRIANGULATED GRAPHS AS INTERSECTION GRAPHS 49
The resulting overall complexity is O(jV j + P v 2 V jAdj (v)j + P v 2 V jA(v)j), where
jA(v)j is the last (maximal) size of A(v), including the duplicates. For each v 2 V , Adj (v)
adds vertices only to one list A(u), and the number of added vertices is at most jAdj (v) ; 1j.
Therefore P v 2 V jA(v)j < P v 2 V jAdj (v)j = O(jE j) and the overall time complexity is
O(jV j + jE j).
Corollary 4.5 There is a linear time algorithm for checking whether a given graph is tri-
angulated.
4.2 Triangulated Graphs as Intersection Graphs

In Lecture 1 it was shown that every interval graph is triangulated, and that interval graphs
are characterized as intersection graphs. Is there such a characterization for triangulated
graphs as well? We shall see that a graph G is triangulated if and only if G is an intersection
graph of a family of subtrees of a tree. The tree containing all the subtrees is called the host
tree.
f
h
g f
c g
e
i d
c
h
d e i
a b b
Figure 4.9: The intersection graph of a set of subtrees of a host tree is triangulated
4.2.1 Evolutionary Trees

An application for this characterization of triangulated graphs is the evolutionary tree, also
called the phylogenetic tree. This model is formulated in the study of evolution, where we
wish to determine the way a group of species evolved in time to their current properties. We
use the following assumptions :
There exists a single ancient species, from which all the other species have evolved.
Each ancient species might have split into two subspecies, because of a mutation chang-
ing a property of one of the subspecies.
Mutations are rare, so each mutation can occur only once in the tree.
We can therefore dene the evolutionary tree, which is a rooted binary tree. In this tree,
every vertex corresponds to a species (every leaf corresponds to a current species). Because
of the third assumption for every property the set of all the species having that property
corresponds to a set of tree vertices that induces a connected subgraph, i.e. a subtree.
We can build a graph with vertices that match properties. Two vertices will be adjacent
if their properties appear together in a certain species (as in gure 4.10). This graph must
be triangulated.
A B C D
A1
1to2 1 2 3
A1 A2
AtoB AtoD
C1 D2
BtoC 2to3
A1 B1 C1 A2 D2 D3
Figure 4.10: An evolutionary tree describing the properties A B C D 1 2 3, and the cor-
responding triangulated intersection graph
4.2.2 Triangulated Graphs as Intersection Graphs
Denition Let S = fTigi I be a family of subsets of a set T . S has the Helly property if
and only if for every J I , if 8i j 2 J : Ti \ Tj 6= then j\J Tj 6= 0. (gure)
2
Figure 4.11: The left family has the Helly property, while the right family does not
Lemma 4.6 Suppose that fTigi S is a set of subtrees of T , and that there exist points a b c
on T , such for each i 2 S , Ti contains at least two of them, then i\S Ti 6=
2
Proof: Denote: P1 the simple path from a to b in T , P2 the simple path from b to c in
T , and P3 the simple path from c to a in T . Since T is a tree, P1 \ P2 \ P3 6= (otherwise
P1 ; P2 ; P3 contains a cycle). Every Ti contains at least one of the paths Pi i = 1 2 3.
Hence, i\S Ti P1 \ P2 \ P3 6= .
2
Common intersection point
a
b
Figure 4.12: The paths P1, P2, and P3 must intersect

Theorem 4.7 A family of subtrees of a host tree has the Helly property.
Proof: We shall prove the claim by induction on k, the size of the family J . For k = 2
the theorem is trivially true. Assume correctness for jJ j k. Let J be a family of k + 1
trees : fT1 : : : Tk+1g, each pair of which having a non-empty intersection. By the induction
hypothesis:
j=1\k Tj 6= so there exists a 2 j=1\k Tj .
k+1 k+1
j=2
\ Tj 6= so there exists b 2 j=2
\ Tj .
In addition, by assumption, T1 \ Tk+1 6= so there exists c 2 T1 \ Tk+1.
k+1
Lemma 4.6 applied for fT1 : : : Tk+1g and the points a b c, yields j=1\ Tj 6= .
Theorem 4.8 (Walter 72, Gavril 74, Buneman 74)
Let G(V E ) be an undirected graph. The following statements are equivalent :
(1) G is triangulated.
(2) G is the intersection graph of a family of subtrees of a tree.
(3) There exists a tree T = (K E ) (as in gure 4.13) such that K is the set of maximal
cliques in G, and for every v 2 V , if K v = fC 2 K jv 2 C g, then TK v , the subgraph
induced by K v , is connected (hence a subtree).
Proof:
(3) ) (2) Let T be a tree satisfying (3). We shall prove that vw 2 E , TK v \ TK w 6= :
) If v w 2 V , vw 2 E , then for some maximal clique C 2 K , v w 2 C so TK v \
TK w 6= .
( If TK v \ TK w 6= then there are maximal cliques that contain both w and v, so
vw 2 E .
It follows that G is an intersection graph of the family fTK v gv V of subtrees of T .
2
(2) ) (1) Let fTK v gv V be a family of subtrees of a tree T , for which vw 2 E , TK v \ TK w 6= .

2
Suppose that G contains an induced chordless cycle v0 v1 : : : vk 1 v0], for k > 3. Let
the corresponding subtrees be T0 T1 : : :Tk 1, respectively. It is obvious that ji ; j j
;
1(modk) , Ti \ Tj 6= . Let ai be an arbitrary point in Ti \ Ti+1, for i = 0 : : : k ; 1

;
(throughout, indices are mod k). We would like to prove that the ai's form a cycle,
10 2 5 7
1 9
4 11 8
K2
2,10 1,2,3 2,3,5 3,5,6
K3 K6
1,3,4 3,11 6,7,8,9
E )
Figure 4.13: The tree T = (K
but we have to \throw away" possible branches, which are not on the main cycle, but
that the ai's may lie on. Let bi be the last common point in the simple paths ai ! ai 1
and ai ! ai+1. The path ai ! ai 1 is in Ti and the path ai ! ai+1 is in Ti+1, therefore
;
bi 2 Ti \ Ti+1.
;
Denote by Pi the simple path from bi 1 to bi. It follows that Pi Ti, hence Pi \ Pj =
for every ji ; j j > 1. On the other hand, Pi \ Pi+1 = fbig for every i. We have shown
;
k
that i=1Pi is a simple cycle in T , contradicting T being a tree.
T
1
a a
1 0 T
a b 0
3 3
b
1
P
1 b
T 0
3 P
3 P
2
T b a
2 2 2
Figure 4.14: The paths P0 : : : Pk 1 form a cycle

;
B
A
Figure 4.15: The case B 6= Y
(1) ) (3) We shall prove the claim by induction on n, the number of vertices in G. Assume
correctness for every graph with less then n vertices.
Figure 4.16: The case B = Y
{ If G is the complete graph then K = V , and T contains one vertex making the
claim trivial (this applies also for n = 1).
{ If G is not connected with connected components G1 : : : Gk , then by the induc-
tion hypothesis for each Gi there is a host tree Ti satisfying (3), for i = 1 : : : k.
For each i = 1 : : : k ; 1, connect Ti to Ti+1 by adding an arbitrary edge. The
resulting graph is a tree T satisfying (3).
{ If G is a connected, not complete, triangulated graph, then let a be a simplicial
vertex in G. The set A = fag Adj (a) is a maximal clique in G. Denote
U = fu 2 AjAdj (u) Ag and Y = A n U .
U 6= because a 2 U .
Y 6= because G is not complete.
V n A 6= because G is connected and not complete.
The graph G = GV U is triangulated and contains less than n vertices, therefore,
0
by the induction hypothesis there is a tree T = (K E ), whose vertices corre-

n
0 0 0
spond to the maximal cliques in G , in a way that for each vertex u 2 V n U , the
0
set K v = fC 2 K jv 2 C g induces a connected subtree of T .

0 0 0
Let us examine the connection between K and K , the sets of maximal cliques in
0
G and in G , respectively. Let B be the maximal clique in G that contains Y .

0 0
There are two cases :

B = Y In that case T can be generated from T by renaming B as A. For each y 2 Y
0
: K y = K y + fAg n fB g and K y induces the same subtree as K y , because all

0 0
we did is renaming B as A (as in gure 4.16).

B 6= Y In that case T can be generated from T by adding a vertex A, and connecting
0
it to B . For each y 2 Y : K y = K y + fAg and K y induces a connected subtree

0
(as in gure 4.15).
In both cases, for each vertex u 2 U , fAg = K u , and for each vertex v 2 V n A,
K v = K v , and these sets induce connected subtrees in T , respectively.
0
Chapter 5
Lecture 5
Scribe : Lea Epstein
Lecture 5 : May 19
5.1 Triangulated Graphs Are Perfect

5.1.1 Proving This Property
In some graph families, the minimum graph coloring and the maximum clique problem, can
be simplied using the principle of separation into pieces, as follows:
Theorem 5.1 (Berge, 1973) Let G be a connected, undirected graph, let S be a vertex
separator in G, and let GA1 GA2 ::: GAt be the connected components of GV S . If S is a
;
clique (not necessarily maximal), then

(G) = maxi(GS+Ai )
w(G) = maxi w(GS+Ai )
Proof: Clearly (G) (GS +Ai ) for each i, (because GS +Ai is an induced subgraph of
G) , so
(G) k = maxi (GS+Ai ) On the other hand, G can be colored using exactly k colors, as
follows:
First, color GS , (S is a clique, so there is only one way to do it).
57
Figure 5.1: The Graph

Then, independently extend the coloring to each piece GS+Ai . Each piece can be
colored with at most k colors, and S is always colored using jS j colors , so it is possible
to extend the coloring on S to GS+Ai (renaming the colors if necessary).
There will be no contradictions because if v 2 Ai u 2 Aj i 6= j have the same color
they are disconnected (because S separates them).
Next, consider the case of clique. Again w(G) w(GS+Ai ) for each i, so
w(G) maxi w(GS+Ai ) = m.
Let X be a maximum clique of G, i.e. jX j = w(G). It is impossible that two vertices of X
lie in GAi and GAj , i 6= j , since they are not connected . Thus, X is wholly contained in one
of the pieces, say GS+Ar . Hence m w(GS+Ar ) jX j = w(G) .
Therefore, w(G) = m.
Corollary 5.2 Let S be a separating set in a connected graph G = (V E ), and let
GA1 ,...,GAt be the connected components of GV S . If S is a clique, and if each subgraph
;
GS+Ai is perfect, then G is perfect.

Proof: The proof is by induction on n, the number of vertices in the graph.
Claim 5.3 It is sucient to prove (G) = w(G)
Proof: It is trivial if n = 1. Suppose jV (G)j = n and the claim is true for all graphs with
less then n vertices. Suppose S is a separating set, and a clique in G, and each GS+Ai is
perfect. Pick v 2 V , and let G = G ; fvg. If v 2 S , dene S = S ; fvg. In G S is also
0 0 0 0
a clique , and also each GS +Ai is perfect (because this is an induced subgraph of a perfect
0
graph). If v 2 Ai , S is still a clique , and GS+Ai v is perfect. In both cases, from the
;f g
induction hypothesis , G is perfect . This implies that for every proper induced subgraph
0
H , (H ) = w(H ) .
Now prove that (G) = w(G) .
(G) = maxi (GS+Ai ) (using theorem 5.1).
maxi (GS+Ai ) = maxi w(GS+Ai ) (because every GS+Ai is perfect),
w(G) = maxi w(GS+Ai ) (using theorem 5.1 again),
Thus, (G) = w(G) .
5.1. TRIANGULATED GRAPHS ARE PERFECT 59
Theorem 5.4 (Berge 1960, Hajnal-Suranyi 1958)
Every triangulated graph is perfect .
Proof: By induction on the number of vertices in the graph. Let G be a triangulated
graph. We shall assume it is connected, otherwise we consider each component individually.
If G is complete, it is perfect.
Otherwise let S be a minimal vertex separator for a pair a b of nonadjacent vertices.
By theorem 3.3, S is a clique. Each of the subgraphs GS+Ai (dened in theorem 5.1) is
triangulated (a hereditary property) and thus by induction hypothesis using corollary
5.2, G is perfect.
5.1.2 Other Results

Denition A Meyniel graph is a graph in which there are at least two chords in every odd
cycle of length 5.
Figure 5.2: Some examples of Meyniel graphs (which are not chordal!)
Theorem 5.5 (Meyniel 1976) Every Meyniel graph is perfect.

(The proof can be found in Golumbic's book section 6.4).
Denition G is weakly chordal graph if G and G do not contain an induced Ck , k 5 .

Theorem 5.6 (Hayward 1984) Weakly Chordal graphs are perfect.
V1 V2 V3 V4
P
Figure 5.3: Example-A Chain Graph
5.2 Computing the Minimum Fill In

5.2.1 The Problem
The Minimum ll-in problem is dened as follows: Given a graph, nd the minimum number
of edges (ll in) whose addition makes the graph triangulated. i.e. for G = (V E ) nd F ,
such that jF j is minimal and G = (V E + F ) is triangulated. This problem has to do with
0
the Gaussian elimination of symmetric matrices. (When looking for an elimination with
a maximal number of zeros in the matrix in each step - the problem arises in numerical
algebra). In the book of Garey and Johnson, the minimum ll in problem (called there
Chordal Graph Completion) is mentioned as a major open problem, shortly after the book
was published, Yannakakis showed that the problem is NP complete.
5.2.2 Fill In is NP-Hard

Chain Graphs
Denition A bipartite graph G = (P Q E ) is a chain graph if one of the parts, say P , can
be ordered v1,v2,..,vk so that Adj (v1) Adj (v2) ::: Adj (vk) (the chain property).
This property is hereditary, (if we take out a vertex from P , the chain of containment
doesn't change, if we take out a vertex from Q, the containment order does not change).
Claim 5.7 If a bipartite graph is a chain graph, both parts of the graph have the chain
property.
Proof: We have to nd an order for Q. Let Qi = \j i Adj (vj ) In this way,
Q1 Q2 ::: Qk . Let Ri = Qi ; Qi+1, i = 1 ::: k ; 1, Rk = Qk . Then R1 + R2 + ::: + Rk

is a partition of Q. Impose a linear order on the vertices in Q such that if u 2 Ri,v 2 Rj

and i < j then v u. Thus if v u, u v 2 Q then either u v 2 Ri and Adj (u) = Adj (v) =
fv1 ::: vig, or u 2 Ri,v 2 Rj and i < j , in which case Adj (v) = fv1 ::: vj g fv1 ::: vig =
Adj (u).
5.2. COMPUTING THE MINIMUM FILL IN 61
V1 V2 V3 V4
P
Q
U3 U1 U2 U4
a b c d
Figure 5.4: Ordering Q, in the same chain graph
For the example in Figure 5.4, we have:

Q1 = fa b c dg R1 = fdg
Q2 = fa b cg R2 = fag
Q3 = fb cg R3 = fb cg
Q4 = R4 =
Denition Two edges ij ,kl are independent if the graph induction on fi j k lg induces
2K2. (see gure 5.5)
i j
k l
Figure 5.5: Independent edges (broken lines denote non-edges)
Theorem 5.8 A bipartite graph G is a chain graph i G does not contain two independent
edges.
Proof: (=)) Let ij and kl be independent edges. Then Adj (i) 6 Adj (k), Adj (i) 6 Adj (k)
and G is not a chain graph.
((=) If this isn't a chain graph, there must be two vertices: i k 2 P so that
u x
v y
Figure 5.6: Independent edges that form an induced C4
Adj (i) 6 Adj (k) and Adj (i) 6 Adj (k)

there exist: j 2 Adj (i) ; Adj (k) and l 2 Adj (k) ; Adj (i)
and the edges ij and kl are independent.
Denition For a bipartite graph G = (P Q E ) we dene a new graph C (G) = (N E ), 0
where N = P Q, E = E + fuv j u v 2 P g + fuv j u v 2 Qg . (i.e. we place the

0
independent set on each side of the clique. See gure 5.7).
V
4 π(4)
V
3
V
1
V
2 π(1)
Figure 5.7: C (G) for the graph G of Figure 5.3
Lemma 5.9 G is a chain graph i C (G) is triangulated.

Proof: (=)) If G is a chain graph, by lemma G contains two independent edges: ux , vy
.
Thus, fu v y xg is an induced C4 in C (G) .
((=) Let G be a chain graph, let be an ordering of P , such that
Adj ((1)) Adj ((2)) ::: Adj ((p))
where p = jP j . Since the property of being a chain graph is hereditary, it suces to show
that C (G) has a simplicial vertex. The neighborhood Adj ((p)) of (p) in C (G) is
0
Adj ((p)) = AdjG((p)) + P ; (p)

0
5 4
3 4 5 1 2
3 2 1
3 5 4 2 1
Figure 5.8: a graph (left) with two possible layouts of cost 8 (center) and 6 (right)
AdjG((p)) Q and Q is a clique, so AdjG((p)) is a clique in C (G). P ; (p) P and P
is a clique, so P ; (p) is a clique in C (G). Also, since Adj ((p)) Adj (v) for every v 2 P ,
all the nodes of P are adjacent to all nodes of Adj ((p)) . Therefore Adj ((p)) is a clique
0
in C (G), and (p) is a simplicial vertex of C (G).
Optimal Linear Arrangement (OLA)

A linear arrangement (or a layout) of a graph G = (N E ) is a 1 ; 1 mapping
: N ! f1 2 :: jN jg. For an edge e =Puv of G , let (e ) = j 1(u); 1(v)j. The cost c()
; ;
of the linear arrangement is c() = e E (e ). The optimal linear arrangement problem

2
is to decide, given a graph G and an integer k, whether there exists a linear arrangement
of G with cost c() k . This problem is NP-complete.
Chain Graph Completion (CGC)

Given a bipartite graph G = (P Q E ) and a constant k, is there a set of non-edges F such
that jF j k and (P Q E + F ) is a chain graph?
Theorem 5.10 ( Yannakakis , 1981)
Chain graph completion is NP-complete.
Proof: Reduction from OLA: Given an instance of OLA: a graph G = (N E ) and an
integer k , we build an instance of CGC: a bipartite graph G = (P Q E ) as follows: P = N
0 0
, and Q contains two vertices e1 and e2 for every edge e of G, and a set R(v) of n ; d(v)
vertices for every vertex v of N , where n = jN j and d(v) is the degree of v in G. If e = (u v)
is an edge of G, then the nodes e1 e2 that correspond to e are adjacent to u and v. The
nodes in R(v) are adjacent to v. In Fig 5.9 we see an example of this construction.
Figure 5.9: the graph G (left) and G' (right)
Let A = n (n2 1) ] ; 2m, where m is the number of edges in G. The constant for CGC will
2 ;
be k = k + A. Let l(G) be the minimum cost of a linear arrangement of G, and h(G ) the
0 0
minimum number of edges whose addition to G gives a chain graph. We shall prove that
0
h(G ) = l(G) + A Thus l(G) k () h(G ) k + A. which will prove the correctness of the
0 0
reduction. First observe that an ordering of N species uniquely a minimal set H () of

edges whose addition makes G into a chain graph with the neighborhoods of the nodes in
0
P ordered according to . For every node x 2 Q, let

(x) = maxfi j (x (i)) 2 E g. Then 0
v u v
P
e
u Q
1 2 3 4 e1 e2
Figure 5.10: Example-Edges we have to add

H () = f(x (j )) j x 2 Q j < 1(x)g ; E .
; 0
Conversely, suppose that F is a set of edges such that G (F ) = (P Q E F ) is a chain graph
0 0
and let be an ordering of the nodes in P according to their neighborhoods in G(F ). It is 0
easy to see that F H (), and therefore if F is a minimal augmentation then F = H ().

Let h() = jH ()j. In order to complete the proof, it suces to show that for every ordering
of N , h() = c() + 2 ] ; 2m,where c() is the cost of the linear arrangement of G.
n 2(n 1) ;
Let be an ordering of N . For every v 2 N and x 2 R(v), (see Figure 5.11) H () contains
1(v) ; 1 edges incident to x. Let e = uv be an edge of G, and suppose without loss of
;
generality that 1(u) < 1(v). The number of edges of H () incident to each of the two
; ;
nodes e1,e2 that correspond to e is
Q P
x
Figure 5.11: x and v
1(v) ; 2 = 1(u) + 1(v) ; 1(u)] ; 2

; ; ; ;
= 1(u) + (e ) ; 2
;
thus the number of edges of H () incident to e1 and e2 is

2( 1(v) ; 2) = 1(v) ; 2 + 1(u) + (e ) ; 2
; ; ;
= 1(u) + 1(v) + (e ) ; 4. Consequently:

; ;
X X
h() = 1(v) ; 1]
;
(5.1)
v N x R(v)
2 2
X
+ 1(v) + 1(u) + (e ) ; 4]
; ;
(5.2)
e=uv E2
X
= (n ; d(v))( 1(v) ; 1);
(5.3)
v N
2
X X
+ d(v)( 1(v)) +
;
(e ) ; 4m (5.4)
v N
2 e E
2
X X
= n 1(v) ; 1] +
;
d(v) + c() ; 4m (5.5)
v N
2 v N
2
= c() + n (n2; 1) ] ; 2m
2
(5.6)
the last equality follows X
d(v) = 2m (5.7)
v N
2
and
1(v) ; 1] = 0 + 1 + 2 + ::: + (n ; 1) = n(n ; 1) ]
X ;
(5.8)
v 2 N 2
The result for Fill in

Theorem 5.11 Minimum Fill in is NP-hard
Proof: Reduction from CGC: Given G = (P Q E ) and a constant k we build the graph
C (G), and give the same k ( as an instance of the minimum ll in problem). By Lemma 5.8
Adding edges to C (G) creates a triangulated graph i adding the same edges to G creates a
chain graph.
5.2.3 Problems
Maximum Chordal subgraph: Remove minimum number of edges from a graph,
until you get a triangulated graph. This problem is still open. There exists an algorithm
for nding a maximal ll in (not maximum), it's time complexity is O(mn) where
m = jE j , n = jV j (Ohtsuki, 1976).
Approximating the ll in: Find in polynomial time a ll-in set (i.e. a set F such 0
that (V E +F ) is chordal) such that if F is an optimal ll in, E + F ] f (n). The best
0
p
0 j j j j
E+Fj j j j
f(n) known today is O( klog4n) for graph with max degree k (Klein et al., 1990). Can
one get better approximation, or prove lower bounds on the achievable approximation,
assuming P 6= NP ?
parametric ll-in problem: How fast can we solve the ll in problem for xed k?
Simple enumeration of all possible sets of size takes O(n2k m). Two better algorithms
are known (Kaplan,Shamir,Tarjan 94) of complexity O(ck m) and O(k5mn + Ck ). Both
Ck and ck are, of course, exponential in k.
Chapter 6
Lecture 6
Scribe : Rotem Erlich
Lecture 6 : May 26
6.1 Some Optimization Algorithms on Triangulated

Graphs
6.1.1 Computing the chromatic number and all maximal cliques
Theorem 6.1 (Fulkerson and Gross 65) Let G = (V E ) be a triangulated graph, and let
be a perfect elimination order for G. Every maximal clique in G is of the form: Xu fug
where Xu = fx 2 Adj (u)j
1(u) <
1(x)g.
; ;
4 8 7 5 3
1 6 2
Figure 6.1: An example for maximal cliques in a graph

Example: In the chordal graph in Figure 6.1, vertices are numbered according to p.e.o
and we have: X1 f1g = f1 6 7 8g X2 f2g = f2 5 6g X3 f3g = f3 5g X4 f4g =
67
f4 8g X5 f5g = f5 6 7g - maximal cliques, and: X6 f6g = f6 7 8g X7 f7g = f7 8g -
non maximal cliques.
Proof: Let C be an arbitrary maximal clique in G, and let v be the rst vertex according
to the order
which is contained in C . Let w 2 G
1(w) >
1(v). If w 2= Xv fvg, then
; ;
wv 2= E , therefore w 2= C , so we get: C Xv fvg.

On the other hand, since G is triangulated Xv fvg is a clique, and because C is maximal
we get: C = Xv fvg.
Corollary 6.2 A triangulated graph on n vertices has at most n maximal cliques, with
equality if and only if the graph has no edges.
Proof: Every maximal clique is characterized uniquely by the rst vertex in
which is
contained in it.
From Theorem 6.1, a naive algorithm for calculating the maximum clique in a triangu-
lated graph is as follows:
Find fvg Xv for each v.
Filter out the non-maximal cliques.
A more ecient algorithm is based on the algorithm TEST for testing whether a given
vertices permutation
is a PEO as shown in lecture no. 4. In that algorithm Xv is calculated
for each v.
Theorem 6.3 When algorithm TEST operates on a chordal graph and a p.e.o.
, fug Xu
is not a maximal clique i Xu is concatenated to A(u).
Proof: Reminder: In the phase of eliminating v (phase
1(v ) = i of TEST), Xv ; fug
;
is concatenated to A(u) if u is the rst neighbor of v after it in

.
(=)) If Xu fug is not a maximal clique then there exists a vertex w
1(w) <
1(u)
; ;
such that fug Xu fwg Xw . Let us take v such that v =

(maxf
1(w)j
1(w) <
; ;
1(u)g).
;
In phase
1(v) Xu = Xv ; fug is concatenated to A(u).
;
((=) Assume that at phase

1(v) Xu = Xv ; fug is concatenated to A(u). Then
;
Xv = Xu fug. So Xv fvg is a clique and Xu fug Xv fvg, therefore Xu fug is not

maximal.
See the modied algorithm in gure 6.2.
Theorem 6.4 The above algorithm correctly calculates the chromatic number and all max-
imal cliques of a triangulated graph G = (V E ) in O(jV j + jE j) time.
The proof is similar to that of the algorithm in lecture no. 4.
6.1. SOME OPTIMIZATION ALGORITHMS ON TRIANGULATED GRAPHS 69
1. 1
2. for all vertices v do S (v) 0
3. for i 1 to n do
begin
4. v
(i).
5. X fx 2Adj(v)j
1(v) <
1(x)g
; ;
6. if Adj(v) = then print fvg.

7. if X = then goto 13
8. u
(minf
1(x)jx 2 X g)
;
9. S (u) maxfS (u) jX j ; 1g

10. if S (v) < jX j then do
begin
11. print fvg X
12. =maxf 1 + jX jg
end
13. end
14. print "The chromatic number is ",.
Figure 6.2: Chromatic Number Calculation
6.1.2 Finding and k
Next we tackle the problem of nding the stability number (G) of a triangulated graph G.
Better yet, since G is perfect, the clique cover size k(G) = (G).
Let
be a perfect elimination scheme for G = (V E ). We dene inductively a sequence
of vertices y1 y2 : : : yt in the following manner:
(1) y1 =
(1)
(2) yi =
(minf
1(x)j
1(x) >
1(yi 1) and x 2= Xy1 Xy2 : : : Xyi 1 g)
; ; ;
yi is the rst vertex in

which appears after yi 1 in
, and which is not in Xy1 Xy2
; ;
Xyi 1 . All vertices following yt are in Xy1 Xy2 Xyt.

;
1 5 8 6
3 7 4
Figure 6.3: An example for nding maximal stable set

For example, in the graph in Figure 6.3:
y1 = 1 Xy1 = f3 5g
y2 = 2 Xy2 = f8g
y3 = 4 Xy3 = f6 7g
Here fy1 y2 y3g is a maximum stable set, and:
y1 Xy1 = f1 3 5g y2 Xy2 = f2 8g y3 Xy3 = f4 6 7g are a minimum clique cover.
Theorem 6.5 (Gavril, 1972) The set fy1 y2 : : : ytg is a maximum stable set of G, and the
collection of sets Yi = fyig Xyi i = 1 2 : : : t comprises a minimum clique cover of G.
Proof: The set y1 y2 : : : yt is stable since if yi yj 2 E for i < j , then yj 2 Xyi , which
cannot be. Thus (G) t.
On the other hand, since fY1 Y2 : : : Ytg is a set of cliques which together cover G, we
get k(G) t.
Since (G) k(G) we get: (G) = k(G) = t, and we have produced the desired
maximum stable set and minimum clique cover.
The method above, with a slight modication in the algorithm of section 6.1.1, nds a
maximum stable set and a minimum clique cover in time complexity of O(jV j + jE j).
6.2. COMPARABILITY GRAPHS 71
6.2 Comparability Graphs
6.2.1 Some Denitions
An undirected graph G = (V E ) is a comparability graph if there exists an orientation (V F )
of G satisfying
F \ F 1 = ,
;
F + F 1 = E
;
F 2 F where F 2 = facjab bc 2 F for some vertex bg

F is also called a transitive orientation of G, and G is also called transitively orientable
(TRO).
yes yes no
Figure 6.4: some comparability graphs

The relation F is a strict partial ordering (irreexive, transitive). Comparability graphs
are also known as partially orderable graphs.
Dene the binary relation ; on the edges of an undirected graph G = (V E ) as follows:
ab;a b () a = a and bb 2= E or b = b and aa 2= E .
0 0 0 0 0 0
We say that ab directly forces a b whenever ab;a b .

0 0 0 0
The reexive, transitive closure of ; is ; . It is easily shown that ; is 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 i there exists a sequence of edges
ab = a0b0;a1b1; : : : ;ak bk = cd k 0. Such a sequence is called a ;-chain from ab to cd,
and we say that ab (eventually) forces cd whenever ab; cd.
c d e
b
Figure 6.5: An example for implication classes
For example, the 8 implication classes in the graph of gure 6.5 are: A1 = fabg, A2 =
fcd1g, A3 = fac ad aeg, A4 = fbc bd beg, A1 1 = fbag, A2 1 = fdcg, A3 1 = fca da eag,
; ; ;
A4 = fcb db ebg.

;
Notice that ba;dc () ab;cd and ba; dc () ab; cd, by denition. Notice also that

ab;ab (by our convention that G contains no self-loops), but not ab;ba necessarily.
Let I (G) denote the collection of implication classes of G. We dene I^(G) = fA^jA 2
I (G)g, where A^ = A A 1 (the symmetric closure of A^): The members of I^(G) are called
;
the color classes of G.
6.2.2 Implication Classes

Theorem 6.6 Let A be an implication class of an undirected graph G. If G has a transitive
orientation F , then either F \ A^ = A or F \ A^ = A 1 and, in either case, A \ A^ = .
;
Proof: We dened ; in order to capture the fact that, for any transitive orientation F
of G, if ab;a b and ab 2 F , then a b 2 F . Applying this property repeatedly, we obtain
0 0 0 0
F \ A = or A F .
If F \ A = ) A F 1 ) A 1 F ) F \ A^ = A 1
; ; ;
If A F ) A 1 F 1 ) F \ A 1 = (since F \ F 1 = ) ) F \ A^ = A.
; ; ; ;
In both cases A \ A 1 = (since F \ F 1 = ):

; ;
We shall prove later that the converse of Theorem 6.6 is also valid, namely if A \ A 1 =
;
for every implication class A, then G has a transitive orientation.

Note that the existence of an arbitrary union of implication classes F = iAi satisfying
F \ F 1 = and F + F 1 = E does not necessarily mean that F is a transitive orientation
; ;
of G. For example, in a simple triangle abc, fabg fbcg and fcag are implication classes
satisfying the condition but the resulting orientation is cyclic.
Lemma 6.7 If ab; cd, then there exists a ;-chain from ab to cd of the form

ab = a0b0;a1b0;a1b1;a2b1 : : : ;ak bk = cd
Proof: In the ;-chain in the lemma the indices change for the ai -s in the odd ";"-s, and
for the bi-s in the even ones. Let us look at any ;-chain between ab and cd. If in that chain
there is a group of 3 or more sequential edges with the same index changed - in particular in
that group the rst and last edges directly force each other. We shall take for "our" chain
those rst and last edges - so we got only 2 edges in a row with the same index changed.
We shall do that for every such group of 3 or more edges in a row, and will get the desired
;-chain.
A ;-chain satisfying the lemma is called a canonical ;-chain.
6.2.3 The Triangle Lemma
Lemma 6.8 (The Triangle Lemma) Let A B and C be implication classes of an undi-
rected graph G = (V E ) with A =6 B and A =6 C 1 and having edges ab 2 C , ac 2 B and
;
bc 2 A
(i) If b c 2 A then ab 2 C and ac 2 B
0 0 0 0
(ii) If b c 2 A and a b 2 C , then a c 2 B

0 0 0 0 0 0
(iii) No edge in A touches the vertex a
a
a
b0 c0
b0 c0
2
a 1 1
b 1
b1 c 1 b1 c 1
2
a 2 b2
a k
bk ck
bk ck
B
(1) A (2)
C
Figure 6.6: The triangle lemma

Proof: If b c 2 A, by Lemma 6.7 there exists a canonical ;-chain
0 0
bc = b0c0;b1c0;b1c1 : : : ;bk ck = b c 0 0
(see Figure 6.6(1)). Let a = a0.

a0c0 2 B doesn't force b1c0 2 A ) a0b1 2 E .
a0b0 2 C b0b1 2= E (since b0c0;b1c0) ) a0b1 2 C .
a0b1 2 C b1c1 2 A ) a0c1 2 E (otherwise b1a0;b1c1, contradiction to C 1 6= A). ;
c0c1 2= E a0c0 2 B ) a0c1 2 B . By induction on the same argument we get: a0bk 2

C a0ck 2 B . This proves (i). In particular (i))(iii).
Next, assume b c 2 A a b 2 C . Consider a new canonical ;-chain
0 0 0 0
ab = a0b0;a1b0;a1b1; : : : albl = a b 0 0
(see Figure 6.6(2)). If ca1 2= E then ba1;bc but C 1 6= A therefore ca1 2 E .

;
aa1 2= E ac 2 B ) a1c 2 B .
The triangle a1bc has the same characteristics as the original triangle abc. By induction
we get that the triangle albc = a bc has also the same characteristics. From (i) (with: a
0 0
instead of a) we get a c 2 B .0 0
Theorem 6.9 Let A be an implication class of an undirected graph G = (V E ). Exactly

one of the following alternatives holds:
(i) A = A^ = A 1 ;
(ii) A \ A 1 = A and A 1 are transitive orientations of A^.

; ;
Proof: (i) Assume A \ A 1 6= . Let ab 2 A \ A 1. So ab; ba. According to the denition

; ;
of ; for any cd 2 A cd; ab i dc; ba. Since ; is an equivalence relation and ab; ba, cd; dc

and therefore dc 2 A. Thus A = A^.

(ii) Assume A \ A 1 = and let ab bc 2 A. We shall now prove ac 2 A. Now ac 2= E )
;
ab;cb ) cb 2 A ) bc 2 A 1, a contradiction. Thus ac 2 E . Let B be the implication class

;
of G containing ac, and suppose A 6= B .

Since A 6= A 1 and ab 2 A, the triangle lemma (iii) on triangle abc (with bc 2 A ac 2
;
B ab 2 C = A) gives us that there is no edge in A touching a.

But ab 2 C = A touches a. A contradiction.
Therefore ac 2 A ) A is transitive ) A 1 is transitive. ;
Finally, A is an implication class of A^, so by Theorem 6.6 A and A 1 are the only ;
transitive orientations of A^.

Corollary 6.10 Each color class of an undirected graph G either has exactly two transitive
orientations, one being the reversal of the other, or has no transitive orientation. If in G
there is a color class having no transitive orientation, then G is not a comparability graph.
6.2.4 Decomposition of Graphs

Let H0 be a graph with n vertices v1 v2 : : : vn and let H1 H2 : : : Hn be n disjoint graphs.
The composition graph H = H0H1 H2 : : : Hn] is formed as follows:
For all 1 i j n, replace vertex vi in H0 with the graph Hi and make each vertex Hi
adjacent to each vertex of Hj whenever vi is adjacent to vj in H0 .
Formally, for Hi = (Vi Ei) we dene H = (V E ) as follows:
V = i 1Vi

E = i 1 Ei fxyjx 2 Vi y 2 Vj and vivj 2 E0g

We may also denote E = E0E1 E2 : : : En ]

H0 is called the outer factor of H .
Hi i = 1 : : : n are called the inner factors of H .
Theorem 6.11 Let G = G0G1 G2 : : : Gn ] where the Gi -s are disjoint undirected graphs.
Then G is a comparability graph i each Gi (0 i n) is a comparability graph.
Proof:
(=)) It is obvious that if one of the Gi -s is not a TRO graph then G is not TRO.
((=) Let F0 F1 : : : Fn be transitive orientations of G0 G1 : : : Gn respectively. Let us
show that F = F0F1 F2 : : : Fn] is transitive: Let ab bc 2 F : if ab bc 2 Fi then ac 2 Fi by
transitivity. If ab 2 Fi bc 2 Fj j 6= i then we cannot have i 6= 0 j 6= 0 since inner factors
are disjoint. If ab 2 Fi bc 2 F0 then a 2 Gi so, since every vertex in Gi is connected to every
vertices in the factor in which c is contained in, then ac 2 F0. If ab 2 F0 bc 2 Fi then c 2 Gi
so ac 2 F0,]
and in both cases we get transitivity.
A graph is called decomposable if it can be expressed as a non-trivial composition of some
of its induced subgraphs.
(For each graph there is the trivial composition: G = k1G].)
Formally, G = (V E ) is decomposable if there exists a partition V = V1 + V2 + + Vr
of the vertices into nonempty pairwise disjoint subsets with 1 < r < jV j such that G =
GRGV 1 GV 2 : : : GV r ] for any set of representatives R = fx1 x2 : : : xr g xi 2 Vi .
Such a partition is said to induce a proper decomposition of G.
Chapter 7
Lecture 7
Scribe : Sariel Har-Peled
Lecture 7 : June 2
7.1 Uniquely Partially Orderable Graphs

Denition Given a graph H0 with
h n vertices labeled
i fv1 : : : vng and n disjoint graphs
H1 : : : Hn , we denote by H = H0 H1 H2 : : : Hn the composition graph, resulting from
replacing each vertex in H0, vi 1 i n by the graph Hi, and replacing all edges ij 2 E (H0 )
by the edges connecting all the vertices of Hi with the vertices of Hj . That is V (H ) = ni=1Vi
and E (H ) = ni=1 Ei f(v u)jv 2 Vi u 2 Vj and (i j ) 2 E0g.
The graph H0 is called the outer factor of H , and the graphs H1 : : : Hn are called the
inner factors of H .
Example 7.1 See gure 7.1 for an undirected example of a composition graph. See gure
7.2 for a directed example of a composition graph.
h i
Theorem 7.2 Given n + 1 undirected graphs G0 : : : Gn , let G = G0 G1 : : : Gn denote
their composition graph. Then, G is a comparability graph if and only if, all the graphs
Gi 0 i n are comparability graphs.
Proof: ((=) Clearly, if one of the Gi i = 0 : : : n is not a comparability graph, then G
can not be a comparability graph.
77
Figure 7.1: The composition graph of P3 K3 K1 C4]
F1 F2 F3 F4
Figure 7.2: The composition graph of F1F2 F3 F4]

7.1. UNIQUELY PARTIALLY ORDERABLE GRAPHS 79
(=)h ) Let F0 : :i: Fn be transitive orientations of the graphs G0 : : : Gn , let
F = F0 F1 : : : Fn be an orientation of G. Given two edges a ! b b ! c 2 F . We have to
show that a ! c 2 F . One of the following hold:
The vertices a b c 2 Vi 1 i n. Since Fi is a transitive orientation it follows that
a ! c 2 F.
The vertices a b 2 Vi c 2 Vj i =6 j . Then, it follows that i ! j 2 F0, but then
a ! c 2 F.
The vertices a 2 Vi b c 2 Vj i =6 j , then i ! j 2 F0, thus a ! c 2 F .
The vertices a 2 Vi b 2 Vj c 2 Vk , then i ! j j ! k 2 F0, since F0 is transitive it
follows i ! k 2 F0, thus a ! c 2 F .
Thus F is a transitive orientation of G.
Denition A graph G = (V E ) V = fv1 : : : vng is decomposable if it can be presented
as a non-trivial
h i composition
h of inducedi subgraphs of G. The trivial decompositions of G
are G = K1 G , G = G fv1g : : : fvng . A graph which has only trivial decompositions is
called prime.
If there exists ha partition of ithe vertex set of G: V = V1 + V2 + + Vr , 1 < r < jV j,
such that G = GR GV1 : : : GVr , where R is a set of r representatives from V1 : : : Vr , then
we say the the partition induces a proper decomposition of G.
Example 7.3 See gure 7.3 for an example of various decompositions of a graph.
Theorem 7.4 Let G0 : : : Gn be n disjoint, undirected graphs, and let G = G0G1 : : : Gn ]
be their composition. Let Ab be a color class of G, then exactly one of the following holds:
Ab Ej for a single value of 1 j n.
Ab \ Ej = for all 1 j n.
Proof: Any color-class in G is, by denition, a connected subgraph of G. Assume that

Ab \ Ej 6 for a vertex 1 j n. Look at an edge ab 2 Ab \ Ej , and let cd 2 E (G) be an
=
edge, such that cd;ab. There are three possibilities:
1. The edge cd 2 Ek k 6= j . This is not possible, because internal edges of distinct inner
factors do not share a vertex.
1 1
3 4 5 3 4 5
2 G 2
P[
2 , ]
1 1
3 4 5 3 4 5
2 2
P[
3 , , ]
P[
2 , ]
Figure 7.3: The decomposition of a graph
2. The edge cd is an external edge. This is not possible, because then the set C =
fa bg fc dg having three vertices, induce a triangle in G, contradicting the fact that
cd;ab.
Thus the only remaining possibility is that cd 2 Ej , and by the connectivity of Ab , it follows
that Ab Ej .
Denition Let G = (V E ) be an undirected graph. A set Y V is called a module or
a partitive set if for all x 2 V n Y either Y \ Adj (x) = or Y Adj (x). The set Y is a
non-trivial module if 1 < jY j < jV j.
Figure 7.4: Module in a graph
Example 7.5 See gure 7.4 for an example of a module in a graph.

Lemma 7.6 The graph G possess a non-trivial module Y V , if and only if G is decom-
posable.
Proof: ((=) Let G = G0G1 : : : Gn ] is a decomposition of G, with Gi = (Vi Ei ) a non-
trivial inner factor. If x 2 Gj j 6= i then either Vi adj (x) (if ij 2 E0) or vi \ Adj (x) =
(if ij 2= E0). Hence Gi is a module.
(=)) If Y is a non-trivial module of G, then the division -
V = fv1g + fv2g + + fvk g + Y
induce a proper decomposition of G.
Corollary 7.7 A set of vertices Y V in the graph G = (V E ) with 1 < jY j < jV j is a
module, if and only if there exists a proper decomposition of G = G0 G1 : : : Gn ], such that,
there is an inner factor of G, Gi 1 i n, such that Gi = GY .
Claim 7.8 If Y is a set of vertices spanned by a color-class Ab in the undirected graph
G = (V E ), then Y is a module.
Proof: If Y = V or jY j = 1, then the set Y is a trivial module of G.
Else, let a be any vertex in V n Y , and assume that b 2 Y \ Adj (a), then ab 2 E n Ab.
Since Y is dened by edges there exists a vertex c 2 Y such that bc 2 Ab. The edge ac 2 E
exists (else, ba 2 Ab). The edge ac 2= Ab (else, a 2 Y ).
Given any vertex d 2 Y , let de 2 Ab be an edge connected to d. Now, using the triangle
lemma, it follows that ad 2 E , thus Y Adj (a).
Claim 7.9 An undirected graph G = (V E ) has at most one color-class which spans all of
V.
Proof: Assume that A b Bb are two distinct color-classes that both span V . Given any
vertex b 2 V , there are edges ab 2 Bb bc 2 Ab. Since Ab 6= Bb the edge ac 2 E .
Let Cb be the color-class that contain the edge ac. If Cb 6= A b Cb 6= Bb , from the triangle
b
lemma (iii), it follows that no edge in A is incident to the vertex a. A contradiction.
Thus Cb = Ab 6= Bb , and by the same argument no edge from Bb is incident to the vertex c.
A contradiction.
Denition A comparability graph is uniquely partially orderable graph (abbreviated UPO
graph), if it has only two transitive orientations, where each orientation is the reversal of the
other.
Figure 7.5: Uniquely partially orderable graph
Example 7.10 See gure 7.5 for an example of a uniquely orderable graph
Corollary 7.11 A graph G is UPO i G have a single color-class.
Theorem 7.12 (Shevrin-Filippov '70) Let G be a connected comparability graph. The
following conditions are equivalent:
1. The graph G is UPO.
2. Every non-trivial module in G is an independent set.
3. In any proper decomposition of G, every inner-factor is an independent set (Explicitly,
all the edges in G are external).
Figure 7.6: A graph with a color class that induce all vertices
Proof: (2) () (3) Follow immediately from the natural equivalence between a module
and a inner factor in a decomposition, as stated in corollary 7.7.
If there is a non-trivial module Y , which is not stable, there is a decomposition of G,
with inner factor GY , which is not stable.
If GY is an inner factor in a given proper decomposition, then Y is a module, thus if GY
is not stable, then the set Y is a non-stable module.
(1) =) (3) If G is UPO, then in G there is a color-class, Ab, which contains all the
edges of G. Thus by theorem 7.4, it follows that all the edges are external in any non-trivial
decomposition of G (because no non-trivial inner factor can contain all the edges of Ab, thus
such an inner factor, has a n empty intersection with Ab = E (G)).
(2) =) (1) Assume, for the sake of contradiction, that G is not UPO. Then G has more
than one color-classes (Since G is a comparability graph). From claim 7.9 it follows that
there is a color-class which spans a proper subset Y of V . By claim 7.8 it follows that Y is
a module of G. But Y is not a stable set (Y is dened as these vertices that are spanned by
a color-class, which is connected). A contradiction.
Corollary 7.13 A prime comparability graph G is UPO.

Proof: If G is not decomposable, then it must be connected. Condition (3) in theorem
7.12 holds trivially thus G is UPO.
7.2 Characterizations and Recognition Algorithms
Denition For the undirected graph G = (V E ), a partition of E into k sets: E =
Bc1 + Bc2 + + Bck , is called a G-decomposition if Bi is an implication class of the graph
G = (V Bci + Bdi+1 + + B
ck ). See gure 7.7 for an example.
A sequence of edges x1y1 : : : xk yk ] is a decomposition scheme for G, if there exists a
G-decomposition E = Bc1 + Bc2 + + Bck , such that xiyi 2 Bci, i = 1 : : : k.
2 5
1 3
4 G B1 B2 B3
C1 C2 C3
Figure 7.7: A graph G-decomposition
Example 7.14 The following are examples for the decomposition schemes for gure 7.7:
2 3
35
6 24 13 7
For the top G-decomposition: 664 12 14 23
7
5.
7
43
h i
For the bottom G-decomposition: 12 13 43
Remark 7.15 A G-decomposition of a graph can have several decomposition-schemes, but
the decomposition scheme uniquely denes the G-decomposition. We will present algorithms
to calculate both.
See gure 7.8 for a description of the G-decomposition algorithm.
7.2. CHARACTERIZATIONS AND RECOGNITION ALGORITHMS 85
Algorithm: construct G-decomposition

Input: A graph G V E
= ( ).
G
Output: A -decomposition: A number k
k sets of edges of G: Bc1 : : : Bck .
begin
Let E1 E , i 1.
Loop:
Arbitrarily pick an edge ei = xiyi 2 Ei .
Enumerate the implication class Bi of Ei containing xiyi.
Let Ei+1 Ei n Bci.
if Ei+1 = then
Let k i
stop.
else
Increase i by 1
go to Loop
end if
end construct G-decomposition
Figure 7.8: Algorithm to construct G-decomposition

Theorem 7.16 (Golumbic 77) Let A be an implication class of an directed graph G =
(V E ). Let D be an implication class in E n Ab. One of the following holds:
1. D is an implication class in E , and A is an implication class in E n Dc.
2. D = B + C , where C B are implication classes in E , and G contain a triangle with

edge taken from each implication class: A B C .
Proof: In the graph remaining after the removal of Ab several implication classes of G
may had been united. Let D be an implication class in E n Ab. Assume that D is a union of
several implication classes of E , including B and C .
b b
A C A C
c B a c B a
Figure 7.9: A triangle forcing the merge of two implication classes

Since B C are two distinct implication classes in E , that were united in E n Ab, there
exist edges ab ac 2 E n Ab such that ab 2 C ac 2 B bc 2 Ab or ba 2 C ca 2 B bc 2 Ab (see
gure 7.9). Assume without limiting generality that the former holds, since the second case
is the inverse of the rst case.
If B = C 1 then ba 2 B . It is known that ac 2 B but bc 2= B . Thus B is not transitive,
;
and using theorem 6: it follows that C 1 = Bb , explicitly B = C . A contradiction.

;
Thus Bb \ Cb = and thus the triangle 4abc have three-colors.

Is it possible that D is a union of more than 2 implication classes? Assume that there is
another implication class F , distinct from B C , in E n Ab that is contained in D. From the
b F B as its colors (See
above discussion, it follows that there is a three color triangle with A
gure 7.10). Now, using the triangle lemma it follows that triangles containing edges from
both Ab Bb , have their third edge from the same color-class. Thus D = B + C .
For the other case, suppose D is a single implication class in E . If A is not an implication
class in E n D c, then there is a triangle with A and D as two of its edges. This triangle
exist because where removing D from E , the implication class A was merged with another
implication class. Let A1 denote the implication class in E ; D c that contains the third edge
of the triangle (see gure 7.11).
But when removing A from E , this triangle implies the merge of the implication classes
A1 D into a single implication class. This contradicts the fact that D is an implication class
in E n Ab. Thus A is an implication class in E n D c.
A F
A C
B B
b Cb must be equal.
Figure 7.10: The classes F
A C
c B a
Figure 7.11: Ab + Ac1 is an implication class in E n D

c.
Corollary 7.17 If A is an implication class, such that A = Ab, then there is no triangle with
three colors such that Ab is one of its colors.
We leave this proof as an exercise. This claim implies, that for any triangle containing
Ab, the two other edges in the triangle are colored by the same color.
Corollary 7.18 If A is an implication of G = (V E ) such that A = Ab then all the other
implication classes of G are implication classes in E n Ab.
Proof: By theorem 7.16 if E n Ab contains two implication classes from E that where
merged, then these two class with Ab induce a three colored triangle in G, which is impossible
by Corollary 7.17.
Theorem 7.19 (characterizations of comparability graphs) (Gilmore and Ho -
man 1964, Ghouilla-Houri 1962)
Let G = (V E ) be an undirected graph with G-decomposition, E = Bc1 + + Bck . The
following conditions are equivalent:
1. G is a comparability graph.
2. For all implication classes A of E , A \ A 1 = .
;
3. Bi \ Bi 1 = for i = 1 : : : k .
;
4. Any circuit v1v2 v2v3 : : : vq v1 in E (not necessarily simple), such that vq 1 v1 2= E ,
vq v2 2= E and vi 1vi+1 2= E i = 2 : : : q ; 1 has even length,
;
Dene a short chord or a triangular chord in a cycle to be a chord connecting two

vertices which are distance 2 along the cycle. Condition 4 is equivalent to requiring
that every odd circuit in G contains a short chord.
Proof: (1) =) (2) Proved in theorem 6.6.

(2) =) (3) By induction of k. For k = 1, B1 is an implication class in G and it is given
that B1 \ B1 1 = .
;
We assume the claim holds for all G-decomposition of graphs that include less than k
sets. Specically, the claim holds for E n Bci. Let D by an implication class of E n Bci. By
theorem 7.16 two cases are possible:
The class D is an implication class of E , then by the induction hypothesis it follows
D \ D 1 = .
;
D = B + C where B C are implication classes in E . Then, D \ D 1 = (B C ) T(B 1

; ;
C 1) = (B \ B 1) (C \ C 1) = , because B \ C = and B \ B 1 = and C \ C 1 =

; ; ; ; ;
by the induction hypothesis.

(3) =) (1) Let E = Bc1 + + Bck a G-decomposition of E such that Bi \ Bi 1 = . ;
By theorem 6:9 it follows that B1 is transitive. If k = 1 then we are done, else assume
correctness to all G-decomposition of less than k sets. It follows that F = B2 + + Bk is a
transitive orientation of E n Bc1. We have to prove that B1 + F is transitive orientation of G.
Let ab bc 2 B1 + F . If both edges belong to B1 or to F , then the edge ac exists and must
belong to the corresponding set by the fact that both B1 F are transitive. Thus, assume
that ab 2 B1 and bc 2 F . Since B1 \ B1 1 = , if ac 2= E then ab;cb thus cb 2 B1 thus
;
bc 2 Bc1 but bc 2 F by assumption, a contradiction. Thus ac 2 E .

Assume that ac 2= B1 + F , then ca 2 B1 + F , there are two possibilities :
If ca 2 B1, since ab 2 B1 then cb 2 B1, a contradiction.
If ca 2 F , since bc 2 F then ba 2 F , a contradiction.
Thus ac 2 B1 + F . In the same manner, if we assume that ab 2 F bc 2 B1 it follows that
ac 2 B1 + F . Thus B1 + + Bk is a transitive orientation of E .
(1) () (4) Assume that G is not a comparability graph. By the fact that (1) () (2)
it follows that there exists v1v2 2 A \ A 1 =;
6 . By the lemma ensuring the existence of a
canonical implication chain, there is a canonical implication chain such that -
v1v2;v3v2;v3v4;v5v4 ;vq vq 1;vq vq+1 = v2v1:
;
The length of the chain is odd (because of its special structure), and there is no short
chord between two vertices in cycle, because then vivi+1 6 ;vi+2vi+1.
For the other direction, assume that there is an odd cycle in the graph with no short
chords, then it holds -
v1v2;v2v3;v3v4 ;vq vq 1;vq v1;v2v1
;
is a canonical implication chain. Thus v1v2 2 A \ A 1 6= and G is not a comparability

;
graph.
Denition Given an undirected graph G = (V E ) we dene a graph G = (V E ) in the
0 0 0
following manner -
V = f(i j ) (j i)jij 2 E g
0
E = f(x y) $ (y z)jxz 2= E g

0
(Note that by our convention xx 2= E so (x y) $ (y x) for every xy 2 E .) An orientation

of G such that (x ! y y ! z) imply that xz 2 E is called quasi-transitive orientation of
0
G.
Example 7.20 See gure 7.12 for an example of a graph and the resulting quasi-transitive
graph. The gure omits some the edges of the form (ij ji).
Theorem 7.21 (Ghouilla-Houri, 1962) The following claims are equivalent:
1. The graph G is a comparability graph
2. The graph G have a quasi-transitive orientation.
3. The graph G is bipartite.
0
Proof: (1) =) (2) Trivial, since G has a transitive orientation and this orientation is also
quasi-transitive.
(2) =) (1) Let F be a quasi transitive of G, and let C be an odd cycle. The cycle C
must contain two adjacent edges in F : y ! z x ! y and then because of the quasi-transitive
property xz 2 E . Thus (1) follows from theorem 7.19 since we found a short chord.
(1) () (3) We will use GH characterization (Any odd cycle contains a short-chord).
The vertices of a cycle in G correspond to adjacent edges in G (with common vertex), that
0
induce a cycle in G with no short chords. Thus G is bipartite () G is a comparability

0
graph.
By the above theorem it is possible to check if a graph is a comparability graph, as
follows: Construct the graph G , and check if it is bipartite. The size of G is : jV j = 2jE j,
0 0 0
and -
12 23 34
25 56
3
G
1 2 4
54 45
G
6 5
65 52
43 32 21
Figure 7.12: A graph G and its quasi-transitive graph G (in G , edges (x y) $ (y x) are
0 0
not drawn)
X n o
jE j = jE j + 2
0
w x wv 2 E vx 2 E wx 2= E = O(jV j3)
v V
2
Thus the complexity of constructing the graph is O(jV j3). Check if G is bipartite can
0
be performed in linear time in the size of G , i.e. in O(jV j3).

0
One can obtain a transitive orientation of G by taking the orientation of edges whose
vertices belong to one part of the bipartite graph G . In particular, G is UPO i G is
0 0
connected and bipartite. We shell give more ecient algorithms for recognizing comparability
graph later in the course.
Chapter 8
Lecture 8
Lecturer : Martin Golumbic
Scribe : Shimon Shahar
Lecture 8 : June 5
8.1 Some interesting graph families characterized by

intersection
8.1.1 Introduction
The topic of this lecture is intersection graphs. We shall focus mainly on F-graphs and
tolerance graphs.
Denition Let S = fTiji 2 I g be a family of subsets of a set T. The intersection graph of
S is G = (S E ), where
S = fTigj 2I
E = f(i,j) jTi \ Tj 6= g.
It was proved (Marczewski 1945]) that if we allow S to be an arbitrary family, then
we obtain all the undirected graphs. Usually we are interested in families of sets with
specic topological properties. Moreover we should notice that some graphs were dened as
intersection graphs (like line graphs), and others were dened by their properties, and only
afterwards were proved to be an intersection graph (for example triangulated graphs, were
originally studied as Ck -free graphs (k > 3) and only later characterized as intersection of
subtrees).
93
8.1.2 Permutation graphs
Denition A matching diagram consists of n points, on each of two parallel lines, and n
straight line segments matching the points. The intersection graph of the line segments is
called a permutation graph (See Fig. 8.1). If the endpoints on the line are numbered 1,2,: : :n,
then the endpoints on the other side of the line are numbered by a permutation on f1,2: : :,ng
denoted
.
1 4
2 2
3 6
2 1 3
4 1
5 3 4 6 5
6 5
matching diagram Corresponding permutation graph
Figure 8.1: A permutation graph for the permutation 4,2,6,1,3,5]
Remark 8.1 It is easy to build the permutation graph directly from the permutation.
8i j i < j and
(i) >
(j )] , (i j ) 2 E .
Theorem 8.2 A permutation graph G = (V E ) is transitively orientable.
Proof: Let
be the permutation of f1 : : : ng generating G = (V E ), where V =
f1 2 : : : ng. For every (i j ) 2 E , dene F as follows :
(i j ) 2 F i i < j .
We shall prove that D = (V F ) is transitively orientable.
Let (i j ) 2 F (j k) 2 F , then by Remark 8.1, since (i j ) 2 E , then
1(j ) <
1(i), and
; ;
1(k) <
1(j ). Hence
1(k) <
1(i), and therefore (i k) 2 F .
; ; ; ;
which implies F is a transitive orientation (See g. 8.2).

8.8.1. SOME GRAPH FAMILIES CHARACTERIZED BY INTERSECTION 95
1 4
2 2
3 6
2 1 3
4 1
5 3 4 6 5
6 5
matching diagram Corresponding permutation graph
Figure 8.2: An orientation on a permutation graph
Denition If
is a permutation, we say that
is the opposite permutation, meaning that
(i) =
(n + 1 ; i). For example
= 5 6 7 2 3 8 1 9 4],
= 4 9 1 8 3 2 7 6 5].
Theorem 8.3 If G is a permutation graph generated by
, then G is also a permutation
graph, generated by
.
Proof: We shall prove that the graph H = (V F ) generated by
is G .
For every i,j2V, (i j ) 2 E i (i-j)(
1 (i) ;
1(j )) < 0. But since
1(i) <
1(j ) i
; ; ; ;
(i) >
1(j ) and (i,j)2F i (i-j)(
1(i) ;
1(j )) >0 , we obtain that (i,j)2E , (i,j)62F,
; 1 ; ; ;
so H=G .
Theorem 8.4 (Pnueli Lempel Even 1971) G is a permutation graph i both G and G are
transitively orientable.
Proof:
) Trivial from the previous two theorems.
( Wasn't proved in class.
1 4
2 2
4 3
3 6
4 1 1 6
2 5
5 5 Corresponding F-graph
6 3
Function Diagram
Figure 8.3: An F-graph
8.1.3 F -Graphs
Denition An F-diagram consists of n points, on each of two parallel lines, and n continuous
function curves ending in the points. The intersection graph of the functions segments is
called a function graph, or an F-graph (See Fig. 8.3). Note the requirement that each curve
segment is a function (See Fig. 8.4 for negative example).
\The air controllers strike"

Given two collections of cities lying on two parallel lines, and a collection of ight paths
between them, assign an altitude to each ight so that no crashes occur. It is easy to see
that if we take the intersection graph of the paths, we obtain an F-graph (assuming that
no ight is going backwards). We should assign dierent altitude only for those ights that
intersect. Therefore we can consider each altitude to be a color, and we should solve the
coloring problem for that F-graph (See Fig. 8.5). Note that if all ight paths are straight
lines, then the resulting graph is a permutation graph.
A composition of permutation diagram.
We can think of F-graph as if it was constructed from few partial orders (each of them is a
permutation on the vertices). Therefore we can construct the F-graph from the composition
of these partial orders. (See Fig. 8.6).
1 1
2 5
3 A 6
4 4
5 2
6 3
Not a function diagram -point A has two values
Figure 8.4: Not an F-graph,point A have two values of the function.

Theorem 8.5 (Golumbic Rotem Urrutia 1982)
Let G be an undirected graph. The following statements are equivalent.
(1) G is an F-graph.
(2) G is the intersection graph of a composition of permutation diagrams.
(3) G is transitively orientable.
Proof:
(2) ) (1) Trivial, each piecewise-linear path (connecting the point numbered i in all the orders),
is a continuous function. Therefore it is an F-Graph.
(1) ) (3) We shall dene H = (V F ) an orientation of G as follows : (i,j) 2F if the i-th path in
the diagram is completely below the j-th path. More formally :
(i,j)2F if 8 x fi(x) < fj (x) and (j,i) if 8 x fj (x) < fi(x), otherwise the edge is not in
F.
If (i j ) 2 E , than obviously the function fi (x) ; fj (x) changes sign at least once.
Therefore (i,j) 62 F , and vice versa. So F = E . We should now prove that F is
transitive orientation for G .
Milwakee Albany
Newark
Chicago
Philadelphia
Indianapolis Baltimore
Richmond
Nashville
Wilmington
Birmingham Charleston
Figure 8.5: The air controllers strike problem.

If (i j ) 2 F , and (j k) 2 F , then : 8xfi(x) < fj (x) < fk (x).
) (i k) 2 F .
(3) ) (2) Let F be a transitive orientation of G . We can choose l1 l2 : : : lk ] to be a realizer of
F. For each line order li we assign the points of its permutation in the corresponding
order on the i-th horizontal line (See g. 8.6). For i=1,2,: : :,n connect all points
corresponding to i by straight line segment. The resulting structure is a function
diagram.
Now, by the denition of the realizer, (i,j)2F i for every order lk i,j appear in the
same order (i.e. either for all k i <lk j or for all k j <lk i). Hence (i,j)62F i the paths
of i and j intersect. Hence G is an F-graph.
8.1.4 Tolerance graphs

Denition Let I = fI1 I2 : : : Ing be a collection of line segments, and let T = ft1 t2 : : : tng
a set of positive numbers. The tolerance graph G = (V E ) of (I,T) is dened as follows :
V = fv1 : : : vng
E = f(i j )jjIi \ Ij j min(ti tj )g (See Fig. 8.7)
Interval graphs as a subset of tolerance graphs.

The family of tolerance graphs is a generalization of the interval graphs. They are usually
used for more general scheduling problems, where we allow things to happen simultaneously.
Given an interval graph G = (V E ) whose realization is I = fI1 I2 : : : Ing, we can
construct an isomorphic tolerance graph in the following manner :
T = (V E ) the tolerance graph of (I X ):
dene X (i j ) = jIi \ Ij j, if Ii \ Ij 6=
C = minfX(i j )j(i j ) 2 E g Now let I=fI1 I2 : : : Ing and T=ft1 t2 : : : tng where
ti = C=2 for all i. We claim that G is the tolerance graph of (I,T). Two line segments
intersect in the original interval graph, i the intersection length is bigger than C/2, therefore
the edge sets of the two graphs coincide.
The following lemma will be used in the next theorem :
Denition G=(V,E) is an interval containment graph if there exists a set of intervals
fIugu V such that (u,v)2E, IuisinIv or vice versa].
2
Lemma 8.6 G is a permutation graph if and only if it is a interval containment graph.

4
7 7
5
6 4
1 7
5 3 1 5 7
1
4 5
2 2
3
2 6 4
6 6 Corresponding graph
2
3 3
1
Composition of permutations
Figure 8.6: The composition of permutation diagrams.
0 1 2 3 4 5 6 7 8 9 10 E D
Line realization
The intervals of vertices A-E appear in A B C

their alphabetic order, from top to bottom.
Corresponding graph.
Tolerances : { T(A) = 2, T(B) = 1, T(C) = 0,

T(D) = 0.5, T(E) = 1.5}
Notice : (A,E) is not an edge, because its
intersection is less than min(2,1.5) .
Figure 8.7: A tolerance graph.

Proof: Both sides wasn't proved in class.
Theorem 8.7 G is a tolerance graph with ti = jIij, if and only if G is a permutation graph.
Proof:
) The proof is based on a the previous lemma. So all we have to do is to prove that G
is an interval containment graph.
Let G be a tolerance graph with ti dened as above. If (i j ) 2 E then either Ij contains
Ii, or vice versa. Using the previous theorem, our theorem is proved.
( If G is a permutation graph, than it is a line containment graph, therefore we can
realize it with a tolerance graph with tolerances dened as above.
8.1.5 Bounded and unbounded tolerance graphs.

If for a vertex i ti > jIij, than we can enlarge ti as much as we want, without aecting the
graph. Any interval which intersects Ii will have a common intersection with Ii not longer
than ti, so in this case an edge (i,j) will be in the graph if and only if jIi \ Ij j tj . Vi will
be called a vertex with unbounded tolerance. If G is a tolerance graph, in which no vi is of
unbounded tolerance, we call G a bounded tolerance graph.
Theorem 8.8 Let G be a bounded tolerance graph, then is an F-graph.
Proof: Wasn't proved in class.
0 1 2 3 4 5 6 7 8 9 10 E D
Line realization
A B C
The intervals of vertices A-E appear in
alphabetic order, from top to bottom. Corresponding graph.
Tolerances : { T(A) = 2, T(B) = 1, T(C) = 0,

T(D) = 0.5, T(E) = 5}
Notice : (A,E) is not an edge, because its
intersection is less than T(A)
Figure 8.8: An unbounded tolerance graph.

Chapter 9
Lecture 9
Scribe : Muafaq Tannous
Lecture 9 : June 9
9.1 Comparability Graph Recognition

In lecture 6 we discussed characterization of comparability graphs, and we proved a theorem
- TRO theorem - that characterizes comparability graphs. We described an algorithm that
constructs G-decomposition. By combining the TRO theorem with the G-decomposition
construction we obtain an algorithm that recognizes comparability graphs and generates a
transitive orientation the graph if it is TRO. The algorithm is shown in gure 9.1
0. initialize i ; 1 Ei ; E F ;
1. arbitrarily pick an edge xiyi 2 Ei
2. enumerate the implication class Bi of Ei containing xiyi
if Bi \ Bi 1 = 0 then add Bi to F
;
else stop (G is not TRO)

3. Ei+1 ; Ei ; B^i
4. if Ei+1 = 0 then output F and stop
else i ; i + 1 go to 1
Figure 9.1: Code of The TRO Algorithm
103
b
a e
d c
Figure 9.2: Example 1
Example: Let us run the algorithm on the graphs shown in gures 9.2, 9.3 and 9.4. The
run of the algorithm on the rst graph will be as the following :
a
c d e

Initialize : E1 = E = fab ba ae ea affa ad da bccb ceec cffc eaaegF = . Af-
ter the rst iteration we have x1y2 = ab B1 = fab ad af ae cb cf cecdg F = B1 E2 = ,
so the graph is TRO.
The execution of the algorithm on the second graph will be the following :
d
e c
a b

Initialize :
E1 = fac ca ae ea ad daab ba cbbc ebbecd dc db bdg F = :
9.2. THE COMPLEXITY OF COMPARABILITY GRAPH RECOGNITION 105
After the rst iteration we have:
x1y1 = ac B1 = fac ab ad aeg F = fac ab ad aeg E2 = fdc cd ba ab be eb de edg:
After the second iteration
x2y2 = bc B2 = fbc bd ba beg E2 = fcdg F = fac ab ad ae bc bd ba begE2 = fcdg:
After the third iteration
x3y3 = dc B3 = fdcg F = fac ab ad ae bc bd ba be dcg E2 = :
Thus the graph is TRO.
The execution of the algorithm on the second graph will be the following :
Initialize : E1 = fab ba bc cb cd dc de ed eaaegF = . After the rst iteration we
have x1y1 = ab B1 = fab ae cb cd ed eababc dc deg B1 \ B1 1 6= , so the graph is not
;
TRO.
The edges we choose in step (1) in the algorithm determine which transitive orientation
of the set of transitive orientations of the graph will be the output of the algorithm. If we
run the algorithm several times, and in every time we choose dierent edges in step (1) we
may get dierent transitive orientations, but one can prove that the number of iterations of
the algorithm in every run will be the same.
9.2 The Complexity of Comparability Graph Recogni-

tion
9.2.1 Implementation
We shall show that the complexity of the above algorithm is O( jE j) time and O(jE j + jV j)
space, where is the maximum degree of a vertex, and we'll give the implementation of
this algorithm that gives such complexity. Let G(V E ) be an undirected graph with V =
fv1 v2 : : : vng. The algorithm uses a function called CLASS (i j ), which is dened as the
following : CLASS (i j ) = 0, if vivj 62 E CLASS (i j ) = k, if vivj has been assigned to
Bk CLASS (i j ) = ;k, if vj vi has been assigned to Bk or undened, if vivj has not been
assigned yet.
Let di be the out-degree of vi after directing the graph G.
The algorithm gets as an input graph G(V E ) such that V = fv1 v2 : : : vng and its
adjacency sets such that j 2 Adj (i) i vivj 2 E . The output of the algorithm :
1. a variable FLAG which tells if G is a comparability graph. FLAG = 0 i G is
comparability graph.
2. if FLAG = 0 then the transitive orientation that was found by the algorithm is the
union of all edges having positive CLASS .
The algorithm consists of two subroutines, the EXPLORE procedure which is shown in
gure 9.5, and the main procedure which is shown in gure 9.6.
The set of edges explored in iteration k is denoted Bk . In iteration k, the algorithm
chooses an edge e that has not been assigned to any Bi i < k, and start exploring from e.
If Bk \ Bk 1 6= then the value of the variable FLAG will be changed from 0 to 1.
;
9.2.2 Complexity Analysis

The input graph G is represented by adjacency lists. The adjacency sets are stored as linked
lists sorted in increasing order and every element of the list represents a vertex. Each element
in the linked list Adj (i) is described by four elds. For every edge vivj in G there are two
elements as shown in gure 9.7.
1. j
2. CLASS (i j )
3. pointer to CLASS (j i)
4. pointer to next element on Adj (i)
This data structure requires O(jV j + jE j) space. Let us compute the time complexity
of the algorithm: In executing the 'for' command in EXPLORE (i j ) we use two pointers:
one points to Adj (i) and the other points to Adj (j ). These two pointers will scan these two
sets in parallel looking for a vertex m which satises the condition of the 'for' command and
then the value of CLASS will be updated. The loop will take O(di + dj ). The complexity
of the second loop is similar because it is executed in the same way. The algorithm calls the
procedure EXPLORE twice for every edge, hence the time complexity of the algorithm is
X
C (di + dj ) 2 O( jE j)
vi vj E
2
In summary :
Theorem 9.1 Comparability graph recognition and nding a transitive orientation can be
done in O( jE j) time and O(jV j + jE j) space, where is the maximum degree of a vertex.
9.2. THE COMPLEXITY OF COMPARABILITY GRAPH RECOGNITION 107
procedure EXPLORE(i,j):
for each m 2 Adj (i) such that m 62 Adj (j ) or jCLASS (j m)j < k do
begin
if CLASS (i m) is undened then
begin
CLASS (i m) ; k CLASS (m i) ; ;k
EXPLORE(i m)
end
else
if CLASS (i m) = ;k then
begin
CLASS (i m) ; k FLAG ; 1
EXPLORE(i m)
end
end
for each m 2 Adj (j ) such that m 62 Adj (i) or jCLASS (i m)j < k do
begin
if CLASS (m j ) is undened then
begin
CLASS (m j ) ; k CLASS (j m) ; ;k
EXPLORE(m j )
end
else
if CLASS (m j ) = ;k then
begin
CLASS (m j ) ; k FLAG ; 1
EXPLORE(m j )
end
end
return
Figure 9.5: Procedure EXPLORE

begin
initialize : k ; 0 FLAG ; 0
for each edge vivj in E do
if CLASS (i j ) is undened then
begin
k ; k + 1
CLASS (i j ) ; k CLASS (j i) ; ;k
EXPLORE(i j )
end
end
Figure 9.6: The main procedure of the algorithm
ADJ(i) j k
CLASS(i,j) = k
ADJ(j) i -k
Figure 9.7: Two elements for each Edge-

9.3. COLORING AND OTHER PROBLEMS ON COMPARABILITY GRAPHS 109
As we dene the procedure EXPLORE it explores the edges in a depth-rst search, we
can change the algorithm to breadth-rst search by storing the edges in queue (instead of
stack), then it will run in the same time and space complexity.
There are several dierent algorithms that solve the TRO recognition problem. Until
recently, the time complexity of the best known algorithm was due to Spinrad and has
complexity O(n2:5), In 1994 McConnel and Spinrad found an algorithm that in O(m+n log n)
constructs an orientation F in the graph G such that if G is a comparability graph then F
is transitive. There is an algorithm that decide if F is transitive in O(n2:371 ).

9.3 Coloring and Other Problems on Comparability

Graphs
9.3.1 Comparability Graphs Are Perfect
On a directed and acyclic graph G(V F ), dene a strict partial ordering of vertices, namely,
x > y i there is nontrivial path in F from x to y. A height function h can be dened on
such graphs : h(v) = 0 if out-degree of v is 0, h(v) = 1+ maxfh(w)jvw 2 F g. An example of
h(v) is shown in gure 9.8. Here in subsequent gures we use Hasse diagram of the partially
ordered set, i.e. all edges are pointing upwards and transitive edges are omitted.
0
1 1
3 2

h(v) is equal to the number of vertices in the longest path from one of G roots to v. h(v)
can be computed in linear time using BFS. The function h gives a proper vertex coloring
of G, because for every v and u if vu 2 F then h(v) < h(u), but it is not necessarily a
minimum coloring. For a graph G which is undirected, for any acyclic orientation F , the
corresponding h will be a proper vertex coloring of G. We now prove that if G is transitive
then h is an optimal coloring.
Theorem 9.2 Every comparability graph is a perfect graph.
Proof: Let F be a transitive orientation of a comparability graph G(V E ). Every path
in F corresponds to a clique of G, because of transitivity, and every clique corresponds to
a path, hence the function h is a proper vertex coloring of G that uses exactly !(G) colors,
this is an optimal vertex coloring because !. Because of the hereditary property of
comparability graphs we nd that !(GA ) = (GA ) for all induced subgraph GA of G, so G
is a perfect graph.
Let (X ) be a partially ordered set. A subset of X that is completely ordered is called
chain. A subset in X that doesn't contain two members comparable by is called antichain.
The sets fc e ag fd g bg ff bg are chains in the partial order shown in gure 9.9, the
sets fc dg fe f gg are antichains in the same partial order.
a e b g
c f d

A Hasse diagram is a compact representation of a transitive order. For a partially ordered
set (V F ), the Hasse diagram is (V H ) where H F and uv 2 H i uv 2 F and there is
no m 2 V s.t. um 2 V mv 2 V . The edges of Hasse diagram is drawn without the direction
arrow, with the convention that all edges are pointing up. For example the Hasse diagram
of the graph in gure 9.9 is in gure 9.10.
a e b g
c f d
Theorem 9.3 (Dilworth 1950) Let (X ) be a partially ordered set, the minimum number
of chains needed to partition (cover) X is equal to the maximum cardinality of a antichain
in X .
Proof: (X ) corresponds to comparability graph G which is a perfect graph, hence
k(G) = (G). k is equal to the minimum number of chains needed to partition X , because
9.3. COLORING AND OTHER PROBLEMS ON COMPARABILITY GRAPHS 111
a vertex set S induces a clique in G i there is a path in X passing through all vertices in S .
is equal to the maximum cardinality of an antichain. So, the minimum number of chains
needed to partition X is equal to the maximum cardinality of antichain.
For example consider the partial order that is shown in gure 9.9. The number of chains
which partition the partial order is 3, also, the maximum cardinality of antichains is 3.
Corollary 9.4 In a comparability graph with transitive orientation F , optimal proper col-
oring and maximum clique can be calculated in O(jE j + jV j) time.
We shall illustrate this by solving a slightly more general problem.
9.3.2 Maximum Weighted Clique

For a given undirected graph G and an assignment of a weights w(v) to each vertex, the
problem is to nd a clique in G for which the sum of the weights of its vertices is largest.
The algorithm calculates for every vertex v the value W (v) which is the sum of weights of
every vertex in the heaviest path from v to one of the graph sinks. A pointer is assigned to
v designating its successor on that heaviest path.
Input : a transitive orientation F of a comparability graph G(V E ) and a weight function
w dened on V .
Output : a clique K of G whose weight is maximum.
The algorithm uses two subroutine, the EXPLORE procedure which is shown in gure
9.11, and the main procedure which is shown in gure 9.12.
Proving the correctness of the algorithm and displaying an implementation whose com-
plexity is linear are left as exercises.
After running the algorithm on the graph that its Hasse diagram is shown in gure 9.13
we will have W (a) = 5 W (b) = 6 W (c) = 12 W (d) = 2 W (e) = 11 W (f ) = 4 W (g) =
14 POINTER(a) = null POINTER(b) = a POINTER(C ) = b POINTER(d) = null
POINTER(e) = f POINTER(F ) = null POINTER(g) = e and the maximum weight
clique is f e g.
9.3.3 Calculating (G)

We shall show a polynomial-time method for nding (G) for a comparability graph G. The
method gets as an input a graph G(V E ) and a transitive orientation F . We transform F
into a transportation network by adding two new vertices s and t and edges sx and yt for
each source x and sink y. Then we assign a lower bound 1 on the ow through each vertex.
We compute a minimum-ow, which is a polynomial problem. The value of such ow will
procedure EXPLORE(v):
if Adj (v) = then
W (v) = w(v)
POINTER(v) ; ^
return
else for all x 2 Adj (v) do
if x is unexplored then
EXPLORE(x)
end for all
select y 2 Adj (v) such that W (y) = maxfW (x)jx 2 Adj (v)g
W (v) ; w(v) + W (y)
POINTER(v) ; y
return
end.
Figure 9.11: Procedure EXPLORE
procedure MAXWEIGHT CLIQUE(V F ):

for all v 2 V do
if v is unexplored then
EXPLORE(v)
end for all
select y 2 V such that W (y) = maxfW (v)jv 2 V g
K ; y
y ; POINTER(y)
while y 6= ^ do
K ; K fyg
y ; POINTER(y)
return
end.
Figure 9.12: The main procedure of the algorithm

9.4. THE DIMENSION OF PARTIAL ORDERS 113
c w=1 g w=3
b w=6 w=7 f
e w=4
a w=5 d w=2
equal to k(G), the size of the smallest covering of the vertices by cliques. This value is
equal to (G), the maximum cardinality independent set since every comparability graph is
perfect.
9.4 The Dimension of Partial Orders

Every partial order (X P ) can be extended to a linear order (also called its topological order).
Such an order is called a linear extension of P . Let L(P ) be the set of linear extensions of
P . A subset L L(P ) that satises Tl L l = P is called a realizer of P , and its size is jLj.
2
Example: The orders L1 = (a b c d e) L2 = (a c b e d) L3 = (a b c e d) are linear
extensions of the partial order in gure 9.14. fL1 L2g is realizer of this partial order.
d e
b c

A realizer of partial order P with the smallest size is called a minimum realizer. The
dimension of partial order P , dim(P ) is the size of a minimum realizer of P .
The dimension of the partial order in gure 9.14 is 2.
Remark 9.5 P is a complete order i dim(P ) = 1.
f f’
c d e c’ d’ e’
b b’
a
a’
P P’
x
Example: The orders

L1 = (a x b c d e f a b e d c f ) L2 = (a x b c d e f a b e d c f )
0 0 0 0 0 0 0 0 0 0 0 0
are linear extensions of the partial order P in gure 9.15, fL1 L2g is the realizer of P ,
dim(P ) = 2. And
L1 = (b e d c f ) L2 = (b c d e f )
0 0 0 0 0 0 0 0 0 0 0 0
are linear extensions of P in gure 9.15, fL1 L2g is the realizer of P , dim(P ) = 2.
0 0 0 0 0
Example: The orders L1 = (a b c c b a ) L2 = (b c a a c b ) L3 = (c a b b a c )

0 0 0 0 0 0 0 0 0
are linear extensions of the partial order P in gure 9.16, fL1 L2 L3g is realizer of P , and
dim(P ) = 3. To Show this, let us assume that dim(P ) = 2, then there are two linear
extensions l1 l2, that in one of them, say l1, the elements (a b c) will be ordered: a b c.
Then in l2 these three elements must be ordered: c b a . l1 must contain the subsequent
a b c b a and l2 must contain the subsequent c b a b so we see that both extensions have
0 0 0
b b , a contradiction.
0
Lemma 9.6 Let (X P ) be partial order. For every subset Y of X , Y X we have
dim(PY ) dim(PX )
Proof: Let L be a realizer of P . Restricting L to the elements of Y yields a realizer of
Py . L is a minimum realizer for P so dim(PY ) dim(P )
Theorem 9.7 (Hiraguchi 1951) Let P = P0P1 P2 : : : Pk ] be a composition of a disjoint
partial orders (Xi Pi) i = 0 1 : : : k then dim(P ) = maxfdim(Pi)j0 i k g
c’ b’ a’
a b c

Proof: Let m = maxfdim(Pi)j0 i kg and fLi1 Li2 : : : Lim g be a realizer of Pi .
Dene ^i = Loj L1j L2j : : : Lkj ], then f^j jj = 1 2 : : : mg is a realizer of P , so dim(P )
m, every Pi is a subset of P , so dim(P ) maxfdim(Pi) 0 i kg ;! dim(P ) =
maxfdim(Pi) 0 i kg
For an example see gure 9.17.
a b 2 3 4
P1 P2
P0
1 5 6
L01 ={ab} L11 = {1,2,3} L21 ={5,6,4}
L02 ={ba} L12 = {1,3,2} L22 ={6,5,4}
2 3 4
P 1 5 6
L1 = 1,2,3,5,6,4
L2 = 6,5,4,1,3,2
Theorem 9.8 (Dushnik-Miller, 1941) Let G be a comparability graph of poset P dim(P )

2 i G the complement graph of G, is a comparability graph.
Proof:
) : Let F be a transitive orientation of G. Consider the graph G (V E ) when E = P + F . 0 0 0
This graph satises :

1. for every a b 2 V either ab 2 E or ba 2 E (but not both)
0 0
2. G is transitive
0
To prove 2, note that if ab bc ca 2 E , it is obvious that ab bc ca cannot be in the
0
same set (all in F or all in P ), because F P are transitive. So, two of these three edges
are in the same set (F or P ), say bc ab ab 2 F and by transitivity of F , we see that
ac 2 F so ca 62 P . (1) and (2) imply that G is a complete order on V . Because of the
0
same reasons we know that G (V E ) when E = P + F 1 is a complete order on V .

00 00 00 ;
L = fP + F P + F 1g is a realizer of P because if ab 2 P then ab 2 P + F P + F 1,

; ;
and if ab 62 P then either ab 2 F and ab 62 F 1, or ab 2 F 1 and ab 62 F . Hence

; ;
dim(P ) <= 2.
( : Assume that dim(P ) 2. Let L = fl1 l2g be a realizer of P . Dene F = l1 ; P =
(l2 ; P 1) (the equality holds because L is a realizer. To calculate l1 ; P we look at
;
l1 as a complete directed graph and remove all the edges from P , see gure 9.18)
f
f
c d e c d e
P b b
y x a g y x a y
l =gaxybcdef F = li - P
L = { l , l‘}
l‘ = y x a g b e d c f

We will show that F is a transitive orientation of G, assume that ab bc 2 F and ac 62 F .
Then ac 2 P . ab bc 2 F ;! ba cb 2 l2 ;! ca 2 l2 ;! ca 2 P contradiction , hence F is
transitive.
Corollary 9.9 Let G be a comparability graph of poset P . dim(P ) 2 i G is a permutation
graph.
Proof: We will give two proofs :
1. dim(P ) 2 from the previous theorem ! G is a comparability graph. From the
theorem of Pnuely-Even-Lempel (1971), G G are comparability graphs ! G is a
permutation graph.
2. let us look at the matching diagram of permutation graph(see gure 9.19).
1 4
1 2 3 4 5
2 1
3 3 2 5 3 1 4
1 5
4 5
5 2 2 3 4

(i j ) 2 E ! 1(i) > 1(j ) i < j ].Hence, the complete orders l1 = (n n ; 1 : : : 1)
; ;
and l2 = ( 1(1) 1(2) : : : 1(n)) is a realizer of P , so dim(P ) 2. Conversely, from the

; ; ;
realizer fl1 l2g we can construct the matching diagram in the same way.
The theorem of Dushnik-Miller tells that for two partial orders which have the same
comparability graph, the dimension of both 2 or the dimension of both 2.
a b
a
a c
c
d e e e
d d
b G b
P P‘
c

Theorem 9.10 (Trotter, Moore, Sumner 1976, Gysin 1977) Two partial orders
which have the same comparability graph have the same dimension.
Figure 9.20 gives an example of two partial orders that have the same comparability
graph.
The problem of calculating the dimension of partial order mentioned as a major open
problem in the Garey and Johnson book. In 1982, Yannakakis proved that the problem is
NPC. In addition he proved deciding if dim(P ) k, is NPC, for every k 3.
<
Chapter 10
Lecture 10
Scribe : Michael Seltser
Lecture 10 : June 16
10.1 Comparability invariants

In lecture # 9 we quoted the following theorem:
Theorem 10.1 ( Trotter, Moore, Sumner 1976], Gysin 1977] ) If two partial
orders have the same comparability graph then their dimensions are same.
Hence, the dimensions of all posets which have the same comparability graph is the same.
This leads to the following denition:
Denition Let be a property of partially ordered sets (posets). Then for poset P
(P) := 1 if holds on P and (P ) := 0 otherwise. Property is called comparability
invariant (CI) if for every two posets P P G(P ) = G(P ) implies (P ) = (P ).
0 0 0
Some of the known comparability invariants are: the dimension, length of maximum
chain (because it is !(G)), minimum number of chains ( or anti-chains ) needed for covering
of P (because they are k(G) (G) in correspondence), and the number of linear extensions
of P .
In 1991 Brightwell and Winkler showed that the problem of counting the number of linear
extensions of a poset is #P - complete.
Denition Let P = (V E ) be poset, L = (V E ) is linear extension. Without loss of
0
generality order for E is v1 < v2 < : : : < vn: The jump number (JP) of L is the number of
0
119
pairs (vi vi+1) such that fvi vi+1g 62 E .
More generally:
J (P L) =j fi j (L 1(i) L 1(i + 1)) 62 E g j
; ;
The jump number of P is the least jump number over all linear extensions of P i.e.
J (P ) = minfJ (P L) j L is a linear extension of P g:
The jump number is known to be a comparability invariant. In 1991 Syslo proved that
calculating J (P ) is NP - complete.
Theorem 10.2 ( Habib, 1982] ) Property is comparability invariant i the following
is true:
(1) (P ) = (P 1) 8 poset P .
;
(2) Let P = P0 P1 : : : Pm ] Q = Q0Q1 : : : Qm ] be proper decompositions and

G(Pi ) = G(Qi ) 0 i m: If (Pi) = (Qi) 8i then (P ) = (Q).
Denition Let P1 = ( V1, <1 ), P2 = ( V2 <2 ) be posets where V1 V2 are disjoint. Then
we'll dene a parallel composition of P1 P2 by P1 + P2 = ( V1 + V2 <+ ) where
x <+ y , ((x y 2 V1 x <1 y) _ (x y 2 V2 x <2 y)):
The serial composition of P1 P2 is P1 P2 = (V1 + V2 < ) where

x < y , ((x y 2 V1 x <1 y) _ (x y 2 V2 x <2 y) _ (x 2 v1 y 2 v2)):

Thus, P1 + P2 2K1 P1 P2], P1 P2 P1 P1 P2]:

Denition A poset P = (V , <) is called series parallel (SP) if either j V j= 1 or one can
derive it from any series parallel posets P1 P2 by parallel or series compositions.
For examples see Figure 10.1.
Denition A graph is called cograph if it is P4 - free.
Theorem 10.3 G is a cograph i for every induced subgraph H of G either H or H is
disconnected.
Proof: Exercise.
Theorem 10.4 Every cograph is a permutation graph.
Theorem 10.5 A is a cograph i A is the comparability graph of a series-parallel poset.
Corollary 10.6 The property "being a series-parallel poset" is a comparability invariant.
10.2. INTERVAL GRAPHS 121
a) b) c)
d) e)
Figure 10.1: a) not SP poset b) - e) SP posets
10.2 Interval graphs

Denition graph G = (V E ) is called an interval graph if oneTcan assign to each vertex
v 2 V an interval on the real line Iv such that (v1 v2) 2 E , Iv1 Iv2 6= :
Interval graph were introduced by Hajos(1957) andBenzer(1958).
a) b) c)
d) e)
Figure 10.2: a) interval graph b) it's interval representation c) chordal, but not interval
graph
For examples see Figure 10.2.
Theorem 10.7 (Gilmore - Ho man 1964]) The following are equivalent:
(1) G is an interval graph.
(2) G contains no induced C4 and G is a comparability graph.
(3) It is possible to order the maximal cliques in G so that for every vertex v 2 V the
cliques that contain v appear consecutively in this order.
Proof:
(1) )(2) Immediate, because it is easy to see that an interval graph G contains no induced C4
and a transitive orientation for G is obtained by taking an interval realization of G,
fIv gv V and setting (u v) 2 E (G ) , Iu is completely to the left of Iv .
2
(2) )(3) Let F be a transitive orientation for G .

Lemma: Let A1 A2 be distinct non-empty maximal cliques of G, and let (V F ) be a
transitive orientation of G . Then 1) there exists an arc from F between A1 and A2
2) all such arcs between A1 and A2 have the same direction.
Proof:
(1) It is trivial because otherwise B := A1 S A2 is a clique in G and j B j>j A1 j j A2 j,
in contradiction to the denition of A1 A2.
(2) Assume that a c 2 A1 b d 2 A2 (a b) (d c) 2 F . For example see Figure 10.3.
If a c (or b d) then we get (d,b) 2 F , a contradiction to A2 being a clique
in G. Therefore the vertices are dierent. The situation fa dg 2 E and fb cg 2
E is impossible, because otherwise G contains an induced C4. Without loss of
generality assume that fadg 62 E . Then we have two possible cases now:
(i) (a d) 2 F .Then the transitivity of F implies (a c) 2 F .
(ii) (d a) 2 F . Transitivity of F implies (d b) 2 F .
In both cases we get a contradiction.
A1 a A2
b
c d
Figure 10.3: example for lemma
Interval graphs 123
We now return to the theorem. Dene a relation among maximal cliques in G as

follows: for max.cliques A1 and A2, A1 < A2 , there is exists arc from F whose tail is
in A1 and it's head is in A2. According to lemma, this relation denes a tournament on
the set of maximal cliques of G. We'll prove that it is transitive and therefore denes
a complete order on G. Assume A1 < A2 < A3. Then 9 (w x) 2 F : w 2 A1 x 2 A2
and 9 (y z) 2 F : y 2 A2 z 2 A3: For example see Figure 10.4.
A1
A2 A
x 3
w z
y
Figure 10.4: example for theorem of Gilmore-Homan
( ) if fx zg 62 E then the lemma implies that (x z) 2 F and because F is a transitive
orientation for G we get (w z) 2 F and, hence, A1 < A2.
( ) if fw yg 62 E then analogously we get (w y) 2 F and A1 < A2.
( ) fx zg fw yg 2 E . G contains no C4 and therefore fw zg 62 E . It follows that,
since F is transitive, (w z) 2 F . Hence, A1 < A3.
We have shown that < is complete order (linear) on the maximal cliques of G.
Let x 2 V . Suppose that there exist Ai < Aj < Ak such that x 2 Ai Ak and
x 62 Aj :
(x 62 Aj ) ) 9y 2 Aj : fx yg 62 E
(Ai < Aj ) ) (x y) 2 F
(Aj < Ak) ) (y x) 2 F
We have a contradiction. Therefore the claim is true.
(3) )(1) Let I (x) be the set of maximal cliques of G that contain x. By (3) each such set is
an interval
T in complete order (C <) of maximal cliques. 8x y 2 V fx yg 2 E ,
I (x) I (y) 6= , because two vertices are neighbors i they are contained together in
some maximal clique.
Corollary 10.8 An undirected G is an interval graph i G is chordal and G is comparability.

Proof: If G is interval then it is chordal by the theorem of Hajos (1958) (lecture # 1).The
rest holds according to the last theorem. Conversely, if G is chordal then it contains no C4
and therefore if G is a comparability graph then G is interval.
Reminder: The clique matrix of graph is a (0,1) matrix M such that its rows correspond
to the maximal cliques of graph, its columns to the vertices and Mij = 1 i vj belongs to
the i-th maximal clique. Theorem 10.4 implies that if rows are ordered in the order < then
the ones in every column will appear contiguously. This property is called the consecutive
ones property for columns. For example see Figure 10.5.
Theorem 10.9 (Fulkerson-Gross 1965]) An undirected graph G is interval i its clique
matrix has the consecutive ones property.
Proof: Immediately follows from the third characterization of Theorem 10.4.
a) 2 4
6 7
1
5
3
1 2 3 4 5 6 7
{1, 2, 3} 1 1 1 0 0 0 0
0 1 1 1 1 0 0
b) {2, 3, 4, 5}
{3, 4, 5, 6} 0 0 1 1 1 1 0
{6, 7} 0 0 0 0 0 1 1
Figure 10.5: a) interval graph b) it's clique matrix.
Denition In an undirected graph G a triplet of vertices < x y z > is called an asteroidal

triplet if G<x y z> is stable and for each pair from fx y zg there exists path between these
two vertices which does not pass through neighbors of the third vertex.
Interval graphs 125
Theorem 10.10 (Lekkerkerker-Boland 1962]) G is an interval graph i it is chordal
and asteroidal triplet-free.
Proof: If G interval then it is chordal. Consider an interval realization of G: Assume that
< z y x > is asteroidal triplet. Without loss of generalization suppose that the left endpoint
of Iz is smallest among left endpoints of the intervals of the triplet in this realization and
right endpoint of Ix is biggest among right endpoints of this intervals. Then there exists
sequence of intervals z =Tv1 : : : vk = x such that 8 1 i k ; 1 Ivi T Ivi+1 6= . On other
hand, 8 2 j k ; 1 Iy Ivi = . It is contradiction. Therefore G is asteroidal triplet-free.
For example see Figure 10.6. Proof of other direction is omitted here.
I I
z x
? Iy
Figure 10.6: Example for Lekkerkerker - Boland theorem

Chapter 11
Lecture 11
Scribe : Vered Zilkha
11.1 Preference and Indierence

Let V be a set of possibilities. A decision maker denes a preference relation on this set:
for every pair of distinct members of V , he either prefers one over the other, or he feels
indierent about them.
We construct an oriented graph H = (V P ), and an undirected graph G = (V E ) as
follows:
8x y 2 V
xy 2 P , x is preferred over y
xy 2 E , indierence is felt between x and y
By denition, (V P + P 1 + E ) is complete.
;
We assume that the preference relation is acyclic. In fact we would want it to be transitive.
Thus we require that P be a partial order.
One way of formalizing preference relations (in utility theory) is the notion of Semiorder.
Denition Given > 0, a real-valued function u : V ! R is called a utility function of
semiorder if 8x y 2 V , xy 2 P , u(x) u(y ) +
Remark: If P has a utility function of semiorder then P is a partial order (irreex-
ive,transitive).
127
Let us now characterize the partial orders that have a utility function of semiorder. We
rst state without proof a fundamental result from utility theory:
Theorem 11.1 ( Scott - Suppes 1958] ) A binary relation (V P ) has a utility function
of semiorder i 8x y z w 2 V
(S1) P is irreexive
(S2) If xy 2 P , zw 2 P then either xw 2 P or zy 2 P
(S3) If xu 2 P , yz 2 P then either xw 2 P or wz 2 P
Denition An interval order is a binary relation (V P ) such that 8i 2 V ,we can match an
interval Ii on the real line, such that ij 2 P , Ii < Ij (i.e Ii \ Ij = and Ii is on the left
side of Ij )
Claim 11.2 A partial order (V P ) is an interval order i it satises (S1),(S2).
(The proof of this claim is left as an exercise)
The claim implies: If (V P ) is an oriented graph, (V E ) is an undirected graph, and
P + P 1 + E is the edge set of the complete graph on V then if (V P ) is a semiorder then
;
(V E ) is an interval graph.

(The converse is not true. For example see gure 11.1).
a b c a b c
d d
G G
a b c
d
Figure 11.1: Example
Denition An interval graph G = (V E ) is called a proper interval graph if it has a
realization fIv gv V such that no interval properly contains another.
2
11.1. PREFERENCE AND INDIFFERENCE 129
Theorem 11.3 ( Roberts 1969] ) Let G = (V E ) be an undirected graph. The following
statements are equivalent:
(1) There exists a real valued function u : V ! R such that 8xy 2 V , x 6= y , xy 2 E ,j
u(x) ; u(y) j< 1
(2) There exists a semiorder (V P ) such that E = P + P 1 ;
(3) G is a comparability graph and every transitive orientation of G = (V E ) is a semiorder

(4) G is an interval graph that doesn't contain an induced K1 3
(5) G is a proper interval graph
(6) G is a unit interval graph
Proof:
(1) ) (6) Let u be a function that satises (1). We will match to each vertex x 2 V an open
interval Ix = (u(x) ; 21 u(x) + 12 ). Then 8x y x 6= y, Ix \ Iy 6= ,j u(x) ; u(y) j<
1 , xy 2 E . Therefore fIxgx V is a unit-interval realization of G.
2
(6) ) (5) In a unit-interval realization of G, no interval can properly contain another, because
they all have length 1. Therefore this realization will be proper.
(5) ) (4) Let fIxgx V be a proper interval realization of G. Suppose G contains an induced sub-
graph G y z1 z2 z3 isomorphic to K1 3 where fz1 z2 z3g is a stable set and y is adjacent
2
f g
to each zi. Without lost of generality, suppose that Iz1 < Iz2 < Iz3 . Then Iy must
properly contain Iz2 , a contradiction. Thus, G can have no induced copy of K1 3.
(4) ) (3) Since G is an interval graph, G = (V E ) is a comparability graph. Let F be a transitive
orientation of G . It is easy to show that F satises (S 1) (S 2). Now we will see that it
also satises (S 3). By Scott-Suppes Theorem this will imply F is a semiorder.
Assume to the contrary that xy 2 F , yz 2 F , xw 62 F , wz 62 F . By the transitivity
of F it follows that: wx 62 F , zw 62 F , wy 62 F , yw 62 F . But xz 2 F . Thus Gx y z w
induces a K1 3, a contradiction.
(3) ) (2) Immediate.
(2) ) (1) If (V P ) is a semiorder, then by denition there exist a real valued function u~ : V ! R
and a constant > 0 such that xy 2 P , u~(x) ; u~(y) . Denoting u(x) = u~(x) we get
xy 2 P , u(x) ; u(y) 1, and because P + P 1 = E we get that xy 2 E , xy 62 P
;
and xy 62 P 1 ,j u(x) ; u(y) j< 1.

;
Additional results on unit interval graphs:

Denition A (0,1)-valued matrix satises the consecutive 1's property for the columns if
its rows can be permuted such that on each column, the 1's will appear consecutively.
Denition A layout of a graph G = (V E ) is a one-to-one function (a permutation)
L : V ! f1 : : : j V jg. We will denote L - the linear order L 1(1) < L 1(2) < : : : < L 1(n).
; ; ;
Let G = (V E ) be an interval graph. Then other equivalent conditions to G being unit
interval are:
The clique matrix of G satises the consecutive 1's property for rows and for columns.
Let A (G) be the adjacency matrix of G when we add 1's on the diagonal. Then A (G)

satises the consecutive 1's property for columns. ( Roberts 1968 )

G is a triangulated graph and it doesn't contain as an induced subgraph any of the
graphs appearing in gure 11.2.
Figure 11.2: Wegner 1967
There exists a layout L of G such that for every vertex v, L(Adj (v)) = fL(v);kv L(v);
kv + 1 : : : L(v) ; 1 L(v) + 1 : : : L(v) + lv g ( Deng, Hell, Huang 1992 )
Some hard optimization problems involving unit interval graphs are:
Adding a minimum number of edges to a graph to get a unit interval graph is NP -
complete. ( Golumbic, Kaplan, Shamir 1993 )
Removing a minimum number of edges to get a unit interval graph is NP -complete.
( Golumbic, Kaplan, Shamir 1993 )
11.2. RECOGNIZING INTERVAL GRAPHS 131
11.2 Recognizing interval graphs
Theorem 11.4 G = (V E ) is an interval graph i it has a layout L that satises:
L(u) < L(v) < L(w) uw 2 E ] ) vw 2 E (11.1)
(The proof of this theorem is left as an exercise).
Theorem 11.5 Let G = (V E ) be a co-comparability graph, and let F be a transitive ori-
entation of G . We denote d(u) the out-degree of u in (V F ). If L is a layout of G such
that
d(u) > d(v) ) L(u) < L(v) (11.2)
then G is an interval graph i L satises condition 11.1
Proof:
) If condition 11.1 holds then according to Theorem 11.4 G is an interval graph.
( Assuming, to the contrary that L satises condition 11.2 but not condition 11.1, we
will prove that G contains an induced C4:
(a) if uv 2 F then by transitivity of F , d(u) > d(v), and thus L(u) < L(v).
(b) if L(u) < L(v) and uv 62 E then uv 2 F (else vu 2 F and from (a) we get
L(v) < L(u)).
From (b) and the transitivity of F it follows that:
L(u) < L(v) < L(w) uv 62 E vw 62 E ] ) uw 62 E (11.3)
L(u) < L(v) < L(w) uw 2 E uv 62 E ] ) vw 2 E (11.4)

Now, suppose L does not satisfy 11.1. Thus there exist u v w such that
L(u) < L(v) < L(w) uw 2 E vw 62 E (11.5)
From (b) it follows that vw 2 F , and from condition 11.3 it follows that uv 2 E .
Since L(u) < L(v), it follows that d(u) d(v). vw 2 F uw 62 F , thus there exists
y such that uy 2 F , vy 62 F . Moreover, yv 62 F (else uy yv 2 F ) uv 2 F , a
contraction), thus vy 2 E . consider that
If yw 2 F then uy yw 2 F ) uw 2 F , a contradiction.
If wy 2 F then vw wy 2 F ) vy 2 F , a contradiction.
Thus yw 2 E , and G u v w y induces a C4.
f g
Theorem 11.6 Let G = (V E ) be a graph with a layout L, and let us dene functions
l h : V ! f0 : : : j V jg:
l(u) = vmin
N u]
2
L(v)
8
>
>
>
>
>
max L(v) fv j L(v) < L(u) v 62 N (u)g 6=
< v N u]
h(u) = > L(v)<L(u)
62
>
>
>
>
: 0 otherwise
then L satises condition 11.1 i 8w 2 V h(w) < l(w)
l(v) 1 1 3 2 3
h(v) 0 0 2 0 2
Figure 11.3: Example of function l h
Proof:
) If L does not satisfy condition 11.1 then there exist u v w such that L(u) < L(v) <
L(w), and uw 2 E , but vw 62 E . Thus l(w) L(u) < L(v) h(w).
( If there exists w 2 V such that l(w) < h(w) then taking u = L 1(l(w)), v = L 1 (h(w))
; ;
gives us vertices u v w such that L(u) < L(v) < L(w), and uw 2 E , but vw 62 E , thus
contradicting condition 11.1.
11.2.1 A quadratic algorithm 1

We rst describe an elementary O(n2) algorithm for recognizing interval graph. This algo-
rithm is due to Ramalingam and Rangan 1990].
Let G = (V E ) be an undirected graph.
1. Find a transitive orientation of G , and call it F .
2. Compute a layout of G according to a decreasing order of out-degrees of vertices in
(V F ).
3. For each vertex, compute l(u) h(u).
G is an interval graph i 8w 2 V , h(w) < l(w).
correctness:
1. If G is an interval graph then G has a transitive orientation. From theorem 11.5, the
layout L computed in step 2 of the algorithm satises condition 11.1. From theorem
11.6 the condition h < l holds.
2. If G is not an interval graph then no layout L satises condition 11.1, and from theorem
11.6 the condition h < l fails.
complexity:
1. Finding a transitive orientation of G : O(n2) (Spinrad 1985]), O(n + m log n) (Mc-
Connell - Spinrad 1994). (We don't need to check that the orientation is transitive)
2. Sorting by out-degrees of vertices: O(m + n log n). Where m =j E (G ) j].
3. Computing l(u) for all vertices: O(m ).
Computing h(u) for all vertices: O(m).
We can see that the complexity does not exceed O(n2 ).
Remarks:
* We can produce a realization of an interval graph, using the layout L.(this follows from
the proof of theorem 11.4, which is constructive.
* The proof implies that every transitive orientation of G has a linear extension L (one
or more) that satises condition 11.1.
11.2.2 A Linear Algorithm

To check whether a graph G = (V E ) is an interval graph we can use the following algorithm:
1. Check whether G is triangulated
2. Produce all the maximal cliques in G, and then produce the clique matrix M of G
3. Check whether the matrix M satises the consecutive 1's property
The rst two stages can be implemented in a linear time: O(j V j + j E j). We will see
now that the third stage can be implemented in a linear time, by an algorithm of Booth and
Leuker 1975].
The data structure being used is called PQ - trees: a PQ-tree is a tree with a root, that
has vertices of two kinds: P and Q. The sons of a P -vertex are an unordered set. The sons
of a Q-vertex are an ordered set. The leaves of the tree are labeled by 1-1 correspondence
with the members of the set X . The frontier of the tree is the permutation of X , obtained
by reading the labels on the leaves from left to right.
Two PQ-trees, T , T are equivalent, denoted T T , if one can be obtained from the
0 0
other by applying a sequence of the following transformation rules:

Arbitrarily permute the children of a P -vertex.
Reverse the order of children of a Q-vertex.
Any frontier obtainable from a PQ-tree equivalent with T is said to be consistent with
T . the set of permutations on X consistent with T is denoted by:
CONSISTENT (T ) = fFRONTIER(T ) j T T g 0 0
J I A B
G H E D C
Figure 11.4: a PQ-tree. The FRONTIER here is F J I G H A B E D C]
The null tree T0 has no nodes and CONSISTENT (T0) = . The universal tree has one
internal P -vertex, the root, and a leaf for every member of X .
The relation between the consecutive arrangement problem and PQ-trees is as follows:
Let ' be a collection of subsets of a set X , and let )(') be the collection of all permu-
tations of X such that the members of each I 2 ' occur consecutively in .
Theorem 11.7 ( Booth - Leuker 1976] )
1. For every collection of subsets ' of X there exists a PQ-tree T such that )(') =
CONSISTENT (T ).
2. For every PQ-tree there exists a collection of subsets ' such that
)(') = CONSISTENT (T ).
The role of the Q-vertices is to restrict the number of permutations by making some of
the brother relationships rigid.
The main approach of the algorithm is as follows:
Begin with the collection of all possible permutations. Introduce the subjects one at a
time, reducing the tree to those permutations that also satisfy the constraints of the next
subset. Below is an implementation of the algorithm:
Procedure CONSECUTIVE(X '):
begin
T universal tree
for each I 2 ' do
T REDUCE(T,I)
return T
end
The routine REDUCE attempts to apply on the existing tree, one of 11 templates,
according to the tree and the set I , and replaces a part of the tree accordingly. The templates
are applied from the bottom to the top of the tree. We omit the details here. gure 11.5
gives an example of the execution of the algorithm.
Theorem 11.8 ( Booth - Leuker 1976] ) The collection of permutations P )(') can be
calculated (when implemented with a PQ-tree) in time O(j ' j + j X j + I ' j I j).
2
Corollary 11.9 A (0 1)-valued matrix Am n with f 1's can be checked for the consecutive

1's property in O(m + n + f ) steps.

Corollary 11.10 Interval graphs can be recognized in linear time.
Denition A (0,1)-valued matrix on satises the circular 1's property in the columns if its
rows can be rearranged such that on each column, the 1's will appear cyclicly consecutive.
Remarks:
A (0,1)-valued matrix satises the circular 1's property i it satises the circular 0's
property.
720 permutations
A B C D E F
16 permutations
I={A,D,F}
E
D
144 permutations
A F B C
B C E
A D F I={B,E}
I={B,C,D}
D C B E
A F 4 permutations
Figure 11.5: Example of execution of CONSECUTIVE
11.3. MORE ABOUT THE CONSECUTIVE 1'S PROPERTY 137
If a matrix satises the circular 1's property for the columns, then after switching 0's
with 1's in a column, the property still holds.
Theorem 11.11 (Tucker 1971) ] Given a (0,1)-valued matrix M , switch 0's with 1's in
every column that has 1 in the k -th row. Denote the new matrix M . 0
M satises the circular 1's property i M satises the consecutive 1's property.
0
Proof:
) If M satises the circular 1's property then M satises the circular 1's property. Move
0
the k-th row to the rst row by rotating the rows cyclicly. In this row there will be 0's
only in M . Hence, M satises the consecutive 1's property.
0 0
( If M satises the consecutive 1's property then it satises the circular 1's property,
0
and therefore M satises the circular 1's property.
Theorem 11.12 ( Booth 1975] ) A (0,1)-valued matrix Am n, with f 1's can be checked

for the circular 1's property in time O(m + n + f ), assuming that the matrix is given as a
list of the 1's for each row.
Proof: Let k be the row with the smallest number of 1's, and let M be the matrix
0
obtained from M by switching 0's with 1's in every column that has 1 in the k-th row.
M has at most 2f 1's, because in every row the number of 1's is at most doubled (the
0
number of 0's that become 1's is not larger than the original number of 1's in the row).
Then apply the algorithm for consecutive 1's on M . According to the previous theorem, M
0 0
satises the consecutive 1's property i M satises the circular 1's property.
11.3 More about the consecutive 1's property

Theorem 11.13 ( Booth 1975] ) Given a (0,1)-valued matrix Mr c and an integer k,
deciding whether or not there exists an r
k submatrix of M satisfying the consecutive 1's

property (or circular 1's property) is NP -complete.

Proof: It is easy to show that the problem is NP . To prove that the problem is NP
complete we can show a reduction from the Hamiltonian-path problem: Given an undirected
graph G = (V E ), we produce the adjacency-matrix A of size j V j
j E j (The rows
represent vertices and the columns represent edges). Then it is easy to see that A has a
submatrix of size j V j
j V j that satises the consecutive 1's property , There exists a
Hamiltonian path in G. For circular 1's property, reduce from Hamiltonian cycle instead.
Theorem 11.14 ( Kou 1977] ) Given a (0,1)-valued matrix, nding a permutation of the
rows that minimizes the total number of consecutive blocks of 1's appearing in the columns
is NP -complete.
Theorem 11.15 ( Kou 1977] ) Given a (0,1)-valued matrix, nding a permutation of
the rows that minimizes the number of times a row must be split into two pieces to obtain
consecutive 1's in the columns is NP -complete.
Proof: First, we will introduce an equivalent problem: Given a nite alphabet ; and a
collection of subsets s1 : : : sk such that si ;, we want to nd a chain
2 ; of length B ,
such that for each si there exists a consecutive interval of length j si j in G which contains
all letters in si.
It is obvious that the problem is NP . to show that it is NP -complete we will show
a reduction from the Hamiltonian-path problem: S given a graph G = (V E ) we will dene
; = V + E , B = 2 j E j +1, si = f< vi > x Adj(vi)(vi x)g 8i = 1 : : : n.
2
) If there exists a Hamiltonian path, without lost of generality v1 : : : vn, then we can
produce the chain:
v1 (|v1 ) (v1 ){z : : : (v1 v2)} v2 (|v2 ) : :{z: (v2 v3)} v3 : : : vn (|vn ) :{z: : (vn )} :
all edges of v1 the other edges of v2 the other edges of vn
The length of the chain: j V j +2 j E j ;(j V j ;1) = B .
( If there exists a chain with length B , since for every two sets si sj , si 6 sj , j si \ sj j
1, it follows that the length of the chain is at least:
n
X
j si j ;(n ; 1) = n + 2 j E j ;(n ; 1) = B
i=1
Therefore: if there exists a chain of length B, it matches a Hamiltonian cycle (there is
an edge between each pair of adjacent si sj ).
Remarks
The problem is NP -complete even if 8i, j si j= 4 (the same reduction from a 3-regular
graph).
the problem is polynomial if 8i, j si j= 2. (exercise)
Finding a polynomial approximation is an open problem.
Theorem 11.16 (Goldberg, Golumbic, Kaplan, Shamir 1993) ] Given a
(0 1)-valued matrix, deciding there exists a permutation of the rows that gives at most k
blocks of 1's in every column is NP -complete, for every k 2.
Chapter 12
Lecture 12
Scribe : Michael Seltser
12.1 Temporal Reasoning and Interval algebras

In this section we deal with problems in temporal reasoning. Let us start with an example.
Example 12.1 A paleonthologist wants to determine the relative span of existence of the
species Brontosaurus, Parasaurulophus, Stegosaurus, and Tyrannosaurus. His assumptions:
B and P did not exist simultaneously
P perished before S emerged
S emerged before T and perished not after it
Either P perished before T emerged or P emerged when T perished
B emerged and perished during T 's existence
Is there a scenario satisfying all his assumptions ?
See gure 12.1 for a possible scenario consistent with the assumptions.
Allen in 1983 has dened a model for temporal logic whose elements are time intervals
(corresponding to events that happen during a certain period of time) and used the following
13 primitive relations:
139
P S T
B
Figure 12.1: A consistent scenario
f m m 1 o o 1 s s 1 f f 1 d d 1 g
; ; ; ; ;
to express the relative position of two intervals (See Figure 12.2). We will call this the
13-valued interval algebra A13 .
RELATION NOTATION INTERPRETATION

x before y
y after x
x meets y m
y met-by x m 1;
x overlaps y o
y overlapped-by x o 1
;
x starts y s
y started-by x s 1
;
x during y d
y includes x d 1;
x nishes y f
y nished-by x f 1;
x equals y
Figure 12.2: The 13-valued interval algebra A13. Single line: x interval. Double line: y interval.
From these we dene the 3-valued interval algebra A3 whose elements are called its atomic
relations
A3 : f \g
Here X \Y denotes that events ( intervals ) x and y have a non-empty intersection, i.e.,
any one of the eleven relations excluding and in Allen's algebra may hold between x
and y. The choice of which algebra to use depends on the nature of the application.
12.2. INTERVAL SATISFIABILITY PROBLEMS 141
12.2 Interval satisability problems
We now dene our temporal reasoning problems in the context of constraint satisability.
For each pair of events x and y, let D(x y) be a set of atomic relations in the algebra Ai.
The semantics here is that we do not know precisely the relationship between x and y, but
it must be one of those in the set D(x y). (In the language of constraint satisability, there
is a variable v(x y) representing the relation between x and y, and its value must be taken
from those in the set D(x y) corresponding to its domain.)
For example, we read D(x y) = f og as x is either before or overlaps y. We also call a
set of atomic relations a relation set. Occasionally we shall use the notation xD(x y)y for
short. For example, D(x y) = f d 1 g and xf d 1 gy both mean \either x ends
; ;
before y starts, or y is during x, or the two are equal". We omit braces and commas when
there is no ambiguity, e.g., xy or x\y.
In particular, relation sets in A3 will be presented by concatenation of the atomic rela-
tions. Hence, xy denotes x and y are disjoint, x\y denotes xf \gy, etc.
An instance (or input) for all the problems studied in this paper consists of n events,
and a relation set D(x y) for every pair of distinct events x y. Without loss of generality,
we assume that for each pair of events x and y, the relation sets D(x y) and D(y x) given
as input are consistent, i.e., for each atomic relation R, R 2 D(x y) , R 1 2 D(y x):
;
Otherwise, it is a simple matter to restrict the relation sets further so that they satisfy these
properties or are shown to be unsatisable. Hence, we can assume that the input contains
for each pair of events x y only one relation set, say D(x y), and that the other relation set
D(y x) is given implicitly by the above rule.
Remark 12.2 The size of an instance is the number s of non-trivial relation sets in it, and
is linearly equivalent to the binary input length for each Ai . We will assume that relation
sets which are trivial (i.e., contain all the atomic relations) are not listed explicitly as part
of a problem input, and that each event is involved in at least one non-trivial relation set.
Hence, the input size s of a problem with n events satises s = O(n2 ) and s = *(n).
An interval realization for the instance fD(x y)g is an assignment of intervals on the real
line to events, so that for each pair of events, one of the atomic relations in their relation
set holds. Since the algebras discussed here are concerned with topological properties only,
a realization can be viewed also as a complete weak order of the interval endpoints, thereby
identifying all realizations which dier in metric only. Two realizations are distinct if the
orders of the endpoints in them dier. The input data fD(x y)g is consistent (or satisable)
if it admits at least one realization.
The interval satisability problem (ISAT) is determining the existence of (and nding)
one instantiation that is consistent with the input data fD(x y)g. Occasionally, we shall
make an additional distinction between the decision version of ISAT, whose output is only
a yes/no answer, and the constructive version, whose output is the answer 'no' if the input
is inconsistent and a solution if it is consistent.
The minimal labeling problem (MLP) is to determine the minimal sets D (x y) D(x y)
0
such that the set of realizations is unchanged, and every remaining atomic relation partici-
pates in some solution.
Example 12.3 xf m ogy, yf gz, zff sgx.
Here xoy yz zsx, and xy yz zfx are both consistent with the input, as shown in
Figure 12.3. On the other hand, y z and z y are impossible. The minimal labeling for
this problem is xf m ogy , yz, zff sgx.
x x
z y z y
I: xoy yz zsx II: xy yz zfx
Figure 12.3: Two interval realizations for Example 12.2
The all-consistent-realizations problem (ACRP) is nding of all the consistent realizations.
Here there may be exponentially many solutions, but we would like to generate them in time
polynomial per each new realization.
A fragment of an algebra (A3 or A13 ) is a collection of relation sets in that algebra. A
problem is restricted to a fragment F if all relation sets participating (implicitly or explicitly)
in the instance belong to F . We will denote the problem A restricted to fragment F by A(F ).
Example 12.4 In the problem ISAT ( ), for every pair of events x, y either xy or
xy. Clearly the input is consistent i the relation is a linear order, as no intersections
are allowed. This can be checked in linear time.
Example 12.5 ISAT ( ) is polynomial by a similar argument. Here the instance is
consistent i is a partial order.
Example 12.6 ISAT ( \) is equivalent to interval graph recognition: Form a graph G
with a vertex for each event with an edge between x and y i x\y . The instance has a
realization i G is an interval graph. Hence the problem is solvable in linear time by the
algorithm of Booth and Leuker (1976).
Theorem 12.7 (Golumbic - Shamir, 93) For each of A13 A3, MLP and ACRP are poly-
nomially equivalent to ISAT.
12.3. INTERVAL SANDWICH PROBLEMS AND ISAT 143
J.Allen dened a constraint propagation polynomial heuristic for solving ISAT on A13.
If the algorithm returns a negative answer then it is correct, but a positive answer may be
incorrect. He conjectured that ISAT is NP-complete on A13. This was proved by Vilain and
Kautz (1986). Vilain, Kautz and van Beek (1989) have shown that Allen's algorithm always
gives the correct answer if the input data is restricted to a fragment of A13. Nokel (1988)
has given another polynomial algorithm on a dierent fragment of A13.
Applications of interval satisability problems are numerous. They include planning,
molecular biology (DNA), knowledge representation, natural language processing, informa-
tion systems (databases), articial intelligence, behavioral psychology, archaeology (seri-
ation), graph theory and others.
12.3 Interval sandwich problems and ISAT

Let E 1 and E 2 be two disjoint sets of edges dened on the same vertex-set V . A graph
G(V E ) with E 1 E E 1 E 2 is called a sandwich graph for (E 1 E 2). The interval graph
sandwich (IGS) problem is the following:
Interval Graph Sandwich problem:
INPUT: Two graphs G1 = (V E 1) and G2 = (V E 2) with the same set of vertices,
and disjoint edge-sets E 1 and E 2.
QUESTION: Is there a sandwich graph for (E 1 E 2) which is an interval graph?
Dene F = fE 1 E 2g (i. e., F is the set of non-edges in the graph (V E 1 E 2).) When
E 1 = or F = , the answer is trivially yes. When E 2 = , the problem is equivalent to
recognizing if G(V E 1) is interval and thus is polynomial by the algorithm of Booth and
Lueker. We shall show that in the general case, the problem is NP-complete.
Given a set of intervals on the real line, one can dene a partial order on the intervals by
a b if and only if the interval a is completely to the left of interval b. We use this interval
order to give a reduction from the following problem:
BETWEENNESS:
INPUT: A set of elements S = fa1 : : : ang, and a set T = fT1 : : : Tmg of ordered
triplets of elements from S , where Ti = (ai1 ai2 ai3 ) i = 1 : : : m .
QUESTION: Does there exist a one-to-one function f : S ! f1 2 : : : ng such that
either f (ai1 ) < f (ai2 ) < f (ai3 ) or f (ai1 ) > f (ai2 ) > f (ai3 ) for i = 1 : : : m?
The problem was shown to be NP-complete by Opatrny (1980).
Theorem 12.8 The interval graph sandwich problem is NP-complete.
Proof: Membership in NP follows since a given interval realization can be checked to t
the data in polynomial time. Given an instance of BETWEENNESS, we transform it to an
interval sandwich instance as follows: Dene a vertex vi for each element ai, and two vertices
x1j and x2j for triplet Tj . The vertex set is V = fv1 : : : vng fx1j x2j j j = 1 : : : mg. The
edges sets are
E 1 = f(vi1 x1i ) (x1i vi2 )
(vi2 x2i ) (x2i vi3 ) j i = 1 : : : mg
F = f(vivj ) j i 6= j g
f(vi1 x2i ) (vi3 x1i ) (x1i x2i ) j i = 1 : : : mg
In other words, to each triplet corresponds a 5-chain (see gure 12.4(a)) which should
also appear as a 5-chain in the sandwich graph. In addition, all the vi-s should form an
independent set in that graph. The reduction is clearly polynomial.
The key to the reduction is the following simple observation: in any interval realization
of a 5-chain (a x b y c), either a b c or c b a (see gure 12.4(b)).
(b) vi1 vi2 vi3
vvi1 xv1i vvi2 xv2i vvi3
(a) x1i x2i
Figure 12.4: (a) The 5-chain (vi1 x1i vi2 x2i vi3 ). (b) An interval realization of the 5-chain.
Suppose there is an ordering f of the set S which satises the betweenness conditions.
Create an interval realization for an interval sandwich as follows: Draw n disjoint intervals
corresponding to the vi-s in the same order as the ai-s appear in f . That is, I (vi) I (vj )
if and only if f (ai) < f (aj ). Since the order conditions on triplets are satised, additional
intervals corresponding to x1i and x2i can be added for each triplet Ti such that the subgraph
induced by the ve vertices fvi1 vi2 vi3 x1i x2i g is a 5-chain. The resulting interval graph on
V is a sandwich graph.
For the converse, suppose there exists an interval sandwich solution with realization I .
Since in that graph the set fv1 : : : vng induces an independent set, a complete ordering
on S is generated: Number the elements in the order of their intervals from left to right,
namely, f (ai) < f (aj ) if and only if I (vi) I (vj ). Since for each triplet Ti the ve vertices
fvi1 vi2 vi3 x1i x2i g must induce a 5-chain in every sandwich graph, by the above observation,
every triplet satises the betweenness condition.
Interestingly, we can also show that a restricted version of the above problem is NP-
complete: A function c : V ! Z is called a coloring of the graph G = (V E ) if for every
(u v) 2 E , c(u) 6= c(v). The problem is dened as follows:
12.4. A LINEAR TIME ALGORITHM FOR ISAT(1) 145
COLORED INTERVAL SANDWICH:
INPUT: A graph G = (V E ) and a coloring c of G.
QUESTION: Does there exist a set E so that E E , G = (V E ) is an interval
0 0 0 0
graph and c is also a coloring for G ?0
Notice that this is a special case of the sandwich problem where E 1 = E and F =
f(u v) j c(u) = c(v)g is given implicitly by the coloring.
Theorem 12.9 The colored interval sandwich problem is NP-complete.
Proof: Reduction from BETWEENNESS, where the vertex set and E = E 1 is dened
as in Theorem 12.8 and the coloring is c(x1i ) = c(x2i ) = i (i = 1 : : : m) and c(v1) = c(v2) =
= c(vn) = m + 1. A slight modication of the arguments in Theorem 12.8 proves the
result.
Remark 12.10 Golumbic, Kaplan, and Shamir (1993) have shown that the analog problems
to Theorem 12.8 and 12.9, with the additional condition that all intervals have the same
length, are still NP-complete.
Corollary 12.11 ISAT( \ \ ) is NP-complete.
Proof: Just consider V = f events g, (i j ) 2 E 1 , i\j , (i j ) 2 F , ij ,
(i j ) 2 E 2 , i\j .
Corollary 12.12 ISAT( A3 ) is NP-complete.
Proof: Follows from Corollary 12.11 since ISAT( \ \) is a restriction of
ISAT( A3 ).
Corollary 12.13 ISAT, MLP, ACRP are NP-hard for A3 and A13.
Theorem 12.14 ISAT( \ ) is polynomial.
Proof: exercise.
12.4 A linear time algorithm for ISAT( 1)

In this section we deal with ISAT problems on the fragment:
1 = f \ \ \ \g:
That is, A3 ; . We shall give ecient algorithms for ISAT, MLP and ARP on this
fragment, and they will apply immediately to any subfragment of 1. Hence, by excluding
just a single relation set from A3 the problems become tractable.
The polynomial algorithm for ISAT(1) utilizes a directed graph, which will be described
later, with vertices corresponding to the endpoints of event intervals and arcs representing
the relative order of endpoints. The key observation here is that every relation in 1 is
equivalent to a certain order requirement between a pair (or two pairs) of such endpoints.
This observation is proved below:
Lemma 12.15 Let i and j be the event intervals li ri] and lj rj ], respectively. In each of
the following cases, the intervals satisfy the set of relations if and only if their endpoints
satisfy the corresponding inequalities:
ij , ri < lj ij , rj < li (12.1)
i\j , li rj i\j , lj ri (12.2)
i\j , li rj and lj ri (12.3)
If i\j , no constraint is imposed.
Proof: (1) is immediate. The conditions in (2) are the negation of those in (1). (3) is
equivalent to the intersection of the two conditions in (2).
The graph is now constructed as follows: For an instance J of ISAT(1) with n events,
form a directed graph G(J ) = G(V E ) with vertex set V = fr1 : : :rn l1 : : :lng. The arc set
E consists of two disjoint subsets, E0 and E1.
The former will represent weak inequality and the latter will represent strict inequality
between pairs of endpoints.
The arcs are dened as follows:
(li ri) 2 E0 i = 1 : : : n (12.4)
(ri lj ) 2 E1 8i j s:t: ij (12.5)
(li rj ) 2 E0 8i j s:t: i\j (12.6)
(li rj ) 2 E0 and (lj ri) 2 E0 8i j s:t: i\j (12.7)
For pairs i j with the relation i\j , no arc is introduced. Dene now E = E0 E1.
We call the arcs in E0 (resp., E1) the weak arcs (resp., strict arcs). Note that the graph G is
bipartite. Denote the two parts of the vertex set by R = fr1 : : : rng and L = fl1 : : : lng,
and call an arc an RL-arcs (resp., LR-arc) if it is directed from R to L (resp., from L to R).
Remark 12.16 In the graph G all the RL-arcs are strict and all the LR-arcs are weak.
Hence, we need not record explicitly the type of each arc since it is implied by its direction.
Lemma 12.17 Every cycle in G contains a strict arc.
12.4. A LINEAR TIME ALGORITHM FOR ISAT(1) 147
Algorithm: 1-CONSISTENCY
begin
1. construct G(J ) according to rules (12.4) - (12.7).
2. if G(J ) contains a cycle - J is not satisable. Otherwise, it is.
end construct G-decomposition
Figure 12.5: Algorithm to construct Delta1-CONSISTENCY
Proof: Since G is bipartite, a cycle must contain vertices both from R and from L. In
particular, it must contain an RL-arc, so the claim follows by Remark 12.16.
An algorithm for solving ISAT(1) is described in gure 12.5
Lemma 12.18 Suppose G(J ) = (V E ) is acyclic. Then a linear order on V is consistent
with the partial order G(J ) if and only if it is a realization of J .
Proof: Take any linear order P which extends the partial order G. P is an ordering of
all the endpoints, which by Lemma 12.15 satises all the relations in the input, so P gives
a realization for J . On the other hand, every realization of J gives a linear order of the
endpoints, in which each of the input relations must be satised, by Lemma 12.15. Hence,
the linear order must be consistent with the partial order G(J ).
Theorem 12.19 Algorithm 1-CONSISTENCY correctly recognizes if an instance of
ISAT(1) is satisable in linear time.
Proof: Validity: By Lemma 12.15, each arc reects the order relation of a pair of interval
endpoints as prescribed by the input relations.
If G contains a cycle, then by Lemma 12.17 that cycle contains a strict arc. Hence,
that cycle must satisfy ri < lj : : : ri, which implies that the input relations cannot be
satised. In case G is acyclic,
Lemma 12.18 implies that J is satisable.
Complexity: Constructing G(J ) requires O(s) steps, where s is the number of input
relation sets, since the eort is constant per relation. Checking if G is acyclic can be done,
for example, by depth rst search, in time linear in jE j, the number of arcs.
Since jE j 2s + n, using remark 12.2 we conclude that the algorithm is linear.
In a similar way one can construct algorithms for MLP(1) and ACRP(1) in the fol-
lowing manner: An algorithm for MLP is obtained by calculation of transitive closure of G.
The algorithm for ACRP(1) is based on nding of all linear extensions of G.
Remark 12.20 ISAT is polynomial if all events are points ( A3 = A13 in that case ).
Proof: Exercise.
Remark 12.21 One can handle additional metric constraints in A3 ; , by solving the
linear programming problem dened by (12.1) - (12.3) with additional metric constraints on
the endpoints.
\ \
; T T Poly
interval graph
Poly Poly Poly Poly
linear order interval order acyclic graph
\ T T NPC* NPC*
\ T T T NPC
interval sandwich
\ Poly Poly NPC* NPC*
\ Poly Poly Poly NPC
\ \ T T NPC NPC
\ \ Poly Poly NPC NPC
Figure 12.6: Complexity of ISAT on all fragments of A3. T : satised by setting all intervals
disjoint or identical. The four asterisk cases were shown to be NPC by Webber (1994).
Figure 12.6 summarizes results from Golumbic - Shamir (1993) and Webber (1994) on
the complexity of ISAT on all fragments of A3.
12.5 Bandwidth, Pathwidth and Proper Pathwidth

We now switch to dierent problems related to interval and proper interval graphs. The
motivation studying to these problems arises in physical mapping of DNA, a major problem
in the Human Genome Project.
Lemma 12.22 Every interval graph (and every proper interval graph) has an interval (resp.,
proper interval) representation on the real line in which all endpoints are distinct, and the
left endpoints are assigned to the integers 1 2 : : : jV j.
We shall call such a representation canonical.
Denition Pathwidth and Proper Pathwidth A path decomposition of a given graph
G = (V E ), is a sequence of subsets of V , X = (X1 : : : Xl ) such that
12.5. BANDWIDTH, PATHWIDTH AND PROPER PATHWIDTH 149
(1) V = i Xi
(2) For each edge (u v) 2 E , there exists some i 2 f1 : : : lg so that both u and v belong
to Xi .
(3) For each v 2 V there exist some s(v) e(v) 2 f1 : : : lg so that s(v) e(v), and v 2 Xj
if and only if j 2 fs(v) s(v) + 1 : : : e(v)g.
The width of X is dened by pwX (G) = maxfjXij j i = 1 : : : lg ; 1. The pathwidth
of G, denoted pw(G), is the minimum value of pwX (G) over all path decompositions, i.e.,
pw(G) = minfpwX (G) j X is a path decomposition of Gg. The notion of pathwidth was
originally introduced by Robertson and Seymour (1983).
Lemma 12.23 (e.g., Mohring, 91) For any path decomposition X = (X1 : : : Xl ) of a
graph G, every clique in G must be contained in some Xi .
The PATHWIDTH problem is to decide for a given graph G and a given integer k if
pw(G) k. Equivalent problems arise in various areas, including VLSI layout, processor
management, node searching and vertex separation.
Lemma 12.24 For every graph G, the pathwidth of G is one less than the least clique size
of any interval supergraph of G.
We introduce here the notion of a proper path decomposition of G It is as a path decom-
position X which satises (1)-(3) and also:
(4) For every u v 2 V , fs(u) s(u) + 1 : : : e(u)g 6 fs(v) s(v) + 1 : : : e(v)g.
(As usual, denotes strict containment.) The proper pathwidth of G, denoted ppw(G), is
the minimum value of pwX (G) over all proper path decompositions of G, namely ppw(G) =
minfpwX (G) j X is a proper path decomposition of Gg.
Clearly, pw(G) ppw(G), but note that pw(G) and ppw(G) can dier dramatically: For
the star graph with 2n + 1 vertices, G = K1 2n, pw(G) = 1 but ppw(G) = n. An easy way
to see the last equality is using the characterizations we shall provide in the next section.
Bandwidth: Another measure of graph width which we shall use is bandwidth: For G =
(V E ) with jV j = n, a linear layout of the graph is a 1-1 function L : V ! f1 : : : ng. The
bandwidth of L is bwL(G) = max(u v) E fjL(u) ; L(v )jg. The bandwidth of G is the minimum
bandwidth over all layouts of G, namely, bw(G) = minfbwL(G) j L is a layout of Gg. The
2
BANDWIDTH problem is to decide for a given graph G and integer k, if bw(G) k . This
problem has been studied intensely because of its application to sparse matrix algebra. It
is known to be NP-complete even for binary trees ( Garey, Graham, Johnson, and Knuth,
1978 ) and for caterpillars with hair length at most three ( Monien, 1986 ). On the other
hand, it is solvable in O(nk ) for arbitrary k and in linear time for k = 2 ( Garey, Graham,
Johnson, and Knuth, 1978 ).
Lemma 12.25 ( Kaplan - Shamir, 94 ) For every graph G, the proper pathwidth of G is
one less than the least clique size of any proper interval supergraph of G.
Proof: Suppose X = (X1 : : : Xl) is a proper path decomposition of G and pwX (G) =
ppw(G) = k. On the linear order 1 < 2 < < l, let I (v) be the interval s(v) e(v)]. Then
fI (v)gv V is a representation of an interval graph G . G is proper since no interval is strictly
2
0 0
contained in the other, by property (4) in the denition of a proper path decomposition. G is 0
a supergraph of G, by property (2). Since X is also a path decomposition of G , Lemma 12.23

0
implies that every clique in G must be contained in some Xi . Hence, !(G ) k + 1.

0 0
Conversely, suppose k + 1 is the least clique size of any proper interval supergraph of
G. Let fI (v)gv V be a proper interval representation of such supergraph G = (V E ) with
2
0 0
!(G ) = k + 1, in which all endpoints are distinct, and I (v) = l(v) r(v)]. Order the 2n
0
endpoints from left to right p1 : : : p2n. Dene Xi = fv 2 V j pi 2 I (v)g for i = 1 : : : 2n.
(Here and throughout n = jV j). We claim that (X1 : : : X2n ) is a proper path decomposition
of G. Checking the four requirements from a proper path decomposition one can observe
that (1) is immediate. (2) follows since E E , and if (u v) 2 E and l(u) < l(v) then
0 0
l(v) 2 I (v) \ I (u). In other words, every edge in E is represented in some Xi. Since G
0 0
is a proper interval graph, it follows that s(v) = l(v) and e(v) = r(v) satisfy (3) and (4).
Since the clique size of G is k + 1, each point pi can meet at most k + 1 intervals in the
0
representation. Hence, maxi jXi j k + 1 so ppw(G) k.

Theorem 12.26 ( Kaplan - Shamir, 94 ) The bandwidth of a graph equals its proper
pathwidth.
Proof: For the graph G = (V E ) with ppw(G) = k , by Lemma 12.25 there exists a
supergraph G = (V E ) of G which is proper interval with clique size k + 1. Let fI (v)gv V
0 0
2
be a canonical representation for G , where I (v) = l(v) r(v)]. Dene a layout L on G by

0
L(u) = l(u) for all u 2 V . Suppose L(u) ; L(v) > k for some (u v) 2 E , where L(v) = i.
Then fL 1 (i) : : : L 1(i + k + 1)g form a clique of size k + 2 in G , a contradiction. Hence,
; ; 0
bw(G) bwL(G) ppw(G).

Conversely, let L be a layout of G so that bwL(G) = bw(G) = k. Form a set of equal-
length intervals on the real line fI (v)gv V where I (v) = L(v) L(v) + k]. Let G = (V E )0 0
be the intersection graph of that set of intervals. Clearly E E since if (u v) 2 E then
2
0
jL(u) ; L(v)j k, so I (u) \ I (v) 6= . Since the maximum clique size for G is at most k + 1,
0
using Lemma 12.25, ppw(G) k. Hence, bw(G) ppw(G).

This interesting - and somewhat surprising - equivalence between bandwidth and proper
pathwidth allows us to draw several immediate conclusions from the results on bandwidth:
Corollary 12.27
A. Finding the proper pathwidth is NP-hard, even for binary trees and for caterpillars with
12.5. BANDWIDTH, PATHWIDTH AND PROPER PATHWIDTH 151
hair length at most three.
B. Problem A can be solved in O(f (k)nk 1 ) time, and in particular is polynomial when k is
;
xed.
C. Deciding whether the proper pathwidth of a graph is not greater than two can be done in
linear time.
These results follow immediately from known results on bandwidth, using Theorem 12.26.
We nish by mentioning two parametric results. The rst deals with the unit interval
sandwich problem mentioned above:
Theorem 12.28 ( Kaplan-Shamir 94 ) Deciding if a sandwich instance admits a proper
interval sandwich graph with clique size at most k can be done in O(Ck nk ) time, and in
particular in polynomial time when k is xed. Furthermore, it cannot be solved in O(Ck n )
where is independent of k , unless deciding if a given graph has an independent set of size
k can be solved in O(Ck n ).
0
0
In contrast to that negative result, the completion problem for proper interval graphs is
tractable, in the parametric sense:
Theorem 12.29 ( Kaplan, Shamir, Tarjan 94 ) Given a graph G = (V E ), one can
decide if there exists a proper interval graph G = (V E ) with E E and jE n E j k in
0 0 0 0
time O(Ck (jV j + jE j)).

Index
G-Decomposition, 84 complete, 12
H = G h, 27 Complete bipartite graph, 15
:, 59, 60, 62 Composition, 77
compositions, 120
a-b separator, 38 Connected graph, 15
adjacency set, 11 consecutive 1's property, 130
adjacent, 11 Cycle, 15
;Perfect property, 23
antichain, 110 Decomposition, 79
asteroidal triplet, 124 dimension, 113
duality, 15
Bipartite graph, 15
edge, 11
c-colorable, 14 edge graph, 23
c-coloring, 14 endpoints, 11
Chain, 15
chain, 110 F-diagram, 96
chordal, 37 Hasse diagram, 109, 110
chordless, 15 helly, 51
Chordless cycle, 15 Helly property, 19
chromatic number, 14 hereditary property, 20
circle graphs, 17
circular 1's property, 135 independent set, 14
clique, 12 integral, 36
clique cover, 14 intersection graph, 15, 93
clique cover number, 14 Interval containment graph, 99
clique matrix, 36 interval graph, 15, 121
clique number, 14 interval order, 128
closed neighborhood, 11 interval representation, 19
co-graph, 120 isomorphic, 11
comparability graph, 21
comparability invariants, 119 join, 26
complement, 12 jump number, 119
152
INDEX 153
layout, 130 Strongly connected graph, 15
line graph, 23 subgraph, 12
linear extension, 113 subgraph induced by A, 12
symmetric closure, 12
maximal, 14
maximum, 14 Tolerance graph, 99
minimal vertex separator, 38 triangle, 15
minimum realizer, 113 triangulated, 37
Module, 81 triangulated graphs, 20
monotone transitive, 37
undirected, 12
odd anti-hole, 36 union, 26
odd hole, 36 unit interval graph, 15
open neighborhood, 11 UPO, 82
opposite permutation, 95 utility function, 127
Orientation, 12
oriented graph, 12 Vertex Multiplication, 27
Extended, 27
p-critical, 35 vertex separator, 38
Path, 15 vertices, 11
Pathwidth and Proper Pathwidth, 148
perfect elimination, 37 X-free, 37
perfect elimination order, 38
permutation diagram, 17, 94
permutation graph, 17
proper circular arc graph, 17
proper interval graph, 128
Quasi-Transitive Orientation, 89
r-clique, 14
realizer, 113
reversal, 12
rigid-circuit, 37
series-parallel, 120
simple, 15
Simple cycle, 15
simplicial, 37
stability number, 14
stable set, 14
strong perfect graph conjecture, 36

Advanced Topics in Algorithms Shamir

Uploaded by

Advanced Topics in Algorithms Shamir

Uploaded by

Advanced Topics in Graph Algorithms

Based on Notes from the Course

1.1 Introduction to Graph Theory

is a bijection f : V ! V satisfying for all x y 2 V ,

(x y) 2 E $ (f (x) f (y)) 2 E

Figure 1.1: Three pictures of the same graph.

said to be the reversal of G, where E 1 = f(x y) j (y x) 2 E g, that is,

We dene symmetric closure of G to be the graph G^ = (V E^ ), where E^ = E  E 1. ;

A graph G = (V E ) is called undirected if its adjacency relation is symmetric i.e., if

A graph is called an oriented graph if E \ E 1 = .

Orientation of an undirected graph is a choice of a direction on every edge in it.

Given a subset A V of the vertices, we dene the subgraph induced by A to be GA =

(a) (b) (c)

1.1.2 Intersection Graphs

(a) (c) (d)

1.1.3 Circular-arc Graphs

b c G , the circular-arc graph.

1.1.4 Interval Graphs

or strictly decreasing sequence. Therefore, it is impossible for I0 and Il 1 to intersect,

contradicting the criterion that V0 Vl 1 is an edge of G.

The A graph The suspension bridge graph The house graph

Figure 1.12: Transitive orientations of Three comparability graphs.

An undirected graph which is transitively orientable is sometimes called a comparability

C The bull’s head graph

1.1.5 Line graphs of bipartite graphs

Figure 1.14: (a) A graph G.(b) The line graph of G.

2.1 Perfect Graphs

2.1.1 Denition of perfect graph

Proof: Denote = (G). Let U be a maximum independent set of G. By the assumption

It follows that GV D contains a maximum independent set of size , therefore G contains

2.1.2 Some Denitions and Properties

of x: G  x = (V  fx g E = E  fx uju 2 Adj (x)g).

set: V + x ; y. In both graphs, uv is an edge i one of the following exist:

Denition More generally, if x1 : : : xn are the vertices of G and h = (h1 : : : hn ) is a

2.1.3 Perfect Graph Theorem

in !(G  x) colors. Hence, G  x satises (1).

Thus, (G  x) = k(G  x).

with the lowest possible number of cliques, because always k , so

Let H be the subgraph of H induced by X = K1  : : :  Kq h+1  fu1g. Thus

jX j = p(q ; h + 1) + 1 h<2 pq + 1 equation

equation 2:7 equation 2:6

Let S be a stable set of H of cardinality q ; h + 2. Certainly u1 2 S , for otherwise S

would contain two vertices of a clique (by the denition of H ). Therefore, S = S  U = 0 0

S + (U ; fu g) is a stable set of H with (q ; h + 2) + (h ; 1) = q + 1 vertices, contradicting

Construct a (0,1)-matrix M with rows corresponding to vertices, and a column i corre-

matrix is illustrated in gure 2.1.

In the column of T In each column at most

K(S 1) K(S 2) K(S |I| )

Since every clique in G  x contains at most one "copy" of vertex x 2 G, we have

T is a stable set and K (S ) is a clique, so

3.1 Perfect Graphs

Figure 3.1: A graph and its clique matrix

3.1.3 The Strong Perfect Graph Conjecture

3.2 Triangulated Graphs

3.2.2 Characterizing Triangulated Graphs

S is an a-b separator, S is called minimal vertex separator.

contain a b respectively. We shall look at GA+S :

less vertices, so by induction it has a perfect elimination order 

1. assign label 0 to each vertex

Figure 3.3: Code for MCS algorithm

3.2.3 Recognizing Triangulated Graphs

4.1 Recognizing Triangulated Graphs

otherwise = x vj : : : vk w x] is a chordless cycle of length 4, contradicting G's

path = w vk : : : vj x] satises 4.1, contradicting the maximality of

It follows that there is no path satisfying 4.1.

Figure 4.2: The vertices x and vj

Figure 4.3: The path 0

Figure 4.4: The path 00

Figure 4.5: A forbidden structure in a chordal graph

there exist u w 2 Adj (v) such that uw 62 E , and

4.1.2 Testing an Elimination Scheme

7. concatenate X n fug to A(u)

Figure 4.6: Code for PEO testing algorithm

Figure 4.7: An input chordal graph and elimination order

Correctness of the Algorithm

is not met, so every x 2 Xv n fug is a neighbor of u.

We dene symmetric closure of G to be the graph G^ = (V E^ ), where E^ = E E 1. ;

Given a subset A V of the vertices, we dene the subgraph induced by A to be GA =

of x: G x = (V fx g E = E fx uju 2 Adj (x)g).

set: V + x ; y. In both graphs, uv is an edge i one of the following exist:

Denition More generally, if x1 : : : xn are the vertices of G and h = (h1 : : : hn ) is a

in !(G x) colors. Hence, G x satises (1).

Thus, (G x) = k(G x).

Let H be the subgraph of H induced by X = K1 : : : Kq h+1 fu1g. Thus

would contain two vertices of a clique (by the denition of H ). Therefore, S = S U = 0 0

matrix is illustrated in gure 2.1.

Since every clique in G x contains at most one "copy" of vertex x 2 G, we have

less vertices, so by induction it has a perfect elimination order

otherwise = x vj : : : vk w x] is a chordless cycle of length 4, contradicting G's

path = w vk : : : vj x] satises 4.1, contradicting the maximality of

by the induction hypothesis there is a tree T = (K E ), whose vertices corre-

set K v = fC 2 K jv 2 C g induces a connected subtree of T .

: K y = K y + fAg n fB g and K y induces the same subtree as K y , because all

we did is renaming B as A (as in gure 4.16).

it to B . For each y 2 Y : K y = K y + fAg and K y induces a connected subtree

Denition G is weakly chordal graph if G and G do not contain an induced Ck , k 5 .

is a partition of Q. Impose a linear order on the vertices in Q such that if u 2 Ri,v 2 Rj

Adj (i) 6 Adj (k) and Adj (i) 6 Adj (k)

where N = P Q, E = E + fuv j u v 2 P g + fuv j u v 2 Qg . (i.e. we place the

independent set on each side of the clique. See gure 5.7).

Lemma 5.9 G is a chain graph i C (G) is triangulated.

Adj ((p)) = AdjG((p)) + P ; (p)

of the linear arrangement is c() = e E (e ). The optimal linear arrangement problem

reduction. First observe that an ordering of N species uniquely a minimal set H () of

easy to see that F H (), and therefore if F is a minimal augmentation then F = H ().

1(v) ; 2 = 1(u) + 1(v) ; 1(u)] ; 2

thus the number of edges of H () incident to e1 and e2 is

wv 2= E , therefore w 2= C , so we get: C Xv fvg.

is concatenated to A(u) if u is the rst neighbor of v after it in

such that fug Xu fwg Xw . Let us take v such that v =

Xv = Xu fug. So Xv fvg is a clique and Xu fug Xv fvg, therefore Xu fug is not

9. S (u) maxfS (u) jX j ; 1g

yi is the rst vertex in

Xyi 1 . All vertices following yt are in Xy1 Xy2 Xyt.

Figure 6.3: An example for nding maximal stable set